]>
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 | */ | |
698b1b30 | 10 | #include <linux/cpu.h> |
748446bb MG |
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> | |
76ab0f53 | 16 | #include <linux/sysctl.h> |
ed4a6d7f | 17 | #include <linux/sysfs.h> |
bf6bddf1 | 18 | #include <linux/balloon_compaction.h> |
194159fb | 19 | #include <linux/page-isolation.h> |
b8c73fc2 | 20 | #include <linux/kasan.h> |
698b1b30 VB |
21 | #include <linux/kthread.h> |
22 | #include <linux/freezer.h> | |
748446bb MG |
23 | #include "internal.h" |
24 | ||
010fc29a MK |
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 | ||
ff9543fd MN |
40 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
41 | ||
b7aba698 MG |
42 | #define CREATE_TRACE_POINTS |
43 | #include <trace/events/compaction.h> | |
44 | ||
06b6640a VB |
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 | ||
748446bb MG |
50 | static unsigned long release_freepages(struct list_head *freelist) |
51 | { | |
52 | struct page *page, *next; | |
6bace090 | 53 | unsigned long high_pfn = 0; |
748446bb MG |
54 | |
55 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 56 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
57 | list_del(&page->lru); |
58 | __free_page(page); | |
6bace090 VB |
59 | if (pfn > high_pfn) |
60 | high_pfn = pfn; | |
748446bb MG |
61 | } |
62 | ||
6bace090 | 63 | return high_pfn; |
748446bb MG |
64 | } |
65 | ||
ff9543fd MN |
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); | |
b8c73fc2 | 73 | kasan_alloc_pages(page, 0); |
ff9543fd MN |
74 | } |
75 | } | |
76 | ||
47118af0 MN |
77 | static inline bool migrate_async_suitable(int migratetype) |
78 | { | |
79 | return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE; | |
80 | } | |
81 | ||
bb13ffeb | 82 | #ifdef CONFIG_COMPACTION |
24e2716f JK |
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 | ||
bb13ffeb MG |
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 | ||
02333641 VB |
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; | |
623446e4 | 168 | zone->compact_cached_free_pfn = |
06b6640a | 169 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
170 | } |
171 | ||
bb13ffeb MG |
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 | */ | |
62997027 | 177 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
178 | { |
179 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 180 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
181 | unsigned long pfn; |
182 | ||
62997027 | 183 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
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 | } | |
02333641 VB |
200 | |
201 | reset_cached_positions(zone); | |
bb13ffeb MG |
202 | } |
203 | ||
62997027 MG |
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 | ||
bb13ffeb MG |
219 | /* |
220 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 221 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 222 | */ |
c89511ab MG |
223 | static void update_pageblock_skip(struct compact_control *cc, |
224 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 225 | bool migrate_scanner) |
bb13ffeb | 226 | { |
c89511ab | 227 | struct zone *zone = cc->zone; |
35979ef3 | 228 | unsigned long pfn; |
6815bf3f JK |
229 | |
230 | if (cc->ignore_skip_hint) | |
231 | return; | |
232 | ||
bb13ffeb MG |
233 | if (!page) |
234 | return; | |
235 | ||
35979ef3 DR |
236 | if (nr_isolated) |
237 | return; | |
238 | ||
edc2ca61 | 239 | set_pageblock_skip(page); |
c89511ab | 240 | |
35979ef3 DR |
241 | pfn = page_to_pfn(page); |
242 | ||
243 | /* Update where async and sync compaction should restart */ | |
244 | if (migrate_scanner) { | |
35979ef3 DR |
245 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
246 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
247 | if (cc->mode != MIGRATE_ASYNC && |
248 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
249 | zone->compact_cached_migrate_pfn[1] = pfn; |
250 | } else { | |
35979ef3 DR |
251 | if (pfn < zone->compact_cached_free_pfn) |
252 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 253 | } |
bb13ffeb MG |
254 | } |
255 | #else | |
256 | static inline bool isolation_suitable(struct compact_control *cc, | |
257 | struct page *page) | |
258 | { | |
259 | return true; | |
260 | } | |
261 | ||
c89511ab MG |
262 | static void update_pageblock_skip(struct compact_control *cc, |
263 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 264 | bool migrate_scanner) |
bb13ffeb MG |
265 | { |
266 | } | |
267 | #endif /* CONFIG_COMPACTION */ | |
268 | ||
8b44d279 VB |
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) | |
2a1402aa | 279 | { |
8b44d279 VB |
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 | } | |
1f9efdef | 288 | |
8b44d279 | 289 | return true; |
2a1402aa MG |
290 | } |
291 | ||
c67fe375 MG |
292 | /* |
293 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
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. | |
c67fe375 | 301 | * |
8b44d279 VB |
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) | |
c67fe375 | 306 | */ |
8b44d279 VB |
307 | static bool compact_unlock_should_abort(spinlock_t *lock, |
308 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 309 | { |
8b44d279 VB |
310 | if (*locked) { |
311 | spin_unlock_irqrestore(lock, flags); | |
312 | *locked = false; | |
313 | } | |
1f9efdef | 314 | |
8b44d279 VB |
315 | if (fatal_signal_pending(current)) { |
316 | cc->contended = COMPACT_CONTENDED_SCHED; | |
317 | return true; | |
318 | } | |
c67fe375 | 319 | |
8b44d279 | 320 | if (need_resched()) { |
e0b9daeb | 321 | if (cc->mode == MIGRATE_ASYNC) { |
8b44d279 VB |
322 | cc->contended = COMPACT_CONTENDED_SCHED; |
323 | return true; | |
c67fe375 | 324 | } |
c67fe375 | 325 | cond_resched(); |
c67fe375 MG |
326 | } |
327 | ||
8b44d279 | 328 | return false; |
c67fe375 MG |
329 | } |
330 | ||
be976572 VB |
331 | /* |
332 | * Aside from avoiding lock contention, compaction also periodically checks | |
333 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 334 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
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) { | |
1f9efdef | 345 | cc->contended = COMPACT_CONTENDED_SCHED; |
be976572 VB |
346 | return true; |
347 | } | |
348 | ||
349 | cond_resched(); | |
350 | } | |
351 | ||
352 | return false; | |
353 | } | |
354 | ||
85aa125f | 355 | /* |
9e4be470 JM |
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). | |
85aa125f | 359 | */ |
f40d1e42 | 360 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 361 | unsigned long *start_pfn, |
85aa125f MN |
362 | unsigned long end_pfn, |
363 | struct list_head *freelist, | |
364 | bool strict) | |
748446bb | 365 | { |
b7aba698 | 366 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 367 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 368 | unsigned long flags = 0; |
f40d1e42 | 369 | bool locked = false; |
e14c720e | 370 | unsigned long blockpfn = *start_pfn; |
748446bb | 371 | |
748446bb MG |
372 | cursor = pfn_to_page(blockpfn); |
373 | ||
f40d1e42 | 374 | /* Isolate free pages. */ |
748446bb MG |
375 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
376 | int isolated, i; | |
377 | struct page *page = cursor; | |
378 | ||
8b44d279 VB |
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 | ||
b7aba698 | 389 | nr_scanned++; |
f40d1e42 | 390 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
391 | goto isolate_fail; |
392 | ||
bb13ffeb MG |
393 | if (!valid_page) |
394 | valid_page = page; | |
9fcd6d2e VB |
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 | ||
f40d1e42 | 413 | if (!PageBuddy(page)) |
2af120bc | 414 | goto isolate_fail; |
f40d1e42 MG |
415 | |
416 | /* | |
69b7189f VB |
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. | |
f40d1e42 | 422 | */ |
69b7189f VB |
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 | */ | |
8b44d279 VB |
432 | locked = compact_trylock_irqsave(&cc->zone->lock, |
433 | &flags, cc); | |
69b7189f VB |
434 | if (!locked) |
435 | break; | |
f40d1e42 | 436 | |
69b7189f VB |
