]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/compaction.c | |
3 | * | |
4 | * Memory compaction for the reduction of external fragmentation. Note that | |
5 | * this heavily depends upon page migration to do all the real heavy | |
6 | * lifting | |
7 | * | |
8 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
9 | */ | |
10 | #include <linux/cpu.h> | |
11 | #include <linux/swap.h> | |
12 | #include <linux/migrate.h> | |
13 | #include <linux/compaction.h> | |
14 | #include <linux/mm_inline.h> | |
15 | #include <linux/sched/signal.h> | |
16 | #include <linux/backing-dev.h> | |
17 | #include <linux/sysctl.h> | |
18 | #include <linux/sysfs.h> | |
19 | #include <linux/page-isolation.h> | |
20 | #include <linux/kasan.h> | |
21 | #include <linux/kthread.h> | |
22 | #include <linux/freezer.h> | |
23 | #include <linux/page_owner.h> | |
24 | #include "internal.h" | |
25 | ||
26 | #ifdef CONFIG_COMPACTION | |
27 | static inline void count_compact_event(enum vm_event_item item) | |
28 | { | |
29 | count_vm_event(item); | |
30 | } | |
31 | ||
32 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
33 | { | |
34 | count_vm_events(item, delta); | |
35 | } | |
36 | #else | |
37 | #define count_compact_event(item) do { } while (0) | |
38 | #define count_compact_events(item, delta) do { } while (0) | |
39 | #endif | |
40 | ||
41 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA | |
42 | ||
43 | #define CREATE_TRACE_POINTS | |
44 | #include <trace/events/compaction.h> | |
45 | ||
46 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) | |
47 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
48 | #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order) | |
49 | #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order) | |
50 | ||
51 | static unsigned long release_freepages(struct list_head *freelist) | |
52 | { | |
53 | struct page *page, *next; | |
54 | unsigned long high_pfn = 0; | |
55 | ||
56 | list_for_each_entry_safe(page, next, freelist, lru) { | |
57 | unsigned long pfn = page_to_pfn(page); | |
58 | list_del(&page->lru); | |
59 | __free_page(page); | |
60 | if (pfn > high_pfn) | |
61 | high_pfn = pfn; | |
62 | } | |
63 | ||
64 | return high_pfn; | |
65 | } | |
66 | ||
67 | static void map_pages(struct list_head *list) | |
68 | { | |
69 | unsigned int i, order, nr_pages; | |
70 | struct page *page, *next; | |
71 | LIST_HEAD(tmp_list); | |
72 | ||
73 | list_for_each_entry_safe(page, next, list, lru) { | |
74 | list_del(&page->lru); | |
75 | ||
76 | order = page_private(page); | |
77 | nr_pages = 1 << order; | |
78 | ||
79 | post_alloc_hook(page, order, __GFP_MOVABLE); | |
80 | if (order) | |
81 | split_page(page, order); | |
82 | ||
83 | for (i = 0; i < nr_pages; i++) { | |
84 | list_add(&page->lru, &tmp_list); | |
85 | page++; | |
86 | } | |
87 | } | |
88 | ||
89 | list_splice(&tmp_list, list); | |
90 | } | |
91 | ||
92 | #ifdef CONFIG_COMPACTION | |
93 | ||
94 | int PageMovable(struct page *page) | |
95 | { | |
96 | struct address_space *mapping; | |
97 | ||
98 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
99 | if (!__PageMovable(page)) | |
100 | return 0; | |
101 | ||
102 | mapping = page_mapping(page); | |
103 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
104 | return 1; | |
105 | ||
106 | return 0; | |
107 | } | |
108 | EXPORT_SYMBOL(PageMovable); | |
109 | ||
110 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
111 | { | |
112 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
113 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
114 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
115 | } | |
116 | EXPORT_SYMBOL(__SetPageMovable); | |
117 | ||
118 | void __ClearPageMovable(struct page *page) | |
119 | { | |
120 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
121 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
122 | /* | |
123 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
124 | * flag so that VM can catch up released page by driver after isolation. | |
125 | * With it, VM migration doesn't try to put it back. | |
126 | */ | |
127 | page->mapping = (void *)((unsigned long)page->mapping & | |
128 | PAGE_MAPPING_MOVABLE); | |
129 | } | |
130 | EXPORT_SYMBOL(__ClearPageMovable); | |
131 | ||
132 | /* Do not skip compaction more than 64 times */ | |
133 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
134 | ||
135 | /* | |
136 | * Compaction is deferred when compaction fails to result in a page | |
137 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
138 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
139 | */ | |
140 | void defer_compaction(struct zone *zone, int order) | |
141 | { | |
142 | zone->compact_considered = 0; | |
143 | zone->compact_defer_shift++; | |
144 | ||
145 | if (order < zone->compact_order_failed) | |
146 | zone->compact_order_failed = order; | |
147 | ||
148 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
149 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
150 | ||
151 | trace_mm_compaction_defer_compaction(zone, order); | |
152 | } | |
153 | ||
154 | /* Returns true if compaction should be skipped this time */ | |
155 | bool compaction_deferred(struct zone *zone, int order) | |
156 | { | |
157 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
158 | ||
159 | if (order < zone->compact_order_failed) | |
160 | return false; | |
161 | ||
162 | /* Avoid possible overflow */ | |
163 | if (++zone->compact_considered > defer_limit) | |
164 | zone->compact_considered = defer_limit; | |
165 | ||
166 | if (zone->compact_considered >= defer_limit) | |
167 | return false; | |
168 | ||
169 | trace_mm_compaction_deferred(zone, order); | |
170 | ||
171 | return true; | |
172 | } | |
173 | ||
174 | /* | |
175 | * Update defer tracking counters after successful compaction of given order, | |
176 | * which means an allocation either succeeded (alloc_success == true) or is | |
177 | * expected to succeed. | |
178 | */ | |
179 | void compaction_defer_reset(struct zone *zone, int order, | |
180 | bool alloc_success) | |
181 | { | |
182 | if (alloc_success) { | |
183 | zone->compact_considered = 0; | |
184 | zone->compact_defer_shift = 0; | |
185 | } | |
186 | if (order >= zone->compact_order_failed) | |
187 | zone->compact_order_failed = order + 1; | |
188 | ||
189 | trace_mm_compaction_defer_reset(zone, order); | |
190 | } | |
191 | ||
192 | /* Returns true if restarting compaction after many failures */ | |
193 | bool compaction_restarting(struct zone *zone, int order) | |
194 | { | |
195 | if (order < zone->compact_order_failed) | |
196 | return false; | |
197 | ||
198 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
199 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
200 | } | |
201 | ||
202 | /* Returns true if the pageblock should be scanned for pages to isolate. */ | |
203 | static inline bool isolation_suitable(struct compact_control *cc, | |
204 | struct page *page) | |
205 | { | |
206 | if (cc->ignore_skip_hint) | |
207 | return true; | |
208 | ||
209 | return !get_pageblock_skip(page); | |
210 | } | |
211 | ||
212 | static void reset_cached_positions(struct zone *zone) | |
213 | { | |
214 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
215 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
216 | zone->compact_cached_free_pfn = | |
217 | pageblock_start_pfn(zone_end_pfn(zone) - 1); | |
218 | } | |
219 | ||
220 | /* | |
221 | * This function is called to clear all cached information on pageblocks that | |
222 | * should be skipped for page isolation when the migrate and free page scanner | |
223 | * meet. | |
224 | */ | |
225 | static void __reset_isolation_suitable(struct zone *zone) | |
226 | { | |
227 | unsigned long start_pfn = zone->zone_start_pfn; | |
228 | unsigned long end_pfn = zone_end_pfn(zone); | |
229 | unsigned long pfn; | |
230 | ||
231 | zone->compact_blockskip_flush = false; | |
232 | ||
233 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
234 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
235 | struct page *page; | |
236 | ||
237 | cond_resched(); | |
238 | ||
239 | page = pfn_to_online_page(pfn); | |
240 | if (!page) | |
241 | continue; | |
242 | if (zone != page_zone(page)) | |
243 | continue; | |
244 | ||
245 | clear_pageblock_skip(page); | |
246 | } | |
247 | ||
248 | reset_cached_positions(zone); | |
249 | } | |
250 | ||
251 | void reset_isolation_suitable(pg_data_t *pgdat) | |
252 | { | |
253 | int zoneid; | |
254 | ||
255 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
256 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
257 | if (!populated_zone(zone)) | |
258 | continue; | |
259 | ||
260 | /* Only flush if a full compaction finished recently */ | |
261 | if (zone->compact_blockskip_flush) | |
262 | __reset_isolation_suitable(zone); | |
263 | } | |
264 | } | |
265 | ||
266 | /* | |
267 | * If no pages were isolated then mark this pageblock to be skipped in the | |
268 | * future. The information is later cleared by __reset_isolation_suitable(). | |
269 | */ | |
270 | static void update_pageblock_skip(struct compact_control *cc, | |
271 | struct page *page, unsigned long nr_isolated, | |
272 | bool migrate_scanner) | |
273 | { | |
274 | struct zone *zone = cc->zone; | |
275 | unsigned long pfn; | |
276 | ||
277 | if (cc->ignore_skip_hint) | |
278 | return; | |
279 | ||
280 | if (!page) | |
281 | return; | |
282 | ||
283 | if (nr_isolated) | |
284 | return; | |
285 | ||
286 | set_pageblock_skip(page); | |
