Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
748446bb MG |
2 | /* |
3 | * linux/mm/compaction.c | |
4 | * | |
5 | * Memory compaction for the reduction of external fragmentation. Note that | |
6 | * this heavily depends upon page migration to do all the real heavy | |
7 | * lifting | |
8 | * | |
9 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
10 | */ | |
698b1b30 | 11 | #include <linux/cpu.h> |
748446bb MG |
12 | #include <linux/swap.h> |
13 | #include <linux/migrate.h> | |
14 | #include <linux/compaction.h> | |
15 | #include <linux/mm_inline.h> | |
174cd4b1 | 16 | #include <linux/sched/signal.h> |
748446bb | 17 | #include <linux/backing-dev.h> |
76ab0f53 | 18 | #include <linux/sysctl.h> |
ed4a6d7f | 19 | #include <linux/sysfs.h> |
194159fb | 20 | #include <linux/page-isolation.h> |
b8c73fc2 | 21 | #include <linux/kasan.h> |
698b1b30 VB |
22 | #include <linux/kthread.h> |
23 | #include <linux/freezer.h> | |
83358ece | 24 | #include <linux/page_owner.h> |
eb414681 | 25 | #include <linux/psi.h> |
748446bb MG |
26 | #include "internal.h" |
27 | ||
010fc29a | 28 | #ifdef CONFIG_COMPACTION |
31ca72fa CTK |
29 | /* |
30 | * Fragmentation score check interval for proactive compaction purposes. | |
31 | */ | |
32 | #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500) | |
33 | ||
010fc29a MK |
34 | static inline void count_compact_event(enum vm_event_item item) |
35 | { | |
36 | count_vm_event(item); | |
37 | } | |
38 | ||
39 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
40 | { | |
41 | count_vm_events(item, delta); | |
42 | } | |
43 | #else | |
44 | #define count_compact_event(item) do { } while (0) | |
45 | #define count_compact_events(item, delta) do { } while (0) | |
46 | #endif | |
47 | ||
ff9543fd MN |
48 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
49 | ||
b7aba698 MG |
50 | #define CREATE_TRACE_POINTS |
51 | #include <trace/events/compaction.h> | |
52 | ||
06b6640a VB |
53 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
54 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
06b6640a | 55 | |
facdaa91 NG |
56 | /* |
57 | * Page order with-respect-to which proactive compaction | |
58 | * calculates external fragmentation, which is used as | |
59 | * the "fragmentation score" of a node/zone. | |
60 | */ | |
61 | #if defined CONFIG_TRANSPARENT_HUGEPAGE | |
62 | #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER | |
25788738 | 63 | #elif defined CONFIG_HUGETLBFS |
facdaa91 NG |
64 | #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER |
65 | #else | |
66 | #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
67 | #endif | |
68 | ||
748446bb MG |
69 | static unsigned long release_freepages(struct list_head *freelist) |
70 | { | |
71 | struct page *page, *next; | |
6bace090 | 72 | unsigned long high_pfn = 0; |
748446bb MG |
73 | |
74 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 75 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
76 | list_del(&page->lru); |
77 | __free_page(page); | |
6bace090 VB |
78 | if (pfn > high_pfn) |
79 | high_pfn = pfn; | |
748446bb MG |
80 | } |
81 | ||
6bace090 | 82 | return high_pfn; |
748446bb MG |
83 | } |
84 | ||
4469ab98 | 85 | static void split_map_pages(struct list_head *list) |
ff9543fd | 86 | { |
66c64223 JK |
87 | unsigned int i, order, nr_pages; |
88 | struct page *page, *next; | |
89 | LIST_HEAD(tmp_list); | |
90 | ||
91 | list_for_each_entry_safe(page, next, list, lru) { | |
92 | list_del(&page->lru); | |
93 | ||
94 | order = page_private(page); | |
95 | nr_pages = 1 << order; | |
66c64223 | 96 | |
46f24fd8 | 97 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
98 | if (order) |
99 | split_page(page, order); | |
ff9543fd | 100 | |
66c64223 JK |
101 | for (i = 0; i < nr_pages; i++) { |
102 | list_add(&page->lru, &tmp_list); | |
103 | page++; | |
104 | } | |
ff9543fd | 105 | } |
66c64223 JK |
106 | |
107 | list_splice(&tmp_list, list); | |
ff9543fd MN |
108 | } |
109 | ||
bb13ffeb | 110 | #ifdef CONFIG_COMPACTION |
68f2736a | 111 | bool PageMovable(struct page *page) |
bda807d4 | 112 | { |
68f2736a | 113 | const struct movable_operations *mops; |
bda807d4 MK |
114 | |
115 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
116 | if (!__PageMovable(page)) | |
68f2736a | 117 | return false; |
bda807d4 | 118 | |
68f2736a MWO |
119 | mops = page_movable_ops(page); |
120 | if (mops) | |
121 | return true; | |
bda807d4 | 122 | |
68f2736a | 123 | return false; |
bda807d4 | 124 | } |
bda807d4 | 125 | |
68f2736a | 126 | void __SetPageMovable(struct page *page, const struct movable_operations *mops) |
bda807d4 MK |
127 | { |
128 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
68f2736a MWO |
129 | VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page); |
130 | page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE); | |
bda807d4 MK |
131 | } |
132 | EXPORT_SYMBOL(__SetPageMovable); | |
133 | ||
134 | void __ClearPageMovable(struct page *page) | |
135 | { | |
bda807d4 MK |
136 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
137 | /* | |
68f2736a MWO |
138 | * This page still has the type of a movable page, but it's |
139 | * actually not movable any more. | |
bda807d4 | 140 | */ |
68f2736a | 141 | page->mapping = (void *)PAGE_MAPPING_MOVABLE; |
bda807d4 MK |
142 | } |
143 | EXPORT_SYMBOL(__ClearPageMovable); | |
144 | ||
24e2716f JK |
145 | /* Do not skip compaction more than 64 times */ |
146 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
147 | ||
148 | /* | |
149 | * Compaction is deferred when compaction fails to result in a page | |
860b3272 | 150 | * allocation success. 1 << compact_defer_shift, compactions are skipped up |
24e2716f JK |
151 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT |
152 | */ | |
2271b016 | 153 | static void defer_compaction(struct zone *zone, int order) |
24e2716f JK |
154 | { |
155 | zone->compact_considered = 0; | |
156 | zone->compact_defer_shift++; | |
157 | ||
158 | if (order < zone->compact_order_failed) | |
159 | zone->compact_order_failed = order; | |
160 | ||
161 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
162 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
163 | ||
164 | trace_mm_compaction_defer_compaction(zone, order); | |
165 | } | |
166 | ||
167 | /* Returns true if compaction should be skipped this time */ | |
2271b016 | 168 | static bool compaction_deferred(struct zone *zone, int order) |
24e2716f JK |
169 | { |
170 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
171 | ||
172 | if (order < zone->compact_order_failed) | |
173 | return false; | |
174 | ||
175 | /* Avoid possible overflow */ | |
62b35fe0 | 176 | if (++zone->compact_considered >= defer_limit) { |
24e2716f | 177 | zone->compact_considered = defer_limit; |
24e2716f | 178 | return false; |
62b35fe0 | 179 | } |
24e2716f JK |
180 | |
181 | trace_mm_compaction_deferred(zone, order); | |
182 | ||
183 | return true; | |
184 | } | |
185 | ||
186 | /* | |
187 | * Update defer tracking counters after successful compaction of given order, | |
188 | * which means an allocation either succeeded (alloc_success == true) or is | |
189 | * expected to succeed. | |
190 | */ | |
191 | void compaction_defer_reset(struct zone *zone, int order, | |
192 | bool alloc_success) | |
193 | { | |
194 | if (alloc_success) { | |
195 | zone->compact_considered = 0; | |
196 | zone->compact_defer_shift = 0; | |
197 | } | |
198 | if (order >= zone->compact_order_failed) | |
199 | zone->compact_order_failed = order + 1; | |
200 | ||
201 | trace_mm_compaction_defer_reset(zone, order); | |
202 | } | |
203 | ||
204 | /* Returns true if restarting compaction after many failures */ | |
2271b016 | 205 | static bool compaction_restarting(struct zone *zone, int order) |
24e2716f JK |
206 | { |
207 | if (order < zone->compact_order_failed) | |
208 | return false; | |
209 | ||
210 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
211 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
212 | } | |
213 | ||
bb13ffeb MG |
214 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
215 | static inline bool isolation_suitable(struct compact_control *cc, | |
216 | struct page *page) | |
217 | { | |
218 | if (cc->ignore_skip_hint) | |
219 | return true; | |
220 | ||
221 | return !get_pageblock_skip(page); | |
222 | } | |
223 | ||
02333641 VB |
224 | static void reset_cached_positions(struct zone *zone) |
225 | { | |
226 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
227 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 228 | zone->compact_cached_free_pfn = |
06b6640a | 229 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
230 | } |
231 | ||
9721fd82 BW |
232 | #ifdef CONFIG_SPARSEMEM |
233 | /* | |
234 | * If the PFN falls into an offline section, return the start PFN of the | |
235 | * next online section. If the PFN falls into an online section or if | |
236 | * there is no next online section, return 0. | |
237 | */ | |
238 | static unsigned long skip_offline_sections(unsigned long start_pfn) | |
239 | { | |
240 | unsigned long start_nr = pfn_to_section_nr(start_pfn); | |
241 | ||
242 | if (online_section_nr(start_nr)) | |
243 | return 0; | |
244 | ||
245 | while (++start_nr <= __highest_present_section_nr) { | |
246 | if (online_section_nr(start_nr)) | |
247 | return section_nr_to_pfn(start_nr); | |
248 | } | |
249 | ||
250 | return 0; | |
251 | } | |
e6e0c767 BW |
252 | |
253 | /* | |
254 | * If the PFN falls into an offline section, return the end PFN of the | |
255 | * next online section in reverse. If the PFN falls into an online section | |
256 | * or if there is no next online section in reverse, return 0. | |
257 | */ | |
258 | static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) | |
259 | { | |
260 | unsigned long start_nr = pfn_to_section_nr(start_pfn); | |
261 | ||
262 | if (!start_nr || online_section_nr(start_nr)) | |
263 | return 0; | |
264 | ||
265 | while (start_nr-- > 0) { | |
266 | if (online_section_nr(start_nr)) | |
267 | return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION; | |
268 | } | |
269 | ||
270 | return 0; | |
271 | } | |
9721fd82 BW |
272 | #else |
273 | static unsigned long skip_offline_sections(unsigned long start_pfn) | |
274 | { | |
275 | return 0; | |
276 | } | |
e6e0c767 BW |
277 | |
278 | static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) | |
279 | { | |
280 | return 0; | |
281 | } | |
9721fd82 BW |
282 | #endif |
283 | ||
21dc7e02 | 284 | /* |
2271b016 | 285 | * Compound pages of >= pageblock_order should consistently be skipped until |
b527cfe5 VB |
286 | * released. It is always pointless to compact pages of such order (if they are |
287 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 288 | */ |
b527cfe5 | 289 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 290 | { |
b527cfe5 | 291 | if (!PageCompound(page)) |
21dc7e02 | 292 | return false; |
b527cfe5 VB |
293 | |
294 | page = compound_head(page); | |
295 | ||
296 | if (compound_order(page) >= pageblock_order) | |
297 | return true; | |
298 | ||
299 | return false; | |
21dc7e02 DR |
300 | } |
301 | ||
e332f741 MG |
302 | static bool |
303 | __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, | |
304 | bool check_target) | |
305 | { | |
306 | struct page *page = pfn_to_online_page(pfn); | |
6b0868c8 | 307 | struct page *block_page; |
e332f741 MG |
308 | struct page *end_page; |
309 | unsigned long block_pfn; | |
310 | ||
311 | if (!page) | |
312 | return false; | |
313 | if (zone != page_zone(page)) | |
314 | return false; | |
315 | if (pageblock_skip_persistent(page)) | |
316 | return false; | |
317 | ||
318 | /* | |
319 | * If skip is already cleared do no further checking once the | |
320 | * restart points have been set. | |
321 | */ | |
322 | if (check_source && check_target && !get_pageblock_skip(page)) | |
323 | return true; | |
324 | ||
325 | /* | |
326 | * If clearing skip for the target scanner, do not select a | |
327 | * non-movable pageblock as the starting point. | |
328 | */ | |
329 | if (!check_source && check_target && | |
330 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
331 | return false; | |
332 | ||
6b0868c8 MG |
333 | /* Ensure the start of the pageblock or zone is online and valid */ |
334 | block_pfn = pageblock_start_pfn(pfn); | |
a2e9a5af VB |
335 | block_pfn = max(block_pfn, zone->zone_start_pfn); |
336 | block_page = pfn_to_online_page(block_pfn); | |
6b0868c8 MG |
337 | if (block_page) { |
338 | page = block_page; | |
339 | pfn = block_pfn; | |
340 | } | |
341 | ||
342 | /* Ensure the end of the pageblock or zone is online and valid */ | |
a2e9a5af | 343 | block_pfn = pageblock_end_pfn(pfn) - 1; |
6b0868c8 MG |
344 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); |
345 | end_page = pfn_to_online_page(block_pfn); | |
346 | if (!end_page) | |
347 | return false; | |
348 | ||
e332f741 MG |
349 | /* |
350 | * Only clear the hint if a sample indicates there is either a | |
351 | * free page or an LRU page in the block. One or other condition | |
352 | * is necessary for the block to be a migration source/target. | |
353 | */ | |
e332f741 | 354 | do { |
859a85dd MR |
355 | if (check_source && PageLRU(page)) { |
356 | clear_pageblock_skip(page); | |
357 | return true; | |
358 | } | |
e332f741 | 359 | |
859a85dd MR |
360 | if (check_target && PageBuddy(page)) { |
361 | clear_pageblock_skip(page); | |
362 | return true; | |
e332f741 MG |
363 | } |
364 | ||
365 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
a2e9a5af | 366 | } while (page <= end_page); |
e332f741 MG |
367 | |
368 | return false; | |
369 | } | |
370 | ||
bb13ffeb MG |
371 | /* |
372 | * This function is called to clear all cached information on pageblocks that | |
373 | * should be skipped for page isolation when the migrate and free page scanner | |
374 | * meet. | |
375 | */ | |
62997027 | 376 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 377 | { |
e332f741 | 378 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 379 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
380 | unsigned long reset_migrate = free_pfn; |
381 | unsigned long reset_free = migrate_pfn; | |
382 | bool source_set = false; | |
383 | bool free_set = false; | |
384 | ||
8df4e28c | 385 | /* Only flush if a full compaction finished recently */ |
e332f741 MG |
386 | if (!zone->compact_blockskip_flush) |
387 | return; | |
bb13ffeb | 388 | |
62997027 | 389 | zone->compact_blockskip_flush = false; |
bb13ffeb | 390 | |
e332f741 MG |
391 | /* |
392 | * Walk the zone and update pageblock skip information. Source looks | |
393 | * for PageLRU while target looks for PageBuddy. When the scanner | |
394 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
395 | * is suitable as both source and target. | |
396 | */ | |
397 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
398 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
399 | cond_resched(); |
400 | ||
e332f741 MG |
401 | /* Update the migrate PFN */ |
402 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
403 | migrate_pfn < reset_migrate) { | |
404 | source_set = true; | |
405 | reset_migrate = migrate_pfn; | |
406 | zone->compact_init_migrate_pfn = reset_migrate; | |
407 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
408 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
409 | } | |
bb13ffeb | 410 | |
e332f741 MG |
411 | /* Update the free PFN */ |
412 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
413 | free_pfn > reset_free) { | |
414 | free_set = true; | |
415 | reset_free = free_pfn; | |
416 | zone->compact_init_free_pfn = reset_free; | |
417 | zone->compact_cached_free_pfn = reset_free; | |
418 | } | |
bb13ffeb | 419 | } |
02333641 | 420 | |
e332f741 MG |
421 | /* Leave no distance if no suitable block was reset */ |
422 | if (reset_migrate >= reset_free) { | |
423 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
424 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
425 | zone->compact_cached_free_pfn = free_pfn; | |
426 | } | |
bb13ffeb MG |
427 | } |
428 | ||
62997027 MG |
429 | void reset_isolation_suitable(pg_data_t *pgdat) |
430 | { | |
431 | int zoneid; | |
432 | ||
433 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
434 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