437 | /* Recheck this is a buddy page under lock */ |
438 | if (!PageBuddy(page)) | |
439 | goto isolate_fail; | |
440 | } | |
748446bb MG |
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) { | |
932ff6bb JK |
452 | cc->nr_freepages += isolated; |
453 | if (!strict && | |
454 | cc->nr_migratepages <= cc->nr_freepages) { | |
455 | blockpfn += isolated; | |
456 | break; | |
457 | } | |
458 | ||
748446bb MG |
459 | blockpfn += isolated - 1; |
460 | cursor += isolated - 1; | |
2af120bc | 461 | continue; |
748446bb | 462 | } |
2af120bc LA |
463 | |
464 | isolate_fail: | |
465 | if (strict) | |
466 | break; | |
467 | else | |
468 | continue; | |
469 | ||
748446bb MG |
470 | } |
471 | ||
9fcd6d2e VB |
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 | ||
e34d85f0 JK |
479 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
480 | nr_scanned, total_isolated); | |
481 | ||
e14c720e VB |
482 | /* Record how far we have got within the block */ |
483 | *start_pfn = blockpfn; | |
484 | ||
f40d1e42 MG |
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 | */ | |
2af120bc | 490 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
491 | total_isolated = 0; |
492 | ||
493 | if (locked) | |
494 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
495 | ||
bb13ffeb MG |
496 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
497 | if (blockpfn == end_pfn) | |
edc2ca61 | 498 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 499 | |
010fc29a | 500 | count_compact_events(COMPACTFREE_SCANNED, nr_scanned); |
397487db | 501 | if (total_isolated) |
010fc29a | 502 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
503 | return total_isolated; |
504 | } | |
505 | ||
85aa125f MN |
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 | */ | |
ff9543fd | 519 | unsigned long |
bb13ffeb MG |
520 | isolate_freepages_range(struct compact_control *cc, |
521 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 522 | { |
e1409c32 | 523 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
524 | LIST_HEAD(freelist); |
525 | ||
7d49d886 | 526 | pfn = start_pfn; |
06b6640a | 527 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
528 | if (block_start_pfn < cc->zone->zone_start_pfn) |
529 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 530 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
531 | |
532 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 533 | block_start_pfn = block_end_pfn, |
7d49d886 | 534 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
535 | /* Protect pfn from changing by isolate_freepages_block */ |
536 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 537 | |
85aa125f MN |
538 | block_end_pfn = min(block_end_pfn, end_pfn); |
539 | ||
58420016 JK |
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) { | |
06b6640a VB |
546 | block_start_pfn = pageblock_start_pfn(pfn); |
547 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
548 | block_end_pfn = min(block_end_pfn, end_pfn); |
549 | } | |
550 | ||
e1409c32 JK |
551 | if (!pageblock_pfn_to_page(block_start_pfn, |
552 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
553 | break; |
554 | ||
e14c720e VB |
555 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
556 | block_end_pfn, &freelist, true); | |
85aa125f MN |
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 | ||
748446bb | 586 | /* Update the number of anon and file isolated pages in the zone */ |
edc2ca61 | 587 | static void acct_isolated(struct zone *zone, struct compact_control *cc) |
748446bb MG |
588 | { |
589 | struct page *page; | |
b9e84ac1 | 590 | unsigned int count[2] = { 0, }; |
748446bb | 591 | |
edc2ca61 VB |
592 | if (list_empty(&cc->migratepages)) |
593 | return; | |
594 | ||
b9e84ac1 MK |
595 | list_for_each_entry(page, &cc->migratepages, lru) |
596 | count[!!page_is_file_cache(page)]++; | |
748446bb | 597 | |
edc2ca61 VB |
598 | mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
599 | mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
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 | { | |
bc693045 | 605 | unsigned long active, inactive, isolated; |
748446bb MG |
606 | |
607 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
608 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
609 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
610 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
611 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
612 | zone_page_state(zone, NR_ISOLATED_ANON); | |
613 | ||
bc693045 | 614 | return isolated > (inactive + active) / 2; |
748446bb MG |
615 | } |
616 | ||
2fe86e00 | 617 | /** |
edc2ca61 VB |
618 | * isolate_migratepages_block() - isolate all migrate-able pages within |
619 | * a single pageblock | |
2fe86e00 | 620 | * @cc: Compaction control structure. |
edc2ca61 VB |
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. | |
2fe86e00 MN |
624 | * |
625 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
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). | |
2fe86e00 | 630 | * |
edc2ca61 VB |
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. | |
748446bb | 634 | */ |
edc2ca61 VB |
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) | |
748446bb | 638 | { |
edc2ca61 | 639 | struct zone *zone = cc->zone; |
b7aba698 | 640 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 641 | struct list_head *migratelist = &cc->migratepages; |
fa9add64 | 642 | struct lruvec *lruvec; |
b8b2d825 | 643 | unsigned long flags = 0; |
2a1402aa | 644 | bool locked = false; |
bb13ffeb | 645 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 646 | unsigned long start_pfn = low_pfn; |
748446bb | 647 | |
748446bb MG |
648 | /* |
649 | * Ensure that there are not too many pages isolated from the LRU | |
650 | * list by either parallel reclaimers or compaction. If there are, | |
651 | * delay for some time until fewer pages are isolated | |
652 | */ | |
653 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 654 | /* async migration should just abort */ |
e0b9daeb | 655 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 656 | return 0; |
f9e35b3b | 657 | |
748446bb MG |
658 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
659 | ||
660 | if (fatal_signal_pending(current)) | |
2fe86e00 | 661 | return 0; |
748446bb MG |
662 | } |
663 | ||
be976572 VB |
664 | if (compact_should_abort(cc)) |
665 | return 0; | |
aeef4b83 | 666 | |
748446bb | 667 | /* Time to isolate some pages for migration */ |
748446bb | 668 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 VB |
669 | bool is_lru; |
670 | ||
8b44d279 VB |
671 | /* |
672 | * Periodically drop the lock (if held) regardless of its | |
673 | * contention, to give chance to IRQs. Abort async compaction | |
674 | * if contended. | |
675 | */ | |
676 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
677 | && compact_unlock_should_abort(&zone->lru_lock, flags, | |
678 | &locked, cc)) | |
679 | break; | |
c67fe375 | 680 | |
748446bb MG |
681 | if (!pfn_valid_within(low_pfn)) |
682 | continue; | |
b7aba698 | 683 | nr_scanned++; |
748446bb | 684 | |
748446bb | 685 | page = pfn_to_page(low_pfn); |
dc908600 | 686 | |
bb13ffeb MG |
687 | if (!valid_page) |
688 | valid_page = page; | |
689 | ||
6c14466c | 690 | /* |
99c0fd5e VB |
691 | * Skip if free. We read page order here without zone lock |
692 | * which is generally unsafe, but the race window is small and | |
693 | * the worst thing that can happen is that we skip some | |
694 | * potential isolation targets. | |
6c14466c | 695 | */ |
99c0fd5e VB |
696 | if (PageBuddy(page)) { |
697 | unsigned long freepage_order = page_order_unsafe(page); | |
698 | ||
699 | /* | |
700 | * Without lock, we cannot be sure that what we got is | |
701 | * a valid page order. Consider only values in the | |
702 | * valid order range to prevent low_pfn overflow. | |
703 | */ | |
704 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
705 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 706 | continue; |
99c0fd5e | 707 | } |
748446bb | 708 | |
bf6bddf1 RA |
709 | /* |
