1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/memory_hotplug.c
4  *
5  *  Copyright (C)
6  */
7 
8 #include <linux/stddef.h>
9 #include <linux/mm.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/pagevec.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
20 #include <linux/cpu.h>
21 #include <linux/memory.h>
22 #include <linux/memremap.h>
23 #include <linux/memory_hotplug.h>
24 #include <linux/vmalloc.h>
25 #include <linux/ioport.h>
26 #include <linux/delay.h>
27 #include <linux/migrate.h>
28 #include <linux/page-isolation.h>
29 #include <linux/pfn.h>
30 #include <linux/suspend.h>
31 #include <linux/mm_inline.h>
32 #include <linux/firmware-map.h>
33 #include <linux/stop_machine.h>
34 #include <linux/hugetlb.h>
35 #include <linux/memblock.h>
36 #include <linux/compaction.h>
37 #include <linux/rmap.h>
38 #include <linux/module.h>
39 
40 #include <asm/tlbflush.h>
41 
42 #include "internal.h"
43 #include "shuffle.h"
44 
45 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
46 /*
47  * memory_hotplug.memmap_on_memory parameter
48  */
49 static bool memmap_on_memory __ro_after_init;
50 module_param(memmap_on_memory, bool, 0444);
51 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
52 
mhp_memmap_on_memory(void)53 static inline bool mhp_memmap_on_memory(void)
54 {
55 	return memmap_on_memory;
56 }
57 #else
mhp_memmap_on_memory(void)58 static inline bool mhp_memmap_on_memory(void)
59 {
60 	return false;
61 }
62 #endif
63 
64 enum {
65 	ONLINE_POLICY_CONTIG_ZONES = 0,
66 	ONLINE_POLICY_AUTO_MOVABLE,
67 };
68 
69 static const char * const online_policy_to_str[] = {
70 	[ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
71 	[ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
72 };
73 
set_online_policy(const char * val,const struct kernel_param * kp)74 static int set_online_policy(const char *val, const struct kernel_param *kp)
75 {
76 	int ret = sysfs_match_string(online_policy_to_str, val);
77 
78 	if (ret < 0)
79 		return ret;
80 	*((int *)kp->arg) = ret;
81 	return 0;
82 }
83 
get_online_policy(char * buffer,const struct kernel_param * kp)84 static int get_online_policy(char *buffer, const struct kernel_param *kp)
85 {
86 	return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
87 }
88 
89 /*
90  * memory_hotplug.online_policy: configure online behavior when onlining without
91  * specifying a zone (MMOP_ONLINE)
92  *
93  * "contig-zones": keep zone contiguous
94  * "auto-movable": online memory to ZONE_MOVABLE if the configuration
95  *                 (auto_movable_ratio, auto_movable_numa_aware) allows for it
96  */
97 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
98 static const struct kernel_param_ops online_policy_ops = {
99 	.set = set_online_policy,
100 	.get = get_online_policy,
101 };
102 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
103 MODULE_PARM_DESC(online_policy,
104 		"Set the online policy (\"contig-zones\", \"auto-movable\") "
105 		"Default: \"contig-zones\"");
106 
107 /*
108  * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
109  *
110  * The ratio represent an upper limit and the kernel might decide to not
111  * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
112  * doesn't allow for more MOVABLE memory.
113  */
114 static unsigned int auto_movable_ratio __read_mostly = 301;
115 module_param(auto_movable_ratio, uint, 0644);
116 MODULE_PARM_DESC(auto_movable_ratio,
117 		"Set the maximum ratio of MOVABLE:KERNEL memory in the system "
118 		"in percent for \"auto-movable\" online policy. Default: 301");
119 
120 /*
121  * memory_hotplug.auto_movable_numa_aware: consider numa node stats
122  */
123 #ifdef CONFIG_NUMA
124 static bool auto_movable_numa_aware __read_mostly = true;
125 module_param(auto_movable_numa_aware, bool, 0644);
126 MODULE_PARM_DESC(auto_movable_numa_aware,
127 		"Consider numa node stats in addition to global stats in "
128 		"\"auto-movable\" online policy. Default: true");
129 #endif /* CONFIG_NUMA */
130 
131 /*
132  * online_page_callback contains pointer to current page onlining function.
133  * Initially it is generic_online_page(). If it is required it could be
134  * changed by calling set_online_page_callback() for callback registration
135  * and restore_online_page_callback() for generic callback restore.
136  */
137 
138 static online_page_callback_t online_page_callback = generic_online_page;
139 static DEFINE_MUTEX(online_page_callback_lock);
140 
141 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
142 
get_online_mems(void)143 void get_online_mems(void)
144 {
145 	percpu_down_read(&mem_hotplug_lock);
146 }
147 
put_online_mems(void)148 void put_online_mems(void)
149 {
150 	percpu_up_read(&mem_hotplug_lock);
151 }
152 
153 bool movable_node_enabled = false;
154 
155 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
156 int mhp_default_online_type = MMOP_OFFLINE;
157 #else
158 int mhp_default_online_type = MMOP_ONLINE;
159 #endif
160 
setup_memhp_default_state(char * str)161 static int __init setup_memhp_default_state(char *str)
162 {
163 	const int online_type = mhp_online_type_from_str(str);
164 
165 	if (online_type >= 0)
166 		mhp_default_online_type = online_type;
167 
168 	return 1;
169 }
170 __setup("memhp_default_state=", setup_memhp_default_state);
171 
mem_hotplug_begin(void)172 void mem_hotplug_begin(void)
173 {
174 	cpus_read_lock();
175 	percpu_down_write(&mem_hotplug_lock);
176 }
177 
mem_hotplug_done(void)178 void mem_hotplug_done(void)
179 {
180 	percpu_up_write(&mem_hotplug_lock);
181 	cpus_read_unlock();
182 }
183 
184 u64 max_mem_size = U64_MAX;
185 
186 /* add this memory to iomem resource */
register_memory_resource(u64 start,u64 size,const char * resource_name)187 static struct resource *register_memory_resource(u64 start, u64 size,
188 						 const char *resource_name)
189 {
190 	struct resource *res;
191 	unsigned long flags =  IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
192 
193 	if (strcmp(resource_name, "System RAM"))
194 		flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
195 
196 	if (!mhp_range_allowed(start, size, true))
197 		return ERR_PTR(-E2BIG);
198 
199 	/*
200 	 * Make sure value parsed from 'mem=' only restricts memory adding
201 	 * while booting, so that memory hotplug won't be impacted. Please
202 	 * refer to document of 'mem=' in kernel-parameters.txt for more
203 	 * details.
204 	 */
205 	if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
206 		return ERR_PTR(-E2BIG);
207 
208 	/*
209 	 * Request ownership of the new memory range.  This might be
210 	 * a child of an existing resource that was present but
211 	 * not marked as busy.
212 	 */
213 	res = __request_region(&iomem_resource, start, size,
214 			       resource_name, flags);
215 
216 	if (!res) {
217 		pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
218 				start, start + size);
219 		return ERR_PTR(-EEXIST);
220 	}
221 	return res;
222 }
223 
release_memory_resource(struct resource * res)224 static void release_memory_resource(struct resource *res)
225 {
226 	if (!res)
227 		return;
228 	release_resource(res);
229 	kfree(res);
230 }
231 
check_pfn_span(unsigned long pfn,unsigned long nr_pages)232 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
233 {
234 	/*
235 	 * Disallow all operations smaller than a sub-section and only
236 	 * allow operations smaller than a section for
237 	 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
238 	 * enforces a larger memory_block_size_bytes() granularity for
239 	 * memory that will be marked online, so this check should only
240 	 * fire for direct arch_{add,remove}_memory() users outside of
241 	 * add_memory_resource().
242 	 */
243 	unsigned long min_align;
244 
245 	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
246 		min_align = PAGES_PER_SUBSECTION;
247 	else
248 		min_align = PAGES_PER_SECTION;
249 	if (!IS_ALIGNED(pfn | nr_pages, min_align))
250 		return -EINVAL;
251 	return 0;
252 }
253 
254 /*
255  * Return page for the valid pfn only if the page is online. All pfn
256  * walkers which rely on the fully initialized page->flags and others
257  * should use this rather than pfn_valid && pfn_to_page
258  */
pfn_to_online_page(unsigned long pfn)259 struct page *pfn_to_online_page(unsigned long pfn)
260 {
261 	unsigned long nr = pfn_to_section_nr(pfn);
262 	struct dev_pagemap *pgmap;
263 	struct mem_section *ms;
264 
265 	if (nr >= NR_MEM_SECTIONS)
266 		return NULL;
267 
268 	ms = __nr_to_section(nr);
269 	if (!online_section(ms))
270 		return NULL;
271 
272 	/*
273 	 * Save some code text when online_section() +
274 	 * pfn_section_valid() are sufficient.
275 	 */
276 	if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
277 		return NULL;
278 
279 	if (!pfn_section_valid(ms, pfn))
280 		return NULL;
281 
282 	if (!online_device_section(ms))
283 		return pfn_to_page(pfn);
284 
285 	/*
286 	 * Slowpath: when ZONE_DEVICE collides with
287 	 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
288 	 * the section may be 'offline' but 'valid'. Only
289 	 * get_dev_pagemap() can determine sub-section online status.
