]>
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 | ||
e332f741 MG |
240 | static bool |
241 | __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, | |
242 | bool check_target) | |
243 | { | |
244 | struct page *page = pfn_to_online_page(pfn); | |
6b0868c8 | 245 | struct page *block_page; |
e332f741 MG |
246 | struct page *end_page; |
247 | unsigned long block_pfn; | |
248 | ||
249 | if (!page) | |
250 | return false; | |
251 | if (zone != page_zone(page)) | |
252 | return false; | |
253 | if (pageblock_skip_persistent(page)) | |
254 | return false; | |
255 | ||
256 | /* | |
257 | * If skip is already cleared do no further checking once the | |
258 | * restart points have been set. | |
259 | */ | |
260 | if (check_source && check_target && !get_pageblock_skip(page)) | |
261 | return true; | |
262 | ||
263 | /* | |
264 | * If clearing skip for the target scanner, do not select a | |
265 | * non-movable pageblock as the starting point. | |
266 | */ | |
267 | if (!check_source && check_target && | |
268 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
269 | return false; | |
270 | ||
6b0868c8 MG |
271 | /* Ensure the start of the pageblock or zone is online and valid */ |
272 | block_pfn = pageblock_start_pfn(pfn); | |
273 | block_page = pfn_to_online_page(max(block_pfn, zone->zone_start_pfn)); | |
274 | if (block_page) { | |
275 | page = block_page; | |
276 | pfn = block_pfn; | |
277 | } | |
278 | ||
279 | /* Ensure the end of the pageblock or zone is online and valid */ | |
280 | block_pfn += pageblock_nr_pages; | |
281 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); | |
282 | end_page = pfn_to_online_page(block_pfn); | |
283 | if (!end_page) | |
284 | return false; | |
285 | ||
e332f741 MG |
286 | /* |
287 | * Only clear the hint if a sample indicates there is either a | |
288 | * free page or an LRU page in the block. One or other condition | |
289 | * is necessary for the block to be a migration source/target. | |
290 | */ | |
e332f741 MG |
291 | do { |
292 | if (pfn_valid_within(pfn)) { | |
293 | if (check_source && PageLRU(page)) { | |
294 | clear_pageblock_skip(page); | |
295 | return true; | |
296 | } | |
297 | ||
298 | if (check_target && PageBuddy(page)) { | |
299 | clear_pageblock_skip(page); | |
300 | return true; | |
301 | } | |
302 | } | |
303 | ||
304 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
305 | pfn += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
306 | } while (page < end_page); | |
307 | ||
308 | return false; | |
309 | } | |
310 | ||
bb13ffeb MG |
311 | /* |
312 | * This function is called to clear all cached information on pageblocks that | |
313 | * should be skipped for page isolation when the migrate and free page scanner | |
314 | * meet. | |
315 | */ | |
62997027 | 316 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 317 | { |
e332f741 | 318 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 319 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
320 | unsigned long reset_migrate = free_pfn; |
321 | unsigned long reset_free = migrate_pfn; | |
322 | bool source_set = false; | |
323 | bool free_set = false; | |
324 | ||
325 | if (!zone->compact_blockskip_flush) | |
326 | return; | |
bb13ffeb | 327 | |
62997027 | 328 | zone->compact_blockskip_flush = false; |
bb13ffeb | 329 | |
e332f741 MG |
330 | /* |
331 | * Walk the zone and update pageblock skip information. Source looks | |
332 | * for PageLRU while target looks for PageBuddy. When the scanner | |
333 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
334 | * is suitable as both source and target. | |
335 | */ | |
336 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
337 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
338 | cond_resched(); |
339 | ||
e332f741 MG |
340 | /* Update the migrate PFN */ |
341 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
342 | migrate_pfn < reset_migrate) { | |
343 | source_set = true; | |
344 | reset_migrate = migrate_pfn; | |
345 | zone->compact_init_migrate_pfn = reset_migrate; | |
346 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
347 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
348 | } | |
bb13ffeb | 349 | |
e332f741 MG |
350 | /* Update the free PFN */ |
351 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
352 | free_pfn > reset_free) { | |
353 | free_set = true; | |
354 | reset_free = free_pfn; | |
355 | zone->compact_init_free_pfn = reset_free; | |
356 | zone->compact_cached_free_pfn = reset_free; | |
357 | } | |
bb13ffeb | 358 | } |
02333641 | 359 | |
e332f741 MG |
360 | /* Leave no distance if no suitable block was reset */ |
361 | if (reset_migrate >= reset_free) { | |
362 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
363 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
364 | zone->compact_cached_free_pfn = free_pfn; | |
365 | } | |
bb13ffeb MG |
366 | } |
367 | ||
62997027 MG |
368 | void reset_isolation_suitable(pg_data_t *pgdat) |
369 | { | |
370 | int zoneid; | |
371 | ||
372 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
373 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
374 | if (!populated_zone(zone)) | |
375 | continue; | |
376 | ||
377 | /* Only flush if a full compaction finished recently */ | |
378 | if (zone->compact_blockskip_flush) | |
379 | __reset_isolation_suitable(zone); | |
380 | } | |
381 | } | |
382 | ||
e380bebe MG |
383 | /* |
384 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
385 | * locks are not required for read/writers. Returns true if it was already set. | |
386 | */ | |
387 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
388 | unsigned long pfn) | |
389 | { | |
390 | bool skip; | |
391 | ||
392 | /* Do no update if skip hint is being ignored */ | |
393 | if (cc->ignore_skip_hint) | |
394 | return false; | |
395 | ||
396 | if (!IS_ALIGNED(pfn, pageblock_nr_pages)) | |
397 | return false; | |
398 | ||
399 | skip = get_pageblock_skip(page); | |
400 | if (!skip && !cc->no_set_skip_hint) | |
401 | set_pageblock_skip(page); | |
402 | ||
403 | return skip; | |
404 | } | |
405 | ||
406 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
407 | { | |
408 | struct zone *zone = cc->zone; | |
409 | ||
410 | pfn = pageblock_end_pfn(pfn); | |
411 | ||
412 | /* Set for isolation rather than compaction */ | |
413 | if (cc->no_set_skip_hint) | |
414 | return; | |
415 | ||
416 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
417 | zone->compact_cached_migrate_pfn[0] = pfn; | |
418 | if (cc->mode != MIGRATE_ASYNC && | |
419 | pfn > zone->compact_cached_migrate_pfn[1]) | |
420 | zone->compact_cached_migrate_pfn[1] = pfn; | |
421 | } | |
422 | ||
bb13ffeb MG |
423 | /* |
424 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 425 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 426 | */ |
c89511ab | 427 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 428 | struct page *page, unsigned long pfn) |
bb13ffeb | 429 | { |
c89511ab | 430 | struct zone *zone = cc->zone; |
6815bf3f | 431 | |
2583d671 | 432 | if (cc->no_set_skip_hint) |
6815bf3f JK |
433 | return; |
434 | ||
bb13ffeb MG |
435 | if (!page) |
436 | return; | |
437 | ||
edc2ca61 | 438 | set_pageblock_skip(page); |
c89511ab | 439 | |
35979ef3 | 440 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
441 | if (pfn < zone->compact_cached_free_pfn) |
442 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
443 | } |
444 | #else | |
445 | static inline bool isolation_suitable(struct compact_control *cc, | |
446 | struct page *page) | |
447 | { | |
448 | return true; | |
449 | } | |
450 | ||
b527cfe5 | 451 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
452 | { |
453 | return false; | |
454 | } | |
455 | ||
456 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 457 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
458 | { |
459 | } | |
e380bebe MG |
460 | |
461 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
462 | { | |
463 | } | |
464 | ||
465 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
466 | unsigned long pfn) | |
467 | { | |
468 | return false; | |
469 | } | |
bb13ffeb MG |
470 | #endif /* CONFIG_COMPACTION */ |
471 | ||
8b44d279 VB |
472 | /* |
473 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
474 | * very heavily contended. For async compaction, trylock and record if the |
475 | * lock is contended. The lock will still be acquired but compaction will | |
476 | * abort when the current block is finished regardless of success rate. | |
477 | * Sync compaction acquires the lock. | |
8b44d279 | 478 | * |
cb2dcaf0 | 479 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 480 | */ |
cb2dcaf0 | 481 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 482 | struct compact_control *cc) |
2a1402aa | 483 | { |
cb2dcaf0 MG |
484 | /* Track if the lock is contended in async mode */ |
485 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
486 | if (spin_trylock_irqsave(lock, *flags)) | |
487 | return true; | |
488 | ||
489 | cc->contended = true; | |
8b44d279 | 490 | } |
1f9efdef | 491 | |
cb2dcaf0 | 492 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 493 | return true; |
2a1402aa MG |
494 | } |
495 | ||
c67fe375 MG |
496 | /* |
497 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
498 | * very heavily contended. The lock should be periodically unlocked to avoid |
499 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
500 | * the lock. It might also be that allowing the IRQs will result in | |
501 | * need_resched() becoming true. If scheduling is needed, async compaction | |
502 | * aborts. Sync compaction schedules. | |
503 | * Either compaction type will also abort if a fatal signal is pending. | |
504 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 505 | * |
8b44d279 VB |
506 | * Returns true if compaction should abort due to fatal signal pending, or |
507 | * async compaction due to need_resched() | |
508 | * Returns false when compaction can continue (sync compaction might have | |
509 | * scheduled) | |
c67fe375 | 510 | */ |
8b44d279 VB |
511 | static bool compact_unlock_should_abort(spinlock_t *lock, |
512 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 513 | { |
8b44d279 VB |
514 | if (*locked) { |
515 | spin_unlock_irqrestore(lock, flags); | |
516 | *locked = false; | |
517 | } | |
1f9efdef | 518 | |
8b44d279 | 519 | if (fatal_signal_pending(current)) { |
c3486f53 | 520 | cc->contended = true; |
8b44d279 VB |
521 | return true; |
522 | } | |
c67fe375 | 523 | |
cf66f070 | 524 | cond_resched(); |
be976572 VB |
525 | |
526 | return false; | |
527 | } | |
528 | ||
85aa125f | 529 | /* |
9e4be470 JM |
530 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
531 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
532 | * (even though it may still end up isolating some pages). | |
85aa125f | 533 | */ |
f40d1e42 | 534 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 535 | unsigned long *start_pfn, |
85aa125f MN |
536 | unsigned long end_pfn, |
537 | struct list_head *freelist, | |
4fca9730 | 538 | unsigned int stride, |
85aa125f | 539 | bool strict) |
748446bb | 540 | { |
b7aba698 | 541 | int nr_scanned = 0, total_isolated = 0; |
d097a6f6 | 542 | struct page *cursor; |
b8b2d825 | 543 | unsigned long flags = 0; |
f40d1e42 | 544 | bool locked = false; |
e14c720e | 545 | unsigned long blockpfn = *start_pfn; |
66c64223 | 546 | unsigned int order; |
748446bb | 547 | |
4fca9730 MG |
548 | /* Strict mode is for isolation, speed is secondary */ |
549 | if (strict) | |
550 | stride = 1; | |
551 | ||
748446bb MG |
552 | cursor = pfn_to_page(blockpfn); |
553 | ||
f40d1e42 | 554 | /* Isolate free pages. */ |
4fca9730 | 555 | for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { |
66c64223 | 556 | int isolated; |
748446bb MG |
557 | struct page *page = cursor; |
558 | ||
8b44d279 VB |
559 | /* |
560 | * Periodically drop the lock (if held) regardless of its | |
561 | * contention, to give chance to IRQs. Abort if fatal signal | |
562 | * pending or async compaction detects need_resched() | |
563 | */ | |
564 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
