]>
Commit | Line | Data |
---|---|---|
748446bb MG |
1 | /* |
2 | * linux/mm/compaction.c | |
3 | * | |
4 | * Memory compaction for the reduction of external fragmentation. Note that | |
5 | * this heavily depends upon page migration to do all the real heavy | |
6 | * lifting | |
7 | * | |
8 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
9 | */ | |
10 | #include <linux/swap.h> | |
11 | #include <linux/migrate.h> | |
12 | #include <linux/compaction.h> | |
13 | #include <linux/mm_inline.h> | |
14 | #include <linux/backing-dev.h> | |
76ab0f53 | 15 | #include <linux/sysctl.h> |
ed4a6d7f | 16 | #include <linux/sysfs.h> |
bf6bddf1 | 17 | #include <linux/balloon_compaction.h> |
194159fb | 18 | #include <linux/page-isolation.h> |
b8c73fc2 | 19 | #include <linux/kasan.h> |
748446bb MG |
20 | #include "internal.h" |
21 | ||
010fc29a MK |
22 | #ifdef CONFIG_COMPACTION |
23 | static inline void count_compact_event(enum vm_event_item item) | |
24 | { | |
25 | count_vm_event(item); | |
26 | } | |
27 | ||
28 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
29 | { | |
30 | count_vm_events(item, delta); | |
31 | } | |
32 | #else | |
33 | #define count_compact_event(item) do { } while (0) | |
34 | #define count_compact_events(item, delta) do { } while (0) | |
35 | #endif | |
36 | ||
ff9543fd MN |
37 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
38 | ||
b7aba698 MG |
39 | #define CREATE_TRACE_POINTS |
40 | #include <trace/events/compaction.h> | |
41 | ||
748446bb MG |
42 | static unsigned long release_freepages(struct list_head *freelist) |
43 | { | |
44 | struct page *page, *next; | |
6bace090 | 45 | unsigned long high_pfn = 0; |
748446bb MG |
46 | |
47 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 48 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
49 | list_del(&page->lru); |
50 | __free_page(page); | |
6bace090 VB |
51 | if (pfn > high_pfn) |
52 | high_pfn = pfn; | |
748446bb MG |
53 | } |
54 | ||
6bace090 | 55 | return high_pfn; |
748446bb MG |
56 | } |
57 | ||
ff9543fd MN |
58 | static void map_pages(struct list_head *list) |
59 | { | |
60 | struct page *page; | |
61 | ||
62 | list_for_each_entry(page, list, lru) { | |
63 | arch_alloc_page(page, 0); | |
64 | kernel_map_pages(page, 1, 1); | |
b8c73fc2 | 65 | kasan_alloc_pages(page, 0); |
ff9543fd MN |
66 | } |
67 | } | |
68 | ||
47118af0 MN |
69 | static inline bool migrate_async_suitable(int migratetype) |
70 | { | |
71 | return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE; | |
72 | } | |
73 | ||
7d49d886 VB |
74 | /* |
75 | * Check that the whole (or subset of) a pageblock given by the interval of | |
76 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it | |
77 | * with the migration of free compaction scanner. The scanners then need to | |
78 | * use only pfn_valid_within() check for arches that allow holes within | |
79 | * pageblocks. | |
80 | * | |
81 | * Return struct page pointer of start_pfn, or NULL if checks were not passed. | |
82 | * | |
83 | * It's possible on some configurations to have a setup like node0 node1 node0 | |
84 | * i.e. it's possible that all pages within a zones range of pages do not | |
85 | * belong to a single zone. We assume that a border between node0 and node1 | |
86 | * can occur within a single pageblock, but not a node0 node1 node0 | |
87 | * interleaving within a single pageblock. It is therefore sufficient to check | |
88 | * the first and last page of a pageblock and avoid checking each individual | |
89 | * page in a pageblock. | |
90 | */ | |
91 | static struct page *pageblock_pfn_to_page(unsigned long start_pfn, | |
92 | unsigned long end_pfn, struct zone *zone) | |
93 | { | |
94 | struct page *start_page; | |
95 | struct page *end_page; | |
96 | ||
97 | /* end_pfn is one past the range we are checking */ | |
98 | end_pfn--; | |
99 | ||
100 | if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn)) | |
101 | return NULL; | |
102 | ||
103 | start_page = pfn_to_page(start_pfn); | |
104 | ||
105 | if (page_zone(start_page) != zone) | |
106 | return NULL; | |
107 | ||
108 | end_page = pfn_to_page(end_pfn); | |
109 | ||
110 | /* This gives a shorter code than deriving page_zone(end_page) */ | |
111 | if (page_zone_id(start_page) != page_zone_id(end_page)) | |
112 | return NULL; | |
113 | ||
114 | return start_page; | |
115 | } | |
116 | ||
bb13ffeb | 117 | #ifdef CONFIG_COMPACTION |
24e2716f JK |
118 | |
119 | /* Do not skip compaction more than 64 times */ | |
120 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
121 | ||
122 | /* | |
123 | * Compaction is deferred when compaction fails to result in a page | |
124 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
125 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
126 | */ | |
127 | void defer_compaction(struct zone *zone, int order) | |
128 | { | |
129 | zone->compact_considered = 0; | |
130 | zone->compact_defer_shift++; | |
131 | ||
132 | if (order < zone->compact_order_failed) | |
133 | zone->compact_order_failed = order; | |
134 | ||
135 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
136 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
137 | ||
138 | trace_mm_compaction_defer_compaction(zone, order); | |
139 | } | |
140 | ||
141 | /* Returns true if compaction should be skipped this time */ | |
142 | bool compaction_deferred(struct zone *zone, int order) | |
143 | { | |
144 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
145 | ||
146 | if (order < zone->compact_order_failed) | |
147 | return false; | |
148 | ||
149 | /* Avoid possible overflow */ | |
150 | if (++zone->compact_considered > defer_limit) | |
151 | zone->compact_considered = defer_limit; | |
152 | ||
153 | if (zone->compact_considered >= defer_limit) | |
154 | return false; | |
155 | ||
156 | trace_mm_compaction_deferred(zone, order); | |
157 | ||
158 | return true; | |
159 | } | |
160 | ||
161 | /* | |
162 | * Update defer tracking counters after successful compaction of given order, | |
163 | * which means an allocation either succeeded (alloc_success == true) or is | |
164 | * expected to succeed. | |
165 | */ | |
166 | void compaction_defer_reset(struct zone *zone, int order, | |
167 | bool alloc_success) | |
168 | { | |
169 | if (alloc_success) { | |
170 | zone->compact_considered = 0; | |
171 | zone->compact_defer_shift = 0; | |
172 | } | |
173 | if (order >= zone->compact_order_failed) | |
174 | zone->compact_order_failed = order + 1; | |
175 | ||
176 | trace_mm_compaction_defer_reset(zone, order); | |
177 | } | |
178 | ||
179 | /* Returns true if restarting compaction after many failures */ | |
180 | bool compaction_restarting(struct zone *zone, int order) | |
181 | { | |
182 | if (order < zone->compact_order_failed) | |
183 | return false; | |
184 | ||
185 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
186 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
187 | } | |
188 | ||
bb13ffeb MG |
189 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
190 | static inline bool isolation_suitable(struct compact_control *cc, | |
191 | struct page *page) | |
192 | { | |
193 | if (cc->ignore_skip_hint) | |
194 | return true; | |
195 | ||
196 | return !get_pageblock_skip(page); | |
197 | } | |
198 | ||
02333641 VB |
199 | static void reset_cached_positions(struct zone *zone) |
200 | { | |
201 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
202 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
8b315860 JK |
203 | zone->compact_cached_free_pfn = |
204 | round_down(zone_end_pfn(zone) - 1, pageblock_nr_pages); | |
02333641 VB |
205 | } |
206 | ||
bb13ffeb MG |
207 | /* |
208 | * This function is called to clear all cached information on pageblocks that | |
209 | * should be skipped for page isolation when the migrate and free page scanner | |
210 | * meet. | |
211 | */ | |
62997027 | 212 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
213 | { |
214 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 215 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
216 | unsigned long pfn; |
217 | ||
62997027 | 218 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
219 | |
220 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
221 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
222 | struct page *page; | |
223 | ||
224 | cond_resched(); | |
225 | ||
226 | if (!pfn_valid(pfn)) | |
227 | continue; | |
228 | ||
229 | page = pfn_to_page(pfn); | |
230 | if (zone != page_zone(page)) | |
231 | continue; | |
232 | ||
233 | clear_pageblock_skip(page); | |
234 | } | |
02333641 VB |
235 | |
236 | reset_cached_positions(zone); | |
bb13ffeb MG |
237 | } |
238 | ||
62997027 MG |
239 | void reset_isolation_suitable(pg_data_t *pgdat) |
240 | { | |
241 | int zoneid; | |
242 | ||
243 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
244 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
245 | if (!populated_zone(zone)) | |
246 | continue; | |
