1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/page_alloc.c
4  *
5  *  Manages the free list, the system allocates free pages here.
6  *  Note that kmalloc() lives in slab.c
7  *
8  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
9  *  Swap reorganised 29.12.95, Stephen Tweedie
10  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11  *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13  *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14  *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15  *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16  */
17 
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/kmsan.h>
31 #include <linux/module.h>
32 #include <linux/suspend.h>
33 #include <linux/pagevec.h>
34 #include <linux/blkdev.h>
35 #include <linux/slab.h>
36 #include <linux/ratelimit.h>
37 #include <linux/oom.h>
38 #include <linux/topology.h>
39 #include <linux/sysctl.h>
40 #include <linux/cpu.h>
41 #include <linux/cpuset.h>
42 #include <linux/memory_hotplug.h>
43 #include <linux/nodemask.h>
44 #include <linux/vmalloc.h>
45 #include <linux/vmstat.h>
46 #include <linux/mempolicy.h>
47 #include <linux/memremap.h>
48 #include <linux/stop_machine.h>
49 #include <linux/random.h>
50 #include <linux/sort.h>
51 #include <linux/pfn.h>
52 #include <linux/backing-dev.h>
53 #include <linux/fault-inject.h>
54 #include <linux/page-isolation.h>
55 #include <linux/debugobjects.h>
56 #include <linux/kmemleak.h>
57 #include <linux/compaction.h>
58 #include <trace/events/kmem.h>
59 #include <trace/events/oom.h>
60 #include <linux/prefetch.h>
61 #include <linux/mm_inline.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/sched/rt.h>
66 #include <linux/sched/mm.h>
67 #include <linux/page_owner.h>
68 #include <linux/page_table_check.h>
69 #include <linux/kthread.h>
70 #include <linux/memcontrol.h>
71 #include <linux/ftrace.h>
72 #include <linux/lockdep.h>
73 #include <linux/nmi.h>
74 #include <linux/psi.h>
75 #include <linux/padata.h>
76 #include <linux/khugepaged.h>
77 #include <linux/buffer_head.h>
78 #include <linux/delayacct.h>
79 #include <asm/sections.h>
80 #include <asm/tlbflush.h>
81 #include <asm/div64.h>
82 #include "internal.h"
83 #include "shuffle.h"
84 #include "page_reporting.h"
85 #include "swap.h"
86 
87 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
88 typedef int __bitwise fpi_t;
89 
90 /* No special request */
91 #define FPI_NONE		((__force fpi_t)0)
92 
93 /*
94  * Skip free page reporting notification for the (possibly merged) page.
95  * This does not hinder free page reporting from grabbing the page,
96  * reporting it and marking it "reported" -  it only skips notifying
97  * the free page reporting infrastructure about a newly freed page. For
98  * example, used when temporarily pulling a page from a freelist and
99  * putting it back unmodified.
100  */
101 #define FPI_SKIP_REPORT_NOTIFY	((__force fpi_t)BIT(0))
102 
103 /*
104  * Place the (possibly merged) page to the tail of the freelist. Will ignore
105  * page shuffling (relevant code - e.g., memory onlining - is expected to
106  * shuffle the whole zone).
107  *
108  * Note: No code should rely on this flag for correctness - it's purely
109  *       to allow for optimizations when handing back either fresh pages
110  *       (memory onlining) or untouched pages (page isolation, free page
111  *       reporting).
112  */
113 #define FPI_TO_TAIL		((__force fpi_t)BIT(1))
114 
115 /*
116  * Don't poison memory with KASAN (only for the tag-based modes).
117  * During boot, all non-reserved memblock memory is exposed to page_alloc.
118  * Poisoning all that memory lengthens boot time, especially on systems with
119  * large amount of RAM. This flag is used to skip that poisoning.
120  * This is only done for the tag-based KASAN modes, as those are able to
121  * detect memory corruptions with the memory tags assigned by default.
122  * All memory allocated normally after boot gets poisoned as usual.
123  */
124 #define FPI_SKIP_KASAN_POISON	((__force fpi_t)BIT(2))
125 
126 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
127 static DEFINE_MUTEX(pcp_batch_high_lock);
128 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
129 
130 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
131 /*
132  * On SMP, spin_trylock is sufficient protection.
133  * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
134  */
135 #define pcp_trylock_prepare(flags)	do { } while (0)
136 #define pcp_trylock_finish(flag)	do { } while (0)
137 #else
138 
139 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
140 #define pcp_trylock_prepare(flags)	local_irq_save(flags)
141 #define pcp_trylock_finish(flags)	local_irq_restore(flags)
142 #endif
143 
144 /*
145  * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
146  * a migration causing the wrong PCP to be locked and remote memory being
147  * potentially allocated, pin the task to the CPU for the lookup+lock.
148  * preempt_disable is used on !RT because it is faster than migrate_disable.
149  * migrate_disable is used on RT because otherwise RT spinlock usage is
150  * interfered with and a high priority task cannot preempt the allocator.
151  */
152 #ifndef CONFIG_PREEMPT_RT
153 #define pcpu_task_pin()		preempt_disable()
154 #define pcpu_task_unpin()	preempt_enable()
155 #else
156 #define pcpu_task_pin()		migrate_disable()
157 #define pcpu_task_unpin()	migrate_enable()
158 #endif
159 
160 /*
161  * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
162  * Return value should be used with equivalent unlock helper.
163  */
164 #define pcpu_spin_lock(type, member, ptr)				\
165 ({									\
166 	type *_ret;							\
167 	pcpu_task_pin();						\
168 	_ret = this_cpu_ptr(ptr);					\
169 	spin_lock(&_ret->member);					\
170 	_ret;								\
171 })
172 
173 #define pcpu_spin_lock_irqsave(type, member, ptr, flags)		\
174 ({									\
175 	type *_ret;							\
176 	pcpu_task_pin();						\
177 	_ret = this_cpu_ptr(ptr);					\
178 	spin_lock_irqsave(&_ret->member, flags);			\
179 	_ret;								\
180 })
181 
182 #define pcpu_spin_trylock_irqsave(type, member, ptr, flags)		\
183 ({									\
184 	type *_ret;							\
185 	pcpu_task_pin();						\
186 	_ret = this_cpu_ptr(ptr);					\
187 	if (!spin_trylock_irqsave(&_ret->member, flags)) {		\
188 		pcpu_task_unpin();					\
189 		_ret = NULL;						\
190 	}								\
191 	_ret;								\
192 })
193 
194 #define pcpu_spin_unlock(member, ptr)					\
195 ({									\
196 	spin_unlock(&ptr->member);					\
197 	pcpu_task_unpin();						\
198 })
199 
200 #define pcpu_spin_unlock_irqrestore(member, ptr, flags)			\
201 ({									\
202 	spin_unlock_irqrestore(&ptr->member, flags);			\
203 	pcpu_task_unpin();						\
204 })
205 
206 /* struct per_cpu_pages specific helpers. */
207 #define pcp_spin_lock(ptr)						\
208 	pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
209 
210 #define pcp_spin_lock_irqsave(ptr, flags)				\
211 	pcpu_spin_lock_irqsave(struct per_cpu_pages, lock, ptr, flags)
212 
213 #define pcp_spin_trylock_irqsave(ptr, flags)				\
214 	pcpu_spin_trylock_irqsave(struct per_cpu_pages, lock, ptr, flags)
215 
216 #define pcp_spin_unlock(ptr)						\
217 	pcpu_spin_unlock(lock, ptr)
218 
219 #define pcp_spin_unlock_irqrestore(ptr, flags)				\
220 	pcpu_spin_unlock_irqrestore(lock, ptr, flags)
221 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
222 DEFINE_PER_CPU(int, numa_node);
223 EXPORT_PER_CPU_SYMBOL(numa_node);
224 #endif
225 
226 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
227 
228 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
229 /*
230  * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
231  * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
232  * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
233  * defined in <linux/topology.h>.
234  */
235 DEFINE_PER_CPU(int, _numa_mem_);		/* Kernel "local memory" node */
236 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
237 #endif
238 
239 static DEFINE_MUTEX(pcpu_drain_mutex);
240 
241 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
242 volatile unsigned long latent_entropy __latent_entropy;
243 EXPORT_SYMBOL(latent_entropy);
244 #endif
245 
246 /*
247  * Array of node states.
248  */
249 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
250 	[N_POSSIBLE] = NODE_MASK_ALL,
251 	[N_ONLINE] = { { [0] = 1UL } },
252 #ifndef CONFIG_NUMA
253 	[N_NORMAL_MEMORY] = { { [0] = 1UL } },
254 #ifdef CONFIG_HIGHMEM
255 	[N_HIGH_MEMORY] = { { [0] = 1UL } },
256 #endif
257 	[N_MEMORY] = { { [0] = 1UL } },
258 	[N_CPU] = { { [0] = 1UL } },
259 #endif	/* NUMA */
260 };
261 EXPORT_SYMBOL(node_states);
262 
263 atomic_long_t _totalram_pages __read_mostly;
264 EXPORT_SYMBOL(_totalram_pages);
265 unsigned long totalreserve_pages __read_mostly;
266 unsigned long totalcma_pages __read_mostly;
267 
268 int percpu_pagelist_high_fraction;
269 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
270 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
271 EXPORT_SYMBOL(init_on_alloc);
272 
273 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
274 EXPORT_SYMBOL(init_on_free);
275 
276 static bool _init_on_alloc_enabled_early __read_mostly
277 				= IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
early_init_on_alloc(char * buf)278 static int __init early_init_on_alloc(char *buf)
279 {
280 
281 	return kstrtobool(buf, &_init_on_alloc_enabled_early);
282 }
283 early_param("init_on_alloc", early_init_on_alloc);
284 
285 static bool _init_on_free_enabled_early __read_mostly
286 				= IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
early_init_on_free(char * buf)287 static int __init early_init_on_free(char *buf)
288 {
289 	return kstrtobool(buf, &_init_on_free_enabled_early);
290 }
291 early_param("init_on_free", early_init_on_free);
292 
293 /*
294  * A cached value of the page's pageblock's migratetype, used when the page is
295  * put on a pcplist. Used to avoid the pageblock migratetype lookup when
296  * freeing from pcplists in most cases, at the cost of possibly becoming stale.
297  * Also the migratetype set in the page does not necessarily match the pcplist
298  * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
299  * other index - this ensures that it will be put on the correct CMA freelist.
300  */
get_pcppage_migratetype(struct page * page)301 static inline int get_pcppage_migratetype(struct page *page)
302 {
303 	return page->index;
304 }
305 
set_pcppage_migratetype(struct page * page,int migratetype)306 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
307 {
308 	page->index = migratetype;
309 }
310 
311 #ifdef CONFIG_PM_SLEEP
312 /*
313  * The following functions are used by the suspend/hibernate code to temporarily
314  * change gfp_allowed_mask in order to avoid using I/O during memory allocations
315  * while devices are suspended.  To avoid races with the suspend/hibernate code,
316  * they should always be called with system_transition_mutex held
317  * (gfp_allowed_mask also should only be modified with system_transition_mutex
318  * held, unless the suspend/hibernate code is guaranteed not to run in parallel
319  * with that modification).
320  */
321 
322 static gfp_t saved_gfp_mask;
323 
pm_restore_gfp_mask(void)324 void pm_restore_gfp_mask(void)
325 {
326 	WARN_ON(!mutex_is_locked(&system_transition_mutex));
327 	if (saved_gfp_mask) {
328 		gfp_allowed_mask = saved_gfp_mask;
329 		saved_gfp_mask = 0;
330 	}
331 }
332 
pm_restrict_gfp_mask(void)333 void pm_restrict_gfp_mask(void)
334 {
335 	WARN_ON(!mutex_is_locked(&system_transition_mutex));
336 	WARN_ON(saved_gfp_mask);
337 	saved_gfp_mask = gfp_allowed_mask;
338 	gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
339 }
340 
pm_suspended_storage(void)341 bool pm_suspended_storage(void)
342 {
343 	if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
344 		return false;
345 	return true;
346 }
347 #endif /* CONFIG_PM_SLEEP */
348 
349 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
350 unsigned int pageblock_order __read_mostly;
351 #endif
352 
353 static void __free_pages_ok(struct page *page, unsigned int order,
354 			    fpi_t fpi_flags);
355 
356 /*
357  * results with 256, 32 in the lowmem_reserve sysctl:
358  *	1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
359  *	1G machine -> (16M dma, 784M normal, 224M high)
360  *	NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
361  *	HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
362  *	HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
363  *
364  * TBD: should special case ZONE_DMA32 machines here - in those we normally
365  * don't need any ZONE_NORMAL reservation
366  */
367 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
368 #ifdef CONFIG_ZONE_DMA
369 	[ZONE_DMA] = 256,
370 #endif
371 #ifdef CONFIG_ZONE_DMA32
372 	[ZONE_DMA32] = 256,
373 #endif
374 	[ZONE_NORMAL] = 32,
375 #ifdef CONFIG_HIGHMEM
376 	[ZONE_HIGHMEM] = 0,
377 #endif
378 	[ZONE_MOVABLE] = 0,
379 };
380 
381 static char * const zone_names[MAX_NR_ZONES] = {
382 #ifdef CONFIG_ZONE_DMA
383 	 "DMA",
384 #endif
385 #ifdef CONFIG_ZONE_DMA32
386 	 "DMA32",
387 #endif
388 	 "Normal",
389 #ifdef CONFIG_HIGHMEM
390 	 "HighMem",
391 #endif
392 	 "Movable",
393 #ifdef CONFIG_ZONE_DEVICE
394 	 "Device",
395 #endif
396 };
397 
398 const char * const migratetype_names[MIGRATE_TYPES] = {
399 	"Unmovable",
400 	"Movable",
401 	"Reclaimable",
402 	"HighAtomic",
403 #ifdef CONFIG_CMA
404 	"CMA",
405 #endif
406 #ifdef CONFIG_MEMORY_ISOLATION
407 	"Isolate",
408 #endif
409 };
410 
411 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
412 	[NULL_COMPOUND_DTOR] = NULL,
413 	[COMPOUND_PAGE_DTOR] = free_compound_page,
414 #ifdef CONFIG_HUGETLB_PAGE
415 	[HUGETLB_PAGE_DTOR] = free_huge_page,
416 #endif
417 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
418 	[TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
419 #endif
420 };
421 
422 int min_free_kbytes = 1024;
423 int user_min_free_kbytes = -1;
424 int watermark_boost_factor __read_mostly = 15000;
425 int watermark_scale_factor = 10;
426 
427 static unsigned long nr_kernel_pages __initdata;
428 static unsigned long nr_all_pages __initdata;
429 static unsigned long dma_reserve __initdata;
430 
431 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
432 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
433 static unsigned long required_kernelcore __initdata;
434 static unsigned long required_kernelcore_percent __initdata;
435 static unsigned long required_movablecore __initdata;
436 static unsigned long required_movablecore_percent __initdata;
437 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
438 bool mirrored_kernelcore __initdata_memblock;
439 
440 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
441 int movable_zone;
442 EXPORT_SYMBOL(movable_zone);
443 
444 #if MAX_NUMNODES > 1
445 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
446 unsigned int nr_online_nodes __read_mostly = 1;
447 EXPORT_SYMBOL(nr_node_ids);
448 EXPORT_SYMBOL(nr_online_nodes);
449 #endif
450 
451 int page_group_by_mobility_disabled __read_mostly;
452 
453 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
454 /*
455  * During boot we initialize deferred pages on-demand, as needed, but once
456  * page_alloc_init_late() has finished, the deferred pages are all initialized,
457  * and we can permanently disable that path.
458  */
459 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
460 
deferred_pages_enabled(void)461 static inline bool deferred_pages_enabled(void)
462 {
463 	return static_branch_unlikely(&deferred_pages);
464 }
465 
466 /* Returns true if the struct page for the pfn is uninitialised */
early_page_uninitialised(unsigned long pfn)467 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
468 {
469 	int nid = early_pfn_to_nid(pfn);
470 
471 	if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
472 		return true;
473 
474 	return false;
475 }
476 
477 /*
478  * Returns true when the remaining initialisation should be deferred until
479  * later in the boot cycle when it can be parallelised.
480  */
481 static bool __meminit
defer_init(int nid,unsigned long pfn,unsigned long end_pfn)482 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
483 {
484 	static unsigned long prev_end_pfn, nr_initialised;
485 
486 	if (early_page_ext_enabled())
487 		return false;
488 	/*
489 	 * prev_end_pfn static that contains the end of previous zone
490 	 * No need to protect because called very early in boot before smp_init.
491 	 */
492 	if (prev_end_pfn != end_pfn) {
493 		prev_end_pfn = end_pfn;
494 		nr_initialised = 0;
495 	}
496 
497 	/* Always populate low zones for address-constrained allocations */
498 	if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
499 		return false;
500 
501 	if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
502 		return true;
503 	/*
504 	 * We start only with one section of pages, more pages are added as
505 	 * needed until the rest of deferred pages are initialized.
506 	 */
507 	nr_initialised++;
508 	if ((nr_initialised > PAGES_PER_SECTION) &&
509 	    (pfn & (PAGES_PER_SECTION - 1)) == 0) {
510 		NODE_DATA(nid)->first_deferred_pfn = pfn;
511 		return true;
512 	}
513 	return false;
514 }
515 #else
deferred_pages_enabled(void)516 static inline bool deferred_pages_enabled(void)
517 {
518 	return false;
519 }
520 
early_page_uninitialised(unsigned long pfn)521 static inline bool early_page_uninitialised(unsigned long pfn)
522 {
523 	return false;
524 }
525 
defer_init(int nid,unsigned long pfn,unsigned long end_pfn)526 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
527 {
528 	return false;
529 }
530 #endif
531 
532 /* Return a pointer to the bitmap storing bits affecting a block of pages */
get_pageblock_bitmap(const struct page * page,unsigned long pfn)533 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
534 							unsigned long pfn)
535 {
536 #ifdef CONFIG_SPARSEMEM
537 	return section_to_usemap(__pfn_to_section(pfn));
538 #else
539 	return page_zone(page)->pageblock_flags;
540 #endif /* CONFIG_SPARSEMEM */
541 }
542 
pfn_to_bitidx(const struct page * page,unsigned long pfn)543 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
544 {
545 #ifdef CONFIG_SPARSEMEM
546 	pfn &= (PAGES_PER_SECTION-1);
547 #else
548 	pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
549 #endif /* CONFIG_SPARSEMEM */
550 	return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
551 }
552 
553 static __always_inline
__get_pfnblock_flags_mask(const struct page * page,unsigned long pfn,unsigned long mask)554 unsigned long __get_pfnblock_flags_mask(const struct page *page,
555 					unsigned long pfn,
556 					unsigned long mask)
557 {
558 	unsigned long *bitmap;
559 	unsigned long bitidx, word_bitidx;
560 	unsigned long word;
561 
562 	bitmap = get_pageblock_bitmap(page, pfn);
563 	bitidx = pfn_to_bitidx(page, pfn);
564 	word_bitidx = bitidx / BITS_PER_LONG;
565 	bitidx &= (BITS_PER_LONG-1);
566 	/*
567 	 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
568 	 * a consistent read of the memory array, so that results, even though
569 	 * racy, are not corrupted.
570 	 */
571 	word = READ_ONCE(bitmap[word_bitidx]);
572 	return (word >> bitidx) & mask;
573 }
574 
575 /**
576  * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
577  * @page: The page within the block of interest
578  * @pfn: The target page frame number
579  * @mask: mask of bits that the caller is interested in
580  *
581  * Return: pageblock_bits flags
582  */
get_pfnblock_flags_mask(const struct page * page,unsigned long pfn,unsigned long mask)583 unsigned long get_pfnblock_flags_mask(const struct page *page,
584 					unsigned long pfn, unsigned long mask)
585 {
586 	return __get_pfnblock_flags_mask(page, pfn, mask);
587 }
588 
get_pfnblock_migratetype(const struct page * page,unsigned long pfn)589 static __always_inline int get_pfnblock_migratetype(const struct page *page,
590 					unsigned long pfn)
591 {
592 	return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
593 }
594 
595 /**
596  * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
597  * @page: The page within the block of interest
598  * @flags: The flags to set
599  * @pfn: The target page frame number
600  * @mask: mask of bits that the caller is interested in
601  */
set_pfnblock_flags_mask(struct page * page,unsigned long flags,unsigned long pfn,unsigned long mask)602 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
603 					unsigned long pfn,
604 					unsigned long mask)
605 {
606 	unsigned long *bitmap;
607 	unsigned long bitidx, word_bitidx;
608 	unsigned long word;
609 
610 	BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
611 	BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
612 
613 	bitmap = get_pageblock_bitmap(page, pfn);
614 	bitidx = pfn_to_bitidx(page, pfn);
615 	word_bitidx = bitidx / BITS_PER_LONG;
616 	bitidx &= (BITS_PER_LONG-1);
617 
618 	VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
619 
620 	mask <<= bitidx;
621 	flags <<= bitidx;
622 
623 	word = READ_ONCE(bitmap[word_bitidx]);
624 	do {
625 	} while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
626 }
627 
set_pageblock_migratetype(struct page * page,int migratetype)628 void set_pageblock_migratetype(struct page *page, int migratetype)
629 {
630 	if (unlikely(page_group_by_mobility_disabled &&
631 		     migratetype < MIGRATE_PCPTYPES))
632 		migratetype = MIGRATE_UNMOVABLE;
633 
634 	set_pfnblock_flags_mask(page, (unsigned long)migratetype,
635 				page_to_pfn(page), MIGRATETYPE_MASK);
636 }
637 
638 #ifdef CONFIG_DEBUG_VM
page_outside_zone_boundaries(struct zone * zone,struct page * page)639 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
640 {
641 	int ret = 0;
642 	unsigned seq;
643 	unsigned long pfn = page_to_pfn(page);
644 	unsigned long sp, start_pfn;
645 
646 	do {
647 		seq = zone_span_seqbegin(zone);
648 		start_pfn = zone->zone_start_pfn;
649 		sp = zone->spanned_pages;
650 		if (!zone_spans_pfn(zone, pfn))
651 			ret = 1;
652 	} while (zone_span_seqretry(zone, seq));
653 
654 	if (ret)
655 		pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
656 			pfn, zone_to_nid(zone), zone->name,
657 			start_pfn, start_pfn + sp);
658 
659 	return ret;
660 }
661 
page_is_consistent(struct zone * zone,struct page * page)662 static int page_is_consistent(struct zone *zone, struct page *page)
663 {
664 	if (zone != page_zone(page))
665 		return 0;
666 
667 	return 1;
668 }
669 /*
670  * Temporary debugging check for pages not lying within a given zone.
671  */
bad_range(struct zone * zone,struct page * page)672 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
673 {
674 	if (page_outside_zone_boundaries(zone, page))
675 		return 1;
676 	if (!page_is_consistent(zone, page))
677 		return 1;
678 
679 	return 0;
680 }
681 #else
bad_range(struct zone * zone,struct page * page)682 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
683 {
684 	return 0;
685 }
686 #endif
687 
bad_page(struct page * page,const char * reason)688 static void bad_page(struct page *page, const char *reason)
689 {
690 	static unsigned long resume;
691 	static unsigned long nr_shown;
692 	static unsigned long nr_unshown;
693 
694 	/*
695 	 * Allow a burst of 60 reports, then keep quiet for that minute;
696 	 * or allow a steady drip of one report per second.
697 	 */
698 	if (nr_shown == 60) {
699 		if (time_before(jiffies, resume)) {
700 			nr_unshown++;
701 			goto out;
702 		}
703 		if (nr_unshown) {
704 			pr_alert(
705 			      "BUG: Bad page state: %lu messages suppressed\n",
706 				nr_unshown);
707 			nr_unshown = 0;
708 		}
709 		nr_shown = 0;
710 	}
711 	if (nr_shown++ == 0)
712 		resume = jiffies + 60 * HZ;
713 
714 	pr_alert("BUG: Bad page state in process %s  pfn:%05lx\n",
715 		current->comm, page_to_pfn(page));
716 	dump_page(page, reason);
717 
718 	print_modules();
719 	dump_stack();
720 out:
721 	/* Leave bad fields for debug, except PageBuddy could make trouble */
722 	page_mapcount_reset(page); /* remove PageBuddy */
723 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
724 }
725 
order_to_pindex(int migratetype,int order)726 static inline unsigned int order_to_pindex(int migratetype, int order)
727 {
728 	int base = order;
729 
730 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
731 	if (order > PAGE_ALLOC_COSTLY_ORDER) {
732 		VM_BUG_ON(order != pageblock_order);
733 		return NR_LOWORDER_PCP_LISTS;
734 	}
735 #else
736 	VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
737 #endif
738 
739 	return (MIGRATE_PCPTYPES * base) + migratetype;
740 }
741 
pindex_to_order(unsigned int pindex)742 static inline int pindex_to_order(unsigned int pindex)
743 {
744 	int order = pindex / MIGRATE_PCPTYPES;
745 
746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
747 	if (pindex == NR_LOWORDER_PCP_LISTS)
748 		order = pageblock_order;
749 #else
750 	VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
751 #endif
752 
753 	return order;
754 }
755 
pcp_allowed_order(unsigned int order)756 static inline bool pcp_allowed_order(unsigned int order)
757 {
758 	if (order <= PAGE_ALLOC_COSTLY_ORDER)
759 		return true;
760 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
761 	if (order == pageblock_order)
762 		return true;
763 #endif
764 	return false;
765 }
766 
free_the_page(struct page * page,unsigned int order)767 static inline void free_the_page(struct page *page, unsigned int order)
768 {
769 	if (pcp_allowed_order(order))		/* Via pcp? */
770 		free_unref_page(page, order);
771 	else
772 		__free_pages_ok(page, order, FPI_NONE);
773 }
774 
775 /*
776  * Higher-order pages are called "compound pages".  They are structured thusly:
777  *
778  * The first PAGE_SIZE page is called the "head page" and have PG_head set.
779  *
780  * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
781  * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
782  *
783  * The first tail page's ->compound_dtor holds the offset in array of compound
784  * page destructors. See compound_page_dtors.
785  *
786  * The first tail page's ->compound_order holds the order of allocation.
787  * This usage means that zero-order pages may not be compound.
788  */
789 
free_compound_page(struct page * page)790 void free_compound_page(struct page *page)
791 {
792 	mem_cgroup_uncharge(page_folio(page));
793 	free_the_page(page, compound_order(page));
794 }
795 
prep_compound_head(struct page * page,unsigned int order)796 static void prep_compound_head(struct page *page, unsigned int order)
797 {
798 	set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
799 	set_compound_order(page, order);
800 	atomic_set(compound_mapcount_ptr(page), -1);
801 	atomic_set(compound_pincount_ptr(page), 0);
802 }
803 
prep_compound_tail(struct page * head,int tail_idx)804 static void prep_compound_tail(struct page *head, int tail_idx)
805 {
806 	struct page *p = head + tail_idx;
807 
808 	p->mapping = TAIL_MAPPING;
809 	set_compound_head(p, head);
810 	set_page_private(p, 0);
811 }
812 
prep_compound_page(struct page * page,unsigned int order)813 void prep_compound_page(struct page *page, unsigned int order)
814 {
815 	int i;
816 	int nr_pages = 1 << order;
817 
818 	__SetPageHead(page);
819 	for (i = 1; i < nr_pages; i++)
820 		prep_compound_tail(page, i);
821 
822 	prep_compound_head(page, order);
823 }
824 
destroy_large_folio(struct folio * folio)825 void destroy_large_folio(struct folio *folio)
826 {
827 	enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
828 
829 	VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
830 	compound_page_dtors[dtor](&folio->page);
831 }
832 
833 #ifdef CONFIG_DEBUG_PAGEALLOC
834 unsigned int _debug_guardpage_minorder;
835 
836 bool _debug_pagealloc_enabled_early __read_mostly
837 			= IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
838 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
839 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
840 EXPORT_SYMBOL(_debug_pagealloc_enabled);
841 
842 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
843 
early_debug_pagealloc(char * buf)844 static int __init early_debug_pagealloc(char *buf)
845 {
846 	return kstrtobool(buf, &_debug_pagealloc_enabled_early);
847 }
848 early_param("debug_pagealloc", early_debug_pagealloc);
849 
debug_guardpage_minorder_setup(char * buf)850 static int __init debug_guardpage_minorder_setup(char *buf)
851 {
852 	unsigned long res;
853 
854 	if (kstrtoul(buf, 10, &res) < 0 ||  res > MAX_ORDER / 2) {
855 		pr_err("Bad debug_guardpage_minorder value\n");
856 		return 0;
857 	}
858 	_debug_guardpage_minorder = res;
859 	pr_info("Setting debug_guardpage_minorder to %lu\n", res);
860 	return 0;
861 }
862 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
863 
set_page_guard(struct zone * zone,struct page * page,unsigned int order,int migratetype)864 static inline bool set_page_guard(struct zone *zone, struct page *page,
865 				unsigned int order, int migratetype)
866 {
867 	if (!debug_guardpage_enabled())
868 		return false;
869 
870 	if (order >= debug_guardpage_minorder())
871 		return false;
872 
873 	__SetPageGuard(page);
874 	INIT_LIST_HEAD(&page->buddy_list);
875 	set_page_private(page, order);
876 	/* Guard pages are not available for any usage */
877 	if (!is_migrate_isolate(migratetype))
878 		__mod_zone_freepage_state(zone, -(1 << order), migratetype);
879 
880 	return true;
881 }
882 
clear_page_guard(struct zone * zone,struct page * page,unsigned int order,int migratetype)883 static inline void clear_page_guard(struct zone *zone, struct page *page,
884 				unsigned int order, int migratetype)
885 {
886 	if (!debug_guardpage_enabled())
887 		return;
888 
889 	__ClearPageGuard(page);
890 
891 	set_page_private(page, 0);
892 	if (!is_migrate_isolate(migratetype))
893 		__mod_zone_freepage_state(zone, (1 << order), migratetype);
894 }
895 #else
set_page_guard(struct zone * zone,struct page * page,unsigned int order,int migratetype)896 static inline bool set_page_guard(struct zone *zone, struct page *page,
897 			unsigned int order, int migratetype) { return false; }
clear_page_guard(struct zone * zone,struct page * page,unsigned int order,int migratetype)898 static inline void clear_page_guard(struct zone *zone, struct page *page,
899 				unsigned int order, int migratetype) {}
900 #endif
901 
902 /*
903  * Enable static keys related to various memory debugging and hardening options.
904  * Some override others, and depend on early params that are evaluated in the
905  * order of appearance. So we need to first gather the full picture of what was
906  * enabled, and then make decisions.
907  */
init_mem_debugging_and_hardening(void)908 void __init init_mem_debugging_and_hardening(void)
909 {
910 	bool page_poisoning_requested = false;
911 
912 #ifdef CONFIG_PAGE_POISONING
913 	/*
914 	 * Page poisoning is debug page alloc for some arches. If
915 	 * either of those options are enabled, enable poisoning.
916 	 */
917 	if (page_poisoning_enabled() ||
918 	     (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
919 	      debug_pagealloc_enabled())) {
920 		static_branch_enable(&_page_poisoning_enabled);
921 		page_poisoning_requested = true;
922 	}
923 #endif
924 
925 	if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
926 	    page_poisoning_requested) {
927 		pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
928 			"will take precedence over init_on_alloc and init_on_free\n");
929 		_init_on_alloc_enabled_early = false;
930 		_init_on_free_enabled_early = false;
931 	}
932 
933 	if (_init_on_alloc_enabled_early)
934 		static_branch_enable(&init_on_alloc);
935 	else
936 		static_branch_disable(&init_on_alloc);
937 
938 	if (_init_on_free_enabled_early)
939 		static_branch_enable(&init_on_free);
940 	else
941 		static_branch_disable(&init_on_free);
942 
943 	if (IS_ENABLED(CONFIG_KMSAN) &&
944 	    (_init_on_alloc_enabled_early || _init_on_free_enabled_early))
945 		pr_info("mem auto-init: please make sure init_on_alloc and init_on_free are disabled when running KMSAN\n");
946 
947 #ifdef CONFIG_DEBUG_PAGEALLOC
948 	if (!debug_pagealloc_enabled())
949 		return;
950 
951 	static_branch_enable(&_debug_pagealloc_enabled);
952 
953 	if (!debug_guardpage_minorder())
954 		return;
955 
956 	static_branch_enable(&_debug_guardpage_enabled);
957 #endif
958 }
959 
set_buddy_order(struct page * page,unsigned int order)960 static inline void set_buddy_order(struct page *page, unsigned int order)
961 {
962 	set_page_private(page, order);
963 	__SetPageBuddy(page);
964 }
965 
966 #ifdef CONFIG_COMPACTION
task_capc(struct zone * zone)967 static inline struct capture_control *task_capc(struct zone *zone)
968 {
969 	struct capture_control *capc = current->capture_control;
970 
971 	return unlikely(capc) &&
972 		!(current->flags & PF_KTHREAD) &&
973 		!capc->page &&
974 		capc->cc->zone == zone ? capc : NULL;
975 }
976 
977 static inline bool
compaction_capture(struct capture_control * capc,struct page * page,int order,int migratetype)978 compaction_capture(struct capture_control *capc, struct page *page,
979 		   int order, int migratetype)
980 {
981 	if (!capc || order != capc->cc->order)
982 		return false;
983 
984 	/* Do not accidentally pollute CMA or isolated regions*/
985 	if (is_migrate_cma(migratetype) ||
986 	    is_migrate_isolate(migratetype))
987 		return false;
988 
989 	/*
990 	 * Do not let lower order allocations pollute a movable pageblock.
991 	 * This might let an unmovable request use a reclaimable pageblock
992 	 * and vice-versa but no more than normal fallback logic which can
993 	 * have trouble finding a high-order free page.
994 	 */
995 	if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
996 		return false;
997 
998 	capc->page = page;
999 	return true;
1000 }
1001 
1002 #else
task_capc(struct zone * zone)1003 static inline struct capture_control *task_capc(struct zone *zone)
1004 {
1005 	return NULL;
1006 }
1007 
1008 static inline bool
compaction_capture(struct capture_control * capc,struct page * page,int order,int migratetype)1009 compaction_capture(struct capture_control *capc, struct page *page,
1010 		   int order, int migratetype)
1011 {
1012 	return false;
1013 }
1014 #endif /* CONFIG_COMPACTION */
1015 
1016 /* Used for pages not on another list */
add_to_free_list(struct page * page,struct zone * zone,unsigned int order,int migratetype)1017 static inline void add_to_free_list(struct page *page, struct zone *zone,
1018 				    unsigned int order, int migratetype)
1019 {
1020 	struct free_area *area = &zone->free_area[order];
1021 
1022 	list_add(&page->buddy_list, &area->free_list[migratetype]);
1023 	area->nr_free++;
1024 }
1025 
1026 /* Used for pages not on another list */
add_to_free_list_tail(struct page * page,struct zone * zone,unsigned int order,int migratetype)1027 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
1028 					 unsigned int order, int migratetype)
1029 {
1030 	struct free_area *area = &zone->free_area[order];
1031 
1032 	list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
1033 	area->nr_free++;
1034 }
1035 
1036 /*
1037  * Used for pages which are on another list. Move the pages to the tail
1038  * of the list - so the moved pages won't immediately be considered for
1039  * allocation again (e.g., optimization for memory onlining).
1040  */
move_to_free_list(struct page * page,struct zone * zone,unsigned int order,int migratetype)1041 static inline void move_to_free_list(struct page *page, struct zone *zone,
1042 				     unsigned int order, int migratetype)
1043 {
1044 	struct free_area *area = &zone->free_area[order];
1045 
1046 	list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
1047 }
1048 
del_page_from_free_list(struct page * page,struct zone * zone,unsigned int order)1049 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1050 					   unsigned int order)
1051 {
1052 	/* clear reported state and update reported page count */
1053 	if (page_reported(page))
1054 		__ClearPageReported(page);
1055 
1056 	list_del(&page->buddy_list);
1057 	__ClearPageBuddy(page);
1058 	set_page_private(page, 0);
1059 	zone->free_area[order].nr_free--;
1060 }
1061 
1062 /*
1063  * If this is not the largest possible page, check if the buddy
1064  * of the next-highest order is free. If it is, it's possible
1065  * that pages are being freed that will coalesce soon. In case,
1066  * that is happening, add the free page to the tail of the list
1067  * so it's less likely to be used soon and more likely to be merged
1068  * as a higher order page
1069  */
1070 static inline bool
buddy_merge_likely(unsigned long pfn,unsigned long buddy_pfn,struct page * page,unsigned int order)1071 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1072 		   struct page *page, unsigned int order)
1073 {
1074 	unsigned long higher_page_pfn;
1075 	struct page *higher_page;
1076 
1077 	if (order >= MAX_ORDER - 2)
1078 		return false;
1079 
1080 	higher_page_pfn = buddy_pfn & pfn;
1081 	higher_page = page + (higher_page_pfn - pfn);
1082 
1083 	return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1084 			NULL) != NULL;
1085 }
1086 
1087 /*
1088  * Freeing function for a buddy system allocator.
1089  *
1090  * The concept of a buddy system is to maintain direct-mapped table
1091  * (containing bit values) for memory blocks of various "orders".
1092  * The bottom level table contains the map for the smallest allocatable
1093  * units of memory (here, pages), and each level above it describes
1094  * pairs of units from the levels below, hence, "buddies".
1095  * At a high level, all that happens here is marking the table entry
1096  * at the bottom level available, and propagating the changes upward
1097  * as necessary, plus some accounting needed to play nicely with other
1098  * parts of the VM system.
1099  * At each level, we keep a list of pages, which are heads of continuous
1100  * free pages of length of (1 << order) and marked with PageBuddy.
1101  * Page's order is recorded in page_private(page) field.
1102  * So when we are allocating or freeing one, we can derive the state of the
1103  * other.  That is, if we allocate a small block, and both were
1104  * free, the remainder of the region must be split into blocks.
1105  * If a block is freed, and its buddy is also free, then this
1106  * triggers coalescing into a block of larger size.
1107  *
1108  * -- nyc
1109  */
1110 
__free_one_page(struct page * page,unsigned long pfn,struct zone * zone,unsigned int order,int migratetype,fpi_t fpi_flags)1111 static inline void __free_one_page(struct page *page,
1112 		unsigned long pfn,
1113 		struct zone *zone, unsigned int order,
1114 		int migratetype, fpi_t fpi_flags)
1115 {
1116 	struct capture_control *capc = task_capc(zone);
1117 	unsigned long buddy_pfn = 0;
1118 	unsigned long combined_pfn;
1119 	struct page *buddy;
1120 	bool to_tail;
1121 
1122 	VM_BUG_ON(!zone_is_initialized(zone));
1123 	VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1124 
1125 	VM_BUG_ON(migratetype == -1);
1126 	if (likely(!is_migrate_isolate(migratetype)))
1127 		__mod_zone_freepage_state(zone, 1 << order, migratetype);
1128 
1129 	VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1130 	VM_BUG_ON_PAGE(bad_range(zone, page), page);
1131 
1132 	while (order < MAX_ORDER - 1) {
1133 		if (compaction_capture(capc, page, order, migratetype)) {
1134 			__mod_zone_freepage_state(zone, -(1 << order),
1135 								migratetype);
1136 			return;
1137 		}
1138 
1139 		buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1140 		if (!buddy)
1141 			goto done_merging;
1142 
1143 		if (unlikely(order >= pageblock_order)) {
1144 			/*
1145 			 * We want to prevent merge between freepages on pageblock
1146 			 * without fallbacks and normal pageblock. Without this,
1147 			 * pageblock isolation could cause incorrect freepage or CMA
1148 			 * accounting or HIGHATOMIC accounting.
1149 			 */
1150 			int buddy_mt = get_pageblock_migratetype(buddy);
1151 
1152 			if (migratetype != buddy_mt
1153 					&& (!migratetype_is_mergeable(migratetype) ||
1154 						!migratetype_is_mergeable(buddy_mt)))
1155 				goto done_merging;
1156 		}
1157 
1158 		/*
1159 		 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1160 		 * merge with it and move up one order.
1161 		 */
1162 		if (page_is_guard(buddy))
1163 			clear_page_guard(zone, buddy, order, migratetype);
1164 		else
1165 			del_page_from_free_list(buddy, zone, order);
1166 		combined_pfn = buddy_pfn & pfn;
1167 		page = page + (combined_pfn - pfn);
1168 		pfn = combined_pfn;
1169 		order++;
1170 	}
1171 
1172 done_merging:
1173 	set_buddy_order(page, order);
1174 
1175 	if (fpi_flags & FPI_TO_TAIL)
1176 		to_tail = true;
1177 	else if (is_shuffle_order(order))
1178 		to_tail = shuffle_pick_tail();
1179 	else
1180 		to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1181 
1182 	if (to_tail)
1183 		add_to_free_list_tail(page, zone, order, migratetype);
1184 	else
1185 		add_to_free_list(page, zone, order, migratetype);
1186 
1187 	/* Notify page reporting subsystem of freed page */
1188 	if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1189 		page_reporting_notify_free(order);
1190 }
1191 
1192 /**
1193  * split_free_page() -- split a free page at split_pfn_offset
1194  * @free_page:		the original free page
1195  * @order:		the order of the page
1196  * @split_pfn_offset:	split offset within the page
1197  *
1198  * Return -ENOENT if the free page is changed, otherwise 0
1199  *
1200  * It is used when the free page crosses two pageblocks with different migratetypes
1201  * at split_pfn_offset within the page. The split free page will be put into
1202  * separate migratetype lists afterwards. Otherwise, the function achieves
1203  * nothing.
1204  */
split_free_page(struct page * free_page,unsigned int order,unsigned long split_pfn_offset)1205 int split_free_page(struct page *free_page,
1206 			unsigned int order, unsigned long split_pfn_offset)
1207 {
1208 	struct zone *zone = page_zone(free_page);
1209 	unsigned long free_page_pfn = page_to_pfn(free_page);
1210 	unsigned long pfn;
1211 	unsigned long flags;
1212 	int free_page_order;
1213 	int mt;
1214 	int ret = 0;
1215 
1216 	if (split_pfn_offset == 0)
1217 		return ret;
1218 
1219 	spin_lock_irqsave(&zone->lock, flags);
1220 
1221 	if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1222 		ret = -ENOENT;
1223 		goto out;
1224 	}
1225 
1226 	mt = get_pageblock_migratetype(free_page);
1227 	if (likely(!is_migrate_isolate(mt)))
1228 		__mod_zone_freepage_state(zone, -(1UL << order), mt);
1229 
1230 	del_page_from_free_list(free_page, zone, order);
1231 	for (pfn = free_page_pfn;
1232 	     pfn < free_page_pfn + (1UL << order);) {
1233 		int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1234 
1235 		free_page_order = min_t(unsigned int,
1236 					pfn ? __ffs(pfn) : order,
1237 					__fls(split_pfn_offset));
1238 		__free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1239 				mt, FPI_NONE);
1240 		pfn += 1UL << free_page_order;
1241 		split_pfn_offset -= (1UL << free_page_order);
1242 		/* we have done the first part, now switch to second part */
1243 		if (split_pfn_offset == 0)
1244 			split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1245 	}
1246 out:
1247 	spin_unlock_irqrestore(&zone->lock, flags);
1248 	return ret;
1249 }
1250 /*
1251  * A bad page could be due to a number of fields. Instead of multiple branches,
1252  * try and check multiple fields with one check. The caller must do a detailed
1253  * check if necessary.
