page_isolation.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * linux/mm/page_isolation.c
  4 */
  5
  6#include <linux/mm.h>
  7#include <linux/page-isolation.h>
  8#include <linux/pageblock-flags.h>
  9#include <linux/memory.h>
 10#include <linux/hugetlb.h>
 11#include <linux/page_owner.h>
 12#include <linux/migrate.h>
 13#include "internal.h"
 14
 15#define CREATE_TRACE_POINTS
 16#include <trace/events/page_isolation.h>
 17
 18/*
 19 * This function checks whether the range [start_pfn, end_pfn) includes
 20 * unmovable pages or not. The range must fall into a single pageblock and
 21 * consequently belong to a single zone.
 22 *
 23 * PageLRU check without isolation or lru_lock could race so that
 24 * MIGRATE_MOVABLE block might include unmovable pages. Similarly, pages
 25 * with movable_ops can only be identified some time after they were
 26 * allocated. So you can't expect this function should be exact.
 27 *
 28 * Returns a page without holding a reference. If the caller wants to
 29 * dereference that page (e.g., dumping), it has to make sure that it
 30 * cannot get removed (e.g., via memory unplug) concurrently.
 31 *
 32 */
 33static struct page *has_unmovable_pages(unsigned long start_pfn, unsigned long end_pfn,
 34				enum pb_isolate_mode mode)
 35{
 36	struct page *page = pfn_to_page(start_pfn);
 37	struct zone *zone = page_zone(page);
 38	unsigned long pfn;
 39
 40	VM_BUG_ON(pageblock_start_pfn(start_pfn) !=
 41		  pageblock_start_pfn(end_pfn - 1));
 42
 43	if (is_migrate_cma_page(page)) {
 44		/*
 45		 * CMA allocations (alloc_contig_range) really need to mark
 46		 * isolate CMA pageblocks even when they are not movable in fact
 47		 * so consider them movable here.
 48		 */
 49		if (mode == PB_ISOLATE_MODE_CMA_ALLOC)
 50			return NULL;
 51
 52		return page;
 53	}
 54
 55	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
 56		page = pfn_to_page(pfn);
 57
 58		/*
 59		 * Both, bootmem allocations and memory holes are marked
 60		 * PG_reserved and are unmovable. We can even have unmovable
 61		 * allocations inside ZONE_MOVABLE, for example when
 62		 * specifying "movablecore".
 63		 */
 64		if (PageReserved(page))
 65			return page;
 66
 67		/*
 68		 * If the zone is movable and we have ruled out all reserved
 69		 * pages then it should be reasonably safe to assume the rest
 70		 * is movable.
 71		 */
 72		if (zone_idx(zone) == ZONE_MOVABLE)
 73			continue;
 74
 75		/*
 76		 * Hugepages are not in LRU lists, but they're movable.
 77		 * THPs are on the LRU, but need to be counted as #small pages.
 78		 * We need not scan over tail pages because we don't
 79		 * handle each tail page individually in migration.
 80		 */
 81		if (PageHuge(page) || PageTransCompound(page)) {
 82			struct folio *folio = page_folio(page);
 83			unsigned int skip_pages;
 84
 85			if (PageHuge(page)) {
 86				struct hstate *h;
 87
 88				/*
 89				 * The huge page may be freed so can not
 90				 * use folio_hstate() directly.
 91				 */
 92				h = size_to_hstate(folio_size(folio));
 93				if (h && !hugepage_migration_supported(h))
 94					return page;
 95			} else if (!folio_test_lru(folio)) {
 96				return page;
 97			}
 98
 99			skip_pages = folio_nr_pages(folio) - folio_page_idx(folio, page);
100			pfn += skip_pages - 1;
101			continue;
102		}
103
104		/*
105		 * We can't use page_count without pin a page
106		 * because another CPU can free compound page.
107		 * This check already skips compound tails of THP
108		 * because their page->_refcount is zero at all time.
