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 | 83 | |
bda807d4 MK |
84 | int PageMovable(struct page *page) |
85 | { | |
86 | struct address_space *mapping; | |
87 | ||
88 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
89 | if (!__PageMovable(page)) | |
90 | return 0; | |
91 | ||
92 | mapping = page_mapping(page); | |
93 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
94 | return 1; | |
95 | ||
96 | return 0; | |
97 | } | |
98 | EXPORT_SYMBOL(PageMovable); | |
99 | ||
100 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
101 | { | |
102 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
103 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
104 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
105 | } | |
106 | EXPORT_SYMBOL(__SetPageMovable); | |
107 | ||
108 | void __ClearPageMovable(struct page *page) | |
109 | { | |
110 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
111 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
112 | /* | |
113 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
114 | * flag so that VM can catch up released page by driver after isolation. | |
115 | * With it, VM migration doesn't try to put it back. | |
116 | */ | |
117 | page->mapping = (void *)((unsigned long)page->mapping & | |
118 | PAGE_MAPPING_MOVABLE); | |
119 | } | |
120 | EXPORT_SYMBOL(__ClearPageMovable); | |
121 | ||
24e2716f JK |
122 | /* Do not skip compaction more than 64 times */ |
123 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
124 | ||
125 | /* | |
126 | * Compaction is deferred when compaction fails to result in a page | |
127 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
128 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
129 | */ | |
130 | void defer_compaction(struct zone *zone, int order) | |
131 | { | |
132 | zone->compact_considered = 0; | |
133 | zone->compact_defer_shift++; | |
134 | ||
135 | if (order < zone->compact_order_failed) | |
136 | zone->compact_order_failed = order; | |
137 | ||
138 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
139 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
140 | ||
141 | trace_mm_compaction_defer_compaction(zone, order); | |
142 | } | |
143 | ||
144 | /* Returns true if compaction should be skipped this time */ | |
145 | bool compaction_deferred(struct zone *zone, int order) | |
146 | { | |
147 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
148 | ||
149 | if (order < zone->compact_order_failed) | |
150 | return false; | |
151 | ||
152 | /* Avoid possible overflow */ | |
153 | if (++zone->compact_considered > defer_limit) | |
154 | zone->compact_considered = defer_limit; | |
155 | ||
156 | if (zone->compact_considered >= defer_limit) | |
157 | return false; | |
158 | ||
159 | trace_mm_compaction_deferred(zone, order); | |
160 | ||
161 | return true; | |
162 | } | |
163 | ||
164 | /* | |
165 | * Update defer tracking counters after successful compaction of given order, | |
166 | * which means an allocation either succeeded (alloc_success == true) or is | |
167 | * expected to succeed. | |
168 | */ | |
169 | void compaction_defer_reset(struct zone *zone, int order, | |
170 | bool alloc_success) | |
171 | { | |
172 | if (alloc_success) { | |
173 | zone->compact_considered = 0; | |
174 | zone->compact_defer_shift = 0; | |
175 | } | |
176 | if (order >= zone->compact_order_failed) | |
177 | zone->compact_order_failed = order + 1; | |
178 | ||
179 | trace_mm_compaction_defer_reset(zone, order); | |
180 | } | |
181 | ||
182 | /* Returns true if restarting compaction after many failures */ | |
183 | bool compaction_restarting(struct zone *zone, int order) | |
184 | { | |
185 | if (order < zone->compact_order_failed) | |
186 | return false; | |
187 | ||
188 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
189 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
190 | } | |
191 | ||
bb13ffeb MG |
192 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
193 | static inline bool isolation_suitable(struct compact_control *cc, | |
194 | struct page *page) | |
195 | { | |
196 | if (cc->ignore_skip_hint) | |
197 | return true; | |
198 | ||
199 | return !get_pageblock_skip(page); | |
200 | } | |
201 | ||
02333641 VB |
202 | static void reset_cached_positions(struct zone *zone) |
203 | { | |
204 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
205 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 206 | zone->compact_cached_free_pfn = |
06b6640a | 207 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
208 | } |
209 | ||
bb13ffeb MG |
210 | /* |
211 | * This function is called to clear all cached information on pageblocks that | |
212 | * should be skipped for page isolation when the migrate and free page scanner | |
213 | * meet. | |
214 | */ | |
62997027 | 215 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
216 | { |
217 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 218 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
219 | unsigned long pfn; |
220 | ||
62997027 | 221 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
222 | |
223 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
224 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
225 | struct page *page; | |
226 | ||
227 | cond_resched(); | |
228 | ||
229 | if (!pfn_valid(pfn)) | |
230 | continue; | |
231 | ||
232 | page = pfn_to_page(pfn); | |
233 | if (zone != page_zone(page)) | |
234 | continue; | |
235 | ||
236 | clear_pageblock_skip(page); | |
237 | } | |
02333641 VB |
238 | |
239 | reset_cached_positions(zone); | |
bb13ffeb MG |
240 | } |
241 | ||
62997027 MG |
242 | void reset_isolation_suitable(pg_data_t *pgdat) |
243 | { | |
244 | int zoneid; | |
245 | ||
246 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
247 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
248 | if (!populated_zone(zone)) | |
249 | continue; | |
250 | ||
251 | /* Only flush if a full compaction finished recently */ | |
252 | if (zone->compact_blockskip_flush) | |
253 | __reset_isolation_suitable(zone); | |
254 | } | |
255 | } | |
256 | ||
bb13ffeb MG |
257 | /* |
258 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 259 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 260 | */ |
c89511ab MG |
261 | static void update_pageblock_skip(struct compact_control *cc, |
262 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 263 | bool migrate_scanner) |
bb13ffeb | 264 | { |
c89511ab | 265 | struct zone *zone = cc->zone; |
35979ef3 | 266 | unsigned long pfn; |
6815bf3f JK |
267 | |
268 | if (cc->ignore_skip_hint) | |
269 | return; | |
270 | ||
bb13ffeb MG |
271 | if (!page) |
272 | return; | |
273 | ||
35979ef3 DR |
274 | if (nr_isolated) |
275 | return; | |
276 | ||
edc2ca61 | 277 | set_pageblock_skip(page); |
c89511ab | 278 | |
35979ef3 DR |
279 | pfn = page_to_pfn(page); |
280 | ||
281 | /* Update where async and sync compaction should restart */ | |
282 | if (migrate_scanner) { | |
35979ef3 DR |
283 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
284 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
285 | if (cc->mode != MIGRATE_ASYNC && |
286 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
287 | zone->compact_cached_migrate_pfn[1] = pfn; |
288 | } else { | |
35979ef3 DR |
289 | if (pfn < zone->compact_cached_free_pfn) |
290 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 291 | } |
bb13ffeb MG |
292 | } |
293 | #else | |
294 | static inline bool isolation_suitable(struct compact_control *cc, | |
295 | struct page *page) | |
296 | { | |
297 | return true; | |
298 | } | |
299 | ||
c89511ab MG |
300 | static void update_pageblock_skip(struct compact_control *cc, |
301 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 302 | bool migrate_scanner) |
bb13ffeb MG |
303 | { |
304 | } | |
305 | #endif /* CONFIG_COMPACTION */ | |
306 | ||
8b44d279 VB |
307 | /* |
308 | * Compaction requires the taking of some coarse locks that are potentially | |
309 | * very heavily contended. For async compaction, back out if the lock cannot | |
310 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
311 | * | |
312 | * Returns true if the lock is held | |
313 | * Returns false if the lock is not held and compaction should abort | |
314 | */ | |
315 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
316 | struct compact_control *cc) | |
2a1402aa | 317 | { |
8b44d279 VB |
318 | if (cc->mode == MIGRATE_ASYNC) { |
319 | if (!spin_trylock_irqsave(lock, *flags)) { | |
320 | cc->contended = COMPACT_CONTENDED_LOCK; | |
321 | return false; | |
322 | } | |
323 | } else { | |
324 | spin_lock_irqsave(lock, *flags); | |
325 | } | |
1f9efdef | 326 | |
8b44d279 | 327 | return true; |
2a1402aa MG |
328 | } |
329 | ||
c67fe375 MG |
330 | /* |
331 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
332 | * very heavily contended. The lock should be periodically unlocked to avoid |
333 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
334 | * the lock. It might also be that allowing the IRQs will result in | |
335 | * need_resched() becoming true. If scheduling is needed, async compaction | |
336 | * aborts. Sync compaction schedules. | |
337 | * Either compaction type will also abort if a fatal signal is pending. | |
338 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 339 | * |
8b44d279 VB |
340 | * Returns true if compaction should abort due to fatal signal pending, or |
341 | * async compaction due to need_resched() | |
342 | * Returns false when compaction can continue (sync compaction might have | |
343 | * scheduled) | |
c67fe375 | 344 | */ |
8b44d279 VB |
345 | static bool compact_unlock_should_abort(spinlock_t *lock, |
346 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 347 | { |
8b44d279 VB |
348 | if (*locked) { |
