]>
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); | |
a2e9a5af VB |
273 | block_pfn = max(block_pfn, zone->zone_start_pfn); |
274 | block_page = pfn_to_online_page(block_pfn); | |
6b0868c8 MG |
275 | if (block_page) { |
276 | page = block_page; | |
277 | pfn = block_pfn; | |
278 | } | |
279 | ||
280 | /* Ensure the end of the pageblock or zone is online and valid */ | |
a2e9a5af | 281 | block_pfn = pageblock_end_pfn(pfn) - 1; |
6b0868c8 MG |
282 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); |
283 | end_page = pfn_to_online_page(block_pfn); | |
284 | if (!end_page) | |
285 | return false; | |
286 | ||
e332f741 MG |
287 | /* |
288 | * Only clear the hint if a sample indicates there is either a | |
289 | * free page or an LRU page in the block. One or other condition | |
290 | * is necessary for the block to be a migration source/target. | |
291 | */ | |
e332f741 MG |
292 | do { |
293 | if (pfn_valid_within(pfn)) { | |
294 | if (check_source && PageLRU(page)) { | |
295 | clear_pageblock_skip(page); | |
296 | return true; | |
297 | } | |
298 | ||
299 | if (check_target && PageBuddy(page)) { | |
300 | clear_pageblock_skip(page); | |
301 | return true; | |
302 | } | |
303 | } | |
304 | ||
305 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
306 | pfn += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
a2e9a5af | 307 | } while (page <= end_page); |
e332f741 MG |
308 | |
309 | return false; | |
310 | } | |
311 | ||
bb13ffeb MG |
312 | /* |
313 | * This function is called to clear all cached information on pageblocks that | |
314 | * should be skipped for page isolation when the migrate and free page scanner | |
315 | * meet. | |
316 | */ | |
62997027 | 317 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 318 | { |
e332f741 | 319 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 320 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
321 | unsigned long reset_migrate = free_pfn; |
322 | unsigned long reset_free = migrate_pfn; | |
323 | bool source_set = false; | |
324 | bool free_set = false; | |
325 | ||
326 | if (!zone->compact_blockskip_flush) | |
327 | return; | |
bb13ffeb | 328 | |
62997027 | 329 | zone->compact_blockskip_flush = false; |
bb13ffeb | 330 | |
e332f741 MG |
331 | /* |
332 | * Walk the zone and update pageblock skip information. Source looks | |
333 | * for PageLRU while target looks for PageBuddy. When the scanner | |
334 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
335 | * is suitable as both source and target. | |
336 | */ | |
337 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
338 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
339 | cond_resched(); |
340 | ||
e332f741 MG |
341 | /* Update the migrate PFN */ |
342 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
343 | migrate_pfn < reset_migrate) { | |
344 | source_set = true; | |
345 | reset_migrate = migrate_pfn; | |
346 | zone->compact_init_migrate_pfn = reset_migrate; | |
347 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
348 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
349 | } | |
bb13ffeb | 350 | |
e332f741 MG |
351 | /* Update the free PFN */ |
352 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
353 | free_pfn > reset_free) { | |
354 | free_set = true; | |
355 | reset_free = free_pfn; | |
356 | zone->compact_init_free_pfn = reset_free; | |
357 | zone->compact_cached_free_pfn = reset_free; | |
358 | } | |
bb13ffeb | 359 | } |
02333641 | 360 | |
e332f741 MG |
361 | /* Leave no distance if no suitable block was reset */ |
362 | if (reset_migrate >= reset_free) { | |
363 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
364 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
365 | zone->compact_cached_free_pfn = free_pfn; | |
366 | } | |
bb13ffeb MG |
367 | } |
368 | ||
62997027 MG |
369 | void reset_isolation_suitable(pg_data_t *pgdat) |
370 | { | |
371 | int zoneid; | |
372 | ||
373 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
374 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
375 | if (!populated_zone(zone)) | |
376 | continue; | |
377 | ||
378 | /* Only flush if a full compaction finished recently */ | |
379 | if (zone->compact_blockskip_flush) | |
380 | __reset_isolation_suitable(zone); | |
381 | } | |
382 | } | |
383 | ||
e380bebe MG |
384 | /* |
385 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
386 | * locks are not required for read/writers. Returns true if it was already set. | |
387 | */ | |
388 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
389 | unsigned long pfn) | |
390 | { | |
391 | bool skip; | |
392 | ||
393 | /* Do no update if skip hint is being ignored */ | |
394 | if (cc->ignore_skip_hint) | |
395 | return false; | |
396 | ||
397 | if (!IS_ALIGNED(pfn, pageblock_nr_pages)) | |
398 | return false; | |
399 | ||
400 | skip = get_pageblock_skip(page); | |
401 | if (!skip && !cc->no_set_skip_hint) | |
402 | set_pageblock_skip(page); | |
403 | ||
404 | return skip; | |
405 | } | |
406 | ||
407 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
408 | { | |
409 | struct zone *zone = cc->zone; | |
410 | ||
411 | pfn = pageblock_end_pfn(pfn); | |
412 | ||
413 | /* Set for isolation rather than compaction */ | |
414 | if (cc->no_set_skip_hint) | |
415 | return; | |
416 | ||
417 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
418 | zone->compact_cached_migrate_pfn[0] = pfn; | |
419 | if (cc->mode != MIGRATE_ASYNC && | |
420 | pfn > zone->compact_cached_migrate_pfn[1]) | |
421 | zone->compact_cached_migrate_pfn[1] = pfn; | |
422 | } | |
423 | ||
bb13ffeb MG |
424 | /* |
425 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 426 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 427 | */ |
c89511ab | 428 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 429 | struct page *page, unsigned long pfn) |
bb13ffeb | 430 | { |
c89511ab | 431 | struct zone *zone = cc->zone; |
6815bf3f | 432 | |
2583d671 | 433 | if (cc->no_set_skip_hint) |
6815bf3f JK |
434 | return; |
435 | ||
bb13ffeb MG |
436 | if (!page) |
437 | return; | |
438 | ||
edc2ca61 | 439 | set_pageblock_skip(page); |
c89511ab | 440 | |
35979ef3 | 441 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
442 | if (pfn < zone->compact_cached_free_pfn) |
443 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
444 | } |
445 | #else | |
446 | static inline bool isolation_suitable(struct compact_control *cc, | |
447 | struct page *page) | |
448 | { | |
449 | return true; | |
450 | } | |
451 | ||
b527cfe5 | 452 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
453 | { |
454 | return false; | |
455 | } | |
456 | ||
457 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 458 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
459 | { |
460 | } | |
e380bebe MG |
461 | |
462 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
463 | { | |
464 | } | |
465 | ||
466 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
467 | unsigned long pfn) | |
468 | { | |
469 | return false; | |
470 | } | |
bb13ffeb MG |
471 | #endif /* CONFIG_COMPACTION */ |
472 | ||
8b44d279 VB |
473 | /* |
474 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
475 | * very heavily contended. For async compaction, trylock and record if the |
476 | * lock is contended. The lock will still be acquired but compaction will | |
477 | * abort when the current block is finished regardless of success rate. | |
478 | * Sync compaction acquires the lock. | |
8b44d279 | 479 | * |
cb2dcaf0 | 480 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 481 | */ |
cb2dcaf0 | 482 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 483 | struct compact_control *cc) |
77337ede | 484 | __acquires(lock) |
2a1402aa | 485 | { |
cb2dcaf0 MG |
486 | /* Track if the lock is contended in async mode */ |
487 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
488 | if (spin_trylock_irqsave(lock, *flags)) | |
489 | return true; | |
490 | ||
491 | cc->contended = true; | |
8b44d279 | 492 | } |
1f9efdef | 493 | |
cb2dcaf0 | 494 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 495 | return true; |
2a1402aa MG |
496 | } |
497 | ||
c67fe375 MG |
498 | /* |
499 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
500 | * very heavily contended. The lock should be periodically unlocked to avoid |
501 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
502 | * the lock. It might also be that allowing the IRQs will result in | |
503 | * need_resched() becoming true. If scheduling is needed, async compaction | |
504 | * aborts. Sync compaction schedules. | |
505 | * Either compaction type will also abort if a fatal signal is pending. | |
506 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 507 | * |
8b44d279 VB |
508 | * Returns true if compaction should abort due to fatal signal pending, or |
509 | * async compaction due to need_resched() | |
510 | * Returns false when compaction can continue (sync compaction might have | |
511 | * scheduled) | |
c67fe375 | 512 | */ |
8b44d279 VB |
513 | static bool compact_unlock_should_abort(spinlock_t *lock, |
514 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 515 | { |
8b44d279 VB |
516 | if (*locked) { |
517 | spin_unlock_irqrestore(lock, flags); | |
518 | *locked = false; | |
519 | } | |
1f9efdef | 520 | |
8b44d279 | 521 | if (fatal_signal_pending(current)) { |
c3486f53 | 522 | cc->contended = true; |
8b44d279 VB |
523 | return true; |
524 | } | |
c67fe375 | 525 | |
cf66f070 | 526 | cond_resched(); |
be976572 VB |
527 | |
528 | return false; | |
529 | } | |
530 | ||
85aa125f | 531 | /* |
9e4be470 JM |
532 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
533 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
534 | * (even though it may still end up isolating some pages). | |
85aa125f | 535 | */ |
f40d1e42 | 536 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 537 | unsigned long *start_pfn, |
85aa125f MN |
538 | unsigned long end_pfn, |
539 | struct list_head *freelist, | |
4fca9730 | 540 | unsigned int stride, |
85aa125f | 541 | bool strict) |
748446bb | 542 | { |
b7aba698 | 543 | int nr_scanned = 0, total_isolated = 0; |
d097a6f6 | 544 | struct page *cursor; |
b8b2d825 | 545 | unsigned long flags = 0; |
f40d1e42 | 546 | bool locked = false; |
e14c720e | 547 | unsigned long blockpfn = *start_pfn; |
66c64223 | 548 | unsigned int order; |
748446bb | 549 | |
4fca9730 MG |
550 | /* Strict mode is for isolation, speed is secondary */ |
551 | if (strict) | |
552 | stride = 1; | |
553 | ||
748446bb MG |
554 | cursor = pfn_to_page(blockpfn); |
555 | ||
f40d1e42 | 556 | /* Isolate free pages. */ |
4fca9730 | 557 | for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { |
66c64223 | 558 | int isolated; |
748446bb MG |
559 | struct page *page = cursor; |
560 | ||
8b44d279 VB |
561 | /* |
562 | * Periodically drop the lock (if held) regardless of its | |
563 | * contention, to give chance to IRQs. Abort if fatal signal | |
564 | * pending or async compaction detects need_resched() | |
565 | */ | |
566 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
567 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
568 | &locked, cc)) | |
569 | break; | |
570 | ||
b7aba698 | 571 | nr_scanned++; |
f40d1e42 | 572 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
573 | goto isolate_fail; |
574 | ||
9fcd6d2e VB |
575 | /* |
576 | * For compound pages such as THP and hugetlbfs, we can save | |
577 | * potentially a lot of iterations if we skip them at once. | |
578 | * The check is racy, but we can consider only valid values | |
579 | * and the only danger is skipping too much. | |
580 | */ | |
581 | if (PageCompound(page)) { | |
21dc7e02 DR |
582 | const unsigned int order = compound_order(page); |
583 | ||
d3c85bad | 584 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
585 | blockpfn += (1UL << order) - 1; |
586 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 587 | } |
9fcd6d2e VB |
588 | goto isolate_fail; |
589 | } | |
590 | ||
f40d1e42 | 591 | if (!PageBuddy(page)) |