109		 */
110		if (!page_ref_count(page)) {
111			if (PageBuddy(page))
112				pfn += (1 << buddy_order(page)) - 1;
113			continue;
114		}
115
116		/*
117		 * The HWPoisoned page may be not in buddy system, and
118		 * page_count() is not 0.
119		 */
120		if ((mode == PB_ISOLATE_MODE_MEM_OFFLINE) && PageHWPoison(page))
121			continue;
122
123		/*
124		 * We treat all PageOffline() pages as movable when offlining
125		 * to give drivers a chance to decrement their reference count
126		 * in MEM_GOING_OFFLINE in order to indicate that these pages
127		 * can be offlined as there are no direct references anymore.
128		 * For actually unmovable PageOffline() where the driver does
129		 * not support this, we will fail later when trying to actually
130		 * move these pages that still have a reference count > 0.
131		 * (false negatives in this function only)
132		 */
133		if ((mode == PB_ISOLATE_MODE_MEM_OFFLINE) && PageOffline(page))
134			continue;
135
136		if (PageLRU(page) || page_has_movable_ops(page))
137			continue;
138
139		/*
140		 * If there are RECLAIMABLE pages, we need to check
141		 * it.  But now, memory offline itself doesn't call
142		 * shrink_node_slabs() and it still to be fixed.
143		 */
144		return page;
145	}
146	return NULL;
147}
148
149/*
150 * This function set pageblock migratetype to isolate if no unmovable page is
151 * present in [start_pfn, end_pfn). The pageblock must intersect with
152 * [start_pfn, end_pfn).
153 */
154static int set_migratetype_isolate(struct page *page, enum pb_isolate_mode mode,
155			unsigned long start_pfn, unsigned long end_pfn)
156{
157	struct zone *zone = page_zone(page);
158	struct page *unmovable;
159	unsigned long flags;
160	unsigned long check_unmovable_start, check_unmovable_end;
161
162	if (PageUnaccepted(page))
163		accept_page(page);
164
165	spin_lock_irqsave(&zone->lock, flags);
166
167	/*
168	 * We assume the caller intended to SET migrate type to isolate.
169	 * If it is already set, then someone else must have raced and
170	 * set it before us.
171	 */
172	if (is_migrate_isolate_page(page)) {
173		spin_unlock_irqrestore(&zone->lock, flags);
174		return -EBUSY;
175	}
176
177	/*
178	 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
179	 * We just check MOVABLE pages.
180	 *
181	 * Pass the intersection of [start_pfn, end_pfn) and the page's pageblock
182	 * to avoid redundant checks.
183	 */
184	check_unmovable_start = max(page_to_pfn(page), start_pfn);
185	check_unmovable_end = min(pageblock_end_pfn(page_to_pfn(page)),
186				  end_pfn);
187
188	unmovable = has_unmovable_pages(check_unmovable_start, check_unmovable_end,
189			mode);
190	if (!unmovable) {
191		if (!pageblock_isolate_and_move_free_pages(zone, page)) {
192			spin_unlock_irqrestore(&zone->lock, flags);
193			return -EBUSY;
194		}
195		zone->nr_isolate_pageblock++;
196		spin_unlock_irqrestore(&zone->lock, flags);
197		return 0;
198	}
199
200	spin_unlock_irqrestore(&zone->lock, flags);
201	if (mode == PB_ISOLATE_MODE_MEM_OFFLINE) {
202		/*
203		 * printk() with zone->lock held will likely trigger a
204		 * lockdep splat, so defer it here.
205		 */
206		dump_page(unmovable, "unmovable page");
207	}
208
209	return -EBUSY;
210}
211
212static void unset_migratetype_isolate(struct page *page)
213{
214	struct zone *zone;
215	unsigned long flags;
216	bool isolated_page = false;
217	unsigned int order;
218	struct page *buddy;
219
220	zone = page_zone(page);
221	spin_lock_irqsave(&zone->lock, flags);
222	if (!is_migrate_isolate_page(page))
223		goto out;
224
225	/*
226	 * Because freepage with more than pageblock_order on isolated
227	 * pageblock is restricted to merge due to freepage counting problem,
228	 * it is possible that there is free buddy page.
