Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
748446bb MG |
2 | /* |
3 | * linux/mm/compaction.c | |
4 | * | |
5 | * Memory compaction for the reduction of external fragmentation. Note that | |
6 | * this heavily depends upon page migration to do all the real heavy | |
7 | * lifting | |
8 | * | |
9 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
10 | */ | |
698b1b30 | 11 | #include <linux/cpu.h> |
748446bb MG |
12 | #include <linux/swap.h> |
13 | #include <linux/migrate.h> | |
14 | #include <linux/compaction.h> | |
15 | #include <linux/mm_inline.h> | |
174cd4b1 | 16 | #include <linux/sched/signal.h> |
748446bb | 17 | #include <linux/backing-dev.h> |
76ab0f53 | 18 | #include <linux/sysctl.h> |
ed4a6d7f | 19 | #include <linux/sysfs.h> |
194159fb | 20 | #include <linux/page-isolation.h> |
b8c73fc2 | 21 | #include <linux/kasan.h> |
698b1b30 VB |
22 | #include <linux/kthread.h> |
23 | #include <linux/freezer.h> | |
83358ece | 24 | #include <linux/page_owner.h> |
eb414681 | 25 | #include <linux/psi.h> |
748446bb MG |
26 | #include "internal.h" |
27 | ||
010fc29a MK |
28 | #ifdef CONFIG_COMPACTION |
29 | static inline void count_compact_event(enum vm_event_item item) | |
30 | { | |
31 | count_vm_event(item); | |
32 | } | |
33 | ||
34 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
35 | { | |
36 | count_vm_events(item, delta); | |
37 | } | |
38 | #else | |
39 | #define count_compact_event(item) do { } while (0) | |
40 | #define count_compact_events(item, delta) do { } while (0) | |
41 | #endif | |
42 | ||
ff9543fd MN |
43 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
44 | ||
b7aba698 MG |
45 | #define CREATE_TRACE_POINTS |
46 | #include <trace/events/compaction.h> | |
47 | ||
06b6640a VB |
48 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
49 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
50 | #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order) | |
51 | #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order) | |
52 | ||
facdaa91 NG |
53 | /* |
54 | * Fragmentation score check interval for proactive compaction purposes. | |
55 | */ | |
d34c0a75 | 56 | static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500; |
facdaa91 NG |
57 | |
58 | /* | |
59 | * Page order with-respect-to which proactive compaction | |
60 | * calculates external fragmentation, which is used as | |
61 | * the "fragmentation score" of a node/zone. | |
62 | */ | |
63 | #if defined CONFIG_TRANSPARENT_HUGEPAGE | |
64 | #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER | |
25788738 | 65 | #elif defined CONFIG_HUGETLBFS |
facdaa91 NG |
66 | #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER |
67 | #else | |
68 | #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
69 | #endif | |
70 | ||
748446bb MG |
71 | static unsigned long release_freepages(struct list_head *freelist) |
72 | { | |
73 | struct page *page, *next; | |
6bace090 | 74 | unsigned long high_pfn = 0; |
748446bb MG |
75 | |
76 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 77 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
78 | list_del(&page->lru); |
79 | __free_page(page); | |
6bace090 VB |
80 | if (pfn > high_pfn) |
81 | high_pfn = pfn; | |
748446bb MG |
82 | } |
83 | ||
6bace090 | 84 | return high_pfn; |
748446bb MG |
85 | } |
86 | ||
4469ab98 | 87 | static void split_map_pages(struct list_head *list) |
ff9543fd | 88 | { |
66c64223 JK |
89 | unsigned int i, order, nr_pages; |
90 | struct page *page, *next; | |
91 | LIST_HEAD(tmp_list); | |
92 | ||
93 | list_for_each_entry_safe(page, next, list, lru) { | |
94 | list_del(&page->lru); | |
95 | ||
96 | order = page_private(page); | |
97 | nr_pages = 1 << order; | |
66c64223 | 98 | |
46f24fd8 | 99 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
100 | if (order) |
101 | split_page(page, order); | |
ff9543fd | 102 | |
66c64223 JK |
103 | for (i = 0; i < nr_pages; i++) { |
104 | list_add(&page->lru, &tmp_list); | |
105 | page++; | |
106 | } | |
ff9543fd | 107 | } |
66c64223 JK |
108 | |
109 | list_splice(&tmp_list, list); | |
ff9543fd MN |
110 | } |
111 | ||
bb13ffeb | 112 | #ifdef CONFIG_COMPACTION |
24e2716f | 113 | |
bda807d4 MK |
114 | int PageMovable(struct page *page) |
115 | { | |
116 | struct address_space *mapping; | |
117 | ||
118 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
119 | if (!__PageMovable(page)) | |
120 | return 0; | |
121 | ||
122 | mapping = page_mapping(page); | |
123 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
124 | return 1; | |
125 | ||
126 | return 0; | |
127 | } | |
128 | EXPORT_SYMBOL(PageMovable); | |
129 | ||
130 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
131 | { | |
132 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
133 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
134 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
135 | } | |
136 | EXPORT_SYMBOL(__SetPageMovable); | |
137 | ||
138 | void __ClearPageMovable(struct page *page) | |
139 | { | |
bda807d4 MK |
140 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
141 | /* | |
142 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
143 | * flag so that VM can catch up released page by driver after isolation. | |
144 | * With it, VM migration doesn't try to put it back. | |
145 | */ | |
146 | page->mapping = (void *)((unsigned long)page->mapping & | |
147 | PAGE_MAPPING_MOVABLE); | |
148 | } | |
149 | EXPORT_SYMBOL(__ClearPageMovable); | |
150 | ||
24e2716f JK |
151 | /* Do not skip compaction more than 64 times */ |
152 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
153 | ||
154 | /* | |
155 | * Compaction is deferred when compaction fails to result in a page | |
860b3272 | 156 | * allocation success. 1 << compact_defer_shift, compactions are skipped up |
24e2716f JK |
157 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT |
158 | */ | |
2271b016 | 159 | static void defer_compaction(struct zone *zone, int order) |
24e2716f JK |
160 | { |
161 | zone->compact_considered = 0; | |
162 | zone->compact_defer_shift++; | |
163 | ||
164 | if (order < zone->compact_order_failed) | |
165 | zone->compact_order_failed = order; | |
166 | ||
167 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
168 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
169 | ||
170 | trace_mm_compaction_defer_compaction(zone, order); | |
171 | } | |
172 | ||
173 | /* Returns true if compaction should be skipped this time */ | |
2271b016 | 174 | static bool compaction_deferred(struct zone *zone, int order) |
24e2716f JK |
175 | { |
176 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
177 | ||
178 | if (order < zone->compact_order_failed) | |
179 | return false; | |
180 | ||
181 | /* Avoid possible overflow */ | |
62b35fe0 | 182 | if (++zone->compact_considered >= defer_limit) { |
24e2716f | 183 | zone->compact_considered = defer_limit; |
24e2716f | 184 | return false; |
62b35fe0 | 185 | } |
24e2716f JK |
186 | |
187 | trace_mm_compaction_deferred(zone, order); | |
188 | ||
189 | return true; | |
190 | } | |
191 | ||
192 | /* | |
193 | * Update defer tracking counters after successful compaction of given order, | |
194 | * which means an allocation either succeeded (alloc_success == true) or is | |
195 | * expected to succeed. | |
196 | */ | |
197 | void compaction_defer_reset(struct zone *zone, int order, | |
198 | bool alloc_success) | |
199 | { | |
200 | if (alloc_success) { | |
201 | zone->compact_considered = 0; | |
202 | zone->compact_defer_shift = 0; | |
203 | } | |
204 | if (order >= zone->compact_order_failed) | |
205 | zone->compact_order_failed = order + 1; | |
206 | ||
207 | trace_mm_compaction_defer_reset(zone, order); | |
208 | } | |
209 | ||
210 | /* Returns true if restarting compaction after many failures */ | |
2271b016 | 211 | static bool compaction_restarting(struct zone *zone, int order) |
24e2716f JK |
212 | { |
213 | if (order < zone->compact_order_failed) | |
214 | return false; | |
215 | ||
216 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
217 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
218 | } | |
219 | ||
bb13ffeb MG |
220 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
221 | static inline bool isolation_suitable(struct compact_control *cc, | |
222 | struct page *page) | |
223 | { | |
224 | if (cc->ignore_skip_hint) | |
225 | return true; | |
226 | ||
227 | return !get_pageblock_skip(page); | |
228 | } | |
229 | ||
02333641 VB |
230 | static void reset_cached_positions(struct zone *zone) |
231 | { | |
232 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
233 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 234 | zone->compact_cached_free_pfn = |
06b6640a | 235 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
236 | } |
237 | ||
21dc7e02 | 238 | /* |
2271b016 | 239 | * Compound pages of >= pageblock_order should consistently be skipped until |
b527cfe5 VB |
240 | * released. It is always pointless to compact pages of such order (if they are |
241 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 242 | */ |
b527cfe5 | 243 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 244 | { |
b527cfe5 | 245 | if (!PageCompound(page)) |
21dc7e02 | 246 | return false; |
b527cfe5 VB |
247 | |
248 | page = compound_head(page); | |
249 | ||
250 | if (compound_order(page) >= pageblock_order) | |
251 | return true; | |
252 | ||
253 | return false; | |
21dc7e02 DR |
254 | } |
255 | ||
e332f741 MG |
256 | static bool |
257 | __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, | |
258 | bool check_target) | |
259 | { | |
260 | struct page *page = pfn_to_online_page(pfn); | |
6b0868c8 | 261 | struct page *block_page; |
e332f741 MG |
262 | struct page *end_page; |
263 | unsigned long block_pfn; | |
264 | ||
265 | if (!page) | |
266 | return false; | |
267 | if (zone != page_zone(page)) | |
268 | return false; | |
269 | if (pageblock_skip_persistent(page)) | |
270 | return false; | |
271 | ||
272 | /* | |
273 | * If skip is already cleared do no further checking once the | |
274 | * restart points have been set. | |
275 | */ | |
276 | if (check_source && check_target && !get_pageblock_skip(page)) | |
277 | return true; | |
278 | ||
279 | /* | |
280 | * If clearing skip for the target scanner, do not select a | |
281 | * non-movable pageblock as the starting point. | |
282 | */ | |
283 | if (!check_source && check_target && | |
284 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
285 | return false; | |
286 | ||
6b0868c8 MG |
287 | /* Ensure the start of the pageblock or zone is online and valid */ |
288 | block_pfn = pageblock_start_pfn(pfn); | |
a2e9a5af VB |
289 | block_pfn = max(block_pfn, zone->zone_start_pfn); |
290 | block_page = pfn_to_online_page(block_pfn); | |
6b0868c8 MG |
291 | if (block_page) { |
292 | page = block_page; | |
293 | pfn = block_pfn; | |
294 | } | |
295 | ||
296 | /* Ensure the end of the pageblock or zone is online and valid */ | |
a2e9a5af | 297 | block_pfn = pageblock_end_pfn(pfn) - 1; |
6b0868c8 MG |
298 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); |
299 | end_page = pfn_to_online_page(block_pfn); | |
300 | if (!end_page) | |
301 | return false; | |
302 | ||
e332f741 MG |
303 | /* |
304 | * Only clear the hint if a sample indicates there is either a | |
305 | * free page or an LRU page in the block. One or other condition | |
306 | * is necessary for the block to be a migration source/target. | |
307 | */ | |
e332f741 | 308 | do { |
859a85dd MR |
309 | if (check_source && PageLRU(page)) { |
310 | clear_pageblock_skip(page); | |
311 | return true; | |
312 | } | |
e332f741 | 313 | |
859a85dd MR |
314 | if (check_target && PageBuddy(page)) { |
315 | clear_pageblock_skip(page); | |
316 | return true; | |
e332f741 MG |
317 | } |
318 | ||
319 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
320 | pfn += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
a2e9a5af | 321 | } while (page <= end_page); |
e332f741 MG |
322 | |
323 | return false; | |
324 | } | |
325 | ||
bb13ffeb MG |
326 | /* |
327 | * This function is called to clear all cached information on pageblocks that | |
328 | * should be skipped for page isolation when the migrate and free page scanner | |
329 | * meet. | |
330 | */ | |
62997027 | 331 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 332 | { |
e332f741 | 333 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 334 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
335 | unsigned long reset_migrate = free_pfn; |
336 | unsigned long reset_free = migrate_pfn; | |
337 | bool source_set = false; | |
338 | bool free_set = false; | |
339 | ||
340 | if (!zone->compact_blockskip_flush) | |
341 | return; | |
bb13ffeb | 342 | |
62997027 | 343 | zone->compact_blockskip_flush = false; |
bb13ffeb | 344 | |
e332f741 MG |
345 | /* |
346 | * Walk the zone and update pageblock skip information. Source looks | |
347 | * for PageLRU while target looks for PageBuddy. When the scanner | |
348 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
349 | * is suitable as both source and target. | |
350 | */ | |
351 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
352 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
353 | cond_resched(); |
354 | ||
e332f741 MG |
355 | /* Update the migrate PFN */ |
356 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
357 | migrate_pfn < reset_migrate) { | |
358 | source_set = true; | |
359 | reset_migrate = migrate_pfn; | |
360 | zone->compact_init_migrate_pfn = reset_migrate; | |
361 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
362 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
363 | } | |
bb13ffeb | 364 | |
e332f741 MG |
365 | /* Update the free PFN */ |
366 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
367 | free_pfn > reset_free) { | |
368 | free_set = true; | |
369 | reset_free = free_pfn; | |
370 | zone->compact_init_free_pfn = reset_free; | |
371 | zone->compact_cached_free_pfn = reset_free; | |
372 | } | |
bb13ffeb | 373 | } |
02333641 | 374 | |
e332f741 MG |
375 | /* Leave no distance if no suitable block was reset */ |
376 | if (reset_migrate >= reset_free) { | |
377 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
378 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
379 | zone->compact_cached_free_pfn = free_pfn; | |
380 | } | |
bb13ffeb MG |
381 | } |
382 | ||
62997027 MG |
383 | void reset_isolation_suitable(pg_data_t *pgdat) |
384 | { | |
385 | int zoneid; | |
386 | ||
387 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
388 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
389 | if (!populated_zone(zone)) | |
390 | continue; | |
391 | ||
392 | /* Only flush if a full compaction finished recently */ | |
393 | if (zone->compact_blockskip_flush) | |
394 | __reset_isolation_suitable(zone); | |
395 | } | |
396 | } | |
397 | ||
e380bebe MG |
398 | /* |
399 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
400 | * locks are not required for read/writers. Returns true if it was already set. | |
401 | */ | |
402 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
403 | unsigned long pfn) | |
404 | { | |
405 | bool skip; | |
406 | ||