565 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
566 | &locked, cc)) | |
567 | break; | |
568 | ||
b7aba698 | 569 | nr_scanned++; |
f40d1e42 | 570 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
571 | goto isolate_fail; |
572 | ||
9fcd6d2e VB |
573 | /* |
574 | * For compound pages such as THP and hugetlbfs, we can save | |
575 | * potentially a lot of iterations if we skip them at once. | |
576 | * The check is racy, but we can consider only valid values | |
577 | * and the only danger is skipping too much. | |
578 | */ | |
579 | if (PageCompound(page)) { | |
21dc7e02 DR |
580 | const unsigned int order = compound_order(page); |
581 | ||
d3c85bad | 582 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
583 | blockpfn += (1UL << order) - 1; |
584 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 585 | } |
9fcd6d2e VB |
586 | goto isolate_fail; |
587 | } | |
588 | ||
f40d1e42 | 589 | if (!PageBuddy(page)) |
2af120bc | 590 | goto isolate_fail; |
f40d1e42 MG |
591 | |
592 | /* | |
69b7189f VB |
593 | * If we already hold the lock, we can skip some rechecking. |
594 | * Note that if we hold the lock now, checked_pageblock was | |
595 | * already set in some previous iteration (or strict is true), | |
596 | * so it is correct to skip the suitable migration target | |
597 | * recheck as well. | |
f40d1e42 | 598 | */ |
69b7189f | 599 | if (!locked) { |
cb2dcaf0 | 600 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 601 | &flags, cc); |
f40d1e42 | 602 | |
69b7189f VB |
603 | /* Recheck this is a buddy page under lock */ |
604 | if (!PageBuddy(page)) | |
605 | goto isolate_fail; | |
606 | } | |
748446bb | 607 | |
66c64223 JK |
608 | /* Found a free page, will break it into order-0 pages */ |
609 | order = page_order(page); | |
610 | isolated = __isolate_free_page(page, order); | |
a4f04f2c DR |
611 | if (!isolated) |
612 | break; | |
66c64223 | 613 | set_page_private(page, order); |
a4f04f2c | 614 | |
748446bb | 615 | total_isolated += isolated; |
a4f04f2c | 616 | cc->nr_freepages += isolated; |
66c64223 JK |
617 | list_add_tail(&page->lru, freelist); |
618 | ||
a4f04f2c DR |
619 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
620 | blockpfn += isolated; | |
621 | break; | |
748446bb | 622 | } |
a4f04f2c DR |
623 | /* Advance to the end of split page */ |
624 | blockpfn += isolated - 1; | |
625 | cursor += isolated - 1; | |
626 | continue; | |
2af120bc LA |
627 | |
628 | isolate_fail: | |
629 | if (strict) | |
630 | break; | |
631 | else | |
632 | continue; | |
633 | ||
748446bb MG |
634 | } |
635 | ||
a4f04f2c DR |
636 | if (locked) |
637 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
638 | ||
9fcd6d2e VB |
639 | /* |
640 | * There is a tiny chance that we have read bogus compound_order(), | |
641 | * so be careful to not go outside of the pageblock. | |
642 | */ | |
643 | if (unlikely(blockpfn > end_pfn)) | |
644 | blockpfn = end_pfn; | |
645 | ||
e34d85f0 JK |
646 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
647 | nr_scanned, total_isolated); | |
648 | ||
e14c720e VB |
649 | /* Record how far we have got within the block */ |
650 | *start_pfn = blockpfn; | |
651 | ||
f40d1e42 MG |
652 | /* |
653 | * If strict isolation is requested by CMA then check that all the | |
654 | * pages requested were isolated. If there were any failures, 0 is | |
655 | * returned and CMA will fail. | |
656 | */ | |
2af120bc | 657 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
658 | total_isolated = 0; |
659 | ||
7f354a54 | 660 | cc->total_free_scanned += nr_scanned; |
397487db | 661 | if (total_isolated) |
010fc29a | 662 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
663 | return total_isolated; |
664 | } | |
665 | ||
85aa125f MN |
666 | /** |
667 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 668 | * @cc: Compaction control structure. |
85aa125f MN |
669 | * @start_pfn: The first PFN to start isolating. |
670 | * @end_pfn: The one-past-last PFN. | |
671 | * | |
672 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
673 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
674 | * undo its actions and return zero. | |
675 | * | |
676 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
677 | * (which may be greater then end_pfn if end fell in a middle of | |
678 | * a free page). | |
679 | */ | |
ff9543fd | 680 | unsigned long |
bb13ffeb MG |
681 | isolate_freepages_range(struct compact_control *cc, |
682 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 683 | { |
e1409c32 | 684 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
685 | LIST_HEAD(freelist); |
686 | ||
7d49d886 | 687 | pfn = start_pfn; |
06b6640a | 688 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
689 | if (block_start_pfn < cc->zone->zone_start_pfn) |
690 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 691 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
692 | |
693 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 694 | block_start_pfn = block_end_pfn, |
7d49d886 | 695 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
696 | /* Protect pfn from changing by isolate_freepages_block */ |
697 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 698 | |
85aa125f MN |
699 | block_end_pfn = min(block_end_pfn, end_pfn); |
700 | ||
58420016 JK |
701 | /* |
702 | * pfn could pass the block_end_pfn if isolated freepage | |
703 | * is more than pageblock order. In this case, we adjust | |
704 | * scanning range to right one. | |
705 | */ | |
706 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
707 | block_start_pfn = pageblock_start_pfn(pfn); |
708 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
709 | block_end_pfn = min(block_end_pfn, end_pfn); |
710 | } | |
711 | ||
e1409c32 JK |
712 | if (!pageblock_pfn_to_page(block_start_pfn, |
713 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
714 | break; |
715 | ||
e14c720e | 716 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
4fca9730 | 717 | block_end_pfn, &freelist, 0, true); |
85aa125f MN |
718 | |
719 | /* | |
720 | * In strict mode, isolate_freepages_block() returns 0 if | |
721 | * there are any holes in the block (ie. invalid PFNs or | |
722 | * non-free pages). | |
723 | */ | |
724 | if (!isolated) | |
725 | break; | |
726 | ||
727 | /* | |
728 | * If we managed to isolate pages, it is always (1 << n) * | |
729 | * pageblock_nr_pages for some non-negative n. (Max order | |
730 | * page may span two pageblocks). | |
731 | */ | |
732 | } | |
733 | ||
66c64223 | 734 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 735 | split_map_pages(&freelist); |
85aa125f MN |
736 | |
737 | if (pfn < end_pfn) { | |
738 | /* Loop terminated early, cleanup. */ | |
739 | release_freepages(&freelist); | |
740 | return 0; | |
741 | } | |
742 | ||
743 | /* We don't use freelists for anything. */ | |
744 | return pfn; | |
745 | } | |
746 | ||
748446bb | 747 | /* Similar to reclaim, but different enough that they don't share logic */ |
5f438eee | 748 | static bool too_many_isolated(pg_data_t *pgdat) |
748446bb | 749 | { |
bc693045 | 750 | unsigned long active, inactive, isolated; |
748446bb | 751 | |
5f438eee AR |
752 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + |
753 | node_page_state(pgdat, NR_INACTIVE_ANON); | |
754 | active = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
755 | node_page_state(pgdat, NR_ACTIVE_ANON); | |
756 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + | |
757 | node_page_state(pgdat, NR_ISOLATED_ANON); | |
748446bb | 758 | |
bc693045 | 759 | return isolated > (inactive + active) / 2; |
748446bb MG |
760 | } |
761 | ||
2fe86e00 | 762 | /** |
edc2ca61 VB |
763 | * isolate_migratepages_block() - isolate all migrate-able pages within |
764 | * a single pageblock | |
2fe86e00 | 765 | * @cc: Compaction control structure. |
edc2ca61 VB |
766 | * @low_pfn: The first PFN to isolate |
767 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
768 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
769 | * |
770 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
771 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
772 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
773 | * first page that was not scanned (which may be both less, equal to or more | |
774 | * than end_pfn). | |
2fe86e00 | 775 | * |
edc2ca61 VB |
776 | * The pages are isolated on cc->migratepages list (not required to be empty), |
777 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
778 | * is neither read nor updated. | |
748446bb | 779 | */ |
edc2ca61 VB |
780 | static unsigned long |
781 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
782 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 783 | { |
5f438eee | 784 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 785 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 786 | struct lruvec *lruvec; |
b8b2d825 | 787 | unsigned long flags = 0; |
2a1402aa | 788 | bool locked = false; |
bb13ffeb | 789 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 790 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
791 | bool skip_on_failure = false; |
792 | unsigned long next_skip_pfn = 0; | |
e380bebe | 793 | bool skip_updated = false; |
748446bb | 794 | |
748446bb MG |
795 | /* |
796 | * Ensure that there are not too many pages isolated from the LRU | |
797 | * list by either parallel reclaimers or compaction. If there are, | |
798 | * delay for some time until fewer pages are isolated | |
799 | */ | |
5f438eee | 800 | while (unlikely(too_many_isolated(pgdat))) { |
f9e35b3b | 801 | /* async migration should just abort */ |
e0b9daeb | 802 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 803 | return 0; |
f9e35b3b | 804 | |
748446bb MG |
805 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
806 | ||
807 | if (fatal_signal_pending(current)) | |
2fe86e00 | 808 | return 0; |
748446bb MG |
809 | } |
810 | ||
cf66f070 | 811 | cond_resched(); |
aeef4b83 | 812 | |
fdd048e1 VB |
813 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
814 | skip_on_failure = true; | |
815 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
816 | } | |
817 | ||
748446bb | 818 | /* Time to isolate some pages for migration */ |
748446bb | 819 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 820 | |
fdd048e1 VB |
821 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
822 | /* | |
823 | * We have isolated all migration candidates in the | |
824 | * previous order-aligned block, and did not skip it due | |
825 | * to failure. We should migrate the pages now and | |
826 | * hopefully succeed compaction. | |
827 | */ | |
828 | if (nr_isolated) | |
829 | break; | |
830 | ||
831 | /* | |
832 | * We failed to isolate in the previous order-aligned | |
833 | * block. Set the new boundary to the end of the | |
834 | * current block. Note we can't simply increase | |
835 | * next_skip_pfn by 1 << order, as low_pfn might have | |
836 | * been incremented by a higher number due to skipping | |
837 | * a compound or a high-order buddy page in the | |
838 | * previous loop iteration. | |
839 | */ | |
840 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
841 | } | |
842 | ||
8b44d279 VB |
843 | /* |
844 | * Periodically drop the lock (if held) regardless of its | |
845 | * contention, to give chance to IRQs. Abort async compaction | |
846 | * if contended. | |
847 | */ | |
848 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
f4b7e272 AR |
849 | && compact_unlock_should_abort(&pgdat->lru_lock, |
850 | flags, &locked, cc)) | |
8b44d279 | 851 | break; |
c67fe375 | 852 | |
748446bb | 853 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 854 | goto isolate_fail; |
b7aba698 | 855 | nr_scanned++; |
748446bb | 856 | |
748446bb | 857 | page = pfn_to_page(low_pfn); |
dc908600 | 858 | |
e380bebe MG |
859 | /* |
860 | * Check if the pageblock has already been marked skipped. | |
861 | * Only the aligned PFN is checked as the caller isolates | |
862 | * COMPACT_CLUSTER_MAX at a time so the second call must | |
863 | * not falsely conclude that the block should be skipped. | |