229	 * move_freepages_block() doesn't care of merge so we need other
230	 * approach in order to merge them. Isolation and free will make
231	 * these pages to be merged.
232	 */
233	if (PageBuddy(page)) {
234		order = buddy_order(page);
235		if (order >= pageblock_order && order < MAX_PAGE_ORDER) {
236			buddy = find_buddy_page_pfn(page, page_to_pfn(page),
237						    order, NULL);
238			if (buddy && !is_migrate_isolate_page(buddy)) {
239				isolated_page = !!__isolate_free_page(page, order);
240				/*
241				 * Isolating a free page in an isolated pageblock
242				 * is expected to always work as watermarks don't
243				 * apply here.
244				 */
245				VM_WARN_ON(!isolated_page);
246			}
247		}
248	}
249
250	/*
251	 * If we isolate freepage with more than pageblock_order, there
252	 * should be no freepage in the range, so we could avoid costly
253	 * pageblock scanning for freepage moving.
254	 *
255	 * We didn't actually touch any of the isolated pages, so place them
256	 * to the tail of the freelist. This is an optimization for memory
257	 * onlining - just onlined memory won't immediately be considered for
258	 * allocation.
259	 */
260	if (!isolated_page) {
261		/*
262		 * Isolating this block already succeeded, so this
263		 * should not fail on zone boundaries.
264		 */
265		WARN_ON_ONCE(!pageblock_unisolate_and_move_free_pages(zone, page));
266	} else {
267		clear_pageblock_isolate(page);
268		__putback_isolated_page(page, order, get_pageblock_migratetype(page));
269	}
270	zone->nr_isolate_pageblock--;
271out:
272	spin_unlock_irqrestore(&zone->lock, flags);
273}
274
275static inline struct page *
276__first_valid_page(unsigned long pfn, unsigned long nr_pages)
277{
278	int i;
279
280	for (i = 0; i < nr_pages; i++) {
281		struct page *page;
282
283		page = pfn_to_online_page(pfn + i);
284		if (!page)
285			continue;
286		return page;
287	}
288	return NULL;
289}
290
291/**
292 * isolate_single_pageblock() -- tries to isolate a pageblock that might be
293 * within a free or in-use page.
294 * @boundary_pfn:		pageblock-aligned pfn that a page might cross
295 * @mode:			isolation mode
296 * @isolate_before:	isolate the pageblock before the boundary_pfn
297 * @skip_isolation:	the flag to skip the pageblock isolation in second
298 *			isolate_single_pageblock()
299 *
300 * Free and in-use pages can be as big as MAX_PAGE_ORDER and contain more than one
301 * pageblock. When not all pageblocks within a page are isolated at the same
302 * time, free page accounting can go wrong. For example, in the case of
303 * MAX_PAGE_ORDER = pageblock_order + 1, a MAX_PAGE_ORDER page has two
304 * pagelbocks.
305 * [      MAX_PAGE_ORDER         ]
306 * [  pageblock0  |  pageblock1  ]
307 * When either pageblock is isolated, if it is a free page, the page is not
308 * split into separate migratetype lists, which is supposed to; if it is an
309 * in-use page and freed later, __free_one_page() does not split the free page
310 * either. The function handles this by splitting the free page or migrating
311 * the in-use page then splitting the free page.
312 */
313static int isolate_single_pageblock(unsigned long boundary_pfn,
314			enum pb_isolate_mode mode, bool isolate_before,
315			bool skip_isolation)
316{
317	unsigned long start_pfn;
318	unsigned long isolate_pageblock;
319	unsigned long pfn;
320	struct zone *zone;
321	int ret;
322
323	VM_BUG_ON(!pageblock_aligned(boundary_pfn));
324
325	if (isolate_before)
326		isolate_pageblock = boundary_pfn - pageblock_nr_pages;
327	else
328		isolate_pageblock = boundary_pfn;
329
330	/*
331	 * scan at the beginning of MAX_ORDER_NR_PAGES aligned range to avoid
332	 * only isolating a subset of pageblocks from a bigger than pageblock
333	 * free or in-use page. Also make sure all to-be-isolated pageblocks
334	 * are within the same zone.