1254  */
page_expected_state(struct page * page,unsigned long check_flags)1255 static inline bool page_expected_state(struct page *page,
1256 					unsigned long check_flags)
1257 {
1258 	if (unlikely(atomic_read(&page->_mapcount) != -1))
1259 		return false;
1260 
1261 	if (unlikely((unsigned long)page->mapping |
1262 			page_ref_count(page) |
1263 #ifdef CONFIG_MEMCG
1264 			page->memcg_data |
1265 #endif
1266 			(page->flags & check_flags)))
1267 		return false;
1268 
1269 	return true;
1270 }
1271 
page_bad_reason(struct page * page,unsigned long flags)1272 static const char *page_bad_reason(struct page *page, unsigned long flags)
1273 {
1274 	const char *bad_reason = NULL;
1275 
1276 	if (unlikely(atomic_read(&page->_mapcount) != -1))
1277 		bad_reason = "nonzero mapcount";
1278 	if (unlikely(page->mapping != NULL))
1279 		bad_reason = "non-NULL mapping";
1280 	if (unlikely(page_ref_count(page) != 0))
1281 		bad_reason = "nonzero _refcount";
1282 	if (unlikely(page->flags & flags)) {
1283 		if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1284 			bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1285 		else
1286 			bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1287 	}
1288 #ifdef CONFIG_MEMCG
1289 	if (unlikely(page->memcg_data))
1290 		bad_reason = "page still charged to cgroup";
1291 #endif
1292 	return bad_reason;
1293 }
1294 
free_page_is_bad_report(struct page * page)1295 static void free_page_is_bad_report(struct page *page)
1296 {
1297 	bad_page(page,
1298 		 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1299 }
1300 
free_page_is_bad(struct page * page)1301 static inline bool free_page_is_bad(struct page *page)
1302 {
1303 	if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1304 		return false;
1305 
1306 	/* Something has gone sideways, find it */
1307 	free_page_is_bad_report(page);
1308 	return true;
1309 }
1310 
free_tail_pages_check(struct page * head_page,struct page * page)1311 static int free_tail_pages_check(struct page *head_page, struct page *page)
1312 {
1313 	int ret = 1;
1314 
1315 	/*
1316 	 * We rely page->lru.next never has bit 0 set, unless the page
1317 	 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1318 	 */
1319 	BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1320 
1321 	if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1322 		ret = 0;
1323 		goto out;
1324 	}
1325 	switch (page - head_page) {
1326 	case 1:
1327 		/* the first tail page: ->mapping may be compound_mapcount() */
1328 		if (unlikely(compound_mapcount(page))) {
1329 			bad_page(page, "nonzero compound_mapcount");
1330 			goto out;
1331 		}
1332 		break;
1333 	case 2:
1334 		/*
1335 		 * the second tail page: ->mapping is
1336 		 * deferred_list.next -- ignore value.
1337 		 */
1338 		break;
1339 	default:
1340 		if (page->mapping != TAIL_MAPPING) {
1341 			bad_page(page, "corrupted mapping in tail page");
1342 			goto out;
1343 		}
1344 		break;
1345 	}
1346 	if (unlikely(!PageTail(page))) {
1347 		bad_page(page, "PageTail not set");
1348 		goto out;
1349 	}
1350 	if (unlikely(compound_head(page) != head_page)) {
1351 		bad_page(page, "compound_head not consistent");
1352 		goto out;
1353 	}
1354 	ret = 0;
1355 out:
1356 	page->mapping = NULL;
1357 	clear_compound_head(page);
1358 	return ret;
1359 }
1360 
1361 /*
1362  * Skip KASAN memory poisoning when either:
1363  *
1364  * 1. Deferred memory initialization has not yet completed,
1365  *    see the explanation below.
1366  * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1367  *    see the comment next to it.
1368  * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1369  *    see the comment next to it.
1370  *
1371  * Poisoning pages during deferred memory init will greatly lengthen the
1372  * process and cause problem in large memory systems as the deferred pages
1373  * initialization is done with interrupt disabled.
1374  *
1375  * Assuming that there will be no reference to those newly initialized
1376  * pages before they are ever allocated, this should have no effect on
1377  * KASAN memory tracking as the poison will be properly inserted at page
1378  * allocation time. The only corner case is when pages are allocated by
1379  * on-demand allocation and then freed again before the deferred pages
1380  * initialization is done, but this is not likely to happen.
1381  */
should_skip_kasan_poison(struct page * page,fpi_t fpi_flags)1382 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1383 {
1384 	return deferred_pages_enabled() ||
1385 	       (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1386 		(fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1387 	       PageSkipKASanPoison(page);
1388 }
1389 
kernel_init_pages(struct page * page,int numpages)1390 static void kernel_init_pages(struct page *page, int numpages)
1391 {
1392 	int i;
1393 
1394 	/* s390's use of memset() could override KASAN redzones. */
1395 	kasan_disable_current();
1396 	for (i = 0; i < numpages; i++)
1397 		clear_highpage_kasan_tagged(page + i);
1398 	kasan_enable_current();
1399 }
1400 
free_pages_prepare(struct page * page,unsigned int order,bool check_free,fpi_t fpi_flags)1401 static __always_inline bool free_pages_prepare(struct page *page,
1402 			unsigned int order, bool check_free, fpi_t fpi_flags)
1403 {
1404 	int bad = 0;
1405 	bool init = want_init_on_free();
1406 
1407 	VM_BUG_ON_PAGE(PageTail(page), page);
1408 
1409 	trace_mm_page_free(page, order);
1410 	kmsan_free_page(page, order);
1411 
1412 	if (unlikely(PageHWPoison(page)) && !order) {
1413 		/*
1414 		 * Do not let hwpoison pages hit pcplists/buddy
1415 		 * Untie memcg state and reset page's owner
1416 		 */
1417 		if (memcg_kmem_enabled() && PageMemcgKmem(page))
1418 			__memcg_kmem_uncharge_page(page, order);
1419 		reset_page_owner(page, order);
1420 		page_table_check_free(page, order);
1421 		return false;
1422 	}
1423 
1424 	/*
1425 	 * Check tail pages before head page information is cleared to
1426 	 * avoid checking PageCompound for order-0 pages.
1427 	 */
1428 	if (unlikely(order)) {
1429 		bool compound = PageCompound(page);
1430 		int i;
1431 
1432 		VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1433 
1434 		if (compound) {
1435 			ClearPageDoubleMap(page);
1436 			ClearPageHasHWPoisoned(page);
1437 		}
1438 		for (i = 1; i < (1 << order); i++) {
1439 			if (compound)
1440 				bad += free_tail_pages_check(page, page + i);
1441 			if (unlikely(free_page_is_bad(page + i))) {
1442 				bad++;
1443 				continue;
1444 			}
1445 			(page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1446 		}
1447 	}
1448 	if (PageMappingFlags(page))
1449 		page->mapping = NULL;
1450 	if (memcg_kmem_enabled() && PageMemcgKmem(page))
1451 		__memcg_kmem_uncharge_page(page, order);
1452 	if (check_free && free_page_is_bad(page))
1453 		bad++;
1454 	if (bad)
1455 		return false;
1456 
1457 	page_cpupid_reset_last(page);
1458 	page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1459 	reset_page_owner(page, order);
1460 	page_table_check_free(page, order);
1461 
1462 	if (!PageHighMem(page)) {
1463 		debug_check_no_locks_freed(page_address(page),
1464 					   PAGE_SIZE << order);
1465 		debug_check_no_obj_freed(page_address(page),
1466 					   PAGE_SIZE << order);
1467 	}
1468 
1469 	kernel_poison_pages(page, 1 << order);
1470 
1471 	/*
1472 	 * As memory initialization might be integrated into KASAN,
1473 	 * KASAN poisoning and memory initialization code must be
1474 	 * kept together to avoid discrepancies in behavior.
1475 	 *
1476 	 * With hardware tag-based KASAN, memory tags must be set before the
1477 	 * page becomes unavailable via debug_pagealloc or arch_free_page.
1478 	 */
1479 	if (!should_skip_kasan_poison(page, fpi_flags)) {
1480 		kasan_poison_pages(page, order, init);
1481 
1482 		/* Memory is already initialized if KASAN did it internally. */
1483 		if (kasan_has_integrated_init())
1484 			init = false;
1485 	}
1486 	if (init)
1487 		kernel_init_pages(page, 1 << order);
1488 
1489 	/*
1490 	 * arch_free_page() can make the page's contents inaccessible.  s390
1491 	 * does this.  So nothing which can access the page's contents should
1492 	 * happen after this.
1493 	 */
1494 	arch_free_page(page, order);
1495 
1496 	debug_pagealloc_unmap_pages(page, 1 << order);
1497 
1498 	return true;
1499 }
1500 
1501 #ifdef CONFIG_DEBUG_VM
1502 /*
1503  * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1504  * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1505  * moved from pcp lists to free lists.
1506  */
free_pcp_prepare(struct page * page,unsigned int order)1507 static bool free_pcp_prepare(struct page *page, unsigned int order)
1508 {
1509 	return free_pages_prepare(page, order, true, FPI_NONE);
1510 }
1511 
1512 /* return true if this page has an inappropriate state */
bulkfree_pcp_prepare(struct page * page)1513 static bool bulkfree_pcp_prepare(struct page *page)
1514 {
1515 	if (debug_pagealloc_enabled_static())
1516 		return free_page_is_bad(page);
1517 	else
1518 		return false;
1519 }
1520 #else
1521 /*
1522  * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1523  * moving from pcp lists to free list in order to reduce overhead. With
1524  * debug_pagealloc enabled, they are checked also immediately when being freed
1525  * to the pcp lists.
1526  */
free_pcp_prepare(struct page * page,unsigned int order)1527 static bool free_pcp_prepare(struct page *page, unsigned int order)
1528 {
1529 	if (debug_pagealloc_enabled_static())
1530 		return free_pages_prepare(page, order, true, FPI_NONE);
1531 	else
1532 		return free_pages_prepare(page, order, false, FPI_NONE);
1533 }
1534 
bulkfree_pcp_prepare(struct page * page)1535 static bool bulkfree_pcp_prepare(struct page *page)
1536 {
1537 	return free_page_is_bad(page);
1538 }
1539 #endif /* CONFIG_DEBUG_VM */
1540 
1541 /*
1542  * Frees a number of pages from the PCP lists
1543  * Assumes all pages on list are in same zone.
1544  * count is the number of pages to free.
1545  */
free_pcppages_bulk(struct zone * zone,int count,struct per_cpu_pages * pcp,int pindex)1546 static void free_pcppages_bulk(struct zone *zone, int count,
1547 					struct per_cpu_pages *pcp,
1548 					int pindex)
1549 {
1550 	int min_pindex = 0;
1551 	int max_pindex = NR_PCP_LISTS - 1;
1552 	unsigned int order;
1553 	bool isolated_pageblocks;
1554 	struct page *page;
1555 
1556 	/*
1557 	 * Ensure proper count is passed which otherwise would stuck in the
1558 	 * below while (list_empty(list)) loop.
1559 	 */
1560 	count = min(pcp->count, count);
1561 
1562 	/* Ensure requested pindex is drained first. */
1563 	pindex = pindex - 1;
1564 
1565 	/* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
1566 	spin_lock(&zone->lock);
1567 	isolated_pageblocks = has_isolate_pageblock(zone);
1568 
1569 	while (count > 0) {
1570 		struct list_head *list;
1571 		int nr_pages;
1572 
1573 		/* Remove pages from lists in a round-robin fashion. */
1574 		do {
1575 			if (++pindex > max_pindex)
1576 				pindex = min_pindex;
1577 			list = &pcp->lists[pindex];
1578 			if (!list_empty(list))
1579 				break;
1580 
1581 			if (pindex == max_pindex)
1582 				max_pindex--;
1583 			if (pindex == min_pindex)
1584 				min_pindex++;
1585 		} while (1);
1586 
1587 		order = pindex_to_order(pindex);
1588 		nr_pages = 1 << order;
1589 		do {
1590 			int mt;
1591 
1592 			page = list_last_entry(list, struct page, pcp_list);
1593 			mt = get_pcppage_migratetype(page);
1594 
1595 			/* must delete to avoid corrupting pcp list */
1596 			list_del(&page->pcp_list);
1597 			count -= nr_pages;
1598 			pcp->count -= nr_pages;
1599 
1600 			if (bulkfree_pcp_prepare(page))
1601 				continue;
1602 
1603 			/* MIGRATE_ISOLATE page should not go to pcplists */
1604 			VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1605 			/* Pageblock could have been isolated meanwhile */
1606 			if (unlikely(isolated_pageblocks))
1607 				mt = get_pageblock_migratetype(page);
1608 
1609 			__free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1610 			trace_mm_page_pcpu_drain(page, order, mt);
1611 		} while (count > 0 && !list_empty(list));
1612 	}
1613 
1614 	spin_unlock(&zone->lock);
1615 }
1616 
free_one_page(struct zone * zone,struct page * page,unsigned long pfn,unsigned int order,int migratetype,fpi_t fpi_flags)1617 static void free_one_page(struct zone *zone,
1618 				struct page *page, unsigned long pfn,
1619 				unsigned int order,
1620 				int migratetype, fpi_t fpi_flags)
1621 {
1622 	unsigned long flags;
1623 
1624 	spin_lock_irqsave(&zone->lock, flags);
1625 	if (unlikely(has_isolate_pageblock(zone) ||
1626 		is_migrate_isolate(migratetype))) {
1627 		migratetype = get_pfnblock_migratetype(page, pfn);
1628 	}
1629 	__free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1630 	spin_unlock_irqrestore(&zone->lock, flags);
1631 }
1632 
__init_single_page(struct page * page,unsigned long pfn,unsigned long zone,int nid)1633 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1634 				unsigned long zone, int nid)
1635 {
1636 	mm_zero_struct_page(page);
1637 	set_page_links(page, zone, nid, pfn);
1638 	init_page_count(page);
1639 	page_mapcount_reset(page);
1640 	page_cpupid_reset_last(page);
1641 	page_kasan_tag_reset(page);
1642 
1643 	INIT_LIST_HEAD(&page->lru);
1644 #ifdef WANT_PAGE_VIRTUAL
1645 	/* The shift won't overflow because ZONE_NORMAL is below 4G. */
1646 	if (!is_highmem_idx(zone))
1647 		set_page_address(page, __va(pfn << PAGE_SHIFT));
1648 #endif
1649 }
1650 
1651 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
init_reserved_page(unsigned long pfn)1652 static void __meminit init_reserved_page(unsigned long pfn)
1653 {
1654 	pg_data_t *pgdat;
1655 	int nid, zid;
1656 
1657 	if (!early_page_uninitialised(pfn))
1658 		return;
1659 
1660 	nid = early_pfn_to_nid(pfn);
1661 	pgdat = NODE_DATA(nid);
1662 
1663 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1664 		struct zone *zone = &pgdat->node_zones[zid];
1665 
1666 		if (zone_spans_pfn(zone, pfn))
1667 			break;
1668 	}
1669 	__init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1670 }
1671 #else
init_reserved_page(unsigned long pfn)1672 static inline void init_reserved_page(unsigned long pfn)
1673 {
1674 }
1675 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1676 
1677 /*
1678  * Initialised pages do not have PageReserved set. This function is
1679  * called for each range allocated by the bootmem allocator and
1680  * marks the pages PageReserved. The remaining valid pages are later
1681  * sent to the buddy page allocator.
1682  */
reserve_bootmem_region(phys_addr_t start,phys_addr_t end)1683 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1684 {
1685 	unsigned long start_pfn = PFN_DOWN(start);
1686 	unsigned long end_pfn = PFN_UP(end);
1687 
1688 	for (; start_pfn < end_pfn; start_pfn++) {
1689 		if (pfn_valid(start_pfn)) {
1690 			struct page *page = pfn_to_page(start_pfn);
1691 
1692 			init_reserved_page(start_pfn);
1693 
1694 			/* Avoid false-positive PageTail() */
1695 			INIT_LIST_HEAD(&page->lru);
1696 
1697 			/*
1698 			 * no need for atomic set_bit because the struct
1699 			 * page is not visible yet so nobody should
1700 			 * access it yet.
1701 			 */
1702 			__SetPageReserved(page);
1703 		}
1704 	}
1705 }
1706 
__free_pages_ok(struct page * page,unsigned int order,fpi_t fpi_flags)1707 static void __free_pages_ok(struct page *page, unsigned int order,
1708 			    fpi_t fpi_flags)
1709 {
1710 	unsigned long flags;
1711 	int migratetype;
1712 	unsigned long pfn = page_to_pfn(page);
1713 	struct zone *zone = page_zone(page);
1714 
1715 	if (!free_pages_prepare(page, order, true, fpi_flags))
1716 		return;
1717 
1718 	migratetype = get_pfnblock_migratetype(page, pfn);
1719 
1720 	spin_lock_irqsave(&zone->lock, flags);
1721 	if (unlikely(has_isolate_pageblock(zone) ||
1722 		is_migrate_isolate(migratetype))) {
1723 		migratetype = get_pfnblock_migratetype(page, pfn);
1724 	}
1725 	__free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1726 	spin_unlock_irqrestore(&zone->lock, flags);
1727 
1728 	__count_vm_events(PGFREE, 1 << order);
1729 }
1730 
__free_pages_core(struct page * page,unsigned int order)1731 void __free_pages_core(struct page *page, unsigned int order)
1732 {
1733 	unsigned int nr_pages = 1 << order;
1734 	struct page *p = page;
1735 	unsigned int loop;
1736 
1737 	/*
1738 	 * When initializing the memmap, __init_single_page() sets the refcount
1739 	 * of all pages to 1 ("allocated"/"not free"). We have to set the
1740 	 * refcount of all involved pages to 0.
1741 	 */
1742 	prefetchw(p);
1743 	for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1744 		prefetchw(p + 1);
1745 		__ClearPageReserved(p);
1746 		set_page_count(p, 0);
1747 	}
1748 	__ClearPageReserved(p);
1749 	set_page_count(p, 0);
1750 
1751 	atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1752 
1753 	/*
1754 	 * Bypass PCP and place fresh pages right to the tail, primarily
1755 	 * relevant for memory onlining.
1756 	 */
1757 	__free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1758 }
1759 
1760 #ifdef CONFIG_NUMA
1761 
1762 /*
1763  * During memory init memblocks map pfns to nids. The search is expensive and
1764  * this caches recent lookups. The implementation of __early_pfn_to_nid
1765  * treats start/end as pfns.
1766  */
1767 struct mminit_pfnnid_cache {
1768 	unsigned long last_start;
1769 	unsigned long last_end;
1770 	int last_nid;
1771 };
1772 
1773 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1774 
1775 /*
1776  * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1777  */
__early_pfn_to_nid(unsigned long pfn,struct mminit_pfnnid_cache * state)1778 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1779 					struct mminit_pfnnid_cache *state)
1780 {
1781 	unsigned long start_pfn, end_pfn;
1782 	int nid;
1783 
1784 	if (state->last_start <= pfn && pfn < state->last_end)
1785 		return state->last_nid;
1786 
1787 	nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1788 	if (nid != NUMA_NO_NODE) {
1789 		state->last_start = start_pfn;
1790 		state->last_end = end_pfn;
1791 		state->last_nid = nid;
1792 	}
1793 
1794 	return nid;
1795 }
1796 
early_pfn_to_nid(unsigned long pfn)1797 int __meminit early_pfn_to_nid(unsigned long pfn)
1798 {
1799 	static DEFINE_SPINLOCK(early_pfn_lock);
1800 	int nid;
1801 
1802 	spin_lock(&early_pfn_lock);
1803 	nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1804 	if (nid < 0)
1805 		nid = first_online_node;
1806 	spin_unlock(&early_pfn_lock);
1807 
1808 	return nid;
1809 }
1810 #endif /* CONFIG_NUMA */
1811 
memblock_free_pages(struct page * page,unsigned long pfn,unsigned int order)1812 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1813 							unsigned int order)
1814 {
1815 	if (early_page_uninitialised(pfn))
1816 		return;
1817 	if (!kmsan_memblock_free_pages(page, order)) {
1818 		/* KMSAN will take care of these pages. */
1819 		return;
1820 	}
1821 	__free_pages_core(page, order);
1822 }
1823 
1824 /*
1825  * Check that the whole (or subset of) a pageblock given by the interval of
1826  * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1827  * with the migration of free compaction scanner.
1828  *
1829  * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1830  *
1831  * It's possible on some configurations to have a setup like node0 node1 node0
1832  * i.e. it's possible that all pages within a zones range of pages do not
1833  * belong to a single zone. We assume that a border between node0 and node1
1834  * can occur within a single pageblock, but not a node0 node1 node0
1835  * interleaving within a single pageblock. It is therefore sufficient to check
1836  * the first and last page of a pageblock and avoid checking each individual
1837  * page in a pageblock.
1838  */
__pageblock_pfn_to_page(unsigned long start_pfn,unsigned long end_pfn,struct zone * zone)1839 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1840 				     unsigned long end_pfn, struct zone *zone)
1841 {
1842 	struct page *start_page;
1843 	struct page *end_page;
1844 
1845 	/* end_pfn is one past the range we are checking */
1846 	end_pfn--;
1847 
1848 	if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1849 		return NULL;
1850 
1851 	start_page = pfn_to_online_page(start_pfn);
1852 	if (!start_page)
1853 		return NULL;
1854 
1855 	if (page_zone(start_page) != zone)
1856 		return NULL;
1857 
1858 	end_page = pfn_to_page(end_pfn);
1859 
1860 	/* This gives a shorter code than deriving page_zone(end_page) */
1861 	if (page_zone_id(start_page) != page_zone_id(end_page))
1862 		return NULL;
1863 
1864 	return start_page;
1865 }
1866 
set_zone_contiguous(struct zone * zone)1867 void set_zone_contiguous(struct zone *zone)
1868 {
1869 	unsigned long block_start_pfn = zone->zone_start_pfn;
1870 	unsigned long block_end_pfn;
1871 
1872 	block_end_pfn = pageblock_end_pfn(block_start_pfn);
1873 	for (; block_start_pfn < zone_end_pfn(zone);
1874 			block_start_pfn = block_end_pfn,
1875 			 block_end_pfn += pageblock_nr_pages) {
1876 
1877 		block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1878 
1879 		if (!__pageblock_pfn_to_page(block_start_pfn,
1880 					     block_end_pfn, zone))
1881 			return;
1882 		cond_resched();
1883 	}
1884 
1885 	/* We confirm that there is no hole */
1886 	zone->contiguous = true;
1887 }
1888 
clear_zone_contiguous(struct zone * zone)1889 void clear_zone_contiguous(struct zone *zone)
1890 {
1891 	zone->contiguous = false;
1892 }
1893 
1894 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
deferred_free_range(unsigned long pfn,unsigned long nr_pages)1895 static void __init deferred_free_range(unsigned long pfn,
1896 				       unsigned long nr_pages)
1897 {
1898 	struct page *page;
1899 	unsigned long i;
1900 
1901 	if (!nr_pages)
1902 		return;
1903 
1904 	page = pfn_to_page(pfn);
1905 
1906 	/* Free a large naturally-aligned chunk if possible */
1907 	if (nr_pages == pageblock_nr_pages && pageblock_aligned(pfn)) {
1908 		set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1909 		__free_pages_core(page, pageblock_order);
1910 		return;
1911 	}
1912 
1913 	for (i = 0; i < nr_pages; i++, page++, pfn++) {
1914 		if (pageblock_aligned(pfn))
1915 			set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1916 		__free_pages_core(page, 0);
1917 	}
1918 }
1919 
1920 /* Completion tracking for deferred_init_memmap() threads */
1921 static atomic_t pgdat_init_n_undone __initdata;
1922 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1923 
pgdat_init_report_one_done(void)1924 static inline void __init pgdat_init_report_one_done(void)
1925 {
1926 	if (atomic_dec_and_test(&pgdat_init_n_undone))
1927 		complete(&pgdat_init_all_done_comp);
1928 }
1929 
1930 /*
1931  * Returns true if page needs to be initialized or freed to buddy allocator.
1932  *
1933  * We check if a current large page is valid by only checking the validity
1934  * of the head pfn.
1935  */
deferred_pfn_valid(unsigned long pfn)1936 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1937 {
1938 	if (pageblock_aligned(pfn) && !pfn_valid(pfn))
1939 		return false;
1940 	return true;
1941 }
1942 
1943 /*
1944  * Free pages to buddy allocator. Try to free aligned pages in
1945  * pageblock_nr_pages sizes.
1946  */
deferred_free_pages(unsigned long pfn,unsigned long end_pfn)1947 static void __init deferred_free_pages(unsigned long pfn,
1948 				       unsigned long end_pfn)
1949 {
1950 	unsigned long nr_free = 0;
1951 
1952 	for (; pfn < end_pfn; pfn++) {
1953 		if (!deferred_pfn_valid(pfn)) {
1954 			deferred_free_range(pfn - nr_free, nr_free);
1955 			nr_free = 0;
1956 		} else if (pageblock_aligned(pfn)) {
1957 			deferred_free_range(pfn - nr_free, nr_free);
1958 			nr_free = 1;
1959 		} else {
1960 			nr_free++;
1961 		}
1962 	}
1963 	/* Free the last block of pages to allocator */
1964 	deferred_free_range(pfn - nr_free, nr_free);
1965 }
1966 
1967 /*
1968  * Initialize struct pages.  We minimize pfn page lookups and scheduler checks
1969  * by performing it only once every pageblock_nr_pages.
1970  * Return number of pages initialized.
1971  */
deferred_init_pages(struct zone * zone,unsigned long pfn,unsigned long end_pfn)1972 static unsigned long  __init deferred_init_pages(struct zone *zone,
1973 						 unsigned long pfn,
1974 						 unsigned long end_pfn)
1975 {
1976 	int nid = zone_to_nid(zone);
1977 	unsigned long nr_pages = 0;
1978 	int zid = zone_idx(zone);
1979 	struct page *page = NULL;
1980 
1981 	for (; pfn < end_pfn; pfn++) {
1982 		if (!deferred_pfn_valid(pfn)) {
1983 			page = NULL;
1984 			continue;
1985 		} else if (!page || pageblock_aligned(pfn)) {
1986 			page = pfn_to_page(pfn);
1987 		} else {
1988 			page++;
1989 		}
1990 		__init_single_page(page, pfn, zid, nid);
1991 		nr_pages++;
1992 	}
1993 	return (nr_pages);
1994 }
1995 
1996 /*
1997  * This function is meant to pre-load the iterator for the zone init.
1998  * Specifically it walks through the ranges until we are caught up to the
1999  * first_init_pfn value and exits there. If we never encounter the value we
2000  * return false indicating there are no valid ranges left.
2001  */
2002 static bool __init
deferred_init_mem_pfn_range_in_zone(u64 * i,struct zone * zone,unsigned long * spfn,unsigned long * epfn,unsigned long first_init_pfn)2003 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
2004 				    unsigned long *spfn, unsigned long *epfn,
2005 				    unsigned long first_init_pfn)
2006 {
2007 	u64 j;
2008 
2009 	/*
2010 	 * Start out by walking through the ranges in this zone that have
2011 	 * already been initialized. We don't need to do anything with them
2012 	 * so we just need to flush them out of the system.
2013 	 */
2014 	for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
2015 		if (*epfn <= first_init_pfn)
2016 			continue;
2017 		if (*spfn < first_init_pfn)
2018 			*spfn = first_init_pfn;
2019 		*i = j;
2020 		return true;
2021 	}
2022 
2023 	return false;
2024 }
2025 
2026 /*
2027  * Initialize and free pages. We do it in two loops: first we initialize
2028  * struct page, then free to buddy allocator, because while we are
2029  * freeing pages we can access pages that are ahead (computing buddy
2030  * page in __free_one_page()).
2031  *
2032  * In order to try and keep some memory in the cache we have the loop
2033  * broken along max page order boundaries. This way we will not cause
2034  * any issues with the buddy page computation.
2035  */
2036 static unsigned long __init
deferred_init_maxorder(u64 * i,struct zone * zone,unsigned long * start_pfn,unsigned long * end_pfn)2037 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2038 		       unsigned long *end_pfn)
2039 {
2040 	unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2041 	unsigned long spfn = *start_pfn, epfn = *end_pfn;
2042 	unsigned long nr_pages = 0;
2043 	u64 j = *i;
2044 
2045 	/* First we loop through and initialize the page values */
2046 	for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2047 		unsigned long t;
2048 
2049 		if (mo_pfn <= *start_pfn)
2050 			break;
2051 
2052 		t = min(mo_pfn, *end_pfn);
2053 		nr_pages += deferred_init_pages(zone, *start_pfn, t);
2054 
2055 		if (mo_pfn < *end_pfn) {
2056 			*start_pfn = mo_pfn;
2057 			break;
2058 		}
2059 	}
2060 
2061 	/* Reset values and now loop through freeing pages as needed */
2062 	swap(j, *i);
2063 
2064 	for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2065 		unsigned long t;
2066 
2067 		if (mo_pfn <= spfn)
2068 			break;
2069 
2070 		t = min(mo_pfn, epfn);
2071 		deferred_free_pages(spfn, t);
2072 
2073 		if (mo_pfn <= epfn)
2074 			break;
2075 	}
2076 
2077 	return nr_pages;
2078 }
2079 
2080 static void __init
deferred_init_memmap_chunk(unsigned long start_pfn,unsigned long end_pfn,void * arg)2081 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2082 			   void *arg)
2083 {
2084 	unsigned long spfn, epfn;
2085 	struct zone *zone = arg;
2086 	u64 i;
2087 
2088 	deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2089 
2090 	/*
2091 	 * Initialize and free pages in MAX_ORDER sized increments so that we
2092 	 * can avoid introducing any issues with the buddy allocator.
2093 	 */
2094 	while (spfn < end_pfn) {
2095 		deferred_init_maxorder(&i, zone, &spfn, &epfn);
2096 		cond_resched();
2097 	}
2098 }
2099 
2100 /* An arch may override for more concurrency. */
2101 __weak int __init
deferred_page_init_max_threads(const struct cpumask * node_cpumask)2102 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2103 {
2104 	return 1;
2105 }
2106 
2107 /* Initialise remaining memory on a node */
deferred_init_memmap(void * data)2108 static int __init deferred_init_memmap(void *data)
2109 {
2110 	pg_data_t *pgdat = data;
2111 	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2112 	unsigned long spfn = 0, epfn = 0;
2113 	unsigned long first_init_pfn, flags;
2114 	unsigned long start = jiffies;
2115 	struct zone *zone;
2116 	int zid, max_threads;
2117 	u64 i;
2118 
2119 	/* Bind memory initialisation thread to a local node if possible */
2120 	if (!cpumask_empty(cpumask))
2121 		set_cpus_allowed_ptr(current, cpumask);
2122 
2123 	pgdat_resize_lock(pgdat, &flags);
2124 	first_init_pfn = pgdat->first_deferred_pfn;
2125 	if (first_init_pfn == ULONG_MAX) {
2126 		pgdat_resize_unlock(pgdat, &flags);
2127 		pgdat_init_report_one_done();
2128 		return 0;
2129 	}
2130 
2131 	/* Sanity check boundaries */
2132 	BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2133 	BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2134 	pgdat->first_deferred_pfn = ULONG_MAX;
2135 
2136 	/*
2137 	 * Once we unlock here, the zone cannot be grown anymore, thus if an
2138 	 * interrupt thread must allocate this early in boot, zone must be
2139 	 * pre-grown prior to start of deferred page initialization.
2140 	 */
2141 	pgdat_resize_unlock(pgdat, &flags);
2142 
2143 	/* Only the highest zone is deferred so find it */
2144 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2145 		zone = pgdat->node_zones + zid;
2146 		if (first_init_pfn < zone_end_pfn(zone))
2147 			break;
2148 	}
2149 
2150 	/* If the zone is empty somebody else may have cleared out the zone */
2151 	if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2152 						 first_init_pfn))
2153 		goto zone_empty;
2154 
2155 	max_threads = deferred_page_init_max_threads(cpumask);
2156 
2157 	while (spfn < epfn) {
2158 		unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2159 		struct padata_mt_job job = {
2160 			.thread_fn   = deferred_init_memmap_chunk,
2161 			.fn_arg      = zone,
2162 			.start       = spfn,
2163 			.size        = epfn_align - spfn,
2164 			.align       = PAGES_PER_SECTION,
2165 			.min_chunk   = PAGES_PER_SECTION,
2166 			.max_threads = max_threads,
2167 		};
2168 
2169 		padata_do_multithreaded(&job);
2170 		deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2171 						    epfn_align);
2172 	}
2173 zone_empty:
2174 	/* Sanity check that the next zone really is unpopulated */
2175 	WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2176 
2177 	pr_info("node %d deferred pages initialised in %ums\n",
2178 		pgdat->node_id, jiffies_to_msecs(jiffies - start));
2179 
2180 	pgdat_init_report_one_done();
2181 	return 0;
2182 }
2183 
2184 /*
2185  * If this zone has deferred pages, try to grow it by initializing enough
2186  * deferred pages to satisfy the allocation specified by order, rounded up to
2187  * the nearest PAGES_PER_SECTION boundary.  So we're adding memory in increments
2188  * of SECTION_SIZE bytes by initializing struct pages in increments of
2189  * PAGES_PER_SECTION * sizeof(struct page) bytes.
2190  *
2191  * Return true when zone was grown, otherwise return false. We return true even
2192  * when we grow less than requested, to let the caller decide if there are
2193  * enough pages to satisfy the allocation.
2194  *
2195  * Note: We use noinline because this function is needed only during boot, and
2196  * it is called from a __ref function _deferred_grow_zone. This way we are
2197  * making sure that it is not inlined into permanent text section.
2198  */
2199 static noinline bool __init
deferred_grow_zone(struct zone * zone,unsigned int order)2200 deferred_grow_zone(struct zone *zone, unsigned int order)
2201 {
2202 	unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2203 	pg_data_t *pgdat = zone->zone_pgdat;
2204 	unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2205 	unsigned long spfn, epfn, flags;
2206 	unsigned long nr_pages = 0;
2207 	u64 i;
2208 
2209 	/* Only the last zone may have deferred pages */
2210 	if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2211 		return false;
2212 
2213 	pgdat_resize_lock(pgdat, &flags);
2214 
2215 	/*
2216 	 * If someone grew this zone while we were waiting for spinlock, return
2217 	 * true, as there might be enough pages already.
2218 	 */
2219 	if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2220 		pgdat_resize_unlock(pgdat, &flags);
2221 		return true;
2222 	}
2223 
2224 	/* If the zone is empty somebody else may have cleared out the zone */
2225 	if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2226 						 first_deferred_pfn)) {
2227 		pgdat->first_deferred_pfn = ULONG_MAX;
2228 		pgdat_resize_unlock(pgdat, &flags);
2229 		/* Retry only once. */
2230 		return first_deferred_pfn != ULONG_MAX;
2231 	}
2232 
2233 	/*
2234 	 * Initialize and free pages in MAX_ORDER sized increments so
2235 	 * that we can avoid introducing any issues with the buddy
2236 	 * allocator.
2237 	 */
2238 	while (spfn < epfn) {
2239 		/* update our first deferred PFN for this section */
2240 		first_deferred_pfn = spfn;
2241 
2242 		nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2243 		touch_nmi_watchdog();
2244 
2245 		/* We should only stop along section boundaries */
2246 		if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2247 			continue;
2248 
2249 		/* If our quota has been met we can stop here */
2250 		if (nr_pages >= nr_pages_needed)
2251 			break;
2252 	}
2253 
2254 	pgdat->first_deferred_pfn = spfn;
2255 	pgdat_resize_unlock(pgdat, &flags);
2256 
2257 	return nr_pages > 0;
2258 }
2259 
2260 /*
2261  * deferred_grow_zone() is __init, but it is called from
2262  * get_page_from_freelist() during early boot until deferred_pages permanently
2263  * disables this call. This is why we have refdata wrapper to avoid warning,
2264  * and to ensure that the function body gets unloaded.
2265  */
2266 static bool __ref
_deferred_grow_zone(struct zone * zone,unsigned int order)2267 _deferred_grow_zone(struct zone *zone, unsigned int order)
2268 {
2269 	return deferred_grow_zone(zone, order);
2270 }
2271 
2272 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2273 
page_alloc_init_late(void)2274 void __init page_alloc_init_late(void)
2275 {
2276 	struct zone *zone;
2277 	int nid;
2278 
2279 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2280 
2281 	/* There will be num_node_state(N_MEMORY) threads */
2282 	atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2283 	for_each_node_state(nid, N_MEMORY) {
2284 		kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2285 	}
2286 
2287 	/* Block until all are initialised */
2288 	wait_for_completion(&pgdat_init_all_done_comp);
2289 
2290 	/*
2291 	 * We initialized the rest of the deferred pages.  Permanently disable
2292 	 * on-demand struct page initialization.
2293 	 */
2294 	static_branch_disable(&deferred_pages);
2295 
2296 	/* Reinit limits that are based on free pages after the kernel is up */
2297 	files_maxfiles_init();
2298 #endif
2299 
2300 	buffer_init();
2301 
2302 	/* Discard memblock private memory */
2303 	memblock_discard();
2304 
2305 	for_each_node_state(nid, N_MEMORY)
2306 		shuffle_free_memory(NODE_DATA(nid));
2307 
2308 	for_each_populated_zone(zone)
2309 		set_zone_contiguous(zone);
2310 }
2311 
2312 #ifdef CONFIG_CMA
2313 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
init_cma_reserved_pageblock(struct page * page)2314 void __init init_cma_reserved_pageblock(struct page *page)
2315 {
2316 	unsigned i = pageblock_nr_pages;
2317 	struct page *p = page;
2318 
2319 	do {
2320 		__ClearPageReserved(p);
2321 		set_page_count(p, 0);
2322 	} while (++p, --i);
2323 
2324 	set_pageblock_migratetype(page, MIGRATE_CMA);
2325 	set_page_refcounted(page);
2326 	__free_pages(page, pageblock_order);
2327 
2328 	adjust_managed_page_count(page, pageblock_nr_pages);
2329 	page_zone(page)->cma_pages += pageblock_nr_pages;
2330 }
2331 #endif
2332 
2333 /*
2334  * The order of subdivision here is critical for the IO subsystem.
2335  * Please do not alter this order without good reasons and regression
2336  * testing. Specifically, as large blocks of memory are subdivided,
2337  * the order in which smaller blocks are delivered depends on the order
2338  * they're subdivided in this function. This is the primary factor
2339  * influencing the order in which pages are delivered to the IO
2340  * subsystem according to empirical testing, and this is also justified
2341  * by considering the behavior of a buddy system containing a single
2342  * large block of memory acted on by a series of small allocations.
2343  * This behavior is a critical factor in sglist merging's success.
2344  *
2345  * -- nyc
2346  */
expand(struct zone * zone,struct page * page,int low,int high,int migratetype)2347 static inline void expand(struct zone *zone, struct page *page,
2348 	int low, int high, int migratetype)
2349 {
2350 	unsigned long size = 1 << high;
2351 
2352 	while (high > low) {
2353 		high--;
2354 		size >>= 1;
2355 		VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2356 
2357 		/*
2358 		 * Mark as guard pages (or page), that will allow to
2359 		 * merge back to allocator when buddy will be freed.
2360 		 * Corresponding page table entries will not be touched,
2361 		 * pages will stay not present in virtual address space
2362 		 */
2363 		if (set_page_guard(zone, &page[size], high, migratetype))
2364 			continue;
2365 
2366 		add_to_free_list(&page[size], zone, high, migratetype);
2367 		set_buddy_order(&page[size], high);
2368 	}
2369 }
2370 
check_new_page_bad(struct page * page)2371 static void check_new_page_bad(struct page *page)
2372 {
2373 	if (unlikely(page->flags & __PG_HWPOISON)) {
2374 		/* Don't complain about hwpoisoned pages */
2375 		page_mapcount_reset(page); /* remove PageBuddy */
2376 		return;
2377 	}
2378 
2379 	bad_page(page,
2380 		 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2381 }
2382 
2383 /*
2384  * This page is about to be returned from the page allocator
2385  */
check_new_page(struct page * page)2386 static inline int check_new_page(struct page *page)
2387 {
2388 	if (likely(page_expected_state(page,
2389 				PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2390 		return 0;
2391 
2392 	check_new_page_bad(page);
2393 	return 1;
2394 }
2395 
check_new_pages(struct page * page,unsigned int order)2396 static bool check_new_pages(struct page *page, unsigned int order)
2397 {
2398 	int i;
2399 	for (i = 0; i < (1 << order); i++) {
2400 		struct page *p = page + i;
2401 
2402 		if (unlikely(check_new_page(p)))
2403 			return true;
2404 	}
2405 
2406 	return false;
2407 }
2408 
2409 #ifdef CONFIG_DEBUG_VM
2410 /*
2411  * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2412  * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2413  * also checked when pcp lists are refilled from the free lists.
2414  */
check_pcp_refill(struct page * page,unsigned int order)2415 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2416 {
2417 	if (debug_pagealloc_enabled_static())
2418 		return check_new_pages(page, order);
2419 	else
2420 		return false;
2421 }
2422 
check_new_pcp(struct page * page,unsigned int order)2423 static inline bool check_new_pcp(struct page *page, unsigned int order)
2424 {
2425 	return check_new_pages(page, order);
2426 }
2427 #else
2428 /*
2429  * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2430  * when pcp lists are being refilled from the free lists. With debug_pagealloc
2431  * enabled, they are also checked when being allocated from the pcp lists.
2432  */
check_pcp_refill(struct page * page,unsigned int order)2433 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2434 {
2435 	return check_new_pages(page, order);
2436 }
check_new_pcp(struct page * page,unsigned int order)2437 static inline bool check_new_pcp(struct page *page, unsigned int order)
2438 {
2439 	if (debug_pagealloc_enabled_static())
2440 		return check_new_pages(page, order);
2441 	else
2442 		return false;
2443 }
2444 #endif /* CONFIG_DEBUG_VM */
2445 
should_skip_kasan_unpoison(gfp_t flags)2446 static inline bool should_skip_kasan_unpoison(gfp_t flags)
2447 {
2448 	/* Don't skip if a software KASAN mode is enabled. */
2449 	if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2450 	    IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2451 		return false;
2452 
2453 	/* Skip, if hardware tag-based KASAN is not enabled. */
2454 	if (!kasan_hw_tags_enabled())
2455 		return true;
2456 
2457 	/*
2458 	 * With hardware tag-based KASAN enabled, skip if this has been
2459 	 * requested via __GFP_SKIP_KASAN_UNPOISON.