864 | */ | |
865 | if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) { | |
866 | if (!cc->ignore_skip_hint && get_pageblock_skip(page)) { | |
867 | low_pfn = end_pfn; | |
868 | goto isolate_abort; | |
869 | } | |
bb13ffeb | 870 | valid_page = page; |
e380bebe | 871 | } |
bb13ffeb | 872 | |
6c14466c | 873 | /* |
99c0fd5e VB |
874 | * Skip if free. We read page order here without zone lock |
875 | * which is generally unsafe, but the race window is small and | |
876 | * the worst thing that can happen is that we skip some | |
877 | * potential isolation targets. | |
6c14466c | 878 | */ |
99c0fd5e VB |
879 | if (PageBuddy(page)) { |
880 | unsigned long freepage_order = page_order_unsafe(page); | |
881 | ||
882 | /* | |
883 | * Without lock, we cannot be sure that what we got is | |
884 | * a valid page order. Consider only values in the | |
885 | * valid order range to prevent low_pfn overflow. | |
886 | */ | |
887 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
888 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 889 | continue; |
99c0fd5e | 890 | } |
748446bb | 891 | |
bc835011 | 892 | /* |
29c0dde8 VB |
893 | * Regardless of being on LRU, compound pages such as THP and |
894 | * hugetlbfs are not to be compacted. We can potentially save | |
895 | * a lot of iterations if we skip them at once. The check is | |
896 | * racy, but we can consider only valid values and the only | |
897 | * danger is skipping too much. | |
bc835011 | 898 | */ |
29c0dde8 | 899 | if (PageCompound(page)) { |
21dc7e02 | 900 | const unsigned int order = compound_order(page); |
edc2ca61 | 901 | |
d3c85bad | 902 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 903 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 904 | goto isolate_fail; |
2a1402aa MG |
905 | } |
906 | ||
bda807d4 MK |
907 | /* |
908 | * Check may be lockless but that's ok as we recheck later. | |
909 | * It's possible to migrate LRU and non-lru movable pages. | |
910 | * Skip any other type of page | |
911 | */ | |
912 | if (!PageLRU(page)) { | |
bda807d4 MK |
913 | /* |
914 | * __PageMovable can return false positive so we need | |
915 | * to verify it under page_lock. | |
916 | */ | |
917 | if (unlikely(__PageMovable(page)) && | |
918 | !PageIsolated(page)) { | |
919 | if (locked) { | |
f4b7e272 | 920 | spin_unlock_irqrestore(&pgdat->lru_lock, |
bda807d4 MK |
921 | flags); |
922 | locked = false; | |
923 | } | |
924 | ||
9e5bcd61 | 925 | if (!isolate_movable_page(page, isolate_mode)) |
bda807d4 MK |
926 | goto isolate_success; |
927 | } | |
928 | ||
fdd048e1 | 929 | goto isolate_fail; |
bda807d4 | 930 | } |
29c0dde8 | 931 | |
119d6d59 DR |
932 | /* |
933 | * Migration will fail if an anonymous page is pinned in memory, | |
934 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
935 | * admittedly racy check. | |
936 | */ | |
937 | if (!page_mapping(page) && | |
938 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 939 | goto isolate_fail; |
119d6d59 | 940 | |
73e64c51 MH |
941 | /* |
942 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
943 | * because those do not depend on fs locks. | |
944 | */ | |
945 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
946 | goto isolate_fail; | |
947 | ||
69b7189f VB |
948 | /* If we already hold the lock, we can skip some rechecking */ |
949 | if (!locked) { | |
f4b7e272 | 950 | locked = compact_lock_irqsave(&pgdat->lru_lock, |
8b44d279 | 951 | &flags, cc); |
e380bebe | 952 | |
e380bebe MG |
953 | /* Try get exclusive access under lock */ |
954 | if (!skip_updated) { | |
955 | skip_updated = true; | |
956 | if (test_and_set_skip(cc, page, low_pfn)) | |
957 | goto isolate_abort; | |
958 | } | |
2a1402aa | 959 | |
29c0dde8 | 960 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 961 | if (!PageLRU(page)) |
fdd048e1 | 962 | goto isolate_fail; |
29c0dde8 VB |
963 | |
964 | /* | |
965 | * Page become compound since the non-locked check, | |
966 | * and it's on LRU. It can only be a THP so the order | |
967 | * is safe to read and it's 0 for tail pages. | |
968 | */ | |
969 | if (unlikely(PageCompound(page))) { | |
d3c85bad | 970 | low_pfn += (1UL << compound_order(page)) - 1; |
fdd048e1 | 971 | goto isolate_fail; |
69b7189f | 972 | } |
bc835011 AA |
973 | } |
974 | ||
f4b7e272 | 975 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 976 | |
748446bb | 977 | /* Try isolate the page */ |
edc2ca61 | 978 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 979 | goto isolate_fail; |
748446bb | 980 | |
29c0dde8 | 981 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 982 | |
748446bb | 983 | /* Successfully isolated */ |
fa9add64 | 984 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
6afcf8ef ML |
985 | inc_node_page_state(page, |
986 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
b6c75016 JK |
987 | |
988 | isolate_success: | |
fdd048e1 | 989 | list_add(&page->lru, &cc->migratepages); |
748446bb | 990 | cc->nr_migratepages++; |
b7aba698 | 991 | nr_isolated++; |
748446bb | 992 | |
804d3121 MG |
993 | /* |
994 | * Avoid isolating too much unless this block is being | |
cb2dcaf0 MG |
995 | * rescanned (e.g. dirty/writeback pages, parallel allocation) |
996 | * or a lock is contended. For contention, isolate quickly to | |
997 | * potentially remove one source of contention. | |
804d3121 | 998 | */ |
cb2dcaf0 MG |
999 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX && |
1000 | !cc->rescan && !cc->contended) { | |
31b8384a | 1001 | ++low_pfn; |
748446bb | 1002 | break; |
31b8384a | 1003 | } |
fdd048e1 VB |
1004 | |
1005 | continue; | |
1006 | isolate_fail: | |
1007 | if (!skip_on_failure) | |
1008 | continue; | |
1009 | ||
1010 | /* | |
1011 | * We have isolated some pages, but then failed. Release them | |
1012 | * instead of migrating, as we cannot form the cc->order buddy | |
1013 | * page anyway. | |
1014 | */ | |
1015 | if (nr_isolated) { | |
1016 | if (locked) { | |
f4b7e272 | 1017 | spin_unlock_irqrestore(&pgdat->lru_lock, flags); |
fdd048e1 VB |
1018 | locked = false; |
1019 | } | |
fdd048e1 VB |
1020 | putback_movable_pages(&cc->migratepages); |
1021 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1022 | nr_isolated = 0; |
1023 | } | |
1024 | ||
1025 | if (low_pfn < next_skip_pfn) { | |
1026 | low_pfn = next_skip_pfn - 1; | |
1027 | /* | |
1028 | * The check near the loop beginning would have updated | |
1029 | * next_skip_pfn too, but this is a bit simpler. | |
1030 | */ | |
1031 | next_skip_pfn += 1UL << cc->order; | |
1032 | } | |
748446bb MG |
1033 | } |
1034 | ||
99c0fd5e VB |
1035 | /* |
1036 | * The PageBuddy() check could have potentially brought us outside | |
1037 | * the range to be scanned. | |
1038 | */ | |
1039 | if (unlikely(low_pfn > end_pfn)) | |
1040 | low_pfn = end_pfn; | |
1041 | ||
e380bebe | 1042 | isolate_abort: |
c67fe375 | 1043 | if (locked) |
f4b7e272 | 1044 | spin_unlock_irqrestore(&pgdat->lru_lock, flags); |
748446bb | 1045 | |
50b5b094 | 1046 | /* |
804d3121 MG |
1047 | * Updated the cached scanner pfn once the pageblock has been scanned |
1048 | * Pages will either be migrated in which case there is no point | |
1049 | * scanning in the near future or migration failed in which case the | |
1050 | * failure reason may persist. The block is marked for skipping if | |
1051 | * there were no pages isolated in the block or if the block is | |
1052 | * rescanned twice in a row. | |
50b5b094 | 1053 | */ |
804d3121 | 1054 | if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) { |
e380bebe MG |
1055 | if (valid_page && !skip_updated) |
1056 | set_pageblock_skip(valid_page); | |
1057 | update_cached_migrate(cc, low_pfn); | |
1058 | } | |
bb13ffeb | 1059 | |
e34d85f0 JK |
1060 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1061 | nr_scanned, nr_isolated); | |
b7aba698 | 1062 | |
7f354a54 | 1063 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1064 | if (nr_isolated) |
010fc29a | 1065 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1066 | |
2fe86e00 MN |
1067 | return low_pfn; |
1068 | } | |
1069 | ||
edc2ca61 VB |
1070 | /** |
1071 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1072 | * @cc: Compaction control structure. | |
1073 | * @start_pfn: The first PFN to start isolating. | |
1074 | * @end_pfn: The one-past-last PFN. | |
1075 | * | |
1076 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
1077 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
1078 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
1079 | */ | |
1080 | unsigned long | |
1081 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
1082 | unsigned long end_pfn) | |
1083 | { | |
e1409c32 | 1084 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
1085 | |
1086 | /* Scan block by block. First and last block may be incomplete */ | |
1087 | pfn = start_pfn; | |
06b6640a | 1088 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1089 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1090 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1091 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1092 | |
1093 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1094 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1095 | block_end_pfn += pageblock_nr_pages) { |
1096 | ||
1097 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1098 | ||
e1409c32 JK |
1099 | if (!pageblock_pfn_to_page(block_start_pfn, |
1100 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1101 | continue; |
1102 | ||
1103 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
1104 | ISOLATE_UNEVICTABLE); | |
1105 | ||
14af4a5e | 1106 | if (!pfn) |
edc2ca61 | 1107 | break; |
6ea41c0c JK |
1108 | |
1109 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
1110 | break; | |
edc2ca61 | 1111 | } |
edc2ca61 VB |
1112 | |
1113 | return pfn; | |
1114 | } | |
1115 | ||
ff9543fd MN |
1116 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1117 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1118 | |
b682debd VB |
1119 | static bool suitable_migration_source(struct compact_control *cc, |
1120 | struct page *page) | |
1121 | { | |
282722b0 VB |
1122 | int block_mt; |
1123 | ||
9bebefd5 MG |
1124 | if (pageblock_skip_persistent(page)) |
1125 | return false; | |
1126 | ||
282722b0 | 1127 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1128 | return true; |
1129 | ||
282722b0 VB |
1130 | block_mt = get_pageblock_migratetype(page); |
1131 | ||
1132 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1133 | return is_migrate_movable(block_mt); | |
1134 | else | |
1135 | return block_mt == cc->migratetype; | |
b682debd VB |
1136 | } |
1137 | ||
018e9a49 | 1138 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1139 | static bool suitable_migration_target(struct compact_control *cc, |
1140 | struct page *page) | |
018e9a49 AM |
1141 | { |
1142 | /* If the page is a large free page, then disallow migration */ | |
1143 | if (PageBuddy(page)) { | |
1144 | /* | |
1145 | * We are checking page_order without zone->lock taken. But | |
1146 | * the only small danger is that we skip a potentially suitable | |
1147 | * pageblock, so it's not worth to check order for valid range. | |
1148 | */ | |
1149 | if (page_order_unsafe(page) >= pageblock_order) | |
1150 | return false; | |
1151 | } | |
1152 | ||
1ef36db2 YX |
1153 | if (cc->ignore_block_suitable) |
1154 | return true; | |
1155 | ||
018e9a49 | 1156 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1157 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1158 | return true; |
1159 | ||
1160 | /* Otherwise skip the block */ | |
1161 | return false; | |
1162 | } | |
1163 | ||
70b44595 MG |
1164 | static inline unsigned int |
1165 | freelist_scan_limit(struct compact_control *cc) | |
1166 | { | |
1167 | return (COMPACT_CLUSTER_MAX >> cc->fast_search_fail) + 1; | |
1168 | } | |
1169 | ||
f2849aa0 VB |
1170 | /* |
1171 | * Test whether the free scanner has reached the same or lower pageblock than | |
1172 | * the migration scanner, and compaction should thus terminate. | |
1173 | */ | |
1174 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1175 | { | |