247 | ||
248 | /* Only flush if a full compaction finished recently */ | |
249 | if (zone->compact_blockskip_flush) | |
250 | __reset_isolation_suitable(zone); | |
251 | } | |
252 | } | |
253 | ||
bb13ffeb MG |
254 | /* |
255 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 256 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 257 | */ |
c89511ab MG |
258 | static void update_pageblock_skip(struct compact_control *cc, |
259 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 260 | bool migrate_scanner) |
bb13ffeb | 261 | { |
c89511ab | 262 | struct zone *zone = cc->zone; |
35979ef3 | 263 | unsigned long pfn; |
6815bf3f JK |
264 | |
265 | if (cc->ignore_skip_hint) | |
266 | return; | |
267 | ||
bb13ffeb MG |
268 | if (!page) |
269 | return; | |
270 | ||
35979ef3 DR |
271 | if (nr_isolated) |
272 | return; | |
273 | ||
edc2ca61 | 274 | set_pageblock_skip(page); |
c89511ab | 275 | |
35979ef3 DR |
276 | pfn = page_to_pfn(page); |
277 | ||
278 | /* Update where async and sync compaction should restart */ | |
279 | if (migrate_scanner) { | |
35979ef3 DR |
280 | if (pfn > zone->compact_cached_migrate_pfn[0]) |
281 | zone->compact_cached_migrate_pfn[0] = pfn; | |
e0b9daeb DR |
282 | if (cc->mode != MIGRATE_ASYNC && |
283 | pfn > zone->compact_cached_migrate_pfn[1]) | |
35979ef3 DR |
284 | zone->compact_cached_migrate_pfn[1] = pfn; |
285 | } else { | |
35979ef3 DR |
286 | if (pfn < zone->compact_cached_free_pfn) |
287 | zone->compact_cached_free_pfn = pfn; | |
c89511ab | 288 | } |
bb13ffeb MG |
289 | } |
290 | #else | |
291 | static inline bool isolation_suitable(struct compact_control *cc, | |
292 | struct page *page) | |
293 | { | |
294 | return true; | |
295 | } | |
296 | ||
c89511ab MG |
297 | static void update_pageblock_skip(struct compact_control *cc, |
298 | struct page *page, unsigned long nr_isolated, | |
edc2ca61 | 299 | bool migrate_scanner) |
bb13ffeb MG |
300 | { |
301 | } | |
302 | #endif /* CONFIG_COMPACTION */ | |
303 | ||
8b44d279 VB |
304 | /* |
305 | * Compaction requires the taking of some coarse locks that are potentially | |
306 | * very heavily contended. For async compaction, back out if the lock cannot | |
307 | * be taken immediately. For sync compaction, spin on the lock if needed. | |
308 | * | |
309 | * Returns true if the lock is held | |
310 | * Returns false if the lock is not held and compaction should abort | |
311 | */ | |
312 | static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags, | |
313 | struct compact_control *cc) | |
2a1402aa | 314 | { |
8b44d279 VB |
315 | if (cc->mode == MIGRATE_ASYNC) { |
316 | if (!spin_trylock_irqsave(lock, *flags)) { | |
317 | cc->contended = COMPACT_CONTENDED_LOCK; | |
318 | return false; | |
319 | } | |
320 | } else { | |
321 | spin_lock_irqsave(lock, *flags); | |
322 | } | |
1f9efdef | 323 | |
8b44d279 | 324 | return true; |
2a1402aa MG |
325 | } |
326 | ||
c67fe375 MG |
327 | /* |
328 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
329 | * very heavily contended. The lock should be periodically unlocked to avoid |
330 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
331 | * the lock. It might also be that allowing the IRQs will result in | |
332 | * need_resched() becoming true. If scheduling is needed, async compaction | |
333 | * aborts. Sync compaction schedules. | |
334 | * Either compaction type will also abort if a fatal signal is pending. | |
335 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 336 | * |
8b44d279 VB |
337 | * Returns true if compaction should abort due to fatal signal pending, or |
338 | * async compaction due to need_resched() | |
339 | * Returns false when compaction can continue (sync compaction might have | |
340 | * scheduled) | |
c67fe375 | 341 | */ |
8b44d279 VB |
342 | static bool compact_unlock_should_abort(spinlock_t *lock, |
343 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 344 | { |
8b44d279 VB |
345 | if (*locked) { |
346 | spin_unlock_irqrestore(lock, flags); | |
347 | *locked = false; | |
348 | } | |
1f9efdef | 349 | |
8b44d279 VB |
350 | if (fatal_signal_pending(current)) { |
351 | cc->contended = COMPACT_CONTENDED_SCHED; | |
352 | return true; | |
353 | } | |
c67fe375 | 354 | |
8b44d279 | 355 | if (need_resched()) { |
e0b9daeb | 356 | if (cc->mode == MIGRATE_ASYNC) { |
8b44d279 VB |
357 | cc->contended = COMPACT_CONTENDED_SCHED; |
358 | return true; | |
c67fe375 | 359 | } |
c67fe375 | 360 | cond_resched(); |
c67fe375 MG |
361 | } |
362 | ||
8b44d279 | 363 | return false; |
c67fe375 MG |
364 | } |
365 | ||
be976572 VB |
366 | /* |
367 | * Aside from avoiding lock contention, compaction also periodically checks | |
368 | * need_resched() and either schedules in sync compaction or aborts async | |
8b44d279 | 369 | * compaction. This is similar to what compact_unlock_should_abort() does, but |
be976572 VB |
370 | * is used where no lock is concerned. |
371 | * | |
372 | * Returns false when no scheduling was needed, or sync compaction scheduled. | |
373 | * Returns true when async compaction should abort. | |
374 | */ | |
375 | static inline bool compact_should_abort(struct compact_control *cc) | |
376 | { | |
377 | /* async compaction aborts if contended */ | |
378 | if (need_resched()) { | |
379 | if (cc->mode == MIGRATE_ASYNC) { | |
1f9efdef | 380 | cc->contended = COMPACT_CONTENDED_SCHED; |
be976572 VB |
381 | return true; |
382 | } | |
383 | ||
384 | cond_resched(); | |
385 | } | |
386 | ||
387 | return false; | |
388 | } | |
389 | ||
85aa125f | 390 | /* |
9e4be470 JM |
391 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
392 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
393 | * (even though it may still end up isolating some pages). | |
85aa125f | 394 | */ |
f40d1e42 | 395 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 396 | unsigned long *start_pfn, |
85aa125f MN |
397 | unsigned long end_pfn, |
398 | struct list_head *freelist, | |
399 | bool strict) | |
748446bb | 400 | { |
b7aba698 | 401 | int nr_scanned = 0, total_isolated = 0; |
bb13ffeb | 402 | struct page *cursor, *valid_page = NULL; |
b8b2d825 | 403 | unsigned long flags = 0; |
f40d1e42 | 404 | bool locked = false; |
e14c720e | 405 | unsigned long blockpfn = *start_pfn; |
748446bb | 406 | |
748446bb MG |
407 | cursor = pfn_to_page(blockpfn); |
408 | ||
f40d1e42 | 409 | /* Isolate free pages. */ |
748446bb MG |
410 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
411 | int isolated, i; | |
412 | struct page *page = cursor; | |
413 | ||
8b44d279 VB |
414 | /* |
415 | * Periodically drop the lock (if held) regardless of its | |
416 | * contention, to give chance to IRQs. Abort if fatal signal | |
417 | * pending or async compaction detects need_resched() | |
418 | */ | |
419 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
420 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
421 | &locked, cc)) | |
422 | break; | |
423 | ||
b7aba698 | 424 | nr_scanned++; |
f40d1e42 | 425 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
426 | goto isolate_fail; |
427 | ||
bb13ffeb MG |
428 | if (!valid_page) |
429 | valid_page = page; | |
9fcd6d2e VB |
430 | |
431 | /* | |
432 | * For compound pages such as THP and hugetlbfs, we can save | |
433 | * potentially a lot of iterations if we skip them at once. | |
434 | * The check is racy, but we can consider only valid values | |
435 | * and the only danger is skipping too much. | |
436 | */ | |
437 | if (PageCompound(page)) { | |
438 | unsigned int comp_order = compound_order(page); | |
439 | ||
440 | if (likely(comp_order < MAX_ORDER)) { | |
441 | blockpfn += (1UL << comp_order) - 1; | |
442 | cursor += (1UL << comp_order) - 1; | |
443 | } | |
444 | ||
445 | goto isolate_fail; | |
446 | } | |
447 | ||
f40d1e42 | 448 | if (!PageBuddy(page)) |
2af120bc | 449 | goto isolate_fail; |
f40d1e42 MG |
450 | |
451 | /* | |
69b7189f VB |
452 | * If we already hold the lock, we can skip some rechecking. |
453 | * Note that if we hold the lock now, checked_pageblock was | |
454 | * already set in some previous iteration (or strict is true), | |
455 | * so it is correct to skip the suitable migration target | |
456 | * recheck as well. | |
f40d1e42 | 457 | */ |
69b7189f VB |
458 | if (!locked) { |
459 | /* | |
460 | * The zone lock must be held to isolate freepages. | |
461 | * Unfortunately this is a very coarse lock and can be | |
462 | * heavily contended if there are parallel allocations | |
463 | * or parallel compactions. For async compaction do not | |
464 | * spin on the lock and we acquire the lock as late as | |
465 | * possible. | |
466 | */ | |
8b44d279 VB |
467 | locked = compact_trylock_irqsave(&cc->zone->lock, |
468 | &flags, cc); | |
69b7189f VB |
469 | if (!locked) |
470 | break; | |
f40d1e42 | 471 | |
69b7189f VB |
472 | /* Recheck this is a buddy page under lock */ |
473 | if (!PageBuddy(page)) | |
474 | goto isolate_fail; | |
475 | } | |
748446bb MG |
476 | |