710 | * Check may be lockless but that's ok as we recheck later. | |
711 | * It's possible to migrate LRU pages and balloon pages | |
712 | * Skip any other type of page | |
713 | */ | |
29c0dde8 VB |
714 | is_lru = PageLRU(page); |
715 | if (!is_lru) { | |
bf6bddf1 | 716 | if (unlikely(balloon_page_movable(page))) { |
d6d86c0a | 717 | if (balloon_page_isolate(page)) { |
bf6bddf1 | 718 | /* Successfully isolated */ |
b6c75016 | 719 | goto isolate_success; |
bf6bddf1 RA |
720 | } |
721 | } | |
bf6bddf1 | 722 | } |
bc835011 AA |
723 | |
724 | /* | |
29c0dde8 VB |
725 | * Regardless of being on LRU, compound pages such as THP and |
726 | * hugetlbfs are not to be compacted. We can potentially save | |
727 | * a lot of iterations if we skip them at once. The check is | |
728 | * racy, but we can consider only valid values and the only | |
729 | * danger is skipping too much. | |
bc835011 | 730 | */ |
29c0dde8 VB |
731 | if (PageCompound(page)) { |
732 | unsigned int comp_order = compound_order(page); | |
733 | ||
734 | if (likely(comp_order < MAX_ORDER)) | |
735 | low_pfn += (1UL << comp_order) - 1; | |
edc2ca61 | 736 | |
2a1402aa MG |
737 | continue; |
738 | } | |
739 | ||
29c0dde8 VB |
740 | if (!is_lru) |
741 | continue; | |
742 | ||
119d6d59 DR |
743 | /* |
744 | * Migration will fail if an anonymous page is pinned in memory, | |
745 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
746 | * admittedly racy check. | |
747 | */ | |
748 | if (!page_mapping(page) && | |
749 | page_count(page) > page_mapcount(page)) | |
750 | continue; | |
751 | ||
69b7189f VB |
752 | /* If we already hold the lock, we can skip some rechecking */ |
753 | if (!locked) { | |
8b44d279 VB |
754 | locked = compact_trylock_irqsave(&zone->lru_lock, |
755 | &flags, cc); | |
69b7189f VB |
756 | if (!locked) |
757 | break; | |
2a1402aa | 758 | |
29c0dde8 | 759 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f VB |
760 | if (!PageLRU(page)) |
761 | continue; | |
29c0dde8 VB |
762 | |
763 | /* | |
764 | * Page become compound since the non-locked check, | |
765 | * and it's on LRU. It can only be a THP so the order | |
766 | * is safe to read and it's 0 for tail pages. | |
767 | */ | |
768 | if (unlikely(PageCompound(page))) { | |
769 | low_pfn += (1UL << compound_order(page)) - 1; | |
69b7189f VB |
770 | continue; |
771 | } | |
bc835011 AA |
772 | } |
773 | ||
fa9add64 HD |
774 | lruvec = mem_cgroup_page_lruvec(page, zone); |
775 | ||
748446bb | 776 | /* Try isolate the page */ |
edc2ca61 | 777 | if (__isolate_lru_page(page, isolate_mode) != 0) |
748446bb MG |
778 | continue; |
779 | ||
29c0dde8 | 780 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 781 | |
748446bb | 782 | /* Successfully isolated */ |
fa9add64 | 783 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
b6c75016 JK |
784 | |
785 | isolate_success: | |
748446bb | 786 | list_add(&page->lru, migratelist); |
748446bb | 787 | cc->nr_migratepages++; |
b7aba698 | 788 | nr_isolated++; |
748446bb MG |
789 | |
790 | /* Avoid isolating too much */ | |
31b8384a HD |
791 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
792 | ++low_pfn; | |
748446bb | 793 | break; |
31b8384a | 794 | } |
748446bb MG |
795 | } |
796 | ||
99c0fd5e VB |
797 | /* |
798 | * The PageBuddy() check could have potentially brought us outside | |
799 | * the range to be scanned. | |
800 | */ | |
801 | if (unlikely(low_pfn > end_pfn)) | |
802 | low_pfn = end_pfn; | |
803 | ||
c67fe375 MG |
804 | if (locked) |
805 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
748446bb | 806 | |
50b5b094 VB |
807 | /* |
808 | * Update the pageblock-skip information and cached scanner pfn, | |
809 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 810 | */ |
35979ef3 | 811 | if (low_pfn == end_pfn) |
edc2ca61 | 812 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 813 | |
e34d85f0 JK |
814 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
815 | nr_scanned, nr_isolated); | |
b7aba698 | 816 | |
010fc29a | 817 | count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned); |
397487db | 818 | if (nr_isolated) |
010fc29a | 819 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 820 | |
2fe86e00 MN |
821 | return low_pfn; |
822 | } | |
823 | ||
edc2ca61 VB |
824 | /** |
825 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
826 | * @cc: Compaction control structure. | |
827 | * @start_pfn: The first PFN to start isolating. | |
828 | * @end_pfn: The one-past-last PFN. | |
829 | * | |
830 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
831 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
832 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
833 | */ | |
834 | unsigned long | |
835 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
836 | unsigned long end_pfn) | |
837 | { | |
e1409c32 | 838 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
839 | |
840 | /* Scan block by block. First and last block may be incomplete */ | |
841 | pfn = start_pfn; | |
06b6640a | 842 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
843 | if (block_start_pfn < cc->zone->zone_start_pfn) |
844 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 845 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
846 | |
847 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 848 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
849 | block_end_pfn += pageblock_nr_pages) { |
850 | ||
851 | block_end_pfn = min(block_end_pfn, end_pfn); | |
852 | ||
e1409c32 JK |
853 | if (!pageblock_pfn_to_page(block_start_pfn, |
854 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
855 | continue; |
856 | ||
857 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
858 | ISOLATE_UNEVICTABLE); | |
859 | ||
14af4a5e | 860 | if (!pfn) |
edc2ca61 | 861 | break; |
6ea41c0c JK |
862 | |
863 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
864 | break; | |
edc2ca61 VB |
865 | } |
866 | acct_isolated(cc->zone, cc); | |
867 | ||
868 | return pfn; | |
869 | } | |
870 | ||
ff9543fd MN |
871 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
872 | #ifdef CONFIG_COMPACTION | |
018e9a49 AM |
873 | |
874 | /* Returns true if the page is within a block suitable for migration to */ | |
875 | static bool suitable_migration_target(struct page *page) | |
876 | { | |
877 | /* If the page is a large free page, then disallow migration */ | |
878 | if (PageBuddy(page)) { | |
879 | /* | |
880 | * We are checking page_order without zone->lock taken. But | |
881 | * the only small danger is that we skip a potentially suitable | |
882 | * pageblock, so it's not worth to check order for valid range. | |
883 | */ | |
884 | if (page_order_unsafe(page) >= pageblock_order) | |
885 | return false; | |
886 | } | |
887 | ||
888 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ | |
889 | if (migrate_async_suitable(get_pageblock_migratetype(page))) | |
890 | return true; | |
891 | ||
892 | /* Otherwise skip the block */ | |
893 | return false; | |
894 | } | |
895 | ||
f2849aa0 VB |
896 | /* |
897 | * Test whether the free scanner has reached the same or lower pageblock than | |
898 | * the migration scanner, and compaction should thus terminate. | |
899 | */ | |
900 | static inline bool compact_scanners_met(struct compact_control *cc) | |
901 | { | |
902 | return (cc->free_pfn >> pageblock_order) | |
903 | <= (cc->migrate_pfn >> pageblock_order); | |
904 | } | |
905 | ||
2fe86e00 | 906 | /* |
ff9543fd MN |
907 | * Based on information in the current compact_control, find blocks |
908 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 909 | */ |
edc2ca61 | 910 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 911 | { |
edc2ca61 | 912 | struct zone *zone = cc->zone; |
ff9543fd | 913 | struct page *page; |
c96b9e50 | 914 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 915 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
916 | unsigned long block_end_pfn; /* end of current pageblock */ |
917 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 918 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 919 | |
