Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
748446bb MG |
2 | /* |
3 | * linux/mm/compaction.c | |
4 | * | |
5 | * Memory compaction for the reduction of external fragmentation. Note that | |
6 | * this heavily depends upon page migration to do all the real heavy | |
7 | * lifting | |
8 | * | |
9 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
10 | */ | |
698b1b30 | 11 | #include <linux/cpu.h> |
748446bb MG |
12 | #include <linux/swap.h> |
13 | #include <linux/migrate.h> | |
14 | #include <linux/compaction.h> | |
15 | #include <linux/mm_inline.h> | |
174cd4b1 | 16 | #include <linux/sched/signal.h> |
748446bb | 17 | #include <linux/backing-dev.h> |
76ab0f53 | 18 | #include <linux/sysctl.h> |
ed4a6d7f | 19 | #include <linux/sysfs.h> |
194159fb | 20 | #include <linux/page-isolation.h> |
b8c73fc2 | 21 | #include <linux/kasan.h> |
698b1b30 VB |
22 | #include <linux/kthread.h> |
23 | #include <linux/freezer.h> | |
83358ece | 24 | #include <linux/page_owner.h> |
748446bb MG |
25 | #include "internal.h" |
26 | ||
010fc29a MK |
27 | #ifdef CONFIG_COMPACTION |
28 | static inline void count_compact_event(enum vm_event_item item) | |
29 | { | |
30 | count_vm_event(item); | |
31 | } | |
32 | ||
33 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
34 | { | |
35 | count_vm_events(item, delta); | |
36 | } | |
37 | #else | |
38 | #define count_compact_event(item) do { } while (0) | |
39 | #define count_compact_events(item, delta) do { } while (0) | |
40 | #endif | |
41 | ||
ff9543fd MN |
42 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
43 | ||
b7aba698 MG |
44 | #define CREATE_TRACE_POINTS |
45 | #include <trace/events/compaction.h> | |
46 | ||
06b6640a VB |
47 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
48 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
49 | #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order) | |
50 | #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order) | |
51 | ||
748446bb MG |
52 | static unsigned long release_freepages(struct list_head *freelist) |
53 | { | |
54 | struct page *page, *next; | |
6bace090 | 55 | unsigned long high_pfn = 0; |
748446bb MG |
56 | |
57 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 58 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
59 | list_del(&page->lru); |
60 | __free_page(page); | |
6bace090 VB |
61 | if (pfn > high_pfn) |
62 | high_pfn = pfn; | |
748446bb MG |
63 | } |
64 | ||
6bace090 | 65 | return high_pfn; |
748446bb MG |
66 | } |
67 | ||
ff9543fd MN |
68 | static void map_pages(struct list_head *list) |
69 | { | |
66c64223 JK |
70 | unsigned int i, order, nr_pages; |
71 | struct page *page, *next; | |
72 | LIST_HEAD(tmp_list); | |
73 | ||
74 | list_for_each_entry_safe(page, next, list, lru) { | |
75 | list_del(&page->lru); | |
76 | ||
77 | order = page_private(page); | |
78 | nr_pages = 1 << order; | |
66c64223 | 79 | |
46f24fd8 | 80 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
81 | if (order) |
82 | split_page(page, order); | |
ff9543fd | 83 | |
66c64223 JK |
84 | for (i = 0; i < nr_pages; i++) { |
85 | list_add(&page->lru, &tmp_list); | |
86 | page++; | |
87 | } | |
ff9543fd | 88 | } |
66c64223 JK |
89 | |
90 | list_splice(&tmp_list, list); | |
ff9543fd MN |
91 | } |
92 | ||
bb13ffeb | 93 | #ifdef CONFIG_COMPACTION |
24e2716f | 94 | |
bda807d4 MK |
95 | int PageMovable(struct page *page) |
96 | { | |
97 | struct address_space *mapping; | |
98 | ||
99 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
100 | if (!__PageMovable(page)) | |
101 | return 0; | |
102 | ||
103 | mapping = page_mapping(page); | |
104 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
105 | return 1; | |
106 | ||
107 | return 0; | |
108 | } | |
109 | EXPORT_SYMBOL(PageMovable); | |
110 | ||
111 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
112 | { | |
113 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
114 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
115 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
116 | } | |
117 | EXPORT_SYMBOL(__SetPageMovable); | |
118 | ||
119 | void __ClearPageMovable(struct page *page) | |
120 | { | |
121 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
122 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
123 | /* | |
124 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
125 | * flag so that VM can catch up released page by driver after isolation. | |
126 | * With it, VM migration doesn't try to put it back. | |
127 | */ | |
128 | page->mapping = (void *)((unsigned long)page->mapping & | |
129 | PAGE_MAPPING_MOVABLE); | |
130 | } | |
131 | EXPORT_SYMBOL(__ClearPageMovable); | |
132 | ||
24e2716f JK |
133 | /* Do not skip compaction more than 64 times */ |
134 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
135 | ||
136 | /* | |
137 | * Compaction is deferred when compaction fails to result in a page | |
138 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
139 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
140 | */ | |
141 | void defer_compaction(struct zone *zone, int order) | |
142 | { | |
143 | zone->compact_considered = 0; | |
144 | zone->compact_defer_shift++; | |
145 | ||
146 | if (order < zone->compact_order_failed) | |
147 | zone->compact_order_failed = order; | |
148 | ||
149 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
150 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
151 | ||
152 | trace_mm_compaction_defer_compaction(zone, order); | |
153 | } | |
154 | ||
155 | /* Returns true if compaction should be skipped this time */ | |
156 | bool compaction_deferred(struct zone *zone, int order) | |
157 | { | |
158 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
159 | ||
160 | if (order < zone->compact_order_failed) | |
161 | return false; | |
162 | ||
163 | /* Avoid possible overflow */ | |
164 | if (++zone->compact_considered > defer_limit) | |
165 | zone->compact_considered = defer_limit; | |
166 | ||
167 | if (zone->compact_considered >= defer_limit) | |
168 | return false; | |
169 | ||
170 | trace_mm_compaction_deferred(zone, order); | |
171 | ||
172 | return true; | |
173 | } | |
174 | ||
175 | /* | |
176 | * Update defer tracking counters after successful compaction of given order, | |
177 | * which means an allocation either succeeded (alloc_success == true) or is | |
178 | * expected to succeed. | |
179 | */ | |
180 | void compaction_defer_reset(struct zone *zone, int order, | |
181 | bool alloc_success) | |
182 | { | |
183 | if (alloc_success) { | |
184 | zone->compact_considered = 0; | |
185 | zone->compact_defer_shift = 0; | |
186 | } | |
187 | if (order >= zone->compact_order_failed) | |
188 | zone->compact_order_failed = order + 1; | |
189 | ||
190 | trace_mm_compaction_defer_reset(zone, order); | |
191 | } | |
192 | ||
193 | /* Returns true if restarting compaction after many failures */ | |
194 | bool compaction_restarting(struct zone *zone, int order) | |
195 | { | |
196 | if (order < zone->compact_order_failed) | |
197 | return false; | |
198 | ||
199 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
200 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
201 | } | |
202 | ||
bb13ffeb MG |
203 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
204 | static inline bool isolation_suitable(struct compact_control *cc, | |
205 | struct page *page) | |
206 | { | |
207 | if (cc->ignore_skip_hint) | |
208 | return true; | |
209 | ||
210 | return !get_pageblock_skip(page); | |
211 | } | |
212 | ||
02333641 VB |
213 | static void reset_cached_positions(struct zone *zone) |
214 | { | |
215 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
216 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 217 | zone->compact_cached_free_pfn = |
06b6640a | 218 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
219 | } |
220 | ||
21dc7e02 DR |
221 | /* |
222 | * Hugetlbfs pages should consistenly be skipped until updated by the hugetlb | |
223 | * subsystem. It is always pointless to compact pages of pageblock_order and | |
224 | * the free scanner can reconsider when no longer huge. | |
225 | */ | |
226 | static bool pageblock_skip_persistent(struct page *page, unsigned int order) | |
227 | { | |
228 | if (!PageHuge(page)) | |
229 | return false; | |
230 | if (order != pageblock_order) | |
231 | return false; | |
232 | return true; | |
233 | } | |
234 | ||
bb13ffeb MG |
235 | /* |
236 | * This function is called to clear all cached information on pageblocks that | |
237 | * should be skipped for page isolation when the migrate and free page scanner | |
238 | * meet. | |
239 | */ | |
62997027 | 240 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
241 | { |
242 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 243 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
244 | unsigned long pfn; |
245 | ||
62997027 | 246 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
247 | |
248 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
249 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
250 | struct page *page; | |
251 | ||
252 | cond_resched(); | |
253 | ||
ccbe1e4d MH |
254 | page = pfn_to_online_page(pfn); |
255 | if (!page) | |
bb13ffeb | 256 | continue; |
bb13ffeb MG |
257 | if (zone != page_zone(page)) |
258 | continue; | |
21dc7e02 DR |
259 | if (pageblock_skip_persistent(page, compound_order(page))) |
260 | continue; | |
bb13ffeb MG |
261 | |
262 | clear_pageblock_skip(page); | |
263 | } | |
02333641 VB |
264 | |
265 | reset_cached_positions(zone); | |
bb13ffeb MG |
266 | } |
267 | ||
62997027 MG |
268 | void reset_isolation_suitable(pg_data_t *pgdat) |
269 | { | |
270 | int zoneid; | |
271 | ||
272 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
273 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
274 | if (!populated_zone(zone)) | |
275 | continue; | |
276 | ||
277 | /* Only flush if a full compaction finished recently */ | |
278 | if (zone->compact_blockskip_flush) | |
279 | __reset_isolation_suitable(zone); | |
280 | } | |
281 | } | |
282 | ||
bb13ffeb MG |
283 | /* |
284 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 285 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 286 | */ |
c89511ab MG |
287 | static void update_pageblock_skip(struct compact_control *cc, |
288 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 289 | bool migrate_scanner) |
bb13ffeb | 290 | { |
c89511ab | 291 | struct zone *zone = cc->zone; |
