]>
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 | } | |
ee6f62fd ZY |
43 | |
44 | /* | |
45 | * order == -1 is expected when compacting proactively via | |
46 | * 1. /proc/sys/vm/compact_memory | |
47 | * 2. /sys/devices/system/node/nodex/compact | |
48 | * 3. /proc/sys/vm/compaction_proactiveness | |
49 | */ | |
50 | static inline bool is_via_compact_memory(int order) | |
51 | { | |
52 | return order == -1; | |
53 | } | |
54 | ||
010fc29a MK |
55 | #else |
56 | #define count_compact_event(item) do { } while (0) | |
57 | #define count_compact_events(item, delta) do { } while (0) | |
ee6f62fd | 58 | static inline bool is_via_compact_memory(int order) { return false; } |
010fc29a MK |
59 | #endif |
60 | ||
ff9543fd MN |
61 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
62 | ||
b7aba698 MG |
63 | #define CREATE_TRACE_POINTS |
64 | #include <trace/events/compaction.h> | |
65 | ||
06b6640a VB |
66 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
67 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
06b6640a | 68 | |
facdaa91 NG |
69 | /* |
70 | * Page order with-respect-to which proactive compaction | |
71 | * calculates external fragmentation, which is used as | |
72 | * the "fragmentation score" of a node/zone. | |
73 | */ | |
74 | #if defined CONFIG_TRANSPARENT_HUGEPAGE | |
75 | #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER | |
25788738 | 76 | #elif defined CONFIG_HUGETLBFS |
facdaa91 NG |
77 | #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER |
78 | #else | |
79 | #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
80 | #endif | |
81 | ||
733aea0b | 82 | static void split_map_pages(struct list_head *freepages) |
748446bb | 83 | { |
733aea0b | 84 | unsigned int i, order; |
748446bb | 85 | struct page *page, *next; |
733aea0b | 86 | LIST_HEAD(tmp_list); |
748446bb | 87 | |
733aea0b ZY |
88 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
89 | list_for_each_entry_safe(page, next, &freepages[order], lru) { | |
90 | unsigned int nr_pages; | |
748446bb | 91 | |
733aea0b ZY |
92 | list_del(&page->lru); |
93 | ||
94 | nr_pages = 1 << order; | |
95 | ||
96 | post_alloc_hook(page, order, __GFP_MOVABLE); | |
97 | if (order) | |
98 | split_page(page, order); | |
99 | ||
100 | for (i = 0; i < nr_pages; i++) { | |
101 | list_add(&page->lru, &tmp_list); | |
102 | page++; | |
103 | } | |
104 | } | |
105 | list_splice_init(&tmp_list, &freepages[0]); | |
106 | } | |
748446bb MG |
107 | } |
108 | ||
733aea0b | 109 | static unsigned long release_free_list(struct list_head *freepages) |
ff9543fd | 110 | { |
733aea0b ZY |
111 | int order; |
112 | unsigned long high_pfn = 0; | |
66c64223 | 113 | |
733aea0b ZY |
114 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
115 | struct page *page, *next; | |
66c64223 | 116 | |
733aea0b ZY |
117 | list_for_each_entry_safe(page, next, &freepages[order], lru) { |
118 | unsigned long pfn = page_to_pfn(page); | |
ff9543fd | 119 | |
733aea0b ZY |
120 | list_del(&page->lru); |
121 | /* | |
122 | * Convert free pages into post allocation pages, so | |
123 | * that we can free them via __free_page. | |
124 | */ | |
125 | post_alloc_hook(page, order, __GFP_MOVABLE); | |
126 | __free_pages(page, order); | |
127 | if (pfn > high_pfn) | |
128 | high_pfn = pfn; | |
66c64223 | 129 | } |
ff9543fd | 130 | } |
733aea0b | 131 | return high_pfn; |
ff9543fd MN |
132 | } |
133 | ||
bb13ffeb | 134 | #ifdef CONFIG_COMPACTION |
68f2736a | 135 | bool PageMovable(struct page *page) |
bda807d4 | 136 | { |
68f2736a | 137 | const struct movable_operations *mops; |
bda807d4 MK |
138 | |
139 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
140 | if (!__PageMovable(page)) | |
68f2736a | 141 | return false; |
bda807d4 | 142 | |
68f2736a MWO |
143 | mops = page_movable_ops(page); |
144 | if (mops) | |
145 | return true; | |
bda807d4 | 146 | |
68f2736a | 147 | return false; |
bda807d4 | 148 | } |
bda807d4 | 149 | |
68f2736a | 150 | void __SetPageMovable(struct page *page, const struct movable_operations *mops) |
bda807d4 MK |
151 | { |
152 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
68f2736a MWO |
153 | VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page); |
154 | page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE); | |
bda807d4 MK |
155 | } |
156 | EXPORT_SYMBOL(__SetPageMovable); | |
157 | ||
158 | void __ClearPageMovable(struct page *page) | |
159 | { | |
bda807d4 MK |
160 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
161 | /* | |
68f2736a MWO |
162 | * This page still has the type of a movable page, but it's |
163 | * actually not movable any more. | |
bda807d4 | 164 | */ |
68f2736a | 165 | page->mapping = (void *)PAGE_MAPPING_MOVABLE; |
bda807d4 MK |
166 | } |
167 | EXPORT_SYMBOL(__ClearPageMovable); | |
168 | ||
24e2716f JK |
169 | /* Do not skip compaction more than 64 times */ |
170 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
171 | ||
172 | /* | |
173 | * Compaction is deferred when compaction fails to result in a page | |
860b3272 | 174 | * allocation success. 1 << compact_defer_shift, compactions are skipped up |
24e2716f JK |
175 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT |
176 | */ | |
2271b016 | 177 | static void defer_compaction(struct zone *zone, int order) |
24e2716f JK |
178 | { |
179 | zone->compact_considered = 0; | |
180 | zone->compact_defer_shift++; | |
181 | ||
182 | if (order < zone->compact_order_failed) | |
183 | zone->compact_order_failed = order; | |
184 | ||
185 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
186 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
187 | ||
188 | trace_mm_compaction_defer_compaction(zone, order); | |
189 | } | |
190 | ||
191 | /* Returns true if compaction should be skipped this time */ | |
2271b016 | 192 | static bool compaction_deferred(struct zone *zone, int order) |
24e2716f JK |
193 | { |
194 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
195 | ||
196 | if (order < zone->compact_order_failed) | |
197 | return false; | |
198 | ||
199 | /* Avoid possible overflow */ | |
62b35fe0 | 200 | if (++zone->compact_considered >= defer_limit) { |
24e2716f | 201 | zone->compact_considered = defer_limit; |
24e2716f | 202 | return false; |
62b35fe0 | 203 | } |
24e2716f JK |
204 | |
205 | trace_mm_compaction_deferred(zone, order); | |
206 | ||
207 | return true; | |
208 | } | |
209 | ||
210 | /* | |
211 | * Update defer tracking counters after successful compaction of given order, | |
212 | * which means an allocation either succeeded (alloc_success == true) or is | |
213 | * expected to succeed. | |
214 | */ | |
215 | void compaction_defer_reset(struct zone *zone, int order, | |
216 | bool alloc_success) | |
217 | { | |
218 | if (alloc_success) { | |
219 | zone->compact_considered = 0; | |
220 | zone->compact_defer_shift = 0; | |
221 | } | |
222 | if (order >= zone->compact_order_failed) | |
223 | zone->compact_order_failed = order + 1; | |
224 | ||
225 | trace_mm_compaction_defer_reset(zone, order); | |
226 | } | |
227 | ||
228 | /* Returns true if restarting compaction after many failures */ | |
2271b016 | 229 | static bool compaction_restarting(struct zone *zone, int order) |
24e2716f JK |
230 | { |
231 | if (order < zone->compact_order_failed) | |
232 | return false; | |
233 | ||
234 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
235 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
236 | } | |
237 | ||
bb13ffeb MG |
238 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
239 | static inline bool isolation_suitable(struct compact_control *cc, | |
240 | struct page *page) | |
241 | { | |
242 | if (cc->ignore_skip_hint) | |
243 | return true; | |
244 | ||
245 | return !get_pageblock_skip(page); | |
246 | } | |
247 | ||
02333641 VB |
248 | static void reset_cached_positions(struct zone *zone) |
249 | { | |
250 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
251 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 252 | zone->compact_cached_free_pfn = |
06b6640a | 253 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
254 | } |
255 | ||
9721fd82 BW |
256 | #ifdef CONFIG_SPARSEMEM |
257 | /* | |
258 | * If the PFN falls into an offline section, return the start PFN of the | |
259 | * next online section. If the PFN falls into an online section or if | |
260 | * there is no next online section, return 0. | |
261 | */ | |
262 | static unsigned long skip_offline_sections(unsigned long start_pfn) | |
263 | { | |
264 | unsigned long start_nr = pfn_to_section_nr(start_pfn); | |
265 | ||
266 | if (online_section_nr(start_nr)) | |
267 | return 0; | |
268 | ||
269 | while (++start_nr <= __highest_present_section_nr) { | |
270 | if (online_section_nr(start_nr)) | |
271 | return section_nr_to_pfn(start_nr); | |
272 | } | |
273 | ||
274 | return 0; | |
275 | } | |
e6e0c767 BW |
276 | |
277 | /* | |
278 | * If the PFN falls into an offline section, return the end PFN of the | |
279 | * next online section in reverse. If the PFN falls into an online section | |
280 | * or if there is no next online section in reverse, return 0. | |
281 | */ | |
282 | static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) | |
283 | { | |
284 | unsigned long start_nr = pfn_to_section_nr(start_pfn); | |
285 | ||
286 | if (!start_nr || online_section_nr(start_nr)) | |
287 | return 0; | |
288 | ||
289 | while (start_nr-- > 0) { | |
290 | if (online_section_nr(start_nr)) | |
291 | return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION; | |
292 | } | |
293 | ||
294 | return 0; | |
295 | } | |
9721fd82 BW |
296 | #else |
297 | static unsigned long skip_offline_sections(unsigned long start_pfn) | |
298 | { | |
299 | return 0; | |
300 | } | |
e6e0c767 BW |
301 | |
302 | static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) | |
303 | { | |
304 | return 0; | |
305 | } | |
9721fd82 BW |
306 | #endif |
307 | ||
21dc7e02 | 308 | /* |
2271b016 | 309 | * Compound pages of >= pageblock_order should consistently be skipped until |
b527cfe5 VB |
310 | * released. It is always pointless to compact pages of such order (if they are |
311 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 312 | */ |
b527cfe5 | 313 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 314 | { |
b527cfe5 | 315 | if (!PageCompound(page)) |
21dc7e02 | 316 | return false; |
b527cfe5 VB |
317 | |
318 | page = compound_head(page); | |
319 | ||
320 | if (compound_order(page) >= pageblock_order) | |
321 | return true; | |
322 | ||
323 | return false; | |
21dc7e02 DR |
324 | } |
325 | ||
e332f741 MG |
326 | static bool |
327 | __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, | |
328 | bool check_target) | |
329 | { | |
330 | struct page *page = pfn_to_online_page(pfn); | |
6b0868c8 | 331 | struct page *block_page; |
e332f741 MG |
332 | struct page *end_page; |
333 | unsigned long block_pfn; | |
334 | ||
335 | if (!page) | |
336 | return false; | |
337 | if (zone != page_zone(page)) | |
338 | return false; | |
339 | if (pageblock_skip_persistent(page)) | |
340 | return false; | |
341 | ||
342 | /* | |
343 | * If skip is already cleared do no further checking once the | |
344 | * restart points have been set. | |
345 | */ | |
346 | if (check_source && check_target && !get_pageblock_skip(page)) | |
347 | return true; | |
348 | ||
349 | /* | |
350 | * If clearing skip for the target scanner, do not select a | |
351 | * non-movable pageblock as the starting point. | |
352 | */ | |
353 | if (!check_source && check_target && | |
354 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
355 | return false; | |
356 | ||
6b0868c8 MG |
357 | /* Ensure the start of the pageblock or zone is online and valid */ |
358 | block_pfn = pageblock_start_pfn(pfn); | |
a2e9a5af VB |
359 | block_pfn = max(block_pfn, zone->zone_start_pfn); |
360 | block_page = pfn_to_online_page(block_pfn); | |
6b0868c8 MG |
361 | if (block_page) { |
362 | page = block_page; | |
363 | pfn = block_pfn; | |
364 | } | |
365 | ||
366 | /* Ensure the end of the pageblock or zone is online and valid */ | |
a2e9a5af | 367 | block_pfn = pageblock_end_pfn(pfn) - 1; |
6b0868c8 MG |
368 | block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); |
369 | end_page = pfn_to_online_page(block_pfn); | |
370 | if (!end_page) | |
371 | return false; | |
372 | ||
e332f741 MG |
373 | /* |
374 | * Only clear the hint if a sample indicates there is either a | |
375 | * free page or an LRU page in the block. One or other condition | |
376 | * is necessary for the block to be a migration source/target. | |
377 | */ | |
e332f741 | 378 | do { |
859a85dd MR |
379 | if (check_source && PageLRU(page)) { |
380 | clear_pageblock_skip(page); | |
381 | return true; | |
382 | } | |
e332f741 | 383 | |
859a85dd MR |
384 | if (check_target && PageBuddy(page)) { |
385 | clear_pageblock_skip(page); | |
386 | return true; | |
e332f741 MG |
387 | } |
388 | ||
389 | page += (1 << PAGE_ALLOC_COSTLY_ORDER); | |
a2e9a5af | 390 | } while (page <= end_page); |
e332f741 MG |
391 | |
392 | return false; | |
393 | } | |
394 | ||
bb13ffeb MG |
395 | /* |
396 | * This function is called to clear all cached information on pageblocks that | |
397 | * should be skipped for page isolation when the migrate and free page scanner | |
398 | * meet. | |
399 | */ | |
62997027 | 400 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb | 401 | { |
e332f741 | 402 | unsigned long migrate_pfn = zone->zone_start_pfn; |
6b0868c8 | 403 | unsigned long free_pfn = zone_end_pfn(zone) - 1; |
e332f741 MG |
404 | unsigned long reset_migrate = free_pfn; |
405 | unsigned long reset_free = migrate_pfn; | |
406 | bool source_set = false; | |
407 | bool free_set = false; | |
408 | ||
8df4e28c | 409 | /* Only flush if a full compaction finished recently */ |
e332f741 MG |
410 | if (!zone->compact_blockskip_flush) |
411 | return; | |
bb13ffeb | 412 | |
62997027 | 413 | zone->compact_blockskip_flush = false; |
bb13ffeb | 414 | |
e332f741 MG |
415 | /* |
416 | * Walk the zone and update pageblock skip information. Source looks | |
417 | * for PageLRU while target looks for PageBuddy. When the scanner | |
418 | * is found, both PageBuddy and PageLRU are checked as the pageblock | |
419 | * is suitable as both source and target. | |
420 | */ | |
421 | for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, | |
422 | free_pfn -= pageblock_nr_pages) { | |
bb13ffeb MG |
423 | cond_resched(); |
424 | ||
e332f741 MG |
425 | /* Update the migrate PFN */ |
426 | if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && | |
427 | migrate_pfn < reset_migrate) { | |
428 | source_set = true; | |
429 | reset_migrate = migrate_pfn; | |
430 | zone->compact_init_migrate_pfn = reset_migrate; | |
431 | zone->compact_cached_migrate_pfn[0] = reset_migrate; | |
432 | zone->compact_cached_migrate_pfn[1] = reset_migrate; | |
433 | } | |
bb13ffeb | 434 | |
e332f741 MG |
435 | /* Update the free PFN */ |
436 | if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && | |
437 | free_pfn > reset_free) { | |
438 | free_set = true; | |
439 | reset_free = free_pfn; | |
440 | zone->compact_init_free_pfn = reset_free; | |
441 | zone->compact_cached_free_pfn = reset_free; | |
442 | } | |
bb13ffeb | 443 | } |
02333641 | 444 | |
e332f741 MG |
445 | /* Leave no distance if no suitable block was reset */ |
446 | if (reset_migrate >= reset_free) { | |
447 | zone->compact_cached_migrate_pfn[0] = migrate_pfn; | |
448 | zone->compact_cached_migrate_pfn[1] = migrate_pfn; | |
449 | zone->compact_cached_free_pfn = free_pfn; | |
450 | } | |
bb13ffeb MG |
451 | } |
452 | ||
62997027 MG |
453 | void reset_isolation_suitable(pg_data_t *pgdat) |
454 | { | |
455 | int zoneid; | |
456 | ||
457 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
458 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
459 | if (!populated_zone(zone)) | |
460 | continue; | |
461 | ||
8df4e28c | 462 | __reset_isolation_suitable(zone); |
62997027 MG |
463 | } |
464 | } | |
465 | ||
e380bebe MG |
466 | /* |
467 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
468 | * locks are not required for read/writers. Returns true if it was already set. | |
469 | */ | |
590ccea8 | 470 | static bool test_and_set_skip(struct compact_control *cc, struct page *page) |
e380bebe MG |
471 | { |
472 | bool skip; | |
473 | ||
590ccea8 | 474 | /* Do not update if skip hint is being ignored */ |
e380bebe MG |
475 | if (cc->ignore_skip_hint) |
476 | return false; | |
477 | ||
e380bebe MG |
478 | skip = get_pageblock_skip(page); |
479 | if (!skip && !cc->no_set_skip_hint) | |
480 | set_pageblock_skip(page); | |
481 | ||
482 | return skip; | |
483 | } | |
484 | ||