287 | ||
288 | pfn = page_to_pfn(page); | |
289 | ||
290 | /* Update where async and sync compaction should restart */ | |
291 | if (migrate_scanner) { | |
292 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
293 | zone->compact_cached_migrate_pfn[0] = pfn; | |
294 | if (cc->mode != MIGRATE_ASYNC && | |
295 | pfn > zone->compact_cached_migrate_pfn[1]) | |
296 | zone->compact_cached_migrate_pfn[1] = pfn; | |
297 | } else { | |
298 | if (pfn < zone->compact_cached_free_pfn) | |
299 | zone->compact_cached_free_pfn = pfn; | |
300 | } | |
301 | } | |
302 | #else | |
303 | static inline bool isolation_suitable(struct compact_control *cc, | |
304 | struct page *page) | |
305 | { | |
306 | return true; | |
307 | } | |
308 | ||
309 | static void update_pageblock_skip(struct compact_control *cc, | |
310 | struct page *page, unsigned long nr_isolated, | |
311 | bool migrate_scanner) | |
312 | { | |
313 | } | |
314 | #endif /* CONFIG_COMPACTION */ | |
315 | ||
316 | /* | |
317 | * Compaction requires the taking of some coarse locks that are potentially | |
318 | * very heavily contended. For async compaction, back out if the lock cannot | |
319 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
320 | * | |
321 | * Returns true if the lock is held | |
322 | * Returns false if the lock is not held and compaction should abort | |
323 | */ | |
324 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
325 | struct compact_control *cc) | |
326 | { | |
327 | if (cc->mode == MIGRATE_ASYNC) { | |
328 | if (!spin_trylock_irqsave(lock, *flags)) { | |
329 | cc->contended = true; | |
330 | return false; | |
331 | } | |
332 | } else { | |
333 | spin_lock_irqsave(lock, *flags); | |
334 | } | |
335 | ||
336 | return true; | |
337 | } | |
338 | ||
339 | /* | |
340 | * Compaction requires the taking of some coarse locks that are potentially | |
341 | * very heavily contended. The lock should be periodically unlocked to avoid | |
342 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
343 | * the lock. It might also be that allowing the IRQs will result in | |
344 | * need_resched() becoming true. If scheduling is needed, async compaction | |
345 | * aborts. Sync compaction schedules. | |
346 | * Either compaction type will also abort if a fatal signal is pending. | |
347 | * In either case if the lock was locked, it is dropped and not regained. | |
348 | * | |
349 | * Returns true if compaction should abort due to fatal signal pending, or | |
350 | * async compaction due to need_resched() | |
351 | * Returns false when compaction can continue (sync compaction might have | |
352 | * scheduled) | |
353 | */ | |
354 | static bool compact_unlock_should_abort(spinlock_t *lock, | |
355 | unsigned long flags, bool *locked, struct compact_control *cc) | |
356 | { | |
357 | if (*locked) { | |
358 | spin_unlock_irqrestore(lock, flags); | |
359 | *locked = false; | |
360 | } | |
361 | ||
362 | if (fatal_signal_pending(current)) { | |
363 | cc->contended = true; | |
364 | return true; | |
365 | } | |
366 | ||
367 | if (need_resched()) { | |
368 | if (cc->mode == MIGRATE_ASYNC) { | |
369 | cc->contended = true; | |
370 | return true; | |
371 | } | |
372 | cond_resched(); | |
373 | } | |
374 | ||
375 | return false; | |
376 | } | |
377 | ||
378 | /* | |
379 | * Aside from avoiding lock contention, compaction also periodically checks | |
380 | * need_resched() and either schedules in sync compaction or aborts async | |
381 | * compaction. This is similar to what compact_unlock_should_abort() does, but | |
382 | * is used where no lock is concerned. | |
383 | * | |
384 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
385 | * Returns true when async compaction should abort. | |
386 | */ | |
387 | static inline bool compact_should_abort(struct compact_control *cc) | |
388 | { | |
389 | /* async compaction aborts if contended */ | |
390 | if (need_resched()) { | |
391 | if (cc->mode == MIGRATE_ASYNC) { | |
392 | cc->contended = true; | |
393 | return true; | |
394 | } | |
395 | ||
396 | cond_resched(); | |
397 | } | |
398 | ||
399 | return false; | |
400 | } | |
401 | ||
402 | /* | |
403 | * Isolate free pages onto a private freelist. If @strict is true, will abort | |
404 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
405 | * (even though it may still end up isolating some pages). | |
406 | */ | |
407 | static unsigned long isolate_freepages_block(struct compact_control *cc, | |
408 | unsigned long *start_pfn, | |
409 | unsigned long end_pfn, | |
410 | struct list_head *freelist, | |
411 | bool strict) | |
412 | { | |
413 | int nr_scanned = 0, total_isolated = 0; | |
414 | struct page *cursor, *valid_page = NULL; | |
415 | unsigned long flags = 0; | |
416 | bool locked = false; | |
417 | unsigned long blockpfn = *start_pfn; | |
418 | unsigned int order; | |
419 | ||
420 | cursor = pfn_to_page(blockpfn); | |
421 | ||
422 | /* Isolate free pages. */ | |
423 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { | |
424 | int isolated; | |
425 | struct page *page = cursor; | |
426 | ||
427 | /* | |
428 | * Periodically drop the lock (if held) regardless of its | |
429 | * contention, to give chance to IRQs. Abort if fatal signal | |
430 | * pending or async compaction detects need_resched() | |
431 | */ | |
432 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
433 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
434 | &locked, cc)) | |
435 | break; | |
436 | ||
437 | nr_scanned++; | |
438 | if (!pfn_valid_within(blockpfn)) | |
439 | goto isolate_fail; | |
440 | ||
441 | if (!valid_page) | |
442 | valid_page = page; | |
443 | ||
444 | /* | |
445 | * For compound pages such as THP and hugetlbfs, we can save | |
446 | * potentially a lot of iterations if we skip them at once. | |
447 | * The check is racy, but we can consider only valid values | |
448 | * and the only danger is skipping too much. | |
449 | */ | |
450 | if (PageCompound(page)) { | |
451 | unsigned int comp_order = compound_order(page); | |
452 | ||
453 | if (likely(comp_order < MAX_ORDER)) { | |
454 | blockpfn += (1UL << comp_order) - 1; | |
455 | cursor += (1UL << comp_order) - 1; | |
456 | } | |
457 | ||
458 | goto isolate_fail; | |
459 | } | |
460 | ||
461 | if (!PageBuddy(page)) | |
462 | goto isolate_fail; | |
463 | ||
464 | /* | |
465 | * If we already hold the lock, we can skip some rechecking. | |
466 | * Note that if we hold the lock now, checked_pageblock was | |
467 | * already set in some previous iteration (or strict is true), | |
468 | * so it is correct to skip the suitable migration target | |
469 | * recheck as well. | |
470 | */ | |
471 | if (!locked) { | |
472 | /* | |
473 | * The zone lock must be held to isolate freepages. | |
474 | * Unfortunately this is a very coarse lock and can be | |
475 | * heavily contended if there are parallel allocations | |
476 | * or parallel compactions. For async compaction do not | |
477 | * spin on the lock and we acquire the lock as late as | |
478 | * possible. | |
479 | */ | |
480 | locked = compact_trylock_irqsave(&cc->zone->lock, | |
481 | &flags, cc); | |
482 | if (!locked) | |
483 | break; | |
484 | ||
485 | /* Recheck this is a buddy page under lock */ | |
486 | if (!PageBuddy(page)) | |
487 | goto isolate_fail; | |
488 | } | |
489 | ||
490 | /* Found a free page, will break it into order-0 pages */ | |
491 | order = page_order(page); | |
492 | isolated = __isolate_free_page(page, order); | |
493 | if (!isolated) | |
494 | break; | |
495 | set_page_private(page, order); | |
496 | ||
497 | total_isolated += isolated; | |
498 | cc->nr_freepages += isolated; | |
499 | list_add_tail(&page->lru, freelist); | |
500 | ||
501 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { | |
502 | blockpfn += isolated; | |
503 | break; | |
504 | } | |
505 | /* Advance to the end of split page */ | |
506 | blockpfn += isolated - 1; | |
507 | cursor += isolated - 1; | |
508 | continue; | |
509 | ||
510 | isolate_fail: | |
511 | if (strict) | |
512 | break; | |
513 | else | |
514 | continue; | |
515 | ||
516 | } | |
517 | ||
518 | if (locked) | |
519 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
520 | ||
521 | /* | |
522 | * There is a tiny chance that we have read bogus compound_order(), | |
523 | * so be careful to not go outside of the pageblock. | |
524 | */ | |
525 | if (unlikely(blockpfn > end_pfn)) | |
526 | blockpfn = end_pfn; | |
527 | ||
528 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, | |
529 | nr_scanned, total_isolated); | |
530 | ||
531 | /* Record how far we have got within the block */ | |
532 | *start_pfn = blockpfn; | |
533 | ||
534 | /* | |
535 | * If strict isolation is requested by CMA then check that all the | |
536 | * pages requested were isolated. If there were any failures, 0 is | |
537 | * returned and CMA will fail. | |
538 | */ | |
539 | if (strict && blockpfn < end_pfn) | |
540 | total_isolated = 0; | |
541 | ||
542 | /* Update the pageblock-skip if the whole pageblock was scanned */ | |
543 | if (blockpfn == end_pfn) | |
544 | update_pageblock_skip(cc, valid_page, total_isolated, false); | |
545 | ||
546 | cc->total_free_scanned += nr_scanned; | |
547 | if (total_isolated) | |
548 | count_compact_events(COMPACTISOLATED, total_isolated); | |
549 | return total_isolated; | |
550 | } | |
551 | ||
552 | /** | |
553 | * isolate_freepages_range() - isolate free pages. | |
554 | * @start_pfn: The first PFN to start isolating. | |