335	 */
336	zone  = page_zone(pfn_to_page(isolate_pageblock));
337	start_pfn  = max(ALIGN_DOWN(isolate_pageblock, MAX_ORDER_NR_PAGES),
338				      zone->zone_start_pfn);
339
340	if (skip_isolation) {
341		VM_BUG_ON(!get_pageblock_isolate(pfn_to_page(isolate_pageblock)));
342	} else {
343		ret = set_migratetype_isolate(pfn_to_page(isolate_pageblock),
344				mode, isolate_pageblock,
345				isolate_pageblock + pageblock_nr_pages);
346
347		if (ret)
348			return ret;
349	}
350
351	/*
352	 * Bail out early when the to-be-isolated pageblock does not form
353	 * a free or in-use page across boundary_pfn:
354	 *
355	 * 1. isolate before boundary_pfn: the page after is not online
356	 * 2. isolate after boundary_pfn: the page before is not online
357	 *
358	 * This also ensures correctness. Without it, when isolate after
359	 * boundary_pfn and [start_pfn, boundary_pfn) are not online,
360	 * __first_valid_page() will return unexpected NULL in the for loop
361	 * below.
362	 */
363	if (isolate_before) {
364		if (!pfn_to_online_page(boundary_pfn))
365			return 0;
366	} else {
367		if (!pfn_to_online_page(boundary_pfn - 1))
368			return 0;
369	}
370
371	for (pfn = start_pfn; pfn < boundary_pfn;) {
372		struct page *page = __first_valid_page(pfn, boundary_pfn - pfn);
373
374		VM_BUG_ON(!page);
375		pfn = page_to_pfn(page);
376
377		if (PageUnaccepted(page)) {
378			pfn += MAX_ORDER_NR_PAGES;
379			continue;
380		}
381
382		if (PageBuddy(page)) {
383			int order = buddy_order(page);
384
385			/* pageblock_isolate_and_move_free_pages() handled this */
386			VM_WARN_ON_ONCE(pfn + (1 << order) > boundary_pfn);
387
388			pfn += 1UL << order;
389			continue;
390		}
391
392		/*
393		 * If a compound page is straddling our block, attempt
394		 * to migrate it out of the way.
395		 *
396		 * We don't have to worry about this creating a large
397		 * free page that straddles into our block: gigantic
398		 * pages are freed as order-0 chunks, and LRU pages
399		 * (currently) do not exceed pageblock_order.
400		 *
401		 * The block of interest has already been marked
402		 * MIGRATE_ISOLATE above, so when migration is done it
403		 * will free its pages onto the correct freelists.
404		 */
405		if (PageCompound(page)) {
406			struct page *head = compound_head(page);
407			unsigned long head_pfn = page_to_pfn(head);
408			unsigned long nr_pages = compound_nr(head);
409
410			if (head_pfn + nr_pages <= boundary_pfn ||
411			    PageHuge(page)) {
412				pfn = head_pfn + nr_pages;
413				continue;
414			}
415
416			/*
417			 * These pages are movable too, but they're
418			 * not expected to exceed pageblock_order.
419			 *
420			 * Let us know when they do, so we can add
421			 * proper free and split handling for them.
422			 */
423			VM_WARN_ON_ONCE_PAGE(PageLRU(page), page);
424			VM_WARN_ON_ONCE_PAGE(page_has_movable_ops(page), page);
425
426			goto failed;
427		}
428
429		pfn++;
430	}
431	return 0;
432failed:
433	/* restore the original migratetype */
434	if (!skip_isolation)
435		unset_migratetype_isolate(pfn_to_page(isolate_pageblock));
436	return -EBUSY;
437}
438
439/**
440 * start_isolate_page_range() - mark page range MIGRATE_ISOLATE
441 * @start_pfn:		The first PFN of the range to be isolated.