2af120bc | 592 | goto isolate_fail; |
f40d1e42 MG |
593 | |
594 | /* | |
69b7189f VB |
595 | * If we already hold the lock, we can skip some rechecking. |
596 | * Note that if we hold the lock now, checked_pageblock was | |
597 | * already set in some previous iteration (or strict is true), | |
598 | * so it is correct to skip the suitable migration target | |
599 | * recheck as well. | |
f40d1e42 | 600 | */ |
69b7189f | 601 | if (!locked) { |
cb2dcaf0 | 602 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 603 | &flags, cc); |
f40d1e42 | 604 | |
69b7189f VB |
605 | /* Recheck this is a buddy page under lock */ |
606 | if (!PageBuddy(page)) | |
607 | goto isolate_fail; | |
608 | } | |
748446bb | 609 | |
66c64223 JK |
610 | /* Found a free page, will break it into order-0 pages */ |
611 | order = page_order(page); | |
612 | isolated = __isolate_free_page(page, order); | |
a4f04f2c DR |
613 | if (!isolated) |
614 | break; | |
66c64223 | 615 | set_page_private(page, order); |
a4f04f2c | 616 | |
748446bb | 617 | total_isolated += isolated; |
a4f04f2c | 618 | cc->nr_freepages += isolated; |
66c64223 JK |
619 | list_add_tail(&page->lru, freelist); |
620 | ||
a4f04f2c DR |
621 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
622 | blockpfn += isolated; | |
623 | break; | |
748446bb | 624 | } |
a4f04f2c DR |
625 | /* Advance to the end of split page */ |
626 | blockpfn += isolated - 1; | |
627 | cursor += isolated - 1; | |
628 | continue; | |
2af120bc LA |
629 | |
630 | isolate_fail: | |
631 | if (strict) | |
632 | break; | |
633 | else | |
634 | continue; | |
635 | ||
748446bb MG |
636 | } |
637 | ||
a4f04f2c DR |
638 | if (locked) |
639 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
640 | ||
9fcd6d2e VB |
641 | /* |
642 | * There is a tiny chance that we have read bogus compound_order(), | |
643 | * so be careful to not go outside of the pageblock. | |
644 | */ | |
645 | if (unlikely(blockpfn > end_pfn)) | |
646 | blockpfn = end_pfn; | |
647 | ||
e34d85f0 JK |
648 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
649 | nr_scanned, total_isolated); | |
650 | ||
e14c720e VB |
651 | /* Record how far we have got within the block */ |
652 | *start_pfn = blockpfn; | |
653 | ||
f40d1e42 MG |
654 | /* |
655 | * If strict isolation is requested by CMA then check that all the | |
656 | * pages requested were isolated. If there were any failures, 0 is | |
657 | * returned and CMA will fail. | |
658 | */ | |
2af120bc | 659 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
660 | total_isolated = 0; |
661 | ||
7f354a54 | 662 | cc->total_free_scanned += nr_scanned; |
397487db | 663 | if (total_isolated) |
010fc29a | 664 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
665 | return total_isolated; |
666 | } | |
667 | ||
85aa125f MN |
668 | /** |
669 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 670 | * @cc: Compaction control structure. |
85aa125f MN |
671 | * @start_pfn: The first PFN to start isolating. |
672 | * @end_pfn: The one-past-last PFN. | |
673 | * | |
674 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
675 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
676 | * undo its actions and return zero. | |
677 | * | |
678 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
679 | * (which may be greater then end_pfn if end fell in a middle of | |
680 | * a free page). | |
681 | */ | |
ff9543fd | 682 | unsigned long |
bb13ffeb MG |
683 | isolate_freepages_range(struct compact_control *cc, |
684 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 685 | { |
e1409c32 | 686 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
687 | LIST_HEAD(freelist); |
688 | ||
7d49d886 | 689 | pfn = start_pfn; |
06b6640a | 690 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
691 | if (block_start_pfn < cc->zone->zone_start_pfn) |
692 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 693 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
694 | |
695 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 696 | block_start_pfn = block_end_pfn, |
7d49d886 | 697 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
698 | /* Protect pfn from changing by isolate_freepages_block */ |
699 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 700 | |
85aa125f MN |
701 | block_end_pfn = min(block_end_pfn, end_pfn); |
702 | ||
58420016 JK |
703 | /* |
704 | * pfn could pass the block_end_pfn if isolated freepage | |
705 | * is more than pageblock order. In this case, we adjust | |
706 | * scanning range to right one. | |
707 | */ | |
708 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
709 | block_start_pfn = pageblock_start_pfn(pfn); |
710 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
711 | block_end_pfn = min(block_end_pfn, end_pfn); |
712 | } | |
713 | ||
e1409c32 JK |
714 | if (!pageblock_pfn_to_page(block_start_pfn, |
715 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
716 | break; |
717 | ||
e14c720e | 718 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
4fca9730 | 719 | block_end_pfn, &freelist, 0, true); |
85aa125f MN |
720 | |
721 | /* | |
722 | * In strict mode, isolate_freepages_block() returns 0 if | |
723 | * there are any holes in the block (ie. invalid PFNs or | |
724 | * non-free pages). | |
725 | */ | |
726 | if (!isolated) | |
727 | break; | |
728 | ||
729 | /* | |
730 | * If we managed to isolate pages, it is always (1 << n) * | |
731 | * pageblock_nr_pages for some non-negative n. (Max order | |
732 | * page may span two pageblocks). | |
733 | */ | |
734 | } | |
735 | ||
66c64223 | 736 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 737 | split_map_pages(&freelist); |
85aa125f MN |
738 | |
739 | if (pfn < end_pfn) { | |
740 | /* Loop terminated early, cleanup. */ | |
741 | release_freepages(&freelist); | |
742 | return 0; | |
743 | } | |
744 | ||
745 | /* We don't use freelists for anything. */ | |
746 | return pfn; | |
747 | } | |
748 | ||
748446bb | 749 | /* Similar to reclaim, but different enough that they don't share logic */ |
5f438eee | 750 | static bool too_many_isolated(pg_data_t *pgdat) |
748446bb | 751 | { |
bc693045 | 752 | unsigned long active, inactive, isolated; |
748446bb | 753 | |
5f438eee AR |
754 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + |
755 | node_page_state(pgdat, NR_INACTIVE_ANON); | |
756 | active = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
757 | node_page_state(pgdat, NR_ACTIVE_ANON); | |
758 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + | |
759 | node_page_state(pgdat, NR_ISOLATED_ANON); | |
748446bb | 760 | |
bc693045 | 761 | return isolated > (inactive + active) / 2; |
748446bb MG |
762 | } |
763 | ||
2fe86e00 | 764 | /** |
edc2ca61 VB |
765 | * isolate_migratepages_block() - isolate all migrate-able pages within |
766 | * a single pageblock | |
2fe86e00 | 767 | * @cc: Compaction control structure. |
edc2ca61 VB |
768 | * @low_pfn: The first PFN to isolate |
769 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
770 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
771 | * |
772 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
773 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
774 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
775 | * first page that was not scanned (which may be both less, equal to or more | |
776 | * than end_pfn). | |
2fe86e00 | 777 | * |
edc2ca61 VB |
778 | * The pages are isolated on cc->migratepages list (not required to be empty), |
779 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
780 | * is neither read nor updated. | |
748446bb | 781 | */ |
edc2ca61 VB |
782 | static unsigned long |
783 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
784 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 785 | { |
5f438eee | 786 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 787 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 788 | struct lruvec *lruvec; |
b8b2d825 | 789 | unsigned long flags = 0; |
2a1402aa | 790 | bool locked = false; |
bb13ffeb | 791 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 792 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
793 | bool skip_on_failure = false; |
794 | unsigned long next_skip_pfn = 0; | |
e380bebe | 795 | bool skip_updated = false; |
748446bb | 796 | |
748446bb MG |
797 | /* |
798 | * Ensure that there are not too many pages isolated from the LRU | |
799 | * list by either parallel reclaimers or compaction. If there are, | |
800 | * delay for some time until fewer pages are isolated | |
801 | */ | |
5f438eee | 802 | while (unlikely(too_many_isolated(pgdat))) { |
f9e35b3b | 803 | /* async migration should just abort */ |
e0b9daeb | 804 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 805 | return 0; |
f9e35b3b | 806 | |
748446bb MG |
807 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
808 | ||
809 | if (fatal_signal_pending(current)) | |
2fe86e00 | 810 | return 0; |
748446bb MG |
811 | } |
812 | ||
cf66f070 | 813 | cond_resched(); |
aeef4b83 | 814 | |
fdd048e1 VB |
815 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
816 | skip_on_failure = true; | |
817 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
818 | } | |
819 | ||
748446bb | 820 | /* Time to isolate some pages for migration */ |
748446bb | 821 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 822 | |
fdd048e1 VB |
823 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
824 | /* | |
825 | * We have isolated all migration candidates in the | |
826 | * previous order-aligned block, and did not skip it due | |
827 | * to failure. We should migrate the pages now and | |
828 | * hopefully succeed compaction. | |
829 | */ | |
830 | if (nr_isolated) | |
831 | break; | |
832 | ||
833 | /* | |
834 | * We failed to isolate in the previous order-aligned | |
835 | * block. Set the new boundary to the end of the | |
836 | * current block. Note we can't simply increase | |
837 | * next_skip_pfn by 1 << order, as low_pfn might have | |
838 | * been incremented by a higher number due to skipping | |
839 | * a compound or a high-order buddy page in the | |
840 | * previous loop iteration. | |
841 | */ | |
842 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
843 | } | |
844 | ||
8b44d279 VB |
845 | /* |
846 | * Periodically drop the lock (if held) regardless of its | |
670105a2 MG |
847 | * contention, to give chance to IRQs. Abort completely if |
848 | * a fatal signal is pending. | |
8b44d279 VB |
849 | */ |
850 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
f4b7e272 | 851 | && compact_unlock_should_abort(&pgdat->lru_lock, |
670105a2 MG |
852 | flags, &locked, cc)) { |
853 | low_pfn = 0; | |
854 | goto fatal_pending; | |
855 | } | |
c67fe375 | 856 | |
748446bb | 857 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 858 | goto isolate_fail; |
b7aba698 | 859 | nr_scanned++; |
748446bb | 860 | |
748446bb | 861 | page = pfn_to_page(low_pfn); |
dc908600 | 862 | |
e380bebe MG |
863 | /* |
864 | * Check if the pageblock has already been marked skipped. | |
865 | * Only the aligned PFN is checked as the caller isolates | |
866 | * COMPACT_CLUSTER_MAX at a time so the second call must | |
867 | * not falsely conclude that the block should be skipped. | |
868 | */ | |
869 | if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) { | |
870 | if (!cc->ignore_skip_hint && get_pageblock_skip(page)) { | |
871 | low_pfn = end_pfn; | |
872 | goto isolate_abort; | |
873 | } | |
bb13ffeb | 874 | valid_page = page; |
e380bebe | 875 | } |
bb13ffeb | 876 | |
6c14466c | 877 | /* |
99c0fd5e VB |
878 | * Skip if free. We read page order here without zone lock |
879 | * which is generally unsafe, but the race window is small and | |
880 | * the worst thing that can happen is that we skip some | |
881 | * potential isolation targets. | |
6c14466c | 882 | */ |
99c0fd5e VB |
883 | if (PageBuddy(page)) { |
884 | unsigned long freepage_order = page_order_unsafe(page); | |
885 | ||
886 | /* | |
887 | * Without lock, we cannot be sure that what we got is | |
