]>
Commit | Line | Data |
---|---|---|
748446bb MG |
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/swap.h> | |
11 | #include <linux/migrate.h> | |
12 | #include <linux/compaction.h> | |
13 | #include <linux/mm_inline.h> | |
14 | #include <linux/backing-dev.h> | |
76ab0f53 | 15 | #include <linux/sysctl.h> |
ed4a6d7f | 16 | #include <linux/sysfs.h> |
bf6bddf1 | 17 | #include <linux/balloon_compaction.h> |
194159fb | 18 | #include <linux/page-isolation.h> |
b8c73fc2 | 19 | #include <linux/kasan.h> |
748446bb MG |
20 | #include "internal.h" |
21 | ||
010fc29a MK |
22 | #ifdef CONFIG_COMPACTION |
23 | static inline void count_compact_event(enum vm_event_item item) | |
24 | { | |
25 | count_vm_event(item); | |
26 | } | |
27 | ||
28 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
29 | { | |
30 | count_vm_events(item, delta); | |
31 | } | |
32 | #else | |
33 | #define count_compact_event(item) do { } while (0) | |
34 | #define count_compact_events(item, delta) do { } while (0) | |
35 | #endif | |
36 | ||
ff9543fd | 37 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
16c4a097 JK |
38 | #ifdef CONFIG_TRACEPOINTS |
39 | static const char *const compaction_status_string[] = { | |
40 | "deferred", | |
41 | "skipped", | |
42 | "continue", | |
43 | "partial", | |
44 | "complete", | |
837d026d JK |
45 | "no_suitable_page", |
46 | "not_suitable_zone", | |
16c4a097 JK |
47 | }; |
48 | #endif | |
ff9543fd | 49 | |
b7aba698 MG |
50 | #define CREATE_TRACE_POINTS |
51 | #include <trace/events/compaction.h> | |
52 | ||
748446bb MG |
53 | static unsigned long release_freepages(struct list_head *freelist) |
54 | { | |
55 | struct page *page, *next; | |
6bace090 | 56 | unsigned long high_pfn = 0; |
748446bb MG |
57 | |
58 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 59 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
60 | list_del(&page->lru); |
61 | __free_page(page); | |
6bace090 VB |
62 | if (pfn > high_pfn) |
63 | high_pfn = pfn; | |
748446bb MG |
64 | } |
65 | ||
6bace090 | 66 | return high_pfn; |
748446bb MG |
67 | } |
68 | ||
ff9543fd MN |
69 | static void map_pages(struct list_head *list) |
70 | { | |
71 | struct page *page; | |
72 | ||
73 | list_for_each_entry(page, list, lru) { | |
74 | arch_alloc_page(page, 0); | |
75 | kernel_map_pages(page, 1, 1); | |
b8c73fc2 | 76 | kasan_alloc_pages(page, 0); |
ff9543fd MN |
77 | } |
78 | } | |
79 | ||
47118af0 MN |
80 | static inline bool migrate_async_suitable(int migratetype) |
81 | { | |
82 | return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE; | |
83 | } | |
84 | ||
7d49d886 VB |
85 | /* |
86 | * Check that the whole (or subset of) a pageblock given by the interval of | |
87 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it | |
88 | * with the migration of free compaction scanner. The scanners then need to | |
89 | * use only pfn_valid_within() check for arches that allow holes within | |
90 | * pageblocks. | |
91 | * | |
92 | * Return struct page pointer of start_pfn, or NULL if checks were not passed. | |
93 | * | |
94 | * It's possible on some configurations to have a setup like node0 node1 node0 | |
95 | * i.e. it's possible that all pages within a zones range of pages do not | |
96 | * belong to a single zone. We assume that a border between node0 and node1 | |
97 | * can occur within a single pageblock, but not a node0 node1 node0 | |
98 | * interleaving within a single pageblock. It is therefore sufficient to check | |
99 | * the first and last page of a pageblock and avoid checking each individual | |
100 | * page in a pageblock. | |
101 | */ | |
102 | static struct page *pageblock_pfn_to_page(unsigned long start_pfn, | |
103 | unsigned long end_pfn, struct zone *zone) | |
104 | { | |
105 | struct page *start_page; | |
106 | struct page *end_page; | |
107 | ||
108 | /* end_pfn is one past the range we are checking */ | |
109 | end_pfn--; | |
110 | ||
111 | if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn)) | |
112 | return NULL; | |
113 | ||
114 | start_page = pfn_to_page(start_pfn); | |
115 | ||
116 | if (page_zone(start_page) != zone) | |
117 | return NULL; | |
118 | ||
119 | end_page = pfn_to_page(end_pfn); | |
120 | ||
121 | /* This gives a shorter code than deriving page_zone(end_page) */ | |
122 | if (page_zone_id(start_page) != page_zone_id(end_page)) | |
123 | return NULL; | |
124 | ||
125 | return start_page; | |
126 | } | |
127 | ||
bb13ffeb | 128 | #ifdef CONFIG_COMPACTION |
24e2716f JK |
129 | |
130 | /* Do not skip compaction more than 64 times */ | |
131 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
132 | ||
133 | /* | |
134 | * Compaction is deferred when compaction fails to result in a page | |
135 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
136 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
137 | */ | |
138 | void defer_compaction(struct zone *zone, int order) | |
139 | { | |
140 | zone->compact_considered = 0; | |
141 | zone->compact_defer_shift++; | |
142 | ||
143 | if (order < zone->compact_order_failed) | |
144 | zone->compact_order_failed = order; | |
145 | ||
146 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
147 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
148 | ||
149 | trace_mm_compaction_defer_compaction(zone, order); | |
150 | } | |
151 | ||
152 | /* Returns true if compaction should be skipped this time */ | |
153 | bool compaction_deferred(struct zone *zone, int order) | |
154 | { | |
155 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
156 | ||
157 | if (order < zone->compact_order_failed) | |
158 | return false; | |
159 | ||
160 | /* Avoid possible overflow */ | |
161 | if (++zone->compact_considered > defer_limit) | |
162 | zone->compact_considered = defer_limit; | |
163 | ||
164 | if (zone->compact_considered >= defer_limit) | |
165 | return false; | |
166 | ||
167 | trace_mm_compaction_deferred(zone, order); | |
168 | ||
169 | return true; | |
170 | } | |
171 | ||
172 | /* | |
173 | * Update defer tracking counters after successful compaction of given order, | |
174 | * which means an allocation either succeeded (alloc_success == true) or is | |
175 | * expected to succeed. | |
176 | */ | |
177 | void compaction_defer_reset(struct zone *zone, int order, | |
178 | bool alloc_success) | |
179 | { | |
180 | if (alloc_success) { | |
181 | zone->compact_considered = 0; | |
182 | zone->compact_defer_shift = 0; | |
183 | } | |
184 | if (order >= zone->compact_order_failed) | |
185 | zone->compact_order_failed = order + 1; | |
186 | ||
187 | trace_mm_compaction_defer_reset(zone, order); | |
188 | } | |
189 | ||
190 | /* Returns true if restarting compaction after many failures */ | |
191 | bool compaction_restarting(struct zone *zone, int order) | |
192 | { | |
193 | if (order < zone->compact_order_failed) | |
194 | return false; | |
195 | ||
196 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
197 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
198 | } | |
199 | ||
bb13ffeb MG |
200 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
201 | static inline bool isolation_suitable(struct compact_control *cc, | |
202 | struct page *page) | |
203 | { | |
204 | if (cc->ignore_skip_hint) | |
205 | return true; | |
206 | ||
207 | return !get_pageblock_skip(page); | |
208 | } | |
209 | ||
02333641 VB |
210 | static void reset_cached_positions(struct zone *zone) |
211 | { | |
212 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
213 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
214 | zone->compact_cached_free_pfn = zone_end_pfn(zone); | |
215 | } | |
216 | ||
bb13ffeb MG |
217 | /* |
218 | * This function is called to clear all cached information on pageblocks that | |
219 | * should be skipped for page isolation when the migrate and free page scanner | |
220 | * meet. | |
221 | */ | |
62997027 | 222 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
223 | { |
224 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 225 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
226 | unsigned long pfn; |
227 | ||
62997027 | 228 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
229 | |
230 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
231 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
232 | struct page *page; | |
233 | ||
234 | cond_resched(); | |
235 | ||
236 | if (!pfn_valid(pfn)) | |
237 | continue; | |
238 | ||