477 | /* Found a free page, break it into order-0 pages */ | |
478 | isolated = split_free_page(page); | |
a6864c9b DR |
479 | if (!isolated) |
480 | break; | |
481 | ||
748446bb | 482 | total_isolated += isolated; |
a6864c9b | 483 | cc->nr_freepages += isolated; |
748446bb MG |
484 | for (i = 0; i < isolated; i++) { |
485 | list_add(&page->lru, freelist); | |
486 | page++; | |
487 | } | |
a6864c9b DR |
488 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
489 | blockpfn += isolated; | |
490 | break; | |
748446bb | 491 | } |
a6864c9b DR |
492 | /* Advance to the end of split page */ |
493 | blockpfn += isolated - 1; | |
494 | cursor += isolated - 1; | |
495 | continue; | |
2af120bc LA |
496 | |
497 | isolate_fail: | |
498 | if (strict) | |
499 | break; | |
500 | else | |
501 | continue; | |
502 | ||
748446bb MG |
503 | } |
504 | ||
a6864c9b DR |
505 | if (locked) |
506 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
507 | ||
9fcd6d2e VB |
508 | /* |
509 | * There is a tiny chance that we have read bogus compound_order(), | |
510 | * so be careful to not go outside of the pageblock. | |
511 | */ | |
512 | if (unlikely(blockpfn > end_pfn)) | |
513 | blockpfn = end_pfn; | |
514 | ||
e34d85f0 JK |
515 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
516 | nr_scanned, total_isolated); | |
517 | ||
e14c720e VB |
518 | /* Record how far we have got within the block */ |
519 | *start_pfn = blockpfn; | |
520 | ||
f40d1e42 MG |
521 | /* |
522 | * If strict isolation is requested by CMA then check that all the | |
523 | * pages requested were isolated. If there were any failures, 0 is | |
524 | * returned and CMA will fail. | |
525 | */ | |
2af120bc | 526 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
527 | total_isolated = 0; |
528 | ||
bb13ffeb MG |
529 | /* Update the pageblock-skip if the whole pageblock was scanned */ |
530 | if (blockpfn == end_pfn) | |
edc2ca61 | 531 | update_pageblock_skip(cc, valid_page, total_isolated, false); |
bb13ffeb | 532 | |
010fc29a | 533 | count_compact_events(COMPACTFREE_SCANNED, nr_scanned); |
397487db | 534 | if (total_isolated) |
010fc29a | 535 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
536 | return total_isolated; |
537 | } | |
538 | ||
85aa125f MN |
539 | /** |
540 | * isolate_freepages_range() - isolate free pages. | |
541 | * @start_pfn: The first PFN to start isolating. | |
542 | * @end_pfn: The one-past-last PFN. | |
543 | * | |
544 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
545 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
546 | * undo its actions and return zero. | |
547 | * | |
548 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
549 | * (which may be greater then end_pfn if end fell in a middle of | |
550 | * a free page). | |
551 | */ | |
ff9543fd | 552 | unsigned long |
bb13ffeb MG |
553 | isolate_freepages_range(struct compact_control *cc, |
554 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 555 | { |
5a115a5a | 556 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
557 | LIST_HEAD(freelist); |
558 | ||
7d49d886 | 559 | pfn = start_pfn; |
5a115a5a JK |
560 | block_start_pfn = pfn & ~(pageblock_nr_pages - 1); |
561 | if (block_start_pfn < cc->zone->zone_start_pfn) | |
562 | block_start_pfn = cc->zone->zone_start_pfn; | |
7d49d886 VB |
563 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); |
564 | ||
565 | for (; pfn < end_pfn; pfn += isolated, | |
5a115a5a | 566 | block_start_pfn = block_end_pfn, |
7d49d886 | 567 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
568 | /* Protect pfn from changing by isolate_freepages_block */ |
569 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 570 | |
85aa125f MN |
571 | block_end_pfn = min(block_end_pfn, end_pfn); |
572 | ||
58420016 JK |
573 | /* |
574 | * pfn could pass the block_end_pfn if isolated freepage | |
575 | * is more than pageblock order. In this case, we adjust | |
576 | * scanning range to right one. | |
577 | */ | |
578 | if (pfn >= block_end_pfn) { | |
5a115a5a | 579 | block_start_pfn = pfn & ~(pageblock_nr_pages - 1); |
58420016 JK |
580 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); |
581 | block_end_pfn = min(block_end_pfn, end_pfn); | |
582 | } | |
583 | ||
5a115a5a JK |
584 | if (!pageblock_pfn_to_page(block_start_pfn, |
585 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
586 | break; |
587 | ||
e14c720e VB |
588 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
589 | block_end_pfn, &freelist, true); | |
85aa125f MN |
590 | |
591 | /* | |
592 | * In strict mode, isolate_freepages_block() returns 0 if | |
593 | * there are any holes in the block (ie. invalid PFNs or | |
594 | * non-free pages). | |
595 | */ | |
596 | if (!isolated) | |
597 | break; | |
598 | ||
599 | /* | |
600 | * If we managed to isolate pages, it is always (1 << n) * | |
601 | * pageblock_nr_pages for some non-negative n. (Max order | |
602 | * page may span two pageblocks). | |
603 | */ | |
604 | } | |
605 | ||
606 | /* split_free_page does not map the pages */ | |
607 | map_pages(&freelist); | |
608 | ||
609 | if (pfn < end_pfn) { | |
610 | /* Loop terminated early, cleanup. */ | |
611 | release_freepages(&freelist); | |
612 | return 0; | |
613 | } | |
614 | ||
615 | /* We don't use freelists for anything. */ | |
616 | return pfn; | |
617 | } | |
618 | ||
748446bb | 619 | /* Update the number of anon and file isolated pages in the zone */ |
edc2ca61 | 620 | static void acct_isolated(struct zone *zone, struct compact_control *cc) |
748446bb MG |
621 | { |
622 | struct page *page; | |
b9e84ac1 | 623 | unsigned int count[2] = { 0, }; |
748446bb | 624 | |
edc2ca61 VB |
625 | if (list_empty(&cc->migratepages)) |
626 | return; | |
627 | ||
b9e84ac1 MK |
628 | list_for_each_entry(page, &cc->migratepages, lru) |
629 | count[!!page_is_file_cache(page)]++; | |
748446bb | 630 | |
edc2ca61 VB |
631 | mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
632 | mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
633 | } |
634 | ||
635 | /* Similar to reclaim, but different enough that they don't share logic */ | |
636 | static bool too_many_isolated(struct zone *zone) | |
637 | { | |
bc693045 | 638 | unsigned long active, inactive, isolated; |
748446bb MG |
639 | |
640 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
641 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
642 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
643 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
644 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
645 | zone_page_state(zone, NR_ISOLATED_ANON); | |
646 | ||
bc693045 | 647 | return isolated > (inactive + active) / 2; |
748446bb MG |
648 | } |
649 | ||
2fe86e00 | 650 | /** |
edc2ca61 VB |
651 | * isolate_migratepages_block() - isolate all migrate-able pages within |
652 | * a single pageblock | |
2fe86e00 | 653 | * @cc: Compaction control structure. |
edc2ca61 VB |
654 | * @low_pfn: The first PFN to isolate |
655 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
656 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
657 | * |
658 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
659 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
660 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
661 | * first page that was not scanned (which may be both less, equal to or more | |
662 | * than end_pfn). | |
2fe86e00 | 663 | * |
edc2ca61 VB |
664 | * The pages are isolated on cc->migratepages list (not required to be empty), |
665 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
666 | * is neither read nor updated. | |
748446bb | 667 | */ |
edc2ca61 VB |
668 | static unsigned long |
669 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
670 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 671 | { |
edc2ca61 | 672 | struct zone *zone = cc->zone; |
b7aba698 | 673 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 674 | struct list_head *migratelist = &cc->migratepages; |
fa9add64 | 675 | struct lruvec *lruvec; |
b8b2d825 | 676 | unsigned long flags = 0; |
2a1402aa | 677 | bool locked = false; |
bb13ffeb | 678 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 679 | unsigned long start_pfn = low_pfn; |
748446bb | 680 | |
748446bb MG |
681 | /* |
682 | * Ensure that there are not too many pages isolated from the LRU | |
683 | * list by either parallel reclaimers or compaction. If there are, | |
684 | * delay for some time until fewer pages are isolated | |
685 | */ | |
686 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 687 | /* async migration should just abort */ |
e0b9daeb | 688 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 689 | return 0; |
f9e35b3b | 690 | |
748446bb MG |
691 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
692 | ||
693 | if (fatal_signal_pending(current)) | |
2fe86e00 | 694 | return 0; |
748446bb MG |
695 | } |
696 | ||
be976572 VB |
697 | if (compact_should_abort(cc)) |