ff9543fd MN |
920 | /* |
921 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 922 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
923 | * zone when isolating for the first time. For looping we also need |
924 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
925 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
926 | * For ending point, take care when isolating in last pageblock of a | |
927 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
928 | * The low boundary is the end of the pageblock the migration scanner |
929 | * is using. | |
ff9543fd | 930 | */ |
e14c720e | 931 | isolate_start_pfn = cc->free_pfn; |
06b6640a | 932 | block_start_pfn = pageblock_start_pfn(cc->free_pfn); |
c96b9e50 VB |
933 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
934 | zone_end_pfn(zone)); | |
06b6640a | 935 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
2fe86e00 | 936 | |
ff9543fd MN |
937 | /* |
938 | * Isolate free pages until enough are available to migrate the | |
939 | * pages on cc->migratepages. We stop searching if the migrate | |
940 | * and free page scanners meet or enough free pages are isolated. | |
941 | */ | |
f5f61a32 | 942 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 943 | block_end_pfn = block_start_pfn, |
e14c720e VB |
944 | block_start_pfn -= pageblock_nr_pages, |
945 | isolate_start_pfn = block_start_pfn) { | |
2fe86e00 | 946 | |
f6ea3adb DR |
947 | /* |
948 | * This can iterate a massively long zone without finding any | |
949 | * suitable migration targets, so periodically check if we need | |
be976572 | 950 | * to schedule, or even abort async compaction. |
f6ea3adb | 951 | */ |
be976572 VB |
952 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
953 | && compact_should_abort(cc)) | |
954 | break; | |
f6ea3adb | 955 | |
7d49d886 VB |
956 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
957 | zone); | |
958 | if (!page) | |
ff9543fd MN |
959 | continue; |
960 | ||
961 | /* Check the block is suitable for migration */ | |
68e3e926 | 962 | if (!suitable_migration_target(page)) |
ff9543fd | 963 | continue; |
68e3e926 | 964 | |
bb13ffeb MG |
965 | /* If isolation recently failed, do not retry */ |
966 | if (!isolation_suitable(cc, page)) | |
967 | continue; | |
968 | ||
e14c720e | 969 | /* Found a block suitable for isolating free pages from. */ |
932ff6bb | 970 | isolate_freepages_block(cc, &isolate_start_pfn, |
c96b9e50 | 971 | block_end_pfn, freelist, false); |
ff9543fd | 972 | |
e14c720e | 973 | /* |
f5f61a32 VB |
974 | * If we isolated enough freepages, or aborted due to async |
975 | * compaction being contended, terminate the loop. | |
e14c720e VB |
976 | * Remember where the free scanner should restart next time, |
977 | * which is where isolate_freepages_block() left off. | |
978 | * But if it scanned the whole pageblock, isolate_start_pfn | |
979 | * now points at block_end_pfn, which is the start of the next | |
980 | * pageblock. | |
981 | * In that case we will however want to restart at the start | |
982 | * of the previous pageblock. | |
983 | */ | |
f5f61a32 VB |
984 | if ((cc->nr_freepages >= cc->nr_migratepages) |
985 | || cc->contended) { | |
986 | if (isolate_start_pfn >= block_end_pfn) | |
987 | isolate_start_pfn = | |
988 | block_start_pfn - pageblock_nr_pages; | |
be976572 | 989 | break; |
f5f61a32 VB |
990 | } else { |
991 | /* | |
992 | * isolate_freepages_block() should not terminate | |
993 | * prematurely unless contended, or isolated enough | |
994 | */ | |
995 | VM_BUG_ON(isolate_start_pfn < block_end_pfn); | |
996 | } | |
ff9543fd MN |
997 | } |
998 | ||
999 | /* split_free_page does not map the pages */ | |
1000 | map_pages(freelist); | |
1001 | ||
7ed695e0 | 1002 | /* |
f5f61a32 VB |
1003 | * Record where the free scanner will restart next time. Either we |
1004 | * broke from the loop and set isolate_start_pfn based on the last | |
1005 | * call to isolate_freepages_block(), or we met the migration scanner | |
1006 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1007 | */ |
f5f61a32 | 1008 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1009 | } |
1010 | ||
1011 | /* | |
1012 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1013 | * from the isolated freelists in the block we are migrating to. | |
1014 | */ | |
1015 | static struct page *compaction_alloc(struct page *migratepage, | |
1016 | unsigned long data, | |
1017 | int **result) | |
1018 | { | |
1019 | struct compact_control *cc = (struct compact_control *)data; | |
1020 | struct page *freepage; | |
1021 | ||
be976572 VB |
1022 | /* |
1023 | * Isolate free pages if necessary, and if we are not aborting due to | |
1024 | * contention. | |
1025 | */ | |
748446bb | 1026 | if (list_empty(&cc->freepages)) { |
be976572 | 1027 | if (!cc->contended) |
edc2ca61 | 1028 | isolate_freepages(cc); |
748446bb MG |
1029 | |
1030 | if (list_empty(&cc->freepages)) | |
1031 | return NULL; | |
1032 | } | |
1033 | ||
1034 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1035 | list_del(&freepage->lru); | |
1036 | cc->nr_freepages--; | |
1037 | ||
1038 | return freepage; | |
1039 | } | |
1040 | ||
1041 | /* | |
d53aea3d DR |
1042 | * This is a migrate-callback that "frees" freepages back to the isolated |
1043 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1044 | * special handling needed for NUMA. | |
1045 | */ | |
1046 | static void compaction_free(struct page *page, unsigned long data) | |
1047 | { | |
1048 | struct compact_control *cc = (struct compact_control *)data; | |
1049 | ||
1050 | list_add(&page->lru, &cc->freepages); | |
1051 | cc->nr_freepages++; | |
1052 | } | |
1053 | ||
ff9543fd MN |
1054 | /* possible outcome of isolate_migratepages */ |
1055 | typedef enum { | |
1056 | ISOLATE_ABORT, /* Abort compaction now */ | |
1057 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1058 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1059 | } isolate_migrate_t; | |
1060 | ||
5bbe3547 EM |
1061 | /* |
1062 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1063 | * compactable pages. | |
1064 | */ | |
1065 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1066 | ||
ff9543fd | 1067 | /* |
edc2ca61 VB |
1068 | * Isolate all pages that can be migrated from the first suitable block, |
1069 | * starting at the block pointed to by the migrate scanner pfn within | |
1070 | * compact_control. | |
ff9543fd MN |
1071 | */ |
1072 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1073 | struct compact_control *cc) | |
1074 | { | |
e1409c32 JK |
1075 | unsigned long block_start_pfn; |
1076 | unsigned long block_end_pfn; | |
1077 | unsigned long low_pfn; | |
1a16718c | 1078 | unsigned long isolate_start_pfn; |
edc2ca61 VB |
1079 | struct page *page; |
1080 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1081 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
edc2ca61 | 1082 | (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1083 | |
edc2ca61 VB |
1084 | /* |
1085 | * Start at where we last stopped, or beginning of the zone as | |
1086 | * initialized by compact_zone() | |
1087 | */ | |
1088 | low_pfn = cc->migrate_pfn; | |
06b6640a | 1089 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1090 | if (block_start_pfn < zone->zone_start_pfn) |
1091 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd MN |
1092 | |
1093 | /* Only scan within a pageblock boundary */ | |
06b6640a | 1094 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1095 | |
edc2ca61 VB |
1096 | /* |
1097 | * Iterate over whole pageblocks until we find the first suitable. | |
1098 | * Do not cross the free scanner. | |
1099 | */ | |
e1409c32 JK |
1100 | for (; block_end_pfn <= cc->free_pfn; |
1101 | low_pfn = block_end_pfn, | |
1102 | block_start_pfn = block_end_pfn, | |
1103 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1104 | |
edc2ca61 VB |
1105 | /* |
1106 | * This can potentially iterate a massively long zone with | |
1107 | * many pageblocks unsuitable, so periodically check if we | |
1108 | * need to schedule, or even abort async compaction. | |