239 | page = pfn_to_page(pfn); | |
240 | if (zone != page_zone(page)) | |
241 | continue; | |
242 | ||
243 | clear_pageblock_skip(page); | |
244 | } | |
02333641 VB |
245 | |
246 | reset_cached_positions(zone); | |
bb13ffeb MG |
247 | } |
248 | ||
62997027 MG |
249 | void reset_isolation_suitable(pg_data_t *pgdat) |
250 | { | |
251 | int zoneid; | |
252 | ||
253 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
254 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
255 | if (!populated_zone(zone)) | |
256 | continue; | |
257 | ||
258 | /* Only flush if a full compaction finished recently */ | |
259 | if (zone->compact_blockskip_flush) | |
260 | __reset_isolation_suitable(zone); | |
261 | } | |
262 | } | |
263 | ||
bb13ffeb MG |
264 | /* |
265 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 266 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 267 | */ |
c89511ab MG |
268 | static void update_pageblock_skip(struct compact_control *cc, |
269 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 270 | bool migrate_scanner) |
bb13ffeb | 271 | { |
c89511ab | 272 | struct zone *zone = cc->zone; |
35979ef3 | 273 | unsigned long pfn; |
6815bf3f JK |
274 | |
275 | if (cc->ignore_skip_hint) | |
276 | return; | |
277 | ||
bb13ffeb MG |
278 | if (!page) |
279 | return; | |
280 | ||
35979ef3 DR |
281 | if (nr_isolated) |
282 | return; | |
283 | ||
edc2ca61 | 284 | set_pageblock_skip(page); |
c89511ab | 285 | |
35979ef3 DR |
286 | pfn = page_to_pfn(page); |
287 | ||
288 | /* Update where async and sync compaction should restart */ | |
289 | if (migrate_scanner) { | |
35979ef3 DR |
290 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
291 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
292 | if (cc->mode != MIGRATE_ASYNC && |
293 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
294 | zone->compact_cached_migrate_pfn[1] = pfn; |
295 | } else { | |
35979ef3 DR |
296 | if (pfn < zone->compact_cached_free_pfn) |
297 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 298 | } |
bb13ffeb MG |
299 | } |
300 | #else | |
301 | static inline bool isolation_suitable(struct compact_control *cc, | |
302 | struct page *page) | |
303 | { | |
304 | return true; | |
305 | } | |
306 | ||
c89511ab MG |
307 | static void update_pageblock_skip(struct compact_control *cc, |
308 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 309 | bool migrate_scanner) |
bb13ffeb MG |
310 | { |
311 | } | |
312 | #endif /* CONFIG_COMPACTION */ | |
313 | ||
8b44d279 VB |
314 | /* |
315 | * Compaction requires the taking of some coarse locks that are potentially | |
316 | * very heavily contended. For async compaction, back out if the lock cannot | |
317 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
318 | * | |
319 | * Returns true if the lock is held | |
320 | * Returns false if the lock is not held and compaction should abort | |
321 | */ | |
322 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
323 | struct compact_control *cc) | |
2a1402aa | 324 | { |
8b44d279 VB |
325 | if (cc->mode == MIGRATE_ASYNC) { |
326 | if (!spin_trylock_irqsave(lock, *flags)) { | |
327 | cc->contended = COMPACT_CONTENDED_LOCK; | |
328 | return false; | |
329 | } | |
330 | } else { | |
331 | spin_lock_irqsave(lock, *flags); | |
332 | } | |
1f9efdef | 333 | |
8b44d279 | 334 | return true; |
2a1402aa MG |
335 | } |
336 | ||
c67fe375 MG |
337 | /* |
338 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
339 | * very heavily contended. The lock should be periodically unlocked to avoid |
340 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
341 | * the lock. It might also be that allowing the IRQs will result in | |
342 | * need_resched() becoming true. If scheduling is needed, async compaction | |
343 | * aborts. Sync compaction schedules. | |
344 | * Either compaction type will also abort if a fatal signal is pending. | |
345 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 346 | * |
8b44d279 VB |
347 | * Returns true if compaction should abort due to fatal signal pending, or |
348 | * async compaction due to need_resched() | |
349 | * Returns false when compaction can continue (sync compaction might have | |
350 | * scheduled) | |
c67fe375 | 351 | */ |
8b44d279 VB |
352 | static bool compact_unlock_should_abort(spinlock_t *lock, |
353 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 354 | { |
8b44d279 VB |
355 | if (*locked) { |
356 | spin_unlock_irqrestore(lock, flags); | |
357 | *locked = false; | |
358 | } | |
1f9efdef | 359 | |
8b44d279 VB |
360 | if (fatal_signal_pending(current)) { |
361 | cc->contended = COMPACT_CONTENDED_SCHED; | |
362 | return true; | |
363 | } | |
c67fe375 | 364 | |
8b44d279 | 365 | if (need_resched()) { |
e0b9daeb | 366 | if (cc->mode == MIGRATE_ASYNC) { |
8b44d279 VB |
367 | cc->contended = COMPACT_CONTENDED_SCHED; |
368 | return true; | |
c67fe375 | 369 | } |
c67fe375 | 370 | cond_resched(); |
c67fe375 MG |
371 | } |
372 | ||
8b44d279 | 373 | return false; |
c67fe375 MG |
374 | } |
375 | ||
be976572 VB |
376 | /* |
377 | * Aside from avoiding lock contention, compaction also periodically checks | |
378 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 379 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
380 | * is used where no lock is concerned. |
381 | * | |
382 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
383 | * Returns true when async compaction should abort. | |
384 | */ | |
385 | static inline bool compact_should_abort(struct compact_control *cc) | |
386 | { | |
387 | /* async compaction aborts if contended */ | |
388 | if (need_resched()) { | |
389 | if (cc->mode == MIGRATE_ASYNC) { | |
1f9efdef | 390 | cc->contended = COMPACT_CONTENDED_SCHED; |
be976572 VB |
391 | return true; |
392 | } | |
393 | ||
394 | cond_resched(); | |
395 | } | |
396 | ||
397 | return false; | |
398 | } | |
399 | ||
85aa125f | 400 | /* |
9e4be470 JM |
401 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
402 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
403 | * (even though it may still end up isolating some pages). | |
85aa125f | 404 | */ |
f40d1e42 | 405 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 406 | unsigned long *start_pfn, |
85aa125f MN |
407 | unsigned long end_pfn, |
408 | struct list_head *freelist, | |
409 | bool strict) | |
748446bb | 410 | { |
b7aba698 | 411 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 412 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 413 | unsigned long flags = 0; |
f40d1e42 | 414 | bool locked = false; |
e14c720e | 415 | unsigned long blockpfn = *start_pfn; |
748446bb | 416 | |
748446bb MG |
417 | cursor = pfn_to_page(blockpfn); |
418 | ||
f40d1e42 | 419 | /* Isolate free pages. */ |
748446bb MG |
420 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
421 | int isolated, i; | |
422 | struct page *page = cursor; | |
423 | ||
8b44d279 VB |
424 | /* |
425 | * Periodically drop the lock (if held) regardless of its | |
426 | * contention, to give chance to IRQs. Abort if fatal signal | |
427 | * pending or async compaction detects need_resched() | |
428 | */ | |
429 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
430 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
431 | &locked, cc)) | |
432 | break; | |
433 | ||
b7aba698 | 434 | nr_scanned++; |
f40d1e42 | 435 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
436 | goto isolate_fail; |
437 | ||
bb13ffeb MG |
438 | if (!valid_page) |
439 | valid_page = page; | |
f40d1e42 | 440 | if (!PageBuddy(page)) |
2af120bc | 441 | goto isolate_fail; |
f40d1e42 MG |
442 | |
443 | /* | |
69b7189f VB |
444 | * If we already hold the lock, we can skip some rechecking. |
445 | * Note that if we hold the lock now, checked_pageblock was | |
446 | * already set in some previous iteration (or strict is true), | |
447 | * so it is correct to skip the suitable migration target | |
448 | * recheck as well. | |
f40d1e42 | 449 | */ |
69b7189f VB |
450 | if (!locked) { |
451 | /* | |
452 | * The zone lock must be held to isolate freepages. | |
453 | * Unfortunately this is a very coarse lock and can be | |
454 | * heavily contended if there are parallel allocations | |
455 | * or parallel compactions. For async compaction do not | |
456 | * spin on the lock and we acquire the lock as late as | |
457 | * possible. | |
458 | */ | |
8b44d279 VB |
459 | locked = compact_trylock_irqsave(&cc->zone->lock, |
