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 | 28 | #ifdef CONFIG_COMPACTION |
31ca72fa CTK |
29 | /* |
30 | * Fragmentation score check interval for proactive compaction purposes. | |
31 | */ | |
32 | #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500) | |
33 | ||
010fc29a MK |
34 | static inline void count_compact_event(enum vm_event_item item) |
35 | { | |
36 | count_vm_event(item); | |
37 | } | |
38 | ||
39 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
40 | { | |
41 | count_vm_events(item, delta); | |
42 | } | |
43 | #else | |
44 | #define count_compact_event(item) do { } while (0) | |
45 | #define count_compact_events(item, delta) do { } while (0) | |
46 | #endif | |
47 | ||
ff9543fd MN |
48 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
49 | ||
b7aba698 MG |
50 | #define CREATE_TRACE_POINTS |
51 | #include <trace/events/compaction.h> | |
52 | ||
06b6640a VB |
53 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
54 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
06b6640a | 55 | |
facdaa91 NG |
56 | /* |
57 | * Page order with-respect-to which proactive compaction | |
58 | * calculates external fragmentation, which is used as | |
59 | * the "fragmentation score" of a node/zone. | |
60 | */ | |
61 | #if defined CONFIG_TRANSPARENT_HUGEPAGE | |
62 | #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER | |
25788738 | 63 | #elif defined CONFIG_HUGETLBFS |
facdaa91 NG |
64 | #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER |
65 | #else | |
66 | #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
67 | #endif | |
68 | ||
748446bb MG |
69 | static unsigned long release_freepages(struct list_head *freelist) |
70 | { | |
71 | struct page *page, *next; | |
6bace090 | 72 | unsigned long high_pfn = 0; |
748446bb MG |
73 | |
74 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 75 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
76 | list_del(&page->lru); |
77 | __free_page(page); | |
6bace090 VB |
78 | if (pfn > high_pfn) |
79 | high_pfn = pfn; | |
748446bb MG |
80 | } |
81 | ||
6bace090 | 82 | return high_pfn; |
748446bb MG |
83 | } |
84 | ||
4469ab98 | 85 | static void split_map_pages(struct list_head *list) |
ff9543fd | 86 | { |
66c64223 JK |
87 | unsigned int i, order, nr_pages; |
88 | struct page *page, *next; | |
89 | LIST_HEAD(tmp_list); | |
90 | ||
91 | list_for_each_entry_safe(page, next, list, lru) { | |
92 | list_del(&page->lru); | |
93 | ||
94 | order = page_private(page); | |
95 | nr_pages = 1 << order; | |
66c64223 | 96 | |
46f24fd8 | 97 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
98 | if (order) |
99 | split_page(page, order); | |
ff9543fd | 100 | |
66c64223 JK |
101 | for (i = 0; i < nr_pages; i++) { |
102 | list_add(&page->lru, &tmp_list); | |
103 | page++; | |
104 | } | |
ff9543fd | 105 | } |
66c64223 JK |
106 | |
107 | list_splice(&tmp_list, list); | |
ff9543fd MN |
108 | } |
109 | ||
bb13ffeb | 110 | #ifdef CONFIG_COMPACTION |
68f2736a | 111 | bool PageMovable(struct page *page) |
bda807d4 | 112 | { |
68f2736a | 113 | const struct movable_operations *mops; |
bda807d4 MK |
114 | |
115 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
116 | if (!__PageMovable(page)) | |
68f2736a | 117 | return false; |
bda807d4 | 118 | |
68f2736a MWO |
119 | mops = page_movable_ops(page); |
120 | if (mops) | |
121 | return true; | |
bda807d4 | 122 | |
68f2736a | 123 | return false; |
bda807d4 MK |
124 | } |
125 | EXPORT_SYMBOL(PageMovable); | |
126 | ||
68f2736a | 127 | void __SetPageMovable(struct page *page, const struct movable_operations *mops) |
bda807d4 MK |
128 | { |
129 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
68f2736a MWO |
130 | VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page); |
131 | page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE); | |
bda807d4 MK |
132 | } |
133 | EXPORT_SYMBOL(__SetPageMovable); | |
134 | ||
135 | void __ClearPageMovable(struct page *page) | |
136 | { | |
bda807d4 MK |
137 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
138 | /* | |
68f2736a MWO |
139 | * This page still has the type of a movable page, but it's |
140 | * actually not movable any more. | |
bda807d4 | 141 | */ |
68f2736a | 142 | page->mapping = (void *)PAGE_MAPPING_MOVABLE; |
bda807d4 MK |
143 | } |
144 | EXPORT_SYMBOL(__ClearPageMovable); | |
145 | ||
24e2716f JK |
146 | /* Do not skip compaction more than 64 times */ |
147 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
148 | ||
149 | /* | |
150 | * Compaction is deferred when compaction fails to result in a page | |
860b3272 | 151 | * allocation success. 1 << compact_defer_shift, compactions are skipped up |
24e2716f JK |
152 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT |
153 | */ | |
2271b016 | 154 | static void defer_compaction(struct zone *zone, int order) |
24e2716f JK |
155 | { |
156 | zone->compact_considered = 0; | |
157 | zone->compact_defer_shift++; | |
158 | ||
159 | if (order < zone->compact_order_failed) | |
160 | zone->compact_order_failed = order; | |
161 | ||
162 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
163 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
164 | ||
165 | trace_mm_compaction_defer_compaction(zone, order); | |
166 | } | |
167 | ||
168 | /* Returns true if compaction should be skipped this time */ | |
2271b016 | 169 | static bool compaction_deferred(struct zone *zone, int order) |
24e2716f JK |
170 | { |
171 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
172 | ||
173 | if (order < zone->compact_order_failed) | |
174 | return false; | |
175 | ||
176 | /* Avoid possible overflow */ | |
62b35fe0 | 177 | if (++zone->compact_considered >= defer_limit) { |
24e2716f | 178 | zone->compact_considered = defer_limit; |
24e2716f | 179 | return false; |
62b35fe0 | 180 | } |
24e2716f JK |
181 | |
182 | trace_mm_compaction_deferred(zone, order); | |
183 | ||
184 | return true; | |
185 | } | |
186 | ||
187 | /* | |
188 | * Update defer tracking counters after successful compaction of given order, | |
189 | * which means an allocation either succeeded (alloc_success == true) or is | |
190 | * expected to succeed. | |
191 | */ | |
192 | void compaction_defer_reset(struct zone *zone, int order, | |
193 | bool alloc_success) | |
194 | { | |
195 | if (alloc_success) { | |
196 | zone->compact_considered = 0; | |
197 | zone->compact_defer_shift = 0; | |
198 | } | |
199 | if (order >= zone->compact_order_failed) | |
200 | zone->compact_order_failed = order + 1; | |
201 | ||
202 | trace_mm_compaction_defer_reset(zone, order); | |
203 | } | |
204 | ||
205 | /* Returns true if restarting compaction after many failures */ | |
2271b016 | 206 | static bool compaction_restarting(struct zone *zone, int order) |
24e2716f JK |
207 | { |
208 | if (order < zone->compact_order_failed) | |
209 | return false; | |
210 | ||
211 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
212 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
213 | } | |
214 | ||
bb13ffeb MG |
215 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
216 | static inline bool isolation_suitable(struct compact_control *cc, | |
217 | struct page *page) | |
218 | { | |
219 | if (cc->ignore_skip_hint) | |
220 | return true; | |
221 | ||
222 | return !get_pageblock_skip(page); | |
223 | } | |
224 | ||
02333641 VB |
225 | static void reset_cached_positions(struct zone *zone) |
226 | { | |
227 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
228 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 229 | zone->compact_cached_free_pfn = |
06b6640a | 230 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
231 | } |
232 | ||
21dc7e02 | 233 | /* |
2271b016 | 234 | * Compound pages of >= pageblock_order should consistently be skipped until |
b527cfe5 VB |
235 | * released. It is always pointless to compact pages of such order (if they are |
236 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 237 | */ |
b527cfe5 | 238 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 239 | { |
b527cfe5 | 240 | if (!PageCompound(page)) |
21dc7e02 | 241 | return false; |
b527cfe5 VB |
242 | |
243 | page = compound_head(page); | |
244 | ||
245 | if (compound_order(page) >= pageblock_order) | |
246 | return true; | |
247 | ||
248 | return false; | |
21dc7e02 DR |
249 | } |
250 | ||
e332f741 MG |
251 | static bool |
252 | __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, | |
253 | bool check_target) | |
254 | { | |
255 | struct page *page = pfn_to_online_page(pfn); | |
6b0868c8 | 256 | struct page *block_page; |
e332f741 MG |
257 | struct page *end_page; |
258 | unsigned long block_pfn; | |
259 | ||
260 | if (!page) | |
261 | return false; | |
262 | if (zone != page_zone(page)) | |
263 | return false; | |
264 | if (pageblock_skip_persistent(page)) | |
265 | return false; | |
266 | ||
267 | /* | |
268 | * If skip is already cleared do no further checking once the | |
269 | * restart points have been set. | |
270 | */ | |
271 | if (check_source && check_target && !get_pageblock_skip(page)) | |
272 | return true; | |
273 | ||
274 | /* | |
275 | * If clearing skip for the target scanner, do not select a | |
276 | * non-movable pageblock as the starting point. | |
277 | */ | |
278 | if (!check_source && check_target && | |
279 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
280 | return false; | |
281 | ||
6b0868c8 MG |
282 | /* Ensure the start of the pageblock or zone is online and valid */ |
283 | block_pfn = pageblock_start_pfn(pfn); | |
a2e9a5af VB |
284 | block_pfn = max(block_pfn, zone->zone_start_pfn); |
285 | block_page = pfn_to_online_page(block_pfn); | |
6b0868c8 MG |
286 | if (block_page) { |
287 | page = block_page; | |
288 | pfn = block_pfn; | |
289 | } | |
290 | ||
291 | /* Ensure the end of the pageblock or zone is online and valid */ | |
a2e9a5af | 292 | block_pfn = pageblock_end_pfn(pfn) - 1; |
6b0868c8 MG |
293 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); |
294 | end_page = pfn_to_online_page(block_pfn); | |
295 | if (!end_page) | |
296 | return false; | |
297 | ||
e332f741 MG |
298 | /* |
299 | * Only clear the hint if a sample indicates there is either a | |
300 | * free page or an LRU page in the block. One or other condition | |
301 | * is necessary for the block to be a migration source/target. | |
302 | */ | |
e332f741 | 303 | do { |
859a85dd MR |
304 | if (check_source && PageLRU(page)) { |
305 | clear_pageblock_skip(page); | |
306 | return true; | |
307 | } | |
e332f741 | 308 | |
859a85dd MR |
309 | if (check_target && PageBuddy(page)) { |
310 | clear_pageblock_skip(page); | |
311 | return true; | |
e332f741 MG |
312 | } |
313 | ||
314 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
a2e9a5af | 315 | } while (page <= end_page); |
e332f741 MG |
316 | |
317 | return false; | |
318 | } | |
319 | ||
bb13ffeb MG |
320 | /* |
321 | * This function is called to clear all cached information on pageblocks that | |
322 | * should be skipped for page isolation when the migrate and free page scanner | |
323 | * meet. | |
324 | */ | |
62997027 | 325 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 326 | { |
e332f741 | 327 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 328 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
329 | unsigned long reset_migrate = free_pfn; |
330 | unsigned long reset_free = migrate_pfn; | |
331 | bool source_set = false; | |
332 | bool free_set = false; | |
333 | ||
334 | if (!zone->compact_blockskip_flush) | |
335 | return; | |
bb13ffeb | 336 | |
62997027 | 337 | zone->compact_blockskip_flush = false; |
bb13ffeb | 338 | |
e332f741 MG |
339 | /* |
340 | * Walk the zone and update pageblock skip information. Source looks | |
341 | * for PageLRU while target looks for PageBuddy. When the scanner | |
342 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
343 | * is suitable as both source and target. | |
344 | */ | |
345 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
346 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
347 | cond_resched(); |
348 | ||
e332f741 MG |
349 | /* Update the migrate PFN */ |
350 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
351 | migrate_pfn < reset_migrate) { | |
352 | source_set = true; | |
353 | reset_migrate = migrate_pfn; | |
354 | zone->compact_init_migrate_pfn = reset_migrate; | |
355 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
356 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
357 | } | |
bb13ffeb | 358 | |
e332f741 MG |
359 | /* Update the free PFN */ |
360 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
361 | free_pfn > reset_free) { | |
362 | free_set = true; | |
363 | reset_free = free_pfn; | |
364 | zone->compact_init_free_pfn = reset_free; | |
365 | zone->compact_cached_free_pfn = reset_free; | |
366 | } | |
bb13ffeb | 367 | } |
02333641 | 368 | |
e332f741 MG |
369 | /* Leave no distance if no suitable block was reset */ |
370 | if (reset_migrate >= reset_free) { | |
371 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
372 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
373 | zone->compact_cached_free_pfn = free_pfn; | |
374 | } | |
bb13ffeb MG |
375 | } |
376 | ||
62997027 MG |
377 | void reset_isolation_suitable(pg_data_t *pgdat) |
378 | { | |
379 | int zoneid; | |
380 | ||
381 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
382 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
383 | if (!populated_zone(zone)) | |
384 | continue; | |
385 | ||
386 | /* Only flush if a full compaction finished recently */ | |
387 | if (zone->compact_blockskip_flush) | |
388 | __reset_isolation_suitable(zone); | |
389 | } | |
390 | } | |
391 | ||
e380bebe MG |
392 | /* |
393 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
394 | * locks are not required for read/writers. Returns true if it was already set. | |
395 | */ | |
396 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
397 | unsigned long pfn) | |
398 | { | |
399 | bool skip; | |
400 | ||
401 | /* Do no update if skip hint is being ignored */ | |
402 | if (cc->ignore_skip_hint) | |
403 | return false; | |
404 | ||
ee0913c4 | 405 | if (!pageblock_aligned(pfn)) |
e380bebe MG |
406 | return false; |
407 | ||
408 | skip = get_pageblock_skip(page); | |
409 | if (!skip && !cc->no_set_skip_hint) | |
410 | set_pageblock_skip(page); | |
411 | ||
412 | return skip; | |
413 | } | |
414 | ||
415 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
416 | { | |
417 | struct zone *zone = cc->zone; | |
418 | ||
419 | pfn = pageblock_end_pfn(pfn); | |
420 | ||
421 | /* Set for isolation rather than compaction */ | |
422 | if (cc->no_set_skip_hint) | |
423 | return; | |
424 | ||
425 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
426 | zone->compact_cached_migrate_pfn[0] = pfn; | |
427 | if (cc->mode != MIGRATE_ASYNC && | |
428 | pfn > zone->compact_cached_migrate_pfn[1]) | |
429 | zone->compact_cached_migrate_pfn[1] = pfn; | |
430 | } | |
431 | ||