442 * @end_pfn:		The last PFN of the range to be isolated.
443 * @mode:		isolation mode
444 *
445 * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
446 * the range will never be allocated. Any free pages and pages freed in the
447 * future will not be allocated again. If specified range includes migrate types
448 * other than MOVABLE or CMA, this will fail with -EBUSY. For isolating all
449 * pages in the range finally, the caller have to free all pages in the range.
450 * test_page_isolated() can be used for test it.
451 *
452 * The function first tries to isolate the pageblocks at the beginning and end
453 * of the range, since there might be pages across the range boundaries.
454 * Afterwards, it isolates the rest of the range.
455 *
456 * There is no high level synchronization mechanism that prevents two threads
457 * from trying to isolate overlapping ranges. If this happens, one thread
458 * will notice pageblocks in the overlapping range already set to isolate.
459 * This happens in set_migratetype_isolate, and set_migratetype_isolate
460 * returns an error. We then clean up by restoring the migration type on
461 * pageblocks we may have modified and return -EBUSY to caller. This
462 * prevents two threads from simultaneously working on overlapping ranges.
463 *
464 * Please note that there is no strong synchronization with the page allocator
465 * either. Pages might be freed while their page blocks are marked ISOLATED.
466 * A call to drain_all_pages() after isolation can flush most of them. However
467 * in some cases pages might still end up on pcp lists and that would allow
468 * for their allocation even when they are in fact isolated already. Depending
469 * on how strong of a guarantee the caller needs, zone_pcp_disable/enable()
470 * might be used to flush and disable pcplist before isolation and enable after
471 * unisolation.
472 *
473 * Return: 0 on success and -EBUSY if any part of range cannot be isolated.
474 */
475int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
476			     enum pb_isolate_mode mode)
477{
478	unsigned long pfn;
479	struct page *page;
480	/* isolation is done at page block granularity */
481	unsigned long isolate_start = pageblock_start_pfn(start_pfn);
482	unsigned long isolate_end = pageblock_align(end_pfn);
483	int ret;
484	bool skip_isolation = false;
485
486	/* isolate [isolate_start, isolate_start + pageblock_nr_pages) pageblock */
487	ret = isolate_single_pageblock(isolate_start, mode, false,
488			skip_isolation);
489	if (ret)
490		return ret;
491
492	if (isolate_start == isolate_end - pageblock_nr_pages)
493		skip_isolation = true;
494
495	/* isolate [isolate_end - pageblock_nr_pages, isolate_end) pageblock */
496	ret = isolate_single_pageblock(isolate_end, mode, true, skip_isolation);
497	if (ret) {
498		unset_migratetype_isolate(pfn_to_page(isolate_start));
499		return ret;
500	}
501
502	/* skip isolated pageblocks at the beginning and end */
503	for (pfn = isolate_start + pageblock_nr_pages;
504	     pfn < isolate_end - pageblock_nr_pages;
505	     pfn += pageblock_nr_pages) {
506		page = __first_valid_page(pfn, pageblock_nr_pages);
507		if (page && set_migratetype_isolate(page, mode, start_pfn,
508					end_pfn)) {
509			undo_isolate_page_range(isolate_start, pfn);
510			unset_migratetype_isolate(
511				pfn_to_page(isolate_end - pageblock_nr_pages));
512			return -EBUSY;
513		}
514	}
515	return 0;
516}
517
518/**
519 * undo_isolate_page_range - undo effects of start_isolate_page_range()
520 * @start_pfn:		The first PFN of the isolated range
521 * @end_pfn:		The last PFN of the isolated range
522 *
523 * This finds and unsets every MIGRATE_ISOLATE page block in the given range
524 */
525void undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn)
526{
527	unsigned long pfn;
528	struct page *page;
529	unsigned long isolate_start = pageblock_start_pfn(start_pfn);
530	unsigned long isolate_end = pageblock_align(end_pfn);
531
532	for (pfn = isolate_start;
533	     pfn < isolate_end;
534	     pfn += pageblock_nr_pages) {
535		page = __first_valid_page(pfn, pageblock_nr_pages);
536		if (!page || !is_migrate_isolate_page(page))
537			continue;
538		unset_migratetype_isolate(page);
539	}
540}
541/*
542 * Test all pages in the range is free(means isolated) or not.
