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