1109 | */ | |
1110 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1111 | && compact_should_abort(cc)) | |
1112 | break; | |
ff9543fd | 1113 | |
e1409c32 JK |
1114 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1115 | zone); | |
7d49d886 | 1116 | if (!page) |
edc2ca61 VB |
1117 | continue; |
1118 | ||
edc2ca61 VB |
1119 | /* If isolation recently failed, do not retry */ |
1120 | if (!isolation_suitable(cc, page)) | |
1121 | continue; | |
1122 | ||
1123 | /* | |
1124 | * For async compaction, also only scan in MOVABLE blocks. | |
1125 | * Async compaction is optimistic to see if the minimum amount | |
1126 | * of work satisfies the allocation. | |
1127 | */ | |
1128 | if (cc->mode == MIGRATE_ASYNC && | |
1129 | !migrate_async_suitable(get_pageblock_migratetype(page))) | |
1130 | continue; | |
1131 | ||
1132 | /* Perform the isolation */ | |
1a16718c | 1133 | isolate_start_pfn = low_pfn; |
e1409c32 JK |
1134 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1135 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1136 | |
ff59909a HD |
1137 | if (!low_pfn || cc->contended) { |
1138 | acct_isolated(zone, cc); | |
edc2ca61 | 1139 | return ISOLATE_ABORT; |
ff59909a | 1140 | } |
edc2ca61 | 1141 | |
1a16718c JK |
1142 | /* |
1143 | * Record where we could have freed pages by migration and not | |
1144 | * yet flushed them to buddy allocator. | |
1145 | * - this is the lowest page that could have been isolated and | |
1146 | * then freed by migration. | |
1147 | */ | |
1148 | if (cc->nr_migratepages && !cc->last_migrated_pfn) | |
1149 | cc->last_migrated_pfn = isolate_start_pfn; | |
1150 | ||
edc2ca61 VB |
1151 | /* |
1152 | * Either we isolated something and proceed with migration. Or | |
1153 | * we failed and compact_zone should decide if we should | |
1154 | * continue or not. | |
1155 | */ | |
1156 | break; | |
1157 | } | |
1158 | ||
1159 | acct_isolated(zone, cc); | |
f2849aa0 VB |
1160 | /* Record where migration scanner will be restarted. */ |
1161 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1162 | |
edc2ca61 | 1163 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1164 | } |
1165 | ||
21c527a3 YB |
1166 | /* |
1167 | * order == -1 is expected when compacting via | |
1168 | * /proc/sys/vm/compact_memory | |
1169 | */ | |
1170 | static inline bool is_via_compact_memory(int order) | |
1171 | { | |
1172 | return order == -1; | |
1173 | } | |
1174 | ||
837d026d | 1175 | static int __compact_finished(struct zone *zone, struct compact_control *cc, |
6d7ce559 | 1176 | const int migratetype) |
748446bb | 1177 | { |
8fb74b9f | 1178 | unsigned int order; |
5a03b051 | 1179 | unsigned long watermark; |
56de7263 | 1180 | |
be976572 | 1181 | if (cc->contended || fatal_signal_pending(current)) |
2d1e1041 | 1182 | return COMPACT_CONTENDED; |
748446bb | 1183 | |
753341a4 | 1184 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1185 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1186 | /* Let the next compaction start anew. */ |
02333641 | 1187 | reset_cached_positions(zone); |
55b7c4c9 | 1188 | |
62997027 MG |
1189 | /* |
1190 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1191 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1192 | * flag itself as the decision to be clear should be directly |
1193 | * based on an allocation request. | |
1194 | */ | |
accf6242 | 1195 | if (cc->direct_compaction) |
62997027 MG |
1196 | zone->compact_blockskip_flush = true; |
1197 | ||
748446bb | 1198 | return COMPACT_COMPLETE; |
bb13ffeb | 1199 | } |
748446bb | 1200 | |
21c527a3 | 1201 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1202 | return COMPACT_CONTINUE; |
1203 | ||
3957c776 MH |
1204 | /* Compaction run is not finished if the watermark is not met */ |
1205 | watermark = low_wmark_pages(zone); | |
3957c776 | 1206 | |
ebff3980 VB |
1207 | if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx, |
1208 | cc->alloc_flags)) | |
3957c776 MH |
1209 | return COMPACT_CONTINUE; |
1210 | ||
56de7263 | 1211 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f MG |
1212 | for (order = cc->order; order < MAX_ORDER; order++) { |
1213 | struct free_area *area = &zone->free_area[order]; | |
2149cdae | 1214 | bool can_steal; |
8fb74b9f MG |
1215 | |
1216 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1217 | if (!list_empty(&area->free_list[migratetype])) |
8fb74b9f MG |
1218 | return COMPACT_PARTIAL; |
1219 | ||
2149cdae JK |
1220 | #ifdef CONFIG_CMA |
1221 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1222 | if (migratetype == MIGRATE_MOVABLE && | |
1223 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
1224 | return COMPACT_PARTIAL; | |
1225 | #endif | |
1226 | /* | |
1227 | * Job done if allocation would steal freepages from | |
1228 | * other migratetype buddy lists. | |
1229 | */ | |
1230 | if (find_suitable_fallback(area, order, migratetype, | |
1231 | true, &can_steal) != -1) | |
56de7263 MG |
1232 | return COMPACT_PARTIAL; |
1233 | } | |
1234 | ||
837d026d JK |
1235 | return COMPACT_NO_SUITABLE_PAGE; |
1236 | } | |
1237 | ||
1238 | static int compact_finished(struct zone *zone, struct compact_control *cc, | |
1239 | const int migratetype) | |
1240 | { | |
1241 | int ret; | |
1242 | ||
1243 | ret = __compact_finished(zone, cc, migratetype); | |
1244 | trace_mm_compaction_finished(zone, cc->order, ret); | |
1245 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1246 | ret = COMPACT_CONTINUE; | |
1247 | ||
1248 | return ret; | |
748446bb MG |
1249 | } |
1250 | ||
3e7d3449 MG |
1251 | /* |
1252 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1253 | * Returns | |
1254 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
1255 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
1256 | * COMPACT_CONTINUE - If compaction should run now | |
1257 | */ | |
837d026d | 1258 | static unsigned long __compaction_suitable(struct zone *zone, int order, |
ebff3980 | 1259 | int alloc_flags, int classzone_idx) |
3e7d3449 MG |
1260 | { |
1261 | int fragindex; | |
1262 | unsigned long watermark; | |
1263 | ||
21c527a3 | 1264 | if (is_via_compact_memory(order)) |
3957c776 MH |
1265 | return COMPACT_CONTINUE; |
1266 | ||
ebff3980 VB |
1267 | watermark = low_wmark_pages(zone); |
1268 | /* | |
1269 | * If watermarks for high-order allocation are already met, there | |
1270 | * should be no need for compaction at all. | |
1271 | */ | |
1272 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1273 | alloc_flags)) | |
1274 | return COMPACT_PARTIAL; | |
1275 | ||
3e7d3449 MG |
1276 | /* |
1277 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
1278 | * This is because during migration, copies of pages need to be | |
1279 | * allocated and for a short time, the footprint is higher | |
1280 | */ | |
ebff3980 VB |
1281 | watermark += (2UL << order); |
1282 | if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags)) | |
3e7d3449 MG |
1283 | return COMPACT_SKIPPED; |
1284 | ||
1285 | /* | |
1286 | * fragmentation index determines if allocation failures are due to | |
1287 | * low memory or external fragmentation | |
1288 | * | |
ebff3980 VB |
1289 | * index of -1000 would imply allocations might succeed depending on |
1290 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1291 | * index towards 0 implies failure is due to lack of memory |
1292 | * index towards 1000 implies failure is due to fragmentation | |
1293 | * | |
1294 | * Only compact if a failure would be due to fragmentation. | |
1295 | */ | |
1296 | fragindex = fragmentation_index(zone, order); | |
1297 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
837d026d | 1298 | return COMPACT_NOT_SUITABLE_ZONE; |
3e7d3449 | 1299 | |
3e7d3449 MG |
1300 | return COMPACT_CONTINUE; |
1301 | } | |
1302 | ||
837d026d JK |
1303 | unsigned long compaction_suitable(struct zone *zone, int order, |
1304 | int alloc_flags, int classzone_idx) | |
1305 | { | |
1306 | unsigned long ret; | |
1307 | ||
1308 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx); | |