543 * all pages in [start_pfn...end_pfn) must be in the same zone.
544 * zone->lock must be held before call this.
545 *
546 * Returns the last tested pfn.
547 */
548static unsigned long
549__test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
550				  enum pb_isolate_mode mode)
551{
552	struct page *page;
553
554	while (pfn < end_pfn) {
555		page = pfn_to_page(pfn);
556		if (PageBuddy(page))
557			/*
558			 * If the page is on a free list, it has to be on
559			 * the correct MIGRATE_ISOLATE freelist. There is no
560			 * simple way to verify that as VM_BUG_ON(), though.
561			 */
562			pfn += 1 << buddy_order(page);
563		else if ((mode == PB_ISOLATE_MODE_MEM_OFFLINE) &&
564			 PageHWPoison(page))
565			/* A HWPoisoned page cannot be also PageBuddy */
566			pfn++;
567		else if ((mode == PB_ISOLATE_MODE_MEM_OFFLINE) &&
568			 PageOffline(page) && !page_count(page))
569			/*
570			 * The responsible driver agreed to skip PageOffline()
571			 * pages when offlining memory by dropping its
572			 * reference in MEM_GOING_OFFLINE.
573			 */
574			pfn++;
575		else
576			break;
577	}
578
579	return pfn;
580}
581
582/**
583 * test_pages_isolated - check if pageblocks in range are isolated
584 * @start_pfn:		The first PFN of the isolated range
585 * @end_pfn:		The first PFN *after* the isolated range
586 * @mode:		Testing mode
587 *
588 * This tests if all in the specified range are free.
589 *
590 * If %PB_ISOLATE_MODE_MEM_OFFLINE specified in @mode, it will consider
591 * poisoned and offlined pages free as well.
592 *
593 * Caller must ensure the requested range doesn't span zones.
594 *
595 * Returns 0 if true, -EBUSY if one or more pages are in use.
596 */
597int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
598			enum pb_isolate_mode mode)
599{
600	unsigned long pfn, flags;
601	struct page *page;
602	struct zone *zone;
603	int ret;
604
605	/*
606	 * Due to the deferred freeing of hugetlb folios, the hugepage folios may
607	 * not immediately release to the buddy system. This can cause PageBuddy()
608	 * to fail in __test_page_isolated_in_pageblock(). To ensure that the
609	 * hugetlb folios are properly released back to the buddy system, we
610	 * invoke the wait_for_freed_hugetlb_folios() function to wait for the
611	 * release to complete.
612	 */
613	wait_for_freed_hugetlb_folios();
614
615	/*
616	 * Note: pageblock_nr_pages != MAX_PAGE_ORDER. Then, chunks of free
617	 * pages are not aligned to pageblock_nr_pages.
618	 * Then we just check migratetype first.
619	 */
620	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
621		page = __first_valid_page(pfn, pageblock_nr_pages);
622		if (page && !is_migrate_isolate_page(page))
623			break;
624	}
625	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
626	if ((pfn < end_pfn) || !page) {
627		ret = -EBUSY;
628		goto out;
629	}
630
631	/* Check all pages are free or marked as ISOLATED */
632	zone = page_zone(page);
633	spin_lock_irqsave(&zone->lock, flags);
634	pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn, mode);
635	spin_unlock_irqrestore(&zone->lock, flags);
636
637	ret = pfn < end_pfn ? -EBUSY : 0;
638
639out:
640	trace_test_pages_isolated(start_pfn, end_pfn, pfn);
641
642	return ret;
643}