35979ef3 | 292 | unsigned long pfn; |
6815bf3f JK |
293 | |
294 | if (cc->ignore_skip_hint) | |
295 | return; | |
296 | ||
bb13ffeb MG |
297 | if (!page) |
298 | return; | |
299 | ||
35979ef3 DR |
300 | if (nr_isolated) |
301 | return; | |
302 | ||
edc2ca61 | 303 | set_pageblock_skip(page); |
c89511ab | 304 | |
35979ef3 DR |
305 | pfn = page_to_pfn(page); |
306 | ||
307 | /* Update where async and sync compaction should restart */ | |
308 | if (migrate_scanner) { | |
35979ef3 DR |
309 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
310 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
311 | if (cc->mode != MIGRATE_ASYNC && |
312 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
313 | zone->compact_cached_migrate_pfn[1] = pfn; |
314 | } else { | |
35979ef3 DR |
315 | if (pfn < zone->compact_cached_free_pfn) |
316 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 317 | } |
bb13ffeb MG |
318 | } |
319 | #else | |
320 | static inline bool isolation_suitable(struct compact_control *cc, | |
321 | struct page *page) | |
322 | { | |
323 | return true; | |
324 | } | |
325 | ||
21dc7e02 DR |
326 | static inline bool pageblock_skip_persistent(struct page *page, |
327 | unsigned int order) | |
328 | { | |
329 | return false; | |
330 | } | |
331 | ||
332 | static inline void update_pageblock_skip(struct compact_control *cc, | |
c89511ab | 333 | struct page *page, unsigned long nr_isolated, |
edc2ca61 | 334 | bool migrate_scanner) |
bb13ffeb MG |
335 | { |
336 | } | |
337 | #endif /* CONFIG_COMPACTION */ | |
338 | ||
8b44d279 VB |
339 | /* |
340 | * Compaction requires the taking of some coarse locks that are potentially | |
341 | * very heavily contended. For async compaction, back out if the lock cannot | |
342 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
343 | * | |
344 | * Returns true if the lock is held | |
345 | * Returns false if the lock is not held and compaction should abort | |
346 | */ | |
347 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
348 | struct compact_control *cc) | |
2a1402aa | 349 | { |
8b44d279 VB |
350 | if (cc->mode == MIGRATE_ASYNC) { |
351 | if (!spin_trylock_irqsave(lock, *flags)) { | |
c3486f53 | 352 | cc->contended = true; |
8b44d279 VB |
353 | return false; |
354 | } | |
355 | } else { | |
356 | spin_lock_irqsave(lock, *flags); | |
357 | } | |
1f9efdef | 358 | |
8b44d279 | 359 | return true; |
2a1402aa MG |
360 | } |
361 | ||
c67fe375 MG |
362 | /* |
363 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
364 | * very heavily contended. The lock should be periodically unlocked to avoid |
365 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
366 | * the lock. It might also be that allowing the IRQs will result in | |
367 | * need_resched() becoming true. If scheduling is needed, async compaction | |
368 | * aborts. Sync compaction schedules. | |
369 | * Either compaction type will also abort if a fatal signal is pending. | |
370 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 371 | * |
8b44d279 VB |
372 | * Returns true if compaction should abort due to fatal signal pending, or |
373 | * async compaction due to need_resched() | |
374 | * Returns false when compaction can continue (sync compaction might have | |
375 | * scheduled) | |
c67fe375 | 376 | */ |
8b44d279 VB |
377 | static bool compact_unlock_should_abort(spinlock_t *lock, |
378 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 379 | { |
8b44d279 VB |
380 | if (*locked) { |
381 | spin_unlock_irqrestore(lock, flags); | |
382 | *locked = false; | |
383 | } | |
1f9efdef | 384 | |
8b44d279 | 385 | if (fatal_signal_pending(current)) { |
c3486f53 | 386 | cc->contended = true; |
8b44d279 VB |
387 | return true; |
388 | } | |
c67fe375 | 389 | |
8b44d279 | 390 | if (need_resched()) { |
e0b9daeb | 391 | if (cc->mode == MIGRATE_ASYNC) { |
c3486f53 | 392 | cc->contended = true; |
8b44d279 | 393 | return true; |
c67fe375 | 394 | } |
c67fe375 | 395 | cond_resched(); |
c67fe375 MG |
396 | } |
397 | ||
8b44d279 | 398 | return false; |
c67fe375 MG |
399 | } |
400 | ||
be976572 VB |
401 | /* |
402 | * Aside from avoiding lock contention, compaction also periodically checks | |
403 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 404 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
405 | * is used where no lock is concerned. |
406 | * | |
407 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
408 | * Returns true when async compaction should abort. | |
409 | */ | |
410 | static inline bool compact_should_abort(struct compact_control *cc) | |
411 | { | |
412 | /* async compaction aborts if contended */ | |
413 | if (need_resched()) { | |
414 | if (cc->mode == MIGRATE_ASYNC) { | |
c3486f53 | 415 | cc->contended = true; |
be976572 VB |
416 | return true; |
417 | } | |
418 | ||
419 | cond_resched(); | |
420 | } | |
421 | ||
422 | return false; | |
423 | } | |
424 | ||
85aa125f | 425 | /* |
9e4be470 JM |
426 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
427 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
428 | * (even though it may still end up isolating some pages). | |
85aa125f | 429 | */ |
f40d1e42 | 430 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 431 | unsigned long *start_pfn, |
85aa125f MN |
432 | unsigned long end_pfn, |
433 | struct list_head *freelist, | |
434 | bool strict) | |
748446bb | 435 | { |
b7aba698 | 436 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 437 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 438 | unsigned long flags = 0; |
f40d1e42 | 439 | bool locked = false; |
e14c720e | 440 | unsigned long blockpfn = *start_pfn; |
66c64223 | 441 | unsigned int order; |
748446bb | 442 | |
748446bb MG |
443 | cursor = pfn_to_page(blockpfn); |
444 | ||
f40d1e42 | 445 | /* Isolate free pages. */ |
748446bb | 446 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
66c64223 | 447 | int isolated; |
748446bb MG |
448 | struct page *page = cursor; |
449 | ||
8b44d279 VB |
450 | /* |
451 | * Periodically drop the lock (if held) regardless of its | |
452 | * contention, to give chance to IRQs. Abort if fatal signal | |
453 | * pending or async compaction detects need_resched() | |
454 | */ | |
455 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
456 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
457 | &locked, cc)) | |
458 | break; | |
459 | ||
b7aba698 | 460 | nr_scanned++; |
f40d1e42 | 461 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
462 | goto isolate_fail; |
463 | ||
bb13ffeb MG |
464 | if (!valid_page) |
465 | valid_page = page; | |
9fcd6d2e VB |
466 | |
467 | /* | |
468 | * For compound pages such as THP and hugetlbfs, we can save | |
469 | * potentially a lot of iterations if we skip them at once. | |
470 | * The check is racy, but we can consider only valid values | |
471 | * and the only danger is skipping too much. | |
472 | */ | |
473 | if (PageCompound(page)) { | |
21dc7e02 DR |
474 | const unsigned int order = compound_order(page); |
475 | ||
476 | if (pageblock_skip_persistent(page, order)) { | |
477 | set_pageblock_skip(page); | |
478 | blockpfn = end_pfn; | |
479 | } else if (likely(order < MAX_ORDER)) { | |
480 | blockpfn += (1UL << order) - 1; | |
481 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 482 | } |
9fcd6d2e VB |
483 | goto isolate_fail; |
484 | } | |
485 | ||
f40d1e42 | 486 | if (!PageBuddy(page)) |
2af120bc | 487 | goto isolate_fail; |
f40d1e42 MG |
488 | |
489 | /* | |
69b7189f VB |
490 | * If we already hold the lock, we can skip some rechecking. |
491 | * Note that if we hold the lock now, checked_pageblock was | |
492 | * already set in some previous iteration (or strict is true), | |
493 | * so it is correct to skip the suitable migration target | |
494 | * recheck as well. | |
f40d1e42 | 495 | */ |
69b7189f VB |
496 | if (!locked) { |
497 | /* | |
498 | * The zone lock must be held to isolate freepages. | |
499 | * Unfortunately this is a very coarse lock and can be | |
500 | * heavily contended if there are parallel allocations | |
501 | * or parallel compactions. For async compaction do not | |
502 | * spin on the lock and we acquire the lock as late as | |
503 | * possible. | |
504 | */ | |
8b44d279 VB |
505 | locked = compact_trylock_irqsave(&cc->zone->lock, |
506 | &flags, cc); | |
69b7189f VB |
507 | if (!locked) |
508 | break; | |
f40d1e42 | 509 | |
69b7189f VB |
510 | /* Recheck this is a buddy page under lock */ |
511 | if (!PageBuddy(page)) | |
512 | goto isolate_fail; | |
513 | } | |
748446bb | 514 | |
66c64223 JK |
515 | /* Found a free page, will break it into order-0 pages */ |
516 | order = page_order(page); | |
517 | isolated = __isolate_free_page(page, order); | |
a4f04f2c DR |
518 | if (!isolated) |
519 | break; | |
66c64223 | 520 | set_page_private(page, order); |
a4f04f2c | 521 | |
748446bb | 522 | total_isolated += isolated; |
a4f04f2c | 523 | cc->nr_freepages += isolated; |
66c64223 JK |
524 | list_add_tail(&page->lru, freelist); |
525 | ||
a4f04f2c DR |
526 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
527 | blockpfn += isolated; | |
528 | break; | |
748446bb | 529 | } |
a4f04f2c DR |
530 | /* Advance to the end of split page */ |
531 | blockpfn += isolated - 1; | |
532 | cursor += isolated - 1; | |
533 | continue; | |
2af120bc LA |
534 | |
535 | isolate_fail: | |
536 | if (strict) | |
537 | break; | |
538 | else | |
539 | continue; | |
540 | ||
748446bb MG |
541 | } |
542 | ||
a4f04f2c DR |
543 | if (locked) |
544 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
545 | ||
9fcd6d2e VB |
546 | /* |
547 | * There is a tiny chance that we have read bogus compound_order(), | |
548 | * so be careful to not go outside of the pageblock. | |
549 | */ | |
550 | if (unlikely(blockpfn > end_pfn)) | |
551 | blockpfn = end_pfn; | |
552 | ||
e34d85f0 JK |
553 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
554 | nr_scanned, total_isolated); | |
555 | ||
e14c720e VB |
556 | /* Record how far we have got within the block */ |
557 | *start_pfn = blockpfn; | |
558 | ||
f40d1e42 MG |
559 | /* |
560 | * If strict isolation is requested by CMA then check that all the | |
561 | * pages requested were isolated. If there were any failures, 0 is | |
562 | * returned and CMA will fail. | |
563 | */ | |
2af120bc | 564 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
565 | total_isolated = 0; |
566 | ||
bb13ffeb MG |
567 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