349 | spin_unlock_irqrestore(lock, flags); | |
350 | *locked = false; | |
351 | } | |
1f9efdef | 352 | |
8b44d279 VB |
353 | if (fatal_signal_pending(current)) { |
354 | cc->contended = COMPACT_CONTENDED_SCHED; | |
355 | return true; | |
356 | } | |
c67fe375 | 357 | |
8b44d279 | 358 | if (need_resched()) { |
e0b9daeb | 359 | if (cc->mode == MIGRATE_ASYNC) { |
8b44d279 VB |
360 | cc->contended = COMPACT_CONTENDED_SCHED; |
361 | return true; | |
c67fe375 | 362 | } |
c67fe375 | 363 | cond_resched(); |
c67fe375 MG |
364 | } |
365 | ||
8b44d279 | 366 | return false; |
c67fe375 MG |
367 | } |
368 | ||
be976572 VB |
369 | /* |
370 | * Aside from avoiding lock contention, compaction also periodically checks | |
371 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 372 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
373 | * is used where no lock is concerned. |
374 | * | |
375 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
376 | * Returns true when async compaction should abort. | |
377 | */ | |
378 | static inline bool compact_should_abort(struct compact_control *cc) | |
379 | { | |
380 | /* async compaction aborts if contended */ | |
381 | if (need_resched()) { | |
382 | if (cc->mode == MIGRATE_ASYNC) { | |
1f9efdef | 383 | cc->contended = COMPACT_CONTENDED_SCHED; |
be976572 VB |
384 | return true; |
385 | } | |
386 | ||
387 | cond_resched(); | |
388 | } | |
389 | ||
390 | return false; | |
391 | } | |
392 | ||
85aa125f | 393 | /* |
9e4be470 JM |
394 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
395 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
396 | * (even though it may still end up isolating some pages). | |
85aa125f | 397 | */ |
f40d1e42 | 398 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 399 | unsigned long *start_pfn, |
85aa125f MN |
400 | unsigned long end_pfn, |
401 | struct list_head *freelist, | |
402 | bool strict) | |
748446bb | 403 | { |
b7aba698 | 404 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 405 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 406 | unsigned long flags = 0; |
f40d1e42 | 407 | bool locked = false; |
e14c720e | 408 | unsigned long blockpfn = *start_pfn; |
748446bb | 409 | |
748446bb MG |
410 | cursor = pfn_to_page(blockpfn); |
411 | ||
f40d1e42 | 412 | /* Isolate free pages. */ |
748446bb MG |
413 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
414 | int isolated, i; | |
415 | struct page *page = cursor; | |
416 | ||
8b44d279 VB |
417 | /* |
418 | * Periodically drop the lock (if held) regardless of its | |
419 | * contention, to give chance to IRQs. Abort if fatal signal | |
420 | * pending or async compaction detects need_resched() | |
421 | */ | |
422 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
423 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
424 | &locked, cc)) | |
425 | break; | |
426 | ||
b7aba698 | 427 | nr_scanned++; |
f40d1e42 | 428 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
429 | goto isolate_fail; |
430 | ||
bb13ffeb MG |
431 | if (!valid_page) |
432 | valid_page = page; | |
9fcd6d2e VB |
433 | |
434 | /* | |
435 | * For compound pages such as THP and hugetlbfs, we can save | |
436 | * potentially a lot of iterations if we skip them at once. | |
437 | * The check is racy, but we can consider only valid values | |
438 | * and the only danger is skipping too much. | |
439 | */ | |
440 | if (PageCompound(page)) { | |
441 | unsigned int comp_order = compound_order(page); | |
442 | ||
443 | if (likely(comp_order < MAX_ORDER)) { | |
444 | blockpfn += (1UL << comp_order) - 1; | |
445 | cursor += (1UL << comp_order) - 1; | |
446 | } | |
447 | ||
448 | goto isolate_fail; | |
449 | } | |
450 | ||
f40d1e42 | 451 | if (!PageBuddy(page)) |
2af120bc | 452 | goto isolate_fail; |
f40d1e42 MG |
453 | |
454 | /* | |
69b7189f VB |
455 | * If we already hold the lock, we can skip some rechecking. |
456 | * Note that if we hold the lock now, checked_pageblock was | |
457 | * already set in some previous iteration (or strict is true), | |
458 | * so it is correct to skip the suitable migration target | |
459 | * recheck as well. | |
f40d1e42 | 460 | */ |
69b7189f VB |
461 | if (!locked) { |
462 | /* | |
463 | * The zone lock must be held to isolate freepages. | |
464 | * Unfortunately this is a very coarse lock and can be | |
465 | * heavily contended if there are parallel allocations | |
466 | * or parallel compactions. For async compaction do not | |
467 | * spin on the lock and we acquire the lock as late as | |
468 | * possible. | |
469 | */ | |
8b44d279 VB |
470 | locked = compact_trylock_irqsave(&cc->zone->lock, |
471 | &flags, cc); | |
69b7189f VB |
472 | if (!locked) |
473 | break; | |
f40d1e42 | 474 | |
69b7189f VB |
475 | /* Recheck this is a buddy page under lock */ |
476 | if (!PageBuddy(page)) | |
477 | goto isolate_fail; | |
478 | } | |
748446bb MG |
479 | |
480 | /* Found a free page, break it into order-0 pages */ | |
481 | isolated = split_free_page(page); | |
a4f04f2c DR |
482 | if (!isolated) |
483 | break; | |
484 | ||
748446bb | 485 | total_isolated += isolated; |
a4f04f2c | 486 | cc->nr_freepages += isolated; |
748446bb MG |
487 | for (i = 0; i < isolated; i++) { |
488 | list_add(&page->lru, freelist); | |
489 | page++; | |
490 | } | |
a4f04f2c DR |
491 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
492 | blockpfn += isolated; | |
493 | break; | |
748446bb | 494 | } |
a4f04f2c DR |
495 | /* Advance to the end of split page */ |
496 | blockpfn += isolated - 1; | |
497 | cursor += isolated - 1; | |
498 | continue; | |
2af120bc LA |
499 | |
500 | isolate_fail: | |
501 | if (strict) | |
502 | break; | |
503 | else | |
504 | continue; | |
505 | ||
748446bb MG |
506 | } |
507 | ||
a4f04f2c DR |
508 | if (locked) |
509 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
510 | ||
9fcd6d2e VB |
511 | /* |
512 | * There is a tiny chance that we have read bogus compound_order(), | |
513 | * so be careful to not go outside of the pageblock. | |
514 | */ | |
515 | if (unlikely(blockpfn > end_pfn)) | |
516 | blockpfn = end_pfn; | |
517 | ||
e34d85f0 JK |
518 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
519 | nr_scanned, total_isolated); | |
520 | ||
e14c720e VB |
521 | /* Record how far we have got within the block */ |
522 | *start_pfn = blockpfn; | |
523 | ||
f40d1e42 MG |
524 | /* |
525 | * If strict isolation is requested by CMA then check that all the | |
526 | * pages requested were isolated. If there were any failures, 0 is | |
527 | * returned and CMA will fail. | |
528 | */ | |
2af120bc | 529 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
530 | total_isolated = 0; |
531 | ||
bb13ffeb MG |
532 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
533 | if (blockpfn == end_pfn) | |
edc2ca61 | 534 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 535 | |
010fc29a | 536 | count_compact_events(COMPACTFREE_SCANNED, nr_scanned); |
397487db | 537 | if (total_isolated) |
010fc29a | 538 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
539 | return total_isolated; |
540 | } | |
541 | ||
85aa125f MN |
542 | /** |
543 | * isolate_freepages_range() - isolate free pages. | |
544 | * @start_pfn: The first PFN to start isolating. | |
545 | * @end_pfn: The one-past-last PFN. | |
546 | * | |
547 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
548 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
549 | * undo its actions and return zero. | |
550 | * | |
551 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
552 | * (which may be greater then end_pfn if end fell in a middle of | |
553 | * a free page). | |
554 | */ | |
ff9543fd | 555 | unsigned long |
bb13ffeb MG |
556 | isolate_freepages_range(struct compact_control *cc, |
557 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 558 | { |
e1409c32 | 559 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
560 | LIST_HEAD(freelist); |
561 | ||
7d49d886 | 562 | pfn = start_pfn; |
06b6640a | 563 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
564 | if (block_start_pfn < cc->zone->zone_start_pfn) |
565 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 566 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
567 | |
568 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 569 | block_start_pfn = block_end_pfn, |
7d49d886 | 570 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
571 | /* Protect pfn from changing by isolate_freepages_block */ |
572 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 573 | |
85aa125f MN |
574 | block_end_pfn = min(block_end_pfn, end_pfn); |
575 | ||
58420016 JK |
576 | /* |
577 | * pfn could pass the block_end_pfn if isolated freepage | |
578 | * is more than pageblock order. In this case, we adjust | |
579 | * scanning range to right one. | |
580 | */ | |
581 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
582 | block_start_pfn = pageblock_start_pfn(pfn); |
583 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
584 | block_end_pfn = min(block_end_pfn, end_pfn); |
585 | } | |
586 | ||
e1409c32 JK |
587 | if (!pageblock_pfn_to_page(block_start_pfn, |
588 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
589 | break; |
590 | ||
e14c720e VB |
591 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
592 | block_end_pfn, &freelist, true); | |
85aa125f MN |
593 | |
594 | /* | |
595 | * In strict mode, isolate_freepages_block() returns 0 if | |
596 | * there are any holes in the block (ie. invalid PFNs or | |
597 | * non-free pages). | |
598 | */ | |
599 | if (!isolated) | |
600 | break; | |
601 | ||
602 | /* | |
603 | * If we managed to isolate pages, it is always (1 << n) * | |
604 | * pageblock_nr_pages for some non-negative n. (Max order | |
605 | * page may span two pageblocks). | |
606 | */ | |
607 | } | |