435 | if (!populated_zone(zone)) | |
436 | continue; | |
437 | ||
8df4e28c | 438 | __reset_isolation_suitable(zone); |
62997027 MG |
439 | } |
440 | } | |
441 | ||
e380bebe MG |
442 | /* |
443 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
444 | * locks are not required for read/writers. Returns true if it was already set. | |
445 | */ | |
590ccea8 | 446 | static bool test_and_set_skip(struct compact_control *cc, struct page *page) |
e380bebe MG |
447 | { |
448 | bool skip; | |
449 | ||
590ccea8 | 450 | /* Do not update if skip hint is being ignored */ |
e380bebe MG |
451 | if (cc->ignore_skip_hint) |
452 | return false; | |
453 | ||
e380bebe MG |
454 | skip = get_pageblock_skip(page); |
455 | if (!skip && !cc->no_set_skip_hint) | |
456 | set_pageblock_skip(page); | |
457 | ||
458 | return skip; | |
459 | } | |
460 | ||
461 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
462 | { | |
463 | struct zone *zone = cc->zone; | |
464 | ||
e380bebe MG |
465 | /* Set for isolation rather than compaction */ |
466 | if (cc->no_set_skip_hint) | |
467 | return; | |
468 | ||
3c099a2b KS |
469 | pfn = pageblock_end_pfn(pfn); |
470 | ||
cf043a00 | 471 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
472 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
473 | zone->compact_cached_migrate_pfn[0] = pfn; | |
474 | if (cc->mode != MIGRATE_ASYNC && | |
475 | pfn > zone->compact_cached_migrate_pfn[1]) | |
476 | zone->compact_cached_migrate_pfn[1] = pfn; | |
477 | } | |
478 | ||
bb13ffeb MG |
479 | /* |
480 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 481 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 482 | */ |
c89511ab | 483 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 484 | struct page *page, unsigned long pfn) |
bb13ffeb | 485 | { |
c89511ab | 486 | struct zone *zone = cc->zone; |
6815bf3f | 487 | |
2583d671 | 488 | if (cc->no_set_skip_hint) |
6815bf3f JK |
489 | return; |
490 | ||
edc2ca61 | 491 | set_pageblock_skip(page); |
c89511ab | 492 | |
e380bebe MG |
493 | if (pfn < zone->compact_cached_free_pfn) |
494 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
495 | } |
496 | #else | |
497 | static inline bool isolation_suitable(struct compact_control *cc, | |
498 | struct page *page) | |
499 | { | |
500 | return true; | |
501 | } | |
502 | ||
b527cfe5 | 503 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
504 | { |
505 | return false; | |
506 | } | |
507 | ||
508 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 509 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
510 | { |
511 | } | |
e380bebe MG |
512 | |
513 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
514 | { | |
515 | } | |
516 | ||
590ccea8 | 517 | static bool test_and_set_skip(struct compact_control *cc, struct page *page) |
e380bebe MG |
518 | { |
519 | return false; | |
520 | } | |
bb13ffeb MG |
521 | #endif /* CONFIG_COMPACTION */ |
522 | ||
8b44d279 VB |
523 | /* |
524 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
525 | * very heavily contended. For async compaction, trylock and record if the |
526 | * lock is contended. The lock will still be acquired but compaction will | |
527 | * abort when the current block is finished regardless of success rate. | |
528 | * Sync compaction acquires the lock. | |
8b44d279 | 529 | * |
cb2dcaf0 | 530 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 531 | */ |
cb2dcaf0 | 532 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 533 | struct compact_control *cc) |
77337ede | 534 | __acquires(lock) |
2a1402aa | 535 | { |
cb2dcaf0 MG |
536 | /* Track if the lock is contended in async mode */ |
537 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
538 | if (spin_trylock_irqsave(lock, *flags)) | |
539 | return true; | |
540 | ||
541 | cc->contended = true; | |
8b44d279 | 542 | } |
1f9efdef | 543 | |
cb2dcaf0 | 544 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 545 | return true; |
2a1402aa MG |
546 | } |
547 | ||
c67fe375 MG |
548 | /* |
549 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
550 | * very heavily contended. The lock should be periodically unlocked to avoid |
551 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
552 | * the lock. It might also be that allowing the IRQs will result in | |
d56c1584 | 553 | * need_resched() becoming true. If scheduling is needed, compaction schedules. |
8b44d279 VB |
554 | * Either compaction type will also abort if a fatal signal is pending. |
555 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 556 | * |
d56c1584 ML |
557 | * Returns true if compaction should abort due to fatal signal pending. |
558 | * Returns false when compaction can continue. | |
c67fe375 | 559 | */ |
8b44d279 VB |
560 | static bool compact_unlock_should_abort(spinlock_t *lock, |
561 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 562 | { |
8b44d279 VB |
563 | if (*locked) { |
564 | spin_unlock_irqrestore(lock, flags); | |
565 | *locked = false; | |
566 | } | |
1f9efdef | 567 | |
8b44d279 | 568 | if (fatal_signal_pending(current)) { |
c3486f53 | 569 | cc->contended = true; |
8b44d279 VB |
570 | return true; |
571 | } | |
c67fe375 | 572 | |
cf66f070 | 573 | cond_resched(); |
be976572 VB |
574 | |
575 | return false; | |
576 | } | |
577 | ||
85aa125f | 578 | /* |
9e4be470 JM |
579 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
580 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
581 | * (even though it may still end up isolating some pages). | |
85aa125f | 582 | */ |
f40d1e42 | 583 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 584 | unsigned long *start_pfn, |
85aa125f MN |
585 | unsigned long end_pfn, |
586 | struct list_head *freelist, | |
4fca9730 | 587 | unsigned int stride, |
85aa125f | 588 | bool strict) |
748446bb | 589 | { |
b7aba698 | 590 | int nr_scanned = 0, total_isolated = 0; |
dc13292c | 591 | struct page *page; |
b8b2d825 | 592 | unsigned long flags = 0; |
f40d1e42 | 593 | bool locked = false; |
e14c720e | 594 | unsigned long blockpfn = *start_pfn; |
66c64223 | 595 | unsigned int order; |
748446bb | 596 | |
4fca9730 MG |
597 | /* Strict mode is for isolation, speed is secondary */ |
598 | if (strict) | |
599 | stride = 1; | |
600 | ||
dc13292c | 601 | page = pfn_to_page(blockpfn); |
748446bb | 602 | |
f40d1e42 | 603 | /* Isolate free pages. */ |
dc13292c | 604 | for (; blockpfn < end_pfn; blockpfn += stride, page += stride) { |
66c64223 | 605 | int isolated; |
748446bb | 606 | |
8b44d279 VB |
607 | /* |
608 | * Periodically drop the lock (if held) regardless of its | |
609 | * contention, to give chance to IRQs. Abort if fatal signal | |
d56c1584 | 610 | * pending. |
8b44d279 | 611 | */ |
c036ddff | 612 | if (!(blockpfn % COMPACT_CLUSTER_MAX) |
8b44d279 VB |
613 | && compact_unlock_should_abort(&cc->zone->lock, flags, |
614 | &locked, cc)) | |
615 | break; | |
616 | ||
b7aba698 | 617 | nr_scanned++; |
2af120bc | 618 | |
9fcd6d2e VB |
619 | /* |
620 | * For compound pages such as THP and hugetlbfs, we can save | |
621 | * potentially a lot of iterations if we skip them at once. | |
622 | * The check is racy, but we can consider only valid values | |
623 | * and the only danger is skipping too much. | |
624 | */ | |
625 | if (PageCompound(page)) { | |
21dc7e02 DR |
626 | const unsigned int order = compound_order(page); |
627 | ||
3da0272a | 628 | if (blockpfn + (1UL << order) <= end_pfn) { |
21dc7e02 | 629 | blockpfn += (1UL << order) - 1; |
dc13292c | 630 | page += (1UL << order) - 1; |
56d48d8d | 631 | nr_scanned += (1UL << order) - 1; |
9fcd6d2e | 632 | } |
3da0272a | 633 | |
9fcd6d2e VB |
634 | goto isolate_fail; |
635 | } | |
636 | ||
f40d1e42 | 637 | if (!PageBuddy(page)) |
2af120bc | 638 | goto isolate_fail; |
f40d1e42 | 639 | |
85f73e6d | 640 | /* If we already hold the lock, we can skip some rechecking. */ |
69b7189f | 641 | if (!locked) { |
cb2dcaf0 | 642 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 643 | &flags, cc); |
f40d1e42 | 644 | |
69b7189f VB |
645 | /* Recheck this is a buddy page under lock */ |
646 | if (!PageBuddy(page)) | |
647 | goto isolate_fail; | |
648 | } | |
748446bb | 649 | |
66c64223 | 650 | /* Found a free page, will break it into order-0 pages */ |
ab130f91 | 651 | order = buddy_order(page); |
66c64223 | 652 | isolated = __isolate_free_page(page, order); |
a4f04f2c DR |
653 | if (!isolated) |
654 | break; | |
66c64223 | 655 | set_page_private(page, order); |
a4f04f2c | 656 | |
b717d6b9 | 657 | nr_scanned += isolated - 1; |
748446bb | 658 | total_isolated += isolated; |
a4f04f2c | 659 | cc->nr_freepages += isolated; |
66c64223 JK |
660 | list_add_tail(&page->lru, freelist); |
661 | ||
a4f04f2c DR |
662 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
663 | blockpfn += isolated; | |
664 | break; | |
748446bb | 665 | } |
a4f04f2c DR |
666 | /* Advance to the end of split page */ |
667 | blockpfn += isolated - 1; | |
dc13292c | 668 | page += isolated - 1; |
a4f04f2c | 669 | continue; |
2af120bc LA |
670 | |
671 | isolate_fail: | |
672 | if (strict) | |
673 | break; | |
2af120bc | 674 | |
748446bb MG |
675 | } |
676 | ||
a4f04f2c DR |
677 | if (locked) |
678 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
679 | ||
9fcd6d2e | 680 | /* |
3da0272a | 681 | * Be careful to not go outside of the pageblock. |
9fcd6d2e VB |
682 | */ |
683 | if (unlikely(blockpfn > end_pfn)) | |
684 | blockpfn = end_pfn; | |
685 | ||
e34d85f0 JK |
686 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
687 | nr_scanned, total_isolated); | |
688 | ||
e14c720e VB |
689 | /* Record how far we have got within the block */ |
690 | *start_pfn = blockpfn; | |
691 | ||
f40d1e42 MG |
692 | /* |
693 | * If strict isolation is requested by CMA then check that all the | |
694 | * pages requested were isolated. If there were any failures, 0 is | |
695 | * returned and CMA will fail. | |
696 | */ | |
2af120bc | 697 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
698 | total_isolated = 0; |
699 | ||
7f354a54 | 700 | cc->total_free_scanned += nr_scanned; |
397487db | 701 | if (total_isolated) |
010fc29a | 702 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
703 | return total_isolated; |
704 | } | |
705 | ||
85aa125f MN |
706 | /** |
707 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 708 | * @cc: Compaction control structure. |
85aa125f MN |
709 | * @start_pfn: The first PFN to start isolating. |
710 | * @end_pfn: The one-past-last PFN. | |
711 | * | |
712 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
713 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
714 | * undo its actions and return zero. | |
715 | * | |
716 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
717 | * (which may be greater then end_pfn if end fell in a middle of | |
718 | * a free page). | |
719 | */ | |
ff9543fd | 720 | unsigned long |
bb13ffeb MG |
721 | isolate_freepages_range(struct compact_control *cc, |
722 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 723 | { |
e1409c32 | 724 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
725 | LIST_HEAD(freelist); |
726 | ||
7d49d886 | 727 | pfn = start_pfn; |
06b6640a | 728 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
729 | if (block_start_pfn < cc->zone->zone_start_pfn) |
730 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 731 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
732 | |
733 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 734 | block_start_pfn = block_end_pfn, |
7d49d886 | 735 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
736 | /* Protect pfn from changing by isolate_freepages_block */ |
737 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 738 | |
58420016 JK |
739 | /* |
740 | * pfn could pass the block_end_pfn if isolated freepage | |
741 | * is more than pageblock order. In this case, we adjust | |
742 | * scanning range to right one. | |
743 | */ | |
744 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
745 | block_start_pfn = pageblock_start_pfn(pfn); |
746 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
747 | } |
748 | ||
a2864a67 KS |
749 | block_end_pfn = min(block_end_pfn, end_pfn); |
750 | ||
e1409c32 JK |
751 | if (!pageblock_pfn_to_page(block_start_pfn, |
752 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
753 | break; |
754 | ||
e14c720e | 755 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
4fca9730 | 756 | block_end_pfn, &freelist, 0, true); |
85aa125f MN |
757 | |
758 | /* | |
759 | * In strict mode, isolate_freepages_block() returns 0 if | |
760 | * there are any holes in the block (ie. invalid PFNs or | |
761 | * non-free pages). | |
762 | */ | |
763 | if (!isolated) | |
764 | break; | |
765 | ||
766 | /* | |
767 | * If we managed to isolate pages, it is always (1 << n) * | |
768 | * pageblock_nr_pages for some non-negative n. (Max order | |
769 | * page may span two pageblocks). | |
770 | */ | |
771 | } | |
772 | ||
66c64223 | 773 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 774 | split_map_pages(&freelist); |
85aa125f MN |
775 | |
776 | if (pfn < end_pfn) { | |
777 | /* Loop terminated early, cleanup. */ | |
778 | release_freepages(&freelist); | |
779 | return 0; | |
780 | } | |
781 | ||
782 | /* We don't use freelists for anything. */ | |
783 | return pfn; | |
784 | } | |
785 | ||
748446bb | 786 | /* Similar to reclaim, but different enough that they don't share logic */ |
4fbbb3fd | 787 | static bool too_many_isolated(struct compact_control *cc) |
748446bb | 788 | { |
4fbbb3fd | 789 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
d818fca1 MG |
790 | bool too_many; |
791 | ||
bc693045 | 792 | unsigned long active, inactive, isolated; |
748446bb | 793 | |
5f438eee AR |
794 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + |
795 | node_page_state(pgdat, NR_INACTIVE_ANON); | |
796 | active = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
797 | node_page_state(pgdat, NR_ACTIVE_ANON); | |
798 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + | |
799 | node_page_state(pgdat, NR_ISOLATED_ANON); | |
748446bb | 800 | |
4fbbb3fd JW |
801 | /* |
802 | * Allow GFP_NOFS to isolate past the limit set for regular | |
803 | * compaction runs. This prevents an ABBA deadlock when other | |
804 | * compactors have already isolated to the limit, but are | |
805 | * blocked on filesystem locks held by the GFP_NOFS thread. | |
806 | */ | |
807 | if (cc->gfp_mask & __GFP_FS) { | |
808 | inactive >>= 3; | |
809 | active >>= 3; | |
810 | } | |
811 | ||
d818fca1 MG |
812 | too_many = isolated > (inactive + active) / 2; |
813 | if (!too_many) | |
814 | wake_throttle_isolated(pgdat); | |
815 | ||
816 | return too_many; | |
748446bb MG |
817 | } |
818 | ||
2fe86e00 | 819 | /** |
edc2ca61 VB |
820 | * isolate_migratepages_block() - isolate all migrate-able pages within |
821 | * a single pageblock | |
2fe86e00 | 822 | * @cc: Compaction control structure. |
edc2ca61 VB |
823 | * @low_pfn: The first PFN to isolate |
824 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
89f6c88a | 825 | * @mode: Isolation mode to be used. |
2fe86e00 MN |
826 | * |
827 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 | 828 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
c2ad7a1f | 829 | * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, |
369fa227 | 830 | * -ENOMEM in case we could not allocate a page, or 0. |
c2ad7a1f | 831 | * cc->migrate_pfn will contain the next pfn to scan. |
2fe86e00 | 832 | * |
edc2ca61 | 833 | * The pages are isolated on cc->migratepages list (not required to be empty), |
c2ad7a1f | 834 | * and cc->nr_migratepages is updated accordingly. |
748446bb | 835 | */ |
c2ad7a1f | 836 | static int |
edc2ca61 | 837 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, |
89f6c88a | 838 | unsigned long end_pfn, isolate_mode_t mode) |
748446bb | 839 | { |
5f438eee | 840 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 841 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 842 | struct lruvec *lruvec; |
b8b2d825 | 843 | unsigned long flags = 0; |
6168d0da | 844 | struct lruvec *locked = NULL; |
56ae0bb3 | 845 | struct folio *folio = NULL; |
bb13ffeb | 846 | struct page *page = NULL, *valid_page = NULL; |
89f6c88a | 847 | struct address_space *mapping; |
e34d85f0 | 848 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
849 | bool skip_on_failure = false; |