485 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
486 | { | |
487 | struct zone *zone = cc->zone; | |
488 | ||
e380bebe MG |
489 | /* Set for isolation rather than compaction */ |
490 | if (cc->no_set_skip_hint) | |
491 | return; | |
492 | ||
3c099a2b KS |
493 | pfn = pageblock_end_pfn(pfn); |
494 | ||
cf043a00 | 495 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
496 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
497 | zone->compact_cached_migrate_pfn[0] = pfn; | |
498 | if (cc->mode != MIGRATE_ASYNC && | |
499 | pfn > zone->compact_cached_migrate_pfn[1]) | |
500 | zone->compact_cached_migrate_pfn[1] = pfn; | |
501 | } | |
502 | ||
bb13ffeb MG |
503 | /* |
504 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 505 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 506 | */ |
c89511ab | 507 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 508 | struct page *page, unsigned long pfn) |
bb13ffeb | 509 | { |
c89511ab | 510 | struct zone *zone = cc->zone; |
6815bf3f | 511 | |
2583d671 | 512 | if (cc->no_set_skip_hint) |
6815bf3f JK |
513 | return; |
514 | ||
edc2ca61 | 515 | set_pageblock_skip(page); |
c89511ab | 516 | |
e380bebe MG |
517 | if (pfn < zone->compact_cached_free_pfn) |
518 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
519 | } |
520 | #else | |
521 | static inline bool isolation_suitable(struct compact_control *cc, | |
522 | struct page *page) | |
523 | { | |
524 | return true; | |
525 | } | |
526 | ||
b527cfe5 | 527 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
528 | { |
529 | return false; | |
530 | } | |
531 | ||
532 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 533 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
534 | { |
535 | } | |
e380bebe MG |
536 | |
537 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
538 | { | |
539 | } | |
540 | ||
590ccea8 | 541 | static bool test_and_set_skip(struct compact_control *cc, struct page *page) |
e380bebe MG |
542 | { |
543 | return false; | |
544 | } | |
bb13ffeb MG |
545 | #endif /* CONFIG_COMPACTION */ |
546 | ||
8b44d279 VB |
547 | /* |
548 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
549 | * very heavily contended. For async compaction, trylock and record if the |
550 | * lock is contended. The lock will still be acquired but compaction will | |
551 | * abort when the current block is finished regardless of success rate. | |
552 | * Sync compaction acquires the lock. | |
8b44d279 | 553 | * |
cb2dcaf0 | 554 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 555 | */ |
cb2dcaf0 | 556 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 557 | struct compact_control *cc) |
77337ede | 558 | __acquires(lock) |
2a1402aa | 559 | { |
cb2dcaf0 MG |
560 | /* Track if the lock is contended in async mode */ |
561 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
562 | if (spin_trylock_irqsave(lock, *flags)) | |
563 | return true; | |
564 | ||
565 | cc->contended = true; | |
8b44d279 | 566 | } |
1f9efdef | 567 | |
cb2dcaf0 | 568 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 569 | return true; |
2a1402aa MG |
570 | } |
571 | ||
c67fe375 MG |
572 | /* |
573 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
574 | * very heavily contended. The lock should be periodically unlocked to avoid |
575 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
576 | * the lock. It might also be that allowing the IRQs will result in | |
d56c1584 | 577 | * need_resched() becoming true. If scheduling is needed, compaction schedules. |
8b44d279 VB |
578 | * Either compaction type will also abort if a fatal signal is pending. |
579 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 580 | * |
d56c1584 ML |
581 | * Returns true if compaction should abort due to fatal signal pending. |
582 | * Returns false when compaction can continue. | |
c67fe375 | 583 | */ |
8b44d279 VB |
584 | static bool compact_unlock_should_abort(spinlock_t *lock, |
585 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 586 | { |
8b44d279 VB |
587 | if (*locked) { |
588 | spin_unlock_irqrestore(lock, flags); | |
589 | *locked = false; | |
590 | } | |
1f9efdef | 591 | |
8b44d279 | 592 | if (fatal_signal_pending(current)) { |
c3486f53 | 593 | cc->contended = true; |
8b44d279 VB |
594 | return true; |
595 | } | |
c67fe375 | 596 | |
cf66f070 | 597 | cond_resched(); |
be976572 VB |
598 | |
599 | return false; | |
600 | } | |
601 | ||
85aa125f | 602 | /* |
9e4be470 JM |
603 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
604 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
605 | * (even though it may still end up isolating some pages). | |
85aa125f | 606 | */ |
f40d1e42 | 607 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 608 | unsigned long *start_pfn, |
85aa125f MN |
609 | unsigned long end_pfn, |
610 | struct list_head *freelist, | |
4fca9730 | 611 | unsigned int stride, |
85aa125f | 612 | bool strict) |
748446bb | 613 | { |
b7aba698 | 614 | int nr_scanned = 0, total_isolated = 0; |
dc13292c | 615 | struct page *page; |
b8b2d825 | 616 | unsigned long flags = 0; |
f40d1e42 | 617 | bool locked = false; |
e14c720e | 618 | unsigned long blockpfn = *start_pfn; |
66c64223 | 619 | unsigned int order; |
748446bb | 620 | |
4fca9730 MG |
621 | /* Strict mode is for isolation, speed is secondary */ |
622 | if (strict) | |
623 | stride = 1; | |
624 | ||
dc13292c | 625 | page = pfn_to_page(blockpfn); |
748446bb | 626 | |
f40d1e42 | 627 | /* Isolate free pages. */ |
dc13292c | 628 | for (; blockpfn < end_pfn; blockpfn += stride, page += stride) { |
66c64223 | 629 | int isolated; |
748446bb | 630 | |
8b44d279 VB |
631 | /* |
632 | * Periodically drop the lock (if held) regardless of its | |
633 | * contention, to give chance to IRQs. Abort if fatal signal | |
d56c1584 | 634 | * pending. |
8b44d279 | 635 | */ |
c036ddff | 636 | if (!(blockpfn % COMPACT_CLUSTER_MAX) |
8b44d279 VB |
637 | && compact_unlock_should_abort(&cc->zone->lock, flags, |
638 | &locked, cc)) | |
639 | break; | |
640 | ||
b7aba698 | 641 | nr_scanned++; |
2af120bc | 642 | |
9fcd6d2e VB |
643 | /* |
644 | * For compound pages such as THP and hugetlbfs, we can save | |
645 | * potentially a lot of iterations if we skip them at once. | |
646 | * The check is racy, but we can consider only valid values | |
647 | * and the only danger is skipping too much. | |
648 | */ | |
649 | if (PageCompound(page)) { | |
21dc7e02 DR |
650 | const unsigned int order = compound_order(page); |
651 | ||
3da0272a | 652 | if (blockpfn + (1UL << order) <= end_pfn) { |
21dc7e02 | 653 | blockpfn += (1UL << order) - 1; |
dc13292c | 654 | page += (1UL << order) - 1; |
56d48d8d | 655 | nr_scanned += (1UL << order) - 1; |
9fcd6d2e | 656 | } |
3da0272a | 657 | |
9fcd6d2e VB |
658 | goto isolate_fail; |
659 | } | |
660 | ||
f40d1e42 | 661 | if (!PageBuddy(page)) |
2af120bc | 662 | goto isolate_fail; |
f40d1e42 | 663 | |
85f73e6d | 664 | /* If we already hold the lock, we can skip some rechecking. */ |
69b7189f | 665 | if (!locked) { |
cb2dcaf0 | 666 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 667 | &flags, cc); |
f40d1e42 | 668 | |
69b7189f VB |
669 | /* Recheck this is a buddy page under lock */ |
670 | if (!PageBuddy(page)) | |
671 | goto isolate_fail; | |
672 | } | |
748446bb | 673 | |
66c64223 | 674 | /* Found a free page, will break it into order-0 pages */ |
ab130f91 | 675 | order = buddy_order(page); |
66c64223 | 676 | isolated = __isolate_free_page(page, order); |
a4f04f2c DR |
677 | if (!isolated) |
678 | break; | |
66c64223 | 679 | set_page_private(page, order); |
a4f04f2c | 680 | |
b717d6b9 | 681 | nr_scanned += isolated - 1; |
748446bb | 682 | total_isolated += isolated; |
a4f04f2c | 683 | cc->nr_freepages += isolated; |
733aea0b | 684 | list_add_tail(&page->lru, &freelist[order]); |
66c64223 | 685 | |
a4f04f2c DR |
686 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
687 | blockpfn += isolated; | |
688 | break; | |
748446bb | 689 | } |
a4f04f2c DR |
690 | /* Advance to the end of split page */ |
691 | blockpfn += isolated - 1; | |
dc13292c | 692 | page += isolated - 1; |
a4f04f2c | 693 | continue; |
2af120bc LA |
694 | |
695 | isolate_fail: | |
696 | if (strict) | |
697 | break; | |
2af120bc | 698 | |
748446bb MG |
699 | } |
700 | ||
a4f04f2c DR |
701 | if (locked) |
702 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
703 | ||
9fcd6d2e | 704 | /* |
3da0272a | 705 | * Be careful to not go outside of the pageblock. |
9fcd6d2e VB |
706 | */ |
707 | if (unlikely(blockpfn > end_pfn)) | |
708 | blockpfn = end_pfn; | |
709 | ||
e34d85f0 JK |
710 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
711 | nr_scanned, total_isolated); | |
712 | ||
e14c720e VB |
713 | /* Record how far we have got within the block */ |
714 | *start_pfn = blockpfn; | |
715 | ||
f40d1e42 MG |
716 | /* |
717 | * If strict isolation is requested by CMA then check that all the | |
718 | * pages requested were isolated. If there were any failures, 0 is | |
719 | * returned and CMA will fail. | |
720 | */ | |
2af120bc | 721 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
722 | total_isolated = 0; |
723 | ||
7f354a54 | 724 | cc->total_free_scanned += nr_scanned; |
397487db | 725 | if (total_isolated) |
010fc29a | 726 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
727 | return total_isolated; |
728 | } | |
729 | ||
85aa125f MN |
730 | /** |
731 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 732 | * @cc: Compaction control structure. |
85aa125f MN |
733 | * @start_pfn: The first PFN to start isolating. |
734 | * @end_pfn: The one-past-last PFN. | |
735 | * | |
736 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
737 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
738 | * undo its actions and return zero. | |
739 | * | |
740 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
741 | * (which may be greater then end_pfn if end fell in a middle of | |
742 | * a free page). | |
743 | */ | |
ff9543fd | 744 | unsigned long |
bb13ffeb MG |
745 | isolate_freepages_range(struct compact_control *cc, |
746 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 747 | { |
e1409c32 | 748 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
733aea0b ZY |
749 | int order; |
750 | struct list_head tmp_freepages[NR_PAGE_ORDERS]; | |
751 | ||
752 | for (order = 0; order < NR_PAGE_ORDERS; order++) | |
753 | INIT_LIST_HEAD(&tmp_freepages[order]); | |
85aa125f | 754 | |
7d49d886 | 755 | pfn = start_pfn; |
06b6640a | 756 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
757 | if (block_start_pfn < cc->zone->zone_start_pfn) |
758 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 759 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
760 | |
761 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 762 | block_start_pfn = block_end_pfn, |
7d49d886 | 763 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
764 | /* Protect pfn from changing by isolate_freepages_block */ |
765 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 766 | |
58420016 JK |
767 | /* |
768 | * pfn could pass the block_end_pfn if isolated freepage | |
769 | * is more than pageblock order. In this case, we adjust | |
770 | * scanning range to right one. | |
771 | */ | |
772 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
773 | block_start_pfn = pageblock_start_pfn(pfn); |
774 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
775 | } |
776 | ||
a2864a67 KS |
777 | block_end_pfn = min(block_end_pfn, end_pfn); |
778 | ||
e1409c32 JK |
779 | if (!pageblock_pfn_to_page(block_start_pfn, |
780 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
781 | break; |
782 | ||
e14c720e | 783 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
733aea0b | 784 | block_end_pfn, tmp_freepages, 0, true); |
85aa125f MN |
785 | |
786 | /* | |
787 | * In strict mode, isolate_freepages_block() returns 0 if | |
788 | * there are any holes in the block (ie. invalid PFNs or | |
789 | * non-free pages). | |
790 | */ | |
791 | if (!isolated) | |
792 | break; | |
793 | ||
794 | /* | |
795 | * If we managed to isolate pages, it is always (1 << n) * | |
796 | * pageblock_nr_pages for some non-negative n. (Max order | |
797 | * page may span two pageblocks). | |
798 | */ | |
799 | } | |
800 | ||
85aa125f MN |
801 | if (pfn < end_pfn) { |
802 | /* Loop terminated early, cleanup. */ | |
733aea0b | 803 | release_free_list(tmp_freepages); |
85aa125f MN |
804 | return 0; |
805 | } | |
806 | ||
733aea0b ZY |
807 | /* __isolate_free_page() does not map the pages */ |
808 | split_map_pages(tmp_freepages); | |
809 | ||
85aa125f MN |
810 | /* We don't use freelists for anything. */ |
811 | return pfn; | |
812 | } | |
813 | ||
748446bb | 814 | /* Similar to reclaim, but different enough that they don't share logic */ |
4fbbb3fd | 815 | static bool too_many_isolated(struct compact_control *cc) |
748446bb | 816 | { |
4fbbb3fd | 817 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
d818fca1 MG |
818 | bool too_many; |
819 | ||
bc693045 | 820 | unsigned long active, inactive, isolated; |
748446bb | 821 | |
5f438eee AR |
822 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + |
823 | node_page_state(pgdat, NR_INACTIVE_ANON); | |
824 | active = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
825 | node_page_state(pgdat, NR_ACTIVE_ANON); | |
826 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + | |
827 | node_page_state(pgdat, NR_ISOLATED_ANON); | |
748446bb | 828 | |
4fbbb3fd JW |
829 | /* |
830 | * Allow GFP_NOFS to isolate past the limit set for regular | |
831 | * compaction runs. This prevents an ABBA deadlock when other | |
832 | * compactors have already isolated to the limit, but are | |
833 | * blocked on filesystem locks held by the GFP_NOFS thread. | |
834 | */ | |
835 | if (cc->gfp_mask & __GFP_FS) { | |
836 | inactive >>= 3; | |
837 | active >>= 3; | |
838 | } | |
839 | ||
d818fca1 MG |
840 | too_many = isolated > (inactive + active) / 2; |
841 | if (!too_many) | |
842 | wake_throttle_isolated(pgdat); | |
843 | ||
844 | return too_many; | |
748446bb MG |
845 | } |
846 | ||
ee6f62fd ZY |
847 | /** |
848 | * skip_isolation_on_order() - determine when to skip folio isolation based on | |
849 | * folio order and compaction target order | |
850 | * @order: to-be-isolated folio order | |
851 | * @target_order: compaction target order | |
852 | * | |
853 | * This avoids unnecessary folio isolations during compaction. | |
854 | */ | |
855 | static bool skip_isolation_on_order(int order, int target_order) | |
856 | { | |
857 | /* | |
858 | * Unless we are performing global compaction (i.e., | |
859 | * is_via_compact_memory), skip any folios that are larger than the | |
860 | * target order: we wouldn't be here if we'd have a free folio with | |
861 | * the desired target_order, so migrating this folio would likely fail | |
862 | * later. | |
863 | */ | |
864 | if (!is_via_compact_memory(target_order) && order >= target_order) | |
865 | return true; | |
866 | /* | |
867 | * We limit memory compaction to pageblocks and won't try | |
868 | * creating free blocks of memory that are larger than that. | |
869 | */ | |
870 | return order >= pageblock_order; | |
871 | } | |
872 | ||
2fe86e00 | 873 | /** |
edc2ca61 VB |
874 | * isolate_migratepages_block() - isolate all migrate-able pages within |
875 | * a single pageblock | |
2fe86e00 | 876 | * @cc: Compaction control structure. |
edc2ca61 VB |
877 | * @low_pfn: The first PFN to isolate |
878 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
89f6c88a | 879 | * @mode: Isolation mode to be used. |
2fe86e00 MN |
880 | * |
881 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 | 882 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
c2ad7a1f | 883 | * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, |
369fa227 | 884 | * -ENOMEM in case we could not allocate a page, or 0. |
c2ad7a1f | 885 | * cc->migrate_pfn will contain the next pfn to scan. |
2fe86e00 | 886 | * |
edc2ca61 | 887 | * The pages are isolated on cc->migratepages list (not required to be empty), |
c2ad7a1f | 888 | * and cc->nr_migratepages is updated accordingly. |
748446bb | 889 | */ |
c2ad7a1f | 890 | static int |
edc2ca61 | 891 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, |
89f6c88a | 892 | unsigned long end_pfn, isolate_mode_t mode) |
748446bb | 893 | { |
5f438eee | 894 | pg_data_t *pgdat = cc->zone->zone_pgdat; |
b7aba698 | 895 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 896 | struct lruvec *lruvec; |
b8b2d825 | 897 | unsigned long flags = 0; |
6168d0da | 898 | struct lruvec *locked = NULL; |
56ae0bb3 | 899 | struct folio *folio = NULL; |
bb13ffeb | 900 | struct page *page = NULL, *valid_page = NULL; |