2460 	 */
2461 	return flags & __GFP_SKIP_KASAN_UNPOISON;
2462 }
2463 
should_skip_init(gfp_t flags)2464 static inline bool should_skip_init(gfp_t flags)
2465 {
2466 	/* Don't skip, if hardware tag-based KASAN is not enabled. */
2467 	if (!kasan_hw_tags_enabled())
2468 		return false;
2469 
2470 	/* For hardware tag-based KASAN, skip if requested. */
2471 	return (flags & __GFP_SKIP_ZERO);
2472 }
2473 
post_alloc_hook(struct page * page,unsigned int order,gfp_t gfp_flags)2474 inline void post_alloc_hook(struct page *page, unsigned int order,
2475 				gfp_t gfp_flags)
2476 {
2477 	bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2478 			!should_skip_init(gfp_flags);
2479 	bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2480 	int i;
2481 
2482 	set_page_private(page, 0);
2483 	set_page_refcounted(page);
2484 
2485 	arch_alloc_page(page, order);
2486 	debug_pagealloc_map_pages(page, 1 << order);
2487 
2488 	/*
2489 	 * Page unpoisoning must happen before memory initialization.
2490 	 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2491 	 * allocations and the page unpoisoning code will complain.
2492 	 */
2493 	kernel_unpoison_pages(page, 1 << order);
2494 
2495 	/*
2496 	 * As memory initialization might be integrated into KASAN,
2497 	 * KASAN unpoisoning and memory initializion code must be
2498 	 * kept together to avoid discrepancies in behavior.
2499 	 */
2500 
2501 	/*
2502 	 * If memory tags should be zeroed (which happens only when memory
2503 	 * should be initialized as well).
2504 	 */
2505 	if (init_tags) {
2506 		/* Initialize both memory and tags. */
2507 		for (i = 0; i != 1 << order; ++i)
2508 			tag_clear_highpage(page + i);
2509 
2510 		/* Note that memory is already initialized by the loop above. */
2511 		init = false;
2512 	}
2513 	if (!should_skip_kasan_unpoison(gfp_flags)) {
2514 		/* Unpoison shadow memory or set memory tags. */
2515 		kasan_unpoison_pages(page, order, init);
2516 
2517 		/* Note that memory is already initialized by KASAN. */
2518 		if (kasan_has_integrated_init())
2519 			init = false;
2520 	} else {
2521 		/* Ensure page_address() dereferencing does not fault. */
2522 		for (i = 0; i != 1 << order; ++i)
2523 			page_kasan_tag_reset(page + i);
2524 	}
2525 	/* If memory is still not initialized, do it now. */
2526 	if (init)
2527 		kernel_init_pages(page, 1 << order);
2528 	/* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2529 	if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2530 		SetPageSkipKASanPoison(page);
2531 
2532 	set_page_owner(page, order, gfp_flags);
2533 	page_table_check_alloc(page, order);
2534 }
2535 
prep_new_page(struct page * page,unsigned int order,gfp_t gfp_flags,unsigned int alloc_flags)2536 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2537 							unsigned int alloc_flags)
2538 {
2539 	post_alloc_hook(page, order, gfp_flags);
2540 
2541 	if (order && (gfp_flags & __GFP_COMP))
2542 		prep_compound_page(page, order);
2543 
2544 	/*
2545 	 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2546 	 * allocate the page. The expectation is that the caller is taking
2547 	 * steps that will free more memory. The caller should avoid the page
2548 	 * being used for !PFMEMALLOC purposes.
2549 	 */
2550 	if (alloc_flags & ALLOC_NO_WATERMARKS)
2551 		set_page_pfmemalloc(page);
2552 	else
2553 		clear_page_pfmemalloc(page);
2554 }
2555 
2556 /*
2557  * Go through the free lists for the given migratetype and remove
2558  * the smallest available page from the freelists
2559  */
2560 static __always_inline
__rmqueue_smallest(struct zone * zone,unsigned int order,int migratetype)2561 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2562 						int migratetype)
2563 {
2564 	unsigned int current_order;
2565 	struct free_area *area;
2566 	struct page *page;
2567 
2568 	/* Find a page of the appropriate size in the preferred list */
2569 	for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2570 		area = &(zone->free_area[current_order]);
2571 		page = get_page_from_free_area(area, migratetype);
2572 		if (!page)
2573 			continue;
2574 		del_page_from_free_list(page, zone, current_order);
2575 		expand(zone, page, order, current_order, migratetype);
2576 		set_pcppage_migratetype(page, migratetype);
2577 		trace_mm_page_alloc_zone_locked(page, order, migratetype,
2578 				pcp_allowed_order(order) &&
2579 				migratetype < MIGRATE_PCPTYPES);
2580 		return page;
2581 	}
2582 
2583 	return NULL;
2584 }
2585 
2586 
2587 /*
2588  * This array describes the order lists are fallen back to when
2589  * the free lists for the desirable migrate type are depleted
2590  *
2591  * The other migratetypes do not have fallbacks.
2592  */
2593 static int fallbacks[MIGRATE_TYPES][3] = {
2594 	[MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_TYPES },
2595 	[MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2596 	[MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_TYPES },
2597 };
2598 
2599 #ifdef CONFIG_CMA
__rmqueue_cma_fallback(struct zone * zone,unsigned int order)2600 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2601 					unsigned int order)
2602 {
2603 	return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2604 }
2605 #else
__rmqueue_cma_fallback(struct zone * zone,unsigned int order)2606 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2607 					unsigned int order) { return NULL; }
2608 #endif
2609 
2610 /*
2611  * Move the free pages in a range to the freelist tail of the requested type.
2612  * Note that start_page and end_pages are not aligned on a pageblock
2613  * boundary. If alignment is required, use move_freepages_block()
2614  */
move_freepages(struct zone * zone,unsigned long start_pfn,unsigned long end_pfn,int migratetype,int * num_movable)2615 static int move_freepages(struct zone *zone,
2616 			  unsigned long start_pfn, unsigned long end_pfn,
2617 			  int migratetype, int *num_movable)
2618 {
2619 	struct page *page;
2620 	unsigned long pfn;
2621 	unsigned int order;
2622 	int pages_moved = 0;
2623 
2624 	for (pfn = start_pfn; pfn <= end_pfn;) {
2625 		page = pfn_to_page(pfn);
2626 		if (!PageBuddy(page)) {
2627 			/*
2628 			 * We assume that pages that could be isolated for
2629 			 * migration are movable. But we don't actually try
2630 			 * isolating, as that would be expensive.
2631 			 */
2632 			if (num_movable &&
2633 					(PageLRU(page) || __PageMovable(page)))
2634 				(*num_movable)++;
2635 			pfn++;
2636 			continue;
2637 		}
2638 
2639 		/* Make sure we are not inadvertently changing nodes */
2640 		VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2641 		VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2642 
2643 		order = buddy_order(page);
2644 		move_to_free_list(page, zone, order, migratetype);
2645 		pfn += 1 << order;
2646 		pages_moved += 1 << order;
2647 	}
2648 
2649 	return pages_moved;
2650 }
2651 
move_freepages_block(struct zone * zone,struct page * page,int migratetype,int * num_movable)2652 int move_freepages_block(struct zone *zone, struct page *page,
2653 				int migratetype, int *num_movable)
2654 {
2655 	unsigned long start_pfn, end_pfn, pfn;
2656 
2657 	if (num_movable)
2658 		*num_movable = 0;
2659 
2660 	pfn = page_to_pfn(page);
2661 	start_pfn = pageblock_start_pfn(pfn);
2662 	end_pfn = pageblock_end_pfn(pfn) - 1;
2663 
2664 	/* Do not cross zone boundaries */
2665 	if (!zone_spans_pfn(zone, start_pfn))
2666 		start_pfn = pfn;
2667 	if (!zone_spans_pfn(zone, end_pfn))
2668 		return 0;
2669 
2670 	return move_freepages(zone, start_pfn, end_pfn, migratetype,
2671 								num_movable);
2672 }
2673 
change_pageblock_range(struct page * pageblock_page,int start_order,int migratetype)2674 static void change_pageblock_range(struct page *pageblock_page,
2675 					int start_order, int migratetype)
2676 {
2677 	int nr_pageblocks = 1 << (start_order - pageblock_order);
2678 
2679 	while (nr_pageblocks--) {
2680 		set_pageblock_migratetype(pageblock_page, migratetype);
2681 		pageblock_page += pageblock_nr_pages;
2682 	}
2683 }
2684 
2685 /*
2686  * When we are falling back to another migratetype during allocation, try to
2687  * steal extra free pages from the same pageblocks to satisfy further
2688  * allocations, instead of polluting multiple pageblocks.
2689  *
2690  * If we are stealing a relatively large buddy page, it is likely there will
2691  * be more free pages in the pageblock, so try to steal them all. For
2692  * reclaimable and unmovable allocations, we steal regardless of page size,
2693  * as fragmentation caused by those allocations polluting movable pageblocks
2694  * is worse than movable allocations stealing from unmovable and reclaimable
2695  * pageblocks.
2696  */
can_steal_fallback(unsigned int order,int start_mt)2697 static bool can_steal_fallback(unsigned int order, int start_mt)
2698 {
2699 	/*
2700 	 * Leaving this order check is intended, although there is
2701 	 * relaxed order check in next check. The reason is that
2702 	 * we can actually steal whole pageblock if this condition met,
2703 	 * but, below check doesn't guarantee it and that is just heuristic
2704 	 * so could be changed anytime.
2705 	 */
2706 	if (order >= pageblock_order)
2707 		return true;
2708 
2709 	if (order >= pageblock_order / 2 ||
2710 		start_mt == MIGRATE_RECLAIMABLE ||
2711 		start_mt == MIGRATE_UNMOVABLE ||
2712 		page_group_by_mobility_disabled)
2713 		return true;
2714 
2715 	return false;
2716 }
2717 
boost_watermark(struct zone * zone)2718 static inline bool boost_watermark(struct zone *zone)
2719 {
2720 	unsigned long max_boost;
2721 
2722 	if (!watermark_boost_factor)
2723 		return false;
2724 	/*
2725 	 * Don't bother in zones that are unlikely to produce results.
2726 	 * On small machines, including kdump capture kernels running
2727 	 * in a small area, boosting the watermark can cause an out of
2728 	 * memory situation immediately.
2729 	 */
2730 	if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2731 		return false;
2732 
2733 	max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2734 			watermark_boost_factor, 10000);
2735 
2736 	/*
2737 	 * high watermark may be uninitialised if fragmentation occurs
2738 	 * very early in boot so do not boost. We do not fall
2739 	 * through and boost by pageblock_nr_pages as failing
2740 	 * allocations that early means that reclaim is not going
2741 	 * to help and it may even be impossible to reclaim the
2742 	 * boosted watermark resulting in a hang.
2743 	 */
2744 	if (!max_boost)
2745 		return false;
2746 
2747 	max_boost = max(pageblock_nr_pages, max_boost);
2748 
2749 	zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2750 		max_boost);
2751 
2752 	return true;
2753 }
2754 
2755 /*
2756  * This function implements actual steal behaviour. If order is large enough,
2757  * we can steal whole pageblock. If not, we first move freepages in this
2758  * pageblock to our migratetype and determine how many already-allocated pages
2759  * are there in the pageblock with a compatible migratetype. If at least half
2760  * of pages are free or compatible, we can change migratetype of the pageblock
2761  * itself, so pages freed in the future will be put on the correct free list.
2762  */
steal_suitable_fallback(struct zone * zone,struct page * page,unsigned int alloc_flags,int start_type,bool whole_block)2763 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2764 		unsigned int alloc_flags, int start_type, bool whole_block)
2765 {
2766 	unsigned int current_order = buddy_order(page);
2767 	int free_pages, movable_pages, alike_pages;
2768 	int old_block_type;
2769 
2770 	old_block_type = get_pageblock_migratetype(page);
2771 
2772 	/*
2773 	 * This can happen due to races and we want to prevent broken
2774 	 * highatomic accounting.
2775 	 */
2776 	if (is_migrate_highatomic(old_block_type))
2777 		goto single_page;
2778 
2779 	/* Take ownership for orders >= pageblock_order */
2780 	if (current_order >= pageblock_order) {
2781 		change_pageblock_range(page, current_order, start_type);
2782 		goto single_page;
2783 	}
2784 
2785 	/*
2786 	 * Boost watermarks to increase reclaim pressure to reduce the
2787 	 * likelihood of future fallbacks. Wake kswapd now as the node
2788 	 * may be balanced overall and kswapd will not wake naturally.
2789 	 */
2790 	if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2791 		set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2792 
2793 	/* We are not allowed to try stealing from the whole block */
2794 	if (!whole_block)
2795 		goto single_page;
2796 
2797 	free_pages = move_freepages_block(zone, page, start_type,
2798 						&movable_pages);
2799 	/*
2800 	 * Determine how many pages are compatible with our allocation.
2801 	 * For movable allocation, it's the number of movable pages which
2802 	 * we just obtained. For other types it's a bit more tricky.
2803 	 */
2804 	if (start_type == MIGRATE_MOVABLE) {
2805 		alike_pages = movable_pages;
2806 	} else {
2807 		/*
2808 		 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2809 		 * to MOVABLE pageblock, consider all non-movable pages as
2810 		 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2811 		 * vice versa, be conservative since we can't distinguish the
2812 		 * exact migratetype of non-movable pages.
2813 		 */
2814 		if (old_block_type == MIGRATE_MOVABLE)
2815 			alike_pages = pageblock_nr_pages
2816 						- (free_pages + movable_pages);
2817 		else
2818 			alike_pages = 0;
2819 	}
2820 
2821 	/* moving whole block can fail due to zone boundary conditions */
2822 	if (!free_pages)
2823 		goto single_page;
2824 
2825 	/*
2826 	 * If a sufficient number of pages in the block are either free or of
2827 	 * comparable migratability as our allocation, claim the whole block.
2828 	 */
2829 	if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2830 			page_group_by_mobility_disabled)
2831 		set_pageblock_migratetype(page, start_type);
2832 
2833 	return;
2834 
2835 single_page:
2836 	move_to_free_list(page, zone, current_order, start_type);
2837 }
2838 
2839 /*
2840  * Check whether there is a suitable fallback freepage with requested order.
2841  * If only_stealable is true, this function returns fallback_mt only if
2842  * we can steal other freepages all together. This would help to reduce
2843  * fragmentation due to mixed migratetype pages in one pageblock.
2844  */
find_suitable_fallback(struct free_area * area,unsigned int order,int migratetype,bool only_stealable,bool * can_steal)2845 int find_suitable_fallback(struct free_area *area, unsigned int order,
2846 			int migratetype, bool only_stealable, bool *can_steal)
2847 {
2848 	int i;
2849 	int fallback_mt;
2850 
2851 	if (area->nr_free == 0)
2852 		return -1;
2853 
2854 	*can_steal = false;
2855 	for (i = 0;; i++) {
2856 		fallback_mt = fallbacks[migratetype][i];
2857 		if (fallback_mt == MIGRATE_TYPES)
2858 			break;
2859 
2860 		if (free_area_empty(area, fallback_mt))
2861 			continue;
2862 
2863 		if (can_steal_fallback(order, migratetype))
2864 			*can_steal = true;
2865 
2866 		if (!only_stealable)
2867 			return fallback_mt;
2868 
2869 		if (*can_steal)
2870 			return fallback_mt;
2871 	}
2872 
2873 	return -1;
2874 }
2875 
2876 /*
2877  * Reserve a pageblock for exclusive use of high-order atomic allocations if
2878  * there are no empty page blocks that contain a page with a suitable order
2879  */
reserve_highatomic_pageblock(struct page * page,struct zone * zone,unsigned int alloc_order)2880 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2881 				unsigned int alloc_order)
2882 {
2883 	int mt;
2884 	unsigned long max_managed, flags;
2885 
2886 	/*
2887 	 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2888 	 * Check is race-prone but harmless.
2889 	 */
2890 	max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2891 	if (zone->nr_reserved_highatomic >= max_managed)
2892 		return;
2893 
2894 	spin_lock_irqsave(&zone->lock, flags);
2895 
2896 	/* Recheck the nr_reserved_highatomic limit under the lock */
2897 	if (zone->nr_reserved_highatomic >= max_managed)
2898 		goto out_unlock;
2899 
2900 	/* Yoink! */
2901 	mt = get_pageblock_migratetype(page);
2902 	/* Only reserve normal pageblocks (i.e., they can merge with others) */
2903 	if (migratetype_is_mergeable(mt)) {
2904 		zone->nr_reserved_highatomic += pageblock_nr_pages;
2905 		set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2906 		move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2907 	}
2908 
2909 out_unlock:
2910 	spin_unlock_irqrestore(&zone->lock, flags);
2911 }
2912 
2913 /*
2914  * Used when an allocation is about to fail under memory pressure. This
2915  * potentially hurts the reliability of high-order allocations when under
2916  * intense memory pressure but failed atomic allocations should be easier
2917  * to recover from than an OOM.
2918  *
2919  * If @force is true, try to unreserve a pageblock even though highatomic
2920  * pageblock is exhausted.
2921  */
unreserve_highatomic_pageblock(const struct alloc_context * ac,bool force)2922 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2923 						bool force)
2924 {
2925 	struct zonelist *zonelist = ac->zonelist;
2926 	unsigned long flags;
2927 	struct zoneref *z;
2928 	struct zone *zone;
2929 	struct page *page;
2930 	int order;
2931 	bool ret;
2932 
2933 	for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2934 								ac->nodemask) {
2935 		/*
2936 		 * Preserve at least one pageblock unless memory pressure
2937 		 * is really high.
2938 		 */
2939 		if (!force && zone->nr_reserved_highatomic <=
2940 					pageblock_nr_pages)
2941 			continue;
2942 
2943 		spin_lock_irqsave(&zone->lock, flags);
2944 		for (order = 0; order < MAX_ORDER; order++) {
2945 			struct free_area *area = &(zone->free_area[order]);
2946 
2947 			page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2948 			if (!page)
2949 				continue;
2950 
2951 			/*
2952 			 * In page freeing path, migratetype change is racy so
2953 			 * we can counter several free pages in a pageblock
2954 			 * in this loop although we changed the pageblock type
2955 			 * from highatomic to ac->migratetype. So we should
2956 			 * adjust the count once.
2957 			 */
2958 			if (is_migrate_highatomic_page(page)) {
2959 				/*
2960 				 * It should never happen but changes to
2961 				 * locking could inadvertently allow a per-cpu
2962 				 * drain to add pages to MIGRATE_HIGHATOMIC
2963 				 * while unreserving so be safe and watch for
2964 				 * underflows.
2965 				 */
2966 				zone->nr_reserved_highatomic -= min(
2967 						pageblock_nr_pages,
2968 						zone->nr_reserved_highatomic);
2969 			}
2970 
2971 			/*
2972 			 * Convert to ac->migratetype and avoid the normal
2973 			 * pageblock stealing heuristics. Minimally, the caller
2974 			 * is doing the work and needs the pages. More
2975 			 * importantly, if the block was always converted to
2976 			 * MIGRATE_UNMOVABLE or another type then the number
2977 			 * of pageblocks that cannot be completely freed
2978 			 * may increase.
2979 			 */
2980 			set_pageblock_migratetype(page, ac->migratetype);
2981 			ret = move_freepages_block(zone, page, ac->migratetype,
2982 									NULL);
2983 			if (ret) {
2984 				spin_unlock_irqrestore(&zone->lock, flags);
2985 				return ret;
2986 			}
2987 		}
2988 		spin_unlock_irqrestore(&zone->lock, flags);
2989 	}
2990 
2991 	return false;
2992 }
2993 
2994 /*
2995  * Try finding a free buddy page on the fallback list and put it on the free
2996  * list of requested migratetype, possibly along with other pages from the same
2997  * block, depending on fragmentation avoidance heuristics. Returns true if
2998  * fallback was found so that __rmqueue_smallest() can grab it.
2999  *
3000  * The use of signed ints for order and current_order is a deliberate
3001  * deviation from the rest of this file, to make the for loop
3002  * condition simpler.
3003  */
3004 static __always_inline bool
__rmqueue_fallback(struct zone * zone,int order,int start_migratetype,unsigned int alloc_flags)3005 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
3006 						unsigned int alloc_flags)
3007 {
3008 	struct free_area *area;
3009 	int current_order;
3010 	int min_order = order;
3011 	struct page *page;
3012 	int fallback_mt;
3013 	bool can_steal;
3014 
3015 	/*
3016 	 * Do not steal pages from freelists belonging to other pageblocks
3017 	 * i.e. orders < pageblock_order. If there are no local zones free,
3018 	 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3019 	 */
3020 	if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
3021 		min_order = pageblock_order;
3022 
3023 	/*
3024 	 * Find the largest available free page in the other list. This roughly
3025 	 * approximates finding the pageblock with the most free pages, which
3026 	 * would be too costly to do exactly.
3027 	 */
3028 	for (current_order = MAX_ORDER - 1; current_order >= min_order;
3029 				--current_order) {
3030 		area = &(zone->free_area[current_order]);
3031 		fallback_mt = find_suitable_fallback(area, current_order,
3032 				start_migratetype, false, &can_steal);
3033 		if (fallback_mt == -1)
3034 			continue;
3035 
3036 		/*
3037 		 * We cannot steal all free pages from the pageblock and the
3038 		 * requested migratetype is movable. In that case it's better to
3039 		 * steal and split the smallest available page instead of the
3040 		 * largest available page, because even if the next movable
3041 		 * allocation falls back into a different pageblock than this
3042 		 * one, it won't cause permanent fragmentation.
3043 		 */
3044 		if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3045 					&& current_order > order)
3046 			goto find_smallest;
3047 
3048 		goto do_steal;
3049 	}
3050 
3051 	return false;
3052 
3053 find_smallest:
3054 	for (current_order = order; current_order < MAX_ORDER;
3055 							current_order++) {
3056 		area = &(zone->free_area[current_order]);
3057 		fallback_mt = find_suitable_fallback(area, current_order,
3058 				start_migratetype, false, &can_steal);
3059 		if (fallback_mt != -1)
3060 			break;
3061 	}
3062 
3063 	/*
3064 	 * This should not happen - we already found a suitable fallback
3065 	 * when looking for the largest page.
3066 	 */
3067 	VM_BUG_ON(current_order == MAX_ORDER);
3068 
3069 do_steal:
3070 	page = get_page_from_free_area(area, fallback_mt);
3071 
3072 	steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3073 								can_steal);
3074 
3075 	trace_mm_page_alloc_extfrag(page, order, current_order,
3076 		start_migratetype, fallback_mt);
3077 
3078 	return true;
3079 
3080 }
3081 
3082 /*
3083  * Do the hard work of removing an element from the buddy allocator.
3084  * Call me with the zone->lock already held.
3085  */
3086 static __always_inline struct page *
__rmqueue(struct zone * zone,unsigned int order,int migratetype,unsigned int alloc_flags)3087 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3088 						unsigned int alloc_flags)
3089 {
3090 	struct page *page;
3091 
3092 	if (IS_ENABLED(CONFIG_CMA)) {
3093 		/*
3094 		 * Balance movable allocations between regular and CMA areas by
3095 		 * allocating from CMA when over half of the zone's free memory
3096 		 * is in the CMA area.
3097 		 */
3098 		if (alloc_flags & ALLOC_CMA &&
3099 		    zone_page_state(zone, NR_FREE_CMA_PAGES) >
3100 		    zone_page_state(zone, NR_FREE_PAGES) / 2) {
3101 			page = __rmqueue_cma_fallback(zone, order);
3102 			if (page)
3103 				return page;
3104 		}
3105 	}
3106 retry:
3107 	page = __rmqueue_smallest(zone, order, migratetype);
3108 	if (unlikely(!page)) {
3109 		if (alloc_flags & ALLOC_CMA)
3110 			page = __rmqueue_cma_fallback(zone, order);
3111 
3112 		if (!page && __rmqueue_fallback(zone, order, migratetype,
3113 								alloc_flags))
3114 			goto retry;
3115 	}
3116 	return page;
3117 }
3118 
3119 /*
3120  * Obtain a specified number of elements from the buddy allocator, all under
3121  * a single hold of the lock, for efficiency.  Add them to the supplied list.
3122  * Returns the number of new pages which were placed at *list.
3123  */
rmqueue_bulk(struct zone * zone,unsigned int order,unsigned long count,struct list_head * list,int migratetype,unsigned int alloc_flags)3124 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3125 			unsigned long count, struct list_head *list,
3126 			int migratetype, unsigned int alloc_flags)
3127 {
3128 	int i, allocated = 0;
3129 
3130 	/* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
3131 	spin_lock(&zone->lock);
3132 	for (i = 0; i < count; ++i) {
3133 		struct page *page = __rmqueue(zone, order, migratetype,
3134 								alloc_flags);
3135 		if (unlikely(page == NULL))
3136 			break;
3137 
3138 		if (unlikely(check_pcp_refill(page, order)))
3139 			continue;
3140 
3141 		/*
3142 		 * Split buddy pages returned by expand() are received here in
3143 		 * physical page order. The page is added to the tail of
3144 		 * caller's list. From the callers perspective, the linked list
3145 		 * is ordered by page number under some conditions. This is
3146 		 * useful for IO devices that can forward direction from the
3147 		 * head, thus also in the physical page order. This is useful
3148 		 * for IO devices that can merge IO requests if the physical
3149 		 * pages are ordered properly.
3150 		 */
3151 		list_add_tail(&page->pcp_list, list);
3152 		allocated++;
3153 		if (is_migrate_cma(get_pcppage_migratetype(page)))
3154 			__mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3155 					      -(1 << order));
3156 	}
3157 
3158 	/*
3159 	 * i pages were removed from the buddy list even if some leak due
3160 	 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3161 	 * on i. Do not confuse with 'allocated' which is the number of
3162 	 * pages added to the pcp list.
3163 	 */
3164 	__mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3165 	spin_unlock(&zone->lock);
3166 	return allocated;
3167 }
3168 
3169 #ifdef CONFIG_NUMA
3170 /*
3171  * Called from the vmstat counter updater to drain pagesets of this
3172  * currently executing processor on remote nodes after they have
3173  * expired.
3174  */
drain_zone_pages(struct zone * zone,struct per_cpu_pages * pcp)3175 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3176 {
3177 	int to_drain, batch;
3178 
3179 	batch = READ_ONCE(pcp->batch);
3180 	to_drain = min(pcp->count, batch);
3181 	if (to_drain > 0) {
3182 		unsigned long flags;
3183 
3184 		/*
3185 		 * free_pcppages_bulk expects IRQs disabled for zone->lock
3186 		 * so even though pcp->lock is not intended to be IRQ-safe,
3187 		 * it's needed in this context.
3188 		 */
3189 		spin_lock_irqsave(&pcp->lock, flags);
3190 		free_pcppages_bulk(zone, to_drain, pcp, 0);
3191 		spin_unlock_irqrestore(&pcp->lock, flags);
3192 	}
3193 }
3194 #endif
3195 
3196 /*
3197  * Drain pcplists of the indicated processor and zone.
3198  */
drain_pages_zone(unsigned int cpu,struct zone * zone)3199 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3200 {
3201 	struct per_cpu_pages *pcp;
3202 
3203 	pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3204 	if (pcp->count) {
3205 		unsigned long flags;
3206 
3207 		/* See drain_zone_pages on why this is disabling IRQs */
3208 		spin_lock_irqsave(&pcp->lock, flags);
3209 		free_pcppages_bulk(zone, pcp->count, pcp, 0);
3210 		spin_unlock_irqrestore(&pcp->lock, flags);
3211 	}
3212 }
3213 
3214 /*
3215  * Drain pcplists of all zones on the indicated processor.
3216  */
drain_pages(unsigned int cpu)3217 static void drain_pages(unsigned int cpu)
3218 {
3219 	struct zone *zone;
3220 
3221 	for_each_populated_zone(zone) {
3222 		drain_pages_zone(cpu, zone);
3223 	}
3224 }
3225 
3226 /*
3227  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3228  */
drain_local_pages(struct zone * zone)3229 void drain_local_pages(struct zone *zone)
3230 {
3231 	int cpu = smp_processor_id();
3232 
3233 	if (zone)
3234 		drain_pages_zone(cpu, zone);
3235 	else
3236 		drain_pages(cpu);
3237 }
3238 
3239 /*
3240  * The implementation of drain_all_pages(), exposing an extra parameter to
3241  * drain on all cpus.
3242  *
3243  * drain_all_pages() is optimized to only execute on cpus where pcplists are
3244  * not empty. The check for non-emptiness can however race with a free to
3245  * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3246  * that need the guarantee that every CPU has drained can disable the
3247  * optimizing racy check.
3248  */
__drain_all_pages(struct zone * zone,bool force_all_cpus)3249 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3250 {
3251 	int cpu;
3252 
3253 	/*
3254 	 * Allocate in the BSS so we won't require allocation in
3255 	 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3256 	 */
3257 	static cpumask_t cpus_with_pcps;
3258 
3259 	/*
3260 	 * Do not drain if one is already in progress unless it's specific to
3261 	 * a zone. Such callers are primarily CMA and memory hotplug and need
3262 	 * the drain to be complete when the call returns.
3263 	 */
3264 	if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3265 		if (!zone)
3266 			return;
3267 		mutex_lock(&pcpu_drain_mutex);
3268 	}
3269 
3270 	/*
3271 	 * We don't care about racing with CPU hotplug event
3272 	 * as offline notification will cause the notified
3273 	 * cpu to drain that CPU pcps and on_each_cpu_mask
3274 	 * disables preemption as part of its processing
3275 	 */
3276 	for_each_online_cpu(cpu) {
3277 		struct per_cpu_pages *pcp;
3278 		struct zone *z;
3279 		bool has_pcps = false;
3280 
3281 		if (force_all_cpus) {
3282 			/*
3283 			 * The pcp.count check is racy, some callers need a
3284 			 * guarantee that no cpu is missed.
3285 			 */
3286 			has_pcps = true;
3287 		} else if (zone) {
3288 			pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3289 			if (pcp->count)
3290 				has_pcps = true;
3291 		} else {
3292 			for_each_populated_zone(z) {
3293 				pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3294 				if (pcp->count) {
3295 					has_pcps = true;
3296 					break;
3297 				}
3298 			}
3299 		}
3300 
3301 		if (has_pcps)
3302 			cpumask_set_cpu(cpu, &cpus_with_pcps);
3303 		else
3304 			cpumask_clear_cpu(cpu, &cpus_with_pcps);
3305 	}
3306 
3307 	for_each_cpu(cpu, &cpus_with_pcps) {
3308 		if (zone)
3309 			drain_pages_zone(cpu, zone);
3310 		else
3311 			drain_pages(cpu);
3312 	}
3313 
3314 	mutex_unlock(&pcpu_drain_mutex);
3315 }
3316 
3317 /*
3318  * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3319  *
3320  * When zone parameter is non-NULL, spill just the single zone's pages.
3321  */
drain_all_pages(struct zone * zone)3322 void drain_all_pages(struct zone *zone)
3323 {
3324 	__drain_all_pages(zone, false);
3325 }
3326 
3327 #ifdef CONFIG_HIBERNATION
3328 
3329 /*
3330  * Touch the watchdog for every WD_PAGE_COUNT pages.
3331  */
3332 #define WD_PAGE_COUNT	(128*1024)
3333 
mark_free_pages(struct zone * zone)3334 void mark_free_pages(struct zone *zone)
3335 {
3336 	unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3337 	unsigned long flags;
3338 	unsigned int order, t;
3339 	struct page *page;
3340 
3341 	if (zone_is_empty(zone))
3342 		return;
3343 
3344 	spin_lock_irqsave(&zone->lock, flags);
3345 
3346 	max_zone_pfn = zone_end_pfn(zone);
3347 	for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3348 		if (pfn_valid(pfn)) {
3349 			page = pfn_to_page(pfn);
3350 
3351 			if (!--page_count) {
3352 				touch_nmi_watchdog();
3353 				page_count = WD_PAGE_COUNT;
3354 			}
3355 
3356 			if (page_zone(page) != zone)
3357 				continue;
3358 
3359 			if (!swsusp_page_is_forbidden(page))
3360 				swsusp_unset_page_free(page);
3361 		}
3362 
3363 	for_each_migratetype_order(order, t) {
3364 		list_for_each_entry(page,
3365 				&zone->free_area[order].free_list[t], buddy_list) {
3366 			unsigned long i;
3367 
3368 			pfn = page_to_pfn(page);
3369 			for (i = 0; i < (1UL << order); i++) {
3370 				if (!--page_count) {
3371 					touch_nmi_watchdog();
3372 					page_count = WD_PAGE_COUNT;
3373 				}
3374 				swsusp_set_page_free(pfn_to_page(pfn + i));
3375 			}
3376 		}
3377 	}
3378 	spin_unlock_irqrestore(&zone->lock, flags);
3379 }
3380 #endif /* CONFIG_PM */
3381 
free_unref_page_prepare(struct page * page,unsigned long pfn,unsigned int order)3382 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3383 							unsigned int order)
3384 {
3385 	int migratetype;
3386 
3387 	if (!free_pcp_prepare(page, order))
3388 		return false;
3389 
3390 	migratetype = get_pfnblock_migratetype(page, pfn);
3391 	set_pcppage_migratetype(page, migratetype);
3392 	return true;
3393 }
3394 
nr_pcp_free(struct per_cpu_pages * pcp,int high,int batch,bool free_high)3395 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3396 		       bool free_high)
3397 {
3398 	int min_nr_free, max_nr_free;
3399 
3400 	/* Free everything if batch freeing high-order pages. */
3401 	if (unlikely(free_high))
3402 		return pcp->count;
3403 
3404 	/* Check for PCP disabled or boot pageset */
3405 	if (unlikely(high < batch))
3406 		return 1;
3407 
3408 	/* Leave at least pcp->batch pages on the list */
3409 	min_nr_free = batch;
3410 	max_nr_free = high - batch;
3411 
3412 	/*
3413 	 * Double the number of pages freed each time there is subsequent
3414 	 * freeing of pages without any allocation.
3415 	 */
3416 	batch <<= pcp->free_factor;
3417 	if (batch < max_nr_free)
3418 		pcp->free_factor++;
3419 	batch = clamp(batch, min_nr_free, max_nr_free);
3420 
3421 	return batch;
3422 }
3423 
nr_pcp_high(struct per_cpu_pages * pcp,struct zone * zone,bool free_high)3424 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3425 		       bool free_high)
3426 {
3427 	int high = READ_ONCE(pcp->high);
3428 
3429 	if (unlikely(!high || free_high))
3430 		return 0;
3431 
3432 	if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3433 		return high;
3434 
3435 	/*
3436 	 * If reclaim is active, limit the number of pages that can be
3437 	 * stored on pcp lists
3438 	 */
3439 	return min(READ_ONCE(pcp->batch) << 2, high);
3440 }
3441 
free_unref_page_commit(struct zone * zone,struct per_cpu_pages * pcp,struct page * page,int migratetype,unsigned int order)3442 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3443 				   struct page *page, int migratetype,
3444 				   unsigned int order)
3445 {
3446 	int high;
3447 	int pindex;
3448 	bool free_high;
3449 
3450 	__count_vm_events(PGFREE, 1 << order);
3451 	pindex = order_to_pindex(migratetype, order);
3452 	list_add(&page->pcp_list, &pcp->lists[pindex]);
3453 	pcp->count += 1 << order;
3454 
3455 	/*
3456 	 * As high-order pages other than THP's stored on PCP can contribute
3457 	 * to fragmentation, limit the number stored when PCP is heavily
3458 	 * freeing without allocation. The remainder after bulk freeing
3459 	 * stops will be drained from vmstat refresh context.
3460 	 */
3461 	free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3462 
3463 	high = nr_pcp_high(pcp, zone, free_high);
3464 	if (pcp->count >= high) {
3465 		int batch = READ_ONCE(pcp->batch);
3466 
3467 		free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3468 	}
3469 }
3470 
3471 /*
3472  * Free a pcp page
3473  */
free_unref_page(struct page * page,unsigned int order)3474 void free_unref_page(struct page *page, unsigned int order)
3475 {
3476 	unsigned long flags;
3477 	unsigned long __maybe_unused UP_flags;
3478 	struct per_cpu_pages *pcp;
3479 	struct zone *zone;
3480 	unsigned long pfn = page_to_pfn(page);
3481 	int migratetype;
3482 
3483 	if (!free_unref_page_prepare(page, pfn, order))
3484 		return;
3485 
3486 	/*
3487 	 * We only track unmovable, reclaimable and movable on pcp lists.
3488 	 * Place ISOLATE pages on the isolated list because they are being
3489 	 * offlined but treat HIGHATOMIC as movable pages so we can get those
3490 	 * areas back if necessary. Otherwise, we may have to free
3491 	 * excessively into the page allocator
3492 	 */
3493 	migratetype = get_pcppage_migratetype(page);
3494 	if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3495 		if (unlikely(is_migrate_isolate(migratetype))) {
3496 			free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3497 			return;
3498 		}
3499 		migratetype = MIGRATE_MOVABLE;
3500 	}
3501 
3502 	zone = page_zone(page);
3503 	pcp_trylock_prepare(UP_flags);
3504 	pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3505 	if (pcp) {
3506 		free_unref_page_commit(zone, pcp, page, migratetype, order);
3507 		pcp_spin_unlock_irqrestore(pcp, flags);
3508 	} else {
3509 		free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3510 	}
3511 	pcp_trylock_finish(UP_flags);
3512 }
3513 
3514 /*
3515  * Free a list of 0-order pages
3516  */
free_unref_page_list(struct list_head * list)3517 void free_unref_page_list(struct list_head *list)
3518 {
3519 	struct page *page, *next;
3520 	struct per_cpu_pages *pcp = NULL;
3521 	struct zone *locked_zone = NULL;
3522 	unsigned long flags;
3523 	int batch_count = 0;
3524 	int migratetype;
3525 
3526 	/* Prepare pages for freeing */
3527 	list_for_each_entry_safe(page, next, list, lru) {
3528 		unsigned long pfn = page_to_pfn(page);
3529 		if (!free_unref_page_prepare(page, pfn, 0)) {
3530 			list_del(&page->lru);
3531 			continue;
3532 		}
3533 
3534 		/*
3535 		 * Free isolated pages directly to the allocator, see
3536 		 * comment in free_unref_page.
3537 		 */
3538 		migratetype = get_pcppage_migratetype(page);
3539 		if (unlikely(is_migrate_isolate(migratetype))) {
3540 			list_del(&page->lru);
3541 			free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3542 			continue;
3543 		}
3544 	}
3545 
3546 	list_for_each_entry_safe(page, next, list, lru) {
3547 		struct zone *zone = page_zone(page);
3548 
3549 		/* Different zone, different pcp lock. */
3550 		if (zone != locked_zone) {
3551 			if (pcp)
3552 				pcp_spin_unlock_irqrestore(pcp, flags);
3553 
3554 			locked_zone = zone;
3555 			pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3556 		}
3557 
3558 		/*
3559 		 * Non-isolated types over MIGRATE_PCPTYPES get added
3560 		 * to the MIGRATE_MOVABLE pcp list.
3561 		 */
3562 		migratetype = get_pcppage_migratetype(page);
3563 		if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3564 			migratetype = MIGRATE_MOVABLE;
3565 
3566 		trace_mm_page_free_batched(page);
3567 		free_unref_page_commit(zone, pcp, page, migratetype, 0);
3568 
3569 		/*
3570 		 * Guard against excessive IRQ disabled times when we get
3571 		 * a large list of pages to free.
3572 		 */
3573 		if (++batch_count == SWAP_CLUSTER_MAX) {
3574 			pcp_spin_unlock_irqrestore(pcp, flags);
3575 			batch_count = 0;
3576 			pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3577 		}
3578 	}
3579 
3580 	if (pcp)
3581 		pcp_spin_unlock_irqrestore(pcp, flags);
3582 }
3583 
3584 /*
3585  * split_page takes a non-compound higher-order page, and splits it into
3586  * n (1<<order) sub-pages: page[0..n]
3587  * Each sub-page must be freed individually.
3588  *
3589  * Note: this is probably too low level an operation for use in drivers.
3590  * Please consult with lkml before using this in your driver.
3591  */
split_page(struct page * page,unsigned int order)3592 void split_page(struct page *page, unsigned int order)
3593 {
3594 	int i;
3595 
3596 	VM_BUG_ON_PAGE(PageCompound(page), page);
3597 	VM_BUG_ON_PAGE(!page_count(page), page);
3598 
3599 	for (i = 1; i < (1 << order); i++)
3600 		set_page_refcounted(page + i);
3601 	split_page_owner(page, 1 << order);
3602 	split_page_memcg(page, 1 << order);
3603 }
3604 EXPORT_SYMBOL_GPL(split_page);
3605 
__isolate_free_page(struct page * page,unsigned int order)3606 int __isolate_free_page(struct page *page, unsigned int order)
3607 {
3608 	struct zone *zone = page_zone(page);
3609 	int mt = get_pageblock_migratetype(page);
3610 
3611 	if (!is_migrate_isolate(mt)) {
3612 		unsigned long watermark;
3613 		/*
3614 		 * Obey watermarks as if the page was being allocated. We can
3615 		 * emulate a high-order watermark check with a raised order-0
3616 		 * watermark, because we already know our high-order page
3617 		 * exists.
3618 		 */
3619 		watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3620 		if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3621 			return 0;
3622 
3623 		__mod_zone_freepage_state(zone, -(1UL << order), mt);
3624 	}
3625 
3626 	del_page_from_free_list(page, zone, order);
3627 
3628 	/*
3629 	 * Set the pageblock if the isolated page is at least half of a
3630 	 * pageblock
3631 	 */
3632 	if (order >= pageblock_order - 1) {
3633 		struct page *endpage = page + (1 << order) - 1;
3634 		for (; page < endpage; page += pageblock_nr_pages) {
3635 			int mt = get_pageblock_migratetype(page);
3636 			/*
3637 			 * Only change normal pageblocks (i.e., they can merge
3638 			 * with others)
3639 			 */
3640 			if (migratetype_is_mergeable(mt))
3641 				set_pageblock_migratetype(page,
3642 							  MIGRATE_MOVABLE);
3643 		}
3644 	}
3645 
3646 	return 1UL << order;
3647 }
3648 
3649 /**
3650  * __putback_isolated_page - Return a now-isolated page back where we got it
3651  * @page: Page that was isolated
3652  * @order: Order of the isolated page
3653  * @mt: The page's pageblock's migratetype
3654  *
3655  * This function is meant to return a page pulled from the free lists via
3656  * __isolate_free_page back to the free lists they were pulled from.