290 	 */
291 	pgmap = get_dev_pagemap(pfn, NULL);
292 	put_dev_pagemap(pgmap);
293 
294 	/* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
295 	if (pgmap)
296 		return NULL;
297 
298 	return pfn_to_page(pfn);
299 }
300 EXPORT_SYMBOL_GPL(pfn_to_online_page);
301 
__add_pages(int nid,unsigned long pfn,unsigned long nr_pages,struct mhp_params * params)302 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
303 		struct mhp_params *params)
304 {
305 	const unsigned long end_pfn = pfn + nr_pages;
306 	unsigned long cur_nr_pages;
307 	int err;
308 	struct vmem_altmap *altmap = params->altmap;
309 
310 	if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
311 		return -EINVAL;
312 
313 	VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
314 
315 	if (altmap) {
316 		/*
317 		 * Validate altmap is within bounds of the total request
318 		 */
319 		if (altmap->base_pfn != pfn
320 				|| vmem_altmap_offset(altmap) > nr_pages) {
321 			pr_warn_once("memory add fail, invalid altmap\n");
322 			return -EINVAL;
323 		}
324 		altmap->alloc = 0;
325 	}
326 
327 	if (check_pfn_span(pfn, nr_pages)) {
328 		WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
329 		return -EINVAL;
330 	}
331 
332 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
333 		/* Select all remaining pages up to the next section boundary */
334 		cur_nr_pages = min(end_pfn - pfn,
335 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
336 		err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
337 					 params->pgmap);
338 		if (err)
339 			break;
340 		cond_resched();
341 	}
342 	vmemmap_populate_print_last();
343 	return err;
344 }
345 
346 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
find_smallest_section_pfn(int nid,struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)347 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
348 				     unsigned long start_pfn,
349 				     unsigned long end_pfn)
350 {
351 	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
352 		if (unlikely(!pfn_to_online_page(start_pfn)))
353 			continue;
354 
355 		if (unlikely(pfn_to_nid(start_pfn) != nid))
356 			continue;
357 
358 		if (zone != page_zone(pfn_to_page(start_pfn)))
359 			continue;
360 
361 		return start_pfn;
362 	}
363 
364 	return 0;
365 }
366 
367 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
find_biggest_section_pfn(int nid,struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)368 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
369 				    unsigned long start_pfn,
370 				    unsigned long end_pfn)
371 {
372 	unsigned long pfn;
373 
374 	/* pfn is the end pfn of a memory section. */
375 	pfn = end_pfn - 1;
376 	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
377 		if (unlikely(!pfn_to_online_page(pfn)))
378 			continue;
379 
380 		if (unlikely(pfn_to_nid(pfn) != nid))
381 			continue;
382 
383 		if (zone != page_zone(pfn_to_page(pfn)))
384 			continue;
385 
386 		return pfn;
387 	}
388 
389 	return 0;
390 }
391 
shrink_zone_span(struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)392 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
393 			     unsigned long end_pfn)
394 {
395 	unsigned long pfn;
396 	int nid = zone_to_nid(zone);
397 
398 	if (zone->zone_start_pfn == start_pfn) {
399 		/*
400 		 * If the section is smallest section in the zone, it need
401 		 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
402 		 * In this case, we find second smallest valid mem_section
403 		 * for shrinking zone.
404 		 */
405 		pfn = find_smallest_section_pfn(nid, zone, end_pfn,
406 						zone_end_pfn(zone));
407 		if (pfn) {
408 			zone->spanned_pages = zone_end_pfn(zone) - pfn;
409 			zone->zone_start_pfn = pfn;
410 		} else {
411 			zone->zone_start_pfn = 0;
412 			zone->spanned_pages = 0;
413 		}
414 	} else if (zone_end_pfn(zone) == end_pfn) {
415 		/*
416 		 * If the section is biggest section in the zone, it need
417 		 * shrink zone->spanned_pages.
418 		 * In this case, we find second biggest valid mem_section for
419 		 * shrinking zone.
420 		 */
421 		pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
422 					       start_pfn);
423 		if (pfn)
424 			zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
425 		else {
426 			zone->zone_start_pfn = 0;
427 			zone->spanned_pages = 0;
428 		}
429 	}
430 }
431 
update_pgdat_span(struct pglist_data * pgdat)432 static void update_pgdat_span(struct pglist_data *pgdat)
433 {
434 	unsigned long node_start_pfn = 0, node_end_pfn = 0;
435 	struct zone *zone;
436 
437 	for (zone = pgdat->node_zones;
438 	     zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
439 		unsigned long end_pfn = zone_end_pfn(zone);
440 
441 		/* No need to lock the zones, they can't change. */
442 		if (!zone->spanned_pages)
443 			continue;
444 		if (!node_end_pfn) {
445 			node_start_pfn = zone->zone_start_pfn;
446 			node_end_pfn = end_pfn;
447 			continue;
448 		}
449 
450 		if (end_pfn > node_end_pfn)
451 			node_end_pfn = end_pfn;
452 		if (zone->zone_start_pfn < node_start_pfn)
453 			node_start_pfn = zone->zone_start_pfn;
454 	}
455 
456 	pgdat->node_start_pfn = node_start_pfn;
457 	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
458 }
459 
remove_pfn_range_from_zone(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)460 void __ref remove_pfn_range_from_zone(struct zone *zone,
461 				      unsigned long start_pfn,
462 				      unsigned long nr_pages)
463 {
464 	const unsigned long end_pfn = start_pfn + nr_pages;
465 	struct pglist_data *pgdat = zone->zone_pgdat;
466 	unsigned long pfn, cur_nr_pages;
467 
468 	/* Poison struct pages because they are now uninitialized again. */
469 	for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
470 		cond_resched();
471 
472 		/* Select all remaining pages up to the next section boundary */
473 		cur_nr_pages =
474 			min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
475 		page_init_poison(pfn_to_page(pfn),
476 				 sizeof(struct page) * cur_nr_pages);
477 	}
478 
479 	/*
480 	 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
481 	 * we will not try to shrink the zones - which is okay as
482 	 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
483 	 */
484 	if (zone_is_zone_device(zone))
485 		return;
486 
487 	clear_zone_contiguous(zone);
488 
489 	shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
490 	update_pgdat_span(pgdat);
491 
492 	set_zone_contiguous(zone);
493 }
494 
__remove_section(unsigned long pfn,unsigned long nr_pages,unsigned long map_offset,struct vmem_altmap * altmap)495 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
496 			     unsigned long map_offset,
497 			     struct vmem_altmap *altmap)
498 {
499 	struct mem_section *ms = __pfn_to_section(pfn);
500 
501 	if (WARN_ON_ONCE(!valid_section(ms)))
502 		return;
503 
504 	sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
505 }
506 
507 /**
508  * __remove_pages() - remove sections of pages
509  * @pfn: starting pageframe (must be aligned to start of a section)
510  * @nr_pages: number of pages to remove (must be multiple of section size)
511  * @altmap: alternative device page map or %NULL if default memmap is used
512  *
513  * Generic helper function to remove section mappings and sysfs entries
514  * for the section of the memory we are removing. Caller needs to make
515  * sure that pages are marked reserved and zones are adjust properly by
516  * calling offline_pages().
517  */
__remove_pages(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)518 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
519 		    struct vmem_altmap *altmap)
520 {
521 	const unsigned long end_pfn = pfn + nr_pages;
522 	unsigned long cur_nr_pages;
523 	unsigned long map_offset = 0;
524 
525 	map_offset = vmem_altmap_offset(altmap);
526 
527 	if (check_pfn_span(pfn, nr_pages)) {
528 		WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
529 		return;
530 	}
531 
532 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
533 		cond_resched();
534 		/* Select all remaining pages up to the next section boundary */
535 		cur_nr_pages = min(end_pfn - pfn,
536 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
537 		__remove_section(pfn, cur_nr_pages, map_offset, altmap);
538 		map_offset = 0;
539 	}
540 }
541 
set_online_page_callback(online_page_callback_t callback)542 int set_online_page_callback(online_page_callback_t callback)
543 {
544 	int rc = -EINVAL;
545 
546 	get_online_mems();
547 	mutex_lock(&online_page_callback_lock);
548 
549 	if (online_page_callback == generic_online_page) {
550 		online_page_callback = callback;
551 		rc = 0;
552 	}
553 
554 	mutex_unlock(&online_page_callback_lock);
555 	put_online_mems();
556 
557 	return rc;
558 }
559 EXPORT_SYMBOL_GPL(set_online_page_callback);
560 
restore_online_page_callback(online_page_callback_t callback)561 int restore_online_page_callback(online_page_callback_t callback)
562 {
563 	int rc = -EINVAL;
564 
565 	get_online_mems();
566 	mutex_lock(&online_page_callback_lock);
567 
568 	if (online_page_callback == callback) {
569 		online_page_callback = generic_online_page;
570 		rc = 0;
571 	}
572 
573 	mutex_unlock(&online_page_callback_lock);
574 	put_online_mems();
575 
576 	return rc;
577 }
578 EXPORT_SYMBOL_GPL(restore_online_page_callback);
579 
generic_online_page(struct page * page,unsigned int order)580 void generic_online_page(struct page *page, unsigned int order)
581 {
582 	/*
583 	 * Freeing the page with debug_pagealloc enabled will try to unmap it,
584 	 * so we should map it first. This is better than introducing a special
585 	 * case in page freeing fast path.
586 	 */
587 	debug_pagealloc_map_pages(page, 1 << order);
588 	__free_pages_core(page, order);
589 	totalram_pages_add(1UL << order);
590 }
591 EXPORT_SYMBOL_GPL(generic_online_page);
592 
online_pages_range(unsigned long start_pfn,unsigned long nr_pages)593 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
594 {
595 	const unsigned long end_pfn = start_pfn + nr_pages;
596 	unsigned long pfn;
597 
598 	/*
599 	 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
600 	 * decide to not expose all pages to the buddy (e.g., expose them
601 	 * later). We account all pages as being online and belonging to this
602 	 * zone ("present").
603 	 * When using memmap_on_memory, the range might not be aligned to
604 	 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
605 	 * this and the first chunk to online will be pageblock_nr_pages.