555 | * @end_pfn: The one-past-last PFN. | |
556 | * | |
557 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
558 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
559 | * undo its actions and return zero. | |
560 | * | |
561 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
562 | * (which may be greater then end_pfn if end fell in a middle of | |
563 | * a free page). | |
564 | */ | |
565 | unsigned long | |
566 | isolate_freepages_range(struct compact_control *cc, | |
567 | unsigned long start_pfn, unsigned long end_pfn) | |
568 | { | |
569 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; | |
570 | LIST_HEAD(freelist); | |
571 | ||
572 | pfn = start_pfn; | |
573 | block_start_pfn = pageblock_start_pfn(pfn); | |
574 | if (block_start_pfn < cc->zone->zone_start_pfn) | |
575 | block_start_pfn = cc->zone->zone_start_pfn; | |
576 | block_end_pfn = pageblock_end_pfn(pfn); | |
577 | ||
578 | for (; pfn < end_pfn; pfn += isolated, | |
579 | block_start_pfn = block_end_pfn, | |
580 | block_end_pfn += pageblock_nr_pages) { | |
581 | /* Protect pfn from changing by isolate_freepages_block */ | |
582 | unsigned long isolate_start_pfn = pfn; | |
583 | ||
584 | block_end_pfn = min(block_end_pfn, end_pfn); | |
585 | ||
586 | /* | |
587 | * pfn could pass the block_end_pfn if isolated freepage | |
588 | * is more than pageblock order. In this case, we adjust | |
589 | * scanning range to right one. | |
590 | */ | |
591 | if (pfn >= block_end_pfn) { | |
592 | block_start_pfn = pageblock_start_pfn(pfn); | |
593 | block_end_pfn = pageblock_end_pfn(pfn); | |
594 | block_end_pfn = min(block_end_pfn, end_pfn); | |
595 | } | |
596 | ||
597 | if (!pageblock_pfn_to_page(block_start_pfn, | |
598 | block_end_pfn, cc->zone)) | |
599 | break; | |
600 | ||
601 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, | |
602 | block_end_pfn, &freelist, true); | |
603 | ||
604 | /* | |
605 | * In strict mode, isolate_freepages_block() returns 0 if | |
606 | * there are any holes in the block (ie. invalid PFNs or | |
607 | * non-free pages). | |
608 | */ | |
609 | if (!isolated) | |
610 | break; | |
611 | ||
612 | /* | |
613 | * If we managed to isolate pages, it is always (1 << n) * | |
614 | * pageblock_nr_pages for some non-negative n. (Max order | |
615 | * page may span two pageblocks). | |
616 | */ | |
617 | } | |
618 | ||
619 | /* __isolate_free_page() does not map the pages */ | |
620 | map_pages(&freelist); | |
621 | ||
622 | if (pfn < end_pfn) { | |
623 | /* Loop terminated early, cleanup. */ | |
624 | release_freepages(&freelist); | |
625 | return 0; | |
626 | } | |
627 | ||
628 | /* We don't use freelists for anything. */ | |
629 | return pfn; | |
630 | } | |
631 | ||
632 | /* Similar to reclaim, but different enough that they don't share logic */ | |
633 | static bool too_many_isolated(struct zone *zone) | |
634 | { | |
635 | unsigned long active, inactive, isolated; | |
636 | ||
637 | inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) + | |
638 | node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON); | |
639 | active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) + | |
640 | node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON); | |
641 | isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) + | |
642 | node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON); | |
643 | ||
644 | return isolated > (inactive + active) / 2; | |
645 | } | |
646 | ||
647 | /** | |
648 | * isolate_migratepages_block() - isolate all migrate-able pages within | |
649 | * a single pageblock | |
650 | * @cc: Compaction control structure. | |
651 | * @low_pfn: The first PFN to isolate | |
652 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
653 | * @isolate_mode: Isolation mode to be used. | |
654 | * | |
655 | * Isolate all pages that can be migrated from the range specified by | |
656 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. | |
657 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
658 | * first page that was not scanned (which may be both less, equal to or more | |
659 | * than end_pfn). | |
660 | * | |
661 | * The pages are isolated on cc->migratepages list (not required to be empty), | |
662 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
663 | * is neither read nor updated. | |
664 | */ | |
665 | static unsigned long | |
666 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
667 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
668 | { | |
669 | struct zone *zone = cc->zone; | |
670 | unsigned long nr_scanned = 0, nr_isolated = 0; | |
671 | struct lruvec *lruvec; | |
672 | unsigned long flags = 0; | |
673 | bool locked = false; | |
674 | struct page *page = NULL, *valid_page = NULL; | |
675 | unsigned long start_pfn = low_pfn; | |
676 | bool skip_on_failure = false; | |
677 | unsigned long next_skip_pfn = 0; | |
678 | ||
679 | /* | |
680 | * Ensure that there are not too many pages isolated from the LRU | |
681 | * list by either parallel reclaimers or compaction. If there are, | |
682 | * delay for some time until fewer pages are isolated | |
683 | */ | |
684 | while (unlikely(too_many_isolated(zone))) { | |
685 | /* async migration should just abort */ | |
686 | if (cc->mode == MIGRATE_ASYNC) | |
687 | return 0; | |
688 | ||
689 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
690 | ||
691 | if (fatal_signal_pending(current)) | |
692 | return 0; | |
693 | } | |
694 | ||
695 | if (compact_should_abort(cc)) | |
696 | return 0; | |
697 | ||
698 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { | |
699 | skip_on_failure = true; | |
700 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
701 | } | |
702 | ||
703 | /* Time to isolate some pages for migration */ | |
704 | for (; low_pfn < end_pfn; low_pfn++) { | |
705 | ||
706 | if (skip_on_failure && low_pfn >= next_skip_pfn) { | |
707 | /* | |
708 | * We have isolated all migration candidates in the | |
709 | * previous order-aligned block, and did not skip it due | |
710 | * to failure. We should migrate the pages now and | |
711 | * hopefully succeed compaction. | |
712 | */ | |
713 | if (nr_isolated) | |
714 | break; | |
715 | ||
716 | /* | |
717 | * We failed to isolate in the previous order-aligned | |
718 | * block. Set the new boundary to the end of the | |
719 | * current block. Note we can't simply increase | |
720 | * next_skip_pfn by 1 << order, as low_pfn might have | |
721 | * been incremented by a higher number due to skipping | |
722 | * a compound or a high-order buddy page in the | |
723 | * previous loop iteration. | |
724 | */ | |
725 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
726 | } | |
727 | ||
728 | /* | |
729 | * Periodically drop the lock (if held) regardless of its | |
730 | * contention, to give chance to IRQs. Abort async compaction | |
731 | * if contended. | |
732 | */ | |
733 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
734 | && compact_unlock_should_abort(zone_lru_lock(zone), flags, | |
735 | &locked, cc)) | |
736 | break; | |
737 | ||
738 | if (!pfn_valid_within(low_pfn)) | |
739 | goto isolate_fail; | |
740 | nr_scanned++; | |
741 | ||
742 | page = pfn_to_page(low_pfn); | |
743 | ||
744 | if (!valid_page) | |
745 | valid_page = page; | |
746 | ||
747 | /* | |
748 | * Skip if free. We read page order here without zone lock | |
749 | * which is generally unsafe, but the race window is small and | |
750 | * the worst thing that can happen is that we skip some | |
751 | * potential isolation targets. | |
752 | */ | |
753 | if (PageBuddy(page)) { | |
754 | unsigned long freepage_order = page_order_unsafe(page); | |
755 | ||
756 | /* | |
757 | * Without lock, we cannot be sure that what we got is | |
758 | * a valid page order. Consider only values in the | |
759 | * valid order range to prevent low_pfn overflow. | |
760 | */ | |
761 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
762 | low_pfn += (1UL << freepage_order) - 1; | |
763 | continue; | |
764 | } | |
765 | ||
766 | /* | |
767 | * Regardless of being on LRU, compound pages such as THP and | |
768 | * hugetlbfs are not to be compacted. We can potentially save | |
769 | * a lot of iterations if we skip them at once. The check is | |
770 | * racy, but we can consider only valid values and the only | |
771 | * danger is skipping too much. | |
772 | */ | |
773 | if (PageCompound(page)) { | |
774 | unsigned int comp_order = compound_order(page); | |
775 | ||
776 | if (likely(comp_order < MAX_ORDER)) | |
777 | low_pfn += (1UL << comp_order) - 1; | |
778 | ||
779 | goto isolate_fail; | |
780 | } | |
781 | ||
782 | /* | |
783 | * Check may be lockless but that's ok as we recheck later. | |
784 | * It's possible to migrate LRU and non-lru movable pages. | |
785 | * Skip any other type of page | |
786 | */ | |
787 | if (!PageLRU(page)) { | |
788 | /* | |
789 | * __PageMovable can return false positive so we need | |
790 | * to verify it under page_lock. | |
791 | */ | |
792 | if (unlikely(__PageMovable(page)) && | |
793 | !PageIsolated(page)) { | |
794 | if (locked) { | |
795 | spin_unlock_irqrestore(zone_lru_lock(zone), | |
796 | flags); | |
797 | locked = false; | |
798 | } | |
799 | ||
800 | if (!isolate_movable_page(page, isolate_mode)) | |
801 | goto isolate_success; | |
802 | } | |
803 | ||
804 | goto isolate_fail; | |
805 | } | |
806 | ||
807 | /* | |
808 | * Migration will fail if an anonymous page is pinned in memory, | |
809 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