1309 | trace_mm_compaction_suitable(zone, order, ret); | |
1310 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1311 | ret = COMPACT_SKIPPED; | |
1312 | ||
1313 | return ret; | |
1314 | } | |
1315 | ||
748446bb MG |
1316 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
1317 | { | |
1318 | int ret; | |
c89511ab | 1319 | unsigned long start_pfn = zone->zone_start_pfn; |
108bcc96 | 1320 | unsigned long end_pfn = zone_end_pfn(zone); |
6d7ce559 | 1321 | const int migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
e0b9daeb | 1322 | const bool sync = cc->mode != MIGRATE_ASYNC; |
748446bb | 1323 | |
ebff3980 VB |
1324 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, |
1325 | cc->classzone_idx); | |
3e7d3449 MG |
1326 | switch (ret) { |
1327 | case COMPACT_PARTIAL: | |
1328 | case COMPACT_SKIPPED: | |
1329 | /* Compaction is likely to fail */ | |
1330 | return ret; | |
1331 | case COMPACT_CONTINUE: | |
1332 | /* Fall through to compaction */ | |
1333 | ; | |
1334 | } | |
1335 | ||
d3132e4b VB |
1336 | /* |
1337 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 1338 | * is about to be retried after being deferred. |
d3132e4b | 1339 | */ |
accf6242 | 1340 | if (compaction_restarting(zone, cc->order)) |
d3132e4b VB |
1341 | __reset_isolation_suitable(zone); |
1342 | ||
c89511ab MG |
1343 | /* |
1344 | * Setup to move all movable pages to the end of the zone. Used cached | |
1345 | * information on where the scanners should start but check that it | |
1346 | * is initialised by ensuring the values are within zone boundaries. | |
1347 | */ | |
e0b9daeb | 1348 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; |
c89511ab | 1349 | cc->free_pfn = zone->compact_cached_free_pfn; |
623446e4 | 1350 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
06b6640a | 1351 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
c89511ab MG |
1352 | zone->compact_cached_free_pfn = cc->free_pfn; |
1353 | } | |
623446e4 | 1354 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { |
c89511ab | 1355 | cc->migrate_pfn = start_pfn; |
35979ef3 DR |
1356 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1357 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
c89511ab | 1358 | } |
1a16718c | 1359 | cc->last_migrated_pfn = 0; |
748446bb | 1360 | |
16c4a097 JK |
1361 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1362 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1363 | |
748446bb MG |
1364 | migrate_prep_local(); |
1365 | ||
6d7ce559 DR |
1366 | while ((ret = compact_finished(zone, cc, migratetype)) == |
1367 | COMPACT_CONTINUE) { | |
9d502c1c | 1368 | int err; |
748446bb | 1369 | |
f9e35b3b MG |
1370 | switch (isolate_migratepages(zone, cc)) { |
1371 | case ISOLATE_ABORT: | |
2d1e1041 | 1372 | ret = COMPACT_CONTENDED; |
5733c7d1 | 1373 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1374 | cc->nr_migratepages = 0; |
f9e35b3b MG |
1375 | goto out; |
1376 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1377 | /* |
1378 | * We haven't isolated and migrated anything, but | |
1379 | * there might still be unflushed migrations from | |
1380 | * previous cc->order aligned block. | |
1381 | */ | |
1382 | goto check_drain; | |
f9e35b3b MG |
1383 | case ISOLATE_SUCCESS: |
1384 | ; | |
1385 | } | |
748446bb | 1386 | |
d53aea3d | 1387 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1388 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1389 | MR_COMPACTION); |
748446bb | 1390 | |
f8c9301f VB |
1391 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1392 | &cc->migratepages); | |
748446bb | 1393 | |
f8c9301f VB |
1394 | /* All pages were either migrated or will be released */ |
1395 | cc->nr_migratepages = 0; | |
9d502c1c | 1396 | if (err) { |
5733c7d1 | 1397 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1398 | /* |
1399 | * migrate_pages() may return -ENOMEM when scanners meet | |
1400 | * and we want compact_finished() to detect it | |
1401 | */ | |
f2849aa0 | 1402 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 1403 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
1404 | goto out; |
1405 | } | |
748446bb | 1406 | } |
fdaf7f5c | 1407 | |
fdaf7f5c VB |
1408 | check_drain: |
1409 | /* | |
1410 | * Has the migration scanner moved away from the previous | |
1411 | * cc->order aligned block where we migrated from? If yes, | |
1412 | * flush the pages that were freed, so that they can merge and | |
1413 | * compact_finished() can detect immediately if allocation | |
1414 | * would succeed. | |
1415 | */ | |
1a16718c | 1416 | if (cc->order > 0 && cc->last_migrated_pfn) { |
fdaf7f5c VB |
1417 | int cpu; |
1418 | unsigned long current_block_start = | |
06b6640a | 1419 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 1420 | |
1a16718c | 1421 | if (cc->last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
1422 | cpu = get_cpu(); |
1423 | lru_add_drain_cpu(cpu); | |
1424 | drain_local_pages(zone); | |
1425 | put_cpu(); | |
1426 | /* No more flushing until we migrate again */ | |
1a16718c | 1427 | cc->last_migrated_pfn = 0; |
fdaf7f5c VB |
1428 | } |
1429 | } | |
1430 | ||
748446bb MG |
1431 | } |
1432 | ||
f9e35b3b | 1433 | out: |
6bace090 VB |
1434 | /* |
1435 | * Release free pages and update where the free scanner should restart, | |
1436 | * so we don't leave any returned pages behind in the next attempt. | |
1437 | */ | |
1438 | if (cc->nr_freepages > 0) { | |
1439 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1440 | ||
1441 | cc->nr_freepages = 0; | |
1442 | VM_BUG_ON(free_pfn == 0); | |
1443 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 1444 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
1445 | /* |
1446 | * Only go back, not forward. The cached pfn might have been | |
1447 | * already reset to zone end in compact_finished() | |
1448 | */ | |
1449 | if (free_pfn > zone->compact_cached_free_pfn) | |
1450 | zone->compact_cached_free_pfn = free_pfn; | |
1451 | } | |
748446bb | 1452 | |
16c4a097 JK |
1453 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1454 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1455 | |
2d1e1041 VB |
1456 | if (ret == COMPACT_CONTENDED) |
1457 | ret = COMPACT_PARTIAL; | |
1458 | ||
748446bb MG |
1459 | return ret; |
1460 | } | |
76ab0f53 | 1461 | |
e0b9daeb | 1462 | static unsigned long compact_zone_order(struct zone *zone, int order, |
ebff3980 VB |
1463 | gfp_t gfp_mask, enum migrate_mode mode, int *contended, |
1464 | int alloc_flags, int classzone_idx) | |
56de7263 | 1465 | { |
e64c5237 | 1466 | unsigned long ret; |
56de7263 MG |
1467 | struct compact_control cc = { |
1468 | .nr_freepages = 0, | |
1469 | .nr_migratepages = 0, | |
1470 | .order = order, | |
6d7ce559 | 1471 | .gfp_mask = gfp_mask, |
56de7263 | 1472 | .zone = zone, |
e0b9daeb | 1473 | .mode = mode, |
ebff3980 VB |
1474 | .alloc_flags = alloc_flags, |
1475 | .classzone_idx = classzone_idx, | |
accf6242 | 1476 | .direct_compaction = true, |
56de7263 MG |
1477 | }; |
1478 | INIT_LIST_HEAD(&cc.freepages); | |
1479 | INIT_LIST_HEAD(&cc.migratepages); | |
1480 | ||
e64c5237 SL |
1481 | ret = compact_zone(zone, &cc); |
1482 | ||
1483 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1484 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1485 | ||
1486 | *contended = cc.contended; | |
1487 | return ret; | |
56de7263 MG |
1488 | } |
1489 | ||
5e771905 MG |
1490 | int sysctl_extfrag_threshold = 500; |
1491 | ||
56de7263 MG |
1492 | /** |
1493 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1494 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1495 | * @order: The order of the current allocation |
1496 | * @alloc_flags: The allocation flags of the current allocation | |
1497 | * @ac: The context of current allocation | |
e0b9daeb | 1498 | * @mode: The migration mode for async, sync light, or sync migration |
1f9efdef VB |
1499 | * @contended: Return value that determines if compaction was aborted due to |
1500 | * need_resched() or lock contention | |
56de7263 MG |
1501 | * |
1502 | * This is the main entry point for direct page compaction. | |