407 | /* Do no update if skip hint is being ignored */ | |
408 | if (cc->ignore_skip_hint) | |
409 | return false; | |
410 | ||
411 | if (!IS_ALIGNED(pfn, pageblock_nr_pages)) | |
412 | return false; | |
413 | ||
414 | skip = get_pageblock_skip(page); | |
415 | if (!skip && !cc->no_set_skip_hint) | |
416 | set_pageblock_skip(page); | |
417 | ||
418 | return skip; | |
419 | } | |
420 | ||
421 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
422 | { | |
423 | struct zone *zone = cc->zone; | |
424 | ||
425 | pfn = pageblock_end_pfn(pfn); | |
426 | ||
427 | /* Set for isolation rather than compaction */ | |
428 | if (cc->no_set_skip_hint) | |
429 | return; | |
430 | ||
431 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
432 | zone->compact_cached_migrate_pfn[0] = pfn; | |
433 | if (cc->mode != MIGRATE_ASYNC && | |
434 | pfn > zone->compact_cached_migrate_pfn[1]) | |
435 | zone->compact_cached_migrate_pfn[1] = pfn; | |
436 | } | |
437 | ||
bb13ffeb MG |
438 | /* |
439 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 440 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 441 | */ |
c89511ab | 442 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 443 | struct page *page, unsigned long pfn) |
bb13ffeb | 444 | { |
c89511ab | 445 | struct zone *zone = cc->zone; |
6815bf3f | 446 | |
2583d671 | 447 | if (cc->no_set_skip_hint) |
6815bf3f JK |
448 | return; |
449 | ||
bb13ffeb MG |
450 | if (!page) |
451 | return; | |
452 | ||
edc2ca61 | 453 | set_pageblock_skip(page); |
c89511ab | 454 | |
35979ef3 | 455 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
456 | if (pfn < zone->compact_cached_free_pfn) |
457 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
458 | } |
459 | #else | |
460 | static inline bool isolation_suitable(struct compact_control *cc, | |
461 | struct page *page) | |
462 | { | |
463 | return true; | |
464 | } | |
465 | ||
b527cfe5 | 466 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
467 | { |
468 | return false; | |
469 | } | |
470 | ||
471 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 472 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
473 | { |
474 | } | |
e380bebe MG |
475 | |
476 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
477 | { | |
478 | } | |
479 | ||
480 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
481 | unsigned long pfn) | |
482 | { | |
483 | return false; | |
484 | } | |
bb13ffeb MG |
485 | #endif /* CONFIG_COMPACTION */ |
486 | ||
8b44d279 VB |
487 | /* |
488 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
489 | * very heavily contended. For async compaction, trylock and record if the |
490 | * lock is contended. The lock will still be acquired but compaction will | |
491 | * abort when the current block is finished regardless of success rate. | |
492 | * Sync compaction acquires the lock. | |
8b44d279 | 493 | * |
cb2dcaf0 | 494 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 495 | */ |
cb2dcaf0 | 496 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 497 | struct compact_control *cc) |
77337ede | 498 | __acquires(lock) |
2a1402aa | 499 | { |
cb2dcaf0 MG |
500 | /* Track if the lock is contended in async mode */ |
501 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
502 | if (spin_trylock_irqsave(lock, *flags)) | |
503 | return true; | |
504 | ||
505 | cc->contended = true; | |
8b44d279 | 506 | } |
1f9efdef | 507 | |
cb2dcaf0 | 508 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 509 | return true; |
2a1402aa MG |
510 | } |
511 | ||
c67fe375 MG |
512 | /* |
513 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
514 | * very heavily contended. The lock should be periodically unlocked to avoid |
515 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
516 | * the lock. It might also be that allowing the IRQs will result in | |
517 | * need_resched() becoming true. If scheduling is needed, async compaction | |
518 | * aborts. Sync compaction schedules. | |
519 | * Either compaction type will also abort if a fatal signal is pending. | |
520 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 521 | * |
8b44d279 VB |
522 | * Returns true if compaction should abort due to fatal signal pending, or |
523 | * async compaction due to need_resched() | |
524 | * Returns false when compaction can continue (sync compaction might have | |
525 | * scheduled) | |
c67fe375 | 526 | */ |
8b44d279 VB |
527 | static bool compact_unlock_should_abort(spinlock_t *lock, |
528 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 529 | { |
8b44d279 VB |
530 | if (*locked) { |
531 | spin_unlock_irqrestore(lock, flags); | |
532 | *locked = false; | |
533 | } | |
1f9efdef | 534 | |
8b44d279 | 535 | if (fatal_signal_pending(current)) { |
c3486f53 | 536 | cc->contended = true; |
8b44d279 VB |
537 | return true; |
538 | } | |
c67fe375 | 539 | |
cf66f070 | 540 | cond_resched(); |
be976572 VB |
541 | |
542 | return false; | |
543 | } | |
544 | ||
85aa125f | 545 | /* |
9e4be470 JM |
546 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
547 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
548 | * (even though it may still end up isolating some pages). | |
85aa125f | 549 | */ |
f40d1e42 | 550 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 551 | unsigned long *start_pfn, |
85aa125f MN |
552 | unsigned long end_pfn, |
553 | struct list_head *freelist, | |
4fca9730 | 554 | unsigned int stride, |
85aa125f | 555 | bool strict) |
748446bb | 556 | { |
b7aba698 | 557 | int nr_scanned = 0, total_isolated = 0; |
d097a6f6 | 558 | struct page *cursor; |
b8b2d825 | 559 | unsigned long flags = 0; |
f40d1e42 | 560 | bool locked = false; |
e14c720e | 561 | unsigned long blockpfn = *start_pfn; |
66c64223 | 562 | unsigned int order; |
748446bb | 563 | |
4fca9730 MG |
564 | /* Strict mode is for isolation, speed is secondary */ |
565 | if (strict) | |
566 | stride = 1; | |
567 | ||
748446bb MG |
568 | cursor = pfn_to_page(blockpfn); |
569 | ||
f40d1e42 | 570 | /* Isolate free pages. */ |
4fca9730 | 571 | for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { |
66c64223 | 572 | int isolated; |
748446bb MG |
573 | struct page *page = cursor; |
574 | ||
8b44d279 VB |
575 | /* |
576 | * Periodically drop the lock (if held) regardless of its | |
577 | * contention, to give chance to IRQs. Abort if fatal signal | |
578 | * pending or async compaction detects need_resched() | |
579 | */ | |
580 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
581 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
582 | &locked, cc)) | |
583 | break; | |
584 | ||
b7aba698 | 585 | nr_scanned++; |
2af120bc | 586 | |
9fcd6d2e VB |
587 | /* |
588 | * For compound pages such as THP and hugetlbfs, we can save | |
589 | * potentially a lot of iterations if we skip them at once. | |
590 | * The check is racy, but we can consider only valid values | |
591 | * and the only danger is skipping too much. | |
592 | */ | |
593 | if (PageCompound(page)) { | |
21dc7e02 DR |
594 | const unsigned int order = compound_order(page); |
595 | ||
d3c85bad | 596 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
597 | blockpfn += (1UL << order) - 1; |
598 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 599 | } |
9fcd6d2e VB |
600 | goto isolate_fail; |
601 | } | |
602 | ||
f40d1e42 | 603 | if (!PageBuddy(page)) |
2af120bc | 604 | goto isolate_fail; |
f40d1e42 MG |
605 | |
606 | /* | |
69b7189f VB |
607 | * If we already hold the lock, we can skip some rechecking. |
608 | * Note that if we hold the lock now, checked_pageblock was | |
609 | * already set in some previous iteration (or strict is true), | |
610 | * so it is correct to skip the suitable migration target | |
611 | * recheck as well. | |
f40d1e42 | 612 | */ |
69b7189f | 613 | if (!locked) { |
cb2dcaf0 | 614 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 615 | &flags, cc); |
f40d1e42 | 616 | |
69b7189f VB |
617 | /* Recheck this is a buddy page under lock */ |
618 | if (!PageBuddy(page)) | |
619 | goto isolate_fail; | |
620 | } | |
748446bb | 621 | |
66c64223 | 622 | /* Found a free page, will break it into order-0 pages */ |
ab130f91 | 623 | order = buddy_order(page); |
66c64223 | 624 | isolated = __isolate_free_page(page, order); |
a4f04f2c DR |
625 | if (!isolated) |
626 | break; | |
66c64223 | 627 | set_page_private(page, order); |
a4f04f2c | 628 | |
748446bb | 629 | total_isolated += isolated; |
a4f04f2c | 630 | cc->nr_freepages += isolated; |
66c64223 JK |
631 | list_add_tail(&page->lru, freelist); |
632 | ||
a4f04f2c DR |
633 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
634 | blockpfn += isolated; | |
635 | break; | |
748446bb | 636 | } |
a4f04f2c DR |
637 | /* Advance to the end of split page */ |
638 | blockpfn += isolated - 1; | |
639 | cursor += isolated - 1; | |
640 | continue; | |
2af120bc LA |
641 | |
642 | isolate_fail: | |
643 | if (strict) | |
644 | break; | |
645 | else | |
646 | continue; | |
647 | ||
748446bb MG |
648 | } |
649 | ||
a4f04f2c DR |
650 | if (locked) |
651 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
652 | ||
9fcd6d2e VB |
653 | /* |
654 | * There is a tiny chance that we have read bogus compound_order(), | |
655 | * so be careful to not go outside of the pageblock. | |
656 | */ | |
657 | if (unlikely(blockpfn > end_pfn)) | |
658 | blockpfn = end_pfn; | |
659 | ||
e34d85f0 JK |
660 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
661 | nr_scanned, total_isolated); | |
662 | ||
e14c720e VB |
663 | /* Record how far we have got within the block */ |
664 | *start_pfn = blockpfn; | |
665 | ||
f40d1e42 MG |
666 | /* |
667 | * If strict isolation is requested by CMA then check that all the | |
668 | * pages requested were isolated. If there were any failures, 0 is | |
669 | * returned and CMA will fail. | |
670 | */ | |
2af120bc | 671 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
672 | total_isolated = 0; |
673 | ||
7f354a54 | 674 | cc->total_free_scanned += nr_scanned; |
397487db | 675 | if (total_isolated) |
010fc29a | 676 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
677 | return total_isolated; |
678 | } | |
679 | ||
85aa125f MN |
680 | /** |
681 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 682 | * @cc: Compaction control structure. |
85aa125f MN |
683 | * @start_pfn: The first PFN to start isolating. |
684 | * @end_pfn: The one-past-last PFN. | |
685 | * | |
686 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
687 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
688 | * undo its actions and return zero. | |
689 | * | |
690 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
691 | * (which may be greater then end_pfn if end fell in a middle of | |
692 | * a free page). | |
693 | */ | |
ff9543fd | 694 | unsigned long |
bb13ffeb MG |
695 | isolate_freepages_range(struct compact_control *cc, |
696 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 697 | { |
e1409c32 | 698 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
699 | LIST_HEAD(freelist); |
700 | ||
7d49d886 | 701 | pfn = start_pfn; |
06b6640a | 702 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
703 | if (block_start_pfn < cc->zone->zone_start_pfn) |
704 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 705 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
706 | |
707 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 708 | block_start_pfn = block_end_pfn, |
7d49d886 | 709 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
710 | /* Protect pfn from changing by isolate_freepages_block */ |
711 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 712 | |
85aa125f MN |
713 | block_end_pfn = min(block_end_pfn, end_pfn); |
714 | ||
58420016 JK |
715 | /* |
716 | * pfn could pass the block_end_pfn if isolated freepage | |
717 | * is more than pageblock order. In this case, we adjust | |
718 | * scanning range to right one. | |
719 | */ | |
720 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
721 | block_start_pfn = pageblock_start_pfn(pfn); |
722 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
723 | block_end_pfn = min(block_end_pfn, end_pfn); |
724 | } | |
725 | ||
e1409c32 JK |
726 | if (!pageblock_pfn_to_page(block_start_pfn, |
727 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
728 | break; |
729 | ||
e14c720e | 730 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
4fca9730 | 731 | block_end_pfn, &freelist, 0, true); |
85aa125f MN |
732 | |
733 | /* | |
734 | * In strict mode, isolate_freepages_block() returns 0 if | |
735 | * there are any holes in the block (ie. invalid PFNs or | |
736 | * non-free pages). | |
737 | */ | |
738 | if (!isolated) | |
739 | break; | |
740 | ||
741 | /* | |
742 | * If we managed to isolate pages, it is always (1 << n) * | |
743 | * pageblock_nr_pages for some non-negative n. (Max order | |
744 | * page may span two pageblocks). | |
745 | */ | |
746 | } | |
747 | ||
66c64223 | 748 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 749 | split_map_pages(&freelist); |
85aa125f MN |
750 | |
751 | if (pfn < end_pfn) { | |
752 | /* Loop terminated early, cleanup. */ | |
753 | release_freepages(&freelist); | |
754 | return 0; | |
755 | } | |
756 | ||
757 | /* We don't use freelists for anything. */ | |
758 | return pfn; | |
759 | } | |
760 | ||
748446bb | 761 | /* Similar to reclaim, but different enough that they don't share logic */ |
5f438eee | 762 | static bool too_many_isolated(pg_data_t *pgdat) |
748446bb | 763 | { |
d818fca1 MG |
764 | bool too_many; |
765 | ||
bc693045 | 766 | unsigned long active, inactive, isolated; |
748446bb | 767 | |
5f438eee AR |
768 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + |
769 | node_page_state(pgdat, NR_INACTIVE_ANON); | |
770 | active = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
771 | node_page_state(pgdat, NR_ACTIVE_ANON); | |
772 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + | |
773 | node_page_state(pgdat, NR_ISOLATED_ANON); | |
748446bb | 774 | |
d818fca1 MG |
775 | too_many = isolated > (inactive + active) / 2; |
776 | if (!too_many) | |
777 | wake_throttle_isolated(pgdat); | |
778 | ||
779 | return too_many; | |
748446bb MG |
780 | } |
781 | ||
2fe86e00 | 782 | /** |
edc2ca61 VB |
783 | * isolate_migratepages_block() - isolate all migrate-able pages within |
784 | * a single pageblock | |
2fe86e00 | 785 | * @cc: Compaction control structure. |
edc2ca61 VB |
786 | * @low_pfn: The first PFN to isolate |
787 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
89f6c88a | 788 | * @mode: Isolation mode to be used. |
2fe86e00 MN |
789 | * |
790 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 | 791 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
c2ad7a1f | 792 | * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, |
369fa227 | 793 | * -ENOMEM in case we could not allocate a page, or 0. |
c2ad7a1f | 794 | * cc->migrate_pfn will contain the next pfn to scan. |
2fe86e00 | 795 | * |
edc2ca61 | 796 | * The pages are isolated on cc->migratepages list (not required to be empty), |
c2ad7a1f | 797 | * and cc->nr_migratepages is updated accordingly. |
748446bb | 798 | */ |
c2ad7a1f | 799 | static int |
