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> |
eb414681 | 25 | #include <linux/psi.h> |
748446bb MG |
26 | #include "internal.h" |
27 | ||
010fc29a MK |
28 | #ifdef CONFIG_COMPACTION |
29 | static inline void count_compact_event(enum vm_event_item item) | |
30 | { | |
31 | count_vm_event(item); | |
32 | } | |
33 | ||
34 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
35 | { | |
36 | count_vm_events(item, delta); | |
37 | } | |
38 | #else | |
39 | #define count_compact_event(item) do { } while (0) | |
40 | #define count_compact_events(item, delta) do { } while (0) | |
41 | #endif | |
42 | ||
ff9543fd MN |
43 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
44 | ||
b7aba698 MG |
45 | #define CREATE_TRACE_POINTS |
46 | #include <trace/events/compaction.h> | |
47 | ||
06b6640a VB |
48 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
49 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
50 | #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order) | |
51 | #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order) | |
52 | ||
748446bb MG |
53 | static unsigned long release_freepages(struct list_head *freelist) |
54 | { | |
55 | struct page *page, *next; | |
6bace090 | 56 | unsigned long high_pfn = 0; |
748446bb MG |
57 | |
58 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 59 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
60 | list_del(&page->lru); |
61 | __free_page(page); | |
6bace090 VB |
62 | if (pfn > high_pfn) |
63 | high_pfn = pfn; | |
748446bb MG |
64 | } |
65 | ||
6bace090 | 66 | return high_pfn; |
748446bb MG |
67 | } |
68 | ||
4469ab98 | 69 | static void split_map_pages(struct list_head *list) |
ff9543fd | 70 | { |
66c64223 JK |
71 | unsigned int i, order, nr_pages; |
72 | struct page *page, *next; | |
73 | LIST_HEAD(tmp_list); | |
74 | ||
75 | list_for_each_entry_safe(page, next, list, lru) { | |
76 | list_del(&page->lru); | |
77 | ||
78 | order = page_private(page); | |
79 | nr_pages = 1 << order; | |
66c64223 | 80 | |
46f24fd8 | 81 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
82 | if (order) |
83 | split_page(page, order); | |
ff9543fd | 84 | |
66c64223 JK |
85 | for (i = 0; i < nr_pages; i++) { |
86 | list_add(&page->lru, &tmp_list); | |
87 | page++; | |
88 | } | |
ff9543fd | 89 | } |
66c64223 JK |
90 | |
91 | list_splice(&tmp_list, list); | |
ff9543fd MN |
92 | } |
93 | ||
bb13ffeb | 94 | #ifdef CONFIG_COMPACTION |
24e2716f | 95 | |
bda807d4 MK |
96 | int PageMovable(struct page *page) |
97 | { | |
98 | struct address_space *mapping; | |
99 | ||
100 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
101 | if (!__PageMovable(page)) | |
102 | return 0; | |
103 | ||
104 | mapping = page_mapping(page); | |
105 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
106 | return 1; | |
107 | ||
108 | return 0; | |
109 | } | |
110 | EXPORT_SYMBOL(PageMovable); | |
111 | ||
112 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
113 | { | |
114 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
115 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
116 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
117 | } | |
118 | EXPORT_SYMBOL(__SetPageMovable); | |
119 | ||
120 | void __ClearPageMovable(struct page *page) | |
121 | { | |
122 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
123 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
124 | /* | |
125 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
126 | * flag so that VM can catch up released page by driver after isolation. | |
127 | * With it, VM migration doesn't try to put it back. | |
128 | */ | |
129 | page->mapping = (void *)((unsigned long)page->mapping & | |
130 | PAGE_MAPPING_MOVABLE); | |
131 | } | |
132 | EXPORT_SYMBOL(__ClearPageMovable); | |
133 | ||
24e2716f JK |
134 | /* Do not skip compaction more than 64 times */ |
135 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
136 | ||
137 | /* | |
138 | * Compaction is deferred when compaction fails to result in a page | |
139 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
140 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
141 | */ | |
142 | void defer_compaction(struct zone *zone, int order) | |
143 | { | |
144 | zone->compact_considered = 0; | |
145 | zone->compact_defer_shift++; | |
146 | ||
147 | if (order < zone->compact_order_failed) | |
148 | zone->compact_order_failed = order; | |
149 | ||
150 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
151 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
152 | ||
153 | trace_mm_compaction_defer_compaction(zone, order); | |
154 | } | |
155 | ||
156 | /* Returns true if compaction should be skipped this time */ | |
157 | bool compaction_deferred(struct zone *zone, int order) | |
158 | { | |
159 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
160 | ||
161 | if (order < zone->compact_order_failed) | |
162 | return false; | |
163 | ||
164 | /* Avoid possible overflow */ | |
165 | if (++zone->compact_considered > defer_limit) | |
166 | zone->compact_considered = defer_limit; | |
167 | ||
168 | if (zone->compact_considered >= defer_limit) | |
169 | return false; | |
170 | ||
171 | trace_mm_compaction_deferred(zone, order); | |
172 | ||
173 | return true; | |
174 | } | |
175 | ||
176 | /* | |
177 | * Update defer tracking counters after successful compaction of given order, | |
178 | * which means an allocation either succeeded (alloc_success == true) or is | |
179 | * expected to succeed. | |
180 | */ | |
181 | void compaction_defer_reset(struct zone *zone, int order, | |
182 | bool alloc_success) | |
183 | { | |
184 | if (alloc_success) { | |
185 | zone->compact_considered = 0; | |
186 | zone->compact_defer_shift = 0; | |
187 | } | |
188 | if (order >= zone->compact_order_failed) | |
189 | zone->compact_order_failed = order + 1; | |
190 | ||
191 | trace_mm_compaction_defer_reset(zone, order); | |
192 | } | |
193 | ||
194 | /* Returns true if restarting compaction after many failures */ | |
195 | bool compaction_restarting(struct zone *zone, int order) | |
196 | { | |
197 | if (order < zone->compact_order_failed) | |
198 | return false; | |
199 | ||
200 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
201 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
202 | } | |
203 | ||
bb13ffeb MG |
204 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
205 | static inline bool isolation_suitable(struct compact_control *cc, | |
206 | struct page *page) | |
207 | { | |
208 | if (cc->ignore_skip_hint) | |
209 | return true; | |
210 | ||
211 | return !get_pageblock_skip(page); | |
212 | } | |
213 | ||
02333641 VB |
214 | static void reset_cached_positions(struct zone *zone) |
215 | { | |
216 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
217 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 218 | zone->compact_cached_free_pfn = |
06b6640a | 219 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
220 | } |
221 | ||
21dc7e02 | 222 | /* |
b527cfe5 VB |
223 | * Compound pages of >= pageblock_order should consistenly be skipped until |
224 | * released. It is always pointless to compact pages of such order (if they are | |
225 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 226 | */ |
b527cfe5 | 227 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 228 | { |
b527cfe5 | 229 | if (!PageCompound(page)) |
21dc7e02 | 230 | return false; |
b527cfe5 VB |
231 | |
232 | page = compound_head(page); | |
233 | ||
234 | if (compound_order(page) >= pageblock_order) | |
235 | return true; | |
236 | ||
237 | return false; | |
21dc7e02 DR |
238 | } |
239 | ||
bb13ffeb MG |
240 | /* |
241 | * This function is called to clear all cached information on pageblocks that | |
242 | * should be skipped for page isolation when the migrate and free page scanner | |
243 | * meet. | |
244 | */ | |
62997027 | 245 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
246 | { |
247 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 248 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
249 | unsigned long pfn; |
250 | ||
62997027 | 251 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
252 | |
253 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
254 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
255 | struct page *page; | |
256 | ||
257 | cond_resched(); | |
258 | ||
ccbe1e4d MH |
259 | page = pfn_to_online_page(pfn); |
260 | if (!page) | |
bb13ffeb | 261 | continue; |
bb13ffeb MG |
262 | if (zone != page_zone(page)) |
263 | continue; | |
b527cfe5 | 264 | if (pageblock_skip_persistent(page)) |
21dc7e02 | 265 | continue; |
bb13ffeb MG |
266 | |
267 | clear_pageblock_skip(page); | |
268 | } | |
02333641 VB |
269 | |
270 | reset_cached_positions(zone); | |
bb13ffeb MG |
271 | } |
272 | ||
62997027 MG |
273 | void reset_isolation_suitable(pg_data_t *pgdat) |
274 | { | |
275 | int zoneid; | |
276 | ||
277 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
278 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
279 | if (!populated_zone(zone)) | |
280 | continue; | |
281 | ||
282 | /* Only flush if a full compaction finished recently */ | |
283 | if (zone->compact_blockskip_flush) | |
284 | __reset_isolation_suitable(zone); | |
285 | } | |
286 | } | |
287 | ||
bb13ffeb MG |
288 | /* |
289 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 290 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 291 | */ |
c89511ab MG |
292 | static void update_pageblock_skip(struct compact_control *cc, |
293 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 294 | bool migrate_scanner) |
bb13ffeb | 295 | { |
c89511ab | 296 | struct zone *zone = cc->zone; |
35979ef3 | 297 | unsigned long pfn; |
6815bf3f | 298 | |
2583d671 | 299 | if (cc->no_set_skip_hint) |
6815bf3f JK |
300 | return; |
301 | ||
bb13ffeb MG |
302 | if (!page) |
303 | return; | |
304 | ||
35979ef3 DR |
305 | if (nr_isolated) |
306 | return; | |
307 | ||
edc2ca61 | 308 | set_pageblock_skip(page); |
c89511ab | 309 | |
35979ef3 DR |
310 | pfn = page_to_pfn(page); |
311 | ||
312 | /* Update where async and sync compaction should restart */ | |
313 | if (migrate_scanner) { | |
35979ef3 DR |
314 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
315 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