1503 | */ | |
1a6d53a1 VB |
1504 | unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
1505 | int alloc_flags, const struct alloc_context *ac, | |
1506 | enum migrate_mode mode, int *contended) | |
56de7263 | 1507 | { |
56de7263 MG |
1508 | int may_enter_fs = gfp_mask & __GFP_FS; |
1509 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
1510 | struct zoneref *z; |
1511 | struct zone *zone; | |
53853e2d | 1512 | int rc = COMPACT_DEFERRED; |
1f9efdef VB |
1513 | int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */ |
1514 | ||
1515 | *contended = COMPACT_CONTENDED_NONE; | |
56de7263 | 1516 | |
4ffb6335 | 1517 | /* Check if the GFP flags allow compaction */ |
c5a73c3d | 1518 | if (!order || !may_enter_fs || !may_perform_io) |
53853e2d | 1519 | return COMPACT_SKIPPED; |
56de7263 | 1520 | |
837d026d JK |
1521 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode); |
1522 | ||
56de7263 | 1523 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1524 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1525 | ac->nodemask) { | |
56de7263 | 1526 | int status; |
1f9efdef | 1527 | int zone_contended; |
56de7263 | 1528 | |
53853e2d VB |
1529 | if (compaction_deferred(zone, order)) |
1530 | continue; | |
1531 | ||
e0b9daeb | 1532 | status = compact_zone_order(zone, order, gfp_mask, mode, |
1a6d53a1 VB |
1533 | &zone_contended, alloc_flags, |
1534 | ac->classzone_idx); | |
56de7263 | 1535 | rc = max(status, rc); |
1f9efdef VB |
1536 | /* |
1537 | * It takes at least one zone that wasn't lock contended | |
1538 | * to clear all_zones_contended. | |
1539 | */ | |
1540 | all_zones_contended &= zone_contended; | |
56de7263 | 1541 | |
3e7d3449 | 1542 | /* If a normal allocation would succeed, stop compacting */ |
ebff3980 | 1543 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), |
1a6d53a1 | 1544 | ac->classzone_idx, alloc_flags)) { |
53853e2d VB |
1545 | /* |
1546 | * We think the allocation will succeed in this zone, | |
1547 | * but it is not certain, hence the false. The caller | |
1548 | * will repeat this with true if allocation indeed | |
1549 | * succeeds in this zone. | |
1550 | */ | |
1551 | compaction_defer_reset(zone, order, false); | |
1f9efdef VB |
1552 | /* |
1553 | * It is possible that async compaction aborted due to | |
1554 | * need_resched() and the watermarks were ok thanks to | |
1555 | * somebody else freeing memory. The allocation can | |
1556 | * however still fail so we better signal the | |
1557 | * need_resched() contention anyway (this will not | |
1558 | * prevent the allocation attempt). | |
1559 | */ | |
1560 | if (zone_contended == COMPACT_CONTENDED_SCHED) | |
1561 | *contended = COMPACT_CONTENDED_SCHED; | |
1562 | ||
1563 | goto break_loop; | |
1564 | } | |
1565 | ||
f8669795 | 1566 | if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) { |
53853e2d VB |
1567 | /* |
1568 | * We think that allocation won't succeed in this zone | |
1569 | * so we defer compaction there. If it ends up | |
1570 | * succeeding after all, it will be reset. | |
1571 | */ | |
1572 | defer_compaction(zone, order); | |
1573 | } | |
1f9efdef VB |
1574 | |
1575 | /* | |
1576 | * We might have stopped compacting due to need_resched() in | |
1577 | * async compaction, or due to a fatal signal detected. In that | |
1578 | * case do not try further zones and signal need_resched() | |
1579 | * contention. | |
1580 | */ | |
1581 | if ((zone_contended == COMPACT_CONTENDED_SCHED) | |
1582 | || fatal_signal_pending(current)) { | |
1583 | *contended = COMPACT_CONTENDED_SCHED; | |
1584 | goto break_loop; | |
1585 | } | |
1586 | ||
1587 | continue; | |
1588 | break_loop: | |
1589 | /* | |
1590 | * We might not have tried all the zones, so be conservative | |
1591 | * and assume they are not all lock contended. | |
1592 | */ | |
1593 | all_zones_contended = 0; | |
1594 | break; | |
56de7263 MG |
1595 | } |
1596 | ||
1f9efdef VB |
1597 | /* |
1598 | * If at least one zone wasn't deferred or skipped, we report if all | |
1599 | * zones that were tried were lock contended. | |
1600 | */ | |
1601 | if (rc > COMPACT_SKIPPED && all_zones_contended) | |
1602 | *contended = COMPACT_CONTENDED_LOCK; | |
1603 | ||
56de7263 MG |
1604 | return rc; |
1605 | } | |
1606 | ||
1607 | ||
76ab0f53 | 1608 | /* Compact all zones within a node */ |
7103f16d | 1609 | static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
1610 | { |
1611 | int zoneid; | |
76ab0f53 MG |
1612 | struct zone *zone; |
1613 | ||
76ab0f53 | 1614 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1615 | |
1616 | zone = &pgdat->node_zones[zoneid]; | |
1617 | if (!populated_zone(zone)) | |
1618 | continue; | |
1619 | ||
7be62de9 RR |
1620 | cc->nr_freepages = 0; |
1621 | cc->nr_migratepages = 0; | |
1622 | cc->zone = zone; | |
1623 | INIT_LIST_HEAD(&cc->freepages); | |
1624 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 1625 | |
195b0c60 GK |
1626 | /* |
1627 | * When called via /proc/sys/vm/compact_memory | |
1628 | * this makes sure we compact the whole zone regardless of | |
1629 | * cached scanner positions. | |
1630 | */ | |
21c527a3 | 1631 | if (is_via_compact_memory(cc->order)) |
195b0c60 GK |
1632 | __reset_isolation_suitable(zone); |
1633 | ||
21c527a3 YB |
1634 | if (is_via_compact_memory(cc->order) || |
1635 | !compaction_deferred(zone, cc->order)) | |
7be62de9 | 1636 | compact_zone(zone, cc); |
76ab0f53 | 1637 | |
7be62de9 RR |
1638 | VM_BUG_ON(!list_empty(&cc->freepages)); |
1639 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
75469345 JK |
1640 | |
1641 | if (is_via_compact_memory(cc->order)) | |
1642 | continue; | |
1643 | ||
1644 | if (zone_watermark_ok(zone, cc->order, | |
1645 | low_wmark_pages(zone), 0, 0)) | |
1646 | compaction_defer_reset(zone, cc->order, false); | |
76ab0f53 | 1647 | } |
76ab0f53 MG |
1648 | } |
1649 | ||
7103f16d | 1650 | void compact_pgdat(pg_data_t *pgdat, int order) |
7be62de9 RR |
1651 | { |
1652 | struct compact_control cc = { | |
1653 | .order = order, | |
e0b9daeb | 1654 | .mode = MIGRATE_ASYNC, |
7be62de9 RR |
1655 | }; |
1656 | ||
3a7200af MG |
1657 | if (!order) |
1658 | return; | |
1659 | ||
7103f16d | 1660 | __compact_pgdat(pgdat, &cc); |
7be62de9 RR |
1661 | } |
1662 | ||
7103f16d | 1663 | static void compact_node(int nid) |
7be62de9 | 1664 | { |
7be62de9 RR |
1665 | struct compact_control cc = { |
1666 | .order = -1, | |
e0b9daeb | 1667 | .mode = MIGRATE_SYNC, |
91ca9186 | 1668 | .ignore_skip_hint = true, |
7be62de9 RR |
1669 | }; |
1670 | ||
7103f16d | 1671 | __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
1672 | } |
1673 | ||
76ab0f53 | 1674 | /* Compact all nodes in the system */ |
7964c06d | 1675 | static void compact_nodes(void) |
76ab0f53 MG |
1676 | { |
1677 | int nid; | |
1678 | ||
8575ec29 HD |
1679 | /* Flush pending updates to the LRU lists */ |
1680 | lru_add_drain_all(); | |
1681 | ||
76ab0f53 MG |
1682 | for_each_online_node(nid) |
1683 | compact_node(nid); | |
76ab0f53 MG |
1684 | } |
1685 | ||
1686 | /* The written value is actually unused, all memory is compacted */ | |
1687 | int sysctl_compact_memory; | |
1688 | ||
fec4eb2c YB |
1689 | /* |
1690 | * This is the entry point for compacting all nodes via | |
1691 | * /proc/sys/vm/compact_memory | |
1692 | */ | |
76ab0f53 MG |
1693 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
1694 | void __user *buffer, size_t *length, loff_t *ppos) | |
1695 | { | |
1696 | if (write) | |
7964c06d | 1697 | compact_nodes(); |
76ab0f53 MG |
1698 | |
1699 | return 0; | |
1700 | } | |
ed4a6d7f | 1701 | |
5e771905 MG |
1702 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
1703 | void __user *buffer, size_t *length, loff_t *ppos) | |
1704 | { | |
1705 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1706 | ||
1707 | return 0; | |
1708 | } | |
1709 | ||
ed4a6d7f | 1710 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
74e77fb9 | 1711 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1712 | struct device_attribute *attr, |
ed4a6d7f MG |
1713 | const char *buf, size_t count) |