edc2ca61 | 800 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, |
89f6c88a | 801 | unsigned long end_pfn, isolate_mode_t mode) |
748446bb | 802 | { |
5f438eee | 803 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 804 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 805 | struct lruvec *lruvec; |
b8b2d825 | 806 | unsigned long flags = 0; |
6168d0da | 807 | struct lruvec *locked = NULL; |
bb13ffeb | 808 | struct page *page = NULL, *valid_page = NULL; |
89f6c88a | 809 | struct address_space *mapping; |
e34d85f0 | 810 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
811 | bool skip_on_failure = false; |
812 | unsigned long next_skip_pfn = 0; | |
e380bebe | 813 | bool skip_updated = false; |
c2ad7a1f OS |
814 | int ret = 0; |
815 | ||
816 | cc->migrate_pfn = low_pfn; | |
748446bb | 817 | |
748446bb MG |
818 | /* |
819 | * Ensure that there are not too many pages isolated from the LRU | |
820 | * list by either parallel reclaimers or compaction. If there are, | |
821 | * delay for some time until fewer pages are isolated | |
822 | */ | |
5f438eee | 823 | while (unlikely(too_many_isolated(pgdat))) { |
d20bdd57 ZY |
824 | /* stop isolation if there are still pages not migrated */ |
825 | if (cc->nr_migratepages) | |
c2ad7a1f | 826 | return -EAGAIN; |
d20bdd57 | 827 | |
f9e35b3b | 828 | /* async migration should just abort */ |
e0b9daeb | 829 | if (cc->mode == MIGRATE_ASYNC) |
c2ad7a1f | 830 | return -EAGAIN; |
f9e35b3b | 831 | |
c3f4a9a2 | 832 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
748446bb MG |
833 | |
834 | if (fatal_signal_pending(current)) | |
c2ad7a1f | 835 | return -EINTR; |
748446bb MG |
836 | } |
837 | ||
cf66f070 | 838 | cond_resched(); |
aeef4b83 | 839 | |
fdd048e1 VB |
840 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
841 | skip_on_failure = true; | |
842 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
843 | } | |
844 | ||
748446bb | 845 | /* Time to isolate some pages for migration */ |
748446bb | 846 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 847 | |
fdd048e1 VB |
848 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
849 | /* | |
850 | * We have isolated all migration candidates in the | |
851 | * previous order-aligned block, and did not skip it due | |
852 | * to failure. We should migrate the pages now and | |
853 | * hopefully succeed compaction. | |
854 | */ | |
855 | if (nr_isolated) | |
856 | break; | |
857 | ||
858 | /* | |
859 | * We failed to isolate in the previous order-aligned | |
860 | * block. Set the new boundary to the end of the | |
861 | * current block. Note we can't simply increase | |
862 | * next_skip_pfn by 1 << order, as low_pfn might have | |
863 | * been incremented by a higher number due to skipping | |
864 | * a compound or a high-order buddy page in the | |
865 | * previous loop iteration. | |
866 | */ | |
867 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
868 | } | |
869 | ||
8b44d279 VB |
870 | /* |
871 | * Periodically drop the lock (if held) regardless of its | |
670105a2 MG |
872 | * contention, to give chance to IRQs. Abort completely if |
873 | * a fatal signal is pending. | |
8b44d279 | 874 | */ |
6168d0da AS |
875 | if (!(low_pfn % SWAP_CLUSTER_MAX)) { |
876 | if (locked) { | |
877 | unlock_page_lruvec_irqrestore(locked, flags); | |
878 | locked = NULL; | |
879 | } | |
880 | ||
881 | if (fatal_signal_pending(current)) { | |
882 | cc->contended = true; | |
c2ad7a1f | 883 | ret = -EINTR; |
6168d0da | 884 | |
6168d0da AS |
885 | goto fatal_pending; |
886 | } | |
887 | ||
888 | cond_resched(); | |
670105a2 | 889 | } |
c67fe375 | 890 | |
b7aba698 | 891 | nr_scanned++; |
748446bb | 892 | |
748446bb | 893 | page = pfn_to_page(low_pfn); |
dc908600 | 894 | |
e380bebe MG |
895 | /* |
896 | * Check if the pageblock has already been marked skipped. | |
897 | * Only the aligned PFN is checked as the caller isolates | |
898 | * COMPACT_CLUSTER_MAX at a time so the second call must | |
899 | * not falsely conclude that the block should be skipped. | |
900 | */ | |
901 | if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) { | |
902 | if (!cc->ignore_skip_hint && get_pageblock_skip(page)) { | |
903 | low_pfn = end_pfn; | |
9df41314 | 904 | page = NULL; |
e380bebe MG |
905 | goto isolate_abort; |
906 | } | |
bb13ffeb | 907 | valid_page = page; |
e380bebe | 908 | } |
bb13ffeb | 909 | |
369fa227 | 910 | if (PageHuge(page) && cc->alloc_contig) { |
ae37c7ff | 911 | ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); |
369fa227 OS |
912 | |
913 | /* | |
914 | * Fail isolation in case isolate_or_dissolve_huge_page() | |
915 | * reports an error. In case of -ENOMEM, abort right away. | |
916 | */ | |
917 | if (ret < 0) { | |
918 | /* Do not report -EBUSY down the chain */ | |
919 | if (ret == -EBUSY) | |
920 | ret = 0; | |
921 | low_pfn += (1UL << compound_order(page)) - 1; | |
922 | goto isolate_fail; | |
923 | } | |
924 | ||
ae37c7ff OS |
925 | if (PageHuge(page)) { |
926 | /* | |
927 | * Hugepage was successfully isolated and placed | |
928 | * on the cc->migratepages list. | |
929 | */ | |
930 | low_pfn += compound_nr(page) - 1; | |
931 | goto isolate_success_no_list; | |
932 | } | |
933 | ||
369fa227 OS |
934 | /* |
935 | * Ok, the hugepage was dissolved. Now these pages are | |
936 | * Buddy and cannot be re-allocated because they are | |
937 | * isolated. Fall-through as the check below handles | |
938 | * Buddy pages. | |
939 | */ | |
940 | } | |
941 | ||
6c14466c | 942 | /* |
99c0fd5e VB |
943 | * Skip if free. We read page order here without zone lock |
944 | * which is generally unsafe, but the race window is small and | |
945 | * the worst thing that can happen is that we skip some | |
946 | * potential isolation targets. | |
6c14466c | 947 | */ |
99c0fd5e | 948 | if (PageBuddy(page)) { |
ab130f91 | 949 | unsigned long freepage_order = buddy_order_unsafe(page); |
99c0fd5e VB |
950 | |
951 | /* | |
952 | * Without lock, we cannot be sure that what we got is | |
953 | * a valid page order. Consider only values in the | |
954 | * valid order range to prevent low_pfn overflow. | |
955 | */ | |
956 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
957 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 958 | continue; |
99c0fd5e | 959 | } |
748446bb | 960 | |
bc835011 | 961 | /* |
29c0dde8 | 962 | * Regardless of being on LRU, compound pages such as THP and |
1da2f328 RR |
963 | * hugetlbfs are not to be compacted unless we are attempting |
964 | * an allocation much larger than the huge page size (eg CMA). | |
965 | * We can potentially save a lot of iterations if we skip them | |
966 | * at once. The check is racy, but we can consider only valid | |
967 | * values and the only danger is skipping too much. | |
bc835011 | 968 | */ |
1da2f328 | 969 | if (PageCompound(page) && !cc->alloc_contig) { |
21dc7e02 | 970 | const unsigned int order = compound_order(page); |
edc2ca61 | 971 | |
d3c85bad | 972 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 973 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 974 | goto isolate_fail; |
2a1402aa MG |
975 | } |
976 | ||
bda807d4 MK |
977 | /* |
978 | * Check may be lockless but that's ok as we recheck later. | |
979 | * It's possible to migrate LRU and non-lru movable pages. | |
980 | * Skip any other type of page | |
981 | */ | |
982 | if (!PageLRU(page)) { | |
bda807d4 MK |
983 | /* |
984 | * __PageMovable can return false positive so we need | |
985 | * to verify it under page_lock. | |
986 | */ | |
987 | if (unlikely(__PageMovable(page)) && | |
988 | !PageIsolated(page)) { | |
989 | if (locked) { | |
6168d0da AS |
990 | unlock_page_lruvec_irqrestore(locked, flags); |
991 | locked = NULL; | |
bda807d4 MK |
992 | } |
993 | ||
89f6c88a | 994 | if (!isolate_movable_page(page, mode)) |
bda807d4 MK |
995 | goto isolate_success; |
996 | } | |
997 | ||
fdd048e1 | 998 | goto isolate_fail; |
bda807d4 | 999 | } |
29c0dde8 | 1000 | |
119d6d59 DR |
1001 | /* |
1002 | * Migration will fail if an anonymous page is pinned in memory, | |
1003 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
1004 | * admittedly racy check. | |
1005 | */ | |
89f6c88a HD |
1006 | mapping = page_mapping(page); |
1007 | if (!mapping && page_count(page) > page_mapcount(page)) | |
fdd048e1 | 1008 | goto isolate_fail; |
119d6d59 | 1009 | |
73e64c51 MH |
1010 | /* |
1011 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
1012 | * because those do not depend on fs locks. | |
1013 | */ | |
89f6c88a | 1014 | if (!(cc->gfp_mask & __GFP_FS) && mapping) |
73e64c51 MH |
1015 | goto isolate_fail; |
1016 | ||
9df41314 AS |
1017 | /* |
1018 | * Be careful not to clear PageLRU until after we're | |
1019 | * sure the page is not being freed elsewhere -- the | |
1020 | * page release code relies on it. | |
1021 | */ | |
1022 | if (unlikely(!get_page_unless_zero(page))) | |
1023 | goto isolate_fail; | |
1024 | ||
89f6c88a HD |
1025 | /* Only take pages on LRU: a check now makes later tests safe */ |
1026 | if (!PageLRU(page)) | |
1027 | goto isolate_fail_put; | |
1028 | ||
1029 | /* Compaction might skip unevictable pages but CMA takes them */ | |
1030 | if (!(mode & ISOLATE_UNEVICTABLE) && PageUnevictable(page)) | |
1031 | goto isolate_fail_put; | |
1032 | ||
1033 | /* | |
1034 | * To minimise LRU disruption, the caller can indicate with | |
1035 | * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages | |
1036 | * it will be able to migrate without blocking - clean pages | |
1037 | * for the most part. PageWriteback would require blocking. | |
1038 | */ | |
1039 | if ((mode & ISOLATE_ASYNC_MIGRATE) && PageWriteback(page)) | |
9df41314 AS |
1040 | goto isolate_fail_put; |
1041 | ||
89f6c88a HD |
1042 | if ((mode & ISOLATE_ASYNC_MIGRATE) && PageDirty(page)) { |
1043 | bool migrate_dirty; | |
1044 | ||
1045 | /* | |
1046 | * Only pages without mappings or that have a | |
1047 | * ->migratepage callback are possible to migrate | |
1048 | * without blocking. However, we can be racing with | |
1049 | * truncation so it's necessary to lock the page | |
1050 | * to stabilise the mapping as truncation holds | |
1051 | * the page lock until after the page is removed | |
1052 | * from the page cache. | |
1053 | */ | |
1054 | if (!trylock_page(page)) | |
1055 | goto isolate_fail_put; | |
1056 | ||
1057 | mapping = page_mapping(page); | |
1058 | migrate_dirty = !mapping || mapping->a_ops->migratepage; | |
1059 | unlock_page(page); | |
1060 | if (!migrate_dirty) | |
1061 | goto isolate_fail_put; | |
1062 | } | |
1063 | ||
9df41314 AS |
1064 | /* Try isolate the page */ |
1065 | if (!TestClearPageLRU(page)) | |
1066 | goto isolate_fail_put; | |
1067 | ||
b1baabd9 | 1068 | lruvec = folio_lruvec(page_folio(page)); |
6168d0da | 1069 | |
69b7189f | 1070 | /* If we already hold the lock, we can skip some rechecking */ |
6168d0da AS |
1071 | if (lruvec != locked) { |
1072 | if (locked) | |
1073 | unlock_page_lruvec_irqrestore(locked, flags); | |
1074 | ||
1075 | compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); | |
1076 | locked = lruvec; | |
6168d0da | 1077 | |
e809c3fe | 1078 | lruvec_memcg_debug(lruvec, page_folio(page)); |
e380bebe | 1079 | |
e380bebe MG |
1080 | /* Try get exclusive access under lock */ |
1081 | if (!skip_updated) { | |
1082 | skip_updated = true; | |
1083 | if (test_and_set_skip(cc, page, low_pfn)) | |
1084 | goto isolate_abort; | |
1085 | } | |
2a1402aa | 1086 | |
29c0dde8 VB |
1087 | /* |
1088 | * Page become compound since the non-locked check, | |
1089 | * and it's on LRU. It can only be a THP so the order | |
1090 | * is safe to read and it's 0 for tail pages. | |
1091 | */ | |
1da2f328 | 1092 | if (unlikely(PageCompound(page) && !cc->alloc_contig)) { |
d8c6546b | 1093 | low_pfn += compound_nr(page) - 1; |
9df41314 AS |
1094 | SetPageLRU(page); |
1095 | goto isolate_fail_put; | |
69b7189f | 1096 | } |
d99fd5fe | 1097 | } |
fa9add64 | 1098 | |
1da2f328 RR |
1099 | /* The whole page is taken off the LRU; skip the tail pages. */ |
1100 | if (PageCompound(page)) | |
1101 | low_pfn += compound_nr(page) - 1; | |
bc835011 | 1102 | |
748446bb | 1103 | /* Successfully isolated */ |
46ae6b2c | 1104 | del_page_from_lru_list(page, lruvec); |
1da2f328 | 1105 | mod_node_page_state(page_pgdat(page), |
9de4f22a | 1106 | NR_ISOLATED_ANON + page_is_file_lru(page), |
6c357848 | 1107 | thp_nr_pages(page)); |
b6c75016 JK |
1108 | |
1109 | isolate_success: | |
fdd048e1 | 1110 | list_add(&page->lru, &cc->migratepages); |
ae37c7ff | 1111 | isolate_success_no_list: |
38935861 ZY |
1112 | cc->nr_migratepages += compound_nr(page); |
1113 | nr_isolated += compound_nr(page); | |
748446bb | 1114 | |
804d3121 MG |
1115 | /* |
1116 | * Avoid isolating too much unless this block is being | |
cb2dcaf0 MG |
1117 | * rescanned (e.g. dirty/writeback pages, parallel allocation) |
1118 | * or a lock is contended. For contention, isolate quickly to | |
1119 | * potentially remove one source of contention. | |
804d3121 | 1120 | */ |
38935861 | 1121 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && |
cb2dcaf0 | 1122 | !cc->rescan && !cc->contended) { |
31b8384a | 1123 | ++low_pfn; |
748446bb | 1124 | break; |
31b8384a | 1125 | } |
fdd048e1 VB |
1126 | |
1127 | continue; | |
9df41314 AS |
1128 | |
1129 | isolate_fail_put: | |
1130 | /* Avoid potential deadlock in freeing page under lru_lock */ | |
1131 | if (locked) { | |
6168d0da AS |
1132 | unlock_page_lruvec_irqrestore(locked, flags); |
1133 | locked = NULL; | |
9df41314 AS |
1134 | } |
1135 | put_page(page); | |
1136 | ||
fdd048e1 | 1137 | isolate_fail: |
369fa227 | 1138 | if (!skip_on_failure && ret != -ENOMEM) |
fdd048e1 VB |
1139 | continue; |
1140 | ||
1141 | /* | |
1142 | * We have isolated some pages, but then failed. Release them | |
1143 | * instead of migrating, as we cannot form the cc->order buddy | |
1144 | * page anyway. | |
1145 | */ | |
1146 | if (nr_isolated) { | |
1147 | if (locked) { | |
6168d0da AS |
1148 | unlock_page_lruvec_irqrestore(locked, flags); |
1149 | locked = NULL; | |
fdd048e1 | 1150 | } |
fdd048e1 VB |
1151 | putback_movable_pages(&cc->migratepages); |
1152 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1153 | nr_isolated = 0; |
1154 | } | |
1155 | ||
1156 | if (low_pfn < next_skip_pfn) { | |
1157 | low_pfn = next_skip_pfn - 1; | |
1158 | /* | |
1159 | * The check near the loop beginning would have updated | |
1160 | * next_skip_pfn too, but this is a bit simpler. | |
1161 | */ | |
1162 | next_skip_pfn += 1UL << cc->order; | |
1163 | } | |
369fa227 OS |
1164 | |
1165 | if (ret == -ENOMEM) | |
1166 | break; | |
748446bb MG |
1167 | } |
1168 | ||
99c0fd5e VB |
1169 | /* |
1170 | * The PageBuddy() check could have potentially brought us outside | |
1171 | * the range to be scanned. | |
1172 | */ | |
1173 | if (unlikely(low_pfn > end_pfn)) | |
1174 | low_pfn = end_pfn; | |
1175 | ||
9df41314 AS |
1176 | page = NULL; |
1177 | ||
e380bebe | 1178 | isolate_abort: |
c67fe375 | 1179 | if (locked) |
6168d0da | 1180 | unlock_page_lruvec_irqrestore(locked, flags); |
9df41314 AS |
1181 | if (page) { |
1182 | SetPageLRU(page); | |