568 | if (blockpfn == end_pfn) | |
edc2ca61 | 569 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 570 | |
7f354a54 | 571 | cc->total_free_scanned += nr_scanned; |
397487db | 572 | if (total_isolated) |
010fc29a | 573 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
574 | return total_isolated; |
575 | } | |
576 | ||
85aa125f MN |
577 | /** |
578 | * isolate_freepages_range() - isolate free pages. | |
579 | * @start_pfn: The first PFN to start isolating. | |
580 | * @end_pfn: The one-past-last PFN. | |
581 | * | |
582 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
583 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
584 | * undo its actions and return zero. | |
585 | * | |
586 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
587 | * (which may be greater then end_pfn if end fell in a middle of | |
588 | * a free page). | |
589 | */ | |
ff9543fd | 590 | unsigned long |
bb13ffeb MG |
591 | isolate_freepages_range(struct compact_control *cc, |
592 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 593 | { |
e1409c32 | 594 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
595 | LIST_HEAD(freelist); |
596 | ||
7d49d886 | 597 | pfn = start_pfn; |
06b6640a | 598 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
599 | if (block_start_pfn < cc->zone->zone_start_pfn) |
600 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 601 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
602 | |
603 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 604 | block_start_pfn = block_end_pfn, |
7d49d886 | 605 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
606 | /* Protect pfn from changing by isolate_freepages_block */ |
607 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 608 | |
85aa125f MN |
609 | block_end_pfn = min(block_end_pfn, end_pfn); |
610 | ||
58420016 JK |
611 | /* |
612 | * pfn could pass the block_end_pfn if isolated freepage | |
613 | * is more than pageblock order. In this case, we adjust | |
614 | * scanning range to right one. | |
615 | */ | |
616 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
617 | block_start_pfn = pageblock_start_pfn(pfn); |
618 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
619 | block_end_pfn = min(block_end_pfn, end_pfn); |
620 | } | |
621 | ||
e1409c32 JK |
622 | if (!pageblock_pfn_to_page(block_start_pfn, |
623 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
624 | break; |
625 | ||
e14c720e VB |
626 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
627 | block_end_pfn, &freelist, true); | |
85aa125f MN |
628 | |
629 | /* | |
630 | * In strict mode, isolate_freepages_block() returns 0 if | |
631 | * there are any holes in the block (ie. invalid PFNs or | |
632 | * non-free pages). | |
633 | */ | |
634 | if (!isolated) | |
635 | break; | |
636 | ||
637 | /* | |
638 | * If we managed to isolate pages, it is always (1 << n) * | |
639 | * pageblock_nr_pages for some non-negative n. (Max order | |
640 | * page may span two pageblocks). | |
641 | */ | |
642 | } | |
643 | ||
66c64223 | 644 | /* __isolate_free_page() does not map the pages */ |
85aa125f MN |
645 | map_pages(&freelist); |
646 | ||
647 | if (pfn < end_pfn) { | |
648 | /* Loop terminated early, cleanup. */ | |
649 | release_freepages(&freelist); | |
650 | return 0; | |
651 | } | |
652 | ||
653 | /* We don't use freelists for anything. */ | |
654 | return pfn; | |
655 | } | |
656 | ||
748446bb MG |
657 | /* Similar to reclaim, but different enough that they don't share logic */ |
658 | static bool too_many_isolated(struct zone *zone) | |
659 | { | |
bc693045 | 660 | unsigned long active, inactive, isolated; |
748446bb | 661 | |
599d0c95 MG |
662 | inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) + |
663 | node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON); | |
664 | active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) + | |
665 | node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON); | |
666 | isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) + | |
667 | node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON); | |
748446bb | 668 | |
bc693045 | 669 | return isolated > (inactive + active) / 2; |
748446bb MG |
670 | } |
671 | ||
2fe86e00 | 672 | /** |
edc2ca61 VB |
673 | * isolate_migratepages_block() - isolate all migrate-able pages within |
674 | * a single pageblock | |
2fe86e00 | 675 | * @cc: Compaction control structure. |
edc2ca61 VB |
676 | * @low_pfn: The first PFN to isolate |
677 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
678 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
679 | * |
680 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
681 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
682 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
683 | * first page that was not scanned (which may be both less, equal to or more | |
684 | * than end_pfn). | |
2fe86e00 | 685 | * |
edc2ca61 VB |
686 | * The pages are isolated on cc->migratepages list (not required to be empty), |
687 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
688 | * is neither read nor updated. | |
748446bb | 689 | */ |
edc2ca61 VB |
690 | static unsigned long |
691 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
692 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 693 | { |
edc2ca61 | 694 | struct zone *zone = cc->zone; |
b7aba698 | 695 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 696 | struct lruvec *lruvec; |
b8b2d825 | 697 | unsigned long flags = 0; |
2a1402aa | 698 | bool locked = false; |
bb13ffeb | 699 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 700 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
701 | bool skip_on_failure = false; |
702 | unsigned long next_skip_pfn = 0; | |
748446bb | 703 | |
748446bb MG |
704 | /* |
705 | * Ensure that there are not too many pages isolated from the LRU | |
706 | * list by either parallel reclaimers or compaction. If there are, | |
707 | * delay for some time until fewer pages are isolated | |
708 | */ | |
709 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 710 | /* async migration should just abort */ |
e0b9daeb | 711 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 712 | return 0; |
f9e35b3b | 713 | |
748446bb MG |
714 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
715 | ||
716 | if (fatal_signal_pending(current)) | |
2fe86e00 | 717 | return 0; |
748446bb MG |
718 | } |
719 | ||
be976572 VB |
720 | if (compact_should_abort(cc)) |
721 | return 0; | |
aeef4b83 | 722 | |
fdd048e1 VB |
723 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
724 | skip_on_failure = true; | |
725 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
726 | } | |
727 | ||
748446bb | 728 | /* Time to isolate some pages for migration */ |
748446bb | 729 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 730 | |
fdd048e1 VB |
731 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
732 | /* | |
733 | * We have isolated all migration candidates in the | |
734 | * previous order-aligned block, and did not skip it due | |
735 | * to failure. We should migrate the pages now and | |
736 | * hopefully succeed compaction. | |
737 | */ | |
738 | if (nr_isolated) | |
739 | break; | |
740 | ||
741 | /* | |
742 | * We failed to isolate in the previous order-aligned | |
743 | * block. Set the new boundary to the end of the | |
744 | * current block. Note we can't simply increase | |
745 | * next_skip_pfn by 1 << order, as low_pfn might have | |
746 | * been incremented by a higher number due to skipping | |
747 | * a compound or a high-order buddy page in the | |
748 | * previous loop iteration. | |
749 | */ | |
750 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
751 | } | |
752 | ||
8b44d279 VB |
753 | /* |
754 | * Periodically drop the lock (if held) regardless of its | |
755 | * contention, to give chance to IRQs. Abort async compaction | |
756 | * if contended. | |
757 | */ | |
758 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
a52633d8 | 759 | && compact_unlock_should_abort(zone_lru_lock(zone), flags, |
8b44d279 VB |
760 | &locked, cc)) |
761 | break; | |
c67fe375 | 762 | |
748446bb | 763 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 764 | goto isolate_fail; |
b7aba698 | 765 | nr_scanned++; |
748446bb | 766 | |
748446bb | 767 | page = pfn_to_page(low_pfn); |
dc908600 | 768 | |
bb13ffeb MG |
769 | if (!valid_page) |
770 | valid_page = page; | |
771 | ||
6c14466c | 772 | /* |
99c0fd5e VB |
773 | * Skip if free. We read page order here without zone lock |
774 | * which is generally unsafe, but the race window is small and | |
775 | * the worst thing that can happen is that we skip some | |
776 | * potential isolation targets. | |
6c14466c | 777 | */ |
99c0fd5e VB |
778 | if (PageBuddy(page)) { |
779 | unsigned long freepage_order = page_order_unsafe(page); | |
780 | ||
781 | /* | |
782 | * Without lock, we cannot be sure that what we got is | |
783 | * a valid page order. Consider only values in the | |
784 | * valid order range to prevent low_pfn overflow. | |
785 | */ | |
786 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
787 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 788 | continue; |
99c0fd5e | 789 | } |
748446bb | 790 | |
bc835011 | 791 | /* |
29c0dde8 VB |
792 | * Regardless of being on LRU, compound pages such as THP and |
793 | * hugetlbfs are not to be compacted. We can potentially save | |
794 | * a lot of iterations if we skip them at once. The check is | |
795 | * racy, but we can consider only valid values and the only | |
796 | * danger is skipping too much. | |
bc835011 | 797 | */ |
29c0dde8 | 798 | if (PageCompound(page)) { |
21dc7e02 | 799 | const unsigned int order = compound_order(page); |
edc2ca61 | 800 | |
21dc7e02 DR |
801 | if (pageblock_skip_persistent(page, order)) { |
802 | set_pageblock_skip(page); | |
803 | low_pfn = end_pfn; | |
804 | } else if (likely(order < MAX_ORDER)) | |
805 | low_pfn += (1UL << order) - 1; | |
fdd048e1 | 806 | goto isolate_fail; |
2a1402aa MG |
807 | } |
808 | ||
bda807d4 MK |
809 | /* |
810 | * Check may be lockless but that's ok as we recheck later. | |
811 | * It's possible to migrate LRU and non-lru movable pages. | |
812 | * Skip any other type of page | |
813 | */ | |
814 | if (!PageLRU(page)) { | |
bda807d4 MK |
815 | /* |
816 | * __PageMovable can return false positive so we need | |
817 | * to verify it under page_lock. | |
818 | */ | |