bb13ffeb MG |
432 | /* |
433 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 434 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 435 | */ |
c89511ab | 436 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 437 | struct page *page, unsigned long pfn) |
bb13ffeb | 438 | { |
c89511ab | 439 | struct zone *zone = cc->zone; |
6815bf3f | 440 | |
2583d671 | 441 | if (cc->no_set_skip_hint) |
6815bf3f JK |
442 | return; |
443 | ||
bb13ffeb MG |
444 | if (!page) |
445 | return; | |
446 | ||
edc2ca61 | 447 | set_pageblock_skip(page); |
c89511ab | 448 | |
35979ef3 | 449 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
450 | if (pfn < zone->compact_cached_free_pfn) |
451 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
452 | } |
453 | #else | |
454 | static inline bool isolation_suitable(struct compact_control *cc, | |
455 | struct page *page) | |
456 | { | |
457 | return true; | |
458 | } | |
459 | ||
b527cfe5 | 460 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
461 | { |
462 | return false; | |
463 | } | |
464 | ||
465 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 466 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
467 | { |
468 | } | |
e380bebe MG |
469 | |
470 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
471 | { | |
472 | } | |
473 | ||
474 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
475 | unsigned long pfn) | |
476 | { | |
477 | return false; | |
478 | } | |
bb13ffeb MG |
479 | #endif /* CONFIG_COMPACTION */ |
480 | ||
8b44d279 VB |
481 | /* |
482 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
483 | * very heavily contended. For async compaction, trylock and record if the |
484 | * lock is contended. The lock will still be acquired but compaction will | |
485 | * abort when the current block is finished regardless of success rate. | |
486 | * Sync compaction acquires the lock. | |
8b44d279 | 487 | * |
cb2dcaf0 | 488 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 489 | */ |
cb2dcaf0 | 490 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 491 | struct compact_control *cc) |
77337ede | 492 | __acquires(lock) |
2a1402aa | 493 | { |
cb2dcaf0 MG |
494 | /* Track if the lock is contended in async mode */ |
495 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
496 | if (spin_trylock_irqsave(lock, *flags)) | |
497 | return true; | |
498 | ||
499 | cc->contended = true; | |
8b44d279 | 500 | } |
1f9efdef | 501 | |
cb2dcaf0 | 502 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 503 | return true; |
2a1402aa MG |
504 | } |
505 | ||
c67fe375 MG |
506 | /* |
507 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
508 | * very heavily contended. The lock should be periodically unlocked to avoid |
509 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
510 | * the lock. It might also be that allowing the IRQs will result in | |
d56c1584 | 511 | * need_resched() becoming true. If scheduling is needed, compaction schedules. |
8b44d279 VB |
512 | * Either compaction type will also abort if a fatal signal is pending. |
513 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 514 | * |
d56c1584 ML |
515 | * Returns true if compaction should abort due to fatal signal pending. |
516 | * Returns false when compaction can continue. | |
c67fe375 | 517 | */ |
8b44d279 VB |
518 | static bool compact_unlock_should_abort(spinlock_t *lock, |
519 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 520 | { |
8b44d279 VB |
521 | if (*locked) { |
522 | spin_unlock_irqrestore(lock, flags); | |
523 | *locked = false; | |
524 | } | |
1f9efdef | 525 | |
8b44d279 | 526 | if (fatal_signal_pending(current)) { |
c3486f53 | 527 | cc->contended = true; |
8b44d279 VB |
528 | return true; |
529 | } | |
c67fe375 | 530 | |
cf66f070 | 531 | cond_resched(); |
be976572 VB |
532 | |
533 | return false; | |
534 | } | |
535 | ||
85aa125f | 536 | /* |
9e4be470 JM |
537 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
538 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
539 | * (even though it may still end up isolating some pages). | |
85aa125f | 540 | */ |
f40d1e42 | 541 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 542 | unsigned long *start_pfn, |
85aa125f MN |
543 | unsigned long end_pfn, |
544 | struct list_head *freelist, | |
4fca9730 | 545 | unsigned int stride, |
85aa125f | 546 | bool strict) |
748446bb | 547 | { |
b7aba698 | 548 | int nr_scanned = 0, total_isolated = 0; |
d097a6f6 | 549 | struct page *cursor; |
b8b2d825 | 550 | unsigned long flags = 0; |
f40d1e42 | 551 | bool locked = false; |
e14c720e | 552 | unsigned long blockpfn = *start_pfn; |
66c64223 | 553 | unsigned int order; |
748446bb | 554 | |
4fca9730 MG |
555 | /* Strict mode is for isolation, speed is secondary */ |
556 | if (strict) | |
557 | stride = 1; | |
558 | ||
748446bb MG |
559 | cursor = pfn_to_page(blockpfn); |
560 | ||
f40d1e42 | 561 | /* Isolate free pages. */ |
4fca9730 | 562 | for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { |
66c64223 | 563 | int isolated; |
748446bb MG |
564 | struct page *page = cursor; |
565 | ||
8b44d279 VB |
566 | /* |
567 | * Periodically drop the lock (if held) regardless of its | |
568 | * contention, to give chance to IRQs. Abort if fatal signal | |
d56c1584 | 569 | * pending. |
8b44d279 | 570 | */ |
c036ddff | 571 | if (!(blockpfn % COMPACT_CLUSTER_MAX) |
8b44d279 VB |
572 | && compact_unlock_should_abort(&cc->zone->lock, flags, |
573 | &locked, cc)) | |
574 | break; | |
575 | ||
b7aba698 | 576 | nr_scanned++; |
2af120bc | 577 | |
9fcd6d2e VB |
578 | /* |
579 | * For compound pages such as THP and hugetlbfs, we can save | |
580 | * potentially a lot of iterations if we skip them at once. | |
581 | * The check is racy, but we can consider only valid values | |
582 | * and the only danger is skipping too much. | |
583 | */ | |
584 | if (PageCompound(page)) { | |
21dc7e02 DR |
585 | const unsigned int order = compound_order(page); |
586 | ||
d3c85bad | 587 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
588 | blockpfn += (1UL << order) - 1; |
589 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 590 | } |
9fcd6d2e VB |
591 | goto isolate_fail; |
592 | } | |
593 | ||
f40d1e42 | 594 | if (!PageBuddy(page)) |
2af120bc | 595 | goto isolate_fail; |
f40d1e42 | 596 | |
85f73e6d | 597 | /* If we already hold the lock, we can skip some rechecking. */ |
69b7189f | 598 | if (!locked) { |
cb2dcaf0 | 599 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 600 | &flags, cc); |
f40d1e42 | 601 | |
69b7189f VB |
602 | /* Recheck this is a buddy page under lock */ |
603 | if (!PageBuddy(page)) | |
604 | goto isolate_fail; | |
605 | } | |
748446bb | 606 | |
66c64223 | 607 | /* Found a free page, will break it into order-0 pages */ |
ab130f91 | 608 | order = buddy_order(page); |
66c64223 | 609 | isolated = __isolate_free_page(page, order); |
a4f04f2c DR |
610 | if (!isolated) |
611 | break; | |
66c64223 | 612 | set_page_private(page, order); |
a4f04f2c | 613 | |
b717d6b9 | 614 | nr_scanned += isolated - 1; |
748446bb | 615 | total_isolated += isolated; |
a4f04f2c | 616 | cc->nr_freepages += isolated; |
66c64223 JK |
617 | list_add_tail(&page->lru, freelist); |
618 | ||
a4f04f2c DR |
619 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
620 | blockpfn += isolated; | |
621 | break; | |
748446bb | 622 | } |
a4f04f2c DR |
623 | /* Advance to the end of split page */ |
624 | blockpfn += isolated - 1; | |
625 | cursor += isolated - 1; | |
626 | continue; | |
2af120bc LA |
627 | |
628 | isolate_fail: | |
629 | if (strict) | |
630 | break; | |
631 | else | |
632 | continue; | |
633 | ||
748446bb MG |
634 | } |
635 | ||
a4f04f2c DR |
636 | if (locked) |
637 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
638 | ||
9fcd6d2e VB |
639 | /* |
640 | * There is a tiny chance that we have read bogus compound_order(), | |
641 | * so be careful to not go outside of the pageblock. | |
642 | */ | |
643 | if (unlikely(blockpfn > end_pfn)) | |
644 | blockpfn = end_pfn; | |
645 | ||
e34d85f0 JK |
646 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
647 | nr_scanned, total_isolated); | |
648 | ||
e14c720e VB |
649 | /* Record how far we have got within the block */ |
650 | *start_pfn = blockpfn; | |
651 | ||
f40d1e42 MG |
652 | /* |
653 | * If strict isolation is requested by CMA then check that all the | |
654 | * pages requested were isolated. If there were any failures, 0 is | |
655 | * returned and CMA will fail. | |
656 | */ | |
2af120bc | 657 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
658 | total_isolated = 0; |
659 | ||
7f354a54 | 660 | cc->total_free_scanned += nr_scanned; |
397487db | 661 | if (total_isolated) |
010fc29a | 662 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
663 | return total_isolated; |
664 | } | |
665 | ||
85aa125f MN |
666 | /** |
667 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 668 | * @cc: Compaction control structure. |
85aa125f MN |
669 | * @start_pfn: The first PFN to start isolating. |
670 | * @end_pfn: The one-past-last PFN. | |
671 | * | |
672 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
673 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
674 | * undo its actions and return zero. | |
675 | * | |
676 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
677 | * (which may be greater then end_pfn if end fell in a middle of | |
678 | * a free page). | |
679 | */ | |
ff9543fd | 680 | unsigned long |
bb13ffeb MG |
681 | isolate_freepages_range(struct compact_control *cc, |
682 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 683 | { |
e1409c32 | 684 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
685 | LIST_HEAD(freelist); |
686 | ||
7d49d886 | 687 | pfn = start_pfn; |
06b6640a | 688 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
689 | if (block_start_pfn < cc->zone->zone_start_pfn) |
690 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 691 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
692 | |
693 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 694 | block_start_pfn = block_end_pfn, |
7d49d886 | 695 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
696 | /* Protect pfn from changing by isolate_freepages_block */ |
697 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 698 | |
85aa125f MN |
699 | block_end_pfn = min(block_end_pfn, end_pfn); |
700 | ||
58420016 JK |
701 | /* |
702 | * pfn could pass the block_end_pfn if isolated freepage | |
703 | * is more than pageblock order. In this case, we adjust | |
704 | * scanning range to right one. | |
705 | */ | |
706 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
707 | block_start_pfn = pageblock_start_pfn(pfn); |
708 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
709 | block_end_pfn = min(block_end_pfn, end_pfn); |
710 | } | |
711 | ||
e1409c32 JK |
712 | if (!pageblock_pfn_to_page(block_start_pfn, |
713 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
714 | break; |
715 | ||
e14c720e | 716 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
4fca9730 | 717 | block_end_pfn, &freelist, 0, true); |
85aa125f MN |
718 | |
719 | /* | |
720 | * In strict mode, isolate_freepages_block() returns 0 if | |
721 | * there are any holes in the block (ie. invalid PFNs or | |
722 | * non-free pages). | |
723 | */ | |
724 | if (!isolated) | |
725 | break; | |
726 | ||
727 | /* | |
728 | * If we managed to isolate pages, it is always (1 << n) * | |
729 | * pageblock_nr_pages for some non-negative n. (Max order | |
730 | * page may span two pageblocks). | |
731 | */ | |
732 | } | |
733 | ||
66c64223 | 734 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 735 | split_map_pages(&freelist); |
85aa125f MN |
736 | |
737 | if (pfn < end_pfn) { | |
738 | /* Loop terminated early, cleanup. */ | |
739 | release_freepages(&freelist); | |
740 | return 0; | |
741 | } | |
742 | ||
743 | /* We don't use freelists for anything. */ | |
744 | return pfn; | |
745 | } | |
746 | ||
748446bb | 747 | /* Similar to reclaim, but different enough that they don't share logic */ |
5f438eee | 748 | static bool too_many_isolated(pg_data_t *pgdat) |
748446bb | 749 | { |
d818fca1 MG |
750 | bool too_many; |
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 | |
d818fca1 MG |
761 | too_many = isolated > (inactive + active) / 2; |
762 | if (!too_many) | |
763 | wake_throttle_isolated(pgdat); | |
764 | ||
765 | return too_many; | |
748446bb MG |
766 | } |
767 | ||
2fe86e00 | 768 | /** |
edc2ca61 VB |
769 | * isolate_migratepages_block() - isolate all migrate-able pages within |
770 | * a single pageblock | |
2fe86e00 | 771 | * @cc: Compaction control structure. |
edc2ca61 VB |
772 | * @low_pfn: The first PFN to isolate |
773 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
89f6c88a | 774 | * @mode: Isolation mode to be used. |
2fe86e00 MN |
775 | * |
776 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 | 777 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
c2ad7a1f | 778 | * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, |
369fa227 | 779 | * -ENOMEM in case we could not allocate a page, or 0. |
c2ad7a1f | 780 | * cc->migrate_pfn will contain the next pfn to scan. |
2fe86e00 | 781 | * |
edc2ca61 | 782 | * The pages are isolated on cc->migratepages list (not required to be empty), |
c2ad7a1f | 783 | * and cc->nr_migratepages is updated accordingly. |
748446bb | 784 | */ |
c2ad7a1f | 785 | static int |
edc2ca61 | 786 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, |
89f6c88a | 787 | unsigned long end_pfn, isolate_mode_t mode) |
748446bb | 788 | { |
5f438eee | 789 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 790 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 791 | struct lruvec *lruvec; |
b8b2d825 | 792 | unsigned long flags = 0; |
6168d0da | 793 | struct lruvec *locked = NULL; |
bb13ffeb | 794 | struct page *page = NULL, *valid_page = NULL; |
89f6c88a | 795 | struct address_space *mapping; |
e34d85f0 | 796 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
797 | bool skip_on_failure = false; |
798 | unsigned long next_skip_pfn = 0; | |
e380bebe | 799 | bool skip_updated = false; |
c2ad7a1f OS |
800 | int ret = 0; |
801 | ||
802 | cc->migrate_pfn = low_pfn; | |
748446bb | 803 | |
748446bb MG |
804 | /* |
805 | * Ensure that there are not too many pages isolated from the LRU | |
806 | * list by either parallel reclaimers or compaction. If there are, | |
807 | * delay for some time until fewer pages are isolated | |
808 | */ | |
5f438eee | 809 | while (unlikely(too_many_isolated(pgdat))) { |
d20bdd57 ZY |
810 | /* stop isolation if there are still pages not migrated */ |
811 | if (cc->nr_migratepages) | |
c2ad7a1f | 812 | return -EAGAIN; |
d20bdd57 | 813 | |
f9e35b3b | 814 | /* async migration should just abort */ |
e0b9daeb | 815 | if (cc->mode == MIGRATE_ASYNC) |
c2ad7a1f | 816 | return -EAGAIN; |
f9e35b3b | 817 | |
c3f4a9a2 | 818 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