89f6c88a | 901 | struct address_space *mapping; |
e34d85f0 | 902 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
903 | bool skip_on_failure = false; |
904 | unsigned long next_skip_pfn = 0; | |
e380bebe | 905 | bool skip_updated = false; |
c2ad7a1f OS |
906 | int ret = 0; |
907 | ||
908 | cc->migrate_pfn = low_pfn; | |
748446bb | 909 | |
748446bb MG |
910 | /* |
911 | * Ensure that there are not too many pages isolated from the LRU | |
912 | * list by either parallel reclaimers or compaction. If there are, | |
913 | * delay for some time until fewer pages are isolated | |
914 | */ | |
4fbbb3fd | 915 | while (unlikely(too_many_isolated(cc))) { |
d20bdd57 ZY |
916 | /* stop isolation if there are still pages not migrated */ |
917 | if (cc->nr_migratepages) | |
c2ad7a1f | 918 | return -EAGAIN; |
d20bdd57 | 919 | |
f9e35b3b | 920 | /* async migration should just abort */ |
e0b9daeb | 921 | if (cc->mode == MIGRATE_ASYNC) |
c2ad7a1f | 922 | return -EAGAIN; |
f9e35b3b | 923 | |
c3f4a9a2 | 924 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
748446bb MG |
925 | |
926 | if (fatal_signal_pending(current)) | |
c2ad7a1f | 927 | return -EINTR; |
748446bb MG |
928 | } |
929 | ||
cf66f070 | 930 | cond_resched(); |
aeef4b83 | 931 | |
fdd048e1 VB |
932 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
933 | skip_on_failure = true; | |
934 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
935 | } | |
936 | ||
748446bb | 937 | /* Time to isolate some pages for migration */ |
748446bb | 938 | for (; low_pfn < end_pfn; low_pfn++) { |
0003e2a4 | 939 | bool is_dirty, is_unevictable; |
29c0dde8 | 940 | |
fdd048e1 VB |
941 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
942 | /* | |
943 | * We have isolated all migration candidates in the | |
944 | * previous order-aligned block, and did not skip it due | |
945 | * to failure. We should migrate the pages now and | |
946 | * hopefully succeed compaction. | |
947 | */ | |
948 | if (nr_isolated) | |
949 | break; | |
950 | ||
951 | /* | |
952 | * We failed to isolate in the previous order-aligned | |
953 | * block. Set the new boundary to the end of the | |
954 | * current block. Note we can't simply increase | |
955 | * next_skip_pfn by 1 << order, as low_pfn might have | |
956 | * been incremented by a higher number due to skipping | |
957 | * a compound or a high-order buddy page in the | |
958 | * previous loop iteration. | |
959 | */ | |
960 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
961 | } | |
962 | ||
8b44d279 VB |
963 | /* |
964 | * Periodically drop the lock (if held) regardless of its | |
670105a2 MG |
965 | * contention, to give chance to IRQs. Abort completely if |
966 | * a fatal signal is pending. | |
8b44d279 | 967 | */ |
c036ddff | 968 | if (!(low_pfn % COMPACT_CLUSTER_MAX)) { |
6168d0da AS |
969 | if (locked) { |
970 | unlock_page_lruvec_irqrestore(locked, flags); | |
971 | locked = NULL; | |
972 | } | |
973 | ||
974 | if (fatal_signal_pending(current)) { | |
975 | cc->contended = true; | |
c2ad7a1f | 976 | ret = -EINTR; |
6168d0da | 977 | |
6168d0da AS |
978 | goto fatal_pending; |
979 | } | |
980 | ||
981 | cond_resched(); | |
670105a2 | 982 | } |
c67fe375 | 983 | |
b7aba698 | 984 | nr_scanned++; |
748446bb | 985 | |
748446bb | 986 | page = pfn_to_page(low_pfn); |
dc908600 | 987 | |
e380bebe MG |
988 | /* |
989 | * Check if the pageblock has already been marked skipped. | |
493614da | 990 | * Only the first PFN is checked as the caller isolates |
e380bebe MG |
991 | * COMPACT_CLUSTER_MAX at a time so the second call must |
992 | * not falsely conclude that the block should be skipped. | |
993 | */ | |
493614da JW |
994 | if (!valid_page && (pageblock_aligned(low_pfn) || |
995 | low_pfn == cc->zone->zone_start_pfn)) { | |
4af12d04 | 996 | if (!isolation_suitable(cc, page)) { |
e380bebe | 997 | low_pfn = end_pfn; |
56ae0bb3 | 998 | folio = NULL; |
e380bebe MG |
999 | goto isolate_abort; |
1000 | } | |
bb13ffeb | 1001 | valid_page = page; |
e380bebe | 1002 | } |
bb13ffeb | 1003 | |
ee6f62fd ZY |
1004 | if (PageHuge(page)) { |
1005 | /* | |
1006 | * skip hugetlbfs if we are not compacting for pages | |
1007 | * bigger than its order. THPs and other compound pages | |
1008 | * are handled below. | |
1009 | */ | |
1010 | if (!cc->alloc_contig) { | |
1011 | const unsigned int order = compound_order(page); | |
1012 | ||
1013 | if (order <= MAX_PAGE_ORDER) { | |
1014 | low_pfn += (1UL << order) - 1; | |
1015 | nr_scanned += (1UL << order) - 1; | |
1016 | } | |
1017 | goto isolate_fail; | |
1018 | } | |
1019 | /* for alloc_contig case */ | |
1c06b6a5 BW |
1020 | if (locked) { |
1021 | unlock_page_lruvec_irqrestore(locked, flags); | |
1022 | locked = NULL; | |
1023 | } | |
1024 | ||
ae37c7ff | 1025 | ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); |
369fa227 OS |
1026 | |
1027 | /* | |
1028 | * Fail isolation in case isolate_or_dissolve_huge_page() | |
1029 | * reports an error. In case of -ENOMEM, abort right away. | |
1030 | */ | |
1031 | if (ret < 0) { | |
1032 | /* Do not report -EBUSY down the chain */ | |
1033 | if (ret == -EBUSY) | |
1034 | ret = 0; | |
66fe1cf7 | 1035 | low_pfn += compound_nr(page) - 1; |
56d48d8d | 1036 | nr_scanned += compound_nr(page) - 1; |
369fa227 OS |
1037 | goto isolate_fail; |
1038 | } | |
1039 | ||
ae37c7ff OS |
1040 | if (PageHuge(page)) { |
1041 | /* | |
1042 | * Hugepage was successfully isolated and placed | |
1043 | * on the cc->migratepages list. | |
1044 | */ | |
56ae0bb3 KW |
1045 | folio = page_folio(page); |
1046 | low_pfn += folio_nr_pages(folio) - 1; | |
ae37c7ff OS |
1047 | goto isolate_success_no_list; |
1048 | } | |
1049 | ||
369fa227 OS |
1050 | /* |
1051 | * Ok, the hugepage was dissolved. Now these pages are | |
1052 | * Buddy and cannot be re-allocated because they are | |
1053 | * isolated. Fall-through as the check below handles | |
1054 | * Buddy pages. | |
1055 | */ | |
1056 | } | |
1057 | ||
6c14466c | 1058 | /* |
99c0fd5e VB |
1059 | * Skip if free. We read page order here without zone lock |
1060 | * which is generally unsafe, but the race window is small and | |
1061 | * the worst thing that can happen is that we skip some | |
1062 | * potential isolation targets. | |
6c14466c | 1063 | */ |
99c0fd5e | 1064 | if (PageBuddy(page)) { |
ab130f91 | 1065 | unsigned long freepage_order = buddy_order_unsafe(page); |
99c0fd5e VB |
1066 | |
1067 | /* | |
1068 | * Without lock, we cannot be sure that what we got is | |
1069 | * a valid page order. Consider only values in the | |
1070 | * valid order range to prevent low_pfn overflow. | |
1071 | */ | |
5e0a760b | 1072 | if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) { |
99c0fd5e | 1073 | low_pfn += (1UL << freepage_order) - 1; |
56d48d8d BW |
1074 | nr_scanned += (1UL << freepage_order) - 1; |
1075 | } | |
748446bb | 1076 | continue; |
99c0fd5e | 1077 | } |
748446bb | 1078 | |
bc835011 | 1079 | /* |
ee6f62fd ZY |
1080 | * Regardless of being on LRU, compound pages such as THP |
1081 | * (hugetlbfs is handled above) are not to be compacted unless | |
1082 | * we are attempting an allocation larger than the compound | |
1083 | * page size. We can potentially save a lot of iterations if we | |
1084 | * skip them at once. The check is racy, but we can consider | |
1085 | * only valid values and the only danger is skipping too much. | |
bc835011 | 1086 | */ |
1da2f328 | 1087 | if (PageCompound(page) && !cc->alloc_contig) { |
21dc7e02 | 1088 | const unsigned int order = compound_order(page); |
edc2ca61 | 1089 | |
ee6f62fd ZY |
1090 | /* Skip based on page order and compaction target order. */ |
1091 | if (skip_isolation_on_order(order, cc->order)) { | |
1092 | if (order <= MAX_PAGE_ORDER) { | |
1093 | low_pfn += (1UL << order) - 1; | |
1094 | nr_scanned += (1UL << order) - 1; | |
1095 | } | |
1096 | goto isolate_fail; | |
56d48d8d | 1097 | } |
2a1402aa MG |
1098 | } |
1099 | ||
bda807d4 MK |
1100 | /* |
1101 | * Check may be lockless but that's ok as we recheck later. | |
1102 | * It's possible to migrate LRU and non-lru movable pages. | |
1103 | * Skip any other type of page | |
1104 | */ | |
1105 | if (!PageLRU(page)) { | |
bda807d4 MK |
1106 | /* |
1107 | * __PageMovable can return false positive so we need | |
1108 | * to verify it under page_lock. | |
1109 | */ | |
1110 | if (unlikely(__PageMovable(page)) && | |
1111 | !PageIsolated(page)) { | |
1112 | if (locked) { | |
6168d0da AS |
1113 | unlock_page_lruvec_irqrestore(locked, flags); |
1114 | locked = NULL; | |
bda807d4 MK |
1115 | } |
1116 | ||
56ae0bb3 KW |
1117 | if (isolate_movable_page(page, mode)) { |
1118 | folio = page_folio(page); | |
bda807d4 | 1119 | goto isolate_success; |
56ae0bb3 | 1120 | } |
bda807d4 MK |
1121 | } |
1122 | ||
fdd048e1 | 1123 | goto isolate_fail; |
bda807d4 | 1124 | } |
29c0dde8 | 1125 | |
829ae0f8 GS |
1126 | /* |
1127 | * Be careful not to clear PageLRU until after we're | |
1128 | * sure the page is not being freed elsewhere -- the | |
1129 | * page release code relies on it. | |
1130 | */ | |
56ae0bb3 KW |
1131 | folio = folio_get_nontail_page(page); |
1132 | if (unlikely(!folio)) | |
829ae0f8 GS |
1133 | goto isolate_fail; |
1134 | ||
119d6d59 DR |
1135 | /* |
1136 | * Migration will fail if an anonymous page is pinned in memory, | |
1137 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
1138 | * admittedly racy check. | |
1139 | */ | |
56ae0bb3 KW |
1140 | mapping = folio_mapping(folio); |
1141 | if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio)) | |
829ae0f8 | 1142 | goto isolate_fail_put; |
119d6d59 | 1143 | |
73e64c51 MH |
1144 | /* |
1145 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
1146 | * because those do not depend on fs locks. | |
1147 | */ | |
89f6c88a | 1148 | if (!(cc->gfp_mask & __GFP_FS) && mapping) |
829ae0f8 | 1149 | goto isolate_fail_put; |
9df41314 | 1150 | |
89f6c88a | 1151 | /* Only take pages on LRU: a check now makes later tests safe */ |
56ae0bb3 | 1152 | if (!folio_test_lru(folio)) |
89f6c88a HD |
1153 | goto isolate_fail_put; |
1154 | ||
0003e2a4 SC |
1155 | is_unevictable = folio_test_unevictable(folio); |
1156 | ||
89f6c88a | 1157 | /* Compaction might skip unevictable pages but CMA takes them */ |
0003e2a4 | 1158 | if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable) |
89f6c88a HD |
1159 | goto isolate_fail_put; |
1160 | ||
1161 | /* | |
1162 | * To minimise LRU disruption, the caller can indicate with | |
1163 | * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages | |
1164 | * it will be able to migrate without blocking - clean pages | |
1165 | * for the most part. PageWriteback would require blocking. | |
1166 | */ | |
56ae0bb3 | 1167 | if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio)) |
9df41314 AS |
1168 | goto isolate_fail_put; |
1169 | ||
0003e2a4 SC |
1170 | is_dirty = folio_test_dirty(folio); |
1171 | ||
1172 | if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) || | |
1173 | (mapping && is_unevictable)) { | |
1174 | bool migrate_dirty = true; | |
1175 | bool is_unmovable; | |
89f6c88a HD |
1176 | |
1177 | /* | |
866ff801 | 1178 | * Only folios without mappings or that have |
0003e2a4 SC |
1179 | * a ->migrate_folio callback are possible to migrate |
1180 | * without blocking. | |
1181 | * | |
1182 | * Folios from unmovable mappings are not migratable. | |
1183 | * | |
1184 | * However, we can be racing with truncation, which can | |
1185 | * free the mapping that we need to check. Truncation | |
1186 | * holds the folio lock until after the folio is removed | |
1187 | * from the page so holding it ourselves is sufficient. | |
1188 | * | |
1189 | * To avoid locking the folio just to check unmovable, | |
1190 | * assume every unmovable folio is also unevictable, | |
1191 | * which is a cheaper test. If our assumption goes | |
1192 | * wrong, it's not a correctness bug, just potentially | |
1193 | * wasted cycles. | |
89f6c88a | 1194 | */ |
56ae0bb3 | 1195 | if (!folio_trylock(folio)) |
89f6c88a HD |
1196 | goto isolate_fail_put; |
1197 | ||
56ae0bb3 | 1198 | mapping = folio_mapping(folio); |
0003e2a4 SC |
1199 | if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) { |
1200 | migrate_dirty = !mapping || | |
1201 | mapping->a_ops->migrate_folio; | |
1202 | } | |
1203 | is_unmovable = mapping && mapping_unmovable(mapping); | |
56ae0bb3 | 1204 | folio_unlock(folio); |
0003e2a4 | 1205 | if (!migrate_dirty || is_unmovable) |
89f6c88a HD |
1206 | goto isolate_fail_put; |
1207 | } | |
1208 | ||
56ae0bb3 KW |
1209 | /* Try isolate the folio */ |
1210 | if (!folio_test_clear_lru(folio)) | |
9df41314 AS |
1211 | goto isolate_fail_put; |
1212 | ||
56ae0bb3 | 1213 | lruvec = folio_lruvec(folio); |
6168d0da | 1214 | |
69b7189f | 1215 | /* If we already hold the lock, we can skip some rechecking */ |
6168d0da AS |
1216 | if (lruvec != locked) { |
1217 | if (locked) | |
1218 | unlock_page_lruvec_irqrestore(locked, flags); | |
1219 | ||
1220 | compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); | |
1221 | locked = lruvec; | |
6168d0da | 1222 | |
56ae0bb3 | 1223 | lruvec_memcg_debug(lruvec, folio); |
e380bebe | 1224 | |
590ccea8 MG |
1225 | /* |
1226 | * Try get exclusive access under lock. If marked for | |
1227 | * skip, the scan is aborted unless the current context | |
1228 | * is a rescan to reach the end of the pageblock. | |
1229 | */ | |
1230 | if (!skip_updated && valid_page) { | |
e380bebe | 1231 | skip_updated = true; |
590ccea8 MG |
1232 | if (test_and_set_skip(cc, valid_page) && |
1233 | !cc->finish_pageblock) { | |
7545e2f2 | 1234 | low_pfn = end_pfn; |
e380bebe | 1235 | goto isolate_abort; |
590ccea8 | 1236 | } |
e380bebe | 1237 | } |
2a1402aa | 1238 | |
29c0dde8 | 1239 | /* |
ee6f62fd | 1240 | * Check LRU folio order under the lock |
29c0dde8 | 1241 | */ |
ee6f62fd ZY |
1242 | if (unlikely(skip_isolation_on_order(folio_order(folio), |
1243 | cc->order) && | |
1244 | !cc->alloc_contig)) { | |
56ae0bb3 KW |
1245 | low_pfn += folio_nr_pages(folio) - 1; |
1246 | nr_scanned += folio_nr_pages(folio) - 1; | |
1247 | folio_set_lru(folio); | |
9df41314 | 1248 | goto isolate_fail_put; |
69b7189f | 1249 | } |
d99fd5fe | 1250 | } |
fa9add64 | 1251 | |
56ae0bb3 KW |
1252 | /* The folio is taken off the LRU */ |
1253 | if (folio_test_large(folio)) | |
1254 | low_pfn += folio_nr_pages(folio) - 1; | |
bc835011 | 1255 | |
748446bb | 1256 | /* Successfully isolated */ |
56ae0bb3 KW |
1257 | lruvec_del_folio(lruvec, folio); |
1258 | node_stat_mod_folio(folio, | |
1259 | NR_ISOLATED_ANON + folio_is_file_lru(folio), | |
1260 | folio_nr_pages(folio)); | |
b6c75016 JK |
1261 | |
1262 | isolate_success: | |
56ae0bb3 | 1263 | list_add(&folio->lru, &cc->migratepages); |
ae37c7ff | 1264 | isolate_success_no_list: |
56ae0bb3 KW |
1265 | cc->nr_migratepages += folio_nr_pages(folio); |
1266 | nr_isolated += folio_nr_pages(folio); | |
1267 | nr_scanned += folio_nr_pages(folio) - 1; | |
748446bb | 1268 | |
804d3121 MG |
1269 | /* |
1270 | * Avoid isolating too much unless this block is being | |
48731c84 | 1271 | * fully scanned (e.g. dirty/writeback pages, parallel allocation) |
cb2dcaf0 MG |
1272 | * or a lock is contended. For contention, isolate quickly to |
1273 | * potentially remove one source of contention. | |
804d3121 | 1274 | */ |
38935861 | 1275 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && |
48731c84 | 1276 | !cc->finish_pageblock && !cc->contended) { |
31b8384a | 1277 | ++low_pfn; |
748446bb | 1278 | break; |
31b8384a | 1279 | } |
fdd048e1 VB |
1280 | |
1281 | continue; | |
9df41314 AS |
1282 | |
1283 | isolate_fail_put: | |
1284 | /* Avoid potential deadlock in freeing page under lru_lock */ | |
1285 | if (locked) { | |
6168d0da AS |
1286 | unlock_page_lruvec_irqrestore(locked, flags); |
1287 | locked = NULL; | |
9df41314 | 1288 | } |
56ae0bb3 | 1289 | folio_put(folio); |
9df41314 | 1290 | |
fdd048e1 | 1291 | isolate_fail: |
369fa227 | 1292 | if (!skip_on_failure && ret != -ENOMEM) |
fdd048e1 VB |
1293 | continue; |
1294 | ||
1295 | /* | |
1296 | * We have isolated some pages, but then failed. Release them | |
1297 | * instead of migrating, as we cannot form the cc->order buddy | |
1298 | * page anyway. | |
1299 | */ | |
1300 | if (nr_isolated) { | |
1301 | if (locked) { | |
6168d0da AS |
1302 | unlock_page_lruvec_irqrestore(locked, flags); |
1303 | locked = NULL; | |
fdd048e1 | 1304 | } |
fdd048e1 VB |
1305 | putback_movable_pages(&cc->migratepages); |
1306 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
1307 | nr_isolated = 0; |
1308 | } | |
1309 | ||
1310 | if (low_pfn < next_skip_pfn) { | |