888 | * a valid page order. Consider only values in the | |
889 | * valid order range to prevent low_pfn overflow. | |
890 | */ | |
891 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
892 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 893 | continue; |
99c0fd5e | 894 | } |
748446bb | 895 | |
bc835011 | 896 | /* |
29c0dde8 | 897 | * Regardless of being on LRU, compound pages such as THP and |
1da2f328 RR |
898 | * hugetlbfs are not to be compacted unless we are attempting |
899 | * an allocation much larger than the huge page size (eg CMA). | |
900 | * We can potentially save a lot of iterations if we skip them | |
901 | * at once. The check is racy, but we can consider only valid | |
902 | * values and the only danger is skipping too much. | |
bc835011 | 903 | */ |
1da2f328 | 904 | if (PageCompound(page) && !cc->alloc_contig) { |
21dc7e02 | 905 | const unsigned int order = compound_order(page); |
edc2ca61 | 906 | |
d3c85bad | 907 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 908 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 909 | goto isolate_fail; |
2a1402aa MG |
910 | } |
911 | ||
bda807d4 MK |
912 | /* |
913 | * Check may be lockless but that's ok as we recheck later. | |
914 | * It's possible to migrate LRU and non-lru movable pages. | |
915 | * Skip any other type of page | |
916 | */ | |
917 | if (!PageLRU(page)) { | |
bda807d4 MK |
918 | /* |
919 | * __PageMovable can return false positive so we need | |
920 | * to verify it under page_lock. | |
921 | */ | |
922 | if (unlikely(__PageMovable(page)) && | |
923 | !PageIsolated(page)) { | |
924 | if (locked) { | |
f4b7e272 | 925 | spin_unlock_irqrestore(&pgdat->lru_lock, |
bda807d4 MK |
926 | flags); |
927 | locked = false; | |
928 | } | |
929 | ||
9e5bcd61 | 930 | if (!isolate_movable_page(page, isolate_mode)) |
bda807d4 MK |
931 | goto isolate_success; |
932 | } | |
933 | ||
fdd048e1 | 934 | goto isolate_fail; |
bda807d4 | 935 | } |
29c0dde8 | 936 | |
119d6d59 DR |
937 | /* |
938 | * Migration will fail if an anonymous page is pinned in memory, | |
939 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
940 | * admittedly racy check. | |
941 | */ | |
942 | if (!page_mapping(page) && | |
943 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 944 | goto isolate_fail; |
119d6d59 | 945 | |
73e64c51 MH |
946 | /* |
947 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
948 | * because those do not depend on fs locks. | |
949 | */ | |
950 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
951 | goto isolate_fail; | |
952 | ||
69b7189f VB |
953 | /* If we already hold the lock, we can skip some rechecking */ |
954 | if (!locked) { | |
f4b7e272 | 955 | locked = compact_lock_irqsave(&pgdat->lru_lock, |
8b44d279 | 956 | &flags, cc); |
e380bebe | 957 | |
e380bebe MG |
958 | /* Try get exclusive access under lock */ |
959 | if (!skip_updated) { | |
960 | skip_updated = true; | |
961 | if (test_and_set_skip(cc, page, low_pfn)) | |
962 | goto isolate_abort; | |
963 | } | |
2a1402aa | 964 | |
29c0dde8 | 965 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 966 | if (!PageLRU(page)) |
fdd048e1 | 967 | goto isolate_fail; |
29c0dde8 VB |
968 | |
969 | /* | |
970 | * Page become compound since the non-locked check, | |
971 | * and it's on LRU. It can only be a THP so the order | |
972 | * is safe to read and it's 0 for tail pages. | |
973 | */ | |
1da2f328 | 974 | if (unlikely(PageCompound(page) && !cc->alloc_contig)) { |
d8c6546b | 975 | low_pfn += compound_nr(page) - 1; |
fdd048e1 | 976 | goto isolate_fail; |
69b7189f | 977 | } |
bc835011 AA |
978 | } |
979 | ||
f4b7e272 | 980 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 981 | |
748446bb | 982 | /* Try isolate the page */ |
edc2ca61 | 983 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 984 | goto isolate_fail; |
748446bb | 985 | |
1da2f328 RR |
986 | /* The whole page is taken off the LRU; skip the tail pages. */ |
987 | if (PageCompound(page)) | |
988 | low_pfn += compound_nr(page) - 1; | |
bc835011 | 989 | |
748446bb | 990 | /* Successfully isolated */ |
fa9add64 | 991 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
1da2f328 | 992 | mod_node_page_state(page_pgdat(page), |
9de4f22a | 993 | NR_ISOLATED_ANON + page_is_file_lru(page), |
1da2f328 | 994 | hpage_nr_pages(page)); |
b6c75016 JK |
995 | |
996 | isolate_success: | |
fdd048e1 | 997 | list_add(&page->lru, &cc->migratepages); |
748446bb | 998 | cc->nr_migratepages++; |
b7aba698 | 999 | nr_isolated++; |
748446bb | 1000 | |
804d3121 MG |
1001 | /* |
1002 | * Avoid isolating too much unless this block is being | |
cb2dcaf0 MG |
1003 | * rescanned (e.g. dirty/writeback pages, parallel allocation) |
1004 | * or a lock is contended. For contention, isolate quickly to | |
1005 | * potentially remove one source of contention. | |
804d3121 | 1006 | */ |
cb2dcaf0 MG |
1007 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX && |
1008 | !cc->rescan && !cc->contended) { | |
31b8384a | 1009 | ++low_pfn; |
748446bb | 1010 | break; |
31b8384a | 1011 | } |
fdd048e1 VB |
1012 | |
1013 | continue; | |
1014 | isolate_fail: | |
1015 | if (!skip_on_failure) | |
1016 | continue; | |
1017 | ||
1018 | /* | |
1019 | * We have isolated some pages, but then failed. Release them | |
1020 | * instead of migrating, as we cannot form the cc->order buddy | |
1021 | * page anyway. | |
1022 | */ | |
1023 | if (nr_isolated) { | |
1024 | if (locked) { | |
f4b7e272 | 1025 | spin_unlock_irqrestore(&pgdat->lru_lock, flags); |
fdd048e1 VB |
1026 | locked = false; |
1027 | } | |
fdd048e1 VB |
1028 | putback_movable_pages(&cc->migratepages); |
1029 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1030 | nr_isolated = 0; |
1031 | } | |
1032 | ||
1033 | if (low_pfn < next_skip_pfn) { | |
1034 | low_pfn = next_skip_pfn - 1; | |
1035 | /* | |
1036 | * The check near the loop beginning would have updated | |
1037 | * next_skip_pfn too, but this is a bit simpler. | |
1038 | */ | |
1039 | next_skip_pfn += 1UL << cc->order; | |
1040 | } | |
748446bb MG |
1041 | } |
1042 | ||
99c0fd5e VB |
1043 | /* |
1044 | * The PageBuddy() check could have potentially brought us outside | |
1045 | * the range to be scanned. | |
1046 | */ | |
1047 | if (unlikely(low_pfn > end_pfn)) | |
1048 | low_pfn = end_pfn; | |
1049 | ||
e380bebe | 1050 | isolate_abort: |
c67fe375 | 1051 | if (locked) |
f4b7e272 | 1052 | spin_unlock_irqrestore(&pgdat->lru_lock, flags); |
748446bb | 1053 | |
50b5b094 | 1054 | /* |
804d3121 MG |
1055 | * Updated the cached scanner pfn once the pageblock has been scanned |
1056 | * Pages will either be migrated in which case there is no point | |
1057 | * scanning in the near future or migration failed in which case the | |
1058 | * failure reason may persist. The block is marked for skipping if | |
1059 | * there were no pages isolated in the block or if the block is | |
1060 | * rescanned twice in a row. | |
50b5b094 | 1061 | */ |
804d3121 | 1062 | if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) { |
e380bebe MG |
1063 | if (valid_page && !skip_updated) |
1064 | set_pageblock_skip(valid_page); | |
1065 | update_cached_migrate(cc, low_pfn); | |
1066 | } | |
bb13ffeb | 1067 | |
e34d85f0 JK |
1068 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1069 | nr_scanned, nr_isolated); | |
b7aba698 | 1070 | |
670105a2 | 1071 | fatal_pending: |
7f354a54 | 1072 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1073 | if (nr_isolated) |
010fc29a | 1074 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1075 | |
2fe86e00 MN |
1076 | return low_pfn; |
1077 | } | |
1078 | ||
edc2ca61 VB |
1079 | /** |
1080 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1081 | * @cc: Compaction control structure. | |
1082 | * @start_pfn: The first PFN to start isolating. | |
1083 | * @end_pfn: The one-past-last PFN. | |
1084 | * | |
1085 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
1086 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
1087 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
1088 | */ | |
1089 | unsigned long | |
1090 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
1091 | unsigned long end_pfn) | |
1092 | { | |
e1409c32 | 1093 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
1094 | |
1095 | /* Scan block by block. First and last block may be incomplete */ | |
1096 | pfn = start_pfn; | |
06b6640a | 1097 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1098 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1099 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1100 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1101 | |
1102 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1103 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1104 | block_end_pfn += pageblock_nr_pages) { |
1105 | ||
1106 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1107 | ||
e1409c32 JK |
1108 | if (!pageblock_pfn_to_page(block_start_pfn, |
1109 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1110 | continue; |
1111 | ||
1112 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
1113 | ISOLATE_UNEVICTABLE); | |
1114 | ||
14af4a5e | 1115 | if (!pfn) |
edc2ca61 | 1116 | break; |
6ea41c0c JK |
1117 | |
1118 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
1119 | break; | |
edc2ca61 | 1120 | } |
edc2ca61 VB |
1121 | |
1122 | return pfn; | |
1123 | } | |
1124 | ||
ff9543fd MN |
1125 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1126 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1127 | |
b682debd VB |
1128 | static bool suitable_migration_source(struct compact_control *cc, |
1129 | struct page *page) | |
1130 | { | |
282722b0 VB |
1131 | int block_mt; |
1132 | ||
9bebefd5 MG |
1133 | if (pageblock_skip_persistent(page)) |
1134 | return false; | |
1135 | ||
282722b0 | 1136 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1137 | return true; |
1138 | ||
282722b0 VB |
1139 | block_mt = get_pageblock_migratetype(page); |
1140 | ||
1141 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1142 | return is_migrate_movable(block_mt); | |
1143 | else | |
1144 | return block_mt == cc->migratetype; | |
b682debd VB |
1145 | } |
1146 | ||
018e9a49 | 1147 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1148 | static bool suitable_migration_target(struct compact_control *cc, |
1149 | struct page *page) | |
018e9a49 AM |
1150 | { |
1151 | /* If the page is a large free page, then disallow migration */ | |
1152 | if (PageBuddy(page)) { | |
1153 | /* | |
1154 | * We are checking page_order without zone->lock taken. But | |
1155 | * the only small danger is that we skip a potentially suitable | |
1156 | * pageblock, so it's not worth to check order for valid range. | |
1157 | */ | |
1158 | if (page_order_unsafe(page) >= pageblock_order) | |
1159 | return false; | |
1160 | } | |
1161 | ||
1ef36db2 YX |
1162 | if (cc->ignore_block_suitable) |
1163 | return true; | |
1164 | ||
018e9a49 | 1165 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1166 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1167 | return true; |
1168 | ||
1169 | /* Otherwise skip the block */ | |
1170 | return false; | |
1171 | } | |
1172 | ||
70b44595 MG |
1173 | static inline unsigned int |
1174 | freelist_scan_limit(struct compact_control *cc) | |
1175 | { | |
dd7ef7bd QC |
1176 | unsigned short shift = BITS_PER_LONG - 1; |
1177 | ||
1178 | return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; | |
70b44595 MG |
1179 | } |
1180 | ||
f2849aa0 VB |
1181 | /* |
1182 | * Test whether the free scanner has reached the same or lower pageblock than | |
1183 | * the migration scanner, and compaction should thus terminate. | |
1184 | */ | |
1185 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1186 | { | |
1187 | return (cc->free_pfn >> pageblock_order) | |
1188 | <= (cc->migrate_pfn >> pageblock_order); | |
1189 | } | |
1190 | ||
5a811889 MG |
1191 | /* |
1192 | * Used when scanning for a suitable migration target which scans freelists | |
