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