850 | unsigned long next_skip_pfn = 0; | |
e380bebe | 851 | bool skip_updated = false; |
c2ad7a1f OS |
852 | int ret = 0; |
853 | ||
854 | cc->migrate_pfn = low_pfn; | |
748446bb | 855 | |
748446bb MG |
856 | /* |
857 | * Ensure that there are not too many pages isolated from the LRU | |
858 | * list by either parallel reclaimers or compaction. If there are, | |
859 | * delay for some time until fewer pages are isolated | |
860 | */ | |
4fbbb3fd | 861 | while (unlikely(too_many_isolated(cc))) { |
d20bdd57 ZY |
862 | /* stop isolation if there are still pages not migrated */ |
863 | if (cc->nr_migratepages) | |
c2ad7a1f | 864 | return -EAGAIN; |
d20bdd57 | 865 | |
f9e35b3b | 866 | /* async migration should just abort */ |
e0b9daeb | 867 | if (cc->mode == MIGRATE_ASYNC) |
c2ad7a1f | 868 | return -EAGAIN; |
f9e35b3b | 869 | |
c3f4a9a2 | 870 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
748446bb MG |
871 | |
872 | if (fatal_signal_pending(current)) | |
c2ad7a1f | 873 | return -EINTR; |
748446bb MG |
874 | } |
875 | ||
cf66f070 | 876 | cond_resched(); |
aeef4b83 | 877 | |
fdd048e1 VB |
878 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
879 | skip_on_failure = true; | |
880 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
881 | } | |
882 | ||
748446bb | 883 | /* Time to isolate some pages for migration */ |
748446bb | 884 | for (; low_pfn < end_pfn; low_pfn++) { |
0003e2a4 | 885 | bool is_dirty, is_unevictable; |
29c0dde8 | 886 | |
fdd048e1 VB |
887 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
888 | /* | |
889 | * We have isolated all migration candidates in the | |
890 | * previous order-aligned block, and did not skip it due | |
891 | * to failure. We should migrate the pages now and | |
892 | * hopefully succeed compaction. | |
893 | */ | |
894 | if (nr_isolated) | |
895 | break; | |
896 | ||
897 | /* | |
898 | * We failed to isolate in the previous order-aligned | |
899 | * block. Set the new boundary to the end of the | |
900 | * current block. Note we can't simply increase | |
901 | * next_skip_pfn by 1 << order, as low_pfn might have | |
902 | * been incremented by a higher number due to skipping | |
903 | * a compound or a high-order buddy page in the | |
904 | * previous loop iteration. | |
905 | */ | |
906 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
907 | } | |
908 | ||
8b44d279 VB |
909 | /* |
910 | * Periodically drop the lock (if held) regardless of its | |
670105a2 MG |
911 | * contention, to give chance to IRQs. Abort completely if |
912 | * a fatal signal is pending. | |
8b44d279 | 913 | */ |
c036ddff | 914 | if (!(low_pfn % COMPACT_CLUSTER_MAX)) { |
6168d0da AS |
915 | if (locked) { |
916 | unlock_page_lruvec_irqrestore(locked, flags); | |
917 | locked = NULL; | |
918 | } | |
919 | ||
920 | if (fatal_signal_pending(current)) { | |
921 | cc->contended = true; | |
c2ad7a1f | 922 | ret = -EINTR; |
6168d0da | 923 | |
6168d0da AS |
924 | goto fatal_pending; |
925 | } | |
926 | ||
927 | cond_resched(); | |
670105a2 | 928 | } |
c67fe375 | 929 | |
b7aba698 | 930 | nr_scanned++; |
748446bb | 931 | |
748446bb | 932 | page = pfn_to_page(low_pfn); |
dc908600 | 933 | |
e380bebe MG |
934 | /* |
935 | * Check if the pageblock has already been marked skipped. | |
493614da | 936 | * Only the first PFN is checked as the caller isolates |
e380bebe MG |
937 | * COMPACT_CLUSTER_MAX at a time so the second call must |
938 | * not falsely conclude that the block should be skipped. | |
939 | */ | |
493614da JW |
940 | if (!valid_page && (pageblock_aligned(low_pfn) || |
941 | low_pfn == cc->zone->zone_start_pfn)) { | |
4af12d04 | 942 | if (!isolation_suitable(cc, page)) { |
e380bebe | 943 | low_pfn = end_pfn; |
56ae0bb3 | 944 | folio = NULL; |
e380bebe MG |
945 | goto isolate_abort; |
946 | } | |
bb13ffeb | 947 | valid_page = page; |
e380bebe | 948 | } |
bb13ffeb | 949 | |
369fa227 | 950 | if (PageHuge(page) && cc->alloc_contig) { |
1c06b6a5 BW |
951 | if (locked) { |
952 | unlock_page_lruvec_irqrestore(locked, flags); | |
953 | locked = NULL; | |
954 | } | |
955 | ||
ae37c7ff | 956 | ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); |
369fa227 OS |
957 | |
958 | /* | |
959 | * Fail isolation in case isolate_or_dissolve_huge_page() | |
960 | * reports an error. In case of -ENOMEM, abort right away. | |
961 | */ | |
962 | if (ret < 0) { | |
963 | /* Do not report -EBUSY down the chain */ | |
964 | if (ret == -EBUSY) | |
965 | ret = 0; | |
66fe1cf7 | 966 | low_pfn += compound_nr(page) - 1; |
56d48d8d | 967 | nr_scanned += compound_nr(page) - 1; |
369fa227 OS |
968 | goto isolate_fail; |
969 | } | |
970 | ||
ae37c7ff OS |
971 | if (PageHuge(page)) { |
972 | /* | |
973 | * Hugepage was successfully isolated and placed | |
974 | * on the cc->migratepages list. | |
975 | */ | |
56ae0bb3 KW |
976 | folio = page_folio(page); |
977 | low_pfn += folio_nr_pages(folio) - 1; | |
ae37c7ff OS |
978 | goto isolate_success_no_list; |
979 | } | |
980 | ||
369fa227 OS |
981 | /* |
982 | * Ok, the hugepage was dissolved. Now these pages are | |
983 | * Buddy and cannot be re-allocated because they are | |
984 | * isolated. Fall-through as the check below handles | |
985 | * Buddy pages. | |
986 | */ | |
987 | } | |
988 | ||
6c14466c | 989 | /* |
99c0fd5e VB |
990 | * Skip if free. We read page order here without zone lock |
991 | * which is generally unsafe, but the race window is small and | |
992 | * the worst thing that can happen is that we skip some | |
993 | * potential isolation targets. | |
6c14466c | 994 | */ |
99c0fd5e | 995 | if (PageBuddy(page)) { |
ab130f91 | 996 | unsigned long freepage_order = buddy_order_unsafe(page); |
99c0fd5e VB |
997 | |
998 | /* | |
999 | * Without lock, we cannot be sure that what we got is | |
1000 | * a valid page order. Consider only values in the | |
1001 | * valid order range to prevent low_pfn overflow. | |
1002 | */ | |
5e0a760b | 1003 | if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) { |
99c0fd5e | 1004 | low_pfn += (1UL << freepage_order) - 1; |
56d48d8d BW |
1005 | nr_scanned += (1UL << freepage_order) - 1; |
1006 | } | |
748446bb | 1007 | continue; |
99c0fd5e | 1008 | } |
748446bb | 1009 | |
bc835011 | 1010 | /* |
29c0dde8 | 1011 | * Regardless of being on LRU, compound pages such as THP and |
1da2f328 RR |
1012 | * hugetlbfs are not to be compacted unless we are attempting |
1013 | * an allocation much larger than the huge page size (eg CMA). | |
1014 | * We can potentially save a lot of iterations if we skip them | |
1015 | * at once. The check is racy, but we can consider only valid | |
1016 | * values and the only danger is skipping too much. | |
bc835011 | 1017 | */ |
1da2f328 | 1018 | if (PageCompound(page) && !cc->alloc_contig) { |
21dc7e02 | 1019 | const unsigned int order = compound_order(page); |
edc2ca61 | 1020 | |
5e0a760b | 1021 | if (likely(order <= MAX_PAGE_ORDER)) { |
21dc7e02 | 1022 | low_pfn += (1UL << order) - 1; |
56d48d8d BW |
1023 | nr_scanned += (1UL << order) - 1; |
1024 | } | |
fdd048e1 | 1025 | goto isolate_fail; |
2a1402aa MG |
1026 | } |
1027 | ||
bda807d4 MK |
1028 | /* |
1029 | * Check may be lockless but that's ok as we recheck later. | |
1030 | * It's possible to migrate LRU and non-lru movable pages. | |
1031 | * Skip any other type of page | |
1032 | */ | |
1033 | if (!PageLRU(page)) { | |
bda807d4 MK |
1034 | /* |
1035 | * __PageMovable can return false positive so we need | |
1036 | * to verify it under page_lock. | |
1037 | */ | |
1038 | if (unlikely(__PageMovable(page)) && | |
1039 | !PageIsolated(page)) { | |
1040 | if (locked) { | |
6168d0da AS |
1041 | unlock_page_lruvec_irqrestore(locked, flags); |
1042 | locked = NULL; | |
bda807d4 MK |
1043 | } |
1044 | ||
56ae0bb3 KW |
1045 | if (isolate_movable_page(page, mode)) { |
1046 | folio = page_folio(page); | |
bda807d4 | 1047 | goto isolate_success; |
56ae0bb3 | 1048 | } |
bda807d4 MK |
1049 | } |
1050 | ||
fdd048e1 | 1051 | goto isolate_fail; |
bda807d4 | 1052 | } |
29c0dde8 | 1053 | |
829ae0f8 GS |
1054 | /* |
1055 | * Be careful not to clear PageLRU until after we're | |
1056 | * sure the page is not being freed elsewhere -- the | |
1057 | * page release code relies on it. | |
1058 | */ | |
56ae0bb3 KW |
1059 | folio = folio_get_nontail_page(page); |
1060 | if (unlikely(!folio)) | |
829ae0f8 GS |
1061 | goto isolate_fail; |
1062 | ||
119d6d59 DR |
1063 | /* |
1064 | * Migration will fail if an anonymous page is pinned in memory, | |
1065 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
1066 | * admittedly racy check. | |
1067 | */ | |
56ae0bb3 KW |
1068 | mapping = folio_mapping(folio); |
1069 | if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio)) | |
829ae0f8 | 1070 | goto isolate_fail_put; |
119d6d59 | 1071 | |
73e64c51 MH |
1072 | /* |
1073 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
1074 | * because those do not depend on fs locks. | |
1075 | */ | |
89f6c88a | 1076 | if (!(cc->gfp_mask & __GFP_FS) && mapping) |
829ae0f8 | 1077 | goto isolate_fail_put; |
9df41314 | 1078 | |
89f6c88a | 1079 | /* Only take pages on LRU: a check now makes later tests safe */ |
56ae0bb3 | 1080 | if (!folio_test_lru(folio)) |
89f6c88a HD |
1081 | goto isolate_fail_put; |
1082 | ||
0003e2a4 SC |
1083 | is_unevictable = folio_test_unevictable(folio); |
1084 | ||
89f6c88a | 1085 | /* Compaction might skip unevictable pages but CMA takes them */ |
0003e2a4 | 1086 | if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable) |
89f6c88a HD |
1087 | goto isolate_fail_put; |
1088 | ||
1089 | /* | |
1090 | * To minimise LRU disruption, the caller can indicate with | |
1091 | * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages | |
1092 | * it will be able to migrate without blocking - clean pages | |
1093 | * for the most part. PageWriteback would require blocking. | |
1094 | */ | |
56ae0bb3 | 1095 | if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio)) |
9df41314 AS |
1096 | goto isolate_fail_put; |
1097 | ||
0003e2a4 SC |
1098 | is_dirty = folio_test_dirty(folio); |
1099 | ||
1100 | if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) || | |
1101 | (mapping && is_unevictable)) { | |
1102 | bool migrate_dirty = true; | |
1103 | bool is_unmovable; | |
89f6c88a HD |
1104 | |
1105 | /* | |
866ff801 | 1106 | * Only folios without mappings or that have |
0003e2a4 SC |
1107 | * a ->migrate_folio callback are possible to migrate |
1108 | * without blocking. | |
1109 | * | |
1110 | * Folios from unmovable mappings are not migratable. | |
1111 | * | |
1112 | * However, we can be racing with truncation, which can | |
1113 | * free the mapping that we need to check. Truncation | |
1114 | * holds the folio lock until after the folio is removed | |
1115 | * from the page so holding it ourselves is sufficient. | |
1116 | * | |
1117 | * To avoid locking the folio just to check unmovable, | |
1118 | * assume every unmovable folio is also unevictable, | |
1119 | * which is a cheaper test. If our assumption goes | |
1120 | * wrong, it's not a correctness bug, just potentially | |
1121 | * wasted cycles. | |
89f6c88a | 1122 | */ |
56ae0bb3 | 1123 | if (!folio_trylock(folio)) |
89f6c88a HD |
1124 | goto isolate_fail_put; |
1125 | ||
56ae0bb3 | 1126 | mapping = folio_mapping(folio); |
0003e2a4 SC |
1127 | if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) { |
1128 | migrate_dirty = !mapping || | |
1129 | mapping->a_ops->migrate_folio; | |
1130 | } | |
1131 | is_unmovable = mapping && mapping_unmovable(mapping); | |
56ae0bb3 | 1132 | folio_unlock(folio); |
0003e2a4 | 1133 | if (!migrate_dirty || is_unmovable) |
89f6c88a HD |
1134 | goto isolate_fail_put; |
1135 | } | |
1136 | ||
56ae0bb3 KW |
1137 | /* Try isolate the folio */ |
1138 | if (!folio_test_clear_lru(folio)) | |
9df41314 AS |
1139 | goto isolate_fail_put; |
1140 | ||
56ae0bb3 | 1141 | lruvec = folio_lruvec(folio); |
6168d0da | 1142 | |
69b7189f | 1143 | /* If we already hold the lock, we can skip some rechecking */ |
6168d0da AS |
1144 | if (lruvec != locked) { |
1145 | if (locked) | |
1146 | unlock_page_lruvec_irqrestore(locked, flags); | |
1147 | ||
1148 | compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); | |
1149 | locked = lruvec; | |
6168d0da | 1150 | |
56ae0bb3 | 1151 | lruvec_memcg_debug(lruvec, folio); |
e380bebe | 1152 | |
590ccea8 MG |
1153 | /* |
1154 | * Try get exclusive access under lock. If marked for | |
1155 | * skip, the scan is aborted unless the current context | |
1156 | * is a rescan to reach the end of the pageblock. | |
1157 | */ | |
1158 | if (!skip_updated && valid_page) { | |
e380bebe | 1159 | skip_updated = true; |
590ccea8 MG |
1160 | if (test_and_set_skip(cc, valid_page) && |
1161 | !cc->finish_pageblock) { | |
7545e2f2 | 1162 | low_pfn = end_pfn; |
e380bebe | 1163 | goto isolate_abort; |
590ccea8 | 1164 | } |
e380bebe | 1165 | } |
2a1402aa | 1166 | |
29c0dde8 | 1167 | /* |
56ae0bb3 KW |
1168 | * folio become large since the non-locked check, |
1169 | * and it's on LRU. | |
29c0dde8 | 1170 | */ |
56ae0bb3 KW |
1171 | if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) { |
1172 | low_pfn += folio_nr_pages(folio) - 1; | |
1173 | nr_scanned += folio_nr_pages(folio) - 1; | |
1174 | folio_set_lru(folio); | |
9df41314 | 1175 | goto isolate_fail_put; |
69b7189f | 1176 | } |
d99fd5fe | 1177 | } |
fa9add64 | 1178 | |
56ae0bb3 KW |
1179 | /* The folio is taken off the LRU */ |
1180 | if (folio_test_large(folio)) | |
1181 | low_pfn += folio_nr_pages(folio) - 1; | |
bc835011 | 1182 | |
748446bb | 1183 | /* Successfully isolated */ |
56ae0bb3 KW |
1184 | lruvec_del_folio(lruvec, folio); |
1185 | node_stat_mod_folio(folio, | |
1186 | NR_ISOLATED_ANON + folio_is_file_lru(folio), | |
1187 | folio_nr_pages(folio)); | |
b6c75016 JK |
1188 | |
1189 | isolate_success: | |
56ae0bb3 | 1190 | list_add(&folio->lru, &cc->migratepages); |
ae37c7ff | 1191 | isolate_success_no_list: |
56ae0bb3 KW |
1192 | cc->nr_migratepages += folio_nr_pages(folio); |
1193 | nr_isolated += folio_nr_pages(folio); | |
1194 | nr_scanned += folio_nr_pages(folio) - 1; | |
748446bb | 1195 | |
804d3121 MG |
1196 | /* |
1197 | * Avoid isolating too much unless this block is being | |
48731c84 | 1198 | * fully scanned (e.g. dirty/writeback pages, parallel allocation) |
cb2dcaf0 MG |
1199 | * or a lock is contended. For contention, isolate quickly to |
1200 | * potentially remove one source of contention. | |
804d3121 | 1201 | */ |
38935861 | 1202 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && |
48731c84 | 1203 | !cc->finish_pageblock && !cc->contended) { |
31b8384a | 1204 | ++low_pfn; |
748446bb | 1205 | break; |
31b8384a | 1206 | } |
fdd048e1 VB |
1207 | |
1208 | continue; | |
9df41314 AS |
1209 | |
1210 | isolate_fail_put: | |
1211 | /* Avoid potential deadlock in freeing page under lru_lock */ | |
1212 | if (locked) { | |
6168d0da AS |
1213 | unlock_page_lruvec_irqrestore(locked, flags); |
1214 | locked = NULL; | |
9df41314 | 1215 | } |
56ae0bb3 | 1216 | folio_put(folio); |
9df41314 | 1217 | |
fdd048e1 | 1218 | isolate_fail: |
369fa227 | 1219 | if (!skip_on_failure && ret != -ENOMEM) |
fdd048e1 VB |
1220 | continue; |
1221 | ||
1222 | /* | |
1223 | * We have isolated some pages, but then failed. Release them | |
1224 | * instead of migrating, as we cannot form the cc->order buddy | |
1225 | * page anyway. | |
1226 | */ | |
1227 | if (nr_isolated) { | |
1228 | if (locked) { | |
6168d0da AS |
1229 | unlock_page_lruvec_irqrestore(locked, flags); |
1230 | locked = NULL; | |
fdd048e1 | 1231 | } |
fdd048e1 VB |
1232 | putback_movable_pages(&cc->migratepages); |
1233 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1234 | nr_isolated = 0; |
1235 | } | |
1236 | ||
1237 | if (low_pfn < next_skip_pfn) { | |
1238 | low_pfn = next_skip_pfn - 1; | |
1239 | /* | |
1240 | * The check near the loop beginning would have updated | |
1241 | * next_skip_pfn too, but this is a bit simpler. | |
1242 | */ | |
1243 | next_skip_pfn += 1UL << cc->order; | |
1244 | } | |
369fa227 OS |
1245 | |
1246 | if (ret == -ENOMEM) | |
1247 | break; | |
748446bb MG |
1248 | } |
1249 | ||
99c0fd5e VB |
1250 | /* |
1251 | * The PageBuddy() check could have potentially brought us outside | |
1252 | * the range to be scanned. | |
1253 | */ | |