1176 | return (cc->free_pfn >> pageblock_order) | |
1177 | <= (cc->migrate_pfn >> pageblock_order); | |
1178 | } | |
1179 | ||
5a811889 MG |
1180 | /* |
1181 | * Used when scanning for a suitable migration target which scans freelists | |
1182 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1183 | * first on the next iteration of the free scanner | |
1184 | */ | |
1185 | static void | |
1186 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1187 | { | |
1188 | LIST_HEAD(sublist); | |
1189 | ||
1190 | if (!list_is_last(freelist, &freepage->lru)) { | |
1191 | list_cut_before(&sublist, freelist, &freepage->lru); | |
1192 | if (!list_empty(&sublist)) | |
1193 | list_splice_tail(&sublist, freelist); | |
1194 | } | |
1195 | } | |
1196 | ||
1197 | /* | |
1198 | * Similar to move_freelist_head except used by the migration scanner | |
1199 | * when scanning forward. It's possible for these list operations to | |
1200 | * move against each other if they search the free list exactly in | |
1201 | * lockstep. | |
1202 | */ | |
70b44595 MG |
1203 | static void |
1204 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1205 | { | |
1206 | LIST_HEAD(sublist); | |
1207 | ||
1208 | if (!list_is_first(freelist, &freepage->lru)) { | |
1209 | list_cut_position(&sublist, freelist, &freepage->lru); | |
1210 | if (!list_empty(&sublist)) | |
1211 | list_splice_tail(&sublist, freelist); | |
1212 | } | |
1213 | } | |
1214 | ||
5a811889 MG |
1215 | static void |
1216 | fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated) | |
1217 | { | |
1218 | unsigned long start_pfn, end_pfn; | |
1219 | struct page *page = pfn_to_page(pfn); | |
1220 | ||
1221 | /* Do not search around if there are enough pages already */ | |
1222 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1223 | return; | |
1224 | ||
1225 | /* Minimise scanning during async compaction */ | |
1226 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1227 | return; | |
1228 | ||
1229 | /* Pageblock boundaries */ | |
1230 | start_pfn = pageblock_start_pfn(pfn); | |
1231 | end_pfn = min(start_pfn + pageblock_nr_pages, zone_end_pfn(cc->zone)); | |
1232 | ||
1233 | /* Scan before */ | |
1234 | if (start_pfn != pfn) { | |
4fca9730 | 1235 | isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false); |
5a811889 MG |
1236 | if (cc->nr_freepages >= cc->nr_migratepages) |
1237 | return; | |
1238 | } | |
1239 | ||
1240 | /* Scan after */ | |
1241 | start_pfn = pfn + nr_isolated; | |
1242 | if (start_pfn != end_pfn) | |
4fca9730 | 1243 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); |
5a811889 MG |
1244 | |
1245 | /* Skip this pageblock in the future as it's full or nearly full */ | |
1246 | if (cc->nr_freepages < cc->nr_migratepages) | |
1247 | set_pageblock_skip(page); | |
1248 | } | |
1249 | ||
dbe2d4e4 MG |
1250 | /* Search orders in round-robin fashion */ |
1251 | static int next_search_order(struct compact_control *cc, int order) | |
1252 | { | |
1253 | order--; | |
1254 | if (order < 0) | |
1255 | order = cc->order - 1; | |
1256 | ||
1257 | /* Search wrapped around? */ | |
1258 | if (order == cc->search_order) { | |
1259 | cc->search_order--; | |
1260 | if (cc->search_order < 0) | |
1261 | cc->search_order = cc->order - 1; | |
1262 | return -1; | |
1263 | } | |
1264 | ||
1265 | return order; | |
1266 | } | |
1267 | ||
5a811889 MG |
1268 | static unsigned long |
1269 | fast_isolate_freepages(struct compact_control *cc) | |
1270 | { | |
1271 | unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1); | |
1272 | unsigned int nr_scanned = 0; | |
1273 | unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0; | |
1274 | unsigned long nr_isolated = 0; | |
1275 | unsigned long distance; | |
1276 | struct page *page = NULL; | |
1277 | bool scan_start = false; | |
1278 | int order; | |
1279 | ||
1280 | /* Full compaction passes in a negative order */ | |
1281 | if (cc->order <= 0) | |
1282 | return cc->free_pfn; | |
1283 | ||
1284 | /* | |
1285 | * If starting the scan, use a deeper search and use the highest | |
1286 | * PFN found if a suitable one is not found. | |
1287 | */ | |
e332f741 | 1288 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1289 | limit = pageblock_nr_pages >> 1; |
1290 | scan_start = true; | |
1291 | } | |
1292 | ||
1293 | /* | |
1294 | * Preferred point is in the top quarter of the scan space but take | |
1295 | * a pfn from the top half if the search is problematic. | |
1296 | */ | |
1297 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1298 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1299 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1300 | ||
1301 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1302 | low_pfn = min_pfn; | |
1303 | ||
dbe2d4e4 MG |
1304 | /* |
1305 | * Search starts from the last successful isolation order or the next | |
1306 | * order to search after a previous failure | |
1307 | */ | |
1308 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1309 | ||
1310 | for (order = cc->search_order; | |
1311 | !page && order >= 0; | |
1312 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1313 | struct free_area *area = &cc->zone->free_area[order]; |
1314 | struct list_head *freelist; | |
1315 | struct page *freepage; | |
1316 | unsigned long flags; | |
1317 | unsigned int order_scanned = 0; | |
1318 | ||
1319 | if (!area->nr_free) | |
1320 | continue; | |
1321 | ||
1322 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1323 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1324 | list_for_each_entry_reverse(freepage, freelist, lru) { | |
1325 | unsigned long pfn; | |
1326 | ||
1327 | order_scanned++; | |
1328 | nr_scanned++; | |
1329 | pfn = page_to_pfn(freepage); | |
1330 | ||
1331 | if (pfn >= highest) | |
1332 | highest = pageblock_start_pfn(pfn); | |
1333 | ||
1334 | if (pfn >= low_pfn) { | |
1335 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1336 | cc->search_order = order; |
5a811889 MG |
1337 | page = freepage; |
1338 | break; | |
1339 | } | |
1340 | ||
1341 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1342 | high_pfn = pfn; | |
1343 | ||
1344 | /* Shorten the scan if a candidate is found */ | |
1345 | limit >>= 1; | |
1346 | } | |
1347 | ||
1348 | if (order_scanned >= limit) | |
1349 | break; | |
1350 | } | |
1351 | ||
1352 | /* Use a minimum pfn if a preferred one was not found */ | |
1353 | if (!page && high_pfn) { | |
1354 | page = pfn_to_page(high_pfn); | |
1355 | ||
1356 | /* Update freepage for the list reorder below */ | |
1357 | freepage = page; | |
1358 | } | |
1359 | ||
1360 | /* Reorder to so a future search skips recent pages */ | |
1361 | move_freelist_head(freelist, freepage); | |
1362 | ||
1363 | /* Isolate the page if available */ | |
1364 | if (page) { | |
1365 | if (__isolate_free_page(page, order)) { | |
1366 | set_page_private(page, order); | |
1367 | nr_isolated = 1 << order; | |
1368 | cc->nr_freepages += nr_isolated; | |
1369 | list_add_tail(&page->lru, &cc->freepages); | |
1370 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1371 | } else { | |
1372 | /* If isolation fails, abort the search */ | |
5b56d996 | 1373 | order = cc->search_order + 1; |
5a811889 MG |
1374 | page = NULL; |
1375 | } | |
1376 | } | |
1377 | ||
1378 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1379 | ||
1380 | /* | |
1381 | * Smaller scan on next order so the total scan ig related | |
1382 | * to freelist_scan_limit. | |
1383 | */ | |
1384 | if (order_scanned >= limit) | |
1385 | limit = min(1U, limit >> 1); | |
1386 | } | |
1387 | ||
1388 | if (!page) { | |
1389 | cc->fast_search_fail++; | |
1390 | if (scan_start) { | |
1391 | /* | |
1392 | * Use the highest PFN found above min. If one was | |
1393 | * not found, be pessemistic for direct compaction | |
1394 | * and use the min mark. | |
1395 | */ | |
1396 | if (highest) { | |
1397 | page = pfn_to_page(highest); | |
1398 | cc->free_pfn = highest; | |
1399 | } else { | |
1400 | if (cc->direct_compaction) { | |
1401 | page = pfn_to_page(min_pfn); | |
1402 | cc->free_pfn = min_pfn; | |
1403 | } | |
1404 | } | |
1405 | } | |
1406 | } | |
1407 | ||
d097a6f6 MG |
1408 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1409 | highest -= pageblock_nr_pages; | |
5a811889 | 1410 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1411 | } |
5a811889 MG |
1412 | |
1413 | cc->total_free_scanned += nr_scanned; | |
1414 | if (!page) | |
1415 | return cc->free_pfn; | |
1416 | ||
1417 | low_pfn = page_to_pfn(page); | |
1418 | fast_isolate_around(cc, low_pfn, nr_isolated); | |
1419 | return low_pfn; | |
1420 | } | |
1421 | ||
2fe86e00 | 1422 | /* |
ff9543fd MN |
1423 | * Based on information in the current compact_control, find blocks |
1424 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1425 | */ |
edc2ca61 | 1426 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1427 | { |
edc2ca61 | 1428 | struct zone *zone = cc->zone; |
ff9543fd | 1429 | struct page *page; |
c96b9e50 | 1430 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1431 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1432 | unsigned long block_end_pfn; /* end of current pageblock */ |
1433 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1434 | struct list_head *freelist = &cc->freepages; |
4fca9730 | 1435 | unsigned int stride; |
2fe86e00 | 1436 | |
5a811889 MG |
1437 | /* Try a small search of the free lists for a candidate */ |
1438 | isolate_start_pfn = fast_isolate_freepages(cc); | |
1439 | if (cc->nr_freepages) | |
1440 | goto splitmap; | |
1441 | ||
ff9543fd MN |
1442 | /* |
1443 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1444 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1445 | * zone when isolating for the first time. For looping we also need |
1446 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1447 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1448 | * For ending point, take care when isolating in last pageblock of a | |
1449 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1450 | * The low boundary is the end of the pageblock the migration scanner |
1451 | * is using. | |
ff9543fd | 1452 | */ |
e14c720e | 1453 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1454 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1455 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1456 | zone_end_pfn(zone)); | |
06b6640a | 1457 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1458 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1459 | |
ff9543fd MN |
1460 | /* |
1461 | * Isolate free pages until enough are available to migrate the | |
1462 | * pages on cc->migratepages. We stop searching if the migrate | |
1463 | * and free page scanners meet or enough free pages are isolated. | |
1464 | */ | |
f5f61a32 | 1465 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1466 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1467 | block_start_pfn -= pageblock_nr_pages, |
1468 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1469 | unsigned long nr_isolated; |
1470 | ||
f6ea3adb DR |
1471 | /* |
1472 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1473 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1474 | */ |
cb810ad2 | 1475 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1476 | cond_resched(); |
f6ea3adb | 1477 | |
7d49d886 VB |
1478 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1479 | zone); | |
1480 | if (!page) | |
ff9543fd MN |
1481 | continue; |
1482 | ||
1483 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1484 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1485 | continue; |
68e3e926 | 1486 | |
bb13ffeb MG |
1487 | /* If isolation recently failed, do not retry */ |
1488 | if (!isolation_suitable(cc, page)) | |
1489 | continue; | |
1490 | ||
e14c720e | 1491 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 MG |
1492 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
1493 | block_end_pfn, freelist, stride, false); | |
ff9543fd | 1494 | |
d097a6f6 MG |
1495 | /* Update the skip hint if the full pageblock was scanned */ |