810 | * admittedly racy check. | |
811 | */ | |
812 | if (!page_mapping(page) && | |
813 | page_count(page) > page_mapcount(page)) | |
814 | goto isolate_fail; | |
815 | ||
816 | /* | |
817 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
818 | * because those do not depend on fs locks. | |
819 | */ | |
820 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
821 | goto isolate_fail; | |
822 | ||
823 | /* If we already hold the lock, we can skip some rechecking */ | |
824 | if (!locked) { | |
825 | locked = compact_trylock_irqsave(zone_lru_lock(zone), | |
826 | &flags, cc); | |
827 | if (!locked) | |
828 | break; | |
829 | ||
830 | /* Recheck PageLRU and PageCompound under lock */ | |
831 | if (!PageLRU(page)) | |
832 | goto isolate_fail; | |
833 | ||
834 | /* | |
835 | * Page become compound since the non-locked check, | |
836 | * and it's on LRU. It can only be a THP so the order | |
837 | * is safe to read and it's 0 for tail pages. | |
838 | */ | |
839 | if (unlikely(PageCompound(page))) { | |
840 | low_pfn += (1UL << compound_order(page)) - 1; | |
841 | goto isolate_fail; | |
842 | } | |
843 | } | |
844 | ||
845 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); | |
846 | ||
847 | /* Try isolate the page */ | |
848 | if (__isolate_lru_page(page, isolate_mode) != 0) | |
849 | goto isolate_fail; | |
850 | ||
851 | VM_BUG_ON_PAGE(PageCompound(page), page); | |
852 | ||
853 | /* Successfully isolated */ | |
854 | del_page_from_lru_list(page, lruvec, page_lru(page)); | |
855 | inc_node_page_state(page, | |
856 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
857 | ||
858 | isolate_success: | |
859 | list_add(&page->lru, &cc->migratepages); | |
860 | cc->nr_migratepages++; | |
861 | nr_isolated++; | |
862 | ||
863 | /* | |
864 | * Record where we could have freed pages by migration and not | |
865 | * yet flushed them to buddy allocator. | |
866 | * - this is the lowest page that was isolated and likely be | |
867 | * then freed by migration. | |
868 | */ | |
869 | if (!cc->last_migrated_pfn) | |
870 | cc->last_migrated_pfn = low_pfn; | |
871 | ||
872 | /* Avoid isolating too much */ | |
873 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { | |
874 | ++low_pfn; | |
875 | break; | |
876 | } | |
877 | ||
878 | continue; | |
879 | isolate_fail: | |
880 | if (!skip_on_failure) | |
881 | continue; | |
882 | ||
883 | /* | |
884 | * We have isolated some pages, but then failed. Release them | |
885 | * instead of migrating, as we cannot form the cc->order buddy | |
886 | * page anyway. | |
887 | */ | |
888 | if (nr_isolated) { | |
889 | if (locked) { | |
890 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); | |
891 | locked = false; | |
892 | } | |
893 | putback_movable_pages(&cc->migratepages); | |
894 | cc->nr_migratepages = 0; | |
895 | cc->last_migrated_pfn = 0; | |
896 | nr_isolated = 0; | |
897 | } | |
898 | ||
899 | if (low_pfn < next_skip_pfn) { | |
900 | low_pfn = next_skip_pfn - 1; | |
901 | /* | |
902 | * The check near the loop beginning would have updated | |
903 | * next_skip_pfn too, but this is a bit simpler. | |
904 | */ | |
905 | next_skip_pfn += 1UL << cc->order; | |
906 | } | |
907 | } | |
908 | ||
909 | /* | |
910 | * The PageBuddy() check could have potentially brought us outside | |
911 | * the range to be scanned. | |
912 | */ | |
913 | if (unlikely(low_pfn > end_pfn)) | |
914 | low_pfn = end_pfn; | |
915 | ||
916 | if (locked) | |
917 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); | |
918 | ||
919 | /* | |
920 | * Update the pageblock-skip information and cached scanner pfn, | |
921 | * if the whole pageblock was scanned without isolating any page. | |
922 | */ | |
923 | if (low_pfn == end_pfn) | |
924 | update_pageblock_skip(cc, valid_page, nr_isolated, true); | |
925 | ||
926 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, | |
927 | nr_scanned, nr_isolated); | |
928 | ||
929 | cc->total_migrate_scanned += nr_scanned; | |
930 | if (nr_isolated) | |
931 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
932 | ||
933 | return low_pfn; | |
934 | } | |
935 | ||
936 | /** | |
937 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
938 | * @cc: Compaction control structure. | |
939 | * @start_pfn: The first PFN to start isolating. | |
940 | * @end_pfn: The one-past-last PFN. | |
941 | * | |
942 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
943 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
944 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
945 | */ | |
946 | unsigned long | |
947 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
948 | unsigned long end_pfn) | |
949 | { | |
950 | unsigned long pfn, block_start_pfn, block_end_pfn; | |
951 | ||
952 | /* Scan block by block. First and last block may be incomplete */ | |
953 | pfn = start_pfn; | |
954 | block_start_pfn = pageblock_start_pfn(pfn); | |
955 | if (block_start_pfn < cc->zone->zone_start_pfn) | |
956 | block_start_pfn = cc->zone->zone_start_pfn; | |
957 | block_end_pfn = pageblock_end_pfn(pfn); | |
958 | ||
959 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
960 | block_start_pfn = block_end_pfn, | |
961 | block_end_pfn += pageblock_nr_pages) { | |
962 | ||
963 | block_end_pfn = min(block_end_pfn, end_pfn); | |
964 | ||
965 | if (!pageblock_pfn_to_page(block_start_pfn, | |
966 | block_end_pfn, cc->zone)) | |
967 | continue; | |
968 | ||
969 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
970 | ISOLATE_UNEVICTABLE); | |
971 | ||
972 | if (!pfn) | |
973 | break; | |
974 | ||
975 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
976 | break; | |
977 | } | |
978 | ||
979 | return pfn; | |
980 | } | |
981 | ||
982 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ | |
983 | #ifdef CONFIG_COMPACTION | |
984 | ||
985 | static bool suitable_migration_source(struct compact_control *cc, | |
986 | struct page *page) | |
987 | { | |
988 | int block_mt; | |
989 | ||
990 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) | |
991 | return true; | |
992 | ||
993 | block_mt = get_pageblock_migratetype(page); | |
994 | ||
995 | if (cc->migratetype == MIGRATE_MOVABLE) | |
996 | return is_migrate_movable(block_mt); | |
997 | else | |
998 | return block_mt == cc->migratetype; | |
999 | } | |
1000 | ||
1001 | /* Returns true if the page is within a block suitable for migration to */ | |
1002 | static bool suitable_migration_target(struct compact_control *cc, | |
1003 | struct page *page) | |
1004 | { | |
1005 | /* If the page is a large free page, then disallow migration */ | |
1006 | if (PageBuddy(page)) { | |
1007 | /* | |
1008 | * We are checking page_order without zone->lock taken. But | |
1009 | * the only small danger is that we skip a potentially suitable | |
1010 | * pageblock, so it's not worth to check order for valid range. | |
1011 | */ | |
1012 | if (page_order_unsafe(page) >= pageblock_order) | |
1013 | return false; | |
1014 | } | |
1015 | ||
1016 | if (cc->ignore_block_suitable) | |
1017 | return true; | |
1018 | ||
1019 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ | |
1020 | if (is_migrate_movable(get_pageblock_migratetype(page))) | |
1021 | return true; | |
1022 | ||
1023 | /* Otherwise skip the block */ | |
1024 | return false; | |
1025 | } | |
1026 | ||
1027 | /* | |
1028 | * Test whether the free scanner has reached the same or lower pageblock than | |
1029 | * the migration scanner, and compaction should thus terminate. | |
1030 | */ | |
1031 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1032 | { | |
1033 | return (cc->free_pfn >> pageblock_order) | |
1034 | <= (cc->migrate_pfn >> pageblock_order); | |
1035 | } | |
1036 | ||
1037 | /* | |
1038 | * Based on information in the current compact_control, find blocks | |
1039 | * suitable for isolating free pages from and then isolate them. | |
1040 | */ | |
1041 | static void isolate_freepages(struct compact_control *cc) | |
1042 | { | |
1043 | struct zone *zone = cc->zone; | |
1044 | struct page *page; | |
1045 | unsigned long block_start_pfn; /* start of current pageblock */ | |
1046 | unsigned long isolate_start_pfn; /* exact pfn we start at */ | |
1047 | unsigned long block_end_pfn; /* end of current pageblock */ | |
1048 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
1049 | struct list_head *freelist = &cc->freepages; | |
1050 | ||
1051 | /* | |
1052 | * Initialise the free scanner. The starting point is where we last | |
1053 | * successfully isolated from, zone-cached value, or the end of the | |
1054 | * zone when isolating for the first time. For looping we also need | |
1055 | * this pfn aligned down to the pageblock boundary, because we do | |
1056 | * block_start_pfn -= pageblock_nr_pages in the for loop. | |
1057 | * For ending point, take care when isolating in last pageblock of a | |
1058 | * a zone which ends in the middle of a pageblock. | |
1059 | * The low boundary is the end of the pageblock the migration scanner | |
1060 | * is using. | |
1061 | */ | |
1062 | isolate_start_pfn = cc->free_pfn; | |
1063 | block_start_pfn = pageblock_start_pfn(cc->free_pfn); | |
1064 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, | |
1065 | zone_end_pfn(zone)); | |
1066 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); | |
1067 | ||
1068 | /* | |
1069 | * Isolate free pages until enough are available to migrate the | |
1070 | * pages on cc->migratepages. We stop searching if the migrate | |