316 | if (cc->mode != MIGRATE_ASYNC && |
317 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
318 | zone->compact_cached_migrate_pfn[1] = pfn; |
319 | } else { | |
35979ef3 DR |
320 | if (pfn < zone->compact_cached_free_pfn) |
321 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 322 | } |
bb13ffeb MG |
323 | } |
324 | #else | |
325 | static inline bool isolation_suitable(struct compact_control *cc, | |
326 | struct page *page) | |
327 | { | |
328 | return true; | |
329 | } | |
330 | ||
b527cfe5 | 331 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
332 | { |
333 | return false; | |
334 | } | |
335 | ||
336 | static inline void update_pageblock_skip(struct compact_control *cc, | |
c89511ab | 337 | struct page *page, unsigned long nr_isolated, |
edc2ca61 | 338 | bool migrate_scanner) |
bb13ffeb MG |
339 | { |
340 | } | |
341 | #endif /* CONFIG_COMPACTION */ | |
342 | ||
8b44d279 VB |
343 | /* |
344 | * Compaction requires the taking of some coarse locks that are potentially | |
345 | * very heavily contended. For async compaction, back out if the lock cannot | |
346 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
347 | * | |
348 | * Returns true if the lock is held | |
349 | * Returns false if the lock is not held and compaction should abort | |
350 | */ | |
351 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
352 | struct compact_control *cc) | |
2a1402aa | 353 | { |
8b44d279 VB |
354 | if (cc->mode == MIGRATE_ASYNC) { |
355 | if (!spin_trylock_irqsave(lock, *flags)) { | |
c3486f53 | 356 | cc->contended = true; |
8b44d279 VB |
357 | return false; |
358 | } | |
359 | } else { | |
360 | spin_lock_irqsave(lock, *flags); | |
361 | } | |
1f9efdef | 362 | |
8b44d279 | 363 | return true; |
2a1402aa MG |
364 | } |
365 | ||
c67fe375 MG |
366 | /* |
367 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
368 | * very heavily contended. The lock should be periodically unlocked to avoid |
369 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
370 | * the lock. It might also be that allowing the IRQs will result in | |
371 | * need_resched() becoming true. If scheduling is needed, async compaction | |
372 | * aborts. Sync compaction schedules. | |
373 | * Either compaction type will also abort if a fatal signal is pending. | |
374 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 375 | * |
8b44d279 VB |
376 | * Returns true if compaction should abort due to fatal signal pending, or |
377 | * async compaction due to need_resched() | |
378 | * Returns false when compaction can continue (sync compaction might have | |
379 | * scheduled) | |
c67fe375 | 380 | */ |
8b44d279 VB |
381 | static bool compact_unlock_should_abort(spinlock_t *lock, |
382 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 383 | { |
8b44d279 VB |
384 | if (*locked) { |
385 | spin_unlock_irqrestore(lock, flags); | |
386 | *locked = false; | |
387 | } | |
1f9efdef | 388 | |
8b44d279 | 389 | if (fatal_signal_pending(current)) { |
c3486f53 | 390 | cc->contended = true; |
8b44d279 VB |
391 | return true; |
392 | } | |
c67fe375 | 393 | |
8b44d279 | 394 | if (need_resched()) { |
e0b9daeb | 395 | if (cc->mode == MIGRATE_ASYNC) { |
c3486f53 | 396 | cc->contended = true; |
8b44d279 | 397 | return true; |
c67fe375 | 398 | } |
c67fe375 | 399 | cond_resched(); |
c67fe375 MG |
400 | } |
401 | ||
8b44d279 | 402 | return false; |
c67fe375 MG |
403 | } |
404 | ||
be976572 VB |
405 | /* |
406 | * Aside from avoiding lock contention, compaction also periodically checks | |
407 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 408 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
409 | * is used where no lock is concerned. |
410 | * | |
411 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
412 | * Returns true when async compaction should abort. | |
413 | */ | |
414 | static inline bool compact_should_abort(struct compact_control *cc) | |
415 | { | |
416 | /* async compaction aborts if contended */ | |
417 | if (need_resched()) { | |
418 | if (cc->mode == MIGRATE_ASYNC) { | |
c3486f53 | 419 | cc->contended = true; |
be976572 VB |
420 | return true; |
421 | } | |
422 | ||
423 | cond_resched(); | |
424 | } | |
425 | ||
426 | return false; | |
427 | } | |
428 | ||
85aa125f | 429 | /* |
9e4be470 JM |
430 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
431 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
432 | * (even though it may still end up isolating some pages). | |
85aa125f | 433 | */ |
f40d1e42 | 434 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 435 | unsigned long *start_pfn, |
85aa125f MN |
436 | unsigned long end_pfn, |
437 | struct list_head *freelist, | |
438 | bool strict) | |
748446bb | 439 | { |
b7aba698 | 440 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 441 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 442 | unsigned long flags = 0; |
f40d1e42 | 443 | bool locked = false; |
e14c720e | 444 | unsigned long blockpfn = *start_pfn; |
66c64223 | 445 | unsigned int order; |
748446bb | 446 | |
748446bb MG |
447 | cursor = pfn_to_page(blockpfn); |
448 | ||
f40d1e42 | 449 | /* Isolate free pages. */ |
748446bb | 450 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
66c64223 | 451 | int isolated; |
748446bb MG |
452 | struct page *page = cursor; |
453 | ||
8b44d279 VB |
454 | /* |
455 | * Periodically drop the lock (if held) regardless of its | |
456 | * contention, to give chance to IRQs. Abort if fatal signal | |
457 | * pending or async compaction detects need_resched() | |
458 | */ | |
459 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
460 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
461 | &locked, cc)) | |
462 | break; | |
463 | ||
b7aba698 | 464 | nr_scanned++; |
f40d1e42 | 465 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
466 | goto isolate_fail; |
467 | ||
bb13ffeb MG |
468 | if (!valid_page) |
469 | valid_page = page; | |
9fcd6d2e VB |
470 | |
471 | /* | |
472 | * For compound pages such as THP and hugetlbfs, we can save | |
473 | * potentially a lot of iterations if we skip them at once. | |
474 | * The check is racy, but we can consider only valid values | |
475 | * and the only danger is skipping too much. | |
476 | */ | |
477 | if (PageCompound(page)) { | |
21dc7e02 DR |
478 | const unsigned int order = compound_order(page); |
479 | ||
d3c85bad | 480 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
481 | blockpfn += (1UL << order) - 1; |
482 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 483 | } |
9fcd6d2e VB |
484 | goto isolate_fail; |
485 | } | |
486 | ||
f40d1e42 | 487 | if (!PageBuddy(page)) |
2af120bc | 488 | goto isolate_fail; |
f40d1e42 MG |
489 | |
490 | /* | |
69b7189f VB |
491 | * If we already hold the lock, we can skip some rechecking. |
492 | * Note that if we hold the lock now, checked_pageblock was | |
493 | * already set in some previous iteration (or strict is true), | |
494 | * so it is correct to skip the suitable migration target | |
495 | * recheck as well. | |
f40d1e42 | 496 | */ |
69b7189f VB |
497 | if (!locked) { |
498 | /* | |
499 | * The zone lock must be held to isolate freepages. | |
500 | * Unfortunately this is a very coarse lock and can be | |
501 | * heavily contended if there are parallel allocations | |
502 | * or parallel compactions. For async compaction do not | |
503 | * spin on the lock and we acquire the lock as late as | |
504 | * possible. | |
505 | */ | |
8b44d279 VB |
506 | locked = compact_trylock_irqsave(&cc->zone->lock, |
507 | &flags, cc); | |
69b7189f VB |
508 | if (!locked) |
509 | break; | |
f40d1e42 | 510 | |
69b7189f VB |
511 | /* Recheck this is a buddy page under lock */ |
512 | if (!PageBuddy(page)) | |
513 | goto isolate_fail; | |
514 | } | |
748446bb | 515 | |
66c64223 JK |
516 | /* Found a free page, will break it into order-0 pages */ |
517 | order = page_order(page); | |
518 | isolated = __isolate_free_page(page, order); | |
a4f04f2c DR |
519 | if (!isolated) |
520 | break; | |
66c64223 | 521 | set_page_private(page, order); |
a4f04f2c | 522 | |
748446bb | 523 | total_isolated += isolated; |
a4f04f2c | 524 | cc->nr_freepages += isolated; |
66c64223 JK |
525 | list_add_tail(&page->lru, freelist); |
526 | ||
a4f04f2c DR |
527 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
528 | blockpfn += isolated; | |
529 | break; | |
748446bb | 530 | } |
a4f04f2c DR |
531 | /* Advance to the end of split page */ |
532 | blockpfn += isolated - 1; | |
533 | cursor += isolated - 1; | |
534 | continue; | |
2af120bc LA |
535 | |
536 | isolate_fail: | |
537 | if (strict) | |
538 | break; | |
539 | else | |
540 | continue; | |
541 | ||
748446bb MG |
542 | } |
543 | ||
a4f04f2c DR |
544 | if (locked) |
545 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
546 | ||
9fcd6d2e VB |
547 | /* |
548 | * There is a tiny chance that we have read bogus compound_order(), | |
549 | * so be careful to not go outside of the pageblock. | |
550 | */ | |
551 | if (unlikely(blockpfn > end_pfn)) | |
552 | blockpfn = end_pfn; | |
553 | ||
e34d85f0 JK |
554 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
555 | nr_scanned, total_isolated); | |
556 | ||
e14c720e VB |
557 | /* Record how far we have got within the block */ |
558 | *start_pfn = blockpfn; | |
559 | ||
f40d1e42 MG |
560 | /* |
561 | * If strict isolation is requested by CMA then check that all the | |
562 | * pages requested were isolated. If there were any failures, 0 is | |
563 | * returned and CMA will fail. | |
564 | */ | |
2af120bc | 565 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
566 | total_isolated = 0; |
567 | ||
bb13ffeb MG |
568 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
569 | if (blockpfn == end_pfn) | |
edc2ca61 | 570 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 571 | |
7f354a54 | 572 | cc->total_free_scanned += nr_scanned; |
397487db | 573 | if (total_isolated) |
010fc29a | 574 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
575 | return total_isolated; |
576 | } | |
577 | ||
85aa125f MN |
578 | /** |
579 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 580 | * @cc: Compaction control structure. |
85aa125f MN |
581 | * @start_pfn: The first PFN to start isolating. |
582 | * @end_pfn: The one-past-last PFN. | |
583 | * | |
