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