3657  */
__putback_isolated_page(struct page * page,unsigned int order,int mt)3658 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3659 {
3660 	struct zone *zone = page_zone(page);
3661 
3662 	/* zone lock should be held when this function is called */
3663 	lockdep_assert_held(&zone->lock);
3664 
3665 	/* Return isolated page to tail of freelist. */
3666 	__free_one_page(page, page_to_pfn(page), zone, order, mt,
3667 			FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3668 }
3669 
3670 /*
3671  * Update NUMA hit/miss statistics
3672  */
zone_statistics(struct zone * preferred_zone,struct zone * z,long nr_account)3673 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3674 				   long nr_account)
3675 {
3676 #ifdef CONFIG_NUMA
3677 	enum numa_stat_item local_stat = NUMA_LOCAL;
3678 
3679 	/* skip numa counters update if numa stats is disabled */
3680 	if (!static_branch_likely(&vm_numa_stat_key))
3681 		return;
3682 
3683 	if (zone_to_nid(z) != numa_node_id())
3684 		local_stat = NUMA_OTHER;
3685 
3686 	if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3687 		__count_numa_events(z, NUMA_HIT, nr_account);
3688 	else {
3689 		__count_numa_events(z, NUMA_MISS, nr_account);
3690 		__count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3691 	}
3692 	__count_numa_events(z, local_stat, nr_account);
3693 #endif
3694 }
3695 
3696 static __always_inline
rmqueue_buddy(struct zone * preferred_zone,struct zone * zone,unsigned int order,unsigned int alloc_flags,int migratetype)3697 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3698 			   unsigned int order, unsigned int alloc_flags,
3699 			   int migratetype)
3700 {
3701 	struct page *page;
3702 	unsigned long flags;
3703 
3704 	do {
3705 		page = NULL;
3706 		spin_lock_irqsave(&zone->lock, flags);
3707 		/*
3708 		 * order-0 request can reach here when the pcplist is skipped
3709 		 * due to non-CMA allocation context. HIGHATOMIC area is
3710 		 * reserved for high-order atomic allocation, so order-0
3711 		 * request should skip it.
3712 		 */
3713 		if (order > 0 && alloc_flags & ALLOC_HARDER)
3714 			page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3715 		if (!page) {
3716 			page = __rmqueue(zone, order, migratetype, alloc_flags);
3717 			if (!page) {
3718 				spin_unlock_irqrestore(&zone->lock, flags);
3719 				return NULL;
3720 			}
3721 		}
3722 		__mod_zone_freepage_state(zone, -(1 << order),
3723 					  get_pcppage_migratetype(page));
3724 		spin_unlock_irqrestore(&zone->lock, flags);
3725 	} while (check_new_pages(page, order));
3726 
3727 	__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3728 	zone_statistics(preferred_zone, zone, 1);
3729 
3730 	return page;
3731 }
3732 
3733 /* Remove page from the per-cpu list, caller must protect the list */
3734 static inline
__rmqueue_pcplist(struct zone * zone,unsigned int order,int migratetype,unsigned int alloc_flags,struct per_cpu_pages * pcp,struct list_head * list)3735 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3736 			int migratetype,
3737 			unsigned int alloc_flags,
3738 			struct per_cpu_pages *pcp,
3739 			struct list_head *list)
3740 {
3741 	struct page *page;
3742 
3743 	do {
3744 		if (list_empty(list)) {
3745 			int batch = READ_ONCE(pcp->batch);
3746 			int alloced;
3747 
3748 			/*
3749 			 * Scale batch relative to order if batch implies
3750 			 * free pages can be stored on the PCP. Batch can
3751 			 * be 1 for small zones or for boot pagesets which
3752 			 * should never store free pages as the pages may
3753 			 * belong to arbitrary zones.
3754 			 */
3755 			if (batch > 1)
3756 				batch = max(batch >> order, 2);
3757 			alloced = rmqueue_bulk(zone, order,
3758 					batch, list,
3759 					migratetype, alloc_flags);
3760 
3761 			pcp->count += alloced << order;
3762 			if (unlikely(list_empty(list)))
3763 				return NULL;
3764 		}
3765 
3766 		page = list_first_entry(list, struct page, pcp_list);
3767 		list_del(&page->pcp_list);
3768 		pcp->count -= 1 << order;
3769 	} while (check_new_pcp(page, order));
3770 
3771 	return page;
3772 }
3773 
3774 /* Lock and remove page from the per-cpu list */
rmqueue_pcplist(struct zone * preferred_zone,struct zone * zone,unsigned int order,int migratetype,unsigned int alloc_flags)3775 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3776 			struct zone *zone, unsigned int order,
3777 			int migratetype, unsigned int alloc_flags)
3778 {
3779 	struct per_cpu_pages *pcp;
3780 	struct list_head *list;
3781 	struct page *page;
3782 	unsigned long flags;
3783 	unsigned long __maybe_unused UP_flags;
3784 
3785 	/*
3786 	 * spin_trylock may fail due to a parallel drain. In the future, the
3787 	 * trylock will also protect against IRQ reentrancy.
3788 	 */
3789 	pcp_trylock_prepare(UP_flags);
3790 	pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3791 	if (!pcp) {
3792 		pcp_trylock_finish(UP_flags);
3793 		return NULL;
3794 	}
3795 
3796 	/*
3797 	 * On allocation, reduce the number of pages that are batch freed.
3798 	 * See nr_pcp_free() where free_factor is increased for subsequent
3799 	 * frees.
3800 	 */
3801 	pcp->free_factor >>= 1;
3802 	list = &pcp->lists[order_to_pindex(migratetype, order)];
3803 	page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3804 	pcp_spin_unlock_irqrestore(pcp, flags);
3805 	pcp_trylock_finish(UP_flags);
3806 	if (page) {
3807 		__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3808 		zone_statistics(preferred_zone, zone, 1);
3809 	}
3810 	return page;
3811 }
3812 
3813 /*
3814  * Allocate a page from the given zone.
3815  * Use pcplists for THP or "cheap" high-order allocations.
3816  */
3817 
3818 /*
3819  * Do not instrument rmqueue() with KMSAN. This function may call
3820  * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3821  * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3822  * may call rmqueue() again, which will result in a deadlock.
3823  */
3824 __no_sanitize_memory
3825 static inline
rmqueue(struct zone * preferred_zone,struct zone * zone,unsigned int order,gfp_t gfp_flags,unsigned int alloc_flags,int migratetype)3826 struct page *rmqueue(struct zone *preferred_zone,
3827 			struct zone *zone, unsigned int order,
3828 			gfp_t gfp_flags, unsigned int alloc_flags,
3829 			int migratetype)
3830 {
3831 	struct page *page;
3832 
3833 	/*
3834 	 * We most definitely don't want callers attempting to
3835 	 * allocate greater than order-1 page units with __GFP_NOFAIL.
3836 	 */
3837 	WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3838 
3839 	if (likely(pcp_allowed_order(order))) {
3840 		/*
3841 		 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3842 		 * we need to skip it when CMA area isn't allowed.
3843 		 */
3844 		if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3845 				migratetype != MIGRATE_MOVABLE) {
3846 			page = rmqueue_pcplist(preferred_zone, zone, order,
3847 					migratetype, alloc_flags);
3848 			if (likely(page))
3849 				goto out;
3850 		}
3851 	}
3852 
3853 	page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3854 							migratetype);
3855 
3856 out:
3857 	/* Separate test+clear to avoid unnecessary atomics */
3858 	if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3859 		clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3860 		wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3861 	}
3862 
3863 	VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3864 	return page;
3865 }
3866 
3867 #ifdef CONFIG_FAIL_PAGE_ALLOC
3868 
3869 static struct {
3870 	struct fault_attr attr;
3871 
3872 	bool ignore_gfp_highmem;
3873 	bool ignore_gfp_reclaim;
3874 	u32 min_order;
3875 } fail_page_alloc = {
3876 	.attr = FAULT_ATTR_INITIALIZER,
3877 	.ignore_gfp_reclaim = true,
3878 	.ignore_gfp_highmem = true,
3879 	.min_order = 1,
3880 };
3881 
setup_fail_page_alloc(char * str)3882 static int __init setup_fail_page_alloc(char *str)
3883 {
3884 	return setup_fault_attr(&fail_page_alloc.attr, str);
3885 }
3886 __setup("fail_page_alloc=", setup_fail_page_alloc);
3887 
__should_fail_alloc_page(gfp_t gfp_mask,unsigned int order)3888 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3889 {
3890 	int flags = 0;
3891 
3892 	if (order < fail_page_alloc.min_order)
3893 		return false;
3894 	if (gfp_mask & __GFP_NOFAIL)
3895 		return false;
3896 	if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3897 		return false;
3898 	if (fail_page_alloc.ignore_gfp_reclaim &&
3899 			(gfp_mask & __GFP_DIRECT_RECLAIM))
3900 		return false;
3901 
3902 	/* See comment in __should_failslab() */
3903 	if (gfp_mask & __GFP_NOWARN)
3904 		flags |= FAULT_NOWARN;
3905 
3906 	return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3907 }
3908 
3909 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3910 
fail_page_alloc_debugfs(void)3911 static int __init fail_page_alloc_debugfs(void)
3912 {
3913 	umode_t mode = S_IFREG | 0600;
3914 	struct dentry *dir;
3915 
3916 	dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3917 					&fail_page_alloc.attr);
3918 
3919 	debugfs_create_bool("ignore-gfp-wait", mode, dir,
3920 			    &fail_page_alloc.ignore_gfp_reclaim);
3921 	debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3922 			    &fail_page_alloc.ignore_gfp_highmem);
3923 	debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3924 
3925 	return 0;
3926 }
3927 
3928 late_initcall(fail_page_alloc_debugfs);
3929 
3930 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3931 
3932 #else /* CONFIG_FAIL_PAGE_ALLOC */
3933 
__should_fail_alloc_page(gfp_t gfp_mask,unsigned int order)3934 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3935 {
3936 	return false;
3937 }
3938 
3939 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3940 
should_fail_alloc_page(gfp_t gfp_mask,unsigned int order)3941 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3942 {
3943 	return __should_fail_alloc_page(gfp_mask, order);
3944 }
3945 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3946 
__zone_watermark_unusable_free(struct zone * z,unsigned int order,unsigned int alloc_flags)3947 static inline long __zone_watermark_unusable_free(struct zone *z,
3948 				unsigned int order, unsigned int alloc_flags)
3949 {
3950 	const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3951 	long unusable_free = (1 << order) - 1;
3952 
3953 	/*
3954 	 * If the caller does not have rights to ALLOC_HARDER then subtract
3955 	 * the high-atomic reserves. This will over-estimate the size of the
3956 	 * atomic reserve but it avoids a search.
3957 	 */
3958 	if (likely(!alloc_harder))
3959 		unusable_free += z->nr_reserved_highatomic;
3960 
3961 #ifdef CONFIG_CMA
3962 	/* If allocation can't use CMA areas don't use free CMA pages */
3963 	if (!(alloc_flags & ALLOC_CMA))
3964 		unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3965 #endif
3966 
3967 	return unusable_free;
3968 }
3969 
3970 /*
3971  * Return true if free base pages are above 'mark'. For high-order checks it
3972  * will return true of the order-0 watermark is reached and there is at least
3973  * one free page of a suitable size. Checking now avoids taking the zone lock
3974  * to check in the allocation paths if no pages are free.
3975  */
__zone_watermark_ok(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags,long free_pages)3976 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3977 			 int highest_zoneidx, unsigned int alloc_flags,
3978 			 long free_pages)
3979 {
3980 	long min = mark;
3981 	int o;
3982 	const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3983 
3984 	/* free_pages may go negative - that's OK */
3985 	free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3986 
3987 	if (alloc_flags & ALLOC_HIGH)
3988 		min -= min / 2;
3989 
3990 	if (unlikely(alloc_harder)) {
3991 		/*
3992 		 * OOM victims can try even harder than normal ALLOC_HARDER
3993 		 * users on the grounds that it's definitely going to be in
3994 		 * the exit path shortly and free memory. Any allocation it
3995 		 * makes during the free path will be small and short-lived.
3996 		 */
3997 		if (alloc_flags & ALLOC_OOM)
3998 			min -= min / 2;
3999 		else
4000 			min -= min / 4;
4001 	}
4002 
4003 	/*
4004 	 * Check watermarks for an order-0 allocation request. If these
4005 	 * are not met, then a high-order request also cannot go ahead
4006 	 * even if a suitable page happened to be free.
4007 	 */
4008 	if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
4009 		return false;
4010 
4011 	/* If this is an order-0 request then the watermark is fine */
4012 	if (!order)
4013 		return true;
4014 
4015 	/* For a high-order request, check at least one suitable page is free */
4016 	for (o = order; o < MAX_ORDER; o++) {
4017 		struct free_area *area = &z->free_area[o];
4018 		int mt;
4019 
4020 		if (!area->nr_free)
4021 			continue;
4022 
4023 		for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4024 			if (!free_area_empty(area, mt))
4025 				return true;
4026 		}
4027 
4028 #ifdef CONFIG_CMA
4029 		if ((alloc_flags & ALLOC_CMA) &&
4030 		    !free_area_empty(area, MIGRATE_CMA)) {
4031 			return true;
4032 		}
4033 #endif
4034 		if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4035 			return true;
4036 	}
4037 	return false;
4038 }
4039 
zone_watermark_ok(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags)4040 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4041 		      int highest_zoneidx, unsigned int alloc_flags)
4042 {
4043 	return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4044 					zone_page_state(z, NR_FREE_PAGES));
4045 }
4046 
zone_watermark_fast(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags,gfp_t gfp_mask)4047 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4048 				unsigned long mark, int highest_zoneidx,
4049 				unsigned int alloc_flags, gfp_t gfp_mask)
4050 {
4051 	long free_pages;
4052 
4053 	free_pages = zone_page_state(z, NR_FREE_PAGES);
4054 
4055 	/*
4056 	 * Fast check for order-0 only. If this fails then the reserves
4057 	 * need to be calculated.
4058 	 */
4059 	if (!order) {
4060 		long usable_free;
4061 		long reserved;
4062 
4063 		usable_free = free_pages;
4064 		reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
4065 
4066 		/* reserved may over estimate high-atomic reserves. */
4067 		usable_free -= min(usable_free, reserved);
4068 		if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
4069 			return true;
4070 	}
4071 
4072 	if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4073 					free_pages))
4074 		return true;
4075 	/*
4076 	 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4077 	 * when checking the min watermark. The min watermark is the
4078 	 * point where boosting is ignored so that kswapd is woken up
4079 	 * when below the low watermark.
4080 	 */
4081 	if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4082 		&& ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4083 		mark = z->_watermark[WMARK_MIN];
4084 		return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4085 					alloc_flags, free_pages);
4086 	}
4087 
4088 	return false;
4089 }
4090 
zone_watermark_ok_safe(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx)4091 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4092 			unsigned long mark, int highest_zoneidx)
4093 {
4094 	long free_pages = zone_page_state(z, NR_FREE_PAGES);
4095 
4096 	if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4097 		free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4098 
4099 	return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4100 								free_pages);
4101 }
4102 
4103 #ifdef CONFIG_NUMA
4104 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4105 
zone_allows_reclaim(struct zone * local_zone,struct zone * zone)4106 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4107 {
4108 	return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4109 				node_reclaim_distance;
4110 }
4111 #else	/* CONFIG_NUMA */
zone_allows_reclaim(struct zone * local_zone,struct zone * zone)4112 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4113 {
4114 	return true;
4115 }
4116 #endif	/* CONFIG_NUMA */
4117 
4118 /*
4119  * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4120  * fragmentation is subtle. If the preferred zone was HIGHMEM then
4121  * premature use of a lower zone may cause lowmem pressure problems that
4122  * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4123  * probably too small. It only makes sense to spread allocations to avoid
4124  * fragmentation between the Normal and DMA32 zones.
4125  */
4126 static inline unsigned int
alloc_flags_nofragment(struct zone * zone,gfp_t gfp_mask)4127 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4128 {
4129 	unsigned int alloc_flags;
4130 
4131 	/*
4132 	 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4133 	 * to save a branch.
4134 	 */
4135 	alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4136 
4137 #ifdef CONFIG_ZONE_DMA32
4138 	if (!zone)
4139 		return alloc_flags;
4140 
4141 	if (zone_idx(zone) != ZONE_NORMAL)
4142 		return alloc_flags;
4143 
4144 	/*
4145 	 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4146 	 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4147 	 * on UMA that if Normal is populated then so is DMA32.
4148 	 */
4149 	BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4150 	if (nr_online_nodes > 1 && !populated_zone(--zone))
4151 		return alloc_flags;
4152 
4153 	alloc_flags |= ALLOC_NOFRAGMENT;
4154 #endif /* CONFIG_ZONE_DMA32 */
4155 	return alloc_flags;
4156 }
4157 
4158 /* Must be called after current_gfp_context() which can change gfp_mask */
gfp_to_alloc_flags_cma(gfp_t gfp_mask,unsigned int alloc_flags)4159 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4160 						  unsigned int alloc_flags)
4161 {
4162 #ifdef CONFIG_CMA
4163 	if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4164 		alloc_flags |= ALLOC_CMA;
4165 #endif
4166 	return alloc_flags;
4167 }
4168 
4169 /*
4170  * get_page_from_freelist goes through the zonelist trying to allocate
4171  * a page.
4172  */
4173 static struct page *
get_page_from_freelist(gfp_t gfp_mask,unsigned int order,int alloc_flags,const struct alloc_context * ac)4174 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4175 						const struct alloc_context *ac)
4176 {
4177 	struct zoneref *z;
4178 	struct zone *zone;
4179 	struct pglist_data *last_pgdat = NULL;
4180 	bool last_pgdat_dirty_ok = false;
4181 	bool no_fallback;
4182 
4183 retry:
4184 	/*
4185 	 * Scan zonelist, looking for a zone with enough free.
4186 	 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
4187 	 */
4188 	no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4189 	z = ac->preferred_zoneref;
4190 	for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4191 					ac->nodemask) {
4192 		struct page *page;
4193 		unsigned long mark;
4194 
4195 		if (cpusets_enabled() &&
4196 			(alloc_flags & ALLOC_CPUSET) &&
4197 			!__cpuset_zone_allowed(zone, gfp_mask))
4198 				continue;
4199 		/*
4200 		 * When allocating a page cache page for writing, we
4201 		 * want to get it from a node that is within its dirty
4202 		 * limit, such that no single node holds more than its
4203 		 * proportional share of globally allowed dirty pages.
4204 		 * The dirty limits take into account the node's
4205 		 * lowmem reserves and high watermark so that kswapd
4206 		 * should be able to balance it without having to
4207 		 * write pages from its LRU list.
4208 		 *
4209 		 * XXX: For now, allow allocations to potentially
4210 		 * exceed the per-node dirty limit in the slowpath
4211 		 * (spread_dirty_pages unset) before going into reclaim,
4212 		 * which is important when on a NUMA setup the allowed
4213 		 * nodes are together not big enough to reach the
4214 		 * global limit.  The proper fix for these situations
4215 		 * will require awareness of nodes in the
4216 		 * dirty-throttling and the flusher threads.
4217 		 */
4218 		if (ac->spread_dirty_pages) {
4219 			if (last_pgdat != zone->zone_pgdat) {
4220 				last_pgdat = zone->zone_pgdat;
4221 				last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4222 			}
4223 
4224 			if (!last_pgdat_dirty_ok)
4225 				continue;
4226 		}
4227 
4228 		if (no_fallback && nr_online_nodes > 1 &&
4229 		    zone != ac->preferred_zoneref->zone) {
4230 			int local_nid;
4231 
4232 			/*
4233 			 * If moving to a remote node, retry but allow
4234 			 * fragmenting fallbacks. Locality is more important
4235 			 * than fragmentation avoidance.
4236 			 */
4237 			local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4238 			if (zone_to_nid(zone) != local_nid) {
4239 				alloc_flags &= ~ALLOC_NOFRAGMENT;
4240 				goto retry;
4241 			}
4242 		}
4243 
4244 		mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4245 		if (!zone_watermark_fast(zone, order, mark,
4246 				       ac->highest_zoneidx, alloc_flags,
4247 				       gfp_mask)) {
4248 			int ret;
4249 
4250 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4251 			/*
4252 			 * Watermark failed for this zone, but see if we can
4253 			 * grow this zone if it contains deferred pages.
4254 			 */
4255 			if (static_branch_unlikely(&deferred_pages)) {
4256 				if (_deferred_grow_zone(zone, order))
4257 					goto try_this_zone;
4258 			}
4259 #endif
4260 			/* Checked here to keep the fast path fast */
4261 			BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4262 			if (alloc_flags & ALLOC_NO_WATERMARKS)
4263 				goto try_this_zone;
4264 
4265 			if (!node_reclaim_enabled() ||
4266 			    !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4267 				continue;
4268 
4269 			ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4270 			switch (ret) {
4271 			case NODE_RECLAIM_NOSCAN:
4272 				/* did not scan */
4273 				continue;
4274 			case NODE_RECLAIM_FULL:
4275 				/* scanned but unreclaimable */
4276 				continue;
4277 			default:
4278 				/* did we reclaim enough */
4279 				if (zone_watermark_ok(zone, order, mark,
4280 					ac->highest_zoneidx, alloc_flags))
4281 					goto try_this_zone;
4282 
4283 				continue;
4284 			}
4285 		}
4286 
4287 try_this_zone:
4288 		page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4289 				gfp_mask, alloc_flags, ac->migratetype);
4290 		if (page) {
4291 			prep_new_page(page, order, gfp_mask, alloc_flags);
4292 
4293 			/*
4294 			 * If this is a high-order atomic allocation then check
4295 			 * if the pageblock should be reserved for the future
4296 			 */
4297 			if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4298 				reserve_highatomic_pageblock(page, zone, order);
4299 
4300 			return page;
4301 		} else {
4302 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4303 			/* Try again if zone has deferred pages */
4304 			if (static_branch_unlikely(&deferred_pages)) {
4305 				if (_deferred_grow_zone(zone, order))
4306 					goto try_this_zone;
4307 			}
4308 #endif
4309 		}
4310 	}
4311 
4312 	/*
4313 	 * It's possible on a UMA machine to get through all zones that are
4314 	 * fragmented. If avoiding fragmentation, reset and try again.
4315 	 */
4316 	if (no_fallback) {
4317 		alloc_flags &= ~ALLOC_NOFRAGMENT;
4318 		goto retry;
4319 	}
4320 
4321 	return NULL;
4322 }
4323 
warn_alloc_show_mem(gfp_t gfp_mask,nodemask_t * nodemask)4324 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4325 {
4326 	unsigned int filter = SHOW_MEM_FILTER_NODES;
4327 
4328 	/*
4329 	 * This documents exceptions given to allocations in certain
4330 	 * contexts that are allowed to allocate outside current's set
4331 	 * of allowed nodes.
4332 	 */
4333 	if (!(gfp_mask & __GFP_NOMEMALLOC))
4334 		if (tsk_is_oom_victim(current) ||
4335 		    (current->flags & (PF_MEMALLOC | PF_EXITING)))
4336 			filter &= ~SHOW_MEM_FILTER_NODES;
4337 	if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4338 		filter &= ~SHOW_MEM_FILTER_NODES;
4339 
4340 	__show_mem(filter, nodemask, gfp_zone(gfp_mask));
4341 }
4342 
warn_alloc(gfp_t gfp_mask,nodemask_t * nodemask,const char * fmt,...)4343 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4344 {
4345 	struct va_format vaf;
4346 	va_list args;
4347 	static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4348 
4349 	if ((gfp_mask & __GFP_NOWARN) ||
4350 	     !__ratelimit(&nopage_rs) ||
4351 	     ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4352 		return;
4353 
4354 	va_start(args, fmt);
4355 	vaf.fmt = fmt;
4356 	vaf.va = &args;
4357 	pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4358 			current->comm, &vaf, gfp_mask, &gfp_mask,
4359 			nodemask_pr_args(nodemask));
4360 	va_end(args);
4361 
4362 	cpuset_print_current_mems_allowed();
4363 	pr_cont("\n");
4364 	dump_stack();
4365 	warn_alloc_show_mem(gfp_mask, nodemask);
4366 }
4367 
4368 static inline struct page *
__alloc_pages_cpuset_fallback(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac)4369 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4370 			      unsigned int alloc_flags,
4371 			      const struct alloc_context *ac)
4372 {
4373 	struct page *page;
4374 
4375 	page = get_page_from_freelist(gfp_mask, order,
4376 			alloc_flags|ALLOC_CPUSET, ac);
4377 	/*
4378 	 * fallback to ignore cpuset restriction if our nodes
4379 	 * are depleted
4380 	 */
4381 	if (!page)
4382 		page = get_page_from_freelist(gfp_mask, order,
4383 				alloc_flags, ac);
4384 
4385 	return page;
4386 }
4387 
4388 static inline struct page *
__alloc_pages_may_oom(gfp_t gfp_mask,unsigned int order,const struct alloc_context * ac,unsigned long * did_some_progress)4389 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4390 	const struct alloc_context *ac, unsigned long *did_some_progress)
4391 {
4392 	struct oom_control oc = {
4393 		.zonelist = ac->zonelist,
4394 		.nodemask = ac->nodemask,
4395 		.memcg = NULL,
4396 		.gfp_mask = gfp_mask,
4397 		.order = order,
4398 	};
4399 	struct page *page;
4400 
4401 	*did_some_progress = 0;
4402 
4403 	/*
4404 	 * Acquire the oom lock.  If that fails, somebody else is
4405 	 * making progress for us.
4406 	 */
4407 	if (!mutex_trylock(&oom_lock)) {
4408 		*did_some_progress = 1;
4409 		schedule_timeout_uninterruptible(1);
4410 		return NULL;
4411 	}
4412 
4413 	/*
4414 	 * Go through the zonelist yet one more time, keep very high watermark
4415 	 * here, this is only to catch a parallel oom killing, we must fail if
4416 	 * we're still under heavy pressure. But make sure that this reclaim
4417 	 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4418 	 * allocation which will never fail due to oom_lock already held.
4419 	 */
4420 	page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4421 				      ~__GFP_DIRECT_RECLAIM, order,
4422 				      ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4423 	if (page)
4424 		goto out;
4425 
4426 	/* Coredumps can quickly deplete all memory reserves */
4427 	if (current->flags & PF_DUMPCORE)
4428 		goto out;
4429 	/* The OOM killer will not help higher order allocs */
4430 	if (order > PAGE_ALLOC_COSTLY_ORDER)
4431 		goto out;
4432 	/*
4433 	 * We have already exhausted all our reclaim opportunities without any
4434 	 * success so it is time to admit defeat. We will skip the OOM killer
4435 	 * because it is very likely that the caller has a more reasonable
4436 	 * fallback than shooting a random task.
4437 	 *
4438 	 * The OOM killer may not free memory on a specific node.
4439 	 */
4440 	if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4441 		goto out;
4442 	/* The OOM killer does not needlessly kill tasks for lowmem */
4443 	if (ac->highest_zoneidx < ZONE_NORMAL)
4444 		goto out;
4445 	if (pm_suspended_storage())
4446 		goto out;
4447 	/*
4448 	 * XXX: GFP_NOFS allocations should rather fail than rely on
4449 	 * other request to make a forward progress.
4450 	 * We are in an unfortunate situation where out_of_memory cannot
4451 	 * do much for this context but let's try it to at least get
4452 	 * access to memory reserved if the current task is killed (see
4453 	 * out_of_memory). Once filesystems are ready to handle allocation
4454 	 * failures more gracefully we should just bail out here.
4455 	 */
4456 
4457 	/* Exhausted what can be done so it's blame time */
4458 	if (out_of_memory(&oc) ||
4459 	    WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4460 		*did_some_progress = 1;
4461 
4462 		/*
4463 		 * Help non-failing allocations by giving them access to memory
4464 		 * reserves
4465 		 */
4466 		if (gfp_mask & __GFP_NOFAIL)
4467 			page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4468 					ALLOC_NO_WATERMARKS, ac);
4469 	}
4470 out:
4471 	mutex_unlock(&oom_lock);
4472 	return page;
4473 }
4474 
4475 /*
4476  * Maximum number of compaction retries with a progress before OOM
4477  * killer is consider as the only way to move forward.
4478  */
4479 #define MAX_COMPACT_RETRIES 16
4480 
4481 #ifdef CONFIG_COMPACTION
4482 /* Try memory compaction for high-order allocations before reclaim */
4483 static struct page *
__alloc_pages_direct_compact(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac,enum compact_priority prio,enum compact_result * compact_result)4484 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4485 		unsigned int alloc_flags, const struct alloc_context *ac,
4486 		enum compact_priority prio, enum compact_result *compact_result)
4487 {
4488 	struct page *page = NULL;
4489 	unsigned long pflags;
4490 	unsigned int noreclaim_flag;
4491 
4492 	if (!order)
4493 		return NULL;
4494 
4495 	psi_memstall_enter(&pflags);
4496 	delayacct_compact_start();
4497 	noreclaim_flag = memalloc_noreclaim_save();
4498 
4499 	*compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4500 								prio, &page);
4501 
4502 	memalloc_noreclaim_restore(noreclaim_flag);
4503 	psi_memstall_leave(&pflags);
4504 	delayacct_compact_end();
4505 
4506 	if (*compact_result == COMPACT_SKIPPED)
4507 		return NULL;
4508 	/*
4509 	 * At least in one zone compaction wasn't deferred or skipped, so let's
4510 	 * count a compaction stall
4511 	 */
4512 	count_vm_event(COMPACTSTALL);
4513 
4514 	/* Prep a captured page if available */
4515 	if (page)
4516 		prep_new_page(page, order, gfp_mask, alloc_flags);
4517 
4518 	/* Try get a page from the freelist if available */
4519 	if (!page)
4520 		page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4521 
4522 	if (page) {
4523 		struct zone *zone = page_zone(page);
4524 
4525 		zone->compact_blockskip_flush = false;
4526 		compaction_defer_reset(zone, order, true);
4527 		count_vm_event(COMPACTSUCCESS);
4528 		return page;
4529 	}
4530 
4531 	/*
4532 	 * It's bad if compaction run occurs and fails. The most likely reason
4533 	 * is that pages exist, but not enough to satisfy watermarks.
4534 	 */
4535 	count_vm_event(COMPACTFAIL);
4536 
4537 	cond_resched();
4538 
4539 	return NULL;
4540 }
4541 
4542 static inline bool
should_compact_retry(struct alloc_context * ac,int order,int alloc_flags,enum compact_result compact_result,enum compact_priority * compact_priority,int * compaction_retries)4543 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4544 		     enum compact_result compact_result,
4545 		     enum compact_priority *compact_priority,
4546 		     int *compaction_retries)
4547 {
4548 	int max_retries = MAX_COMPACT_RETRIES;
4549 	int min_priority;
4550 	bool ret = false;
4551 	int retries = *compaction_retries;
4552 	enum compact_priority priority = *compact_priority;
4553 
4554 	if (!order)
4555 		return false;
4556 
4557 	if (fatal_signal_pending(current))
4558 		return false;
4559 
4560 	if (compaction_made_progress(compact_result))
4561 		(*compaction_retries)++;
4562 
4563 	/*
4564 	 * compaction considers all the zone as desperately out of memory
4565 	 * so it doesn't really make much sense to retry except when the
4566 	 * failure could be caused by insufficient priority
4567 	 */
4568 	if (compaction_failed(compact_result))
4569 		goto check_priority;
4570 
4571 	/*
4572 	 * compaction was skipped because there are not enough order-0 pages
4573 	 * to work with, so we retry only if it looks like reclaim can help.
4574 	 */
4575 	if (compaction_needs_reclaim(compact_result)) {
4576 		ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4577 		goto out;
4578 	}
4579 
4580 	/*
4581 	 * make sure the compaction wasn't deferred or didn't bail out early
4582 	 * due to locks contention before we declare that we should give up.
4583 	 * But the next retry should use a higher priority if allowed, so
4584 	 * we don't just keep bailing out endlessly.
4585 	 */
4586 	if (compaction_withdrawn(compact_result)) {
4587 		goto check_priority;
4588 	}
4589 
4590 	/*
4591 	 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4592 	 * costly ones because they are de facto nofail and invoke OOM
4593 	 * killer to move on while costly can fail and users are ready
4594 	 * to cope with that. 1/4 retries is rather arbitrary but we
4595 	 * would need much more detailed feedback from compaction to
4596 	 * make a better decision.
4597 	 */
4598 	if (order > PAGE_ALLOC_COSTLY_ORDER)
4599 		max_retries /= 4;
4600 	if (*compaction_retries <= max_retries) {
4601 		ret = true;
4602 		goto out;
4603 	}
4604 
4605 	/*
4606 	 * Make sure there are attempts at the highest priority if we exhausted
4607 	 * all retries or failed at the lower priorities.
4608 	 */
4609 check_priority:
4610 	min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4611 			MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4612 
4613 	if (*compact_priority > min_priority) {
4614 		(*compact_priority)--;
4615 		*compaction_retries = 0;
4616 		ret = true;
4617 	}
4618 out:
4619 	trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4620 	return ret;
4621 }
4622 #else
4623 static inline struct page *
__alloc_pages_direct_compact(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac,enum compact_priority prio,enum compact_result * compact_result)4624 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4625 		unsigned int alloc_flags, const struct alloc_context *ac,
4626 		enum compact_priority prio, enum compact_result *compact_result)
4627 {
4628 	*compact_result = COMPACT_SKIPPED;
4629 	return NULL;
4630 }
4631 
4632 static inline bool
should_compact_retry(struct alloc_context * ac,unsigned int order,int alloc_flags,enum compact_result compact_result,enum compact_priority * compact_priority,int * compaction_retries)4633 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4634 		     enum compact_result compact_result,
4635 		     enum compact_priority *compact_priority,
4636 		     int *compaction_retries)
4637 {
4638 	struct zone *zone;
4639 	struct zoneref *z;
4640 
4641 	if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4642 		return false;
4643 
4644 	/*
4645 	 * There are setups with compaction disabled which would prefer to loop
4646 	 * inside the allocator rather than hit the oom killer prematurely.
4647 	 * Let's give them a good hope and keep retrying while the order-0
4648 	 * watermarks are OK.
4649 	 */
4650 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4651 				ac->highest_zoneidx, ac->nodemask) {
4652 		if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4653 					ac->highest_zoneidx, alloc_flags))
4654 			return true;
4655 	}
4656 	return false;
4657 }
4658 #endif /* CONFIG_COMPACTION */
4659 
4660 #ifdef CONFIG_LOCKDEP
4661 static struct lockdep_map __fs_reclaim_map =
4662 	STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4663 
__need_reclaim(gfp_t gfp_mask)4664 static bool __need_reclaim(gfp_t gfp_mask)
4665 {
4666 	/* no reclaim without waiting on it */
4667 	if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4668 		return false;
4669 
4670 	/* this guy won't enter reclaim */
4671 	if (current->flags & PF_MEMALLOC)
4672 		return false;
4673 
4674 	if (gfp_mask & __GFP_NOLOCKDEP)
4675 		return false;
4676 
4677 	return true;
4678 }
4679 
__fs_reclaim_acquire(unsigned long ip)4680 void __fs_reclaim_acquire(unsigned long ip)
4681 {
4682 	lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4683 }
4684 
__fs_reclaim_release(unsigned long ip)4685 void __fs_reclaim_release(unsigned long ip)
4686 {
4687 	lock_release(&__fs_reclaim_map, ip);
4688 }
4689 
fs_reclaim_acquire(gfp_t gfp_mask)4690 void fs_reclaim_acquire(gfp_t gfp_mask)
4691 {
4692 	gfp_mask = current_gfp_context(gfp_mask);
4693 
4694 	if (__need_reclaim(gfp_mask)) {
4695 		if (gfp_mask & __GFP_FS)
4696 			__fs_reclaim_acquire(_RET_IP_);
4697 
4698 #ifdef CONFIG_MMU_NOTIFIER
4699 		lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4700 		lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4701 #endif
4702 
4703 	}
4704 }
4705 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4706 
fs_reclaim_release(gfp_t gfp_mask)4707 void fs_reclaim_release(gfp_t gfp_mask)
4708 {
4709 	gfp_mask = current_gfp_context(gfp_mask);
4710 
4711 	if (__need_reclaim(gfp_mask)) {
4712 		if (gfp_mask & __GFP_FS)
4713 			__fs_reclaim_release(_RET_IP_);
4714 	}
4715 }
4716 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4717 #endif
4718 
4719 /*
4720  * Zonelists may change due to hotplug during allocation. Detect when zonelists
4721  * have been rebuilt so allocation retries. Reader side does not lock and
4722  * retries the allocation if zonelist changes. Writer side is protected by the
4723  * embedded spin_lock.
4724  */
4725 static DEFINE_SEQLOCK(zonelist_update_seq);
4726 
zonelist_iter_begin(void)4727 static unsigned int zonelist_iter_begin(void)
4728 {
4729 	if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4730 		return read_seqbegin(&zonelist_update_seq);
4731 
4732 	return 0;
4733 }
4734 
check_retry_zonelist(unsigned int seq)4735 static unsigned int check_retry_zonelist(unsigned int seq)
4736 {
4737 	if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4738 		return read_seqretry(&zonelist_update_seq, seq);
4739 
4740 	return seq;
4741 }
4742 
4743 /* Perform direct synchronous page reclaim */
4744 static unsigned long
__perform_reclaim(gfp_t gfp_mask,unsigned int order,const struct alloc_context * ac)4745 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4746 					const struct alloc_context *ac)
4747 {
4748 	unsigned int noreclaim_flag;
4749 	unsigned long progress;
4750 
4751 	cond_resched();
4752 
4753 	/* We now go into synchronous reclaim */
4754 	cpuset_memory_pressure_bump();
4755 	fs_reclaim_acquire(gfp_mask);
4756 	noreclaim_flag = memalloc_noreclaim_save();
4757 
4758 	progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4759 								ac->nodemask);
4760 
4761 	memalloc_noreclaim_restore(noreclaim_flag);
4762 	fs_reclaim_release(gfp_mask);
4763 
4764 	cond_resched();
4765 
4766 	return progress;
4767 }
4768 
4769 /* The really slow allocator path where we enter direct reclaim */
4770 static inline struct page *
__alloc_pages_direct_reclaim(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac,unsigned long * did_some_progress)4771 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4772 		unsigned int alloc_flags, const struct alloc_context *ac,
4773 		unsigned long *did_some_progress)
4774 {
4775 	struct page *page = NULL;
4776 	unsigned long pflags;
4777 	bool drained = false;
4778 
4779 	psi_memstall_enter(&pflags);
4780 	*did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4781 	if (unlikely(!(*did_some_progress)))
4782 		goto out;
4783 
4784 retry:
4785 	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4786 
4787 	/*
4788 	 * If an allocation failed after direct reclaim, it could be because
4789 	 * pages are pinned on the per-cpu lists or in high alloc reserves.
4790 	 * Shrink them and try again
4791 	 */
4792 	if (!page && !drained) {
4793 		unreserve_highatomic_pageblock(ac, false);
4794 		drain_all_pages(NULL);
4795 		drained = true;
4796 		goto retry;
4797 	}
4798 out:
4799 	psi_memstall_leave(&pflags);
4800 
4801 	return page;
4802 }
4803 
wake_all_kswapds(unsigned int order,gfp_t gfp_mask,const struct alloc_context * ac)4804 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4805 			     const struct alloc_context *ac)
4806 {
4807 	struct zoneref *z;
4808 	struct zone *zone;
4809 	pg_data_t *last_pgdat = NULL;
4810 	enum zone_type highest_zoneidx = ac->highest_zoneidx;
4811 
4812 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4813 					ac->nodemask) {
4814 		if (!managed_zone(zone))
4815 			continue;
4816 		if (last_pgdat != zone->zone_pgdat) {
4817 			wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4818 			last_pgdat = zone->zone_pgdat;
4819 		}
4820 	}
4821 }
4822 
4823 static inline unsigned int
gfp_to_alloc_flags(gfp_t gfp_mask)4824 gfp_to_alloc_flags(gfp_t gfp_mask)
4825 {
4826 	unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4827 
4828 	/*
4829 	 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4830 	 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4831 	 * to save two branches.
4832 	 */
4833 	BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4834 	BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4835 
4836 	/*
4837 	 * The caller may dip into page reserves a bit more if the caller
4838 	 * cannot run direct reclaim, or if the caller has realtime scheduling
4839 	 * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
4840 	 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4841 	 */
4842 	alloc_flags |= (__force int)
4843 		(gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4844 
4845 	if (gfp_mask & __GFP_ATOMIC) {
4846 		/*
4847 		 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4848 		 * if it can't schedule.
4849 		 */
4850 		if (!(gfp_mask & __GFP_NOMEMALLOC))
4851 			alloc_flags |= ALLOC_HARDER;
4852 		/*
4853 		 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4854 		 * comment for __cpuset_node_allowed().
4855 		 */
4856 		alloc_flags &= ~ALLOC_CPUSET;
4857 	} else if (unlikely(rt_task(current)) && in_task())
4858 		alloc_flags |= ALLOC_HARDER;
4859 
4860 	alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4861 
4862 	return alloc_flags;
4863 }
4864 
oom_reserves_allowed(struct task_struct * tsk)4865 static bool oom_reserves_allowed(struct task_struct *tsk)
4866 {
4867 	if (!tsk_is_oom_victim(tsk))
4868 		return false;
4869 
4870 	/*
4871 	 * !MMU doesn't have oom reaper so give access to memory reserves
4872 	 * only to the thread with TIF_MEMDIE set
4873 	 */
4874 	if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4875 		return false;
4876 
4877 	return true;
4878 }
4879 
4880 /*
4881  * Distinguish requests which really need access to full memory
4882  * reserves from oom victims which can live with a portion of it
4883  */
__gfp_pfmemalloc_flags(gfp_t gfp_mask)4884 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4885 {
4886 	if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4887 		return 0;
4888 	if (gfp_mask & __GFP_MEMALLOC)
4889 		return ALLOC_NO_WATERMARKS;
4890 	if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4891 		return ALLOC_NO_WATERMARKS;
4892 	if (!in_interrupt()) {
4893 		if (current->flags & PF_MEMALLOC)
4894 			return ALLOC_NO_WATERMARKS;
4895 		else if (oom_reserves_allowed(current))
4896 			return ALLOC_OOM;
4897 	}
4898 
4899 	return 0;
4900 }
4901 
gfp_pfmemalloc_allowed(gfp_t gfp_mask)4902 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4903 {
4904 	return !!__gfp_pfmemalloc_flags(gfp_mask);
4905 }
4906 
4907 /*
4908  * Checks whether it makes sense to retry the reclaim to make a forward progress
4909  * for the given allocation request.
4910  *
4911  * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4912  * without success, or when we couldn't even meet the watermark if we
4913  * reclaimed all remaining pages on the LRU lists.
4914  *
4915  * Returns true if a retry is viable or false to enter the oom path.
4916  */
4917 static inline bool
should_reclaim_retry(gfp_t gfp_mask,unsigned order,struct alloc_context * ac,int alloc_flags,bool did_some_progress,int * no_progress_loops)4918 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4919 		     struct alloc_context *ac, int alloc_flags,
4920 		     bool did_some_progress, int *no_progress_loops)
4921 {
4922 	struct zone *zone;
4923 	struct zoneref *z;
4924 	bool ret = false;
4925 
4926 	/*
4927 	 * Costly allocations might have made a progress but this doesn't mean
4928 	 * their order will become available due to high fragmentation so
4929 	 * always increment the no progress counter for them
4930 	 */
4931 	if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4932 		*no_progress_loops = 0;
4933 	else
4934 		(*no_progress_loops)++;
4935 
4936 	/*
4937 	 * Make sure we converge to OOM if we cannot make any progress
4938 	 * several times in the row.