1714 | { | |
8575ec29 HD |
1715 | int nid = dev->id; |
1716 | ||
1717 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1718 | /* Flush pending updates to the LRU lists */ | |
1719 | lru_add_drain_all(); | |
1720 | ||
1721 | compact_node(nid); | |
1722 | } | |
ed4a6d7f MG |
1723 | |
1724 | return count; | |
1725 | } | |
10fbcf4c | 1726 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1727 | |
1728 | int compaction_register_node(struct node *node) | |
1729 | { | |
10fbcf4c | 1730 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1731 | } |
1732 | ||
1733 | void compaction_unregister_node(struct node *node) | |
1734 | { | |
10fbcf4c | 1735 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1736 | } |
1737 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 1738 | |
698b1b30 VB |
1739 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
1740 | { | |
172400c6 | 1741 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
1742 | } |
1743 | ||
1744 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
1745 | { | |
1746 | int zoneid; | |
1747 | struct zone *zone; | |
1748 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
1749 | ||
1750 | for (zoneid = 0; zoneid < classzone_idx; zoneid++) { | |
1751 | zone = &pgdat->node_zones[zoneid]; | |
1752 | ||
1753 | if (!populated_zone(zone)) | |
1754 | continue; | |
1755 | ||
1756 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
1757 | classzone_idx) == COMPACT_CONTINUE) | |
1758 | return true; | |
1759 | } | |
1760 | ||
1761 | return false; | |
1762 | } | |
1763 | ||
1764 | static void kcompactd_do_work(pg_data_t *pgdat) | |
1765 | { | |
1766 | /* | |
1767 | * With no special task, compact all zones so that a page of requested | |
1768 | * order is allocatable. | |
1769 | */ | |
1770 | int zoneid; | |
1771 | struct zone *zone; | |
1772 | struct compact_control cc = { | |
1773 | .order = pgdat->kcompactd_max_order, | |
1774 | .classzone_idx = pgdat->kcompactd_classzone_idx, | |
1775 | .mode = MIGRATE_SYNC_LIGHT, | |
1776 | .ignore_skip_hint = true, | |
1777 | ||
1778 | }; | |
1779 | bool success = false; | |
1780 | ||
1781 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, | |
1782 | cc.classzone_idx); | |
1783 | count_vm_event(KCOMPACTD_WAKE); | |
1784 | ||
1785 | for (zoneid = 0; zoneid < cc.classzone_idx; zoneid++) { | |
1786 | int status; | |
1787 | ||
1788 | zone = &pgdat->node_zones[zoneid]; | |
1789 | if (!populated_zone(zone)) | |
1790 | continue; | |
1791 | ||
1792 | if (compaction_deferred(zone, cc.order)) | |
1793 | continue; | |
1794 | ||
1795 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
1796 | COMPACT_CONTINUE) | |
1797 | continue; | |
1798 | ||
1799 | cc.nr_freepages = 0; | |
1800 | cc.nr_migratepages = 0; | |
1801 | cc.zone = zone; | |
1802 | INIT_LIST_HEAD(&cc.freepages); | |
1803 | INIT_LIST_HEAD(&cc.migratepages); | |
1804 | ||
172400c6 VB |
1805 | if (kthread_should_stop()) |
1806 | return; | |
698b1b30 VB |
1807 | status = compact_zone(zone, &cc); |
1808 | ||
1809 | if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone), | |
1810 | cc.classzone_idx, 0)) { | |
1811 | success = true; | |
1812 | compaction_defer_reset(zone, cc.order, false); | |
1813 | } else if (status == COMPACT_COMPLETE) { | |
1814 | /* | |
1815 | * We use sync migration mode here, so we defer like | |
1816 | * sync direct compaction does. | |
1817 | */ | |
1818 | defer_compaction(zone, cc.order); | |
1819 | } | |
1820 | ||
1821 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1822 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1823 | } | |
1824 | ||
1825 | /* | |
1826 | * Regardless of success, we are done until woken up next. But remember | |
1827 | * the requested order/classzone_idx in case it was higher/tighter than | |
1828 | * our current ones | |
1829 | */ | |
1830 | if (pgdat->kcompactd_max_order <= cc.order) | |
1831 | pgdat->kcompactd_max_order = 0; | |
1832 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
1833 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
1834 | } | |
1835 | ||
1836 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
1837 | { | |
1838 | if (!order) | |
1839 | return; | |
1840 | ||
1841 | if (pgdat->kcompactd_max_order < order) | |
1842 | pgdat->kcompactd_max_order = order; | |
1843 | ||
1844 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
1845 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
1846 | ||
1847 | if (!waitqueue_active(&pgdat->kcompactd_wait)) | |
1848 | return; | |
1849 | ||
1850 | if (!kcompactd_node_suitable(pgdat)) | |
1851 | return; | |
1852 | ||
1853 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
1854 | classzone_idx); | |
1855 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
1856 | } | |
1857 | ||
1858 | /* | |
1859 | * The background compaction daemon, started as a kernel thread | |
1860 | * from the init process. | |
1861 | */ | |
1862 | static int kcompactd(void *p) | |
1863 | { | |
1864 | pg_data_t *pgdat = (pg_data_t*)p; | |
1865 | struct task_struct *tsk = current; | |
1866 | ||
1867 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
1868 | ||
1869 | if (!cpumask_empty(cpumask)) | |
1870 | set_cpus_allowed_ptr(tsk, cpumask); | |
1871 | ||
1872 | set_freezable(); | |
1873 | ||
1874 | pgdat->kcompactd_max_order = 0; | |
1875 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
1876 | ||
1877 | while (!kthread_should_stop()) { | |
1878 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); | |
1879 | wait_event_freezable(pgdat->kcompactd_wait, | |
1880 | kcompactd_work_requested(pgdat)); | |
1881 | ||
1882 | kcompactd_do_work(pgdat); | |
1883 | } | |
1884 | ||
1885 | return 0; | |
1886 | } | |
1887 | ||
1888 | /* | |
1889 | * This kcompactd start function will be called by init and node-hot-add. | |
1890 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
1891 | */ | |
1892 | int kcompactd_run(int nid) | |
1893 | { | |
1894 | pg_data_t *pgdat = NODE_DATA(nid); | |
1895 | int ret = 0; | |
1896 | ||
1897 | if (pgdat->kcompactd) | |
1898 | return 0; | |
1899 | ||
1900 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
1901 | if (IS_ERR(pgdat->kcompactd)) { | |
1902 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
1903 | ret = PTR_ERR(pgdat->kcompactd); | |
1904 | pgdat->kcompactd = NULL; | |
1905 | } | |
1906 | return ret; | |
1907 | } | |
1908 | ||
1909 | /* | |
1910 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
1911 | * hold mem_hotplug_begin/end(). | |
1912 | */ | |
1913 | void kcompactd_stop(int nid) | |
1914 | { | |
1915 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
1916 | ||
1917 | if (kcompactd) { | |
1918 | kthread_stop(kcompactd); | |
1919 | NODE_DATA(nid)->kcompactd = NULL; | |
1920 | } | |
1921 | } | |
1922 | ||
1923 | /* | |
1924 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
1925 | * not required for correctness. So if the last cpu in a node goes | |
1926 | * away, we get changed to run anywhere: as the first one comes back, | |
1927 | * restore their cpu bindings. | |
1928 | */ | |
1929 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, | |
1930 | void *hcpu) | |
1931 | { | |
1932 | int nid; | |
1933 | ||
1934 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
1935 | for_each_node_state(nid, N_MEMORY) { | |
1936 | pg_data_t *pgdat = NODE_DATA(nid); | |
1937 | const struct cpumask *mask; | |
1938 | ||
1939 | mask = cpumask_of_node(pgdat->node_id); | |
1940 | ||
1941 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
1942 | /* One of our CPUs online: restore mask */ | |
1943 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
1944 | } | |
1945 | } | |
1946 | return NOTIFY_OK; | |
1947 | } | |
1948 | ||
1949 | static int __init kcompactd_init(void) | |
1950 | { | |
1951 | int nid; | |
1952 | ||
1953 | for_each_node_state(nid, N_MEMORY) | |
1954 | kcompactd_run(nid); | |
1955 | hotcpu_notifier(cpu_callback, 0); | |
1956 | return 0; | |
1957 | } | |
1958 | subsys_initcall(kcompactd_init) | |
1959 | ||
ff9543fd | 1960 | #endif /* CONFIG_COMPACTION */ |