606 	 */
607 	for (pfn = start_pfn; pfn < end_pfn;) {
608 		int order = min(MAX_ORDER - 1UL, __ffs(pfn));
609 
610 		(*online_page_callback)(pfn_to_page(pfn), order);
611 		pfn += (1UL << order);
612 	}
613 
614 	/* mark all involved sections as online */
615 	online_mem_sections(start_pfn, end_pfn);
616 }
617 
618 /* check which state of node_states will be changed when online memory */
node_states_check_changes_online(unsigned long nr_pages,struct zone * zone,struct memory_notify * arg)619 static void node_states_check_changes_online(unsigned long nr_pages,
620 	struct zone *zone, struct memory_notify *arg)
621 {
622 	int nid = zone_to_nid(zone);
623 
624 	arg->status_change_nid = NUMA_NO_NODE;
625 	arg->status_change_nid_normal = NUMA_NO_NODE;
626 
627 	if (!node_state(nid, N_MEMORY))
628 		arg->status_change_nid = nid;
629 	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
630 		arg->status_change_nid_normal = nid;
631 }
632 
node_states_set_node(int node,struct memory_notify * arg)633 static void node_states_set_node(int node, struct memory_notify *arg)
634 {
635 	if (arg->status_change_nid_normal >= 0)
636 		node_set_state(node, N_NORMAL_MEMORY);
637 
638 	if (arg->status_change_nid >= 0)
639 		node_set_state(node, N_MEMORY);
640 }
641 
resize_zone_range(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)642 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
643 		unsigned long nr_pages)
644 {
645 	unsigned long old_end_pfn = zone_end_pfn(zone);
646 
647 	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
648 		zone->zone_start_pfn = start_pfn;
649 
650 	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
651 }
652 
resize_pgdat_range(struct pglist_data * pgdat,unsigned long start_pfn,unsigned long nr_pages)653 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
654                                      unsigned long nr_pages)
655 {
656 	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
657 
658 	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
659 		pgdat->node_start_pfn = start_pfn;
660 
661 	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
662 
663 }
664 
665 #ifdef CONFIG_ZONE_DEVICE
section_taint_zone_device(unsigned long pfn)666 static void section_taint_zone_device(unsigned long pfn)
667 {
668 	struct mem_section *ms = __pfn_to_section(pfn);
669 
670 	ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
671 }
672 #else
section_taint_zone_device(unsigned long pfn)673 static inline void section_taint_zone_device(unsigned long pfn)
674 {
675 }
676 #endif
677 
678 /*
679  * Associate the pfn range with the given zone, initializing the memmaps
680  * and resizing the pgdat/zone data to span the added pages. After this
681  * call, all affected pages are PG_reserved.
682  *
683  * All aligned pageblocks are initialized to the specified migratetype
684  * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
685  * zone stats (e.g., nr_isolate_pageblock) are touched.
686  */
move_pfn_range_to_zone(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages,struct vmem_altmap * altmap,int migratetype)687 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
688 				  unsigned long nr_pages,
689 				  struct vmem_altmap *altmap, int migratetype)
690 {
691 	struct pglist_data *pgdat = zone->zone_pgdat;
692 	int nid = pgdat->node_id;
693 
694 	clear_zone_contiguous(zone);
695 
696 	if (zone_is_empty(zone))
697 		init_currently_empty_zone(zone, start_pfn, nr_pages);
698 	resize_zone_range(zone, start_pfn, nr_pages);
699 	resize_pgdat_range(pgdat, start_pfn, nr_pages);
700 
701 	/*
702 	 * Subsection population requires care in pfn_to_online_page().
703 	 * Set the taint to enable the slow path detection of
704 	 * ZONE_DEVICE pages in an otherwise  ZONE_{NORMAL,MOVABLE}
705 	 * section.
706 	 */
707 	if (zone_is_zone_device(zone)) {
708 		if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
709 			section_taint_zone_device(start_pfn);
710 		if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
711 			section_taint_zone_device(start_pfn + nr_pages);
712 	}
713 
714 	/*
715 	 * TODO now we have a visible range of pages which are not associated
716 	 * with their zone properly. Not nice but set_pfnblock_flags_mask
717 	 * expects the zone spans the pfn range. All the pages in the range
718 	 * are reserved so nobody should be touching them so we should be safe
719 	 */
720 	memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
721 			 MEMINIT_HOTPLUG, altmap, migratetype);
722 
723 	set_zone_contiguous(zone);
724 }
725 
726 struct auto_movable_stats {
727 	unsigned long kernel_early_pages;
728 	unsigned long movable_pages;
729 };
730 
auto_movable_stats_account_zone(struct auto_movable_stats * stats,struct zone * zone)731 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
732 					    struct zone *zone)
733 {
734 	if (zone_idx(zone) == ZONE_MOVABLE) {
735 		stats->movable_pages += zone->present_pages;
736 	} else {
737 		stats->kernel_early_pages += zone->present_early_pages;
738 #ifdef CONFIG_CMA
739 		/*
740 		 * CMA pages (never on hotplugged memory) behave like
741 		 * ZONE_MOVABLE.
742 		 */
743 		stats->movable_pages += zone->cma_pages;
744 		stats->kernel_early_pages -= zone->cma_pages;
745 #endif /* CONFIG_CMA */
746 	}
747 }
748 struct auto_movable_group_stats {
749 	unsigned long movable_pages;
750 	unsigned long req_kernel_early_pages;
751 };
752 
auto_movable_stats_account_group(struct memory_group * group,void * arg)753 static int auto_movable_stats_account_group(struct memory_group *group,
754 					   void *arg)
755 {
756 	const int ratio = READ_ONCE(auto_movable_ratio);
757 	struct auto_movable_group_stats *stats = arg;
758 	long pages;
759 
760 	/*
761 	 * We don't support modifying the config while the auto-movable online
762 	 * policy is already enabled. Just avoid the division by zero below.
763 	 */
764 	if (!ratio)
765 		return 0;
766 
767 	/*
768 	 * Calculate how many early kernel pages this group requires to
769 	 * satisfy the configured zone ratio.
770 	 */
771 	pages = group->present_movable_pages * 100 / ratio;
772 	pages -= group->present_kernel_pages;
773 
774 	if (pages > 0)
775 		stats->req_kernel_early_pages += pages;
776 	stats->movable_pages += group->present_movable_pages;
777 	return 0;
778 }
779 
auto_movable_can_online_movable(int nid,struct memory_group * group,unsigned long nr_pages)780 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
781 					    unsigned long nr_pages)
782 {
783 	unsigned long kernel_early_pages, movable_pages;
784 	struct auto_movable_group_stats group_stats = {};
785 	struct auto_movable_stats stats = {};
786 	pg_data_t *pgdat = NODE_DATA(nid);
787 	struct zone *zone;
788 	int i;
789 
790 	/* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
791 	if (nid == NUMA_NO_NODE) {
792 		/* TODO: cache values */
793 		for_each_populated_zone(zone)
794 			auto_movable_stats_account_zone(&stats, zone);
795 	} else {
796 		for (i = 0; i < MAX_NR_ZONES; i++) {
797 			zone = pgdat->node_zones + i;
798 			if (populated_zone(zone))
799 				auto_movable_stats_account_zone(&stats, zone);
800 		}
801 	}
802 
803 	kernel_early_pages = stats.kernel_early_pages;
804 	movable_pages = stats.movable_pages;
805 
806 	/*
807 	 * Kernel memory inside dynamic memory group allows for more MOVABLE
808 	 * memory within the same group. Remove the effect of all but the
809 	 * current group from the stats.
810 	 */
811 	walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
812 				   group, &group_stats);
813 	if (kernel_early_pages <= group_stats.req_kernel_early_pages)
814 		return false;
815 	kernel_early_pages -= group_stats.req_kernel_early_pages;
816 	movable_pages -= group_stats.movable_pages;
817 
818 	if (group && group->is_dynamic)
819 		kernel_early_pages += group->present_kernel_pages;
820 
821 	/*
822 	 * Test if we could online the given number of pages to ZONE_MOVABLE
823 	 * and still stay in the configured ratio.
824 	 */
825 	movable_pages += nr_pages;
826 	return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
827 }
828 
829 /*
830  * Returns a default kernel memory zone for the given pfn range.
831  * If no kernel zone covers this pfn range it will automatically go
832  * to the ZONE_NORMAL.
833  */
default_kernel_zone_for_pfn(int nid,unsigned long start_pfn,unsigned long nr_pages)834 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
835 		unsigned long nr_pages)
836 {
837 	struct pglist_data *pgdat = NODE_DATA(nid);
838 	int zid;
839 
840 	for (zid = 0; zid < ZONE_NORMAL; zid++) {
841 		struct zone *zone = &pgdat->node_zones[zid];
842 
843 		if (zone_intersects(zone, start_pfn, nr_pages))
844 			return zone;
845 	}
846 
847 	return &pgdat->node_zones[ZONE_NORMAL];
848 }
849 
850 /*
851  * Determine to which zone to online memory dynamically based on user
852  * configuration and system stats. We care about the following ratio:
853  *
854  *   MOVABLE : KERNEL
855  *
856  * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
857  * one of the kernel zones. CMA pages inside one of the kernel zones really
858  * behaves like ZONE_MOVABLE, so we treat them accordingly.
859  *
860  * We don't allow for hotplugged memory in a KERNEL zone to increase the
861  * amount of MOVABLE memory we can have, so we end up with:
862  *
863  *   MOVABLE : KERNEL_EARLY
864  *
865  * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
866  * boot. We base our calculation on KERNEL_EARLY internally, because:
867  *
868  * a) Hotplugged memory in one of the kernel zones can sometimes still get
869  *    hotunplugged, especially when hot(un)plugging individual memory blocks.
870  *    There is no coordination across memory devices, therefore "automatic"
871  *    hotunplugging, as implemented in hypervisors, could result in zone
872  *    imbalances.
873  * b) Early/boot memory in one of the kernel zones can usually not get
874  *    hotunplugged again (e.g., no firmware interface to unplug, fragmented
875  *    with unmovable allocations). While there are corner cases where it might
876  *    still work, it is barely relevant in practice.
877  *
878  * Exceptions are dynamic memory groups, which allow for more MOVABLE
879  * memory within the same memory group -- because in that case, there is
880  * coordination within the single memory device managed by a single driver.