1496 | if (isolate_start_pfn == block_end_pfn) | |
1497 | update_pageblock_skip(cc, page, block_start_pfn); | |
1498 | ||
cb2dcaf0 MG |
1499 | /* Are enough freepages isolated? */ |
1500 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1501 | if (isolate_start_pfn >= block_end_pfn) { |
1502 | /* | |
1503 | * Restart at previous pageblock if more | |
1504 | * freepages can be isolated next time. | |
1505 | */ | |
f5f61a32 VB |
1506 | isolate_start_pfn = |
1507 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1508 | } |
be976572 | 1509 | break; |
a46cbf3b | 1510 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1511 | /* |
a46cbf3b DR |
1512 | * If isolation failed early, do not continue |
1513 | * needlessly. | |
f5f61a32 | 1514 | */ |
a46cbf3b | 1515 | break; |
f5f61a32 | 1516 | } |
4fca9730 MG |
1517 | |
1518 | /* Adjust stride depending on isolation */ | |
1519 | if (nr_isolated) { | |
1520 | stride = 1; | |
1521 | continue; | |
1522 | } | |
1523 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1524 | } |
1525 | ||
7ed695e0 | 1526 | /* |
f5f61a32 VB |
1527 | * Record where the free scanner will restart next time. Either we |
1528 | * broke from the loop and set isolate_start_pfn based on the last | |
1529 | * call to isolate_freepages_block(), or we met the migration scanner | |
1530 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1531 | */ |
f5f61a32 | 1532 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1533 | |
1534 | splitmap: | |
1535 | /* __isolate_free_page() does not map the pages */ | |
1536 | split_map_pages(freelist); | |
748446bb MG |
1537 | } |
1538 | ||
1539 | /* | |
1540 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1541 | * from the isolated freelists in the block we are migrating to. | |
1542 | */ | |
1543 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1544 | unsigned long data) |
748446bb MG |
1545 | { |
1546 | struct compact_control *cc = (struct compact_control *)data; | |
1547 | struct page *freepage; | |
1548 | ||
748446bb | 1549 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1550 | isolate_freepages(cc); |
748446bb MG |
1551 | |
1552 | if (list_empty(&cc->freepages)) | |
1553 | return NULL; | |
1554 | } | |
1555 | ||
1556 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1557 | list_del(&freepage->lru); | |
1558 | cc->nr_freepages--; | |
1559 | ||
1560 | return freepage; | |
1561 | } | |
1562 | ||
1563 | /* | |
d53aea3d DR |
1564 | * This is a migrate-callback that "frees" freepages back to the isolated |
1565 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1566 | * special handling needed for NUMA. | |
1567 | */ | |
1568 | static void compaction_free(struct page *page, unsigned long data) | |
1569 | { | |
1570 | struct compact_control *cc = (struct compact_control *)data; | |
1571 | ||
1572 | list_add(&page->lru, &cc->freepages); | |
1573 | cc->nr_freepages++; | |
1574 | } | |
1575 | ||
ff9543fd MN |
1576 | /* possible outcome of isolate_migratepages */ |
1577 | typedef enum { | |
1578 | ISOLATE_ABORT, /* Abort compaction now */ | |
1579 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1580 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1581 | } isolate_migrate_t; | |
1582 | ||
5bbe3547 EM |
1583 | /* |
1584 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1585 | * compactable pages. | |
1586 | */ | |
1587 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1588 | ||
70b44595 MG |
1589 | static inline void |
1590 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1591 | { | |
1592 | if (cc->fast_start_pfn == ULONG_MAX) | |
1593 | return; | |
1594 | ||
1595 | if (!cc->fast_start_pfn) | |
1596 | cc->fast_start_pfn = pfn; | |
1597 | ||
1598 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1599 | } | |
1600 | ||
1601 | static inline unsigned long | |
1602 | reinit_migrate_pfn(struct compact_control *cc) | |
1603 | { | |
1604 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1605 | return cc->migrate_pfn; | |
1606 | ||
1607 | cc->migrate_pfn = cc->fast_start_pfn; | |
1608 | cc->fast_start_pfn = ULONG_MAX; | |
1609 | ||
1610 | return cc->migrate_pfn; | |
1611 | } | |
1612 | ||
1613 | /* | |
1614 | * Briefly search the free lists for a migration source that already has | |
1615 | * some free pages to reduce the number of pages that need migration | |
1616 | * before a pageblock is free. | |
1617 | */ | |
1618 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1619 | { | |
1620 | unsigned int limit = freelist_scan_limit(cc); | |
1621 | unsigned int nr_scanned = 0; | |
1622 | unsigned long distance; | |
1623 | unsigned long pfn = cc->migrate_pfn; | |
1624 | unsigned long high_pfn; | |
1625 | int order; | |
1626 | ||
1627 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1628 | if (cc->ignore_skip_hint) | |
1629 | return pfn; | |
1630 | ||
1631 | /* | |
1632 | * If the migrate_pfn is not at the start of a zone or the start | |
1633 | * of a pageblock then assume this is a continuation of a previous | |
1634 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1635 | */ | |
1636 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1637 | return pfn; | |
1638 | ||
1639 | /* | |
1640 | * For smaller orders, just linearly scan as the number of pages | |
1641 | * to migrate should be relatively small and does not necessarily | |
1642 | * justify freeing up a large block for a small allocation. | |
1643 | */ | |
1644 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1645 | return pfn; | |
1646 | ||
1647 | /* | |
1648 | * Only allow kcompactd and direct requests for movable pages to | |
1649 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1650 | * reduces the risk that a large movable pageblock is freed for | |
1651 | * an unmovable/reclaimable small allocation. | |
1652 | */ | |
1653 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1654 | return pfn; | |
1655 | ||
1656 | /* | |
1657 | * When starting the migration scanner, pick any pageblock within the | |
1658 | * first half of the search space. Otherwise try and pick a pageblock | |
1659 | * within the first eighth to reduce the chances that a migration | |
1660 | * target later becomes a source. | |
1661 | */ | |
1662 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1663 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1664 | distance >>= 2; | |
1665 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1666 | ||
1667 | for (order = cc->order - 1; | |
1668 | order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; | |
1669 | order--) { | |
1670 | struct free_area *area = &cc->zone->free_area[order]; | |
1671 | struct list_head *freelist; | |
1672 | unsigned long flags; | |
1673 | struct page *freepage; | |
1674 | ||
1675 | if (!area->nr_free) | |
1676 | continue; | |
1677 | ||
1678 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1679 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1680 | list_for_each_entry(freepage, freelist, lru) { | |
1681 | unsigned long free_pfn; | |
1682 | ||
1683 | nr_scanned++; | |
1684 | free_pfn = page_to_pfn(freepage); | |
1685 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1686 | /* |
1687 | * Avoid if skipped recently. Ideally it would | |
1688 | * move to the tail but even safe iteration of | |
1689 | * the list assumes an entry is deleted, not | |
1690 | * reordered. | |
1691 | */ | |
1692 | if (get_pageblock_skip(freepage)) { | |
1693 | if (list_is_last(freelist, &freepage->lru)) | |
1694 | break; | |
1695 | ||
1696 | continue; | |
1697 | } | |
1698 | ||
1699 | /* Reorder to so a future search skips recent pages */ | |
1700 | move_freelist_tail(freelist, freepage); | |
1701 | ||
e380bebe | 1702 | update_fast_start_pfn(cc, free_pfn); |
70b44595 MG |
1703 | pfn = pageblock_start_pfn(free_pfn); |
1704 | cc->fast_search_fail = 0; | |
1705 | set_pageblock_skip(freepage); | |
1706 | break; | |
1707 | } | |
1708 | ||
1709 | if (nr_scanned >= limit) { | |
1710 | cc->fast_search_fail++; | |
1711 | move_freelist_tail(freelist, freepage); | |
1712 | break; | |
1713 | } | |
1714 | } | |
1715 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1716 | } | |
1717 | ||
1718 | cc->total_migrate_scanned += nr_scanned; | |
1719 | ||
1720 | /* | |
1721 | * If fast scanning failed then use a cached entry for a page block | |
1722 | * that had free pages as the basis for starting a linear scan. | |
1723 | */ | |
1724 | if (pfn == cc->migrate_pfn) | |
1725 | pfn = reinit_migrate_pfn(cc); | |
1726 | ||
1727 | return pfn; | |
1728 | } | |
1729 | ||
ff9543fd | 1730 | /* |
edc2ca61 VB |
1731 | * Isolate all pages that can be migrated from the first suitable block, |
1732 | * starting at the block pointed to by the migrate scanner pfn within | |
1733 | * compact_control. | |
ff9543fd MN |
1734 | */ |
1735 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1736 | struct compact_control *cc) | |
1737 | { | |
e1409c32 JK |
1738 | unsigned long block_start_pfn; |
1739 | unsigned long block_end_pfn; | |
1740 | unsigned long low_pfn; | |
edc2ca61 VB |
1741 | struct page *page; |
1742 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1743 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1744 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 1745 | bool fast_find_block; |
ff9543fd | 1746 | |
edc2ca61 VB |
1747 | /* |
1748 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
1749 | * initialized by compact_zone(). The first failure will use |
1750 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 1751 | */ |
70b44595 | 1752 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 1753 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1754 | if (block_start_pfn < zone->zone_start_pfn) |
1755 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd | 1756 | |
70b44595 MG |
1757 | /* |
1758 | * fast_find_migrateblock marks a pageblock skipped so to avoid | |
1759 | * the isolation_suitable check below, check whether the fast | |
1760 | * search was successful. | |
1761 | */ | |
1762 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
1763 | ||
ff9543fd | 1764 | /* Only scan within a pageblock boundary */ |
06b6640a | 1765 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1766 | |
edc2ca61 VB |
1767 | /* |
1768 | * Iterate over whole pageblocks until we find the first suitable. | |
1769 | * Do not cross the free scanner. | |
1770 | */ | |
e1409c32 | 1771 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 1772 | fast_find_block = false, |
e1409c32 JK |
1773 | low_pfn = block_end_pfn, |
1774 | block_start_pfn = block_end_pfn, | |
1775 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1776 | |
edc2ca61 VB |
1777 | /* |
1778 | * This can potentially iterate a massively long zone with | |
1779 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 1780 | * need to schedule. |
edc2ca61 | 1781 | */ |
cb810ad2 | 1782 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1783 | cond_resched(); |
ff9543fd | 1784 | |
e1409c32 JK |
1785 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1786 | zone); | |
7d49d886 | 1787 | if (!page) |
edc2ca61 VB |
1788 | continue; |
1789 | ||
e380bebe MG |
1790 | /* |
1791 | * If isolation recently failed, do not retry. Only check the | |
1792 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
1793 | * to be visited multiple times. Assume skip was checked | |
1794 | * before making it "skip" so other compaction instances do | |
1795 | * not scan the same block. | |
1796 | */ | |
1797 | if (IS_ALIGNED(low_pfn, pageblock_nr_pages) && | |
1798 | !fast_find_block && !isolation_suitable(cc, page)) | |
edc2ca61 VB |
1799 | continue; |
1800 | ||
1801 | /* | |
9bebefd5 MG |
1802 | * For async compaction, also only scan in MOVABLE blocks |
1803 | * without huge pages. Async compaction is optimistic to see | |
1804 | * if the minimum amount of work satisfies the allocation. | |
1805 | * The cached PFN is updated as it's possible that all | |
1806 | * remaining blocks between source and target are unsuitable | |
1807 | * and the compaction scanners fail to meet. | |
edc2ca61 | 1808 | */ |
9bebefd5 MG |