460 | &flags, cc); | |
69b7189f VB |
461 | if (!locked) |
462 | break; | |
f40d1e42 | 463 | |
69b7189f VB |
464 | /* Recheck this is a buddy page under lock */ |
465 | if (!PageBuddy(page)) | |
466 | goto isolate_fail; | |
467 | } | |
748446bb MG |
468 | |
469 | /* Found a free page, break it into order-0 pages */ | |
470 | isolated = split_free_page(page); | |
471 | total_isolated += isolated; | |
472 | for (i = 0; i < isolated; i++) { | |
473 | list_add(&page->lru, freelist); | |
474 | page++; | |
475 | } | |
476 | ||
477 | /* If a page was split, advance to the end of it */ | |
478 | if (isolated) { | |
932ff6bb JK |
479 | cc->nr_freepages += isolated; |
480 | if (!strict && | |
481 | cc->nr_migratepages <= cc->nr_freepages) { | |
482 | blockpfn += isolated; | |
483 | break; | |
484 | } | |
485 | ||
748446bb MG |
486 | blockpfn += isolated - 1; |
487 | cursor += isolated - 1; | |
2af120bc | 488 | continue; |
748446bb | 489 | } |
2af120bc LA |
490 | |
491 | isolate_fail: | |
492 | if (strict) | |
493 | break; | |
494 | else | |
495 | continue; | |
496 | ||
748446bb MG |
497 | } |
498 | ||
e34d85f0 JK |
499 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
500 | nr_scanned, total_isolated); | |
501 | ||
e14c720e VB |
502 | /* Record how far we have got within the block */ |
503 | *start_pfn = blockpfn; | |
504 | ||
f40d1e42 MG |
505 | /* |
506 | * If strict isolation is requested by CMA then check that all the | |
507 | * pages requested were isolated. If there were any failures, 0 is | |
508 | * returned and CMA will fail. | |
509 | */ | |
2af120bc | 510 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
511 | total_isolated = 0; |
512 | ||
513 | if (locked) | |
514 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
515 | ||
bb13ffeb MG |
516 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
517 | if (blockpfn == end_pfn) | |
edc2ca61 | 518 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 519 | |
010fc29a | 520 | count_compact_events(COMPACTFREE_SCANNED, nr_scanned); |
397487db | 521 | if (total_isolated) |
010fc29a | 522 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
523 | return total_isolated; |
524 | } | |
525 | ||
85aa125f MN |
526 | /** |
527 | * isolate_freepages_range() - isolate free pages. | |
528 | * @start_pfn: The first PFN to start isolating. | |
529 | * @end_pfn: The one-past-last PFN. | |
530 | * | |
531 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
532 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
533 | * undo its actions and return zero. | |
534 | * | |
535 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
536 | * (which may be greater then end_pfn if end fell in a middle of | |
537 | * a free page). | |
538 | */ | |
ff9543fd | 539 | unsigned long |
bb13ffeb MG |
540 | isolate_freepages_range(struct compact_control *cc, |
541 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 542 | { |
f40d1e42 | 543 | unsigned long isolated, pfn, block_end_pfn; |
85aa125f MN |
544 | LIST_HEAD(freelist); |
545 | ||
7d49d886 VB |
546 | pfn = start_pfn; |
547 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
548 | ||
549 | for (; pfn < end_pfn; pfn += isolated, | |
550 | block_end_pfn += pageblock_nr_pages) { | |
e14c720e VB |
551 | /* Protect pfn from changing by isolate_freepages_block */ |
552 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 553 | |
85aa125f MN |
554 | block_end_pfn = min(block_end_pfn, end_pfn); |
555 | ||
58420016 JK |
556 | /* |
557 | * pfn could pass the block_end_pfn if isolated freepage | |
558 | * is more than pageblock order. In this case, we adjust | |
559 | * scanning range to right one. | |
560 | */ | |
561 | if (pfn >= block_end_pfn) { | |
562 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
563 | block_end_pfn = min(block_end_pfn, end_pfn); | |
564 | } | |
565 | ||
7d49d886 VB |
566 | if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone)) |
567 | break; | |
568 | ||
e14c720e VB |
569 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
570 | block_end_pfn, &freelist, true); | |
85aa125f MN |
571 | |
572 | /* | |
573 | * In strict mode, isolate_freepages_block() returns 0 if | |
574 | * there are any holes in the block (ie. invalid PFNs or | |
575 | * non-free pages). | |
576 | */ | |
577 | if (!isolated) | |
578 | break; | |
579 | ||
580 | /* | |
581 | * If we managed to isolate pages, it is always (1 << n) * | |
582 | * pageblock_nr_pages for some non-negative n. (Max order | |
583 | * page may span two pageblocks). | |
584 | */ | |
585 | } | |
586 | ||
587 | /* split_free_page does not map the pages */ | |
588 | map_pages(&freelist); | |
589 | ||
590 | if (pfn < end_pfn) { | |
591 | /* Loop terminated early, cleanup. */ | |
592 | release_freepages(&freelist); | |
593 | return 0; | |
594 | } | |
595 | ||
596 | /* We don't use freelists for anything. */ | |
597 | return pfn; | |
598 | } | |
599 | ||
748446bb | 600 | /* Update the number of anon and file isolated pages in the zone */ |
edc2ca61 | 601 | static void acct_isolated(struct zone *zone, struct compact_control *cc) |
748446bb MG |
602 | { |
603 | struct page *page; | |
b9e84ac1 | 604 | unsigned int count[2] = { 0, }; |
748446bb | 605 | |
edc2ca61 VB |
606 | if (list_empty(&cc->migratepages)) |
607 | return; | |
608 | ||
b9e84ac1 MK |
609 | list_for_each_entry(page, &cc->migratepages, lru) |
610 | count[!!page_is_file_cache(page)]++; | |
748446bb | 611 | |
edc2ca61 VB |
612 | mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
613 | mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
614 | } |
615 | ||
616 | /* Similar to reclaim, but different enough that they don't share logic */ | |
617 | static bool too_many_isolated(struct zone *zone) | |
618 | { | |
bc693045 | 619 | unsigned long active, inactive, isolated; |
748446bb MG |
620 | |
621 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
622 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
623 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
624 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
625 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
626 | zone_page_state(zone, NR_ISOLATED_ANON); | |
627 | ||
bc693045 | 628 | return isolated > (inactive + active) / 2; |
748446bb MG |
629 | } |
630 | ||
2fe86e00 | 631 | /** |
edc2ca61 VB |
632 | * isolate_migratepages_block() - isolate all migrate-able pages within |
633 | * a single pageblock | |
2fe86e00 | 634 | * @cc: Compaction control structure. |
edc2ca61 VB |
635 | * @low_pfn: The first PFN to isolate |
636 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
637 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
638 | * |
639 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
640 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
641 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
642 | * first page that was not scanned (which may be both less, equal to or more | |
643 | * than end_pfn). | |
2fe86e00 | 644 | * |
edc2ca61 VB |
645 | * The pages are isolated on cc->migratepages list (not required to be empty), |
646 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
647 | * is neither read nor updated. | |
748446bb | 648 | */ |
edc2ca61 VB |
649 | static unsigned long |
650 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
651 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 652 | { |
edc2ca61 | 653 | struct zone *zone = cc->zone; |
b7aba698 | 654 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 655 | struct list_head *migratelist = &cc->migratepages; |
fa9add64 | 656 | struct lruvec *lruvec; |
b8b2d825 | 657 | unsigned long flags = 0; |
2a1402aa | 658 | bool locked = false; |
bb13ffeb | 659 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 660 | unsigned long start_pfn = low_pfn; |
748446bb | 661 | |
748446bb MG |
662 | /* |
663 | * Ensure that there are not too many pages isolated from the LRU | |
664 | * list by either parallel reclaimers or compaction. If there are, | |
665 | * delay for some time until fewer pages are isolated | |
666 | */ | |
667 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 668 | /* async migration should just abort */ |
e0b9daeb | 669 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 670 | return 0; |
f9e35b3b | 671 | |
748446bb MG |
672 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
673 | ||
674 | if (fatal_signal_pending(current)) | |
2fe86e00 | 675 | return 0; |