819 | if (unlikely(__PageMovable(page)) && | |
820 | !PageIsolated(page)) { | |
821 | if (locked) { | |
a52633d8 | 822 | spin_unlock_irqrestore(zone_lru_lock(zone), |
bda807d4 MK |
823 | flags); |
824 | locked = false; | |
825 | } | |
826 | ||
9e5bcd61 | 827 | if (!isolate_movable_page(page, isolate_mode)) |
bda807d4 MK |
828 | goto isolate_success; |
829 | } | |
830 | ||
fdd048e1 | 831 | goto isolate_fail; |
bda807d4 | 832 | } |
29c0dde8 | 833 | |
119d6d59 DR |
834 | /* |
835 | * Migration will fail if an anonymous page is pinned in memory, | |
836 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
837 | * admittedly racy check. | |
838 | */ | |
839 | if (!page_mapping(page) && | |
840 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 841 | goto isolate_fail; |
119d6d59 | 842 | |
73e64c51 MH |
843 | /* |
844 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
845 | * because those do not depend on fs locks. | |
846 | */ | |
847 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
848 | goto isolate_fail; | |
849 | ||
69b7189f VB |
850 | /* If we already hold the lock, we can skip some rechecking */ |
851 | if (!locked) { | |
a52633d8 | 852 | locked = compact_trylock_irqsave(zone_lru_lock(zone), |
8b44d279 | 853 | &flags, cc); |
69b7189f VB |
854 | if (!locked) |
855 | break; | |
2a1402aa | 856 | |
29c0dde8 | 857 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 858 | if (!PageLRU(page)) |
fdd048e1 | 859 | goto isolate_fail; |
29c0dde8 VB |
860 | |
861 | /* | |
862 | * Page become compound since the non-locked check, | |
863 | * and it's on LRU. It can only be a THP so the order | |
864 | * is safe to read and it's 0 for tail pages. | |
865 | */ | |
866 | if (unlikely(PageCompound(page))) { | |
21dc7e02 DR |
867 | const unsigned int order = compound_order(page); |
868 | ||
869 | if (pageblock_skip_persistent(page, order)) { | |
870 | set_pageblock_skip(page); | |
871 | low_pfn = end_pfn; | |
872 | } else | |
873 | low_pfn += (1UL << order) - 1; | |
fdd048e1 | 874 | goto isolate_fail; |
69b7189f | 875 | } |
bc835011 AA |
876 | } |
877 | ||
599d0c95 | 878 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
fa9add64 | 879 | |
748446bb | 880 | /* Try isolate the page */ |
edc2ca61 | 881 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 882 | goto isolate_fail; |
748446bb | 883 | |
29c0dde8 | 884 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 885 | |
748446bb | 886 | /* Successfully isolated */ |
fa9add64 | 887 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
6afcf8ef ML |
888 | inc_node_page_state(page, |
889 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
b6c75016 JK |
890 | |
891 | isolate_success: | |
fdd048e1 | 892 | list_add(&page->lru, &cc->migratepages); |
748446bb | 893 | cc->nr_migratepages++; |
b7aba698 | 894 | nr_isolated++; |
748446bb | 895 | |
a34753d2 VB |
896 | /* |
897 | * Record where we could have freed pages by migration and not | |
898 | * yet flushed them to buddy allocator. | |
899 | * - this is the lowest page that was isolated and likely be | |
900 | * then freed by migration. | |
901 | */ | |
902 | if (!cc->last_migrated_pfn) | |
903 | cc->last_migrated_pfn = low_pfn; | |
904 | ||
748446bb | 905 | /* Avoid isolating too much */ |
31b8384a HD |
906 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
907 | ++low_pfn; | |
748446bb | 908 | break; |
31b8384a | 909 | } |
fdd048e1 VB |
910 | |
911 | continue; | |
912 | isolate_fail: | |
913 | if (!skip_on_failure) | |
914 | continue; | |
915 | ||
916 | /* | |
917 | * We have isolated some pages, but then failed. Release them | |
918 | * instead of migrating, as we cannot form the cc->order buddy | |
919 | * page anyway. | |
920 | */ | |
921 | if (nr_isolated) { | |
922 | if (locked) { | |
a52633d8 | 923 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
fdd048e1 VB |
924 | locked = false; |
925 | } | |
fdd048e1 VB |
926 | putback_movable_pages(&cc->migratepages); |
927 | cc->nr_migratepages = 0; | |
928 | cc->last_migrated_pfn = 0; | |
929 | nr_isolated = 0; | |
930 | } | |
931 | ||
932 | if (low_pfn < next_skip_pfn) { | |
933 | low_pfn = next_skip_pfn - 1; | |
934 | /* | |
935 | * The check near the loop beginning would have updated | |
936 | * next_skip_pfn too, but this is a bit simpler. | |
937 | */ | |
938 | next_skip_pfn += 1UL << cc->order; | |
939 | } | |
748446bb MG |
940 | } |
941 | ||
99c0fd5e VB |
942 | /* |
943 | * The PageBuddy() check could have potentially brought us outside | |
944 | * the range to be scanned. | |
945 | */ | |
946 | if (unlikely(low_pfn > end_pfn)) | |
947 | low_pfn = end_pfn; | |
948 | ||
c67fe375 | 949 | if (locked) |
a52633d8 | 950 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
748446bb | 951 | |
50b5b094 VB |
952 | /* |
953 | * Update the pageblock-skip information and cached scanner pfn, | |
954 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 955 | */ |
35979ef3 | 956 | if (low_pfn == end_pfn) |
edc2ca61 | 957 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 958 | |
e34d85f0 JK |
959 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
960 | nr_scanned, nr_isolated); | |
b7aba698 | 961 | |
7f354a54 | 962 | cc->total_migrate_scanned += nr_scanned; |
397487db | 963 | if (nr_isolated) |
010fc29a | 964 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 965 | |
2fe86e00 MN |
966 | return low_pfn; |
967 | } | |
968 | ||
edc2ca61 VB |
969 | /** |
970 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
971 | * @cc: Compaction control structure. | |
972 | * @start_pfn: The first PFN to start isolating. | |
973 | * @end_pfn: The one-past-last PFN. | |
974 | * | |
975 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
976 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
977 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
978 | */ | |
979 | unsigned long | |
980 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
981 | unsigned long end_pfn) | |
982 | { | |
e1409c32 | 983 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
984 | |
985 | /* Scan block by block. First and last block may be incomplete */ | |
986 | pfn = start_pfn; | |
06b6640a | 987 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
988 | if (block_start_pfn < cc->zone->zone_start_pfn) |
989 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 990 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
991 | |
992 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 993 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
994 | block_end_pfn += pageblock_nr_pages) { |
995 | ||
996 | block_end_pfn = min(block_end_pfn, end_pfn); | |
997 | ||
e1409c32 JK |
998 | if (!pageblock_pfn_to_page(block_start_pfn, |
999 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1000 | continue; |
1001 | ||
1002 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
1003 | ISOLATE_UNEVICTABLE); | |
1004 | ||
14af4a5e | 1005 | if (!pfn) |
edc2ca61 | 1006 | break; |
6ea41c0c JK |
1007 | |
1008 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
1009 | break; | |
edc2ca61 | 1010 | } |
edc2ca61 VB |
1011 | |
1012 | return pfn; | |
1013 | } | |
1014 | ||
ff9543fd MN |
1015 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1016 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1017 | |
b682debd VB |
1018 | static bool suitable_migration_source(struct compact_control *cc, |
1019 | struct page *page) | |
1020 | { | |
282722b0 VB |
1021 | int block_mt; |
1022 | ||
1023 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) | |
b682debd VB |
1024 | return true; |
1025 | ||
282722b0 VB |
1026 | block_mt = get_pageblock_migratetype(page); |
1027 | ||
1028 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1029 | return is_migrate_movable(block_mt); | |
1030 | else | |
1031 | return block_mt == cc->migratetype; | |
b682debd VB |
1032 | } |
1033 | ||
018e9a49 | 1034 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1035 | static bool suitable_migration_target(struct compact_control *cc, |
1036 | struct page *page) | |
018e9a49 AM |
1037 | { |
1038 | /* If the page is a large free page, then disallow migration */ | |
1039 | if (PageBuddy(page)) { | |
1040 | /* | |
1041 | * We are checking page_order without zone->lock taken. But | |
1042 | * the only small danger is that we skip a potentially suitable | |
1043 | * pageblock, so it's not worth to check order for valid range. | |
1044 | */ | |
1045 | if (page_order_unsafe(page) >= pageblock_order) | |
1046 | return false; | |
1047 | } | |
1048 | ||
1ef36db2 YX |
1049 | if (cc->ignore_block_suitable) |
1050 | return true; | |
1051 | ||
018e9a49 | 1052 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1053 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1054 | return true; |
1055 | ||
1056 | /* Otherwise skip the block */ | |
1057 | return false; | |
1058 | } | |
1059 | ||
f2849aa0 VB |
1060 | /* |
1061 | * Test whether the free scanner has reached the same or lower pageblock than | |
1062 | * the migration scanner, and compaction should thus terminate. | |
1063 | */ | |
1064 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1065 | { | |
1066 | return (cc->free_pfn >> pageblock_order) | |
1067 | <= (cc->migrate_pfn >> pageblock_order); | |
1068 | } | |
1069 | ||
2fe86e00 | 1070 | /* |
ff9543fd MN |
1071 | * Based on information in the current compact_control, find blocks |
1072 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1073 | */ |
edc2ca61 | 1074 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1075 | { |
edc2ca61 | 1076 | struct zone *zone = cc->zone; |
ff9543fd | 1077 | struct page *page; |
c96b9e50 | 1078 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1079 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1080 | unsigned long block_end_pfn; /* end of current pageblock */ |
1081 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1082 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 1083 | |
ff9543fd MN |
1084 | /* |
1085 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1086 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1087 | * zone when isolating for the first time. For looping we also need |
1088 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1089 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1090 | * For ending point, take care when isolating in last pageblock of a | |
1091 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1092 | * The low boundary is the end of the pageblock the migration scanner |