881  *
882  * We rely on "present pages" instead of "managed pages", as the latter is
883  * highly unreliable and dynamic in virtualized environments, and does not
884  * consider boot time allocations. For example, memory ballooning adjusts the
885  * managed pages when inflating/deflating the balloon, and balloon compaction
886  * can even migrate inflated pages between zones.
887  *
888  * Using "present pages" is better but some things to keep in mind are:
889  *
890  * a) Some memblock allocations, such as for the crashkernel area, are
891  *    effectively unused by the kernel, yet they account to "present pages".
892  *    Fortunately, these allocations are comparatively small in relevant setups
893  *    (e.g., fraction of system memory).
894  * b) Some hotplugged memory blocks in virtualized environments, esecially
895  *    hotplugged by virtio-mem, look like they are completely present, however,
896  *    only parts of the memory block are actually currently usable.
897  *    "present pages" is an upper limit that can get reached at runtime. As
898  *    we base our calculations on KERNEL_EARLY, this is not an issue.
899  */
auto_movable_zone_for_pfn(int nid,struct memory_group * group,unsigned long pfn,unsigned long nr_pages)900 static struct zone *auto_movable_zone_for_pfn(int nid,
901 					      struct memory_group *group,
902 					      unsigned long pfn,
903 					      unsigned long nr_pages)
904 {
905 	unsigned long online_pages = 0, max_pages, end_pfn;
906 	struct page *page;
907 
908 	if (!auto_movable_ratio)
909 		goto kernel_zone;
910 
911 	if (group && !group->is_dynamic) {
912 		max_pages = group->s.max_pages;
913 		online_pages = group->present_movable_pages;
914 
915 		/* If anything is !MOVABLE online the rest !MOVABLE. */
916 		if (group->present_kernel_pages)
917 			goto kernel_zone;
918 	} else if (!group || group->d.unit_pages == nr_pages) {
919 		max_pages = nr_pages;
920 	} else {
921 		max_pages = group->d.unit_pages;
922 		/*
923 		 * Take a look at all online sections in the current unit.
924 		 * We can safely assume that all pages within a section belong
925 		 * to the same zone, because dynamic memory groups only deal
926 		 * with hotplugged memory.
927 		 */
928 		pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
929 		end_pfn = pfn + group->d.unit_pages;
930 		for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
931 			page = pfn_to_online_page(pfn);
932 			if (!page)
933 				continue;
934 			/* If anything is !MOVABLE online the rest !MOVABLE. */
935 			if (!is_zone_movable_page(page))
936 				goto kernel_zone;
937 			online_pages += PAGES_PER_SECTION;
938 		}
939 	}
940 
941 	/*
942 	 * Online MOVABLE if we could *currently* online all remaining parts
943 	 * MOVABLE. We expect to (add+) online them immediately next, so if
944 	 * nobody interferes, all will be MOVABLE if possible.
945 	 */
946 	nr_pages = max_pages - online_pages;
947 	if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
948 		goto kernel_zone;
949 
950 #ifdef CONFIG_NUMA
951 	if (auto_movable_numa_aware &&
952 	    !auto_movable_can_online_movable(nid, group, nr_pages))
953 		goto kernel_zone;
954 #endif /* CONFIG_NUMA */
955 
956 	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
957 kernel_zone:
958 	return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
959 }
960 
default_zone_for_pfn(int nid,unsigned long start_pfn,unsigned long nr_pages)961 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
962 		unsigned long nr_pages)
963 {
964 	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
965 			nr_pages);
966 	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
967 	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
968 	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
969 
970 	/*
971 	 * We inherit the existing zone in a simple case where zones do not
972 	 * overlap in the given range
973 	 */
974 	if (in_kernel ^ in_movable)
975 		return (in_kernel) ? kernel_zone : movable_zone;
976 
977 	/*
978 	 * If the range doesn't belong to any zone or two zones overlap in the
979 	 * given range then we use movable zone only if movable_node is
980 	 * enabled because we always online to a kernel zone by default.
981 	 */
982 	return movable_node_enabled ? movable_zone : kernel_zone;
983 }
984 
zone_for_pfn_range(int online_type,int nid,struct memory_group * group,unsigned long start_pfn,unsigned long nr_pages)985 struct zone *zone_for_pfn_range(int online_type, int nid,
986 		struct memory_group *group, unsigned long start_pfn,
987 		unsigned long nr_pages)
988 {
989 	if (online_type == MMOP_ONLINE_KERNEL)
990 		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
991 
992 	if (online_type == MMOP_ONLINE_MOVABLE)
993 		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
994 
995 	if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
996 		return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
997 
998 	return default_zone_for_pfn(nid, start_pfn, nr_pages);
999 }
1000 
1001 /*
1002  * This function should only be called by memory_block_{online,offline},
1003  * and {online,offline}_pages.
1004  */
adjust_present_page_count(struct page * page,struct memory_group * group,long nr_pages)1005 void adjust_present_page_count(struct page *page, struct memory_group *group,
1006 			       long nr_pages)
1007 {
1008 	struct zone *zone = page_zone(page);
1009 	const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1010 
1011 	/*
1012 	 * We only support onlining/offlining/adding/removing of complete
1013 	 * memory blocks; therefore, either all is either early or hotplugged.
1014 	 */
1015 	if (early_section(__pfn_to_section(page_to_pfn(page))))
1016 		zone->present_early_pages += nr_pages;
1017 	zone->present_pages += nr_pages;
1018 	zone->zone_pgdat->node_present_pages += nr_pages;
1019 
1020 	if (group && movable)
1021 		group->present_movable_pages += nr_pages;
1022 	else if (group && !movable)
1023 		group->present_kernel_pages += nr_pages;
1024 }
1025 
mhp_init_memmap_on_memory(unsigned long pfn,unsigned long nr_pages,struct zone * zone)1026 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1027 			      struct zone *zone)
1028 {
1029 	unsigned long end_pfn = pfn + nr_pages;
1030 	int ret, i;
1031 
1032 	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1033 	if (ret)
1034 		return ret;
1035 
1036 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1037 
1038 	for (i = 0; i < nr_pages; i++)
1039 		SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1040 
1041 	/*
1042 	 * It might be that the vmemmap_pages fully span sections. If that is
1043 	 * the case, mark those sections online here as otherwise they will be
1044 	 * left offline.
1045 	 */
1046 	if (nr_pages >= PAGES_PER_SECTION)
1047 	        online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1048 
1049 	return ret;
1050 }
1051 
mhp_deinit_memmap_on_memory(unsigned long pfn,unsigned long nr_pages)1052 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1053 {
1054 	unsigned long end_pfn = pfn + nr_pages;
1055 
1056 	/*
1057 	 * It might be that the vmemmap_pages fully span sections. If that is
1058 	 * the case, mark those sections offline here as otherwise they will be
1059 	 * left online.
1060 	 */
1061 	if (nr_pages >= PAGES_PER_SECTION)
1062 		offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1063 
1064         /*
1065 	 * The pages associated with this vmemmap have been offlined, so
1066 	 * we can reset its state here.
1067 	 */
1068 	remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1069 	kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1070 }
1071 
online_pages(unsigned long pfn,unsigned long nr_pages,struct zone * zone,struct memory_group * group)1072 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1073 		       struct zone *zone, struct memory_group *group)
1074 {
1075 	unsigned long flags;
1076 	int need_zonelists_rebuild = 0;
1077 	const int nid = zone_to_nid(zone);
1078 	int ret;
1079 	struct memory_notify arg;
1080 
1081 	/*
1082 	 * {on,off}lining is constrained to full memory sections (or more
1083 	 * precisely to memory blocks from the user space POV).
1084 	 * memmap_on_memory is an exception because it reserves initial part
1085 	 * of the physical memory space for vmemmaps. That space is pageblock
1086 	 * aligned.
1087 	 */
1088 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1089 			 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1090 		return -EINVAL;
1091 
1092 	mem_hotplug_begin();
1093 
1094 	/* associate pfn range with the zone */
1095 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1096 
1097 	arg.start_pfn = pfn;
1098 	arg.nr_pages = nr_pages;
1099 	node_states_check_changes_online(nr_pages, zone, &arg);
1100 
1101 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
1102 	ret = notifier_to_errno(ret);
1103 	if (ret)
1104 		goto failed_addition;
1105 
1106 	/*
1107 	 * Fixup the number of isolated pageblocks before marking the sections
1108 	 * onlining, such that undo_isolate_page_range() works correctly.
1109 	 */
1110 	spin_lock_irqsave(&zone->lock, flags);
1111 	zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1112 	spin_unlock_irqrestore(&zone->lock, flags);
1113 
1114 	/*
1115 	 * If this zone is not populated, then it is not in zonelist.
1116 	 * This means the page allocator ignores this zone.
1117 	 * So, zonelist must be updated after online.
1118 	 */
1119 	if (!populated_zone(zone)) {
1120 		need_zonelists_rebuild = 1;
1121 		setup_zone_pageset(zone);
1122 	}
1123 
1124 	online_pages_range(pfn, nr_pages);
1125 	adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1126 
1127 	node_states_set_node(nid, &arg);
1128 	if (need_zonelists_rebuild)
1129 		build_all_zonelists(NULL);
1130 
1131 	/* Basic onlining is complete, allow allocation of onlined pages. */
1132 	undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1133 
1134 	/*
1135 	 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1136 	 * the tail of the freelist when undoing isolation). Shuffle the whole
1137 	 * zone to make sure the just onlined pages are properly distributed
1138 	 * across the whole freelist - to create an initial shuffle.