748446bb MG |
676 | } |
677 | ||
be976572 VB |
678 | if (compact_should_abort(cc)) |
679 | return 0; | |
aeef4b83 | 680 | |
748446bb | 681 | /* Time to isolate some pages for migration */ |
748446bb | 682 | for (; low_pfn < end_pfn; low_pfn++) { |
8b44d279 VB |
683 | /* |
684 | * Periodically drop the lock (if held) regardless of its | |
685 | * contention, to give chance to IRQs. Abort async compaction | |
686 | * if contended. | |
687 | */ | |
688 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
689 | && compact_unlock_should_abort(&zone->lru_lock, flags, | |
690 | &locked, cc)) | |
691 | break; | |
c67fe375 | 692 | |
748446bb MG |
693 | if (!pfn_valid_within(low_pfn)) |
694 | continue; | |
b7aba698 | 695 | nr_scanned++; |
748446bb | 696 | |
748446bb | 697 | page = pfn_to_page(low_pfn); |
dc908600 | 698 | |
bb13ffeb MG |
699 | if (!valid_page) |
700 | valid_page = page; | |
701 | ||
6c14466c | 702 | /* |
99c0fd5e VB |
703 | * Skip if free. We read page order here without zone lock |
704 | * which is generally unsafe, but the race window is small and | |
705 | * the worst thing that can happen is that we skip some | |
706 | * potential isolation targets. | |
6c14466c | 707 | */ |
99c0fd5e VB |
708 | if (PageBuddy(page)) { |
709 | unsigned long freepage_order = page_order_unsafe(page); | |
710 | ||
711 | /* | |
712 | * Without lock, we cannot be sure that what we got is | |
713 | * a valid page order. Consider only values in the | |
714 | * valid order range to prevent low_pfn overflow. | |
715 | */ | |
716 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
717 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 718 | continue; |
99c0fd5e | 719 | } |
748446bb | 720 | |
bf6bddf1 RA |
721 | /* |
722 | * Check may be lockless but that's ok as we recheck later. | |
723 | * It's possible to migrate LRU pages and balloon pages | |
724 | * Skip any other type of page | |
725 | */ | |
726 | if (!PageLRU(page)) { | |
727 | if (unlikely(balloon_page_movable(page))) { | |
d6d86c0a | 728 | if (balloon_page_isolate(page)) { |
bf6bddf1 | 729 | /* Successfully isolated */ |
b6c75016 | 730 | goto isolate_success; |
bf6bddf1 RA |
731 | } |
732 | } | |
bc835011 | 733 | continue; |
bf6bddf1 | 734 | } |
bc835011 AA |
735 | |
736 | /* | |
2a1402aa MG |
737 | * PageLRU is set. lru_lock normally excludes isolation |
738 | * splitting and collapsing (collapsing has already happened | |
739 | * if PageLRU is set) but the lock is not necessarily taken | |
740 | * here and it is wasteful to take it just to check transhuge. | |
741 | * Check TransHuge without lock and skip the whole pageblock if | |
742 | * it's either a transhuge or hugetlbfs page, as calling | |
743 | * compound_order() without preventing THP from splitting the | |
744 | * page underneath us may return surprising results. | |
bc835011 | 745 | */ |
2a1402aa MG |
746 | if (PageTransHuge(page)) { |
747 | if (!locked) | |
edc2ca61 VB |
748 | low_pfn = ALIGN(low_pfn + 1, |
749 | pageblock_nr_pages) - 1; | |
750 | else | |
751 | low_pfn += (1 << compound_order(page)) - 1; | |
752 | ||
2a1402aa MG |
753 | continue; |
754 | } | |
755 | ||
119d6d59 DR |
756 | /* |
757 | * Migration will fail if an anonymous page is pinned in memory, | |
758 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
759 | * admittedly racy check. | |
760 | */ | |
761 | if (!page_mapping(page) && | |
762 | page_count(page) > page_mapcount(page)) | |
763 | continue; | |
764 | ||
69b7189f VB |
765 | /* If we already hold the lock, we can skip some rechecking */ |
766 | if (!locked) { | |
8b44d279 VB |
767 | locked = compact_trylock_irqsave(&zone->lru_lock, |
768 | &flags, cc); | |
69b7189f VB |
769 | if (!locked) |
770 | break; | |
2a1402aa | 771 | |
69b7189f VB |
772 | /* Recheck PageLRU and PageTransHuge under lock */ |
773 | if (!PageLRU(page)) | |
774 | continue; | |
775 | if (PageTransHuge(page)) { | |
776 | low_pfn += (1 << compound_order(page)) - 1; | |
777 | continue; | |
778 | } | |
bc835011 AA |
779 | } |
780 | ||
fa9add64 HD |
781 | lruvec = mem_cgroup_page_lruvec(page, zone); |
782 | ||
748446bb | 783 | /* Try isolate the page */ |
edc2ca61 | 784 | if (__isolate_lru_page(page, isolate_mode) != 0) |
748446bb MG |
785 | continue; |
786 | ||
309381fe | 787 | VM_BUG_ON_PAGE(PageTransCompound(page), page); |
bc835011 | 788 | |
748446bb | 789 | /* Successfully isolated */ |
fa9add64 | 790 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
b6c75016 JK |
791 | |
792 | isolate_success: | |
748446bb | 793 | list_add(&page->lru, migratelist); |
748446bb | 794 | cc->nr_migratepages++; |
b7aba698 | 795 | nr_isolated++; |
748446bb MG |
796 | |
797 | /* Avoid isolating too much */ | |
31b8384a HD |
798 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
799 | ++low_pfn; | |
748446bb | 800 | break; |
31b8384a | 801 | } |
748446bb MG |
802 | } |
803 | ||
99c0fd5e VB |
804 | /* |
805 | * The PageBuddy() check could have potentially brought us outside | |
806 | * the range to be scanned. | |
807 | */ | |
808 | if (unlikely(low_pfn > end_pfn)) | |
809 | low_pfn = end_pfn; | |
810 | ||
c67fe375 MG |
811 | if (locked) |
812 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
748446bb | 813 | |
50b5b094 VB |
814 | /* |
815 | * Update the pageblock-skip information and cached scanner pfn, | |
816 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 817 | */ |
35979ef3 | 818 | if (low_pfn == end_pfn) |
edc2ca61 | 819 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 820 | |
e34d85f0 JK |
821 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
822 | nr_scanned, nr_isolated); | |
b7aba698 | 823 | |
010fc29a | 824 | count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned); |
397487db | 825 | if (nr_isolated) |
010fc29a | 826 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 827 | |
2fe86e00 MN |
828 | return low_pfn; |
829 | } | |
830 | ||
edc2ca61 VB |
831 | /** |
832 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
833 | * @cc: Compaction control structure. | |
834 | * @start_pfn: The first PFN to start isolating. | |
835 | * @end_pfn: The one-past-last PFN. | |
836 | * | |
837 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
838 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
839 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
840 | */ | |
841 | unsigned long | |
842 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
843 | unsigned long end_pfn) | |
844 | { | |
845 | unsigned long pfn, block_end_pfn; | |
846 | ||
847 | /* Scan block by block. First and last block may be incomplete */ | |
848 | pfn = start_pfn; | |
849 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
850 | ||
851 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
852 | block_end_pfn += pageblock_nr_pages) { | |
853 | ||
854 | block_end_pfn = min(block_end_pfn, end_pfn); | |
855 | ||
7d49d886 | 856 | if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone)) |
edc2ca61 VB |
857 | continue; |
858 | ||
859 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
860 | ISOLATE_UNEVICTABLE); | |
861 | ||
862 | /* | |
863 | * In case of fatal failure, release everything that might | |
864 | * have been isolated in the previous iteration, and signal | |
865 | * the failure back to caller. | |
866 | */ | |
867 | if (!pfn) { | |
868 | putback_movable_pages(&cc->migratepages); | |
869 | cc->nr_migratepages = 0; | |
870 | break; | |
871 | } | |
6ea41c0c JK |
872 | |
873 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
874 | break; | |
edc2ca61 VB |
875 | } |
876 | acct_isolated(cc->zone, cc); | |
877 | ||
878 | return pfn; | |
879 | } | |
880 | ||
ff9543fd MN |
881 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
882 | #ifdef CONFIG_COMPACTION | |
018e9a49 AM |
883 | |
884 | /* Returns true if the page is within a block suitable for migration to */ | |
885 | static bool suitable_migration_target(struct page *page) | |
886 | { | |
887 | /* If the page is a large free page, then disallow migration */ | |
888 | if (PageBuddy(page)) { | |
889 | /* | |
890 | * We are checking page_order without zone->lock taken. But | |
891 | * the only small danger is that we skip a potentially suitable | |
892 | * pageblock, so it's not worth to check order for valid range. | |
893 | */ | |
894 | if (page_order_unsafe(page) >= pageblock_order) | |
895 | return false; | |
896 | } | |
897 | ||