1809 | if (!suitable_migration_source(cc, page)) { |
1810 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 1811 | continue; |
9bebefd5 | 1812 | } |
edc2ca61 VB |
1813 | |
1814 | /* Perform the isolation */ | |
e1409c32 JK |
1815 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1816 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1817 | |
cb2dcaf0 | 1818 | if (!low_pfn) |
edc2ca61 VB |
1819 | return ISOLATE_ABORT; |
1820 | ||
1821 | /* | |
1822 | * Either we isolated something and proceed with migration. Or | |
1823 | * we failed and compact_zone should decide if we should | |
1824 | * continue or not. | |
1825 | */ | |
1826 | break; | |
1827 | } | |
1828 | ||
f2849aa0 VB |
1829 | /* Record where migration scanner will be restarted. */ |
1830 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1831 | |
edc2ca61 | 1832 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1833 | } |
1834 | ||
21c527a3 YB |
1835 | /* |
1836 | * order == -1 is expected when compacting via | |
1837 | * /proc/sys/vm/compact_memory | |
1838 | */ | |
1839 | static inline bool is_via_compact_memory(int order) | |
1840 | { | |
1841 | return order == -1; | |
1842 | } | |
1843 | ||
40cacbcb | 1844 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 1845 | { |
8fb74b9f | 1846 | unsigned int order; |
d39773a0 | 1847 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 1848 | int ret; |
748446bb | 1849 | |
753341a4 | 1850 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1851 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1852 | /* Let the next compaction start anew. */ |
40cacbcb | 1853 | reset_cached_positions(cc->zone); |
55b7c4c9 | 1854 | |
62997027 MG |
1855 | /* |
1856 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1857 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1858 | * flag itself as the decision to be clear should be directly |
1859 | * based on an allocation request. | |
1860 | */ | |
accf6242 | 1861 | if (cc->direct_compaction) |
40cacbcb | 1862 | cc->zone->compact_blockskip_flush = true; |
62997027 | 1863 | |
c8f7de0b MH |
1864 | if (cc->whole_zone) |
1865 | return COMPACT_COMPLETE; | |
1866 | else | |
1867 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1868 | } |
748446bb | 1869 | |
21c527a3 | 1870 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1871 | return COMPACT_CONTINUE; |
1872 | ||
efe771c7 MG |
1873 | /* |
1874 | * Always finish scanning a pageblock to reduce the possibility of | |
1875 | * fallbacks in the future. This is particularly important when | |
1876 | * migration source is unmovable/reclaimable but it's not worth | |
1877 | * special casing. | |
1878 | */ | |
1879 | if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1880 | return COMPACT_CONTINUE; | |
baf6a9a1 | 1881 | |
56de7263 | 1882 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 1883 | ret = COMPACT_NO_SUITABLE_PAGE; |
8fb74b9f | 1884 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 1885 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 1886 | bool can_steal; |
8fb74b9f MG |
1887 | |
1888 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1889 | if (!list_empty(&area->free_list[migratetype])) |
cf378319 | 1890 | return COMPACT_SUCCESS; |
8fb74b9f | 1891 | |
2149cdae JK |
1892 | #ifdef CONFIG_CMA |
1893 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1894 | if (migratetype == MIGRATE_MOVABLE && | |
1895 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
cf378319 | 1896 | return COMPACT_SUCCESS; |
2149cdae JK |
1897 | #endif |
1898 | /* | |
1899 | * Job done if allocation would steal freepages from | |
1900 | * other migratetype buddy lists. | |
1901 | */ | |
1902 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
1903 | true, &can_steal) != -1) { |
1904 | ||
1905 | /* movable pages are OK in any pageblock */ | |
1906 | if (migratetype == MIGRATE_MOVABLE) | |
1907 | return COMPACT_SUCCESS; | |
1908 | ||
1909 | /* | |
1910 | * We are stealing for a non-movable allocation. Make | |
1911 | * sure we finish compacting the current pageblock | |
1912 | * first so it is as free as possible and we won't | |
1913 | * have to steal another one soon. This only applies | |
1914 | * to sync compaction, as async compaction operates | |
1915 | * on pageblocks of the same migratetype. | |
1916 | */ | |
1917 | if (cc->mode == MIGRATE_ASYNC || | |
1918 | IS_ALIGNED(cc->migrate_pfn, | |
1919 | pageblock_nr_pages)) { | |
1920 | return COMPACT_SUCCESS; | |
1921 | } | |
1922 | ||
cb2dcaf0 MG |
1923 | ret = COMPACT_CONTINUE; |
1924 | break; | |
baf6a9a1 | 1925 | } |
56de7263 MG |
1926 | } |
1927 | ||
cb2dcaf0 MG |
1928 | if (cc->contended || fatal_signal_pending(current)) |
1929 | ret = COMPACT_CONTENDED; | |
1930 | ||
1931 | return ret; | |
837d026d JK |
1932 | } |
1933 | ||
40cacbcb | 1934 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
1935 | { |
1936 | int ret; | |
1937 | ||
40cacbcb MG |
1938 | ret = __compact_finished(cc); |
1939 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
1940 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
1941 | ret = COMPACT_CONTINUE; | |
1942 | ||
1943 | return ret; | |
748446bb MG |
1944 | } |
1945 | ||
3e7d3449 MG |
1946 | /* |
1947 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1948 | * Returns | |
1949 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
cf378319 | 1950 | * COMPACT_SUCCESS - If the allocation would succeed without compaction |
3e7d3449 MG |
1951 | * COMPACT_CONTINUE - If compaction should run now |
1952 | */ | |
ea7ab982 | 1953 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1954 | unsigned int alloc_flags, |
86a294a8 MH |
1955 | int classzone_idx, |
1956 | unsigned long wmark_target) | |
3e7d3449 | 1957 | { |
3e7d3449 MG |
1958 | unsigned long watermark; |
1959 | ||
21c527a3 | 1960 | if (is_via_compact_memory(order)) |
3957c776 MH |
1961 | return COMPACT_CONTINUE; |
1962 | ||
a9214443 | 1963 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
1964 | /* |
1965 | * If watermarks for high-order allocation are already met, there | |
1966 | * should be no need for compaction at all. | |
1967 | */ | |
1968 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1969 | alloc_flags)) | |
cf378319 | 1970 | return COMPACT_SUCCESS; |
ebff3980 | 1971 | |
3e7d3449 | 1972 | /* |
9861a62c | 1973 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
1974 | * isolate free pages for migration targets. This means that the |
1975 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1976 | * the check in __isolate_free_page(). We don't use the direct | |
1977 | * compactor's alloc_flags, as they are not relevant for freepage | |
1978 | * isolation. We however do use the direct compactor's classzone_idx to | |
1979 | * skip over zones where lowmem reserves would prevent allocation even | |
1980 | * if compaction succeeds. | |
8348faf9 VB |
1981 | * For costly orders, we require low watermark instead of min for |
1982 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
1983 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
1984 | * suitable migration targets | |
3e7d3449 | 1985 | */ |
8348faf9 VB |
1986 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
1987 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
1988 | watermark += compact_gap(order); | |
86a294a8 | 1989 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
d883c6cf | 1990 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
1991 | return COMPACT_SKIPPED; |
1992 | ||
cc5c9f09 VB |
1993 | return COMPACT_CONTINUE; |
1994 | } | |
1995 | ||
1996 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
1997 | unsigned int alloc_flags, | |
1998 | int classzone_idx) | |
1999 | { | |
2000 | enum compact_result ret; | |
2001 | int fragindex; | |
2002 | ||
2003 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
2004 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
2005 | /* |
2006 | * fragmentation index determines if allocation failures are due to | |
2007 | * low memory or external fragmentation | |
2008 | * | |
ebff3980 VB |
2009 | * index of -1000 would imply allocations might succeed depending on |
2010 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2011 | * index towards 0 implies failure is due to lack of memory |
2012 | * index towards 1000 implies failure is due to fragmentation | |
2013 | * | |
20311420 VB |
2014 | * Only compact if a failure would be due to fragmentation. Also |
2015 | * ignore fragindex for non-costly orders where the alternative to | |
2016 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2017 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2018 | * excessive compaction for costly orders, but it should not be at the | |
2019 | * expense of system stability. | |
3e7d3449 | 2020 | */ |
20311420 | 2021 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
2022 | fragindex = fragmentation_index(zone, order); |
2023 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
2024 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
2025 | } | |
837d026d | 2026 | |
837d026d JK |
2027 | trace_mm_compaction_suitable(zone, order, ret); |
2028 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
2029 | ret = COMPACT_SKIPPED; | |
2030 | ||
2031 | return ret; | |
2032 | } | |
2033 | ||
86a294a8 MH |
2034 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2035 | int alloc_flags) | |
2036 | { | |
2037 | struct zone *zone; | |
2038 | struct zoneref *z; | |
2039 | ||
2040 | /* | |
2041 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2042 | * retrying the reclaim. | |
2043 | */ | |
2044 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
2045 | ac->nodemask) { | |
2046 | unsigned long available; | |
2047 | enum compact_result compact_result; | |
2048 | ||
2049 | /* | |
2050 | * Do not consider all the reclaimable memory because we do not | |
2051 | * want to trash just for a single high order allocation which | |
2052 | * is even not guaranteed to appear even if __compaction_suitable | |
2053 | * is happy about the watermark check. | |
2054 | */ | |
5a1c84b4 | 2055 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
2056 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
2057 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
2058 | ac_classzone_idx(ac), available); | |
cc5c9f09 | 2059 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
2060 | return true; |
2061 | } | |
2062 | ||
2063 | return false; | |
2064 | } | |
2065 | ||
5e1f0f09 MG |
2066 | static enum compact_result |
2067 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2068 | { |
ea7ab982 | 2069 | enum compact_result ret; |
40cacbcb MG |
2070 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2071 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2072 | unsigned long last_migrated_pfn; |
e0b9daeb | 2073 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2074 | bool update_cached; |
748446bb | 2075 | |
d39773a0 | 2076 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
40cacbcb | 2077 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
ebff3980 | 2078 | cc->classzone_idx); |
c46649de | 2079 | /* Compaction is likely to fail */ |
cf378319 | 2080 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 2081 | return ret; |
c46649de MH |
2082 | |
2083 | /* huh, compaction_suitable is returning something unexpected */ | |
2084 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 2085 | |
d3132e4b VB |
2086 | /* |
2087 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2088 | * is about to be retried after being deferred. |
d3132e4b | 2089 | */ |
40cacbcb MG |
2090 | if (compaction_restarting(cc->zone, cc->order)) |
2091 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2092 | |
c89511ab MG |
2093 | /* |
2094 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2095 | * information on where the scanners should start (unless we explicitly |
2096 | * want to compact the whole zone), but check that it is initialised | |
2097 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2098 | */ |
70b44595 | 2099 | cc->fast_start_pfn = 0; |
06ed2998 | 2100 | if (cc->whole_zone) { |