748446bb MG |
819 | |
820 | if (fatal_signal_pending(current)) | |
c2ad7a1f | 821 | return -EINTR; |
748446bb MG |
822 | } |
823 | ||
cf66f070 | 824 | cond_resched(); |
aeef4b83 | 825 | |
fdd048e1 VB |
826 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
827 | skip_on_failure = true; | |
828 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
829 | } | |
830 | ||
748446bb | 831 | /* Time to isolate some pages for migration */ |
748446bb | 832 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 833 | |
fdd048e1 VB |
834 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
835 | /* | |
836 | * We have isolated all migration candidates in the | |
837 | * previous order-aligned block, and did not skip it due | |
838 | * to failure. We should migrate the pages now and | |
839 | * hopefully succeed compaction. | |
840 | */ | |
841 | if (nr_isolated) | |
842 | break; | |
843 | ||
844 | /* | |
845 | * We failed to isolate in the previous order-aligned | |
846 | * block. Set the new boundary to the end of the | |
847 | * current block. Note we can't simply increase | |
848 | * next_skip_pfn by 1 << order, as low_pfn might have | |
849 | * been incremented by a higher number due to skipping | |
850 | * a compound or a high-order buddy page in the | |
851 | * previous loop iteration. | |
852 | */ | |
853 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
854 | } | |
855 | ||
8b44d279 VB |
856 | /* |
857 | * Periodically drop the lock (if held) regardless of its | |
670105a2 MG |
858 | * contention, to give chance to IRQs. Abort completely if |
859 | * a fatal signal is pending. | |
8b44d279 | 860 | */ |
c036ddff | 861 | if (!(low_pfn % COMPACT_CLUSTER_MAX)) { |
6168d0da AS |
862 | if (locked) { |
863 | unlock_page_lruvec_irqrestore(locked, flags); | |
864 | locked = NULL; | |
865 | } | |
866 | ||
867 | if (fatal_signal_pending(current)) { | |
868 | cc->contended = true; | |
c2ad7a1f | 869 | ret = -EINTR; |
6168d0da | 870 | |
6168d0da AS |
871 | goto fatal_pending; |
872 | } | |
873 | ||
874 | cond_resched(); | |
670105a2 | 875 | } |
c67fe375 | 876 | |
b7aba698 | 877 | nr_scanned++; |
748446bb | 878 | |
748446bb | 879 | page = pfn_to_page(low_pfn); |
dc908600 | 880 | |
e380bebe MG |
881 | /* |
882 | * Check if the pageblock has already been marked skipped. | |
883 | * Only the aligned PFN is checked as the caller isolates | |
884 | * COMPACT_CLUSTER_MAX at a time so the second call must | |
885 | * not falsely conclude that the block should be skipped. | |
886 | */ | |
ee0913c4 | 887 | if (!valid_page && pageblock_aligned(low_pfn)) { |
4af12d04 | 888 | if (!isolation_suitable(cc, page)) { |
e380bebe | 889 | low_pfn = end_pfn; |
9df41314 | 890 | page = NULL; |
e380bebe MG |
891 | goto isolate_abort; |
892 | } | |
bb13ffeb | 893 | valid_page = page; |
e380bebe | 894 | } |
bb13ffeb | 895 | |
369fa227 | 896 | if (PageHuge(page) && cc->alloc_contig) { |
ae37c7ff | 897 | ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); |
369fa227 OS |
898 | |
899 | /* | |
900 | * Fail isolation in case isolate_or_dissolve_huge_page() | |
901 | * reports an error. In case of -ENOMEM, abort right away. | |
902 | */ | |
903 | if (ret < 0) { | |
904 | /* Do not report -EBUSY down the chain */ | |
905 | if (ret == -EBUSY) | |
906 | ret = 0; | |
66fe1cf7 | 907 | low_pfn += compound_nr(page) - 1; |
369fa227 OS |
908 | goto isolate_fail; |
909 | } | |
910 | ||
ae37c7ff OS |
911 | if (PageHuge(page)) { |
912 | /* | |
913 | * Hugepage was successfully isolated and placed | |
914 | * on the cc->migratepages list. | |
915 | */ | |
916 | low_pfn += compound_nr(page) - 1; | |
917 | goto isolate_success_no_list; | |
918 | } | |
919 | ||
369fa227 OS |
920 | /* |
921 | * Ok, the hugepage was dissolved. Now these pages are | |
922 | * Buddy and cannot be re-allocated because they are | |
923 | * isolated. Fall-through as the check below handles | |
924 | * Buddy pages. | |
925 | */ | |
926 | } | |
927 | ||
6c14466c | 928 | /* |
99c0fd5e VB |
929 | * Skip if free. We read page order here without zone lock |
930 | * which is generally unsafe, but the race window is small and | |
931 | * the worst thing that can happen is that we skip some | |
932 | * potential isolation targets. | |
6c14466c | 933 | */ |
99c0fd5e | 934 | if (PageBuddy(page)) { |
ab130f91 | 935 | unsigned long freepage_order = buddy_order_unsafe(page); |
99c0fd5e VB |
936 | |
937 | /* | |
938 | * Without lock, we cannot be sure that what we got is | |
939 | * a valid page order. Consider only values in the | |
940 | * valid order range to prevent low_pfn overflow. | |
941 | */ | |
942 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
943 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 944 | continue; |
99c0fd5e | 945 | } |
748446bb | 946 | |
bc835011 | 947 | /* |
29c0dde8 | 948 | * Regardless of being on LRU, compound pages such as THP and |
1da2f328 RR |
949 | * hugetlbfs are not to be compacted unless we are attempting |
950 | * an allocation much larger than the huge page size (eg CMA). | |
951 | * We can potentially save a lot of iterations if we skip them | |
952 | * at once. The check is racy, but we can consider only valid | |
953 | * values and the only danger is skipping too much. | |
bc835011 | 954 | */ |
1da2f328 | 955 | if (PageCompound(page) && !cc->alloc_contig) { |
21dc7e02 | 956 | const unsigned int order = compound_order(page); |
edc2ca61 | 957 | |
d3c85bad | 958 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 959 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 960 | goto isolate_fail; |
2a1402aa MG |
961 | } |
962 | ||
bda807d4 MK |
963 | /* |
964 | * Check may be lockless but that's ok as we recheck later. | |
965 | * It's possible to migrate LRU and non-lru movable pages. | |
966 | * Skip any other type of page | |
967 | */ | |
968 | if (!PageLRU(page)) { | |
bda807d4 MK |
969 | /* |
970 | * __PageMovable can return false positive so we need | |
971 | * to verify it under page_lock. | |
972 | */ | |
973 | if (unlikely(__PageMovable(page)) && | |
974 | !PageIsolated(page)) { | |
975 | if (locked) { | |
6168d0da AS |
976 | unlock_page_lruvec_irqrestore(locked, flags); |
977 | locked = NULL; | |
bda807d4 MK |
978 | } |
979 | ||
89f6c88a | 980 | if (!isolate_movable_page(page, mode)) |
bda807d4 MK |
981 | goto isolate_success; |
982 | } | |
983 | ||
fdd048e1 | 984 | goto isolate_fail; |
bda807d4 | 985 | } |
29c0dde8 | 986 | |
119d6d59 DR |
987 | /* |
988 | * Migration will fail if an anonymous page is pinned in memory, | |
989 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
990 | * admittedly racy check. | |
991 | */ | |
89f6c88a HD |
992 | mapping = page_mapping(page); |
993 | if (!mapping && page_count(page) > page_mapcount(page)) | |
fdd048e1 | 994 | goto isolate_fail; |
119d6d59 | 995 | |
73e64c51 MH |
996 | /* |
997 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
998 | * because those do not depend on fs locks. | |
999 | */ | |
89f6c88a | 1000 | if (!(cc->gfp_mask & __GFP_FS) && mapping) |
73e64c51 MH |
1001 | goto isolate_fail; |
1002 | ||
9df41314 AS |
1003 | /* |
1004 | * Be careful not to clear PageLRU until after we're | |
1005 | * sure the page is not being freed elsewhere -- the | |
1006 | * page release code relies on it. | |
1007 | */ | |
1008 | if (unlikely(!get_page_unless_zero(page))) | |
1009 | goto isolate_fail; | |
1010 | ||
89f6c88a HD |
1011 | /* Only take pages on LRU: a check now makes later tests safe */ |
1012 | if (!PageLRU(page)) | |
1013 | goto isolate_fail_put; | |
1014 | ||
1015 | /* Compaction might skip unevictable pages but CMA takes them */ | |
1016 | if (!(mode & ISOLATE_UNEVICTABLE) && PageUnevictable(page)) | |
1017 | goto isolate_fail_put; | |
1018 | ||
1019 | /* | |
1020 | * To minimise LRU disruption, the caller can indicate with | |
1021 | * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages | |
1022 | * it will be able to migrate without blocking - clean pages | |
1023 | * for the most part. PageWriteback would require blocking. | |
1024 | */ | |
1025 | if ((mode & ISOLATE_ASYNC_MIGRATE) && PageWriteback(page)) | |
9df41314 AS |
1026 | goto isolate_fail_put; |
1027 | ||
89f6c88a HD |
1028 | if ((mode & ISOLATE_ASYNC_MIGRATE) && PageDirty(page)) { |
1029 | bool migrate_dirty; | |
1030 | ||
1031 | /* | |
1032 | * Only pages without mappings or that have a | |
9d0ddc0c | 1033 | * ->migrate_folio callback are possible to migrate |
89f6c88a HD |
1034 | * without blocking. However, we can be racing with |
1035 | * truncation so it's necessary to lock the page | |
1036 | * to stabilise the mapping as truncation holds | |
1037 | * the page lock until after the page is removed | |
1038 | * from the page cache. | |
1039 | */ | |
1040 | if (!trylock_page(page)) | |
1041 | goto isolate_fail_put; | |
1042 | ||
1043 | mapping = page_mapping(page); | |
5490da4f | 1044 | migrate_dirty = !mapping || |
9d0ddc0c | 1045 | mapping->a_ops->migrate_folio; |
89f6c88a HD |
1046 | unlock_page(page); |
1047 | if (!migrate_dirty) | |
1048 | goto isolate_fail_put; | |
1049 | } | |
1050 | ||
9df41314 AS |
1051 | /* Try isolate the page */ |
1052 | if (!TestClearPageLRU(page)) | |
1053 | goto isolate_fail_put; | |
1054 | ||
b1baabd9 | 1055 | lruvec = folio_lruvec(page_folio(page)); |
6168d0da | 1056 | |
69b7189f | 1057 | /* If we already hold the lock, we can skip some rechecking */ |
6168d0da AS |
1058 | if (lruvec != locked) { |
1059 | if (locked) | |
1060 | unlock_page_lruvec_irqrestore(locked, flags); | |
1061 | ||
1062 | compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); | |
1063 | locked = lruvec; | |
6168d0da | 1064 | |
e809c3fe | 1065 | lruvec_memcg_debug(lruvec, page_folio(page)); |
e380bebe | 1066 | |
e380bebe MG |
1067 | /* Try get exclusive access under lock */ |
1068 | if (!skip_updated) { | |
1069 | skip_updated = true; | |
1070 | if (test_and_set_skip(cc, page, low_pfn)) | |
1071 | goto isolate_abort; | |
1072 | } | |
2a1402aa | 1073 | |
29c0dde8 VB |
1074 | /* |
1075 | * Page become compound since the non-locked check, | |
1076 | * and it's on LRU. It can only be a THP so the order | |
1077 | * is safe to read and it's 0 for tail pages. | |
1078 | */ | |
1da2f328 | 1079 | if (unlikely(PageCompound(page) && !cc->alloc_contig)) { |
d8c6546b | 1080 | low_pfn += compound_nr(page) - 1; |
9df41314 AS |
1081 | SetPageLRU(page); |
1082 | goto isolate_fail_put; | |
69b7189f | 1083 | } |
d99fd5fe | 1084 | } |
fa9add64 | 1085 | |
1da2f328 RR |
1086 | /* The whole page is taken off the LRU; skip the tail pages. */ |
1087 | if (PageCompound(page)) | |
1088 | low_pfn += compound_nr(page) - 1; | |
bc835011 | 1089 | |
748446bb | 1090 | /* Successfully isolated */ |
46ae6b2c | 1091 | del_page_from_lru_list(page, lruvec); |
1da2f328 | 1092 | mod_node_page_state(page_pgdat(page), |
9de4f22a | 1093 | NR_ISOLATED_ANON + page_is_file_lru(page), |
6c357848 | 1094 | thp_nr_pages(page)); |
b6c75016 JK |
1095 | |
1096 | isolate_success: | |
fdd048e1 | 1097 | list_add(&page->lru, &cc->migratepages); |
ae37c7ff | 1098 | isolate_success_no_list: |
38935861 ZY |
1099 | cc->nr_migratepages += compound_nr(page); |
1100 | nr_isolated += compound_nr(page); | |
b717d6b9 | 1101 | nr_scanned += compound_nr(page) - 1; |
748446bb | 1102 | |
804d3121 MG |
1103 | /* |
1104 | * Avoid isolating too much unless this block is being | |
cb2dcaf0 MG |
1105 | * rescanned (e.g. dirty/writeback pages, parallel allocation) |
1106 | * or a lock is contended. For contention, isolate quickly to | |
1107 | * potentially remove one source of contention. | |
804d3121 | 1108 | */ |
38935861 | 1109 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && |
cb2dcaf0 | 1110 | !cc->rescan && !cc->contended) { |
31b8384a | 1111 | ++low_pfn; |
748446bb | 1112 | break; |
31b8384a | 1113 | } |
fdd048e1 VB |
1114 | |
1115 | continue; | |
9df41314 AS |
1116 | |
1117 | isolate_fail_put: | |
1118 | /* Avoid potential deadlock in freeing page under lru_lock */ | |
1119 | if (locked) { | |
6168d0da AS |
1120 | unlock_page_lruvec_irqrestore(locked, flags); |
1121 | locked = NULL; | |
9df41314 AS |
1122 | } |
1123 | put_page(page); | |
1124 | ||
fdd048e1 | 1125 | isolate_fail: |
369fa227 | 1126 | if (!skip_on_failure && ret != -ENOMEM) |
fdd048e1 VB |
1127 | continue; |
1128 | ||
1129 | /* | |
1130 | * We have isolated some pages, but then failed. Release them | |
1131 | * instead of migrating, as we cannot form the cc->order buddy | |
1132 | * page anyway. | |
1133 | */ | |
1134 | if (nr_isolated) { | |
1135 | if (locked) { | |
6168d0da AS |
1136 | unlock_page_lruvec_irqrestore(locked, flags); |
1137 | locked = NULL; | |
fdd048e1 | 1138 | } |
fdd048e1 VB |
1139 | putback_movable_pages(&cc->migratepages); |
1140 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1141 | nr_isolated = 0; |
1142 | } | |
1143 | ||
1144 | if (low_pfn < next_skip_pfn) { | |
1145 | low_pfn = next_skip_pfn - 1; | |
1146 | /* | |
1147 | * The check near the loop beginning would have updated | |
1148 | * next_skip_pfn too, but this is a bit simpler. | |
1149 | */ | |
1150 | next_skip_pfn += 1UL << cc->order; | |
1151 | } | |
369fa227 OS |
1152 | |
1153 | if (ret == -ENOMEM) | |
1154 | break; | |
748446bb MG |
1155 | } |
1156 | ||
99c0fd5e VB |
1157 | /* |
1158 | * The PageBuddy() check could have potentially brought us outside | |
1159 | * the range to be scanned. | |
1160 | */ | |
1161 | if (unlikely(low_pfn > end_pfn)) | |
1162 | low_pfn = end_pfn; | |
1163 | ||
9df41314 AS |
1164 | page = NULL; |
1165 | ||
e380bebe | 1166 | isolate_abort: |
c67fe375 | 1167 | if (locked) |
6168d0da | 1168 | unlock_page_lruvec_irqrestore(locked, flags); |
9df41314 AS |
1169 | if (page) { |
1170 | SetPageLRU(page); | |
1171 | put_page(page); | |
1172 | } | |
748446bb | 1173 | |
50b5b094 | 1174 | /* |
804d3121 MG |
1175 | * Updated the cached scanner pfn once the pageblock has been scanned |
1176 | * Pages will either be migrated in which case there is no point | |
1177 | * scanning in the near future or migration failed in which case the | |
1178 | * failure reason may persist. The block is marked for skipping if | |
1179 | * there were no pages isolated in the block or if the block is | |
1180 | * rescanned twice in a row. | |
50b5b094 | 1181 | */ |
804d3121 | 1182 | if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) { |
e380bebe MG |
1183 | if (valid_page && !skip_updated) |
1184 | set_pageblock_skip(valid_page); | |
1185 | update_cached_migrate(cc, low_pfn); | |
1186 | } | |
bb13ffeb | 1187 | |
e34d85f0 JK |
1188 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1189 | nr_scanned, nr_isolated); | |
b7aba698 | 1190 | |
670105a2 | 1191 | fatal_pending: |
7f354a54 | 1192 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1193 | if (nr_isolated) |
010fc29a | 1194 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1195 | |
c2ad7a1f OS |
1196 | cc->migrate_pfn = low_pfn; |
1197 | ||