898 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ | |
899 | if (migrate_async_suitable(get_pageblock_migratetype(page))) | |
900 | return true; | |
901 | ||
902 | /* Otherwise skip the block */ | |
903 | return false; | |
904 | } | |
905 | ||
f2849aa0 VB |
906 | /* |
907 | * Test whether the free scanner has reached the same or lower pageblock than | |
908 | * the migration scanner, and compaction should thus terminate. | |
909 | */ | |
910 | static inline bool compact_scanners_met(struct compact_control *cc) | |
911 | { | |
912 | return (cc->free_pfn >> pageblock_order) | |
913 | <= (cc->migrate_pfn >> pageblock_order); | |
914 | } | |
915 | ||
2fe86e00 | 916 | /* |
ff9543fd MN |
917 | * Based on information in the current compact_control, find blocks |
918 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 919 | */ |
edc2ca61 | 920 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 921 | { |
edc2ca61 | 922 | struct zone *zone = cc->zone; |
ff9543fd | 923 | struct page *page; |
c96b9e50 | 924 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 925 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
926 | unsigned long block_end_pfn; /* end of current pageblock */ |
927 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 928 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 929 | |
ff9543fd MN |
930 | /* |
931 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 932 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
933 | * zone when isolating for the first time. For looping we also need |
934 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
935 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
936 | * For ending point, take care when isolating in last pageblock of a | |
937 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
938 | * The low boundary is the end of the pageblock the migration scanner |
939 | * is using. | |
ff9543fd | 940 | */ |
e14c720e | 941 | isolate_start_pfn = cc->free_pfn; |
c96b9e50 VB |
942 | block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1); |
943 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, | |
944 | zone_end_pfn(zone)); | |
7ed695e0 | 945 | low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages); |
2fe86e00 | 946 | |
ff9543fd MN |
947 | /* |
948 | * Isolate free pages until enough are available to migrate the | |
949 | * pages on cc->migratepages. We stop searching if the migrate | |
950 | * and free page scanners meet or enough free pages are isolated. | |
951 | */ | |
f5f61a32 | 952 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 953 | block_end_pfn = block_start_pfn, |
e14c720e VB |
954 | block_start_pfn -= pageblock_nr_pages, |
955 | isolate_start_pfn = block_start_pfn) { | |
2fe86e00 | 956 | |
f6ea3adb DR |
957 | /* |
958 | * This can iterate a massively long zone without finding any | |
959 | * suitable migration targets, so periodically check if we need | |
be976572 | 960 | * to schedule, or even abort async compaction. |
f6ea3adb | 961 | */ |
be976572 VB |
962 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
963 | && compact_should_abort(cc)) | |
964 | break; | |
f6ea3adb | 965 | |
7d49d886 VB |
966 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
967 | zone); | |
968 | if (!page) | |
ff9543fd MN |
969 | continue; |
970 | ||
971 | /* Check the block is suitable for migration */ | |
68e3e926 | 972 | if (!suitable_migration_target(page)) |
ff9543fd | 973 | continue; |
68e3e926 | 974 | |
bb13ffeb MG |
975 | /* If isolation recently failed, do not retry */ |
976 | if (!isolation_suitable(cc, page)) | |
977 | continue; | |
978 | ||
e14c720e | 979 | /* Found a block suitable for isolating free pages from. */ |
932ff6bb | 980 | isolate_freepages_block(cc, &isolate_start_pfn, |
c96b9e50 | 981 | block_end_pfn, freelist, false); |
ff9543fd | 982 | |
e14c720e | 983 | /* |
f5f61a32 VB |
984 | * If we isolated enough freepages, or aborted due to async |
985 | * compaction being contended, terminate the loop. | |
e14c720e VB |
986 | * Remember where the free scanner should restart next time, |
987 | * which is where isolate_freepages_block() left off. | |
988 | * But if it scanned the whole pageblock, isolate_start_pfn | |
989 | * now points at block_end_pfn, which is the start of the next | |
990 | * pageblock. | |
991 | * In that case we will however want to restart at the start | |
992 | * of the previous pageblock. | |
993 | */ | |
f5f61a32 VB |
994 | if ((cc->nr_freepages >= cc->nr_migratepages) |
995 | || cc->contended) { | |
996 | if (isolate_start_pfn >= block_end_pfn) | |
997 | isolate_start_pfn = | |
998 | block_start_pfn - pageblock_nr_pages; | |
be976572 | 999 | break; |
f5f61a32 VB |
1000 | } else { |
1001 | /* | |
1002 | * isolate_freepages_block() should not terminate | |
1003 | * prematurely unless contended, or isolated enough | |
1004 | */ | |
1005 | VM_BUG_ON(isolate_start_pfn < block_end_pfn); | |
1006 | } | |
ff9543fd MN |
1007 | } |
1008 | ||
1009 | /* split_free_page does not map the pages */ | |
1010 | map_pages(freelist); | |
1011 | ||
7ed695e0 | 1012 | /* |
f5f61a32 VB |
1013 | * Record where the free scanner will restart next time. Either we |
1014 | * broke from the loop and set isolate_start_pfn based on the last | |
1015 | * call to isolate_freepages_block(), or we met the migration scanner | |
1016 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1017 | */ |
f5f61a32 | 1018 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1019 | } |
1020 | ||
1021 | /* | |
1022 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1023 | * from the isolated freelists in the block we are migrating to. | |
1024 | */ | |
1025 | static struct page *compaction_alloc(struct page *migratepage, | |
1026 | unsigned long data, | |
1027 | int **result) | |
1028 | { | |
1029 | struct compact_control *cc = (struct compact_control *)data; | |
1030 | struct page *freepage; | |
1031 | ||
be976572 VB |
1032 | /* |
1033 | * Isolate free pages if necessary, and if we are not aborting due to | |
1034 | * contention. | |
1035 | */ | |
748446bb | 1036 | if (list_empty(&cc->freepages)) { |
be976572 | 1037 | if (!cc->contended) |
edc2ca61 | 1038 | isolate_freepages(cc); |
748446bb MG |
1039 | |
1040 | if (list_empty(&cc->freepages)) | |
1041 | return NULL; | |
1042 | } | |
1043 | ||
1044 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1045 | list_del(&freepage->lru); | |
1046 | cc->nr_freepages--; | |
1047 | ||
1048 | return freepage; | |
1049 | } | |
1050 | ||
1051 | /* | |
d53aea3d DR |
1052 | * This is a migrate-callback that "frees" freepages back to the isolated |
1053 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1054 | * special handling needed for NUMA. | |
1055 | */ | |
1056 | static void compaction_free(struct page *page, unsigned long data) | |
1057 | { | |
1058 | struct compact_control *cc = (struct compact_control *)data; | |
1059 | ||
1060 | list_add(&page->lru, &cc->freepages); | |
1061 | cc->nr_freepages++; | |
1062 | } | |
1063 | ||
ff9543fd MN |
1064 | /* possible outcome of isolate_migratepages */ |
1065 | typedef enum { | |
1066 | ISOLATE_ABORT, /* Abort compaction now */ | |
1067 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1068 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1069 | } isolate_migrate_t; | |
1070 | ||
5bbe3547 EM |
1071 | /* |
1072 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1073 | * compactable pages. | |
1074 | */ | |
1075 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1076 | ||
ff9543fd | 1077 | /* |
edc2ca61 VB |
1078 | * Isolate all pages that can be migrated from the first suitable block, |
1079 | * starting at the block pointed to by the migrate scanner pfn within | |
1080 | * compact_control. | |
ff9543fd MN |
1081 | */ |
1082 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1083 | struct compact_control *cc) | |
1084 | { | |
1085 | unsigned long low_pfn, end_pfn; | |
edc2ca61 VB |
1086 | struct page *page; |
1087 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1088 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
edc2ca61 | 1089 | (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1090 | |
edc2ca61 VB |
1091 | /* |
1092 | * Start at where we last stopped, or beginning of the zone as | |
1093 | * initialized by compact_zone() | |
1094 | */ | |
1095 | low_pfn = cc->migrate_pfn; | |
ff9543fd MN |
1096 | |
1097 | /* Only scan within a pageblock boundary */ | |