698 | return 0; | |
aeef4b83 | 699 | |
748446bb | 700 | /* Time to isolate some pages for migration */ |
748446bb | 701 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 VB |
702 | bool is_lru; |
703 | ||
8b44d279 VB |
704 | /* |
705 | * Periodically drop the lock (if held) regardless of its | |
706 | * contention, to give chance to IRQs. Abort async compaction | |
707 | * if contended. | |
708 | */ | |
709 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
710 | && compact_unlock_should_abort(&zone->lru_lock, flags, | |
711 | &locked, cc)) | |
712 | break; | |
c67fe375 | 713 | |
748446bb MG |
714 | if (!pfn_valid_within(low_pfn)) |
715 | continue; | |
b7aba698 | 716 | nr_scanned++; |
748446bb | 717 | |
748446bb | 718 | page = pfn_to_page(low_pfn); |
dc908600 | 719 | |
bb13ffeb MG |
720 | if (!valid_page) |
721 | valid_page = page; | |
722 | ||
6c14466c | 723 | /* |
99c0fd5e VB |
724 | * Skip if free. We read page order here without zone lock |
725 | * which is generally unsafe, but the race window is small and | |
726 | * the worst thing that can happen is that we skip some | |
727 | * potential isolation targets. | |
6c14466c | 728 | */ |
99c0fd5e VB |
729 | if (PageBuddy(page)) { |
730 | unsigned long freepage_order = page_order_unsafe(page); | |
731 | ||
732 | /* | |
733 | * Without lock, we cannot be sure that what we got is | |
734 | * a valid page order. Consider only values in the | |
735 | * valid order range to prevent low_pfn overflow. | |
736 | */ | |
737 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
738 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 739 | continue; |
99c0fd5e | 740 | } |
748446bb | 741 | |
bf6bddf1 RA |
742 | /* |
743 | * Check may be lockless but that's ok as we recheck later. | |
744 | * It's possible to migrate LRU pages and balloon pages | |
745 | * Skip any other type of page | |
746 | */ | |
29c0dde8 VB |
747 | is_lru = PageLRU(page); |
748 | if (!is_lru) { | |
bf6bddf1 | 749 | if (unlikely(balloon_page_movable(page))) { |
d6d86c0a | 750 | if (balloon_page_isolate(page)) { |
bf6bddf1 | 751 | /* Successfully isolated */ |
b6c75016 | 752 | goto isolate_success; |
bf6bddf1 RA |
753 | } |
754 | } | |
bf6bddf1 | 755 | } |
bc835011 AA |
756 | |
757 | /* | |
29c0dde8 VB |
758 | * Regardless of being on LRU, compound pages such as THP and |
759 | * hugetlbfs are not to be compacted. We can potentially save | |
760 | * a lot of iterations if we skip them at once. The check is | |
761 | * racy, but we can consider only valid values and the only | |
762 | * danger is skipping too much. | |
bc835011 | 763 | */ |
29c0dde8 VB |
764 | if (PageCompound(page)) { |
765 | unsigned int comp_order = compound_order(page); | |
766 | ||
767 | if (likely(comp_order < MAX_ORDER)) | |
768 | low_pfn += (1UL << comp_order) - 1; | |
edc2ca61 | 769 | |
2a1402aa MG |
770 | continue; |
771 | } | |
772 | ||
29c0dde8 VB |
773 | if (!is_lru) |
774 | continue; | |
775 | ||
119d6d59 DR |
776 | /* |
777 | * Migration will fail if an anonymous page is pinned in memory, | |
778 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
779 | * admittedly racy check. | |
780 | */ | |
781 | if (!page_mapping(page) && | |
782 | page_count(page) > page_mapcount(page)) | |
783 | continue; | |
784 | ||
69b7189f VB |
785 | /* If we already hold the lock, we can skip some rechecking */ |
786 | if (!locked) { | |
8b44d279 VB |
787 | locked = compact_trylock_irqsave(&zone->lru_lock, |
788 | &flags, cc); | |
69b7189f VB |
789 | if (!locked) |
790 | break; | |
2a1402aa | 791 | |
29c0dde8 | 792 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f VB |
793 | if (!PageLRU(page)) |
794 | continue; | |
29c0dde8 VB |
795 | |
796 | /* | |
797 | * Page become compound since the non-locked check, | |
798 | * and it's on LRU. It can only be a THP so the order | |
799 | * is safe to read and it's 0 for tail pages. | |
800 | */ | |
801 | if (unlikely(PageCompound(page))) { | |
802 | low_pfn += (1UL << compound_order(page)) - 1; | |
69b7189f VB |
803 | continue; |
804 | } | |
bc835011 AA |
805 | } |
806 | ||
fa9add64 HD |
807 | lruvec = mem_cgroup_page_lruvec(page, zone); |
808 | ||
748446bb | 809 | /* Try isolate the page */ |
edc2ca61 | 810 | if (__isolate_lru_page(page, isolate_mode) != 0) |
748446bb MG |
811 | continue; |
812 | ||
29c0dde8 | 813 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 814 | |
748446bb | 815 | /* Successfully isolated */ |
fa9add64 | 816 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
b6c75016 JK |
817 | |
818 | isolate_success: | |
748446bb | 819 | list_add(&page->lru, migratelist); |
748446bb | 820 | cc->nr_migratepages++; |
b7aba698 | 821 | nr_isolated++; |
748446bb MG |
822 | |
823 | /* Avoid isolating too much */ | |
31b8384a HD |
824 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
825 | ++low_pfn; | |
748446bb | 826 | break; |
31b8384a | 827 | } |
748446bb MG |
828 | } |
829 | ||
99c0fd5e VB |
830 | /* |
831 | * The PageBuddy() check could have potentially brought us outside | |
832 | * the range to be scanned. | |
833 | */ | |
834 | if (unlikely(low_pfn > end_pfn)) | |
835 | low_pfn = end_pfn; | |
836 | ||
c67fe375 MG |
837 | if (locked) |
838 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
748446bb | 839 | |
50b5b094 VB |
840 | /* |
841 | * Update the pageblock-skip information and cached scanner pfn, | |
842 | * if the whole pageblock was scanned without isolating any page. | |
50b5b094 | 843 | */ |
35979ef3 | 844 | if (low_pfn == end_pfn) |
edc2ca61 | 845 | update_pageblock_skip(cc, valid_page, nr_isolated, true); |
bb13ffeb | 846 | |
e34d85f0 JK |
847 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
848 | nr_scanned, nr_isolated); | |
b7aba698 | 849 | |
010fc29a | 850 | count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned); |
397487db | 851 | if (nr_isolated) |
010fc29a | 852 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 853 | |
2fe86e00 MN |
854 | return low_pfn; |
855 | } | |
856 | ||
edc2ca61 VB |
857 | /** |
858 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
859 | * @cc: Compaction control structure. | |
860 | * @start_pfn: The first PFN to start isolating. | |
861 | * @end_pfn: The one-past-last PFN. | |
862 | * | |
863 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
864 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
865 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
866 | */ | |
867 | unsigned long | |
868 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
869 | unsigned long end_pfn) | |
870 | { | |
5a115a5a | 871 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
872 | |
873 | /* Scan block by block. First and last block may be incomplete */ | |
874 | pfn = start_pfn; | |
5a115a5a JK |
875 | block_start_pfn = pfn & ~(pageblock_nr_pages - 1); |
876 | if (block_start_pfn < cc->zone->zone_start_pfn) | |
877 | block_start_pfn = cc->zone->zone_start_pfn; | |
edc2ca61 VB |
878 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); |
879 | ||
880 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
5a115a5a | 881 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
882 | block_end_pfn += pageblock_nr_pages) { |
883 | ||
884 | block_end_pfn = min(block_end_pfn, end_pfn); | |
885 | ||
5a115a5a JK |
886 | if (!pageblock_pfn_to_page(block_start_pfn, |
887 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
888 | continue; |
889 | ||
890 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
891 | ISOLATE_UNEVICTABLE); | |
892 | ||
1d3c1a9c | 893 | if (!pfn) |
edc2ca61 | 894 | break; |
6ea41c0c JK |
895 | |
896 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
897 | break; | |
edc2ca61 VB |
898 | } |
899 | acct_isolated(cc->zone, cc); | |
900 | ||
901 | return pfn; | |
902 | } | |
903 | ||
ff9543fd MN |
904 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
905 | #ifdef CONFIG_COMPACTION | |
018e9a49 AM |
906 | |
907 | /* Returns true if the page is within a block suitable for migration to */ | |
908 | static bool suitable_migration_target(struct page *page) | |
909 | { | |
910 | /* If the page is a large free page, then disallow migration */ | |
911 | if (PageBuddy(page)) { | |
912 | /* | |
913 | * We are checking page_order without zone->lock taken. But | |
914 | * the only small danger is that we skip a potentially suitable | |
915 | * pageblock, so it's not worth to check order for valid range. | |
916 | */ | |
917 | if (page_order_unsafe(page) >= pageblock_order) | |
918 | return false; | |
919 | } | |
920 | ||
921 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ | |
922 | if (migrate_async_suitable(get_pageblock_migratetype(page))) | |
923 | return true; | |
924 | ||
925 | /* Otherwise skip the block */ | |
926 | return false; | |
927 | } | |
928 | ||
f2849aa0 VB |
929 | /* |
930 | * Test whether the free scanner has reached the same or lower pageblock than | |
931 | * the migration scanner, and compaction should thus terminate. | |