584 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
585 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
586 | * undo its actions and return zero. | |
587 | * | |
588 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
589 | * (which may be greater then end_pfn if end fell in a middle of | |
590 | * a free page). | |
591 | */ | |
ff9543fd | 592 | unsigned long |
bb13ffeb MG |
593 | isolate_freepages_range(struct compact_control *cc, |
594 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 595 | { |
e1409c32 | 596 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
597 | LIST_HEAD(freelist); |
598 | ||
7d49d886 | 599 | pfn = start_pfn; |
06b6640a | 600 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
601 | if (block_start_pfn < cc->zone->zone_start_pfn) |
602 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 603 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
604 | |
605 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 606 | block_start_pfn = block_end_pfn, |
7d49d886 | 607 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
608 | /* Protect pfn from changing by isolate_freepages_block */ |
609 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 610 | |
85aa125f MN |
611 | block_end_pfn = min(block_end_pfn, end_pfn); |
612 | ||
58420016 JK |
613 | /* |
614 | * pfn could pass the block_end_pfn if isolated freepage | |
615 | * is more than pageblock order. In this case, we adjust | |
616 | * scanning range to right one. | |
617 | */ | |
618 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
619 | block_start_pfn = pageblock_start_pfn(pfn); |
620 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
621 | block_end_pfn = min(block_end_pfn, end_pfn); |
622 | } | |
623 | ||
e1409c32 JK |
624 | if (!pageblock_pfn_to_page(block_start_pfn, |
625 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
626 | break; |
627 | ||
e14c720e VB |
628 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
629 | block_end_pfn, &freelist, true); | |
85aa125f MN |
630 | |
631 | /* | |
632 | * In strict mode, isolate_freepages_block() returns 0 if | |
633 | * there are any holes in the block (ie. invalid PFNs or | |
634 | * non-free pages). | |
635 | */ | |
636 | if (!isolated) | |
637 | break; | |
638 | ||
639 | /* | |
640 | * If we managed to isolate pages, it is always (1 << n) * | |
641 | * pageblock_nr_pages for some non-negative n. (Max order | |
642 | * page may span two pageblocks). | |
643 | */ | |
644 | } | |
645 | ||
66c64223 | 646 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 647 | split_map_pages(&freelist); |
85aa125f MN |
648 | |
649 | if (pfn < end_pfn) { | |
650 | /* Loop terminated early, cleanup. */ | |
651 | release_freepages(&freelist); | |
652 | return 0; | |
653 | } | |
654 | ||
655 | /* We don't use freelists for anything. */ | |
656 | return pfn; | |
657 | } | |
658 | ||
748446bb MG |
659 | /* Similar to reclaim, but different enough that they don't share logic */ |
660 | static bool too_many_isolated(struct zone *zone) | |
661 | { | |
bc693045 | 662 | unsigned long active, inactive, isolated; |
748446bb | 663 | |
599d0c95 MG |
664 | inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) + |
665 | node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON); | |
666 | active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) + | |
667 | node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON); | |
668 | isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) + | |
669 | node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON); | |
748446bb | 670 | |
bc693045 | 671 | return isolated > (inactive + active) / 2; |
748446bb MG |
672 | } |
673 | ||
2fe86e00 | 674 | /** |
edc2ca61 VB |
675 | * isolate_migratepages_block() - isolate all migrate-able pages within |
676 | * a single pageblock | |
2fe86e00 | 677 | * @cc: Compaction control structure. |
edc2ca61 VB |
678 | * @low_pfn: The first PFN to isolate |
679 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
680 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
681 | * |
682 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
683 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
684 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
685 | * first page that was not scanned (which may be both less, equal to or more | |
686 | * than end_pfn). | |
2fe86e00 | 687 | * |
edc2ca61 VB |
688 | * The pages are isolated on cc->migratepages list (not required to be empty), |
689 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
690 | * is neither read nor updated. | |
748446bb | 691 | */ |
edc2ca61 VB |
692 | static unsigned long |
693 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
694 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 695 | { |
edc2ca61 | 696 | struct zone *zone = cc->zone; |
b7aba698 | 697 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 698 | struct lruvec *lruvec; |
b8b2d825 | 699 | unsigned long flags = 0; |
2a1402aa | 700 | bool locked = false; |
bb13ffeb | 701 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 702 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
703 | bool skip_on_failure = false; |
704 | unsigned long next_skip_pfn = 0; | |
748446bb | 705 | |
748446bb MG |
706 | /* |
707 | * Ensure that there are not too many pages isolated from the LRU | |
708 | * list by either parallel reclaimers or compaction. If there are, | |
709 | * delay for some time until fewer pages are isolated | |
710 | */ | |
711 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 712 | /* async migration should just abort */ |
e0b9daeb | 713 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 714 | return 0; |
f9e35b3b | 715 | |
748446bb MG |
716 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
717 | ||
718 | if (fatal_signal_pending(current)) | |
2fe86e00 | 719 | return 0; |
748446bb MG |
720 | } |
721 | ||
be976572 VB |
722 | if (compact_should_abort(cc)) |
723 | return 0; | |
aeef4b83 | 724 | |
fdd048e1 VB |
725 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
726 | skip_on_failure = true; | |
727 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
728 | } | |
729 | ||
748446bb | 730 | /* Time to isolate some pages for migration */ |
748446bb | 731 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 732 | |
fdd048e1 VB |
733 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
734 | /* | |
735 | * We have isolated all migration candidates in the | |
736 | * previous order-aligned block, and did not skip it due | |
737 | * to failure. We should migrate the pages now and | |
738 | * hopefully succeed compaction. | |
739 | */ | |
740 | if (nr_isolated) | |
741 | break; | |
742 | ||
743 | /* | |
744 | * We failed to isolate in the previous order-aligned | |
745 | * block. Set the new boundary to the end of the | |
746 | * current block. Note we can't simply increase | |
747 | * next_skip_pfn by 1 << order, as low_pfn might have | |
748 | * been incremented by a higher number due to skipping | |
749 | * a compound or a high-order buddy page in the | |
750 | * previous loop iteration. | |
751 | */ | |
752 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
753 | } | |
754 | ||
8b44d279 VB |
755 | /* |
756 | * Periodically drop the lock (if held) regardless of its | |
757 | * contention, to give chance to IRQs. Abort async compaction | |
758 | * if contended. | |
759 | */ | |
760 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
a52633d8 | 761 | && compact_unlock_should_abort(zone_lru_lock(zone), flags, |
8b44d279 VB |
762 | &locked, cc)) |
763 | break; | |
c67fe375 | 764 | |
748446bb | 765 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 766 | goto isolate_fail; |
b7aba698 | 767 | nr_scanned++; |
748446bb | 768 | |
748446bb | 769 | page = pfn_to_page(low_pfn); |
dc908600 | 770 | |
bb13ffeb MG |
771 | if (!valid_page) |
772 | valid_page = page; | |
773 | ||
6c14466c | 774 | /* |
99c0fd5e VB |
775 | * Skip if free. We read page order here without zone lock |
776 | * which is generally unsafe, but the race window is small and | |
777 | * the worst thing that can happen is that we skip some | |
778 | * potential isolation targets. | |
6c14466c | 779 | */ |
99c0fd5e VB |
780 | if (PageBuddy(page)) { |
781 | unsigned long freepage_order = page_order_unsafe(page); | |
782 | ||
783 | /* | |
784 | * Without lock, we cannot be sure that what we got is | |
785 | * a valid page order. Consider only values in the | |
786 | * valid order range to prevent low_pfn overflow. | |
787 | */ | |
788 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
789 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 790 | continue; |
99c0fd5e | 791 | } |
748446bb | 792 | |
bc835011 | 793 | /* |
29c0dde8 VB |
794 | * Regardless of being on LRU, compound pages such as THP and |
795 | * hugetlbfs are not to be compacted. We can potentially save | |
796 | * a lot of iterations if we skip them at once. The check is | |
797 | * racy, but we can consider only valid values and the only | |
798 | * danger is skipping too much. | |
bc835011 | 799 | */ |
29c0dde8 | 800 | if (PageCompound(page)) { |
21dc7e02 | 801 | const unsigned int order = compound_order(page); |
edc2ca61 | 802 | |
d3c85bad | 803 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 804 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 805 | goto isolate_fail; |
2a1402aa MG |
806 | } |
807 | ||
bda807d4 MK |
808 | /* |
809 | * Check may be lockless but that's ok as we recheck later. | |
810 | * It's possible to migrate LRU and non-lru movable pages. | |
811 | * Skip any other type of page | |
812 | */ | |
813 | if (!PageLRU(page)) { | |
bda807d4 MK |
814 | /* |
815 | * __PageMovable can return false positive so we need | |
816 | * to verify it under page_lock. | |
817 | */ | |
818 | if (unlikely(__PageMovable(page)) && | |
819 | !PageIsolated(page)) { | |
820 | if (locked) { | |
a52633d8 | 821 | spin_unlock_irqrestore(zone_lru_lock(zone), |
bda807d4 MK |
822 | flags); |
823 | locked = false; | |
824 | } | |
825 | ||
9e5bcd61 | 826 | if (!isolate_movable_page(page, isolate_mode)) |
bda807d4 MK |
827 | goto isolate_success; |
828 | } | |
829 | ||
fdd048e1 | 830 | goto isolate_fail; |
bda807d4 | 831 | } |
29c0dde8 | 832 | |
119d6d59 DR |
833 | /* |