1311 | low_pfn = next_skip_pfn - 1; | |
1312 | /* | |
1313 | * The check near the loop beginning would have updated | |
1314 | * next_skip_pfn too, but this is a bit simpler. | |
1315 | */ | |
1316 | next_skip_pfn += 1UL << cc->order; | |
1317 | } | |
369fa227 OS |
1318 | |
1319 | if (ret == -ENOMEM) | |
1320 | break; | |
748446bb MG |
1321 | } |
1322 | ||
99c0fd5e VB |
1323 | /* |
1324 | * The PageBuddy() check could have potentially brought us outside | |
1325 | * the range to be scanned. | |
1326 | */ | |
1327 | if (unlikely(low_pfn > end_pfn)) | |
1328 | low_pfn = end_pfn; | |
1329 | ||
56ae0bb3 | 1330 | folio = NULL; |
9df41314 | 1331 | |
e380bebe | 1332 | isolate_abort: |
c67fe375 | 1333 | if (locked) |
6168d0da | 1334 | unlock_page_lruvec_irqrestore(locked, flags); |
56ae0bb3 KW |
1335 | if (folio) { |
1336 | folio_set_lru(folio); | |
1337 | folio_put(folio); | |
9df41314 | 1338 | } |
748446bb | 1339 | |
50b5b094 | 1340 | /* |
48731c84 | 1341 | * Update the cached scanner pfn once the pageblock has been scanned. |
804d3121 MG |
1342 | * Pages will either be migrated in which case there is no point |
1343 | * scanning in the near future or migration failed in which case the | |
1344 | * failure reason may persist. The block is marked for skipping if | |
1345 | * there were no pages isolated in the block or if the block is | |
1346 | * rescanned twice in a row. | |
50b5b094 | 1347 | */ |
48731c84 | 1348 | if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) { |
8b71b499 | 1349 | if (!cc->no_set_skip_hint && valid_page && !skip_updated) |
e380bebe MG |
1350 | set_pageblock_skip(valid_page); |
1351 | update_cached_migrate(cc, low_pfn); | |
1352 | } | |
bb13ffeb | 1353 | |
e34d85f0 JK |
1354 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
1355 | nr_scanned, nr_isolated); | |
b7aba698 | 1356 | |
670105a2 | 1357 | fatal_pending: |
7f354a54 | 1358 | cc->total_migrate_scanned += nr_scanned; |
397487db | 1359 | if (nr_isolated) |
010fc29a | 1360 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 1361 | |
c2ad7a1f OS |
1362 | cc->migrate_pfn = low_pfn; |
1363 | ||
1364 | return ret; | |
2fe86e00 MN |
1365 | } |
1366 | ||
edc2ca61 VB |
1367 | /** |
1368 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
1369 | * @cc: Compaction control structure. | |
1370 | * @start_pfn: The first PFN to start isolating. | |
1371 | * @end_pfn: The one-past-last PFN. | |
1372 | * | |
369fa227 OS |
1373 | * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM |
1374 | * in case we could not allocate a page, or 0. | |
edc2ca61 | 1375 | */ |
c2ad7a1f | 1376 | int |
edc2ca61 VB |
1377 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, |
1378 | unsigned long end_pfn) | |
1379 | { | |
e1409c32 | 1380 | unsigned long pfn, block_start_pfn, block_end_pfn; |
c2ad7a1f | 1381 | int ret = 0; |
edc2ca61 VB |
1382 | |
1383 | /* Scan block by block. First and last block may be incomplete */ | |
1384 | pfn = start_pfn; | |
06b6640a | 1385 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
1386 | if (block_start_pfn < cc->zone->zone_start_pfn) |
1387 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 1388 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
1389 | |
1390 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 1391 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
1392 | block_end_pfn += pageblock_nr_pages) { |
1393 | ||
1394 | block_end_pfn = min(block_end_pfn, end_pfn); | |
1395 | ||
e1409c32 JK |
1396 | if (!pageblock_pfn_to_page(block_start_pfn, |
1397 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1398 | continue; |
1399 | ||
c2ad7a1f OS |
1400 | ret = isolate_migratepages_block(cc, pfn, block_end_pfn, |
1401 | ISOLATE_UNEVICTABLE); | |
edc2ca61 | 1402 | |
c2ad7a1f | 1403 | if (ret) |
edc2ca61 | 1404 | break; |
6ea41c0c | 1405 | |
38935861 | 1406 | if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) |
6ea41c0c | 1407 | break; |
edc2ca61 | 1408 | } |
edc2ca61 | 1409 | |
c2ad7a1f | 1410 | return ret; |
edc2ca61 VB |
1411 | } |
1412 | ||
ff9543fd MN |
1413 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1414 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1415 | |
b682debd VB |
1416 | static bool suitable_migration_source(struct compact_control *cc, |
1417 | struct page *page) | |
1418 | { | |
282722b0 VB |
1419 | int block_mt; |
1420 | ||
9bebefd5 MG |
1421 | if (pageblock_skip_persistent(page)) |
1422 | return false; | |
1423 | ||
282722b0 | 1424 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1425 | return true; |
1426 | ||
282722b0 VB |
1427 | block_mt = get_pageblock_migratetype(page); |
1428 | ||
1429 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1430 | return is_migrate_movable(block_mt); | |
1431 | else | |
1432 | return block_mt == cc->migratetype; | |
b682debd VB |
1433 | } |
1434 | ||
018e9a49 | 1435 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1436 | static bool suitable_migration_target(struct compact_control *cc, |
1437 | struct page *page) | |
018e9a49 AM |
1438 | { |
1439 | /* If the page is a large free page, then disallow migration */ | |
1440 | if (PageBuddy(page)) { | |
1883e8ac BW |
1441 | int order = cc->order > 0 ? cc->order : pageblock_order; |
1442 | ||
018e9a49 AM |
1443 | /* |
1444 | * We are checking page_order without zone->lock taken. But | |
1445 | * the only small danger is that we skip a potentially suitable | |
1446 | * pageblock, so it's not worth to check order for valid range. | |
1447 | */ | |
1883e8ac | 1448 | if (buddy_order_unsafe(page) >= order) |
018e9a49 AM |
1449 | return false; |
1450 | } | |
1451 | ||
1ef36db2 YX |
1452 | if (cc->ignore_block_suitable) |
1453 | return true; | |
1454 | ||
018e9a49 | 1455 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1456 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1457 | return true; |
1458 | ||
1459 | /* Otherwise skip the block */ | |
1460 | return false; | |
1461 | } | |
1462 | ||
70b44595 MG |
1463 | static inline unsigned int |
1464 | freelist_scan_limit(struct compact_control *cc) | |
1465 | { | |
dd7ef7bd QC |
1466 | unsigned short shift = BITS_PER_LONG - 1; |
1467 | ||
1468 | return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; | |
70b44595 MG |
1469 | } |
1470 | ||
f2849aa0 VB |
1471 | /* |
1472 | * Test whether the free scanner has reached the same or lower pageblock than | |
1473 | * the migration scanner, and compaction should thus terminate. | |
1474 | */ | |
1475 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1476 | { | |
1477 | return (cc->free_pfn >> pageblock_order) | |
1478 | <= (cc->migrate_pfn >> pageblock_order); | |
1479 | } | |
1480 | ||
5a811889 MG |
1481 | /* |
1482 | * Used when scanning for a suitable migration target which scans freelists | |
1483 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1484 | * first on the next iteration of the free scanner | |
1485 | */ | |
1486 | static void | |
1487 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1488 | { | |
1489 | LIST_HEAD(sublist); | |
1490 | ||
4c179891 | 1491 | if (!list_is_first(&freepage->buddy_list, freelist)) { |
bbefa0fc | 1492 | list_cut_before(&sublist, freelist, &freepage->buddy_list); |
d2155fe5 | 1493 | list_splice_tail(&sublist, freelist); |
5a811889 MG |
1494 | } |
1495 | } | |
1496 | ||
1497 | /* | |
1498 | * Similar to move_freelist_head except used by the migration scanner | |
1499 | * when scanning forward. It's possible for these list operations to | |
1500 | * move against each other if they search the free list exactly in | |
1501 | * lockstep. | |
1502 | */ | |
70b44595 MG |
1503 | static void |
1504 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1505 | { | |
1506 | LIST_HEAD(sublist); | |
1507 | ||
4c179891 | 1508 | if (!list_is_last(&freepage->buddy_list, freelist)) { |
bbefa0fc | 1509 | list_cut_position(&sublist, freelist, &freepage->buddy_list); |
d2155fe5 | 1510 | list_splice_tail(&sublist, freelist); |
70b44595 MG |
1511 | } |
1512 | } | |
1513 | ||
5a811889 | 1514 | static void |
be21b32a | 1515 | fast_isolate_around(struct compact_control *cc, unsigned long pfn) |
5a811889 MG |
1516 | { |
1517 | unsigned long start_pfn, end_pfn; | |
6e2b7044 | 1518 | struct page *page; |
5a811889 MG |
1519 | |
1520 | /* Do not search around if there are enough pages already */ | |
1521 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1522 | return; | |
1523 | ||
1524 | /* Minimise scanning during async compaction */ | |
1525 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1526 | return; | |
1527 | ||
1528 | /* Pageblock boundaries */ | |
6e2b7044 VB |
1529 | start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); |
1530 | end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); | |
1531 | ||
1532 | page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); | |
1533 | if (!page) | |
1534 | return; | |
5a811889 | 1535 | |
733aea0b | 1536 | isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false); |
5a811889 MG |
1537 | |
1538 | /* Skip this pageblock in the future as it's full or nearly full */ | |
18c59d58 | 1539 | if (start_pfn == end_pfn && !cc->no_set_skip_hint) |
5a811889 MG |
1540 | set_pageblock_skip(page); |
1541 | } | |
1542 | ||
dbe2d4e4 MG |
1543 | /* Search orders in round-robin fashion */ |
1544 | static int next_search_order(struct compact_control *cc, int order) | |
1545 | { | |
1546 | order--; | |
1547 | if (order < 0) | |
1548 | order = cc->order - 1; | |
1549 | ||
1550 | /* Search wrapped around? */ | |
1551 | if (order == cc->search_order) { | |
1552 | cc->search_order--; | |
1553 | if (cc->search_order < 0) | |
1554 | cc->search_order = cc->order - 1; | |
1555 | return -1; | |
1556 | } | |
1557 | ||
1558 | return order; | |
1559 | } | |
1560 | ||
2dbd9005 | 1561 | static void fast_isolate_freepages(struct compact_control *cc) |
5a811889 | 1562 | { |
b55ca526 | 1563 | unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); |
447ba886 | 1564 | unsigned int nr_scanned = 0, total_isolated = 0; |
74e21484 | 1565 | unsigned long low_pfn, min_pfn, highest = 0; |
5a811889 MG |
1566 | unsigned long nr_isolated = 0; |
1567 | unsigned long distance; | |
1568 | struct page *page = NULL; | |
1569 | bool scan_start = false; | |
1570 | int order; | |
1571 | ||
1572 | /* Full compaction passes in a negative order */ | |
1573 | if (cc->order <= 0) | |
2dbd9005 | 1574 | return; |
5a811889 MG |
1575 | |
1576 | /* | |
1577 | * If starting the scan, use a deeper search and use the highest | |
1578 | * PFN found if a suitable one is not found. | |
1579 | */ | |
e332f741 | 1580 | if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { |
5a811889 MG |
1581 | limit = pageblock_nr_pages >> 1; |
1582 | scan_start = true; | |
1583 | } | |
1584 | ||
1585 | /* | |
1586 | * Preferred point is in the top quarter of the scan space but take | |
1587 | * a pfn from the top half if the search is problematic. | |
1588 | */ | |
1589 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1590 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1591 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1592 | ||
1593 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1594 | low_pfn = min_pfn; | |
1595 | ||
dbe2d4e4 MG |
1596 | /* |
1597 | * Search starts from the last successful isolation order or the next | |
1598 | * order to search after a previous failure | |
1599 | */ | |
1600 | cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); | |
1601 | ||
1602 | for (order = cc->search_order; | |
1603 | !page && order >= 0; | |
1604 | order = next_search_order(cc, order)) { | |
5a811889 MG |
1605 | struct free_area *area = &cc->zone->free_area[order]; |
1606 | struct list_head *freelist; | |
1607 | struct page *freepage; | |
1608 | unsigned long flags; | |
1609 | unsigned int order_scanned = 0; | |
74e21484 | 1610 | unsigned long high_pfn = 0; |
5a811889 MG |
1611 | |
1612 | if (!area->nr_free) | |
1613 | continue; | |
1614 | ||
1615 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1616 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
94ec2003 | 1617 | list_for_each_entry_reverse(freepage, freelist, buddy_list) { |
5a811889 MG |
1618 | unsigned long pfn; |
1619 | ||
1620 | order_scanned++; | |
1621 | nr_scanned++; | |
1622 | pfn = page_to_pfn(freepage); | |
1623 | ||
1624 | if (pfn >= highest) | |
6e2b7044 VB |
1625 | highest = max(pageblock_start_pfn(pfn), |
1626 | cc->zone->zone_start_pfn); | |
5a811889 MG |
1627 | |
1628 | if (pfn >= low_pfn) { | |
1629 | cc->fast_search_fail = 0; | |
dbe2d4e4 | 1630 | cc->search_order = order; |
5a811889 MG |
1631 | page = freepage; |
1632 | break; | |
1633 | } | |
1634 | ||
1635 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1636 | high_pfn = pfn; | |
1637 | ||
1638 | /* Shorten the scan if a candidate is found */ | |
1639 | limit >>= 1; | |
1640 | } | |
1641 | ||
1642 | if (order_scanned >= limit) | |
1643 | break; | |
1644 | } | |
1645 | ||
e6bd14ec | 1646 | /* Use a maximum candidate pfn if a preferred one was not found */ |
5a811889 MG |
1647 | if (!page && high_pfn) { |
1648 | page = pfn_to_page(high_pfn); | |
1649 | ||
1650 | /* Update freepage for the list reorder below */ | |
1651 | freepage = page; | |
1652 | } | |
1653 | ||
1654 | /* Reorder to so a future search skips recent pages */ | |
1655 | move_freelist_head(freelist, freepage); | |
1656 | ||
1657 | /* Isolate the page if available */ | |
1658 | if (page) { | |
1659 | if (__isolate_free_page(page, order)) { | |
1660 | set_page_private(page, order); | |
1661 | nr_isolated = 1 << order; | |
b717d6b9 | 1662 | nr_scanned += nr_isolated - 1; |
447ba886 | 1663 | total_isolated += nr_isolated; |
5a811889 | 1664 | cc->nr_freepages += nr_isolated; |
733aea0b | 1665 | list_add_tail(&page->lru, &cc->freepages[order]); |
5a811889 MG |
1666 | count_compact_events(COMPACTISOLATED, nr_isolated); |
1667 | } else { | |
1668 | /* If isolation fails, abort the search */ | |
5b56d996 | 1669 | order = cc->search_order + 1; |
5a811889 MG |
1670 | page = NULL; |
1671 | } | |
1672 | } | |
1673 | ||
1674 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1675 | ||
a8d13355 BW |
1676 | /* Skip fast search if enough freepages isolated */ |
1677 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1678 | break; | |
1679 | ||
5a811889 | 1680 | /* |
b55ca526 | 1681 | * Smaller scan on next order so the total scan is related |
5a811889 MG |
1682 | * to freelist_scan_limit. |
1683 | */ | |
1684 | if (order_scanned >= limit) | |
b55ca526 | 1685 | limit = max(1U, limit >> 1); |
5a811889 MG |
1686 | } |
1687 | ||
447ba886 BW |
1688 | trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn, |
1689 | nr_scanned, total_isolated); | |
1690 | ||
5a811889 MG |
1691 | if (!page) { |
1692 | cc->fast_search_fail++; | |
1693 | if (scan_start) { | |
1694 | /* | |
1695 | * Use the highest PFN found above min. If one was | |
f3867755 | 1696 | * not found, be pessimistic for direct compaction |
5a811889 MG |
1697 | * and use the min mark. |
1698 | */ | |
ca2864e5 | 1699 | if (highest >= min_pfn) { |
5a811889 MG |
1700 | page = pfn_to_page(highest); |
1701 | cc->free_pfn = highest; | |
1702 | } else { | |
e577c8b6 | 1703 | if (cc->direct_compaction && pfn_valid(min_pfn)) { |
73a6e474 | 1704 | page = pageblock_pfn_to_page(min_pfn, |
6e2b7044 VB |
1705 | min(pageblock_end_pfn(min_pfn), |
1706 | zone_end_pfn(cc->zone)), | |
73a6e474 | 1707 | cc->zone); |
d19b1a17 BS |
1708 | if (page && !suitable_migration_target(cc, page)) |
1709 | page = NULL; | |
1710 | ||
5a811889 MG |
1711 | cc->free_pfn = min_pfn; |
1712 | } | |
1713 | } | |
1714 | } | |
1715 | } | |
1716 | ||
d097a6f6 MG |
1717 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1718 | highest -= pageblock_nr_pages; | |
5a811889 | 1719 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1720 | } |
5a811889 MG |
1721 | |
1722 | cc->total_free_scanned += nr_scanned; | |
1723 | if (!page) | |
2dbd9005 | 1724 | return; |
5a811889 MG |
1725 | |
1726 | low_pfn = page_to_pfn(page); | |
be21b32a | 1727 | fast_isolate_around(cc, low_pfn); |
5a811889 MG |
1728 | } |
1729 | ||
2fe86e00 | 1730 | /* |
ff9543fd MN |
1731 | * Based on information in the current compact_control, find blocks |
1732 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1733 | */ |
edc2ca61 | 1734 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1735 | { |
edc2ca61 | 1736 | struct zone *zone = cc->zone; |
ff9543fd | 1737 | struct page *page; |
c96b9e50 | 1738 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1739 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1740 | unsigned long block_end_pfn; /* end of current pageblock */ |