4939 	 */
4940 	if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4941 		/* Before OOM, exhaust highatomic_reserve */
4942 		return unreserve_highatomic_pageblock(ac, true);
4943 	}
4944 
4945 	/*
4946 	 * Keep reclaiming pages while there is a chance this will lead
4947 	 * somewhere.  If none of the target zones can satisfy our allocation
4948 	 * request even if all reclaimable pages are considered then we are
4949 	 * screwed and have to go OOM.
4950 	 */
4951 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4952 				ac->highest_zoneidx, ac->nodemask) {
4953 		unsigned long available;
4954 		unsigned long reclaimable;
4955 		unsigned long min_wmark = min_wmark_pages(zone);
4956 		bool wmark;
4957 
4958 		available = reclaimable = zone_reclaimable_pages(zone);
4959 		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4960 
4961 		/*
4962 		 * Would the allocation succeed if we reclaimed all
4963 		 * reclaimable pages?
4964 		 */
4965 		wmark = __zone_watermark_ok(zone, order, min_wmark,
4966 				ac->highest_zoneidx, alloc_flags, available);
4967 		trace_reclaim_retry_zone(z, order, reclaimable,
4968 				available, min_wmark, *no_progress_loops, wmark);
4969 		if (wmark) {
4970 			ret = true;
4971 			break;
4972 		}
4973 	}
4974 
4975 	/*
4976 	 * Memory allocation/reclaim might be called from a WQ context and the
4977 	 * current implementation of the WQ concurrency control doesn't
4978 	 * recognize that a particular WQ is congested if the worker thread is
4979 	 * looping without ever sleeping. Therefore we have to do a short sleep
4980 	 * here rather than calling cond_resched().
4981 	 */
4982 	if (current->flags & PF_WQ_WORKER)
4983 		schedule_timeout_uninterruptible(1);
4984 	else
4985 		cond_resched();
4986 	return ret;
4987 }
4988 
4989 static inline bool
check_retry_cpuset(int cpuset_mems_cookie,struct alloc_context * ac)4990 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4991 {
4992 	/*
4993 	 * It's possible that cpuset's mems_allowed and the nodemask from
4994 	 * mempolicy don't intersect. This should be normally dealt with by
4995 	 * policy_nodemask(), but it's possible to race with cpuset update in
4996 	 * such a way the check therein was true, and then it became false
4997 	 * before we got our cpuset_mems_cookie here.
4998 	 * This assumes that for all allocations, ac->nodemask can come only
4999 	 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
5000 	 * when it does not intersect with the cpuset restrictions) or the
5001 	 * caller can deal with a violated nodemask.
5002 	 */
5003 	if (cpusets_enabled() && ac->nodemask &&
5004 			!cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
5005 		ac->nodemask = NULL;
5006 		return true;
5007 	}
5008 
5009 	/*
5010 	 * When updating a task's mems_allowed or mempolicy nodemask, it is
5011 	 * possible to race with parallel threads in such a way that our
5012 	 * allocation can fail while the mask is being updated. If we are about
5013 	 * to fail, check if the cpuset changed during allocation and if so,
5014 	 * retry.
5015 	 */
5016 	if (read_mems_allowed_retry(cpuset_mems_cookie))
5017 		return true;
5018 
5019 	return false;
5020 }
5021 
5022 static inline struct page *
__alloc_pages_slowpath(gfp_t gfp_mask,unsigned int order,struct alloc_context * ac)5023 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
5024 						struct alloc_context *ac)
5025 {
5026 	bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
5027 	const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
5028 	struct page *page = NULL;
5029 	unsigned int alloc_flags;
5030 	unsigned long did_some_progress;
5031 	enum compact_priority compact_priority;
5032 	enum compact_result compact_result;
5033 	int compaction_retries;
5034 	int no_progress_loops;
5035 	unsigned int cpuset_mems_cookie;
5036 	unsigned int zonelist_iter_cookie;
5037 	int reserve_flags;
5038 
5039 	/*
5040 	 * We also sanity check to catch abuse of atomic reserves being used by
5041 	 * callers that are not in atomic context.
5042 	 */
5043 	if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5044 				(__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5045 		gfp_mask &= ~__GFP_ATOMIC;
5046 
5047 restart:
5048 	compaction_retries = 0;
5049 	no_progress_loops = 0;
5050 	compact_priority = DEF_COMPACT_PRIORITY;
5051 	cpuset_mems_cookie = read_mems_allowed_begin();
5052 	zonelist_iter_cookie = zonelist_iter_begin();
5053 
5054 	/*
5055 	 * The fast path uses conservative alloc_flags to succeed only until
5056 	 * kswapd needs to be woken up, and to avoid the cost of setting up
5057 	 * alloc_flags precisely. So we do that now.
5058 	 */
5059 	alloc_flags = gfp_to_alloc_flags(gfp_mask);
5060 
5061 	/*
5062 	 * We need to recalculate the starting point for the zonelist iterator
5063 	 * because we might have used different nodemask in the fast path, or
5064 	 * there was a cpuset modification and we are retrying - otherwise we
5065 	 * could end up iterating over non-eligible zones endlessly.
5066 	 */
5067 	ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5068 					ac->highest_zoneidx, ac->nodemask);
5069 	if (!ac->preferred_zoneref->zone)
5070 		goto nopage;
5071 
5072 	/*
5073 	 * Check for insane configurations where the cpuset doesn't contain
5074 	 * any suitable zone to satisfy the request - e.g. non-movable
5075 	 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5076 	 */
5077 	if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5078 		struct zoneref *z = first_zones_zonelist(ac->zonelist,
5079 					ac->highest_zoneidx,
5080 					&cpuset_current_mems_allowed);
5081 		if (!z->zone)
5082 			goto nopage;
5083 	}
5084 
5085 	if (alloc_flags & ALLOC_KSWAPD)
5086 		wake_all_kswapds(order, gfp_mask, ac);
5087 
5088 	/*
5089 	 * The adjusted alloc_flags might result in immediate success, so try
5090 	 * that first
5091 	 */
5092 	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5093 	if (page)
5094 		goto got_pg;
5095 
5096 	/*
5097 	 * For costly allocations, try direct compaction first, as it's likely
5098 	 * that we have enough base pages and don't need to reclaim. For non-
5099 	 * movable high-order allocations, do that as well, as compaction will
5100 	 * try prevent permanent fragmentation by migrating from blocks of the
5101 	 * same migratetype.
5102 	 * Don't try this for allocations that are allowed to ignore
5103 	 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5104 	 */
5105 	if (can_direct_reclaim &&
5106 			(costly_order ||
5107 			   (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5108 			&& !gfp_pfmemalloc_allowed(gfp_mask)) {
5109 		page = __alloc_pages_direct_compact(gfp_mask, order,
5110 						alloc_flags, ac,
5111 						INIT_COMPACT_PRIORITY,
5112 						&compact_result);
5113 		if (page)
5114 			goto got_pg;
5115 
5116 		/*
5117 		 * Checks for costly allocations with __GFP_NORETRY, which
5118 		 * includes some THP page fault allocations
5119 		 */
5120 		if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5121 			/*
5122 			 * If allocating entire pageblock(s) and compaction
5123 			 * failed because all zones are below low watermarks
5124 			 * or is prohibited because it recently failed at this
5125 			 * order, fail immediately unless the allocator has
5126 			 * requested compaction and reclaim retry.
5127 			 *
5128 			 * Reclaim is
5129 			 *  - potentially very expensive because zones are far
5130 			 *    below their low watermarks or this is part of very
5131 			 *    bursty high order allocations,
5132 			 *  - not guaranteed to help because isolate_freepages()
5133 			 *    may not iterate over freed pages as part of its
5134 			 *    linear scan, and
5135 			 *  - unlikely to make entire pageblocks free on its
5136 			 *    own.
5137 			 */
5138 			if (compact_result == COMPACT_SKIPPED ||
5139 			    compact_result == COMPACT_DEFERRED)
5140 				goto nopage;
5141 
5142 			/*
5143 			 * Looks like reclaim/compaction is worth trying, but
5144 			 * sync compaction could be very expensive, so keep
5145 			 * using async compaction.
5146 			 */
5147 			compact_priority = INIT_COMPACT_PRIORITY;
5148 		}
5149 	}
5150 
5151 retry:
5152 	/* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5153 	if (alloc_flags & ALLOC_KSWAPD)
5154 		wake_all_kswapds(order, gfp_mask, ac);
5155 
5156 	reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5157 	if (reserve_flags)
5158 		alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
5159 					  (alloc_flags & ALLOC_KSWAPD);
5160 
5161 	/*
5162 	 * Reset the nodemask and zonelist iterators if memory policies can be
5163 	 * ignored. These allocations are high priority and system rather than
5164 	 * user oriented.
5165 	 */
5166 	if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5167 		ac->nodemask = NULL;
5168 		ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5169 					ac->highest_zoneidx, ac->nodemask);
5170 	}
5171 
5172 	/* Attempt with potentially adjusted zonelist and alloc_flags */
5173 	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5174 	if (page)
5175 		goto got_pg;
5176 
5177 	/* Caller is not willing to reclaim, we can't balance anything */
5178 	if (!can_direct_reclaim)
5179 		goto nopage;
5180 
5181 	/* Avoid recursion of direct reclaim */
5182 	if (current->flags & PF_MEMALLOC)
5183 		goto nopage;
5184 
5185 	/* Try direct reclaim and then allocating */
5186 	page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5187 							&did_some_progress);
5188 	if (page)
5189 		goto got_pg;
5190 
5191 	/* Try direct compaction and then allocating */
5192 	page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5193 					compact_priority, &compact_result);
5194 	if (page)
5195 		goto got_pg;
5196 
5197 	/* Do not loop if specifically requested */
5198 	if (gfp_mask & __GFP_NORETRY)
5199 		goto nopage;
5200 
5201 	/*
5202 	 * Do not retry costly high order allocations unless they are
5203 	 * __GFP_RETRY_MAYFAIL
5204 	 */
5205 	if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5206 		goto nopage;
5207 
5208 	if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5209 				 did_some_progress > 0, &no_progress_loops))
5210 		goto retry;
5211 
5212 	/*
5213 	 * It doesn't make any sense to retry for the compaction if the order-0
5214 	 * reclaim is not able to make any progress because the current
5215 	 * implementation of the compaction depends on the sufficient amount
5216 	 * of free memory (see __compaction_suitable)
5217 	 */
5218 	if (did_some_progress > 0 &&
5219 			should_compact_retry(ac, order, alloc_flags,
5220 				compact_result, &compact_priority,
5221 				&compaction_retries))
5222 		goto retry;
5223 
5224 
5225 	/*
5226 	 * Deal with possible cpuset update races or zonelist updates to avoid
5227 	 * a unnecessary OOM kill.
5228 	 */
5229 	if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5230 	    check_retry_zonelist(zonelist_iter_cookie))
5231 		goto restart;
5232 
5233 	/* Reclaim has failed us, start killing things */
5234 	page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5235 	if (page)
5236 		goto got_pg;
5237 
5238 	/* Avoid allocations with no watermarks from looping endlessly */
5239 	if (tsk_is_oom_victim(current) &&
5240 	    (alloc_flags & ALLOC_OOM ||
5241 	     (gfp_mask & __GFP_NOMEMALLOC)))
5242 		goto nopage;
5243 
5244 	/* Retry as long as the OOM killer is making progress */
5245 	if (did_some_progress) {
5246 		no_progress_loops = 0;
5247 		goto retry;
5248 	}
5249 
5250 nopage:
5251 	/*
5252 	 * Deal with possible cpuset update races or zonelist updates to avoid
5253 	 * a unnecessary OOM kill.
5254 	 */
5255 	if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5256 	    check_retry_zonelist(zonelist_iter_cookie))
5257 		goto restart;
5258 
5259 	/*
5260 	 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5261 	 * we always retry
5262 	 */
5263 	if (gfp_mask & __GFP_NOFAIL) {
5264 		/*
5265 		 * All existing users of the __GFP_NOFAIL are blockable, so warn
5266 		 * of any new users that actually require GFP_NOWAIT
5267 		 */
5268 		if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5269 			goto fail;
5270 
5271 		/*
5272 		 * PF_MEMALLOC request from this context is rather bizarre
5273 		 * because we cannot reclaim anything and only can loop waiting
5274 		 * for somebody to do a work for us
5275 		 */
5276 		WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5277 
5278 		/*
5279 		 * non failing costly orders are a hard requirement which we
5280 		 * are not prepared for much so let's warn about these users
5281 		 * so that we can identify them and convert them to something
5282 		 * else.
5283 		 */
5284 		WARN_ON_ONCE_GFP(costly_order, gfp_mask);
5285 
5286 		/*
5287 		 * Help non-failing allocations by giving them access to memory
5288 		 * reserves but do not use ALLOC_NO_WATERMARKS because this
5289 		 * could deplete whole memory reserves which would just make
5290 		 * the situation worse
5291 		 */
5292 		page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5293 		if (page)
5294 			goto got_pg;
5295 
5296 		cond_resched();
5297 		goto retry;
5298 	}
5299 fail:
5300 	warn_alloc(gfp_mask, ac->nodemask,
5301 			"page allocation failure: order:%u", order);
5302 got_pg:
5303 	return page;
5304 }
5305 
prepare_alloc_pages(gfp_t gfp_mask,unsigned int order,int preferred_nid,nodemask_t * nodemask,struct alloc_context * ac,gfp_t * alloc_gfp,unsigned int * alloc_flags)5306 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5307 		int preferred_nid, nodemask_t *nodemask,
5308 		struct alloc_context *ac, gfp_t *alloc_gfp,
5309 		unsigned int *alloc_flags)
5310 {
5311 	ac->highest_zoneidx = gfp_zone(gfp_mask);
5312 	ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5313 	ac->nodemask = nodemask;
5314 	ac->migratetype = gfp_migratetype(gfp_mask);
5315 
5316 	if (cpusets_enabled()) {
5317 		*alloc_gfp |= __GFP_HARDWALL;
5318 		/*
5319 		 * When we are in the interrupt context, it is irrelevant
5320 		 * to the current task context. It means that any node ok.
5321 		 */
5322 		if (in_task() && !ac->nodemask)
5323 			ac->nodemask = &cpuset_current_mems_allowed;
5324 		else
5325 			*alloc_flags |= ALLOC_CPUSET;
5326 	}
5327 
5328 	might_alloc(gfp_mask);
5329 
5330 	if (should_fail_alloc_page(gfp_mask, order))
5331 		return false;
5332 
5333 	*alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5334 
5335 	/* Dirty zone balancing only done in the fast path */
5336 	ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5337 
5338 	/*
5339 	 * The preferred zone is used for statistics but crucially it is
5340 	 * also used as the starting point for the zonelist iterator. It
5341 	 * may get reset for allocations that ignore memory policies.
5342 	 */
5343 	ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5344 					ac->highest_zoneidx, ac->nodemask);
5345 
5346 	return true;
5347 }
5348 
5349 /*
5350  * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5351  * @gfp: GFP flags for the allocation
5352  * @preferred_nid: The preferred NUMA node ID to allocate from
5353  * @nodemask: Set of nodes to allocate from, may be NULL
5354  * @nr_pages: The number of pages desired on the list or array
5355  * @page_list: Optional list to store the allocated pages
5356  * @page_array: Optional array to store the pages
5357  *
5358  * This is a batched version of the page allocator that attempts to
5359  * allocate nr_pages quickly. Pages are added to page_list if page_list
5360  * is not NULL, otherwise it is assumed that the page_array is valid.
5361  *
5362  * For lists, nr_pages is the number of pages that should be allocated.
5363  *
5364  * For arrays, only NULL elements are populated with pages and nr_pages
5365  * is the maximum number of pages that will be stored in the array.
5366  *
5367  * Returns the number of pages on the list or array.
5368  */
__alloc_pages_bulk(gfp_t gfp,int preferred_nid,nodemask_t * nodemask,int nr_pages,struct list_head * page_list,struct page ** page_array)5369 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5370 			nodemask_t *nodemask, int nr_pages,
5371 			struct list_head *page_list,
5372 			struct page **page_array)
5373 {
5374 	struct page *page;
5375 	unsigned long flags;
5376 	unsigned long __maybe_unused UP_flags;
5377 	struct zone *zone;
5378 	struct zoneref *z;
5379 	struct per_cpu_pages *pcp;
5380 	struct list_head *pcp_list;
5381 	struct alloc_context ac;
5382 	gfp_t alloc_gfp;
5383 	unsigned int alloc_flags = ALLOC_WMARK_LOW;
5384 	int nr_populated = 0, nr_account = 0;
5385 
5386 	/*
5387 	 * Skip populated array elements to determine if any pages need
5388 	 * to be allocated before disabling IRQs.
5389 	 */
5390 	while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5391 		nr_populated++;
5392 
5393 	/* No pages requested? */
5394 	if (unlikely(nr_pages <= 0))
5395 		goto out;
5396 
5397 	/* Already populated array? */
5398 	if (unlikely(page_array && nr_pages - nr_populated == 0))
5399 		goto out;
5400 
5401 	/* Bulk allocator does not support memcg accounting. */
5402 	if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5403 		goto failed;
5404 
5405 	/* Use the single page allocator for one page. */
5406 	if (nr_pages - nr_populated == 1)
5407 		goto failed;
5408 
5409 #ifdef CONFIG_PAGE_OWNER
5410 	/*
5411 	 * PAGE_OWNER may recurse into the allocator to allocate space to
5412 	 * save the stack with pagesets.lock held. Releasing/reacquiring
5413 	 * removes much of the performance benefit of bulk allocation so
5414 	 * force the caller to allocate one page at a time as it'll have
5415 	 * similar performance to added complexity to the bulk allocator.
5416 	 */
5417 	if (static_branch_unlikely(&page_owner_inited))
5418 		goto failed;
5419 #endif
5420 
5421 	/* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5422 	gfp &= gfp_allowed_mask;
5423 	alloc_gfp = gfp;
5424 	if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5425 		goto out;
5426 	gfp = alloc_gfp;
5427 
5428 	/* Find an allowed local zone that meets the low watermark. */
5429 	for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5430 		unsigned long mark;
5431 
5432 		if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5433 		    !__cpuset_zone_allowed(zone, gfp)) {
5434 			continue;
5435 		}
5436 
5437 		if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5438 		    zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5439 			goto failed;
5440 		}
5441 
5442 		mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5443 		if (zone_watermark_fast(zone, 0,  mark,
5444 				zonelist_zone_idx(ac.preferred_zoneref),
5445 				alloc_flags, gfp)) {
5446 			break;
5447 		}
5448 	}
5449 
5450 	/*
5451 	 * If there are no allowed local zones that meets the watermarks then
5452 	 * try to allocate a single page and reclaim if necessary.
5453 	 */
5454 	if (unlikely(!zone))
5455 		goto failed;
5456 
5457 	/* Is a parallel drain in progress? */
5458 	pcp_trylock_prepare(UP_flags);
5459 	pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
5460 	if (!pcp)
5461 		goto failed_irq;
5462 
5463 	/* Attempt the batch allocation */
5464 	pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5465 	while (nr_populated < nr_pages) {
5466 
5467 		/* Skip existing pages */
5468 		if (page_array && page_array[nr_populated]) {
5469 			nr_populated++;
5470 			continue;
5471 		}
5472 
5473 		page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5474 								pcp, pcp_list);
5475 		if (unlikely(!page)) {
5476 			/* Try and allocate at least one page */
5477 			if (!nr_account) {
5478 				pcp_spin_unlock_irqrestore(pcp, flags);
5479 				goto failed_irq;
5480 			}
5481 			break;
5482 		}
5483 		nr_account++;
5484 
5485 		prep_new_page(page, 0, gfp, 0);
5486 		if (page_list)
5487 			list_add(&page->lru, page_list);
5488 		else
5489 			page_array[nr_populated] = page;
5490 		nr_populated++;
5491 	}
5492 
5493 	pcp_spin_unlock_irqrestore(pcp, flags);
5494 	pcp_trylock_finish(UP_flags);
5495 
5496 	__count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5497 	zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5498 
5499 out:
5500 	return nr_populated;
5501 
5502 failed_irq:
5503 	pcp_trylock_finish(UP_flags);
5504 
5505 failed:
5506 	page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5507 	if (page) {
5508 		if (page_list)
5509 			list_add(&page->lru, page_list);
5510 		else
5511 			page_array[nr_populated] = page;
5512 		nr_populated++;
5513 	}
5514 
5515 	goto out;
5516 }
5517 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5518 
5519 /*
5520  * This is the 'heart' of the zoned buddy allocator.
5521  */
__alloc_pages(gfp_t gfp,unsigned int order,int preferred_nid,nodemask_t * nodemask)5522 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5523 							nodemask_t *nodemask)
5524 {
5525 	struct page *page;
5526 	unsigned int alloc_flags = ALLOC_WMARK_LOW;
5527 	gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5528 	struct alloc_context ac = { };
5529 
5530 	/*
5531 	 * There are several places where we assume that the order value is sane
5532 	 * so bail out early if the request is out of bound.
5533 	 */
5534 	if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5535 		return NULL;
5536 
5537 	gfp &= gfp_allowed_mask;
5538 	/*
5539 	 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5540 	 * resp. GFP_NOIO which has to be inherited for all allocation requests
5541 	 * from a particular context which has been marked by
5542 	 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5543 	 * movable zones are not used during allocation.
5544 	 */
5545 	gfp = current_gfp_context(gfp);
5546 	alloc_gfp = gfp;
5547 	if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5548 			&alloc_gfp, &alloc_flags))
5549 		return NULL;
5550 
5551 	/*
5552 	 * Forbid the first pass from falling back to types that fragment
5553 	 * memory until all local zones are considered.
5554 	 */
5555 	alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5556 
5557 	/* First allocation attempt */
5558 	page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5559 	if (likely(page))
5560 		goto out;
5561 
5562 	alloc_gfp = gfp;
5563 	ac.spread_dirty_pages = false;
5564 
5565 	/*
5566 	 * Restore the original nodemask if it was potentially replaced with
5567 	 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5568 	 */
5569 	ac.nodemask = nodemask;
5570 
5571 	page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5572 
5573 out:
5574 	if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5575 	    unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5576 		__free_pages(page, order);
5577 		page = NULL;
5578 	}
5579 
5580 	trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5581 	kmsan_alloc_page(page, order, alloc_gfp);
5582 
5583 	return page;
5584 }
5585 EXPORT_SYMBOL(__alloc_pages);
5586 
__folio_alloc(gfp_t gfp,unsigned int order,int preferred_nid,nodemask_t * nodemask)5587 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5588 		nodemask_t *nodemask)
5589 {
5590 	struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5591 			preferred_nid, nodemask);
5592 
5593 	if (page && order > 1)
5594 		prep_transhuge_page(page);
5595 	return (struct folio *)page;
5596 }
5597 EXPORT_SYMBOL(__folio_alloc);
5598 
5599 /*
5600  * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5601  * address cannot represent highmem pages. Use alloc_pages and then kmap if
5602  * you need to access high mem.
5603  */
__get_free_pages(gfp_t gfp_mask,unsigned int order)5604 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5605 {
5606 	struct page *page;
5607 
5608 	page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5609 	if (!page)
5610 		return 0;
5611 	return (unsigned long) page_address(page);
5612 }
5613 EXPORT_SYMBOL(__get_free_pages);
5614 
get_zeroed_page(gfp_t gfp_mask)5615 unsigned long get_zeroed_page(gfp_t gfp_mask)
5616 {
5617 	return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5618 }
5619 EXPORT_SYMBOL(get_zeroed_page);
5620 
5621 /**
5622  * __free_pages - Free pages allocated with alloc_pages().
5623  * @page: The page pointer returned from alloc_pages().
5624  * @order: The order of the allocation.
5625  *
5626  * This function can free multi-page allocations that are not compound
5627  * pages.  It does not check that the @order passed in matches that of
5628  * the allocation, so it is easy to leak memory.  Freeing more memory
5629  * than was allocated will probably emit a warning.
5630  *
5631  * If the last reference to this page is speculative, it will be released
5632  * by put_page() which only frees the first page of a non-compound
5633  * allocation.  To prevent the remaining pages from being leaked, we free
5634  * the subsequent pages here.  If you want to use the page's reference
5635  * count to decide when to free the allocation, you should allocate a
5636  * compound page, and use put_page() instead of __free_pages().
5637  *
5638  * Context: May be called in interrupt context or while holding a normal
5639  * spinlock, but not in NMI context or while holding a raw spinlock.
5640  */
__free_pages(struct page * page,unsigned int order)5641 void __free_pages(struct page *page, unsigned int order)
5642 {
5643 	if (put_page_testzero(page))
5644 		free_the_page(page, order);
5645 	else if (!PageHead(page))
5646 		while (order-- > 0)
5647 			free_the_page(page + (1 << order), order);
5648 }
5649 EXPORT_SYMBOL(__free_pages);
5650 
free_pages(unsigned long addr,unsigned int order)5651 void free_pages(unsigned long addr, unsigned int order)
5652 {
5653 	if (addr != 0) {
5654 		VM_BUG_ON(!virt_addr_valid((void *)addr));
5655 		__free_pages(virt_to_page((void *)addr), order);
5656 	}
5657 }
5658 
5659 EXPORT_SYMBOL(free_pages);
5660 
5661 /*
5662  * Page Fragment:
5663  *  An arbitrary-length arbitrary-offset area of memory which resides
5664  *  within a 0 or higher order page.  Multiple fragments within that page
5665  *  are individually refcounted, in the page's reference counter.
5666  *
5667  * The page_frag functions below provide a simple allocation framework for
5668  * page fragments.  This is used by the network stack and network device
5669  * drivers to provide a backing region of memory for use as either an
5670  * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5671  */
__page_frag_cache_refill(struct page_frag_cache * nc,gfp_t gfp_mask)5672 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5673 					     gfp_t gfp_mask)
5674 {
5675 	struct page *page = NULL;
5676 	gfp_t gfp = gfp_mask;
5677 
5678 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5679 	gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5680 		    __GFP_NOMEMALLOC;
5681 	page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5682 				PAGE_FRAG_CACHE_MAX_ORDER);
5683 	nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5684 #endif
5685 	if (unlikely(!page))
5686 		page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5687 
5688 	nc->va = page ? page_address(page) : NULL;
5689 
5690 	return page;
5691 }
5692 
__page_frag_cache_drain(struct page * page,unsigned int count)5693 void __page_frag_cache_drain(struct page *page, unsigned int count)
5694 {
5695 	VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5696 
5697 	if (page_ref_sub_and_test(page, count))
5698 		free_the_page(page, compound_order(page));
5699 }
5700 EXPORT_SYMBOL(__page_frag_cache_drain);
5701 
page_frag_alloc_align(struct page_frag_cache * nc,unsigned int fragsz,gfp_t gfp_mask,unsigned int align_mask)5702 void *page_frag_alloc_align(struct page_frag_cache *nc,
5703 		      unsigned int fragsz, gfp_t gfp_mask,
5704 		      unsigned int align_mask)
5705 {
5706 	unsigned int size = PAGE_SIZE;
5707 	struct page *page;
5708 	int offset;
5709 
5710 	if (unlikely(!nc->va)) {
5711 refill:
5712 		page = __page_frag_cache_refill(nc, gfp_mask);
5713 		if (!page)
5714 			return NULL;
5715 
5716 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5717 		/* if size can vary use size else just use PAGE_SIZE */
5718 		size = nc->size;
5719 #endif
5720 		/* Even if we own the page, we do not use atomic_set().
5721 		 * This would break get_page_unless_zero() users.
5722 		 */
5723 		page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5724 
5725 		/* reset page count bias and offset to start of new frag */
5726 		nc->pfmemalloc = page_is_pfmemalloc(page);
5727 		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5728 		nc->offset = size;
5729 	}
5730 
5731 	offset = nc->offset - fragsz;
5732 	if (unlikely(offset < 0)) {
5733 		page = virt_to_page(nc->va);
5734 
5735 		if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5736 			goto refill;
5737 
5738 		if (unlikely(nc->pfmemalloc)) {
5739 			free_the_page(page, compound_order(page));
5740 			goto refill;
5741 		}
5742 
5743 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5744 		/* if size can vary use size else just use PAGE_SIZE */
5745 		size = nc->size;
5746 #endif
5747 		/* OK, page count is 0, we can safely set it */
5748 		set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5749 
5750 		/* reset page count bias and offset to start of new frag */
5751 		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5752 		offset = size - fragsz;
5753 		if (unlikely(offset < 0)) {
5754 			/*
5755 			 * The caller is trying to allocate a fragment
5756 			 * with fragsz > PAGE_SIZE but the cache isn't big
5757 			 * enough to satisfy the request, this may
5758 			 * happen in low memory conditions.
5759 			 * We don't release the cache page because
5760 			 * it could make memory pressure worse
5761 			 * so we simply return NULL here.
5762 			 */
5763 			return NULL;
5764 		}
5765 	}
5766 
5767 	nc->pagecnt_bias--;
5768 	offset &= align_mask;
5769 	nc->offset = offset;
5770 
5771 	return nc->va + offset;
5772 }
5773 EXPORT_SYMBOL(page_frag_alloc_align);
5774 
5775 /*
5776  * Frees a page fragment allocated out of either a compound or order 0 page.
5777  */
page_frag_free(void * addr)5778 void page_frag_free(void *addr)
5779 {
5780 	struct page *page = virt_to_head_page(addr);
5781 
5782 	if (unlikely(put_page_testzero(page)))
5783 		free_the_page(page, compound_order(page));
5784 }
5785 EXPORT_SYMBOL(page_frag_free);
5786 
make_alloc_exact(unsigned long addr,unsigned int order,size_t size)5787 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5788 		size_t size)
5789 {
5790 	if (addr) {
5791 		unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5792 		struct page *page = virt_to_page((void *)addr);
5793 		struct page *last = page + nr;
5794 
5795 		split_page_owner(page, 1 << order);
5796 		split_page_memcg(page, 1 << order);
5797 		while (page < --last)
5798 			set_page_refcounted(last);
5799 
5800 		last = page + (1UL << order);
5801 		for (page += nr; page < last; page++)
5802 			__free_pages_ok(page, 0, FPI_TO_TAIL);
5803 	}
5804 	return (void *)addr;
5805 }
5806 
5807 /**
5808  * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5809  * @size: the number of bytes to allocate
5810  * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5811  *
5812  * This function is similar to alloc_pages(), except that it allocates the
5813  * minimum number of pages to satisfy the request.  alloc_pages() can only
5814  * allocate memory in power-of-two pages.
5815  *
5816  * This function is also limited by MAX_ORDER.
5817  *
5818  * Memory allocated by this function must be released by free_pages_exact().
5819  *
5820  * Return: pointer to the allocated area or %NULL in case of error.
5821  */
alloc_pages_exact(size_t size,gfp_t gfp_mask)5822 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5823 {
5824 	unsigned int order = get_order(size);
5825 	unsigned long addr;
5826 
5827 	if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5828 		gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5829 
5830 	addr = __get_free_pages(gfp_mask, order);
5831 	return make_alloc_exact(addr, order, size);
5832 }
5833 EXPORT_SYMBOL(alloc_pages_exact);
5834 
5835 /**
5836  * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5837  *			   pages on a node.
5838  * @nid: the preferred node ID where memory should be allocated
5839  * @size: the number of bytes to allocate
5840  * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5841  *
5842  * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5843  * back.
5844  *
5845  * Return: pointer to the allocated area or %NULL in case of error.
5846  */
alloc_pages_exact_nid(int nid,size_t size,gfp_t gfp_mask)5847 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5848 {
5849 	unsigned int order = get_order(size);
5850 	struct page *p;
5851 
5852 	if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5853 		gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5854 
5855 	p = alloc_pages_node(nid, gfp_mask, order);
5856 	if (!p)
5857 		return NULL;
5858 	return make_alloc_exact((unsigned long)page_address(p), order, size);
5859 }
5860 
5861 /**
5862  * free_pages_exact - release memory allocated via alloc_pages_exact()
5863  * @virt: the value returned by alloc_pages_exact.
5864  * @size: size of allocation, same value as passed to alloc_pages_exact().
5865  *
5866  * Release the memory allocated by a previous call to alloc_pages_exact.
5867  */
free_pages_exact(void * virt,size_t size)5868 void free_pages_exact(void *virt, size_t size)
5869 {
5870 	unsigned long addr = (unsigned long)virt;
5871 	unsigned long end = addr + PAGE_ALIGN(size);
5872 
5873 	while (addr < end) {
5874 		free_page(addr);
5875 		addr += PAGE_SIZE;
5876 	}
5877 }
5878 EXPORT_SYMBOL(free_pages_exact);
5879 
5880 /**
5881  * nr_free_zone_pages - count number of pages beyond high watermark
5882  * @offset: The zone index of the highest zone
5883  *
5884  * nr_free_zone_pages() counts the number of pages which are beyond the
5885  * high watermark within all zones at or below a given zone index.  For each
5886  * zone, the number of pages is calculated as:
5887  *
5888  *     nr_free_zone_pages = managed_pages - high_pages
5889  *
5890  * Return: number of pages beyond high watermark.
5891  */
nr_free_zone_pages(int offset)5892 static unsigned long nr_free_zone_pages(int offset)
5893 {
5894 	struct zoneref *z;
5895 	struct zone *zone;
5896 
5897 	/* Just pick one node, since fallback list is circular */
5898 	unsigned long sum = 0;
5899 
5900 	struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5901 
5902 	for_each_zone_zonelist(zone, z, zonelist, offset) {
5903 		unsigned long size = zone_managed_pages(zone);
5904 		unsigned long high = high_wmark_pages(zone);
5905 		if (size > high)
5906 			sum += size - high;
5907 	}
5908 
5909 	return sum;
5910 }
5911 
5912 /**
5913  * nr_free_buffer_pages - count number of pages beyond high watermark
5914  *
5915  * nr_free_buffer_pages() counts the number of pages which are beyond the high
5916  * watermark within ZONE_DMA and ZONE_NORMAL.
5917  *
5918  * Return: number of pages beyond high watermark within ZONE_DMA and
5919  * ZONE_NORMAL.
5920  */
nr_free_buffer_pages(void)5921 unsigned long nr_free_buffer_pages(void)
5922 {
5923 	return nr_free_zone_pages(gfp_zone(GFP_USER));
5924 }
5925 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5926 
show_node(struct zone * zone)5927 static inline void show_node(struct zone *zone)
5928 {
5929 	if (IS_ENABLED(CONFIG_NUMA))
5930 		printk("Node %d ", zone_to_nid(zone));
5931 }
5932 
si_mem_available(void)5933 long si_mem_available(void)
5934 {
5935 	long available;
5936 	unsigned long pagecache;
5937 	unsigned long wmark_low = 0;
5938 	unsigned long pages[NR_LRU_LISTS];
5939 	unsigned long reclaimable;
5940 	struct zone *zone;
5941 	int lru;
5942 
5943 	for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5944 		pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5945 
5946 	for_each_zone(zone)
5947 		wmark_low += low_wmark_pages(zone);
5948 
5949 	/*
5950 	 * Estimate the amount of memory available for userspace allocations,
5951 	 * without causing swapping or OOM.
5952 	 */
5953 	available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5954 
5955 	/*
5956 	 * Not all the page cache can be freed, otherwise the system will
5957 	 * start swapping or thrashing. Assume at least half of the page
5958 	 * cache, or the low watermark worth of cache, needs to stay.
5959 	 */
5960 	pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5961 	pagecache -= min(pagecache / 2, wmark_low);
5962 	available += pagecache;
5963 
5964 	/*
5965 	 * Part of the reclaimable slab and other kernel memory consists of
5966 	 * items that are in use, and cannot be freed. Cap this estimate at the
5967 	 * low watermark.
5968 	 */
5969 	reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5970 		global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5971 	available += reclaimable - min(reclaimable / 2, wmark_low);
5972 
5973 	if (available < 0)
5974 		available = 0;
5975 	return available;
5976 }
5977 EXPORT_SYMBOL_GPL(si_mem_available);
5978 
si_meminfo(struct sysinfo * val)5979 void si_meminfo(struct sysinfo *val)
5980 {
5981 	val->totalram = totalram_pages();
5982 	val->sharedram = global_node_page_state(NR_SHMEM);
5983 	val->freeram = global_zone_page_state(NR_FREE_PAGES);
5984 	val->bufferram = nr_blockdev_pages();
5985 	val->totalhigh = totalhigh_pages();
5986 	val->freehigh = nr_free_highpages();
5987 	val->mem_unit = PAGE_SIZE;
5988 }
5989 
5990 EXPORT_SYMBOL(si_meminfo);
5991 
5992 #ifdef CONFIG_NUMA
si_meminfo_node(struct sysinfo * val,int nid)5993 void si_meminfo_node(struct sysinfo *val, int nid)
5994 {
5995 	int zone_type;		/* needs to be signed */
5996 	unsigned long managed_pages = 0;
5997 	unsigned long managed_highpages = 0;
5998 	unsigned long free_highpages = 0;
5999 	pg_data_t *pgdat = NODE_DATA(nid);
6000 
6001 	for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
6002 		managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
6003 	val->totalram = managed_pages;
6004 	val->sharedram = node_page_state(pgdat, NR_SHMEM);
6005 	val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
6006 #ifdef CONFIG_HIGHMEM
6007 	for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
6008 		struct zone *zone = &pgdat->node_zones[zone_type];
6009 
6010 		if (is_highmem(zone)) {
6011 			managed_highpages += zone_managed_pages(zone);
6012 			free_highpages += zone_page_state(zone, NR_FREE_PAGES);
6013 		}
6014 	}
6015 	val->totalhigh = managed_highpages;
6016 	val->freehigh = free_highpages;
6017 #else
6018 	val->totalhigh = managed_highpages;
6019 	val->freehigh = free_highpages;
6020 #endif
6021 	val->mem_unit = PAGE_SIZE;
6022 }
6023 #endif
6024 
6025 /*
6026  * Determine whether the node should be displayed or not, depending on whether
6027  * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
6028  */
show_mem_node_skip(unsigned int flags,int nid,nodemask_t * nodemask)6029 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
6030 {
6031 	if (!(flags & SHOW_MEM_FILTER_NODES))
6032 		return false;
6033 
6034 	/*
6035 	 * no node mask - aka implicit memory numa policy. Do not bother with
6036 	 * the synchronization - read_mems_allowed_begin - because we do not
6037 	 * have to be precise here.
6038 	 */
6039 	if (!nodemask)
6040 		nodemask = &cpuset_current_mems_allowed;
6041 
6042 	return !node_isset(nid, *nodemask);
6043 }
6044 
6045 #define K(x) ((x) << (PAGE_SHIFT-10))
6046 
show_migration_types(unsigned char type)6047 static void show_migration_types(unsigned char type)
6048 {
6049 	static const char types[MIGRATE_TYPES] = {
6050 		[MIGRATE_UNMOVABLE]	= 'U',
6051 		[MIGRATE_MOVABLE]	= 'M',
6052 		[MIGRATE_RECLAIMABLE]	= 'E',
6053 		[MIGRATE_HIGHATOMIC]	= 'H',
6054 #ifdef CONFIG_CMA
6055 		[MIGRATE_CMA]		= 'C',
6056 #endif
6057 #ifdef CONFIG_MEMORY_ISOLATION
6058 		[MIGRATE_ISOLATE]	= 'I',
6059 #endif
6060 	};
6061 	char tmp[MIGRATE_TYPES + 1];
6062 	char *p = tmp;
6063 	int i;
6064 
6065 	for (i = 0; i < MIGRATE_TYPES; i++) {
6066 		if (type & (1 << i))
6067 			*p++ = types[i];
6068 	}
6069 
6070 	*p = '\0';
6071 	printk(KERN_CONT "(%s) ", tmp);
6072 }
6073 
node_has_managed_zones(pg_data_t * pgdat,int max_zone_idx)6074 static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
6075 {
6076 	int zone_idx;
6077 	for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
6078 		if (zone_managed_pages(pgdat->node_zones + zone_idx))
6079 			return true;
6080 	return false;
6081 }
6082 
6083 /*
6084  * Show free area list (used inside shift_scroll-lock stuff)
6085  * We also calculate the percentage fragmentation. We do this by counting the
6086  * memory on each free list with the exception of the first item on the list.
6087  *
6088  * Bits in @filter:
6089  * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6090  *   cpuset.