c89511ab | 2101 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2102 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2103 | } else { | |
40cacbcb MG |
2104 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2105 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2106 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2107 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2108 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2109 | } |
2110 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2111 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2112 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2113 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2114 | } |
c8f7de0b | 2115 | |
e332f741 | 2116 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2117 | cc->whole_zone = true; |
2118 | } | |
c8f7de0b | 2119 | |
566e54e1 | 2120 | last_migrated_pfn = 0; |
748446bb | 2121 | |
8854c55f MG |
2122 | /* |
2123 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2124 | * the basis that some migrations will fail in ASYNC mode. However, | |
2125 | * if the cached PFNs match and pageblocks are skipped due to having | |
2126 | * no isolation candidates, then the sync state does not matter. | |
2127 | * Until a pageblock with isolation candidates is found, keep the | |
2128 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2129 | */ | |
2130 | update_cached = !sync && | |
2131 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2132 | ||
16c4a097 JK |
2133 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
2134 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 2135 | |
748446bb MG |
2136 | migrate_prep_local(); |
2137 | ||
40cacbcb | 2138 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2139 | int err; |
566e54e1 | 2140 | unsigned long start_pfn = cc->migrate_pfn; |
748446bb | 2141 | |
804d3121 MG |
2142 | /* |
2143 | * Avoid multiple rescans which can happen if a page cannot be | |
2144 | * isolated (dirty/writeback in async mode) or if the migrated | |
2145 | * pages are being allocated before the pageblock is cleared. | |
2146 | * The first rescan will capture the entire pageblock for | |
2147 | * migration. If it fails, it'll be marked skip and scanning | |
2148 | * will proceed as normal. | |
2149 | */ | |
2150 | cc->rescan = false; | |
2151 | if (pageblock_start_pfn(last_migrated_pfn) == | |
2152 | pageblock_start_pfn(start_pfn)) { | |
2153 | cc->rescan = true; | |
2154 | } | |
2155 | ||
40cacbcb | 2156 | switch (isolate_migratepages(cc->zone, cc)) { |
f9e35b3b | 2157 | case ISOLATE_ABORT: |
2d1e1041 | 2158 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2159 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2160 | cc->nr_migratepages = 0; |
566e54e1 | 2161 | last_migrated_pfn = 0; |
f9e35b3b MG |
2162 | goto out; |
2163 | case ISOLATE_NONE: | |
8854c55f MG |
2164 | if (update_cached) { |
2165 | cc->zone->compact_cached_migrate_pfn[1] = | |
2166 | cc->zone->compact_cached_migrate_pfn[0]; | |
2167 | } | |
2168 | ||
fdaf7f5c VB |
2169 | /* |
2170 | * We haven't isolated and migrated anything, but | |
2171 | * there might still be unflushed migrations from | |
2172 | * previous cc->order aligned block. | |
2173 | */ | |
2174 | goto check_drain; | |
f9e35b3b | 2175 | case ISOLATE_SUCCESS: |
8854c55f | 2176 | update_cached = false; |
566e54e1 | 2177 | last_migrated_pfn = start_pfn; |
f9e35b3b MG |
2178 | ; |
2179 | } | |
748446bb | 2180 | |
d53aea3d | 2181 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2182 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 2183 | MR_COMPACTION); |
748446bb | 2184 | |
f8c9301f VB |
2185 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
2186 | &cc->migratepages); | |
748446bb | 2187 | |
f8c9301f VB |
2188 | /* All pages were either migrated or will be released */ |
2189 | cc->nr_migratepages = 0; | |
9d502c1c | 2190 | if (err) { |
5733c7d1 | 2191 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2192 | /* |
2193 | * migrate_pages() may return -ENOMEM when scanners meet | |
2194 | * and we want compact_finished() to detect it | |
2195 | */ | |
f2849aa0 | 2196 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2197 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2198 | goto out; |
2199 | } | |
fdd048e1 VB |
2200 | /* |
2201 | * We failed to migrate at least one page in the current | |
2202 | * order-aligned block, so skip the rest of it. | |
2203 | */ | |
2204 | if (cc->direct_compaction && | |
2205 | (cc->mode == MIGRATE_ASYNC)) { | |
2206 | cc->migrate_pfn = block_end_pfn( | |
2207 | cc->migrate_pfn - 1, cc->order); | |
2208 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 2209 | last_migrated_pfn = 0; |
fdd048e1 | 2210 | } |
748446bb | 2211 | } |
fdaf7f5c | 2212 | |
fdaf7f5c VB |
2213 | check_drain: |
2214 | /* | |
2215 | * Has the migration scanner moved away from the previous | |
2216 | * cc->order aligned block where we migrated from? If yes, | |
2217 | * flush the pages that were freed, so that they can merge and | |
2218 | * compact_finished() can detect immediately if allocation | |
2219 | * would succeed. | |
2220 | */ | |
566e54e1 | 2221 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c VB |
2222 | int cpu; |
2223 | unsigned long current_block_start = | |
06b6640a | 2224 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2225 | |
566e54e1 | 2226 | if (last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
2227 | cpu = get_cpu(); |
2228 | lru_add_drain_cpu(cpu); | |
40cacbcb | 2229 | drain_local_pages(cc->zone); |
fdaf7f5c VB |
2230 | put_cpu(); |
2231 | /* No more flushing until we migrate again */ | |
566e54e1 | 2232 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2233 | } |
2234 | } | |
2235 | ||
5e1f0f09 MG |
2236 | /* Stop if a page has been captured */ |
2237 | if (capc && capc->page) { | |
2238 | ret = COMPACT_SUCCESS; | |
2239 | break; | |
2240 | } | |
748446bb MG |
2241 | } |
2242 | ||
f9e35b3b | 2243 | out: |
6bace090 VB |
2244 | /* |
2245 | * Release free pages and update where the free scanner should restart, | |
2246 | * so we don't leave any returned pages behind in the next attempt. | |
2247 | */ | |
2248 | if (cc->nr_freepages > 0) { | |
2249 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2250 | ||
2251 | cc->nr_freepages = 0; | |
2252 | VM_BUG_ON(free_pfn == 0); | |
2253 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2254 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2255 | /* |
2256 | * Only go back, not forward. The cached pfn might have been | |
2257 | * already reset to zone end in compact_finished() | |
2258 | */ | |
40cacbcb MG |
2259 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2260 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2261 | } |
748446bb | 2262 | |
7f354a54 DR |
2263 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2264 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2265 | ||
16c4a097 JK |
2266 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
2267 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 2268 | |
748446bb MG |
2269 | return ret; |
2270 | } | |
76ab0f53 | 2271 | |
ea7ab982 | 2272 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2273 | gfp_t gfp_mask, enum compact_priority prio, |
5e1f0f09 MG |
2274 | unsigned int alloc_flags, int classzone_idx, |
2275 | struct page **capture) | |
56de7263 | 2276 | { |
ea7ab982 | 2277 | enum compact_result ret; |
56de7263 MG |
2278 | struct compact_control cc = { |
2279 | .nr_freepages = 0, | |
2280 | .nr_migratepages = 0, | |
7f354a54 DR |
2281 | .total_migrate_scanned = 0, |
2282 | .total_free_scanned = 0, | |
56de7263 | 2283 | .order = order, |
dbe2d4e4 | 2284 | .search_order = order, |
6d7ce559 | 2285 | .gfp_mask = gfp_mask, |
56de7263 | 2286 | .zone = zone, |
a5508cd8 VB |
2287 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2288 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 VB |
2289 | .alloc_flags = alloc_flags, |
2290 | .classzone_idx = classzone_idx, | |
accf6242 | 2291 | .direct_compaction = true, |
a8e025e5 | 2292 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2293 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2294 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2295 | }; |
5e1f0f09 MG |
2296 | struct capture_control capc = { |
2297 | .cc = &cc, | |
2298 | .page = NULL, | |
2299 | }; | |
2300 | ||
2301 | if (capture) | |
2302 | current->capture_control = &capc; | |
56de7263 MG |
2303 | INIT_LIST_HEAD(&cc.freepages); |
2304 | INIT_LIST_HEAD(&cc.migratepages); | |
2305 | ||
5e1f0f09 | 2306 | ret = compact_zone(&cc, &capc); |
e64c5237 SL |
2307 | |
2308 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2309 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2310 | ||
5e1f0f09 MG |
2311 | *capture = capc.page; |
2312 | current->capture_control = NULL; | |
2313 | ||
e64c5237 | 2314 | return ret; |
56de7263 MG |
2315 | } |
2316 | ||
5e771905 MG |
2317 | int sysctl_extfrag_threshold = 500; |
2318 | ||
56de7263 MG |
2319 | /** |
2320 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2321 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2322 | * @order: The order of the current allocation |
2323 | * @alloc_flags: The allocation flags of the current allocation | |
2324 | * @ac: The context of current allocation | |
112d2d29 | 2325 | * @prio: Determines how hard direct compaction should try to succeed |
56de7263 MG |
2326 | * |
2327 | * This is the main entry point for direct page compaction. | |
2328 | */ | |
ea7ab982 | 2329 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2330 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2331 | enum compact_priority prio, struct page **capture) |
56de7263 | 2332 | { |
56de7263 | 2333 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
2334 | struct zoneref *z; |
2335 | struct zone *zone; | |
1d4746d3 | 2336 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2337 | |
73e64c51 MH |
2338 | /* |
2339 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2340 | * tricky context because the migration might require IO | |
2341 | */ | |
2342 | if (!may_perform_io) | |
53853e2d | 2343 | return COMPACT_SKIPPED; |
56de7263 | 2344 | |
a5508cd8 | 2345 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2346 | |
56de7263 | 2347 | /* Compact each zone in the list */ |
1a6d53a1 VB |
2348 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
2349 | ac->nodemask) { | |
ea7ab982 | 2350 | enum compact_result status; |
56de7263 | 2351 | |
a8e025e5 VB |
2352 | if (prio > MIN_COMPACT_PRIORITY |
2353 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2354 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2355 | continue; |
1d4746d3 | 2356 | } |
53853e2d | 2357 | |
a5508cd8 | 2358 | status = compact_zone_order(zone, order, gfp_mask, prio, |
5e1f0f09 | 2359 | alloc_flags, ac_classzone_idx(ac), capture); |
56de7263 MG |
2360 | rc = max(status, rc); |
2361 | ||
7ceb009a VB |
2362 | /* The allocation should succeed, stop compacting */ |
2363 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2364 | /* |
2365 | * We think the allocation will succeed in this zone, | |
2366 | * but it is not certain, hence the false. The caller | |
2367 | * will repeat this with true if allocation indeed | |
2368 | * succeeds in this zone. | |
2369 | */ | |
2370 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2371 | |
c3486f53 | 2372 | break; |
1f9efdef VB |
2373 | } |
2374 | ||
a5508cd8 | 2375 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2376 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2377 | /* |
2378 | * We think that allocation won't succeed in this zone | |
2379 | * so we defer compaction there. If it ends up | |
2380 | * succeeding after all, it will be reset. | |
2381 | */ | |
2382 | defer_compaction(zone, order); | |
1f9efdef VB |
2383 | |
2384 | /* | |
2385 | * We might have stopped compacting due to need_resched() in | |
2386 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2387 | * case do not try further zones |
1f9efdef | 2388 | */ |
c3486f53 VB |