1139 	 */
1140 	shuffle_zone(zone);
1141 
1142 	/* reinitialise watermarks and update pcp limits */
1143 	init_per_zone_wmark_min();
1144 
1145 	kswapd_run(nid);
1146 	kcompactd_run(nid);
1147 
1148 	writeback_set_ratelimit();
1149 
1150 	memory_notify(MEM_ONLINE, &arg);
1151 	mem_hotplug_done();
1152 	return 0;
1153 
1154 failed_addition:
1155 	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1156 		 (unsigned long long) pfn << PAGE_SHIFT,
1157 		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1158 	memory_notify(MEM_CANCEL_ONLINE, &arg);
1159 	remove_pfn_range_from_zone(zone, pfn, nr_pages);
1160 	mem_hotplug_done();
1161 	return ret;
1162 }
1163 
reset_node_present_pages(pg_data_t * pgdat)1164 static void reset_node_present_pages(pg_data_t *pgdat)
1165 {
1166 	struct zone *z;
1167 
1168 	for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1169 		z->present_pages = 0;
1170 
1171 	pgdat->node_present_pages = 0;
1172 }
1173 
1174 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
hotadd_init_pgdat(int nid)1175 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1176 {
1177 	struct pglist_data *pgdat;
1178 
1179 	/*
1180 	 * NODE_DATA is preallocated (free_area_init) but its internal
1181 	 * state is not allocated completely. Add missing pieces.
1182 	 * Completely offline nodes stay around and they just need
1183 	 * reintialization.
1184 	 */
1185 	pgdat = NODE_DATA(nid);
1186 
1187 	/* init node's zones as empty zones, we don't have any present pages.*/
1188 	free_area_init_core_hotplug(pgdat);
1189 
1190 	/*
1191 	 * The node we allocated has no zone fallback lists. For avoiding
1192 	 * to access not-initialized zonelist, build here.
1193 	 */
1194 	build_all_zonelists(pgdat);
1195 
1196 	/*
1197 	 * When memory is hot-added, all the memory is in offline state. So
1198 	 * clear all zones' present_pages because they will be updated in
1199 	 * online_pages() and offline_pages().
1200 	 * TODO: should be in free_area_init_core_hotplug?
1201 	 */
1202 	reset_node_managed_pages(pgdat);
1203 	reset_node_present_pages(pgdat);
1204 
1205 	return pgdat;
1206 }
1207 
1208 /*
1209  * __try_online_node - online a node if offlined
1210  * @nid: the node ID
1211  * @set_node_online: Whether we want to online the node
1212  * called by cpu_up() to online a node without onlined memory.
1213  *
1214  * Returns:
1215  * 1 -> a new node has been allocated
1216  * 0 -> the node is already online
1217  * -ENOMEM -> the node could not be allocated
1218  */
__try_online_node(int nid,bool set_node_online)1219 static int __try_online_node(int nid, bool set_node_online)
1220 {
1221 	pg_data_t *pgdat;
1222 	int ret = 1;
1223 
1224 	if (node_online(nid))
1225 		return 0;
1226 
1227 	pgdat = hotadd_init_pgdat(nid);
1228 	if (!pgdat) {
1229 		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1230 		ret = -ENOMEM;
1231 		goto out;
1232 	}
1233 
1234 	if (set_node_online) {
1235 		node_set_online(nid);
1236 		ret = register_one_node(nid);
1237 		BUG_ON(ret);
1238 	}
1239 out:
1240 	return ret;
1241 }
1242 
1243 /*
1244  * Users of this function always want to online/register the node
1245  */
try_online_node(int nid)1246 int try_online_node(int nid)
1247 {
1248 	int ret;
1249 
1250 	mem_hotplug_begin();
1251 	ret =  __try_online_node(nid, true);
1252 	mem_hotplug_done();
1253 	return ret;
1254 }
1255 
check_hotplug_memory_range(u64 start,u64 size)1256 static int check_hotplug_memory_range(u64 start, u64 size)
1257 {
1258 	/* memory range must be block size aligned */
1259 	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1260 	    !IS_ALIGNED(size, memory_block_size_bytes())) {
1261 		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1262 		       memory_block_size_bytes(), start, size);
1263 		return -EINVAL;
1264 	}
1265 
1266 	return 0;
1267 }
1268 
online_memory_block(struct memory_block * mem,void * arg)1269 static int online_memory_block(struct memory_block *mem, void *arg)
1270 {
1271 	mem->online_type = mhp_default_online_type;
1272 	return device_online(&mem->dev);
1273 }
1274 
mhp_supports_memmap_on_memory(unsigned long size)1275 bool mhp_supports_memmap_on_memory(unsigned long size)
1276 {
1277 	unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1278 	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1279 	unsigned long remaining_size = size - vmemmap_size;
1280 
1281 	/*
1282 	 * Besides having arch support and the feature enabled at runtime, we
1283 	 * need a few more assumptions to hold true:
1284 	 *
1285 	 * a) We span a single memory block: memory onlining/offlinin;g happens
1286 	 *    in memory block granularity. We don't want the vmemmap of online
1287 	 *    memory blocks to reside on offline memory blocks. In the future,
1288 	 *    we might want to support variable-sized memory blocks to make the
1289 	 *    feature more versatile.
1290 	 *
1291 	 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1292 	 *    to populate memory from the altmap for unrelated parts (i.e.,
1293 	 *    other memory blocks)
1294 	 *
1295 	 * c) The vmemmap pages (and thereby the pages that will be exposed to
1296 	 *    the buddy) have to cover full pageblocks: memory onlining/offlining
1297 	 *    code requires applicable ranges to be page-aligned, for example, to
1298 	 *    set the migratetypes properly.
1299 	 *
1300 	 * TODO: Although we have a check here to make sure that vmemmap pages
1301 	 *       fully populate a PMD, it is not the right place to check for
1302 	 *       this. A much better solution involves improving vmemmap code
1303 	 *       to fallback to base pages when trying to populate vmemmap using
1304 	 *       altmap as an alternative source of memory, and we do not exactly
1305 	 *       populate a single PMD.
1306 	 */
1307 	return mhp_memmap_on_memory() &&
1308 	       size == memory_block_size_bytes() &&
1309 	       IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1310 	       IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1311 }
1312 
1313 /*
1314  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1315  * and online/offline operations (triggered e.g. by sysfs).
1316  *
1317  * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1318  */
add_memory_resource(int nid,struct resource * res,mhp_t mhp_flags)1319 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1320 {
1321 	struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1322 	enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1323 	struct vmem_altmap mhp_altmap = {};
1324 	struct memory_group *group = NULL;
1325 	u64 start, size;
1326 	bool new_node = false;
1327 	int ret;
1328 
1329 	start = res->start;
1330 	size = resource_size(res);
1331 
1332 	ret = check_hotplug_memory_range(start, size);
1333 	if (ret)
1334 		return ret;
1335 
1336 	if (mhp_flags & MHP_NID_IS_MGID) {
1337 		group = memory_group_find_by_id(nid);
1338 		if (!group)
1339 			return -EINVAL;
1340 		nid = group->nid;
1341 	}
1342 
1343 	if (!node_possible(nid)) {
1344 		WARN(1, "node %d was absent from the node_possible_map\n", nid);
1345 		return -EINVAL;
1346 	}
1347 
1348 	mem_hotplug_begin();
1349 
1350 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1351 		if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1352 			memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1353 		ret = memblock_add_node(start, size, nid, memblock_flags);
1354 		if (ret)
1355 			goto error_mem_hotplug_end;
1356 	}
1357 
1358 	ret = __try_online_node(nid, false);
1359 	if (ret < 0)
1360 		goto error;
1361 	new_node = ret;
1362 
1363 	/*
1364 	 * Self hosted memmap array
1365 	 */
1366 	if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1367 		if (!mhp_supports_memmap_on_memory(size)) {
1368 			ret = -EINVAL;
1369 			goto error;
1370 		}
1371 		mhp_altmap.free = PHYS_PFN(size);
1372 		mhp_altmap.base_pfn = PHYS_PFN(start);
1373 		params.altmap = &mhp_altmap;
1374 	}
1375 
1376 	/* call arch's memory hotadd */
1377 	ret = arch_add_memory(nid, start, size, &params);
1378 	if (ret < 0)
1379 		goto error;
1380 
1381 	/* create memory block devices after memory was added */
1382 	ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1383 					  group);
1384 	if (ret) {
1385 		arch_remove_memory(start, size, NULL);
1386 		goto error;
1387 	}
1388 
1389 	if (new_node) {
1390 		/* If sysfs file of new node can't be created, cpu on the node
1391 		 * can't be hot-added. There is no rollback way now.
1392 		 * So, check by BUG_ON() to catch it reluctantly..
1393 		 * We online node here. We can't roll back from here.
1394 		 */
1395 		node_set_online(nid);
1396 		ret = __register_one_node(nid);
1397 		BUG_ON(ret);
1398 	}
1399 
1400 	register_memory_blocks_under_node(nid, PFN_DOWN(start),
1401 					  PFN_UP(start + size - 1),
1402 					  MEMINIT_HOTPLUG);
1403 
1404 	/* create new memmap entry */
1405 	if (!strcmp(res->name, "System RAM"))
1406 		firmware_map_add_hotplug(start, start + size, "System RAM");
1407 
1408 	/* device_online() will take the lock when calling online_pages() */
1409 	mem_hotplug_done();
1410 
1411 	/*
1412 	 * In case we're allowed to merge the resource, flag it and trigger
1413 	 * merging now that adding succeeded.