608 | ||
609 | /* split_free_page does not map the pages */ | |
610 | map_pages(&freelist); | |
611 | ||
612 | if (pfn < end_pfn) { | |
613 | /* Loop terminated early, cleanup. */ | |
614 | release_freepages(&freelist); | |
615 | return 0; | |
616 | } | |
617 | ||
618 | /* We don't use freelists for anything. */ | |
619 | return pfn; | |
620 | } | |
621 | ||
748446bb | 622 | /* Update the number of anon and file isolated pages in the zone */ |
edc2ca61 | 623 | static void acct_isolated(struct zone *zone, struct compact_control *cc) |
748446bb MG |
624 | { |
625 | struct page *page; | |
b9e84ac1 | 626 | unsigned int count[2] = { 0, }; |
748446bb | 627 | |
edc2ca61 VB |
628 | if (list_empty(&cc->migratepages)) |
629 | return; | |
630 | ||
b9e84ac1 MK |
631 | list_for_each_entry(page, &cc->migratepages, lru) |
632 | count[!!page_is_file_cache(page)]++; | |
748446bb | 633 | |
edc2ca61 VB |
634 | mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
635 | mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
636 | } |
637 | ||
638 | /* Similar to reclaim, but different enough that they don't share logic */ | |
639 | static bool too_many_isolated(struct zone *zone) | |
640 | { | |
bc693045 | 641 | unsigned long active, inactive, isolated; |
748446bb MG |
642 | |
643 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
644 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
645 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
646 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
647 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
648 | zone_page_state(zone, NR_ISOLATED_ANON); | |
649 | ||
bc693045 | 650 | return isolated > (inactive + active) / 2; |
748446bb MG |
651 | } |
652 | ||
2fe86e00 | 653 | /** |
edc2ca61 VB |
654 | * isolate_migratepages_block() - isolate all migrate-able pages within |
655 | * a single pageblock | |
2fe86e00 | 656 | * @cc: Compaction control structure. |
edc2ca61 VB |
657 | * @low_pfn: The first PFN to isolate |
658 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
659 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
660 | * |
661 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
662 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
663 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
664 | * first page that was not scanned (which may be both less, equal to or more | |
665 | * than end_pfn). | |
2fe86e00 | 666 | * |
edc2ca61 VB |
667 | * The pages are isolated on cc->migratepages list (not required to be empty), |
668 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
669 | * is neither read nor updated. | |
748446bb | 670 | */ |
edc2ca61 VB |
671 | static unsigned long |
672 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
673 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 674 | { |
edc2ca61 | 675 | struct zone *zone = cc->zone; |
b7aba698 | 676 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 677 | struct lruvec *lruvec; |
b8b2d825 | 678 | unsigned long flags = 0; |
2a1402aa | 679 | bool locked = false; |
bb13ffeb | 680 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 681 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
682 | bool skip_on_failure = false; |
683 | unsigned long next_skip_pfn = 0; | |
748446bb | 684 | |
748446bb MG |
685 | /* |
686 | * Ensure that there are not too many pages isolated from the LRU | |
687 | * list by either parallel reclaimers or compaction. If there are, | |
688 | * delay for some time until fewer pages are isolated | |
689 | */ | |
690 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 691 | /* async migration should just abort */ |
e0b9daeb | 692 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 693 | return 0; |
f9e35b3b | 694 | |
748446bb MG |
695 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
696 | ||
697 | if (fatal_signal_pending(current)) | |
2fe86e00 | 698 | return 0; |
748446bb MG |
699 | } |
700 | ||
be976572 VB |
701 | if (compact_should_abort(cc)) |
702 | return 0; | |
aeef4b83 | 703 | |
fdd048e1 VB |
704 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
705 | skip_on_failure = true; | |
706 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
707 | } | |
708 | ||
748446bb | 709 | /* Time to isolate some pages for migration */ |
748446bb | 710 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 711 | |
fdd048e1 VB |
712 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
713 | /* | |
714 | * We have isolated all migration candidates in the | |
715 | * previous order-aligned block, and did not skip it due | |
716 | * to failure. We should migrate the pages now and | |
717 | * hopefully succeed compaction. | |
718 | */ | |
719 | if (nr_isolated) | |
720 | break; | |
721 | ||
722 | /* | |
723 | * We failed to isolate in the previous order-aligned | |
724 | * block. Set the new boundary to the end of the | |
725 | * current block. Note we can't simply increase | |
726 | * next_skip_pfn by 1 << order, as low_pfn might have | |
727 | * been incremented by a higher number due to skipping | |
728 | * a compound or a high-order buddy page in the | |
729 | * previous loop iteration. | |
730 | */ | |
731 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
732 | } | |
733 | ||
8b44d279 VB |
734 | /* |
735 | * Periodically drop the lock (if held) regardless of its | |
736 | * contention, to give chance to IRQs. Abort async compaction | |
737 | * if contended. | |
738 | */ | |
739 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
740 | && compact_unlock_should_abort(&zone->lru_lock, flags, | |
741 | &locked, cc)) | |
742 | break; | |
c67fe375 | 743 | |
748446bb | 744 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 745 | goto isolate_fail; |
b7aba698 | 746 | nr_scanned++; |
748446bb | 747 | |
748446bb | 748 | page = pfn_to_page(low_pfn); |
dc908600 | 749 | |
bb13ffeb MG |
750 | if (!valid_page) |
751 | valid_page = page; | |
752 | ||
6c14466c | 753 | /* |
99c0fd5e VB |
754 | * Skip if free. We read page order here without zone lock |
755 | * which is generally unsafe, but the race window is small and | |
756 | * the worst thing that can happen is that we skip some | |
757 | * potential isolation targets. | |
6c14466c | 758 | */ |
99c0fd5e VB |
759 | if (PageBuddy(page)) { |
760 | unsigned long freepage_order = page_order_unsafe(page); | |
761 | ||
762 | /* | |
763 | * Without lock, we cannot be sure that what we got is | |
764 | * a valid page order. Consider only values in the | |
765 | * valid order range to prevent low_pfn overflow. | |
766 | */ | |
767 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
768 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 769 | continue; |
99c0fd5e | 770 | } |
748446bb | 771 | |
bc835011 | 772 | /* |
29c0dde8 VB |
773 | * Regardless of being on LRU, compound pages such as THP and |
774 | * hugetlbfs are not to be compacted. We can potentially save | |
775 | * a lot of iterations if we skip them at once. The check is | |
776 | * racy, but we can consider only valid values and the only | |
777 | * danger is skipping too much. | |
bc835011 | 778 | */ |
29c0dde8 VB |
779 | if (PageCompound(page)) { |
780 | unsigned int comp_order = compound_order(page); | |
781 | ||
782 | if (likely(comp_order < MAX_ORDER)) | |
783 | low_pfn += (1UL << comp_order) - 1; | |
edc2ca61 | 784 | |
fdd048e1 | 785 | goto isolate_fail; |
2a1402aa MG |
786 | } |
787 | ||
bda807d4 MK |
788 | /* |
789 | * Check may be lockless but that's ok as we recheck later. | |
790 | * It's possible to migrate LRU and non-lru movable pages. | |
791 | * Skip any other type of page | |
792 | */ | |
793 | if (!PageLRU(page)) { | |
794 | if (unlikely(balloon_page_movable(page))) { | |
795 | if (balloon_page_isolate(page)) { | |
796 | /* Successfully isolated */ | |
797 | goto isolate_success; | |
798 | } | |
799 | } | |
800 | ||
801 | /* | |
802 | * __PageMovable can return false positive so we need | |
803 | * to verify it under page_lock. | |
804 | */ | |
805 | if (unlikely(__PageMovable(page)) && | |
806 | !PageIsolated(page)) { | |
807 | if (locked) { | |
808 | spin_unlock_irqrestore(&zone->lru_lock, | |
809 | flags); | |
810 | locked = false; | |
811 | } | |
812 | ||
813 | if (isolate_movable_page(page, isolate_mode)) | |
814 | goto isolate_success; | |
815 | } | |
816 | ||
fdd048e1 | 817 | goto isolate_fail; |
bda807d4 | 818 | } |
29c0dde8 | 819 | |
119d6d59 DR |
820 | /* |
821 | * Migration will fail if an anonymous page is pinned in memory, | |
822 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
823 | * admittedly racy check. | |
824 | */ | |
825 | if (!page_mapping(page) && | |
826 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 827 | goto isolate_fail; |
119d6d59 | 828 | |
69b7189f VB |
829 | /* If we already hold the lock, we can skip some rechecking */ |
830 | if (!locked) { | |
8b44d279 VB |
831 | locked = compact_trylock_irqsave(&zone->lru_lock, |
832 | &flags, cc); | |
69b7189f VB |
833 | if (!locked) |
834 | break; | |
2a1402aa | 835 | |
29c0dde8 | 836 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 837 | if (!PageLRU(page)) |
fdd048e1 | 838 | goto isolate_fail; |
29c0dde8 VB |
839 | |
840 | /* | |
841 | * Page become compound since the non-locked check, | |
842 | * and it's on LRU. It can only be a THP so the order | |
843 | * is safe to read and it's 0 for tail pages. | |
844 | */ | |
845 | if (unlikely(PageCompound(page))) { | |
846 | low_pfn += (1UL << compound_order(page)) - 1; | |
fdd048e1 | 847 | goto isolate_fail; |
69b7189f | 848 | } |
bc835011 AA |
849 | } |
850 | ||
fa9add64 HD |
851 | lruvec = mem_cgroup_page_lruvec(page, zone); |
852 | ||
748446bb | 853 | /* Try isolate the page */ |
edc2ca61 | 854 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 855 | goto isolate_fail; |