a9aacbcc | 1098 | end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages); |
ff9543fd | 1099 | |
edc2ca61 VB |
1100 | /* |
1101 | * Iterate over whole pageblocks until we find the first suitable. | |
1102 | * Do not cross the free scanner. | |
1103 | */ | |
1104 | for (; end_pfn <= cc->free_pfn; | |
1105 | low_pfn = end_pfn, end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1106 | |
edc2ca61 VB |
1107 | /* |
1108 | * This can potentially iterate a massively long zone with | |
1109 | * many pageblocks unsuitable, so periodically check if we | |
1110 | * need to schedule, or even abort async compaction. | |
1111 | */ | |
1112 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1113 | && compact_should_abort(cc)) | |
1114 | break; | |
ff9543fd | 1115 | |
7d49d886 VB |
1116 | page = pageblock_pfn_to_page(low_pfn, end_pfn, zone); |
1117 | if (!page) | |
edc2ca61 VB |
1118 | continue; |
1119 | ||
edc2ca61 VB |
1120 | /* If isolation recently failed, do not retry */ |
1121 | if (!isolation_suitable(cc, page)) | |
1122 | continue; | |
1123 | ||
1124 | /* | |
1125 | * For async compaction, also only scan in MOVABLE blocks. | |
1126 | * Async compaction is optimistic to see if the minimum amount | |
1127 | * of work satisfies the allocation. | |
1128 | */ | |
1129 | if (cc->mode == MIGRATE_ASYNC && | |
1130 | !migrate_async_suitable(get_pageblock_migratetype(page))) | |
1131 | continue; | |
1132 | ||
1133 | /* Perform the isolation */ | |
1134 | low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn, | |
1135 | isolate_mode); | |
1136 | ||
ff59909a HD |
1137 | if (!low_pfn || cc->contended) { |
1138 | acct_isolated(zone, cc); | |
edc2ca61 | 1139 | return ISOLATE_ABORT; |
ff59909a | 1140 | } |
edc2ca61 VB |
1141 | |
1142 | /* | |
1143 | * Either we isolated something and proceed with migration. Or | |
1144 | * we failed and compact_zone should decide if we should | |
1145 | * continue or not. | |
1146 | */ | |
1147 | break; | |
1148 | } | |
1149 | ||
1150 | acct_isolated(zone, cc); | |
f2849aa0 VB |
1151 | /* Record where migration scanner will be restarted. */ |
1152 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1153 | |
edc2ca61 | 1154 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1155 | } |
1156 | ||
837d026d | 1157 | static int __compact_finished(struct zone *zone, struct compact_control *cc, |
6d7ce559 | 1158 | const int migratetype) |
748446bb | 1159 | { |
8fb74b9f | 1160 | unsigned int order; |
5a03b051 | 1161 | unsigned long watermark; |
56de7263 | 1162 | |
be976572 | 1163 | if (cc->contended || fatal_signal_pending(current)) |
748446bb MG |
1164 | return COMPACT_PARTIAL; |
1165 | ||
753341a4 | 1166 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1167 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1168 | /* Let the next compaction start anew. */ |
02333641 | 1169 | reset_cached_positions(zone); |
55b7c4c9 | 1170 | |
62997027 MG |
1171 | /* |
1172 | * Mark that the PG_migrate_skip information should be cleared | |
1173 | * by kswapd when it goes to sleep. kswapd does not set the | |
1174 | * flag itself as the decision to be clear should be directly | |
1175 | * based on an allocation request. | |
1176 | */ | |
1177 | if (!current_is_kswapd()) | |
1178 | zone->compact_blockskip_flush = true; | |
1179 | ||
748446bb | 1180 | return COMPACT_COMPLETE; |
bb13ffeb | 1181 | } |
748446bb | 1182 | |
82478fb7 JW |
1183 | /* |
1184 | * order == -1 is expected when compacting via | |
1185 | * /proc/sys/vm/compact_memory | |
1186 | */ | |
56de7263 MG |
1187 | if (cc->order == -1) |
1188 | return COMPACT_CONTINUE; | |
1189 | ||
3957c776 MH |
1190 | /* Compaction run is not finished if the watermark is not met */ |
1191 | watermark = low_wmark_pages(zone); | |
3957c776 | 1192 | |
ebff3980 VB |
1193 | if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx, |
1194 | cc->alloc_flags)) | |
3957c776 MH |
1195 | return COMPACT_CONTINUE; |
1196 | ||
56de7263 | 1197 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f MG |
1198 | for (order = cc->order; order < MAX_ORDER; order++) { |
1199 | struct free_area *area = &zone->free_area[order]; | |
2149cdae | 1200 | bool can_steal; |
8fb74b9f MG |
1201 | |
1202 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1203 | if (!list_empty(&area->free_list[migratetype])) |
8fb74b9f MG |
1204 | return COMPACT_PARTIAL; |
1205 | ||
2149cdae JK |
1206 | #ifdef CONFIG_CMA |
1207 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1208 | if (migratetype == MIGRATE_MOVABLE && | |
1209 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
1210 | return COMPACT_PARTIAL; | |
1211 | #endif | |
1212 | /* | |
1213 | * Job done if allocation would steal freepages from | |
1214 | * other migratetype buddy lists. | |
1215 | */ | |
1216 | if (find_suitable_fallback(area, order, migratetype, | |
1217 | true, &can_steal) != -1) | |
56de7263 MG |
1218 | return COMPACT_PARTIAL; |
1219 | } | |
1220 | ||
837d026d JK |
1221 | return COMPACT_NO_SUITABLE_PAGE; |
1222 | } | |
1223 | ||
1224 | static int compact_finished(struct zone *zone, struct compact_control *cc, | |
1225 | const int migratetype) | |
1226 | { | |
1227 | int ret; | |
1228 | ||
1229 | ret = __compact_finished(zone, cc, migratetype); | |
1230 | trace_mm_compaction_finished(zone, cc->order, ret); | |
1231 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1232 | ret = COMPACT_CONTINUE; | |
1233 | ||
1234 | return ret; | |
748446bb MG |
1235 | } |
1236 | ||
3e7d3449 MG |
1237 | /* |
1238 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1239 | * Returns | |
1240 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
1241 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
1242 | * COMPACT_CONTINUE - If compaction should run now | |
1243 | */ | |
837d026d | 1244 | static unsigned long __compaction_suitable(struct zone *zone, int order, |
ebff3980 | 1245 | int alloc_flags, int classzone_idx) |
3e7d3449 MG |
1246 | { |
1247 | int fragindex; | |
1248 | unsigned long watermark; | |
1249 | ||
3957c776 MH |
1250 | /* |
1251 | * order == -1 is expected when compacting via | |
1252 | * /proc/sys/vm/compact_memory | |
1253 | */ | |
1254 | if (order == -1) | |
1255 | return COMPACT_CONTINUE; | |
1256 | ||
ebff3980 VB |
1257 | watermark = low_wmark_pages(zone); |
1258 | /* | |
1259 | * If watermarks for high-order allocation are already met, there | |
1260 | * should be no need for compaction at all. | |
1261 | */ | |
1262 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1263 | alloc_flags)) | |
1264 | return COMPACT_PARTIAL; | |
1265 | ||
3e7d3449 MG |
1266 | /* |
1267 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
1268 | * This is because during migration, copies of pages need to be | |
1269 | * allocated and for a short time, the footprint is higher | |
1270 | */ | |
ebff3980 VB |
1271 | watermark += (2UL << order); |
1272 | if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags)) | |
3e7d3449 MG |
1273 | return COMPACT_SKIPPED; |
1274 | ||
1275 | /* | |
1276 | * fragmentation index determines if allocation failures are due to | |
1277 | * low memory or external fragmentation | |
1278 | * | |
ebff3980 VB |
1279 | * index of -1000 would imply allocations might succeed depending on |
1280 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1281 | * index towards 0 implies failure is due to lack of memory |
1282 | * index towards 1000 implies failure is due to fragmentation | |
1283 | * | |
1284 | * Only compact if a failure would be due to fragmentation. | |
1285 | */ | |
1286 | fragindex = fragmentation_index(zone, order); | |
1287 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
837d026d | 1288 | return COMPACT_NOT_SUITABLE_ZONE; |
3e7d3449 | 1289 | |
3e7d3449 MG |
1290 | return COMPACT_CONTINUE; |
1291 | } | |
1292 | ||
837d026d JK |
1293 | unsigned long compaction_suitable(struct zone *zone, int order, |
1294 | int alloc_flags, int classzone_idx) | |
1295 | { | |
1296 | unsigned long ret; | |
1297 | ||
1298 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx); | |
1299 | trace_mm_compaction_suitable(zone, order, ret); | |
1300 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1301 | ret = COMPACT_SKIPPED; | |
1302 | ||
1303 | return ret; | |
1304 | } | |
1305 | ||
748446bb MG |
1306 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
1307 | { | |
1308 | int ret; | |