932 | */ | |
933 | static inline bool compact_scanners_met(struct compact_control *cc) | |
934 | { | |
935 | return (cc->free_pfn >> pageblock_order) | |
936 | <= (cc->migrate_pfn >> pageblock_order); | |
937 | } | |
938 | ||
2fe86e00 | 939 | /* |
ff9543fd MN |
940 | * Based on information in the current compact_control, find blocks |
941 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 942 | */ |
edc2ca61 | 943 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 944 | { |
edc2ca61 | 945 | struct zone *zone = cc->zone; |
ff9543fd | 946 | struct page *page; |
c96b9e50 | 947 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 948 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
949 | unsigned long block_end_pfn; /* end of current pageblock */ |
950 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 951 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 952 | |
ff9543fd MN |
953 | /* |
954 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 955 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
956 | * zone when isolating for the first time. For looping we also need |
957 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
958 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
959 | * For ending point, take care when isolating in last pageblock of a | |
960 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
961 | * The low boundary is the end of the pageblock the migration scanner |
962 | * is using. | |
ff9543fd | 963 | */ |
e14c720e | 964 | isolate_start_pfn = cc->free_pfn; |
c96b9e50 VB |
965 | block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1); |
966 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, | |
967 | zone_end_pfn(zone)); | |
7ed695e0 | 968 | low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages); |
2fe86e00 | 969 | |
ff9543fd MN |
970 | /* |
971 | * Isolate free pages until enough are available to migrate the | |
972 | * pages on cc->migratepages. We stop searching if the migrate | |
973 | * and free page scanners meet or enough free pages are isolated. | |
974 | */ | |
f5f61a32 | 975 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 976 | block_end_pfn = block_start_pfn, |
e14c720e VB |
977 | block_start_pfn -= pageblock_nr_pages, |
978 | isolate_start_pfn = block_start_pfn) { | |
f6ea3adb DR |
979 | /* |
980 | * This can iterate a massively long zone without finding any | |
981 | * suitable migration targets, so periodically check if we need | |
be976572 | 982 | * to schedule, or even abort async compaction. |
f6ea3adb | 983 | */ |
be976572 VB |
984 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) |
985 | && compact_should_abort(cc)) | |
986 | break; | |
f6ea3adb | 987 | |
7d49d886 VB |
988 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
989 | zone); | |
990 | if (!page) | |
ff9543fd MN |
991 | continue; |
992 | ||
993 | /* Check the block is suitable for migration */ | |
68e3e926 | 994 | if (!suitable_migration_target(page)) |
ff9543fd | 995 | continue; |
68e3e926 | 996 | |
bb13ffeb MG |
997 | /* If isolation recently failed, do not retry */ |
998 | if (!isolation_suitable(cc, page)) | |
999 | continue; | |
1000 | ||
e14c720e | 1001 | /* Found a block suitable for isolating free pages from. */ |
3b4b3cb8 DR |
1002 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, |
1003 | freelist, false); | |
ff9543fd | 1004 | |
e14c720e | 1005 | /* |
3b4b3cb8 DR |
1006 | * If we isolated enough freepages, or aborted due to lock |
1007 | * contention, terminate. | |
e14c720e | 1008 | */ |
f5f61a32 VB |
1009 | if ((cc->nr_freepages >= cc->nr_migratepages) |
1010 | || cc->contended) { | |
3b4b3cb8 DR |
1011 | if (isolate_start_pfn >= block_end_pfn) { |
1012 | /* | |
1013 | * Restart at previous pageblock if more | |
1014 | * freepages can be isolated next time. | |
1015 | */ | |
f5f61a32 VB |
1016 | isolate_start_pfn = |
1017 | block_start_pfn - pageblock_nr_pages; | |
3b4b3cb8 | 1018 | } |
be976572 | 1019 | break; |
3b4b3cb8 | 1020 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1021 | /* |
3b4b3cb8 DR |
1022 | * If isolation failed early, do not continue |
1023 | * needlessly. | |
f5f61a32 | 1024 | */ |
3b4b3cb8 | 1025 | break; |
f5f61a32 | 1026 | } |
ff9543fd MN |
1027 | } |
1028 | ||
1029 | /* split_free_page does not map the pages */ | |
1030 | map_pages(freelist); | |
1031 | ||
7ed695e0 | 1032 | /* |
f5f61a32 VB |
1033 | * Record where the free scanner will restart next time. Either we |
1034 | * broke from the loop and set isolate_start_pfn based on the last | |
1035 | * call to isolate_freepages_block(), or we met the migration scanner | |
1036 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1037 | */ |
f5f61a32 | 1038 | cc->free_pfn = isolate_start_pfn; |
748446bb MG |
1039 | } |
1040 | ||
1041 | /* | |
1042 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1043 | * from the isolated freelists in the block we are migrating to. | |
1044 | */ | |
1045 | static struct page *compaction_alloc(struct page *migratepage, | |
1046 | unsigned long data, | |
1047 | int **result) | |
1048 | { | |
1049 | struct compact_control *cc = (struct compact_control *)data; | |
1050 | struct page *freepage; | |
1051 | ||
be976572 VB |
1052 | /* |
1053 | * Isolate free pages if necessary, and if we are not aborting due to | |
1054 | * contention. | |
1055 | */ | |
748446bb | 1056 | if (list_empty(&cc->freepages)) { |
be976572 | 1057 | if (!cc->contended) |
edc2ca61 | 1058 | isolate_freepages(cc); |
748446bb MG |
1059 | |
1060 | if (list_empty(&cc->freepages)) | |
1061 | return NULL; | |
1062 | } | |
1063 | ||
1064 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1065 | list_del(&freepage->lru); | |
1066 | cc->nr_freepages--; | |
1067 | ||
1068 | return freepage; | |
1069 | } | |
1070 | ||
1071 | /* | |
d53aea3d DR |
1072 | * This is a migrate-callback that "frees" freepages back to the isolated |
1073 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1074 | * special handling needed for NUMA. | |
1075 | */ | |
1076 | static void compaction_free(struct page *page, unsigned long data) | |
1077 | { | |
1078 | struct compact_control *cc = (struct compact_control *)data; | |
1079 | ||
1080 | list_add(&page->lru, &cc->freepages); | |
1081 | cc->nr_freepages++; | |
1082 | } | |
1083 | ||
ff9543fd MN |
1084 | /* possible outcome of isolate_migratepages */ |
1085 | typedef enum { | |
1086 | ISOLATE_ABORT, /* Abort compaction now */ | |
1087 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1088 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1089 | } isolate_migrate_t; | |
1090 | ||
5bbe3547 EM |
1091 | /* |
1092 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1093 | * compactable pages. | |
1094 | */ | |
1095 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1096 | ||
ff9543fd | 1097 | /* |
edc2ca61 VB |
1098 | * Isolate all pages that can be migrated from the first suitable block, |
1099 | * starting at the block pointed to by the migrate scanner pfn within | |
1100 | * compact_control. | |
ff9543fd MN |
1101 | */ |
1102 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1103 | struct compact_control *cc) | |
1104 | { | |
5a115a5a JK |
1105 | unsigned long block_start_pfn; |
1106 | unsigned long block_end_pfn; | |
1107 | unsigned long low_pfn; | |
1a16718c | 1108 | unsigned long isolate_start_pfn; |
edc2ca61 VB |
1109 | struct page *page; |
1110 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1111 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
edc2ca61 | 1112 | (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
ff9543fd | 1113 | |
edc2ca61 VB |
1114 | /* |
1115 | * Start at where we last stopped, or beginning of the zone as | |
1116 | * initialized by compact_zone() | |
1117 | */ | |
1118 | low_pfn = cc->migrate_pfn; | |
5a115a5a JK |
1119 | block_start_pfn = cc->migrate_pfn & ~(pageblock_nr_pages - 1); |
1120 | if (block_start_pfn < zone->zone_start_pfn) | |
1121 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd MN |
1122 | |
1123 | /* Only scan within a pageblock boundary */ | |
5a115a5a | 1124 | block_end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages); |
ff9543fd | 1125 | |
edc2ca61 VB |
1126 | /* |
1127 | * Iterate over whole pageblocks until we find the first suitable. | |
1128 | * Do not cross the free scanner. | |
1129 | */ | |
5a115a5a JK |
1130 | for (; block_end_pfn <= cc->free_pfn; |
1131 | low_pfn = block_end_pfn, | |
1132 | block_start_pfn = block_end_pfn, | |
1133 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1134 | |
edc2ca61 VB |
1135 | /* |
1136 | * This can potentially iterate a massively long zone with | |
1137 | * many pageblocks unsuitable, so periodically check if we | |
1138 | * need to schedule, or even abort async compaction. | |
1139 | */ | |
1140 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)) | |
1141 | && compact_should_abort(cc)) | |