1183 | put_page(page); | |
1184 | } | |
748446bb | 1185 | |
50b5b094 | 1186 | /* |
804d3121 MG |
1187 | * Updated the cached scanner pfn once the pageblock has been scanned |
1188 | * Pages will either be migrated in which case there is no point | |
1189 | * scanning in the near future or migration failed in which case the | |
1190 | * failure reason may persist. The block is marked for skipping if | |
1191 | * there were no pages isolated in the block or if the block is | |
1192 | * rescanned twice in a row. | |
50b5b094 | 1193 | */ |
804d3121 | 1194 | if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) { |
e380bebe MG |
1195 | if (valid_page && !skip_updated) |
1196 | set_pageblock_skip(valid_page); | |
1197 | update_cached_migrate(cc, low_pfn); | |
1198 | } | |
bb13ffeb | 1199 | |
e34d85f0 JK |
1200 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1201 | nr_scanned, nr_isolated); | |
b7aba698 | 1202 | |
670105a2 | 1203 | fatal_pending: |
7f354a54 | 1204 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1205 | if (nr_isolated) |
010fc29a | 1206 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1207 | |
c2ad7a1f OS |
1208 | cc->migrate_pfn = low_pfn; |
1209 | ||
1210 | return ret; | |
2fe86e00 MN |
1211 | } |
1212 | ||
edc2ca61 VB |
1213 | /** |
1214 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1215 | * @cc: Compaction control structure. | |
1216 | * @start_pfn: The first PFN to start isolating. | |
1217 | * @end_pfn: The one-past-last PFN. | |
1218 | * | |
369fa227 OS |
1219 | * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM |
1220 | * in case we could not allocate a page, or 0. | |
edc2ca61 | 1221 | */ |
c2ad7a1f | 1222 | int |
edc2ca61 VB |
1223 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, |
1224 | unsigned long end_pfn) | |
1225 | { | |
e1409c32 | 1226 | unsigned long pfn, block_start_pfn, block_end_pfn; |
c2ad7a1f | 1227 | int ret = 0; |
edc2ca61 VB |
1228 | |
1229 | /* Scan block by block. First and last block may be incomplete */ | |
1230 | pfn = start_pfn; | |
06b6640a | 1231 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1232 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1233 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1234 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1235 | |
1236 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1237 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1238 | block_end_pfn += pageblock_nr_pages) { |
1239 | ||
1240 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1241 | ||
e1409c32 JK |
1242 | if (!pageblock_pfn_to_page(block_start_pfn, |
1243 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1244 | continue; |
1245 | ||
c2ad7a1f OS |
1246 | ret = isolate_migratepages_block(cc, pfn, block_end_pfn, |
1247 | ISOLATE_UNEVICTABLE); | |
edc2ca61 | 1248 | |
c2ad7a1f | 1249 | if (ret) |
edc2ca61 | 1250 | break; |
6ea41c0c | 1251 | |
38935861 | 1252 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) |
6ea41c0c | 1253 | break; |
edc2ca61 | 1254 | } |
edc2ca61 | 1255 | |
c2ad7a1f | 1256 | return ret; |
edc2ca61 VB |
1257 | } |
1258 | ||
ff9543fd MN |
1259 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1260 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1261 | |
b682debd VB |
1262 | static bool suitable_migration_source(struct compact_control *cc, |
1263 | struct page *page) | |
1264 | { | |
282722b0 VB |
1265 | int block_mt; |
1266 | ||
9bebefd5 MG |
1267 | if (pageblock_skip_persistent(page)) |
1268 | return false; | |
1269 | ||
282722b0 | 1270 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1271 | return true; |
1272 | ||
282722b0 VB |
1273 | block_mt = get_pageblock_migratetype(page); |
1274 | ||
1275 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1276 | return is_migrate_movable(block_mt); | |
1277 | else | |
1278 | return block_mt == cc->migratetype; | |
b682debd VB |
1279 | } |
1280 | ||
018e9a49 | 1281 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1282 | static bool suitable_migration_target(struct compact_control *cc, |
1283 | struct page *page) | |
018e9a49 AM |
1284 | { |
1285 | /* If the page is a large free page, then disallow migration */ | |
1286 | if (PageBuddy(page)) { | |
1287 | /* | |
1288 | * We are checking page_order without zone->lock taken. But | |
1289 | * the only small danger is that we skip a potentially suitable | |
1290 | * pageblock, so it's not worth to check order for valid range. | |
1291 | */ | |
ab130f91 | 1292 | if (buddy_order_unsafe(page) >= pageblock_order) |
018e9a49 AM |
1293 | return false; |
1294 | } | |
1295 | ||
1ef36db2 YX |
1296 | if (cc->ignore_block_suitable) |
1297 | return true; | |
1298 | ||
018e9a49 | 1299 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1300 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1301 | return true; |
1302 | ||
1303 | /* Otherwise skip the block */ | |
1304 | return false; | |
1305 | } | |
1306 | ||
70b44595 MG |
1307 | static inline unsigned int |
1308 | freelist_scan_limit(struct compact_control *cc) | |
1309 | { | |
dd7ef7bd QC |
1310 | unsigned short shift = BITS_PER_LONG - 1; |
1311 | ||
1312 | return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; | |
70b44595 MG |
1313 | } |
1314 | ||
f2849aa0 VB |
1315 | /* |
1316 | * Test whether the free scanner has reached the same or lower pageblock than | |
1317 | * the migration scanner, and compaction should thus terminate. | |
1318 | */ | |
1319 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1320 | { | |
1321 | return (cc->free_pfn >> pageblock_order) | |
1322 | <= (cc->migrate_pfn >> pageblock_order); | |
1323 | } | |
1324 | ||
5a811889 MG |
1325 | /* |
1326 | * Used when scanning for a suitable migration target which scans freelists | |
1327 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1328 | * first on the next iteration of the free scanner | |
1329 | */ | |
1330 | static void | |
1331 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1332 | { | |
1333 | LIST_HEAD(sublist); | |
1334 | ||
1335 | if (!list_is_last(freelist, &freepage->lru)) { | |
1336 | list_cut_before(&sublist, freelist, &freepage->lru); | |
d2155fe5 | 1337 | list_splice_tail(&sublist, freelist); |
5a811889 MG |
1338 | } |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * Similar to move_freelist_head except used by the migration scanner | |
1343 | * when scanning forward. It's possible for these list operations to | |
1344 | * move against each other if they search the free list exactly in | |
1345 | * lockstep. | |
1346 | */ | |
70b44595 MG |
1347 | static void |
1348 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1349 | { | |
1350 | LIST_HEAD(sublist); | |
1351 | ||
1352 | if (!list_is_first(freelist, &freepage->lru)) { | |
1353 | list_cut_position(&sublist, freelist, &freepage->lru); | |
d2155fe5 | 1354 | list_splice_tail(&sublist, freelist); |
70b44595 MG |
1355 | } |
1356 | } | |
1357 | ||
5a811889 MG |
1358 | static void |
1359 | fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated) | |
1360 | { | |
1361 | unsigned long start_pfn, end_pfn; | |
6e2b7044 | 1362 | struct page *page; |
5a811889 MG |
1363 | |
1364 | /* Do not search around if there are enough pages already */ | |
1365 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1366 | return; | |
1367 | ||
1368 | /* Minimise scanning during async compaction */ | |
1369 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1370 | return; | |
1371 | ||
1372 | /* Pageblock boundaries */ | |
6e2b7044 VB |
1373 | start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); |
1374 | end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); | |
1375 | ||
1376 | page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); | |
1377 | if (!page) | |
1378 | return; | |
5a811889 MG |
1379 | |
1380 | /* Scan before */ | |
1381 | if (start_pfn != pfn) { | |
4fca9730 | 1382 | isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false); |
5a811889 MG |
1383 | if (cc->nr_freepages >= cc->nr_migratepages) |
1384 | return; | |
1385 | } | |
1386 | ||
1387 | /* Scan after */ | |
1388 | start_pfn = pfn + nr_isolated; | |
60fce36a | 1389 | if (start_pfn < end_pfn) |
4fca9730 | 1390 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); |
5a811889 MG |
1391 | |
1392 | /* Skip this pageblock in the future as it's full or nearly full */ | |
1393 | if (cc->nr_freepages < cc->nr_migratepages) | |
1394 | set_pageblock_skip(page); | |
1395 | } | |
1396 | ||
dbe2d4e4 MG |
1397 | /* Search orders in round-robin fashion */ |
1398 | static int next_search_order(struct compact_control *cc, int order) | |
1399 | { | |
1400 | order--; | |
1401 | if (order < 0) | |
1402 | order = cc->order - 1; | |
1403 | ||
1404 | /* Search wrapped around? */ | |
1405 | if (order == cc->search_order) { | |
1406 | cc->search_order--; | |
1407 | if (cc->search_order < 0) | |
1408 | cc->search_order = cc->order - 1; | |
1409 | return -1; | |
1410 | } | |
1411 | ||
1412 | return order; | |
1413 | } | |
1414 | ||
5a811889 MG |
1415 | static unsigned long |
1416 | fast_isolate_freepages(struct compact_control *cc) | |
1417 | { | |
b55ca526 | 1418 | unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); |
5a811889 | 1419 | unsigned int nr_scanned = 0; |
74e21484 | 1420 | unsigned long low_pfn, min_pfn, highest = 0; |
5a811889 MG |
1421 | unsigned long nr_isolated = 0; |
1422 | unsigned long distance; | |
1423 | struct page *page = NULL; | |
1424 | bool scan_start = false; | |
1425 | int order; | |
1426 | ||
1427 | /* Full compaction passes in a negative order */ | |
1428 | if (cc->order <= 0) | |
1429 | return cc->free_pfn; | |
1430 | ||
1431 | /* | |
1432 | * If starting the scan, use a deeper search and use the highest | |
1433 | * PFN found if a suitable one is not found. | |
1434 | */ | |
e332f741 | 1435 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1436 | limit = pageblock_nr_pages >> 1; |
1437 | scan_start = true; | |
1438 | } | |
1439 | ||
1440 | /* | |
1441 | * Preferred point is in the top quarter of the scan space but take | |
1442 | * a pfn from the top half if the search is problematic. | |
1443 | */ | |
1444 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1445 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1446 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1447 | ||
1448 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1449 | low_pfn = min_pfn; | |
1450 | ||
dbe2d4e4 MG |
1451 | /* |
1452 | * Search starts from the last successful isolation order or the next | |
1453 | * order to search after a previous failure | |
1454 | */ | |
1455 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1456 | ||
1457 | for (order = cc->search_order; | |
1458 | !page && order >= 0; | |
1459 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1460 | struct free_area *area = &cc->zone->free_area[order]; |
1461 | struct list_head *freelist; | |
1462 | struct page *freepage; | |
1463 | unsigned long flags; | |
1464 | unsigned int order_scanned = 0; | |
74e21484 | 1465 | unsigned long high_pfn = 0; |
5a811889 MG |
1466 | |
1467 | if (!area->nr_free) | |
1468 | continue; | |
1469 | ||
1470 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1471 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1472 | list_for_each_entry_reverse(freepage, freelist, lru) { | |
1473 | unsigned long pfn; | |
1474 | ||
1475 | order_scanned++; | |
1476 | nr_scanned++; | |
1477 | pfn = page_to_pfn(freepage); | |
1478 | ||
1479 | if (pfn >= highest) | |
6e2b7044 VB |
1480 | highest = max(pageblock_start_pfn(pfn), |
1481 | cc->zone->zone_start_pfn); | |
5a811889 MG |
1482 | |
1483 | if (pfn >= low_pfn) { | |
1484 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1485 | cc->search_order = order; |
5a811889 MG |
1486 | page = freepage; |
1487 | break; | |
1488 | } | |
1489 | ||
1490 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1491 | high_pfn = pfn; | |
1492 | ||
1493 | /* Shorten the scan if a candidate is found */ | |
1494 | limit >>= 1; | |
1495 | } | |
1496 | ||
1497 | if (order_scanned >= limit) | |
1498 | break; | |
1499 | } | |
1500 | ||
1501 | /* Use a minimum pfn if a preferred one was not found */ | |
1502 | if (!page && high_pfn) { | |
1503 | page = pfn_to_page(high_pfn); | |
1504 | ||
1505 | /* Update freepage for the list reorder below */ | |
1506 | freepage = page; | |
1507 | } | |
1508 | ||
1509 | /* Reorder to so a future search skips recent pages */ | |
1510 | move_freelist_head(freelist, freepage); | |
1511 | ||
1512 | /* Isolate the page if available */ | |
1513 | if (page) { | |
1514 | if (__isolate_free_page(page, order)) { | |
1515 | set_page_private(page, order); | |
1516 | nr_isolated = 1 << order; | |
1517 | cc->nr_freepages += nr_isolated; | |
1518 | list_add_tail(&page->lru, &cc->freepages); | |
1519 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1520 | } else { | |
1521 | /* If isolation fails, abort the search */ | |
5b56d996 | 1522 | order = cc->search_order + 1; |
5a811889 MG |
1523 | page = NULL; |
1524 | } | |
1525 | } | |
1526 | ||
1527 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1528 | ||
1529 | /* | |
b55ca526 | 1530 | * Smaller scan on next order so the total scan is related |
5a811889 MG |
1531 | * to freelist_scan_limit. |
1532 | */ | |
1533 | if (order_scanned >= limit) | |
b55ca526 | 1534 | limit = max(1U, limit >> 1); |
5a811889 MG |
1535 | } |
1536 | ||
1537 | if (!page) { | |
1538 | cc->fast_search_fail++; | |
1539 | if (scan_start) { | |
1540 | /* | |
1541 | * Use the highest PFN found above min. If one was | |
f3867755 | 1542 | * not found, be pessimistic for direct compaction |
5a811889 MG |
1543 | * and use the min mark. |
1544 | */ | |
1545 | if (highest) { | |
1546 | page = pfn_to_page(highest); | |
1547 | cc->free_pfn = highest; | |
1548 | } else { | |
e577c8b6 | 1549 | if (cc->direct_compaction && pfn_valid(min_pfn)) { |
73a6e474 | 1550 | page = pageblock_pfn_to_page(min_pfn, |
6e2b7044 VB |
1551 | min(pageblock_end_pfn(min_pfn), |
1552 | zone_end_pfn(cc->zone)), | |
73a6e474 | 1553 | cc->zone); |
5a811889 MG |
1554 | cc->free_pfn = min_pfn; |
1555 | } | |
1556 | } | |
1557 | } | |
1558 | } | |
1559 | ||
d097a6f6 MG |
1560 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1561 | highest -= pageblock_nr_pages; | |
5a811889 | 1562 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1563 | } |
5a811889 MG |
1564 | |
1565 | cc->total_free_scanned += nr_scanned; | |
1566 | if (!page) | |
1567 | return cc->free_pfn; | |
1568 | ||
1569 | low_pfn = page_to_pfn(page); | |
1570 | fast_isolate_around(cc, low_pfn, nr_isolated); | |
1571 | return low_pfn; | |
1572 | } | |
1573 | ||
2fe86e00 | 1574 | /* |
ff9543fd MN |
1575 | * Based on information in the current compact_control, find blocks |
1576 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1577 | */ |
edc2ca61 | 1578 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1579 | { |
edc2ca61 | 1580 | struct zone *zone = cc->zone; |
ff9543fd | 1581 | struct page *page; |
c96b9e50 | 1582 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1583 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1584 | unsigned long block_end_pfn; /* end of current pageblock */ |