1093 | * is using. | |
ff9543fd | 1094 | */ |
e14c720e | 1095 | isolate_start_pfn = cc->free_pfn; |
06b6640a | 1096 | block_start_pfn = pageblock_start_pfn(cc->free_pfn); |
c96b9e50 VB |
1097 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1098 | zone_end_pfn(zone)); | |
06b6640a | 1099 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
2fe86e00 | 1100 | |
ff9543fd MN |
1101 | /* |
1102 | * Isolate free pages until enough are available to migrate the | |
1103 | * pages on cc->migratepages. We stop searching if the migrate | |
1104 | * and free page scanners meet or enough free pages are isolated. | |
1105 | */ | |
f5f61a32 | 1106 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1107 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1108 | block_start_pfn -= pageblock_nr_pages, |
1109 | isolate_start_pfn = block_start_pfn) { | |
f6ea3adb DR |
1110 | /* |
1111 | * This can iterate a massively long zone without finding any | |
1112 | * suitable migration targets, so periodically check if we need | |
be976572 | 1113 | * to schedule, or even abort async compaction. |
f6ea3adb | 1114 | */ |
be976572 VB |
1115 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
1116 | && compact_should_abort(cc)) | |
1117 | break; | |
f6ea3adb | 1118 | |
7d49d886 VB |
1119 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1120 | zone); | |
1121 | if (!page) | |
ff9543fd MN |
1122 | continue; |
1123 | ||
1124 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1125 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1126 | continue; |
68e3e926 | 1127 | |
bb13ffeb MG |
1128 | /* If isolation recently failed, do not retry */ |
1129 | if (!isolation_suitable(cc, page)) | |
1130 | continue; | |
1131 | ||
e14c720e | 1132 | /* Found a block suitable for isolating free pages from. */ |
a46cbf3b DR |
1133 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, |
1134 | freelist, false); | |
ff9543fd | 1135 | |
e14c720e | 1136 | /* |
a46cbf3b DR |
1137 | * If we isolated enough freepages, or aborted due to lock |
1138 | * contention, terminate. | |
e14c720e | 1139 | */ |
f5f61a32 VB |
1140 | if ((cc->nr_freepages >= cc->nr_migratepages) |
1141 | || cc->contended) { | |
a46cbf3b DR |
1142 | if (isolate_start_pfn >= block_end_pfn) { |
1143 | /* | |
1144 | * Restart at previous pageblock if more | |
1145 | * freepages can be isolated next time. | |
1146 | */ | |
f5f61a32 VB |
1147 | isolate_start_pfn = |
1148 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1149 | } |
be976572 | 1150 | break; |
a46cbf3b | 1151 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1152 | /* |
a46cbf3b DR |
1153 | * If isolation failed early, do not continue |
1154 | * needlessly. | |
f5f61a32 | 1155 | */ |
a46cbf3b | 1156 | break; |
f5f61a32 | 1157 | } |
ff9543fd MN |
1158 | } |
1159 | ||
66c64223 | 1160 | /* __isolate_free_page() does not map the pages */ |
ff9543fd MN |
1161 | map_pages(freelist); |
1162 | ||
7ed695e0 | 1163 | /* |
f5f61a32 VB |
1164 | * Record where the free scanner will restart next time. Either we |
1165 | * broke from the loop and set isolate_start_pfn based on the last | |
1166 | * call to isolate_freepages_block(), or we met the migration scanner | |
1167 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1168 | */ |
f5f61a32 | 1169 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1170 | } |
1171 | ||
1172 | /* | |
1173 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1174 | * from the isolated freelists in the block we are migrating to. | |
1175 | */ | |
1176 | static struct page *compaction_alloc(struct page *migratepage, | |
1177 | unsigned long data, | |
1178 | int **result) | |
1179 | { | |
1180 | struct compact_control *cc = (struct compact_control *)data; | |
1181 | struct page *freepage; | |
1182 | ||
be976572 VB |
1183 | /* |
1184 | * Isolate free pages if necessary, and if we are not aborting due to | |
1185 | * contention. | |
1186 | */ | |
748446bb | 1187 | if (list_empty(&cc->freepages)) { |
be976572 | 1188 | if (!cc->contended) |
edc2ca61 | 1189 | isolate_freepages(cc); |
748446bb MG |
1190 | |
1191 | if (list_empty(&cc->freepages)) | |
1192 | return NULL; | |
1193 | } | |
1194 | ||
1195 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1196 | list_del(&freepage->lru); | |
1197 | cc->nr_freepages--; | |
1198 | ||
1199 | return freepage; | |
1200 | } | |
1201 | ||
1202 | /* | |
d53aea3d DR |
1203 | * This is a migrate-callback that "frees" freepages back to the isolated |
1204 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1205 | * special handling needed for NUMA. | |
1206 | */ | |
1207 | static void compaction_free(struct page *page, unsigned long data) | |
1208 | { | |
1209 | struct compact_control *cc = (struct compact_control *)data; | |
1210 | ||
1211 | list_add(&page->lru, &cc->freepages); | |
1212 | cc->nr_freepages++; | |
1213 | } | |
1214 | ||
ff9543fd MN |
1215 | /* possible outcome of isolate_migratepages */ |
1216 | typedef enum { | |
1217 | ISOLATE_ABORT, /* Abort compaction now */ | |
1218 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1219 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1220 | } isolate_migrate_t; | |
1221 | ||
5bbe3547 EM |
1222 | /* |
1223 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1224 | * compactable pages. | |
1225 | */ | |
1226 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1227 | ||
ff9543fd | 1228 | /* |
edc2ca61 VB |
1229 | * Isolate all pages that can be migrated from the first suitable block, |
1230 | * starting at the block pointed to by the migrate scanner pfn within | |
1231 | * compact_control. | |
ff9543fd MN |
1232 | */ |
1233 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1234 | struct compact_control *cc) | |
1235 | { | |
e1409c32 JK |
1236 | unsigned long block_start_pfn; |
1237 | unsigned long block_end_pfn; | |
1238 | unsigned long low_pfn; | |
edc2ca61 VB |
1239 | struct page *page; |
1240 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1241 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1242 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1243 | |
edc2ca61 VB |
1244 | /* |
1245 | * Start at where we last stopped, or beginning of the zone as | |
1246 | * initialized by compact_zone() | |
1247 | */ | |
1248 | low_pfn = cc->migrate_pfn; | |
06b6640a | 1249 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1250 | if (block_start_pfn < zone->zone_start_pfn) |
1251 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd MN |
1252 | |
1253 | /* Only scan within a pageblock boundary */ | |
06b6640a | 1254 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1255 | |
edc2ca61 VB |
1256 | /* |
1257 | * Iterate over whole pageblocks until we find the first suitable. | |
1258 | * Do not cross the free scanner. | |
1259 | */ | |
e1409c32 JK |
1260 | for (; block_end_pfn <= cc->free_pfn; |
1261 | low_pfn = block_end_pfn, | |
1262 | block_start_pfn = block_end_pfn, | |
1263 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1264 | |
edc2ca61 VB |
1265 | /* |
1266 | * This can potentially iterate a massively long zone with | |
1267 | * many pageblocks unsuitable, so periodically check if we | |
1268 | * need to schedule, or even abort async compaction. | |
1269 | */ | |
1270 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1271 | && compact_should_abort(cc)) | |
1272 | break; | |
ff9543fd | 1273 | |
e1409c32 JK |
1274 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1275 | zone); | |
7d49d886 | 1276 | if (!page) |
edc2ca61 VB |
1277 | continue; |
1278 | ||
edc2ca61 VB |
1279 | /* If isolation recently failed, do not retry */ |
1280 | if (!isolation_suitable(cc, page)) | |
1281 | continue; | |
1282 | ||
1283 | /* | |
1284 | * For async compaction, also only scan in MOVABLE blocks. | |
1285 | * Async compaction is optimistic to see if the minimum amount | |
1286 | * of work satisfies the allocation. | |
1287 | */ | |
b682debd | 1288 | if (!suitable_migration_source(cc, page)) |
edc2ca61 VB |
1289 | continue; |
1290 | ||
1291 | /* Perform the isolation */ | |
e1409c32 JK |
1292 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1293 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1294 | |
6afcf8ef | 1295 | if (!low_pfn || cc->contended) |
edc2ca61 VB |
1296 | return ISOLATE_ABORT; |
1297 | ||
1298 | /* | |
1299 | * Either we isolated something and proceed with migration. Or | |
1300 | * we failed and compact_zone should decide if we should | |
1301 | * continue or not. | |
1302 | */ | |
1303 | break; | |
1304 | } | |
1305 | ||
f2849aa0 VB |
1306 | /* Record where migration scanner will be restarted. */ |
1307 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1308 | |
edc2ca61 | 1309 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1310 | } |
1311 | ||
21c527a3 YB |
1312 | /* |
1313 | * order == -1 is expected when compacting via | |
1314 | * /proc/sys/vm/compact_memory | |
1315 | */ | |
1316 | static inline bool is_via_compact_memory(int order) | |
1317 | { | |
1318 | return order == -1; | |
1319 | } | |
1320 | ||
d39773a0 VB |
1321 | static enum compact_result __compact_finished(struct zone *zone, |
1322 | struct compact_control *cc) | |
748446bb | 1323 | { |
8fb74b9f | 1324 | unsigned int order; |
d39773a0 | 1325 | const int migratetype = cc->migratetype; |
56de7263 | 1326 | |
be976572 | 1327 | if (cc->contended || fatal_signal_pending(current)) |
2d1e1041 | 1328 | return COMPACT_CONTENDED; |
748446bb | 1329 | |
753341a4 | 1330 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1331 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1332 | /* Let the next compaction start anew. */ |
02333641 | 1333 | reset_cached_positions(zone); |
55b7c4c9 | 1334 | |
62997027 MG |
1335 | /* |
1336 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1337 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1338 | * flag itself as the decision to be clear should be directly |
1339 | * based on an allocation request. | |
1340 | */ | |
accf6242 | 1341 | if (cc->direct_compaction) |
62997027 MG |
1342 | zone->compact_blockskip_flush = true; |
1343 | ||
c8f7de0b MH |
1344 | if (cc->whole_zone) |
1345 | return COMPACT_COMPLETE; | |
1346 | else | |
1347 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1348 | } |
748446bb | 1349 | |
21c527a3 | 1350 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1351 | return COMPACT_CONTINUE; |