1198 | return ret; | |
2fe86e00 MN |
1199 | } |
1200 | ||
edc2ca61 VB |
1201 | /** |
1202 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1203 | * @cc: Compaction control structure. | |
1204 | * @start_pfn: The first PFN to start isolating. | |
1205 | * @end_pfn: The one-past-last PFN. | |
1206 | * | |
369fa227 OS |
1207 | * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM |
1208 | * in case we could not allocate a page, or 0. | |
edc2ca61 | 1209 | */ |
c2ad7a1f | 1210 | int |
edc2ca61 VB |
1211 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, |
1212 | unsigned long end_pfn) | |
1213 | { | |
e1409c32 | 1214 | unsigned long pfn, block_start_pfn, block_end_pfn; |
c2ad7a1f | 1215 | int ret = 0; |
edc2ca61 VB |
1216 | |
1217 | /* Scan block by block. First and last block may be incomplete */ | |
1218 | pfn = start_pfn; | |
06b6640a | 1219 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1220 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1221 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1222 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1223 | |
1224 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1225 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1226 | block_end_pfn += pageblock_nr_pages) { |
1227 | ||
1228 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1229 | ||
e1409c32 JK |
1230 | if (!pageblock_pfn_to_page(block_start_pfn, |
1231 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1232 | continue; |
1233 | ||
c2ad7a1f OS |
1234 | ret = isolate_migratepages_block(cc, pfn, block_end_pfn, |
1235 | ISOLATE_UNEVICTABLE); | |
edc2ca61 | 1236 | |
c2ad7a1f | 1237 | if (ret) |
edc2ca61 | 1238 | break; |
6ea41c0c | 1239 | |
38935861 | 1240 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) |
6ea41c0c | 1241 | break; |
edc2ca61 | 1242 | } |
edc2ca61 | 1243 | |
c2ad7a1f | 1244 | return ret; |
edc2ca61 VB |
1245 | } |
1246 | ||
ff9543fd MN |
1247 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1248 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1249 | |
b682debd VB |
1250 | static bool suitable_migration_source(struct compact_control *cc, |
1251 | struct page *page) | |
1252 | { | |
282722b0 VB |
1253 | int block_mt; |
1254 | ||
9bebefd5 MG |
1255 | if (pageblock_skip_persistent(page)) |
1256 | return false; | |
1257 | ||
282722b0 | 1258 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1259 | return true; |
1260 | ||
282722b0 VB |
1261 | block_mt = get_pageblock_migratetype(page); |
1262 | ||
1263 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1264 | return is_migrate_movable(block_mt); | |
1265 | else | |
1266 | return block_mt == cc->migratetype; | |
b682debd VB |
1267 | } |
1268 | ||
018e9a49 | 1269 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1270 | static bool suitable_migration_target(struct compact_control *cc, |
1271 | struct page *page) | |
018e9a49 AM |
1272 | { |
1273 | /* If the page is a large free page, then disallow migration */ | |
1274 | if (PageBuddy(page)) { | |
1275 | /* | |
1276 | * We are checking page_order without zone->lock taken. But | |
1277 | * the only small danger is that we skip a potentially suitable | |
1278 | * pageblock, so it's not worth to check order for valid range. | |
1279 | */ | |
ab130f91 | 1280 | if (buddy_order_unsafe(page) >= pageblock_order) |
018e9a49 AM |
1281 | return false; |
1282 | } | |
1283 | ||
1ef36db2 YX |
1284 | if (cc->ignore_block_suitable) |
1285 | return true; | |
1286 | ||
018e9a49 | 1287 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1288 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1289 | return true; |
1290 | ||
1291 | /* Otherwise skip the block */ | |
1292 | return false; | |
1293 | } | |
1294 | ||
70b44595 MG |
1295 | static inline unsigned int |
1296 | freelist_scan_limit(struct compact_control *cc) | |
1297 | { | |
dd7ef7bd QC |
1298 | unsigned short shift = BITS_PER_LONG - 1; |
1299 | ||
1300 | return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; | |
70b44595 MG |
1301 | } |
1302 | ||
f2849aa0 VB |
1303 | /* |
1304 | * Test whether the free scanner has reached the same or lower pageblock than | |
1305 | * the migration scanner, and compaction should thus terminate. | |
1306 | */ | |
1307 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1308 | { | |
1309 | return (cc->free_pfn >> pageblock_order) | |
1310 | <= (cc->migrate_pfn >> pageblock_order); | |
1311 | } | |
1312 | ||
5a811889 MG |
1313 | /* |
1314 | * Used when scanning for a suitable migration target which scans freelists | |
1315 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1316 | * first on the next iteration of the free scanner | |
1317 | */ | |
1318 | static void | |
1319 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1320 | { | |
1321 | LIST_HEAD(sublist); | |
1322 | ||
1323 | if (!list_is_last(freelist, &freepage->lru)) { | |
1324 | list_cut_before(&sublist, freelist, &freepage->lru); | |
d2155fe5 | 1325 | list_splice_tail(&sublist, freelist); |
5a811889 MG |
1326 | } |
1327 | } | |
1328 | ||
1329 | /* | |
1330 | * Similar to move_freelist_head except used by the migration scanner | |
1331 | * when scanning forward. It's possible for these list operations to | |
1332 | * move against each other if they search the free list exactly in | |
1333 | * lockstep. | |
1334 | */ | |
70b44595 MG |
1335 | static void |
1336 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1337 | { | |
1338 | LIST_HEAD(sublist); | |
1339 | ||
1340 | if (!list_is_first(freelist, &freepage->lru)) { | |
1341 | list_cut_position(&sublist, freelist, &freepage->lru); | |
d2155fe5 | 1342 | list_splice_tail(&sublist, freelist); |
70b44595 MG |
1343 | } |
1344 | } | |
1345 | ||
5a811889 MG |
1346 | static void |
1347 | fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated) | |
1348 | { | |
1349 | unsigned long start_pfn, end_pfn; | |
6e2b7044 | 1350 | struct page *page; |
5a811889 MG |
1351 | |
1352 | /* Do not search around if there are enough pages already */ | |
1353 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1354 | return; | |
1355 | ||
1356 | /* Minimise scanning during async compaction */ | |
1357 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1358 | return; | |
1359 | ||
1360 | /* Pageblock boundaries */ | |
6e2b7044 VB |
1361 | start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); |
1362 | end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); | |
1363 | ||
1364 | page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); | |
1365 | if (!page) | |
1366 | return; | |
5a811889 MG |
1367 | |
1368 | /* Scan before */ | |
1369 | if (start_pfn != pfn) { | |
4fca9730 | 1370 | isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false); |
5a811889 MG |
1371 | if (cc->nr_freepages >= cc->nr_migratepages) |
1372 | return; | |
1373 | } | |
1374 | ||
1375 | /* Scan after */ | |
1376 | start_pfn = pfn + nr_isolated; | |
60fce36a | 1377 | if (start_pfn < end_pfn) |
4fca9730 | 1378 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); |
5a811889 MG |
1379 | |
1380 | /* Skip this pageblock in the future as it's full or nearly full */ | |
1381 | if (cc->nr_freepages < cc->nr_migratepages) | |
1382 | set_pageblock_skip(page); | |
1383 | } | |
1384 | ||
dbe2d4e4 MG |
1385 | /* Search orders in round-robin fashion */ |
1386 | static int next_search_order(struct compact_control *cc, int order) | |
1387 | { | |
1388 | order--; | |
1389 | if (order < 0) | |
1390 | order = cc->order - 1; | |
1391 | ||
1392 | /* Search wrapped around? */ | |
1393 | if (order == cc->search_order) { | |
1394 | cc->search_order--; | |
1395 | if (cc->search_order < 0) | |
1396 | cc->search_order = cc->order - 1; | |
1397 | return -1; | |
1398 | } | |
1399 | ||
1400 | return order; | |
1401 | } | |
1402 | ||
5a811889 MG |
1403 | static unsigned long |
1404 | fast_isolate_freepages(struct compact_control *cc) | |
1405 | { | |
b55ca526 | 1406 | unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); |
5a811889 | 1407 | unsigned int nr_scanned = 0; |
74e21484 | 1408 | unsigned long low_pfn, min_pfn, highest = 0; |
5a811889 MG |
1409 | unsigned long nr_isolated = 0; |
1410 | unsigned long distance; | |
1411 | struct page *page = NULL; | |
1412 | bool scan_start = false; | |
1413 | int order; | |
1414 | ||
1415 | /* Full compaction passes in a negative order */ | |
1416 | if (cc->order <= 0) | |
1417 | return cc->free_pfn; | |
1418 | ||
1419 | /* | |
1420 | * If starting the scan, use a deeper search and use the highest | |
1421 | * PFN found if a suitable one is not found. | |
1422 | */ | |
e332f741 | 1423 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1424 | limit = pageblock_nr_pages >> 1; |
1425 | scan_start = true; | |
1426 | } | |
1427 | ||
1428 | /* | |
1429 | * Preferred point is in the top quarter of the scan space but take | |
1430 | * a pfn from the top half if the search is problematic. | |
1431 | */ | |
1432 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1433 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1434 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1435 | ||
1436 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1437 | low_pfn = min_pfn; | |
1438 | ||
dbe2d4e4 MG |
1439 | /* |
1440 | * Search starts from the last successful isolation order or the next | |
1441 | * order to search after a previous failure | |
1442 | */ | |
1443 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1444 | ||
1445 | for (order = cc->search_order; | |
1446 | !page && order >= 0; | |
1447 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1448 | struct free_area *area = &cc->zone->free_area[order]; |
1449 | struct list_head *freelist; | |
1450 | struct page *freepage; | |
1451 | unsigned long flags; | |
1452 | unsigned int order_scanned = 0; | |
74e21484 | 1453 | unsigned long high_pfn = 0; |
5a811889 MG |
1454 | |
1455 | if (!area->nr_free) | |
1456 | continue; | |
1457 | ||
1458 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1459 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1460 | list_for_each_entry_reverse(freepage, freelist, lru) { | |
1461 | unsigned long pfn; | |
1462 | ||
1463 | order_scanned++; | |
1464 | nr_scanned++; | |
1465 | pfn = page_to_pfn(freepage); | |
1466 | ||
1467 | if (pfn >= highest) | |
6e2b7044 VB |
1468 | highest = max(pageblock_start_pfn(pfn), |
1469 | cc->zone->zone_start_pfn); | |
5a811889 MG |
1470 | |
1471 | if (pfn >= low_pfn) { | |
1472 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1473 | cc->search_order = order; |
5a811889 MG |
1474 | page = freepage; |
1475 | break; | |
1476 | } | |
1477 | ||
1478 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1479 | high_pfn = pfn; | |
1480 | ||
1481 | /* Shorten the scan if a candidate is found */ | |
1482 | limit >>= 1; | |
1483 | } | |
1484 | ||
1485 | if (order_scanned >= limit) | |
1486 | break; | |
1487 | } | |
1488 | ||
1489 | /* Use a minimum pfn if a preferred one was not found */ | |
1490 | if (!page && high_pfn) { | |
1491 | page = pfn_to_page(high_pfn); | |
1492 | ||
1493 | /* Update freepage for the list reorder below */ | |
1494 | freepage = page; | |
1495 | } | |
1496 | ||
1497 | /* Reorder to so a future search skips recent pages */ | |
1498 | move_freelist_head(freelist, freepage); | |
1499 | ||
1500 | /* Isolate the page if available */ | |
1501 | if (page) { | |
1502 | if (__isolate_free_page(page, order)) { | |
1503 | set_page_private(page, order); | |
1504 | nr_isolated = 1 << order; | |
b717d6b9 | 1505 | nr_scanned += nr_isolated - 1; |
5a811889 MG |
1506 | cc->nr_freepages += nr_isolated; |
1507 | list_add_tail(&page->lru, &cc->freepages); | |
1508 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1509 | } else { | |
1510 | /* If isolation fails, abort the search */ | |
5b56d996 | 1511 | order = cc->search_order + 1; |
5a811889 MG |
1512 | page = NULL; |
1513 | } | |
1514 | } | |
1515 | ||
1516 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1517 | ||
1518 | /* | |
b55ca526 | 1519 | * Smaller scan on next order so the total scan is related |
5a811889 MG |
1520 | * to freelist_scan_limit. |
1521 | */ | |
1522 | if (order_scanned >= limit) | |
b55ca526 | 1523 | limit = max(1U, limit >> 1); |
5a811889 MG |
1524 | } |
1525 | ||
1526 | if (!page) { | |
1527 | cc->fast_search_fail++; | |
1528 | if (scan_start) { | |
1529 | /* | |
1530 | * Use the highest PFN found above min. If one was | |
f3867755 | 1531 | * not found, be pessimistic for direct compaction |
5a811889 MG |
1532 | * and use the min mark. |
1533 | */ | |
ca2864e5 | 1534 | if (highest >= min_pfn) { |
5a811889 MG |
1535 | page = pfn_to_page(highest); |
1536 | cc->free_pfn = highest; | |
1537 | } else { | |
e577c8b6 | 1538 | if (cc->direct_compaction && pfn_valid(min_pfn)) { |
73a6e474 | 1539 | page = pageblock_pfn_to_page(min_pfn, |
6e2b7044 VB |
1540 | min(pageblock_end_pfn(min_pfn), |
1541 | zone_end_pfn(cc->zone)), | |
73a6e474 | 1542 | cc->zone); |
5a811889 MG |
1543 | cc->free_pfn = min_pfn; |
1544 | } | |
1545 | } | |
1546 | } | |
1547 | } | |
1548 | ||
d097a6f6 MG |
1549 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1550 | highest -= pageblock_nr_pages; | |
5a811889 | 1551 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1552 | } |
5a811889 MG |
1553 | |
1554 | cc->total_free_scanned += nr_scanned; | |
1555 | if (!page) | |
1556 | return cc->free_pfn; | |
1557 | ||
1558 | low_pfn = page_to_pfn(page); | |
1559 | fast_isolate_around(cc, low_pfn, nr_isolated); | |
1560 | return low_pfn; | |
1561 | } | |
1562 | ||
2fe86e00 | 1563 | /* |
ff9543fd MN |
1564 | * Based on information in the current compact_control, find blocks |
1565 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1566 | */ |
edc2ca61 | 1567 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1568 | { |
edc2ca61 | 1569 | struct zone *zone = cc->zone; |
ff9543fd | 1570 | struct page *page; |
c96b9e50 | 1571 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1572 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1573 | unsigned long block_end_pfn; /* end of current pageblock */ |
1574 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1575 | struct list_head *freelist = &cc->freepages; |
4fca9730 | 1576 | unsigned int stride; |
2fe86e00 | 1577 | |
5a811889 | 1578 | /* Try a small search of the free lists for a candidate */ |
00bc102f | 1579 | fast_isolate_freepages(cc); |
5a811889 MG |
1580 | if (cc->nr_freepages) |
1581 | goto splitmap; | |
1582 | ||
ff9543fd MN |
1583 | /* |
1584 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1585 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1586 | * zone when isolating for the first time. For looping we also need |
1587 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1588 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1589 | * For ending point, take care when isolating in last pageblock of a | |
a1c1dbeb | 1590 | * zone which ends in the middle of a pageblock. |