1193 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1194 | * first on the next iteration of the free scanner | |
1195 | */ | |
1196 | static void | |
1197 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1198 | { | |
1199 | LIST_HEAD(sublist); | |
1200 | ||
1201 | if (!list_is_last(freelist, &freepage->lru)) { | |
1202 | list_cut_before(&sublist, freelist, &freepage->lru); | |
1203 | if (!list_empty(&sublist)) | |
1204 | list_splice_tail(&sublist, freelist); | |
1205 | } | |
1206 | } | |
1207 | ||
1208 | /* | |
1209 | * Similar to move_freelist_head except used by the migration scanner | |
1210 | * when scanning forward. It's possible for these list operations to | |
1211 | * move against each other if they search the free list exactly in | |
1212 | * lockstep. | |
1213 | */ | |
70b44595 MG |
1214 | static void |
1215 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1216 | { | |
1217 | LIST_HEAD(sublist); | |
1218 | ||
1219 | if (!list_is_first(freelist, &freepage->lru)) { | |
1220 | list_cut_position(&sublist, freelist, &freepage->lru); | |
1221 | if (!list_empty(&sublist)) | |
1222 | list_splice_tail(&sublist, freelist); | |
1223 | } | |
1224 | } | |
1225 | ||
5a811889 MG |
1226 | static void |
1227 | fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated) | |
1228 | { | |
1229 | unsigned long start_pfn, end_pfn; | |
1230 | struct page *page = pfn_to_page(pfn); | |
1231 | ||
1232 | /* Do not search around if there are enough pages already */ | |
1233 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1234 | return; | |
1235 | ||
1236 | /* Minimise scanning during async compaction */ | |
1237 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1238 | return; | |
1239 | ||
1240 | /* Pageblock boundaries */ | |
1241 | start_pfn = pageblock_start_pfn(pfn); | |
60fce36a | 1242 | end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1; |
5a811889 MG |
1243 | |
1244 | /* Scan before */ | |
1245 | if (start_pfn != pfn) { | |
4fca9730 | 1246 | isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false); |
5a811889 MG |
1247 | if (cc->nr_freepages >= cc->nr_migratepages) |
1248 | return; | |
1249 | } | |
1250 | ||
1251 | /* Scan after */ | |
1252 | start_pfn = pfn + nr_isolated; | |
60fce36a | 1253 | if (start_pfn < end_pfn) |
4fca9730 | 1254 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); |
5a811889 MG |
1255 | |
1256 | /* Skip this pageblock in the future as it's full or nearly full */ | |
1257 | if (cc->nr_freepages < cc->nr_migratepages) | |
1258 | set_pageblock_skip(page); | |
1259 | } | |
1260 | ||
dbe2d4e4 MG |
1261 | /* Search orders in round-robin fashion */ |
1262 | static int next_search_order(struct compact_control *cc, int order) | |
1263 | { | |
1264 | order--; | |
1265 | if (order < 0) | |
1266 | order = cc->order - 1; | |
1267 | ||
1268 | /* Search wrapped around? */ | |
1269 | if (order == cc->search_order) { | |
1270 | cc->search_order--; | |
1271 | if (cc->search_order < 0) | |
1272 | cc->search_order = cc->order - 1; | |
1273 | return -1; | |
1274 | } | |
1275 | ||
1276 | return order; | |
1277 | } | |
1278 | ||
5a811889 MG |
1279 | static unsigned long |
1280 | fast_isolate_freepages(struct compact_control *cc) | |
1281 | { | |
1282 | unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1); | |
1283 | unsigned int nr_scanned = 0; | |
1284 | unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0; | |
1285 | unsigned long nr_isolated = 0; | |
1286 | unsigned long distance; | |
1287 | struct page *page = NULL; | |
1288 | bool scan_start = false; | |
1289 | int order; | |
1290 | ||
1291 | /* Full compaction passes in a negative order */ | |
1292 | if (cc->order <= 0) | |
1293 | return cc->free_pfn; | |
1294 | ||
1295 | /* | |
1296 | * If starting the scan, use a deeper search and use the highest | |
1297 | * PFN found if a suitable one is not found. | |
1298 | */ | |
e332f741 | 1299 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1300 | limit = pageblock_nr_pages >> 1; |
1301 | scan_start = true; | |
1302 | } | |
1303 | ||
1304 | /* | |
1305 | * Preferred point is in the top quarter of the scan space but take | |
1306 | * a pfn from the top half if the search is problematic. | |
1307 | */ | |
1308 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1309 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1310 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1311 | ||
1312 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1313 | low_pfn = min_pfn; | |
1314 | ||
dbe2d4e4 MG |
1315 | /* |
1316 | * Search starts from the last successful isolation order or the next | |
1317 | * order to search after a previous failure | |
1318 | */ | |
1319 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1320 | ||
1321 | for (order = cc->search_order; | |
1322 | !page && order >= 0; | |
1323 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1324 | struct free_area *area = &cc->zone->free_area[order]; |
1325 | struct list_head *freelist; | |
1326 | struct page *freepage; | |
1327 | unsigned long flags; | |
1328 | unsigned int order_scanned = 0; | |
1329 | ||
1330 | if (!area->nr_free) | |
1331 | continue; | |
1332 | ||
1333 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1334 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1335 | list_for_each_entry_reverse(freepage, freelist, lru) { | |
1336 | unsigned long pfn; | |
1337 | ||
1338 | order_scanned++; | |
1339 | nr_scanned++; | |
1340 | pfn = page_to_pfn(freepage); | |
1341 | ||
1342 | if (pfn >= highest) | |
1343 | highest = pageblock_start_pfn(pfn); | |
1344 | ||
1345 | if (pfn >= low_pfn) { | |
1346 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1347 | cc->search_order = order; |
5a811889 MG |
1348 | page = freepage; |
1349 | break; | |
1350 | } | |
1351 | ||
1352 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1353 | high_pfn = pfn; | |
1354 | ||
1355 | /* Shorten the scan if a candidate is found */ | |
1356 | limit >>= 1; | |
1357 | } | |
1358 | ||
1359 | if (order_scanned >= limit) | |
1360 | break; | |
1361 | } | |
1362 | ||
1363 | /* Use a minimum pfn if a preferred one was not found */ | |
1364 | if (!page && high_pfn) { | |
1365 | page = pfn_to_page(high_pfn); | |
1366 | ||
1367 | /* Update freepage for the list reorder below */ | |
1368 | freepage = page; | |
1369 | } | |
1370 | ||
1371 | /* Reorder to so a future search skips recent pages */ | |
1372 | move_freelist_head(freelist, freepage); | |
1373 | ||
1374 | /* Isolate the page if available */ | |
1375 | if (page) { | |
1376 | if (__isolate_free_page(page, order)) { | |
1377 | set_page_private(page, order); | |
1378 | nr_isolated = 1 << order; | |
1379 | cc->nr_freepages += nr_isolated; | |
1380 | list_add_tail(&page->lru, &cc->freepages); | |
1381 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1382 | } else { | |
1383 | /* If isolation fails, abort the search */ | |
5b56d996 | 1384 | order = cc->search_order + 1; |
5a811889 MG |
1385 | page = NULL; |
1386 | } | |
1387 | } | |
1388 | ||
1389 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1390 | ||
1391 | /* | |
1392 | * Smaller scan on next order so the total scan ig related | |
1393 | * to freelist_scan_limit. | |
1394 | */ | |
1395 | if (order_scanned >= limit) | |
1396 | limit = min(1U, limit >> 1); | |
1397 | } | |
1398 | ||
1399 | if (!page) { | |
1400 | cc->fast_search_fail++; | |
1401 | if (scan_start) { | |
1402 | /* | |
1403 | * Use the highest PFN found above min. If one was | |
1404 | * not found, be pessemistic for direct compaction | |
1405 | * and use the min mark. | |
1406 | */ | |
1407 | if (highest) { | |
1408 | page = pfn_to_page(highest); | |
1409 | cc->free_pfn = highest; | |
1410 | } else { | |
e577c8b6 | 1411 | if (cc->direct_compaction && pfn_valid(min_pfn)) { |
5a811889 MG |
1412 | page = pfn_to_page(min_pfn); |
1413 | cc->free_pfn = min_pfn; | |
1414 | } | |
1415 | } | |
1416 | } | |
1417 | } | |
1418 | ||
d097a6f6 MG |
1419 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1420 | highest -= pageblock_nr_pages; | |
5a811889 | 1421 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1422 | } |
5a811889 MG |
1423 | |
1424 | cc->total_free_scanned += nr_scanned; | |
1425 | if (!page) | |
1426 | return cc->free_pfn; | |
1427 | ||
1428 | low_pfn = page_to_pfn(page); | |
1429 | fast_isolate_around(cc, low_pfn, nr_isolated); | |
1430 | return low_pfn; | |
1431 | } | |
1432 | ||
2fe86e00 | 1433 | /* |
ff9543fd MN |
1434 | * Based on information in the current compact_control, find blocks |
1435 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1436 | */ |
edc2ca61 | 1437 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1438 | { |
edc2ca61 | 1439 | struct zone *zone = cc->zone; |
ff9543fd | 1440 | struct page *page; |
c96b9e50 | 1441 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1442 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1443 | unsigned long block_end_pfn; /* end of current pageblock */ |
1444 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1445 | struct list_head *freelist = &cc->freepages; |
4fca9730 | 1446 | unsigned int stride; |
2fe86e00 | 1447 | |
5a811889 MG |
1448 | /* Try a small search of the free lists for a candidate */ |
1449 | isolate_start_pfn = fast_isolate_freepages(cc); | |
1450 | if (cc->nr_freepages) | |
1451 | goto splitmap; | |
1452 | ||
ff9543fd MN |
1453 | /* |
1454 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1455 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1456 | * zone when isolating for the first time. For looping we also need |
1457 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1458 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1459 | * For ending point, take care when isolating in last pageblock of a | |
1460 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1461 | * The low boundary is the end of the pageblock the migration scanner |
1462 | * is using. | |
ff9543fd | 1463 | */ |
e14c720e | 1464 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1465 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1466 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1467 | zone_end_pfn(zone)); | |
06b6640a | 1468 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1469 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1470 | |
ff9543fd MN |
1471 | /* |
1472 | * Isolate free pages until enough are available to migrate the | |
1473 | * pages on cc->migratepages. We stop searching if the migrate | |
1474 | * and free page scanners meet or enough free pages are isolated. | |
1475 | */ | |
f5f61a32 | 1476 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1477 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1478 | block_start_pfn -= pageblock_nr_pages, |
1479 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1480 | unsigned long nr_isolated; |
1481 | ||
f6ea3adb DR |
1482 | /* |
1483 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1484 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1485 | */ |
cb810ad2 | 1486 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1487 | cond_resched(); |
f6ea3adb | 1488 | |
7d49d886 VB |
1489 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1490 | zone); | |
1491 | if (!page) | |
ff9543fd MN |
1492 | continue; |
1493 | ||
1494 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1495 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1496 | continue; |
68e3e926 | 1497 | |
bb13ffeb MG |
1498 | /* If isolation recently failed, do not retry */ |
1499 | if (!isolation_suitable(cc, page)) | |
1500 | continue; | |
1501 | ||
e14c720e | 1502 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 MG |
1503 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
1504 | block_end_pfn, freelist, stride, false); | |
ff9543fd | 1505 | |
d097a6f6 MG |
1506 | /* Update the skip hint if the full pageblock was scanned */ |
1507 | if (isolate_start_pfn == block_end_pfn) | |
1508 | update_pageblock_skip(cc, page, block_start_pfn); | |
1509 | ||
cb2dcaf0 MG |
1510 | /* Are enough freepages isolated? */ |