748446bb | 856 | |
29c0dde8 | 857 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 858 | |
748446bb | 859 | /* Successfully isolated */ |
fa9add64 | 860 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
b6c75016 JK |
861 | |
862 | isolate_success: | |
fdd048e1 | 863 | list_add(&page->lru, &cc->migratepages); |
748446bb | 864 | cc->nr_migratepages++; |
b7aba698 | 865 | nr_isolated++; |
748446bb | 866 | |
a34753d2 VB |
867 | /* |
868 | * Record where we could have freed pages by migration and not | |
869 | * yet flushed them to buddy allocator. | |
870 | * - this is the lowest page that was isolated and likely be | |
871 | * then freed by migration. | |
872 | */ | |
873 | if (!cc->last_migrated_pfn) | |
874 | cc->last_migrated_pfn = low_pfn; | |
875 | ||
748446bb | 876 | /* Avoid isolating too much */ |
31b8384a HD |
877 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
878 | ++low_pfn; | |
748446bb | 879 | break; |
31b8384a | 880 | } |
fdd048e1 VB |
881 | |
882 | continue; | |
883 | isolate_fail: | |
884 | if (!skip_on_failure) | |
885 | continue; | |
886 | ||
887 | /* | |
888 | * We have isolated some pages, but then failed. Release them | |
889 | * instead of migrating, as we cannot form the cc->order buddy | |
890 | * page anyway. | |
891 | */ | |
892 | if (nr_isolated) { | |
893 | if (locked) { | |
894 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
895 | locked = false; | |
896 | } | |
897 | acct_isolated(zone, cc); | |
898 | putback_movable_pages(&cc->migratepages); | |
899 | cc->nr_migratepages = 0; | |
900 | cc->last_migrated_pfn = 0; | |
901 | nr_isolated = 0; | |
902 | } | |
903 | ||
904 | if (low_pfn < next_skip_pfn) { | |
905 | low_pfn = next_skip_pfn - 1; | |
906 | /* | |
907 | * The check near the loop beginning would have updated | |
908 | * next_skip_pfn too, but this is a bit simpler. | |
909 | */ | |
910 | next_skip_pfn += 1UL << cc->order; | |
911 | } | |
748446bb MG |
912 | } |
913 | ||
99c0fd5e VB |
914 | /* |
915 | * The PageBuddy() check could have potentially brought us outside | |
916 | * the range to be scanned. | |
917 | */ | |
918 | if (unlikely(low_pfn > end_pfn)) | |
919 | low_pfn = end_pfn; | |
920 | ||
c67fe375 MG |
921 | if (locked) |
922 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
748446bb | 923 | |
50b5b094 VB |
924 | /* |
925 | * Update the pageblock-skip information and cached scanner pfn, | |
926 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 927 | */ |
35979ef3 | 928 | if (low_pfn == end_pfn) |
edc2ca61 | 929 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 930 | |
e34d85f0 JK |
931 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
932 | nr_scanned, nr_isolated); | |
b7aba698 | 933 | |
010fc29a | 934 | count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned); |
397487db | 935 | if (nr_isolated) |
010fc29a | 936 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 937 | |
2fe86e00 MN |
938 | return low_pfn; |
939 | } | |
940 | ||
edc2ca61 VB |
941 | /** |
942 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
943 | * @cc: Compaction control structure. | |
944 | * @start_pfn: The first PFN to start isolating. | |
945 | * @end_pfn: The one-past-last PFN. | |
946 | * | |
947 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
948 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
949 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
950 | */ | |
951 | unsigned long | |
952 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
953 | unsigned long end_pfn) | |
954 | { | |
e1409c32 | 955 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
956 | |
957 | /* Scan block by block. First and last block may be incomplete */ | |
958 | pfn = start_pfn; | |
06b6640a | 959 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
960 | if (block_start_pfn < cc->zone->zone_start_pfn) |
961 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 962 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
963 | |
964 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 965 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
966 | block_end_pfn += pageblock_nr_pages) { |
967 | ||
968 | block_end_pfn = min(block_end_pfn, end_pfn); | |
969 | ||
e1409c32 JK |
970 | if (!pageblock_pfn_to_page(block_start_pfn, |
971 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
972 | continue; |
973 | ||
974 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
975 | ISOLATE_UNEVICTABLE); | |
976 | ||
14af4a5e | 977 | if (!pfn) |
edc2ca61 | 978 | break; |
6ea41c0c JK |
979 | |
980 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
981 | break; | |
edc2ca61 VB |
982 | } |
983 | acct_isolated(cc->zone, cc); | |
984 | ||
985 | return pfn; | |
986 | } | |
987 | ||
ff9543fd MN |
988 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
989 | #ifdef CONFIG_COMPACTION | |
018e9a49 AM |
990 | |
991 | /* Returns true if the page is within a block suitable for migration to */ | |
992 | static bool suitable_migration_target(struct page *page) | |
993 | { | |
994 | /* If the page is a large free page, then disallow migration */ | |
995 | if (PageBuddy(page)) { | |
996 | /* | |
997 | * We are checking page_order without zone->lock taken. But | |
998 | * the only small danger is that we skip a potentially suitable | |
999 | * pageblock, so it's not worth to check order for valid range. | |
1000 | */ | |
1001 | if (page_order_unsafe(page) >= pageblock_order) | |
1002 | return false; | |
1003 | } | |
1004 | ||
1005 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ | |
1006 | if (migrate_async_suitable(get_pageblock_migratetype(page))) | |
1007 | return true; | |
1008 | ||
1009 | /* Otherwise skip the block */ | |
1010 | return false; | |
1011 | } | |
1012 | ||
f2849aa0 VB |
1013 | /* |
1014 | * Test whether the free scanner has reached the same or lower pageblock than | |
1015 | * the migration scanner, and compaction should thus terminate. | |
1016 | */ | |
1017 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1018 | { | |
1019 | return (cc->free_pfn >> pageblock_order) | |
1020 | <= (cc->migrate_pfn >> pageblock_order); | |
1021 | } | |
1022 | ||
2fe86e00 | 1023 | /* |
ff9543fd MN |
1024 | * Based on information in the current compact_control, find blocks |
1025 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1026 | */ |
edc2ca61 | 1027 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1028 | { |
edc2ca61 | 1029 | struct zone *zone = cc->zone; |
ff9543fd | 1030 | struct page *page; |
c96b9e50 | 1031 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1032 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1033 | unsigned long block_end_pfn; /* end of current pageblock */ |
1034 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1035 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 1036 | |
ff9543fd MN |
1037 | /* |
1038 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1039 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1040 | * zone when isolating for the first time. For looping we also need |
1041 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1042 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1043 | * For ending point, take care when isolating in last pageblock of a | |
1044 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1045 | * The low boundary is the end of the pageblock the migration scanner |
1046 | * is using. | |
ff9543fd | 1047 | */ |
e14c720e | 1048 | isolate_start_pfn = cc->free_pfn; |
06b6640a | 1049 | block_start_pfn = pageblock_start_pfn(cc->free_pfn); |
c96b9e50 VB |
1050 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1051 | zone_end_pfn(zone)); | |
06b6640a | 1052 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
2fe86e00 | 1053 | |
ff9543fd MN |
1054 | /* |
1055 | * Isolate free pages until enough are available to migrate the | |
1056 | * pages on cc->migratepages. We stop searching if the migrate | |
1057 | * and free page scanners meet or enough free pages are isolated. | |
1058 | */ | |
f5f61a32 | 1059 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1060 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1061 | block_start_pfn -= pageblock_nr_pages, |
1062 | isolate_start_pfn = block_start_pfn) { | |
f6ea3adb DR |
1063 | /* |
1064 | * This can iterate a massively long zone without finding any | |
1065 | * suitable migration targets, so periodically check if we need | |
be976572 | 1066 | * to schedule, or even abort async compaction. |
f6ea3adb | 1067 | */ |
be976572 VB |
1068 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
1069 | && compact_should_abort(cc)) | |
1070 | break; | |
f6ea3adb | 1071 | |
7d49d886 VB |
1072 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1073 | zone); | |
1074 | if (!page) | |
ff9543fd MN |
1075 | continue; |
1076 | ||
1077 | /* Check the block is suitable for migration */ | |
68e3e926 | 1078 | if (!suitable_migration_target(page)) |
ff9543fd | 1079 | continue; |
68e3e926 | 1080 | |
bb13ffeb MG |
1081 | /* If isolation recently failed, do not retry */ |
1082 | if (!isolation_suitable(cc, page)) | |
1083 | continue; | |
1084 | ||
e14c720e | 1085 | /* Found a block suitable for isolating free pages from. */ |
a46cbf3b DR |
1086 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, |
1087 | freelist, false); | |
ff9543fd | 1088 | |
e14c720e | 1089 | /* |
a46cbf3b DR |
1090 | * If we isolated enough freepages, or aborted due to lock |
1091 | * contention, terminate. | |
e14c720e | 1092 | */ |
f5f61a32 VB |
1093 | if ((cc->nr_freepages >= cc->nr_migratepages) |
1094 | || cc->contended) { | |
a46cbf3b DR |
1095 | if (isolate_start_pfn >= block_end_pfn) { |
1096 | /* | |
1097 | * Restart at previous pageblock if more | |