1741 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
4fca9730 | 1742 | unsigned int stride; |
2fe86e00 | 1743 | |
5a811889 | 1744 | /* Try a small search of the free lists for a candidate */ |
00bc102f | 1745 | fast_isolate_freepages(cc); |
5a811889 | 1746 | if (cc->nr_freepages) |
733aea0b | 1747 | return; |
5a811889 | 1748 | |
ff9543fd MN |
1749 | /* |
1750 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1751 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1752 | * zone when isolating for the first time. For looping we also need |
1753 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1754 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1755 | * For ending point, take care when isolating in last pageblock of a | |
a1c1dbeb | 1756 | * zone which ends in the middle of a pageblock. |
49e068f0 VB |
1757 | * The low boundary is the end of the pageblock the migration scanner |
1758 | * is using. | |
ff9543fd | 1759 | */ |
e14c720e | 1760 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1761 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1762 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1763 | zone_end_pfn(zone)); | |
06b6640a | 1764 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
4fca9730 | 1765 | stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; |
2fe86e00 | 1766 | |
ff9543fd MN |
1767 | /* |
1768 | * Isolate free pages until enough are available to migrate the | |
1769 | * pages on cc->migratepages. We stop searching if the migrate | |
1770 | * and free page scanners meet or enough free pages are isolated. | |
1771 | */ | |
f5f61a32 | 1772 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1773 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1774 | block_start_pfn -= pageblock_nr_pages, |
1775 | isolate_start_pfn = block_start_pfn) { | |
4fca9730 MG |
1776 | unsigned long nr_isolated; |
1777 | ||
f6ea3adb DR |
1778 | /* |
1779 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1780 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1781 | */ |
c036ddff | 1782 | if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1783 | cond_resched(); |
f6ea3adb | 1784 | |
7d49d886 VB |
1785 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1786 | zone); | |
e6e0c767 BW |
1787 | if (!page) { |
1788 | unsigned long next_pfn; | |
1789 | ||
1790 | next_pfn = skip_offline_sections_reverse(block_start_pfn); | |
1791 | if (next_pfn) | |
1792 | block_start_pfn = max(next_pfn, low_pfn); | |
1793 | ||
ff9543fd | 1794 | continue; |
e6e0c767 | 1795 | } |
ff9543fd MN |
1796 | |
1797 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1798 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1799 | continue; |
68e3e926 | 1800 | |
bb13ffeb MG |
1801 | /* If isolation recently failed, do not retry */ |
1802 | if (!isolation_suitable(cc, page)) | |
1803 | continue; | |
1804 | ||
e14c720e | 1805 | /* Found a block suitable for isolating free pages from. */ |
4fca9730 | 1806 | nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
733aea0b | 1807 | block_end_pfn, cc->freepages, stride, false); |
ff9543fd | 1808 | |
d097a6f6 MG |
1809 | /* Update the skip hint if the full pageblock was scanned */ |
1810 | if (isolate_start_pfn == block_end_pfn) | |
16951789 KS |
1811 | update_pageblock_skip(cc, page, block_start_pfn - |
1812 | pageblock_nr_pages); | |
d097a6f6 | 1813 | |
cb2dcaf0 MG |
1814 | /* Are enough freepages isolated? */ |
1815 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1816 | if (isolate_start_pfn >= block_end_pfn) { |
1817 | /* | |
1818 | * Restart at previous pageblock if more | |
1819 | * freepages can be isolated next time. | |
1820 | */ | |
f5f61a32 VB |
1821 | isolate_start_pfn = |
1822 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1823 | } |
be976572 | 1824 | break; |
a46cbf3b | 1825 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1826 | /* |
a46cbf3b DR |
1827 | * If isolation failed early, do not continue |
1828 | * needlessly. | |
f5f61a32 | 1829 | */ |
a46cbf3b | 1830 | break; |
f5f61a32 | 1831 | } |
4fca9730 MG |
1832 | |
1833 | /* Adjust stride depending on isolation */ | |
1834 | if (nr_isolated) { | |
1835 | stride = 1; | |
1836 | continue; | |
1837 | } | |
1838 | stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); | |
ff9543fd MN |
1839 | } |
1840 | ||
7ed695e0 | 1841 | /* |
f5f61a32 VB |
1842 | * Record where the free scanner will restart next time. Either we |
1843 | * broke from the loop and set isolate_start_pfn based on the last | |
1844 | * call to isolate_freepages_block(), or we met the migration scanner | |
1845 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1846 | */ |
f5f61a32 | 1847 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1848 | } |
1849 | ||
1850 | /* | |
1851 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1852 | * from the isolated freelists in the block we are migrating to. | |
1853 | */ | |
4e096ae1 | 1854 | static struct folio *compaction_alloc(struct folio *src, unsigned long data) |
748446bb MG |
1855 | { |
1856 | struct compact_control *cc = (struct compact_control *)data; | |
4e096ae1 | 1857 | struct folio *dst; |
733aea0b | 1858 | int order = folio_order(src); |
73318e2c ZY |
1859 | bool has_isolated_pages = false; |
1860 | int start_order; | |
1861 | struct page *freepage; | |
1862 | unsigned long size; | |
1863 | ||
1864 | again: | |
1865 | for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++) | |
1866 | if (!list_empty(&cc->freepages[start_order])) | |
1867 | break; | |
748446bb | 1868 | |
73318e2c ZY |
1869 | /* no free pages in the list */ |
1870 | if (start_order == NR_PAGE_ORDERS) { | |
1871 | if (has_isolated_pages) | |
748446bb | 1872 | return NULL; |
73318e2c ZY |
1873 | isolate_freepages(cc); |
1874 | has_isolated_pages = true; | |
1875 | goto again; | |
748446bb MG |
1876 | } |
1877 | ||
73318e2c ZY |
1878 | freepage = list_first_entry(&cc->freepages[start_order], struct page, |
1879 | lru); | |
1880 | size = 1 << start_order; | |
1881 | ||
1882 | list_del(&freepage->lru); | |
748446bb | 1883 | |
73318e2c ZY |
1884 | while (start_order > order) { |
1885 | start_order--; | |
1886 | size >>= 1; | |
1887 | ||
1888 | list_add(&freepage[size].lru, &cc->freepages[start_order]); | |
1889 | set_page_private(&freepage[size], start_order); | |
748446bb | 1890 | } |
73318e2c | 1891 | dst = (struct folio *)freepage; |
748446bb | 1892 | |
733aea0b ZY |
1893 | post_alloc_hook(&dst->page, order, __GFP_MOVABLE); |
1894 | if (order) | |
1895 | prep_compound_page(&dst->page, order); | |
1896 | cc->nr_freepages -= 1 << order; | |
1897 | cc->nr_migratepages -= 1 << order; | |
1898 | return page_rmappable_folio(&dst->page); | |
748446bb MG |
1899 | } |
1900 | ||
1901 | /* | |
d53aea3d DR |
1902 | * This is a migrate-callback that "frees" freepages back to the isolated |
1903 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1904 | * special handling needed for NUMA. | |
1905 | */ | |
4e096ae1 | 1906 | static void compaction_free(struct folio *dst, unsigned long data) |
d53aea3d DR |
1907 | { |
1908 | struct compact_control *cc = (struct compact_control *)data; | |
733aea0b ZY |
1909 | int order = folio_order(dst); |
1910 | struct page *page = &dst->page; | |
d53aea3d | 1911 | |
733aea0b ZY |
1912 | if (folio_put_testzero(dst)) { |
1913 | free_pages_prepare(page, order); | |
1914 | list_add(&dst->lru, &cc->freepages[order]); | |
1915 | cc->nr_freepages += 1 << order; | |
1916 | } | |
1917 | cc->nr_migratepages += 1 << order; | |
1918 | /* | |
1919 | * someone else has referenced the page, we cannot take it back to our | |
1920 | * free list. | |
1921 | */ | |
d53aea3d DR |
1922 | } |
1923 | ||
ff9543fd MN |
1924 | /* possible outcome of isolate_migratepages */ |
1925 | typedef enum { | |
1926 | ISOLATE_ABORT, /* Abort compaction now */ | |
1927 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1928 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1929 | } isolate_migrate_t; | |
1930 | ||
5bbe3547 EM |
1931 | /* |
1932 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1933 | * compactable pages. | |
1934 | */ | |
48fe8ab8 MC |
1935 | static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT; |
1936 | /* | |
1937 | * Tunable for proactive compaction. It determines how | |
1938 | * aggressively the kernel should compact memory in the | |
1939 | * background. It takes values in the range [0, 100]. | |
1940 | */ | |
1941 | static unsigned int __read_mostly sysctl_compaction_proactiveness = 20; | |
1942 | static int sysctl_extfrag_threshold = 500; | |
8b9167cd | 1943 | static int __read_mostly sysctl_compact_memory; |
5bbe3547 | 1944 | |
70b44595 MG |
1945 | static inline void |
1946 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1947 | { | |
1948 | if (cc->fast_start_pfn == ULONG_MAX) | |
1949 | return; | |
1950 | ||
1951 | if (!cc->fast_start_pfn) | |
1952 | cc->fast_start_pfn = pfn; | |
1953 | ||
1954 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1955 | } | |
1956 | ||
1957 | static inline unsigned long | |
1958 | reinit_migrate_pfn(struct compact_control *cc) | |
1959 | { | |
1960 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1961 | return cc->migrate_pfn; | |
1962 | ||
1963 | cc->migrate_pfn = cc->fast_start_pfn; | |
1964 | cc->fast_start_pfn = ULONG_MAX; | |
1965 | ||
1966 | return cc->migrate_pfn; | |
1967 | } | |
1968 | ||
1969 | /* | |
1970 | * Briefly search the free lists for a migration source that already has | |
1971 | * some free pages to reduce the number of pages that need migration | |
1972 | * before a pageblock is free. | |
1973 | */ | |
1974 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1975 | { | |
1976 | unsigned int limit = freelist_scan_limit(cc); | |
1977 | unsigned int nr_scanned = 0; | |
1978 | unsigned long distance; | |
1979 | unsigned long pfn = cc->migrate_pfn; | |
1980 | unsigned long high_pfn; | |
1981 | int order; | |
15d28d0d | 1982 | bool found_block = false; |
70b44595 MG |
1983 | |
1984 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1985 | if (cc->ignore_skip_hint) | |
1986 | return pfn; | |
1987 | ||
f9d7fc1a MG |
1988 | /* |
1989 | * If the pageblock should be finished then do not select a different | |
1990 | * pageblock. | |
1991 | */ | |
1992 | if (cc->finish_pageblock) | |
1993 | return pfn; | |
1994 | ||
70b44595 MG |
1995 | /* |
1996 | * If the migrate_pfn is not at the start of a zone or the start | |
1997 | * of a pageblock then assume this is a continuation of a previous | |
1998 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1999 | */ | |
2000 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
2001 | return pfn; | |
2002 | ||
2003 | /* | |
2004 | * For smaller orders, just linearly scan as the number of pages | |
2005 | * to migrate should be relatively small and does not necessarily | |
2006 | * justify freeing up a large block for a small allocation. | |
2007 | */ | |
2008 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2009 | return pfn; | |
2010 | ||
2011 | /* | |
2012 | * Only allow kcompactd and direct requests for movable pages to | |
2013 | * quickly clear out a MOVABLE pageblock for allocation. This | |
2014 | * reduces the risk that a large movable pageblock is freed for | |
2015 | * an unmovable/reclaimable small allocation. | |
2016 | */ | |
2017 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
2018 | return pfn; | |
2019 | ||
2020 | /* | |
2021 | * When starting the migration scanner, pick any pageblock within the | |
2022 | * first half of the search space. Otherwise try and pick a pageblock | |
2023 | * within the first eighth to reduce the chances that a migration | |
2024 | * target later becomes a source. | |
2025 | */ | |
2026 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
2027 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
2028 | distance >>= 2; | |
2029 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
2030 | ||
2031 | for (order = cc->order - 1; | |
15d28d0d | 2032 | order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; |
70b44595 MG |
2033 | order--) { |
2034 | struct free_area *area = &cc->zone->free_area[order]; | |
2035 | struct list_head *freelist; | |
2036 | unsigned long flags; | |
2037 | struct page *freepage; | |
2038 | ||
2039 | if (!area->nr_free) | |
2040 | continue; | |
2041 | ||
2042 | spin_lock_irqsave(&cc->zone->lock, flags); | |
2043 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
94ec2003 | 2044 | list_for_each_entry(freepage, freelist, buddy_list) { |
70b44595 MG |
2045 | unsigned long free_pfn; |
2046 | ||
15d28d0d WY |
2047 | if (nr_scanned++ >= limit) { |
2048 | move_freelist_tail(freelist, freepage); | |
2049 | break; | |
2050 | } | |
2051 | ||
70b44595 MG |
2052 | free_pfn = page_to_pfn(freepage); |
2053 | if (free_pfn < high_pfn) { | |
70b44595 MG |
2054 | /* |
2055 | * Avoid if skipped recently. Ideally it would | |
2056 | * move to the tail but even safe iteration of | |
2057 | * the list assumes an entry is deleted, not | |
2058 | * reordered. | |
2059 | */ | |
15d28d0d | 2060 | if (get_pageblock_skip(freepage)) |
70b44595 | 2061 | continue; |
70b44595 MG |
2062 | |
2063 | /* Reorder to so a future search skips recent pages */ | |
2064 | move_freelist_tail(freelist, freepage); | |
2065 | ||
e380bebe | 2066 | update_fast_start_pfn(cc, free_pfn); |
70b44595 | 2067 | pfn = pageblock_start_pfn(free_pfn); |
bbe832b9 RY |
2068 | if (pfn < cc->zone->zone_start_pfn) |
2069 | pfn = cc->zone->zone_start_pfn; | |
70b44595 | 2070 | cc->fast_search_fail = 0; |
15d28d0d | 2071 | found_block = true; |
70b44595 MG |
2072 | break; |
2073 | } | |
70b44595 MG |
2074 | } |
2075 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
2076 | } | |
2077 | ||
2078 | cc->total_migrate_scanned += nr_scanned; | |
2079 | ||
2080 | /* | |
2081 | * If fast scanning failed then use a cached entry for a page block | |
2082 | * that had free pages as the basis for starting a linear scan. | |
2083 | */ | |
15d28d0d WY |
2084 | if (!found_block) { |
2085 | cc->fast_search_fail++; | |
70b44595 | 2086 | pfn = reinit_migrate_pfn(cc); |
15d28d0d | 2087 | } |
70b44595 MG |
2088 | return pfn; |
2089 | } | |
2090 | ||
ff9543fd | 2091 | /* |
edc2ca61 VB |
2092 | * Isolate all pages that can be migrated from the first suitable block, |
2093 | * starting at the block pointed to by the migrate scanner pfn within | |
2094 | * compact_control. | |
ff9543fd | 2095 | */ |
32aaf055 | 2096 | static isolate_migrate_t isolate_migratepages(struct compact_control *cc) |
ff9543fd | 2097 | { |
e1409c32 JK |
2098 | unsigned long block_start_pfn; |
2099 | unsigned long block_end_pfn; | |
2100 | unsigned long low_pfn; | |
edc2ca61 VB |
2101 | struct page *page; |
2102 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 2103 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 2104 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 2105 | bool fast_find_block; |
ff9543fd | 2106 | |
edc2ca61 VB |
2107 | /* |
2108 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
2109 | * initialized by compact_zone(). The first failure will use |
2110 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 2111 | */ |
70b44595 | 2112 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 2113 | block_start_pfn = pageblock_start_pfn(low_pfn); |
32aaf055 PL |
2114 | if (block_start_pfn < cc->zone->zone_start_pfn) |
2115 | block_start_pfn = cc->zone->zone_start_pfn; | |
ff9543fd | 2116 | |
70b44595 | 2117 | /* |
0aa8ea3c KS |
2118 | * fast_find_migrateblock() has already ensured the pageblock is not |
2119 | * set with a skipped flag, so to avoid the isolation_suitable check | |
2120 | * below again, check whether the fast search was successful. | |
70b44595 MG |
2121 | */ |
2122 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
2123 | ||
ff9543fd | 2124 | /* Only scan within a pageblock boundary */ |
06b6640a | 2125 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 2126 | |
edc2ca61 VB |
2127 | /* |
2128 | * Iterate over whole pageblocks until we find the first suitable. | |
2129 | * Do not cross the free scanner. | |
2130 | */ | |
e1409c32 | 2131 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 2132 | fast_find_block = false, |
c2ad7a1f | 2133 | cc->migrate_pfn = low_pfn = block_end_pfn, |
e1409c32 JK |
2134 | block_start_pfn = block_end_pfn, |
2135 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 2136 | |
edc2ca61 VB |
2137 | /* |
2138 | * This can potentially iterate a massively long zone with | |
2139 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 2140 | * need to schedule. |
edc2ca61 | 2141 | */ |
c036ddff | 2142 | if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 2143 | cond_resched(); |