49e068f0 VB |
1591 | * The low boundary is the end of the pageblock the migration scanner |
1592 | * is using. | |
ff9543fd | 1593 | */ |
e14c720e | 1594 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1595 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1596 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1597 | zone_end_pfn(zone)); | |
06b6640a | 1598 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1599 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1600 | |
ff9543fd MN |
1601 | /* |
1602 | * Isolate free pages until enough are available to migrate the | |
1603 | * pages on cc->migratepages. We stop searching if the migrate | |
1604 | * and free page scanners meet or enough free pages are isolated. | |
1605 | */ | |
f5f61a32 | 1606 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1607 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1608 | block_start_pfn -= pageblock_nr_pages, |
1609 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1610 | unsigned long nr_isolated; |
1611 | ||
f6ea3adb DR |
1612 | /* |
1613 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1614 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1615 | */ |
c036ddff | 1616 | if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1617 | cond_resched(); |
f6ea3adb | 1618 | |
7d49d886 VB |
1619 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1620 | zone); | |
1621 | if (!page) | |
ff9543fd MN |
1622 | continue; |
1623 | ||
1624 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1625 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1626 | continue; |
68e3e926 | 1627 | |
bb13ffeb MG |
1628 | /* If isolation recently failed, do not retry */ |
1629 | if (!isolation_suitable(cc, page)) | |
1630 | continue; | |
1631 | ||
e14c720e | 1632 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 MG |
1633 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
1634 | block_end_pfn, freelist, stride, false); | |
ff9543fd | 1635 | |
d097a6f6 MG |
1636 | /* Update the skip hint if the full pageblock was scanned */ |
1637 | if (isolate_start_pfn == block_end_pfn) | |
1638 | update_pageblock_skip(cc, page, block_start_pfn); | |
1639 | ||
cb2dcaf0 MG |
1640 | /* Are enough freepages isolated? */ |
1641 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1642 | if (isolate_start_pfn >= block_end_pfn) { |
1643 | /* | |
1644 | * Restart at previous pageblock if more | |
1645 | * freepages can be isolated next time. | |
1646 | */ | |
f5f61a32 VB |
1647 | isolate_start_pfn = |
1648 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1649 | } |
be976572 | 1650 | break; |
a46cbf3b | 1651 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1652 | /* |
a46cbf3b DR |
1653 | * If isolation failed early, do not continue |
1654 | * needlessly. | |
f5f61a32 | 1655 | */ |
a46cbf3b | 1656 | break; |
f5f61a32 | 1657 | } |
4fca9730 MG |
1658 | |
1659 | /* Adjust stride depending on isolation */ | |
1660 | if (nr_isolated) { | |
1661 | stride = 1; | |
1662 | continue; | |
1663 | } | |
1664 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1665 | } |
1666 | ||
7ed695e0 | 1667 | /* |
f5f61a32 VB |
1668 | * Record where the free scanner will restart next time. Either we |
1669 | * broke from the loop and set isolate_start_pfn based on the last | |
1670 | * call to isolate_freepages_block(), or we met the migration scanner | |
1671 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1672 | */ |
f5f61a32 | 1673 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1674 | |
1675 | splitmap: | |
1676 | /* __isolate_free_page() does not map the pages */ | |
1677 | split_map_pages(freelist); | |
748446bb MG |
1678 | } |
1679 | ||
1680 | /* | |
1681 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1682 | * from the isolated freelists in the block we are migrating to. | |
1683 | */ | |
1684 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1685 | unsigned long data) |
748446bb MG |
1686 | { |
1687 | struct compact_control *cc = (struct compact_control *)data; | |
1688 | struct page *freepage; | |
1689 | ||
748446bb | 1690 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1691 | isolate_freepages(cc); |
748446bb MG |
1692 | |
1693 | if (list_empty(&cc->freepages)) | |
1694 | return NULL; | |
1695 | } | |
1696 | ||
1697 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1698 | list_del(&freepage->lru); | |
1699 | cc->nr_freepages--; | |
1700 | ||
1701 | return freepage; | |
1702 | } | |
1703 | ||
1704 | /* | |
d53aea3d DR |
1705 | * This is a migrate-callback that "frees" freepages back to the isolated |
1706 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1707 | * special handling needed for NUMA. | |
1708 | */ | |
1709 | static void compaction_free(struct page *page, unsigned long data) | |
1710 | { | |
1711 | struct compact_control *cc = (struct compact_control *)data; | |
1712 | ||
1713 | list_add(&page->lru, &cc->freepages); | |
1714 | cc->nr_freepages++; | |
1715 | } | |
1716 | ||
ff9543fd MN |
1717 | /* possible outcome of isolate_migratepages */ |
1718 | typedef enum { | |
1719 | ISOLATE_ABORT, /* Abort compaction now */ | |
1720 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1721 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1722 | } isolate_migrate_t; | |
1723 | ||
5bbe3547 EM |
1724 | /* |
1725 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1726 | * compactable pages. | |
1727 | */ | |
c7e0b3d0 | 1728 | int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT; |
5bbe3547 | 1729 | |
70b44595 MG |
1730 | static inline void |
1731 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1732 | { | |
1733 | if (cc->fast_start_pfn == ULONG_MAX) | |
1734 | return; | |
1735 | ||
1736 | if (!cc->fast_start_pfn) | |
1737 | cc->fast_start_pfn = pfn; | |
1738 | ||
1739 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1740 | } | |
1741 | ||
1742 | static inline unsigned long | |
1743 | reinit_migrate_pfn(struct compact_control *cc) | |
1744 | { | |
1745 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1746 | return cc->migrate_pfn; | |
1747 | ||
1748 | cc->migrate_pfn = cc->fast_start_pfn; | |
1749 | cc->fast_start_pfn = ULONG_MAX; | |
1750 | ||
1751 | return cc->migrate_pfn; | |
1752 | } | |
1753 | ||
1754 | /* | |
1755 | * Briefly search the free lists for a migration source that already has | |
1756 | * some free pages to reduce the number of pages that need migration | |
1757 | * before a pageblock is free. | |
1758 | */ | |
1759 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1760 | { | |
1761 | unsigned int limit = freelist_scan_limit(cc); | |
1762 | unsigned int nr_scanned = 0; | |
1763 | unsigned long distance; | |
1764 | unsigned long pfn = cc->migrate_pfn; | |
1765 | unsigned long high_pfn; | |
1766 | int order; | |
15d28d0d | 1767 | bool found_block = false; |
70b44595 MG |
1768 | |
1769 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1770 | if (cc->ignore_skip_hint) | |
1771 | return pfn; | |
1772 | ||
1773 | /* | |
1774 | * If the migrate_pfn is not at the start of a zone or the start | |
1775 | * of a pageblock then assume this is a continuation of a previous | |
1776 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1777 | */ | |
1778 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1779 | return pfn; | |
1780 | ||
1781 | /* | |
1782 | * For smaller orders, just linearly scan as the number of pages | |
1783 | * to migrate should be relatively small and does not necessarily | |
1784 | * justify freeing up a large block for a small allocation. | |
1785 | */ | |
1786 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1787 | return pfn; | |
1788 | ||
1789 | /* | |
1790 | * Only allow kcompactd and direct requests for movable pages to | |
1791 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1792 | * reduces the risk that a large movable pageblock is freed for | |
1793 | * an unmovable/reclaimable small allocation. | |
1794 | */ | |
1795 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1796 | return pfn; | |
1797 | ||
1798 | /* | |
1799 | * When starting the migration scanner, pick any pageblock within the | |
1800 | * first half of the search space. Otherwise try and pick a pageblock | |
1801 | * within the first eighth to reduce the chances that a migration | |
1802 | * target later becomes a source. | |
1803 | */ | |
1804 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1805 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1806 | distance >>= 2; | |
1807 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1808 | ||
1809 | for (order = cc->order - 1; | |
15d28d0d | 1810 | order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; |
70b44595 MG |
1811 | order--) { |
1812 | struct free_area *area = &cc->zone->free_area[order]; | |
1813 | struct list_head *freelist; | |
1814 | unsigned long flags; | |
1815 | struct page *freepage; | |
1816 | ||
1817 | if (!area->nr_free) | |
1818 | continue; | |
1819 | ||
1820 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1821 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1822 | list_for_each_entry(freepage, freelist, lru) { | |
1823 | unsigned long free_pfn; | |
1824 | ||
15d28d0d WY |
1825 | if (nr_scanned++ >= limit) { |
1826 | move_freelist_tail(freelist, freepage); | |
1827 | break; | |
1828 | } | |
1829 | ||
70b44595 MG |
1830 | free_pfn = page_to_pfn(freepage); |
1831 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1832 | /* |
1833 | * Avoid if skipped recently. Ideally it would | |
1834 | * move to the tail but even safe iteration of | |
1835 | * the list assumes an entry is deleted, not | |
1836 | * reordered. | |
1837 | */ | |
15d28d0d | 1838 | if (get_pageblock_skip(freepage)) |
70b44595 | 1839 | continue; |
70b44595 MG |
1840 | |
1841 | /* Reorder to so a future search skips recent pages */ | |
1842 | move_freelist_tail(freelist, freepage); | |
1843 | ||
e380bebe | 1844 | update_fast_start_pfn(cc, free_pfn); |
70b44595 | 1845 | pfn = pageblock_start_pfn(free_pfn); |
bbe832b9 RY |
1846 | if (pfn < cc->zone->zone_start_pfn) |
1847 | pfn = cc->zone->zone_start_pfn; | |
70b44595 | 1848 | cc->fast_search_fail = 0; |
15d28d0d | 1849 | found_block = true; |
70b44595 MG |
1850 | break; |
1851 | } | |
70b44595 MG |
1852 | } |
1853 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1854 | } | |
1855 | ||
1856 | cc->total_migrate_scanned += nr_scanned; | |
1857 | ||
1858 | /* | |
1859 | * If fast scanning failed then use a cached entry for a page block | |
1860 | * that had free pages as the basis for starting a linear scan. | |
1861 | */ | |
15d28d0d WY |
1862 | if (!found_block) { |
1863 | cc->fast_search_fail++; | |
70b44595 | 1864 | pfn = reinit_migrate_pfn(cc); |
15d28d0d | 1865 | } |
70b44595 MG |
1866 | return pfn; |
1867 | } | |
1868 | ||
ff9543fd | 1869 | /* |
edc2ca61 VB |
1870 | * Isolate all pages that can be migrated from the first suitable block, |
1871 | * starting at the block pointed to by the migrate scanner pfn within | |
1872 | * compact_control. | |
ff9543fd | 1873 | */ |
32aaf055 | 1874 | static isolate_migrate_t isolate_migratepages(struct compact_control *cc) |
ff9543fd | 1875 | { |
e1409c32 JK |
1876 | unsigned long block_start_pfn; |
1877 | unsigned long block_end_pfn; | |
1878 | unsigned long low_pfn; | |
edc2ca61 VB |
1879 | struct page *page; |
1880 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1881 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1882 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 1883 | bool fast_find_block; |
ff9543fd | 1884 | |
edc2ca61 VB |
1885 | /* |
1886 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
1887 | * initialized by compact_zone(). The first failure will use |
1888 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 1889 | */ |
70b44595 | 1890 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 1891 | block_start_pfn = pageblock_start_pfn(low_pfn); |
32aaf055 PL |
1892 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1893 | block_start_pfn = cc->zone->zone_start_pfn; | |
ff9543fd | 1894 | |
70b44595 MG |
1895 | /* |
1896 | * fast_find_migrateblock marks a pageblock skipped so to avoid | |
1897 | * the isolation_suitable check below, check whether the fast | |
1898 | * search was successful. | |
1899 | */ | |
1900 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
1901 | ||
ff9543fd | 1902 | /* Only scan within a pageblock boundary */ |
06b6640a | 1903 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1904 | |
edc2ca61 VB |
1905 | /* |
1906 | * Iterate over whole pageblocks until we find the first suitable. | |
1907 | * Do not cross the free scanner. | |
1908 | */ | |
e1409c32 | 1909 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 1910 | fast_find_block = false, |
c2ad7a1f | 1911 | cc->migrate_pfn = low_pfn = block_end_pfn, |
e1409c32 JK |
1912 | block_start_pfn = block_end_pfn, |
1913 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1914 | |
edc2ca61 VB |
1915 | /* |
1916 | * This can potentially iterate a massively long zone with | |
1917 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 1918 | * need to schedule. |
edc2ca61 | 1919 | */ |
c036ddff | 1920 | if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1921 | cond_resched(); |
ff9543fd | 1922 | |
32aaf055 PL |
1923 | page = pageblock_pfn_to_page(block_start_pfn, |
1924 | block_end_pfn, cc->zone); | |
7d49d886 | 1925 | if (!page) |
edc2ca61 VB |
1926 | continue; |
1927 | ||
e380bebe MG |
1928 | /* |
1929 | * If isolation recently failed, do not retry. Only check the | |
1930 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
1931 | * to be visited multiple times. Assume skip was checked | |
1932 | * before making it "skip" so other compaction instances do | |
1933 | * not scan the same block. | |
1934 | */ | |
ee0913c4 | 1935 | if (pageblock_aligned(low_pfn) && |
e380bebe | 1936 | !fast_find_block && !isolation_suitable(cc, page)) |
edc2ca61 VB |
1937 | continue; |
1938 | ||
1939 | /* | |
556162bf ML |
1940 | * For async direct compaction, only scan the pageblocks of the |
1941 | * same migratetype without huge pages. Async direct compaction | |
1942 | * is optimistic to see if the minimum amount of work satisfies | |
1943 | * the allocation. The cached PFN is updated as it's possible | |
1944 | * that all remaining blocks between source and target are | |
1945 | * unsuitable and the compaction scanners fail to meet. | |
edc2ca61 | 1946 | */ |
9bebefd5 MG |
1947 | if (!suitable_migration_source(cc, page)) { |
1948 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 1949 | continue; |
9bebefd5 | 1950 | } |
edc2ca61 VB |
1951 | |
1952 | /* Perform the isolation */ | |
c2ad7a1f OS |
1953 | if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, |
1954 | isolate_mode)) | |
edc2ca61 VB |
1955 | return ISOLATE_ABORT; |
1956 | ||
1957 | /* | |
1958 | * Either we isolated something and proceed with migration. Or | |