1071 | * and free page scanners meet or enough free pages are isolated. | |
1072 | */ | |
1073 | for (; block_start_pfn >= low_pfn; | |
1074 | block_end_pfn = block_start_pfn, | |
1075 | block_start_pfn -= pageblock_nr_pages, | |
1076 | isolate_start_pfn = block_start_pfn) { | |
1077 | /* | |
1078 | * This can iterate a massively long zone without finding any | |
1079 | * suitable migration targets, so periodically check if we need | |
1080 | * to schedule, or even abort async compaction. | |
1081 | */ | |
1082 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1083 | && compact_should_abort(cc)) | |
1084 | break; | |
1085 | ||
1086 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, | |
1087 | zone); | |
1088 | if (!page) | |
1089 | continue; | |
1090 | ||
1091 | /* Check the block is suitable for migration */ | |
1092 | if (!suitable_migration_target(cc, page)) | |
1093 | continue; | |
1094 | ||
1095 | /* If isolation recently failed, do not retry */ | |
1096 | if (!isolation_suitable(cc, page)) | |
1097 | continue; | |
1098 | ||
1099 | /* Found a block suitable for isolating free pages from. */ | |
1100 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, | |
1101 | freelist, false); | |
1102 | ||
1103 | /* | |
1104 | * If we isolated enough freepages, or aborted due to lock | |
1105 | * contention, terminate. | |
1106 | */ | |
1107 | if ((cc->nr_freepages >= cc->nr_migratepages) | |
1108 | || cc->contended) { | |
1109 | if (isolate_start_pfn >= block_end_pfn) { | |
1110 | /* | |
1111 | * Restart at previous pageblock if more | |
1112 | * freepages can be isolated next time. | |
1113 | */ | |
1114 | isolate_start_pfn = | |
1115 | block_start_pfn - pageblock_nr_pages; | |
1116 | } | |
1117 | break; | |
1118 | } else if (isolate_start_pfn < block_end_pfn) { | |
1119 | /* | |
1120 | * If isolation failed early, do not continue | |
1121 | * needlessly. | |
1122 | */ | |
1123 | break; | |
1124 | } | |
1125 | } | |
1126 | ||
1127 | /* __isolate_free_page() does not map the pages */ | |
1128 | map_pages(freelist); | |
1129 | ||
1130 | /* | |
1131 | * Record where the free scanner will restart next time. Either we | |
1132 | * broke from the loop and set isolate_start_pfn based on the last | |
1133 | * call to isolate_freepages_block(), or we met the migration scanner | |
1134 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
1135 | */ | |
1136 | cc->free_pfn = isolate_start_pfn; | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1141 | * from the isolated freelists in the block we are migrating to. | |
1142 | */ | |
1143 | static struct page *compaction_alloc(struct page *migratepage, | |
1144 | unsigned long data, | |
1145 | int **result) | |
1146 | { | |
1147 | struct compact_control *cc = (struct compact_control *)data; | |
1148 | struct page *freepage; | |
1149 | ||
1150 | /* | |
1151 | * Isolate free pages if necessary, and if we are not aborting due to | |
1152 | * contention. | |
1153 | */ | |
1154 | if (list_empty(&cc->freepages)) { | |
1155 | if (!cc->contended) | |
1156 | isolate_freepages(cc); | |
1157 | ||
1158 | if (list_empty(&cc->freepages)) | |
1159 | return NULL; | |
1160 | } | |
1161 | ||
1162 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1163 | list_del(&freepage->lru); | |
1164 | cc->nr_freepages--; | |
1165 | ||
1166 | return freepage; | |
1167 | } | |
1168 | ||
1169 | /* | |
1170 | * This is a migrate-callback that "frees" freepages back to the isolated | |
1171 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1172 | * special handling needed for NUMA. | |
1173 | */ | |
1174 | static void compaction_free(struct page *page, unsigned long data) | |
1175 | { | |
1176 | struct compact_control *cc = (struct compact_control *)data; | |
1177 | ||
1178 | list_add(&page->lru, &cc->freepages); | |
1179 | cc->nr_freepages++; | |
1180 | } | |
1181 | ||
1182 | /* possible outcome of isolate_migratepages */ | |
1183 | typedef enum { | |
1184 | ISOLATE_ABORT, /* Abort compaction now */ | |
1185 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1186 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1187 | } isolate_migrate_t; | |
1188 | ||
1189 | /* | |
1190 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1191 | * compactable pages. | |
1192 | */ | |
1193 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1194 | ||
1195 | /* | |
1196 | * Isolate all pages that can be migrated from the first suitable block, | |
1197 | * starting at the block pointed to by the migrate scanner pfn within | |
1198 | * compact_control. | |
1199 | */ | |
1200 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1201 | struct compact_control *cc) | |
1202 | { | |
1203 | unsigned long block_start_pfn; | |
1204 | unsigned long block_end_pfn; | |
1205 | unsigned long low_pfn; | |
1206 | struct page *page; | |
1207 | const isolate_mode_t isolate_mode = | |
1208 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | | |
1209 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); | |
1210 | ||
1211 | /* | |
1212 | * Start at where we last stopped, or beginning of the zone as | |
1213 | * initialized by compact_zone() | |
1214 | */ | |
1215 | low_pfn = cc->migrate_pfn; | |
1216 | block_start_pfn = pageblock_start_pfn(low_pfn); | |
1217 | if (block_start_pfn < zone->zone_start_pfn) | |
1218 | block_start_pfn = zone->zone_start_pfn; | |
1219 | ||
1220 | /* Only scan within a pageblock boundary */ | |
1221 | block_end_pfn = pageblock_end_pfn(low_pfn); | |
1222 | ||
1223 | /* | |
1224 | * Iterate over whole pageblocks until we find the first suitable. | |
1225 | * Do not cross the free scanner. | |
1226 | */ | |
1227 | for (; block_end_pfn <= cc->free_pfn; | |
1228 | low_pfn = block_end_pfn, | |
1229 | block_start_pfn = block_end_pfn, | |
1230 | block_end_pfn += pageblock_nr_pages) { | |
1231 | ||
1232 | /* | |
1233 | * This can potentially iterate a massively long zone with | |
1234 | * many pageblocks unsuitable, so periodically check if we | |
1235 | * need to schedule, or even abort async compaction. | |
1236 | */ | |
1237 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1238 | && compact_should_abort(cc)) | |
1239 | break; | |
1240 | ||
1241 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, | |
1242 | zone); | |
1243 | if (!page) | |
1244 | continue; | |
1245 | ||
1246 | /* If isolation recently failed, do not retry */ | |
1247 | if (!isolation_suitable(cc, page)) | |
1248 | continue; | |
1249 | ||
1250 | /* | |
1251 | * For async compaction, also only scan in MOVABLE blocks. | |
1252 | * Async compaction is optimistic to see if the minimum amount | |
1253 | * of work satisfies the allocation. | |
1254 | */ | |
1255 | if (!suitable_migration_source(cc, page)) | |
1256 | continue; | |
1257 | ||
1258 | /* Perform the isolation */ | |
1259 | low_pfn = isolate_migratepages_block(cc, low_pfn, | |
1260 | block_end_pfn, isolate_mode); | |
1261 | ||
1262 | if (!low_pfn || cc->contended) | |
1263 | return ISOLATE_ABORT; | |
1264 | ||
1265 | /* | |
1266 | * Either we isolated something and proceed with migration. Or | |
1267 | * we failed and compact_zone should decide if we should | |
1268 | * continue or not. | |
1269 | */ | |
1270 | break; | |
1271 | } | |
1272 | ||
1273 | /* Record where migration scanner will be restarted. */ | |
1274 | cc->migrate_pfn = low_pfn; | |
1275 | ||
1276 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; | |
1277 | } | |
1278 | ||
1279 | /* | |
1280 | * order == -1 is expected when compacting via | |
1281 | * /proc/sys/vm/compact_memory | |
1282 | */ | |
1283 | static inline bool is_via_compact_memory(int order) | |
1284 | { | |
1285 | return order == -1; | |
1286 | } | |
1287 | ||
1288 | static enum compact_result __compact_finished(struct zone *zone, | |
1289 | struct compact_control *cc) | |
1290 | { | |
1291 | unsigned int order; | |
1292 | const int migratetype = cc->migratetype; | |
1293 | ||
1294 | if (cc->contended || fatal_signal_pending(current)) | |
1295 | return COMPACT_CONTENDED; | |
1296 | ||
1297 | /* Compaction run completes if the migrate and free scanner meet */ | |
1298 | if (compact_scanners_met(cc)) { | |
1299 | /* Let the next compaction start anew. */ | |
1300 | reset_cached_positions(zone); | |
1301 | ||
1302 | /* | |
1303 | * Mark that the PG_migrate_skip information should be cleared | |
1304 | * by kswapd when it goes to sleep. kcompactd does not set the | |
1305 | * flag itself as the decision to be clear should be directly | |
1306 | * based on an allocation request. | |
1307 | */ | |
1308 | if (cc->direct_compaction) | |
1309 | zone->compact_blockskip_flush = true; | |
1310 | ||
1311 | if (cc->whole_zone) | |
1312 | return COMPACT_COMPLETE; | |
1313 | else | |
1314 | return COMPACT_PARTIAL_SKIPPED; | |
1315 | } | |
1316 | ||
1317 | if (is_via_compact_memory(cc->order)) | |
1318 | return COMPACT_CONTINUE; | |
1319 | ||
1320 | if (cc->finishing_block) { | |
1321 | /* | |
1322 | * We have finished the pageblock, but better check again that | |
1323 | * we really succeeded. | |
1324 | */ | |
1325 | if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1326 | cc->finishing_block = false; | |
1327 | else | |
1328 | return COMPACT_CONTINUE; | |
1329 | } | |
1330 | ||
1331 | /* Direct compactor: Is a suitable page free? */ | |
1332 | for (order = cc->order; order < MAX_ORDER; order++) { | |
1333 | struct free_area *area = &zone->free_area[order]; | |
1334 | bool can_steal; | |
1335 | ||
1336 | /* Job done if page is free of the right migratetype */ | |