1511 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1512 | if (isolate_start_pfn >= block_end_pfn) { |
1513 | /* | |
1514 | * Restart at previous pageblock if more | |
1515 | * freepages can be isolated next time. | |
1516 | */ | |
f5f61a32 VB |
1517 | isolate_start_pfn = |
1518 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1519 | } |
be976572 | 1520 | break; |
a46cbf3b | 1521 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1522 | /* |
a46cbf3b DR |
1523 | * If isolation failed early, do not continue |
1524 | * needlessly. | |
f5f61a32 | 1525 | */ |
a46cbf3b | 1526 | break; |
f5f61a32 | 1527 | } |
4fca9730 MG |
1528 | |
1529 | /* Adjust stride depending on isolation */ | |
1530 | if (nr_isolated) { | |
1531 | stride = 1; | |
1532 | continue; | |
1533 | } | |
1534 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1535 | } |
1536 | ||
7ed695e0 | 1537 | /* |
f5f61a32 VB |
1538 | * Record where the free scanner will restart next time. Either we |
1539 | * broke from the loop and set isolate_start_pfn based on the last | |
1540 | * call to isolate_freepages_block(), or we met the migration scanner | |
1541 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1542 | */ |
f5f61a32 | 1543 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1544 | |
1545 | splitmap: | |
1546 | /* __isolate_free_page() does not map the pages */ | |
1547 | split_map_pages(freelist); | |
748446bb MG |
1548 | } |
1549 | ||
1550 | /* | |
1551 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1552 | * from the isolated freelists in the block we are migrating to. | |
1553 | */ | |
1554 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1555 | unsigned long data) |
748446bb MG |
1556 | { |
1557 | struct compact_control *cc = (struct compact_control *)data; | |
1558 | struct page *freepage; | |
1559 | ||
748446bb | 1560 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1561 | isolate_freepages(cc); |
748446bb MG |
1562 | |
1563 | if (list_empty(&cc->freepages)) | |
1564 | return NULL; | |
1565 | } | |
1566 | ||
1567 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1568 | list_del(&freepage->lru); | |
1569 | cc->nr_freepages--; | |
1570 | ||
1571 | return freepage; | |
1572 | } | |
1573 | ||
1574 | /* | |
d53aea3d DR |
1575 | * This is a migrate-callback that "frees" freepages back to the isolated |
1576 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1577 | * special handling needed for NUMA. | |
1578 | */ | |
1579 | static void compaction_free(struct page *page, unsigned long data) | |
1580 | { | |
1581 | struct compact_control *cc = (struct compact_control *)data; | |
1582 | ||
1583 | list_add(&page->lru, &cc->freepages); | |
1584 | cc->nr_freepages++; | |
1585 | } | |
1586 | ||
ff9543fd MN |
1587 | /* possible outcome of isolate_migratepages */ |
1588 | typedef enum { | |
1589 | ISOLATE_ABORT, /* Abort compaction now */ | |
1590 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1591 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1592 | } isolate_migrate_t; | |
1593 | ||
5bbe3547 EM |
1594 | /* |
1595 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1596 | * compactable pages. | |
1597 | */ | |
6923aa0d SAS |
1598 | #ifdef CONFIG_PREEMPT_RT |
1599 | int sysctl_compact_unevictable_allowed __read_mostly = 0; | |
1600 | #else | |
5bbe3547 | 1601 | int sysctl_compact_unevictable_allowed __read_mostly = 1; |
6923aa0d | 1602 | #endif |
5bbe3547 | 1603 | |
70b44595 MG |
1604 | static inline void |
1605 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1606 | { | |
1607 | if (cc->fast_start_pfn == ULONG_MAX) | |
1608 | return; | |
1609 | ||
1610 | if (!cc->fast_start_pfn) | |
1611 | cc->fast_start_pfn = pfn; | |
1612 | ||
1613 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1614 | } | |
1615 | ||
1616 | static inline unsigned long | |
1617 | reinit_migrate_pfn(struct compact_control *cc) | |
1618 | { | |
1619 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1620 | return cc->migrate_pfn; | |
1621 | ||
1622 | cc->migrate_pfn = cc->fast_start_pfn; | |
1623 | cc->fast_start_pfn = ULONG_MAX; | |
1624 | ||
1625 | return cc->migrate_pfn; | |
1626 | } | |
1627 | ||
1628 | /* | |
1629 | * Briefly search the free lists for a migration source that already has | |
1630 | * some free pages to reduce the number of pages that need migration | |
1631 | * before a pageblock is free. | |
1632 | */ | |
1633 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1634 | { | |
1635 | unsigned int limit = freelist_scan_limit(cc); | |
1636 | unsigned int nr_scanned = 0; | |
1637 | unsigned long distance; | |
1638 | unsigned long pfn = cc->migrate_pfn; | |
1639 | unsigned long high_pfn; | |
1640 | int order; | |
1641 | ||
1642 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1643 | if (cc->ignore_skip_hint) | |
1644 | return pfn; | |
1645 | ||
1646 | /* | |
1647 | * If the migrate_pfn is not at the start of a zone or the start | |
1648 | * of a pageblock then assume this is a continuation of a previous | |
1649 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1650 | */ | |
1651 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1652 | return pfn; | |
1653 | ||
1654 | /* | |
1655 | * For smaller orders, just linearly scan as the number of pages | |
1656 | * to migrate should be relatively small and does not necessarily | |
1657 | * justify freeing up a large block for a small allocation. | |
1658 | */ | |
1659 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1660 | return pfn; | |
1661 | ||
1662 | /* | |
1663 | * Only allow kcompactd and direct requests for movable pages to | |
1664 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1665 | * reduces the risk that a large movable pageblock is freed for | |
1666 | * an unmovable/reclaimable small allocation. | |
1667 | */ | |
1668 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1669 | return pfn; | |
1670 | ||
1671 | /* | |
1672 | * When starting the migration scanner, pick any pageblock within the | |
1673 | * first half of the search space. Otherwise try and pick a pageblock | |
1674 | * within the first eighth to reduce the chances that a migration | |
1675 | * target later becomes a source. | |
1676 | */ | |
1677 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1678 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1679 | distance >>= 2; | |
1680 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1681 | ||
1682 | for (order = cc->order - 1; | |
1683 | order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; | |
1684 | order--) { | |
1685 | struct free_area *area = &cc->zone->free_area[order]; | |
1686 | struct list_head *freelist; | |
1687 | unsigned long flags; | |
1688 | struct page *freepage; | |
1689 | ||
1690 | if (!area->nr_free) | |
1691 | continue; | |
1692 | ||
1693 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1694 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1695 | list_for_each_entry(freepage, freelist, lru) { | |
1696 | unsigned long free_pfn; | |
1697 | ||
1698 | nr_scanned++; | |
1699 | free_pfn = page_to_pfn(freepage); | |
1700 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1701 | /* |
1702 | * Avoid if skipped recently. Ideally it would | |
1703 | * move to the tail but even safe iteration of | |
1704 | * the list assumes an entry is deleted, not | |
1705 | * reordered. | |
1706 | */ | |
1707 | if (get_pageblock_skip(freepage)) { | |
1708 | if (list_is_last(freelist, &freepage->lru)) | |
1709 | break; | |
1710 | ||
1711 | continue; | |
1712 | } | |
1713 | ||
1714 | /* Reorder to so a future search skips recent pages */ | |
1715 | move_freelist_tail(freelist, freepage); | |
1716 | ||
e380bebe | 1717 | update_fast_start_pfn(cc, free_pfn); |
70b44595 MG |
1718 | pfn = pageblock_start_pfn(free_pfn); |
1719 | cc->fast_search_fail = 0; | |
1720 | set_pageblock_skip(freepage); | |
1721 | break; | |
1722 | } | |
1723 | ||
1724 | if (nr_scanned >= limit) { | |
1725 | cc->fast_search_fail++; | |
1726 | move_freelist_tail(freelist, freepage); | |
1727 | break; | |
1728 | } | |
1729 | } | |
1730 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1731 | } | |
1732 | ||
1733 | cc->total_migrate_scanned += nr_scanned; | |
1734 | ||
1735 | /* | |
1736 | * If fast scanning failed then use a cached entry for a page block | |
1737 | * that had free pages as the basis for starting a linear scan. | |
1738 | */ | |
1739 | if (pfn == cc->migrate_pfn) | |
1740 | pfn = reinit_migrate_pfn(cc); | |
1741 | ||
1742 | return pfn; | |
1743 | } | |
1744 | ||
ff9543fd | 1745 | /* |
edc2ca61 VB |
1746 | * Isolate all pages that can be migrated from the first suitable block, |
1747 | * starting at the block pointed to by the migrate scanner pfn within | |
1748 | * compact_control. | |
ff9543fd | 1749 | */ |
32aaf055 | 1750 | static isolate_migrate_t isolate_migratepages(struct compact_control *cc) |
ff9543fd | 1751 | { |
e1409c32 JK |
1752 | unsigned long block_start_pfn; |
1753 | unsigned long block_end_pfn; | |
1754 | unsigned long low_pfn; | |
edc2ca61 VB |
1755 | struct page *page; |
1756 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1757 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1758 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 1759 | bool fast_find_block; |
ff9543fd | 1760 | |
edc2ca61 VB |
1761 | /* |
1762 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
1763 | * initialized by compact_zone(). The first failure will use |
1764 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 1765 | */ |
70b44595 | 1766 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 1767 | block_start_pfn = pageblock_start_pfn(low_pfn); |
32aaf055 PL |
1768 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1769 | block_start_pfn = cc->zone->zone_start_pfn; | |
ff9543fd | 1770 | |
70b44595 MG |
1771 | /* |
1772 | * fast_find_migrateblock marks a pageblock skipped so to avoid | |
1773 | * the isolation_suitable check below, check whether the fast | |
1774 | * search was successful. | |
1775 | */ | |
1776 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
1777 | ||
ff9543fd | 1778 | /* Only scan within a pageblock boundary */ |
06b6640a | 1779 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1780 | |
edc2ca61 VB |
1781 | /* |
1782 | * Iterate over whole pageblocks until we find the first suitable. | |
1783 | * Do not cross the free scanner. | |
1784 | */ | |
e1409c32 | 1785 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 1786 | fast_find_block = false, |
e1409c32 JK |
1787 | low_pfn = block_end_pfn, |
1788 | block_start_pfn = block_end_pfn, | |
1789 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1790 | |
edc2ca61 VB |
1791 | /* |
1792 | * This can potentially iterate a massively long zone with | |
1793 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 1794 | * need to schedule. |
edc2ca61 | 1795 | */ |
cb810ad2 | 1796 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1797 | cond_resched(); |
ff9543fd | 1798 | |
32aaf055 PL |
1799 | page = pageblock_pfn_to_page(block_start_pfn, |
1800 | block_end_pfn, cc->zone); | |
7d49d886 | 1801 | if (!page) |
edc2ca61 VB |
1802 | continue; |
1803 | ||
e380bebe MG |
1804 | /* |
1805 | * If isolation recently failed, do not retry. Only check the | |
1806 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
1807 | * to be visited multiple times. Assume skip was checked | |
1808 | * before making it "skip" so other compaction instances do | |
1809 | * not scan the same block. | |
1810 | */ | |
1811 | if (IS_ALIGNED(low_pfn, pageblock_nr_pages) && | |
1812 | !fast_find_block && !isolation_suitable(cc, page)) | |
edc2ca61 VB |
1813 | continue; |
1814 | ||
1815 | /* | |
9bebefd5 MG |
1816 | * For async compaction, also only scan in MOVABLE blocks |
1817 | * without huge pages. Async compaction is optimistic to see | |
1818 | * if the minimum amount of work satisfies the allocation. | |
1819 | * The cached PFN is updated as it's possible that all | |