1959 | * we failed and compact_zone should decide if we should | |
1960 | * continue or not. | |
1961 | */ | |
1962 | break; | |
1963 | } | |
1964 | ||
edc2ca61 | 1965 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1966 | } |
1967 | ||
21c527a3 YB |
1968 | /* |
1969 | * order == -1 is expected when compacting via | |
1970 | * /proc/sys/vm/compact_memory | |
1971 | */ | |
1972 | static inline bool is_via_compact_memory(int order) | |
1973 | { | |
1974 | return order == -1; | |
1975 | } | |
1976 | ||
b4a0215e KW |
1977 | /* |
1978 | * Determine whether kswapd is (or recently was!) running on this node. | |
1979 | * | |
1980 | * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't | |
1981 | * zero it. | |
1982 | */ | |
facdaa91 NG |
1983 | static bool kswapd_is_running(pg_data_t *pgdat) |
1984 | { | |
b4a0215e KW |
1985 | bool running; |
1986 | ||
1987 | pgdat_kswapd_lock(pgdat); | |
1988 | running = pgdat->kswapd && task_is_running(pgdat->kswapd); | |
1989 | pgdat_kswapd_unlock(pgdat); | |
1990 | ||
1991 | return running; | |
facdaa91 NG |
1992 | } |
1993 | ||
1994 | /* | |
1995 | * A zone's fragmentation score is the external fragmentation wrt to the | |
40d7e203 CTR |
1996 | * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. |
1997 | */ | |
1998 | static unsigned int fragmentation_score_zone(struct zone *zone) | |
1999 | { | |
2000 | return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); | |
2001 | } | |
2002 | ||
2003 | /* | |
2004 | * A weighted zone's fragmentation score is the external fragmentation | |
2005 | * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It | |
2006 | * returns a value in the range [0, 100]. | |
facdaa91 NG |
2007 | * |
2008 | * The scaling factor ensures that proactive compaction focuses on larger | |
2009 | * zones like ZONE_NORMAL, rather than smaller, specialized zones like | |
2010 | * ZONE_DMA32. For smaller zones, the score value remains close to zero, | |
2011 | * and thus never exceeds the high threshold for proactive compaction. | |
2012 | */ | |
40d7e203 | 2013 | static unsigned int fragmentation_score_zone_weighted(struct zone *zone) |
facdaa91 NG |
2014 | { |
2015 | unsigned long score; | |
2016 | ||
40d7e203 | 2017 | score = zone->present_pages * fragmentation_score_zone(zone); |
facdaa91 NG |
2018 | return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); |
2019 | } | |
2020 | ||
2021 | /* | |
2022 | * The per-node proactive (background) compaction process is started by its | |
2023 | * corresponding kcompactd thread when the node's fragmentation score | |
2024 | * exceeds the high threshold. The compaction process remains active till | |
2025 | * the node's score falls below the low threshold, or one of the back-off | |
2026 | * conditions is met. | |
2027 | */ | |
d34c0a75 | 2028 | static unsigned int fragmentation_score_node(pg_data_t *pgdat) |
facdaa91 | 2029 | { |
d34c0a75 | 2030 | unsigned int score = 0; |
facdaa91 NG |
2031 | int zoneid; |
2032 | ||
2033 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2034 | struct zone *zone; | |
2035 | ||
2036 | zone = &pgdat->node_zones[zoneid]; | |
40d7e203 | 2037 | score += fragmentation_score_zone_weighted(zone); |
facdaa91 NG |
2038 | } |
2039 | ||
2040 | return score; | |
2041 | } | |
2042 | ||
d34c0a75 | 2043 | static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low) |
facdaa91 | 2044 | { |
d34c0a75 | 2045 | unsigned int wmark_low; |
facdaa91 NG |
2046 | |
2047 | /* | |
f0953a1b IM |
2048 | * Cap the low watermark to avoid excessive compaction |
2049 | * activity in case a user sets the proactiveness tunable | |
facdaa91 NG |
2050 | * close to 100 (maximum). |
2051 | */ | |
d34c0a75 NG |
2052 | wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); |
2053 | return low ? wmark_low : min(wmark_low + 10, 100U); | |
facdaa91 NG |
2054 | } |
2055 | ||
2056 | static bool should_proactive_compact_node(pg_data_t *pgdat) | |
2057 | { | |
2058 | int wmark_high; | |
2059 | ||
2060 | if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) | |
2061 | return false; | |
2062 | ||
2063 | wmark_high = fragmentation_score_wmark(pgdat, false); | |
2064 | return fragmentation_score_node(pgdat) > wmark_high; | |
2065 | } | |
2066 | ||
40cacbcb | 2067 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 2068 | { |
8fb74b9f | 2069 | unsigned int order; |
d39773a0 | 2070 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 2071 | int ret; |
748446bb | 2072 | |
753341a4 | 2073 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 2074 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 2075 | /* Let the next compaction start anew. */ |
40cacbcb | 2076 | reset_cached_positions(cc->zone); |
55b7c4c9 | 2077 | |
62997027 MG |
2078 | /* |
2079 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 2080 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
2081 | * flag itself as the decision to be clear should be directly |
2082 | * based on an allocation request. | |
2083 | */ | |
accf6242 | 2084 | if (cc->direct_compaction) |
40cacbcb | 2085 | cc->zone->compact_blockskip_flush = true; |
62997027 | 2086 | |
c8f7de0b MH |
2087 | if (cc->whole_zone) |
2088 | return COMPACT_COMPLETE; | |
2089 | else | |
2090 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 2091 | } |
748446bb | 2092 | |
facdaa91 NG |
2093 | if (cc->proactive_compaction) { |
2094 | int score, wmark_low; | |
2095 | pg_data_t *pgdat; | |
2096 | ||
2097 | pgdat = cc->zone->zone_pgdat; | |
2098 | if (kswapd_is_running(pgdat)) | |
2099 | return COMPACT_PARTIAL_SKIPPED; | |
2100 | ||
2101 | score = fragmentation_score_zone(cc->zone); | |
2102 | wmark_low = fragmentation_score_wmark(pgdat, true); | |
2103 | ||
2104 | if (score > wmark_low) | |
2105 | ret = COMPACT_CONTINUE; | |
2106 | else | |
2107 | ret = COMPACT_SUCCESS; | |
2108 | ||
2109 | goto out; | |
2110 | } | |
2111 | ||
21c527a3 | 2112 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
2113 | return COMPACT_CONTINUE; |
2114 | ||
efe771c7 MG |
2115 | /* |
2116 | * Always finish scanning a pageblock to reduce the possibility of | |
2117 | * fallbacks in the future. This is particularly important when | |
2118 | * migration source is unmovable/reclaimable but it's not worth | |
2119 | * special casing. | |
2120 | */ | |
ee0913c4 | 2121 | if (!pageblock_aligned(cc->migrate_pfn)) |
efe771c7 | 2122 | return COMPACT_CONTINUE; |
baf6a9a1 | 2123 | |
56de7263 | 2124 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 2125 | ret = COMPACT_NO_SUITABLE_PAGE; |
8fb74b9f | 2126 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 2127 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 2128 | bool can_steal; |
8fb74b9f MG |
2129 | |
2130 | /* Job done if page is free of the right migratetype */ | |
b03641af | 2131 | if (!free_area_empty(area, migratetype)) |
cf378319 | 2132 | return COMPACT_SUCCESS; |
8fb74b9f | 2133 | |
2149cdae JK |
2134 | #ifdef CONFIG_CMA |
2135 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
2136 | if (migratetype == MIGRATE_MOVABLE && | |
b03641af | 2137 | !free_area_empty(area, MIGRATE_CMA)) |
cf378319 | 2138 | return COMPACT_SUCCESS; |
2149cdae JK |
2139 | #endif |
2140 | /* | |
2141 | * Job done if allocation would steal freepages from | |
2142 | * other migratetype buddy lists. | |
2143 | */ | |
2144 | if (find_suitable_fallback(area, order, migratetype, | |
fa599c44 | 2145 | true, &can_steal) != -1) |
baf6a9a1 | 2146 | /* |
fa599c44 ML |
2147 | * Movable pages are OK in any pageblock. If we are |
2148 | * stealing for a non-movable allocation, make sure | |
2149 | * we finish compacting the current pageblock first | |
2150 | * (which is assured by the above migrate_pfn align | |
2151 | * check) so it is as free as possible and we won't | |
2152 | * have to steal another one soon. | |
baf6a9a1 | 2153 | */ |
fa599c44 | 2154 | return COMPACT_SUCCESS; |
56de7263 MG |
2155 | } |
2156 | ||
facdaa91 | 2157 | out: |
cb2dcaf0 MG |
2158 | if (cc->contended || fatal_signal_pending(current)) |
2159 | ret = COMPACT_CONTENDED; | |
2160 | ||
2161 | return ret; | |
837d026d JK |
2162 | } |
2163 | ||
40cacbcb | 2164 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
2165 | { |
2166 | int ret; | |
2167 | ||
40cacbcb MG |
2168 | ret = __compact_finished(cc); |
2169 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
2170 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
2171 | ret = COMPACT_CONTINUE; | |
2172 | ||
2173 | return ret; | |
748446bb MG |
2174 | } |
2175 | ||
ea7ab982 | 2176 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 2177 | unsigned int alloc_flags, |
97a225e6 | 2178 | int highest_zoneidx, |
86a294a8 | 2179 | unsigned long wmark_target) |
3e7d3449 | 2180 | { |
3e7d3449 MG |
2181 | unsigned long watermark; |
2182 | ||
21c527a3 | 2183 | if (is_via_compact_memory(order)) |
3957c776 MH |
2184 | return COMPACT_CONTINUE; |
2185 | ||
a9214443 | 2186 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
2187 | /* |
2188 | * If watermarks for high-order allocation are already met, there | |
2189 | * should be no need for compaction at all. | |
2190 | */ | |
97a225e6 | 2191 | if (zone_watermark_ok(zone, order, watermark, highest_zoneidx, |
ebff3980 | 2192 | alloc_flags)) |
cf378319 | 2193 | return COMPACT_SUCCESS; |
ebff3980 | 2194 | |
3e7d3449 | 2195 | /* |
9861a62c | 2196 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
2197 | * isolate free pages for migration targets. This means that the |
2198 | * watermark and alloc_flags have to match, or be more pessimistic than | |
2199 | * the check in __isolate_free_page(). We don't use the direct | |
2200 | * compactor's alloc_flags, as they are not relevant for freepage | |
97a225e6 JK |
2201 | * isolation. We however do use the direct compactor's highest_zoneidx |
2202 | * to skip over zones where lowmem reserves would prevent allocation | |
2203 | * even if compaction succeeds. | |
8348faf9 VB |
2204 | * For costly orders, we require low watermark instead of min for |
2205 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
2206 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
2207 | * suitable migration targets | |
3e7d3449 | 2208 | */ |
8348faf9 VB |
2209 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
2210 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
2211 | watermark += compact_gap(order); | |
97a225e6 | 2212 | if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx, |
d883c6cf | 2213 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
2214 | return COMPACT_SKIPPED; |
2215 | ||
cc5c9f09 VB |
2216 | return COMPACT_CONTINUE; |
2217 | } | |
2218 | ||
2b1a20c3 HS |
2219 | /* |
2220 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
2221 | * Returns | |
2222 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
2223 | * COMPACT_SUCCESS - If the allocation would succeed without compaction | |
2224 | * COMPACT_CONTINUE - If compaction should run now | |
2225 | */ | |
cc5c9f09 VB |
2226 | enum compact_result compaction_suitable(struct zone *zone, int order, |
2227 | unsigned int alloc_flags, | |
97a225e6 | 2228 | int highest_zoneidx) |
cc5c9f09 VB |
2229 | { |
2230 | enum compact_result ret; | |
2231 | int fragindex; | |
2232 | ||
97a225e6 | 2233 | ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx, |
cc5c9f09 | 2234 | zone_page_state(zone, NR_FREE_PAGES)); |
3e7d3449 MG |
2235 | /* |
2236 | * fragmentation index determines if allocation failures are due to | |
2237 | * low memory or external fragmentation | |
2238 | * | |
ebff3980 VB |
2239 | * index of -1000 would imply allocations might succeed depending on |
2240 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2241 | * index towards 0 implies failure is due to lack of memory |
2242 | * index towards 1000 implies failure is due to fragmentation | |
2243 | * | |
20311420 VB |
2244 | * Only compact if a failure would be due to fragmentation. Also |
2245 | * ignore fragindex for non-costly orders where the alternative to | |
2246 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2247 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2248 | * excessive compaction for costly orders, but it should not be at the | |
2249 | * expense of system stability. | |
3e7d3449 | 2250 | */ |
20311420 | 2251 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
2252 | fragindex = fragmentation_index(zone, order); |
2253 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
2254 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
2255 | } | |
837d026d | 2256 | |
837d026d JK |
2257 | trace_mm_compaction_suitable(zone, order, ret); |
2258 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
2259 | ret = COMPACT_SKIPPED; | |
2260 | ||
2261 | return ret; | |
2262 | } | |
2263 | ||
86a294a8 MH |
2264 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2265 | int alloc_flags) | |
2266 | { | |
2267 | struct zone *zone; | |
2268 | struct zoneref *z; | |
2269 | ||
2270 | /* | |
2271 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2272 | * retrying the reclaim. | |
2273 | */ | |
97a225e6 JK |
2274 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2275 | ac->highest_zoneidx, ac->nodemask) { | |
86a294a8 MH |
2276 | unsigned long available; |
2277 | enum compact_result compact_result; | |
2278 | ||
2279 | /* | |
2280 | * Do not consider all the reclaimable memory because we do not | |
2281 | * want to trash just for a single high order allocation which | |
2282 | * is even not guaranteed to appear even if __compaction_suitable | |
2283 | * is happy about the watermark check. | |
2284 | */ | |
5a1c84b4 | 2285 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
2286 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
2287 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
97a225e6 | 2288 | ac->highest_zoneidx, available); |
cff387d6 | 2289 | if (compact_result == COMPACT_CONTINUE) |
86a294a8 MH |
2290 | return true; |
2291 | } | |
2292 | ||
2293 | return false; | |
2294 | } | |
2295 | ||
5e1f0f09 MG |
2296 | static enum compact_result |
2297 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2298 | { |
ea7ab982 | 2299 | enum compact_result ret; |
40cacbcb MG |
2300 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2301 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2302 | unsigned long last_migrated_pfn; |
e0b9daeb | 2303 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2304 | bool update_cached; |
84b328aa | 2305 | unsigned int nr_succeeded = 0; |
748446bb | 2306 | |
a94b5252 YS |
2307 | /* |
2308 | * These counters track activities during zone compaction. Initialize | |
2309 | * them before compacting a new zone. | |
2310 | */ | |
2311 | cc->total_migrate_scanned = 0; | |
2312 | cc->total_free_scanned = 0; | |
2313 | cc->nr_migratepages = 0; | |
2314 | cc->nr_freepages = 0; | |
2315 | INIT_LIST_HEAD(&cc->freepages); | |
2316 | INIT_LIST_HEAD(&cc->migratepages); | |
2317 | ||