1585 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1586 | struct list_head *freelist = &cc->freepages; |
4fca9730 | 1587 | unsigned int stride; |
2fe86e00 | 1588 | |
5a811889 MG |
1589 | /* Try a small search of the free lists for a candidate */ |
1590 | isolate_start_pfn = fast_isolate_freepages(cc); | |
1591 | if (cc->nr_freepages) | |
1592 | goto splitmap; | |
1593 | ||
ff9543fd MN |
1594 | /* |
1595 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1596 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1597 | * zone when isolating for the first time. For looping we also need |
1598 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1599 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1600 | * For ending point, take care when isolating in last pageblock of a | |
a1c1dbeb | 1601 | * zone which ends in the middle of a pageblock. |
49e068f0 VB |
1602 | * The low boundary is the end of the pageblock the migration scanner |
1603 | * is using. | |
ff9543fd | 1604 | */ |
e14c720e | 1605 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1606 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1607 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1608 | zone_end_pfn(zone)); | |
06b6640a | 1609 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1610 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1611 | |
ff9543fd MN |
1612 | /* |
1613 | * Isolate free pages until enough are available to migrate the | |
1614 | * pages on cc->migratepages. We stop searching if the migrate | |
1615 | * and free page scanners meet or enough free pages are isolated. | |
1616 | */ | |
f5f61a32 | 1617 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1618 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1619 | block_start_pfn -= pageblock_nr_pages, |
1620 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1621 | unsigned long nr_isolated; |
1622 | ||
f6ea3adb DR |
1623 | /* |
1624 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1625 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1626 | */ |
cb810ad2 | 1627 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1628 | cond_resched(); |
f6ea3adb | 1629 | |
7d49d886 VB |
1630 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1631 | zone); | |
1632 | if (!page) | |
ff9543fd MN |
1633 | continue; |
1634 | ||
1635 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1636 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1637 | continue; |
68e3e926 | 1638 | |
bb13ffeb MG |
1639 | /* If isolation recently failed, do not retry */ |
1640 | if (!isolation_suitable(cc, page)) | |
1641 | continue; | |
1642 | ||
e14c720e | 1643 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 MG |
1644 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
1645 | block_end_pfn, freelist, stride, false); | |
ff9543fd | 1646 | |
d097a6f6 MG |
1647 | /* Update the skip hint if the full pageblock was scanned */ |
1648 | if (isolate_start_pfn == block_end_pfn) | |
1649 | update_pageblock_skip(cc, page, block_start_pfn); | |
1650 | ||
cb2dcaf0 MG |
1651 | /* Are enough freepages isolated? */ |
1652 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1653 | if (isolate_start_pfn >= block_end_pfn) { |
1654 | /* | |
1655 | * Restart at previous pageblock if more | |
1656 | * freepages can be isolated next time. | |
1657 | */ | |
f5f61a32 VB |
1658 | isolate_start_pfn = |
1659 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1660 | } |
be976572 | 1661 | break; |
a46cbf3b | 1662 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1663 | /* |
a46cbf3b DR |
1664 | * If isolation failed early, do not continue |
1665 | * needlessly. | |
f5f61a32 | 1666 | */ |
a46cbf3b | 1667 | break; |
f5f61a32 | 1668 | } |
4fca9730 MG |
1669 | |
1670 | /* Adjust stride depending on isolation */ | |
1671 | if (nr_isolated) { | |
1672 | stride = 1; | |
1673 | continue; | |
1674 | } | |
1675 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1676 | } |
1677 | ||
7ed695e0 | 1678 | /* |
f5f61a32 VB |
1679 | * Record where the free scanner will restart next time. Either we |
1680 | * broke from the loop and set isolate_start_pfn based on the last | |
1681 | * call to isolate_freepages_block(), or we met the migration scanner | |
1682 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1683 | */ |
f5f61a32 | 1684 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1685 | |
1686 | splitmap: | |
1687 | /* __isolate_free_page() does not map the pages */ | |
1688 | split_map_pages(freelist); | |
748446bb MG |
1689 | } |
1690 | ||
1691 | /* | |
1692 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1693 | * from the isolated freelists in the block we are migrating to. | |
1694 | */ | |
1695 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1696 | unsigned long data) |
748446bb MG |
1697 | { |
1698 | struct compact_control *cc = (struct compact_control *)data; | |
1699 | struct page *freepage; | |
1700 | ||
748446bb | 1701 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1702 | isolate_freepages(cc); |
748446bb MG |
1703 | |
1704 | if (list_empty(&cc->freepages)) | |
1705 | return NULL; | |
1706 | } | |
1707 | ||
1708 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1709 | list_del(&freepage->lru); | |
1710 | cc->nr_freepages--; | |
1711 | ||
1712 | return freepage; | |
1713 | } | |
1714 | ||
1715 | /* | |
d53aea3d DR |
1716 | * This is a migrate-callback that "frees" freepages back to the isolated |
1717 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1718 | * special handling needed for NUMA. | |
1719 | */ | |
1720 | static void compaction_free(struct page *page, unsigned long data) | |
1721 | { | |
1722 | struct compact_control *cc = (struct compact_control *)data; | |
1723 | ||
1724 | list_add(&page->lru, &cc->freepages); | |
1725 | cc->nr_freepages++; | |
1726 | } | |
1727 | ||
ff9543fd MN |
1728 | /* possible outcome of isolate_migratepages */ |
1729 | typedef enum { | |
1730 | ISOLATE_ABORT, /* Abort compaction now */ | |
1731 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1732 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1733 | } isolate_migrate_t; | |
1734 | ||
5bbe3547 EM |
1735 | /* |
1736 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1737 | * compactable pages. | |
1738 | */ | |
6923aa0d SAS |
1739 | #ifdef CONFIG_PREEMPT_RT |
1740 | int sysctl_compact_unevictable_allowed __read_mostly = 0; | |
1741 | #else | |
5bbe3547 | 1742 | int sysctl_compact_unevictable_allowed __read_mostly = 1; |
6923aa0d | 1743 | #endif |
5bbe3547 | 1744 | |
70b44595 MG |
1745 | static inline void |
1746 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1747 | { | |
1748 | if (cc->fast_start_pfn == ULONG_MAX) | |
1749 | return; | |
1750 | ||
1751 | if (!cc->fast_start_pfn) | |
1752 | cc->fast_start_pfn = pfn; | |
1753 | ||
1754 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1755 | } | |
1756 | ||
1757 | static inline unsigned long | |
1758 | reinit_migrate_pfn(struct compact_control *cc) | |
1759 | { | |
1760 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1761 | return cc->migrate_pfn; | |
1762 | ||
1763 | cc->migrate_pfn = cc->fast_start_pfn; | |
1764 | cc->fast_start_pfn = ULONG_MAX; | |
1765 | ||
1766 | return cc->migrate_pfn; | |
1767 | } | |
1768 | ||
1769 | /* | |
1770 | * Briefly search the free lists for a migration source that already has | |
1771 | * some free pages to reduce the number of pages that need migration | |
1772 | * before a pageblock is free. | |
1773 | */ | |
1774 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1775 | { | |
1776 | unsigned int limit = freelist_scan_limit(cc); | |
1777 | unsigned int nr_scanned = 0; | |
1778 | unsigned long distance; | |
1779 | unsigned long pfn = cc->migrate_pfn; | |
1780 | unsigned long high_pfn; | |
1781 | int order; | |
15d28d0d | 1782 | bool found_block = false; |
70b44595 MG |
1783 | |
1784 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1785 | if (cc->ignore_skip_hint) | |
1786 | return pfn; | |
1787 | ||
1788 | /* | |
1789 | * If the migrate_pfn is not at the start of a zone or the start | |
1790 | * of a pageblock then assume this is a continuation of a previous | |
1791 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1792 | */ | |
1793 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1794 | return pfn; | |
1795 | ||
1796 | /* | |
1797 | * For smaller orders, just linearly scan as the number of pages | |
1798 | * to migrate should be relatively small and does not necessarily | |
1799 | * justify freeing up a large block for a small allocation. | |
1800 | */ | |
1801 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1802 | return pfn; | |
1803 | ||
1804 | /* | |
1805 | * Only allow kcompactd and direct requests for movable pages to | |
1806 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1807 | * reduces the risk that a large movable pageblock is freed for | |
1808 | * an unmovable/reclaimable small allocation. | |
1809 | */ | |
1810 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1811 | return pfn; | |
1812 | ||
1813 | /* | |
1814 | * When starting the migration scanner, pick any pageblock within the | |
1815 | * first half of the search space. Otherwise try and pick a pageblock | |
1816 | * within the first eighth to reduce the chances that a migration | |
1817 | * target later becomes a source. | |
1818 | */ | |
1819 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1820 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1821 | distance >>= 2; | |
1822 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1823 | ||
1824 | for (order = cc->order - 1; | |
15d28d0d | 1825 | order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; |
70b44595 MG |
1826 | order--) { |
1827 | struct free_area *area = &cc->zone->free_area[order]; | |
1828 | struct list_head *freelist; | |
1829 | unsigned long flags; | |
1830 | struct page *freepage; | |
1831 | ||
1832 | if (!area->nr_free) | |
1833 | continue; | |
1834 | ||
1835 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1836 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1837 | list_for_each_entry(freepage, freelist, lru) { | |
1838 | unsigned long free_pfn; | |
1839 | ||
15d28d0d WY |
1840 | if (nr_scanned++ >= limit) { |
1841 | move_freelist_tail(freelist, freepage); | |
1842 | break; | |
1843 | } | |
1844 | ||
70b44595 MG |
1845 | free_pfn = page_to_pfn(freepage); |
1846 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1847 | /* |
1848 | * Avoid if skipped recently. Ideally it would | |
1849 | * move to the tail but even safe iteration of | |
1850 | * the list assumes an entry is deleted, not | |
1851 | * reordered. | |
1852 | */ | |
15d28d0d | 1853 | if (get_pageblock_skip(freepage)) |
70b44595 | 1854 | continue; |
70b44595 MG |
1855 | |
1856 | /* Reorder to so a future search skips recent pages */ | |
1857 | move_freelist_tail(freelist, freepage); | |
1858 | ||
e380bebe | 1859 | update_fast_start_pfn(cc, free_pfn); |
70b44595 MG |
1860 | pfn = pageblock_start_pfn(free_pfn); |
1861 | cc->fast_search_fail = 0; | |
15d28d0d | 1862 | found_block = true; |
70b44595 MG |
1863 | set_pageblock_skip(freepage); |
1864 | break; | |
1865 | } | |
70b44595 MG |
1866 | } |
1867 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1868 | } | |
1869 | ||
1870 | cc->total_migrate_scanned += nr_scanned; | |
1871 | ||
1872 | /* | |
1873 | * If fast scanning failed then use a cached entry for a page block | |
1874 | * that had free pages as the basis for starting a linear scan. | |
1875 | */ | |
15d28d0d WY |
1876 | if (!found_block) { |
1877 | cc->fast_search_fail++; | |
70b44595 | 1878 | pfn = reinit_migrate_pfn(cc); |
15d28d0d | 1879 | } |
70b44595 MG |
1880 | return pfn; |
1881 | } | |
1882 | ||
ff9543fd | 1883 | /* |
edc2ca61 VB |
1884 | * Isolate all pages that can be migrated from the first suitable block, |
1885 | * starting at the block pointed to by the migrate scanner pfn within | |
1886 | * compact_control. | |
ff9543fd | 1887 | */ |
32aaf055 | 1888 | static isolate_migrate_t isolate_migratepages(struct compact_control *cc) |
ff9543fd | 1889 | { |
e1409c32 JK |
1890 | unsigned long block_start_pfn; |
1891 | unsigned long block_end_pfn; | |
1892 | unsigned long low_pfn; | |
edc2ca61 VB |
1893 | struct page *page; |
1894 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1895 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1896 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 1897 | bool fast_find_block; |
ff9543fd | 1898 | |
edc2ca61 VB |
1899 | /* |
1900 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
1901 | * initialized by compact_zone(). The first failure will use |
1902 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 1903 | */ |
70b44595 | 1904 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 1905 | block_start_pfn = pageblock_start_pfn(low_pfn); |
32aaf055 PL |
1906 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1907 | block_start_pfn = cc->zone->zone_start_pfn; | |
ff9543fd | 1908 | |
70b44595 MG |
1909 | /* |
1910 | * fast_find_migrateblock marks a pageblock skipped so to avoid | |
1911 | * the isolation_suitable check below, check whether the fast | |
1912 | * search was successful. | |
1913 | */ | |
1914 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
1915 | ||
ff9543fd | 1916 | /* Only scan within a pageblock boundary */ |
06b6640a | 1917 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1918 | |
edc2ca61 VB |
1919 | /* |
1920 | * Iterate over whole pageblocks until we find the first suitable. | |
1921 | * Do not cross the free scanner. | |
1922 | */ | |
e1409c32 | 1923 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 1924 | fast_find_block = false, |
c2ad7a1f | 1925 | cc->migrate_pfn = low_pfn = block_end_pfn, |
e1409c32 JK |
1926 | block_start_pfn = block_end_pfn, |
1927 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1928 | |
edc2ca61 VB |
1929 | /* |
1930 | * This can potentially iterate a massively long zone with | |
1931 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 1932 | * need to schedule. |
edc2ca61 | 1933 | */ |
cb810ad2 | 1934 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1935 | cond_resched(); |
ff9543fd | 1936 | |
32aaf055 PL |
1937 | page = pageblock_pfn_to_page(block_start_pfn, |
1938 | block_end_pfn, cc->zone); | |
7d49d886 | 1939 | if (!page) |
edc2ca61 VB |
1940 | continue; |
1941 | ||
e380bebe MG |
1942 | /* |
1943 | * If isolation recently failed, do not retry. Only check the | |
1944 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
1945 | * to be visited multiple times. Assume skip was checked | |
1946 | * before making it "skip" so other compaction instances do | |
1947 | * not scan the same block. | |
1948 | */ | |
1949 | if (IS_ALIGNED(low_pfn, pageblock_nr_pages) && | |
1950 | !fast_find_block && !isolation_suitable(cc, page)) | |
edc2ca61 VB |
1951 | continue; |
1952 | ||
1953 | /* | |
9bebefd5 MG |
1954 | * For async compaction, also only scan in MOVABLE blocks |
1955 | * without huge pages. Async compaction is optimistic to see | |