1142 | break; | |
ff9543fd | 1143 | |
5a115a5a JK |
1144 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1145 | zone); | |
7d49d886 | 1146 | if (!page) |
edc2ca61 VB |
1147 | continue; |
1148 | ||
edc2ca61 VB |
1149 | /* If isolation recently failed, do not retry */ |
1150 | if (!isolation_suitable(cc, page)) | |
1151 | continue; | |
1152 | ||
1153 | /* | |
1154 | * For async compaction, also only scan in MOVABLE blocks. | |
1155 | * Async compaction is optimistic to see if the minimum amount | |
1156 | * of work satisfies the allocation. | |
1157 | */ | |
1158 | if (cc->mode == MIGRATE_ASYNC && | |
1159 | !migrate_async_suitable(get_pageblock_migratetype(page))) | |
1160 | continue; | |
1161 | ||
1162 | /* Perform the isolation */ | |
1a16718c | 1163 | isolate_start_pfn = low_pfn; |
5a115a5a JK |
1164 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1165 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1166 | |
ff59909a HD |
1167 | if (!low_pfn || cc->contended) { |
1168 | acct_isolated(zone, cc); | |
edc2ca61 | 1169 | return ISOLATE_ABORT; |
ff59909a | 1170 | } |
edc2ca61 | 1171 | |
1a16718c JK |
1172 | /* |
1173 | * Record where we could have freed pages by migration and not | |
1174 | * yet flushed them to buddy allocator. | |
1175 | * - this is the lowest page that could have been isolated and | |
1176 | * then freed by migration. | |
1177 | */ | |
1178 | if (cc->nr_migratepages && !cc->last_migrated_pfn) | |
1179 | cc->last_migrated_pfn = isolate_start_pfn; | |
1180 | ||
edc2ca61 VB |
1181 | /* |
1182 | * Either we isolated something and proceed with migration. Or | |
1183 | * we failed and compact_zone should decide if we should | |
1184 | * continue or not. | |
1185 | */ | |
1186 | break; | |
1187 | } | |
1188 | ||
1189 | acct_isolated(zone, cc); | |
f2849aa0 VB |
1190 | /* Record where migration scanner will be restarted. */ |
1191 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1192 | |
edc2ca61 | 1193 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1194 | } |
1195 | ||
21c527a3 YB |
1196 | /* |
1197 | * order == -1 is expected when compacting via | |
1198 | * /proc/sys/vm/compact_memory | |
1199 | */ | |
1200 | static inline bool is_via_compact_memory(int order) | |
1201 | { | |
1202 | return order == -1; | |
1203 | } | |
1204 | ||
ea422501 | 1205 | static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc, |
6d7ce559 | 1206 | const int migratetype) |
748446bb | 1207 | { |
8fb74b9f | 1208 | unsigned int order; |
5a03b051 | 1209 | unsigned long watermark; |
56de7263 | 1210 | |
be976572 | 1211 | if (cc->contended || fatal_signal_pending(current)) |
2d1e1041 | 1212 | return COMPACT_CONTENDED; |
748446bb | 1213 | |
753341a4 | 1214 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1215 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1216 | /* Let the next compaction start anew. */ |
02333641 | 1217 | reset_cached_positions(zone); |
55b7c4c9 | 1218 | |
62997027 MG |
1219 | /* |
1220 | * Mark that the PG_migrate_skip information should be cleared | |
1221 | * by kswapd when it goes to sleep. kswapd does not set the | |
1222 | * flag itself as the decision to be clear should be directly | |
1223 | * based on an allocation request. | |
1224 | */ | |
1225 | if (!current_is_kswapd()) | |
1226 | zone->compact_blockskip_flush = true; | |
1227 | ||
ff732276 MH |
1228 | if (cc->whole_zone) |
1229 | return COMPACT_COMPLETE; | |
1230 | else | |
1231 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1232 | } |
748446bb | 1233 | |
21c527a3 | 1234 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1235 | return COMPACT_CONTINUE; |
1236 | ||
3957c776 MH |
1237 | /* Compaction run is not finished if the watermark is not met */ |
1238 | watermark = low_wmark_pages(zone); | |
3957c776 | 1239 | |
ebff3980 VB |
1240 | if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx, |
1241 | cc->alloc_flags)) | |
3957c776 MH |
1242 | return COMPACT_CONTINUE; |
1243 | ||
56de7263 | 1244 | /* Direct compactor: Is a suitable page free? */ |
8fb74b9f MG |
1245 | for (order = cc->order; order < MAX_ORDER; order++) { |
1246 | struct free_area *area = &zone->free_area[order]; | |
2149cdae | 1247 | bool can_steal; |
8fb74b9f MG |
1248 | |
1249 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1250 | if (!list_empty(&area->free_list[migratetype])) |
8fb74b9f MG |
1251 | return COMPACT_PARTIAL; |
1252 | ||
2149cdae JK |
1253 | #ifdef CONFIG_CMA |
1254 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1255 | if (migratetype == MIGRATE_MOVABLE && | |
1256 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
1257 | return COMPACT_PARTIAL; | |
1258 | #endif | |
1259 | /* | |
1260 | * Job done if allocation would steal freepages from | |
1261 | * other migratetype buddy lists. | |
1262 | */ | |
1263 | if (find_suitable_fallback(area, order, migratetype, | |
1264 | true, &can_steal) != -1) | |
56de7263 MG |
1265 | return COMPACT_PARTIAL; |
1266 | } | |
1267 | ||
837d026d JK |
1268 | return COMPACT_NO_SUITABLE_PAGE; |
1269 | } | |
1270 | ||
ea422501 MH |
1271 | static enum compact_result compact_finished(struct zone *zone, |
1272 | struct compact_control *cc, | |
1273 | const int migratetype) | |
837d026d JK |
1274 | { |
1275 | int ret; | |
1276 | ||
1277 | ret = __compact_finished(zone, cc, migratetype); | |
1278 | trace_mm_compaction_finished(zone, cc->order, ret); | |
1279 | if (ret == COMPACT_NO_SUITABLE_PAGE) | |
1280 | ret = COMPACT_CONTINUE; | |
1281 | ||
1282 | return ret; | |
748446bb MG |
1283 | } |
1284 | ||
3e7d3449 MG |
1285 | /* |
1286 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1287 | * Returns | |
1288 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
1289 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
1290 | * COMPACT_CONTINUE - If compaction should run now | |
1291 | */ | |
ea422501 | 1292 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
09807762 | 1293 | unsigned int alloc_flags, |
8f32a02e MH |
1294 | int classzone_idx, |
1295 | unsigned long wmark_target) | |
3e7d3449 MG |
1296 | { |
1297 | int fragindex; | |
1298 | unsigned long watermark; | |
1299 | ||
21c527a3 | 1300 | if (is_via_compact_memory(order)) |
3957c776 MH |
1301 | return COMPACT_CONTINUE; |
1302 | ||
ebff3980 VB |
1303 | watermark = low_wmark_pages(zone); |
1304 | /* | |
1305 | * If watermarks for high-order allocation are already met, there | |
1306 | * should be no need for compaction at all. | |
1307 | */ | |
1308 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1309 | alloc_flags)) | |
1310 | return COMPACT_PARTIAL; | |
1311 | ||
3e7d3449 MG |
1312 | /* |
1313 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
1314 | * This is because during migration, copies of pages need to be | |
1315 | * allocated and for a short time, the footprint is higher | |
1316 | */ | |
ebff3980 | 1317 | watermark += (2UL << order); |
8f32a02e MH |
1318 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
1319 | alloc_flags, wmark_target)) | |
3e7d3449 MG |
1320 | return COMPACT_SKIPPED; |
1321 | ||
1322 | /* | |
1323 | * fragmentation index determines if allocation failures are due to | |
1324 | * low memory or external fragmentation | |
1325 | * | |
ebff3980 VB |
1326 | * index of -1000 would imply allocations might succeed depending on |
1327 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1328 | * index towards 0 implies failure is due to lack of memory |
1329 | * index towards 1000 implies failure is due to fragmentation | |
1330 | * | |
1331 | * Only compact if a failure would be due to fragmentation. | |
1332 | */ | |
1333 | fragindex = fragmentation_index(zone, order); | |
1334 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
837d026d | 1335 | return COMPACT_NOT_SUITABLE_ZONE; |
3e7d3449 | 1336 | |
3e7d3449 MG |
1337 | return COMPACT_CONTINUE; |
1338 | } | |
1339 | ||
ea422501 | 1340 | enum compact_result compaction_suitable(struct zone *zone, int order, |
09807762 MG |
1341 | unsigned int alloc_flags, |
1342 | int classzone_idx) | |
837d026d | 1343 | { |
ea422501 | 1344 | enum compact_result ret; |
837d026d | 1345 | |
8f32a02e MH |
1346 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, |
1347 | zone_page_state(zone, NR_FREE_PAGES)); | |
837d026d JK |
1348 | trace_mm_compaction_suitable(zone, order, ret); |
1349 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1350 | ret = COMPACT_SKIPPED; | |
1351 | ||
1352 | return ret; | |
1353 | } | |
1354 | ||
8f32a02e MH |
1355 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
1356 | int alloc_flags) | |
1357 | { | |
1358 | struct zone *zone; | |
1359 | struct zoneref *z; | |
1360 | ||
1361 | /* | |
1362 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1363 | * retrying the reclaim. | |
1364 | */ | |
1365 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1366 | ac->nodemask) { | |
1367 | unsigned long available; | |
1368 | enum compact_result compact_result; | |