1352 | ||
baf6a9a1 VB |
1353 | if (cc->finishing_block) { |
1354 | /* | |
1355 | * We have finished the pageblock, but better check again that | |
1356 | * we really succeeded. | |
1357 | */ | |
1358 | if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1359 | cc->finishing_block = false; | |
1360 | else | |
1361 | return COMPACT_CONTINUE; | |
1362 | } | |
1363 | ||
56de7263 | 1364 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f MG |
1365 | for (order = cc->order; order < MAX_ORDER; order++) { |
1366 | struct free_area *area = &zone->free_area[order]; | |
2149cdae | 1367 | bool can_steal; |
8fb74b9f MG |
1368 | |
1369 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1370 | if (!list_empty(&area->free_list[migratetype])) |
cf378319 | 1371 | return COMPACT_SUCCESS; |
8fb74b9f | 1372 | |
2149cdae JK |
1373 | #ifdef CONFIG_CMA |
1374 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1375 | if (migratetype == MIGRATE_MOVABLE && | |
1376 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
cf378319 | 1377 | return COMPACT_SUCCESS; |
2149cdae JK |
1378 | #endif |
1379 | /* | |
1380 | * Job done if allocation would steal freepages from | |
1381 | * other migratetype buddy lists. | |
1382 | */ | |
1383 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
1384 | true, &can_steal) != -1) { |
1385 | ||
1386 | /* movable pages are OK in any pageblock */ | |
1387 | if (migratetype == MIGRATE_MOVABLE) | |
1388 | return COMPACT_SUCCESS; | |
1389 | ||
1390 | /* | |
1391 | * We are stealing for a non-movable allocation. Make | |
1392 | * sure we finish compacting the current pageblock | |
1393 | * first so it is as free as possible and we won't | |
1394 | * have to steal another one soon. This only applies | |
1395 | * to sync compaction, as async compaction operates | |
1396 | * on pageblocks of the same migratetype. | |
1397 | */ | |
1398 | if (cc->mode == MIGRATE_ASYNC || | |
1399 | IS_ALIGNED(cc->migrate_pfn, | |
1400 | pageblock_nr_pages)) { | |
1401 | return COMPACT_SUCCESS; | |
1402 | } | |
1403 | ||
1404 | cc->finishing_block = true; | |
1405 | return COMPACT_CONTINUE; | |
1406 | } | |
56de7263 MG |
1407 | } |
1408 | ||
837d026d JK |
1409 | return COMPACT_NO_SUITABLE_PAGE; |
1410 | } | |
1411 | ||
ea7ab982 | 1412 | static enum compact_result compact_finished(struct zone *zone, |
d39773a0 | 1413 | struct compact_control *cc) |
837d026d JK |
1414 | { |
1415 | int ret; | |
1416 | ||
d39773a0 | 1417 | ret = __compact_finished(zone, cc); |
837d026d JK |
1418 | trace_mm_compaction_finished(zone, cc->order, ret); |
1419 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1420 | ret = COMPACT_CONTINUE; | |
1421 | ||
1422 | return ret; | |
748446bb MG |
1423 | } |
1424 | ||
3e7d3449 MG |
1425 | /* |
1426 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1427 | * Returns | |
1428 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
cf378319 | 1429 | * COMPACT_SUCCESS - If the allocation would succeed without compaction |
3e7d3449 MG |
1430 | * COMPACT_CONTINUE - If compaction should run now |
1431 | */ | |
ea7ab982 | 1432 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1433 | unsigned int alloc_flags, |
86a294a8 MH |
1434 | int classzone_idx, |
1435 | unsigned long wmark_target) | |
3e7d3449 | 1436 | { |
3e7d3449 MG |
1437 | unsigned long watermark; |
1438 | ||
21c527a3 | 1439 | if (is_via_compact_memory(order)) |
3957c776 MH |
1440 | return COMPACT_CONTINUE; |
1441 | ||
f2b8228c | 1442 | watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; |
ebff3980 VB |
1443 | /* |
1444 | * If watermarks for high-order allocation are already met, there | |
1445 | * should be no need for compaction at all. | |
1446 | */ | |
1447 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1448 | alloc_flags)) | |
cf378319 | 1449 | return COMPACT_SUCCESS; |
ebff3980 | 1450 | |
3e7d3449 | 1451 | /* |
9861a62c | 1452 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
1453 | * isolate free pages for migration targets. This means that the |
1454 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1455 | * the check in __isolate_free_page(). We don't use the direct | |
1456 | * compactor's alloc_flags, as they are not relevant for freepage | |
1457 | * isolation. We however do use the direct compactor's classzone_idx to | |
1458 | * skip over zones where lowmem reserves would prevent allocation even | |
1459 | * if compaction succeeds. | |
8348faf9 VB |
1460 | * For costly orders, we require low watermark instead of min for |
1461 | * compaction to proceed to increase its chances. | |
984fdba6 VB |
1462 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
1463 | * suitable migration targets | |
3e7d3449 | 1464 | */ |
8348faf9 VB |
1465 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
1466 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
1467 | watermark += compact_gap(order); | |
86a294a8 | 1468 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
984fdba6 | 1469 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
1470 | return COMPACT_SKIPPED; |
1471 | ||
cc5c9f09 VB |
1472 | return COMPACT_CONTINUE; |
1473 | } | |
1474 | ||
1475 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
1476 | unsigned int alloc_flags, | |
1477 | int classzone_idx) | |
1478 | { | |
1479 | enum compact_result ret; | |
1480 | int fragindex; | |
1481 | ||
1482 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
1483 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
1484 | /* |
1485 | * fragmentation index determines if allocation failures are due to | |
1486 | * low memory or external fragmentation | |
1487 | * | |
ebff3980 VB |
1488 | * index of -1000 would imply allocations might succeed depending on |
1489 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1490 | * index towards 0 implies failure is due to lack of memory |
1491 | * index towards 1000 implies failure is due to fragmentation | |
1492 | * | |
20311420 VB |
1493 | * Only compact if a failure would be due to fragmentation. Also |
1494 | * ignore fragindex for non-costly orders where the alternative to | |
1495 | * a successful reclaim/compaction is OOM. Fragindex and the | |
1496 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
1497 | * excessive compaction for costly orders, but it should not be at the | |
1498 | * expense of system stability. | |
3e7d3449 | 1499 | */ |
20311420 | 1500 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
1501 | fragindex = fragmentation_index(zone, order); |
1502 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
1503 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
1504 | } | |
837d026d | 1505 | |
837d026d JK |
1506 | trace_mm_compaction_suitable(zone, order, ret); |
1507 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1508 | ret = COMPACT_SKIPPED; | |
1509 | ||
1510 | return ret; | |
1511 | } | |
1512 | ||
86a294a8 MH |
1513 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
1514 | int alloc_flags) | |
1515 | { | |
1516 | struct zone *zone; | |
1517 | struct zoneref *z; | |
1518 | ||
1519 | /* | |
1520 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1521 | * retrying the reclaim. | |
1522 | */ | |
1523 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1524 | ac->nodemask) { | |
1525 | unsigned long available; | |
1526 | enum compact_result compact_result; | |
1527 | ||
1528 | /* | |
1529 | * Do not consider all the reclaimable memory because we do not | |
1530 | * want to trash just for a single high order allocation which | |
1531 | * is even not guaranteed to appear even if __compaction_suitable | |
1532 | * is happy about the watermark check. | |
1533 | */ | |
5a1c84b4 | 1534 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
1535 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
1536 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1537 | ac_classzone_idx(ac), available); | |
cc5c9f09 | 1538 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
1539 | return true; |
1540 | } | |
1541 | ||
1542 | return false; | |
1543 | } | |
1544 | ||
ea7ab982 | 1545 | static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc) |
748446bb | 1546 | { |
ea7ab982 | 1547 | enum compact_result ret; |
c89511ab | 1548 | unsigned long start_pfn = zone->zone_start_pfn; |
108bcc96 | 1549 | unsigned long end_pfn = zone_end_pfn(zone); |
e0b9daeb | 1550 | const bool sync = cc->mode != MIGRATE_ASYNC; |
748446bb | 1551 | |
d39773a0 | 1552 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
ebff3980 VB |
1553 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, |
1554 | cc->classzone_idx); | |
c46649de | 1555 | /* Compaction is likely to fail */ |
cf378319 | 1556 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 1557 | return ret; |
c46649de MH |
1558 | |
1559 | /* huh, compaction_suitable is returning something unexpected */ | |
1560 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 1561 | |
d3132e4b VB |
1562 | /* |
1563 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 1564 | * is about to be retried after being deferred. |
d3132e4b | 1565 | */ |
accf6242 | 1566 | if (compaction_restarting(zone, cc->order)) |
d3132e4b VB |
1567 | __reset_isolation_suitable(zone); |
1568 | ||
c89511ab MG |
1569 | /* |
1570 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
1571 | * information on where the scanners should start (unless we explicitly |
1572 | * want to compact the whole zone), but check that it is initialised | |
1573 | * by ensuring the values are within zone boundaries. | |
c89511ab | 1574 | */ |
06ed2998 | 1575 | if (cc->whole_zone) { |
c89511ab | 1576 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
1577 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
1578 | } else { | |
1579 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; | |
1580 | cc->free_pfn = zone->compact_cached_free_pfn; | |
1581 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { | |
1582 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
1583 | zone->compact_cached_free_pfn = cc->free_pfn; | |
1584 | } | |
1585 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
1586 | cc->migrate_pfn = start_pfn; | |
1587 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; | |
1588 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
1589 | } | |
c8f7de0b | 1590 | |
06ed2998 VB |
1591 | if (cc->migrate_pfn == start_pfn) |
1592 | cc->whole_zone = true; | |
1593 | } | |
c8f7de0b | 1594 | |