834 | * Migration will fail if an anonymous page is pinned in memory, | |
835 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
836 | * admittedly racy check. | |
837 | */ | |
838 | if (!page_mapping(page) && | |
839 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 840 | goto isolate_fail; |
119d6d59 | 841 | |
73e64c51 MH |
842 | /* |
843 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
844 | * because those do not depend on fs locks. | |
845 | */ | |
846 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
847 | goto isolate_fail; | |
848 | ||
69b7189f VB |
849 | /* If we already hold the lock, we can skip some rechecking */ |
850 | if (!locked) { | |
a52633d8 | 851 | locked = compact_trylock_irqsave(zone_lru_lock(zone), |
8b44d279 | 852 | &flags, cc); |
69b7189f VB |
853 | if (!locked) |
854 | break; | |
2a1402aa | 855 | |
29c0dde8 | 856 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 857 | if (!PageLRU(page)) |
fdd048e1 | 858 | goto isolate_fail; |
29c0dde8 VB |
859 | |
860 | /* | |
861 | * Page become compound since the non-locked check, | |
862 | * and it's on LRU. It can only be a THP so the order | |
863 | * is safe to read and it's 0 for tail pages. | |
864 | */ | |
865 | if (unlikely(PageCompound(page))) { | |
d3c85bad | 866 | low_pfn += (1UL << compound_order(page)) - 1; |
fdd048e1 | 867 | goto isolate_fail; |
69b7189f | 868 | } |
bc835011 AA |
869 | } |
870 | ||
599d0c95 | 871 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
fa9add64 | 872 | |
748446bb | 873 | /* Try isolate the page */ |
edc2ca61 | 874 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 875 | goto isolate_fail; |
748446bb | 876 | |
29c0dde8 | 877 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 878 | |
748446bb | 879 | /* Successfully isolated */ |
fa9add64 | 880 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
6afcf8ef ML |
881 | inc_node_page_state(page, |
882 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
b6c75016 JK |
883 | |
884 | isolate_success: | |
fdd048e1 | 885 | list_add(&page->lru, &cc->migratepages); |
748446bb | 886 | cc->nr_migratepages++; |
b7aba698 | 887 | nr_isolated++; |
748446bb MG |
888 | |
889 | /* Avoid isolating too much */ | |
31b8384a HD |
890 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
891 | ++low_pfn; | |
748446bb | 892 | break; |
31b8384a | 893 | } |
fdd048e1 VB |
894 | |
895 | continue; | |
896 | isolate_fail: | |
897 | if (!skip_on_failure) | |
898 | continue; | |
899 | ||
900 | /* | |
901 | * We have isolated some pages, but then failed. Release them | |
902 | * instead of migrating, as we cannot form the cc->order buddy | |
903 | * page anyway. | |
904 | */ | |
905 | if (nr_isolated) { | |
906 | if (locked) { | |
a52633d8 | 907 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
fdd048e1 VB |
908 | locked = false; |
909 | } | |
fdd048e1 VB |
910 | putback_movable_pages(&cc->migratepages); |
911 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
912 | nr_isolated = 0; |
913 | } | |
914 | ||
915 | if (low_pfn < next_skip_pfn) { | |
916 | low_pfn = next_skip_pfn - 1; | |
917 | /* | |
918 | * The check near the loop beginning would have updated | |
919 | * next_skip_pfn too, but this is a bit simpler. | |
920 | */ | |
921 | next_skip_pfn += 1UL << cc->order; | |
922 | } | |
748446bb MG |
923 | } |
924 | ||
99c0fd5e VB |
925 | /* |
926 | * The PageBuddy() check could have potentially brought us outside | |
927 | * the range to be scanned. | |
928 | */ | |
929 | if (unlikely(low_pfn > end_pfn)) | |
930 | low_pfn = end_pfn; | |
931 | ||
c67fe375 | 932 | if (locked) |
a52633d8 | 933 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
748446bb | 934 | |
50b5b094 VB |
935 | /* |
936 | * Update the pageblock-skip information and cached scanner pfn, | |
937 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 938 | */ |
35979ef3 | 939 | if (low_pfn == end_pfn) |
edc2ca61 | 940 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 941 | |
e34d85f0 JK |
942 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
943 | nr_scanned, nr_isolated); | |
b7aba698 | 944 | |
7f354a54 | 945 | cc->total_migrate_scanned += nr_scanned; |
397487db | 946 | if (nr_isolated) |
010fc29a | 947 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 948 | |
2fe86e00 MN |
949 | return low_pfn; |
950 | } | |
951 | ||
edc2ca61 VB |
952 | /** |
953 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
954 | * @cc: Compaction control structure. | |
955 | * @start_pfn: The first PFN to start isolating. | |
956 | * @end_pfn: The one-past-last PFN. | |
957 | * | |
958 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
959 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
960 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
961 | */ | |
962 | unsigned long | |
963 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
964 | unsigned long end_pfn) | |
965 | { | |
e1409c32 | 966 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
967 | |
968 | /* Scan block by block. First and last block may be incomplete */ | |
969 | pfn = start_pfn; | |
06b6640a | 970 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
971 | if (block_start_pfn < cc->zone->zone_start_pfn) |
972 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 973 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
974 | |
975 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 976 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
977 | block_end_pfn += pageblock_nr_pages) { |
978 | ||
979 | block_end_pfn = min(block_end_pfn, end_pfn); | |
980 | ||
e1409c32 JK |
981 | if (!pageblock_pfn_to_page(block_start_pfn, |
982 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
983 | continue; |
984 | ||
985 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
986 | ISOLATE_UNEVICTABLE); | |
987 | ||
14af4a5e | 988 | if (!pfn) |
edc2ca61 | 989 | break; |
6ea41c0c JK |
990 | |
991 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
992 | break; | |
edc2ca61 | 993 | } |
edc2ca61 VB |
994 | |
995 | return pfn; | |
996 | } | |
997 | ||
ff9543fd MN |
998 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
999 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1000 | |
b682debd VB |
1001 | static bool suitable_migration_source(struct compact_control *cc, |
1002 | struct page *page) | |
1003 | { | |
282722b0 VB |
1004 | int block_mt; |
1005 | ||
1006 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) | |
b682debd VB |
1007 | return true; |
1008 | ||
282722b0 VB |
1009 | block_mt = get_pageblock_migratetype(page); |
1010 | ||
1011 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1012 | return is_migrate_movable(block_mt); | |
1013 | else | |
1014 | return block_mt == cc->migratetype; | |
b682debd VB |
1015 | } |
1016 | ||
018e9a49 | 1017 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1018 | static bool suitable_migration_target(struct compact_control *cc, |
1019 | struct page *page) | |
018e9a49 AM |
1020 | { |
1021 | /* If the page is a large free page, then disallow migration */ | |
1022 | if (PageBuddy(page)) { | |
1023 | /* | |
1024 | * We are checking page_order without zone->lock taken. But | |
1025 | * the only small danger is that we skip a potentially suitable | |
1026 | * pageblock, so it's not worth to check order for valid range. | |
1027 | */ | |
1028 | if (page_order_unsafe(page) >= pageblock_order) | |
1029 | return false; | |
1030 | } | |
1031 | ||
1ef36db2 YX |
1032 | if (cc->ignore_block_suitable) |
1033 | return true; | |
1034 | ||
018e9a49 | 1035 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1036 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1037 | return true; |
1038 | ||
1039 | /* Otherwise skip the block */ | |
1040 | return false; | |
1041 | } | |
1042 | ||
f2849aa0 VB |
1043 | /* |
1044 | * Test whether the free scanner has reached the same or lower pageblock than | |
1045 | * the migration scanner, and compaction should thus terminate. | |
1046 | */ | |
1047 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1048 | { | |
1049 | return (cc->free_pfn >> pageblock_order) | |
1050 | <= (cc->migrate_pfn >> pageblock_order); | |
1051 | } | |
1052 | ||
2fe86e00 | 1053 | /* |
ff9543fd MN |
1054 | * Based on information in the current compact_control, find blocks |
1055 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1056 | */ |
edc2ca61 | 1057 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1058 | { |
edc2ca61 | 1059 | struct zone *zone = cc->zone; |
ff9543fd | 1060 | struct page *page; |
c96b9e50 | 1061 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1062 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1063 | unsigned long block_end_pfn; /* end of current pageblock */ |
1064 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1065 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 1066 | |
ff9543fd MN |
1067 | /* |
1068 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1069 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1070 | * zone when isolating for the first time. For looping we also need |
1071 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1072 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1073 | * For ending point, take care when isolating in last pageblock of a | |
1074 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1075 | * The low boundary is the end of the pageblock the migration scanner |
1076 | * is using. | |
ff9543fd | 1077 | */ |
e14c720e | 1078 | isolate_start_pfn = cc->free_pfn; |
06b6640a | 1079 | block_start_pfn = pageblock_start_pfn(cc->free_pfn); |
c96b9e50 VB |
1080 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1081 | zone_end_pfn(zone)); | |
06b6640a | 1082 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
2fe86e00 | 1083 | |
ff9543fd MN |
1084 | /* |
1085 | * Isolate free pages until enough are available to migrate the | |
1086 | * pages on cc->migratepages. We stop searching if the migrate | |
1087 | * and free page scanners meet or enough free pages are isolated. | |
1088 | */ | |
f5f61a32 | 1089 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1090 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1091 | block_start_pfn -= pageblock_nr_pages, |
1092 | isolate_start_pfn = block_start_pfn) { | |