ff9543fd | 2144 | |
32aaf055 PL |
2145 | page = pageblock_pfn_to_page(block_start_pfn, |
2146 | block_end_pfn, cc->zone); | |
9721fd82 BW |
2147 | if (!page) { |
2148 | unsigned long next_pfn; | |
2149 | ||
2150 | next_pfn = skip_offline_sections(block_start_pfn); | |
2151 | if (next_pfn) | |
2152 | block_end_pfn = min(next_pfn, cc->free_pfn); | |
edc2ca61 | 2153 | continue; |
9721fd82 | 2154 | } |
edc2ca61 | 2155 | |
e380bebe MG |
2156 | /* |
2157 | * If isolation recently failed, do not retry. Only check the | |
2158 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
2159 | * to be visited multiple times. Assume skip was checked | |
2160 | * before making it "skip" so other compaction instances do | |
2161 | * not scan the same block. | |
2162 | */ | |
493614da JW |
2163 | if ((pageblock_aligned(low_pfn) || |
2164 | low_pfn == cc->zone->zone_start_pfn) && | |
e380bebe | 2165 | !fast_find_block && !isolation_suitable(cc, page)) |
edc2ca61 VB |
2166 | continue; |
2167 | ||
2168 | /* | |
556162bf ML |
2169 | * For async direct compaction, only scan the pageblocks of the |
2170 | * same migratetype without huge pages. Async direct compaction | |
2171 | * is optimistic to see if the minimum amount of work satisfies | |
2172 | * the allocation. The cached PFN is updated as it's possible | |
2173 | * that all remaining blocks between source and target are | |
2174 | * unsuitable and the compaction scanners fail to meet. | |
edc2ca61 | 2175 | */ |
9bebefd5 MG |
2176 | if (!suitable_migration_source(cc, page)) { |
2177 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 2178 | continue; |
9bebefd5 | 2179 | } |
edc2ca61 VB |
2180 | |
2181 | /* Perform the isolation */ | |
c2ad7a1f OS |
2182 | if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, |
2183 | isolate_mode)) | |
edc2ca61 VB |
2184 | return ISOLATE_ABORT; |
2185 | ||
2186 | /* | |
2187 | * Either we isolated something and proceed with migration. Or | |
2188 | * we failed and compact_zone should decide if we should | |
2189 | * continue or not. | |
2190 | */ | |
2191 | break; | |
2192 | } | |
2193 | ||
edc2ca61 | 2194 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
2195 | } |
2196 | ||
b4a0215e KW |
2197 | /* |
2198 | * Determine whether kswapd is (or recently was!) running on this node. | |
2199 | * | |
2200 | * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't | |
2201 | * zero it. | |
2202 | */ | |
facdaa91 NG |
2203 | static bool kswapd_is_running(pg_data_t *pgdat) |
2204 | { | |
b4a0215e KW |
2205 | bool running; |
2206 | ||
2207 | pgdat_kswapd_lock(pgdat); | |
2208 | running = pgdat->kswapd && task_is_running(pgdat->kswapd); | |
2209 | pgdat_kswapd_unlock(pgdat); | |
2210 | ||
2211 | return running; | |
facdaa91 NG |
2212 | } |
2213 | ||
2214 | /* | |
2215 | * A zone's fragmentation score is the external fragmentation wrt to the | |
40d7e203 CTR |
2216 | * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. |
2217 | */ | |
2218 | static unsigned int fragmentation_score_zone(struct zone *zone) | |
2219 | { | |
2220 | return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); | |
2221 | } | |
2222 | ||
2223 | /* | |
2224 | * A weighted zone's fragmentation score is the external fragmentation | |
2225 | * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It | |
2226 | * returns a value in the range [0, 100]. | |
facdaa91 NG |
2227 | * |
2228 | * The scaling factor ensures that proactive compaction focuses on larger | |
2229 | * zones like ZONE_NORMAL, rather than smaller, specialized zones like | |
2230 | * ZONE_DMA32. For smaller zones, the score value remains close to zero, | |
2231 | * and thus never exceeds the high threshold for proactive compaction. | |
2232 | */ | |
40d7e203 | 2233 | static unsigned int fragmentation_score_zone_weighted(struct zone *zone) |
facdaa91 NG |
2234 | { |
2235 | unsigned long score; | |
2236 | ||
40d7e203 | 2237 | score = zone->present_pages * fragmentation_score_zone(zone); |
facdaa91 NG |
2238 | return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); |
2239 | } | |
2240 | ||
2241 | /* | |
2242 | * The per-node proactive (background) compaction process is started by its | |
2243 | * corresponding kcompactd thread when the node's fragmentation score | |
2244 | * exceeds the high threshold. The compaction process remains active till | |
2245 | * the node's score falls below the low threshold, or one of the back-off | |
2246 | * conditions is met. | |
2247 | */ | |
d34c0a75 | 2248 | static unsigned int fragmentation_score_node(pg_data_t *pgdat) |
facdaa91 | 2249 | { |
d34c0a75 | 2250 | unsigned int score = 0; |
facdaa91 NG |
2251 | int zoneid; |
2252 | ||
2253 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2254 | struct zone *zone; | |
2255 | ||
2256 | zone = &pgdat->node_zones[zoneid]; | |
9e552271 BW |
2257 | if (!populated_zone(zone)) |
2258 | continue; | |
40d7e203 | 2259 | score += fragmentation_score_zone_weighted(zone); |
facdaa91 NG |
2260 | } |
2261 | ||
2262 | return score; | |
2263 | } | |
2264 | ||
8fbb92bd | 2265 | static unsigned int fragmentation_score_wmark(bool low) |
facdaa91 | 2266 | { |
d34c0a75 | 2267 | unsigned int wmark_low; |
facdaa91 NG |
2268 | |
2269 | /* | |
f0953a1b IM |
2270 | * Cap the low watermark to avoid excessive compaction |
2271 | * activity in case a user sets the proactiveness tunable | |
facdaa91 NG |
2272 | * close to 100 (maximum). |
2273 | */ | |
d34c0a75 NG |
2274 | wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); |
2275 | return low ? wmark_low : min(wmark_low + 10, 100U); | |
facdaa91 NG |
2276 | } |
2277 | ||
2278 | static bool should_proactive_compact_node(pg_data_t *pgdat) | |
2279 | { | |
2280 | int wmark_high; | |
2281 | ||
2282 | if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) | |
2283 | return false; | |
2284 | ||
8fbb92bd | 2285 | wmark_high = fragmentation_score_wmark(false); |
facdaa91 NG |
2286 | return fragmentation_score_node(pgdat) > wmark_high; |
2287 | } | |
2288 | ||
40cacbcb | 2289 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 2290 | { |
8fb74b9f | 2291 | unsigned int order; |
d39773a0 | 2292 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 2293 | int ret; |
748446bb | 2294 | |
753341a4 | 2295 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 2296 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 2297 | /* Let the next compaction start anew. */ |
40cacbcb | 2298 | reset_cached_positions(cc->zone); |
55b7c4c9 | 2299 | |
62997027 MG |
2300 | /* |
2301 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 2302 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
2303 | * flag itself as the decision to be clear should be directly |
2304 | * based on an allocation request. | |
2305 | */ | |
accf6242 | 2306 | if (cc->direct_compaction) |
40cacbcb | 2307 | cc->zone->compact_blockskip_flush = true; |
62997027 | 2308 | |
c8f7de0b MH |
2309 | if (cc->whole_zone) |
2310 | return COMPACT_COMPLETE; | |
2311 | else | |
2312 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 2313 | } |
748446bb | 2314 | |
facdaa91 NG |
2315 | if (cc->proactive_compaction) { |
2316 | int score, wmark_low; | |
2317 | pg_data_t *pgdat; | |
2318 | ||
2319 | pgdat = cc->zone->zone_pgdat; | |
2320 | if (kswapd_is_running(pgdat)) | |
2321 | return COMPACT_PARTIAL_SKIPPED; | |
2322 | ||
2323 | score = fragmentation_score_zone(cc->zone); | |
8fbb92bd | 2324 | wmark_low = fragmentation_score_wmark(true); |
facdaa91 NG |
2325 | |
2326 | if (score > wmark_low) | |
2327 | ret = COMPACT_CONTINUE; | |
2328 | else | |
2329 | ret = COMPACT_SUCCESS; | |
2330 | ||
2331 | goto out; | |
2332 | } | |
2333 | ||
21c527a3 | 2334 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
2335 | return COMPACT_CONTINUE; |
2336 | ||
efe771c7 MG |
2337 | /* |
2338 | * Always finish scanning a pageblock to reduce the possibility of | |
2339 | * fallbacks in the future. This is particularly important when | |
2340 | * migration source is unmovable/reclaimable but it's not worth | |
2341 | * special casing. | |
2342 | */ | |
ee0913c4 | 2343 | if (!pageblock_aligned(cc->migrate_pfn)) |
efe771c7 | 2344 | return COMPACT_CONTINUE; |
baf6a9a1 | 2345 | |
56de7263 | 2346 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 2347 | ret = COMPACT_NO_SUITABLE_PAGE; |
fd377218 | 2348 | for (order = cc->order; order < NR_PAGE_ORDERS; order++) { |
40cacbcb | 2349 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 2350 | bool can_steal; |
8fb74b9f MG |
2351 | |
2352 | /* Job done if page is free of the right migratetype */ | |
b03641af | 2353 | if (!free_area_empty(area, migratetype)) |
cf378319 | 2354 | return COMPACT_SUCCESS; |
8fb74b9f | 2355 | |
2149cdae JK |
2356 | #ifdef CONFIG_CMA |
2357 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
2358 | if (migratetype == MIGRATE_MOVABLE && | |
b03641af | 2359 | !free_area_empty(area, MIGRATE_CMA)) |
cf378319 | 2360 | return COMPACT_SUCCESS; |
2149cdae JK |
2361 | #endif |
2362 | /* | |
2363 | * Job done if allocation would steal freepages from | |
2364 | * other migratetype buddy lists. | |
2365 | */ | |
2366 | if (find_suitable_fallback(area, order, migratetype, | |
fa599c44 | 2367 | true, &can_steal) != -1) |
baf6a9a1 | 2368 | /* |
fa599c44 ML |
2369 | * Movable pages are OK in any pageblock. If we are |
2370 | * stealing for a non-movable allocation, make sure | |
2371 | * we finish compacting the current pageblock first | |
2372 | * (which is assured by the above migrate_pfn align | |
2373 | * check) so it is as free as possible and we won't | |
2374 | * have to steal another one soon. | |
baf6a9a1 | 2375 | */ |
fa599c44 | 2376 | return COMPACT_SUCCESS; |
56de7263 MG |
2377 | } |
2378 | ||
facdaa91 | 2379 | out: |
cb2dcaf0 MG |
2380 | if (cc->contended || fatal_signal_pending(current)) |
2381 | ret = COMPACT_CONTENDED; | |
2382 | ||
2383 | return ret; | |
837d026d JK |
2384 | } |
2385 | ||
40cacbcb | 2386 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
2387 | { |
2388 | int ret; | |
2389 | ||
40cacbcb MG |
2390 | ret = __compact_finished(cc); |
2391 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
2392 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
2393 | ret = COMPACT_CONTINUE; | |
2394 | ||
2395 | return ret; | |
748446bb MG |
2396 | } |
2397 | ||
3cf04937 JW |
2398 | static bool __compaction_suitable(struct zone *zone, int order, |
2399 | int highest_zoneidx, | |
2400 | unsigned long wmark_target) | |
3e7d3449 | 2401 | { |
3e7d3449 | 2402 | unsigned long watermark; |
3e7d3449 | 2403 | /* |
9861a62c | 2404 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
2405 | * isolate free pages for migration targets. This means that the |
2406 | * watermark and alloc_flags have to match, or be more pessimistic than | |
2407 | * the check in __isolate_free_page(). We don't use the direct | |
2408 | * compactor's alloc_flags, as they are not relevant for freepage | |
97a225e6 JK |
2409 | * isolation. We however do use the direct compactor's highest_zoneidx |
2410 | * to skip over zones where lowmem reserves would prevent allocation | |
2411 | * even if compaction succeeds. | |
8348faf9 VB |
2412 | * For costly orders, we require low watermark instead of min for |
2413 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
2414 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
2415 | * suitable migration targets | |
3e7d3449 | 2416 | */ |
8348faf9 VB |
2417 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
2418 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
2419 | watermark += compact_gap(order); | |
3cf04937 JW |
2420 | return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx, |
2421 | ALLOC_CMA, wmark_target); | |
cc5c9f09 VB |
2422 | } |
2423 | ||
2b1a20c3 HS |
2424 | /* |
2425 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
2b1a20c3 | 2426 | */ |
3cf04937 | 2427 | bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx) |
cc5c9f09 | 2428 | { |
3cf04937 JW |
2429 | enum compact_result compact_result; |
2430 | bool suitable; | |
cc5c9f09 | 2431 | |
3cf04937 JW |
2432 | suitable = __compaction_suitable(zone, order, highest_zoneidx, |
2433 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
2434 | /* |
2435 | * fragmentation index determines if allocation failures are due to | |
2436 | * low memory or external fragmentation | |
2437 | * | |
ebff3980 VB |
2438 | * index of -1000 would imply allocations might succeed depending on |
2439 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
2440 | * index towards 0 implies failure is due to lack of memory |
2441 | * index towards 1000 implies failure is due to fragmentation | |
2442 | * | |
20311420 VB |
2443 | * Only compact if a failure would be due to fragmentation. Also |
2444 | * ignore fragindex for non-costly orders where the alternative to | |
2445 | * a successful reclaim/compaction is OOM. Fragindex and the | |
2446 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
2447 | * excessive compaction for costly orders, but it should not be at the | |
2448 | * expense of system stability. | |
3e7d3449 | 2449 | */ |
3cf04937 JW |
2450 | if (suitable) { |
2451 | compact_result = COMPACT_CONTINUE; | |
2452 | if (order > PAGE_ALLOC_COSTLY_ORDER) { | |
2453 | int fragindex = fragmentation_index(zone, order); | |
2454 | ||
2455 | if (fragindex >= 0 && | |
2456 | fragindex <= sysctl_extfrag_threshold) { | |
2457 | suitable = false; | |
2458 | compact_result = COMPACT_NOT_SUITABLE_ZONE; | |
2459 | } | |
2460 | } | |
2461 | } else { | |
2462 | compact_result = COMPACT_SKIPPED; | |
cc5c9f09 | 2463 | } |
837d026d | 2464 | |
3cf04937 | 2465 | trace_mm_compaction_suitable(zone, order, compact_result); |
837d026d | 2466 | |
3cf04937 | 2467 | return suitable; |
837d026d JK |
2468 | } |
2469 | ||
86a294a8 MH |
2470 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
2471 | int alloc_flags) | |
2472 | { | |
2473 | struct zone *zone; | |
2474 | struct zoneref *z; | |
2475 | ||
2476 | /* | |
2477 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
2478 | * retrying the reclaim. | |
2479 | */ | |
97a225e6 JK |
2480 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2481 | ac->highest_zoneidx, ac->nodemask) { | |
86a294a8 | 2482 | unsigned long available; |
86a294a8 MH |
2483 | |
2484 | /* | |
2485 | * Do not consider all the reclaimable memory because we do not | |
2486 | * want to trash just for a single high order allocation which | |
2487 | * is even not guaranteed to appear even if __compaction_suitable | |
2488 | * is happy about the watermark check. | |
2489 | */ | |
5a1c84b4 | 2490 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 | 2491 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
e8606320 | 2492 | if (__compaction_suitable(zone, order, ac->highest_zoneidx, |
3cf04937 | 2493 | available)) |
86a294a8 MH |
2494 | return true; |
2495 | } | |
2496 | ||
2497 | return false; | |
2498 | } | |
2499 | ||
e19a3f59 KS |
2500 | /* |
2501 | * Should we do compaction for target allocation order. | |
2502 | * Return COMPACT_SUCCESS if allocation for target order can be already | |
2503 | * satisfied | |
2504 | * Return COMPACT_SKIPPED if compaction for target order is likely to fail | |
2505 | * Return COMPACT_CONTINUE if compaction for target order should be ran | |
2506 | */ | |
2507 | static enum compact_result | |
2508 | compaction_suit_allocation_order(struct zone *zone, unsigned int order, | |
2509 | int highest_zoneidx, unsigned int alloc_flags) | |
2510 | { | |
2511 | unsigned long watermark; | |
2512 | ||
2513 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); | |
2514 | if (zone_watermark_ok(zone, order, watermark, highest_zoneidx, | |
2515 | alloc_flags)) | |
2516 | return COMPACT_SUCCESS; | |
2517 | ||
2518 | if (!compaction_suitable(zone, order, highest_zoneidx)) | |
2519 | return COMPACT_SKIPPED; | |
2520 | ||
2521 | return COMPACT_CONTINUE; | |
2522 | } | |
2523 | ||
5e1f0f09 MG |
2524 | static enum compact_result |
2525 | compact_zone(struct compact_control *cc, struct capture_control *capc) | |
748446bb | 2526 | { |
ea7ab982 | 2527 | enum compact_result ret; |
40cacbcb MG |
2528 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
2529 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 2530 | unsigned long last_migrated_pfn; |
e0b9daeb | 2531 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 2532 | bool update_cached; |
ab755bf4 | 2533 | unsigned int nr_succeeded = 0, nr_migratepages; |
733aea0b | 2534 | int order; |
748446bb | 2535 | |
a94b5252 YS |
2536 | /* |
2537 | * These counters track activities during zone compaction. Initialize | |