1a16718c | 1595 | cc->last_migrated_pfn = 0; |
748446bb | 1596 | |
16c4a097 JK |
1597 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1598 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1599 | |
748446bb MG |
1600 | migrate_prep_local(); |
1601 | ||
d39773a0 | 1602 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { |
9d502c1c | 1603 | int err; |
748446bb | 1604 | |
f9e35b3b MG |
1605 | switch (isolate_migratepages(zone, cc)) { |
1606 | case ISOLATE_ABORT: | |
2d1e1041 | 1607 | ret = COMPACT_CONTENDED; |
5733c7d1 | 1608 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1609 | cc->nr_migratepages = 0; |
f9e35b3b MG |
1610 | goto out; |
1611 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1612 | /* |
1613 | * We haven't isolated and migrated anything, but | |
1614 | * there might still be unflushed migrations from | |
1615 | * previous cc->order aligned block. | |
1616 | */ | |
1617 | goto check_drain; | |
f9e35b3b MG |
1618 | case ISOLATE_SUCCESS: |
1619 | ; | |
1620 | } | |
748446bb | 1621 | |
d53aea3d | 1622 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1623 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1624 | MR_COMPACTION); |
748446bb | 1625 | |
f8c9301f VB |
1626 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1627 | &cc->migratepages); | |
748446bb | 1628 | |
f8c9301f VB |
1629 | /* All pages were either migrated or will be released */ |
1630 | cc->nr_migratepages = 0; | |
9d502c1c | 1631 | if (err) { |
5733c7d1 | 1632 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1633 | /* |
1634 | * migrate_pages() may return -ENOMEM when scanners meet | |
1635 | * and we want compact_finished() to detect it | |
1636 | */ | |
f2849aa0 | 1637 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 1638 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
1639 | goto out; |
1640 | } | |
fdd048e1 VB |
1641 | /* |
1642 | * We failed to migrate at least one page in the current | |
1643 | * order-aligned block, so skip the rest of it. | |
1644 | */ | |
1645 | if (cc->direct_compaction && | |
1646 | (cc->mode == MIGRATE_ASYNC)) { | |
1647 | cc->migrate_pfn = block_end_pfn( | |
1648 | cc->migrate_pfn - 1, cc->order); | |
1649 | /* Draining pcplists is useless in this case */ | |
1650 | cc->last_migrated_pfn = 0; | |
1651 | ||
1652 | } | |
748446bb | 1653 | } |
fdaf7f5c | 1654 | |
fdaf7f5c VB |
1655 | check_drain: |
1656 | /* | |
1657 | * Has the migration scanner moved away from the previous | |
1658 | * cc->order aligned block where we migrated from? If yes, | |
1659 | * flush the pages that were freed, so that they can merge and | |
1660 | * compact_finished() can detect immediately if allocation | |
1661 | * would succeed. | |
1662 | */ | |
1a16718c | 1663 | if (cc->order > 0 && cc->last_migrated_pfn) { |
fdaf7f5c VB |
1664 | int cpu; |
1665 | unsigned long current_block_start = | |
06b6640a | 1666 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 1667 | |
1a16718c | 1668 | if (cc->last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
1669 | cpu = get_cpu(); |
1670 | lru_add_drain_cpu(cpu); | |
1671 | drain_local_pages(zone); | |
1672 | put_cpu(); | |
1673 | /* No more flushing until we migrate again */ | |
1a16718c | 1674 | cc->last_migrated_pfn = 0; |
fdaf7f5c VB |
1675 | } |
1676 | } | |
1677 | ||
748446bb MG |
1678 | } |
1679 | ||
f9e35b3b | 1680 | out: |
6bace090 VB |
1681 | /* |
1682 | * Release free pages and update where the free scanner should restart, | |
1683 | * so we don't leave any returned pages behind in the next attempt. | |
1684 | */ | |
1685 | if (cc->nr_freepages > 0) { | |
1686 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1687 | ||
1688 | cc->nr_freepages = 0; | |
1689 | VM_BUG_ON(free_pfn == 0); | |
1690 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 1691 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
1692 | /* |
1693 | * Only go back, not forward. The cached pfn might have been | |
1694 | * already reset to zone end in compact_finished() | |
1695 | */ | |
1696 | if (free_pfn > zone->compact_cached_free_pfn) | |
1697 | zone->compact_cached_free_pfn = free_pfn; | |
1698 | } | |
748446bb | 1699 | |
7f354a54 DR |
1700 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
1701 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
1702 | ||
16c4a097 JK |
1703 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1704 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1705 | |
748446bb MG |
1706 | return ret; |
1707 | } | |
76ab0f53 | 1708 | |
ea7ab982 | 1709 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 1710 | gfp_t gfp_mask, enum compact_priority prio, |
c603844b | 1711 | unsigned int alloc_flags, int classzone_idx) |
56de7263 | 1712 | { |
ea7ab982 | 1713 | enum compact_result ret; |
56de7263 MG |
1714 | struct compact_control cc = { |
1715 | .nr_freepages = 0, | |
1716 | .nr_migratepages = 0, | |
7f354a54 DR |
1717 | .total_migrate_scanned = 0, |
1718 | .total_free_scanned = 0, | |
56de7263 | 1719 | .order = order, |
6d7ce559 | 1720 | .gfp_mask = gfp_mask, |
56de7263 | 1721 | .zone = zone, |
a5508cd8 VB |
1722 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
1723 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 VB |
1724 | .alloc_flags = alloc_flags, |
1725 | .classzone_idx = classzone_idx, | |
accf6242 | 1726 | .direct_compaction = true, |
a8e025e5 | 1727 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
1728 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
1729 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 MG |
1730 | }; |
1731 | INIT_LIST_HEAD(&cc.freepages); | |
1732 | INIT_LIST_HEAD(&cc.migratepages); | |
1733 | ||
e64c5237 SL |
1734 | ret = compact_zone(zone, &cc); |
1735 | ||
1736 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1737 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1738 | ||
e64c5237 | 1739 | return ret; |
56de7263 MG |
1740 | } |
1741 | ||
5e771905 MG |
1742 | int sysctl_extfrag_threshold = 500; |
1743 | ||
56de7263 MG |
1744 | /** |
1745 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1746 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1747 | * @order: The order of the current allocation |
1748 | * @alloc_flags: The allocation flags of the current allocation | |
1749 | * @ac: The context of current allocation | |
e0b9daeb | 1750 | * @mode: The migration mode for async, sync light, or sync migration |
56de7263 MG |
1751 | * |
1752 | * This is the main entry point for direct page compaction. | |
1753 | */ | |
ea7ab982 | 1754 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 1755 | unsigned int alloc_flags, const struct alloc_context *ac, |
c3486f53 | 1756 | enum compact_priority prio) |
56de7263 | 1757 | { |
56de7263 | 1758 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
1759 | struct zoneref *z; |
1760 | struct zone *zone; | |
1d4746d3 | 1761 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 1762 | |
73e64c51 MH |
1763 | /* |
1764 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
1765 | * tricky context because the migration might require IO | |
1766 | */ | |
1767 | if (!may_perform_io) | |
53853e2d | 1768 | return COMPACT_SKIPPED; |
56de7263 | 1769 | |
a5508cd8 | 1770 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 1771 | |
56de7263 | 1772 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1773 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1774 | ac->nodemask) { | |
ea7ab982 | 1775 | enum compact_result status; |
56de7263 | 1776 | |
a8e025e5 VB |
1777 | if (prio > MIN_COMPACT_PRIORITY |
1778 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 1779 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 1780 | continue; |
1d4746d3 | 1781 | } |
53853e2d | 1782 | |
a5508cd8 | 1783 | status = compact_zone_order(zone, order, gfp_mask, prio, |
c3486f53 | 1784 | alloc_flags, ac_classzone_idx(ac)); |
56de7263 MG |
1785 | rc = max(status, rc); |
1786 | ||
7ceb009a VB |
1787 | /* The allocation should succeed, stop compacting */ |
1788 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
1789 | /* |
1790 | * We think the allocation will succeed in this zone, | |
1791 | * but it is not certain, hence the false. The caller | |
1792 | * will repeat this with true if allocation indeed | |
1793 | * succeeds in this zone. | |
1794 | */ | |
1795 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 1796 | |
c3486f53 | 1797 | break; |
1f9efdef VB |
1798 | } |
1799 | ||
a5508cd8 | 1800 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 1801 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
1802 | /* |
1803 | * We think that allocation won't succeed in this zone | |
1804 | * so we defer compaction there. If it ends up | |
1805 | * succeeding after all, it will be reset. | |
1806 | */ | |
1807 | defer_compaction(zone, order); | |
1f9efdef VB |
1808 | |
1809 | /* | |
1810 | * We might have stopped compacting due to need_resched() in | |
1811 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 1812 | * case do not try further zones |
1f9efdef | 1813 | */ |
c3486f53 VB |
1814 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
1815 | || fatal_signal_pending(current)) | |
1816 | break; | |
56de7263 MG |
1817 | } |
1818 | ||
1819 | return rc; | |
1820 | } | |
1821 | ||
1822 | ||
76ab0f53 | 1823 | /* Compact all zones within a node */ |
791cae96 | 1824 | static void compact_node(int nid) |
76ab0f53 | 1825 | { |
791cae96 | 1826 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 1827 | int zoneid; |
76ab0f53 | 1828 | struct zone *zone; |
791cae96 VB |
1829 | struct compact_control cc = { |
1830 | .order = -1, | |
7f354a54 DR |
1831 | .total_migrate_scanned = 0, |
1832 | .total_free_scanned = 0, | |
791cae96 VB |
1833 | .mode = MIGRATE_SYNC, |
1834 | .ignore_skip_hint = true, | |
1835 | .whole_zone = true, | |
73e64c51 | 1836 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
1837 | }; |
1838 | ||
76ab0f53 | 1839 | |
76ab0f53 | 1840 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1841 | |
1842 | zone = &pgdat->node_zones[zoneid]; | |
1843 | if (!populated_zone(zone)) | |
1844 | continue; | |
1845 | ||
791cae96 VB |
1846 | cc.nr_freepages = 0; |
1847 | cc.nr_migratepages = 0; | |
1848 | cc.zone = zone; | |
1849 | INIT_LIST_HEAD(&cc.freepages); | |
1850 | INIT_LIST_HEAD(&cc.migratepages); | |
76ab0f53 | 1851 | |
791cae96 | 1852 | compact_zone(zone, &cc); |
75469345 | 1853 | |
791cae96 VB |
1854 | VM_BUG_ON(!list_empty(&cc.freepages)); |