1414 	 */
1415 	if (mhp_flags & MHP_MERGE_RESOURCE)
1416 		merge_system_ram_resource(res);
1417 
1418 	/* online pages if requested */
1419 	if (mhp_default_online_type != MMOP_OFFLINE)
1420 		walk_memory_blocks(start, size, NULL, online_memory_block);
1421 
1422 	return ret;
1423 error:
1424 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1425 		memblock_remove(start, size);
1426 error_mem_hotplug_end:
1427 	mem_hotplug_done();
1428 	return ret;
1429 }
1430 
1431 /* requires device_hotplug_lock, see add_memory_resource() */
__add_memory(int nid,u64 start,u64 size,mhp_t mhp_flags)1432 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1433 {
1434 	struct resource *res;
1435 	int ret;
1436 
1437 	res = register_memory_resource(start, size, "System RAM");
1438 	if (IS_ERR(res))
1439 		return PTR_ERR(res);
1440 
1441 	ret = add_memory_resource(nid, res, mhp_flags);
1442 	if (ret < 0)
1443 		release_memory_resource(res);
1444 	return ret;
1445 }
1446 
add_memory(int nid,u64 start,u64 size,mhp_t mhp_flags)1447 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1448 {
1449 	int rc;
1450 
1451 	lock_device_hotplug();
1452 	rc = __add_memory(nid, start, size, mhp_flags);
1453 	unlock_device_hotplug();
1454 
1455 	return rc;
1456 }
1457 EXPORT_SYMBOL_GPL(add_memory);
1458 
1459 /*
1460  * Add special, driver-managed memory to the system as system RAM. Such
1461  * memory is not exposed via the raw firmware-provided memmap as system
1462  * RAM, instead, it is detected and added by a driver - during cold boot,
1463  * after a reboot, and after kexec.
1464  *
1465  * Reasons why this memory should not be used for the initial memmap of a
1466  * kexec kernel or for placing kexec images:
1467  * - The booting kernel is in charge of determining how this memory will be
1468  *   used (e.g., use persistent memory as system RAM)
1469  * - Coordination with a hypervisor is required before this memory
1470  *   can be used (e.g., inaccessible parts).
1471  *
1472  * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1473  * memory map") are created. Also, the created memory resource is flagged
1474  * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1475  * this memory as well (esp., not place kexec images onto it).
1476  *
1477  * The resource_name (visible via /proc/iomem) has to have the format
1478  * "System RAM ($DRIVER)".
1479  */
add_memory_driver_managed(int nid,u64 start,u64 size,const char * resource_name,mhp_t mhp_flags)1480 int add_memory_driver_managed(int nid, u64 start, u64 size,
1481 			      const char *resource_name, mhp_t mhp_flags)
1482 {
1483 	struct resource *res;
1484 	int rc;
1485 
1486 	if (!resource_name ||
1487 	    strstr(resource_name, "System RAM (") != resource_name ||
1488 	    resource_name[strlen(resource_name) - 1] != ')')
1489 		return -EINVAL;
1490 
1491 	lock_device_hotplug();
1492 
1493 	res = register_memory_resource(start, size, resource_name);
1494 	if (IS_ERR(res)) {
1495 		rc = PTR_ERR(res);
1496 		goto out_unlock;
1497 	}
1498 
1499 	rc = add_memory_resource(nid, res, mhp_flags);
1500 	if (rc < 0)
1501 		release_memory_resource(res);
1502 
1503 out_unlock:
1504 	unlock_device_hotplug();
1505 	return rc;
1506 }
1507 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1508 
1509 /*
1510  * Platforms should define arch_get_mappable_range() that provides
1511  * maximum possible addressable physical memory range for which the
1512  * linear mapping could be created. The platform returned address
1513  * range must adhere to these following semantics.
1514  *
1515  * - range.start <= range.end
1516  * - Range includes both end points [range.start..range.end]
1517  *
1518  * There is also a fallback definition provided here, allowing the
1519  * entire possible physical address range in case any platform does
1520  * not define arch_get_mappable_range().
1521  */
arch_get_mappable_range(void)1522 struct range __weak arch_get_mappable_range(void)
1523 {
1524 	struct range mhp_range = {
1525 		.start = 0UL,
1526 		.end = -1ULL,
1527 	};
1528 	return mhp_range;
1529 }
1530 
mhp_get_pluggable_range(bool need_mapping)1531 struct range mhp_get_pluggable_range(bool need_mapping)
1532 {
1533 	const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1534 	struct range mhp_range;
1535 
1536 	if (need_mapping) {
1537 		mhp_range = arch_get_mappable_range();
1538 		if (mhp_range.start > max_phys) {
1539 			mhp_range.start = 0;
1540 			mhp_range.end = 0;
1541 		}
1542 		mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1543 	} else {
1544 		mhp_range.start = 0;
1545 		mhp_range.end = max_phys;
1546 	}
1547 	return mhp_range;
1548 }
1549 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1550 
mhp_range_allowed(u64 start,u64 size,bool need_mapping)1551 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1552 {
1553 	struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1554 	u64 end = start + size;
1555 
1556 	if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1557 		return true;
1558 
1559 	pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1560 		start, end, mhp_range.start, mhp_range.end);
1561 	return false;
1562 }
1563 
1564 #ifdef CONFIG_MEMORY_HOTREMOVE
1565 /*
1566  * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1567  * non-lru movable pages and hugepages). Will skip over most unmovable
1568  * pages (esp., pages that can be skipped when offlining), but bail out on
1569  * definitely unmovable pages.
1570  *
1571  * Returns:
1572  *	0 in case a movable page is found and movable_pfn was updated.
1573  *	-ENOENT in case no movable page was found.
1574  *	-EBUSY in case a definitely unmovable page was found.
1575  */
scan_movable_pages(unsigned long start,unsigned long end,unsigned long * movable_pfn)1576 static int scan_movable_pages(unsigned long start, unsigned long end,
1577 			      unsigned long *movable_pfn)
1578 {
1579 	unsigned long pfn;
1580 
1581 	for (pfn = start; pfn < end; pfn++) {
1582 		struct page *page, *head;
1583 		unsigned long skip;
1584 
1585 		if (!pfn_valid(pfn))
1586 			continue;
1587 		page = pfn_to_page(pfn);
1588 		if (PageLRU(page))
1589 			goto found;
1590 		if (__PageMovable(page))
1591 			goto found;
1592 
1593 		/*
1594 		 * PageOffline() pages that are not marked __PageMovable() and
1595 		 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1596 		 * definitely unmovable. If their reference count would be 0,
1597 		 * they could at least be skipped when offlining memory.
1598 		 */
1599 		if (PageOffline(page) && page_count(page))
1600 			return -EBUSY;
1601 
1602 		if (!PageHuge(page))
1603 			continue;
1604 		head = compound_head(page);
1605 		/*
1606 		 * This test is racy as we hold no reference or lock.  The
1607 		 * hugetlb page could have been free'ed and head is no longer
1608 		 * a hugetlb page before the following check.  In such unlikely
1609 		 * cases false positives and negatives are possible.  Calling
1610 		 * code must deal with these scenarios.
1611 		 */
1612 		if (HPageMigratable(head))
1613 			goto found;
1614 		skip = compound_nr(head) - (page - head);
1615 		pfn += skip - 1;
1616 	}
1617 	return -ENOENT;
1618 found:
1619 	*movable_pfn = pfn;
1620 	return 0;
1621 }
1622 
1623 static int
do_migrate_range(unsigned long start_pfn,unsigned long end_pfn)1624 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1625 {
1626 	unsigned long pfn;
1627 	struct page *page, *head;
1628 	int ret = 0;
1629 	LIST_HEAD(source);
1630 	static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1631 				      DEFAULT_RATELIMIT_BURST);
1632 
1633 	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1634 		struct folio *folio;
1635 
1636 		if (!pfn_valid(pfn))
1637 			continue;
1638 		page = pfn_to_page(pfn);
1639 		folio = page_folio(page);
1640 		head = &folio->page;
1641 
1642 		if (PageHuge(page)) {
1643 			pfn = page_to_pfn(head) + compound_nr(head) - 1;
1644 			isolate_hugetlb(head, &source);
1645 			continue;
1646 		} else if (PageTransHuge(page))
1647 			pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1648 
1649 		/*
1650 		 * HWPoison pages have elevated reference counts so the migration would
1651 		 * fail on them. It also doesn't make any sense to migrate them in the
1652 		 * first place. Still try to unmap such a page in case it is still mapped
1653 		 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1654 		 * the unmap as the catch all safety net).
1655 		 */
1656 		if (PageHWPoison(page)) {
1657 			if (WARN_ON(folio_test_lru(folio)))
1658 				folio_isolate_lru(folio);
1659 			if (folio_mapped(folio))
1660 				try_to_unmap(folio, TTU_IGNORE_MLOCK);
1661 			continue;
1662 		}
1663 
1664 		if (!get_page_unless_zero(page))
1665 			continue;
1666 		/*
1667 		 * We can skip free pages. And we can deal with pages on
1668 		 * LRU and non-lru movable pages.
1669 		 */
1670 		if (PageLRU(page))
1671 			ret = isolate_lru_page(page);
1672 		else
1673 			ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1674 		if (!ret) { /* Success */
1675 			list_add_tail(&page->lru, &source);
1676 			if (!__PageMovable(page))
1677 				inc_node_page_state(page, NR_ISOLATED_ANON +
1678 						    page_is_file_lru(page));
1679 
1680 		} else {
1681 			if (__ratelimit(&migrate_rs)) {
1682 				pr_warn("failed to isolate pfn %lx\n", pfn);
1683 				dump_page(page, "isolation failed");
1684 			}
1685 		}
1686 		put_page(page);
1687 	}
1688 	if (!list_empty(&source)) {
1689 		nodemask_t nmask = node_states[N_MEMORY];
1690 		struct migration_target_control mtc = {
1691 			.nmask = &nmask,
1692 			.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1693 		};
1694 
1695 		/*
1696 		 * We have checked that migration range is on a single zone so
1697 		 * we can use the nid of the first page to all the others.