c89511ab | 1309 | unsigned long start_pfn = zone->zone_start_pfn; |
108bcc96 | 1310 | unsigned long end_pfn = zone_end_pfn(zone); |
6d7ce559 | 1311 | const int migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
e0b9daeb | 1312 | const bool sync = cc->mode != MIGRATE_ASYNC; |
fdaf7f5c | 1313 | unsigned long last_migrated_pfn = 0; |
748446bb | 1314 | |
ebff3980 VB |
1315 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, |
1316 | cc->classzone_idx); | |
3e7d3449 MG |
1317 | switch (ret) { |
1318 | case COMPACT_PARTIAL: | |
1319 | case COMPACT_SKIPPED: | |
1320 | /* Compaction is likely to fail */ | |
1321 | return ret; | |
1322 | case COMPACT_CONTINUE: | |
1323 | /* Fall through to compaction */ | |
1324 | ; | |
1325 | } | |
1326 | ||
d3132e4b VB |
1327 | /* |
1328 | * Clear pageblock skip if there were failures recently and compaction | |
1329 | * is about to be retried after being deferred. kswapd does not do | |
1330 | * this reset as it'll reset the cached information when going to sleep. | |
1331 | */ | |
1332 | if (compaction_restarting(zone, cc->order) && !current_is_kswapd()) | |
1333 | __reset_isolation_suitable(zone); | |
1334 | ||
c89511ab MG |
1335 | /* |
1336 | * Setup to move all movable pages to the end of the zone. Used cached | |
1337 | * information on where the scanners should start but check that it | |
1338 | * is initialised by ensuring the values are within zone boundaries. | |
1339 | */ | |
e0b9daeb | 1340 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; |
c89511ab MG |
1341 | cc->free_pfn = zone->compact_cached_free_pfn; |
1342 | if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) { | |
1343 | cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1); | |
1344 | zone->compact_cached_free_pfn = cc->free_pfn; | |
1345 | } | |
1346 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) { | |
1347 | cc->migrate_pfn = start_pfn; | |
35979ef3 DR |
1348 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1349 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
c89511ab | 1350 | } |
748446bb | 1351 | |
16c4a097 JK |
1352 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1353 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1354 | |
748446bb MG |
1355 | migrate_prep_local(); |
1356 | ||
6d7ce559 DR |
1357 | while ((ret = compact_finished(zone, cc, migratetype)) == |
1358 | COMPACT_CONTINUE) { | |
9d502c1c | 1359 | int err; |
fdaf7f5c | 1360 | unsigned long isolate_start_pfn = cc->migrate_pfn; |
748446bb | 1361 | |
f9e35b3b MG |
1362 | switch (isolate_migratepages(zone, cc)) { |
1363 | case ISOLATE_ABORT: | |
1364 | ret = COMPACT_PARTIAL; | |
5733c7d1 | 1365 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1366 | cc->nr_migratepages = 0; |
f9e35b3b MG |
1367 | goto out; |
1368 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1369 | /* |
1370 | * We haven't isolated and migrated anything, but | |
1371 | * there might still be unflushed migrations from | |
1372 | * previous cc->order aligned block. | |
1373 | */ | |
1374 | goto check_drain; | |
f9e35b3b MG |
1375 | case ISOLATE_SUCCESS: |
1376 | ; | |
1377 | } | |
748446bb | 1378 | |
d53aea3d | 1379 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1380 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1381 | MR_COMPACTION); |
748446bb | 1382 | |
f8c9301f VB |
1383 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1384 | &cc->migratepages); | |
748446bb | 1385 | |
f8c9301f VB |
1386 | /* All pages were either migrated or will be released */ |
1387 | cc->nr_migratepages = 0; | |
9d502c1c | 1388 | if (err) { |
5733c7d1 | 1389 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1390 | /* |
1391 | * migrate_pages() may return -ENOMEM when scanners meet | |
1392 | * and we want compact_finished() to detect it | |
1393 | */ | |
f2849aa0 | 1394 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
4bf2bba3 DR |
1395 | ret = COMPACT_PARTIAL; |
1396 | goto out; | |
1397 | } | |
748446bb | 1398 | } |
fdaf7f5c VB |
1399 | |
1400 | /* | |
1401 | * Record where we could have freed pages by migration and not | |
1402 | * yet flushed them to buddy allocator. We use the pfn that | |
1403 | * isolate_migratepages() started from in this loop iteration | |
1404 | * - this is the lowest page that could have been isolated and | |
1405 | * then freed by migration. | |
1406 | */ | |
1407 | if (!last_migrated_pfn) | |
1408 | last_migrated_pfn = isolate_start_pfn; | |
1409 | ||
1410 | check_drain: | |
1411 | /* | |
1412 | * Has the migration scanner moved away from the previous | |
1413 | * cc->order aligned block where we migrated from? If yes, | |
1414 | * flush the pages that were freed, so that they can merge and | |
1415 | * compact_finished() can detect immediately if allocation | |
1416 | * would succeed. | |
1417 | */ | |
1418 | if (cc->order > 0 && last_migrated_pfn) { | |
1419 | int cpu; | |
1420 | unsigned long current_block_start = | |
1421 | cc->migrate_pfn & ~((1UL << cc->order) - 1); | |
1422 | ||
1423 | if (last_migrated_pfn < current_block_start) { | |
1424 | cpu = get_cpu(); | |
1425 | lru_add_drain_cpu(cpu); | |
1426 | drain_local_pages(zone); | |
1427 | put_cpu(); | |
1428 | /* No more flushing until we migrate again */ | |
1429 | last_migrated_pfn = 0; | |
1430 | } | |
1431 | } | |
1432 | ||
748446bb MG |
1433 | } |
1434 | ||
f9e35b3b | 1435 | out: |
6bace090 VB |
1436 | /* |
1437 | * Release free pages and update where the free scanner should restart, | |
1438 | * so we don't leave any returned pages behind in the next attempt. | |
1439 | */ | |
1440 | if (cc->nr_freepages > 0) { | |
1441 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1442 | ||
1443 | cc->nr_freepages = 0; | |
1444 | VM_BUG_ON(free_pfn == 0); | |
1445 | /* The cached pfn is always the first in a pageblock */ | |
1446 | free_pfn &= ~(pageblock_nr_pages-1); | |
1447 | /* | |
1448 | * Only go back, not forward. The cached pfn might have been | |
1449 | * already reset to zone end in compact_finished() | |
1450 | */ | |
1451 | if (free_pfn > zone->compact_cached_free_pfn) | |
1452 | zone->compact_cached_free_pfn = free_pfn; | |
1453 | } | |
748446bb | 1454 | |
16c4a097 JK |
1455 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1456 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1457 | |
748446bb MG |
1458 | return ret; |
1459 | } | |
76ab0f53 | 1460 | |
e0b9daeb | 1461 | static unsigned long compact_zone_order(struct zone *zone, int order, |
ebff3980 VB |
1462 | gfp_t gfp_mask, enum migrate_mode mode, int *contended, |
1463 | int alloc_flags, int classzone_idx) | |
56de7263 | 1464 | { |
e64c5237 | 1465 | unsigned long ret; |
56de7263 MG |
1466 | struct compact_control cc = { |
1467 | .nr_freepages = 0, | |
1468 | .nr_migratepages = 0, | |
1469 | .order = order, | |
6d7ce559 | 1470 | .gfp_mask = gfp_mask, |
56de7263 | 1471 | .zone = zone, |
e0b9daeb | 1472 | .mode = mode, |
ebff3980 VB |
1473 | .alloc_flags = alloc_flags, |
1474 | .classzone_idx = classzone_idx, | |
56de7263 MG |
1475 | }; |
1476 | INIT_LIST_HEAD(&cc.freepages); | |
1477 | INIT_LIST_HEAD(&cc.migratepages); | |
1478 | ||
e64c5237 SL |
1479 | ret = compact_zone(zone, &cc); |
1480 | ||
1481 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1482 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1483 | ||
1484 | *contended = cc.contended; | |
1485 | return ret; | |
56de7263 MG |
1486 | } |
1487 | ||
5e771905 MG |
1488 | int sysctl_extfrag_threshold = 500; |
1489 | ||
56de7263 MG |
1490 | /** |
1491 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1492 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1493 | * @order: The order of the current allocation |
1494 | * @alloc_flags: The allocation flags of the current allocation | |
1495 | * @ac: The context of current allocation | |
e0b9daeb | 1496 | * @mode: The migration mode for async, sync light, or sync migration |
1f9efdef VB |
1497 | * @contended: Return value that determines if compaction was aborted due to |
1498 | * need_resched() or lock contention | |
56de7263 MG |
1499 | * |
1500 | * This is the main entry point for direct page compaction. | |
1501 | */ | |
1a6d53a1 VB |
1502 | unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
1503 | int alloc_flags, const struct alloc_context *ac, | |
1504 | enum migrate_mode mode, int *contended) | |
56de7263 | 1505 | { |
56de7263 MG |
1506 | int may_enter_fs = gfp_mask & __GFP_FS; |