f6ea3adb DR |
1093 | /* |
1094 | * This can iterate a massively long zone without finding any | |
1095 | * suitable migration targets, so periodically check if we need | |
be976572 | 1096 | * to schedule, or even abort async compaction. |
f6ea3adb | 1097 | */ |
be976572 VB |
1098 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
1099 | && compact_should_abort(cc)) | |
1100 | break; | |
f6ea3adb | 1101 | |
7d49d886 VB |
1102 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1103 | zone); | |
1104 | if (!page) | |
ff9543fd MN |
1105 | continue; |
1106 | ||
1107 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1108 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1109 | continue; |
68e3e926 | 1110 | |
bb13ffeb MG |
1111 | /* If isolation recently failed, do not retry */ |
1112 | if (!isolation_suitable(cc, page)) | |
1113 | continue; | |
1114 | ||
e14c720e | 1115 | /* Found a block suitable for isolating free pages from. */ |
a46cbf3b DR |
1116 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, |
1117 | freelist, false); | |
ff9543fd | 1118 | |
e14c720e | 1119 | /* |
a46cbf3b DR |
1120 | * If we isolated enough freepages, or aborted due to lock |
1121 | * contention, terminate. | |
e14c720e | 1122 | */ |
f5f61a32 VB |
1123 | if ((cc->nr_freepages >= cc->nr_migratepages) |
1124 | || cc->contended) { | |
a46cbf3b DR |
1125 | if (isolate_start_pfn >= block_end_pfn) { |
1126 | /* | |
1127 | * Restart at previous pageblock if more | |
1128 | * freepages can be isolated next time. | |
1129 | */ | |
f5f61a32 VB |
1130 | isolate_start_pfn = |
1131 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1132 | } |
be976572 | 1133 | break; |
a46cbf3b | 1134 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1135 | /* |
a46cbf3b DR |
1136 | * If isolation failed early, do not continue |
1137 | * needlessly. | |
f5f61a32 | 1138 | */ |
a46cbf3b | 1139 | break; |
f5f61a32 | 1140 | } |
ff9543fd MN |
1141 | } |
1142 | ||
66c64223 | 1143 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 1144 | split_map_pages(freelist); |
ff9543fd | 1145 | |
7ed695e0 | 1146 | /* |
f5f61a32 VB |
1147 | * Record where the free scanner will restart next time. Either we |
1148 | * broke from the loop and set isolate_start_pfn based on the last | |
1149 | * call to isolate_freepages_block(), or we met the migration scanner | |
1150 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1151 | */ |
f5f61a32 | 1152 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1153 | } |
1154 | ||
1155 | /* | |
1156 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1157 | * from the isolated freelists in the block we are migrating to. | |
1158 | */ | |
1159 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1160 | unsigned long data) |
748446bb MG |
1161 | { |
1162 | struct compact_control *cc = (struct compact_control *)data; | |
1163 | struct page *freepage; | |
1164 | ||
be976572 VB |
1165 | /* |
1166 | * Isolate free pages if necessary, and if we are not aborting due to | |
1167 | * contention. | |
1168 | */ | |
748446bb | 1169 | if (list_empty(&cc->freepages)) { |
be976572 | 1170 | if (!cc->contended) |
edc2ca61 | 1171 | isolate_freepages(cc); |
748446bb MG |
1172 | |
1173 | if (list_empty(&cc->freepages)) | |
1174 | return NULL; | |
1175 | } | |
1176 | ||
1177 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1178 | list_del(&freepage->lru); | |
1179 | cc->nr_freepages--; | |
1180 | ||
1181 | return freepage; | |
1182 | } | |
1183 | ||
1184 | /* | |
d53aea3d DR |
1185 | * This is a migrate-callback that "frees" freepages back to the isolated |
1186 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1187 | * special handling needed for NUMA. | |
1188 | */ | |
1189 | static void compaction_free(struct page *page, unsigned long data) | |
1190 | { | |
1191 | struct compact_control *cc = (struct compact_control *)data; | |
1192 | ||
1193 | list_add(&page->lru, &cc->freepages); | |
1194 | cc->nr_freepages++; | |
1195 | } | |
1196 | ||
ff9543fd MN |
1197 | /* possible outcome of isolate_migratepages */ |
1198 | typedef enum { | |
1199 | ISOLATE_ABORT, /* Abort compaction now */ | |
1200 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1201 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1202 | } isolate_migrate_t; | |
1203 | ||
5bbe3547 EM |
1204 | /* |
1205 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1206 | * compactable pages. | |
1207 | */ | |
1208 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1209 | ||
ff9543fd | 1210 | /* |
edc2ca61 VB |
1211 | * Isolate all pages that can be migrated from the first suitable block, |
1212 | * starting at the block pointed to by the migrate scanner pfn within | |
1213 | * compact_control. | |
ff9543fd MN |
1214 | */ |
1215 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1216 | struct compact_control *cc) | |
1217 | { | |
e1409c32 JK |
1218 | unsigned long block_start_pfn; |
1219 | unsigned long block_end_pfn; | |
1220 | unsigned long low_pfn; | |
edc2ca61 VB |
1221 | struct page *page; |
1222 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1223 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1224 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1225 | |
edc2ca61 VB |
1226 | /* |
1227 | * Start at where we last stopped, or beginning of the zone as | |
1228 | * initialized by compact_zone() | |
1229 | */ | |
1230 | low_pfn = cc->migrate_pfn; | |
06b6640a | 1231 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1232 | if (block_start_pfn < zone->zone_start_pfn) |
1233 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd MN |
1234 | |
1235 | /* Only scan within a pageblock boundary */ | |
06b6640a | 1236 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1237 | |
edc2ca61 VB |
1238 | /* |
1239 | * Iterate over whole pageblocks until we find the first suitable. | |
1240 | * Do not cross the free scanner. | |
1241 | */ | |
e1409c32 JK |
1242 | for (; block_end_pfn <= cc->free_pfn; |
1243 | low_pfn = block_end_pfn, | |
1244 | block_start_pfn = block_end_pfn, | |
1245 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1246 | |
edc2ca61 VB |
1247 | /* |
1248 | * This can potentially iterate a massively long zone with | |
1249 | * many pageblocks unsuitable, so periodically check if we | |
1250 | * need to schedule, or even abort async compaction. | |
1251 | */ | |
1252 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1253 | && compact_should_abort(cc)) | |
1254 | break; | |
ff9543fd | 1255 | |
e1409c32 JK |
1256 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1257 | zone); | |
7d49d886 | 1258 | if (!page) |
edc2ca61 VB |
1259 | continue; |
1260 | ||
edc2ca61 VB |
1261 | /* If isolation recently failed, do not retry */ |
1262 | if (!isolation_suitable(cc, page)) | |
1263 | continue; | |
1264 | ||
1265 | /* | |
1266 | * For async compaction, also only scan in MOVABLE blocks. | |
1267 | * Async compaction is optimistic to see if the minimum amount | |
1268 | * of work satisfies the allocation. | |
1269 | */ | |
b682debd | 1270 | if (!suitable_migration_source(cc, page)) |
edc2ca61 VB |
1271 | continue; |
1272 | ||
1273 | /* Perform the isolation */ | |
e1409c32 JK |
1274 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1275 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1276 | |
6afcf8ef | 1277 | if (!low_pfn || cc->contended) |
edc2ca61 VB |
1278 | return ISOLATE_ABORT; |
1279 | ||
1280 | /* | |
1281 | * Either we isolated something and proceed with migration. Or | |
1282 | * we failed and compact_zone should decide if we should | |
1283 | * continue or not. | |
1284 | */ | |
1285 | break; | |
1286 | } | |
1287 | ||
f2849aa0 VB |
1288 | /* Record where migration scanner will be restarted. */ |
1289 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1290 | |
edc2ca61 | 1291 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1292 | } |
1293 | ||
21c527a3 YB |
1294 | /* |
1295 | * order == -1 is expected when compacting via | |
1296 | * /proc/sys/vm/compact_memory | |
1297 | */ | |
1298 | static inline bool is_via_compact_memory(int order) | |
1299 | { | |
1300 | return order == -1; | |
1301 | } | |
1302 | ||
40cacbcb | 1303 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 1304 | { |
8fb74b9f | 1305 | unsigned int order; |
d39773a0 | 1306 | const int migratetype = cc->migratetype; |
56de7263 | 1307 | |
be976572 | 1308 | if (cc->contended || fatal_signal_pending(current)) |
2d1e1041 | 1309 | return COMPACT_CONTENDED; |
748446bb | 1310 | |
753341a4 | 1311 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1312 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1313 | /* Let the next compaction start anew. */ |
40cacbcb | 1314 | reset_cached_positions(cc->zone); |
55b7c4c9 | 1315 | |
62997027 MG |
1316 | /* |
1317 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1318 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1319 | * flag itself as the decision to be clear should be directly |
1320 | * based on an allocation request. | |
1321 | */ | |
accf6242 | 1322 | if (cc->direct_compaction) |
40cacbcb | 1323 | cc->zone->compact_blockskip_flush = true; |
62997027 | 1324 | |
c8f7de0b MH |
1325 | if (cc->whole_zone) |
1326 | return COMPACT_COMPLETE; | |
1327 | else | |
1328 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1329 | } |
748446bb | 1330 | |
21c527a3 | 1331 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1332 | return COMPACT_CONTINUE; |
1333 | ||
efe771c7 MG |
1334 | /* |
1335 | * Always finish scanning a pageblock to reduce the possibility of | |
1336 | * fallbacks in the future. This is particularly important when | |
1337 | * migration source is unmovable/reclaimable but it's not worth | |
1338 | * special casing. | |
1339 | */ | |
1340 | if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1341 | return COMPACT_CONTINUE; | |
baf6a9a1 | 1342 | |
56de7263 | 1343 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f | 1344 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 1345 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 1346 | bool can_steal; |
8fb74b9f MG |
1347 | |
1348 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1349 | if (!list_empty(&area->free_list[migratetype])) |
cf378319 | 1350 | return COMPACT_SUCCESS; |