2538 | * them before compacting a new zone. | |
2539 | */ | |
2540 | cc->total_migrate_scanned = 0; | |
2541 | cc->total_free_scanned = 0; | |
2542 | cc->nr_migratepages = 0; | |
2543 | cc->nr_freepages = 0; | |
733aea0b ZY |
2544 | for (order = 0; order < NR_PAGE_ORDERS; order++) |
2545 | INIT_LIST_HEAD(&cc->freepages[order]); | |
a94b5252 YS |
2546 | INIT_LIST_HEAD(&cc->migratepages); |
2547 | ||
01c0bfe0 | 2548 | cc->migratetype = gfp_migratetype(cc->gfp_mask); |
e8606320 JW |
2549 | |
2550 | if (!is_via_compact_memory(cc->order)) { | |
e19a3f59 KS |
2551 | ret = compaction_suit_allocation_order(cc->zone, cc->order, |
2552 | cc->highest_zoneidx, | |
2553 | cc->alloc_flags); | |
2554 | if (ret != COMPACT_CONTINUE) | |
2555 | return ret; | |
e8606320 | 2556 | } |
c46649de | 2557 | |
d3132e4b VB |
2558 | /* |
2559 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 2560 | * is about to be retried after being deferred. |
d3132e4b | 2561 | */ |
40cacbcb MG |
2562 | if (compaction_restarting(cc->zone, cc->order)) |
2563 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 2564 | |
c89511ab MG |
2565 | /* |
2566 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
2567 | * information on where the scanners should start (unless we explicitly |
2568 | * want to compact the whole zone), but check that it is initialised | |
2569 | * by ensuring the values are within zone boundaries. | |
c89511ab | 2570 | */ |
70b44595 | 2571 | cc->fast_start_pfn = 0; |
06ed2998 | 2572 | if (cc->whole_zone) { |
c89511ab | 2573 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
2574 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
2575 | } else { | |
40cacbcb MG |
2576 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
2577 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
2578 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
2579 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 2580 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
2581 | } |
2582 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
2583 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
2584 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
2585 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 2586 | } |
c8f7de0b | 2587 | |
e332f741 | 2588 | if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) |
06ed2998 VB |
2589 | cc->whole_zone = true; |
2590 | } | |
c8f7de0b | 2591 | |
566e54e1 | 2592 | last_migrated_pfn = 0; |
748446bb | 2593 | |
8854c55f MG |
2594 | /* |
2595 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
2596 | * the basis that some migrations will fail in ASYNC mode. However, | |
2597 | * if the cached PFNs match and pageblocks are skipped due to having | |
2598 | * no isolation candidates, then the sync state does not matter. | |
2599 | * Until a pageblock with isolation candidates is found, keep the | |
2600 | * cached PFNs in sync to avoid revisiting the same blocks. | |
2601 | */ | |
2602 | update_cached = !sync && | |
2603 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
2604 | ||
abd4349f | 2605 | trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); |
0eb927c0 | 2606 | |
361a2a22 MK |
2607 | /* lru_add_drain_all could be expensive with involving other CPUs */ |
2608 | lru_add_drain(); | |
748446bb | 2609 | |
40cacbcb | 2610 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2611 | int err; |
19d3cf9d | 2612 | unsigned long iteration_start_pfn = cc->migrate_pfn; |
748446bb | 2613 | |
804d3121 | 2614 | /* |
48731c84 MG |
2615 | * Avoid multiple rescans of the same pageblock which can |
2616 | * happen if a page cannot be isolated (dirty/writeback in | |
2617 | * async mode) or if the migrated pages are being allocated | |
2618 | * before the pageblock is cleared. The first rescan will | |
2619 | * capture the entire pageblock for migration. If it fails, | |
2620 | * it'll be marked skip and scanning will proceed as normal. | |
804d3121 | 2621 | */ |
48731c84 | 2622 | cc->finish_pageblock = false; |
804d3121 | 2623 | if (pageblock_start_pfn(last_migrated_pfn) == |
19d3cf9d | 2624 | pageblock_start_pfn(iteration_start_pfn)) { |
48731c84 | 2625 | cc->finish_pageblock = true; |
804d3121 MG |
2626 | } |
2627 | ||
cfccd2e6 | 2628 | rescan: |
32aaf055 | 2629 | switch (isolate_migratepages(cc)) { |
f9e35b3b | 2630 | case ISOLATE_ABORT: |
2d1e1041 | 2631 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2632 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2633 | cc->nr_migratepages = 0; |
f9e35b3b MG |
2634 | goto out; |
2635 | case ISOLATE_NONE: | |
8854c55f MG |
2636 | if (update_cached) { |
2637 | cc->zone->compact_cached_migrate_pfn[1] = | |
2638 | cc->zone->compact_cached_migrate_pfn[0]; | |
2639 | } | |
2640 | ||
fdaf7f5c VB |
2641 | /* |
2642 | * We haven't isolated and migrated anything, but | |
2643 | * there might still be unflushed migrations from | |
2644 | * previous cc->order aligned block. | |
2645 | */ | |
2646 | goto check_drain; | |
f9e35b3b | 2647 | case ISOLATE_SUCCESS: |
8854c55f | 2648 | update_cached = false; |
7c0a84bd KS |
2649 | last_migrated_pfn = max(cc->zone->zone_start_pfn, |
2650 | pageblock_start_pfn(cc->migrate_pfn - 1)); | |
f9e35b3b | 2651 | } |
748446bb | 2652 | |
ab755bf4 BW |
2653 | /* |
2654 | * Record the number of pages to migrate since the | |
2655 | * compaction_alloc/free() will update cc->nr_migratepages | |
2656 | * properly. | |
2657 | */ | |
2658 | nr_migratepages = cc->nr_migratepages; | |
d53aea3d | 2659 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2660 | compaction_free, (unsigned long)cc, cc->mode, |
84b328aa | 2661 | MR_COMPACTION, &nr_succeeded); |
748446bb | 2662 | |
ab755bf4 | 2663 | trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded); |
748446bb | 2664 | |
f8c9301f VB |
2665 | /* All pages were either migrated or will be released */ |
2666 | cc->nr_migratepages = 0; | |
9d502c1c | 2667 | if (err) { |
5733c7d1 | 2668 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2669 | /* |
2670 | * migrate_pages() may return -ENOMEM when scanners meet | |
2671 | * and we want compact_finished() to detect it | |
2672 | */ | |
f2849aa0 | 2673 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2674 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2675 | goto out; |
2676 | } | |
fdd048e1 | 2677 | /* |
cfccd2e6 | 2678 | * If an ASYNC or SYNC_LIGHT fails to migrate a page |
c3750cc7 | 2679 | * within the pageblock_order-aligned block and |
9ecc5fc5 | 2680 | * fast_find_migrateblock may be used then scan the |
cfccd2e6 MG |
2681 | * remainder of the pageblock. This will mark the |
2682 | * pageblock "skip" to avoid rescanning in the near | |
2683 | * future. This will isolate more pages than necessary | |
2684 | * for the request but avoid loops due to | |
2685 | * fast_find_migrateblock revisiting blocks that were | |
2686 | * recently partially scanned. | |
fdd048e1 | 2687 | */ |
539aa041 | 2688 | if (!pageblock_aligned(cc->migrate_pfn) && |
9ecc5fc5 | 2689 | !cc->ignore_skip_hint && !cc->finish_pageblock && |
539aa041 | 2690 | (cc->mode < MIGRATE_SYNC)) { |
cfccd2e6 MG |
2691 | cc->finish_pageblock = true; |
2692 | ||
2693 | /* | |
2694 | * Draining pcplists does not help THP if | |
2695 | * any page failed to migrate. Even after | |
2696 | * drain, the pageblock will not be free. | |
2697 | */ | |
2698 | if (cc->order == COMPACTION_HPAGE_ORDER) | |
2699 | last_migrated_pfn = 0; | |
2700 | ||
2701 | goto rescan; | |
fdd048e1 | 2702 | } |
748446bb | 2703 | } |
fdaf7f5c | 2704 | |
16b3be40 MG |
2705 | /* Stop if a page has been captured */ |
2706 | if (capc && capc->page) { | |
2707 | ret = COMPACT_SUCCESS; | |
2708 | break; | |
2709 | } | |
2710 | ||
fdaf7f5c VB |
2711 | check_drain: |
2712 | /* | |
2713 | * Has the migration scanner moved away from the previous | |
2714 | * cc->order aligned block where we migrated from? If yes, | |
2715 | * flush the pages that were freed, so that they can merge and | |
2716 | * compact_finished() can detect immediately if allocation | |
2717 | * would succeed. | |
2718 | */ | |
566e54e1 | 2719 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c | 2720 | unsigned long current_block_start = |
06b6640a | 2721 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2722 | |
566e54e1 | 2723 | if (last_migrated_pfn < current_block_start) { |
b01b2141 | 2724 | lru_add_drain_cpu_zone(cc->zone); |
fdaf7f5c | 2725 | /* No more flushing until we migrate again */ |
566e54e1 | 2726 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2727 | } |
2728 | } | |
748446bb MG |
2729 | } |
2730 | ||
f9e35b3b | 2731 | out: |
6bace090 VB |
2732 | /* |
2733 | * Release free pages and update where the free scanner should restart, | |
2734 | * so we don't leave any returned pages behind in the next attempt. | |
2735 | */ | |
2736 | if (cc->nr_freepages > 0) { | |
733aea0b | 2737 | unsigned long free_pfn = release_free_list(cc->freepages); |
6bace090 VB |
2738 | |
2739 | cc->nr_freepages = 0; | |
2740 | VM_BUG_ON(free_pfn == 0); | |
2741 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2742 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2743 | /* |
2744 | * Only go back, not forward. The cached pfn might have been | |
2745 | * already reset to zone end in compact_finished() | |
2746 | */ | |
40cacbcb MG |
2747 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2748 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2749 | } |
748446bb | 2750 | |
7f354a54 DR |
2751 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2752 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2753 | ||
abd4349f | 2754 | trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); |
0eb927c0 | 2755 | |
753ec50d BW |
2756 | VM_BUG_ON(!list_empty(&cc->migratepages)); |
2757 | ||
748446bb MG |
2758 | return ret; |
2759 | } | |
76ab0f53 | 2760 | |
ea7ab982 | 2761 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2762 | gfp_t gfp_mask, enum compact_priority prio, |
97a225e6 | 2763 | unsigned int alloc_flags, int highest_zoneidx, |
5e1f0f09 | 2764 | struct page **capture) |
56de7263 | 2765 | { |
ea7ab982 | 2766 | enum compact_result ret; |
56de7263 | 2767 | struct compact_control cc = { |
56de7263 | 2768 | .order = order, |
dbe2d4e4 | 2769 | .search_order = order, |
6d7ce559 | 2770 | .gfp_mask = gfp_mask, |
56de7263 | 2771 | .zone = zone, |
a5508cd8 VB |
2772 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2773 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 | 2774 | .alloc_flags = alloc_flags, |
97a225e6 | 2775 | .highest_zoneidx = highest_zoneidx, |
accf6242 | 2776 | .direct_compaction = true, |
a8e025e5 | 2777 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2778 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2779 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 | 2780 | }; |
5e1f0f09 MG |
2781 | struct capture_control capc = { |
2782 | .cc = &cc, | |
2783 | .page = NULL, | |
2784 | }; | |
2785 | ||
b9e20f0d VB |
2786 | /* |
2787 | * Make sure the structs are really initialized before we expose the | |
2788 | * capture control, in case we are interrupted and the interrupt handler | |
2789 | * frees a page. | |
2790 | */ | |
2791 | barrier(); | |
2792 | WRITE_ONCE(current->capture_control, &capc); | |
56de7263 | 2793 | |
5e1f0f09 | 2794 | ret = compact_zone(&cc, &capc); |
e64c5237 | 2795 | |
b9e20f0d VB |
2796 | /* |
2797 | * Make sure we hide capture control first before we read the captured | |
2798 | * page pointer, otherwise an interrupt could free and capture a page | |
2799 | * and we would leak it. | |
2800 | */ | |
2801 | WRITE_ONCE(current->capture_control, NULL); | |
2802 | *capture = READ_ONCE(capc.page); | |
06dac2f4 CTR |
2803 | /* |
2804 | * Technically, it is also possible that compaction is skipped but | |
2805 | * the page is still captured out of luck(IRQ came and freed the page). | |
2806 | * Returning COMPACT_SUCCESS in such cases helps in properly accounting | |
2807 | * the COMPACT[STALL|FAIL] when compaction is skipped. | |
2808 | */ | |
2809 | if (*capture) | |
2810 | ret = COMPACT_SUCCESS; | |
5e1f0f09 | 2811 | |
e64c5237 | 2812 | return ret; |
56de7263 MG |
2813 | } |
2814 | ||
2815 | /** | |
2816 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2817 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2818 | * @order: The order of the current allocation |
2819 | * @alloc_flags: The allocation flags of the current allocation | |
2820 | * @ac: The context of current allocation | |
112d2d29 | 2821 | * @prio: Determines how hard direct compaction should try to succeed |
6467552c | 2822 | * @capture: Pointer to free page created by compaction will be stored here |
56de7263 MG |
2823 | * |
2824 | * This is the main entry point for direct page compaction. | |
2825 | */ | |
ea7ab982 | 2826 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2827 | unsigned int alloc_flags, const struct alloc_context *ac, |
5e1f0f09 | 2828 | enum compact_priority prio, struct page **capture) |
56de7263 | 2829 | { |
56de7263 MG |
2830 | struct zoneref *z; |
2831 | struct zone *zone; | |
1d4746d3 | 2832 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2833 | |
803de900 | 2834 | if (!gfp_compaction_allowed(gfp_mask)) |
53853e2d | 2835 | return COMPACT_SKIPPED; |
56de7263 | 2836 | |
a5508cd8 | 2837 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2838 | |
56de7263 | 2839 | /* Compact each zone in the list */ |
97a225e6 JK |
2840 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
2841 | ac->highest_zoneidx, ac->nodemask) { | |
ea7ab982 | 2842 | enum compact_result status; |
56de7263 | 2843 | |
a8e025e5 VB |
2844 | if (prio > MIN_COMPACT_PRIORITY |
2845 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2846 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2847 | continue; |
1d4746d3 | 2848 | } |
53853e2d | 2849 | |
a5508cd8 | 2850 | status = compact_zone_order(zone, order, gfp_mask, prio, |
97a225e6 | 2851 | alloc_flags, ac->highest_zoneidx, capture); |
56de7263 MG |
2852 | rc = max(status, rc); |
2853 | ||
7ceb009a VB |
2854 | /* The allocation should succeed, stop compacting */ |
2855 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2856 | /* |
2857 | * We think the allocation will succeed in this zone, | |
2858 | * but it is not certain, hence the false. The caller | |
2859 | * will repeat this with true if allocation indeed | |
2860 | * succeeds in this zone. | |
2861 | */ | |
2862 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2863 | |
c3486f53 | 2864 | break; |
1f9efdef VB |
2865 | } |
2866 | ||
a5508cd8 | 2867 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2868 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2869 | /* |
2870 | * We think that allocation won't succeed in this zone | |
2871 | * so we defer compaction there. If it ends up | |
2872 | * succeeding after all, it will be reset. | |
2873 | */ | |
2874 | defer_compaction(zone, order); | |
1f9efdef VB |
2875 | |
2876 | /* | |
2877 | * We might have stopped compacting due to need_resched() in | |
2878 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2879 | * case do not try further zones |
1f9efdef | 2880 | */ |
c3486f53 VB |
2881 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2882 | || fatal_signal_pending(current)) | |
2883 | break; | |
56de7263 MG |
2884 | } |
2885 | ||
2886 | return rc; | |
2887 | } | |
2888 | ||
facdaa91 | 2889 | /* |
3e40b3f4 KW |
2890 | * compact_node() - compact all zones within a node |
2891 | * @pgdat: The node page data | |
2892 | * @proactive: Whether the compaction is proactive | |
facdaa91 | 2893 | * |
3e40b3f4 KW |
2894 | * For proactive compaction, compact till each zone's fragmentation score |
2895 | * reaches within proactive compaction thresholds (as determined by the | |
2896 | * proactiveness tunable), it is possible that the function returns before | |
2897 | * reaching score targets due to various back-off conditions, such as, | |
2898 | * contention on per-node or per-zone locks. | |
facdaa91 | 2899 | */ |
f6f3f275 | 2900 | static int compact_node(pg_data_t *pgdat, bool proactive) |
facdaa91 NG |
2901 | { |
2902 | int zoneid; | |
2903 | struct zone *zone; | |
2904 | struct compact_control cc = { | |
2905 | .order = -1, | |
3e40b3f4 | 2906 | .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC, |
facdaa91 NG |
2907 | .ignore_skip_hint = true, |
2908 | .whole_zone = true, | |
2909 | .gfp_mask = GFP_KERNEL, | |
3e40b3f4 | 2910 | .proactive_compaction = proactive, |
facdaa91 NG |
2911 | }; |
2912 | ||
2913 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
2914 | zone = &pgdat->node_zones[zoneid]; | |
2915 | if (!populated_zone(zone)) | |