2389 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2390 | || fatal_signal_pending(current)) | |
2391 | break; | |
56de7263 MG |
2392 | } |
2393 | ||
2394 | return rc; | |
2395 | } | |
2396 | ||
2397 | ||
76ab0f53 | 2398 | /* Compact all zones within a node */ |
791cae96 | 2399 | static void compact_node(int nid) |
76ab0f53 | 2400 | { |
791cae96 | 2401 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 2402 | int zoneid; |
76ab0f53 | 2403 | struct zone *zone; |
791cae96 VB |
2404 | struct compact_control cc = { |
2405 | .order = -1, | |
7f354a54 DR |
2406 | .total_migrate_scanned = 0, |
2407 | .total_free_scanned = 0, | |
791cae96 VB |
2408 | .mode = MIGRATE_SYNC, |
2409 | .ignore_skip_hint = true, | |
2410 | .whole_zone = true, | |
73e64c51 | 2411 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
2412 | }; |
2413 | ||
76ab0f53 | 2414 | |
76ab0f53 | 2415 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
2416 | |
2417 | zone = &pgdat->node_zones[zoneid]; | |
2418 | if (!populated_zone(zone)) | |
2419 | continue; | |
2420 | ||
791cae96 VB |
2421 | cc.nr_freepages = 0; |
2422 | cc.nr_migratepages = 0; | |
2423 | cc.zone = zone; | |
2424 | INIT_LIST_HEAD(&cc.freepages); | |
2425 | INIT_LIST_HEAD(&cc.migratepages); | |
76ab0f53 | 2426 | |
5e1f0f09 | 2427 | compact_zone(&cc, NULL); |
75469345 | 2428 | |
791cae96 VB |
2429 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2430 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 2431 | } |
76ab0f53 MG |
2432 | } |
2433 | ||
2434 | /* Compact all nodes in the system */ | |
7964c06d | 2435 | static void compact_nodes(void) |
76ab0f53 MG |
2436 | { |
2437 | int nid; | |
2438 | ||
8575ec29 HD |
2439 | /* Flush pending updates to the LRU lists */ |
2440 | lru_add_drain_all(); | |
2441 | ||
76ab0f53 MG |
2442 | for_each_online_node(nid) |
2443 | compact_node(nid); | |
76ab0f53 MG |
2444 | } |
2445 | ||
2446 | /* The written value is actually unused, all memory is compacted */ | |
2447 | int sysctl_compact_memory; | |
2448 | ||
fec4eb2c YB |
2449 | /* |
2450 | * This is the entry point for compacting all nodes via | |
2451 | * /proc/sys/vm/compact_memory | |
2452 | */ | |
76ab0f53 MG |
2453 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
2454 | void __user *buffer, size_t *length, loff_t *ppos) | |
2455 | { | |
2456 | if (write) | |
7964c06d | 2457 | compact_nodes(); |
76ab0f53 MG |
2458 | |
2459 | return 0; | |
2460 | } | |
ed4a6d7f MG |
2461 | |
2462 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
74e77fb9 | 2463 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 2464 | struct device_attribute *attr, |
ed4a6d7f MG |
2465 | const char *buf, size_t count) |
2466 | { | |
8575ec29 HD |
2467 | int nid = dev->id; |
2468 | ||
2469 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2470 | /* Flush pending updates to the LRU lists */ | |
2471 | lru_add_drain_all(); | |
2472 | ||
2473 | compact_node(nid); | |
2474 | } | |
ed4a6d7f MG |
2475 | |
2476 | return count; | |
2477 | } | |
0825a6f9 | 2478 | static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node); |
ed4a6d7f MG |
2479 | |
2480 | int compaction_register_node(struct node *node) | |
2481 | { | |
10fbcf4c | 2482 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2483 | } |
2484 | ||
2485 | void compaction_unregister_node(struct node *node) | |
2486 | { | |
10fbcf4c | 2487 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2488 | } |
2489 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2490 | |
698b1b30 VB |
2491 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2492 | { | |
172400c6 | 2493 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
2494 | } |
2495 | ||
2496 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2497 | { | |
2498 | int zoneid; | |
2499 | struct zone *zone; | |
2500 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
2501 | ||
6cd9dc3e | 2502 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
2503 | zone = &pgdat->node_zones[zoneid]; |
2504 | ||
2505 | if (!populated_zone(zone)) | |
2506 | continue; | |
2507 | ||
2508 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
2509 | classzone_idx) == COMPACT_CONTINUE) | |
2510 | return true; | |
2511 | } | |
2512 | ||
2513 | return false; | |
2514 | } | |
2515 | ||
2516 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2517 | { | |
2518 | /* | |
2519 | * With no special task, compact all zones so that a page of requested | |
2520 | * order is allocatable. | |
2521 | */ | |
2522 | int zoneid; | |
2523 | struct zone *zone; | |
2524 | struct compact_control cc = { | |
2525 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 2526 | .search_order = pgdat->kcompactd_max_order, |
7f354a54 DR |
2527 | .total_migrate_scanned = 0, |
2528 | .total_free_scanned = 0, | |
698b1b30 VB |
2529 | .classzone_idx = pgdat->kcompactd_classzone_idx, |
2530 | .mode = MIGRATE_SYNC_LIGHT, | |
a0647dc9 | 2531 | .ignore_skip_hint = false, |
73e64c51 | 2532 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2533 | }; |
698b1b30 VB |
2534 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
2535 | cc.classzone_idx); | |
7f354a54 | 2536 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2537 | |
6cd9dc3e | 2538 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
2539 | int status; |
2540 | ||
2541 | zone = &pgdat->node_zones[zoneid]; | |
2542 | if (!populated_zone(zone)) | |
2543 | continue; | |
2544 | ||
2545 | if (compaction_deferred(zone, cc.order)) | |
2546 | continue; | |
2547 | ||
2548 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
2549 | COMPACT_CONTINUE) | |
2550 | continue; | |
2551 | ||
2552 | cc.nr_freepages = 0; | |
2553 | cc.nr_migratepages = 0; | |
7f354a54 DR |
2554 | cc.total_migrate_scanned = 0; |
2555 | cc.total_free_scanned = 0; | |
698b1b30 VB |
2556 | cc.zone = zone; |
2557 | INIT_LIST_HEAD(&cc.freepages); | |
2558 | INIT_LIST_HEAD(&cc.migratepages); | |
2559 | ||
172400c6 VB |
2560 | if (kthread_should_stop()) |
2561 | return; | |
5e1f0f09 | 2562 | status = compact_zone(&cc, NULL); |
698b1b30 | 2563 | |
7ceb009a | 2564 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 2565 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 2566 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
2567 | /* |
2568 | * Buddy pages may become stranded on pcps that could | |
2569 | * otherwise coalesce on the zone's free area for | |
2570 | * order >= cc.order. This is ratelimited by the | |
2571 | * upcoming deferral. | |
2572 | */ | |
2573 | drain_all_pages(zone); | |
2574 | ||
698b1b30 VB |
2575 | /* |
2576 | * We use sync migration mode here, so we defer like | |
2577 | * sync direct compaction does. | |
2578 | */ | |
2579 | defer_compaction(zone, cc.order); | |
2580 | } | |
2581 | ||
7f354a54 DR |
2582 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2583 | cc.total_migrate_scanned); | |
2584 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2585 | cc.total_free_scanned); | |
2586 | ||
698b1b30 VB |
2587 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2588 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2589 | } | |
2590 | ||
2591 | /* | |
2592 | * Regardless of success, we are done until woken up next. But remember | |
2593 | * the requested order/classzone_idx in case it was higher/tighter than | |
2594 | * our current ones | |
2595 | */ | |
2596 | if (pgdat->kcompactd_max_order <= cc.order) | |
2597 | pgdat->kcompactd_max_order = 0; | |
2598 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
2599 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2600 | } | |
2601 | ||
2602 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
2603 | { | |
2604 | if (!order) | |
2605 | return; | |
2606 | ||
2607 | if (pgdat->kcompactd_max_order < order) | |
2608 | pgdat->kcompactd_max_order = order; | |
2609 | ||
2610 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2611 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2612 | ||
6818600f DB |
2613 | /* |
2614 | * Pairs with implicit barrier in wait_event_freezable() | |
2615 | * such that wakeups are not missed. | |
2616 | */ | |
2617 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2618 | return; |
2619 | ||
2620 | if (!kcompactd_node_suitable(pgdat)) | |
2621 | return; | |
2622 | ||
2623 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2624 | classzone_idx); | |
2625 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2626 | } | |
2627 | ||
2628 | /* | |
2629 | * The background compaction daemon, started as a kernel thread | |
2630 | * from the init process. | |
2631 | */ | |
2632 | static int kcompactd(void *p) | |
2633 | { | |
2634 | pg_data_t *pgdat = (pg_data_t*)p; | |
2635 | struct task_struct *tsk = current; | |
2636 | ||
2637 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2638 | ||
2639 | if (!cpumask_empty(cpumask)) | |
2640 | set_cpus_allowed_ptr(tsk, cpumask); | |
2641 | ||
2642 | set_freezable(); | |
2643 | ||
2644 | pgdat->kcompactd_max_order = 0; | |
2645 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2646 | ||
2647 | while (!kthread_should_stop()) { | |
eb414681 JW |
2648 | unsigned long pflags; |
2649 | ||
698b1b30 VB |
2650 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
2651 | wait_event_freezable(pgdat->kcompactd_wait, | |
2652 | kcompactd_work_requested(pgdat)); | |
2653 | ||
eb414681 | 2654 | psi_memstall_enter(&pflags); |
698b1b30 | 2655 | kcompactd_do_work(pgdat); |
eb414681 | 2656 | psi_memstall_leave(&pflags); |
698b1b30 VB |
2657 | } |
2658 | ||
2659 | return 0; | |
2660 | } | |
2661 | ||
2662 | /* | |
2663 | * This kcompactd start function will be called by init and node-hot-add. | |
2664 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2665 | */ | |
2666 | int kcompactd_run(int nid) | |
2667 | { | |
2668 | pg_data_t *pgdat = NODE_DATA(nid); | |
2669 | int ret = 0; | |
2670 | ||
2671 | if (pgdat->kcompactd) | |
2672 | return 0; | |
2673 | ||
2674 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2675 | if (IS_ERR(pgdat->kcompactd)) { | |
2676 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2677 | ret = PTR_ERR(pgdat->kcompactd); | |
2678 | pgdat->kcompactd = NULL; | |
2679 | } | |
2680 | return ret; | |
2681 | } | |
2682 | ||
2683 | /* | |
2684 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2685 | * hold mem_hotplug_begin/end(). | |
2686 | */ | |
2687 | void kcompactd_stop(int nid) | |
2688 | { | |
2689 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2690 | ||
2691 | if (kcompactd) { | |
2692 | kthread_stop(kcompactd); | |
2693 | NODE_DATA(nid)->kcompactd = NULL; | |
2694 | } | |
2695 | } | |
2696 | ||
2697 | /* | |
2698 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2699 | * not required for correctness. So if the last cpu in a node goes | |
2700 | * away, we get changed to run anywhere: as the first one comes back, | |
2701 | * restore their cpu bindings. | |
2702 | */ | |
e46b1db2 | 2703 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
2704 | { |
2705 | int nid; | |
2706 | ||
e46b1db2 AMG |
2707 | for_each_node_state(nid, N_MEMORY) { |
2708 | pg_data_t *pgdat = NODE_DATA(nid); | |
2709 | const struct cpumask *mask; | |
698b1b30 | 2710 | |
e46b1db2 | 2711 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 2712 | |
e46b1db2 AMG |
2713 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
2714 | /* One of our CPUs online: restore mask */ | |
2715 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 2716 | } |
e46b1db2 | 2717 | return 0; |
698b1b30 VB |
2718 | } |
2719 | ||
2720 | static int __init kcompactd_init(void) | |
2721 | { | |
2722 | int nid; | |
e46b1db2 AMG |
2723 | int ret; |
2724 | ||
2725 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2726 | "mm/compaction:online", | |
2727 | kcompactd_cpu_online, NULL); | |
2728 | if (ret < 0) { | |
2729 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2730 | return ret; | |
2731 | } | |
698b1b30 VB |
2732 | |
2733 | for_each_node_state(nid, N_MEMORY) | |
2734 | kcompactd_run(nid); | |
698b1b30 VB |
2735 | return 0; |
2736 | } | |
2737 | subsys_initcall(kcompactd_init) | |
2738 | ||
ff9543fd | 2739 | #endif /* CONFIG_COMPACTION */ |