6091  */
__show_free_areas(unsigned int filter,nodemask_t * nodemask,int max_zone_idx)6092 void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
6093 {
6094 	unsigned long free_pcp = 0;
6095 	int cpu, nid;
6096 	struct zone *zone;
6097 	pg_data_t *pgdat;
6098 
6099 	for_each_populated_zone(zone) {
6100 		if (zone_idx(zone) > max_zone_idx)
6101 			continue;
6102 		if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6103 			continue;
6104 
6105 		for_each_online_cpu(cpu)
6106 			free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6107 	}
6108 
6109 	printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6110 		" active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6111 		" unevictable:%lu dirty:%lu writeback:%lu\n"
6112 		" slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6113 		" mapped:%lu shmem:%lu pagetables:%lu\n"
6114 		" sec_pagetables:%lu bounce:%lu\n"
6115 		" kernel_misc_reclaimable:%lu\n"
6116 		" free:%lu free_pcp:%lu free_cma:%lu\n",
6117 		global_node_page_state(NR_ACTIVE_ANON),
6118 		global_node_page_state(NR_INACTIVE_ANON),
6119 		global_node_page_state(NR_ISOLATED_ANON),
6120 		global_node_page_state(NR_ACTIVE_FILE),
6121 		global_node_page_state(NR_INACTIVE_FILE),
6122 		global_node_page_state(NR_ISOLATED_FILE),
6123 		global_node_page_state(NR_UNEVICTABLE),
6124 		global_node_page_state(NR_FILE_DIRTY),
6125 		global_node_page_state(NR_WRITEBACK),
6126 		global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6127 		global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6128 		global_node_page_state(NR_FILE_MAPPED),
6129 		global_node_page_state(NR_SHMEM),
6130 		global_node_page_state(NR_PAGETABLE),
6131 		global_node_page_state(NR_SECONDARY_PAGETABLE),
6132 		global_zone_page_state(NR_BOUNCE),
6133 		global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6134 		global_zone_page_state(NR_FREE_PAGES),
6135 		free_pcp,
6136 		global_zone_page_state(NR_FREE_CMA_PAGES));
6137 
6138 	for_each_online_pgdat(pgdat) {
6139 		if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6140 			continue;
6141 		if (!node_has_managed_zones(pgdat, max_zone_idx))
6142 			continue;
6143 
6144 		printk("Node %d"
6145 			" active_anon:%lukB"
6146 			" inactive_anon:%lukB"
6147 			" active_file:%lukB"
6148 			" inactive_file:%lukB"
6149 			" unevictable:%lukB"
6150 			" isolated(anon):%lukB"
6151 			" isolated(file):%lukB"
6152 			" mapped:%lukB"
6153 			" dirty:%lukB"
6154 			" writeback:%lukB"
6155 			" shmem:%lukB"
6156 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6157 			" shmem_thp: %lukB"
6158 			" shmem_pmdmapped: %lukB"
6159 			" anon_thp: %lukB"
6160 #endif
6161 			" writeback_tmp:%lukB"
6162 			" kernel_stack:%lukB"
6163 #ifdef CONFIG_SHADOW_CALL_STACK
6164 			" shadow_call_stack:%lukB"
6165 #endif
6166 			" pagetables:%lukB"
6167 			" sec_pagetables:%lukB"
6168 			" all_unreclaimable? %s"
6169 			"\n",
6170 			pgdat->node_id,
6171 			K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6172 			K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6173 			K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6174 			K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6175 			K(node_page_state(pgdat, NR_UNEVICTABLE)),
6176 			K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6177 			K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6178 			K(node_page_state(pgdat, NR_FILE_MAPPED)),
6179 			K(node_page_state(pgdat, NR_FILE_DIRTY)),
6180 			K(node_page_state(pgdat, NR_WRITEBACK)),
6181 			K(node_page_state(pgdat, NR_SHMEM)),
6182 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6183 			K(node_page_state(pgdat, NR_SHMEM_THPS)),
6184 			K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6185 			K(node_page_state(pgdat, NR_ANON_THPS)),
6186 #endif
6187 			K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6188 			node_page_state(pgdat, NR_KERNEL_STACK_KB),
6189 #ifdef CONFIG_SHADOW_CALL_STACK
6190 			node_page_state(pgdat, NR_KERNEL_SCS_KB),
6191 #endif
6192 			K(node_page_state(pgdat, NR_PAGETABLE)),
6193 			K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
6194 			pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6195 				"yes" : "no");
6196 	}
6197 
6198 	for_each_populated_zone(zone) {
6199 		int i;
6200 
6201 		if (zone_idx(zone) > max_zone_idx)
6202 			continue;
6203 		if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6204 			continue;
6205 
6206 		free_pcp = 0;
6207 		for_each_online_cpu(cpu)
6208 			free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6209 
6210 		show_node(zone);
6211 		printk(KERN_CONT
6212 			"%s"
6213 			" free:%lukB"
6214 			" boost:%lukB"
6215 			" min:%lukB"
6216 			" low:%lukB"
6217 			" high:%lukB"
6218 			" reserved_highatomic:%luKB"
6219 			" active_anon:%lukB"
6220 			" inactive_anon:%lukB"
6221 			" active_file:%lukB"
6222 			" inactive_file:%lukB"
6223 			" unevictable:%lukB"
6224 			" writepending:%lukB"
6225 			" present:%lukB"
6226 			" managed:%lukB"
6227 			" mlocked:%lukB"
6228 			" bounce:%lukB"
6229 			" free_pcp:%lukB"
6230 			" local_pcp:%ukB"
6231 			" free_cma:%lukB"
6232 			"\n",
6233 			zone->name,
6234 			K(zone_page_state(zone, NR_FREE_PAGES)),
6235 			K(zone->watermark_boost),
6236 			K(min_wmark_pages(zone)),
6237 			K(low_wmark_pages(zone)),
6238 			K(high_wmark_pages(zone)),
6239 			K(zone->nr_reserved_highatomic),
6240 			K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6241 			K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6242 			K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6243 			K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6244 			K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6245 			K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6246 			K(zone->present_pages),
6247 			K(zone_managed_pages(zone)),
6248 			K(zone_page_state(zone, NR_MLOCK)),
6249 			K(zone_page_state(zone, NR_BOUNCE)),
6250 			K(free_pcp),
6251 			K(this_cpu_read(zone->per_cpu_pageset->count)),
6252 			K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6253 		printk("lowmem_reserve[]:");
6254 		for (i = 0; i < MAX_NR_ZONES; i++)
6255 			printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6256 		printk(KERN_CONT "\n");
6257 	}
6258 
6259 	for_each_populated_zone(zone) {
6260 		unsigned int order;
6261 		unsigned long nr[MAX_ORDER], flags, total = 0;
6262 		unsigned char types[MAX_ORDER];
6263 
6264 		if (zone_idx(zone) > max_zone_idx)
6265 			continue;
6266 		if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6267 			continue;
6268 		show_node(zone);
6269 		printk(KERN_CONT "%s: ", zone->name);
6270 
6271 		spin_lock_irqsave(&zone->lock, flags);
6272 		for (order = 0; order < MAX_ORDER; order++) {
6273 			struct free_area *area = &zone->free_area[order];
6274 			int type;
6275 
6276 			nr[order] = area->nr_free;
6277 			total += nr[order] << order;
6278 
6279 			types[order] = 0;
6280 			for (type = 0; type < MIGRATE_TYPES; type++) {
6281 				if (!free_area_empty(area, type))
6282 					types[order] |= 1 << type;
6283 			}
6284 		}
6285 		spin_unlock_irqrestore(&zone->lock, flags);
6286 		for (order = 0; order < MAX_ORDER; order++) {
6287 			printk(KERN_CONT "%lu*%lukB ",
6288 			       nr[order], K(1UL) << order);
6289 			if (nr[order])
6290 				show_migration_types(types[order]);
6291 		}
6292 		printk(KERN_CONT "= %lukB\n", K(total));
6293 	}
6294 
6295 	for_each_online_node(nid) {
6296 		if (show_mem_node_skip(filter, nid, nodemask))
6297 			continue;
6298 		hugetlb_show_meminfo_node(nid);
6299 	}
6300 
6301 	printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6302 
6303 	show_swap_cache_info();
6304 }
6305 
zoneref_set_zone(struct zone * zone,struct zoneref * zoneref)6306 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6307 {
6308 	zoneref->zone = zone;
6309 	zoneref->zone_idx = zone_idx(zone);
6310 }
6311 
6312 /*
6313  * Builds allocation fallback zone lists.
6314  *
6315  * Add all populated zones of a node to the zonelist.
6316  */
build_zonerefs_node(pg_data_t * pgdat,struct zoneref * zonerefs)6317 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6318 {
6319 	struct zone *zone;
6320 	enum zone_type zone_type = MAX_NR_ZONES;
6321 	int nr_zones = 0;
6322 
6323 	do {
6324 		zone_type--;
6325 		zone = pgdat->node_zones + zone_type;
6326 		if (populated_zone(zone)) {
6327 			zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6328 			check_highest_zone(zone_type);
6329 		}
6330 	} while (zone_type);
6331 
6332 	return nr_zones;
6333 }
6334 
6335 #ifdef CONFIG_NUMA
6336 
__parse_numa_zonelist_order(char * s)6337 static int __parse_numa_zonelist_order(char *s)
6338 {
6339 	/*
6340 	 * We used to support different zonelists modes but they turned
6341 	 * out to be just not useful. Let's keep the warning in place
6342 	 * if somebody still use the cmd line parameter so that we do
6343 	 * not fail it silently
6344 	 */
6345 	if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6346 		pr_warn("Ignoring unsupported numa_zonelist_order value:  %s\n", s);
6347 		return -EINVAL;
6348 	}
6349 	return 0;
6350 }
6351 
6352 char numa_zonelist_order[] = "Node";
6353 
6354 /*
6355  * sysctl handler for numa_zonelist_order
6356  */
numa_zonelist_order_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6357 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6358 		void *buffer, size_t *length, loff_t *ppos)
6359 {
6360 	if (write)
6361 		return __parse_numa_zonelist_order(buffer);
6362 	return proc_dostring(table, write, buffer, length, ppos);
6363 }
6364 
6365 
6366 static int node_load[MAX_NUMNODES];
6367 
6368 /**
6369  * find_next_best_node - find the next node that should appear in a given node's fallback list
6370  * @node: node whose fallback list we're appending
6371  * @used_node_mask: nodemask_t of already used nodes
6372  *
6373  * We use a number of factors to determine which is the next node that should
6374  * appear on a given node's fallback list.  The node should not have appeared
6375  * already in @node's fallback list, and it should be the next closest node
6376  * according to the distance array (which contains arbitrary distance values
6377  * from each node to each node in the system), and should also prefer nodes
6378  * with no CPUs, since presumably they'll have very little allocation pressure
6379  * on them otherwise.
6380  *
6381  * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6382  */
find_next_best_node(int node,nodemask_t * used_node_mask)6383 int find_next_best_node(int node, nodemask_t *used_node_mask)
6384 {
6385 	int n, val;
6386 	int min_val = INT_MAX;
6387 	int best_node = NUMA_NO_NODE;
6388 
6389 	/* Use the local node if we haven't already */
6390 	if (!node_isset(node, *used_node_mask)) {
6391 		node_set(node, *used_node_mask);
6392 		return node;
6393 	}
6394 
6395 	for_each_node_state(n, N_MEMORY) {
6396 
6397 		/* Don't want a node to appear more than once */
6398 		if (node_isset(n, *used_node_mask))
6399 			continue;
6400 
6401 		/* Use the distance array to find the distance */
6402 		val = node_distance(node, n);
6403 
6404 		/* Penalize nodes under us ("prefer the next node") */
6405 		val += (n < node);
6406 
6407 		/* Give preference to headless and unused nodes */
6408 		if (!cpumask_empty(cpumask_of_node(n)))
6409 			val += PENALTY_FOR_NODE_WITH_CPUS;
6410 
6411 		/* Slight preference for less loaded node */
6412 		val *= MAX_NUMNODES;
6413 		val += node_load[n];
6414 
6415 		if (val < min_val) {
6416 			min_val = val;
6417 			best_node = n;
6418 		}
6419 	}
6420 
6421 	if (best_node >= 0)
6422 		node_set(best_node, *used_node_mask);
6423 
6424 	return best_node;
6425 }
6426 
6427 
6428 /*
6429  * Build zonelists ordered by node and zones within node.
6430  * This results in maximum locality--normal zone overflows into local
6431  * DMA zone, if any--but risks exhausting DMA zone.
6432  */
build_zonelists_in_node_order(pg_data_t * pgdat,int * node_order,unsigned nr_nodes)6433 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6434 		unsigned nr_nodes)
6435 {
6436 	struct zoneref *zonerefs;
6437 	int i;
6438 
6439 	zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6440 
6441 	for (i = 0; i < nr_nodes; i++) {
6442 		int nr_zones;
6443 
6444 		pg_data_t *node = NODE_DATA(node_order[i]);
6445 
6446 		nr_zones = build_zonerefs_node(node, zonerefs);
6447 		zonerefs += nr_zones;
6448 	}
6449 	zonerefs->zone = NULL;
6450 	zonerefs->zone_idx = 0;
6451 }
6452 
6453 /*
6454  * Build gfp_thisnode zonelists
6455  */
build_thisnode_zonelists(pg_data_t * pgdat)6456 static void build_thisnode_zonelists(pg_data_t *pgdat)
6457 {
6458 	struct zoneref *zonerefs;
6459 	int nr_zones;
6460 
6461 	zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6462 	nr_zones = build_zonerefs_node(pgdat, zonerefs);
6463 	zonerefs += nr_zones;
6464 	zonerefs->zone = NULL;
6465 	zonerefs->zone_idx = 0;
6466 }
6467 
6468 /*
6469  * Build zonelists ordered by zone and nodes within zones.
6470  * This results in conserving DMA zone[s] until all Normal memory is
6471  * exhausted, but results in overflowing to remote node while memory
6472  * may still exist in local DMA zone.
6473  */
6474 
build_zonelists(pg_data_t * pgdat)6475 static void build_zonelists(pg_data_t *pgdat)
6476 {
6477 	static int node_order[MAX_NUMNODES];
6478 	int node, nr_nodes = 0;
6479 	nodemask_t used_mask = NODE_MASK_NONE;
6480 	int local_node, prev_node;
6481 
6482 	/* NUMA-aware ordering of nodes */
6483 	local_node = pgdat->node_id;
6484 	prev_node = local_node;
6485 
6486 	memset(node_order, 0, sizeof(node_order));
6487 	while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6488 		/*
6489 		 * We don't want to pressure a particular node.
6490 		 * So adding penalty to the first node in same
6491 		 * distance group to make it round-robin.
6492 		 */
6493 		if (node_distance(local_node, node) !=
6494 		    node_distance(local_node, prev_node))
6495 			node_load[node] += 1;
6496 
6497 		node_order[nr_nodes++] = node;
6498 		prev_node = node;
6499 	}
6500 
6501 	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6502 	build_thisnode_zonelists(pgdat);
6503 	pr_info("Fallback order for Node %d: ", local_node);
6504 	for (node = 0; node < nr_nodes; node++)
6505 		pr_cont("%d ", node_order[node]);
6506 	pr_cont("\n");
6507 }
6508 
6509 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6510 /*
6511  * Return node id of node used for "local" allocations.
6512  * I.e., first node id of first zone in arg node's generic zonelist.
6513  * Used for initializing percpu 'numa_mem', which is used primarily
6514  * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6515  */
local_memory_node(int node)6516 int local_memory_node(int node)
6517 {
6518 	struct zoneref *z;
6519 
6520 	z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6521 				   gfp_zone(GFP_KERNEL),
6522 				   NULL);
6523 	return zone_to_nid(z->zone);
6524 }
6525 #endif
6526 
6527 static void setup_min_unmapped_ratio(void);
6528 static void setup_min_slab_ratio(void);
6529 #else	/* CONFIG_NUMA */
6530 
build_zonelists(pg_data_t * pgdat)6531 static void build_zonelists(pg_data_t *pgdat)
6532 {
6533 	int node, local_node;
6534 	struct zoneref *zonerefs;
6535 	int nr_zones;
6536 
6537 	local_node = pgdat->node_id;
6538 
6539 	zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6540 	nr_zones = build_zonerefs_node(pgdat, zonerefs);
6541 	zonerefs += nr_zones;
6542 
6543 	/*
6544 	 * Now we build the zonelist so that it contains the zones
6545 	 * of all the other nodes.
6546 	 * We don't want to pressure a particular node, so when
6547 	 * building the zones for node N, we make sure that the
6548 	 * zones coming right after the local ones are those from
6549 	 * node N+1 (modulo N)
6550 	 */
6551 	for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6552 		if (!node_online(node))
6553 			continue;
6554 		nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6555 		zonerefs += nr_zones;
6556 	}
6557 	for (node = 0; node < local_node; node++) {
6558 		if (!node_online(node))
6559 			continue;
6560 		nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6561 		zonerefs += nr_zones;
6562 	}
6563 
6564 	zonerefs->zone = NULL;
6565 	zonerefs->zone_idx = 0;
6566 }
6567 
6568 #endif	/* CONFIG_NUMA */
6569 
6570 /*
6571  * Boot pageset table. One per cpu which is going to be used for all
6572  * zones and all nodes. The parameters will be set in such a way
6573  * that an item put on a list will immediately be handed over to
6574  * the buddy list. This is safe since pageset manipulation is done
6575  * with interrupts disabled.
6576  *
6577  * The boot_pagesets must be kept even after bootup is complete for
6578  * unused processors and/or zones. They do play a role for bootstrapping
6579  * hotplugged processors.
6580  *
6581  * zoneinfo_show() and maybe other functions do
6582  * not check if the processor is online before following the pageset pointer.
6583  * Other parts of the kernel may not check if the zone is available.
6584  */
6585 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6586 /* These effectively disable the pcplists in the boot pageset completely */
6587 #define BOOT_PAGESET_HIGH	0
6588 #define BOOT_PAGESET_BATCH	1
6589 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6590 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6591 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6592 
__build_all_zonelists(void * data)6593 static void __build_all_zonelists(void *data)
6594 {
6595 	int nid;
6596 	int __maybe_unused cpu;
6597 	pg_data_t *self = data;
6598 
6599 	write_seqlock(&zonelist_update_seq);
6600 
6601 #ifdef CONFIG_NUMA
6602 	memset(node_load, 0, sizeof(node_load));
6603 #endif
6604 
6605 	/*
6606 	 * This node is hotadded and no memory is yet present.   So just
6607 	 * building zonelists is fine - no need to touch other nodes.
6608 	 */
6609 	if (self && !node_online(self->node_id)) {
6610 		build_zonelists(self);
6611 	} else {
6612 		/*
6613 		 * All possible nodes have pgdat preallocated
6614 		 * in free_area_init
6615 		 */
6616 		for_each_node(nid) {
6617 			pg_data_t *pgdat = NODE_DATA(nid);
6618 
6619 			build_zonelists(pgdat);
6620 		}
6621 
6622 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6623 		/*
6624 		 * We now know the "local memory node" for each node--
6625 		 * i.e., the node of the first zone in the generic zonelist.
6626 		 * Set up numa_mem percpu variable for on-line cpus.  During
6627 		 * boot, only the boot cpu should be on-line;  we'll init the
6628 		 * secondary cpus' numa_mem as they come on-line.  During
6629 		 * node/memory hotplug, we'll fixup all on-line cpus.
6630 		 */
6631 		for_each_online_cpu(cpu)
6632 			set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6633 #endif
6634 	}
6635 
6636 	write_sequnlock(&zonelist_update_seq);
6637 }
6638 
6639 static noinline void __init
build_all_zonelists_init(void)6640 build_all_zonelists_init(void)
6641 {
6642 	int cpu;
6643 
6644 	__build_all_zonelists(NULL);
6645 
6646 	/*
6647 	 * Initialize the boot_pagesets that are going to be used
6648 	 * for bootstrapping processors. The real pagesets for
6649 	 * each zone will be allocated later when the per cpu
6650 	 * allocator is available.
6651 	 *
6652 	 * boot_pagesets are used also for bootstrapping offline
6653 	 * cpus if the system is already booted because the pagesets
6654 	 * are needed to initialize allocators on a specific cpu too.
6655 	 * F.e. the percpu allocator needs the page allocator which
6656 	 * needs the percpu allocator in order to allocate its pagesets
6657 	 * (a chicken-egg dilemma).
6658 	 */
6659 	for_each_possible_cpu(cpu)
6660 		per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6661 
6662 	mminit_verify_zonelist();
6663 	cpuset_init_current_mems_allowed();
6664 }
6665 
6666 /*
6667  * unless system_state == SYSTEM_BOOTING.
6668  *
6669  * __ref due to call of __init annotated helper build_all_zonelists_init
6670  * [protected by SYSTEM_BOOTING].
6671  */
build_all_zonelists(pg_data_t * pgdat)6672 void __ref build_all_zonelists(pg_data_t *pgdat)
6673 {
6674 	unsigned long vm_total_pages;
6675 
6676 	if (system_state == SYSTEM_BOOTING) {
6677 		build_all_zonelists_init();
6678 	} else {
6679 		__build_all_zonelists(pgdat);
6680 		/* cpuset refresh routine should be here */
6681 	}
6682 	/* Get the number of free pages beyond high watermark in all zones. */
6683 	vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6684 	/*
6685 	 * Disable grouping by mobility if the number of pages in the
6686 	 * system is too low to allow the mechanism to work. It would be
6687 	 * more accurate, but expensive to check per-zone. This check is
6688 	 * made on memory-hotadd so a system can start with mobility
6689 	 * disabled and enable it later
6690 	 */
6691 	if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6692 		page_group_by_mobility_disabled = 1;
6693 	else
6694 		page_group_by_mobility_disabled = 0;
6695 
6696 	pr_info("Built %u zonelists, mobility grouping %s.  Total pages: %ld\n",
6697 		nr_online_nodes,
6698 		page_group_by_mobility_disabled ? "off" : "on",
6699 		vm_total_pages);
6700 #ifdef CONFIG_NUMA
6701 	pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6702 #endif
6703 }
6704 
6705 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6706 static bool __meminit
overlap_memmap_init(unsigned long zone,unsigned long * pfn)6707 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6708 {
6709 	static struct memblock_region *r;
6710 
6711 	if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6712 		if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6713 			for_each_mem_region(r) {
6714 				if (*pfn < memblock_region_memory_end_pfn(r))
6715 					break;
6716 			}
6717 		}
6718 		if (*pfn >= memblock_region_memory_base_pfn(r) &&
6719 		    memblock_is_mirror(r)) {
6720 			*pfn = memblock_region_memory_end_pfn(r);
6721 			return true;
6722 		}
6723 	}
6724 	return false;
6725 }
6726 
6727 /*
6728  * Initially all pages are reserved - free ones are freed
6729  * up by memblock_free_all() once the early boot process is
6730  * done. Non-atomic initialization, single-pass.
6731  *
6732  * All aligned pageblocks are initialized to the specified migratetype
6733  * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6734  * zone stats (e.g., nr_isolate_pageblock) are touched.
6735  */
memmap_init_range(unsigned long size,int nid,unsigned long zone,unsigned long start_pfn,unsigned long zone_end_pfn,enum meminit_context context,struct vmem_altmap * altmap,int migratetype)6736 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6737 		unsigned long start_pfn, unsigned long zone_end_pfn,
6738 		enum meminit_context context,
6739 		struct vmem_altmap *altmap, int migratetype)
6740 {
6741 	unsigned long pfn, end_pfn = start_pfn + size;
6742 	struct page *page;
6743 
6744 	if (highest_memmap_pfn < end_pfn - 1)
6745 		highest_memmap_pfn = end_pfn - 1;
6746 
6747 #ifdef CONFIG_ZONE_DEVICE
6748 	/*
6749 	 * Honor reservation requested by the driver for this ZONE_DEVICE
6750 	 * memory. We limit the total number of pages to initialize to just
6751 	 * those that might contain the memory mapping. We will defer the
6752 	 * ZONE_DEVICE page initialization until after we have released
6753 	 * the hotplug lock.
6754 	 */
6755 	if (zone == ZONE_DEVICE) {
6756 		if (!altmap)
6757 			return;
6758 
6759 		if (start_pfn == altmap->base_pfn)
6760 			start_pfn += altmap->reserve;
6761 		end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6762 	}
6763 #endif
6764 
6765 	for (pfn = start_pfn; pfn < end_pfn; ) {
6766 		/*
6767 		 * There can be holes in boot-time mem_map[]s handed to this
6768 		 * function.  They do not exist on hotplugged memory.
6769 		 */
6770 		if (context == MEMINIT_EARLY) {
6771 			if (overlap_memmap_init(zone, &pfn))
6772 				continue;
6773 			if (defer_init(nid, pfn, zone_end_pfn))
6774 				break;
6775 		}
6776 
6777 		page = pfn_to_page(pfn);
6778 		__init_single_page(page, pfn, zone, nid);
6779 		if (context == MEMINIT_HOTPLUG)
6780 			__SetPageReserved(page);
6781 
6782 		/*
6783 		 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6784 		 * such that unmovable allocations won't be scattered all
6785 		 * over the place during system boot.
6786 		 */
6787 		if (pageblock_aligned(pfn)) {
6788 			set_pageblock_migratetype(page, migratetype);
6789 			cond_resched();
6790 		}
6791 		pfn++;
6792 	}
6793 }
6794 
6795 #ifdef CONFIG_ZONE_DEVICE
__init_zone_device_page(struct page * page,unsigned long pfn,unsigned long zone_idx,int nid,struct dev_pagemap * pgmap)6796 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6797 					  unsigned long zone_idx, int nid,
6798 					  struct dev_pagemap *pgmap)
6799 {
6800 
6801 	__init_single_page(page, pfn, zone_idx, nid);
6802 
6803 	/*
6804 	 * Mark page reserved as it will need to wait for onlining
6805 	 * phase for it to be fully associated with a zone.
6806 	 *
6807 	 * We can use the non-atomic __set_bit operation for setting
6808 	 * the flag as we are still initializing the pages.
6809 	 */
6810 	__SetPageReserved(page);
6811 
6812 	/*
6813 	 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6814 	 * and zone_device_data.  It is a bug if a ZONE_DEVICE page is
6815 	 * ever freed or placed on a driver-private list.
6816 	 */
6817 	page->pgmap = pgmap;
6818 	page->zone_device_data = NULL;
6819 
6820 	/*
6821 	 * Mark the block movable so that blocks are reserved for
6822 	 * movable at startup. This will force kernel allocations
6823 	 * to reserve their blocks rather than leaking throughout
6824 	 * the address space during boot when many long-lived
6825 	 * kernel allocations are made.
6826 	 *
6827 	 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6828 	 * because this is done early in section_activate()
6829 	 */
6830 	if (pageblock_aligned(pfn)) {
6831 		set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6832 		cond_resched();
6833 	}
6834 
6835 	/*
6836 	 * ZONE_DEVICE pages are released directly to the driver page allocator
6837 	 * which will set the page count to 1 when allocating the page.
6838 	 */
6839 	if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
6840 	    pgmap->type == MEMORY_DEVICE_COHERENT)
6841 		set_page_count(page, 0);
6842 }
6843 
6844 /*
6845  * With compound page geometry and when struct pages are stored in ram most
6846  * tail pages are reused. Consequently, the amount of unique struct pages to
6847  * initialize is a lot smaller that the total amount of struct pages being
6848  * mapped. This is a paired / mild layering violation with explicit knowledge
6849  * of how the sparse_vmemmap internals handle compound pages in the lack
6850  * of an altmap. See vmemmap_populate_compound_pages().
6851  */
compound_nr_pages(struct vmem_altmap * altmap,unsigned long nr_pages)6852 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6853 					      unsigned long nr_pages)
6854 {
6855 	return is_power_of_2(sizeof(struct page)) &&
6856 		!altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6857 }
6858 
memmap_init_compound(struct page * head,unsigned long head_pfn,unsigned long zone_idx,int nid,struct dev_pagemap * pgmap,unsigned long nr_pages)6859 static void __ref memmap_init_compound(struct page *head,
6860 				       unsigned long head_pfn,
6861 				       unsigned long zone_idx, int nid,
6862 				       struct dev_pagemap *pgmap,
6863 				       unsigned long nr_pages)
6864 {
6865 	unsigned long pfn, end_pfn = head_pfn + nr_pages;
6866 	unsigned int order = pgmap->vmemmap_shift;
6867 
6868 	__SetPageHead(head);
6869 	for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6870 		struct page *page = pfn_to_page(pfn);
6871 
6872 		__init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6873 		prep_compound_tail(head, pfn - head_pfn);
6874 		set_page_count(page, 0);
6875 
6876 		/*
6877 		 * The first tail page stores compound_mapcount_ptr() and
6878 		 * compound_order() and the second tail page stores
6879 		 * compound_pincount_ptr(). Call prep_compound_head() after
6880 		 * the first and second tail pages have been initialized to
6881 		 * not have the data overwritten.
6882 		 */
6883 		if (pfn == head_pfn + 2)
6884 			prep_compound_head(head, order);
6885 	}
6886 }
6887 
memmap_init_zone_device(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages,struct dev_pagemap * pgmap)6888 void __ref memmap_init_zone_device(struct zone *zone,
6889 				   unsigned long start_pfn,
6890 				   unsigned long nr_pages,
6891 				   struct dev_pagemap *pgmap)
6892 {
6893 	unsigned long pfn, end_pfn = start_pfn + nr_pages;
6894 	struct pglist_data *pgdat = zone->zone_pgdat;
6895 	struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6896 	unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6897 	unsigned long zone_idx = zone_idx(zone);
6898 	unsigned long start = jiffies;
6899 	int nid = pgdat->node_id;
6900 
6901 	if (WARN_ON_ONCE(!pgmap || zone_idx != ZONE_DEVICE))
6902 		return;
6903 
6904 	/*
6905 	 * The call to memmap_init should have already taken care
6906 	 * of the pages reserved for the memmap, so we can just jump to
6907 	 * the end of that region and start processing the device pages.
6908 	 */
6909 	if (altmap) {
6910 		start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6911 		nr_pages = end_pfn - start_pfn;
6912 	}
6913 
6914 	for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6915 		struct page *page = pfn_to_page(pfn);
6916 
6917 		__init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6918 
6919 		if (pfns_per_compound == 1)
6920 			continue;
6921 
6922 		memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6923 				     compound_nr_pages(altmap, pfns_per_compound));
6924 	}
6925 
6926 	pr_info("%s initialised %lu pages in %ums\n", __func__,
6927 		nr_pages, jiffies_to_msecs(jiffies - start));
6928 }
6929 
6930 #endif
zone_init_free_lists(struct zone * zone)6931 static void __meminit zone_init_free_lists(struct zone *zone)
6932 {
6933 	unsigned int order, t;
6934 	for_each_migratetype_order(order, t) {
6935 		INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6936 		zone->free_area[order].nr_free = 0;
6937 	}
6938 }
6939 
6940 /*
6941  * Only struct pages that correspond to ranges defined by memblock.memory
6942  * are zeroed and initialized by going through __init_single_page() during
6943  * memmap_init_zone_range().
6944  *
6945  * But, there could be struct pages that correspond to holes in
6946  * memblock.memory. This can happen because of the following reasons:
6947  * - physical memory bank size is not necessarily the exact multiple of the
6948  *   arbitrary section size
6949  * - early reserved memory may not be listed in memblock.memory
6950  * - memory layouts defined with memmap= kernel parameter may not align
6951  *   nicely with memmap sections
6952  *
6953  * Explicitly initialize those struct pages so that:
6954  * - PG_Reserved is set
6955  * - zone and node links point to zone and node that span the page if the
6956  *   hole is in the middle of a zone
6957  * - zone and node links point to adjacent zone/node if the hole falls on
6958  *   the zone boundary; the pages in such holes will be prepended to the
6959  *   zone/node above the hole except for the trailing pages in the last
6960  *   section that will be appended to the zone/node below.
6961  */
init_unavailable_range(unsigned long spfn,unsigned long epfn,int zone,int node)6962 static void __init init_unavailable_range(unsigned long spfn,
6963 					  unsigned long epfn,
6964 					  int zone, int node)
6965 {
6966 	unsigned long pfn;
6967 	u64 pgcnt = 0;
6968 
6969 	for (pfn = spfn; pfn < epfn; pfn++) {
6970 		if (!pfn_valid(pageblock_start_pfn(pfn))) {
6971 			pfn = pageblock_end_pfn(pfn) - 1;
6972 			continue;
6973 		}
6974 		__init_single_page(pfn_to_page(pfn), pfn, zone, node);
6975 		__SetPageReserved(pfn_to_page(pfn));
6976 		pgcnt++;
6977 	}
6978 
6979 	if (pgcnt)
6980 		pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6981 			node, zone_names[zone], pgcnt);
6982 }
6983 
memmap_init_zone_range(struct zone * zone,unsigned long start_pfn,unsigned long end_pfn,unsigned long * hole_pfn)6984 static void __init memmap_init_zone_range(struct zone *zone,
6985 					  unsigned long start_pfn,
6986 					  unsigned long end_pfn,
6987 					  unsigned long *hole_pfn)
6988 {
6989 	unsigned long zone_start_pfn = zone->zone_start_pfn;
6990 	unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6991 	int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6992 
6993 	start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6994 	end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6995 
6996 	if (start_pfn >= end_pfn)
6997 		return;
6998 
6999 	memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
7000 			  zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
7001 
7002 	if (*hole_pfn < start_pfn)
7003 		init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
7004 
7005 	*hole_pfn = end_pfn;
7006 }
7007 
memmap_init(void)7008 static void __init memmap_init(void)
7009 {
7010 	unsigned long start_pfn, end_pfn;
7011 	unsigned long hole_pfn = 0;
7012 	int i, j, zone_id = 0, nid;
7013 
7014 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7015 		struct pglist_data *node = NODE_DATA(nid);
7016 
7017 		for (j = 0; j < MAX_NR_ZONES; j++) {
7018 			struct zone *zone = node->node_zones + j;
7019 
7020 			if (!populated_zone(zone))
7021 				continue;
7022 
7023 			memmap_init_zone_range(zone, start_pfn, end_pfn,
7024 					       &hole_pfn);
7025 			zone_id = j;
7026 		}
7027 	}
7028 
7029 #ifdef CONFIG_SPARSEMEM
7030 	/*
7031 	 * Initialize the memory map for hole in the range [memory_end,
7032 	 * section_end].
7033 	 * Append the pages in this hole to the highest zone in the last
7034 	 * node.
7035 	 * The call to init_unavailable_range() is outside the ifdef to
7036 	 * silence the compiler warining about zone_id set but not used;
7037 	 * for FLATMEM it is a nop anyway
7038 	 */
7039 	end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
7040 	if (hole_pfn < end_pfn)
7041 #endif
7042 		init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
7043 }
7044 
memmap_alloc(phys_addr_t size,phys_addr_t align,phys_addr_t min_addr,int nid,bool exact_nid)7045 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
7046 			  phys_addr_t min_addr, int nid, bool exact_nid)
7047 {
7048 	void *ptr;
7049 
7050 	if (exact_nid)
7051 		ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
7052 						   MEMBLOCK_ALLOC_ACCESSIBLE,
7053 						   nid);
7054 	else
7055 		ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
7056 						 MEMBLOCK_ALLOC_ACCESSIBLE,
7057 						 nid);
7058 
7059 	if (ptr && size > 0)
7060 		page_init_poison(ptr, size);
7061 
7062 	return ptr;
7063 }
7064 
zone_batchsize(struct zone * zone)7065 static int zone_batchsize(struct zone *zone)
7066 {
7067 #ifdef CONFIG_MMU
7068 	int batch;
7069 
7070 	/*
7071 	 * The number of pages to batch allocate is either ~0.1%
7072 	 * of the zone or 1MB, whichever is smaller. The batch
7073 	 * size is striking a balance between allocation latency
7074 	 * and zone lock contention.
7075 	 */
7076 	batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
7077 	batch /= 4;		/* We effectively *= 4 below */
7078 	if (batch < 1)
7079 		batch = 1;
7080 
7081 	/*
7082 	 * Clamp the batch to a 2^n - 1 value. Having a power
7083 	 * of 2 value was found to be more likely to have
7084 	 * suboptimal cache aliasing properties in some cases.
7085 	 *
7086 	 * For example if 2 tasks are alternately allocating
7087 	 * batches of pages, one task can end up with a lot
7088 	 * of pages of one half of the possible page colors
7089 	 * and the other with pages of the other colors.
7090 	 */
7091 	batch = rounddown_pow_of_two(batch + batch/2) - 1;
7092 
7093 	return batch;
7094 
7095 #else
7096 	/* The deferral and batching of frees should be suppressed under NOMMU
7097 	 * conditions.
7098 	 *
7099 	 * The problem is that NOMMU needs to be able to allocate large chunks
7100 	 * of contiguous memory as there's no hardware page translation to
7101 	 * assemble apparent contiguous memory from discontiguous pages.
7102 	 *
7103 	 * Queueing large contiguous runs of pages for batching, however,
7104 	 * causes the pages to actually be freed in smaller chunks.  As there
7105 	 * can be a significant delay between the individual batches being
7106 	 * recycled, this leads to the once large chunks of space being
7107 	 * fragmented and becoming unavailable for high-order allocations.
7108 	 */
7109 	return 0;
7110 #endif
7111 }
7112 
zone_highsize(struct zone * zone,int batch,int cpu_online)7113 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7114 {
7115 #ifdef CONFIG_MMU
7116 	int high;
7117 	int nr_split_cpus;
7118 	unsigned long total_pages;
7119 
7120 	if (!percpu_pagelist_high_fraction) {
7121 		/*
7122 		 * By default, the high value of the pcp is based on the zone
7123 		 * low watermark so that if they are full then background
7124 		 * reclaim will not be started prematurely.
7125 		 */
7126 		total_pages = low_wmark_pages(zone);
7127 	} else {
7128 		/*
7129 		 * If percpu_pagelist_high_fraction is configured, the high
7130 		 * value is based on a fraction of the managed pages in the
7131 		 * zone.
7132 		 */
7133 		total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7134 	}
7135 
7136 	/*
7137 	 * Split the high value across all online CPUs local to the zone. Note
7138 	 * that early in boot that CPUs may not be online yet and that during
7139 	 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7140 	 * onlined. For memory nodes that have no CPUs, split pcp->high across
7141 	 * all online CPUs to mitigate the risk that reclaim is triggered
7142 	 * prematurely due to pages stored on pcp lists.
7143 	 */
7144 	nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7145 	if (!nr_split_cpus)
7146 		nr_split_cpus = num_online_cpus();
7147 	high = total_pages / nr_split_cpus;
7148 
7149 	/*
7150 	 * Ensure high is at least batch*4. The multiple is based on the
7151 	 * historical relationship between high and batch.
7152 	 */
7153 	high = max(high, batch << 2);
7154 
7155 	return high;
7156 #else
7157 	return 0;
7158 #endif
7159 }
7160 
7161 /*
7162  * pcp->high and pcp->batch values are related and generally batch is lower
7163  * than high. They are also related to pcp->count such that count is lower
7164  * than high, and as soon as it reaches high, the pcplist is flushed.
7165  *
7166  * However, guaranteeing these relations at all times would require e.g. write
7167  * barriers here but also careful usage of read barriers at the read side, and
7168  * thus be prone to error and bad for performance. Thus the update only prevents
7169  * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7170  * can cope with those fields changing asynchronously, and fully trust only the
7171  * pcp->count field on the local CPU with interrupts disabled.
7172  *
7173  * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7174  * outside of boot time (or some other assurance that no concurrent updaters
7175  * exist).
7176  */
pageset_update(struct per_cpu_pages * pcp,unsigned long high,unsigned long batch)7177 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7178 		unsigned long batch)
7179 {
7180 	WRITE_ONCE(pcp->batch, batch);
7181 	WRITE_ONCE(pcp->high, high);
7182 }
7183 
per_cpu_pages_init(struct per_cpu_pages * pcp,struct per_cpu_zonestat * pzstats)7184 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7185 {
7186 	int pindex;
7187 
7188 	memset(pcp, 0, sizeof(*pcp));
7189 	memset(pzstats, 0, sizeof(*pzstats));
7190 
7191 	spin_lock_init(&pcp->lock);
7192 	for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7193 		INIT_LIST_HEAD(&pcp->lists[pindex]);
7194 
7195 	/*
7196 	 * Set batch and high values safe for a boot pageset. A true percpu
7197 	 * pageset's initialization will update them subsequently. Here we don't
7198 	 * need to be as careful as pageset_update() as nobody can access the
7199 	 * pageset yet.
7200 	 */
7201 	pcp->high = BOOT_PAGESET_HIGH;
7202 	pcp->batch = BOOT_PAGESET_BATCH;
7203 	pcp->free_factor = 0;
7204 }
7205 
__zone_set_pageset_high_and_batch(struct zone * zone,unsigned long high,unsigned long batch)7206 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7207 		unsigned long batch)
7208 {
7209 	struct per_cpu_pages *pcp;
7210 	int cpu;
7211 
7212 	for_each_possible_cpu(cpu) {
7213 		pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7214 		pageset_update(pcp, high, batch);
7215 	}
7216 }
7217 
7218 /*
7219  * Calculate and set new high and batch values for all per-cpu pagesets of a
7220  * zone based on the zone's size.
7221  */
zone_set_pageset_high_and_batch(struct zone * zone,int cpu_online)7222 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7223 {
7224 	int new_high, new_batch;
7225 
7226 	new_batch = max(1, zone_batchsize(zone));
7227 	new_high = zone_highsize(zone, new_batch, cpu_online);
7228 
7229 	if (zone->pageset_high == new_high &&
7230 	    zone->pageset_batch == new_batch)
7231 		return;
7232 
7233 	zone->pageset_high = new_high;
7234 	zone->pageset_batch = new_batch;
7235 
7236 	__zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7237 }
7238 
setup_zone_pageset(struct zone * zone)7239 void __meminit setup_zone_pageset(struct zone *zone)
7240 {
7241 	int cpu;
7242 
7243 	/* Size may be 0 on !SMP && !NUMA */
7244 	if (sizeof(struct per_cpu_zonestat) > 0)
7245 		zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7246 
7247 	zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7248 	for_each_possible_cpu(cpu) {
7249 		struct per_cpu_pages *pcp;
7250 		struct per_cpu_zonestat *pzstats;
7251 
7252 		pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7253 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7254 		per_cpu_pages_init(pcp, pzstats);
7255 	}
7256 
7257 	zone_set_pageset_high_and_batch(zone, 0);
7258 }
7259 
7260 /*
7261  * The zone indicated has a new number of managed_pages; batch sizes and percpu
7262  * page high values need to be recalculated.
7263  */
zone_pcp_update(struct zone * zone,int cpu_online)7264 static void zone_pcp_update(struct zone *zone, int cpu_online)
7265 {
7266 	mutex_lock(&pcp_batch_high_lock);
7267 	zone_set_pageset_high_and_batch(zone, cpu_online);
7268 	mutex_unlock(&pcp_batch_high_lock);
7269 }
7270 
7271 /*
7272  * Allocate per cpu pagesets and initialize them.
7273  * Before this call only boot pagesets were available.
7274  */
setup_per_cpu_pageset(void)7275 void __init setup_per_cpu_pageset(void)
7276 {
7277 	struct pglist_data *pgdat;
7278 	struct zone *zone;
7279 	int __maybe_unused cpu;
7280 
7281 	for_each_populated_zone(zone)
7282 		setup_zone_pageset(zone);
7283 
7284 #ifdef CONFIG_NUMA
7285 	/*
7286 	 * Unpopulated zones continue using the boot pagesets.
7287 	 * The numa stats for these pagesets need to be reset.
7288 	 * Otherwise, they will end up skewing the stats of
7289 	 * the nodes these zones are associated with.
7290 	 */
7291 	for_each_possible_cpu(cpu) {
7292 		struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7293 		memset(pzstats->vm_numa_event, 0,
7294 		       sizeof(pzstats->vm_numa_event));
7295 	}
7296 #endif
7297 
7298 	for_each_online_pgdat(pgdat)
7299 		pgdat->per_cpu_nodestats =
7300 			alloc_percpu(struct per_cpu_nodestat);
7301 }
7302 
zone_pcp_init(struct zone * zone)7303 static __meminit void zone_pcp_init(struct zone *zone)
7304 {
7305 	/*
7306 	 * per cpu subsystem is not up at this point. The following code
7307 	 * relies on the ability of the linker to provide the
7308 	 * offset of a (static) per cpu variable into the per cpu area.
7309 	 */
7310 	zone->per_cpu_pageset = &boot_pageset;
7311 	zone->per_cpu_zonestats = &boot_zonestats;
7312 	zone->pageset_high = BOOT_PAGESET_HIGH;
7313 	zone->pageset_batch = BOOT_PAGESET_BATCH;
7314 
7315 	if (populated_zone(zone))
7316 		pr_debug("  %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7317 			 zone->present_pages, zone_batchsize(zone));
7318 }
7319 
init_currently_empty_zone(struct zone * zone,unsigned long zone_start_pfn,unsigned long size)7320 void __meminit init_currently_empty_zone(struct zone *zone,
7321 					unsigned long zone_start_pfn,
7322 					unsigned long size)
7323 {
7324 	struct pglist_data *pgdat = zone->zone_pgdat;
7325 	int zone_idx = zone_idx(zone) + 1;
7326 
7327 	if (zone_idx > pgdat->nr_zones)
7328 		pgdat->nr_zones = zone_idx;
7329 
7330 	zone->zone_start_pfn = zone_start_pfn;
7331 
7332 	mminit_dprintk(MMINIT_TRACE, "memmap_init",
7333 			"Initialising map node %d zone %lu pfns %lu -> %lu\n",
7334 			pgdat->node_id,
7335 			(unsigned long)zone_idx(zone),
7336 			zone_start_pfn, (zone_start_pfn + size));
7337 
7338 	zone_init_free_lists(zone);
7339 	zone->initialized = 1;
7340 }
7341 
7342 /**
7343  * get_pfn_range_for_nid - Return the start and end page frames for a node
7344  * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7345  * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7346  * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7347  *
7348  * It returns the start and end page frame of a node based on information
7349  * provided by memblock_set_node(). If called for a node
7350  * with no available memory, a warning is printed and the start and end
7351  * PFNs will be 0.