8fb74b9f | 1351 | |
2149cdae JK |
1352 | #ifdef CONFIG_CMA |
1353 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1354 | if (migratetype == MIGRATE_MOVABLE && | |
1355 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
cf378319 | 1356 | return COMPACT_SUCCESS; |
2149cdae JK |
1357 | #endif |
1358 | /* | |
1359 | * Job done if allocation would steal freepages from | |
1360 | * other migratetype buddy lists. | |
1361 | */ | |
1362 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
1363 | true, &can_steal) != -1) { |
1364 | ||
1365 | /* movable pages are OK in any pageblock */ | |
1366 | if (migratetype == MIGRATE_MOVABLE) | |
1367 | return COMPACT_SUCCESS; | |
1368 | ||
1369 | /* | |
1370 | * We are stealing for a non-movable allocation. Make | |
1371 | * sure we finish compacting the current pageblock | |
1372 | * first so it is as free as possible and we won't | |
1373 | * have to steal another one soon. This only applies | |
1374 | * to sync compaction, as async compaction operates | |
1375 | * on pageblocks of the same migratetype. | |
1376 | */ | |
1377 | if (cc->mode == MIGRATE_ASYNC || | |
1378 | IS_ALIGNED(cc->migrate_pfn, | |
1379 | pageblock_nr_pages)) { | |
1380 | return COMPACT_SUCCESS; | |
1381 | } | |
1382 | ||
baf6a9a1 VB |
1383 | return COMPACT_CONTINUE; |
1384 | } | |
56de7263 MG |
1385 | } |
1386 | ||
837d026d JK |
1387 | return COMPACT_NO_SUITABLE_PAGE; |
1388 | } | |
1389 | ||
40cacbcb | 1390 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
1391 | { |
1392 | int ret; | |
1393 | ||
40cacbcb MG |
1394 | ret = __compact_finished(cc); |
1395 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
1396 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
1397 | ret = COMPACT_CONTINUE; | |
1398 | ||
1399 | return ret; | |
748446bb MG |
1400 | } |
1401 | ||
3e7d3449 MG |
1402 | /* |
1403 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1404 | * Returns | |
1405 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
cf378319 | 1406 | * COMPACT_SUCCESS - If the allocation would succeed without compaction |
3e7d3449 MG |
1407 | * COMPACT_CONTINUE - If compaction should run now |
1408 | */ | |
ea7ab982 | 1409 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1410 | unsigned int alloc_flags, |
86a294a8 MH |
1411 | int classzone_idx, |
1412 | unsigned long wmark_target) | |
3e7d3449 | 1413 | { |
3e7d3449 MG |
1414 | unsigned long watermark; |
1415 | ||
21c527a3 | 1416 | if (is_via_compact_memory(order)) |
3957c776 MH |
1417 | return COMPACT_CONTINUE; |
1418 | ||
a9214443 | 1419 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
1420 | /* |
1421 | * If watermarks for high-order allocation are already met, there | |
1422 | * should be no need for compaction at all. | |
1423 | */ | |
1424 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1425 | alloc_flags)) | |
cf378319 | 1426 | return COMPACT_SUCCESS; |
ebff3980 | 1427 | |
3e7d3449 | 1428 | /* |
9861a62c | 1429 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
1430 | * isolate free pages for migration targets. This means that the |
1431 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1432 | * the check in __isolate_free_page(). We don't use the direct | |
1433 | * compactor's alloc_flags, as they are not relevant for freepage | |
1434 | * isolation. We however do use the direct compactor's classzone_idx to | |
1435 | * skip over zones where lowmem reserves would prevent allocation even | |
1436 | * if compaction succeeds. | |
8348faf9 VB |
1437 | * For costly orders, we require low watermark instead of min for |
1438 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
1439 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
1440 | * suitable migration targets | |
3e7d3449 | 1441 | */ |
8348faf9 VB |
1442 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
1443 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
1444 | watermark += compact_gap(order); | |
86a294a8 | 1445 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
d883c6cf | 1446 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
1447 | return COMPACT_SKIPPED; |
1448 | ||
cc5c9f09 VB |
1449 | return COMPACT_CONTINUE; |
1450 | } | |
1451 | ||
1452 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
1453 | unsigned int alloc_flags, | |
1454 | int classzone_idx) | |
1455 | { | |
1456 | enum compact_result ret; | |
1457 | int fragindex; | |
1458 | ||
1459 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
1460 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
1461 | /* |
1462 | * fragmentation index determines if allocation failures are due to | |
1463 | * low memory or external fragmentation | |
1464 | * | |
ebff3980 VB |
1465 | * index of -1000 would imply allocations might succeed depending on |
1466 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1467 | * index towards 0 implies failure is due to lack of memory |
1468 | * index towards 1000 implies failure is due to fragmentation | |
1469 | * | |
20311420 VB |
1470 | * Only compact if a failure would be due to fragmentation. Also |
1471 | * ignore fragindex for non-costly orders where the alternative to | |
1472 | * a successful reclaim/compaction is OOM. Fragindex and the | |
1473 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
1474 | * excessive compaction for costly orders, but it should not be at the | |
1475 | * expense of system stability. | |
3e7d3449 | 1476 | */ |
20311420 | 1477 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
1478 | fragindex = fragmentation_index(zone, order); |
1479 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
1480 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
1481 | } | |
837d026d | 1482 | |
837d026d JK |
1483 | trace_mm_compaction_suitable(zone, order, ret); |
1484 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1485 | ret = COMPACT_SKIPPED; | |
1486 | ||
1487 | return ret; | |
1488 | } | |
1489 | ||
86a294a8 MH |
1490 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
1491 | int alloc_flags) | |
1492 | { | |
1493 | struct zone *zone; | |
1494 | struct zoneref *z; | |
1495 | ||
1496 | /* | |
1497 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1498 | * retrying the reclaim. | |
1499 | */ | |
1500 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1501 | ac->nodemask) { | |
1502 | unsigned long available; | |
1503 | enum compact_result compact_result; | |
1504 | ||
1505 | /* | |
1506 | * Do not consider all the reclaimable memory because we do not | |
1507 | * want to trash just for a single high order allocation which | |
1508 | * is even not guaranteed to appear even if __compaction_suitable | |
1509 | * is happy about the watermark check. | |
1510 | */ | |
5a1c84b4 | 1511 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
1512 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
1513 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1514 | ac_classzone_idx(ac), available); | |
cc5c9f09 | 1515 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
1516 | return true; |
1517 | } | |
1518 | ||
1519 | return false; | |
1520 | } | |
1521 | ||
40cacbcb | 1522 | static enum compact_result compact_zone(struct compact_control *cc) |
748446bb | 1523 | { |
ea7ab982 | 1524 | enum compact_result ret; |
40cacbcb MG |
1525 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
1526 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 1527 | unsigned long last_migrated_pfn; |
e0b9daeb | 1528 | const bool sync = cc->mode != MIGRATE_ASYNC; |
748446bb | 1529 | |
d39773a0 | 1530 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
40cacbcb | 1531 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
ebff3980 | 1532 | cc->classzone_idx); |
c46649de | 1533 | /* Compaction is likely to fail */ |
cf378319 | 1534 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 1535 | return ret; |
c46649de MH |
1536 | |
1537 | /* huh, compaction_suitable is returning something unexpected */ | |
1538 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 1539 | |
d3132e4b VB |
1540 | /* |
1541 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 1542 | * is about to be retried after being deferred. |
d3132e4b | 1543 | */ |
40cacbcb MG |
1544 | if (compaction_restarting(cc->zone, cc->order)) |
1545 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 1546 | |
c89511ab MG |
1547 | /* |
1548 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
1549 | * information on where the scanners should start (unless we explicitly |
1550 | * want to compact the whole zone), but check that it is initialised | |
1551 | * by ensuring the values are within zone boundaries. | |
c89511ab | 1552 | */ |
06ed2998 | 1553 | if (cc->whole_zone) { |
c89511ab | 1554 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
1555 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
1556 | } else { | |
40cacbcb MG |
1557 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
1558 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
1559 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
1560 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 1561 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
1562 | } |
1563 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
1564 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
1565 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1566 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 1567 | } |
c8f7de0b | 1568 | |
06ed2998 VB |
1569 | if (cc->migrate_pfn == start_pfn) |
1570 | cc->whole_zone = true; | |
1571 | } | |
c8f7de0b | 1572 | |
566e54e1 | 1573 | last_migrated_pfn = 0; |
748446bb | 1574 | |
16c4a097 JK |
1575 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1576 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1577 | |
748446bb MG |
1578 | migrate_prep_local(); |
1579 | ||
40cacbcb | 1580 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 1581 | int err; |
566e54e1 | 1582 | unsigned long start_pfn = cc->migrate_pfn; |
748446bb | 1583 | |
40cacbcb | 1584 | switch (isolate_migratepages(cc->zone, cc)) { |
f9e35b3b | 1585 | case ISOLATE_ABORT: |
2d1e1041 | 1586 | ret = COMPACT_CONTENDED; |
5733c7d1 | 1587 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1588 | cc->nr_migratepages = 0; |
566e54e1 | 1589 | last_migrated_pfn = 0; |