1098 | * freepages can be isolated next time. | |
1099 | */ | |
f5f61a32 VB |
1100 | isolate_start_pfn = |
1101 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1102 | } |
be976572 | 1103 | break; |
a46cbf3b | 1104 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1105 | /* |
a46cbf3b DR |
1106 | * If isolation failed early, do not continue |
1107 | * needlessly. | |
f5f61a32 | 1108 | */ |
a46cbf3b | 1109 | break; |
f5f61a32 | 1110 | } |
ff9543fd MN |
1111 | } |
1112 | ||
1113 | /* split_free_page does not map the pages */ | |
1114 | map_pages(freelist); | |
1115 | ||
7ed695e0 | 1116 | /* |
f5f61a32 VB |
1117 | * Record where the free scanner will restart next time. Either we |
1118 | * broke from the loop and set isolate_start_pfn based on the last | |
1119 | * call to isolate_freepages_block(), or we met the migration scanner | |
1120 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1121 | */ |
f5f61a32 | 1122 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1123 | } |
1124 | ||
1125 | /* | |
1126 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1127 | * from the isolated freelists in the block we are migrating to. | |
1128 | */ | |
1129 | static struct page *compaction_alloc(struct page *migratepage, | |
1130 | unsigned long data, | |
1131 | int **result) | |
1132 | { | |
1133 | struct compact_control *cc = (struct compact_control *)data; | |
1134 | struct page *freepage; | |
1135 | ||
be976572 VB |
1136 | /* |
1137 | * Isolate free pages if necessary, and if we are not aborting due to | |
1138 | * contention. | |
1139 | */ | |
748446bb | 1140 | if (list_empty(&cc->freepages)) { |
be976572 | 1141 | if (!cc->contended) |
edc2ca61 | 1142 | isolate_freepages(cc); |
748446bb MG |
1143 | |
1144 | if (list_empty(&cc->freepages)) | |
1145 | return NULL; | |
1146 | } | |
1147 | ||
1148 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1149 | list_del(&freepage->lru); | |
1150 | cc->nr_freepages--; | |
1151 | ||
1152 | return freepage; | |
1153 | } | |
1154 | ||
1155 | /* | |
d53aea3d DR |
1156 | * This is a migrate-callback that "frees" freepages back to the isolated |
1157 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1158 | * special handling needed for NUMA. | |
1159 | */ | |
1160 | static void compaction_free(struct page *page, unsigned long data) | |
1161 | { | |
1162 | struct compact_control *cc = (struct compact_control *)data; | |
1163 | ||
1164 | list_add(&page->lru, &cc->freepages); | |
1165 | cc->nr_freepages++; | |
1166 | } | |
1167 | ||
ff9543fd MN |
1168 | /* possible outcome of isolate_migratepages */ |
1169 | typedef enum { | |
1170 | ISOLATE_ABORT, /* Abort compaction now */ | |
1171 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1172 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1173 | } isolate_migrate_t; | |
1174 | ||
5bbe3547 EM |
1175 | /* |
1176 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1177 | * compactable pages. | |
1178 | */ | |
1179 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1180 | ||
ff9543fd | 1181 | /* |
edc2ca61 VB |
1182 | * Isolate all pages that can be migrated from the first suitable block, |
1183 | * starting at the block pointed to by the migrate scanner pfn within | |
1184 | * compact_control. | |
ff9543fd MN |
1185 | */ |
1186 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1187 | struct compact_control *cc) | |
1188 | { | |
e1409c32 JK |
1189 | unsigned long block_start_pfn; |
1190 | unsigned long block_end_pfn; | |
1191 | unsigned long low_pfn; | |
edc2ca61 VB |
1192 | struct page *page; |
1193 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1194 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
edc2ca61 | 1195 | (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1196 | |
edc2ca61 VB |
1197 | /* |
1198 | * Start at where we last stopped, or beginning of the zone as | |
1199 | * initialized by compact_zone() | |
1200 | */ | |
1201 | low_pfn = cc->migrate_pfn; | |
06b6640a | 1202 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1203 | if (block_start_pfn < zone->zone_start_pfn) |
1204 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd MN |
1205 | |
1206 | /* Only scan within a pageblock boundary */ | |
06b6640a | 1207 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1208 | |
edc2ca61 VB |
1209 | /* |
1210 | * Iterate over whole pageblocks until we find the first suitable. | |
1211 | * Do not cross the free scanner. | |
1212 | */ | |
e1409c32 JK |
1213 | for (; block_end_pfn <= cc->free_pfn; |
1214 | low_pfn = block_end_pfn, | |
1215 | block_start_pfn = block_end_pfn, | |
1216 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1217 | |
edc2ca61 VB |
1218 | /* |
1219 | * This can potentially iterate a massively long zone with | |
1220 | * many pageblocks unsuitable, so periodically check if we | |
1221 | * need to schedule, or even abort async compaction. | |
1222 | */ | |
1223 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1224 | && compact_should_abort(cc)) | |
1225 | break; | |
ff9543fd | 1226 | |
e1409c32 JK |
1227 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1228 | zone); | |
7d49d886 | 1229 | if (!page) |
edc2ca61 VB |
1230 | continue; |
1231 | ||
edc2ca61 VB |
1232 | /* If isolation recently failed, do not retry */ |
1233 | if (!isolation_suitable(cc, page)) | |
1234 | continue; | |
1235 | ||
1236 | /* | |
1237 | * For async compaction, also only scan in MOVABLE blocks. | |
1238 | * Async compaction is optimistic to see if the minimum amount | |
1239 | * of work satisfies the allocation. | |
1240 | */ | |
1241 | if (cc->mode == MIGRATE_ASYNC && | |
1242 | !migrate_async_suitable(get_pageblock_migratetype(page))) | |
1243 | continue; | |
1244 | ||
1245 | /* Perform the isolation */ | |
e1409c32 JK |
1246 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1247 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1248 | |
ff59909a HD |
1249 | if (!low_pfn || cc->contended) { |
1250 | acct_isolated(zone, cc); | |
edc2ca61 | 1251 | return ISOLATE_ABORT; |
ff59909a | 1252 | } |
edc2ca61 VB |
1253 | |
1254 | /* | |
1255 | * Either we isolated something and proceed with migration. Or | |
1256 | * we failed and compact_zone should decide if we should | |
1257 | * continue or not. | |
1258 | */ | |
1259 | break; | |
1260 | } | |
1261 | ||
1262 | acct_isolated(zone, cc); | |
f2849aa0 VB |
1263 | /* Record where migration scanner will be restarted. */ |
1264 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1265 | |
edc2ca61 | 1266 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1267 | } |
1268 | ||
21c527a3 YB |
1269 | /* |
1270 | * order == -1 is expected when compacting via | |
1271 | * /proc/sys/vm/compact_memory | |
1272 | */ | |
1273 | static inline bool is_via_compact_memory(int order) | |
1274 | { | |
1275 | return order == -1; | |
1276 | } | |
1277 | ||
ea7ab982 | 1278 | static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc, |
6d7ce559 | 1279 | const int migratetype) |
748446bb | 1280 | { |
8fb74b9f | 1281 | unsigned int order; |
5a03b051 | 1282 | unsigned long watermark; |
56de7263 | 1283 | |
be976572 | 1284 | if (cc->contended || fatal_signal_pending(current)) |
2d1e1041 | 1285 | return COMPACT_CONTENDED; |
748446bb | 1286 | |
753341a4 | 1287 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1288 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1289 | /* Let the next compaction start anew. */ |
02333641 | 1290 | reset_cached_positions(zone); |
55b7c4c9 | 1291 | |
62997027 MG |
1292 | /* |
1293 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1294 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1295 | * flag itself as the decision to be clear should be directly |
1296 | * based on an allocation request. | |
1297 | */ | |
accf6242 | 1298 | if (cc->direct_compaction) |
62997027 MG |
1299 | zone->compact_blockskip_flush = true; |
1300 | ||
c8f7de0b MH |
1301 | if (cc->whole_zone) |
1302 | return COMPACT_COMPLETE; | |
1303 | else | |
1304 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1305 | } |
748446bb | 1306 | |
21c527a3 | 1307 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1308 | return COMPACT_CONTINUE; |
1309 | ||
3957c776 MH |
1310 | /* Compaction run is not finished if the watermark is not met */ |
1311 | watermark = low_wmark_pages(zone); | |
3957c776 | 1312 | |
ebff3980 VB |
1313 | if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx, |
1314 | cc->alloc_flags)) | |
3957c776 MH |
1315 | return COMPACT_CONTINUE; |
1316 | ||
56de7263 | 1317 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f MG |
1318 | for (order = cc->order; order < MAX_ORDER; order++) { |
1319 | struct free_area *area = &zone->free_area[order]; | |
2149cdae | 1320 | bool can_steal; |
8fb74b9f MG |
1321 | |
1322 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1323 | if (!list_empty(&area->free_list[migratetype])) |
8fb74b9f MG |
1324 | return COMPACT_PARTIAL; |
1325 | ||
2149cdae JK |
1326 | #ifdef CONFIG_CMA |
1327 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1328 | if (migratetype == MIGRATE_MOVABLE && | |
1329 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
1330 | return COMPACT_PARTIAL; | |
1331 | #endif | |
1332 | /* | |
1333 | * Job done if allocation would steal freepages from | |
1334 | * other migratetype buddy lists. | |
1335 | */ | |
1336 | if (find_suitable_fallback(area, order, migratetype, | |
1337 | true, &can_steal) != -1) | |
56de7263 MG |
1338 | return COMPACT_PARTIAL; |
1339 | } | |
1340 | ||
837d026d JK |
1341 | return COMPACT_NO_SUITABLE_PAGE; |
1342 | } | |
1343 | ||
ea7ab982 MH |
1344 | static enum compact_result compact_finished(struct zone *zone, |
1345 | struct compact_control *cc, | |
1346 | const int migratetype) | |
837d026d JK |
1347 | { |
1348 | int ret; | |