7352  */
get_pfn_range_for_nid(unsigned int nid,unsigned long * start_pfn,unsigned long * end_pfn)7353 void __init get_pfn_range_for_nid(unsigned int nid,
7354 			unsigned long *start_pfn, unsigned long *end_pfn)
7355 {
7356 	unsigned long this_start_pfn, this_end_pfn;
7357 	int i;
7358 
7359 	*start_pfn = -1UL;
7360 	*end_pfn = 0;
7361 
7362 	for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7363 		*start_pfn = min(*start_pfn, this_start_pfn);
7364 		*end_pfn = max(*end_pfn, this_end_pfn);
7365 	}
7366 
7367 	if (*start_pfn == -1UL)
7368 		*start_pfn = 0;
7369 }
7370 
7371 /*
7372  * This finds a zone that can be used for ZONE_MOVABLE pages. The
7373  * assumption is made that zones within a node are ordered in monotonic
7374  * increasing memory addresses so that the "highest" populated zone is used
7375  */
find_usable_zone_for_movable(void)7376 static void __init find_usable_zone_for_movable(void)
7377 {
7378 	int zone_index;
7379 	for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7380 		if (zone_index == ZONE_MOVABLE)
7381 			continue;
7382 
7383 		if (arch_zone_highest_possible_pfn[zone_index] >
7384 				arch_zone_lowest_possible_pfn[zone_index])
7385 			break;
7386 	}
7387 
7388 	VM_BUG_ON(zone_index == -1);
7389 	movable_zone = zone_index;
7390 }
7391 
7392 /*
7393  * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7394  * because it is sized independent of architecture. Unlike the other zones,
7395  * the starting point for ZONE_MOVABLE is not fixed. It may be different
7396  * in each node depending on the size of each node and how evenly kernelcore
7397  * is distributed. This helper function adjusts the zone ranges
7398  * provided by the architecture for a given node by using the end of the
7399  * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7400  * zones within a node are in order of monotonic increases memory addresses
7401  */
adjust_zone_range_for_zone_movable(int nid,unsigned long zone_type,unsigned long node_start_pfn,unsigned long node_end_pfn,unsigned long * zone_start_pfn,unsigned long * zone_end_pfn)7402 static void __init adjust_zone_range_for_zone_movable(int nid,
7403 					unsigned long zone_type,
7404 					unsigned long node_start_pfn,
7405 					unsigned long node_end_pfn,
7406 					unsigned long *zone_start_pfn,
7407 					unsigned long *zone_end_pfn)
7408 {
7409 	/* Only adjust if ZONE_MOVABLE is on this node */
7410 	if (zone_movable_pfn[nid]) {
7411 		/* Size ZONE_MOVABLE */
7412 		if (zone_type == ZONE_MOVABLE) {
7413 			*zone_start_pfn = zone_movable_pfn[nid];
7414 			*zone_end_pfn = min(node_end_pfn,
7415 				arch_zone_highest_possible_pfn[movable_zone]);
7416 
7417 		/* Adjust for ZONE_MOVABLE starting within this range */
7418 		} else if (!mirrored_kernelcore &&
7419 			*zone_start_pfn < zone_movable_pfn[nid] &&
7420 			*zone_end_pfn > zone_movable_pfn[nid]) {
7421 			*zone_end_pfn = zone_movable_pfn[nid];
7422 
7423 		/* Check if this whole range is within ZONE_MOVABLE */
7424 		} else if (*zone_start_pfn >= zone_movable_pfn[nid])
7425 			*zone_start_pfn = *zone_end_pfn;
7426 	}
7427 }
7428 
7429 /*
7430  * Return the number of pages a zone spans in a node, including holes
7431  * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7432  */
zone_spanned_pages_in_node(int nid,unsigned long zone_type,unsigned long node_start_pfn,unsigned long node_end_pfn,unsigned long * zone_start_pfn,unsigned long * zone_end_pfn)7433 static unsigned long __init zone_spanned_pages_in_node(int nid,
7434 					unsigned long zone_type,
7435 					unsigned long node_start_pfn,
7436 					unsigned long node_end_pfn,
7437 					unsigned long *zone_start_pfn,
7438 					unsigned long *zone_end_pfn)
7439 {
7440 	unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7441 	unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7442 	/* When hotadd a new node from cpu_up(), the node should be empty */
7443 	if (!node_start_pfn && !node_end_pfn)
7444 		return 0;
7445 
7446 	/* Get the start and end of the zone */
7447 	*zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7448 	*zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7449 	adjust_zone_range_for_zone_movable(nid, zone_type,
7450 				node_start_pfn, node_end_pfn,
7451 				zone_start_pfn, zone_end_pfn);
7452 
7453 	/* Check that this node has pages within the zone's required range */
7454 	if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7455 		return 0;
7456 
7457 	/* Move the zone boundaries inside the node if necessary */
7458 	*zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7459 	*zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7460 
7461 	/* Return the spanned pages */
7462 	return *zone_end_pfn - *zone_start_pfn;
7463 }
7464 
7465 /*
7466  * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7467  * then all holes in the requested range will be accounted for.
7468  */
__absent_pages_in_range(int nid,unsigned long range_start_pfn,unsigned long range_end_pfn)7469 unsigned long __init __absent_pages_in_range(int nid,
7470 				unsigned long range_start_pfn,
7471 				unsigned long range_end_pfn)
7472 {
7473 	unsigned long nr_absent = range_end_pfn - range_start_pfn;
7474 	unsigned long start_pfn, end_pfn;
7475 	int i;
7476 
7477 	for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7478 		start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7479 		end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7480 		nr_absent -= end_pfn - start_pfn;
7481 	}
7482 	return nr_absent;
7483 }
7484 
7485 /**
7486  * absent_pages_in_range - Return number of page frames in holes within a range
7487  * @start_pfn: The start PFN to start searching for holes
7488  * @end_pfn: The end PFN to stop searching for holes
7489  *
7490  * Return: the number of pages frames in memory holes within a range.
7491  */
absent_pages_in_range(unsigned long start_pfn,unsigned long end_pfn)7492 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7493 							unsigned long end_pfn)
7494 {
7495 	return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7496 }
7497 
7498 /* Return the number of page frames in holes in a zone on a node */
zone_absent_pages_in_node(int nid,unsigned long zone_type,unsigned long node_start_pfn,unsigned long node_end_pfn)7499 static unsigned long __init zone_absent_pages_in_node(int nid,
7500 					unsigned long zone_type,
7501 					unsigned long node_start_pfn,
7502 					unsigned long node_end_pfn)
7503 {
7504 	unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7505 	unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7506 	unsigned long zone_start_pfn, zone_end_pfn;
7507 	unsigned long nr_absent;
7508 
7509 	/* When hotadd a new node from cpu_up(), the node should be empty */
7510 	if (!node_start_pfn && !node_end_pfn)
7511 		return 0;
7512 
7513 	zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7514 	zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7515 
7516 	adjust_zone_range_for_zone_movable(nid, zone_type,
7517 			node_start_pfn, node_end_pfn,
7518 			&zone_start_pfn, &zone_end_pfn);
7519 	nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7520 
7521 	/*
7522 	 * ZONE_MOVABLE handling.
7523 	 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7524 	 * and vice versa.
7525 	 */
7526 	if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7527 		unsigned long start_pfn, end_pfn;
7528 		struct memblock_region *r;
7529 
7530 		for_each_mem_region(r) {
7531 			start_pfn = clamp(memblock_region_memory_base_pfn(r),
7532 					  zone_start_pfn, zone_end_pfn);
7533 			end_pfn = clamp(memblock_region_memory_end_pfn(r),
7534 					zone_start_pfn, zone_end_pfn);
7535 
7536 			if (zone_type == ZONE_MOVABLE &&
7537 			    memblock_is_mirror(r))
7538 				nr_absent += end_pfn - start_pfn;
7539 
7540 			if (zone_type == ZONE_NORMAL &&
7541 			    !memblock_is_mirror(r))
7542 				nr_absent += end_pfn - start_pfn;
7543 		}
7544 	}
7545 
7546 	return nr_absent;
7547 }
7548 
calculate_node_totalpages(struct pglist_data * pgdat,unsigned long node_start_pfn,unsigned long node_end_pfn)7549 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7550 						unsigned long node_start_pfn,
7551 						unsigned long node_end_pfn)
7552 {
7553 	unsigned long realtotalpages = 0, totalpages = 0;
7554 	enum zone_type i;
7555 
7556 	for (i = 0; i < MAX_NR_ZONES; i++) {
7557 		struct zone *zone = pgdat->node_zones + i;
7558 		unsigned long zone_start_pfn, zone_end_pfn;
7559 		unsigned long spanned, absent;
7560 		unsigned long size, real_size;
7561 
7562 		spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7563 						     node_start_pfn,
7564 						     node_end_pfn,
7565 						     &zone_start_pfn,
7566 						     &zone_end_pfn);
7567 		absent = zone_absent_pages_in_node(pgdat->node_id, i,
7568 						   node_start_pfn,
7569 						   node_end_pfn);
7570 
7571 		size = spanned;
7572 		real_size = size - absent;
7573 
7574 		if (size)
7575 			zone->zone_start_pfn = zone_start_pfn;
7576 		else
7577 			zone->zone_start_pfn = 0;
7578 		zone->spanned_pages = size;
7579 		zone->present_pages = real_size;
7580 #if defined(CONFIG_MEMORY_HOTPLUG)
7581 		zone->present_early_pages = real_size;
7582 #endif
7583 
7584 		totalpages += size;
7585 		realtotalpages += real_size;
7586 	}
7587 
7588 	pgdat->node_spanned_pages = totalpages;
7589 	pgdat->node_present_pages = realtotalpages;
7590 	pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7591 }
7592 
7593 #ifndef CONFIG_SPARSEMEM
7594 /*
7595  * Calculate the size of the zone->blockflags rounded to an unsigned long
7596  * Start by making sure zonesize is a multiple of pageblock_order by rounding
7597  * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7598  * round what is now in bits to nearest long in bits, then return it in
7599  * bytes.
7600  */
usemap_size(unsigned long zone_start_pfn,unsigned long zonesize)7601 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7602 {
7603 	unsigned long usemapsize;
7604 
7605 	zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7606 	usemapsize = roundup(zonesize, pageblock_nr_pages);
7607 	usemapsize = usemapsize >> pageblock_order;
7608 	usemapsize *= NR_PAGEBLOCK_BITS;
7609 	usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7610 
7611 	return usemapsize / 8;
7612 }
7613 
setup_usemap(struct zone * zone)7614 static void __ref setup_usemap(struct zone *zone)
7615 {
7616 	unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7617 					       zone->spanned_pages);
7618 	zone->pageblock_flags = NULL;
7619 	if (usemapsize) {
7620 		zone->pageblock_flags =
7621 			memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7622 					    zone_to_nid(zone));
7623 		if (!zone->pageblock_flags)
7624 			panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7625 			      usemapsize, zone->name, zone_to_nid(zone));
7626 	}
7627 }
7628 #else
setup_usemap(struct zone * zone)7629 static inline void setup_usemap(struct zone *zone) {}
7630 #endif /* CONFIG_SPARSEMEM */
7631 
7632 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7633 
7634 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
set_pageblock_order(void)7635 void __init set_pageblock_order(void)
7636 {
7637 	unsigned int order = MAX_ORDER - 1;
7638 
7639 	/* Check that pageblock_nr_pages has not already been setup */
7640 	if (pageblock_order)
7641 		return;
7642 
7643 	/* Don't let pageblocks exceed the maximum allocation granularity. */
7644 	if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7645 		order = HUGETLB_PAGE_ORDER;
7646 
7647 	/*
7648 	 * Assume the largest contiguous order of interest is a huge page.
7649 	 * This value may be variable depending on boot parameters on IA64 and
7650 	 * powerpc.
7651 	 */
7652 	pageblock_order = order;
7653 }
7654 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7655 
7656 /*
7657  * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7658  * is unused as pageblock_order is set at compile-time. See
7659  * include/linux/pageblock-flags.h for the values of pageblock_order based on
7660  * the kernel config
7661  */
set_pageblock_order(void)7662 void __init set_pageblock_order(void)
7663 {
7664 }
7665 
7666 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7667 
calc_memmap_size(unsigned long spanned_pages,unsigned long present_pages)7668 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7669 						unsigned long present_pages)
7670 {
7671 	unsigned long pages = spanned_pages;
7672 
7673 	/*
7674 	 * Provide a more accurate estimation if there are holes within
7675 	 * the zone and SPARSEMEM is in use. If there are holes within the
7676 	 * zone, each populated memory region may cost us one or two extra
7677 	 * memmap pages due to alignment because memmap pages for each
7678 	 * populated regions may not be naturally aligned on page boundary.
7679 	 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7680 	 */
7681 	if (spanned_pages > present_pages + (present_pages >> 4) &&
7682 	    IS_ENABLED(CONFIG_SPARSEMEM))
7683 		pages = present_pages;
7684 
7685 	return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7686 }
7687 
7688 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
pgdat_init_split_queue(struct pglist_data * pgdat)7689 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7690 {
7691 	struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7692 
7693 	spin_lock_init(&ds_queue->split_queue_lock);
7694 	INIT_LIST_HEAD(&ds_queue->split_queue);
7695 	ds_queue->split_queue_len = 0;
7696 }
7697 #else
pgdat_init_split_queue(struct pglist_data * pgdat)7698 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7699 #endif
7700 
7701 #ifdef CONFIG_COMPACTION
pgdat_init_kcompactd(struct pglist_data * pgdat)7702 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7703 {
7704 	init_waitqueue_head(&pgdat->kcompactd_wait);
7705 }
7706 #else
pgdat_init_kcompactd(struct pglist_data * pgdat)7707 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7708 #endif
7709 
pgdat_init_internals(struct pglist_data * pgdat)7710 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7711 {
7712 	int i;
7713 
7714 	pgdat_resize_init(pgdat);
7715 	pgdat_kswapd_lock_init(pgdat);
7716 
7717 	pgdat_init_split_queue(pgdat);
7718 	pgdat_init_kcompactd(pgdat);
7719 
7720 	init_waitqueue_head(&pgdat->kswapd_wait);
7721 	init_waitqueue_head(&pgdat->pfmemalloc_wait);
7722 
7723 	for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7724 		init_waitqueue_head(&pgdat->reclaim_wait[i]);
7725 
7726 	pgdat_page_ext_init(pgdat);
7727 	lruvec_init(&pgdat->__lruvec);
7728 }
7729 
zone_init_internals(struct zone * zone,enum zone_type idx,int nid,unsigned long remaining_pages)7730 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7731 							unsigned long remaining_pages)
7732 {
7733 	atomic_long_set(&zone->managed_pages, remaining_pages);
7734 	zone_set_nid(zone, nid);
7735 	zone->name = zone_names[idx];
7736 	zone->zone_pgdat = NODE_DATA(nid);
7737 	spin_lock_init(&zone->lock);
7738 	zone_seqlock_init(zone);
7739 	zone_pcp_init(zone);
7740 }
7741 
7742 /*
7743  * Set up the zone data structures
7744  * - init pgdat internals
7745  * - init all zones belonging to this node
7746  *
7747  * NOTE: this function is only called during memory hotplug
7748  */
7749 #ifdef CONFIG_MEMORY_HOTPLUG
free_area_init_core_hotplug(struct pglist_data * pgdat)7750 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7751 {
7752 	int nid = pgdat->node_id;
7753 	enum zone_type z;
7754 	int cpu;
7755 
7756 	pgdat_init_internals(pgdat);
7757 
7758 	if (pgdat->per_cpu_nodestats == &boot_nodestats)
7759 		pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7760 
7761 	/*
7762 	 * Reset the nr_zones, order and highest_zoneidx before reuse.
7763 	 * Note that kswapd will init kswapd_highest_zoneidx properly
7764 	 * when it starts in the near future.
7765 	 */
7766 	pgdat->nr_zones = 0;
7767 	pgdat->kswapd_order = 0;
7768 	pgdat->kswapd_highest_zoneidx = 0;
7769 	pgdat->node_start_pfn = 0;
7770 	for_each_online_cpu(cpu) {
7771 		struct per_cpu_nodestat *p;
7772 
7773 		p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7774 		memset(p, 0, sizeof(*p));
7775 	}
7776 
7777 	for (z = 0; z < MAX_NR_ZONES; z++)
7778 		zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7779 }
7780 #endif
7781 
7782 /*
7783  * Set up the zone data structures:
7784  *   - mark all pages reserved
7785  *   - mark all memory queues empty
7786  *   - clear the memory bitmaps
7787  *
7788  * NOTE: pgdat should get zeroed by caller.
7789  * NOTE: this function is only called during early init.
7790  */
free_area_init_core(struct pglist_data * pgdat)7791 static void __init free_area_init_core(struct pglist_data *pgdat)
7792 {
7793 	enum zone_type j;
7794 	int nid = pgdat->node_id;
7795 
7796 	pgdat_init_internals(pgdat);
7797 	pgdat->per_cpu_nodestats = &boot_nodestats;
7798 
7799 	for (j = 0; j < MAX_NR_ZONES; j++) {
7800 		struct zone *zone = pgdat->node_zones + j;
7801 		unsigned long size, freesize, memmap_pages;
7802 
7803 		size = zone->spanned_pages;
7804 		freesize = zone->present_pages;
7805 
7806 		/*
7807 		 * Adjust freesize so that it accounts for how much memory
7808 		 * is used by this zone for memmap. This affects the watermark
7809 		 * and per-cpu initialisations
7810 		 */
7811 		memmap_pages = calc_memmap_size(size, freesize);
7812 		if (!is_highmem_idx(j)) {
7813 			if (freesize >= memmap_pages) {
7814 				freesize -= memmap_pages;
7815 				if (memmap_pages)
7816 					pr_debug("  %s zone: %lu pages used for memmap\n",
7817 						 zone_names[j], memmap_pages);
7818 			} else
7819 				pr_warn("  %s zone: %lu memmap pages exceeds freesize %lu\n",
7820 					zone_names[j], memmap_pages, freesize);
7821 		}
7822 
7823 		/* Account for reserved pages */
7824 		if (j == 0 && freesize > dma_reserve) {
7825 			freesize -= dma_reserve;
7826 			pr_debug("  %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7827 		}
7828 
7829 		if (!is_highmem_idx(j))
7830 			nr_kernel_pages += freesize;
7831 		/* Charge for highmem memmap if there are enough kernel pages */
7832 		else if (nr_kernel_pages > memmap_pages * 2)
7833 			nr_kernel_pages -= memmap_pages;
7834 		nr_all_pages += freesize;
7835 
7836 		/*
7837 		 * Set an approximate value for lowmem here, it will be adjusted
7838 		 * when the bootmem allocator frees pages into the buddy system.
7839 		 * And all highmem pages will be managed by the buddy system.
7840 		 */
7841 		zone_init_internals(zone, j, nid, freesize);
7842 
7843 		if (!size)
7844 			continue;
7845 
7846 		set_pageblock_order();
7847 		setup_usemap(zone);
7848 		init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7849 	}
7850 }
7851 
7852 #ifdef CONFIG_FLATMEM
alloc_node_mem_map(struct pglist_data * pgdat)7853 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7854 {
7855 	unsigned long __maybe_unused start = 0;
7856 	unsigned long __maybe_unused offset = 0;
7857 
7858 	/* Skip empty nodes */
7859 	if (!pgdat->node_spanned_pages)
7860 		return;
7861 
7862 	start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7863 	offset = pgdat->node_start_pfn - start;
7864 	/* ia64 gets its own node_mem_map, before this, without bootmem */
7865 	if (!pgdat->node_mem_map) {
7866 		unsigned long size, end;
7867 		struct page *map;
7868 
7869 		/*
7870 		 * The zone's endpoints aren't required to be MAX_ORDER
7871 		 * aligned but the node_mem_map endpoints must be in order
7872 		 * for the buddy allocator to function correctly.
7873 		 */
7874 		end = pgdat_end_pfn(pgdat);
7875 		end = ALIGN(end, MAX_ORDER_NR_PAGES);
7876 		size =  (end - start) * sizeof(struct page);
7877 		map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7878 				   pgdat->node_id, false);
7879 		if (!map)
7880 			panic("Failed to allocate %ld bytes for node %d memory map\n",
7881 			      size, pgdat->node_id);
7882 		pgdat->node_mem_map = map + offset;
7883 	}
7884 	pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7885 				__func__, pgdat->node_id, (unsigned long)pgdat,
7886 				(unsigned long)pgdat->node_mem_map);
7887 #ifndef CONFIG_NUMA
7888 	/*
7889 	 * With no DISCONTIG, the global mem_map is just set as node 0's
7890 	 */
7891 	if (pgdat == NODE_DATA(0)) {
7892 		mem_map = NODE_DATA(0)->node_mem_map;
7893 		if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7894 			mem_map -= offset;
7895 	}
7896 #endif
7897 }
7898 #else
alloc_node_mem_map(struct pglist_data * pgdat)7899 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7900 #endif /* CONFIG_FLATMEM */
7901 
7902 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
pgdat_set_deferred_range(pg_data_t * pgdat)7903 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7904 {
7905 	pgdat->first_deferred_pfn = ULONG_MAX;
7906 }
7907 #else
pgdat_set_deferred_range(pg_data_t * pgdat)7908 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7909 #endif
7910 
free_area_init_node(int nid)7911 static void __init free_area_init_node(int nid)
7912 {
7913 	pg_data_t *pgdat = NODE_DATA(nid);
7914 	unsigned long start_pfn = 0;
7915 	unsigned long end_pfn = 0;
7916 
7917 	/* pg_data_t should be reset to zero when it's allocated */
7918 	WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7919 
7920 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7921 
7922 	pgdat->node_id = nid;
7923 	pgdat->node_start_pfn = start_pfn;
7924 	pgdat->per_cpu_nodestats = NULL;
7925 
7926 	if (start_pfn != end_pfn) {
7927 		pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7928 			(u64)start_pfn << PAGE_SHIFT,
7929 			end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7930 	} else {
7931 		pr_info("Initmem setup node %d as memoryless\n", nid);
7932 	}
7933 
7934 	calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7935 
7936 	alloc_node_mem_map(pgdat);
7937 	pgdat_set_deferred_range(pgdat);
7938 
7939 	free_area_init_core(pgdat);
7940 }
7941 
free_area_init_memoryless_node(int nid)7942 static void __init free_area_init_memoryless_node(int nid)
7943 {
7944 	free_area_init_node(nid);
7945 }
7946 
7947 #if MAX_NUMNODES > 1
7948 /*
7949  * Figure out the number of possible node ids.
7950  */
setup_nr_node_ids(void)7951 void __init setup_nr_node_ids(void)
7952 {
7953 	unsigned int highest;
7954 
7955 	highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7956 	nr_node_ids = highest + 1;
7957 }
7958 #endif
7959 
7960 /**
7961  * node_map_pfn_alignment - determine the maximum internode alignment
7962  *
7963  * This function should be called after node map is populated and sorted.
7964  * It calculates the maximum power of two alignment which can distinguish
7965  * all the nodes.
7966  *
7967  * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7968  * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
7969  * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
7970  * shifted, 1GiB is enough and this function will indicate so.
7971  *
7972  * This is used to test whether pfn -> nid mapping of the chosen memory
7973  * model has fine enough granularity to avoid incorrect mapping for the
7974  * populated node map.
7975  *
7976  * Return: the determined alignment in pfn's.  0 if there is no alignment
7977  * requirement (single node).
7978  */
node_map_pfn_alignment(void)7979 unsigned long __init node_map_pfn_alignment(void)
7980 {
7981 	unsigned long accl_mask = 0, last_end = 0;
7982 	unsigned long start, end, mask;
7983 	int last_nid = NUMA_NO_NODE;
7984 	int i, nid;
7985 
7986 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7987 		if (!start || last_nid < 0 || last_nid == nid) {
7988 			last_nid = nid;
7989 			last_end = end;
7990 			continue;
7991 		}
7992 
7993 		/*
7994 		 * Start with a mask granular enough to pin-point to the
7995 		 * start pfn and tick off bits one-by-one until it becomes
7996 		 * too coarse to separate the current node from the last.
7997 		 */
7998 		mask = ~((1 << __ffs(start)) - 1);
7999 		while (mask && last_end <= (start & (mask << 1)))
8000 			mask <<= 1;
8001 
8002 		/* accumulate all internode masks */
8003 		accl_mask |= mask;
8004 	}
8005 
8006 	/* convert mask to number of pages */
8007 	return ~accl_mask + 1;
8008 }
8009 
8010 /*
8011  * early_calculate_totalpages()
8012  * Sum pages in active regions for movable zone.
8013  * Populate N_MEMORY for calculating usable_nodes.
8014  */
early_calculate_totalpages(void)8015 static unsigned long __init early_calculate_totalpages(void)
8016 {
8017 	unsigned long totalpages = 0;
8018 	unsigned long start_pfn, end_pfn;
8019 	int i, nid;
8020 
8021 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8022 		unsigned long pages = end_pfn - start_pfn;
8023 
8024 		totalpages += pages;
8025 		if (pages)
8026 			node_set_state(nid, N_MEMORY);
8027 	}
8028 	return totalpages;
8029 }
8030 
8031 /*
8032  * Find the PFN the Movable zone begins in each node. Kernel memory
8033  * is spread evenly between nodes as long as the nodes have enough
8034  * memory. When they don't, some nodes will have more kernelcore than
8035  * others
8036  */
find_zone_movable_pfns_for_nodes(void)8037 static void __init find_zone_movable_pfns_for_nodes(void)
8038 {
8039 	int i, nid;
8040 	unsigned long usable_startpfn;
8041 	unsigned long kernelcore_node, kernelcore_remaining;
8042 	/* save the state before borrow the nodemask */
8043 	nodemask_t saved_node_state = node_states[N_MEMORY];
8044 	unsigned long totalpages = early_calculate_totalpages();
8045 	int usable_nodes = nodes_weight(node_states[N_MEMORY]);
8046 	struct memblock_region *r;
8047 
8048 	/* Need to find movable_zone earlier when movable_node is specified. */
8049 	find_usable_zone_for_movable();
8050 
8051 	/*
8052 	 * If movable_node is specified, ignore kernelcore and movablecore
8053 	 * options.
8054 	 */
8055 	if (movable_node_is_enabled()) {
8056 		for_each_mem_region(r) {
8057 			if (!memblock_is_hotpluggable(r))
8058 				continue;
8059 
8060 			nid = memblock_get_region_node(r);
8061 
8062 			usable_startpfn = PFN_DOWN(r->base);
8063 			zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8064 				min(usable_startpfn, zone_movable_pfn[nid]) :
8065 				usable_startpfn;
8066 		}
8067 
8068 		goto out2;
8069 	}
8070 
8071 	/*
8072 	 * If kernelcore=mirror is specified, ignore movablecore option
8073 	 */
8074 	if (mirrored_kernelcore) {
8075 		bool mem_below_4gb_not_mirrored = false;
8076 
8077 		for_each_mem_region(r) {
8078 			if (memblock_is_mirror(r))
8079 				continue;
8080 
8081 			nid = memblock_get_region_node(r);
8082 
8083 			usable_startpfn = memblock_region_memory_base_pfn(r);
8084 
8085 			if (usable_startpfn < PHYS_PFN(SZ_4G)) {
8086 				mem_below_4gb_not_mirrored = true;
8087 				continue;
8088 			}
8089 
8090 			zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8091 				min(usable_startpfn, zone_movable_pfn[nid]) :
8092 				usable_startpfn;
8093 		}
8094 
8095 		if (mem_below_4gb_not_mirrored)
8096 			pr_warn("This configuration results in unmirrored kernel memory.\n");
8097 
8098 		goto out2;
8099 	}
8100 
8101 	/*
8102 	 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8103 	 * amount of necessary memory.
8104 	 */
8105 	if (required_kernelcore_percent)
8106 		required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8107 				       10000UL;
8108 	if (required_movablecore_percent)
8109 		required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8110 					10000UL;
8111 
8112 	/*
8113 	 * If movablecore= was specified, calculate what size of
8114 	 * kernelcore that corresponds so that memory usable for
8115 	 * any allocation type is evenly spread. If both kernelcore
8116 	 * and movablecore are specified, then the value of kernelcore
8117 	 * will be used for required_kernelcore if it's greater than
8118 	 * what movablecore would have allowed.
8119 	 */
8120 	if (required_movablecore) {
8121 		unsigned long corepages;
8122 
8123 		/*
8124 		 * Round-up so that ZONE_MOVABLE is at least as large as what
8125 		 * was requested by the user
8126 		 */
8127 		required_movablecore =
8128 			roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8129 		required_movablecore = min(totalpages, required_movablecore);
8130 		corepages = totalpages - required_movablecore;
8131 
8132 		required_kernelcore = max(required_kernelcore, corepages);
8133 	}
8134 
8135 	/*
8136 	 * If kernelcore was not specified or kernelcore size is larger
8137 	 * than totalpages, there is no ZONE_MOVABLE.
8138 	 */
8139 	if (!required_kernelcore || required_kernelcore >= totalpages)
8140 		goto out;
8141 
8142 	/* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8143 	usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8144 
8145 restart:
8146 	/* Spread kernelcore memory as evenly as possible throughout nodes */
8147 	kernelcore_node = required_kernelcore / usable_nodes;
8148 	for_each_node_state(nid, N_MEMORY) {
8149 		unsigned long start_pfn, end_pfn;
8150 
8151 		/*
8152 		 * Recalculate kernelcore_node if the division per node
8153 		 * now exceeds what is necessary to satisfy the requested
8154 		 * amount of memory for the kernel
8155 		 */
8156 		if (required_kernelcore < kernelcore_node)
8157 			kernelcore_node = required_kernelcore / usable_nodes;
8158 
8159 		/*
8160 		 * As the map is walked, we track how much memory is usable
8161 		 * by the kernel using kernelcore_remaining. When it is
8162 		 * 0, the rest of the node is usable by ZONE_MOVABLE
8163 		 */
8164 		kernelcore_remaining = kernelcore_node;
8165 
8166 		/* Go through each range of PFNs within this node */
8167 		for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8168 			unsigned long size_pages;
8169 
8170 			start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8171 			if (start_pfn >= end_pfn)
8172 				continue;
8173 
8174 			/* Account for what is only usable for kernelcore */
8175 			if (start_pfn < usable_startpfn) {
8176 				unsigned long kernel_pages;
8177 				kernel_pages = min(end_pfn, usable_startpfn)
8178 								- start_pfn;
8179 
8180 				kernelcore_remaining -= min(kernel_pages,
8181 							kernelcore_remaining);
8182 				required_kernelcore -= min(kernel_pages,
8183 							required_kernelcore);
8184 
8185 				/* Continue if range is now fully accounted */
8186 				if (end_pfn <= usable_startpfn) {
8187 
8188 					/*
8189 					 * Push zone_movable_pfn to the end so
8190 					 * that if we have to rebalance
8191 					 * kernelcore across nodes, we will
8192 					 * not double account here
8193 					 */
8194 					zone_movable_pfn[nid] = end_pfn;
8195 					continue;
8196 				}
8197 				start_pfn = usable_startpfn;
8198 			}
8199 
8200 			/*
8201 			 * The usable PFN range for ZONE_MOVABLE is from
8202 			 * start_pfn->end_pfn. Calculate size_pages as the
8203 			 * number of pages used as kernelcore
8204 			 */
8205 			size_pages = end_pfn - start_pfn;
8206 			if (size_pages > kernelcore_remaining)
8207 				size_pages = kernelcore_remaining;
8208 			zone_movable_pfn[nid] = start_pfn + size_pages;
8209 
8210 			/*
8211 			 * Some kernelcore has been met, update counts and
8212 			 * break if the kernelcore for this node has been
8213 			 * satisfied
8214 			 */
8215 			required_kernelcore -= min(required_kernelcore,
8216 								size_pages);
8217 			kernelcore_remaining -= size_pages;
8218 			if (!kernelcore_remaining)
8219 				break;
8220 		}
8221 	}
8222 
8223 	/*
8224 	 * If there is still required_kernelcore, we do another pass with one
8225 	 * less node in the count. This will push zone_movable_pfn[nid] further
8226 	 * along on the nodes that still have memory until kernelcore is
8227 	 * satisfied
8228 	 */
8229 	usable_nodes--;
8230 	if (usable_nodes && required_kernelcore > usable_nodes)
8231 		goto restart;
8232 
8233 out2:
8234 	/* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8235 	for (nid = 0; nid < MAX_NUMNODES; nid++) {
8236 		unsigned long start_pfn, end_pfn;
8237 
8238 		zone_movable_pfn[nid] =
8239 			roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8240 
8241 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8242 		if (zone_movable_pfn[nid] >= end_pfn)
8243 			zone_movable_pfn[nid] = 0;
8244 	}
8245 
8246 out:
8247 	/* restore the node_state */
8248 	node_states[N_MEMORY] = saved_node_state;
8249 }
8250 
8251 /* Any regular or high memory on that node ? */
check_for_memory(pg_data_t * pgdat,int nid)8252 static void check_for_memory(pg_data_t *pgdat, int nid)
8253 {
8254 	enum zone_type zone_type;
8255 
8256 	for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8257 		struct zone *zone = &pgdat->node_zones[zone_type];
8258 		if (populated_zone(zone)) {
8259 			if (IS_ENABLED(CONFIG_HIGHMEM))
8260 				node_set_state(nid, N_HIGH_MEMORY);
8261 			if (zone_type <= ZONE_NORMAL)
8262 				node_set_state(nid, N_NORMAL_MEMORY);
8263 			break;
8264 		}
8265 	}
8266 }
8267 
8268 /*
8269  * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8270  * such cases we allow max_zone_pfn sorted in the descending order
8271  */
arch_has_descending_max_zone_pfns(void)8272 bool __weak arch_has_descending_max_zone_pfns(void)
8273 {
8274 	return false;
8275 }
8276 
8277 /**
8278  * free_area_init - Initialise all pg_data_t and zone data
8279  * @max_zone_pfn: an array of max PFNs for each zone
8280  *
8281  * This will call free_area_init_node() for each active node in the system.
8282  * Using the page ranges provided by memblock_set_node(), the size of each
8283  * zone in each node and their holes is calculated. If the maximum PFN
8284  * between two adjacent zones match, it is assumed that the zone is empty.
8285  * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8286  * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8287  * starts where the previous one ended. For example, ZONE_DMA32 starts
8288  * at arch_max_dma_pfn.
8289  */
free_area_init(unsigned long * max_zone_pfn)8290 void __init free_area_init(unsigned long *max_zone_pfn)
8291 {
8292 	unsigned long start_pfn, end_pfn;
8293 	int i, nid, zone;
8294 	bool descending;
8295 
8296 	/* Record where the zone boundaries are */
8297 	memset(arch_zone_lowest_possible_pfn, 0,
8298 				sizeof(arch_zone_lowest_possible_pfn));
8299 	memset(arch_zone_highest_possible_pfn, 0,
8300 				sizeof(arch_zone_highest_possible_pfn));
8301 
8302 	start_pfn = PHYS_PFN(memblock_start_of_DRAM());
8303 	descending = arch_has_descending_max_zone_pfns();
8304 
8305 	for (i = 0; i < MAX_NR_ZONES; i++) {
8306 		if (descending)
8307 			zone = MAX_NR_ZONES - i - 1;
8308 		else
8309 			zone = i;
8310 
8311 		if (zone == ZONE_MOVABLE)
8312 			continue;
8313 
8314 		end_pfn = max(max_zone_pfn[zone], start_pfn);
8315 		arch_zone_lowest_possible_pfn[zone] = start_pfn;
8316 		arch_zone_highest_possible_pfn[zone] = end_pfn;
8317 
8318 		start_pfn = end_pfn;
8319 	}
8320 
8321 	/* Find the PFNs that ZONE_MOVABLE begins at in each node */
8322 	memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8323 	find_zone_movable_pfns_for_nodes();
8324 
8325 	/* Print out the zone ranges */
8326 	pr_info("Zone ranges:\n");
8327 	for (i = 0; i < MAX_NR_ZONES; i++) {
8328 		if (i == ZONE_MOVABLE)
8329 			continue;
8330 		pr_info("  %-8s ", zone_names[i]);
8331 		if (arch_zone_lowest_possible_pfn[i] ==
8332 				arch_zone_highest_possible_pfn[i])
8333 			pr_cont("empty\n");
8334 		else
8335 			pr_cont("[mem %#018Lx-%#018Lx]\n",
8336 				(u64)arch_zone_lowest_possible_pfn[i]
8337 					<< PAGE_SHIFT,
8338 				((u64)arch_zone_highest_possible_pfn[i]
8339 					<< PAGE_SHIFT) - 1);
8340 	}
8341 
8342 	/* Print out the PFNs ZONE_MOVABLE begins at in each node */
8343 	pr_info("Movable zone start for each node\n");
8344 	for (i = 0; i < MAX_NUMNODES; i++) {
8345 		if (zone_movable_pfn[i])
8346 			pr_info("  Node %d: %#018Lx\n", i,
8347 			       (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8348 	}
8349 
8350 	/*
8351 	 * Print out the early node map, and initialize the
8352 	 * subsection-map relative to active online memory ranges to
8353 	 * enable future "sub-section" extensions of the memory map.
8354 	 */
8355 	pr_info("Early memory node ranges\n");
8356 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8357 		pr_info("  node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8358 			(u64)start_pfn << PAGE_SHIFT,
8359 			((u64)end_pfn << PAGE_SHIFT) - 1);
8360 		subsection_map_init(start_pfn, end_pfn - start_pfn);
8361 	}
8362 
8363 	/* Initialise every node */
8364 	mminit_verify_pageflags_layout();
8365 	setup_nr_node_ids();
8366 	for_each_node(nid) {
8367 		pg_data_t *pgdat;
8368 
8369 		if (!node_online(nid)) {
8370 			pr_info("Initializing node %d as memoryless\n", nid);
8371 
8372 			/* Allocator not initialized yet */
8373 			pgdat = arch_alloc_nodedata(nid);
8374 			if (!pgdat) {
8375 				pr_err("Cannot allocate %zuB for node %d.\n",
8376 						sizeof(*pgdat), nid);
8377 				continue;
8378 			}
8379 			arch_refresh_nodedata(nid, pgdat);
8380 			free_area_init_memoryless_node(nid);
8381 
8382 			/*
8383 			 * We do not want to confuse userspace by sysfs
8384 			 * files/directories for node without any memory
8385 			 * attached to it, so this node is not marked as
8386 			 * N_MEMORY and not marked online so that no sysfs
8387 			 * hierarchy will be created via register_one_node for
8388 			 * it. The pgdat will get fully initialized by
8389 			 * hotadd_init_pgdat() when memory is hotplugged into
8390 			 * this node.
8391 			 */
8392 			continue;
8393 		}
8394 
8395 		pgdat = NODE_DATA(nid);
8396 		free_area_init_node(nid);
8397 
8398 		/* Any memory on that node */
8399 		if (pgdat->node_present_pages)
8400 			node_set_state(nid, N_MEMORY);
8401 		check_for_memory(pgdat, nid);
8402 	}
8403 
8404 	memmap_init();
8405 }
8406 
cmdline_parse_core(char * p,unsigned long * core,unsigned long * percent)8407 static int __init cmdline_parse_core(char *p, unsigned long *core,
8408 				     unsigned long *percent)
8409 {
8410 	unsigned long long coremem;
8411 	char *endptr;
8412 
8413 	if (!p)
8414 		return -EINVAL;
8415 
8416 	/* Value may be a percentage of total memory, otherwise bytes */
8417 	coremem = simple_strtoull(p, &endptr, 0);
8418 	if (*endptr == '%') {
8419 		/* Paranoid check for percent values greater than 100 */
8420 		WARN_ON(coremem > 100);
8421 
8422 		*percent = coremem;
8423 	} else {
8424 		coremem = memparse(p, &p);
8425 		/* Paranoid check that UL is enough for the coremem value */
8426 		WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8427 
8428 		*core = coremem >> PAGE_SHIFT;
8429 		*percent = 0UL;
8430 	}
8431 	return 0;
8432 }
8433 
8434 /*
8435  * kernelcore=size sets the amount of memory for use for allocations that
8436  * cannot be reclaimed or migrated.
8437  */
cmdline_parse_kernelcore(char * p)8438 static int __init cmdline_parse_kernelcore(char *p)
8439 {
8440 	/* parse kernelcore=mirror */
8441 	if (parse_option_str(p, "mirror")) {
8442 		mirrored_kernelcore = true;
8443 		return 0;
8444 	}
8445 
8446 	return cmdline_parse_core(p, &required_kernelcore,
8447 				  &required_kernelcore_percent);
8448 }
8449 
8450 /*
8451  * movablecore=size sets the amount of memory for use for allocations that
8452  * can be reclaimed or migrated.
8453  */
cmdline_parse_movablecore(char * p)8454 static int __init cmdline_parse_movablecore(char *p)
8455 {
8456 	return cmdline_parse_core(p, &required_movablecore,
8457 				  &required_movablecore_percent);
8458 }
8459 
8460 early_param("kernelcore", cmdline_parse_kernelcore);
8461 early_param("movablecore", cmdline_parse_movablecore);
8462 
adjust_managed_page_count(struct page * page,long count)8463 void adjust_managed_page_count(struct page *page, long count)
8464 {
8465 	atomic_long_add(count, &page_zone(page)->managed_pages);
8466 	totalram_pages_add(count);
8467 #ifdef CONFIG_HIGHMEM
8468 	if (PageHighMem(page))
8469 		totalhigh_pages_add(count);
8470 #endif
8471 }
8472 EXPORT_SYMBOL(adjust_managed_page_count);
8473 
free_reserved_area(void * start,void * end,int poison,const char * s)8474 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8475 {
8476 	void *pos;
8477 	unsigned long pages = 0;
8478 
8479 	start = (void *)PAGE_ALIGN((unsigned long)start);
8480 	end = (void *)((unsigned long)end & PAGE_MASK);
8481 	for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8482 		struct page *page = virt_to_page(pos);
8483 		void *direct_map_addr;
8484 
8485 		/*
8486 		 * 'direct_map_addr' might be different from 'pos'
8487 		 * because some architectures' virt_to_page()
8488 		 * work with aliases.  Getting the direct map
8489 		 * address ensures that we get a _writeable_
8490 		 * alias for the memset().
8491 		 */
8492 		direct_map_addr = page_address(page);
8493 		/*
8494 		 * Perform a kasan-unchecked memset() since this memory
8495 		 * has not been initialized.