1254 | if (unlikely(low_pfn > end_pfn)) | |
1255 | low_pfn = end_pfn; | |
1256 | ||
56ae0bb3 | 1257 | folio = NULL; |
9df41314 | 1258 | |
e380bebe | 1259 | isolate_abort: |
c67fe375 | 1260 | if (locked) |
6168d0da | 1261 | unlock_page_lruvec_irqrestore(locked, flags); |
56ae0bb3 KW |
1262 | if (folio) { |
1263 | folio_set_lru(folio); | |
1264 | folio_put(folio); | |
9df41314 | 1265 | } |
748446bb | 1266 | |
50b5b094 | 1267 | /* |
48731c84 | 1268 | * Update the cached scanner pfn once the pageblock has been scanned. |
804d3121 MG |
1269 | * Pages will either be migrated in which case there is no point |
1270 | * scanning in the near future or migration failed in which case the | |
1271 | * failure reason may persist. The block is marked for skipping if | |
1272 | * there were no pages isolated in the block or if the block is | |
1273 | * rescanned twice in a row. | |
50b5b094 | 1274 | */ |
48731c84 | 1275 | if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) { |
8b71b499 | 1276 | if (!cc->no_set_skip_hint && valid_page && !skip_updated) |
e380bebe MG |
1277 | set_pageblock_skip(valid_page); |
1278 | update_cached_migrate(cc, low_pfn); | |
1279 | } | |
bb13ffeb | 1280 | |
e34d85f0 JK |
1281 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1282 | nr_scanned, nr_isolated); | |
b7aba698 | 1283 | |
670105a2 | 1284 | fatal_pending: |
7f354a54 | 1285 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1286 | if (nr_isolated) |
010fc29a | 1287 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1288 | |
c2ad7a1f OS |
1289 | cc->migrate_pfn = low_pfn; |
1290 | ||
1291 | return ret; | |
2fe86e00 MN |
1292 | } |
1293 | ||
edc2ca61 VB |
1294 | /** |
1295 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1296 | * @cc: Compaction control structure. | |
1297 | * @start_pfn: The first PFN to start isolating. | |
1298 | * @end_pfn: The one-past-last PFN. | |
1299 | * | |
369fa227 OS |
1300 | * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM |
1301 | * in case we could not allocate a page, or 0. | |
edc2ca61 | 1302 | */ |
c2ad7a1f | 1303 | int |
edc2ca61 VB |
1304 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, |
1305 | unsigned long end_pfn) | |
1306 | { | |
e1409c32 | 1307 | unsigned long pfn, block_start_pfn, block_end_pfn; |
c2ad7a1f | 1308 | int ret = 0; |
edc2ca61 VB |
1309 | |
1310 | /* Scan block by block. First and last block may be incomplete */ | |
1311 | pfn = start_pfn; | |
06b6640a | 1312 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1313 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1314 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1315 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1316 | |
1317 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1318 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1319 | block_end_pfn += pageblock_nr_pages) { |
1320 | ||
1321 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1322 | ||
e1409c32 JK |
1323 | if (!pageblock_pfn_to_page(block_start_pfn, |
1324 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1325 | continue; |
1326 | ||
c2ad7a1f OS |
1327 | ret = isolate_migratepages_block(cc, pfn, block_end_pfn, |
1328 | ISOLATE_UNEVICTABLE); | |
edc2ca61 | 1329 | |
c2ad7a1f | 1330 | if (ret) |
edc2ca61 | 1331 | break; |
6ea41c0c | 1332 | |
38935861 | 1333 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) |
6ea41c0c | 1334 | break; |
edc2ca61 | 1335 | } |
edc2ca61 | 1336 | |
c2ad7a1f | 1337 | return ret; |
edc2ca61 VB |
1338 | } |
1339 | ||
ff9543fd MN |
1340 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1341 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1342 | |
b682debd VB |
1343 | static bool suitable_migration_source(struct compact_control *cc, |
1344 | struct page *page) | |
1345 | { | |
282722b0 VB |
1346 | int block_mt; |
1347 | ||
9bebefd5 MG |
1348 | if (pageblock_skip_persistent(page)) |
1349 | return false; | |
1350 | ||
282722b0 | 1351 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1352 | return true; |
1353 | ||
282722b0 VB |
1354 | block_mt = get_pageblock_migratetype(page); |
1355 | ||
1356 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1357 | return is_migrate_movable(block_mt); | |
1358 | else | |
1359 | return block_mt == cc->migratetype; | |
b682debd VB |
1360 | } |
1361 | ||
018e9a49 | 1362 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1363 | static bool suitable_migration_target(struct compact_control *cc, |
1364 | struct page *page) | |
018e9a49 AM |
1365 | { |
1366 | /* If the page is a large free page, then disallow migration */ | |
1367 | if (PageBuddy(page)) { | |
1883e8ac BW |
1368 | int order = cc->order > 0 ? cc->order : pageblock_order; |
1369 | ||
018e9a49 AM |
1370 | /* |
1371 | * We are checking page_order without zone->lock taken. But | |
1372 | * the only small danger is that we skip a potentially suitable | |
1373 | * pageblock, so it's not worth to check order for valid range. | |
1374 | */ | |
1883e8ac | 1375 | if (buddy_order_unsafe(page) >= order) |
018e9a49 AM |
1376 | return false; |
1377 | } | |
1378 | ||
1ef36db2 YX |
1379 | if (cc->ignore_block_suitable) |
1380 | return true; | |
1381 | ||
018e9a49 | 1382 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1383 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1384 | return true; |
1385 | ||
1386 | /* Otherwise skip the block */ | |
1387 | return false; | |
1388 | } | |
1389 | ||
70b44595 MG |
1390 | static inline unsigned int |
1391 | freelist_scan_limit(struct compact_control *cc) | |
1392 | { | |
dd7ef7bd QC |
1393 | unsigned short shift = BITS_PER_LONG - 1; |
1394 | ||
1395 | return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; | |
70b44595 MG |
1396 | } |
1397 | ||
f2849aa0 VB |
1398 | /* |
1399 | * Test whether the free scanner has reached the same or lower pageblock than | |
1400 | * the migration scanner, and compaction should thus terminate. | |
1401 | */ | |
1402 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1403 | { | |
1404 | return (cc->free_pfn >> pageblock_order) | |
1405 | <= (cc->migrate_pfn >> pageblock_order); | |
1406 | } | |
1407 | ||
5a811889 MG |
1408 | /* |
1409 | * Used when scanning for a suitable migration target which scans freelists | |
1410 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1411 | * first on the next iteration of the free scanner | |
1412 | */ | |
1413 | static void | |
1414 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1415 | { | |
1416 | LIST_HEAD(sublist); | |
1417 | ||
4c179891 | 1418 | if (!list_is_first(&freepage->buddy_list, freelist)) { |
bbefa0fc | 1419 | list_cut_before(&sublist, freelist, &freepage->buddy_list); |
d2155fe5 | 1420 | list_splice_tail(&sublist, freelist); |
5a811889 MG |
1421 | } |
1422 | } | |
1423 | ||
1424 | /* | |
1425 | * Similar to move_freelist_head except used by the migration scanner | |
1426 | * when scanning forward. It's possible for these list operations to | |
1427 | * move against each other if they search the free list exactly in | |
1428 | * lockstep. | |
1429 | */ | |
70b44595 MG |
1430 | static void |
1431 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1432 | { | |
1433 | LIST_HEAD(sublist); | |
1434 | ||
4c179891 | 1435 | if (!list_is_last(&freepage->buddy_list, freelist)) { |
bbefa0fc | 1436 | list_cut_position(&sublist, freelist, &freepage->buddy_list); |
d2155fe5 | 1437 | list_splice_tail(&sublist, freelist); |
70b44595 MG |
1438 | } |
1439 | } | |
1440 | ||
5a811889 | 1441 | static void |
be21b32a | 1442 | fast_isolate_around(struct compact_control *cc, unsigned long pfn) |
5a811889 MG |
1443 | { |
1444 | unsigned long start_pfn, end_pfn; | |
6e2b7044 | 1445 | struct page *page; |
5a811889 MG |
1446 | |
1447 | /* Do not search around if there are enough pages already */ | |
1448 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1449 | return; | |
1450 | ||
1451 | /* Minimise scanning during async compaction */ | |
1452 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1453 | return; | |
1454 | ||
1455 | /* Pageblock boundaries */ | |
6e2b7044 VB |
1456 | start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); |
1457 | end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); | |
1458 | ||
1459 | page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); | |
1460 | if (!page) | |
1461 | return; | |
5a811889 | 1462 | |
be21b32a | 1463 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); |
5a811889 MG |
1464 | |
1465 | /* Skip this pageblock in the future as it's full or nearly full */ | |
18c59d58 | 1466 | if (start_pfn == end_pfn && !cc->no_set_skip_hint) |
5a811889 MG |
1467 | set_pageblock_skip(page); |
1468 | } | |
1469 | ||
dbe2d4e4 MG |
1470 | /* Search orders in round-robin fashion */ |
1471 | static int next_search_order(struct compact_control *cc, int order) | |
1472 | { | |
1473 | order--; | |
1474 | if (order < 0) | |
1475 | order = cc->order - 1; | |
1476 | ||
1477 | /* Search wrapped around? */ | |
1478 | if (order == cc->search_order) { | |
1479 | cc->search_order--; | |
1480 | if (cc->search_order < 0) | |
1481 | cc->search_order = cc->order - 1; | |
1482 | return -1; | |
1483 | } | |
1484 | ||
1485 | return order; | |
1486 | } | |
1487 | ||
2dbd9005 | 1488 | static void fast_isolate_freepages(struct compact_control *cc) |
5a811889 | 1489 | { |
b55ca526 | 1490 | unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); |
447ba886 | 1491 | unsigned int nr_scanned = 0, total_isolated = 0; |
74e21484 | 1492 | unsigned long low_pfn, min_pfn, highest = 0; |
5a811889 MG |
1493 | unsigned long nr_isolated = 0; |
1494 | unsigned long distance; | |
1495 | struct page *page = NULL; | |
1496 | bool scan_start = false; | |
1497 | int order; | |
1498 | ||
1499 | /* Full compaction passes in a negative order */ | |
1500 | if (cc->order <= 0) | |
2dbd9005 | 1501 | return; |
5a811889 MG |
1502 | |
1503 | /* | |
1504 | * If starting the scan, use a deeper search and use the highest | |
1505 | * PFN found if a suitable one is not found. | |
1506 | */ | |
e332f741 | 1507 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1508 | limit = pageblock_nr_pages >> 1; |
1509 | scan_start = true; | |
1510 | } | |
1511 | ||
1512 | /* | |
1513 | * Preferred point is in the top quarter of the scan space but take | |
1514 | * a pfn from the top half if the search is problematic. | |
1515 | */ | |
1516 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1517 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1518 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1519 | ||
1520 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1521 | low_pfn = min_pfn; | |
1522 | ||
dbe2d4e4 MG |
1523 | /* |
1524 | * Search starts from the last successful isolation order or the next | |
1525 | * order to search after a previous failure | |
1526 | */ | |
1527 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1528 | ||
1529 | for (order = cc->search_order; | |
1530 | !page && order >= 0; | |
1531 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1532 | struct free_area *area = &cc->zone->free_area[order]; |
1533 | struct list_head *freelist; | |
1534 | struct page *freepage; | |
1535 | unsigned long flags; | |
1536 | unsigned int order_scanned = 0; | |
74e21484 | 1537 | unsigned long high_pfn = 0; |
5a811889 MG |
1538 | |
1539 | if (!area->nr_free) | |
1540 | continue; | |
1541 | ||
1542 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1543 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
94ec2003 | 1544 | list_for_each_entry_reverse(freepage, freelist, buddy_list) { |
5a811889 MG |
1545 | unsigned long pfn; |
1546 | ||
1547 | order_scanned++; | |
1548 | nr_scanned++; | |
1549 | pfn = page_to_pfn(freepage); | |
1550 | ||
1551 | if (pfn >= highest) | |
6e2b7044 VB |
1552 | highest = max(pageblock_start_pfn(pfn), |
1553 | cc->zone->zone_start_pfn); | |
5a811889 MG |
1554 | |
1555 | if (pfn >= low_pfn) { | |
1556 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1557 | cc->search_order = order; |
5a811889 MG |
1558 | page = freepage; |
1559 | break; | |
1560 | } | |
1561 | ||
1562 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1563 | high_pfn = pfn; | |
1564 | ||
1565 | /* Shorten the scan if a candidate is found */ | |
1566 | limit >>= 1; | |
1567 | } | |
1568 | ||
1569 | if (order_scanned >= limit) | |
1570 | break; | |
1571 | } | |
1572 | ||
e6bd14ec | 1573 | /* Use a maximum candidate pfn if a preferred one was not found */ |
5a811889 MG |
1574 | if (!page && high_pfn) { |
1575 | page = pfn_to_page(high_pfn); | |
1576 | ||
1577 | /* Update freepage for the list reorder below */ | |
1578 | freepage = page; | |
1579 | } | |
1580 | ||
1581 | /* Reorder to so a future search skips recent pages */ | |
1582 | move_freelist_head(freelist, freepage); | |
1583 | ||
1584 | /* Isolate the page if available */ | |
1585 | if (page) { | |
1586 | if (__isolate_free_page(page, order)) { | |
1587 | set_page_private(page, order); | |
1588 | nr_isolated = 1 << order; | |
b717d6b9 | 1589 | nr_scanned += nr_isolated - 1; |
447ba886 | 1590 | total_isolated += nr_isolated; |
5a811889 MG |
1591 | cc->nr_freepages += nr_isolated; |
1592 | list_add_tail(&page->lru, &cc->freepages); | |
1593 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1594 | } else { | |
1595 | /* If isolation fails, abort the search */ | |
5b56d996 | 1596 | order = cc->search_order + 1; |
5a811889 MG |
1597 | page = NULL; |
1598 | } | |
1599 | } | |
1600 | ||
1601 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1602 | ||
a8d13355 BW |
1603 | /* Skip fast search if enough freepages isolated */ |
1604 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1605 | break; | |
1606 | ||
5a811889 | 1607 | /* |
b55ca526 | 1608 | * Smaller scan on next order so the total scan is related |
5a811889 MG |
1609 | * to freelist_scan_limit. |
1610 | */ | |
1611 | if (order_scanned >= limit) | |
b55ca526 | 1612 | limit = max(1U, limit >> 1); |
5a811889 MG |
1613 | } |
1614 | ||
447ba886 BW |
1615 | trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn, |
1616 | nr_scanned, total_isolated); | |
1617 | ||
5a811889 MG |
1618 | if (!page) { |
1619 | cc->fast_search_fail++; | |
1620 | if (scan_start) { | |
1621 | /* | |
1622 | * Use the highest PFN found above min. If one was | |
f3867755 | 1623 | * not found, be pessimistic for direct compaction |
5a811889 MG |
1624 | * and use the min mark. |
1625 | */ | |
ca2864e5 | 1626 | if (highest >= min_pfn) { |
5a811889 MG |
1627 | page = pfn_to_page(highest); |
1628 | cc->free_pfn = highest; | |
1629 | } else { | |
e577c8b6 | 1630 | if (cc->direct_compaction && pfn_valid(min_pfn)) { |
73a6e474 | 1631 | page = pageblock_pfn_to_page(min_pfn, |
6e2b7044 VB |
1632 | min(pageblock_end_pfn(min_pfn), |
1633 | zone_end_pfn(cc->zone)), | |
73a6e474 | 1634 | cc->zone); |
d19b1a17 BS |
1635 | if (page && !suitable_migration_target(cc, page)) |
1636 | page = NULL; | |
1637 | ||
5a811889 MG |
1638 | cc->free_pfn = min_pfn; |
1639 | } | |
1640 | } | |
1641 | } | |
1642 | } | |
1643 | ||
d097a6f6 MG |
1644 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1645 | highest -= pageblock_nr_pages; | |
5a811889 | 1646 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1647 | } |
5a811889 MG |
1648 | |
1649 | cc->total_free_scanned += nr_scanned; | |
1650 | if (!page) | |
2dbd9005 | 1651 | return; |
5a811889 MG |
1652 | |
1653 | low_pfn = page_to_pfn(page); | |
be21b32a | 1654 | fast_isolate_around(cc, low_pfn); |
5a811889 MG |
1655 | } |
1656 | ||
2fe86e00 | 1657 | /* |
ff9543fd MN |
1658 | * Based on information in the current compact_control, find blocks |
1659 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1660 | */ |
edc2ca61 | 1661 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1662 | { |
edc2ca61 | 1663 | struct zone *zone = cc->zone; |
ff9543fd | 1664 | struct page *page; |
c96b9e50 | 1665 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1666 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1667 | unsigned long block_end_pfn; /* end of current pageblock */ |
1668 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1669 | struct list_head *freelist = &cc->freepages; |
4fca9730 | 1670 | unsigned int stride; |
2fe86e00 | 1671 | |
5a811889 | 1672 | /* Try a small search of the free lists for a candidate */ |