f9e35b3b MG |
1590 | goto out; |
1591 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1592 | /* |
1593 | * We haven't isolated and migrated anything, but | |
1594 | * there might still be unflushed migrations from | |
1595 | * previous cc->order aligned block. | |
1596 | */ | |
1597 | goto check_drain; | |
f9e35b3b | 1598 | case ISOLATE_SUCCESS: |
566e54e1 | 1599 | last_migrated_pfn = start_pfn; |
f9e35b3b MG |
1600 | ; |
1601 | } | |
748446bb | 1602 | |
d53aea3d | 1603 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1604 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1605 | MR_COMPACTION); |
748446bb | 1606 | |
f8c9301f VB |
1607 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1608 | &cc->migratepages); | |
748446bb | 1609 | |
f8c9301f VB |
1610 | /* All pages were either migrated or will be released */ |
1611 | cc->nr_migratepages = 0; | |
9d502c1c | 1612 | if (err) { |
5733c7d1 | 1613 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1614 | /* |
1615 | * migrate_pages() may return -ENOMEM when scanners meet | |
1616 | * and we want compact_finished() to detect it | |
1617 | */ | |
f2849aa0 | 1618 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 1619 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
1620 | goto out; |
1621 | } | |
fdd048e1 VB |
1622 | /* |
1623 | * We failed to migrate at least one page in the current | |
1624 | * order-aligned block, so skip the rest of it. | |
1625 | */ | |
1626 | if (cc->direct_compaction && | |
1627 | (cc->mode == MIGRATE_ASYNC)) { | |
1628 | cc->migrate_pfn = block_end_pfn( | |
1629 | cc->migrate_pfn - 1, cc->order); | |
1630 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 1631 | last_migrated_pfn = 0; |
fdd048e1 | 1632 | } |
748446bb | 1633 | } |
fdaf7f5c | 1634 | |
fdaf7f5c VB |
1635 | check_drain: |
1636 | /* | |
1637 | * Has the migration scanner moved away from the previous | |
1638 | * cc->order aligned block where we migrated from? If yes, | |
1639 | * flush the pages that were freed, so that they can merge and | |
1640 | * compact_finished() can detect immediately if allocation | |
1641 | * would succeed. | |
1642 | */ | |
566e54e1 | 1643 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c VB |
1644 | int cpu; |
1645 | unsigned long current_block_start = | |
06b6640a | 1646 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 1647 | |
566e54e1 | 1648 | if (last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
1649 | cpu = get_cpu(); |
1650 | lru_add_drain_cpu(cpu); | |
40cacbcb | 1651 | drain_local_pages(cc->zone); |
fdaf7f5c VB |
1652 | put_cpu(); |
1653 | /* No more flushing until we migrate again */ | |
566e54e1 | 1654 | last_migrated_pfn = 0; |
fdaf7f5c VB |
1655 | } |
1656 | } | |
1657 | ||
748446bb MG |
1658 | } |
1659 | ||
f9e35b3b | 1660 | out: |
6bace090 VB |
1661 | /* |
1662 | * Release free pages and update where the free scanner should restart, | |
1663 | * so we don't leave any returned pages behind in the next attempt. | |
1664 | */ | |
1665 | if (cc->nr_freepages > 0) { | |
1666 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1667 | ||
1668 | cc->nr_freepages = 0; | |
1669 | VM_BUG_ON(free_pfn == 0); | |
1670 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 1671 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
1672 | /* |
1673 | * Only go back, not forward. The cached pfn might have been | |
1674 | * already reset to zone end in compact_finished() | |
1675 | */ | |
40cacbcb MG |
1676 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
1677 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 1678 | } |
748446bb | 1679 | |
7f354a54 DR |
1680 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
1681 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
1682 | ||
16c4a097 JK |
1683 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1684 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1685 | |
748446bb MG |
1686 | return ret; |
1687 | } | |
76ab0f53 | 1688 | |
ea7ab982 | 1689 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 1690 | gfp_t gfp_mask, enum compact_priority prio, |
c603844b | 1691 | unsigned int alloc_flags, int classzone_idx) |
56de7263 | 1692 | { |
ea7ab982 | 1693 | enum compact_result ret; |
56de7263 MG |
1694 | struct compact_control cc = { |
1695 | .nr_freepages = 0, | |
1696 | .nr_migratepages = 0, | |
7f354a54 DR |
1697 | .total_migrate_scanned = 0, |
1698 | .total_free_scanned = 0, | |
56de7263 | 1699 | .order = order, |
6d7ce559 | 1700 | .gfp_mask = gfp_mask, |
56de7263 | 1701 | .zone = zone, |
a5508cd8 VB |
1702 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
1703 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 VB |
1704 | .alloc_flags = alloc_flags, |
1705 | .classzone_idx = classzone_idx, | |
accf6242 | 1706 | .direct_compaction = true, |
a8e025e5 | 1707 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
1708 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
1709 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 MG |
1710 | }; |
1711 | INIT_LIST_HEAD(&cc.freepages); | |
1712 | INIT_LIST_HEAD(&cc.migratepages); | |
1713 | ||
40cacbcb | 1714 | ret = compact_zone(&cc); |
e64c5237 SL |
1715 | |
1716 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1717 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1718 | ||
e64c5237 | 1719 | return ret; |
56de7263 MG |
1720 | } |
1721 | ||
5e771905 MG |
1722 | int sysctl_extfrag_threshold = 500; |
1723 | ||
56de7263 MG |
1724 | /** |
1725 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1726 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1727 | * @order: The order of the current allocation |
1728 | * @alloc_flags: The allocation flags of the current allocation | |
1729 | * @ac: The context of current allocation | |
112d2d29 | 1730 | * @prio: Determines how hard direct compaction should try to succeed |
56de7263 MG |
1731 | * |
1732 | * This is the main entry point for direct page compaction. | |
1733 | */ | |
ea7ab982 | 1734 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 1735 | unsigned int alloc_flags, const struct alloc_context *ac, |
c3486f53 | 1736 | enum compact_priority prio) |
56de7263 | 1737 | { |
56de7263 | 1738 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
1739 | struct zoneref *z; |
1740 | struct zone *zone; | |
1d4746d3 | 1741 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 1742 | |
73e64c51 MH |
1743 | /* |
1744 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
1745 | * tricky context because the migration might require IO | |
1746 | */ | |
1747 | if (!may_perform_io) | |
53853e2d | 1748 | return COMPACT_SKIPPED; |
56de7263 | 1749 | |
a5508cd8 | 1750 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 1751 | |
56de7263 | 1752 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1753 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1754 | ac->nodemask) { | |
ea7ab982 | 1755 | enum compact_result status; |
56de7263 | 1756 | |
a8e025e5 VB |
1757 | if (prio > MIN_COMPACT_PRIORITY |
1758 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 1759 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 1760 | continue; |
1d4746d3 | 1761 | } |
53853e2d | 1762 | |
a5508cd8 | 1763 | status = compact_zone_order(zone, order, gfp_mask, prio, |
c3486f53 | 1764 | alloc_flags, ac_classzone_idx(ac)); |
56de7263 MG |
1765 | rc = max(status, rc); |
1766 | ||
7ceb009a VB |
1767 | /* The allocation should succeed, stop compacting */ |
1768 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
1769 | /* |
1770 | * We think the allocation will succeed in this zone, | |
1771 | * but it is not certain, hence the false. The caller | |
1772 | * will repeat this with true if allocation indeed | |
1773 | * succeeds in this zone. | |
1774 | */ | |
1775 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 1776 | |
c3486f53 | 1777 | break; |
1f9efdef VB |
1778 | } |
1779 | ||
a5508cd8 | 1780 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 1781 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
1782 | /* |
1783 | * We think that allocation won't succeed in this zone | |
1784 | * so we defer compaction there. If it ends up | |
1785 | * succeeding after all, it will be reset. | |
1786 | */ | |
1787 | defer_compaction(zone, order); | |
1f9efdef VB |
1788 | |
1789 | /* | |
1790 | * We might have stopped compacting due to need_resched() in | |
1791 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 1792 | * case do not try further zones |
1f9efdef | 1793 | */ |
c3486f53 VB |
1794 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
1795 | || fatal_signal_pending(current)) | |
1796 | break; | |
56de7263 MG |
1797 | } |
1798 | ||
1799 | return rc; | |
1800 | } | |
1801 | ||
1802 | ||
76ab0f53 | 1803 | /* Compact all zones within a node */ |
791cae96 | 1804 | static void compact_node(int nid) |
76ab0f53 | 1805 | { |
791cae96 | 1806 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 1807 | int zoneid; |
76ab0f53 | 1808 | struct zone *zone; |
791cae96 VB |
1809 | struct compact_control cc = { |
1810 | .order = -1, | |
7f354a54 DR |
1811 | .total_migrate_scanned = 0, |
1812 | .total_free_scanned = 0, | |
791cae96 VB |
1813 | .mode = MIGRATE_SYNC, |
1814 | .ignore_skip_hint = true, | |
1815 | .whole_zone = true, | |
73e64c51 | 1816 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
1817 | }; |
1818 | ||
76ab0f53 | 1819 | |
76ab0f53 | 1820 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1821 | |
1822 | zone = &pgdat->node_zones[zoneid]; | |
1823 | if (!populated_zone(zone)) | |
1824 | continue; | |
1825 | ||
791cae96 VB |
1826 | cc.nr_freepages = 0; |
1827 | cc.nr_migratepages = 0; | |
1828 | cc.zone = zone; | |
1829 | INIT_LIST_HEAD(&cc.freepages); | |
1830 | INIT_LIST_HEAD(&cc.migratepages); | |
76ab0f53 | 1831 | |
40cacbcb | 1832 | compact_zone(&cc); |
75469345 | 1833 | |
791cae96 VB |
1834 | VM_BUG_ON(!list_empty(&cc.freepages)); |
1835 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 1836 | } |
76ab0f53 MG |
1837 | } |
1838 | ||
1839 | /* Compact all nodes in the system */ | |
7964c06d | 1840 | static void compact_nodes(void) |
76ab0f53 MG |
1841 | { |
1842 | int nid; | |
1843 | ||
8575ec29 HD |
1844 | /* Flush pending updates to the LRU lists */ |
1845 | lru_add_drain_all(); | |