1349 | ||
1350 | ret = __compact_finished(zone, cc, migratetype); | |
1351 | trace_mm_compaction_finished(zone, cc->order, ret); | |
1352 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1353 | ret = COMPACT_CONTINUE; | |
1354 | ||
1355 | return ret; | |
748446bb MG |
1356 | } |
1357 | ||
3e7d3449 MG |
1358 | /* |
1359 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1360 | * Returns | |
1361 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
1362 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
1363 | * COMPACT_CONTINUE - If compaction should run now | |
1364 | */ | |
ea7ab982 | 1365 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1366 | unsigned int alloc_flags, |
86a294a8 MH |
1367 | int classzone_idx, |
1368 | unsigned long wmark_target) | |
3e7d3449 MG |
1369 | { |
1370 | int fragindex; | |
1371 | unsigned long watermark; | |
1372 | ||
21c527a3 | 1373 | if (is_via_compact_memory(order)) |
3957c776 MH |
1374 | return COMPACT_CONTINUE; |
1375 | ||
ebff3980 VB |
1376 | watermark = low_wmark_pages(zone); |
1377 | /* | |
1378 | * If watermarks for high-order allocation are already met, there | |
1379 | * should be no need for compaction at all. | |
1380 | */ | |
1381 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1382 | alloc_flags)) | |
1383 | return COMPACT_PARTIAL; | |
1384 | ||
3e7d3449 MG |
1385 | /* |
1386 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
1387 | * This is because during migration, copies of pages need to be | |
1388 | * allocated and for a short time, the footprint is higher | |
1389 | */ | |
ebff3980 | 1390 | watermark += (2UL << order); |
86a294a8 MH |
1391 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
1392 | alloc_flags, wmark_target)) | |
3e7d3449 MG |
1393 | return COMPACT_SKIPPED; |
1394 | ||
1395 | /* | |
1396 | * fragmentation index determines if allocation failures are due to | |
1397 | * low memory or external fragmentation | |
1398 | * | |
ebff3980 VB |
1399 | * index of -1000 would imply allocations might succeed depending on |
1400 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1401 | * index towards 0 implies failure is due to lack of memory |
1402 | * index towards 1000 implies failure is due to fragmentation | |
1403 | * | |
1404 | * Only compact if a failure would be due to fragmentation. | |
1405 | */ | |
1406 | fragindex = fragmentation_index(zone, order); | |
1407 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
837d026d | 1408 | return COMPACT_NOT_SUITABLE_ZONE; |
3e7d3449 | 1409 | |
3e7d3449 MG |
1410 | return COMPACT_CONTINUE; |
1411 | } | |
1412 | ||
ea7ab982 | 1413 | enum compact_result compaction_suitable(struct zone *zone, int order, |
c603844b MG |
1414 | unsigned int alloc_flags, |
1415 | int classzone_idx) | |
837d026d | 1416 | { |
ea7ab982 | 1417 | enum compact_result ret; |
837d026d | 1418 | |
86a294a8 MH |
1419 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, |
1420 | zone_page_state(zone, NR_FREE_PAGES)); | |
837d026d JK |
1421 | trace_mm_compaction_suitable(zone, order, ret); |
1422 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1423 | ret = COMPACT_SKIPPED; | |
1424 | ||
1425 | return ret; | |
1426 | } | |
1427 | ||
86a294a8 MH |
1428 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
1429 | int alloc_flags) | |
1430 | { | |
1431 | struct zone *zone; | |
1432 | struct zoneref *z; | |
1433 | ||
1434 | /* | |
1435 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1436 | * retrying the reclaim. | |
1437 | */ | |
1438 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1439 | ac->nodemask) { | |
1440 | unsigned long available; | |
1441 | enum compact_result compact_result; | |
1442 | ||
1443 | /* | |
1444 | * Do not consider all the reclaimable memory because we do not | |
1445 | * want to trash just for a single high order allocation which | |
1446 | * is even not guaranteed to appear even if __compaction_suitable | |
1447 | * is happy about the watermark check. | |
1448 | */ | |
1449 | available = zone_reclaimable_pages(zone) / order; | |
1450 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); | |
1451 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1452 | ac_classzone_idx(ac), available); | |
1453 | if (compact_result != COMPACT_SKIPPED && | |
1454 | compact_result != COMPACT_NOT_SUITABLE_ZONE) | |
1455 | return true; | |
1456 | } | |
1457 | ||
1458 | return false; | |
1459 | } | |
1460 | ||
ea7ab982 | 1461 | static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc) |
748446bb | 1462 | { |
ea7ab982 | 1463 | enum compact_result ret; |
c89511ab | 1464 | unsigned long start_pfn = zone->zone_start_pfn; |
108bcc96 | 1465 | unsigned long end_pfn = zone_end_pfn(zone); |
6d7ce559 | 1466 | const int migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
e0b9daeb | 1467 | const bool sync = cc->mode != MIGRATE_ASYNC; |
748446bb | 1468 | |
ebff3980 VB |
1469 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, |
1470 | cc->classzone_idx); | |
c46649de MH |
1471 | /* Compaction is likely to fail */ |
1472 | if (ret == COMPACT_PARTIAL || ret == COMPACT_SKIPPED) | |
3e7d3449 | 1473 | return ret; |
c46649de MH |
1474 | |
1475 | /* huh, compaction_suitable is returning something unexpected */ | |
1476 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 1477 | |
d3132e4b VB |
1478 | /* |
1479 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 1480 | * is about to be retried after being deferred. |
d3132e4b | 1481 | */ |
accf6242 | 1482 | if (compaction_restarting(zone, cc->order)) |
d3132e4b VB |
1483 | __reset_isolation_suitable(zone); |
1484 | ||
c89511ab MG |
1485 | /* |
1486 | * Setup to move all movable pages to the end of the zone. Used cached | |
1487 | * information on where the scanners should start but check that it | |
1488 | * is initialised by ensuring the values are within zone boundaries. | |
1489 | */ | |
e0b9daeb | 1490 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; |
c89511ab | 1491 | cc->free_pfn = zone->compact_cached_free_pfn; |
623446e4 | 1492 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
06b6640a | 1493 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
c89511ab MG |
1494 | zone->compact_cached_free_pfn = cc->free_pfn; |
1495 | } | |
623446e4 | 1496 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { |
c89511ab | 1497 | cc->migrate_pfn = start_pfn; |
35979ef3 DR |
1498 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1499 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
c89511ab | 1500 | } |
c8f7de0b MH |
1501 | |
1502 | if (cc->migrate_pfn == start_pfn) | |
1503 | cc->whole_zone = true; | |
1504 | ||
1a16718c | 1505 | cc->last_migrated_pfn = 0; |
748446bb | 1506 | |
16c4a097 JK |
1507 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1508 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1509 | |
748446bb MG |
1510 | migrate_prep_local(); |
1511 | ||
6d7ce559 DR |
1512 | while ((ret = compact_finished(zone, cc, migratetype)) == |
1513 | COMPACT_CONTINUE) { | |
9d502c1c | 1514 | int err; |
748446bb | 1515 | |
f9e35b3b MG |
1516 | switch (isolate_migratepages(zone, cc)) { |
1517 | case ISOLATE_ABORT: | |
2d1e1041 | 1518 | ret = COMPACT_CONTENDED; |
5733c7d1 | 1519 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1520 | cc->nr_migratepages = 0; |
f9e35b3b MG |
1521 | goto out; |
1522 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1523 | /* |
1524 | * We haven't isolated and migrated anything, but | |
1525 | * there might still be unflushed migrations from | |
1526 | * previous cc->order aligned block. | |
1527 | */ | |
1528 | goto check_drain; | |
f9e35b3b MG |
1529 | case ISOLATE_SUCCESS: |
1530 | ; | |
1531 | } | |
748446bb | 1532 | |
d53aea3d | 1533 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1534 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1535 | MR_COMPACTION); |
748446bb | 1536 | |
f8c9301f VB |
1537 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1538 | &cc->migratepages); | |
748446bb | 1539 | |
f8c9301f VB |
1540 | /* All pages were either migrated or will be released */ |
1541 | cc->nr_migratepages = 0; | |
9d502c1c | 1542 | if (err) { |
5733c7d1 | 1543 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1544 | /* |
1545 | * migrate_pages() may return -ENOMEM when scanners meet | |
1546 | * and we want compact_finished() to detect it | |
1547 | */ | |
f2849aa0 | 1548 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 1549 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
1550 | goto out; |
1551 | } | |
fdd048e1 VB |
1552 | /* |
1553 | * We failed to migrate at least one page in the current | |
1554 | * order-aligned block, so skip the rest of it. | |
1555 | */ | |
1556 | if (cc->direct_compaction && | |
1557 | (cc->mode == MIGRATE_ASYNC)) { | |
1558 | cc->migrate_pfn = block_end_pfn( | |
1559 | cc->migrate_pfn - 1, cc->order); | |
1560 | /* Draining pcplists is useless in this case */ | |
1561 | cc->last_migrated_pfn = 0; | |
1562 | ||
1563 | } | |
748446bb | 1564 | } |
fdaf7f5c | 1565 | |
fdaf7f5c VB |
1566 | check_drain: |
1567 | /* | |
1568 | * Has the migration scanner moved away from the previous | |
1569 | * cc->order aligned block where we migrated from? If yes, | |
1570 | * flush the pages that were freed, so that they can merge and | |
1571 | * compact_finished() can detect immediately if allocation | |
1572 | * would succeed. | |
1573 | */ | |
1a16718c | 1574 | if (cc->order > 0 && cc->last_migrated_pfn) { |
fdaf7f5c VB |
1575 | int cpu; |
1576 | unsigned long current_block_start = | |
06b6640a | 1577 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 1578 | |
1a16718c | 1579 | if (cc->last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
1580 | cpu = get_cpu(); |
1581 | lru_add_drain_cpu(cpu); | |