00bc102f | 1673 | fast_isolate_freepages(cc); |
5a811889 MG |
1674 | if (cc->nr_freepages) |
1675 | goto splitmap; | |
1676 | ||
ff9543fd MN |
1677 | /* |
1678 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1679 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1680 | * zone when isolating for the first time. For looping we also need |
1681 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1682 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1683 | * For ending point, take care when isolating in last pageblock of a | |
a1c1dbeb | 1684 | * zone which ends in the middle of a pageblock. |
49e068f0 VB |
1685 | * The low boundary is the end of the pageblock the migration scanner |
1686 | * is using. | |
ff9543fd | 1687 | */ |
e14c720e | 1688 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1689 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1690 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1691 | zone_end_pfn(zone)); | |
06b6640a | 1692 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1693 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1694 | |
ff9543fd MN |
1695 | /* |
1696 | * Isolate free pages until enough are available to migrate the | |
1697 | * pages on cc->migratepages. We stop searching if the migrate | |
1698 | * and free page scanners meet or enough free pages are isolated. | |
1699 | */ | |
f5f61a32 | 1700 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1701 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1702 | block_start_pfn -= pageblock_nr_pages, |
1703 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1704 | unsigned long nr_isolated; |
1705 | ||
f6ea3adb DR |
1706 | /* |
1707 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1708 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1709 | */ |
c036ddff | 1710 | if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1711 | cond_resched(); |
f6ea3adb | 1712 | |
7d49d886 VB |
1713 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1714 | zone); | |
e6e0c767 BW |
1715 | if (!page) { |
1716 | unsigned long next_pfn; | |
1717 | ||
1718 | next_pfn = skip_offline_sections_reverse(block_start_pfn); | |
1719 | if (next_pfn) | |
1720 | block_start_pfn = max(next_pfn, low_pfn); | |
1721 | ||
ff9543fd | 1722 | continue; |
e6e0c767 | 1723 | } |
ff9543fd MN |
1724 | |
1725 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1726 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1727 | continue; |
68e3e926 | 1728 | |
bb13ffeb MG |
1729 | /* If isolation recently failed, do not retry */ |
1730 | if (!isolation_suitable(cc, page)) | |
1731 | continue; | |
1732 | ||
e14c720e | 1733 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 MG |
1734 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
1735 | block_end_pfn, freelist, stride, false); | |
ff9543fd | 1736 | |
d097a6f6 MG |
1737 | /* Update the skip hint if the full pageblock was scanned */ |
1738 | if (isolate_start_pfn == block_end_pfn) | |
16951789 KS |
1739 | update_pageblock_skip(cc, page, block_start_pfn - |
1740 | pageblock_nr_pages); | |
d097a6f6 | 1741 | |
cb2dcaf0 MG |
1742 | /* Are enough freepages isolated? */ |
1743 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1744 | if (isolate_start_pfn >= block_end_pfn) { |
1745 | /* | |
1746 | * Restart at previous pageblock if more | |
1747 | * freepages can be isolated next time. | |
1748 | */ | |
f5f61a32 VB |
1749 | isolate_start_pfn = |
1750 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1751 | } |
be976572 | 1752 | break; |
a46cbf3b | 1753 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1754 | /* |
a46cbf3b DR |
1755 | * If isolation failed early, do not continue |
1756 | * needlessly. | |
f5f61a32 | 1757 | */ |
a46cbf3b | 1758 | break; |
f5f61a32 | 1759 | } |
4fca9730 MG |
1760 | |
1761 | /* Adjust stride depending on isolation */ | |
1762 | if (nr_isolated) { | |
1763 | stride = 1; | |
1764 | continue; | |
1765 | } | |
1766 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1767 | } |
1768 | ||
7ed695e0 | 1769 | /* |
f5f61a32 VB |
1770 | * Record where the free scanner will restart next time. Either we |
1771 | * broke from the loop and set isolate_start_pfn based on the last | |
1772 | * call to isolate_freepages_block(), or we met the migration scanner | |
1773 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1774 | */ |
f5f61a32 | 1775 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1776 | |
1777 | splitmap: | |
1778 | /* __isolate_free_page() does not map the pages */ | |
1779 | split_map_pages(freelist); | |
748446bb MG |
1780 | } |
1781 | ||
1782 | /* | |
1783 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1784 | * from the isolated freelists in the block we are migrating to. | |
1785 | */ | |
4e096ae1 | 1786 | static struct folio *compaction_alloc(struct folio *src, unsigned long data) |
748446bb MG |
1787 | { |
1788 | struct compact_control *cc = (struct compact_control *)data; | |
4e096ae1 | 1789 | struct folio *dst; |
748446bb | 1790 | |
748446bb | 1791 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1792 | isolate_freepages(cc); |
748446bb MG |
1793 | |
1794 | if (list_empty(&cc->freepages)) | |
1795 | return NULL; | |
1796 | } | |
1797 | ||
4e096ae1 MWO |
1798 | dst = list_entry(cc->freepages.next, struct folio, lru); |
1799 | list_del(&dst->lru); | |
748446bb | 1800 | cc->nr_freepages--; |
ab755bf4 | 1801 | cc->nr_migratepages--; |
748446bb | 1802 | |
4e096ae1 | 1803 | return dst; |
748446bb MG |
1804 | } |
1805 | ||
1806 | /* | |
d53aea3d DR |
1807 | * This is a migrate-callback that "frees" freepages back to the isolated |
1808 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1809 | * special handling needed for NUMA. | |
1810 | */ | |
4e096ae1 | 1811 | static void compaction_free(struct folio *dst, unsigned long data) |
d53aea3d DR |
1812 | { |
1813 | struct compact_control *cc = (struct compact_control *)data; | |
1814 | ||
4e096ae1 | 1815 | list_add(&dst->lru, &cc->freepages); |
d53aea3d | 1816 | cc->nr_freepages++; |
ab755bf4 | 1817 | cc->nr_migratepages++; |
d53aea3d DR |
1818 | } |
1819 | ||
ff9543fd MN |
1820 | /* possible outcome of isolate_migratepages */ |
1821 | typedef enum { | |
1822 | ISOLATE_ABORT, /* Abort compaction now */ | |
1823 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1824 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1825 | } isolate_migrate_t; | |
1826 | ||
5bbe3547 EM |
1827 | /* |
1828 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1829 | * compactable pages. | |
1830 | */ | |
48fe8ab8 MC |
1831 | static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT; |
1832 | /* | |
1833 | * Tunable for proactive compaction. It determines how | |
1834 | * aggressively the kernel should compact memory in the | |
1835 | * background. It takes values in the range [0, 100]. | |
1836 | */ | |
1837 | static unsigned int __read_mostly sysctl_compaction_proactiveness = 20; | |
1838 | static int sysctl_extfrag_threshold = 500; | |
8b9167cd | 1839 | static int __read_mostly sysctl_compact_memory; |
5bbe3547 | 1840 | |
70b44595 MG |
1841 | static inline void |
1842 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1843 | { | |
1844 | if (cc->fast_start_pfn == ULONG_MAX) | |
1845 | return; | |
1846 | ||
1847 | if (!cc->fast_start_pfn) | |
1848 | cc->fast_start_pfn = pfn; | |
1849 | ||
1850 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1851 | } | |
1852 | ||
1853 | static inline unsigned long | |
1854 | reinit_migrate_pfn(struct compact_control *cc) | |
1855 | { | |
1856 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1857 | return cc->migrate_pfn; | |
1858 | ||
1859 | cc->migrate_pfn = cc->fast_start_pfn; | |
1860 | cc->fast_start_pfn = ULONG_MAX; | |
1861 | ||
1862 | return cc->migrate_pfn; | |
1863 | } | |
1864 | ||
1865 | /* | |
1866 | * Briefly search the free lists for a migration source that already has | |
1867 | * some free pages to reduce the number of pages that need migration | |
1868 | * before a pageblock is free. | |
1869 | */ | |
1870 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1871 | { | |
1872 | unsigned int limit = freelist_scan_limit(cc); | |
1873 | unsigned int nr_scanned = 0; | |
1874 | unsigned long distance; | |
1875 | unsigned long pfn = cc->migrate_pfn; | |
1876 | unsigned long high_pfn; | |
1877 | int order; | |
15d28d0d | 1878 | bool found_block = false; |
70b44595 MG |
1879 | |
1880 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1881 | if (cc->ignore_skip_hint) | |
1882 | return pfn; | |
1883 | ||
f9d7fc1a MG |
1884 | /* |
1885 | * If the pageblock should be finished then do not select a different | |
1886 | * pageblock. | |
1887 | */ | |
1888 | if (cc->finish_pageblock) | |
1889 | return pfn; | |
1890 | ||
70b44595 MG |
1891 | /* |
1892 | * If the migrate_pfn is not at the start of a zone or the start | |
1893 | * of a pageblock then assume this is a continuation of a previous | |
1894 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1895 | */ | |
1896 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1897 | return pfn; | |
1898 | ||
1899 | /* | |
1900 | * For smaller orders, just linearly scan as the number of pages | |
1901 | * to migrate should be relatively small and does not necessarily | |
1902 | * justify freeing up a large block for a small allocation. | |
1903 | */ | |
1904 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1905 | return pfn; | |
1906 | ||
1907 | /* | |
1908 | * Only allow kcompactd and direct requests for movable pages to | |
1909 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1910 | * reduces the risk that a large movable pageblock is freed for | |
1911 | * an unmovable/reclaimable small allocation. | |
1912 | */ | |
1913 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1914 | return pfn; | |
1915 | ||
1916 | /* | |
1917 | * When starting the migration scanner, pick any pageblock within the | |
1918 | * first half of the search space. Otherwise try and pick a pageblock | |
1919 | * within the first eighth to reduce the chances that a migration | |
1920 | * target later becomes a source. | |
1921 | */ | |
1922 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1923 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1924 | distance >>= 2; | |
1925 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1926 | ||
1927 | for (order = cc->order - 1; | |
15d28d0d | 1928 | order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; |
70b44595 MG |
1929 | order--) { |
1930 | struct free_area *area = &cc->zone->free_area[order]; | |
1931 | struct list_head *freelist; | |
1932 | unsigned long flags; | |
1933 | struct page *freepage; | |
1934 | ||
1935 | if (!area->nr_free) | |
1936 | continue; | |
1937 | ||
1938 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1939 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
94ec2003 | 1940 | list_for_each_entry(freepage, freelist, buddy_list) { |
70b44595 MG |
1941 | unsigned long free_pfn; |
1942 | ||
15d28d0d WY |
1943 | if (nr_scanned++ >= limit) { |
1944 | move_freelist_tail(freelist, freepage); | |
1945 | break; | |
1946 | } | |
1947 | ||
70b44595 MG |
1948 | free_pfn = page_to_pfn(freepage); |
1949 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1950 | /* |
1951 | * Avoid if skipped recently. Ideally it would | |
1952 | * move to the tail but even safe iteration of | |
1953 | * the list assumes an entry is deleted, not | |
1954 | * reordered. | |
1955 | */ | |
15d28d0d | 1956 | if (get_pageblock_skip(freepage)) |
70b44595 | 1957 | continue; |
70b44595 MG |
1958 | |
1959 | /* Reorder to so a future search skips recent pages */ | |
1960 | move_freelist_tail(freelist, freepage); | |
1961 | ||
e380bebe | 1962 | update_fast_start_pfn(cc, free_pfn); |
70b44595 | 1963 | pfn = pageblock_start_pfn(free_pfn); |
bbe832b9 RY |
1964 | if (pfn < cc->zone->zone_start_pfn) |
1965 | pfn = cc->zone->zone_start_pfn; | |
70b44595 | 1966 | cc->fast_search_fail = 0; |
15d28d0d | 1967 | found_block = true; |
70b44595 MG |
1968 | break; |
1969 | } | |
70b44595 MG |
1970 | } |
1971 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1972 | } | |
1973 | ||
1974 | cc->total_migrate_scanned += nr_scanned; | |
1975 | ||
1976 | /* | |
1977 | * If fast scanning failed then use a cached entry for a page block | |
1978 | * that had free pages as the basis for starting a linear scan. | |
1979 | */ | |
15d28d0d WY |
1980 | if (!found_block) { |
1981 | cc->fast_search_fail++; | |
70b44595 | 1982 | pfn = reinit_migrate_pfn(cc); |
15d28d0d | 1983 | } |
70b44595 MG |
1984 | return pfn; |
1985 | } | |
1986 | ||
ff9543fd | 1987 | /* |
edc2ca61 VB |
1988 | * Isolate all pages that can be migrated from the first suitable block, |
1989 | * starting at the block pointed to by the migrate scanner pfn within | |
1990 | * compact_control. | |
ff9543fd | 1991 | */ |
32aaf055 | 1992 | static isolate_migrate_t isolate_migratepages(struct compact_control *cc) |
ff9543fd | 1993 | { |
e1409c32 JK |
1994 | unsigned long block_start_pfn; |
1995 | unsigned long block_end_pfn; | |
1996 | unsigned long low_pfn; | |
edc2ca61 VB |
1997 | struct page *page; |
1998 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1999 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 2000 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 2001 | bool fast_find_block; |
ff9543fd | 2002 | |
edc2ca61 VB |
2003 | /* |
2004 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
2005 | * initialized by compact_zone(). The first failure will use |
2006 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 2007 | */ |
70b44595 | 2008 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 2009 | block_start_pfn = pageblock_start_pfn(low_pfn); |
32aaf055 PL |
2010 | if (block_start_pfn < cc->zone->zone_start_pfn) |
2011 | block_start_pfn = cc->zone->zone_start_pfn; | |
ff9543fd | 2012 | |
70b44595 | 2013 | /* |
0aa8ea3c KS |
2014 | * fast_find_migrateblock() has already ensured the pageblock is not |
2015 | * set with a skipped flag, so to avoid the isolation_suitable check | |
2016 | * below again, check whether the fast search was successful. | |
70b44595 MG |
2017 | */ |
2018 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
2019 | ||
ff9543fd | 2020 | /* Only scan within a pageblock boundary */ |
06b6640a | 2021 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 2022 | |
edc2ca61 VB |
2023 | /* |
2024 | * Iterate over whole pageblocks until we find the first suitable. | |
2025 | * Do not cross the free scanner. | |
2026 | */ | |
e1409c32 | 2027 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 2028 | fast_find_block = false, |
c2ad7a1f | 2029 | cc->migrate_pfn = low_pfn = block_end_pfn, |
e1409c32 JK |
2030 | block_start_pfn = block_end_pfn, |
2031 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 2032 | |
edc2ca61 VB |
2033 | /* |
2034 | * This can potentially iterate a massively long zone with | |
2035 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 2036 | * need to schedule. |
edc2ca61 | 2037 | */ |
c036ddff | 2038 | if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 2039 | cond_resched(); |
ff9543fd | 2040 | |
32aaf055 PL |
2041 | page = pageblock_pfn_to_page(block_start_pfn, |
2042 | block_end_pfn, cc->zone); | |
9721fd82 BW |
2043 | if (!page) { |
2044 | unsigned long next_pfn; | |
2045 | ||
2046 | next_pfn = skip_offline_sections(block_start_pfn); | |
2047 | if (next_pfn) | |
2048 | block_end_pfn = min(next_pfn, cc->free_pfn); | |
edc2ca61 | 2049 | continue; |
9721fd82 | 2050 | } |
edc2ca61 | 2051 | |
e380bebe MG |
2052 | /* |
2053 | * If isolation recently failed, do not retry. Only check the | |
2054 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
2055 | * to be visited multiple times. Assume skip was checked | |
2056 | * before making it "skip" so other compaction instances do | |
2057 | * not scan the same block. | |
2058 | */ | |
493614da JW |
2059 | if ((pageblock_aligned(low_pfn) || |
2060 | low_pfn == cc->zone->zone_start_pfn) && | |
e380bebe | 2061 | !fast_find_block && !isolation_suitable(cc, page)) |
edc2ca61 VB |
2062 | continue; |
2063 | ||
2064 | /* | |
556162bf ML |
2065 | * For async direct compaction, only scan the pageblocks of the |
2066 | * same migratetype without huge pages. Async direct compaction | |