2916 | continue; | |
2917 | ||
f6f3f275 KW |
2918 | if (fatal_signal_pending(current)) |
2919 | return -EINTR; | |
2920 | ||
facdaa91 NG |
2921 | cc.zone = zone; |
2922 | ||
2923 | compact_zone(&cc, NULL); | |
2924 | ||
3e40b3f4 KW |
2925 | if (proactive) { |
2926 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, | |
2927 | cc.total_migrate_scanned); | |
2928 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2929 | cc.total_free_scanned); | |
2930 | } | |
facdaa91 | 2931 | } |
76ab0f53 | 2932 | |
f6f3f275 | 2933 | return 0; |
76ab0f53 MG |
2934 | } |
2935 | ||
3e40b3f4 | 2936 | /* Compact all zones of all nodes in the system */ |
f6f3f275 | 2937 | static int compact_nodes(void) |
76ab0f53 | 2938 | { |
f6f3f275 | 2939 | int ret, nid; |
76ab0f53 | 2940 | |
8575ec29 HD |
2941 | /* Flush pending updates to the LRU lists */ |
2942 | lru_add_drain_all(); | |
2943 | ||
f6f3f275 KW |
2944 | for_each_online_node(nid) { |
2945 | ret = compact_node(NODE_DATA(nid), false); | |
2946 | if (ret) | |
2947 | return ret; | |
2948 | } | |
2949 | ||
2950 | return 0; | |
76ab0f53 MG |
2951 | } |
2952 | ||
48fe8ab8 | 2953 | static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, |
65d759c8 CTR |
2954 | void *buffer, size_t *length, loff_t *ppos) |
2955 | { | |
2956 | int rc, nid; | |
2957 | ||
2958 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
2959 | if (rc) | |
2960 | return rc; | |
2961 | ||
2962 | if (write && sysctl_compaction_proactiveness) { | |
2963 | for_each_online_node(nid) { | |
2964 | pg_data_t *pgdat = NODE_DATA(nid); | |
2965 | ||
2966 | if (pgdat->proactive_compact_trigger) | |
2967 | continue; | |
2968 | ||
2969 | pgdat->proactive_compact_trigger = true; | |
8fff8b6f BW |
2970 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1, |
2971 | pgdat->nr_zones - 1); | |
65d759c8 CTR |
2972 | wake_up_interruptible(&pgdat->kcompactd_wait); |
2973 | } | |
2974 | } | |
2975 | ||
2976 | return 0; | |
2977 | } | |
2978 | ||
fec4eb2c YB |
2979 | /* |
2980 | * This is the entry point for compacting all nodes via | |
2981 | * /proc/sys/vm/compact_memory | |
2982 | */ | |
48fe8ab8 | 2983 | static int sysctl_compaction_handler(struct ctl_table *table, int write, |
32927393 | 2984 | void *buffer, size_t *length, loff_t *ppos) |
76ab0f53 | 2985 | { |
8b9167cd WY |
2986 | int ret; |
2987 | ||
2988 | ret = proc_dointvec(table, write, buffer, length, ppos); | |
2989 | if (ret) | |
2990 | return ret; | |
2991 | ||
2992 | if (sysctl_compact_memory != 1) | |
2993 | return -EINVAL; | |
2994 | ||
76ab0f53 | 2995 | if (write) |
f6f3f275 | 2996 | ret = compact_nodes(); |
76ab0f53 | 2997 | |
f6f3f275 | 2998 | return ret; |
76ab0f53 | 2999 | } |
ed4a6d7f MG |
3000 | |
3001 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
17adb230 Y |
3002 | static ssize_t compact_store(struct device *dev, |
3003 | struct device_attribute *attr, | |
3004 | const char *buf, size_t count) | |
ed4a6d7f | 3005 | { |
8575ec29 HD |
3006 | int nid = dev->id; |
3007 | ||
3008 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
3009 | /* Flush pending updates to the LRU lists */ | |
3010 | lru_add_drain_all(); | |
3011 | ||
3e40b3f4 | 3012 | compact_node(NODE_DATA(nid), false); |
8575ec29 | 3013 | } |
ed4a6d7f MG |
3014 | |
3015 | return count; | |
3016 | } | |
17adb230 | 3017 | static DEVICE_ATTR_WO(compact); |
ed4a6d7f MG |
3018 | |
3019 | int compaction_register_node(struct node *node) | |
3020 | { | |
10fbcf4c | 3021 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
3022 | } |
3023 | ||
3024 | void compaction_unregister_node(struct node *node) | |
3025 | { | |
f82024cb | 3026 | device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
3027 | } |
3028 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 3029 | |
698b1b30 VB |
3030 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
3031 | { | |
65d759c8 CTR |
3032 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || |
3033 | pgdat->proactive_compact_trigger; | |
698b1b30 VB |
3034 | } |
3035 | ||
3036 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
3037 | { | |
3038 | int zoneid; | |
3039 | struct zone *zone; | |
97a225e6 | 3040 | enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; |
e19a3f59 | 3041 | enum compact_result ret; |
698b1b30 | 3042 | |
97a225e6 | 3043 | for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { |
698b1b30 VB |
3044 | zone = &pgdat->node_zones[zoneid]; |
3045 | ||
3046 | if (!populated_zone(zone)) | |
3047 | continue; | |
3048 | ||
e19a3f59 KS |
3049 | ret = compaction_suit_allocation_order(zone, |
3050 | pgdat->kcompactd_max_order, | |
3051 | highest_zoneidx, ALLOC_WMARK_MIN); | |
3052 | if (ret == COMPACT_CONTINUE) | |
698b1b30 VB |
3053 | return true; |
3054 | } | |
3055 | ||
3056 | return false; | |
3057 | } | |
3058 | ||
3059 | static void kcompactd_do_work(pg_data_t *pgdat) | |
3060 | { | |
3061 | /* | |
3062 | * With no special task, compact all zones so that a page of requested | |
3063 | * order is allocatable. | |
3064 | */ | |
3065 | int zoneid; | |
3066 | struct zone *zone; | |
3067 | struct compact_control cc = { | |
3068 | .order = pgdat->kcompactd_max_order, | |
dbe2d4e4 | 3069 | .search_order = pgdat->kcompactd_max_order, |
97a225e6 | 3070 | .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, |
698b1b30 | 3071 | .mode = MIGRATE_SYNC_LIGHT, |
a0647dc9 | 3072 | .ignore_skip_hint = false, |
73e64c51 | 3073 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 3074 | }; |
e19a3f59 KS |
3075 | enum compact_result ret; |
3076 | ||
698b1b30 | 3077 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
97a225e6 | 3078 | cc.highest_zoneidx); |
7f354a54 | 3079 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 3080 | |
97a225e6 | 3081 | for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { |
698b1b30 VB |
3082 | int status; |
3083 | ||
3084 | zone = &pgdat->node_zones[zoneid]; | |
3085 | if (!populated_zone(zone)) | |
3086 | continue; | |
3087 | ||
3088 | if (compaction_deferred(zone, cc.order)) | |
3089 | continue; | |
3090 | ||
e19a3f59 KS |
3091 | ret = compaction_suit_allocation_order(zone, |
3092 | cc.order, zoneid, ALLOC_WMARK_MIN); | |
3093 | if (ret != COMPACT_CONTINUE) | |
e8606320 | 3094 | continue; |
f98a497e | 3095 | |
172400c6 VB |
3096 | if (kthread_should_stop()) |
3097 | return; | |
a94b5252 YS |
3098 | |
3099 | cc.zone = zone; | |
5e1f0f09 | 3100 | status = compact_zone(&cc, NULL); |
698b1b30 | 3101 | |
7ceb009a | 3102 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 3103 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 3104 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
3105 | /* |
3106 | * Buddy pages may become stranded on pcps that could | |
3107 | * otherwise coalesce on the zone's free area for | |
3108 | * order >= cc.order. This is ratelimited by the | |
3109 | * upcoming deferral. | |
3110 | */ | |
3111 | drain_all_pages(zone); | |
3112 | ||
698b1b30 VB |
3113 | /* |
3114 | * We use sync migration mode here, so we defer like | |
3115 | * sync direct compaction does. | |
3116 | */ | |
3117 | defer_compaction(zone, cc.order); | |
3118 | } | |
3119 | ||
7f354a54 DR |
3120 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
3121 | cc.total_migrate_scanned); | |
3122 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
3123 | cc.total_free_scanned); | |
698b1b30 VB |
3124 | } |
3125 | ||
3126 | /* | |
3127 | * Regardless of success, we are done until woken up next. But remember | |
97a225e6 JK |
3128 | * the requested order/highest_zoneidx in case it was higher/tighter |
3129 | * than our current ones | |
698b1b30 VB |
3130 | */ |
3131 | if (pgdat->kcompactd_max_order <= cc.order) | |
3132 | pgdat->kcompactd_max_order = 0; | |
97a225e6 JK |
3133 | if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) |
3134 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; | |
698b1b30 VB |
3135 | } |
3136 | ||
97a225e6 | 3137 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) |
698b1b30 VB |
3138 | { |
3139 | if (!order) | |
3140 | return; | |
3141 | ||
3142 | if (pgdat->kcompactd_max_order < order) | |
3143 | pgdat->kcompactd_max_order = order; | |
3144 | ||
97a225e6 JK |
3145 | if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) |
3146 | pgdat->kcompactd_highest_zoneidx = highest_zoneidx; | |
698b1b30 | 3147 | |
6818600f DB |
3148 | /* |
3149 | * Pairs with implicit barrier in wait_event_freezable() | |
3150 | * such that wakeups are not missed. | |
3151 | */ | |
3152 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
3153 | return; |
3154 | ||
3155 | if (!kcompactd_node_suitable(pgdat)) | |
3156 | return; | |
3157 | ||
3158 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
97a225e6 | 3159 | highest_zoneidx); |
698b1b30 VB |
3160 | wake_up_interruptible(&pgdat->kcompactd_wait); |
3161 | } | |
3162 | ||
3163 | /* | |
3164 | * The background compaction daemon, started as a kernel thread | |
3165 | * from the init process. | |
3166 | */ | |
3167 | static int kcompactd(void *p) | |
3168 | { | |
68d68ff6 | 3169 | pg_data_t *pgdat = (pg_data_t *)p; |
698b1b30 | 3170 | struct task_struct *tsk = current; |
e1e92bfa CTR |
3171 | long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); |
3172 | long timeout = default_timeout; | |
698b1b30 VB |
3173 | |
3174 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
3175 | ||
3176 | if (!cpumask_empty(cpumask)) | |
3177 | set_cpus_allowed_ptr(tsk, cpumask); | |
3178 | ||
3179 | set_freezable(); | |
3180 | ||
3181 | pgdat->kcompactd_max_order = 0; | |
97a225e6 | 3182 | pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; |
698b1b30 VB |
3183 | |
3184 | while (!kthread_should_stop()) { | |
eb414681 JW |
3185 | unsigned long pflags; |
3186 | ||
65d759c8 CTR |
3187 | /* |
3188 | * Avoid the unnecessary wakeup for proactive compaction | |
3189 | * when it is disabled. | |
3190 | */ | |
3191 | if (!sysctl_compaction_proactiveness) | |
3192 | timeout = MAX_SCHEDULE_TIMEOUT; | |
698b1b30 | 3193 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
facdaa91 | 3194 | if (wait_event_freezable_timeout(pgdat->kcompactd_wait, |
65d759c8 CTR |
3195 | kcompactd_work_requested(pgdat), timeout) && |
3196 | !pgdat->proactive_compact_trigger) { | |
facdaa91 NG |
3197 | |
3198 | psi_memstall_enter(&pflags); | |
3199 | kcompactd_do_work(pgdat); | |
3200 | psi_memstall_leave(&pflags); | |
e1e92bfa CTR |
3201 | /* |
3202 | * Reset the timeout value. The defer timeout from | |
3203 | * proactive compaction is lost here but that is fine | |
3204 | * as the condition of the zone changing substantionally | |
3205 | * then carrying on with the previous defer interval is | |
3206 | * not useful. | |
3207 | */ | |
3208 | timeout = default_timeout; | |
facdaa91 NG |
3209 | continue; |
3210 | } | |
698b1b30 | 3211 | |
e1e92bfa CTR |
3212 | /* |
3213 | * Start the proactive work with default timeout. Based | |
3214 | * on the fragmentation score, this timeout is updated. | |
3215 | */ | |
3216 | timeout = default_timeout; | |
facdaa91 NG |
3217 | if (should_proactive_compact_node(pgdat)) { |
3218 | unsigned int prev_score, score; | |
3219 | ||
facdaa91 | 3220 | prev_score = fragmentation_score_node(pgdat); |
3e40b3f4 | 3221 | compact_node(pgdat, true); |
facdaa91 NG |
3222 | score = fragmentation_score_node(pgdat); |
3223 | /* | |
3224 | * Defer proactive compaction if the fragmentation | |
3225 | * score did not go down i.e. no progress made. | |
3226 | */ | |
e1e92bfa CTR |
3227 | if (unlikely(score >= prev_score)) |
3228 | timeout = | |
3229 | default_timeout << COMPACT_MAX_DEFER_SHIFT; | |
facdaa91 | 3230 | } |
65d759c8 CTR |
3231 | if (unlikely(pgdat->proactive_compact_trigger)) |
3232 | pgdat->proactive_compact_trigger = false; | |
698b1b30 VB |
3233 | } |
3234 | ||
3235 | return 0; | |
3236 | } | |
3237 | ||
3238 | /* | |
3239 | * This kcompactd start function will be called by init and node-hot-add. | |
3240 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
3241 | */ | |
833dfc00 | 3242 | void __meminit kcompactd_run(int nid) |
698b1b30 VB |
3243 | { |
3244 | pg_data_t *pgdat = NODE_DATA(nid); | |
698b1b30 VB |
3245 | |
3246 | if (pgdat->kcompactd) | |
024c61ea | 3247 | return; |
698b1b30 VB |
3248 | |
3249 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
3250 | if (IS_ERR(pgdat->kcompactd)) { | |
3251 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
698b1b30 VB |
3252 | pgdat->kcompactd = NULL; |
3253 | } | |
698b1b30 VB |
3254 | } |
3255 | ||
3256 | /* | |
3257 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
e8da368a | 3258 | * be holding mem_hotplug_begin/done(). |
698b1b30 | 3259 | */ |
833dfc00 | 3260 | void __meminit kcompactd_stop(int nid) |
698b1b30 VB |
3261 | { |
3262 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
3263 | ||
3264 | if (kcompactd) { | |
3265 | kthread_stop(kcompactd); | |
3266 | NODE_DATA(nid)->kcompactd = NULL; | |
3267 | } | |
3268 | } | |
3269 | ||
3270 | /* | |
3271 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
3272 | * not required for correctness. So if the last cpu in a node goes | |
3273 | * away, we get changed to run anywhere: as the first one comes back, | |
3274 | * restore their cpu bindings. | |
3275 | */ | |
e46b1db2 | 3276 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
3277 | { |
3278 | int nid; | |
3279 | ||
e46b1db2 AMG |
3280 | for_each_node_state(nid, N_MEMORY) { |
3281 | pg_data_t *pgdat = NODE_DATA(nid); | |
3282 | const struct cpumask *mask; | |
698b1b30 | 3283 | |
e46b1db2 | 3284 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 3285 | |
e46b1db2 AMG |
3286 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3287 | /* One of our CPUs online: restore mask */ | |
3109de30 ML |
3288 | if (pgdat->kcompactd) |
3289 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 3290 | } |
e46b1db2 | 3291 | return 0; |
698b1b30 VB |
3292 | } |
3293 | ||
48fe8ab8 MC |
3294 | static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table, |
3295 | int write, void *buffer, size_t *lenp, loff_t *ppos) | |
3296 | { | |
3297 | int ret, old; | |
3298 | ||
3299 | if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write) | |
3300 | return proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
3301 | ||
3302 | old = *(int *)table->data; | |
3303 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
3304 | if (ret) | |
3305 | return ret; | |
3306 | if (old != *(int *)table->data) | |
3307 | pr_warn_once("sysctl attribute %s changed by %s[%d]\n", | |
3308 | table->procname, current->comm, | |
3309 | task_pid_nr(current)); | |
3310 | return ret; | |
3311 | } | |
3312 | ||
48fe8ab8 MC |
3313 | static struct ctl_table vm_compaction[] = { |
3314 | { | |
3315 | .procname = "compact_memory", | |
8b9167cd | 3316 | .data = &sysctl_compact_memory, |
48fe8ab8 MC |
3317 | .maxlen = sizeof(int), |
3318 | .mode = 0200, | |
3319 | .proc_handler = sysctl_compaction_handler, | |
3320 | }, | |
3321 | { | |
3322 | .procname = "compaction_proactiveness", | |
3323 | .data = &sysctl_compaction_proactiveness, | |
3324 | .maxlen = sizeof(sysctl_compaction_proactiveness), | |
3325 | .mode = 0644, | |
3326 | .proc_handler = compaction_proactiveness_sysctl_handler, | |
3327 | .extra1 = SYSCTL_ZERO, | |
3328 | .extra2 = SYSCTL_ONE_HUNDRED, | |
3329 | }, | |
3330 | { | |
3331 | .procname = "extfrag_threshold", | |
3332 | .data = &sysctl_extfrag_threshold, | |
3333 | .maxlen = sizeof(int), | |
3334 | .mode = 0644, | |
3335 | .proc_handler = proc_dointvec_minmax, | |
3336 | .extra1 = SYSCTL_ZERO, | |
3337 | .extra2 = SYSCTL_ONE_THOUSAND, | |
3338 | }, | |
3339 | { | |
3340 | .procname = "compact_unevictable_allowed", | |
3341 | .data = &sysctl_compact_unevictable_allowed, | |
3342 | .maxlen = sizeof(int), | |
3343 | .mode = 0644, | |
3344 | .proc_handler = proc_dointvec_minmax_warn_RT_change, | |
3345 | .extra1 = SYSCTL_ZERO, | |
3346 | .extra2 = SYSCTL_ONE, | |
3347 | }, | |
3348 | { } | |
3349 | }; | |
48fe8ab8 | 3350 | |
698b1b30 VB |
3351 | static int __init kcompactd_init(void) |
3352 | { | |
3353 | int nid; | |
e46b1db2 AMG |
3354 | int ret; |
3355 | ||
3356 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
3357 | "mm/compaction:online", | |
3358 | kcompactd_cpu_online, NULL); | |
3359 | if (ret < 0) { | |
3360 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
3361 | return ret; | |
3362 | } | |
698b1b30 VB |
3363 | |
3364 | for_each_node_state(nid, N_MEMORY) | |
3365 | kcompactd_run(nid); | |
48fe8ab8 | 3366 | register_sysctl_init("vm", vm_compaction); |
698b1b30 VB |
3367 | return 0; |
3368 | } | |
3369 | subsys_initcall(kcompactd_init) | |
3370 | ||
ff9543fd | 3371 | #endif /* CONFIG_COMPACTION */ |