1337 | if (!list_empty(&area->free_list[migratetype])) | |
1338 | return COMPACT_SUCCESS; | |
1339 | ||
1340 | #ifdef CONFIG_CMA | |
1341 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1342 | if (migratetype == MIGRATE_MOVABLE && | |
1343 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
1344 | return COMPACT_SUCCESS; | |
1345 | #endif | |
1346 | /* | |
1347 | * Job done if allocation would steal freepages from | |
1348 | * other migratetype buddy lists. | |
1349 | */ | |
1350 | if (find_suitable_fallback(area, order, migratetype, | |
1351 | true, &can_steal) != -1) { | |
1352 | ||
1353 | /* movable pages are OK in any pageblock */ | |
1354 | if (migratetype == MIGRATE_MOVABLE) | |
1355 | return COMPACT_SUCCESS; | |
1356 | ||
1357 | /* | |
1358 | * We are stealing for a non-movable allocation. Make | |
1359 | * sure we finish compacting the current pageblock | |
1360 | * first so it is as free as possible and we won't | |
1361 | * have to steal another one soon. This only applies | |
1362 | * to sync compaction, as async compaction operates | |
1363 | * on pageblocks of the same migratetype. | |
1364 | */ | |
1365 | if (cc->mode == MIGRATE_ASYNC || | |
1366 | IS_ALIGNED(cc->migrate_pfn, | |
1367 | pageblock_nr_pages)) { | |
1368 | return COMPACT_SUCCESS; | |
1369 | } | |
1370 | ||
1371 | cc->finishing_block = true; | |
1372 | return COMPACT_CONTINUE; | |
1373 | } | |
1374 | } | |
1375 | ||
1376 | return COMPACT_NO_SUITABLE_PAGE; | |
1377 | } | |
1378 | ||
1379 | static enum compact_result compact_finished(struct zone *zone, | |
1380 | struct compact_control *cc) | |
1381 | { | |
1382 | int ret; | |
1383 | ||
1384 | ret = __compact_finished(zone, cc); | |
1385 | trace_mm_compaction_finished(zone, cc->order, ret); | |
1386 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1387 | ret = COMPACT_CONTINUE; | |
1388 | ||
1389 | return ret; | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1394 | * Returns | |
1395 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
1396 | * COMPACT_SUCCESS - If the allocation would succeed without compaction | |
1397 | * COMPACT_CONTINUE - If compaction should run now | |
1398 | */ | |
1399 | static enum compact_result __compaction_suitable(struct zone *zone, int order, | |
1400 | unsigned int alloc_flags, | |
1401 | int classzone_idx, | |
1402 | unsigned long wmark_target) | |
1403 | { | |
1404 | unsigned long watermark; | |
1405 | ||
1406 | if (is_via_compact_memory(order)) | |
1407 | return COMPACT_CONTINUE; | |
1408 | ||
1409 | watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; | |
1410 | /* | |
1411 | * If watermarks for high-order allocation are already met, there | |
1412 | * should be no need for compaction at all. | |
1413 | */ | |
1414 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1415 | alloc_flags)) | |
1416 | return COMPACT_SUCCESS; | |
1417 | ||
1418 | /* | |
1419 | * Watermarks for order-0 must be met for compaction to be able to | |
1420 | * isolate free pages for migration targets. This means that the | |
1421 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1422 | * the check in __isolate_free_page(). We don't use the direct | |
1423 | * compactor's alloc_flags, as they are not relevant for freepage | |
1424 | * isolation. We however do use the direct compactor's classzone_idx to | |
1425 | * skip over zones where lowmem reserves would prevent allocation even | |
1426 | * if compaction succeeds. | |
1427 | * For costly orders, we require low watermark instead of min for | |
1428 | * compaction to proceed to increase its chances. | |
1429 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered | |
1430 | * suitable migration targets | |
1431 | */ | |
1432 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? | |
1433 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
1434 | watermark += compact_gap(order); | |
1435 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, | |
1436 | ALLOC_CMA, wmark_target)) | |
1437 | return COMPACT_SKIPPED; | |
1438 | ||
1439 | return COMPACT_CONTINUE; | |
1440 | } | |
1441 | ||
1442 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
1443 | unsigned int alloc_flags, | |
1444 | int classzone_idx) | |
1445 | { | |
1446 | enum compact_result ret; | |
1447 | int fragindex; | |
1448 | ||
1449 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
1450 | zone_page_state(zone, NR_FREE_PAGES)); | |
1451 | /* | |
1452 | * fragmentation index determines if allocation failures are due to | |
1453 | * low memory or external fragmentation | |
1454 | * | |
1455 | * index of -1000 would imply allocations might succeed depending on | |
1456 | * watermarks, but we already failed the high-order watermark check | |
1457 | * index towards 0 implies failure is due to lack of memory | |
1458 | * index towards 1000 implies failure is due to fragmentation | |
1459 | * | |
1460 | * Only compact if a failure would be due to fragmentation. Also | |
1461 | * ignore fragindex for non-costly orders where the alternative to | |
1462 | * a successful reclaim/compaction is OOM. Fragindex and the | |
1463 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
1464 | * excessive compaction for costly orders, but it should not be at the | |
1465 | * expense of system stability. | |
1466 | */ | |
1467 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { | |
1468 | fragindex = fragmentation_index(zone, order); | |
1469 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
1470 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
1471 | } | |
1472 | ||
1473 | trace_mm_compaction_suitable(zone, order, ret); | |
1474 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1475 | ret = COMPACT_SKIPPED; | |
1476 | ||
1477 | return ret; | |
1478 | } | |
1479 | ||
1480 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, | |
1481 | int alloc_flags) | |
1482 | { | |
1483 | struct zone *zone; | |
1484 | struct zoneref *z; | |
1485 | ||
1486 | /* | |
1487 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1488 | * retrying the reclaim. | |
1489 | */ | |
1490 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1491 | ac->nodemask) { | |
1492 | unsigned long available; | |
1493 | enum compact_result compact_result; | |
1494 | ||
1495 | /* | |
1496 | * Do not consider all the reclaimable memory because we do not | |
1497 | * want to trash just for a single high order allocation which | |
1498 | * is even not guaranteed to appear even if __compaction_suitable | |
1499 | * is happy about the watermark check. | |
1500 | */ | |
1501 | available = zone_reclaimable_pages(zone) / order; | |
1502 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); | |
1503 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1504 | ac_classzone_idx(ac), available); | |
1505 | if (compact_result != COMPACT_SKIPPED) | |
1506 | return true; | |
1507 | } | |
1508 | ||
1509 | return false; | |
1510 | } | |
1511 | ||
1512 | static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc) | |
1513 | { | |
1514 | enum compact_result ret; | |
1515 | unsigned long start_pfn = zone->zone_start_pfn; | |
1516 | unsigned long end_pfn = zone_end_pfn(zone); | |
1517 | const bool sync = cc->mode != MIGRATE_ASYNC; | |
1518 | ||
1519 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); | |
1520 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, | |
1521 | cc->classzone_idx); | |
1522 | /* Compaction is likely to fail */ | |
1523 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) | |
1524 | return ret; | |
1525 | ||
1526 | /* huh, compaction_suitable is returning something unexpected */ | |
1527 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
1528 | ||
1529 | /* | |
1530 | * Clear pageblock skip if there were failures recently and compaction | |
1531 | * is about to be retried after being deferred. | |
1532 | */ | |
1533 | if (compaction_restarting(zone, cc->order)) | |
1534 | __reset_isolation_suitable(zone); | |
1535 | ||
1536 | /* | |
1537 | * Setup to move all movable pages to the end of the zone. Used cached | |
1538 | * information on where the scanners should start (unless we explicitly | |
1539 | * want to compact the whole zone), but check that it is initialised | |
1540 | * by ensuring the values are within zone boundaries. | |
1541 | */ | |
1542 | if (cc->whole_zone) { | |
1543 | cc->migrate_pfn = start_pfn; | |
1544 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
1545 | } else { | |
1546 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; | |
1547 | cc->free_pfn = zone->compact_cached_free_pfn; | |
1548 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { | |
1549 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
1550 | zone->compact_cached_free_pfn = cc->free_pfn; | |
1551 | } | |
1552 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
1553 | cc->migrate_pfn = start_pfn; | |
1554 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; | |
1555 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
1556 | } | |
1557 | ||
1558 | if (cc->migrate_pfn == start_pfn) | |
1559 | cc->whole_zone = true; | |
1560 | } | |
1561 | ||
1562 | cc->last_migrated_pfn = 0; | |
1563 | ||
1564 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, | |
1565 | cc->free_pfn, end_pfn, sync); | |
1566 | ||
1567 | migrate_prep_local(); | |
1568 | ||
1569 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { | |
1570 | int err; | |
1571 | ||
1572 | switch (isolate_migratepages(zone, cc)) { | |