1956 | * if the minimum amount of work satisfies the allocation. | |
1957 | * The cached PFN is updated as it's possible that all | |
1958 | * remaining blocks between source and target are unsuitable | |
1959 | * and the compaction scanners fail to meet. | |
edc2ca61 | 1960 | */ |
9bebefd5 MG |
1961 | if (!suitable_migration_source(cc, page)) { |
1962 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 1963 | continue; |
9bebefd5 | 1964 | } |
edc2ca61 VB |
1965 | |
1966 | /* Perform the isolation */ | |
c2ad7a1f OS |
1967 | if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, |
1968 | isolate_mode)) | |
edc2ca61 VB |
1969 | return ISOLATE_ABORT; |
1970 | ||
1971 | /* | |
1972 | * Either we isolated something and proceed with migration. Or | |
1973 | * we failed and compact_zone should decide if we should | |
1974 | * continue or not. | |
1975 | */ | |
1976 | break; | |
1977 | } | |
1978 | ||
edc2ca61 | 1979 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1980 | } |
1981 | ||
21c527a3 YB |
1982 | /* |
1983 | * order == -1 is expected when compacting via | |
1984 | * /proc/sys/vm/compact_memory | |
1985 | */ | |
1986 | static inline bool is_via_compact_memory(int order) | |
1987 | { | |
1988 | return order == -1; | |
1989 | } | |
1990 | ||
facdaa91 NG |
1991 | static bool kswapd_is_running(pg_data_t *pgdat) |
1992 | { | |
b03fbd4f | 1993 | return pgdat->kswapd && task_is_running(pgdat->kswapd); |
facdaa91 NG |
1994 | } |
1995 | ||
1996 | /* | |
1997 | * A zone's fragmentation score is the external fragmentation wrt to the | |
40d7e203 CTR |
1998 | * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. |
1999 | */ | |
2000 | static unsigned int fragmentation_score_zone(struct zone *zone) | |
2001 | { | |
2002 | return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); | |
2003 | } | |
2004 | ||
2005 | /* | |
2006 | * A weighted zone's fragmentation score is the external fragmentation | |
2007 | * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It | |
2008 | * returns a value in the range [0, 100]. | |
facdaa91 NG |
2009 | * |
2010 | * The scaling factor ensures that proactive compaction focuses on larger | |
2011 | * zones like ZONE_NORMAL, rather than smaller, specialized zones like | |
2012 | * ZONE_DMA32. For smaller zones, the score value remains close to zero, | |
2013 | * and thus never exceeds the high threshold for proactive compaction. | |
2014 | */ | |
40d7e203 | 2015 | static unsigned int fragmentation_score_zone_weighted(struct zone *zone) |
facdaa91 NG |
2016 | { |
2017 | unsigned long score; | |
2018 | ||
40d7e203 | 2019 | score = zone->present_pages * fragmentation_score_zone(zone); |
facdaa91 NG |
2020 | return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); |
2021 | } | |
2022 | ||
2023 | /* | |
2024 | * The per-node proactive (background) compaction process is started by its | |
2025 | * corresponding kcompactd thread when the node's fragmentation score | |
2026 | * exceeds the high threshold. The compaction process remains active till | |
2027 | * the node's score falls below the low threshold, or one of the back-off | |
2028 | * conditions is met. | |
2029 | */ | |
d34c0a75 | 2030 | static unsigned int fragmentation_score_node(pg_data_t *pgdat) |
facdaa91 | 2031 | { |
d34c0a75 | 2032 | unsigned int score = 0; |
facdaa91 NG |
2033 | int zoneid; |
2034 | ||
2035 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2036 | struct zone *zone; | |
2037 | ||
2038 | zone = &pgdat->node_zones[zoneid]; | |
40d7e203 | 2039 | score += fragmentation_score_zone_weighted(zone); |
facdaa91 NG |
2040 | } |
2041 | ||
2042 | return score; | |
2043 | } | |
2044 | ||
d34c0a75 | 2045 | static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low) |
facdaa91 | 2046 | { |
d34c0a75 | 2047 | unsigned int wmark_low; |
facdaa91 NG |
2048 | |
2049 | /* | |
f0953a1b IM |
2050 | * Cap the low watermark to avoid excessive compaction |
2051 | * activity in case a user sets the proactiveness tunable | |
facdaa91 NG |
2052 | * close to 100 (maximum). |
2053 | */ | |
d34c0a75 NG |
2054 | wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); |
2055 | return low ? wmark_low : min(wmark_low + 10, 100U); | |
facdaa91 NG |
2056 | } |
2057 | ||
2058 | static bool should_proactive_compact_node(pg_data_t *pgdat) | |
2059 | { | |
2060 | int wmark_high; | |
2061 | ||
2062 | if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) | |
2063 | return false; | |
2064 | ||
2065 | wmark_high = fragmentation_score_wmark(pgdat, false); | |
2066 | return fragmentation_score_node(pgdat) > wmark_high; | |
2067 | } | |
2068 | ||
40cacbcb | 2069 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 2070 | { |
8fb74b9f | 2071 | unsigned int order; |
d39773a0 | 2072 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 2073 | int ret; |
748446bb | 2074 | |
753341a4 | 2075 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 2076 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 2077 | /* Let the next compaction start anew. */ |
40cacbcb | 2078 | reset_cached_positions(cc->zone); |
55b7c4c9 | 2079 | |
62997027 MG |
2080 | /* |
2081 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 2082 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
2083 | * flag itself as the decision to be clear should be directly |
2084 | * based on an allocation request. | |
2085 | */ | |
accf6242 | 2086 | if (cc->direct_compaction) |
40cacbcb | 2087 | cc->zone->compact_blockskip_flush = true; |
62997027 | 2088 | |
c8f7de0b MH |
2089 | if (cc->whole_zone) |
2090 | return COMPACT_COMPLETE; | |
2091 | else | |
2092 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 2093 | } |
748446bb | 2094 | |
facdaa91 NG |
2095 | if (cc->proactive_compaction) { |
2096 | int score, wmark_low; | |
2097 | pg_data_t *pgdat; | |
2098 | ||
2099 | pgdat = cc->zone->zone_pgdat; | |
2100 | if (kswapd_is_running(pgdat)) | |
2101 | return COMPACT_PARTIAL_SKIPPED; | |
2102 | ||
2103 | score = fragmentation_score_zone(cc->zone); | |
2104 | wmark_low = fragmentation_score_wmark(pgdat, true); | |
2105 | ||
2106 | if (score > wmark_low) | |
2107 | ret = COMPACT_CONTINUE; | |
2108 | else | |
2109 | ret = COMPACT_SUCCESS; | |
2110 | ||
2111 | goto out; | |
2112 | } | |
2113 | ||
21c527a3 | 2114 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
2115 | return COMPACT_CONTINUE; |
2116 | ||
efe771c7 MG |
2117 | /* |
2118 | * Always finish scanning a pageblock to reduce the possibility of | |
2119 | * fallbacks in the future. This is particularly important when | |
2120 | * migration source is unmovable/reclaimable but it's not worth | |
2121 | * special casing. | |
2122 | */ | |
2123 | if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
2124 | return COMPACT_CONTINUE; | |
baf6a9a1 | 2125 | |
56de7263 | 2126 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 2127 | ret = COMPACT_NO_SUITABLE_PAGE; |
8fb74b9f | 2128 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 2129 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 2130 | bool can_steal; |
8fb74b9f MG |
2131 | |
2132 | /* Job done if page is free of the right migratetype */ | |
b03641af | 2133 | if (!free_area_empty(area, migratetype)) |
cf378319 | 2134 | return COMPACT_SUCCESS; |
8fb74b9f | 2135 | |
2149cdae JK |
2136 | #ifdef CONFIG_CMA |
2137 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
2138 | if (migratetype == MIGRATE_MOVABLE && | |
b03641af | 2139 | !free_area_empty(area, MIGRATE_CMA)) |
cf378319 | 2140 | return COMPACT_SUCCESS; |
2149cdae JK |
2141 | #endif |
2142 | /* | |
2143 | * Job done if allocation would steal freepages from | |
2144 | * other migratetype buddy lists. | |
2145 | */ | |
2146 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
2147 | true, &can_steal) != -1) { |
2148 | ||
2149 | /* movable pages are OK in any pageblock */ | |
2150 | if (migratetype == MIGRATE_MOVABLE) | |
2151 | return COMPACT_SUCCESS; | |
2152 | ||
2153 | /* | |
2154 | * We are stealing for a non-movable allocation. Make | |
2155 | * sure we finish compacting the current pageblock | |
2156 | * first so it is as free as possible and we won't | |
2157 | * have to steal another one soon. This only applies | |
2158 | * to sync compaction, as async compaction operates | |
2159 | * on pageblocks of the same migratetype. | |
2160 | */ | |
2161 | if (cc->mode == MIGRATE_ASYNC || | |
2162 | IS_ALIGNED(cc->migrate_pfn, | |
2163 | pageblock_nr_pages)) { | |
2164 | return COMPACT_SUCCESS; | |
2165 | } | |
2166 | ||
cb2dcaf0 MG |
2167 | ret = COMPACT_CONTINUE; |
2168 | break; | |
baf6a9a1 | 2169 | } |
56de7263 MG |
2170 | } |
2171 | ||
facdaa91 | 2172 | out: |
cb2dcaf0 MG |
2173 | if (cc->contended || fatal_signal_pending(current)) |
2174 | ret = COMPACT_CONTENDED; | |
2175 | ||
2176 | return ret; | |
837d026d JK |
2177 | } |
2178 | ||
40cacbcb | 2179 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
2180 | { |
2181 | int ret; | |
2182 | ||
40cacbcb MG |
2183 | ret = __compact_finished(cc); |
2184 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
2185 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
2186 | ret = COMPACT_CONTINUE; | |
2187 | ||
2188 | return ret; | |
748446bb MG |
2189 | } |
2190 | ||
ea7ab982 | 2191 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 2192 | unsigned int alloc_flags, |
97a225e6 | 2193 | int highest_zoneidx, |
86a294a8 | 2194 | unsigned long wmark_target) |
3e7d3449 | 2195 | { |
3e7d3449 MG |
2196 | unsigned long watermark; |
2197 | ||
21c527a3 | 2198 | if (is_via_compact_memory(order)) |
3957c776 MH |
2199 | return COMPACT_CONTINUE; |
2200 | ||
a9214443 | 2201 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
2202 | /* |
2203 | * If watermarks for high-order allocation are already met, there | |
2204 | * should be no need for compaction at all. | |
2205 | */ | |
97a225e6 | 2206 | if (zone_watermark_ok(zone, order, watermark, highest_zoneidx, |
ebff3980 | 2207 | alloc_flags)) |
cf378319 | 2208 | return COMPACT_SUCCESS; |
ebff3980 | 2209 | |
3e7d3449 | 2210 | /* |
9861a62c | 2211 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
2212 | * isolate free pages for migration targets. This means that the |
2213 | * watermark and alloc_flags have to match, or be more pessimistic than | |
2214 | * the check in __isolate_free_page(). We don't use the direct | |
2215 | * compactor's alloc_flags, as they are not relevant for freepage | |
97a225e6 JK |
2216 | * isolation. We however do use the direct compactor's highest_zoneidx |
2217 | * to skip over zones where lowmem reserves would prevent allocation | |
2218 | * even if compaction succeeds. | |
8348faf9 VB |
2219 | * For costly orders, we require low watermark instead of min for |
2220 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
2221 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
2222 | * suitable migration targets | |
3e7d3449 | 2223 | */ |
8348faf9 VB |
2224 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
2225 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
2226 | watermark += compact_gap(order); | |
97a225e6 | 2227 | if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx, |
d883c6cf | 2228 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
2229 | return COMPACT_SKIPPED; |
2230 | ||
cc5c9f09 VB |
2231 | return COMPACT_CONTINUE; |
2232 | } | |
2233 | ||
2b1a20c3 HS |
2234 | /* |
2235 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
2236 | * Returns | |
2237 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
2238 | * COMPACT_SUCCESS - If the allocation would succeed without compaction | |
2239 | * COMPACT_CONTINUE - If compaction should run now | |
2240 | */ | |
cc5c9f09 VB |
2241 | enum compact_result compaction_suitable(struct zone *zone, int order, |
2242 | unsigned int alloc_flags, | |
97a225e6 | 2243 | int highest_zoneidx) |
cc5c9f09 VB |
2244 | { |
2245 | enum compact_result ret; | |
2246 | int fragindex; | |
2247 | ||
97a225e6 | 2248 | ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx, |
cc5c9f09 | 2249 | zone_page_state(zone, NR_FREE_PAGES)); |
3e7d3449 MG |
2250 | /* |
2251 | * fragmentation index determines if allocation failures are due to | |
2252 | * low memory or external fragmentation | |
2253 | * | |
ebff3980 VB |
2254 | * index of -1000 would imply allocations might succeed depending on |
2255 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2256 | * index towards 0 implies failure is due to lack of memory |
2257 | * index towards 1000 implies failure is due to fragmentation | |
2258 | * | |
20311420 VB |
2259 | * Only compact if a failure would be due to fragmentation. Also |
2260 | * ignore fragindex for non-costly orders where the alternative to | |
2261 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2262 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2263 | * excessive compaction for costly orders, but it should not be at the | |
2264 | * expense of system stability. | |
3e7d3449 | 2265 | */ |
20311420 | 2266 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
2267 | fragindex = fragmentation_index(zone, order); |
2268 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
2269 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
2270 | } | |
837d026d | 2271 | |
837d026d JK |
2272 | trace_mm_compaction_suitable(zone, order, ret); |
2273 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
2274 | ret = COMPACT_SKIPPED; | |
2275 | ||
2276 | return ret; | |
2277 | } | |
2278 | ||
86a294a8 MH |
2279 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2280 | int alloc_flags) | |
2281 | { | |
2282 | struct zone *zone; | |
2283 | struct zoneref *z; | |
2284 | ||
2285 | /* | |
2286 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2287 | * retrying the reclaim. | |
2288 | */ | |
97a225e6 JK |
2289 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2290 | ac->highest_zoneidx, ac->nodemask) { | |
86a294a8 MH |
2291 | unsigned long available; |
2292 | enum compact_result compact_result; | |
2293 | ||
2294 | /* | |
2295 | * Do not consider all the reclaimable memory because we do not | |
2296 | * want to trash just for a single high order allocation which | |
2297 | * is even not guaranteed to appear even if __compaction_suitable | |
2298 | * is happy about the watermark check. | |
2299 | */ | |
5a1c84b4 | 2300 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
2301 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
2302 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
97a225e6 | 2303 | ac->highest_zoneidx, available); |
cc5c9f09 | 2304 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
2305 | return true; |
2306 | } | |
2307 | ||
2308 | return false; | |
2309 | } | |
2310 | ||
5e1f0f09 MG |
2311 | static enum compact_result |
2312 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2313 | { |
ea7ab982 | 2314 | enum compact_result ret; |
40cacbcb MG |
2315 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2316 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2317 | unsigned long last_migrated_pfn; |
e0b9daeb | 2318 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2319 | bool update_cached; |