2067 | * is optimistic to see if the minimum amount of work satisfies | |
2068 | * the allocation. The cached PFN is updated as it's possible | |
2069 | * that all remaining blocks between source and target are | |
2070 | * unsuitable and the compaction scanners fail to meet. | |
edc2ca61 | 2071 | */ |
9bebefd5 MG |
2072 | if (!suitable_migration_source(cc, page)) { |
2073 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 2074 | continue; |
9bebefd5 | 2075 | } |
edc2ca61 VB |
2076 | |
2077 | /* Perform the isolation */ | |
c2ad7a1f OS |
2078 | if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, |
2079 | isolate_mode)) | |
edc2ca61 VB |
2080 | return ISOLATE_ABORT; |
2081 | ||
2082 | /* | |
2083 | * Either we isolated something and proceed with migration. Or | |
2084 | * we failed and compact_zone should decide if we should | |
2085 | * continue or not. | |
2086 | */ | |
2087 | break; | |
2088 | } | |
2089 | ||
edc2ca61 | 2090 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
2091 | } |
2092 | ||
21c527a3 | 2093 | /* |
9cc17ede KS |
2094 | * order == -1 is expected when compacting proactively via |
2095 | * 1. /proc/sys/vm/compact_memory | |
2096 | * 2. /sys/devices/system/node/nodex/compact | |
2097 | * 3. /proc/sys/vm/compaction_proactiveness | |
21c527a3 YB |
2098 | */ |
2099 | static inline bool is_via_compact_memory(int order) | |
2100 | { | |
2101 | return order == -1; | |
2102 | } | |
2103 | ||
b4a0215e KW |
2104 | /* |
2105 | * Determine whether kswapd is (or recently was!) running on this node. | |
2106 | * | |
2107 | * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't | |
2108 | * zero it. | |
2109 | */ | |
facdaa91 NG |
2110 | static bool kswapd_is_running(pg_data_t *pgdat) |
2111 | { | |
b4a0215e KW |
2112 | bool running; |
2113 | ||
2114 | pgdat_kswapd_lock(pgdat); | |
2115 | running = pgdat->kswapd && task_is_running(pgdat->kswapd); | |
2116 | pgdat_kswapd_unlock(pgdat); | |
2117 | ||
2118 | return running; | |
facdaa91 NG |
2119 | } |
2120 | ||
2121 | /* | |
2122 | * A zone's fragmentation score is the external fragmentation wrt to the | |
40d7e203 CTR |
2123 | * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. |
2124 | */ | |
2125 | static unsigned int fragmentation_score_zone(struct zone *zone) | |
2126 | { | |
2127 | return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); | |
2128 | } | |
2129 | ||
2130 | /* | |
2131 | * A weighted zone's fragmentation score is the external fragmentation | |
2132 | * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It | |
2133 | * returns a value in the range [0, 100]. | |
facdaa91 NG |
2134 | * |
2135 | * The scaling factor ensures that proactive compaction focuses on larger | |
2136 | * zones like ZONE_NORMAL, rather than smaller, specialized zones like | |
2137 | * ZONE_DMA32. For smaller zones, the score value remains close to zero, | |
2138 | * and thus never exceeds the high threshold for proactive compaction. | |
2139 | */ | |
40d7e203 | 2140 | static unsigned int fragmentation_score_zone_weighted(struct zone *zone) |
facdaa91 NG |
2141 | { |
2142 | unsigned long score; | |
2143 | ||
40d7e203 | 2144 | score = zone->present_pages * fragmentation_score_zone(zone); |
facdaa91 NG |
2145 | return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); |
2146 | } | |
2147 | ||
2148 | /* | |
2149 | * The per-node proactive (background) compaction process is started by its | |
2150 | * corresponding kcompactd thread when the node's fragmentation score | |
2151 | * exceeds the high threshold. The compaction process remains active till | |
2152 | * the node's score falls below the low threshold, or one of the back-off | |
2153 | * conditions is met. | |
2154 | */ | |
d34c0a75 | 2155 | static unsigned int fragmentation_score_node(pg_data_t *pgdat) |
facdaa91 | 2156 | { |
d34c0a75 | 2157 | unsigned int score = 0; |
facdaa91 NG |
2158 | int zoneid; |
2159 | ||
2160 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2161 | struct zone *zone; | |
2162 | ||
2163 | zone = &pgdat->node_zones[zoneid]; | |
9e552271 BW |
2164 | if (!populated_zone(zone)) |
2165 | continue; | |
40d7e203 | 2166 | score += fragmentation_score_zone_weighted(zone); |
facdaa91 NG |
2167 | } |
2168 | ||
2169 | return score; | |
2170 | } | |
2171 | ||
8fbb92bd | 2172 | static unsigned int fragmentation_score_wmark(bool low) |
facdaa91 | 2173 | { |
d34c0a75 | 2174 | unsigned int wmark_low; |
facdaa91 NG |
2175 | |
2176 | /* | |
f0953a1b IM |
2177 | * Cap the low watermark to avoid excessive compaction |
2178 | * activity in case a user sets the proactiveness tunable | |
facdaa91 NG |
2179 | * close to 100 (maximum). |
2180 | */ | |
d34c0a75 NG |
2181 | wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); |
2182 | return low ? wmark_low : min(wmark_low + 10, 100U); | |
facdaa91 NG |
2183 | } |
2184 | ||
2185 | static bool should_proactive_compact_node(pg_data_t *pgdat) | |
2186 | { | |
2187 | int wmark_high; | |
2188 | ||
2189 | if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) | |
2190 | return false; | |
2191 | ||
8fbb92bd | 2192 | wmark_high = fragmentation_score_wmark(false); |
facdaa91 NG |
2193 | return fragmentation_score_node(pgdat) > wmark_high; |
2194 | } | |
2195 | ||
40cacbcb | 2196 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 2197 | { |
8fb74b9f | 2198 | unsigned int order; |
d39773a0 | 2199 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 2200 | int ret; |
748446bb | 2201 | |
753341a4 | 2202 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 2203 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 2204 | /* Let the next compaction start anew. */ |
40cacbcb | 2205 | reset_cached_positions(cc->zone); |
55b7c4c9 | 2206 | |
62997027 MG |
2207 | /* |
2208 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 2209 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
2210 | * flag itself as the decision to be clear should be directly |
2211 | * based on an allocation request. | |
2212 | */ | |
accf6242 | 2213 | if (cc->direct_compaction) |
40cacbcb | 2214 | cc->zone->compact_blockskip_flush = true; |
62997027 | 2215 | |
c8f7de0b MH |
2216 | if (cc->whole_zone) |
2217 | return COMPACT_COMPLETE; | |
2218 | else | |
2219 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 2220 | } |
748446bb | 2221 | |
facdaa91 NG |
2222 | if (cc->proactive_compaction) { |
2223 | int score, wmark_low; | |
2224 | pg_data_t *pgdat; | |
2225 | ||
2226 | pgdat = cc->zone->zone_pgdat; | |
2227 | if (kswapd_is_running(pgdat)) | |
2228 | return COMPACT_PARTIAL_SKIPPED; | |
2229 | ||
2230 | score = fragmentation_score_zone(cc->zone); | |
8fbb92bd | 2231 | wmark_low = fragmentation_score_wmark(true); |
facdaa91 NG |
2232 | |
2233 | if (score > wmark_low) | |
2234 | ret = COMPACT_CONTINUE; | |
2235 | else | |
2236 | ret = COMPACT_SUCCESS; | |
2237 | ||
2238 | goto out; | |
2239 | } | |
2240 | ||
21c527a3 | 2241 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
2242 | return COMPACT_CONTINUE; |
2243 | ||
efe771c7 MG |
2244 | /* |
2245 | * Always finish scanning a pageblock to reduce the possibility of | |
2246 | * fallbacks in the future. This is particularly important when | |
2247 | * migration source is unmovable/reclaimable but it's not worth | |
2248 | * special casing. | |
2249 | */ | |
ee0913c4 | 2250 | if (!pageblock_aligned(cc->migrate_pfn)) |
efe771c7 | 2251 | return COMPACT_CONTINUE; |
baf6a9a1 | 2252 | |
56de7263 | 2253 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 2254 | ret = COMPACT_NO_SUITABLE_PAGE; |
fd377218 | 2255 | for (order = cc->order; order < NR_PAGE_ORDERS; order++) { |
40cacbcb | 2256 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 2257 | bool can_steal; |
8fb74b9f MG |
2258 | |
2259 | /* Job done if page is free of the right migratetype */ | |
b03641af | 2260 | if (!free_area_empty(area, migratetype)) |
cf378319 | 2261 | return COMPACT_SUCCESS; |
8fb74b9f | 2262 | |
2149cdae JK |
2263 | #ifdef CONFIG_CMA |
2264 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
2265 | if (migratetype == MIGRATE_MOVABLE && | |
b03641af | 2266 | !free_area_empty(area, MIGRATE_CMA)) |
cf378319 | 2267 | return COMPACT_SUCCESS; |
2149cdae JK |
2268 | #endif |
2269 | /* | |
2270 | * Job done if allocation would steal freepages from | |
2271 | * other migratetype buddy lists. | |
2272 | */ | |
2273 | if (find_suitable_fallback(area, order, migratetype, | |
fa599c44 | 2274 | true, &can_steal) != -1) |
baf6a9a1 | 2275 | /* |
fa599c44 ML |
2276 | * Movable pages are OK in any pageblock. If we are |
2277 | * stealing for a non-movable allocation, make sure | |
2278 | * we finish compacting the current pageblock first | |
2279 | * (which is assured by the above migrate_pfn align | |
2280 | * check) so it is as free as possible and we won't | |
2281 | * have to steal another one soon. | |
baf6a9a1 | 2282 | */ |
fa599c44 | 2283 | return COMPACT_SUCCESS; |
56de7263 MG |
2284 | } |
2285 | ||
facdaa91 | 2286 | out: |
cb2dcaf0 MG |
2287 | if (cc->contended || fatal_signal_pending(current)) |
2288 | ret = COMPACT_CONTENDED; | |
2289 | ||
2290 | return ret; | |
837d026d JK |
2291 | } |
2292 | ||
40cacbcb | 2293 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
2294 | { |
2295 | int ret; | |
2296 | ||
40cacbcb MG |
2297 | ret = __compact_finished(cc); |
2298 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
2299 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
2300 | ret = COMPACT_CONTINUE; | |
2301 | ||
2302 | return ret; | |
748446bb MG |
2303 | } |
2304 | ||
3cf04937 JW |
2305 | static bool __compaction_suitable(struct zone *zone, int order, |
2306 | int highest_zoneidx, | |
2307 | unsigned long wmark_target) | |
3e7d3449 | 2308 | { |
3e7d3449 | 2309 | unsigned long watermark; |
3e7d3449 | 2310 | /* |
9861a62c | 2311 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
2312 | * isolate free pages for migration targets. This means that the |
2313 | * watermark and alloc_flags have to match, or be more pessimistic than | |
2314 | * the check in __isolate_free_page(). We don't use the direct | |
2315 | * compactor's alloc_flags, as they are not relevant for freepage | |
97a225e6 JK |
2316 | * isolation. We however do use the direct compactor's highest_zoneidx |
2317 | * to skip over zones where lowmem reserves would prevent allocation | |
2318 | * even if compaction succeeds. | |
8348faf9 VB |
2319 | * For costly orders, we require low watermark instead of min for |
2320 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
2321 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
2322 | * suitable migration targets | |
3e7d3449 | 2323 | */ |
8348faf9 VB |
2324 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
2325 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
2326 | watermark += compact_gap(order); | |
3cf04937 JW |
2327 | return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx, |
2328 | ALLOC_CMA, wmark_target); | |
cc5c9f09 VB |
2329 | } |
2330 | ||
2b1a20c3 HS |
2331 | /* |
2332 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
2b1a20c3 | 2333 | */ |
3cf04937 | 2334 | bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx) |
cc5c9f09 | 2335 | { |
3cf04937 JW |
2336 | enum compact_result compact_result; |
2337 | bool suitable; | |
cc5c9f09 | 2338 | |
3cf04937 JW |
2339 | suitable = __compaction_suitable(zone, order, highest_zoneidx, |
2340 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
2341 | /* |
2342 | * fragmentation index determines if allocation failures are due to | |
2343 | * low memory or external fragmentation | |
2344 | * | |
ebff3980 VB |
2345 | * index of -1000 would imply allocations might succeed depending on |
2346 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2347 | * index towards 0 implies failure is due to lack of memory |
2348 | * index towards 1000 implies failure is due to fragmentation | |
2349 | * | |
20311420 VB |
2350 | * Only compact if a failure would be due to fragmentation. Also |
2351 | * ignore fragindex for non-costly orders where the alternative to | |
2352 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2353 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2354 | * excessive compaction for costly orders, but it should not be at the | |
2355 | * expense of system stability. | |
3e7d3449 | 2356 | */ |
3cf04937 JW |
2357 | if (suitable) { |
2358 | compact_result = COMPACT_CONTINUE; | |
2359 | if (order > PAGE_ALLOC_COSTLY_ORDER) { | |
2360 | int fragindex = fragmentation_index(zone, order); | |
2361 | ||
2362 | if (fragindex >= 0 && | |
2363 | fragindex <= sysctl_extfrag_threshold) { | |
2364 | suitable = false; | |
2365 | compact_result = COMPACT_NOT_SUITABLE_ZONE; | |
2366 | } | |
2367 | } | |
2368 | } else { | |
2369 | compact_result = COMPACT_SKIPPED; | |
cc5c9f09 | 2370 | } |
837d026d | 2371 | |
3cf04937 | 2372 | trace_mm_compaction_suitable(zone, order, compact_result); |
837d026d | 2373 | |
3cf04937 | 2374 | return suitable; |
837d026d JK |
2375 | } |
2376 | ||
86a294a8 MH |
2377 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2378 | int alloc_flags) | |
2379 | { | |
2380 | struct zone *zone; | |
2381 | struct zoneref *z; | |
2382 | ||
2383 | /* | |
2384 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2385 | * retrying the reclaim. | |
2386 | */ | |
97a225e6 JK |
2387 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2388 | ac->highest_zoneidx, ac->nodemask) { | |
86a294a8 | 2389 | unsigned long available; |
86a294a8 MH |
2390 | |
2391 | /* | |
2392 | * Do not consider all the reclaimable memory because we do not | |
2393 | * want to trash just for a single high order allocation which | |
2394 | * is even not guaranteed to appear even if __compaction_suitable | |
2395 | * is happy about the watermark check. | |
2396 | */ | |
5a1c84b4 | 2397 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 | 2398 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
e8606320 | 2399 | if (__compaction_suitable(zone, order, ac->highest_zoneidx, |
3cf04937 | 2400 | available)) |
86a294a8 MH |
2401 | return true; |
2402 | } | |
2403 | ||
2404 | return false; | |
2405 | } | |
2406 | ||
e19a3f59 KS |
2407 | /* |
2408 | * Should we do compaction for target allocation order. | |
2409 | * Return COMPACT_SUCCESS if allocation for target order can be already | |
2410 | * satisfied | |
2411 | * Return COMPACT_SKIPPED if compaction for target order is likely to fail | |
2412 | * Return COMPACT_CONTINUE if compaction for target order should be ran | |
2413 | */ | |
2414 | static enum compact_result | |
2415 | compaction_suit_allocation_order(struct zone *zone, unsigned int order, | |
2416 | int highest_zoneidx, unsigned int alloc_flags) | |
2417 | { | |
2418 | unsigned long watermark; | |
2419 | ||
2420 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); | |
2421 | if (zone_watermark_ok(zone, order, watermark, highest_zoneidx, | |
2422 | alloc_flags)) | |
2423 | return COMPACT_SUCCESS; | |
2424 | ||
2425 | if (!compaction_suitable(zone, order, highest_zoneidx)) | |
2426 | return COMPACT_SKIPPED; | |
2427 | ||
2428 | return COMPACT_CONTINUE; | |
2429 | } | |
2430 | ||
5e1f0f09 MG |
2431 | static enum compact_result |
2432 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2433 | { |
ea7ab982 | 2434 | enum compact_result ret; |
40cacbcb MG |
2435 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2436 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2437 | unsigned long last_migrated_pfn; |
e0b9daeb | 2438 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2439 | bool update_cached; |
ab755bf4 | 2440 | unsigned int nr_succeeded = 0, nr_migratepages; |
748446bb | 2441 | |
a94b5252 YS |
2442 | /* |
2443 | * These counters track activities during zone compaction. Initialize | |
2444 | * them before compacting a new zone. | |
2445 | */ | |
2446 | cc->total_migrate_scanned = 0; | |
2447 | cc->total_free_scanned = 0; | |
2448 | cc->nr_migratepages = 0; | |
2449 | cc->nr_freepages = 0; | |
2450 | INIT_LIST_HEAD(&cc->freepages); | |
2451 | INIT_LIST_HEAD(&cc->migratepages); | |
2452 | ||
01c0bfe0 | 2453 | cc->migratetype = gfp_migratetype(cc->gfp_mask); |
e8606320 JW |
2454 | |
2455 | if (!is_via_compact_memory(cc->order)) { | |
e19a3f59 KS |