01c0bfe0 | 2318 | cc->migratetype = gfp_migratetype(cc->gfp_mask); |
40cacbcb | 2319 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
97a225e6 | 2320 | cc->highest_zoneidx); |
c46649de | 2321 | /* Compaction is likely to fail */ |
cf378319 | 2322 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 2323 | return ret; |
c46649de MH |
2324 | |
2325 | /* huh, compaction_suitable is returning something unexpected */ | |
2326 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 2327 | |
d3132e4b VB |
2328 | /* |
2329 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2330 | * is about to be retried after being deferred. |
d3132e4b | 2331 | */ |
40cacbcb MG |
2332 | if (compaction_restarting(cc->zone, cc->order)) |
2333 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2334 | |
c89511ab MG |
2335 | /* |
2336 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2337 | * information on where the scanners should start (unless we explicitly |
2338 | * want to compact the whole zone), but check that it is initialised | |
2339 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2340 | */ |
70b44595 | 2341 | cc->fast_start_pfn = 0; |
06ed2998 | 2342 | if (cc->whole_zone) { |
c89511ab | 2343 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2344 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2345 | } else { | |
40cacbcb MG |
2346 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2347 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2348 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2349 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2350 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2351 | } |
2352 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2353 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2354 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2355 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2356 | } |
c8f7de0b | 2357 | |
e332f741 | 2358 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2359 | cc->whole_zone = true; |
2360 | } | |
c8f7de0b | 2361 | |
566e54e1 | 2362 | last_migrated_pfn = 0; |
748446bb | 2363 | |
8854c55f MG |
2364 | /* |
2365 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2366 | * the basis that some migrations will fail in ASYNC mode. However, | |
2367 | * if the cached PFNs match and pageblocks are skipped due to having | |
2368 | * no isolation candidates, then the sync state does not matter. | |
2369 | * Until a pageblock with isolation candidates is found, keep the | |
2370 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2371 | */ | |
2372 | update_cached = !sync && | |
2373 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2374 | ||
abd4349f | 2375 | trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); |
0eb927c0 | 2376 | |
361a2a22 MK |
2377 | /* lru_add_drain_all could be expensive with involving other CPUs */ |
2378 | lru_add_drain(); | |
748446bb | 2379 | |
40cacbcb | 2380 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2381 | int err; |
19d3cf9d | 2382 | unsigned long iteration_start_pfn = cc->migrate_pfn; |
748446bb | 2383 | |
804d3121 MG |
2384 | /* |
2385 | * Avoid multiple rescans which can happen if a page cannot be | |
2386 | * isolated (dirty/writeback in async mode) or if the migrated | |
2387 | * pages are being allocated before the pageblock is cleared. | |
2388 | * The first rescan will capture the entire pageblock for | |
2389 | * migration. If it fails, it'll be marked skip and scanning | |
2390 | * will proceed as normal. | |
2391 | */ | |
2392 | cc->rescan = false; | |
2393 | if (pageblock_start_pfn(last_migrated_pfn) == | |
19d3cf9d | 2394 | pageblock_start_pfn(iteration_start_pfn)) { |
804d3121 MG |
2395 | cc->rescan = true; |
2396 | } | |
2397 | ||
32aaf055 | 2398 | switch (isolate_migratepages(cc)) { |
f9e35b3b | 2399 | case ISOLATE_ABORT: |
2d1e1041 | 2400 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2401 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2402 | cc->nr_migratepages = 0; |
f9e35b3b MG |
2403 | goto out; |
2404 | case ISOLATE_NONE: | |
8854c55f MG |
2405 | if (update_cached) { |
2406 | cc->zone->compact_cached_migrate_pfn[1] = | |
2407 | cc->zone->compact_cached_migrate_pfn[0]; | |
2408 | } | |
2409 | ||
fdaf7f5c VB |
2410 | /* |
2411 | * We haven't isolated and migrated anything, but | |
2412 | * there might still be unflushed migrations from | |
2413 | * previous cc->order aligned block. | |
2414 | */ | |
2415 | goto check_drain; | |
f9e35b3b | 2416 | case ISOLATE_SUCCESS: |
8854c55f | 2417 | update_cached = false; |
19d3cf9d | 2418 | last_migrated_pfn = iteration_start_pfn; |
f9e35b3b | 2419 | } |
748446bb | 2420 | |
d53aea3d | 2421 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2422 | compaction_free, (unsigned long)cc, cc->mode, |
84b328aa | 2423 | MR_COMPACTION, &nr_succeeded); |
748446bb | 2424 | |
abd4349f | 2425 | trace_mm_compaction_migratepages(cc, nr_succeeded); |
748446bb | 2426 | |
f8c9301f VB |
2427 | /* All pages were either migrated or will be released */ |
2428 | cc->nr_migratepages = 0; | |
9d502c1c | 2429 | if (err) { |
5733c7d1 | 2430 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2431 | /* |
2432 | * migrate_pages() may return -ENOMEM when scanners meet | |
2433 | * and we want compact_finished() to detect it | |
2434 | */ | |
f2849aa0 | 2435 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2436 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2437 | goto out; |
2438 | } | |
fdd048e1 VB |
2439 | /* |
2440 | * We failed to migrate at least one page in the current | |
2441 | * order-aligned block, so skip the rest of it. | |
2442 | */ | |
2443 | if (cc->direct_compaction && | |
2444 | (cc->mode == MIGRATE_ASYNC)) { | |
2445 | cc->migrate_pfn = block_end_pfn( | |
2446 | cc->migrate_pfn - 1, cc->order); | |
2447 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 2448 | last_migrated_pfn = 0; |
fdd048e1 | 2449 | } |
748446bb | 2450 | } |
fdaf7f5c | 2451 | |
fdaf7f5c VB |
2452 | check_drain: |
2453 | /* | |
2454 | * Has the migration scanner moved away from the previous | |
2455 | * cc->order aligned block where we migrated from? If yes, | |
2456 | * flush the pages that were freed, so that they can merge and | |
2457 | * compact_finished() can detect immediately if allocation | |
2458 | * would succeed. | |
2459 | */ | |
566e54e1 | 2460 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c | 2461 | unsigned long current_block_start = |
06b6640a | 2462 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2463 | |
566e54e1 | 2464 | if (last_migrated_pfn < current_block_start) { |
b01b2141 | 2465 | lru_add_drain_cpu_zone(cc->zone); |
fdaf7f5c | 2466 | /* No more flushing until we migrate again */ |
566e54e1 | 2467 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2468 | } |
2469 | } | |
2470 | ||
5e1f0f09 MG |
2471 | /* Stop if a page has been captured */ |
2472 | if (capc && capc->page) { | |
2473 | ret = COMPACT_SUCCESS; | |
2474 | break; | |
2475 | } | |
748446bb MG |
2476 | } |
2477 | ||
f9e35b3b | 2478 | out: |
6bace090 VB |
2479 | /* |
2480 | * Release free pages and update where the free scanner should restart, | |
2481 | * so we don't leave any returned pages behind in the next attempt. | |
2482 | */ | |
2483 | if (cc->nr_freepages > 0) { | |
2484 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2485 | ||
2486 | cc->nr_freepages = 0; | |
2487 | VM_BUG_ON(free_pfn == 0); | |
2488 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2489 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2490 | /* |
2491 | * Only go back, not forward. The cached pfn might have been | |
2492 | * already reset to zone end in compact_finished() | |
2493 | */ | |
40cacbcb MG |
2494 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2495 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2496 | } |
748446bb | 2497 | |
7f354a54 DR |
2498 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2499 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2500 | ||
abd4349f | 2501 | trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); |
0eb927c0 | 2502 | |
748446bb MG |
2503 | return ret; |
2504 | } | |
76ab0f53 | 2505 | |
ea7ab982 | 2506 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2507 | gfp_t gfp_mask, enum compact_priority prio, |
97a225e6 | 2508 | unsigned int alloc_flags, int highest_zoneidx, |
5e1f0f09 | 2509 | struct page **capture) |
56de7263 | 2510 | { |
ea7ab982 | 2511 | enum compact_result ret; |
56de7263 | 2512 | struct compact_control cc = { |
56de7263 | 2513 | .order = order, |
dbe2d4e4 | 2514 | .search_order = order, |
6d7ce559 | 2515 | .gfp_mask = gfp_mask, |
56de7263 | 2516 | .zone = zone, |
a5508cd8 VB |
2517 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2518 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 | 2519 | .alloc_flags = alloc_flags, |
97a225e6 | 2520 | .highest_zoneidx = highest_zoneidx, |
accf6242 | 2521 | .direct_compaction = true, |
a8e025e5 | 2522 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2523 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2524 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2525 | }; |
5e1f0f09 MG |
2526 | struct capture_control capc = { |
2527 | .cc = &cc, | |
2528 | .page = NULL, | |
2529 | }; | |
2530 | ||
b9e20f0d VB |
2531 | /* |
2532 | * Make sure the structs are really initialized before we expose the | |
2533 | * capture control, in case we are interrupted and the interrupt handler | |
2534 | * frees a page. | |
2535 | */ | |
2536 | barrier(); | |
2537 | WRITE_ONCE(current->capture_control, &capc); | |
56de7263 | 2538 | |
5e1f0f09 | 2539 | ret = compact_zone(&cc, &capc); |
e64c5237 SL |
2540 | |
2541 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2542 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2543 | ||
b9e20f0d VB |
2544 | /* |
2545 | * Make sure we hide capture control first before we read the captured | |
2546 | * page pointer, otherwise an interrupt could free and capture a page | |
2547 | * and we would leak it. | |
2548 | */ | |
2549 | WRITE_ONCE(current->capture_control, NULL); | |
2550 | *capture = READ_ONCE(capc.page); | |
06dac2f4 CTR |
2551 | /* |
2552 | * Technically, it is also possible that compaction is skipped but | |
2553 | * the page is still captured out of luck(IRQ came and freed the page). | |
2554 | * Returning COMPACT_SUCCESS in such cases helps in properly accounting | |
2555 | * the COMPACT[STALL|FAIL] when compaction is skipped. | |
2556 | */ | |
2557 | if (*capture) | |
2558 | ret = COMPACT_SUCCESS; | |
5e1f0f09 | 2559 | |
e64c5237 | 2560 | return ret; |
56de7263 MG |
2561 | } |
2562 | ||
5e771905 MG |
2563 | int sysctl_extfrag_threshold = 500; |
2564 | ||
56de7263 MG |
2565 | /** |
2566 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2567 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2568 | * @order: The order of the current allocation |
2569 | * @alloc_flags: The allocation flags of the current allocation | |
2570 | * @ac: The context of current allocation | |
112d2d29 | 2571 | * @prio: Determines how hard direct compaction should try to succeed |
6467552c | 2572 | * @capture: Pointer to free page created by compaction will be stored here |
56de7263 MG |
2573 | * |
2574 | * This is the main entry point for direct page compaction. | |
2575 | */ | |
ea7ab982 | 2576 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2577 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2578 | enum compact_priority prio, struct page **capture) |
56de7263 | 2579 | { |
fe573327 | 2580 | int may_perform_io = (__force int)(gfp_mask & __GFP_IO); |
56de7263 MG |
2581 | struct zoneref *z; |
2582 | struct zone *zone; | |
1d4746d3 | 2583 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2584 | |
73e64c51 MH |
2585 | /* |
2586 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2587 | * tricky context because the migration might require IO | |
2588 | */ | |
2589 | if (!may_perform_io) | |
53853e2d | 2590 | return COMPACT_SKIPPED; |
56de7263 | 2591 | |
a5508cd8 | 2592 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2593 | |
56de7263 | 2594 | /* Compact each zone in the list */ |
97a225e6 JK |
2595 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2596 | ac->highest_zoneidx, ac->nodemask) { | |
ea7ab982 | 2597 | enum compact_result status; |
56de7263 | 2598 | |
a8e025e5 VB |
2599 | if (prio > MIN_COMPACT_PRIORITY |
2600 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2601 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2602 | continue; |
1d4746d3 | 2603 | } |
53853e2d | 2604 | |
a5508cd8 | 2605 | status = compact_zone_order(zone, order, gfp_mask, prio, |
97a225e6 | 2606 | alloc_flags, ac->highest_zoneidx, capture); |
56de7263 MG |
2607 | rc = max(status, rc); |
2608 | ||
7ceb009a VB |
2609 | /* The allocation should succeed, stop compacting */ |
2610 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2611 | /* |
2612 | * We think the allocation will succeed in this zone, | |
2613 | * but it is not certain, hence the false. The caller | |
2614 | * will repeat this with true if allocation indeed | |
2615 | * succeeds in this zone. | |
2616 | */ | |
2617 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2618 | |
c3486f53 | 2619 | break; |
1f9efdef VB |
2620 | } |
2621 | ||
a5508cd8 | 2622 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2623 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2624 | /* |
2625 | * We think that allocation won't succeed in this zone | |
2626 | * so we defer compaction there. If it ends up | |
2627 | * succeeding after all, it will be reset. | |
2628 | */ | |
2629 | defer_compaction(zone, order); | |
1f9efdef VB |
2630 | |
2631 | /* | |
2632 | * We might have stopped compacting due to need_resched() in | |
2633 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2634 | * case do not try further zones |
1f9efdef | 2635 | */ |
c3486f53 VB |
2636 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2637 | || fatal_signal_pending(current)) | |
2638 | break; | |
56de7263 MG |
2639 | } |
2640 | ||
2641 | return rc; | |
2642 | } | |
2643 | ||
facdaa91 NG |
2644 | /* |
2645 | * Compact all zones within a node till each zone's fragmentation score | |
2646 | * reaches within proactive compaction thresholds (as determined by the | |
2647 | * proactiveness tunable). | |
2648 | * | |
2649 | * It is possible that the function returns before reaching score targets | |
2650 | * due to various back-off conditions, such as, contention on per-node or | |
2651 | * per-zone locks. | |
2652 | */ | |
2653 | static void proactive_compact_node(pg_data_t *pgdat) | |
2654 | { | |
2655 | int zoneid; | |
2656 | struct zone *zone; | |
2657 | struct compact_control cc = { | |
2658 | .order = -1, | |
2659 | .mode = MIGRATE_SYNC_LIGHT, | |
2660 | .ignore_skip_hint = true, | |
2661 | .whole_zone = true, | |
2662 | .gfp_mask = GFP_KERNEL, | |
2663 | .proactive_compaction = true, | |