1698 		 */
1699 		mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1700 
1701 		/*
1702 		 * try to allocate from a different node but reuse this node
1703 		 * if there are no other online nodes to be used (e.g. we are
1704 		 * offlining a part of the only existing node)
1705 		 */
1706 		node_clear(mtc.nid, nmask);
1707 		if (nodes_empty(nmask))
1708 			node_set(mtc.nid, nmask);
1709 		ret = migrate_pages(&source, alloc_migration_target, NULL,
1710 			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1711 		if (ret) {
1712 			list_for_each_entry(page, &source, lru) {
1713 				if (__ratelimit(&migrate_rs)) {
1714 					pr_warn("migrating pfn %lx failed ret:%d\n",
1715 						page_to_pfn(page), ret);
1716 					dump_page(page, "migration failure");
1717 				}
1718 			}
1719 			putback_movable_pages(&source);
1720 		}
1721 	}
1722 
1723 	return ret;
1724 }
1725 
cmdline_parse_movable_node(char * p)1726 static int __init cmdline_parse_movable_node(char *p)
1727 {
1728 	movable_node_enabled = true;
1729 	return 0;
1730 }
1731 early_param("movable_node", cmdline_parse_movable_node);
1732 
1733 /* check which state of node_states will be changed when offline memory */
node_states_check_changes_offline(unsigned long nr_pages,struct zone * zone,struct memory_notify * arg)1734 static void node_states_check_changes_offline(unsigned long nr_pages,
1735 		struct zone *zone, struct memory_notify *arg)
1736 {
1737 	struct pglist_data *pgdat = zone->zone_pgdat;
1738 	unsigned long present_pages = 0;
1739 	enum zone_type zt;
1740 
1741 	arg->status_change_nid = NUMA_NO_NODE;
1742 	arg->status_change_nid_normal = NUMA_NO_NODE;
1743 
1744 	/*
1745 	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1746 	 * If the memory to be offline is within the range
1747 	 * [0..ZONE_NORMAL], and it is the last present memory there,
1748 	 * the zones in that range will become empty after the offlining,
1749 	 * thus we can determine that we need to clear the node from
1750 	 * node_states[N_NORMAL_MEMORY].
1751 	 */
1752 	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1753 		present_pages += pgdat->node_zones[zt].present_pages;
1754 	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1755 		arg->status_change_nid_normal = zone_to_nid(zone);
1756 
1757 	/*
1758 	 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1759 	 * does not apply as we don't support 32bit.
1760 	 * Here we count the possible pages from ZONE_MOVABLE.
1761 	 * If after having accounted all the pages, we see that the nr_pages
1762 	 * to be offlined is over or equal to the accounted pages,
1763 	 * we know that the node will become empty, and so, we can clear
1764 	 * it for N_MEMORY as well.
1765 	 */
1766 	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1767 
1768 	if (nr_pages >= present_pages)
1769 		arg->status_change_nid = zone_to_nid(zone);
1770 }
1771 
node_states_clear_node(int node,struct memory_notify * arg)1772 static void node_states_clear_node(int node, struct memory_notify *arg)
1773 {
1774 	if (arg->status_change_nid_normal >= 0)
1775 		node_clear_state(node, N_NORMAL_MEMORY);
1776 
1777 	if (arg->status_change_nid >= 0)
1778 		node_clear_state(node, N_MEMORY);
1779 }
1780 
count_system_ram_pages_cb(unsigned long start_pfn,unsigned long nr_pages,void * data)1781 static int count_system_ram_pages_cb(unsigned long start_pfn,
1782 				     unsigned long nr_pages, void *data)
1783 {
1784 	unsigned long *nr_system_ram_pages = data;
1785 
1786 	*nr_system_ram_pages += nr_pages;
1787 	return 0;
1788 }
1789 
offline_pages(unsigned long start_pfn,unsigned long nr_pages,struct zone * zone,struct memory_group * group)1790 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1791 			struct zone *zone, struct memory_group *group)
1792 {
1793 	const unsigned long end_pfn = start_pfn + nr_pages;
1794 	unsigned long pfn, system_ram_pages = 0;
1795 	const int node = zone_to_nid(zone);
1796 	unsigned long flags;
1797 	struct memory_notify arg;
1798 	char *reason;
1799 	int ret;
1800 
1801 	/*
1802 	 * {on,off}lining is constrained to full memory sections (or more
1803 	 * precisely to memory blocks from the user space POV).
1804 	 * memmap_on_memory is an exception because it reserves initial part
1805 	 * of the physical memory space for vmemmaps. That space is pageblock
1806 	 * aligned.
1807 	 */
1808 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1809 			 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1810 		return -EINVAL;
1811 
1812 	mem_hotplug_begin();
1813 
1814 	/*
1815 	 * Don't allow to offline memory blocks that contain holes.
1816 	 * Consequently, memory blocks with holes can never get onlined
1817 	 * via the hotplug path - online_pages() - as hotplugged memory has
1818 	 * no holes. This way, we e.g., don't have to worry about marking
1819 	 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1820 	 * avoid using walk_system_ram_range() later.
1821 	 */
1822 	walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1823 			      count_system_ram_pages_cb);
1824 	if (system_ram_pages != nr_pages) {
1825 		ret = -EINVAL;
1826 		reason = "memory holes";
1827 		goto failed_removal;
1828 	}
1829 
1830 	/*
1831 	 * We only support offlining of memory blocks managed by a single zone,
1832 	 * checked by calling code. This is just a sanity check that we might
1833 	 * want to remove in the future.
1834 	 */
1835 	if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1836 			 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1837 		ret = -EINVAL;
1838 		reason = "multizone range";
1839 		goto failed_removal;
1840 	}
1841 
1842 	/*
1843 	 * Disable pcplists so that page isolation cannot race with freeing
1844 	 * in a way that pages from isolated pageblock are left on pcplists.
1845 	 */
1846 	zone_pcp_disable(zone);
1847 	lru_cache_disable();
1848 
1849 	/* set above range as isolated */
1850 	ret = start_isolate_page_range(start_pfn, end_pfn,
1851 				       MIGRATE_MOVABLE,
1852 				       MEMORY_OFFLINE | REPORT_FAILURE,
1853 				       GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1854 	if (ret) {
1855 		reason = "failure to isolate range";
1856 		goto failed_removal_pcplists_disabled;
1857 	}
1858 
1859 	arg.start_pfn = start_pfn;
1860 	arg.nr_pages = nr_pages;
1861 	node_states_check_changes_offline(nr_pages, zone, &arg);
1862 
1863 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1864 	ret = notifier_to_errno(ret);
1865 	if (ret) {
1866 		reason = "notifier failure";
1867 		goto failed_removal_isolated;
1868 	}
1869 
1870 	do {
1871 		pfn = start_pfn;
1872 		do {
1873 			if (signal_pending(current)) {
1874 				ret = -EINTR;
1875 				reason = "signal backoff";
1876 				goto failed_removal_isolated;
1877 			}
1878 
1879 			cond_resched();
1880 
1881 			ret = scan_movable_pages(pfn, end_pfn, &pfn);
1882 			if (!ret) {
1883 				/*
1884 				 * TODO: fatal migration failures should bail
1885 				 * out
1886 				 */
1887 				do_migrate_range(pfn, end_pfn);
1888 			}
1889 		} while (!ret);
1890 
1891 		if (ret != -ENOENT) {
1892 			reason = "unmovable page";
1893 			goto failed_removal_isolated;
1894 		}
1895 
1896 		/*
1897 		 * Dissolve free hugepages in the memory block before doing
1898 		 * offlining actually in order to make hugetlbfs's object
1899 		 * counting consistent.
1900 		 */
1901 		ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1902 		if (ret) {
1903 			reason = "failure to dissolve huge pages";
1904 			goto failed_removal_isolated;
1905 		}
1906 
1907 		ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1908 
1909 	} while (ret);
1910 
1911 	/* Mark all sections offline and remove free pages from the buddy. */
1912 	__offline_isolated_pages(start_pfn, end_pfn);
1913 	pr_debug("Offlined Pages %ld\n", nr_pages);
1914 
1915 	/*
1916 	 * The memory sections are marked offline, and the pageblock flags
1917 	 * effectively stale; nobody should be touching them. Fixup the number
1918 	 * of isolated pageblocks, memory onlining will properly revert this.
1919 	 */
1920 	spin_lock_irqsave(&zone->lock, flags);
1921 	zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1922 	spin_unlock_irqrestore(&zone->lock, flags);
1923 
1924 	lru_cache_enable();
1925 	zone_pcp_enable(zone);
1926 
1927 	/* removal success */
1928 	adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1929 	adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1930 
1931 	/* reinitialise watermarks and update pcp limits */
1932 	init_per_zone_wmark_min();
1933 
1934 	if (!populated_zone(zone)) {
1935 		zone_pcp_reset(zone);
1936 		build_all_zonelists(NULL);
1937 	}
1938 
1939 	node_states_clear_node(node, &arg);
1940 	if (arg.status_change_nid >= 0) {
1941 		kcompactd_stop(node);
1942 		kswapd_stop(node);
1943 	}
1944 
1945 	writeback_set_ratelimit();
1946 
1947 	memory_notify(MEM_OFFLINE, &arg);
1948 	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1949 	mem_hotplug_done();
1950 	return 0;
1951 
1952 failed_removal_isolated:
1953 	/* pushback to free area */
1954 	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1955 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
1956 failed_removal_pcplists_disabled:
1957 	lru_cache_enable();
1958 	zone_pcp_enable(zone);
1959 failed_removal:
1960 	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1961 		 (unsigned long long) start_pfn << PAGE_SHIFT,
1962 		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1963 		 reason);
1964 	mem_hotplug_done();
1965 	return ret;
1966 }
1967 
check_memblock_offlined_cb(struct memory_block * mem,void * arg)1968 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1969 {
1970 	int *nid = arg;
1971 
1972 	*nid = mem->nid;
1973 	if (unlikely(mem->state != MEM_OFFLINE)) {
1974 		phys_addr_t beginpa, endpa;
1975 
1976 		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1977 		endpa = beginpa + memory_block_size_bytes() - 1;
1978 		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1979 			&beginpa, &endpa);
1980 
1981 		return -EBUSY;
1982 	}
1983 	return 0;
1984 }
1985 
get_nr_vmemmap_pages_cb(struct memory_block * mem,void * arg)1986 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1987 {
1988 	/*
1989 	 * If not set, continue with the next block.