1820 | * remaining blocks between source and target are unsuitable | |
1821 | * and the compaction scanners fail to meet. | |
edc2ca61 | 1822 | */ |
9bebefd5 MG |
1823 | if (!suitable_migration_source(cc, page)) { |
1824 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 1825 | continue; |
9bebefd5 | 1826 | } |
edc2ca61 VB |
1827 | |
1828 | /* Perform the isolation */ | |
e1409c32 JK |
1829 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1830 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1831 | |
cb2dcaf0 | 1832 | if (!low_pfn) |
edc2ca61 VB |
1833 | return ISOLATE_ABORT; |
1834 | ||
1835 | /* | |
1836 | * Either we isolated something and proceed with migration. Or | |
1837 | * we failed and compact_zone should decide if we should | |
1838 | * continue or not. | |
1839 | */ | |
1840 | break; | |
1841 | } | |
1842 | ||
f2849aa0 VB |
1843 | /* Record where migration scanner will be restarted. */ |
1844 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1845 | |
edc2ca61 | 1846 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1847 | } |
1848 | ||
21c527a3 YB |
1849 | /* |
1850 | * order == -1 is expected when compacting via | |
1851 | * /proc/sys/vm/compact_memory | |
1852 | */ | |
1853 | static inline bool is_via_compact_memory(int order) | |
1854 | { | |
1855 | return order == -1; | |
1856 | } | |
1857 | ||
40cacbcb | 1858 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 1859 | { |
8fb74b9f | 1860 | unsigned int order; |
d39773a0 | 1861 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 1862 | int ret; |
748446bb | 1863 | |
753341a4 | 1864 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1865 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1866 | /* Let the next compaction start anew. */ |
40cacbcb | 1867 | reset_cached_positions(cc->zone); |
55b7c4c9 | 1868 | |
62997027 MG |
1869 | /* |
1870 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1871 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1872 | * flag itself as the decision to be clear should be directly |
1873 | * based on an allocation request. | |
1874 | */ | |
accf6242 | 1875 | if (cc->direct_compaction) |
40cacbcb | 1876 | cc->zone->compact_blockskip_flush = true; |
62997027 | 1877 | |
c8f7de0b MH |
1878 | if (cc->whole_zone) |
1879 | return COMPACT_COMPLETE; | |
1880 | else | |
1881 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1882 | } |
748446bb | 1883 | |
21c527a3 | 1884 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1885 | return COMPACT_CONTINUE; |
1886 | ||
efe771c7 MG |
1887 | /* |
1888 | * Always finish scanning a pageblock to reduce the possibility of | |
1889 | * fallbacks in the future. This is particularly important when | |
1890 | * migration source is unmovable/reclaimable but it's not worth | |
1891 | * special casing. | |
1892 | */ | |
1893 | if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1894 | return COMPACT_CONTINUE; | |
baf6a9a1 | 1895 | |
56de7263 | 1896 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 1897 | ret = COMPACT_NO_SUITABLE_PAGE; |
8fb74b9f | 1898 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 1899 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 1900 | bool can_steal; |
8fb74b9f MG |
1901 | |
1902 | /* Job done if page is free of the right migratetype */ | |
b03641af | 1903 | if (!free_area_empty(area, migratetype)) |
cf378319 | 1904 | return COMPACT_SUCCESS; |
8fb74b9f | 1905 | |
2149cdae JK |
1906 | #ifdef CONFIG_CMA |
1907 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1908 | if (migratetype == MIGRATE_MOVABLE && | |
b03641af | 1909 | !free_area_empty(area, MIGRATE_CMA)) |
cf378319 | 1910 | return COMPACT_SUCCESS; |
2149cdae JK |
1911 | #endif |
1912 | /* | |
1913 | * Job done if allocation would steal freepages from | |
1914 | * other migratetype buddy lists. | |
1915 | */ | |
1916 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
1917 | true, &can_steal) != -1) { |
1918 | ||
1919 | /* movable pages are OK in any pageblock */ | |
1920 | if (migratetype == MIGRATE_MOVABLE) | |
1921 | return COMPACT_SUCCESS; | |
1922 | ||
1923 | /* | |
1924 | * We are stealing for a non-movable allocation. Make | |
1925 | * sure we finish compacting the current pageblock | |
1926 | * first so it is as free as possible and we won't | |
1927 | * have to steal another one soon. This only applies | |
1928 | * to sync compaction, as async compaction operates | |
1929 | * on pageblocks of the same migratetype. | |
1930 | */ | |
1931 | if (cc->mode == MIGRATE_ASYNC || | |
1932 | IS_ALIGNED(cc->migrate_pfn, | |
1933 | pageblock_nr_pages)) { | |
1934 | return COMPACT_SUCCESS; | |
1935 | } | |
1936 | ||
cb2dcaf0 MG |
1937 | ret = COMPACT_CONTINUE; |
1938 | break; | |
baf6a9a1 | 1939 | } |
56de7263 MG |
1940 | } |
1941 | ||
cb2dcaf0 MG |
1942 | if (cc->contended || fatal_signal_pending(current)) |
1943 | ret = COMPACT_CONTENDED; | |
1944 | ||
1945 | return ret; | |
837d026d JK |
1946 | } |
1947 | ||
40cacbcb | 1948 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
1949 | { |
1950 | int ret; | |
1951 | ||
40cacbcb MG |
1952 | ret = __compact_finished(cc); |
1953 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
1954 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
1955 | ret = COMPACT_CONTINUE; | |
1956 | ||
1957 | return ret; | |
748446bb MG |
1958 | } |
1959 | ||
3e7d3449 MG |
1960 | /* |
1961 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1962 | * Returns | |
1963 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
cf378319 | 1964 | * COMPACT_SUCCESS - If the allocation would succeed without compaction |
3e7d3449 MG |
1965 | * COMPACT_CONTINUE - If compaction should run now |
1966 | */ | |
ea7ab982 | 1967 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1968 | unsigned int alloc_flags, |
86a294a8 MH |
1969 | int classzone_idx, |
1970 | unsigned long wmark_target) | |
3e7d3449 | 1971 | { |
3e7d3449 MG |
1972 | unsigned long watermark; |
1973 | ||
21c527a3 | 1974 | if (is_via_compact_memory(order)) |
3957c776 MH |
1975 | return COMPACT_CONTINUE; |
1976 | ||
a9214443 | 1977 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
1978 | /* |
1979 | * If watermarks for high-order allocation are already met, there | |
1980 | * should be no need for compaction at all. | |
1981 | */ | |
1982 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1983 | alloc_flags)) | |
cf378319 | 1984 | return COMPACT_SUCCESS; |
ebff3980 | 1985 | |
3e7d3449 | 1986 | /* |
9861a62c | 1987 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
1988 | * isolate free pages for migration targets. This means that the |
1989 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1990 | * the check in __isolate_free_page(). We don't use the direct | |
1991 | * compactor's alloc_flags, as they are not relevant for freepage | |
1992 | * isolation. We however do use the direct compactor's classzone_idx to | |
1993 | * skip over zones where lowmem reserves would prevent allocation even | |
1994 | * if compaction succeeds. | |
8348faf9 VB |
1995 | * For costly orders, we require low watermark instead of min for |
1996 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
1997 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
1998 | * suitable migration targets | |
3e7d3449 | 1999 | */ |
8348faf9 VB |
2000 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
2001 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
2002 | watermark += compact_gap(order); | |
86a294a8 | 2003 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
d883c6cf | 2004 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
2005 | return COMPACT_SKIPPED; |
2006 | ||
cc5c9f09 VB |
2007 | return COMPACT_CONTINUE; |
2008 | } | |
2009 | ||
2010 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
2011 | unsigned int alloc_flags, | |
2012 | int classzone_idx) | |
2013 | { | |
2014 | enum compact_result ret; | |
2015 | int fragindex; | |
2016 | ||
2017 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
2018 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
2019 | /* |
2020 | * fragmentation index determines if allocation failures are due to | |
2021 | * low memory or external fragmentation | |
2022 | * | |
ebff3980 VB |
2023 | * index of -1000 would imply allocations might succeed depending on |
2024 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2025 | * index towards 0 implies failure is due to lack of memory |
2026 | * index towards 1000 implies failure is due to fragmentation | |
2027 | * | |
20311420 VB |
2028 | * Only compact if a failure would be due to fragmentation. Also |
2029 | * ignore fragindex for non-costly orders where the alternative to | |
2030 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2031 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2032 | * excessive compaction for costly orders, but it should not be at the | |
2033 | * expense of system stability. | |
3e7d3449 | 2034 | */ |
20311420 | 2035 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
2036 | fragindex = fragmentation_index(zone, order); |
2037 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
2038 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
2039 | } | |
837d026d | 2040 | |
837d026d JK |
2041 | trace_mm_compaction_suitable(zone, order, ret); |
2042 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
2043 | ret = COMPACT_SKIPPED; | |
2044 | ||
2045 | return ret; | |
2046 | } | |
2047 | ||
86a294a8 MH |
2048 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2049 | int alloc_flags) | |
2050 | { | |
2051 | struct zone *zone; | |
2052 | struct zoneref *z; | |
2053 | ||
2054 | /* | |
2055 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2056 | * retrying the reclaim. | |
2057 | */ | |
2058 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
2059 | ac->nodemask) { | |
2060 | unsigned long available; | |
2061 | enum compact_result compact_result; | |
2062 | ||
2063 | /* | |
2064 | * Do not consider all the reclaimable memory because we do not | |
2065 | * want to trash just for a single high order allocation which | |
2066 | * is even not guaranteed to appear even if __compaction_suitable | |
2067 | * is happy about the watermark check. | |
2068 | */ | |
5a1c84b4 | 2069 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
2070 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
2071 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
2072 | ac_classzone_idx(ac), available); | |
cc5c9f09 | 2073 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
2074 | return true; |
2075 | } | |
2076 | ||
2077 | return false; | |
2078 | } | |
2079 | ||
5e1f0f09 MG |
2080 | static enum compact_result |
2081 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2082 | { |
ea7ab982 | 2083 | enum compact_result ret; |
40cacbcb MG |
2084 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2085 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2086 | unsigned long last_migrated_pfn; |
e0b9daeb | 2087 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2088 | bool update_cached; |
748446bb | 2089 | |
a94b5252 YS |
2090 | /* |
2091 | * These counters track activities during zone compaction. Initialize | |
2092 | * them before compacting a new zone. | |
2093 | */ | |
2094 | cc->total_migrate_scanned = 0; | |
2095 | cc->total_free_scanned = 0; | |
2096 | cc->nr_migratepages = 0; | |
2097 | cc->nr_freepages = 0; | |
2098 | INIT_LIST_HEAD(&cc->freepages); | |
2099 | INIT_LIST_HEAD(&cc->migratepages); | |
2100 | ||
d39773a0 | 2101 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
40cacbcb | 2102 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
ebff3980 | 2103 | cc->classzone_idx); |
c46649de | 2104 | /* Compaction is likely to fail */ |
cf378319 | 2105 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 2106 | return ret; |
c46649de MH |
2107 | |
2108 | /* huh, compaction_suitable is returning something unexpected */ | |
2109 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 2110 | |
d3132e4b VB |
2111 | /* |