2456 | ret = compaction_suit_allocation_order(cc->zone, cc->order, |
2457 | cc->highest_zoneidx, | |
2458 | cc->alloc_flags); | |
2459 | if (ret != COMPACT_CONTINUE) | |
2460 | return ret; | |
e8606320 | 2461 | } |
c46649de | 2462 | |
d3132e4b VB |
2463 | /* |
2464 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2465 | * is about to be retried after being deferred. |
d3132e4b | 2466 | */ |
40cacbcb MG |
2467 | if (compaction_restarting(cc->zone, cc->order)) |
2468 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2469 | |
c89511ab MG |
2470 | /* |
2471 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2472 | * information on where the scanners should start (unless we explicitly |
2473 | * want to compact the whole zone), but check that it is initialised | |
2474 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2475 | */ |
70b44595 | 2476 | cc->fast_start_pfn = 0; |
06ed2998 | 2477 | if (cc->whole_zone) { |
c89511ab | 2478 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2479 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2480 | } else { | |
40cacbcb MG |
2481 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2482 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2483 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2484 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2485 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2486 | } |
2487 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2488 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2489 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2490 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2491 | } |
c8f7de0b | 2492 | |
e332f741 | 2493 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2494 | cc->whole_zone = true; |
2495 | } | |
c8f7de0b | 2496 | |
566e54e1 | 2497 | last_migrated_pfn = 0; |
748446bb | 2498 | |
8854c55f MG |
2499 | /* |
2500 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2501 | * the basis that some migrations will fail in ASYNC mode. However, | |
2502 | * if the cached PFNs match and pageblocks are skipped due to having | |
2503 | * no isolation candidates, then the sync state does not matter. | |
2504 | * Until a pageblock with isolation candidates is found, keep the | |
2505 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2506 | */ | |
2507 | update_cached = !sync && | |
2508 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2509 | ||
abd4349f | 2510 | trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); |
0eb927c0 | 2511 | |
361a2a22 MK |
2512 | /* lru_add_drain_all could be expensive with involving other CPUs */ |
2513 | lru_add_drain(); | |
748446bb | 2514 | |
40cacbcb | 2515 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2516 | int err; |
19d3cf9d | 2517 | unsigned long iteration_start_pfn = cc->migrate_pfn; |
748446bb | 2518 | |
804d3121 | 2519 | /* |
48731c84 MG |
2520 | * Avoid multiple rescans of the same pageblock which can |
2521 | * happen if a page cannot be isolated (dirty/writeback in | |
2522 | * async mode) or if the migrated pages are being allocated | |
2523 | * before the pageblock is cleared. The first rescan will | |
2524 | * capture the entire pageblock for migration. If it fails, | |
2525 | * it'll be marked skip and scanning will proceed as normal. | |
804d3121 | 2526 | */ |
48731c84 | 2527 | cc->finish_pageblock = false; |
804d3121 | 2528 | if (pageblock_start_pfn(last_migrated_pfn) == |
19d3cf9d | 2529 | pageblock_start_pfn(iteration_start_pfn)) { |
48731c84 | 2530 | cc->finish_pageblock = true; |
804d3121 MG |
2531 | } |
2532 | ||
cfccd2e6 | 2533 | rescan: |
32aaf055 | 2534 | switch (isolate_migratepages(cc)) { |
f9e35b3b | 2535 | case ISOLATE_ABORT: |
2d1e1041 | 2536 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2537 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2538 | cc->nr_migratepages = 0; |
f9e35b3b MG |
2539 | goto out; |
2540 | case ISOLATE_NONE: | |
8854c55f MG |
2541 | if (update_cached) { |
2542 | cc->zone->compact_cached_migrate_pfn[1] = | |
2543 | cc->zone->compact_cached_migrate_pfn[0]; | |
2544 | } | |
2545 | ||
fdaf7f5c VB |
2546 | /* |
2547 | * We haven't isolated and migrated anything, but | |
2548 | * there might still be unflushed migrations from | |
2549 | * previous cc->order aligned block. | |
2550 | */ | |
2551 | goto check_drain; | |
f9e35b3b | 2552 | case ISOLATE_SUCCESS: |
8854c55f | 2553 | update_cached = false; |
7c0a84bd KS |
2554 | last_migrated_pfn = max(cc->zone->zone_start_pfn, |
2555 | pageblock_start_pfn(cc->migrate_pfn - 1)); | |
f9e35b3b | 2556 | } |
748446bb | 2557 | |
ab755bf4 BW |
2558 | /* |
2559 | * Record the number of pages to migrate since the | |
2560 | * compaction_alloc/free() will update cc->nr_migratepages | |
2561 | * properly. | |
2562 | */ | |
2563 | nr_migratepages = cc->nr_migratepages; | |
d53aea3d | 2564 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2565 | compaction_free, (unsigned long)cc, cc->mode, |
84b328aa | 2566 | MR_COMPACTION, &nr_succeeded); |
748446bb | 2567 | |
ab755bf4 | 2568 | trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded); |
748446bb | 2569 | |
f8c9301f VB |
2570 | /* All pages were either migrated or will be released */ |
2571 | cc->nr_migratepages = 0; | |
9d502c1c | 2572 | if (err) { |
5733c7d1 | 2573 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2574 | /* |
2575 | * migrate_pages() may return -ENOMEM when scanners meet | |
2576 | * and we want compact_finished() to detect it | |
2577 | */ | |
f2849aa0 | 2578 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2579 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2580 | goto out; |
2581 | } | |
fdd048e1 | 2582 | /* |
cfccd2e6 | 2583 | * If an ASYNC or SYNC_LIGHT fails to migrate a page |
c3750cc7 | 2584 | * within the pageblock_order-aligned block and |
9ecc5fc5 | 2585 | * fast_find_migrateblock may be used then scan the |
cfccd2e6 MG |
2586 | * remainder of the pageblock. This will mark the |
2587 | * pageblock "skip" to avoid rescanning in the near | |
2588 | * future. This will isolate more pages than necessary | |
2589 | * for the request but avoid loops due to | |
2590 | * fast_find_migrateblock revisiting blocks that were | |
2591 | * recently partially scanned. | |
fdd048e1 | 2592 | */ |
539aa041 | 2593 | if (!pageblock_aligned(cc->migrate_pfn) && |
9ecc5fc5 | 2594 | !cc->ignore_skip_hint && !cc->finish_pageblock && |
539aa041 | 2595 | (cc->mode < MIGRATE_SYNC)) { |
cfccd2e6 MG |
2596 | cc->finish_pageblock = true; |
2597 | ||
2598 | /* | |
2599 | * Draining pcplists does not help THP if | |
2600 | * any page failed to migrate. Even after | |
2601 | * drain, the pageblock will not be free. | |
2602 | */ | |
2603 | if (cc->order == COMPACTION_HPAGE_ORDER) | |
2604 | last_migrated_pfn = 0; | |
2605 | ||
2606 | goto rescan; | |
fdd048e1 | 2607 | } |
748446bb | 2608 | } |
fdaf7f5c | 2609 | |
16b3be40 MG |
2610 | /* Stop if a page has been captured */ |
2611 | if (capc && capc->page) { | |
2612 | ret = COMPACT_SUCCESS; | |
2613 | break; | |
2614 | } | |
2615 | ||
fdaf7f5c VB |
2616 | check_drain: |
2617 | /* | |
2618 | * Has the migration scanner moved away from the previous | |
2619 | * cc->order aligned block where we migrated from? If yes, | |
2620 | * flush the pages that were freed, so that they can merge and | |
2621 | * compact_finished() can detect immediately if allocation | |
2622 | * would succeed. | |
2623 | */ | |
566e54e1 | 2624 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c | 2625 | unsigned long current_block_start = |
06b6640a | 2626 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2627 | |
566e54e1 | 2628 | if (last_migrated_pfn < current_block_start) { |
b01b2141 | 2629 | lru_add_drain_cpu_zone(cc->zone); |
fdaf7f5c | 2630 | /* No more flushing until we migrate again */ |
566e54e1 | 2631 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2632 | } |
2633 | } | |
748446bb MG |
2634 | } |
2635 | ||
f9e35b3b | 2636 | out: |
6bace090 VB |
2637 | /* |
2638 | * Release free pages and update where the free scanner should restart, | |
2639 | * so we don't leave any returned pages behind in the next attempt. | |
2640 | */ | |
2641 | if (cc->nr_freepages > 0) { | |
2642 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2643 | ||
2644 | cc->nr_freepages = 0; | |
2645 | VM_BUG_ON(free_pfn == 0); | |
2646 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2647 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2648 | /* |
2649 | * Only go back, not forward. The cached pfn might have been | |
2650 | * already reset to zone end in compact_finished() | |
2651 | */ | |
40cacbcb MG |
2652 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2653 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2654 | } |
748446bb | 2655 | |
7f354a54 DR |
2656 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2657 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2658 | ||
abd4349f | 2659 | trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); |
0eb927c0 | 2660 | |
753ec50d BW |
2661 | VM_BUG_ON(!list_empty(&cc->freepages)); |
2662 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
2663 | ||
748446bb MG |
2664 | return ret; |
2665 | } | |
76ab0f53 | 2666 | |
ea7ab982 | 2667 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2668 | gfp_t gfp_mask, enum compact_priority prio, |
97a225e6 | 2669 | unsigned int alloc_flags, int highest_zoneidx, |
5e1f0f09 | 2670 | struct page **capture) |
56de7263 | 2671 | { |
ea7ab982 | 2672 | enum compact_result ret; |
56de7263 | 2673 | struct compact_control cc = { |
56de7263 | 2674 | .order = order, |
dbe2d4e4 | 2675 | .search_order = order, |
6d7ce559 | 2676 | .gfp_mask = gfp_mask, |
56de7263 | 2677 | .zone = zone, |
a5508cd8 VB |
2678 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2679 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 | 2680 | .alloc_flags = alloc_flags, |
97a225e6 | 2681 | .highest_zoneidx = highest_zoneidx, |
accf6242 | 2682 | .direct_compaction = true, |
a8e025e5 | 2683 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2684 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2685 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2686 | }; |
5e1f0f09 MG |
2687 | struct capture_control capc = { |
2688 | .cc = &cc, | |
2689 | .page = NULL, | |
2690 | }; | |
2691 | ||
b9e20f0d VB |
2692 | /* |
2693 | * Make sure the structs are really initialized before we expose the | |
2694 | * capture control, in case we are interrupted and the interrupt handler | |
2695 | * frees a page. | |
2696 | */ | |
2697 | barrier(); | |
2698 | WRITE_ONCE(current->capture_control, &capc); | |
56de7263 | 2699 | |
5e1f0f09 | 2700 | ret = compact_zone(&cc, &capc); |
e64c5237 | 2701 | |
b9e20f0d VB |
2702 | /* |
2703 | * Make sure we hide capture control first before we read the captured | |
2704 | * page pointer, otherwise an interrupt could free and capture a page | |
2705 | * and we would leak it. | |
2706 | */ | |
2707 | WRITE_ONCE(current->capture_control, NULL); | |
2708 | *capture = READ_ONCE(capc.page); | |
06dac2f4 CTR |
2709 | /* |
2710 | * Technically, it is also possible that compaction is skipped but | |
2711 | * the page is still captured out of luck(IRQ came and freed the page). | |
2712 | * Returning COMPACT_SUCCESS in such cases helps in properly accounting | |
2713 | * the COMPACT[STALL|FAIL] when compaction is skipped. | |
2714 | */ | |
2715 | if (*capture) | |
2716 | ret = COMPACT_SUCCESS; | |
5e1f0f09 | 2717 | |
e64c5237 | 2718 | return ret; |
56de7263 MG |
2719 | } |
2720 | ||
2721 | /** | |
2722 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2723 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2724 | * @order: The order of the current allocation |
2725 | * @alloc_flags: The allocation flags of the current allocation | |
2726 | * @ac: The context of current allocation | |
112d2d29 | 2727 | * @prio: Determines how hard direct compaction should try to succeed |
6467552c | 2728 | * @capture: Pointer to free page created by compaction will be stored here |
56de7263 MG |
2729 | * |
2730 | * This is the main entry point for direct page compaction. | |
2731 | */ | |
ea7ab982 | 2732 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2733 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2734 | enum compact_priority prio, struct page **capture) |
56de7263 | 2735 | { |
fe573327 | 2736 | int may_perform_io = (__force int)(gfp_mask & __GFP_IO); |
56de7263 MG |
2737 | struct zoneref *z; |
2738 | struct zone *zone; | |
1d4746d3 | 2739 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2740 | |
73e64c51 MH |
2741 | /* |
2742 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2743 | * tricky context because the migration might require IO | |
2744 | */ | |
2745 | if (!may_perform_io) | |
53853e2d | 2746 | return COMPACT_SKIPPED; |
56de7263 | 2747 | |
a5508cd8 | 2748 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2749 | |
56de7263 | 2750 | /* Compact each zone in the list */ |
97a225e6 JK |
2751 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2752 | ac->highest_zoneidx, ac->nodemask) { | |
ea7ab982 | 2753 | enum compact_result status; |
56de7263 | 2754 | |
a8e025e5 VB |
2755 | if (prio > MIN_COMPACT_PRIORITY |
2756 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2757 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2758 | continue; |
1d4746d3 | 2759 | } |
53853e2d | 2760 | |
a5508cd8 | 2761 | status = compact_zone_order(zone, order, gfp_mask, prio, |
97a225e6 | 2762 | alloc_flags, ac->highest_zoneidx, capture); |
56de7263 MG |
2763 | rc = max(status, rc); |
2764 | ||
7ceb009a VB |
2765 | /* The allocation should succeed, stop compacting */ |
2766 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2767 | /* |
2768 | * We think the allocation will succeed in this zone, | |
2769 | * but it is not certain, hence the false. The caller | |
2770 | * will repeat this with true if allocation indeed | |
2771 | * succeeds in this zone. | |
2772 | */ | |
2773 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2774 | |
c3486f53 | 2775 | break; |
1f9efdef VB |
2776 | } |
2777 | ||
a5508cd8 | 2778 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2779 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2780 | /* |
2781 | * We think that allocation won't succeed in this zone | |
2782 | * so we defer compaction there. If it ends up | |
2783 | * succeeding after all, it will be reset. | |
2784 | */ | |
2785 | defer_compaction(zone, order); | |
1f9efdef VB |
2786 | |
2787 | /* | |
2788 | * We might have stopped compacting due to need_resched() in | |
2789 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2790 | * case do not try further zones |
1f9efdef | 2791 | */ |
c3486f53 VB |
2792 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2793 | || fatal_signal_pending(current)) | |
2794 | break; | |
56de7263 MG |
2795 | } |
2796 | ||
2797 | return rc; | |
2798 | } | |
2799 | ||
facdaa91 | 2800 | /* |
3e40b3f4 KW |
2801 | * compact_node() - compact all zones within a node |
2802 | * @pgdat: The node page data | |
2803 | * @proactive: Whether the compaction is proactive | |
facdaa91 | 2804 | * |
3e40b3f4 KW |
2805 | * For proactive compaction, compact till each zone's fragmentation score |
2806 | * reaches within proactive compaction thresholds (as determined by the | |
2807 | * proactiveness tunable), it is possible that the function returns before | |
2808 | * reaching score targets due to various back-off conditions, such as, | |
2809 | * contention on per-node or per-zone locks. | |
facdaa91 | 2810 | */ |
3e40b3f4 | 2811 | static void compact_node(pg_data_t *pgdat, bool proactive) |
facdaa91 NG |
2812 | { |
2813 | int zoneid; | |
2814 | struct zone *zone; | |
2815 | struct compact_control cc = { | |
2816 | .order = -1, | |
3e40b3f4 | 2817 | .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC, |
facdaa91 NG |
2818 | .ignore_skip_hint = true, |
2819 | .whole_zone = true, | |
2820 | .gfp_mask = GFP_KERNEL, | |
3e40b3f4 | 2821 | .proactive_compaction = proactive, |
facdaa91 NG |
2822 | }; |
2823 | ||
2824 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2825 | zone = &pgdat->node_zones[zoneid]; | |
2826 | if (!populated_zone(zone)) | |
2827 | continue; | |
2828 | ||
2829 | cc.zone = zone; | |
2830 | ||