1369 | ||
1370 | /* | |
1371 | * Do not consider all the reclaimable memory because we do not | |
1372 | * want to trash just for a single high order allocation which | |
1373 | * is even not guaranteed to appear even if __compaction_suitable | |
1374 | * is happy about the watermark check. | |
1375 | */ | |
1376 | available = zone_reclaimable_pages(zone) / order; | |
1377 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); | |
1378 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1379 | ac->classzone_idx, available); | |
1380 | if (compact_result != COMPACT_SKIPPED && | |
1381 | compact_result != COMPACT_NOT_SUITABLE_ZONE) | |
1382 | return true; | |
1383 | } | |
1384 | ||
1385 | return false; | |
1386 | } | |
1387 | ||
ea422501 | 1388 | static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc) |
748446bb | 1389 | { |
ea422501 | 1390 | enum compact_result ret; |
c89511ab | 1391 | unsigned long start_pfn = zone->zone_start_pfn; |
108bcc96 | 1392 | unsigned long end_pfn = zone_end_pfn(zone); |
6d7ce559 | 1393 | const int migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
e0b9daeb | 1394 | const bool sync = cc->mode != MIGRATE_ASYNC; |
748446bb | 1395 | |
ebff3980 VB |
1396 | ret = compaction_suitable(zone, cc->order, cc->alloc_flags, |
1397 | cc->classzone_idx); | |
3e7d3449 MG |
1398 | switch (ret) { |
1399 | case COMPACT_PARTIAL: | |
1400 | case COMPACT_SKIPPED: | |
1401 | /* Compaction is likely to fail */ | |
1402 | return ret; | |
1403 | case COMPACT_CONTINUE: | |
1404 | /* Fall through to compaction */ | |
1405 | ; | |
1406 | } | |
1407 | ||
d3132e4b VB |
1408 | /* |
1409 | * Clear pageblock skip if there were failures recently and compaction | |
1410 | * is about to be retried after being deferred. kswapd does not do | |
1411 | * this reset as it'll reset the cached information when going to sleep. | |
1412 | */ | |
1413 | if (compaction_restarting(zone, cc->order) && !current_is_kswapd()) | |
1414 | __reset_isolation_suitable(zone); | |
1415 | ||
c89511ab MG |
1416 | /* |
1417 | * Setup to move all movable pages to the end of the zone. Used cached | |
1418 | * information on where the scanners should start but check that it | |
1419 | * is initialised by ensuring the values are within zone boundaries. | |
1420 | */ | |
e0b9daeb | 1421 | cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync]; |
c89511ab | 1422 | cc->free_pfn = zone->compact_cached_free_pfn; |
8b315860 JK |
1423 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
1424 | cc->free_pfn = round_down(end_pfn - 1, pageblock_nr_pages); | |
c89511ab MG |
1425 | zone->compact_cached_free_pfn = cc->free_pfn; |
1426 | } | |
8b315860 | 1427 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { |
c89511ab | 1428 | cc->migrate_pfn = start_pfn; |
35979ef3 DR |
1429 | zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1430 | zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
c89511ab | 1431 | } |
ff732276 MH |
1432 | |
1433 | if (cc->migrate_pfn == start_pfn) | |
1434 | cc->whole_zone = true; | |
1435 | ||
1a16718c | 1436 | cc->last_migrated_pfn = 0; |
748446bb | 1437 | |
16c4a097 JK |
1438 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1439 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1440 | |
748446bb MG |
1441 | migrate_prep_local(); |
1442 | ||
6d7ce559 DR |
1443 | while ((ret = compact_finished(zone, cc, migratetype)) == |
1444 | COMPACT_CONTINUE) { | |
9d502c1c | 1445 | int err; |
748446bb | 1446 | |
f9e35b3b MG |
1447 | switch (isolate_migratepages(zone, cc)) { |
1448 | case ISOLATE_ABORT: | |
2d1e1041 | 1449 | ret = COMPACT_CONTENDED; |
5733c7d1 | 1450 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 1451 | cc->nr_migratepages = 0; |
f9e35b3b MG |
1452 | goto out; |
1453 | case ISOLATE_NONE: | |
fdaf7f5c VB |
1454 | /* |
1455 | * We haven't isolated and migrated anything, but | |
1456 | * there might still be unflushed migrations from | |
1457 | * previous cc->order aligned block. | |
1458 | */ | |
1459 | goto check_drain; | |
f9e35b3b MG |
1460 | case ISOLATE_SUCCESS: |
1461 | ; | |
1462 | } | |
748446bb | 1463 | |
d53aea3d | 1464 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 1465 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 1466 | MR_COMPACTION); |
748446bb | 1467 | |
f8c9301f VB |
1468 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
1469 | &cc->migratepages); | |
748446bb | 1470 | |
f8c9301f VB |
1471 | /* All pages were either migrated or will be released */ |
1472 | cc->nr_migratepages = 0; | |
9d502c1c | 1473 | if (err) { |
5733c7d1 | 1474 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
1475 | /* |
1476 | * migrate_pages() may return -ENOMEM when scanners meet | |
1477 | * and we want compact_finished() to detect it | |
1478 | */ | |
f2849aa0 | 1479 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 1480 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
1481 | goto out; |
1482 | } | |
748446bb | 1483 | } |
fdaf7f5c | 1484 | |
fdaf7f5c VB |
1485 | check_drain: |
1486 | /* | |
1487 | * Has the migration scanner moved away from the previous | |
1488 | * cc->order aligned block where we migrated from? If yes, | |
1489 | * flush the pages that were freed, so that they can merge and | |
1490 | * compact_finished() can detect immediately if allocation | |
1491 | * would succeed. | |
1492 | */ | |
1a16718c | 1493 | if (cc->order > 0 && cc->last_migrated_pfn) { |
fdaf7f5c VB |
1494 | int cpu; |
1495 | unsigned long current_block_start = | |
1496 | cc->migrate_pfn & ~((1UL << cc->order) - 1); | |
1497 | ||
1a16718c | 1498 | if (cc->last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
1499 | cpu = get_cpu(); |
1500 | lru_add_drain_cpu(cpu); | |
1501 | drain_local_pages(zone); | |
1502 | put_cpu(); | |
1503 | /* No more flushing until we migrate again */ | |
1a16718c | 1504 | cc->last_migrated_pfn = 0; |
fdaf7f5c VB |
1505 | } |
1506 | } | |
1507 | ||
748446bb MG |
1508 | } |
1509 | ||
f9e35b3b | 1510 | out: |
6bace090 VB |
1511 | /* |
1512 | * Release free pages and update where the free scanner should restart, | |
1513 | * so we don't leave any returned pages behind in the next attempt. | |
1514 | */ | |
1515 | if (cc->nr_freepages > 0) { | |
1516 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
1517 | ||
1518 | cc->nr_freepages = 0; | |
1519 | VM_BUG_ON(free_pfn == 0); | |
1520 | /* The cached pfn is always the first in a pageblock */ | |
1521 | free_pfn &= ~(pageblock_nr_pages-1); | |
1522 | /* | |
1523 | * Only go back, not forward. The cached pfn might have been | |
1524 | * already reset to zone end in compact_finished() | |
1525 | */ | |
1526 | if (free_pfn > zone->compact_cached_free_pfn) | |
1527 | zone->compact_cached_free_pfn = free_pfn; | |
1528 | } | |
748446bb | 1529 | |
16c4a097 JK |
1530 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
1531 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 1532 | |
2d1e1041 VB |
1533 | if (ret == COMPACT_CONTENDED) |
1534 | ret = COMPACT_PARTIAL; | |
1535 | ||
748446bb MG |
1536 | return ret; |
1537 | } | |
76ab0f53 | 1538 | |
ea422501 | 1539 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
ebff3980 | 1540 | gfp_t gfp_mask, enum migrate_mode mode, int *contended, |
09807762 | 1541 | unsigned int alloc_flags, int classzone_idx) |
56de7263 | 1542 | { |
ea422501 | 1543 | enum compact_result ret; |
56de7263 MG |
1544 | struct compact_control cc = { |
1545 | .nr_freepages = 0, | |
1546 | .nr_migratepages = 0, | |
1547 | .order = order, | |
6d7ce559 | 1548 | .gfp_mask = gfp_mask, |
56de7263 | 1549 | .zone = zone, |
e0b9daeb | 1550 | .mode = mode, |
ebff3980 VB |
1551 | .alloc_flags = alloc_flags, |
1552 | .classzone_idx = classzone_idx, | |
56de7263 MG |
1553 | }; |
1554 | INIT_LIST_HEAD(&cc.freepages); | |
1555 | INIT_LIST_HEAD(&cc.migratepages); | |
1556 | ||
e64c5237 SL |
1557 | ret = compact_zone(zone, &cc); |
1558 | ||
1559 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
1560 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
1561 | ||
1562 | *contended = cc.contended; | |
1563 | return ret; | |
56de7263 MG |
1564 | } |
1565 | ||
5e771905 MG |
1566 | int sysctl_extfrag_threshold = 500; |
1567 | ||
56de7263 MG |
1568 | /** |
1569 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 1570 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
1571 | * @order: The order of the current allocation |
1572 | * @alloc_flags: The allocation flags of the current allocation | |
1573 | * @ac: The context of current allocation | |
e0b9daeb | 1574 | * @mode: The migration mode for async, sync light, or sync migration |
1f9efdef VB |
1575 | * @contended: Return value that determines if compaction was aborted due to |
1576 | * need_resched() or lock contention | |
56de7263 MG |
1577 | * |
1578 | * This is the main entry point for direct page compaction. | |
1579 | */ | |
ea422501 | 1580 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