2112 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2113 | * is about to be retried after being deferred. |
d3132e4b | 2114 | */ |
40cacbcb MG |
2115 | if (compaction_restarting(cc->zone, cc->order)) |
2116 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2117 | |
c89511ab MG |
2118 | /* |
2119 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2120 | * information on where the scanners should start (unless we explicitly |
2121 | * want to compact the whole zone), but check that it is initialised | |
2122 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2123 | */ |
70b44595 | 2124 | cc->fast_start_pfn = 0; |
06ed2998 | 2125 | if (cc->whole_zone) { |
c89511ab | 2126 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2127 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2128 | } else { | |
40cacbcb MG |
2129 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2130 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2131 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2132 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2133 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2134 | } |
2135 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2136 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2137 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2138 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2139 | } |
c8f7de0b | 2140 | |
e332f741 | 2141 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2142 | cc->whole_zone = true; |
2143 | } | |
c8f7de0b | 2144 | |
566e54e1 | 2145 | last_migrated_pfn = 0; |
748446bb | 2146 | |
8854c55f MG |
2147 | /* |
2148 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2149 | * the basis that some migrations will fail in ASYNC mode. However, | |
2150 | * if the cached PFNs match and pageblocks are skipped due to having | |
2151 | * no isolation candidates, then the sync state does not matter. | |
2152 | * Until a pageblock with isolation candidates is found, keep the | |
2153 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2154 | */ | |
2155 | update_cached = !sync && | |
2156 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2157 | ||
16c4a097 JK |
2158 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
2159 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 2160 | |
748446bb MG |
2161 | migrate_prep_local(); |
2162 | ||
40cacbcb | 2163 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2164 | int err; |
566e54e1 | 2165 | unsigned long start_pfn = cc->migrate_pfn; |
748446bb | 2166 | |
804d3121 MG |
2167 | /* |
2168 | * Avoid multiple rescans which can happen if a page cannot be | |
2169 | * isolated (dirty/writeback in async mode) or if the migrated | |
2170 | * pages are being allocated before the pageblock is cleared. | |
2171 | * The first rescan will capture the entire pageblock for | |
2172 | * migration. If it fails, it'll be marked skip and scanning | |
2173 | * will proceed as normal. | |
2174 | */ | |
2175 | cc->rescan = false; | |
2176 | if (pageblock_start_pfn(last_migrated_pfn) == | |
2177 | pageblock_start_pfn(start_pfn)) { | |
2178 | cc->rescan = true; | |
2179 | } | |
2180 | ||
32aaf055 | 2181 | switch (isolate_migratepages(cc)) { |
f9e35b3b | 2182 | case ISOLATE_ABORT: |
2d1e1041 | 2183 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2184 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2185 | cc->nr_migratepages = 0; |
f9e35b3b MG |
2186 | goto out; |
2187 | case ISOLATE_NONE: | |
8854c55f MG |
2188 | if (update_cached) { |
2189 | cc->zone->compact_cached_migrate_pfn[1] = | |
2190 | cc->zone->compact_cached_migrate_pfn[0]; | |
2191 | } | |
2192 | ||
fdaf7f5c VB |
2193 | /* |
2194 | * We haven't isolated and migrated anything, but | |
2195 | * there might still be unflushed migrations from | |
2196 | * previous cc->order aligned block. | |
2197 | */ | |
2198 | goto check_drain; | |
f9e35b3b | 2199 | case ISOLATE_SUCCESS: |
8854c55f | 2200 | update_cached = false; |
566e54e1 | 2201 | last_migrated_pfn = start_pfn; |
f9e35b3b MG |
2202 | ; |
2203 | } | |
748446bb | 2204 | |
d53aea3d | 2205 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2206 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 2207 | MR_COMPACTION); |
748446bb | 2208 | |
f8c9301f VB |
2209 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
2210 | &cc->migratepages); | |
748446bb | 2211 | |
f8c9301f VB |
2212 | /* All pages were either migrated or will be released */ |
2213 | cc->nr_migratepages = 0; | |
9d502c1c | 2214 | if (err) { |
5733c7d1 | 2215 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2216 | /* |
2217 | * migrate_pages() may return -ENOMEM when scanners meet | |
2218 | * and we want compact_finished() to detect it | |
2219 | */ | |
f2849aa0 | 2220 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2221 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2222 | goto out; |
2223 | } | |
fdd048e1 VB |
2224 | /* |
2225 | * We failed to migrate at least one page in the current | |
2226 | * order-aligned block, so skip the rest of it. | |
2227 | */ | |
2228 | if (cc->direct_compaction && | |
2229 | (cc->mode == MIGRATE_ASYNC)) { | |
2230 | cc->migrate_pfn = block_end_pfn( | |
2231 | cc->migrate_pfn - 1, cc->order); | |
2232 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 2233 | last_migrated_pfn = 0; |
fdd048e1 | 2234 | } |
748446bb | 2235 | } |
fdaf7f5c | 2236 | |
fdaf7f5c VB |
2237 | check_drain: |
2238 | /* | |
2239 | * Has the migration scanner moved away from the previous | |
2240 | * cc->order aligned block where we migrated from? If yes, | |
2241 | * flush the pages that were freed, so that they can merge and | |
2242 | * compact_finished() can detect immediately if allocation | |
2243 | * would succeed. | |
2244 | */ | |
566e54e1 | 2245 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c VB |
2246 | int cpu; |
2247 | unsigned long current_block_start = | |
06b6640a | 2248 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2249 | |
566e54e1 | 2250 | if (last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
2251 | cpu = get_cpu(); |
2252 | lru_add_drain_cpu(cpu); | |
40cacbcb | 2253 | drain_local_pages(cc->zone); |
fdaf7f5c VB |
2254 | put_cpu(); |
2255 | /* No more flushing until we migrate again */ | |
566e54e1 | 2256 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2257 | } |
2258 | } | |
2259 | ||
5e1f0f09 MG |
2260 | /* Stop if a page has been captured */ |
2261 | if (capc && capc->page) { | |
2262 | ret = COMPACT_SUCCESS; | |
2263 | break; | |
2264 | } | |
748446bb MG |
2265 | } |
2266 | ||
f9e35b3b | 2267 | out: |
6bace090 VB |
2268 | /* |
2269 | * Release free pages and update where the free scanner should restart, | |
2270 | * so we don't leave any returned pages behind in the next attempt. | |
2271 | */ | |
2272 | if (cc->nr_freepages > 0) { | |
2273 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2274 | ||
2275 | cc->nr_freepages = 0; | |
2276 | VM_BUG_ON(free_pfn == 0); | |
2277 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2278 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2279 | /* |
2280 | * Only go back, not forward. The cached pfn might have been | |
2281 | * already reset to zone end in compact_finished() | |
2282 | */ | |
40cacbcb MG |
2283 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2284 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2285 | } |
748446bb | 2286 | |
7f354a54 DR |
2287 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2288 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2289 | ||
16c4a097 JK |
2290 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
2291 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 2292 | |
748446bb MG |
2293 | return ret; |
2294 | } | |
76ab0f53 | 2295 | |
ea7ab982 | 2296 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2297 | gfp_t gfp_mask, enum compact_priority prio, |
5e1f0f09 MG |
2298 | unsigned int alloc_flags, int classzone_idx, |
2299 | struct page **capture) | |
56de7263 | 2300 | { |
ea7ab982 | 2301 | enum compact_result ret; |
56de7263 | 2302 | struct compact_control cc = { |
56de7263 | 2303 | .order = order, |
dbe2d4e4 | 2304 | .search_order = order, |
6d7ce559 | 2305 | .gfp_mask = gfp_mask, |
56de7263 | 2306 | .zone = zone, |
a5508cd8 VB |
2307 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2308 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 VB |
2309 | .alloc_flags = alloc_flags, |
2310 | .classzone_idx = classzone_idx, | |
accf6242 | 2311 | .direct_compaction = true, |
a8e025e5 | 2312 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2313 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2314 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2315 | }; |
5e1f0f09 MG |
2316 | struct capture_control capc = { |
2317 | .cc = &cc, | |
2318 | .page = NULL, | |
2319 | }; | |
2320 | ||
6467552c | 2321 | current->capture_control = &capc; |
56de7263 | 2322 | |
5e1f0f09 | 2323 | ret = compact_zone(&cc, &capc); |
e64c5237 SL |
2324 | |
2325 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2326 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2327 | ||
5e1f0f09 MG |
2328 | *capture = capc.page; |
2329 | current->capture_control = NULL; | |
2330 | ||
e64c5237 | 2331 | return ret; |
56de7263 MG |
2332 | } |
2333 | ||
5e771905 MG |
2334 | int sysctl_extfrag_threshold = 500; |
2335 | ||
56de7263 MG |
2336 | /** |
2337 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2338 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2339 | * @order: The order of the current allocation |
2340 | * @alloc_flags: The allocation flags of the current allocation | |
2341 | * @ac: The context of current allocation | |
112d2d29 | 2342 | * @prio: Determines how hard direct compaction should try to succeed |
6467552c | 2343 | * @capture: Pointer to free page created by compaction will be stored here |
56de7263 MG |
2344 | * |
2345 | * This is the main entry point for direct page compaction. | |
2346 | */ | |
ea7ab982 | 2347 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2348 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2349 | enum compact_priority prio, struct page **capture) |
56de7263 | 2350 | { |
56de7263 | 2351 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
2352 | struct zoneref *z; |
2353 | struct zone *zone; | |
1d4746d3 | 2354 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2355 | |
73e64c51 MH |
2356 | /* |
2357 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2358 | * tricky context because the migration might require IO | |
2359 | */ | |
2360 | if (!may_perform_io) | |
53853e2d | 2361 | return COMPACT_SKIPPED; |
56de7263 | 2362 | |
a5508cd8 | 2363 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2364 | |
56de7263 | 2365 | /* Compact each zone in the list */ |
1a6d53a1 VB |
2366 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
2367 | ac->nodemask) { | |
ea7ab982 | 2368 | enum compact_result status; |
56de7263 | 2369 | |
a8e025e5 VB |
2370 | if (prio > MIN_COMPACT_PRIORITY |
2371 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2372 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2373 | continue; |
1d4746d3 | 2374 | } |
53853e2d | 2375 | |
a5508cd8 | 2376 | status = compact_zone_order(zone, order, gfp_mask, prio, |
5e1f0f09 | 2377 | alloc_flags, ac_classzone_idx(ac), capture); |
56de7263 MG |
2378 | rc = max(status, rc); |
2379 | ||
7ceb009a VB |
2380 | /* The allocation should succeed, stop compacting */ |
2381 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2382 | /* |
2383 | * We think the allocation will succeed in this zone, | |
2384 | * but it is not certain, hence the false. The caller | |
2385 | * will repeat this with true if allocation indeed | |
2386 | * succeeds in this zone. | |
2387 | */ | |
2388 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2389 | |
c3486f53 | 2390 | break; |
1f9efdef VB |
2391 | } |
2392 | ||
a5508cd8 | 2393 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2394 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2395 | /* |