84b328aa | 2320 | unsigned int nr_succeeded = 0; |
748446bb | 2321 | |
a94b5252 YS |
2322 | /* |
2323 | * These counters track activities during zone compaction. Initialize | |
2324 | * them before compacting a new zone. | |
2325 | */ | |
2326 | cc->total_migrate_scanned = 0; | |
2327 | cc->total_free_scanned = 0; | |
2328 | cc->nr_migratepages = 0; | |
2329 | cc->nr_freepages = 0; | |
2330 | INIT_LIST_HEAD(&cc->freepages); | |
2331 | INIT_LIST_HEAD(&cc->migratepages); | |
2332 | ||
01c0bfe0 | 2333 | cc->migratetype = gfp_migratetype(cc->gfp_mask); |
40cacbcb | 2334 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
97a225e6 | 2335 | cc->highest_zoneidx); |
c46649de | 2336 | /* Compaction is likely to fail */ |
cf378319 | 2337 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 2338 | return ret; |
c46649de MH |
2339 | |
2340 | /* huh, compaction_suitable is returning something unexpected */ | |
2341 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 2342 | |
d3132e4b VB |
2343 | /* |
2344 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2345 | * is about to be retried after being deferred. |
d3132e4b | 2346 | */ |
40cacbcb MG |
2347 | if (compaction_restarting(cc->zone, cc->order)) |
2348 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2349 | |
c89511ab MG |
2350 | /* |
2351 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2352 | * information on where the scanners should start (unless we explicitly |
2353 | * want to compact the whole zone), but check that it is initialised | |
2354 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2355 | */ |
70b44595 | 2356 | cc->fast_start_pfn = 0; |
06ed2998 | 2357 | if (cc->whole_zone) { |
c89511ab | 2358 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2359 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2360 | } else { | |
40cacbcb MG |
2361 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2362 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2363 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2364 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2365 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2366 | } |
2367 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2368 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2369 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2370 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2371 | } |
c8f7de0b | 2372 | |
e332f741 | 2373 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2374 | cc->whole_zone = true; |
2375 | } | |
c8f7de0b | 2376 | |
566e54e1 | 2377 | last_migrated_pfn = 0; |
748446bb | 2378 | |
8854c55f MG |
2379 | /* |
2380 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2381 | * the basis that some migrations will fail in ASYNC mode. However, | |
2382 | * if the cached PFNs match and pageblocks are skipped due to having | |
2383 | * no isolation candidates, then the sync state does not matter. | |
2384 | * Until a pageblock with isolation candidates is found, keep the | |
2385 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2386 | */ | |
2387 | update_cached = !sync && | |
2388 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2389 | ||
abd4349f | 2390 | trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); |
0eb927c0 | 2391 | |
361a2a22 MK |
2392 | /* lru_add_drain_all could be expensive with involving other CPUs */ |
2393 | lru_add_drain(); | |
748446bb | 2394 | |
40cacbcb | 2395 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2396 | int err; |
19d3cf9d | 2397 | unsigned long iteration_start_pfn = cc->migrate_pfn; |
748446bb | 2398 | |
804d3121 MG |
2399 | /* |
2400 | * Avoid multiple rescans which can happen if a page cannot be | |
2401 | * isolated (dirty/writeback in async mode) or if the migrated | |
2402 | * pages are being allocated before the pageblock is cleared. | |
2403 | * The first rescan will capture the entire pageblock for | |
2404 | * migration. If it fails, it'll be marked skip and scanning | |
2405 | * will proceed as normal. | |
2406 | */ | |
2407 | cc->rescan = false; | |
2408 | if (pageblock_start_pfn(last_migrated_pfn) == | |
19d3cf9d | 2409 | pageblock_start_pfn(iteration_start_pfn)) { |
804d3121 MG |
2410 | cc->rescan = true; |
2411 | } | |
2412 | ||
32aaf055 | 2413 | switch (isolate_migratepages(cc)) { |
f9e35b3b | 2414 | case ISOLATE_ABORT: |
2d1e1041 | 2415 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2416 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2417 | cc->nr_migratepages = 0; |
f9e35b3b MG |
2418 | goto out; |
2419 | case ISOLATE_NONE: | |
8854c55f MG |
2420 | if (update_cached) { |
2421 | cc->zone->compact_cached_migrate_pfn[1] = | |
2422 | cc->zone->compact_cached_migrate_pfn[0]; | |
2423 | } | |
2424 | ||
fdaf7f5c VB |
2425 | /* |
2426 | * We haven't isolated and migrated anything, but | |
2427 | * there might still be unflushed migrations from | |
2428 | * previous cc->order aligned block. | |
2429 | */ | |
2430 | goto check_drain; | |
f9e35b3b | 2431 | case ISOLATE_SUCCESS: |
8854c55f | 2432 | update_cached = false; |
19d3cf9d | 2433 | last_migrated_pfn = iteration_start_pfn; |
f9e35b3b | 2434 | } |
748446bb | 2435 | |
d53aea3d | 2436 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2437 | compaction_free, (unsigned long)cc, cc->mode, |
84b328aa | 2438 | MR_COMPACTION, &nr_succeeded); |
748446bb | 2439 | |
abd4349f | 2440 | trace_mm_compaction_migratepages(cc, nr_succeeded); |
748446bb | 2441 | |
f8c9301f VB |
2442 | /* All pages were either migrated or will be released */ |
2443 | cc->nr_migratepages = 0; | |
9d502c1c | 2444 | if (err) { |
5733c7d1 | 2445 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2446 | /* |
2447 | * migrate_pages() may return -ENOMEM when scanners meet | |
2448 | * and we want compact_finished() to detect it | |
2449 | */ | |
f2849aa0 | 2450 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2451 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2452 | goto out; |
2453 | } | |
fdd048e1 VB |
2454 | /* |
2455 | * We failed to migrate at least one page in the current | |
2456 | * order-aligned block, so skip the rest of it. | |
2457 | */ | |
2458 | if (cc->direct_compaction && | |
2459 | (cc->mode == MIGRATE_ASYNC)) { | |
2460 | cc->migrate_pfn = block_end_pfn( | |
2461 | cc->migrate_pfn - 1, cc->order); | |
2462 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 2463 | last_migrated_pfn = 0; |
fdd048e1 | 2464 | } |
748446bb | 2465 | } |
fdaf7f5c | 2466 | |
fdaf7f5c VB |
2467 | check_drain: |
2468 | /* | |
2469 | * Has the migration scanner moved away from the previous | |
2470 | * cc->order aligned block where we migrated from? If yes, | |
2471 | * flush the pages that were freed, so that they can merge and | |
2472 | * compact_finished() can detect immediately if allocation | |
2473 | * would succeed. | |
2474 | */ | |
566e54e1 | 2475 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c | 2476 | unsigned long current_block_start = |
06b6640a | 2477 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2478 | |
566e54e1 | 2479 | if (last_migrated_pfn < current_block_start) { |
b01b2141 | 2480 | lru_add_drain_cpu_zone(cc->zone); |
fdaf7f5c | 2481 | /* No more flushing until we migrate again */ |
566e54e1 | 2482 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2483 | } |
2484 | } | |
2485 | ||
5e1f0f09 MG |
2486 | /* Stop if a page has been captured */ |
2487 | if (capc && capc->page) { | |
2488 | ret = COMPACT_SUCCESS; | |
2489 | break; | |
2490 | } | |
748446bb MG |
2491 | } |
2492 | ||
f9e35b3b | 2493 | out: |
6bace090 VB |
2494 | /* |
2495 | * Release free pages and update where the free scanner should restart, | |
2496 | * so we don't leave any returned pages behind in the next attempt. | |
2497 | */ | |
2498 | if (cc->nr_freepages > 0) { | |
2499 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2500 | ||
2501 | cc->nr_freepages = 0; | |
2502 | VM_BUG_ON(free_pfn == 0); | |
2503 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2504 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2505 | /* |
2506 | * Only go back, not forward. The cached pfn might have been | |
2507 | * already reset to zone end in compact_finished() | |
2508 | */ | |
40cacbcb MG |
2509 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2510 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2511 | } |
748446bb | 2512 | |
7f354a54 DR |
2513 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2514 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2515 | ||
abd4349f | 2516 | trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); |
0eb927c0 | 2517 | |
748446bb MG |
2518 | return ret; |
2519 | } | |
76ab0f53 | 2520 | |
ea7ab982 | 2521 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2522 | gfp_t gfp_mask, enum compact_priority prio, |
97a225e6 | 2523 | unsigned int alloc_flags, int highest_zoneidx, |
5e1f0f09 | 2524 | struct page **capture) |
56de7263 | 2525 | { |
ea7ab982 | 2526 | enum compact_result ret; |
56de7263 | 2527 | struct compact_control cc = { |
56de7263 | 2528 | .order = order, |
dbe2d4e4 | 2529 | .search_order = order, |
6d7ce559 | 2530 | .gfp_mask = gfp_mask, |
56de7263 | 2531 | .zone = zone, |
a5508cd8 VB |
2532 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2533 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 | 2534 | .alloc_flags = alloc_flags, |
97a225e6 | 2535 | .highest_zoneidx = highest_zoneidx, |
accf6242 | 2536 | .direct_compaction = true, |
a8e025e5 | 2537 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2538 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2539 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2540 | }; |
5e1f0f09 MG |
2541 | struct capture_control capc = { |
2542 | .cc = &cc, | |
2543 | .page = NULL, | |
2544 | }; | |
2545 | ||
b9e20f0d VB |
2546 | /* |
2547 | * Make sure the structs are really initialized before we expose the | |
2548 | * capture control, in case we are interrupted and the interrupt handler | |
2549 | * frees a page. | |
2550 | */ | |
2551 | barrier(); | |
2552 | WRITE_ONCE(current->capture_control, &capc); | |
56de7263 | 2553 | |
5e1f0f09 | 2554 | ret = compact_zone(&cc, &capc); |
e64c5237 SL |
2555 | |
2556 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2557 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2558 | ||
b9e20f0d VB |
2559 | /* |
2560 | * Make sure we hide capture control first before we read the captured | |
2561 | * page pointer, otherwise an interrupt could free and capture a page | |
2562 | * and we would leak it. | |
2563 | */ | |
2564 | WRITE_ONCE(current->capture_control, NULL); | |
2565 | *capture = READ_ONCE(capc.page); | |
06dac2f4 CTR |
2566 | /* |
2567 | * Technically, it is also possible that compaction is skipped but | |
2568 | * the page is still captured out of luck(IRQ came and freed the page). | |
2569 | * Returning COMPACT_SUCCESS in such cases helps in properly accounting | |
2570 | * the COMPACT[STALL|FAIL] when compaction is skipped. | |
2571 | */ | |
2572 | if (*capture) | |
2573 | ret = COMPACT_SUCCESS; | |
5e1f0f09 | 2574 | |
e64c5237 | 2575 | return ret; |
56de7263 MG |
2576 | } |
2577 | ||
5e771905 MG |
2578 | int sysctl_extfrag_threshold = 500; |
2579 | ||
56de7263 MG |
2580 | /** |
2581 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2582 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2583 | * @order: The order of the current allocation |
2584 | * @alloc_flags: The allocation flags of the current allocation | |
2585 | * @ac: The context of current allocation | |
112d2d29 | 2586 | * @prio: Determines how hard direct compaction should try to succeed |
6467552c | 2587 | * @capture: Pointer to free page created by compaction will be stored here |
56de7263 MG |
2588 | * |
2589 | * This is the main entry point for direct page compaction. | |
2590 | */ | |
ea7ab982 | 2591 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2592 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2593 | enum compact_priority prio, struct page **capture) |
56de7263 | 2594 | { |
56de7263 | 2595 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
2596 | struct zoneref *z; |
2597 | struct zone *zone; | |
1d4746d3 | 2598 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2599 | |
73e64c51 MH |
2600 | /* |
2601 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2602 | * tricky context because the migration might require IO | |
2603 | */ | |
2604 | if (!may_perform_io) | |
53853e2d | 2605 | return COMPACT_SKIPPED; |
56de7263 | 2606 | |
a5508cd8 | 2607 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2608 | |
56de7263 | 2609 | /* Compact each zone in the list */ |
97a225e6 JK |
2610 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2611 | ac->highest_zoneidx, ac->nodemask) { | |
ea7ab982 | 2612 | enum compact_result status; |
56de7263 | 2613 | |
a8e025e5 VB |
2614 | if (prio > MIN_COMPACT_PRIORITY |
2615 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2616 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2617 | continue; |
1d4746d3 | 2618 | } |
53853e2d | 2619 | |
a5508cd8 | 2620 | status = compact_zone_order(zone, order, gfp_mask, prio, |
97a225e6 | 2621 | alloc_flags, ac->highest_zoneidx, capture); |
56de7263 MG |
2622 | rc = max(status, rc); |
2623 | ||
7ceb009a VB |
2624 | /* The allocation should succeed, stop compacting */ |
2625 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2626 | /* |
2627 | * We think the allocation will succeed in this zone, | |
2628 | * but it is not certain, hence the false. The caller | |
2629 | * will repeat this with true if allocation indeed | |
2630 | * succeeds in this zone. | |
2631 | */ | |
2632 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2633 | |
c3486f53 | 2634 | break; |
1f9efdef VB |
2635 | } |
2636 | ||
a5508cd8 | 2637 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2638 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2639 | /* |
2640 | * We think that allocation won't succeed in this zone | |
2641 | * so we defer compaction there. If it ends up | |
2642 | * succeeding after all, it will be reset. | |
2643 | */ | |
2644 | defer_compaction(zone, order); | |
1f9efdef VB |
2645 | |
2646 | /* | |
2647 | * We might have stopped compacting due to need_resched() in | |
2648 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2649 | * case do not try further zones |
1f9efdef | 2650 | */ |
c3486f53 VB |
2651 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2652 | || fatal_signal_pending(current)) | |
2653 | break; | |
56de7263 MG |
2654 | } |
2655 | ||
2656 | return rc; | |
2657 | } | |
2658 | ||
facdaa91 NG |
2659 | /* |
2660 | * Compact all zones within a node till each zone's fragmentation score | |
2661 | * reaches within proactive compaction thresholds (as determined by the | |
2662 | * proactiveness tunable). | |
2663 | * | |
2664 | * It is possible that the function returns before reaching score targets | |
2665 | * due to various back-off conditions, such as, contention on per-node or | |
2666 | * per-zone locks. | |
2667 | */ | |
2668 | static void proactive_compact_node(pg_data_t *pgdat) | |
2669 | { | |
2670 | int zoneid; | |