1846 | ||
76ab0f53 MG |
1847 | for_each_online_node(nid) |
1848 | compact_node(nid); | |
76ab0f53 MG |
1849 | } |
1850 | ||
1851 | /* The written value is actually unused, all memory is compacted */ | |
1852 | int sysctl_compact_memory; | |
1853 | ||
fec4eb2c YB |
1854 | /* |
1855 | * This is the entry point for compacting all nodes via | |
1856 | * /proc/sys/vm/compact_memory | |
1857 | */ | |
76ab0f53 MG |
1858 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
1859 | void __user *buffer, size_t *length, loff_t *ppos) | |
1860 | { | |
1861 | if (write) | |
7964c06d | 1862 | compact_nodes(); |
76ab0f53 MG |
1863 | |
1864 | return 0; | |
1865 | } | |
ed4a6d7f MG |
1866 | |
1867 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
74e77fb9 | 1868 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1869 | struct device_attribute *attr, |
ed4a6d7f MG |
1870 | const char *buf, size_t count) |
1871 | { | |
8575ec29 HD |
1872 | int nid = dev->id; |
1873 | ||
1874 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1875 | /* Flush pending updates to the LRU lists */ | |
1876 | lru_add_drain_all(); | |
1877 | ||
1878 | compact_node(nid); | |
1879 | } | |
ed4a6d7f MG |
1880 | |
1881 | return count; | |
1882 | } | |
0825a6f9 | 1883 | static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1884 | |
1885 | int compaction_register_node(struct node *node) | |
1886 | { | |
10fbcf4c | 1887 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1888 | } |
1889 | ||
1890 | void compaction_unregister_node(struct node *node) | |
1891 | { | |
10fbcf4c | 1892 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1893 | } |
1894 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 1895 | |
698b1b30 VB |
1896 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
1897 | { | |
172400c6 | 1898 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
1899 | } |
1900 | ||
1901 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
1902 | { | |
1903 | int zoneid; | |
1904 | struct zone *zone; | |
1905 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
1906 | ||
6cd9dc3e | 1907 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
1908 | zone = &pgdat->node_zones[zoneid]; |
1909 | ||
1910 | if (!populated_zone(zone)) | |
1911 | continue; | |
1912 | ||
1913 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
1914 | classzone_idx) == COMPACT_CONTINUE) | |
1915 | return true; | |
1916 | } | |
1917 | ||
1918 | return false; | |
1919 | } | |
1920 | ||
1921 | static void kcompactd_do_work(pg_data_t *pgdat) | |
1922 | { | |
1923 | /* | |
1924 | * With no special task, compact all zones so that a page of requested | |
1925 | * order is allocatable. | |
1926 | */ | |
1927 | int zoneid; | |
1928 | struct zone *zone; | |
1929 | struct compact_control cc = { | |
1930 | .order = pgdat->kcompactd_max_order, | |
7f354a54 DR |
1931 | .total_migrate_scanned = 0, |
1932 | .total_free_scanned = 0, | |
698b1b30 VB |
1933 | .classzone_idx = pgdat->kcompactd_classzone_idx, |
1934 | .mode = MIGRATE_SYNC_LIGHT, | |
a0647dc9 | 1935 | .ignore_skip_hint = false, |
73e64c51 | 1936 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 1937 | }; |
698b1b30 VB |
1938 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
1939 | cc.classzone_idx); | |
7f354a54 | 1940 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 1941 | |
6cd9dc3e | 1942 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
1943 | int status; |
1944 | ||
1945 | zone = &pgdat->node_zones[zoneid]; | |
1946 | if (!populated_zone(zone)) | |
1947 | continue; | |
1948 | ||
1949 | if (compaction_deferred(zone, cc.order)) | |
1950 | continue; | |
1951 | ||
1952 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
1953 | COMPACT_CONTINUE) | |
1954 | continue; | |
1955 | ||
1956 | cc.nr_freepages = 0; | |
1957 | cc.nr_migratepages = 0; | |
7f354a54 DR |
1958 | cc.total_migrate_scanned = 0; |
1959 | cc.total_free_scanned = 0; | |
698b1b30 VB |
1960 | cc.zone = zone; |
1961 | INIT_LIST_HEAD(&cc.freepages); | |
1962 | INIT_LIST_HEAD(&cc.migratepages); | |
1963 | ||
172400c6 VB |
1964 | if (kthread_should_stop()) |
1965 | return; | |
40cacbcb | 1966 | status = compact_zone(&cc); |
698b1b30 | 1967 | |
7ceb009a | 1968 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 1969 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 1970 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
1971 | /* |
1972 | * Buddy pages may become stranded on pcps that could | |
1973 | * otherwise coalesce on the zone's free area for | |
1974 | * order >= cc.order. This is ratelimited by the | |
1975 | * upcoming deferral. | |
1976 | */ | |
1977 | drain_all_pages(zone); | |
1978 | ||
698b1b30 VB |
1979 | /* |
1980 | * We use sync migration mode here, so we defer like | |
1981 | * sync direct compaction does. | |
1982 | */ | |
1983 | defer_compaction(zone, cc.order); | |
1984 | } | |
1985 | ||
7f354a54 DR |
1986 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
1987 | cc.total_migrate_scanned); | |
1988 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
1989 | cc.total_free_scanned); | |
1990 | ||
698b1b30 VB |
1991 | VM_BUG_ON(!list_empty(&cc.freepages)); |
1992 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1993 | } | |
1994 | ||
1995 | /* | |
1996 | * Regardless of success, we are done until woken up next. But remember | |
1997 | * the requested order/classzone_idx in case it was higher/tighter than | |
1998 | * our current ones | |
1999 | */ | |
2000 | if (pgdat->kcompactd_max_order <= cc.order) | |
2001 | pgdat->kcompactd_max_order = 0; | |
2002 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
2003 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2004 | } | |
2005 | ||
2006 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
2007 | { | |
2008 | if (!order) | |
2009 | return; | |
2010 | ||
2011 | if (pgdat->kcompactd_max_order < order) | |
2012 | pgdat->kcompactd_max_order = order; | |
2013 | ||
2014 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2015 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2016 | ||
6818600f DB |
2017 | /* |
2018 | * Pairs with implicit barrier in wait_event_freezable() | |
2019 | * such that wakeups are not missed. | |
2020 | */ | |
2021 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2022 | return; |
2023 | ||
2024 | if (!kcompactd_node_suitable(pgdat)) | |
2025 | return; | |
2026 | ||
2027 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2028 | classzone_idx); | |
2029 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2030 | } | |
2031 | ||
2032 | /* | |
2033 | * The background compaction daemon, started as a kernel thread | |
2034 | * from the init process. | |
2035 | */ | |
2036 | static int kcompactd(void *p) | |
2037 | { | |
2038 | pg_data_t *pgdat = (pg_data_t*)p; | |
2039 | struct task_struct *tsk = current; | |
2040 | ||
2041 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2042 | ||
2043 | if (!cpumask_empty(cpumask)) | |
2044 | set_cpus_allowed_ptr(tsk, cpumask); | |
2045 | ||
2046 | set_freezable(); | |
2047 | ||
2048 | pgdat->kcompactd_max_order = 0; | |
2049 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2050 | ||
2051 | while (!kthread_should_stop()) { | |
eb414681 JW |
2052 | unsigned long pflags; |
2053 | ||
698b1b30 VB |
2054 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
2055 | wait_event_freezable(pgdat->kcompactd_wait, | |
2056 | kcompactd_work_requested(pgdat)); | |
2057 | ||
eb414681 | 2058 | psi_memstall_enter(&pflags); |
698b1b30 | 2059 | kcompactd_do_work(pgdat); |
eb414681 | 2060 | psi_memstall_leave(&pflags); |
698b1b30 VB |
2061 | } |
2062 | ||
2063 | return 0; | |
2064 | } | |
2065 | ||
2066 | /* | |
2067 | * This kcompactd start function will be called by init and node-hot-add. | |
2068 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2069 | */ | |
2070 | int kcompactd_run(int nid) | |
2071 | { | |
2072 | pg_data_t *pgdat = NODE_DATA(nid); | |
2073 | int ret = 0; | |
2074 | ||
2075 | if (pgdat->kcompactd) | |
2076 | return 0; | |
2077 | ||
2078 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2079 | if (IS_ERR(pgdat->kcompactd)) { | |
2080 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2081 | ret = PTR_ERR(pgdat->kcompactd); | |
2082 | pgdat->kcompactd = NULL; | |
2083 | } | |
2084 | return ret; | |
2085 | } | |
2086 | ||
2087 | /* | |
2088 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2089 | * hold mem_hotplug_begin/end(). | |
2090 | */ | |
2091 | void kcompactd_stop(int nid) | |
2092 | { | |
2093 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2094 | ||
2095 | if (kcompactd) { | |
2096 | kthread_stop(kcompactd); | |
2097 | NODE_DATA(nid)->kcompactd = NULL; | |
2098 | } | |
2099 | } | |
2100 | ||
2101 | /* | |
2102 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2103 | * not required for correctness. So if the last cpu in a node goes | |
2104 | * away, we get changed to run anywhere: as the first one comes back, | |
2105 | * restore their cpu bindings. | |
2106 | */ | |
e46b1db2 | 2107 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
2108 | { |
2109 | int nid; | |
2110 | ||
e46b1db2 AMG |
2111 | for_each_node_state(nid, N_MEMORY) { |
2112 | pg_data_t *pgdat = NODE_DATA(nid); | |
2113 | const struct cpumask *mask; | |
698b1b30 | 2114 | |
e46b1db2 | 2115 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 2116 | |
e46b1db2 AMG |
2117 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
2118 | /* One of our CPUs online: restore mask */ | |
2119 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 2120 | } |
e46b1db2 | 2121 | return 0; |
698b1b30 VB |
2122 | } |
2123 | ||
2124 | static int __init kcompactd_init(void) | |
2125 | { | |
2126 | int nid; | |
e46b1db2 AMG |
2127 | int ret; |
2128 | ||
2129 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2130 | "mm/compaction:online", | |
2131 | kcompactd_cpu_online, NULL); | |
2132 | if (ret < 0) { | |
2133 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2134 | return ret; | |
2135 | } | |
698b1b30 VB |
2136 | |
2137 | for_each_node_state(nid, N_MEMORY) | |
2138 | kcompactd_run(nid); | |
698b1b30 VB |
2139 | return 0; |
2140 | } | |
2141 | subsys_initcall(kcompactd_init) | |
2142 | ||
ff9543fd | 2143 | #endif /* CONFIG_COMPACTION */ |