09807762 MG |
1581 | unsigned int alloc_flags, const struct alloc_context *ac, |
1582 | enum migrate_mode mode, int *contended) | |
56de7263 | 1583 | { |
56de7263 MG |
1584 | int may_enter_fs = gfp_mask & __GFP_FS; |
1585 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
1586 | struct zoneref *z; |
1587 | struct zone *zone; | |
0b7dc5f6 | 1588 | enum compact_result rc = COMPACT_SKIPPED; |
1f9efdef VB |
1589 | int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */ |
1590 | ||
1591 | *contended = COMPACT_CONTENDED_NONE; | |
56de7263 | 1592 | |
4ffb6335 | 1593 | /* Check if the GFP flags allow compaction */ |
c5a73c3d | 1594 | if (!order || !may_enter_fs || !may_perform_io) |
53853e2d | 1595 | return COMPACT_SKIPPED; |
56de7263 | 1596 | |
837d026d JK |
1597 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode); |
1598 | ||
56de7263 | 1599 | /* Compact each zone in the list */ |
1a6d53a1 VB |
1600 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
1601 | ac->nodemask) { | |
ea422501 | 1602 | enum compact_result status; |
1f9efdef | 1603 | int zone_contended; |
56de7263 | 1604 | |
0b7dc5f6 MH |
1605 | if (compaction_deferred(zone, order)) { |
1606 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); | |
53853e2d | 1607 | continue; |
0b7dc5f6 | 1608 | } |
53853e2d | 1609 | |
e0b9daeb | 1610 | status = compact_zone_order(zone, order, gfp_mask, mode, |
1a6d53a1 VB |
1611 | &zone_contended, alloc_flags, |
1612 | ac->classzone_idx); | |
56de7263 | 1613 | rc = max(status, rc); |
1f9efdef VB |
1614 | /* |
1615 | * It takes at least one zone that wasn't lock contended | |
1616 | * to clear all_zones_contended. | |
1617 | */ | |
1618 | all_zones_contended &= zone_contended; | |
56de7263 | 1619 | |
3e7d3449 | 1620 | /* If a normal allocation would succeed, stop compacting */ |
ebff3980 | 1621 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), |
1a6d53a1 | 1622 | ac->classzone_idx, alloc_flags)) { |
53853e2d VB |
1623 | /* |
1624 | * We think the allocation will succeed in this zone, | |
1625 | * but it is not certain, hence the false. The caller | |
1626 | * will repeat this with true if allocation indeed | |
1627 | * succeeds in this zone. | |
1628 | */ | |
1629 | compaction_defer_reset(zone, order, false); | |
1f9efdef VB |
1630 | /* |
1631 | * It is possible that async compaction aborted due to | |
1632 | * need_resched() and the watermarks were ok thanks to | |
1633 | * somebody else freeing memory. The allocation can | |
1634 | * however still fail so we better signal the | |
1635 | * need_resched() contention anyway (this will not | |
1636 | * prevent the allocation attempt). | |
1637 | */ | |
1638 | if (zone_contended == COMPACT_CONTENDED_SCHED) | |
1639 | *contended = COMPACT_CONTENDED_SCHED; | |
1640 | ||
1641 | goto break_loop; | |
1642 | } | |
1643 | ||
ff732276 MH |
1644 | if (mode != MIGRATE_ASYNC && (status == COMPACT_COMPLETE || |
1645 | status == COMPACT_PARTIAL_SKIPPED)) { | |
53853e2d VB |
1646 | /* |
1647 | * We think that allocation won't succeed in this zone | |
1648 | * so we defer compaction there. If it ends up | |
1649 | * succeeding after all, it will be reset. | |
1650 | */ | |
1651 | defer_compaction(zone, order); | |
1652 | } | |
1f9efdef VB |
1653 | |
1654 | /* | |
1655 | * We might have stopped compacting due to need_resched() in | |
1656 | * async compaction, or due to a fatal signal detected. In that | |
1657 | * case do not try further zones and signal need_resched() | |
1658 | * contention. | |
1659 | */ | |
1660 | if ((zone_contended == COMPACT_CONTENDED_SCHED) | |
1661 | || fatal_signal_pending(current)) { | |
1662 | *contended = COMPACT_CONTENDED_SCHED; | |
1663 | goto break_loop; | |
1664 | } | |
1665 | ||
1666 | continue; | |
1667 | break_loop: | |
1668 | /* | |
1669 | * We might not have tried all the zones, so be conservative | |
1670 | * and assume they are not all lock contended. | |
1671 | */ | |
1672 | all_zones_contended = 0; | |
1673 | break; | |
56de7263 MG |
1674 | } |
1675 | ||
1f9efdef VB |
1676 | /* |
1677 | * If at least one zone wasn't deferred or skipped, we report if all | |
1678 | * zones that were tried were lock contended. | |
1679 | */ | |
0b7dc5f6 | 1680 | if (rc > COMPACT_INACTIVE && all_zones_contended) |
1f9efdef VB |
1681 | *contended = COMPACT_CONTENDED_LOCK; |
1682 | ||
56de7263 MG |
1683 | return rc; |
1684 | } | |
1685 | ||
1686 | ||
76ab0f53 | 1687 | /* Compact all zones within a node */ |
7103f16d | 1688 | static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
1689 | { |
1690 | int zoneid; | |
76ab0f53 MG |
1691 | struct zone *zone; |
1692 | ||
76ab0f53 | 1693 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
1694 | |
1695 | zone = &pgdat->node_zones[zoneid]; | |
1696 | if (!populated_zone(zone)) | |
1697 | continue; | |
1698 | ||
7be62de9 RR |
1699 | cc->nr_freepages = 0; |
1700 | cc->nr_migratepages = 0; | |
1701 | cc->zone = zone; | |
1702 | INIT_LIST_HEAD(&cc->freepages); | |
1703 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 1704 | |
195b0c60 GK |
1705 | /* |
1706 | * When called via /proc/sys/vm/compact_memory | |
1707 | * this makes sure we compact the whole zone regardless of | |
1708 | * cached scanner positions. | |
1709 | */ | |
21c527a3 | 1710 | if (is_via_compact_memory(cc->order)) |
195b0c60 GK |
1711 | __reset_isolation_suitable(zone); |
1712 | ||
21c527a3 YB |
1713 | if (is_via_compact_memory(cc->order) || |
1714 | !compaction_deferred(zone, cc->order)) | |
7be62de9 | 1715 | compact_zone(zone, cc); |
76ab0f53 | 1716 | |
aff62249 | 1717 | if (cc->order > 0) { |
de6c60a6 VB |
1718 | if (zone_watermark_ok(zone, cc->order, |
1719 | low_wmark_pages(zone), 0, 0)) | |
1720 | compaction_defer_reset(zone, cc->order, false); | |
aff62249 RR |
1721 | } |
1722 | ||
7be62de9 RR |
1723 | VM_BUG_ON(!list_empty(&cc->freepages)); |
1724 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 | 1725 | } |
76ab0f53 MG |
1726 | } |
1727 | ||
7103f16d | 1728 | void compact_pgdat(pg_data_t *pgdat, int order) |
7be62de9 RR |
1729 | { |
1730 | struct compact_control cc = { | |
1731 | .order = order, | |
e0b9daeb | 1732 | .mode = MIGRATE_ASYNC, |
7be62de9 RR |
1733 | }; |
1734 | ||
3a7200af MG |
1735 | if (!order) |
1736 | return; | |
1737 | ||
7103f16d | 1738 | __compact_pgdat(pgdat, &cc); |
7be62de9 RR |
1739 | } |
1740 | ||
7103f16d | 1741 | static void compact_node(int nid) |
7be62de9 | 1742 | { |
7be62de9 RR |
1743 | struct compact_control cc = { |
1744 | .order = -1, | |
e0b9daeb | 1745 | .mode = MIGRATE_SYNC, |
91ca9186 | 1746 | .ignore_skip_hint = true, |
7be62de9 RR |
1747 | }; |
1748 | ||
7103f16d | 1749 | __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
1750 | } |
1751 | ||
76ab0f53 | 1752 | /* Compact all nodes in the system */ |
7964c06d | 1753 | static void compact_nodes(void) |
76ab0f53 MG |
1754 | { |
1755 | int nid; | |
1756 | ||
8575ec29 HD |
1757 | /* Flush pending updates to the LRU lists */ |
1758 | lru_add_drain_all(); | |
1759 | ||
76ab0f53 MG |
1760 | for_each_online_node(nid) |
1761 | compact_node(nid); | |
76ab0f53 MG |
1762 | } |
1763 | ||
1764 | /* The written value is actually unused, all memory is compacted */ | |
1765 | int sysctl_compact_memory; | |
1766 | ||
1767 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
1768 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
1769 | void __user *buffer, size_t *length, loff_t *ppos) | |
1770 | { | |
1771 | if (write) | |
7964c06d | 1772 | compact_nodes(); |
76ab0f53 MG |
1773 | |
1774 | return 0; | |
1775 | } | |
ed4a6d7f | 1776 | |
5e771905 MG |
1777 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
1778 | void __user *buffer, size_t *length, loff_t *ppos) | |
1779 | { | |
1780 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
1781 | ||
1782 | return 0; | |
1783 | } | |
1784 | ||
ed4a6d7f | 1785 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
74e77fb9 | 1786 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 1787 | struct device_attribute *attr, |
ed4a6d7f MG |
1788 | const char *buf, size_t count) |
1789 | { | |
8575ec29 HD |
1790 | int nid = dev->id; |
1791 | ||
1792 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
1793 | /* Flush pending updates to the LRU lists */ | |
1794 | lru_add_drain_all(); | |
1795 | ||
1796 | compact_node(nid); | |
1797 | } | |
ed4a6d7f MG |
1798 | |
1799 | return count; | |
1800 | } | |
10fbcf4c | 1801 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
1802 | |
1803 | int compaction_register_node(struct node *node) | |
1804 | { | |
10fbcf4c | 1805 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1806 | } |
1807 | ||
1808 | void compaction_unregister_node(struct node *node) | |
1809 | { | |
10fbcf4c | 1810 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
1811 | } |
1812 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd MN |
1813 | |
1814 | #endif /* CONFIG_COMPACTION */ |