2671 | struct zone *zone; | |
2672 | struct compact_control cc = { | |
2673 | .order = -1, | |
2674 | .mode = MIGRATE_SYNC_LIGHT, | |
2675 | .ignore_skip_hint = true, | |
2676 | .whole_zone = true, | |
2677 | .gfp_mask = GFP_KERNEL, | |
2678 | .proactive_compaction = true, | |
2679 | }; | |
2680 | ||
2681 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2682 | zone = &pgdat->node_zones[zoneid]; | |
2683 | if (!populated_zone(zone)) | |
2684 | continue; | |
2685 | ||
2686 | cc.zone = zone; | |
2687 | ||
2688 | compact_zone(&cc, NULL); | |
2689 | ||
2690 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2691 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2692 | } | |
2693 | } | |
56de7263 | 2694 | |
76ab0f53 | 2695 | /* Compact all zones within a node */ |
791cae96 | 2696 | static void compact_node(int nid) |
76ab0f53 | 2697 | { |
791cae96 | 2698 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 2699 | int zoneid; |
76ab0f53 | 2700 | struct zone *zone; |
791cae96 VB |
2701 | struct compact_control cc = { |
2702 | .order = -1, | |
2703 | .mode = MIGRATE_SYNC, | |
2704 | .ignore_skip_hint = true, | |
2705 | .whole_zone = true, | |
73e64c51 | 2706 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
2707 | }; |
2708 | ||
76ab0f53 | 2709 | |
76ab0f53 | 2710 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
2711 | |
2712 | zone = &pgdat->node_zones[zoneid]; | |
2713 | if (!populated_zone(zone)) | |
2714 | continue; | |
2715 | ||
791cae96 | 2716 | cc.zone = zone; |
76ab0f53 | 2717 | |
5e1f0f09 | 2718 | compact_zone(&cc, NULL); |
75469345 | 2719 | |
791cae96 VB |
2720 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2721 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 2722 | } |
76ab0f53 MG |
2723 | } |
2724 | ||
2725 | /* Compact all nodes in the system */ | |
7964c06d | 2726 | static void compact_nodes(void) |
76ab0f53 MG |
2727 | { |
2728 | int nid; | |
2729 | ||
8575ec29 HD |
2730 | /* Flush pending updates to the LRU lists */ |
2731 | lru_add_drain_all(); | |
2732 | ||
76ab0f53 MG |
2733 | for_each_online_node(nid) |
2734 | compact_node(nid); | |
76ab0f53 MG |
2735 | } |
2736 | ||
facdaa91 NG |
2737 | /* |
2738 | * Tunable for proactive compaction. It determines how | |
2739 | * aggressively the kernel should compact memory in the | |
2740 | * background. It takes values in the range [0, 100]. | |
2741 | */ | |
d34c0a75 | 2742 | unsigned int __read_mostly sysctl_compaction_proactiveness = 20; |
facdaa91 | 2743 | |
65d759c8 CTR |
2744 | int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, |
2745 | void *buffer, size_t *length, loff_t *ppos) | |
2746 | { | |
2747 | int rc, nid; | |
2748 | ||
2749 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
2750 | if (rc) | |
2751 | return rc; | |
2752 | ||
2753 | if (write && sysctl_compaction_proactiveness) { | |
2754 | for_each_online_node(nid) { | |
2755 | pg_data_t *pgdat = NODE_DATA(nid); | |
2756 | ||
2757 | if (pgdat->proactive_compact_trigger) | |
2758 | continue; | |
2759 | ||
2760 | pgdat->proactive_compact_trigger = true; | |
2761 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2762 | } | |
2763 | } | |
2764 | ||
2765 | return 0; | |
2766 | } | |
2767 | ||
fec4eb2c YB |
2768 | /* |
2769 | * This is the entry point for compacting all nodes via | |
2770 | * /proc/sys/vm/compact_memory | |
2771 | */ | |
76ab0f53 | 2772 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
32927393 | 2773 | void *buffer, size_t *length, loff_t *ppos) |
76ab0f53 MG |
2774 | { |
2775 | if (write) | |
7964c06d | 2776 | compact_nodes(); |
76ab0f53 MG |
2777 | |
2778 | return 0; | |
2779 | } | |
ed4a6d7f MG |
2780 | |
2781 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
17adb230 Y |
2782 | static ssize_t compact_store(struct device *dev, |
2783 | struct device_attribute *attr, | |
2784 | const char *buf, size_t count) | |
ed4a6d7f | 2785 | { |
8575ec29 HD |
2786 | int nid = dev->id; |
2787 | ||
2788 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2789 | /* Flush pending updates to the LRU lists */ | |
2790 | lru_add_drain_all(); | |
2791 | ||
2792 | compact_node(nid); | |
2793 | } | |
ed4a6d7f MG |
2794 | |
2795 | return count; | |
2796 | } | |
17adb230 | 2797 | static DEVICE_ATTR_WO(compact); |
ed4a6d7f MG |
2798 | |
2799 | int compaction_register_node(struct node *node) | |
2800 | { | |
10fbcf4c | 2801 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2802 | } |
2803 | ||
2804 | void compaction_unregister_node(struct node *node) | |
2805 | { | |
10fbcf4c | 2806 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2807 | } |
2808 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2809 | |
698b1b30 VB |
2810 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2811 | { | |
65d759c8 CTR |
2812 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || |
2813 | pgdat->proactive_compact_trigger; | |
698b1b30 VB |
2814 | } |
2815 | ||
2816 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2817 | { | |
2818 | int zoneid; | |
2819 | struct zone *zone; | |
97a225e6 | 2820 | enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; |
698b1b30 | 2821 | |
97a225e6 | 2822 | for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { |
698b1b30 VB |
2823 | zone = &pgdat->node_zones[zoneid]; |
2824 | ||
2825 | if (!populated_zone(zone)) | |
2826 | continue; | |
2827 | ||
2828 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
97a225e6 | 2829 | highest_zoneidx) == COMPACT_CONTINUE) |
698b1b30 VB |
2830 | return true; |
2831 | } | |
2832 | ||
2833 | return false; | |
2834 | } | |
2835 | ||
2836 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2837 | { | |
2838 | /* | |
2839 | * With no special task, compact all zones so that a page of requested | |
2840 | * order is allocatable. | |
2841 | */ | |
2842 | int zoneid; | |
2843 | struct zone *zone; | |
2844 | struct compact_control cc = { | |
2845 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 2846 | .search_order = pgdat->kcompactd_max_order, |
97a225e6 | 2847 | .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, |
698b1b30 | 2848 | .mode = MIGRATE_SYNC_LIGHT, |
a0647dc9 | 2849 | .ignore_skip_hint = false, |
73e64c51 | 2850 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2851 | }; |
698b1b30 | 2852 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
97a225e6 | 2853 | cc.highest_zoneidx); |
7f354a54 | 2854 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2855 | |
97a225e6 | 2856 | for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { |
698b1b30 VB |
2857 | int status; |
2858 | ||
2859 | zone = &pgdat->node_zones[zoneid]; | |
2860 | if (!populated_zone(zone)) | |
2861 | continue; | |
2862 | ||
2863 | if (compaction_deferred(zone, cc.order)) | |
2864 | continue; | |
2865 | ||
2866 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
2867 | COMPACT_CONTINUE) | |
2868 | continue; | |
2869 | ||
172400c6 VB |
2870 | if (kthread_should_stop()) |
2871 | return; | |
a94b5252 YS |
2872 | |
2873 | cc.zone = zone; | |
5e1f0f09 | 2874 | status = compact_zone(&cc, NULL); |
698b1b30 | 2875 | |
7ceb009a | 2876 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 2877 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 2878 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
2879 | /* |
2880 | * Buddy pages may become stranded on pcps that could | |
2881 | * otherwise coalesce on the zone's free area for | |
2882 | * order >= cc.order. This is ratelimited by the | |
2883 | * upcoming deferral. | |
2884 | */ | |
2885 | drain_all_pages(zone); | |
2886 | ||
698b1b30 VB |
2887 | /* |
2888 | * We use sync migration mode here, so we defer like | |
2889 | * sync direct compaction does. | |
2890 | */ | |
2891 | defer_compaction(zone, cc.order); | |
2892 | } | |
2893 | ||
7f354a54 DR |
2894 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2895 | cc.total_migrate_scanned); | |
2896 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2897 | cc.total_free_scanned); | |
2898 | ||
698b1b30 VB |
2899 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2900 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2901 | } | |
2902 | ||
2903 | /* | |
2904 | * Regardless of success, we are done until woken up next. But remember | |
97a225e6 JK |
2905 | * the requested order/highest_zoneidx in case it was higher/tighter |
2906 | * than our current ones | |
698b1b30 VB |
2907 | */ |
2908 | if (pgdat->kcompactd_max_order <= cc.order) | |
2909 | pgdat->kcompactd_max_order = 0; | |
97a225e6 JK |
2910 | if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) |
2911 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; | |
698b1b30 VB |
2912 | } |
2913 | ||
97a225e6 | 2914 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) |
698b1b30 VB |
2915 | { |
2916 | if (!order) | |
2917 | return; | |
2918 | ||
2919 | if (pgdat->kcompactd_max_order < order) | |
2920 | pgdat->kcompactd_max_order = order; | |
2921 | ||
97a225e6 JK |
2922 | if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) |
2923 | pgdat->kcompactd_highest_zoneidx = highest_zoneidx; | |
698b1b30 | 2924 | |
6818600f DB |
2925 | /* |
2926 | * Pairs with implicit barrier in wait_event_freezable() | |
2927 | * such that wakeups are not missed. | |
2928 | */ | |
2929 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2930 | return; |
2931 | ||
2932 | if (!kcompactd_node_suitable(pgdat)) | |
2933 | return; | |
2934 | ||
2935 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
97a225e6 | 2936 | highest_zoneidx); |
698b1b30 VB |
2937 | wake_up_interruptible(&pgdat->kcompactd_wait); |
2938 | } | |
2939 | ||
2940 | /* | |
2941 | * The background compaction daemon, started as a kernel thread | |
2942 | * from the init process. | |
2943 | */ | |
2944 | static int kcompactd(void *p) | |
2945 | { | |
68d68ff6 | 2946 | pg_data_t *pgdat = (pg_data_t *)p; |
698b1b30 | 2947 | struct task_struct *tsk = current; |
e1e92bfa CTR |
2948 | long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); |
2949 | long timeout = default_timeout; | |
698b1b30 VB |
2950 | |
2951 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2952 | ||
2953 | if (!cpumask_empty(cpumask)) | |
2954 | set_cpus_allowed_ptr(tsk, cpumask); | |
2955 | ||
2956 | set_freezable(); | |
2957 | ||
2958 | pgdat->kcompactd_max_order = 0; | |
97a225e6 | 2959 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; |
698b1b30 VB |
2960 | |
2961 | while (!kthread_should_stop()) { | |
eb414681 JW |
2962 | unsigned long pflags; |
2963 | ||
65d759c8 CTR |
2964 | /* |
2965 | * Avoid the unnecessary wakeup for proactive compaction | |
2966 | * when it is disabled. | |
2967 | */ | |
2968 | if (!sysctl_compaction_proactiveness) | |
2969 | timeout = MAX_SCHEDULE_TIMEOUT; | |
698b1b30 | 2970 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
facdaa91 | 2971 | if (wait_event_freezable_timeout(pgdat->kcompactd_wait, |
65d759c8 CTR |
2972 | kcompactd_work_requested(pgdat), timeout) && |
2973 | !pgdat->proactive_compact_trigger) { | |
facdaa91 NG |
2974 | |
2975 | psi_memstall_enter(&pflags); | |
2976 | kcompactd_do_work(pgdat); | |
2977 | psi_memstall_leave(&pflags); | |
e1e92bfa CTR |
2978 | /* |
2979 | * Reset the timeout value. The defer timeout from | |
2980 | * proactive compaction is lost here but that is fine | |
2981 | * as the condition of the zone changing substantionally | |
2982 | * then carrying on with the previous defer interval is | |
2983 | * not useful. | |
2984 | */ | |
2985 | timeout = default_timeout; | |
facdaa91 NG |
2986 | continue; |
2987 | } | |
698b1b30 | 2988 | |
e1e92bfa CTR |
2989 | /* |
2990 | * Start the proactive work with default timeout. Based | |
2991 | * on the fragmentation score, this timeout is updated. | |
2992 | */ | |
2993 | timeout = default_timeout; | |
facdaa91 NG |
2994 | if (should_proactive_compact_node(pgdat)) { |
2995 | unsigned int prev_score, score; | |
2996 | ||
facdaa91 NG |
2997 | prev_score = fragmentation_score_node(pgdat); |
2998 | proactive_compact_node(pgdat); | |
2999 | score = fragmentation_score_node(pgdat); | |
3000 | /* | |
3001 | * Defer proactive compaction if the fragmentation | |
3002 | * score did not go down i.e. no progress made. | |
3003 | */ | |
e1e92bfa CTR |
3004 | if (unlikely(score >= prev_score)) |
3005 | timeout = | |
3006 | default_timeout << COMPACT_MAX_DEFER_SHIFT; | |
facdaa91 | 3007 | } |
65d759c8 CTR |
3008 | if (unlikely(pgdat->proactive_compact_trigger)) |
3009 | pgdat->proactive_compact_trigger = false; | |
698b1b30 VB |
3010 | } |
3011 | ||
3012 | return 0; | |
3013 | } | |
3014 | ||
3015 | /* | |
3016 | * This kcompactd start function will be called by init and node-hot-add. | |
3017 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
3018 | */ | |
3019 | int kcompactd_run(int nid) | |
3020 | { | |
3021 | pg_data_t *pgdat = NODE_DATA(nid); | |
3022 | int ret = 0; | |
3023 | ||
3024 | if (pgdat->kcompactd) | |
3025 | return 0; | |
3026 | ||
3027 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
3028 | if (IS_ERR(pgdat->kcompactd)) { | |
3029 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
3030 | ret = PTR_ERR(pgdat->kcompactd); | |
3031 | pgdat->kcompactd = NULL; | |
3032 | } | |
3033 | return ret; | |
3034 | } | |
3035 | ||
3036 | /* | |
3037 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
3038 | * hold mem_hotplug_begin/end(). | |
3039 | */ | |
3040 | void kcompactd_stop(int nid) | |
3041 | { | |
3042 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
3043 | ||
3044 | if (kcompactd) { | |
3045 | kthread_stop(kcompactd); | |
3046 | NODE_DATA(nid)->kcompactd = NULL; | |
3047 | } | |
3048 | } | |
3049 | ||
3050 | /* | |
3051 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
3052 | * not required for correctness. So if the last cpu in a node goes | |
3053 | * away, we get changed to run anywhere: as the first one comes back, | |
3054 | * restore their cpu bindings. | |
3055 | */ | |
e46b1db2 | 3056 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
3057 | { |
3058 | int nid; | |
3059 | ||
e46b1db2 AMG |
3060 | for_each_node_state(nid, N_MEMORY) { |
3061 | pg_data_t *pgdat = NODE_DATA(nid); | |
3062 | const struct cpumask *mask; | |
698b1b30 | 3063 | |
e46b1db2 | 3064 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 3065 | |
e46b1db2 AMG |
3066 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3067 | /* One of our CPUs online: restore mask */ | |
3068 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 3069 | } |
e46b1db2 | 3070 | return 0; |
698b1b30 VB |
3071 | } |
3072 | ||
3073 | static int __init kcompactd_init(void) | |
3074 | { | |
3075 | int nid; | |
e46b1db2 AMG |
3076 | int ret; |
3077 | ||
3078 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
3079 | "mm/compaction:online", | |
3080 | kcompactd_cpu_online, NULL); | |
3081 | if (ret < 0) { | |
3082 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
3083 | return ret; | |
3084 | } | |
698b1b30 VB |
3085 | |
3086 | for_each_node_state(nid, N_MEMORY) | |
3087 | kcompactd_run(nid); | |
698b1b30 VB |
3088 | return 0; |
3089 | } | |
3090 | subsys_initcall(kcompactd_init) | |
3091 | ||
ff9543fd | 3092 | #endif /* CONFIG_COMPACTION */ |