2664 | }; | |
2665 | ||
2666 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2667 | zone = &pgdat->node_zones[zoneid]; | |
2668 | if (!populated_zone(zone)) | |
2669 | continue; | |
2670 | ||
2671 | cc.zone = zone; | |
2672 | ||
2673 | compact_zone(&cc, NULL); | |
2674 | ||
2675 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2676 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2677 | } | |
2678 | } | |
56de7263 | 2679 | |
76ab0f53 | 2680 | /* Compact all zones within a node */ |
791cae96 | 2681 | static void compact_node(int nid) |
76ab0f53 | 2682 | { |
791cae96 | 2683 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 2684 | int zoneid; |
76ab0f53 | 2685 | struct zone *zone; |
791cae96 VB |
2686 | struct compact_control cc = { |
2687 | .order = -1, | |
2688 | .mode = MIGRATE_SYNC, | |
2689 | .ignore_skip_hint = true, | |
2690 | .whole_zone = true, | |
73e64c51 | 2691 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
2692 | }; |
2693 | ||
76ab0f53 | 2694 | |
76ab0f53 | 2695 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
2696 | |
2697 | zone = &pgdat->node_zones[zoneid]; | |
2698 | if (!populated_zone(zone)) | |
2699 | continue; | |
2700 | ||
791cae96 | 2701 | cc.zone = zone; |
76ab0f53 | 2702 | |
5e1f0f09 | 2703 | compact_zone(&cc, NULL); |
75469345 | 2704 | |
791cae96 VB |
2705 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2706 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 2707 | } |
76ab0f53 MG |
2708 | } |
2709 | ||
2710 | /* Compact all nodes in the system */ | |
7964c06d | 2711 | static void compact_nodes(void) |
76ab0f53 MG |
2712 | { |
2713 | int nid; | |
2714 | ||
8575ec29 HD |
2715 | /* Flush pending updates to the LRU lists */ |
2716 | lru_add_drain_all(); | |
2717 | ||
76ab0f53 MG |
2718 | for_each_online_node(nid) |
2719 | compact_node(nid); | |
76ab0f53 MG |
2720 | } |
2721 | ||
facdaa91 NG |
2722 | /* |
2723 | * Tunable for proactive compaction. It determines how | |
2724 | * aggressively the kernel should compact memory in the | |
2725 | * background. It takes values in the range [0, 100]. | |
2726 | */ | |
d34c0a75 | 2727 | unsigned int __read_mostly sysctl_compaction_proactiveness = 20; |
facdaa91 | 2728 | |
65d759c8 CTR |
2729 | int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, |
2730 | void *buffer, size_t *length, loff_t *ppos) | |
2731 | { | |
2732 | int rc, nid; | |
2733 | ||
2734 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
2735 | if (rc) | |
2736 | return rc; | |
2737 | ||
2738 | if (write && sysctl_compaction_proactiveness) { | |
2739 | for_each_online_node(nid) { | |
2740 | pg_data_t *pgdat = NODE_DATA(nid); | |
2741 | ||
2742 | if (pgdat->proactive_compact_trigger) | |
2743 | continue; | |
2744 | ||
2745 | pgdat->proactive_compact_trigger = true; | |
2746 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2747 | } | |
2748 | } | |
2749 | ||
2750 | return 0; | |
2751 | } | |
2752 | ||
fec4eb2c YB |
2753 | /* |
2754 | * This is the entry point for compacting all nodes via | |
2755 | * /proc/sys/vm/compact_memory | |
2756 | */ | |
76ab0f53 | 2757 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
32927393 | 2758 | void *buffer, size_t *length, loff_t *ppos) |
76ab0f53 MG |
2759 | { |
2760 | if (write) | |
7964c06d | 2761 | compact_nodes(); |
76ab0f53 MG |
2762 | |
2763 | return 0; | |
2764 | } | |
ed4a6d7f MG |
2765 | |
2766 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
17adb230 Y |
2767 | static ssize_t compact_store(struct device *dev, |
2768 | struct device_attribute *attr, | |
2769 | const char *buf, size_t count) | |
ed4a6d7f | 2770 | { |
8575ec29 HD |
2771 | int nid = dev->id; |
2772 | ||
2773 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2774 | /* Flush pending updates to the LRU lists */ | |
2775 | lru_add_drain_all(); | |
2776 | ||
2777 | compact_node(nid); | |
2778 | } | |
ed4a6d7f MG |
2779 | |
2780 | return count; | |
2781 | } | |
17adb230 | 2782 | static DEVICE_ATTR_WO(compact); |
ed4a6d7f MG |
2783 | |
2784 | int compaction_register_node(struct node *node) | |
2785 | { | |
10fbcf4c | 2786 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2787 | } |
2788 | ||
2789 | void compaction_unregister_node(struct node *node) | |
2790 | { | |
10fbcf4c | 2791 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2792 | } |
2793 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2794 | |
698b1b30 VB |
2795 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2796 | { | |
65d759c8 CTR |
2797 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || |
2798 | pgdat->proactive_compact_trigger; | |
698b1b30 VB |
2799 | } |
2800 | ||
2801 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2802 | { | |
2803 | int zoneid; | |
2804 | struct zone *zone; | |
97a225e6 | 2805 | enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; |
698b1b30 | 2806 | |
97a225e6 | 2807 | for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { |
698b1b30 VB |
2808 | zone = &pgdat->node_zones[zoneid]; |
2809 | ||
2810 | if (!populated_zone(zone)) | |
2811 | continue; | |
2812 | ||
2813 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
97a225e6 | 2814 | highest_zoneidx) == COMPACT_CONTINUE) |
698b1b30 VB |
2815 | return true; |
2816 | } | |
2817 | ||
2818 | return false; | |
2819 | } | |
2820 | ||
2821 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2822 | { | |
2823 | /* | |
2824 | * With no special task, compact all zones so that a page of requested | |
2825 | * order is allocatable. | |
2826 | */ | |
2827 | int zoneid; | |
2828 | struct zone *zone; | |
2829 | struct compact_control cc = { | |
2830 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 2831 | .search_order = pgdat->kcompactd_max_order, |
97a225e6 | 2832 | .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, |
698b1b30 | 2833 | .mode = MIGRATE_SYNC_LIGHT, |
a0647dc9 | 2834 | .ignore_skip_hint = false, |
73e64c51 | 2835 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2836 | }; |
698b1b30 | 2837 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
97a225e6 | 2838 | cc.highest_zoneidx); |
7f354a54 | 2839 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2840 | |
97a225e6 | 2841 | for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { |
698b1b30 VB |
2842 | int status; |
2843 | ||
2844 | zone = &pgdat->node_zones[zoneid]; | |
2845 | if (!populated_zone(zone)) | |
2846 | continue; | |
2847 | ||
2848 | if (compaction_deferred(zone, cc.order)) | |
2849 | continue; | |
2850 | ||
2851 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
2852 | COMPACT_CONTINUE) | |
2853 | continue; | |
2854 | ||
172400c6 VB |
2855 | if (kthread_should_stop()) |
2856 | return; | |
a94b5252 YS |
2857 | |
2858 | cc.zone = zone; | |
5e1f0f09 | 2859 | status = compact_zone(&cc, NULL); |
698b1b30 | 2860 | |
7ceb009a | 2861 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 2862 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 2863 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
2864 | /* |
2865 | * Buddy pages may become stranded on pcps that could | |
2866 | * otherwise coalesce on the zone's free area for | |
2867 | * order >= cc.order. This is ratelimited by the | |
2868 | * upcoming deferral. | |
2869 | */ | |
2870 | drain_all_pages(zone); | |
2871 | ||
698b1b30 VB |
2872 | /* |
2873 | * We use sync migration mode here, so we defer like | |
2874 | * sync direct compaction does. | |
2875 | */ | |
2876 | defer_compaction(zone, cc.order); | |
2877 | } | |
2878 | ||
7f354a54 DR |
2879 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2880 | cc.total_migrate_scanned); | |
2881 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2882 | cc.total_free_scanned); | |
2883 | ||
698b1b30 VB |
2884 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2885 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2886 | } | |
2887 | ||
2888 | /* | |
2889 | * Regardless of success, we are done until woken up next. But remember | |
97a225e6 JK |
2890 | * the requested order/highest_zoneidx in case it was higher/tighter |
2891 | * than our current ones | |
698b1b30 VB |
2892 | */ |
2893 | if (pgdat->kcompactd_max_order <= cc.order) | |
2894 | pgdat->kcompactd_max_order = 0; | |
97a225e6 JK |
2895 | if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) |
2896 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; | |
698b1b30 VB |
2897 | } |
2898 | ||
97a225e6 | 2899 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) |
698b1b30 VB |
2900 | { |
2901 | if (!order) | |
2902 | return; | |
2903 | ||
2904 | if (pgdat->kcompactd_max_order < order) | |
2905 | pgdat->kcompactd_max_order = order; | |
2906 | ||
97a225e6 JK |
2907 | if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) |
2908 | pgdat->kcompactd_highest_zoneidx = highest_zoneidx; | |
698b1b30 | 2909 | |
6818600f DB |
2910 | /* |
2911 | * Pairs with implicit barrier in wait_event_freezable() | |
2912 | * such that wakeups are not missed. | |
2913 | */ | |
2914 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2915 | return; |
2916 | ||
2917 | if (!kcompactd_node_suitable(pgdat)) | |
2918 | return; | |
2919 | ||
2920 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
97a225e6 | 2921 | highest_zoneidx); |
698b1b30 VB |
2922 | wake_up_interruptible(&pgdat->kcompactd_wait); |
2923 | } | |
2924 | ||
2925 | /* | |
2926 | * The background compaction daemon, started as a kernel thread | |
2927 | * from the init process. | |
2928 | */ | |
2929 | static int kcompactd(void *p) | |
2930 | { | |
68d68ff6 | 2931 | pg_data_t *pgdat = (pg_data_t *)p; |
698b1b30 | 2932 | struct task_struct *tsk = current; |
e1e92bfa CTR |
2933 | long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); |
2934 | long timeout = default_timeout; | |
698b1b30 VB |
2935 | |
2936 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2937 | ||
2938 | if (!cpumask_empty(cpumask)) | |
2939 | set_cpus_allowed_ptr(tsk, cpumask); | |
2940 | ||
2941 | set_freezable(); | |
2942 | ||
2943 | pgdat->kcompactd_max_order = 0; | |
97a225e6 | 2944 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; |
698b1b30 VB |
2945 | |
2946 | while (!kthread_should_stop()) { | |
eb414681 JW |
2947 | unsigned long pflags; |
2948 | ||
65d759c8 CTR |
2949 | /* |
2950 | * Avoid the unnecessary wakeup for proactive compaction | |
2951 | * when it is disabled. | |
2952 | */ | |
2953 | if (!sysctl_compaction_proactiveness) | |
2954 | timeout = MAX_SCHEDULE_TIMEOUT; | |
698b1b30 | 2955 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
facdaa91 | 2956 | if (wait_event_freezable_timeout(pgdat->kcompactd_wait, |
65d759c8 CTR |
2957 | kcompactd_work_requested(pgdat), timeout) && |
2958 | !pgdat->proactive_compact_trigger) { | |
facdaa91 NG |
2959 | |
2960 | psi_memstall_enter(&pflags); | |
2961 | kcompactd_do_work(pgdat); | |
2962 | psi_memstall_leave(&pflags); | |
e1e92bfa CTR |
2963 | /* |
2964 | * Reset the timeout value. The defer timeout from | |
2965 | * proactive compaction is lost here but that is fine | |
2966 | * as the condition of the zone changing substantionally | |
2967 | * then carrying on with the previous defer interval is | |
2968 | * not useful. | |
2969 | */ | |
2970 | timeout = default_timeout; | |
facdaa91 NG |
2971 | continue; |
2972 | } | |
698b1b30 | 2973 | |
e1e92bfa CTR |
2974 | /* |
2975 | * Start the proactive work with default timeout. Based | |
2976 | * on the fragmentation score, this timeout is updated. | |
2977 | */ | |
2978 | timeout = default_timeout; | |
facdaa91 NG |
2979 | if (should_proactive_compact_node(pgdat)) { |
2980 | unsigned int prev_score, score; | |
2981 | ||
facdaa91 NG |
2982 | prev_score = fragmentation_score_node(pgdat); |
2983 | proactive_compact_node(pgdat); | |
2984 | score = fragmentation_score_node(pgdat); | |
2985 | /* | |
2986 | * Defer proactive compaction if the fragmentation | |
2987 | * score did not go down i.e. no progress made. | |
2988 | */ | |
e1e92bfa CTR |
2989 | if (unlikely(score >= prev_score)) |
2990 | timeout = | |
2991 | default_timeout << COMPACT_MAX_DEFER_SHIFT; | |
facdaa91 | 2992 | } |
65d759c8 CTR |
2993 | if (unlikely(pgdat->proactive_compact_trigger)) |
2994 | pgdat->proactive_compact_trigger = false; | |
698b1b30 VB |
2995 | } |
2996 | ||
2997 | return 0; | |
2998 | } | |
2999 | ||
3000 | /* | |
3001 | * This kcompactd start function will be called by init and node-hot-add. | |
3002 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
3003 | */ | |
024c61ea | 3004 | void kcompactd_run(int nid) |
698b1b30 VB |
3005 | { |
3006 | pg_data_t *pgdat = NODE_DATA(nid); | |
698b1b30 VB |
3007 | |
3008 | if (pgdat->kcompactd) | |
024c61ea | 3009 | return; |
698b1b30 VB |
3010 | |
3011 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
3012 | if (IS_ERR(pgdat->kcompactd)) { | |
3013 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
698b1b30 VB |
3014 | pgdat->kcompactd = NULL; |
3015 | } | |
698b1b30 VB |
3016 | } |
3017 | ||
3018 | /* | |
3019 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
e8da368a | 3020 | * be holding mem_hotplug_begin/done(). |
698b1b30 VB |
3021 | */ |
3022 | void kcompactd_stop(int nid) | |
3023 | { | |
3024 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
3025 | ||
3026 | if (kcompactd) { | |
3027 | kthread_stop(kcompactd); | |
3028 | NODE_DATA(nid)->kcompactd = NULL; | |
3029 | } | |
3030 | } | |
3031 | ||
3032 | /* | |
3033 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
3034 | * not required for correctness. So if the last cpu in a node goes | |
3035 | * away, we get changed to run anywhere: as the first one comes back, | |
3036 | * restore their cpu bindings. | |
3037 | */ | |
e46b1db2 | 3038 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
3039 | { |
3040 | int nid; | |
3041 | ||
e46b1db2 AMG |
3042 | for_each_node_state(nid, N_MEMORY) { |
3043 | pg_data_t *pgdat = NODE_DATA(nid); | |
3044 | const struct cpumask *mask; | |
698b1b30 | 3045 | |
e46b1db2 | 3046 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 3047 | |
e46b1db2 AMG |
3048 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3049 | /* One of our CPUs online: restore mask */ | |
3109de30 ML |
3050 | if (pgdat->kcompactd) |
3051 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 3052 | } |
e46b1db2 | 3053 | return 0; |
698b1b30 VB |
3054 | } |
3055 | ||
3056 | static int __init kcompactd_init(void) | |
3057 | { | |
3058 | int nid; | |
e46b1db2 AMG |
3059 | int ret; |
3060 | ||
3061 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
3062 | "mm/compaction:online", | |
3063 | kcompactd_cpu_online, NULL); | |
3064 | if (ret < 0) { | |
3065 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
3066 | return ret; | |
3067 | } | |
698b1b30 VB |
3068 | |
3069 | for_each_node_state(nid, N_MEMORY) | |
3070 | kcompactd_run(nid); | |
698b1b30 VB |
3071 | return 0; |
3072 | } | |
3073 | subsys_initcall(kcompactd_init) | |
3074 | ||
ff9543fd | 3075 | #endif /* CONFIG_COMPACTION */ |