1990 	 */
1991 	return mem->nr_vmemmap_pages;
1992 }
1993 
check_cpu_on_node(int nid)1994 static int check_cpu_on_node(int nid)
1995 {
1996 	int cpu;
1997 
1998 	for_each_present_cpu(cpu) {
1999 		if (cpu_to_node(cpu) == nid)
2000 			/*
2001 			 * the cpu on this node isn't removed, and we can't
2002 			 * offline this node.
2003 			 */
2004 			return -EBUSY;
2005 	}
2006 
2007 	return 0;
2008 }
2009 
check_no_memblock_for_node_cb(struct memory_block * mem,void * arg)2010 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2011 {
2012 	int nid = *(int *)arg;
2013 
2014 	/*
2015 	 * If a memory block belongs to multiple nodes, the stored nid is not
2016 	 * reliable. However, such blocks are always online (e.g., cannot get
2017 	 * offlined) and, therefore, are still spanned by the node.
2018 	 */
2019 	return mem->nid == nid ? -EEXIST : 0;
2020 }
2021 
2022 /**
2023  * try_offline_node
2024  * @nid: the node ID
2025  *
2026  * Offline a node if all memory sections and cpus of the node are removed.
2027  *
2028  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2029  * and online/offline operations before this call.
2030  */
try_offline_node(int nid)2031 void try_offline_node(int nid)
2032 {
2033 	int rc;
2034 
2035 	/*
2036 	 * If the node still spans pages (especially ZONE_DEVICE), don't
2037 	 * offline it. A node spans memory after move_pfn_range_to_zone(),
2038 	 * e.g., after the memory block was onlined.
2039 	 */
2040 	if (node_spanned_pages(nid))
2041 		return;
2042 
2043 	/*
2044 	 * Especially offline memory blocks might not be spanned by the
2045 	 * node. They will get spanned by the node once they get onlined.
2046 	 * However, they link to the node in sysfs and can get onlined later.
2047 	 */
2048 	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2049 	if (rc)
2050 		return;
2051 
2052 	if (check_cpu_on_node(nid))
2053 		return;
2054 
2055 	/*
2056 	 * all memory/cpu of this node are removed, we can offline this
2057 	 * node now.
2058 	 */
2059 	node_set_offline(nid);
2060 	unregister_one_node(nid);
2061 }
2062 EXPORT_SYMBOL(try_offline_node);
2063 
try_remove_memory(u64 start,u64 size)2064 static int __ref try_remove_memory(u64 start, u64 size)
2065 {
2066 	struct vmem_altmap mhp_altmap = {};
2067 	struct vmem_altmap *altmap = NULL;
2068 	unsigned long nr_vmemmap_pages;
2069 	int rc = 0, nid = NUMA_NO_NODE;
2070 
2071 	BUG_ON(check_hotplug_memory_range(start, size));
2072 
2073 	/*
2074 	 * All memory blocks must be offlined before removing memory.  Check
2075 	 * whether all memory blocks in question are offline and return error
2076 	 * if this is not the case.
2077 	 *
2078 	 * While at it, determine the nid. Note that if we'd have mixed nodes,
2079 	 * we'd only try to offline the last determined one -- which is good
2080 	 * enough for the cases we care about.
2081 	 */
2082 	rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2083 	if (rc)
2084 		return rc;
2085 
2086 	/*
2087 	 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2088 	 * the same granularity it was added - a single memory block.
2089 	 */
2090 	if (mhp_memmap_on_memory()) {
2091 		nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2092 						      get_nr_vmemmap_pages_cb);
2093 		if (nr_vmemmap_pages) {
2094 			if (size != memory_block_size_bytes()) {
2095 				pr_warn("Refuse to remove %#llx - %#llx,"
2096 					"wrong granularity\n",
2097 					start, start + size);
2098 				return -EINVAL;
2099 			}
2100 
2101 			/*
2102 			 * Let remove_pmd_table->free_hugepage_table do the
2103 			 * right thing if we used vmem_altmap when hot-adding
2104 			 * the range.
2105 			 */
2106 			mhp_altmap.alloc = nr_vmemmap_pages;
2107 			altmap = &mhp_altmap;
2108 		}
2109 	}
2110 
2111 	/* remove memmap entry */
2112 	firmware_map_remove(start, start + size, "System RAM");
2113 
2114 	/*
2115 	 * Memory block device removal under the device_hotplug_lock is
2116 	 * a barrier against racing online attempts.
2117 	 */
2118 	remove_memory_block_devices(start, size);
2119 
2120 	mem_hotplug_begin();
2121 
2122 	arch_remove_memory(start, size, altmap);
2123 
2124 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2125 		memblock_phys_free(start, size);
2126 		memblock_remove(start, size);
2127 	}
2128 
2129 	release_mem_region_adjustable(start, size);
2130 
2131 	if (nid != NUMA_NO_NODE)
2132 		try_offline_node(nid);
2133 
2134 	mem_hotplug_done();
2135 	return 0;
2136 }
2137 
2138 /**
2139  * __remove_memory - Remove memory if every memory block is offline
2140  * @start: physical address of the region to remove
2141  * @size: size of the region to remove
2142  *
2143  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2144  * and online/offline operations before this call, as required by
2145  * try_offline_node().
2146  */
__remove_memory(u64 start,u64 size)2147 void __remove_memory(u64 start, u64 size)
2148 {
2149 
2150 	/*
2151 	 * trigger BUG() if some memory is not offlined prior to calling this
2152 	 * function
2153 	 */
2154 	if (try_remove_memory(start, size))
2155 		BUG();
2156 }
2157 
2158 /*
2159  * Remove memory if every memory block is offline, otherwise return -EBUSY is
2160  * some memory is not offline
2161  */
remove_memory(u64 start,u64 size)2162 int remove_memory(u64 start, u64 size)
2163 {
2164 	int rc;
2165 
2166 	lock_device_hotplug();
2167 	rc = try_remove_memory(start, size);
2168 	unlock_device_hotplug();
2169 
2170 	return rc;
2171 }
2172 EXPORT_SYMBOL_GPL(remove_memory);
2173 
try_offline_memory_block(struct memory_block * mem,void * arg)2174 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2175 {
2176 	uint8_t online_type = MMOP_ONLINE_KERNEL;
2177 	uint8_t **online_types = arg;
2178 	struct page *page;
2179 	int rc;
2180 
2181 	/*
2182 	 * Sense the online_type via the zone of the memory block. Offlining
2183 	 * with multiple zones within one memory block will be rejected
2184 	 * by offlining code ... so we don't care about that.
2185 	 */
2186 	page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2187 	if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2188 		online_type = MMOP_ONLINE_MOVABLE;
2189 
2190 	rc = device_offline(&mem->dev);
2191 	/*
2192 	 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2193 	 * so try_reonline_memory_block() can do the right thing.
2194 	 */
2195 	if (!rc)
2196 		**online_types = online_type;
2197 
2198 	(*online_types)++;
2199 	/* Ignore if already offline. */
2200 	return rc < 0 ? rc : 0;
2201 }
2202 
try_reonline_memory_block(struct memory_block * mem,void * arg)2203 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2204 {
2205 	uint8_t **online_types = arg;
2206 	int rc;
2207 
2208 	if (**online_types != MMOP_OFFLINE) {
2209 		mem->online_type = **online_types;
2210 		rc = device_online(&mem->dev);
2211 		if (rc < 0)
2212 			pr_warn("%s: Failed to re-online memory: %d",
2213 				__func__, rc);
2214 	}
2215 
2216 	/* Continue processing all remaining memory blocks. */
2217 	(*online_types)++;
2218 	return 0;
2219 }
2220 
2221 /*
2222  * Try to offline and remove memory. Might take a long time to finish in case
2223  * memory is still in use. Primarily useful for memory devices that logically
2224  * unplugged all memory (so it's no longer in use) and want to offline + remove
2225  * that memory.
2226  */
offline_and_remove_memory(u64 start,u64 size)2227 int offline_and_remove_memory(u64 start, u64 size)
2228 {
2229 	const unsigned long mb_count = size / memory_block_size_bytes();
2230 	uint8_t *online_types, *tmp;
2231 	int rc;
2232 
2233 	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2234 	    !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2235 		return -EINVAL;
2236 
2237 	/*
2238 	 * We'll remember the old online type of each memory block, so we can
2239 	 * try to revert whatever we did when offlining one memory block fails
2240 	 * after offlining some others succeeded.
2241 	 */
2242 	online_types = kmalloc_array(mb_count, sizeof(*online_types),
2243 				     GFP_KERNEL);
2244 	if (!online_types)
2245 		return -ENOMEM;
2246 	/*
2247 	 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2248 	 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2249 	 * try_reonline_memory_block().
2250 	 */
2251 	memset(online_types, MMOP_OFFLINE, mb_count);
2252 
2253 	lock_device_hotplug();
2254 
2255 	tmp = online_types;
2256 	rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2257 
2258 	/*
2259 	 * In case we succeeded to offline all memory, remove it.
2260 	 * This cannot fail as it cannot get onlined in the meantime.
2261 	 */
2262 	if (!rc) {
2263 		rc = try_remove_memory(start, size);
2264 		if (rc)
2265 			pr_err("%s: Failed to remove memory: %d", __func__, rc);
2266 	}
2267 
2268 	/*
2269 	 * Rollback what we did. While memory onlining might theoretically fail
2270 	 * (nacked by a notifier), it barely ever happens.
2271 	 */
2272 	if (rc) {
2273 		tmp = online_types;
2274 		walk_memory_blocks(start, size, &tmp,
2275 				   try_reonline_memory_block);
2276 	}
2277 	unlock_device_hotplug();
2278 
2279 	kfree(online_types);
2280 	return rc;
2281 }
2282 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2283 #endif /* CONFIG_MEMORY_HOTREMOVE */
2284