1855 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 1856 | } |
76ab0f53 MG |
1857 | } |
1858 | ||
1859 | /* Compact all nodes in the system */ | |
7964c06d | 1860 | static void compact_nodes(void) |
76ab0f53 MG |
1861 | { |
1862 | int nid; | |
1863 | ||
8575ec29 HD |
1864 | /* Flush pending updates to the LRU lists */ |
1865 | lru_add_drain_all(); | |
1866 | ||
76ab0f53 MG |
1867 | for_each_online_node(nid) |
1868 | compact_node(nid); | |
76ab0f53 MG |
1869 | } |
1870 | ||
1871 | /* The written value is actually unused, all memory is compacted */ | |
1872 | int sysctl_compact_memory; | |
1873 | ||
fec4eb2c YB |
1874 | /* |
1875 | * This is the entry point for compacting all nodes via | |
1876 | * /proc/sys/vm/compact_memory | |
1877 | */ | |
76ab0f53 MG |
1878 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
1879 | void __user *buffer, size_t *length, loff_t *ppos) | |
1880 | { | |
1881 | if (write) | |
7964c06d | 1882 | compact_nodes(); |
76ab0f53 MG |
1883 | |
1884 | return 0; | |
1885 | } | |
ed4a6d7f | 1886 | |
5e771905 MG |
1887 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
1888 | void __user *buffer, size_t *length, loff_t *ppos) | |
1889 | { | |
1890 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1891 | ||
1892 | return 0; | |
1893 | } | |
1894 | ||
ed4a6d7f | 1895 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
74e77fb9 | 1896 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1897 | struct device_attribute *attr, |
ed4a6d7f MG |
1898 | const char *buf, size_t count) |
1899 | { | |
8575ec29 HD |
1900 | int nid = dev->id; |
1901 | ||
1902 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1903 | /* Flush pending updates to the LRU lists */ | |
1904 | lru_add_drain_all(); | |
1905 | ||
1906 | compact_node(nid); | |
1907 | } | |
ed4a6d7f MG |
1908 | |
1909 | return count; | |
1910 | } | |
10fbcf4c | 1911 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1912 | |
1913 | int compaction_register_node(struct node *node) | |
1914 | { | |
10fbcf4c | 1915 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1916 | } |
1917 | ||
1918 | void compaction_unregister_node(struct node *node) | |
1919 | { | |
10fbcf4c | 1920 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1921 | } |
1922 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 1923 | |
698b1b30 VB |
1924 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
1925 | { | |
172400c6 | 1926 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
1927 | } |
1928 | ||
1929 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
1930 | { | |
1931 | int zoneid; | |
1932 | struct zone *zone; | |
1933 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
1934 | ||
6cd9dc3e | 1935 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
1936 | zone = &pgdat->node_zones[zoneid]; |
1937 | ||
1938 | if (!populated_zone(zone)) | |
1939 | continue; | |
1940 | ||
1941 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
1942 | classzone_idx) == COMPACT_CONTINUE) | |
1943 | return true; | |
1944 | } | |
1945 | ||
1946 | return false; | |
1947 | } | |
1948 | ||
1949 | static void kcompactd_do_work(pg_data_t *pgdat) | |
1950 | { | |
1951 | /* | |
1952 | * With no special task, compact all zones so that a page of requested | |
1953 | * order is allocatable. | |
1954 | */ | |
1955 | int zoneid; | |
1956 | struct zone *zone; | |
1957 | struct compact_control cc = { | |
1958 | .order = pgdat->kcompactd_max_order, | |
7f354a54 DR |
1959 | .total_migrate_scanned = 0, |
1960 | .total_free_scanned = 0, | |
698b1b30 VB |
1961 | .classzone_idx = pgdat->kcompactd_classzone_idx, |
1962 | .mode = MIGRATE_SYNC_LIGHT, | |
a0647dc9 | 1963 | .ignore_skip_hint = false, |
73e64c51 | 1964 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 1965 | }; |
698b1b30 VB |
1966 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
1967 | cc.classzone_idx); | |
7f354a54 | 1968 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 1969 | |
6cd9dc3e | 1970 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
1971 | int status; |
1972 | ||
1973 | zone = &pgdat->node_zones[zoneid]; | |
1974 | if (!populated_zone(zone)) | |
1975 | continue; | |
1976 | ||
1977 | if (compaction_deferred(zone, cc.order)) | |
1978 | continue; | |
1979 | ||
1980 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
1981 | COMPACT_CONTINUE) | |
1982 | continue; | |
1983 | ||
1984 | cc.nr_freepages = 0; | |
1985 | cc.nr_migratepages = 0; | |
7f354a54 DR |
1986 | cc.total_migrate_scanned = 0; |
1987 | cc.total_free_scanned = 0; | |
698b1b30 VB |
1988 | cc.zone = zone; |
1989 | INIT_LIST_HEAD(&cc.freepages); | |
1990 | INIT_LIST_HEAD(&cc.migratepages); | |
1991 | ||
172400c6 VB |
1992 | if (kthread_should_stop()) |
1993 | return; | |
698b1b30 VB |
1994 | status = compact_zone(zone, &cc); |
1995 | ||
7ceb009a | 1996 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 1997 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 1998 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
698b1b30 VB |
1999 | /* |
2000 | * We use sync migration mode here, so we defer like | |
2001 | * sync direct compaction does. | |
2002 | */ | |
2003 | defer_compaction(zone, cc.order); | |
2004 | } | |
2005 | ||
7f354a54 DR |
2006 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2007 | cc.total_migrate_scanned); | |
2008 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2009 | cc.total_free_scanned); | |
2010 | ||
698b1b30 VB |
2011 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2012 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2013 | } | |
2014 | ||
2015 | /* | |
2016 | * Regardless of success, we are done until woken up next. But remember | |
2017 | * the requested order/classzone_idx in case it was higher/tighter than | |
2018 | * our current ones | |
2019 | */ | |
2020 | if (pgdat->kcompactd_max_order <= cc.order) | |
2021 | pgdat->kcompactd_max_order = 0; | |
2022 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
2023 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2024 | } | |
2025 | ||
2026 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
2027 | { | |
2028 | if (!order) | |
2029 | return; | |
2030 | ||
2031 | if (pgdat->kcompactd_max_order < order) | |
2032 | pgdat->kcompactd_max_order = order; | |
2033 | ||
2034 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2035 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2036 | ||
6818600f DB |
2037 | /* |
2038 | * Pairs with implicit barrier in wait_event_freezable() | |
2039 | * such that wakeups are not missed. | |
2040 | */ | |
2041 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2042 | return; |
2043 | ||
2044 | if (!kcompactd_node_suitable(pgdat)) | |
2045 | return; | |
2046 | ||
2047 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2048 | classzone_idx); | |
2049 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2050 | } | |
2051 | ||
2052 | /* | |
2053 | * The background compaction daemon, started as a kernel thread | |
2054 | * from the init process. | |
2055 | */ | |
2056 | static int kcompactd(void *p) | |
2057 | { | |
2058 | pg_data_t *pgdat = (pg_data_t*)p; | |
2059 | struct task_struct *tsk = current; | |
2060 | ||
2061 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2062 | ||
2063 | if (!cpumask_empty(cpumask)) | |
2064 | set_cpus_allowed_ptr(tsk, cpumask); | |
2065 | ||
2066 | set_freezable(); | |
2067 | ||
2068 | pgdat->kcompactd_max_order = 0; | |
2069 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2070 | ||
2071 | while (!kthread_should_stop()) { | |
2072 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); | |
2073 | wait_event_freezable(pgdat->kcompactd_wait, | |
2074 | kcompactd_work_requested(pgdat)); | |
2075 | ||
2076 | kcompactd_do_work(pgdat); | |
2077 | } | |
2078 | ||
2079 | return 0; | |
2080 | } | |
2081 | ||
2082 | /* | |
2083 | * This kcompactd start function will be called by init and node-hot-add. | |
2084 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2085 | */ | |
2086 | int kcompactd_run(int nid) | |
2087 | { | |
2088 | pg_data_t *pgdat = NODE_DATA(nid); | |
2089 | int ret = 0; | |
2090 | ||
2091 | if (pgdat->kcompactd) | |
2092 | return 0; | |
2093 | ||
2094 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2095 | if (IS_ERR(pgdat->kcompactd)) { | |
2096 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2097 | ret = PTR_ERR(pgdat->kcompactd); | |
2098 | pgdat->kcompactd = NULL; | |
2099 | } | |
2100 | return ret; | |
2101 | } | |
2102 | ||
2103 | /* | |
2104 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2105 | * hold mem_hotplug_begin/end(). | |
2106 | */ | |
2107 | void kcompactd_stop(int nid) | |
2108 | { | |
2109 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2110 | ||
2111 | if (kcompactd) { | |
2112 | kthread_stop(kcompactd); | |
2113 | NODE_DATA(nid)->kcompactd = NULL; | |
2114 | } | |
2115 | } | |
2116 | ||
2117 | /* | |
2118 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2119 | * not required for correctness. So if the last cpu in a node goes | |
2120 | * away, we get changed to run anywhere: as the first one comes back, | |
2121 | * restore their cpu bindings. | |
2122 | */ | |
e46b1db2 | 2123 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
2124 | { |
2125 | int nid; | |
2126 | ||
e46b1db2 AMG |
2127 | for_each_node_state(nid, N_MEMORY) { |
2128 | pg_data_t *pgdat = NODE_DATA(nid); | |
2129 | const struct cpumask *mask; | |
698b1b30 | 2130 | |
e46b1db2 | 2131 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 2132 | |
e46b1db2 AMG |
2133 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
2134 | /* One of our CPUs online: restore mask */ | |
2135 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 2136 | } |
e46b1db2 | 2137 | return 0; |
698b1b30 VB |
2138 | } |
2139 | ||
2140 | static int __init kcompactd_init(void) | |
2141 | { | |
2142 | int nid; | |
e46b1db2 AMG |
2143 | int ret; |
2144 | ||
2145 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2146 | "mm/compaction:online", | |
2147 | kcompactd_cpu_online, NULL); | |
2148 | if (ret < 0) { | |
2149 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2150 | return ret; | |
2151 | } | |
698b1b30 VB |
2152 | |
2153 | for_each_node_state(nid, N_MEMORY) | |
2154 | kcompactd_run(nid); | |
698b1b30 VB |
2155 | return 0; |
2156 | } | |
2157 | subsys_initcall(kcompactd_init) | |
2158 | ||
ff9543fd | 2159 | #endif /* CONFIG_COMPACTION */ |