2396 | * We think that allocation won't succeed in this zone | |
2397 | * so we defer compaction there. If it ends up | |
2398 | * succeeding after all, it will be reset. | |
2399 | */ | |
2400 | defer_compaction(zone, order); | |
1f9efdef VB |
2401 | |
2402 | /* | |
2403 | * We might have stopped compacting due to need_resched() in | |
2404 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2405 | * case do not try further zones |
1f9efdef | 2406 | */ |
c3486f53 VB |
2407 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2408 | || fatal_signal_pending(current)) | |
2409 | break; | |
56de7263 MG |
2410 | } |
2411 | ||
2412 | return rc; | |
2413 | } | |
2414 | ||
2415 | ||
76ab0f53 | 2416 | /* Compact all zones within a node */ |
791cae96 | 2417 | static void compact_node(int nid) |
76ab0f53 | 2418 | { |
791cae96 | 2419 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 2420 | int zoneid; |
76ab0f53 | 2421 | struct zone *zone; |
791cae96 VB |
2422 | struct compact_control cc = { |
2423 | .order = -1, | |
2424 | .mode = MIGRATE_SYNC, | |
2425 | .ignore_skip_hint = true, | |
2426 | .whole_zone = true, | |
73e64c51 | 2427 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
2428 | }; |
2429 | ||
76ab0f53 | 2430 | |
76ab0f53 | 2431 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
2432 | |
2433 | zone = &pgdat->node_zones[zoneid]; | |
2434 | if (!populated_zone(zone)) | |
2435 | continue; | |
2436 | ||
791cae96 | 2437 | cc.zone = zone; |
76ab0f53 | 2438 | |
5e1f0f09 | 2439 | compact_zone(&cc, NULL); |
75469345 | 2440 | |
791cae96 VB |
2441 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2442 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 2443 | } |
76ab0f53 MG |
2444 | } |
2445 | ||
2446 | /* Compact all nodes in the system */ | |
7964c06d | 2447 | static void compact_nodes(void) |
76ab0f53 MG |
2448 | { |
2449 | int nid; | |
2450 | ||
8575ec29 HD |
2451 | /* Flush pending updates to the LRU lists */ |
2452 | lru_add_drain_all(); | |
2453 | ||
76ab0f53 MG |
2454 | for_each_online_node(nid) |
2455 | compact_node(nid); | |
76ab0f53 MG |
2456 | } |
2457 | ||
2458 | /* The written value is actually unused, all memory is compacted */ | |
2459 | int sysctl_compact_memory; | |
2460 | ||
fec4eb2c YB |
2461 | /* |
2462 | * This is the entry point for compacting all nodes via | |
2463 | * /proc/sys/vm/compact_memory | |
2464 | */ | |
76ab0f53 MG |
2465 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
2466 | void __user *buffer, size_t *length, loff_t *ppos) | |
2467 | { | |
2468 | if (write) | |
7964c06d | 2469 | compact_nodes(); |
76ab0f53 MG |
2470 | |
2471 | return 0; | |
2472 | } | |
ed4a6d7f MG |
2473 | |
2474 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
74e77fb9 | 2475 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 2476 | struct device_attribute *attr, |
ed4a6d7f MG |
2477 | const char *buf, size_t count) |
2478 | { | |
8575ec29 HD |
2479 | int nid = dev->id; |
2480 | ||
2481 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2482 | /* Flush pending updates to the LRU lists */ | |
2483 | lru_add_drain_all(); | |
2484 | ||
2485 | compact_node(nid); | |
2486 | } | |
ed4a6d7f MG |
2487 | |
2488 | return count; | |
2489 | } | |
0825a6f9 | 2490 | static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node); |
ed4a6d7f MG |
2491 | |
2492 | int compaction_register_node(struct node *node) | |
2493 | { | |
10fbcf4c | 2494 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2495 | } |
2496 | ||
2497 | void compaction_unregister_node(struct node *node) | |
2498 | { | |
10fbcf4c | 2499 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2500 | } |
2501 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2502 | |
698b1b30 VB |
2503 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2504 | { | |
172400c6 | 2505 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
2506 | } |
2507 | ||
2508 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2509 | { | |
2510 | int zoneid; | |
2511 | struct zone *zone; | |
2512 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
2513 | ||
6cd9dc3e | 2514 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
2515 | zone = &pgdat->node_zones[zoneid]; |
2516 | ||
2517 | if (!populated_zone(zone)) | |
2518 | continue; | |
2519 | ||
2520 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
2521 | classzone_idx) == COMPACT_CONTINUE) | |
2522 | return true; | |
2523 | } | |
2524 | ||
2525 | return false; | |
2526 | } | |
2527 | ||
2528 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2529 | { | |
2530 | /* | |
2531 | * With no special task, compact all zones so that a page of requested | |
2532 | * order is allocatable. | |
2533 | */ | |
2534 | int zoneid; | |
2535 | struct zone *zone; | |
2536 | struct compact_control cc = { | |
2537 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 2538 | .search_order = pgdat->kcompactd_max_order, |
698b1b30 VB |
2539 | .classzone_idx = pgdat->kcompactd_classzone_idx, |
2540 | .mode = MIGRATE_SYNC_LIGHT, | |
a0647dc9 | 2541 | .ignore_skip_hint = false, |
73e64c51 | 2542 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2543 | }; |
698b1b30 VB |
2544 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
2545 | cc.classzone_idx); | |
7f354a54 | 2546 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2547 | |
6cd9dc3e | 2548 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
2549 | int status; |
2550 | ||
2551 | zone = &pgdat->node_zones[zoneid]; | |
2552 | if (!populated_zone(zone)) | |
2553 | continue; | |
2554 | ||
2555 | if (compaction_deferred(zone, cc.order)) | |
2556 | continue; | |
2557 | ||
2558 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
2559 | COMPACT_CONTINUE) | |
2560 | continue; | |
2561 | ||
172400c6 VB |
2562 | if (kthread_should_stop()) |
2563 | return; | |
a94b5252 YS |
2564 | |
2565 | cc.zone = zone; | |
5e1f0f09 | 2566 | status = compact_zone(&cc, NULL); |
698b1b30 | 2567 | |
7ceb009a | 2568 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 2569 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 2570 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
2571 | /* |
2572 | * Buddy pages may become stranded on pcps that could | |
2573 | * otherwise coalesce on the zone's free area for | |
2574 | * order >= cc.order. This is ratelimited by the | |
2575 | * upcoming deferral. | |
2576 | */ | |
2577 | drain_all_pages(zone); | |
2578 | ||
698b1b30 VB |
2579 | /* |
2580 | * We use sync migration mode here, so we defer like | |
2581 | * sync direct compaction does. | |
2582 | */ | |
2583 | defer_compaction(zone, cc.order); | |
2584 | } | |
2585 | ||
7f354a54 DR |
2586 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2587 | cc.total_migrate_scanned); | |
2588 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2589 | cc.total_free_scanned); | |
2590 | ||
698b1b30 VB |
2591 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2592 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2593 | } | |
2594 | ||
2595 | /* | |
2596 | * Regardless of success, we are done until woken up next. But remember | |
2597 | * the requested order/classzone_idx in case it was higher/tighter than | |
2598 | * our current ones | |
2599 | */ | |
2600 | if (pgdat->kcompactd_max_order <= cc.order) | |
2601 | pgdat->kcompactd_max_order = 0; | |
2602 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
2603 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2604 | } | |
2605 | ||
2606 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
2607 | { | |
2608 | if (!order) | |
2609 | return; | |
2610 | ||
2611 | if (pgdat->kcompactd_max_order < order) | |
2612 | pgdat->kcompactd_max_order = order; | |
2613 | ||
2614 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2615 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2616 | ||
6818600f DB |
2617 | /* |
2618 | * Pairs with implicit barrier in wait_event_freezable() | |
2619 | * such that wakeups are not missed. | |
2620 | */ | |
2621 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2622 | return; |
2623 | ||
2624 | if (!kcompactd_node_suitable(pgdat)) | |
2625 | return; | |
2626 | ||
2627 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2628 | classzone_idx); | |
2629 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2630 | } | |
2631 | ||
2632 | /* | |
2633 | * The background compaction daemon, started as a kernel thread | |
2634 | * from the init process. | |
2635 | */ | |
2636 | static int kcompactd(void *p) | |
2637 | { | |
2638 | pg_data_t *pgdat = (pg_data_t*)p; | |
2639 | struct task_struct *tsk = current; | |
2640 | ||
2641 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2642 | ||
2643 | if (!cpumask_empty(cpumask)) | |
2644 | set_cpus_allowed_ptr(tsk, cpumask); | |
2645 | ||
2646 | set_freezable(); | |
2647 | ||
2648 | pgdat->kcompactd_max_order = 0; | |
2649 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2650 | ||
2651 | while (!kthread_should_stop()) { | |
eb414681 JW |
2652 | unsigned long pflags; |
2653 | ||
698b1b30 VB |
2654 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
2655 | wait_event_freezable(pgdat->kcompactd_wait, | |
2656 | kcompactd_work_requested(pgdat)); | |
2657 | ||
eb414681 | 2658 | psi_memstall_enter(&pflags); |
698b1b30 | 2659 | kcompactd_do_work(pgdat); |
eb414681 | 2660 | psi_memstall_leave(&pflags); |
698b1b30 VB |
2661 | } |
2662 | ||
2663 | return 0; | |
2664 | } | |
2665 | ||
2666 | /* | |
2667 | * This kcompactd start function will be called by init and node-hot-add. | |
2668 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2669 | */ | |
2670 | int kcompactd_run(int nid) | |
2671 | { | |
2672 | pg_data_t *pgdat = NODE_DATA(nid); | |
2673 | int ret = 0; | |
2674 | ||
2675 | if (pgdat->kcompactd) | |
2676 | return 0; | |
2677 | ||
2678 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2679 | if (IS_ERR(pgdat->kcompactd)) { | |
2680 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2681 | ret = PTR_ERR(pgdat->kcompactd); | |
2682 | pgdat->kcompactd = NULL; | |
2683 | } | |
2684 | return ret; | |
2685 | } | |
2686 | ||
2687 | /* | |
2688 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2689 | * hold mem_hotplug_begin/end(). | |
2690 | */ | |
2691 | void kcompactd_stop(int nid) | |
2692 | { | |
2693 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2694 | ||
2695 | if (kcompactd) { | |
2696 | kthread_stop(kcompactd); | |
2697 | NODE_DATA(nid)->kcompactd = NULL; | |
2698 | } | |
2699 | } | |
2700 | ||
2701 | /* | |
2702 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2703 | * not required for correctness. So if the last cpu in a node goes | |
2704 | * away, we get changed to run anywhere: as the first one comes back, | |
2705 | * restore their cpu bindings. | |
2706 | */ | |
e46b1db2 | 2707 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
2708 | { |
2709 | int nid; | |
2710 | ||
e46b1db2 AMG |
2711 | for_each_node_state(nid, N_MEMORY) { |
2712 | pg_data_t *pgdat = NODE_DATA(nid); | |
2713 | const struct cpumask *mask; | |
698b1b30 | 2714 | |
e46b1db2 | 2715 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 2716 | |
e46b1db2 AMG |
2717 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
2718 | /* One of our CPUs online: restore mask */ | |
2719 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 2720 | } |
e46b1db2 | 2721 | return 0; |
698b1b30 VB |
2722 | } |
2723 | ||
2724 | static int __init kcompactd_init(void) | |
2725 | { | |
2726 | int nid; | |
e46b1db2 AMG |
2727 | int ret; |
2728 | ||
2729 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2730 | "mm/compaction:online", | |
2731 | kcompactd_cpu_online, NULL); | |
2732 | if (ret < 0) { | |
2733 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2734 | return ret; | |
2735 | } | |
698b1b30 VB |
2736 | |
2737 | for_each_node_state(nid, N_MEMORY) | |
2738 | kcompactd_run(nid); | |
698b1b30 VB |
2739 | return 0; |
2740 | } | |
2741 | subsys_initcall(kcompactd_init) | |
2742 | ||
ff9543fd | 2743 | #endif /* CONFIG_COMPACTION */ |