8496 		 */
8497 		direct_map_addr = kasan_reset_tag(direct_map_addr);
8498 		if ((unsigned int)poison <= 0xFF)
8499 			memset(direct_map_addr, poison, PAGE_SIZE);
8500 
8501 		free_reserved_page(page);
8502 	}
8503 
8504 	if (pages && s)
8505 		pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8506 
8507 	return pages;
8508 }
8509 
mem_init_print_info(void)8510 void __init mem_init_print_info(void)
8511 {
8512 	unsigned long physpages, codesize, datasize, rosize, bss_size;
8513 	unsigned long init_code_size, init_data_size;
8514 
8515 	physpages = get_num_physpages();
8516 	codesize = _etext - _stext;
8517 	datasize = _edata - _sdata;
8518 	rosize = __end_rodata - __start_rodata;
8519 	bss_size = __bss_stop - __bss_start;
8520 	init_data_size = __init_end - __init_begin;
8521 	init_code_size = _einittext - _sinittext;
8522 
8523 	/*
8524 	 * Detect special cases and adjust section sizes accordingly:
8525 	 * 1) .init.* may be embedded into .data sections
8526 	 * 2) .init.text.* may be out of [__init_begin, __init_end],
8527 	 *    please refer to arch/tile/kernel/vmlinux.lds.S.
8528 	 * 3) .rodata.* may be embedded into .text or .data sections.
8529 	 */
8530 #define adj_init_size(start, end, size, pos, adj) \
8531 	do { \
8532 		if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8533 			size -= adj; \
8534 	} while (0)
8535 
8536 	adj_init_size(__init_begin, __init_end, init_data_size,
8537 		     _sinittext, init_code_size);
8538 	adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8539 	adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8540 	adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8541 	adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8542 
8543 #undef	adj_init_size
8544 
8545 	pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8546 #ifdef	CONFIG_HIGHMEM
8547 		", %luK highmem"
8548 #endif
8549 		")\n",
8550 		K(nr_free_pages()), K(physpages),
8551 		codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
8552 		(init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
8553 		K(physpages - totalram_pages() - totalcma_pages),
8554 		K(totalcma_pages)
8555 #ifdef	CONFIG_HIGHMEM
8556 		, K(totalhigh_pages())
8557 #endif
8558 		);
8559 }
8560 
8561 /**
8562  * set_dma_reserve - set the specified number of pages reserved in the first zone
8563  * @new_dma_reserve: The number of pages to mark reserved
8564  *
8565  * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8566  * In the DMA zone, a significant percentage may be consumed by kernel image
8567  * and other unfreeable allocations which can skew the watermarks badly. This
8568  * function may optionally be used to account for unfreeable pages in the
8569  * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8570  * smaller per-cpu batchsize.
8571  */
set_dma_reserve(unsigned long new_dma_reserve)8572 void __init set_dma_reserve(unsigned long new_dma_reserve)
8573 {
8574 	dma_reserve = new_dma_reserve;
8575 }
8576 
page_alloc_cpu_dead(unsigned int cpu)8577 static int page_alloc_cpu_dead(unsigned int cpu)
8578 {
8579 	struct zone *zone;
8580 
8581 	lru_add_drain_cpu(cpu);
8582 	mlock_page_drain_remote(cpu);
8583 	drain_pages(cpu);
8584 
8585 	/*
8586 	 * Spill the event counters of the dead processor
8587 	 * into the current processors event counters.
8588 	 * This artificially elevates the count of the current
8589 	 * processor.
8590 	 */
8591 	vm_events_fold_cpu(cpu);
8592 
8593 	/*
8594 	 * Zero the differential counters of the dead processor
8595 	 * so that the vm statistics are consistent.
8596 	 *
8597 	 * This is only okay since the processor is dead and cannot
8598 	 * race with what we are doing.
8599 	 */
8600 	cpu_vm_stats_fold(cpu);
8601 
8602 	for_each_populated_zone(zone)
8603 		zone_pcp_update(zone, 0);
8604 
8605 	return 0;
8606 }
8607 
page_alloc_cpu_online(unsigned int cpu)8608 static int page_alloc_cpu_online(unsigned int cpu)
8609 {
8610 	struct zone *zone;
8611 
8612 	for_each_populated_zone(zone)
8613 		zone_pcp_update(zone, 1);
8614 	return 0;
8615 }
8616 
8617 #ifdef CONFIG_NUMA
8618 int hashdist = HASHDIST_DEFAULT;
8619 
set_hashdist(char * str)8620 static int __init set_hashdist(char *str)
8621 {
8622 	if (!str)
8623 		return 0;
8624 	hashdist = simple_strtoul(str, &str, 0);
8625 	return 1;
8626 }
8627 __setup("hashdist=", set_hashdist);
8628 #endif
8629 
page_alloc_init(void)8630 void __init page_alloc_init(void)
8631 {
8632 	int ret;
8633 
8634 #ifdef CONFIG_NUMA
8635 	if (num_node_state(N_MEMORY) == 1)
8636 		hashdist = 0;
8637 #endif
8638 
8639 	ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8640 					"mm/page_alloc:pcp",
8641 					page_alloc_cpu_online,
8642 					page_alloc_cpu_dead);
8643 	WARN_ON(ret < 0);
8644 }
8645 
8646 /*
8647  * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8648  *	or min_free_kbytes changes.
8649  */
calculate_totalreserve_pages(void)8650 static void calculate_totalreserve_pages(void)
8651 {
8652 	struct pglist_data *pgdat;
8653 	unsigned long reserve_pages = 0;
8654 	enum zone_type i, j;
8655 
8656 	for_each_online_pgdat(pgdat) {
8657 
8658 		pgdat->totalreserve_pages = 0;
8659 
8660 		for (i = 0; i < MAX_NR_ZONES; i++) {
8661 			struct zone *zone = pgdat->node_zones + i;
8662 			long max = 0;
8663 			unsigned long managed_pages = zone_managed_pages(zone);
8664 
8665 			/* Find valid and maximum lowmem_reserve in the zone */
8666 			for (j = i; j < MAX_NR_ZONES; j++) {
8667 				if (zone->lowmem_reserve[j] > max)
8668 					max = zone->lowmem_reserve[j];
8669 			}
8670 
8671 			/* we treat the high watermark as reserved pages. */
8672 			max += high_wmark_pages(zone);
8673 
8674 			if (max > managed_pages)
8675 				max = managed_pages;
8676 
8677 			pgdat->totalreserve_pages += max;
8678 
8679 			reserve_pages += max;
8680 		}
8681 	}
8682 	totalreserve_pages = reserve_pages;
8683 }
8684 
8685 /*
8686  * setup_per_zone_lowmem_reserve - called whenever
8687  *	sysctl_lowmem_reserve_ratio changes.  Ensures that each zone
8688  *	has a correct pages reserved value, so an adequate number of
8689  *	pages are left in the zone after a successful __alloc_pages().
8690  */
setup_per_zone_lowmem_reserve(void)8691 static void setup_per_zone_lowmem_reserve(void)
8692 {
8693 	struct pglist_data *pgdat;
8694 	enum zone_type i, j;
8695 
8696 	for_each_online_pgdat(pgdat) {
8697 		for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8698 			struct zone *zone = &pgdat->node_zones[i];
8699 			int ratio = sysctl_lowmem_reserve_ratio[i];
8700 			bool clear = !ratio || !zone_managed_pages(zone);
8701 			unsigned long managed_pages = 0;
8702 
8703 			for (j = i + 1; j < MAX_NR_ZONES; j++) {
8704 				struct zone *upper_zone = &pgdat->node_zones[j];
8705 
8706 				managed_pages += zone_managed_pages(upper_zone);
8707 
8708 				if (clear)
8709 					zone->lowmem_reserve[j] = 0;
8710 				else
8711 					zone->lowmem_reserve[j] = managed_pages / ratio;
8712 			}
8713 		}
8714 	}
8715 
8716 	/* update totalreserve_pages */
8717 	calculate_totalreserve_pages();
8718 }
8719 
__setup_per_zone_wmarks(void)8720 static void __setup_per_zone_wmarks(void)
8721 {
8722 	unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8723 	unsigned long lowmem_pages = 0;
8724 	struct zone *zone;
8725 	unsigned long flags;
8726 
8727 	/* Calculate total number of !ZONE_HIGHMEM pages */
8728 	for_each_zone(zone) {
8729 		if (!is_highmem(zone))
8730 			lowmem_pages += zone_managed_pages(zone);
8731 	}
8732 
8733 	for_each_zone(zone) {
8734 		u64 tmp;
8735 
8736 		spin_lock_irqsave(&zone->lock, flags);
8737 		tmp = (u64)pages_min * zone_managed_pages(zone);
8738 		do_div(tmp, lowmem_pages);
8739 		if (is_highmem(zone)) {
8740 			/*
8741 			 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8742 			 * need highmem pages, so cap pages_min to a small
8743 			 * value here.
8744 			 *
8745 			 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8746 			 * deltas control async page reclaim, and so should
8747 			 * not be capped for highmem.
8748 			 */
8749 			unsigned long min_pages;
8750 
8751 			min_pages = zone_managed_pages(zone) / 1024;
8752 			min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8753 			zone->_watermark[WMARK_MIN] = min_pages;
8754 		} else {
8755 			/*
8756 			 * If it's a lowmem zone, reserve a number of pages
8757 			 * proportionate to the zone's size.
8758 			 */
8759 			zone->_watermark[WMARK_MIN] = tmp;
8760 		}
8761 
8762 		/*
8763 		 * Set the kswapd watermarks distance according to the
8764 		 * scale factor in proportion to available memory, but
8765 		 * ensure a minimum size on small systems.
8766 		 */
8767 		tmp = max_t(u64, tmp >> 2,
8768 			    mult_frac(zone_managed_pages(zone),
8769 				      watermark_scale_factor, 10000));
8770 
8771 		zone->watermark_boost = 0;
8772 		zone->_watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
8773 		zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8774 		zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8775 
8776 		spin_unlock_irqrestore(&zone->lock, flags);
8777 	}
8778 
8779 	/* update totalreserve_pages */
8780 	calculate_totalreserve_pages();
8781 }
8782 
8783 /**
8784  * setup_per_zone_wmarks - called when min_free_kbytes changes
8785  * or when memory is hot-{added|removed}
8786  *
8787  * Ensures that the watermark[min,low,high] values for each zone are set
8788  * correctly with respect to min_free_kbytes.
8789  */
setup_per_zone_wmarks(void)8790 void setup_per_zone_wmarks(void)
8791 {
8792 	struct zone *zone;
8793 	static DEFINE_SPINLOCK(lock);
8794 
8795 	spin_lock(&lock);
8796 	__setup_per_zone_wmarks();
8797 	spin_unlock(&lock);
8798 
8799 	/*
8800 	 * The watermark size have changed so update the pcpu batch
8801 	 * and high limits or the limits may be inappropriate.
8802 	 */
8803 	for_each_zone(zone)
8804 		zone_pcp_update(zone, 0);
8805 }
8806 
8807 /*
8808  * Initialise min_free_kbytes.
8809  *
8810  * For small machines we want it small (128k min).  For large machines
8811  * we want it large (256MB max).  But it is not linear, because network
8812  * bandwidth does not increase linearly with machine size.  We use
8813  *
8814  *	min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8815  *	min_free_kbytes = sqrt(lowmem_kbytes * 16)
8816  *
8817  * which yields
8818  *
8819  * 16MB:	512k
8820  * 32MB:	724k
8821  * 64MB:	1024k
8822  * 128MB:	1448k
8823  * 256MB:	2048k
8824  * 512MB:	2896k
8825  * 1024MB:	4096k
8826  * 2048MB:	5792k
8827  * 4096MB:	8192k
8828  * 8192MB:	11584k
8829  * 16384MB:	16384k
8830  */
calculate_min_free_kbytes(void)8831 void calculate_min_free_kbytes(void)
8832 {
8833 	unsigned long lowmem_kbytes;
8834 	int new_min_free_kbytes;
8835 
8836 	lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8837 	new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8838 
8839 	if (new_min_free_kbytes > user_min_free_kbytes)
8840 		min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8841 	else
8842 		pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8843 				new_min_free_kbytes, user_min_free_kbytes);
8844 
8845 }
8846 
init_per_zone_wmark_min(void)8847 int __meminit init_per_zone_wmark_min(void)
8848 {
8849 	calculate_min_free_kbytes();
8850 	setup_per_zone_wmarks();
8851 	refresh_zone_stat_thresholds();
8852 	setup_per_zone_lowmem_reserve();
8853 
8854 #ifdef CONFIG_NUMA
8855 	setup_min_unmapped_ratio();
8856 	setup_min_slab_ratio();
8857 #endif
8858 
8859 	khugepaged_min_free_kbytes_update();
8860 
8861 	return 0;
8862 }
postcore_initcall(init_per_zone_wmark_min)8863 postcore_initcall(init_per_zone_wmark_min)
8864 
8865 /*
8866  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8867  *	that we can call two helper functions whenever min_free_kbytes
8868  *	changes.
8869  */
8870 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8871 		void *buffer, size_t *length, loff_t *ppos)
8872 {
8873 	int rc;
8874 
8875 	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8876 	if (rc)
8877 		return rc;
8878 
8879 	if (write) {
8880 		user_min_free_kbytes = min_free_kbytes;
8881 		setup_per_zone_wmarks();
8882 	}
8883 	return 0;
8884 }
8885 
watermark_scale_factor_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)8886 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8887 		void *buffer, size_t *length, loff_t *ppos)
8888 {
8889 	int rc;
8890 
8891 	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8892 	if (rc)
8893 		return rc;
8894 
8895 	if (write)
8896 		setup_per_zone_wmarks();
8897 
8898 	return 0;
8899 }
8900 
8901 #ifdef CONFIG_NUMA
setup_min_unmapped_ratio(void)8902 static void setup_min_unmapped_ratio(void)
8903 {
8904 	pg_data_t *pgdat;
8905 	struct zone *zone;
8906 
8907 	for_each_online_pgdat(pgdat)
8908 		pgdat->min_unmapped_pages = 0;
8909 
8910 	for_each_zone(zone)
8911 		zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8912 						         sysctl_min_unmapped_ratio) / 100;
8913 }
8914 
8915 
sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)8916 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8917 		void *buffer, size_t *length, loff_t *ppos)
8918 {
8919 	int rc;
8920 
8921 	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8922 	if (rc)
8923 		return rc;
8924 
8925 	setup_min_unmapped_ratio();
8926 
8927 	return 0;
8928 }
8929 
setup_min_slab_ratio(void)8930 static void setup_min_slab_ratio(void)
8931 {
8932 	pg_data_t *pgdat;
8933 	struct zone *zone;
8934 
8935 	for_each_online_pgdat(pgdat)
8936 		pgdat->min_slab_pages = 0;
8937 
8938 	for_each_zone(zone)
8939 		zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8940 						     sysctl_min_slab_ratio) / 100;
8941 }
8942 
sysctl_min_slab_ratio_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)8943 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8944 		void *buffer, size_t *length, loff_t *ppos)
8945 {
8946 	int rc;
8947 
8948 	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8949 	if (rc)
8950 		return rc;
8951 
8952 	setup_min_slab_ratio();
8953 
8954 	return 0;
8955 }
8956 #endif
8957 
8958 /*
8959  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8960  *	proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8961  *	whenever sysctl_lowmem_reserve_ratio changes.
8962  *
8963  * The reserve ratio obviously has absolutely no relation with the
8964  * minimum watermarks. The lowmem reserve ratio can only make sense
8965  * if in function of the boot time zone sizes.
8966  */
lowmem_reserve_ratio_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)8967 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8968 		void *buffer, size_t *length, loff_t *ppos)
8969 {
8970 	int i;
8971 
8972 	proc_dointvec_minmax(table, write, buffer, length, ppos);
8973 
8974 	for (i = 0; i < MAX_NR_ZONES; i++) {
8975 		if (sysctl_lowmem_reserve_ratio[i] < 1)
8976 			sysctl_lowmem_reserve_ratio[i] = 0;
8977 	}
8978 
8979 	setup_per_zone_lowmem_reserve();
8980 	return 0;
8981 }
8982 
8983 /*
8984  * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8985  * cpu. It is the fraction of total pages in each zone that a hot per cpu
8986  * pagelist can have before it gets flushed back to buddy allocator.
8987  */
percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)8988 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8989 		int write, void *buffer, size_t *length, loff_t *ppos)
8990 {
8991 	struct zone *zone;
8992 	int old_percpu_pagelist_high_fraction;
8993 	int ret;
8994 
8995 	mutex_lock(&pcp_batch_high_lock);
8996 	old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8997 
8998 	ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8999 	if (!write || ret < 0)
9000 		goto out;
9001 
9002 	/* Sanity checking to avoid pcp imbalance */
9003 	if (percpu_pagelist_high_fraction &&
9004 	    percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
9005 		percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
9006 		ret = -EINVAL;
9007 		goto out;
9008 	}
9009 
9010 	/* No change? */
9011 	if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
9012 		goto out;
9013 
9014 	for_each_populated_zone(zone)
9015 		zone_set_pageset_high_and_batch(zone, 0);
9016 out:
9017 	mutex_unlock(&pcp_batch_high_lock);
9018 	return ret;
9019 }
9020 
9021 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
9022 /*
9023  * Returns the number of pages that arch has reserved but
9024  * is not known to alloc_large_system_hash().
9025  */
arch_reserved_kernel_pages(void)9026 static unsigned long __init arch_reserved_kernel_pages(void)
9027 {
9028 	return 0;
9029 }
9030 #endif
9031 
9032 /*
9033  * Adaptive scale is meant to reduce sizes of hash tables on large memory
9034  * machines. As memory size is increased the scale is also increased but at
9035  * slower pace.  Starting from ADAPT_SCALE_BASE (64G), every time memory
9036  * quadruples the scale is increased by one, which means the size of hash table
9037  * only doubles, instead of quadrupling as well.
9038  * Because 32-bit systems cannot have large physical memory, where this scaling
9039  * makes sense, it is disabled on such platforms.
9040  */
9041 #if __BITS_PER_LONG > 32
9042 #define ADAPT_SCALE_BASE	(64ul << 30)
9043 #define ADAPT_SCALE_SHIFT	2
9044 #define ADAPT_SCALE_NPAGES	(ADAPT_SCALE_BASE >> PAGE_SHIFT)
9045 #endif
9046 
9047 /*
9048  * allocate a large system hash table from bootmem
9049  * - it is assumed that the hash table must contain an exact power-of-2
9050  *   quantity of entries
9051  * - limit is the number of hash buckets, not the total allocation size
9052  */
alloc_large_system_hash(const char * tablename,unsigned long bucketsize,unsigned long numentries,int scale,int flags,unsigned int * _hash_shift,unsigned int * _hash_mask,unsigned long low_limit,unsigned long high_limit)9053 void *__init alloc_large_system_hash(const char *tablename,
9054 				     unsigned long bucketsize,
9055 				     unsigned long numentries,
9056 				     int scale,
9057 				     int flags,
9058 				     unsigned int *_hash_shift,
9059 				     unsigned int *_hash_mask,
9060 				     unsigned long low_limit,
9061 				     unsigned long high_limit)
9062 {
9063 	unsigned long long max = high_limit;
9064 	unsigned long log2qty, size;
9065 	void *table;
9066 	gfp_t gfp_flags;
9067 	bool virt;
9068 	bool huge;
9069 
9070 	/* allow the kernel cmdline to have a say */
9071 	if (!numentries) {
9072 		/* round applicable memory size up to nearest megabyte */
9073 		numentries = nr_kernel_pages;
9074 		numentries -= arch_reserved_kernel_pages();
9075 
9076 		/* It isn't necessary when PAGE_SIZE >= 1MB */
9077 		if (PAGE_SIZE < SZ_1M)
9078 			numentries = round_up(numentries, SZ_1M / PAGE_SIZE);
9079 
9080 #if __BITS_PER_LONG > 32
9081 		if (!high_limit) {
9082 			unsigned long adapt;
9083 
9084 			for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
9085 			     adapt <<= ADAPT_SCALE_SHIFT)
9086 				scale++;
9087 		}
9088 #endif
9089 
9090 		/* limit to 1 bucket per 2^scale bytes of low memory */
9091 		if (scale > PAGE_SHIFT)
9092 			numentries >>= (scale - PAGE_SHIFT);
9093 		else
9094 			numentries <<= (PAGE_SHIFT - scale);
9095 
9096 		/* Make sure we've got at least a 0-order allocation.. */
9097 		if (unlikely(flags & HASH_SMALL)) {
9098 			/* Makes no sense without HASH_EARLY */
9099 			WARN_ON(!(flags & HASH_EARLY));
9100 			if (!(numentries >> *_hash_shift)) {
9101 				numentries = 1UL << *_hash_shift;
9102 				BUG_ON(!numentries);
9103 			}
9104 		} else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9105 			numentries = PAGE_SIZE / bucketsize;
9106 	}
9107 	numentries = roundup_pow_of_two(numentries);
9108 
9109 	/* limit allocation size to 1/16 total memory by default */
9110 	if (max == 0) {
9111 		max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9112 		do_div(max, bucketsize);
9113 	}
9114 	max = min(max, 0x80000000ULL);
9115 
9116 	if (numentries < low_limit)
9117 		numentries = low_limit;
9118 	if (numentries > max)
9119 		numentries = max;
9120 
9121 	log2qty = ilog2(numentries);
9122 
9123 	gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9124 	do {
9125 		virt = false;
9126 		size = bucketsize << log2qty;
9127 		if (flags & HASH_EARLY) {
9128 			if (flags & HASH_ZERO)
9129 				table = memblock_alloc(size, SMP_CACHE_BYTES);
9130 			else
9131 				table = memblock_alloc_raw(size,
9132 							   SMP_CACHE_BYTES);
9133 		} else if (get_order(size) >= MAX_ORDER || hashdist) {
9134 			table = vmalloc_huge(size, gfp_flags);
9135 			virt = true;
9136 			if (table)
9137 				huge = is_vm_area_hugepages(table);
9138 		} else {
9139 			/*
9140 			 * If bucketsize is not a power-of-two, we may free
9141 			 * some pages at the end of hash table which
9142 			 * alloc_pages_exact() automatically does
9143 			 */
9144 			table = alloc_pages_exact(size, gfp_flags);
9145 			kmemleak_alloc(table, size, 1, gfp_flags);
9146 		}
9147 	} while (!table && size > PAGE_SIZE && --log2qty);
9148 
9149 	if (!table)
9150 		panic("Failed to allocate %s hash table\n", tablename);
9151 
9152 	pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9153 		tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9154 		virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9155 
9156 	if (_hash_shift)
9157 		*_hash_shift = log2qty;
9158 	if (_hash_mask)
9159 		*_hash_mask = (1 << log2qty) - 1;
9160 
9161 	return table;
9162 }
9163 
9164 #ifdef CONFIG_CONTIG_ALLOC
9165 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9166 	(defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9167 /* Usage: See admin-guide/dynamic-debug-howto.rst */
alloc_contig_dump_pages(struct list_head * page_list)9168 static void alloc_contig_dump_pages(struct list_head *page_list)
9169 {
9170 	DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9171 
9172 	if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9173 		struct page *page;
9174 
9175 		dump_stack();
9176 		list_for_each_entry(page, page_list, lru)
9177 			dump_page(page, "migration failure");
9178 	}
9179 }
9180 #else
alloc_contig_dump_pages(struct list_head * page_list)9181 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9182 {
9183 }
9184 #endif
9185 
9186 /* [start, end) must belong to a single zone. */
__alloc_contig_migrate_range(struct compact_control * cc,unsigned long start,unsigned long end)9187 int __alloc_contig_migrate_range(struct compact_control *cc,
9188 					unsigned long start, unsigned long end)
9189 {
9190 	/* This function is based on compact_zone() from compaction.c. */
9191 	unsigned int nr_reclaimed;
9192 	unsigned long pfn = start;
9193 	unsigned int tries = 0;
9194 	int ret = 0;
9195 	struct migration_target_control mtc = {
9196 		.nid = zone_to_nid(cc->zone),
9197 		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9198 	};
9199 
9200 	lru_cache_disable();
9201 
9202 	while (pfn < end || !list_empty(&cc->migratepages)) {
9203 		if (fatal_signal_pending(current)) {
9204 			ret = -EINTR;
9205 			break;
9206 		}
9207 
9208 		if (list_empty(&cc->migratepages)) {
9209 			cc->nr_migratepages = 0;
9210 			ret = isolate_migratepages_range(cc, pfn, end);
9211 			if (ret && ret != -EAGAIN)
9212 				break;
9213 			pfn = cc->migrate_pfn;
9214 			tries = 0;
9215 		} else if (++tries == 5) {
9216 			ret = -EBUSY;
9217 			break;
9218 		}
9219 
9220 		nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9221 							&cc->migratepages);
9222 		cc->nr_migratepages -= nr_reclaimed;
9223 
9224 		ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9225 			NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9226 
9227 		/*
9228 		 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9229 		 * to retry again over this error, so do the same here.
9230 		 */
9231 		if (ret == -ENOMEM)
9232 			break;
9233 	}
9234 
9235 	lru_cache_enable();
9236 	if (ret < 0) {
9237 		if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9238 			alloc_contig_dump_pages(&cc->migratepages);
9239 		putback_movable_pages(&cc->migratepages);
9240 		return ret;
9241 	}
9242 	return 0;
9243 }
9244 
9245 /**
9246  * alloc_contig_range() -- tries to allocate given range of pages
9247  * @start:	start PFN to allocate
9248  * @end:	one-past-the-last PFN to allocate
9249  * @migratetype:	migratetype of the underlying pageblocks (either
9250  *			#MIGRATE_MOVABLE or #MIGRATE_CMA).  All pageblocks
9251  *			in range must have the same migratetype and it must
9252  *			be either of the two.
9253  * @gfp_mask:	GFP mask to use during compaction
9254  *
9255  * The PFN range does not have to be pageblock aligned. The PFN range must
9256  * belong to a single zone.
9257  *
9258  * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9259  * pageblocks in the range.  Once isolated, the pageblocks should not
9260  * be modified by others.
9261  *
9262  * Return: zero on success or negative error code.  On success all
9263  * pages which PFN is in [start, end) are allocated for the caller and
9264  * need to be freed with free_contig_range().
9265  */
alloc_contig_range(unsigned long start,unsigned long end,unsigned migratetype,gfp_t gfp_mask)9266 int alloc_contig_range(unsigned long start, unsigned long end,
9267 		       unsigned migratetype, gfp_t gfp_mask)
9268 {
9269 	unsigned long outer_start, outer_end;
9270 	int order;
9271 	int ret = 0;
9272 
9273 	struct compact_control cc = {
9274 		.nr_migratepages = 0,
9275 		.order = -1,
9276 		.zone = page_zone(pfn_to_page(start)),
9277 		.mode = MIGRATE_SYNC,
9278 		.ignore_skip_hint = true,
9279 		.no_set_skip_hint = true,
9280 		.gfp_mask = current_gfp_context(gfp_mask),
9281 		.alloc_contig = true,
9282 	};
9283 	INIT_LIST_HEAD(&cc.migratepages);
9284 
9285 	/*
9286 	 * What we do here is we mark all pageblocks in range as
9287 	 * MIGRATE_ISOLATE.  Because pageblock and max order pages may
9288 	 * have different sizes, and due to the way page allocator
9289 	 * work, start_isolate_page_range() has special handlings for this.
9290 	 *
9291 	 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9292 	 * migrate the pages from an unaligned range (ie. pages that
9293 	 * we are interested in). This will put all the pages in
9294 	 * range back to page allocator as MIGRATE_ISOLATE.
9295 	 *
9296 	 * When this is done, we take the pages in range from page
9297 	 * allocator removing them from the buddy system.  This way
9298 	 * page allocator will never consider using them.
9299 	 *
9300 	 * This lets us mark the pageblocks back as
9301 	 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9302 	 * aligned range but not in the unaligned, original range are
9303 	 * put back to page allocator so that buddy can use them.
9304 	 */
9305 
9306 	ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9307 	if (ret)
9308 		goto done;
9309 
9310 	drain_all_pages(cc.zone);
9311 
9312 	/*
9313 	 * In case of -EBUSY, we'd like to know which page causes problem.
9314 	 * So, just fall through. test_pages_isolated() has a tracepoint
9315 	 * which will report the busy page.
9316 	 *
9317 	 * It is possible that busy pages could become available before
9318 	 * the call to test_pages_isolated, and the range will actually be
9319 	 * allocated.  So, if we fall through be sure to clear ret so that
9320 	 * -EBUSY is not accidentally used or returned to caller.
9321 	 */
9322 	ret = __alloc_contig_migrate_range(&cc, start, end);
9323 	if (ret && ret != -EBUSY)
9324 		goto done;
9325 	ret = 0;
9326 
9327 	/*
9328 	 * Pages from [start, end) are within a pageblock_nr_pages
9329 	 * aligned blocks that are marked as MIGRATE_ISOLATE.  What's
9330 	 * more, all pages in [start, end) are free in page allocator.
9331 	 * What we are going to do is to allocate all pages from
9332 	 * [start, end) (that is remove them from page allocator).
9333 	 *
9334 	 * The only problem is that pages at the beginning and at the
9335 	 * end of interesting range may be not aligned with pages that
9336 	 * page allocator holds, ie. they can be part of higher order
9337 	 * pages.  Because of this, we reserve the bigger range and
9338 	 * once this is done free the pages we are not interested in.
9339 	 *
9340 	 * We don't have to hold zone->lock here because the pages are
9341 	 * isolated thus they won't get removed from buddy.
9342 	 */
9343 
9344 	order = 0;
9345 	outer_start = start;
9346 	while (!PageBuddy(pfn_to_page(outer_start))) {
9347 		if (++order >= MAX_ORDER) {
9348 			outer_start = start;
9349 			break;
9350 		}
9351 		outer_start &= ~0UL << order;
9352 	}
9353 
9354 	if (outer_start != start) {
9355 		order = buddy_order(pfn_to_page(outer_start));
9356 
9357 		/*
9358 		 * outer_start page could be small order buddy page and
9359 		 * it doesn't include start page. Adjust outer_start
9360 		 * in this case to report failed page properly
9361 		 * on tracepoint in test_pages_isolated()
9362 		 */
9363 		if (outer_start + (1UL << order) <= start)
9364 			outer_start = start;
9365 	}
9366 
9367 	/* Make sure the range is really isolated. */
9368 	if (test_pages_isolated(outer_start, end, 0)) {
9369 		ret = -EBUSY;
9370 		goto done;
9371 	}
9372 
9373 	/* Grab isolated pages from freelists. */
9374 	outer_end = isolate_freepages_range(&cc, outer_start, end);
9375 	if (!outer_end) {
9376 		ret = -EBUSY;
9377 		goto done;
9378 	}
9379 
9380 	/* Free head and tail (if any) */
9381 	if (start != outer_start)
9382 		free_contig_range(outer_start, start - outer_start);
9383 	if (end != outer_end)
9384 		free_contig_range(end, outer_end - end);
9385 
9386 done:
9387 	undo_isolate_page_range(start, end, migratetype);
9388 	return ret;
9389 }
9390 EXPORT_SYMBOL(alloc_contig_range);
9391 
__alloc_contig_pages(unsigned long start_pfn,unsigned long nr_pages,gfp_t gfp_mask)9392 static int __alloc_contig_pages(unsigned long start_pfn,
9393 				unsigned long nr_pages, gfp_t gfp_mask)
9394 {
9395 	unsigned long end_pfn = start_pfn + nr_pages;
9396 
9397 	return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9398 				  gfp_mask);
9399 }
9400 
pfn_range_valid_contig(struct zone * z,unsigned long start_pfn,unsigned long nr_pages)9401 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9402 				   unsigned long nr_pages)
9403 {
9404 	unsigned long i, end_pfn = start_pfn + nr_pages;
9405 	struct page *page;
9406 
9407 	for (i = start_pfn; i < end_pfn; i++) {
9408 		page = pfn_to_online_page(i);
9409 		if (!page)
9410 			return false;
9411 
9412 		if (page_zone(page) != z)
9413 			return false;
9414 
9415 		if (PageReserved(page))
9416 			return false;
9417 	}
9418 	return true;
9419 }
9420 
zone_spans_last_pfn(const struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)9421 static bool zone_spans_last_pfn(const struct zone *zone,
9422 				unsigned long start_pfn, unsigned long nr_pages)
9423 {
9424 	unsigned long last_pfn = start_pfn + nr_pages - 1;
9425 
9426 	return zone_spans_pfn(zone, last_pfn);
9427 }
9428 
9429 /**
9430  * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9431  * @nr_pages:	Number of contiguous pages to allocate
9432  * @gfp_mask:	GFP mask to limit search and used during compaction
9433  * @nid:	Target node
9434  * @nodemask:	Mask for other possible nodes
9435  *
9436  * This routine is a wrapper around alloc_contig_range(). It scans over zones
9437  * on an applicable zonelist to find a contiguous pfn range which can then be
9438  * tried for allocation with alloc_contig_range(). This routine is intended
9439  * for allocation requests which can not be fulfilled with the buddy allocator.
9440  *
9441  * The allocated memory is always aligned to a page boundary. If nr_pages is a
9442  * power of two, then allocated range is also guaranteed to be aligned to same
9443  * nr_pages (e.g. 1GB request would be aligned to 1GB).
9444  *
9445  * Allocated pages can be freed with free_contig_range() or by manually calling
9446  * __free_page() on each allocated page.
9447  *
9448  * Return: pointer to contiguous pages on success, or NULL if not successful.
9449  */
alloc_contig_pages(unsigned long nr_pages,gfp_t gfp_mask,int nid,nodemask_t * nodemask)9450 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9451 				int nid, nodemask_t *nodemask)
9452 {
9453 	unsigned long ret, pfn, flags;
9454 	struct zonelist *zonelist;
9455 	struct zone *zone;
9456 	struct zoneref *z;
9457 
9458 	zonelist = node_zonelist(nid, gfp_mask);
9459 	for_each_zone_zonelist_nodemask(zone, z, zonelist,
9460 					gfp_zone(gfp_mask), nodemask) {
9461 		spin_lock_irqsave(&zone->lock, flags);
9462 
9463 		pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9464 		while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9465 			if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9466 				/*
9467 				 * We release the zone lock here because
9468 				 * alloc_contig_range() will also lock the zone
9469 				 * at some point. If there's an allocation
9470 				 * spinning on this lock, it may win the race
9471 				 * and cause alloc_contig_range() to fail...
9472 				 */
9473 				spin_unlock_irqrestore(&zone->lock, flags);
9474 				ret = __alloc_contig_pages(pfn, nr_pages,
9475 							gfp_mask);
9476 				if (!ret)
9477 					return pfn_to_page(pfn);
9478 				spin_lock_irqsave(&zone->lock, flags);
9479 			}
9480 			pfn += nr_pages;
9481 		}
9482 		spin_unlock_irqrestore(&zone->lock, flags);
9483 	}
9484 	return NULL;
9485 }
9486 #endif /* CONFIG_CONTIG_ALLOC */
9487 
free_contig_range(unsigned long pfn,unsigned long nr_pages)9488 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9489 {
9490 	unsigned long count = 0;
9491 
9492 	for (; nr_pages--; pfn++) {
9493 		struct page *page = pfn_to_page(pfn);
9494 
9495 		count += page_count(page) != 1;
9496 		__free_page(page);
9497 	}
9498 	WARN(count != 0, "%lu pages are still in use!\n", count);
9499 }
9500 EXPORT_SYMBOL(free_contig_range);
9501 
9502 /*
9503  * Effectively disable pcplists for the zone by setting the high limit to 0
9504  * and draining all cpus. A concurrent page freeing on another CPU that's about
9505  * to put the page on pcplist will either finish before the drain and the page
9506  * will be drained, or observe the new high limit and skip the pcplist.
9507  *
9508  * Must be paired with a call to zone_pcp_enable().
9509  */
zone_pcp_disable(struct zone * zone)9510 void zone_pcp_disable(struct zone *zone)
9511 {
9512 	mutex_lock(&pcp_batch_high_lock);
9513 	__zone_set_pageset_high_and_batch(zone, 0, 1);
9514 	__drain_all_pages(zone, true);
9515 }
9516 
zone_pcp_enable(struct zone * zone)9517 void zone_pcp_enable(struct zone *zone)
9518 {
9519 	__zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9520 	mutex_unlock(&pcp_batch_high_lock);
9521 }
9522 
zone_pcp_reset(struct zone * zone)9523 void zone_pcp_reset(struct zone *zone)
9524 {
9525 	int cpu;
9526 	struct per_cpu_zonestat *pzstats;
9527 
9528 	if (zone->per_cpu_pageset != &boot_pageset) {
9529 		for_each_online_cpu(cpu) {
9530 			pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9531 			drain_zonestat(zone, pzstats);
9532 		}
9533 		free_percpu(zone->per_cpu_pageset);
9534 		zone->per_cpu_pageset = &boot_pageset;
9535 		if (zone->per_cpu_zonestats != &boot_zonestats) {
9536 			free_percpu(zone->per_cpu_zonestats);
9537 			zone->per_cpu_zonestats = &boot_zonestats;
9538 		}
9539 	}
9540 }
9541 
9542 #ifdef CONFIG_MEMORY_HOTREMOVE
9543 /*
9544  * All pages in the range must be in a single zone, must not contain holes,
9545  * must span full sections, and must be isolated before calling this function.
9546  */
__offline_isolated_pages(unsigned long start_pfn,unsigned long end_pfn)9547 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9548 {
9549 	unsigned long pfn = start_pfn;
9550 	struct page *page;
9551 	struct zone *zone;
9552 	unsigned int order;
9553 	unsigned long flags;
9554 
9555 	offline_mem_sections(pfn, end_pfn);
9556 	zone = page_zone(pfn_to_page(pfn));
9557 	spin_lock_irqsave(&zone->lock, flags);
9558 	while (pfn < end_pfn) {
9559 		page = pfn_to_page(pfn);
9560 		/*
9561 		 * The HWPoisoned page may be not in buddy system, and
9562 		 * page_count() is not 0.
9563 		 */
9564 		if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9565 			pfn++;
9566 			continue;
9567 		}
9568 		/*
9569 		 * At this point all remaining PageOffline() pages have a
9570 		 * reference count of 0 and can simply be skipped.
9571 		 */
9572 		if (PageOffline(page)) {
9573 			BUG_ON(page_count(page));
9574 			BUG_ON(PageBuddy(page));
9575 			pfn++;
9576 			continue;
9577 		}
9578 
9579 		BUG_ON(page_count(page));
9580 		BUG_ON(!PageBuddy(page));
9581 		order = buddy_order(page);
9582 		del_page_from_free_list(page, zone, order);
9583 		pfn += (1 << order);
9584 	}
9585 	spin_unlock_irqrestore(&zone->lock, flags);
9586 }
9587 #endif
9588 
9589 /*
9590  * This function returns a stable result only if called under zone lock.
9591  */
is_free_buddy_page(struct page * page)9592 bool is_free_buddy_page(struct page *page)
9593 {
9594 	unsigned long pfn = page_to_pfn(page);
9595 	unsigned int order;
9596 
9597 	for (order = 0; order < MAX_ORDER; order++) {
9598 		struct page *page_head = page - (pfn & ((1 << order) - 1));
9599 
9600 		if (PageBuddy(page_head) &&
9601 		    buddy_order_unsafe(page_head) >= order)
9602 			break;
9603 	}
9604 
9605 	return order < MAX_ORDER;
9606 }
9607 EXPORT_SYMBOL(is_free_buddy_page);
9608 
9609 #ifdef CONFIG_MEMORY_FAILURE
9610 /*
9611  * Break down a higher-order page in sub-pages, and keep our target out of
9612  * buddy allocator.
9613  */
break_down_buddy_pages(struct zone * zone,struct page * page,struct page * target,int low,int high,int migratetype)9614 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9615 				   struct page *target, int low, int high,
9616 				   int migratetype)
9617 {
9618 	unsigned long size = 1 << high;
9619 	struct page *current_buddy, *next_page;
9620 
9621 	while (high > low) {
9622 		high--;
9623 		size >>= 1;
9624 
9625 		if (target >= &page[size]) {
9626 			next_page = page + size;
9627 			current_buddy = page;
9628 		} else {
9629 			next_page = page;
9630 			current_buddy = page + size;
9631 		}
9632 
9633 		if (set_page_guard(zone, current_buddy, high, migratetype))
9634 			continue;
9635 
9636 		if (current_buddy != target) {
9637 			add_to_free_list(current_buddy, zone, high, migratetype);
9638 			set_buddy_order(current_buddy, high);
9639 			page = next_page;
9640 		}
9641 	}
9642 }
9643 
9644 /*
9645  * Take a page that will be marked as poisoned off the buddy allocator.
9646  */
take_page_off_buddy(struct page * page)9647 bool take_page_off_buddy(struct page *page)
9648 {
9649 	struct zone *zone = page_zone(page);
9650 	unsigned long pfn = page_to_pfn(page);
9651 	unsigned long flags;
9652 	unsigned int order;
9653 	bool ret = false;
9654 
9655 	spin_lock_irqsave(&zone->lock, flags);
9656 	for (order = 0; order < MAX_ORDER; order++) {
9657 		struct page *page_head = page - (pfn & ((1 << order) - 1));
9658 		int page_order = buddy_order(page_head);
9659 
9660 		if (PageBuddy(page_head) && page_order >= order) {
9661 			unsigned long pfn_head = page_to_pfn(page_head);
9662 			int migratetype = get_pfnblock_migratetype(page_head,
9663 								   pfn_head);
9664 
9665 			del_page_from_free_list(page_head, zone, page_order);
9666 			break_down_buddy_pages(zone, page_head, page, 0,
9667 						page_order, migratetype);
9668 			SetPageHWPoisonTakenOff(page);
9669 			if (!is_migrate_isolate(migratetype))
9670 				__mod_zone_freepage_state(zone, -1, migratetype);
9671 			ret = true;
9672 			break;
9673 		}
9674 		if (page_count(page_head) > 0)
9675 			break;
9676 	}
9677 	spin_unlock_irqrestore(&zone->lock, flags);
9678 	return ret;
9679 }
9680 
9681 /*
9682  * Cancel takeoff done by take_page_off_buddy().
9683  */
put_page_back_buddy(struct page * page)9684 bool put_page_back_buddy(struct page *page)
9685 {
9686 	struct zone *zone = page_zone(page);
9687 	unsigned long pfn = page_to_pfn(page);
9688 	unsigned long flags;
9689 	int migratetype = get_pfnblock_migratetype(page, pfn);
9690 	bool ret = false;
9691 
9692 	spin_lock_irqsave(&zone->lock, flags);
9693 	if (put_page_testzero(page)) {
9694 		ClearPageHWPoisonTakenOff(page);
9695 		__free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9696 		if (TestClearPageHWPoison(page)) {
9697 			ret = true;
9698 		}
9699 	}
9700 	spin_unlock_irqrestore(&zone->lock, flags);
9701 
9702 	return ret;
9703 }
9704 #endif
9705 
9706 #ifdef CONFIG_ZONE_DMA
has_managed_dma(void)9707 bool has_managed_dma(void)
9708 {
9709 	struct pglist_data *pgdat;
9710 
9711 	for_each_online_pgdat(pgdat) {
9712 		struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9713 
9714 		if (managed_zone(zone))
9715 			return true;
9716 	}
9717 	return false;
9718 }
9719 #endif /* CONFIG_ZONE_DMA */
9720