1582 | drain_local_pages(zone); | |
1583 | put_cpu(); | |
1584 | /* No more flushing until we migrate again */ | |
1a16718c | 1585 | cc->last_migrated_pfn = 0; |
fdaf7f5c VB |
1586 | } |
1587 | } | |
1588 | ||
748446bb MG |
1589 | } |
1590 | ||
f9e35b3b | 1591 | out: |
6bace090 VB |
1592 | /* |
1593 | * Release free pages and update where the free scanner should restart, | |
1594 | * so we don't leave any returned pages behind in the next attempt. | |
1595 | */ | |
1596 | if (cc->nr_freepages > 0) { | |
1597 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1598 | ||
1599 | cc->nr_freepages = 0; | |
1600 | VM_BUG_ON(free_pfn == 0); | |
1601 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 1602 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
1603 | /* |
1604 | * Only go back, not forward. The cached pfn might have been | |
1605 | * already reset to zone end in compact_finished() | |
1606 | */ | |
1607 | if (free_pfn > zone->compact_cached_free_pfn) | |
1608 | zone->compact_cached_free_pfn = free_pfn; | |
1609 | } | |
748446bb | 1610 | |
16c4a097 JK |
1611 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1612 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1613 | |
2d1e1041 VB |
1614 | if (ret == COMPACT_CONTENDED) |
1615 | ret = COMPACT_PARTIAL; | |
1616 | ||
748446bb MG |
1617 | return ret; |
1618 | } | |
76ab0f53 | 1619 | |
ea7ab982 | 1620 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
ebff3980 | 1621 | gfp_t gfp_mask, enum migrate_mode mode, int *contended, |
c603844b | 1622 | unsigned int alloc_flags, int classzone_idx) |
56de7263 | 1623 | { |
ea7ab982 | 1624 | enum compact_result ret; |
56de7263 MG |
1625 | struct compact_control cc = { |
1626 | .nr_freepages = 0, | |
1627 | .nr_migratepages = 0, | |
1628 | .order = order, | |
6d7ce559 | 1629 | .gfp_mask = gfp_mask, |
56de7263 | 1630 | .zone = zone, |
e0b9daeb | 1631 | .mode = mode, |
ebff3980 VB |
1632 | .alloc_flags = alloc_flags, |
1633 | .classzone_idx = classzone_idx, | |
accf6242 | 1634 | .direct_compaction = true, |
56de7263 MG |
1635 | }; |
1636 | INIT_LIST_HEAD(&cc.freepages); | |
1637 | INIT_LIST_HEAD(&cc.migratepages); | |
1638 | ||
e64c5237 SL |
1639 | ret = compact_zone(zone, &cc); |
1640 | ||
1641 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1642 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1643 | ||
1644 | *contended = cc.contended; | |
1645 | return ret; | |
56de7263 MG |
1646 | } |
1647 | ||
5e771905 MG |
1648 | int sysctl_extfrag_threshold = 500; |
1649 | ||
56de7263 MG |
1650 | /** |
1651 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1652 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1653 | * @order: The order of the current allocation |
1654 | * @alloc_flags: The allocation flags of the current allocation | |
1655 | * @ac: The context of current allocation | |
e0b9daeb | 1656 | * @mode: The migration mode for async, sync light, or sync migration |
1f9efdef VB |
1657 | * @contended: Return value that determines if compaction was aborted due to |
1658 | * need_resched() or lock contention | |
56de7263 MG |
1659 | * |
1660 | * This is the main entry point for direct page compaction. | |
1661 | */ | |
ea7ab982 | 1662 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b MG |
1663 | unsigned int alloc_flags, const struct alloc_context *ac, |
1664 | enum migrate_mode mode, int *contended) | |
56de7263 | 1665 | { |
56de7263 MG |
1666 | int may_enter_fs = gfp_mask & __GFP_FS; |
1667 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
1668 | struct zoneref *z; |
1669 | struct zone *zone; | |
1d4746d3 | 1670 | enum compact_result rc = COMPACT_SKIPPED; |
1f9efdef VB |
1671 | int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */ |
1672 | ||
1673 | *contended = COMPACT_CONTENDED_NONE; | |
56de7263 | 1674 | |
4ffb6335 | 1675 | /* Check if the GFP flags allow compaction */ |
c5a73c3d | 1676 | if (!order || !may_enter_fs || !may_perform_io) |
53853e2d | 1677 | return COMPACT_SKIPPED; |
56de7263 | 1678 | |
837d026d JK |
1679 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode); |
1680 | ||
56de7263 | 1681 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1682 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1683 | ac->nodemask) { | |
ea7ab982 | 1684 | enum compact_result status; |
1f9efdef | 1685 | int zone_contended; |
56de7263 | 1686 | |
1d4746d3 MH |
1687 | if (compaction_deferred(zone, order)) { |
1688 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); | |
53853e2d | 1689 | continue; |
1d4746d3 | 1690 | } |
53853e2d | 1691 | |
e0b9daeb | 1692 | status = compact_zone_order(zone, order, gfp_mask, mode, |
1a6d53a1 | 1693 | &zone_contended, alloc_flags, |
93ea9964 | 1694 | ac_classzone_idx(ac)); |
56de7263 | 1695 | rc = max(status, rc); |
1f9efdef VB |
1696 | /* |
1697 | * It takes at least one zone that wasn't lock contended | |
1698 | * to clear all_zones_contended. | |
1699 | */ | |
1700 | all_zones_contended &= zone_contended; | |
56de7263 | 1701 | |
3e7d3449 | 1702 | /* If a normal allocation would succeed, stop compacting */ |
ebff3980 | 1703 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), |
93ea9964 | 1704 | ac_classzone_idx(ac), alloc_flags)) { |
53853e2d VB |
1705 | /* |
1706 | * We think the allocation will succeed in this zone, | |
1707 | * but it is not certain, hence the false. The caller | |
1708 | * will repeat this with true if allocation indeed | |
1709 | * succeeds in this zone. | |
1710 | */ | |
1711 | compaction_defer_reset(zone, order, false); | |
1f9efdef VB |
1712 | /* |
1713 | * It is possible that async compaction aborted due to | |
1714 | * need_resched() and the watermarks were ok thanks to | |
1715 | * somebody else freeing memory. The allocation can | |
1716 | * however still fail so we better signal the | |
1717 | * need_resched() contention anyway (this will not | |
1718 | * prevent the allocation attempt). | |
1719 | */ | |
1720 | if (zone_contended == COMPACT_CONTENDED_SCHED) | |
1721 | *contended = COMPACT_CONTENDED_SCHED; | |
1722 | ||
1723 | goto break_loop; | |
1724 | } | |
1725 | ||
c8f7de0b MH |
1726 | if (mode != MIGRATE_ASYNC && (status == COMPACT_COMPLETE || |
1727 | status == COMPACT_PARTIAL_SKIPPED)) { | |
53853e2d VB |
1728 | /* |
1729 | * We think that allocation won't succeed in this zone | |
1730 | * so we defer compaction there. If it ends up | |
1731 | * succeeding after all, it will be reset. | |
1732 | */ | |
1733 | defer_compaction(zone, order); | |
1734 | } | |
1f9efdef VB |
1735 | |
1736 | /* | |
1737 | * We might have stopped compacting due to need_resched() in | |
1738 | * async compaction, or due to a fatal signal detected. In that | |
1739 | * case do not try further zones and signal need_resched() | |
1740 | * contention. | |
1741 | */ | |
1742 | if ((zone_contended == COMPACT_CONTENDED_SCHED) | |
1743 | || fatal_signal_pending(current)) { | |
1744 | *contended = COMPACT_CONTENDED_SCHED; | |
1745 | goto break_loop; | |
1746 | } | |
1747 | ||
1748 | continue; | |
1749 | break_loop: | |
1750 | /* | |
1751 | * We might not have tried all the zones, so be conservative | |
1752 | * and assume they are not all lock contended. | |
1753 | */ | |
1754 | all_zones_contended = 0; | |
1755 | break; | |
56de7263 MG |
1756 | } |
1757 | ||
1f9efdef VB |
1758 | /* |
1759 | * If at least one zone wasn't deferred or skipped, we report if all | |
1760 | * zones that were tried were lock contended. | |
1761 | */ | |
1d4746d3 | 1762 | if (rc > COMPACT_INACTIVE && all_zones_contended) |
1f9efdef VB |
1763 | *contended = COMPACT_CONTENDED_LOCK; |
1764 | ||
56de7263 MG |
1765 | return rc; |
1766 | } | |
1767 | ||
1768 | ||
76ab0f53 | 1769 | /* Compact all zones within a node */ |
7103f16d | 1770 | static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
1771 | { |
1772 | int zoneid; | |
76ab0f53 MG |
1773 | struct zone *zone; |
1774 | ||
76ab0f53 | 1775 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1776 | |
1777 | zone = &pgdat->node_zones[zoneid]; | |
1778 | if (!populated_zone(zone)) | |
1779 | continue; | |
1780 | ||
7be62de9 RR |
1781 | cc->nr_freepages = 0; |
1782 | cc->nr_migratepages = 0; | |
1783 | cc->zone = zone; | |
1784 | INIT_LIST_HEAD(&cc->freepages); | |
1785 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 1786 | |
195b0c60 GK |
1787 | /* |
1788 | * When called via /proc/sys/vm/compact_memory | |
1789 | * this makes sure we compact the whole zone regardless of | |
1790 | * cached scanner positions. | |
1791 | */ | |
21c527a3 | 1792 | if (is_via_compact_memory(cc->order)) |
195b0c60 GK |
1793 | __reset_isolation_suitable(zone); |
1794 | ||
21c527a3 YB |
1795 | if (is_via_compact_memory(cc->order) || |
1796 | !compaction_deferred(zone, cc->order)) | |
7be62de9 | 1797 | compact_zone(zone, cc); |
76ab0f53 | 1798 | |
7be62de9 RR |
1799 | VM_BUG_ON(!list_empty(&cc->freepages)); |
1800 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
75469345 JK |
1801 | |
1802 | if (is_via_compact_memory(cc->order)) | |
1803 | continue; | |
1804 | ||
1805 | if (zone_watermark_ok(zone, cc->order, | |
1806 | low_wmark_pages(zone), 0, 0)) | |
1807 | compaction_defer_reset(zone, cc->order, false); | |
76ab0f53 | 1808 | } |
76ab0f53 MG |
1809 | } |
1810 | ||
7103f16d | 1811 | void compact_pgdat(pg_data_t *pgdat, int order) |
7be62de9 RR |
1812 | { |
1813 | struct compact_control cc = { | |
1814 | .order = order, | |
e0b9daeb | 1815 | .mode = MIGRATE_ASYNC, |
7be62de9 RR |
1816 | }; |
1817 | ||
3a7200af MG |
1818 | if (!order) |
1819 | return; | |
1820 | ||
7103f16d | 1821 | __compact_pgdat(pgdat, &cc); |
7be62de9 RR |
1822 | } |
1823 | ||
7103f16d | 1824 | static void compact_node(int nid) |
7be62de9 | 1825 | { |
7be62de9 RR |
1826 | struct compact_control cc = { |