1573 | case ISOLATE_ABORT: | |
1574 | ret = COMPACT_CONTENDED; | |
1575 | putback_movable_pages(&cc->migratepages); | |
1576 | cc->nr_migratepages = 0; | |
1577 | goto out; | |
1578 | case ISOLATE_NONE: | |
1579 | /* | |
1580 | * We haven't isolated and migrated anything, but | |
1581 | * there might still be unflushed migrations from | |
1582 | * previous cc->order aligned block. | |
1583 | */ | |
1584 | goto check_drain; | |
1585 | case ISOLATE_SUCCESS: | |
1586 | ; | |
1587 | } | |
1588 | ||
1589 | err = migrate_pages(&cc->migratepages, compaction_alloc, | |
1590 | compaction_free, (unsigned long)cc, cc->mode, | |
1591 | MR_COMPACTION); | |
1592 | ||
1593 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, | |
1594 | &cc->migratepages); | |
1595 | ||
1596 | /* All pages were either migrated or will be released */ | |
1597 | cc->nr_migratepages = 0; | |
1598 | if (err) { | |
1599 | putback_movable_pages(&cc->migratepages); | |
1600 | /* | |
1601 | * migrate_pages() may return -ENOMEM when scanners meet | |
1602 | * and we want compact_finished() to detect it | |
1603 | */ | |
1604 | if (err == -ENOMEM && !compact_scanners_met(cc)) { | |
1605 | ret = COMPACT_CONTENDED; | |
1606 | goto out; | |
1607 | } | |
1608 | /* | |
1609 | * We failed to migrate at least one page in the current | |
1610 | * order-aligned block, so skip the rest of it. | |
1611 | */ | |
1612 | if (cc->direct_compaction && | |
1613 | (cc->mode == MIGRATE_ASYNC)) { | |
1614 | cc->migrate_pfn = block_end_pfn( | |
1615 | cc->migrate_pfn - 1, cc->order); | |
1616 | /* Draining pcplists is useless in this case */ | |
1617 | cc->last_migrated_pfn = 0; | |
1618 | ||
1619 | } | |
1620 | } | |
1621 | ||
1622 | check_drain: | |
1623 | /* | |
1624 | * Has the migration scanner moved away from the previous | |
1625 | * cc->order aligned block where we migrated from? If yes, | |
1626 | * flush the pages that were freed, so that they can merge and | |
1627 | * compact_finished() can detect immediately if allocation | |
1628 | * would succeed. | |
1629 | */ | |
1630 | if (cc->order > 0 && cc->last_migrated_pfn) { | |
1631 | int cpu; | |
1632 | unsigned long current_block_start = | |
1633 | block_start_pfn(cc->migrate_pfn, cc->order); | |
1634 | ||
1635 | if (cc->last_migrated_pfn < current_block_start) { | |
1636 | cpu = get_cpu(); | |
1637 | lru_add_drain_cpu(cpu); | |
1638 | drain_local_pages(zone); | |
1639 | put_cpu(); | |
1640 | /* No more flushing until we migrate again */ | |
1641 | cc->last_migrated_pfn = 0; | |
1642 | } | |
1643 | } | |
1644 | ||
1645 | } | |
1646 | ||
1647 | out: | |
1648 | /* | |
1649 | * Release free pages and update where the free scanner should restart, | |
1650 | * so we don't leave any returned pages behind in the next attempt. | |
1651 | */ | |
1652 | if (cc->nr_freepages > 0) { | |
1653 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1654 | ||
1655 | cc->nr_freepages = 0; | |
1656 | VM_BUG_ON(free_pfn == 0); | |
1657 | /* The cached pfn is always the first in a pageblock */ | |
1658 | free_pfn = pageblock_start_pfn(free_pfn); | |
1659 | /* | |
1660 | * Only go back, not forward. The cached pfn might have been | |
1661 | * already reset to zone end in compact_finished() | |
1662 | */ | |
1663 | if (free_pfn > zone->compact_cached_free_pfn) | |
1664 | zone->compact_cached_free_pfn = free_pfn; | |
1665 | } | |
1666 | ||
1667 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); | |
1668 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
1669 | ||
1670 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, | |
1671 | cc->free_pfn, end_pfn, sync, ret); | |
1672 | ||
1673 | return ret; | |
1674 | } | |
1675 | ||
1676 | static enum compact_result compact_zone_order(struct zone *zone, int order, | |
1677 | gfp_t gfp_mask, enum compact_priority prio, | |
1678 | unsigned int alloc_flags, int classzone_idx) | |
1679 | { | |
1680 | enum compact_result ret; | |
1681 | struct compact_control cc = { | |
1682 | .nr_freepages = 0, | |
1683 | .nr_migratepages = 0, | |
1684 | .total_migrate_scanned = 0, | |
1685 | .total_free_scanned = 0, | |
1686 | .order = order, | |
1687 | .gfp_mask = gfp_mask, | |
1688 | .zone = zone, | |
1689 | .mode = (prio == COMPACT_PRIO_ASYNC) ? | |
1690 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
1691 | .alloc_flags = alloc_flags, | |
1692 | .classzone_idx = classzone_idx, | |
1693 | .direct_compaction = true, | |
1694 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), | |
1695 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), | |
1696 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
1697 | }; | |
1698 | INIT_LIST_HEAD(&cc.freepages); | |
1699 | INIT_LIST_HEAD(&cc.migratepages); | |
1700 | ||
1701 | ret = compact_zone(zone, &cc); | |
1702 | ||
1703 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1704 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1705 | ||
1706 | return ret; | |
1707 | } | |
1708 | ||
1709 | int sysctl_extfrag_threshold = 500; | |
1710 | ||
1711 | /** | |
1712 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
1713 | * @gfp_mask: The GFP mask of the current allocation | |
1714 | * @order: The order of the current allocation | |
1715 | * @alloc_flags: The allocation flags of the current allocation | |
1716 | * @ac: The context of current allocation | |
1717 | * @mode: The migration mode for async, sync light, or sync migration | |
1718 | * | |
1719 | * This is the main entry point for direct page compaction. | |
1720 | */ | |
1721 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, | |
1722 | unsigned int alloc_flags, const struct alloc_context *ac, | |
1723 | enum compact_priority prio) | |
1724 | { | |
1725 | int may_perform_io = gfp_mask & __GFP_IO; | |
1726 | struct zoneref *z; | |
1727 | struct zone *zone; | |
1728 | enum compact_result rc = COMPACT_SKIPPED; | |
1729 | ||
1730 | /* | |
1731 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
1732 | * tricky context because the migration might require IO | |
1733 | */ | |
1734 | if (!may_perform_io) | |
1735 | return COMPACT_SKIPPED; | |
1736 | ||
1737 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); | |
1738 | ||
1739 | /* Compact each zone in the list */ | |
1740 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1741 | ac->nodemask) { | |
1742 | enum compact_result status; | |
1743 | ||
1744 | if (prio > MIN_COMPACT_PRIORITY | |
1745 | && compaction_deferred(zone, order)) { | |
1746 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); | |
1747 | continue; | |
1748 | } | |
1749 | ||
1750 | status = compact_zone_order(zone, order, gfp_mask, prio, | |
1751 | alloc_flags, ac_classzone_idx(ac)); | |
1752 | rc = max(status, rc); | |
1753 | ||
1754 | /* The allocation should succeed, stop compacting */ | |
1755 | if (status == COMPACT_SUCCESS) { | |
1756 | /* | |
1757 | * We think the allocation will succeed in this zone, | |
1758 | * but it is not certain, hence the false. The caller | |
1759 | * will repeat this with true if allocation indeed | |
1760 | * succeeds in this zone. | |
1761 | */ | |
1762 | compaction_defer_reset(zone, order, false); | |
1763 | ||
1764 | break; | |
1765 | } | |
1766 | ||
1767 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || | |
1768 | status == COMPACT_PARTIAL_SKIPPED)) | |
1769 | /* | |
1770 | * We think that allocation won't succeed in this zone | |
1771 | * so we defer compaction there. If it ends up | |
1772 | * succeeding after all, it will be reset. | |
1773 | */ | |
1774 | defer_compaction(zone, order); | |
1775 | ||
1776 | /* | |
1777 | * We might have stopped compacting due to need_resched() in | |
1778 | * async compaction, or due to a fatal signal detected. In that | |
1779 | * case do not try further zones | |
1780 | */ | |
1781 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) | |
1782 | || fatal_signal_pending(current)) | |
1783 | break; | |
1784 | } | |
1785 | ||
1786 | return rc; | |
1787 | } | |
1788 | ||
1789 | ||
1790 | /* Compact all zones within a node */ | |
1791 | static void compact_node(int nid) | |
1792 | { | |
1793 | pg_data_t *pgdat = NODE_DATA(nid); | |
1794 | int zoneid; | |
1795 | struct zone *zone; | |
1796 | struct compact_control cc = { | |
1797 | .order = -1, | |
1798 | .total_migrate_scanned = 0, | |
1799 | .total_free_scanned = 0, | |
1800 | .mode = MIGRATE_SYNC, | |
1801 | .ignore_skip_hint = true, | |
1802 | .whole_zone = true, | |
1803 | .gfp_mask = GFP_KERNEL, | |
1804 | }; | |
1805 | ||
1806 | ||
1807 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
1808 | ||
1809 | zone = &pgdat->node_zones[zoneid]; | |
1810 | if (!populated_zone(zone)) | |
1811 | continue; | |
1812 | ||
1813 | cc.nr_freepages = 0; | |
1814 | cc.nr_migratepages = 0; | |
1815 | cc.zone = zone; | |
1816 | INIT_LIST_HEAD(&cc.freepages); | |
1817 | INIT_LIST_HEAD(&cc.migratepages); | |
1818 | ||
1819 | compact_zone(zone, &cc); | |
1820 | ||
1821 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1822 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1823 | } | |
1824 | } | |
1825 | ||
1826 | /* Compact all nodes in the system */ | |
1827 | static void compact_nodes(void) | |
1828 | { | |
1829 | int nid; | |
1830 | ||
1831 | /* Flush pending updates to the LRU lists */ | |
1832 | lru_add_drain_all(); | |
1833 | ||
1834 | for_each_online_node(nid) | |
1835 | compact_node(nid); | |
1836 | } | |
1837 | ||
1838 | /* The written value is actually unused, all memory is compacted */ | |
1839 | int sysctl_compact_memory; | |