1507 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
1508 | struct zoneref *z; |
1509 | struct zone *zone; | |
53853e2d | 1510 | int rc = COMPACT_DEFERRED; |
1f9efdef VB |
1511 | int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */ |
1512 | ||
1513 | *contended = COMPACT_CONTENDED_NONE; | |
56de7263 | 1514 | |
4ffb6335 | 1515 | /* Check if the GFP flags allow compaction */ |
c5a73c3d | 1516 | if (!order || !may_enter_fs || !may_perform_io) |
53853e2d | 1517 | return COMPACT_SKIPPED; |
56de7263 | 1518 | |
837d026d JK |
1519 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode); |
1520 | ||
56de7263 | 1521 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1522 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1523 | ac->nodemask) { | |
56de7263 | 1524 | int status; |
1f9efdef | 1525 | int zone_contended; |
56de7263 | 1526 | |
53853e2d VB |
1527 | if (compaction_deferred(zone, order)) |
1528 | continue; | |
1529 | ||
e0b9daeb | 1530 | status = compact_zone_order(zone, order, gfp_mask, mode, |
1a6d53a1 VB |
1531 | &zone_contended, alloc_flags, |
1532 | ac->classzone_idx); | |
56de7263 | 1533 | rc = max(status, rc); |
1f9efdef VB |
1534 | /* |
1535 | * It takes at least one zone that wasn't lock contended | |
1536 | * to clear all_zones_contended. | |
1537 | */ | |
1538 | all_zones_contended &= zone_contended; | |
56de7263 | 1539 | |
3e7d3449 | 1540 | /* If a normal allocation would succeed, stop compacting */ |
ebff3980 | 1541 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), |
1a6d53a1 | 1542 | ac->classzone_idx, alloc_flags)) { |
53853e2d VB |
1543 | /* |
1544 | * We think the allocation will succeed in this zone, | |
1545 | * but it is not certain, hence the false. The caller | |
1546 | * will repeat this with true if allocation indeed | |
1547 | * succeeds in this zone. | |
1548 | */ | |
1549 | compaction_defer_reset(zone, order, false); | |
1f9efdef VB |
1550 | /* |
1551 | * It is possible that async compaction aborted due to | |
1552 | * need_resched() and the watermarks were ok thanks to | |
1553 | * somebody else freeing memory. The allocation can | |
1554 | * however still fail so we better signal the | |
1555 | * need_resched() contention anyway (this will not | |
1556 | * prevent the allocation attempt). | |
1557 | */ | |
1558 | if (zone_contended == COMPACT_CONTENDED_SCHED) | |
1559 | *contended = COMPACT_CONTENDED_SCHED; | |
1560 | ||
1561 | goto break_loop; | |
1562 | } | |
1563 | ||
f8669795 | 1564 | if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) { |
53853e2d VB |
1565 | /* |
1566 | * We think that allocation won't succeed in this zone | |
1567 | * so we defer compaction there. If it ends up | |
1568 | * succeeding after all, it will be reset. | |
1569 | */ | |
1570 | defer_compaction(zone, order); | |
1571 | } | |
1f9efdef VB |
1572 | |
1573 | /* | |
1574 | * We might have stopped compacting due to need_resched() in | |
1575 | * async compaction, or due to a fatal signal detected. In that | |
1576 | * case do not try further zones and signal need_resched() | |
1577 | * contention. | |
1578 | */ | |
1579 | if ((zone_contended == COMPACT_CONTENDED_SCHED) | |
1580 | || fatal_signal_pending(current)) { | |
1581 | *contended = COMPACT_CONTENDED_SCHED; | |
1582 | goto break_loop; | |
1583 | } | |
1584 | ||
1585 | continue; | |
1586 | break_loop: | |
1587 | /* | |
1588 | * We might not have tried all the zones, so be conservative | |
1589 | * and assume they are not all lock contended. | |
1590 | */ | |
1591 | all_zones_contended = 0; | |
1592 | break; | |
56de7263 MG |
1593 | } |
1594 | ||
1f9efdef VB |
1595 | /* |
1596 | * If at least one zone wasn't deferred or skipped, we report if all | |
1597 | * zones that were tried were lock contended. | |
1598 | */ | |
1599 | if (rc > COMPACT_SKIPPED && all_zones_contended) | |
1600 | *contended = COMPACT_CONTENDED_LOCK; | |
1601 | ||
56de7263 MG |
1602 | return rc; |
1603 | } | |
1604 | ||
1605 | ||
76ab0f53 | 1606 | /* Compact all zones within a node */ |
7103f16d | 1607 | static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
1608 | { |
1609 | int zoneid; | |
76ab0f53 MG |
1610 | struct zone *zone; |
1611 | ||
76ab0f53 | 1612 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1613 | |
1614 | zone = &pgdat->node_zones[zoneid]; | |
1615 | if (!populated_zone(zone)) | |
1616 | continue; | |
1617 | ||
7be62de9 RR |
1618 | cc->nr_freepages = 0; |
1619 | cc->nr_migratepages = 0; | |
1620 | cc->zone = zone; | |
1621 | INIT_LIST_HEAD(&cc->freepages); | |
1622 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 1623 | |
195b0c60 GK |
1624 | /* |
1625 | * When called via /proc/sys/vm/compact_memory | |
1626 | * this makes sure we compact the whole zone regardless of | |
1627 | * cached scanner positions. | |
1628 | */ | |
1629 | if (cc->order == -1) | |
1630 | __reset_isolation_suitable(zone); | |
1631 | ||
aad6ec37 | 1632 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
7be62de9 | 1633 | compact_zone(zone, cc); |
76ab0f53 | 1634 | |
aff62249 | 1635 | if (cc->order > 0) { |
de6c60a6 VB |
1636 | if (zone_watermark_ok(zone, cc->order, |
1637 | low_wmark_pages(zone), 0, 0)) | |
1638 | compaction_defer_reset(zone, cc->order, false); | |
aff62249 RR |
1639 | } |
1640 | ||
7be62de9 RR |
1641 | VM_BUG_ON(!list_empty(&cc->freepages)); |
1642 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 | 1643 | } |
76ab0f53 MG |
1644 | } |
1645 | ||
7103f16d | 1646 | void compact_pgdat(pg_data_t *pgdat, int order) |
7be62de9 RR |
1647 | { |
1648 | struct compact_control cc = { | |
1649 | .order = order, | |
e0b9daeb | 1650 | .mode = MIGRATE_ASYNC, |
7be62de9 RR |
1651 | }; |
1652 | ||
3a7200af MG |
1653 | if (!order) |
1654 | return; | |
1655 | ||
7103f16d | 1656 | __compact_pgdat(pgdat, &cc); |
7be62de9 RR |
1657 | } |
1658 | ||
7103f16d | 1659 | static void compact_node(int nid) |
7be62de9 | 1660 | { |
7be62de9 RR |
1661 | struct compact_control cc = { |
1662 | .order = -1, | |
e0b9daeb | 1663 | .mode = MIGRATE_SYNC, |
91ca9186 | 1664 | .ignore_skip_hint = true, |
7be62de9 RR |
1665 | }; |
1666 | ||
7103f16d | 1667 | __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
1668 | } |
1669 | ||
76ab0f53 | 1670 | /* Compact all nodes in the system */ |
7964c06d | 1671 | static void compact_nodes(void) |
76ab0f53 MG |
1672 | { |
1673 | int nid; | |
1674 | ||
8575ec29 HD |
1675 | /* Flush pending updates to the LRU lists */ |
1676 | lru_add_drain_all(); | |
1677 | ||
76ab0f53 MG |
1678 | for_each_online_node(nid) |
1679 | compact_node(nid); | |
76ab0f53 MG |
1680 | } |
1681 | ||
1682 | /* The written value is actually unused, all memory is compacted */ | |
1683 | int sysctl_compact_memory; | |
1684 | ||
1685 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
1686 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
1687 | void __user *buffer, size_t *length, loff_t *ppos) | |
1688 | { | |
1689 | if (write) | |
7964c06d | 1690 | compact_nodes(); |
76ab0f53 MG |
1691 | |
1692 | return 0; | |
1693 | } | |
ed4a6d7f | 1694 | |
5e771905 MG |
1695 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
1696 | void __user *buffer, size_t *length, loff_t *ppos) | |
1697 | { | |
1698 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1699 | ||
1700 | return 0; | |
1701 | } | |
1702 | ||
ed4a6d7f | 1703 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
74e77fb9 | 1704 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1705 | struct device_attribute *attr, |
ed4a6d7f MG |
1706 | const char *buf, size_t count) |
1707 | { | |
8575ec29 HD |
1708 | int nid = dev->id; |
1709 | ||
1710 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1711 | /* Flush pending updates to the LRU lists */ | |
1712 | lru_add_drain_all(); | |
1713 | ||
1714 | compact_node(nid); | |
1715 | } | |
ed4a6d7f MG |
1716 | |
1717 | return count; | |
1718 | } | |
10fbcf4c | 1719 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1720 | |
1721 | int compaction_register_node(struct node *node) | |
1722 | { | |
10fbcf4c | 1723 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1724 | } |
1725 | ||
1726 | void compaction_unregister_node(struct node *node) | |
1727 | { | |
10fbcf4c | 1728 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1729 | } |
1730 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd MN |
1731 | |
1732 | #endif /* CONFIG_COMPACTION */ |