1827 | .order = -1, | |
e0b9daeb | 1828 | .mode = MIGRATE_SYNC, |
91ca9186 | 1829 | .ignore_skip_hint = true, |
7be62de9 RR |
1830 | }; |
1831 | ||
7103f16d | 1832 | __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
1833 | } |
1834 | ||
76ab0f53 | 1835 | /* Compact all nodes in the system */ |
7964c06d | 1836 | static void compact_nodes(void) |
76ab0f53 MG |
1837 | { |
1838 | int nid; | |
1839 | ||
8575ec29 HD |
1840 | /* Flush pending updates to the LRU lists */ |
1841 | lru_add_drain_all(); | |
1842 | ||
76ab0f53 MG |
1843 | for_each_online_node(nid) |
1844 | compact_node(nid); | |
76ab0f53 MG |
1845 | } |
1846 | ||
1847 | /* The written value is actually unused, all memory is compacted */ | |
1848 | int sysctl_compact_memory; | |
1849 | ||
fec4eb2c YB |
1850 | /* |
1851 | * This is the entry point for compacting all nodes via | |
1852 | * /proc/sys/vm/compact_memory | |
1853 | */ | |
76ab0f53 MG |
1854 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
1855 | void __user *buffer, size_t *length, loff_t *ppos) | |
1856 | { | |
1857 | if (write) | |
7964c06d | 1858 | compact_nodes(); |
76ab0f53 MG |
1859 | |
1860 | return 0; | |
1861 | } | |
ed4a6d7f | 1862 | |
5e771905 MG |
1863 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
1864 | void __user *buffer, size_t *length, loff_t *ppos) | |
1865 | { | |
1866 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1867 | ||
1868 | return 0; | |
1869 | } | |
1870 | ||
ed4a6d7f | 1871 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
74e77fb9 | 1872 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1873 | struct device_attribute *attr, |
ed4a6d7f MG |
1874 | const char *buf, size_t count) |
1875 | { | |
8575ec29 HD |
1876 | int nid = dev->id; |
1877 | ||
1878 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1879 | /* Flush pending updates to the LRU lists */ | |
1880 | lru_add_drain_all(); | |
1881 | ||
1882 | compact_node(nid); | |
1883 | } | |
ed4a6d7f MG |
1884 | |
1885 | return count; | |
1886 | } | |
10fbcf4c | 1887 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1888 | |
1889 | int compaction_register_node(struct node *node) | |
1890 | { | |
10fbcf4c | 1891 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1892 | } |
1893 | ||
1894 | void compaction_unregister_node(struct node *node) | |
1895 | { | |
10fbcf4c | 1896 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1897 | } |
1898 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 1899 | |
698b1b30 VB |
1900 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
1901 | { | |
172400c6 | 1902 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
1903 | } |
1904 | ||
1905 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
1906 | { | |
1907 | int zoneid; | |
1908 | struct zone *zone; | |
1909 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
1910 | ||
6cd9dc3e | 1911 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
1912 | zone = &pgdat->node_zones[zoneid]; |
1913 | ||
1914 | if (!populated_zone(zone)) | |
1915 | continue; | |
1916 | ||
1917 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
1918 | classzone_idx) == COMPACT_CONTINUE) | |
1919 | return true; | |
1920 | } | |
1921 | ||
1922 | return false; | |
1923 | } | |
1924 | ||
1925 | static void kcompactd_do_work(pg_data_t *pgdat) | |
1926 | { | |
1927 | /* | |
1928 | * With no special task, compact all zones so that a page of requested | |
1929 | * order is allocatable. | |
1930 | */ | |
1931 | int zoneid; | |
1932 | struct zone *zone; | |
1933 | struct compact_control cc = { | |
1934 | .order = pgdat->kcompactd_max_order, | |
1935 | .classzone_idx = pgdat->kcompactd_classzone_idx, | |
1936 | .mode = MIGRATE_SYNC_LIGHT, | |
1937 | .ignore_skip_hint = true, | |
1938 | ||
1939 | }; | |
1940 | bool success = false; | |
1941 | ||
1942 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, | |
1943 | cc.classzone_idx); | |
1944 | count_vm_event(KCOMPACTD_WAKE); | |
1945 | ||
6cd9dc3e | 1946 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
1947 | int status; |
1948 | ||
1949 | zone = &pgdat->node_zones[zoneid]; | |
1950 | if (!populated_zone(zone)) | |
1951 | continue; | |
1952 | ||
1953 | if (compaction_deferred(zone, cc.order)) | |
1954 | continue; | |
1955 | ||
1956 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
1957 | COMPACT_CONTINUE) | |
1958 | continue; | |
1959 | ||
1960 | cc.nr_freepages = 0; | |
1961 | cc.nr_migratepages = 0; | |
1962 | cc.zone = zone; | |
1963 | INIT_LIST_HEAD(&cc.freepages); | |
1964 | INIT_LIST_HEAD(&cc.migratepages); | |
1965 | ||
172400c6 VB |
1966 | if (kthread_should_stop()) |
1967 | return; | |
698b1b30 VB |
1968 | status = compact_zone(zone, &cc); |
1969 | ||
1970 | if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone), | |
1971 | cc.classzone_idx, 0)) { | |
1972 | success = true; | |
1973 | compaction_defer_reset(zone, cc.order, false); | |
c8f7de0b | 1974 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
698b1b30 VB |
1975 | /* |
1976 | * We use sync migration mode here, so we defer like | |
1977 | * sync direct compaction does. | |
1978 | */ | |
1979 | defer_compaction(zone, cc.order); | |
1980 | } | |
1981 | ||
1982 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1983 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1984 | } | |
1985 | ||
1986 | /* | |
1987 | * Regardless of success, we are done until woken up next. But remember | |
1988 | * the requested order/classzone_idx in case it was higher/tighter than | |
1989 | * our current ones | |
1990 | */ | |
1991 | if (pgdat->kcompactd_max_order <= cc.order) | |
1992 | pgdat->kcompactd_max_order = 0; | |
1993 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
1994 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
1995 | } | |
1996 | ||
1997 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
1998 | { | |
1999 | if (!order) | |
2000 | return; | |
2001 | ||
2002 | if (pgdat->kcompactd_max_order < order) | |
2003 | pgdat->kcompactd_max_order = order; | |
2004 | ||
2005 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2006 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2007 | ||
2008 | if (!waitqueue_active(&pgdat->kcompactd_wait)) | |
2009 | return; | |
2010 | ||
2011 | if (!kcompactd_node_suitable(pgdat)) | |
2012 | return; | |
2013 | ||
2014 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2015 | classzone_idx); | |
2016 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2017 | } | |
2018 | ||
2019 | /* | |
2020 | * The background compaction daemon, started as a kernel thread | |
2021 | * from the init process. | |
2022 | */ | |
2023 | static int kcompactd(void *p) | |
2024 | { | |
2025 | pg_data_t *pgdat = (pg_data_t*)p; | |
2026 | struct task_struct *tsk = current; | |
2027 | ||
2028 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2029 | ||
2030 | if (!cpumask_empty(cpumask)) | |
2031 | set_cpus_allowed_ptr(tsk, cpumask); | |
2032 | ||
2033 | set_freezable(); | |
2034 | ||
2035 | pgdat->kcompactd_max_order = 0; | |
2036 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2037 | ||
2038 | while (!kthread_should_stop()) { | |
2039 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); | |
2040 | wait_event_freezable(pgdat->kcompactd_wait, | |
2041 | kcompactd_work_requested(pgdat)); | |
2042 | ||
2043 | kcompactd_do_work(pgdat); | |
2044 | } | |
2045 | ||
2046 | return 0; | |
2047 | } | |
2048 | ||
2049 | /* | |
2050 | * This kcompactd start function will be called by init and node-hot-add. | |
2051 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2052 | */ | |
2053 | int kcompactd_run(int nid) | |
2054 | { | |
2055 | pg_data_t *pgdat = NODE_DATA(nid); | |
2056 | int ret = 0; | |
2057 | ||
2058 | if (pgdat->kcompactd) | |
2059 | return 0; | |
2060 | ||
2061 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2062 | if (IS_ERR(pgdat->kcompactd)) { | |
2063 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2064 | ret = PTR_ERR(pgdat->kcompactd); | |
2065 | pgdat->kcompactd = NULL; | |
2066 | } | |
2067 | return ret; | |
2068 | } | |
2069 | ||
2070 | /* | |
2071 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2072 | * hold mem_hotplug_begin/end(). | |
2073 | */ | |
2074 | void kcompactd_stop(int nid) | |
2075 | { | |
2076 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2077 | ||
2078 | if (kcompactd) { | |
2079 | kthread_stop(kcompactd); | |
2080 | NODE_DATA(nid)->kcompactd = NULL; | |
2081 | } | |
2082 | } | |
2083 | ||
2084 | /* | |
2085 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2086 | * not required for correctness. So if the last cpu in a node goes | |
2087 | * away, we get changed to run anywhere: as the first one comes back, | |
2088 | * restore their cpu bindings. | |
2089 | */ | |
2090 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, | |
2091 | void *hcpu) | |
2092 | { | |
2093 | int nid; | |
2094 | ||
2095 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
2096 | for_each_node_state(nid, N_MEMORY) { | |
2097 | pg_data_t *pgdat = NODE_DATA(nid); | |
2098 | const struct cpumask *mask; | |
2099 | ||
2100 | mask = cpumask_of_node(pgdat->node_id); | |
2101 | ||
2102 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
2103 | /* One of our CPUs online: restore mask */ | |
2104 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
2105 | } | |
2106 | } | |
2107 | return NOTIFY_OK; | |
2108 | } | |
2109 | ||
2110 | static int __init kcompactd_init(void) | |
2111 | { | |
2112 | int nid; | |
2113 | ||
2114 | for_each_node_state(nid, N_MEMORY) | |
2115 | kcompactd_run(nid); | |
2116 | hotcpu_notifier(cpu_callback, 0); | |
2117 | return 0; | |
2118 | } | |
2119 | subsys_initcall(kcompactd_init) | |
2120 | ||
ff9543fd | 2121 | #endif /* CONFIG_COMPACTION */ |