2831 | compact_zone(&cc, NULL); | |
2832 | ||
3e40b3f4 KW |
2833 | if (proactive) { |
2834 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, | |
2835 | cc.total_migrate_scanned); | |
2836 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2837 | cc.total_free_scanned); | |
2838 | } | |
76ab0f53 | 2839 | } |
76ab0f53 MG |
2840 | } |
2841 | ||
3e40b3f4 | 2842 | /* Compact all zones of all nodes in the system */ |
7964c06d | 2843 | static void compact_nodes(void) |
76ab0f53 MG |
2844 | { |
2845 | int nid; | |
2846 | ||
8575ec29 HD |
2847 | /* Flush pending updates to the LRU lists */ |
2848 | lru_add_drain_all(); | |
2849 | ||
76ab0f53 | 2850 | for_each_online_node(nid) |
3e40b3f4 | 2851 | compact_node(NODE_DATA(nid), false); |
76ab0f53 MG |
2852 | } |
2853 | ||
48fe8ab8 | 2854 | static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, |
65d759c8 CTR |
2855 | void *buffer, size_t *length, loff_t *ppos) |
2856 | { | |
2857 | int rc, nid; | |
2858 | ||
2859 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
2860 | if (rc) | |
2861 | return rc; | |
2862 | ||
2863 | if (write && sysctl_compaction_proactiveness) { | |
2864 | for_each_online_node(nid) { | |
2865 | pg_data_t *pgdat = NODE_DATA(nid); | |
2866 | ||
2867 | if (pgdat->proactive_compact_trigger) | |
2868 | continue; | |
2869 | ||
2870 | pgdat->proactive_compact_trigger = true; | |
8fff8b6f BW |
2871 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1, |
2872 | pgdat->nr_zones - 1); | |
65d759c8 CTR |
2873 | wake_up_interruptible(&pgdat->kcompactd_wait); |
2874 | } | |
2875 | } | |
2876 | ||
2877 | return 0; | |
2878 | } | |
2879 | ||
fec4eb2c YB |
2880 | /* |
2881 | * This is the entry point for compacting all nodes via | |
2882 | * /proc/sys/vm/compact_memory | |
2883 | */ | |
48fe8ab8 | 2884 | static int sysctl_compaction_handler(struct ctl_table *table, int write, |
32927393 | 2885 | void *buffer, size_t *length, loff_t *ppos) |
76ab0f53 | 2886 | { |
8b9167cd WY |
2887 | int ret; |
2888 | ||
2889 | ret = proc_dointvec(table, write, buffer, length, ppos); | |
2890 | if (ret) | |
2891 | return ret; | |
2892 | ||
2893 | if (sysctl_compact_memory != 1) | |
2894 | return -EINVAL; | |
2895 | ||
76ab0f53 | 2896 | if (write) |
7964c06d | 2897 | compact_nodes(); |
76ab0f53 MG |
2898 | |
2899 | return 0; | |
2900 | } | |
ed4a6d7f MG |
2901 | |
2902 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
17adb230 Y |
2903 | static ssize_t compact_store(struct device *dev, |
2904 | struct device_attribute *attr, | |
2905 | const char *buf, size_t count) | |
ed4a6d7f | 2906 | { |
8575ec29 HD |
2907 | int nid = dev->id; |
2908 | ||
2909 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2910 | /* Flush pending updates to the LRU lists */ | |
2911 | lru_add_drain_all(); | |
2912 | ||
3e40b3f4 | 2913 | compact_node(NODE_DATA(nid), false); |
8575ec29 | 2914 | } |
ed4a6d7f MG |
2915 | |
2916 | return count; | |
2917 | } | |
17adb230 | 2918 | static DEVICE_ATTR_WO(compact); |
ed4a6d7f MG |
2919 | |
2920 | int compaction_register_node(struct node *node) | |
2921 | { | |
10fbcf4c | 2922 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2923 | } |
2924 | ||
2925 | void compaction_unregister_node(struct node *node) | |
2926 | { | |
f82024cb | 2927 | device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2928 | } |
2929 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2930 | |
698b1b30 VB |
2931 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2932 | { | |
65d759c8 CTR |
2933 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || |
2934 | pgdat->proactive_compact_trigger; | |
698b1b30 VB |
2935 | } |
2936 | ||
2937 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2938 | { | |
2939 | int zoneid; | |
2940 | struct zone *zone; | |
97a225e6 | 2941 | enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; |
e19a3f59 | 2942 | enum compact_result ret; |
698b1b30 | 2943 | |
97a225e6 | 2944 | for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { |
698b1b30 VB |
2945 | zone = &pgdat->node_zones[zoneid]; |
2946 | ||
2947 | if (!populated_zone(zone)) | |
2948 | continue; | |
2949 | ||
e19a3f59 KS |
2950 | ret = compaction_suit_allocation_order(zone, |
2951 | pgdat->kcompactd_max_order, | |
2952 | highest_zoneidx, ALLOC_WMARK_MIN); | |
2953 | if (ret == COMPACT_CONTINUE) | |
698b1b30 VB |
2954 | return true; |
2955 | } | |
2956 | ||
2957 | return false; | |
2958 | } | |
2959 | ||
2960 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2961 | { | |
2962 | /* | |
2963 | * With no special task, compact all zones so that a page of requested | |
2964 | * order is allocatable. | |
2965 | */ | |
2966 | int zoneid; | |
2967 | struct zone *zone; | |
2968 | struct compact_control cc = { | |
2969 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 2970 | .search_order = pgdat->kcompactd_max_order, |
97a225e6 | 2971 | .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, |
698b1b30 | 2972 | .mode = MIGRATE_SYNC_LIGHT, |
a0647dc9 | 2973 | .ignore_skip_hint = false, |
73e64c51 | 2974 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2975 | }; |
e19a3f59 KS |
2976 | enum compact_result ret; |
2977 | ||
698b1b30 | 2978 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
97a225e6 | 2979 | cc.highest_zoneidx); |
7f354a54 | 2980 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2981 | |
97a225e6 | 2982 | for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { |
698b1b30 VB |
2983 | int status; |
2984 | ||
2985 | zone = &pgdat->node_zones[zoneid]; | |
2986 | if (!populated_zone(zone)) | |
2987 | continue; | |
2988 | ||
2989 | if (compaction_deferred(zone, cc.order)) | |
2990 | continue; | |
2991 | ||
e19a3f59 KS |
2992 | ret = compaction_suit_allocation_order(zone, |
2993 | cc.order, zoneid, ALLOC_WMARK_MIN); | |
2994 | if (ret != COMPACT_CONTINUE) | |
e8606320 | 2995 | continue; |
f98a497e | 2996 | |
172400c6 VB |
2997 | if (kthread_should_stop()) |
2998 | return; | |
a94b5252 YS |
2999 | |
3000 | cc.zone = zone; | |
5e1f0f09 | 3001 | status = compact_zone(&cc, NULL); |
698b1b30 | 3002 | |
7ceb009a | 3003 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 3004 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 3005 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
3006 | /* |
3007 | * Buddy pages may become stranded on pcps that could | |
3008 | * otherwise coalesce on the zone's free area for | |
3009 | * order >= cc.order. This is ratelimited by the | |
3010 | * upcoming deferral. | |
3011 | */ | |
3012 | drain_all_pages(zone); | |
3013 | ||
698b1b30 VB |
3014 | /* |
3015 | * We use sync migration mode here, so we defer like | |
3016 | * sync direct compaction does. | |
3017 | */ | |
3018 | defer_compaction(zone, cc.order); | |
3019 | } | |
3020 | ||
7f354a54 DR |
3021 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
3022 | cc.total_migrate_scanned); | |
3023 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
3024 | cc.total_free_scanned); | |
698b1b30 VB |
3025 | } |
3026 | ||
3027 | /* | |
3028 | * Regardless of success, we are done until woken up next. But remember | |
97a225e6 JK |
3029 | * the requested order/highest_zoneidx in case it was higher/tighter |
3030 | * than our current ones | |
698b1b30 VB |
3031 | */ |
3032 | if (pgdat->kcompactd_max_order <= cc.order) | |
3033 | pgdat->kcompactd_max_order = 0; | |
97a225e6 JK |
3034 | if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) |
3035 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; | |
698b1b30 VB |
3036 | } |
3037 | ||
97a225e6 | 3038 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) |
698b1b30 VB |
3039 | { |
3040 | if (!order) | |
3041 | return; | |
3042 | ||
3043 | if (pgdat->kcompactd_max_order < order) | |
3044 | pgdat->kcompactd_max_order = order; | |
3045 | ||
97a225e6 JK |
3046 | if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) |
3047 | pgdat->kcompactd_highest_zoneidx = highest_zoneidx; | |
698b1b30 | 3048 | |
6818600f DB |
3049 | /* |
3050 | * Pairs with implicit barrier in wait_event_freezable() | |
3051 | * such that wakeups are not missed. | |
3052 | */ | |
3053 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
3054 | return; |
3055 | ||
3056 | if (!kcompactd_node_suitable(pgdat)) | |
3057 | return; | |
3058 | ||
3059 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
97a225e6 | 3060 | highest_zoneidx); |
698b1b30 VB |
3061 | wake_up_interruptible(&pgdat->kcompactd_wait); |
3062 | } | |
3063 | ||
3064 | /* | |
3065 | * The background compaction daemon, started as a kernel thread | |
3066 | * from the init process. | |
3067 | */ | |
3068 | static int kcompactd(void *p) | |
3069 | { | |
68d68ff6 | 3070 | pg_data_t *pgdat = (pg_data_t *)p; |
698b1b30 | 3071 | struct task_struct *tsk = current; |
e1e92bfa CTR |
3072 | long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); |
3073 | long timeout = default_timeout; | |
698b1b30 VB |
3074 | |
3075 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
3076 | ||
3077 | if (!cpumask_empty(cpumask)) | |
3078 | set_cpus_allowed_ptr(tsk, cpumask); | |
3079 | ||
3080 | set_freezable(); | |
3081 | ||
3082 | pgdat->kcompactd_max_order = 0; | |
97a225e6 | 3083 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; |
698b1b30 VB |
3084 | |
3085 | while (!kthread_should_stop()) { | |
eb414681 JW |
3086 | unsigned long pflags; |
3087 | ||
65d759c8 CTR |
3088 | /* |
3089 | * Avoid the unnecessary wakeup for proactive compaction | |
3090 | * when it is disabled. | |
3091 | */ | |
3092 | if (!sysctl_compaction_proactiveness) | |
3093 | timeout = MAX_SCHEDULE_TIMEOUT; | |
698b1b30 | 3094 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
facdaa91 | 3095 | if (wait_event_freezable_timeout(pgdat->kcompactd_wait, |
65d759c8 CTR |
3096 | kcompactd_work_requested(pgdat), timeout) && |
3097 | !pgdat->proactive_compact_trigger) { | |
facdaa91 NG |
3098 | |
3099 | psi_memstall_enter(&pflags); | |
3100 | kcompactd_do_work(pgdat); | |
3101 | psi_memstall_leave(&pflags); | |
e1e92bfa CTR |
3102 | /* |
3103 | * Reset the timeout value. The defer timeout from | |
3104 | * proactive compaction is lost here but that is fine | |
3105 | * as the condition of the zone changing substantionally | |
3106 | * then carrying on with the previous defer interval is | |
3107 | * not useful. | |
3108 | */ | |
3109 | timeout = default_timeout; | |
facdaa91 NG |
3110 | continue; |
3111 | } | |
698b1b30 | 3112 | |
e1e92bfa CTR |
3113 | /* |
3114 | * Start the proactive work with default timeout. Based | |
3115 | * on the fragmentation score, this timeout is updated. | |
3116 | */ | |
3117 | timeout = default_timeout; | |
facdaa91 NG |
3118 | if (should_proactive_compact_node(pgdat)) { |
3119 | unsigned int prev_score, score; | |
3120 | ||
facdaa91 | 3121 | prev_score = fragmentation_score_node(pgdat); |
3e40b3f4 | 3122 | compact_node(pgdat, true); |
facdaa91 NG |
3123 | score = fragmentation_score_node(pgdat); |
3124 | /* | |
3125 | * Defer proactive compaction if the fragmentation | |
3126 | * score did not go down i.e. no progress made. | |
3127 | */ | |
e1e92bfa CTR |
3128 | if (unlikely(score >= prev_score)) |
3129 | timeout = | |
3130 | default_timeout << COMPACT_MAX_DEFER_SHIFT; | |
facdaa91 | 3131 | } |
65d759c8 CTR |
3132 | if (unlikely(pgdat->proactive_compact_trigger)) |
3133 | pgdat->proactive_compact_trigger = false; | |
698b1b30 VB |
3134 | } |
3135 | ||
3136 | return 0; | |
3137 | } | |
3138 | ||
3139 | /* | |
3140 | * This kcompactd start function will be called by init and node-hot-add. | |
3141 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
3142 | */ | |
833dfc00 | 3143 | void __meminit kcompactd_run(int nid) |
698b1b30 VB |
3144 | { |
3145 | pg_data_t *pgdat = NODE_DATA(nid); | |
698b1b30 VB |
3146 | |
3147 | if (pgdat->kcompactd) | |
024c61ea | 3148 | return; |
698b1b30 VB |
3149 | |
3150 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
3151 | if (IS_ERR(pgdat->kcompactd)) { | |
3152 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
698b1b30 VB |
3153 | pgdat->kcompactd = NULL; |
3154 | } | |
698b1b30 VB |
3155 | } |
3156 | ||
3157 | /* | |
3158 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
e8da368a | 3159 | * be holding mem_hotplug_begin/done(). |
698b1b30 | 3160 | */ |
833dfc00 | 3161 | void __meminit kcompactd_stop(int nid) |
698b1b30 VB |
3162 | { |
3163 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
3164 | ||
3165 | if (kcompactd) { | |
3166 | kthread_stop(kcompactd); | |
3167 | NODE_DATA(nid)->kcompactd = NULL; | |
3168 | } | |
3169 | } | |
3170 | ||
3171 | /* | |
3172 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
3173 | * not required for correctness. So if the last cpu in a node goes | |
3174 | * away, we get changed to run anywhere: as the first one comes back, | |
3175 | * restore their cpu bindings. | |
3176 | */ | |
e46b1db2 | 3177 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
3178 | { |
3179 | int nid; | |
3180 | ||
e46b1db2 AMG |
3181 | for_each_node_state(nid, N_MEMORY) { |
3182 | pg_data_t *pgdat = NODE_DATA(nid); | |
3183 | const struct cpumask *mask; | |
698b1b30 | 3184 | |
e46b1db2 | 3185 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 3186 | |
e46b1db2 AMG |
3187 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3188 | /* One of our CPUs online: restore mask */ | |
3109de30 ML |
3189 | if (pgdat->kcompactd) |
3190 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 3191 | } |
e46b1db2 | 3192 | return 0; |
698b1b30 VB |
3193 | } |
3194 | ||
48fe8ab8 MC |
3195 | static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table, |
3196 | int write, void *buffer, size_t *lenp, loff_t *ppos) | |
3197 | { | |
3198 | int ret, old; | |
3199 | ||
3200 | if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write) | |
3201 | return proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
3202 | ||
3203 | old = *(int *)table->data; | |
3204 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
3205 | if (ret) | |
3206 | return ret; | |
3207 | if (old != *(int *)table->data) | |
3208 | pr_warn_once("sysctl attribute %s changed by %s[%d]\n", | |
3209 | table->procname, current->comm, | |
3210 | task_pid_nr(current)); | |
3211 | return ret; | |
3212 | } | |
3213 | ||
48fe8ab8 MC |
3214 | static struct ctl_table vm_compaction[] = { |
3215 | { | |
3216 | .procname = "compact_memory", | |
8b9167cd | 3217 | .data = &sysctl_compact_memory, |
48fe8ab8 MC |
3218 | .maxlen = sizeof(int), |
3219 | .mode = 0200, | |
3220 | .proc_handler = sysctl_compaction_handler, | |
3221 | }, | |
3222 | { | |
3223 | .procname = "compaction_proactiveness", | |
3224 | .data = &sysctl_compaction_proactiveness, | |
3225 | .maxlen = sizeof(sysctl_compaction_proactiveness), | |
3226 | .mode = 0644, | |
3227 | .proc_handler = compaction_proactiveness_sysctl_handler, | |
3228 | .extra1 = SYSCTL_ZERO, | |
3229 | .extra2 = SYSCTL_ONE_HUNDRED, | |
3230 | }, | |
3231 | { | |
3232 | .procname = "extfrag_threshold", | |
3233 | .data = &sysctl_extfrag_threshold, | |
3234 | .maxlen = sizeof(int), | |
3235 | .mode = 0644, | |
3236 | .proc_handler = proc_dointvec_minmax, | |
3237 | .extra1 = SYSCTL_ZERO, | |
3238 | .extra2 = SYSCTL_ONE_THOUSAND, | |
3239 | }, | |
3240 | { | |
3241 | .procname = "compact_unevictable_allowed", | |
3242 | .data = &sysctl_compact_unevictable_allowed, | |
3243 | .maxlen = sizeof(int), | |
3244 | .mode = 0644, | |
3245 | .proc_handler = proc_dointvec_minmax_warn_RT_change, | |
3246 | .extra1 = SYSCTL_ZERO, | |
3247 | .extra2 = SYSCTL_ONE, | |
3248 | }, | |
3249 | { } | |
3250 | }; | |
48fe8ab8 | 3251 | |
698b1b30 VB |
3252 | static int __init kcompactd_init(void) |
3253 | { | |
3254 | int nid; | |
e46b1db2 AMG |
3255 | int ret; |
3256 | ||
3257 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
3258 | "mm/compaction:online", | |
3259 | kcompactd_cpu_online, NULL); | |
3260 | if (ret < 0) { | |
3261 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
3262 | return ret; | |
3263 | } | |
698b1b30 VB |
3264 | |
3265 | for_each_node_state(nid, N_MEMORY) | |
3266 | kcompactd_run(nid); | |
48fe8ab8 | 3267 | register_sysctl_init("vm", vm_compaction); |
698b1b30 VB |
3268 | return 0; |
3269 | } | |
3270 | subsys_initcall(kcompactd_init) | |
3271 | ||
ff9543fd | 3272 | #endif /* CONFIG_COMPACTION */ |