1840 | ||
1841 | /* | |
1842 | * This is the entry point for compacting all nodes via | |
1843 | * /proc/sys/vm/compact_memory | |
1844 | */ | |
1845 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
1846 | void __user *buffer, size_t *length, loff_t *ppos) | |
1847 | { | |
1848 | if (write) | |
1849 | compact_nodes(); | |
1850 | ||
1851 | return 0; | |
1852 | } | |
1853 | ||
1854 | int sysctl_extfrag_handler(struct ctl_table *table, int write, | |
1855 | void __user *buffer, size_t *length, loff_t *ppos) | |
1856 | { | |
1857 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1858 | ||
1859 | return 0; | |
1860 | } | |
1861 | ||
1862 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
1863 | static ssize_t sysfs_compact_node(struct device *dev, | |
1864 | struct device_attribute *attr, | |
1865 | const char *buf, size_t count) | |
1866 | { | |
1867 | int nid = dev->id; | |
1868 | ||
1869 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1870 | /* Flush pending updates to the LRU lists */ | |
1871 | lru_add_drain_all(); | |
1872 | ||
1873 | compact_node(nid); | |
1874 | } | |
1875 | ||
1876 | return count; | |
1877 | } | |
1878 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); | |
1879 | ||
1880 | int compaction_register_node(struct node *node) | |
1881 | { | |
1882 | return device_create_file(&node->dev, &dev_attr_compact); | |
1883 | } | |
1884 | ||
1885 | void compaction_unregister_node(struct node *node) | |
1886 | { | |
1887 | return device_remove_file(&node->dev, &dev_attr_compact); | |
1888 | } | |
1889 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
1890 | ||
1891 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) | |
1892 | { | |
1893 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); | |
1894 | } | |
1895 | ||
1896 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
1897 | { | |
1898 | int zoneid; | |
1899 | struct zone *zone; | |
1900 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
1901 | ||
1902 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { | |
1903 | zone = &pgdat->node_zones[zoneid]; | |
1904 | ||
1905 | if (!populated_zone(zone)) | |
1906 | continue; | |
1907 | ||
1908 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
1909 | classzone_idx) == COMPACT_CONTINUE) | |
1910 | return true; | |
1911 | } | |
1912 | ||
1913 | return false; | |
1914 | } | |
1915 | ||
1916 | static void kcompactd_do_work(pg_data_t *pgdat) | |
1917 | { | |
1918 | /* | |
1919 | * With no special task, compact all zones so that a page of requested | |
1920 | * order is allocatable. | |
1921 | */ | |
1922 | int zoneid; | |
1923 | struct zone *zone; | |
1924 | struct compact_control cc = { | |
1925 | .order = pgdat->kcompactd_max_order, | |
1926 | .total_migrate_scanned = 0, | |
1927 | .total_free_scanned = 0, | |
1928 | .classzone_idx = pgdat->kcompactd_classzone_idx, | |
1929 | .mode = MIGRATE_SYNC_LIGHT, | |
1930 | .ignore_skip_hint = true, | |
1931 | .gfp_mask = GFP_KERNEL, | |
1932 | ||
1933 | }; | |
1934 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, | |
1935 | cc.classzone_idx); | |
1936 | count_compact_event(KCOMPACTD_WAKE); | |
1937 | ||
1938 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { | |
1939 | int status; | |
1940 | ||
1941 | zone = &pgdat->node_zones[zoneid]; | |
1942 | if (!populated_zone(zone)) | |
1943 | continue; | |
1944 | ||
1945 | if (compaction_deferred(zone, cc.order)) | |
1946 | continue; | |
1947 | ||
1948 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
1949 | COMPACT_CONTINUE) | |
1950 | continue; | |
1951 | ||
1952 | cc.nr_freepages = 0; | |
1953 | cc.nr_migratepages = 0; | |
1954 | cc.total_migrate_scanned = 0; | |
1955 | cc.total_free_scanned = 0; | |
1956 | cc.zone = zone; | |
1957 | INIT_LIST_HEAD(&cc.freepages); | |
1958 | INIT_LIST_HEAD(&cc.migratepages); | |
1959 | ||
1960 | if (kthread_should_stop()) | |
1961 | return; | |
1962 | status = compact_zone(zone, &cc); | |
1963 | ||
1964 | if (status == COMPACT_SUCCESS) { | |
1965 | compaction_defer_reset(zone, cc.order, false); | |
1966 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { | |
1967 | /* | |
1968 | * We use sync migration mode here, so we defer like | |
1969 | * sync direct compaction does. | |
1970 | */ | |
1971 | defer_compaction(zone, cc.order); | |
1972 | } | |
1973 | ||
1974 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, | |
1975 | cc.total_migrate_scanned); | |
1976 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
1977 | cc.total_free_scanned); | |
1978 | ||
1979 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1980 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1981 | } | |
1982 | ||
1983 | /* | |
1984 | * Regardless of success, we are done until woken up next. But remember | |
1985 | * the requested order/classzone_idx in case it was higher/tighter than | |
1986 | * our current ones | |
1987 | */ | |
1988 | if (pgdat->kcompactd_max_order <= cc.order) | |
1989 | pgdat->kcompactd_max_order = 0; | |
1990 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
1991 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
1992 | } | |
1993 | ||
1994 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
1995 | { | |
1996 | if (!order) | |
1997 | return; | |
1998 | ||
1999 | if (pgdat->kcompactd_max_order < order) | |
2000 | pgdat->kcompactd_max_order = order; | |
2001 | ||
2002 | /* | |
2003 | * Pairs with implicit barrier in wait_event_freezable() | |
2004 | * such that wakeups are not missed in the lockless | |
2005 | * waitqueue_active() call. | |
2006 | */ | |
2007 | smp_acquire__after_ctrl_dep(); | |
2008 | ||
2009 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2010 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2011 | ||
2012 | if (!waitqueue_active(&pgdat->kcompactd_wait)) | |
2013 | return; | |
2014 | ||
2015 | if (!kcompactd_node_suitable(pgdat)) | |
2016 | return; | |
2017 | ||
2018 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2019 | classzone_idx); | |
2020 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2021 | } | |
2022 | ||
2023 | /* | |
2024 | * The background compaction daemon, started as a kernel thread | |
2025 | * from the init process. | |
2026 | */ | |
2027 | static int kcompactd(void *p) | |
2028 | { | |
2029 | pg_data_t *pgdat = (pg_data_t*)p; | |
2030 | struct task_struct *tsk = current; | |
2031 | ||
2032 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2033 | ||
2034 | if (!cpumask_empty(cpumask)) | |
2035 | set_cpus_allowed_ptr(tsk, cpumask); | |
2036 | ||
2037 | set_freezable(); | |
2038 | ||
2039 | pgdat->kcompactd_max_order = 0; | |
2040 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2041 | ||
2042 | while (!kthread_should_stop()) { | |
2043 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); | |
2044 | wait_event_freezable(pgdat->kcompactd_wait, | |
2045 | kcompactd_work_requested(pgdat)); | |
2046 | ||
2047 | kcompactd_do_work(pgdat); | |
2048 | } | |
2049 | ||
2050 | return 0; | |
2051 | } | |
2052 | ||
2053 | /* | |
2054 | * This kcompactd start function will be called by init and node-hot-add. | |
2055 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2056 | */ | |
2057 | int kcompactd_run(int nid) | |
2058 | { | |
2059 | pg_data_t *pgdat = NODE_DATA(nid); | |
2060 | int ret = 0; | |
2061 | ||
2062 | if (pgdat->kcompactd) | |
2063 | return 0; | |
2064 | ||
2065 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2066 | if (IS_ERR(pgdat->kcompactd)) { | |
2067 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2068 | ret = PTR_ERR(pgdat->kcompactd); | |
2069 | pgdat->kcompactd = NULL; | |
2070 | } | |
2071 | return ret; | |
2072 | } | |
2073 | ||
2074 | /* | |
2075 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2076 | * hold mem_hotplug_begin/end(). | |
2077 | */ | |
2078 | void kcompactd_stop(int nid) | |
2079 | { | |
2080 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2081 | ||
2082 | if (kcompactd) { | |
2083 | kthread_stop(kcompactd); | |
2084 | NODE_DATA(nid)->kcompactd = NULL; | |
2085 | } | |
2086 | } | |
2087 | ||
2088 | /* | |
2089 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2090 | * not required for correctness. So if the last cpu in a node goes | |
2091 | * away, we get changed to run anywhere: as the first one comes back, | |
2092 | * restore their cpu bindings. | |
2093 | */ | |
2094 | static int kcompactd_cpu_online(unsigned int cpu) | |
2095 | { | |
2096 | int nid; | |
2097 | ||
2098 | for_each_node_state(nid, N_MEMORY) { | |
2099 | pg_data_t *pgdat = NODE_DATA(nid); | |
2100 | const struct cpumask *mask; | |
2101 | ||
2102 | mask = cpumask_of_node(pgdat->node_id); | |
2103 | ||
2104 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
2105 | /* One of our CPUs online: restore mask */ | |
2106 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
2107 | } | |
2108 | return 0; | |
2109 | } | |
2110 | ||
2111 | static int __init kcompactd_init(void) | |
2112 | { | |
2113 | int nid; | |
2114 | int ret; | |
2115 | ||
2116 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2117 | "mm/compaction:online", | |
2118 | kcompactd_cpu_online, NULL); | |
2119 | if (ret < 0) { | |
2120 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2121 | return ret; | |
2122 | } | |
2123 | ||
2124 | for_each_node_state(nid, N_MEMORY) | |
2125 | kcompactd_run(nid); | |
2126 | return 0; | |
2127 | } | |
2128 | subsys_initcall(kcompactd_init) | |
2129 | ||
2130 | #endif /* CONFIG_COMPACTION */ |