]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
748446bb MG |
2 | /* |
3 | * linux/mm/compaction.c | |
4 | * | |
5 | * Memory compaction for the reduction of external fragmentation. Note that | |
6 | * this heavily depends upon page migration to do all the real heavy | |
7 | * lifting | |
8 | * | |
9 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
10 | */ | |
698b1b30 | 11 | #include <linux/cpu.h> |
748446bb MG |
12 | #include <linux/swap.h> |
13 | #include <linux/migrate.h> | |
14 | #include <linux/compaction.h> | |
15 | #include <linux/mm_inline.h> | |
174cd4b1 | 16 | #include <linux/sched/signal.h> |
748446bb | 17 | #include <linux/backing-dev.h> |
76ab0f53 | 18 | #include <linux/sysctl.h> |
ed4a6d7f | 19 | #include <linux/sysfs.h> |
194159fb | 20 | #include <linux/page-isolation.h> |
b8c73fc2 | 21 | #include <linux/kasan.h> |
698b1b30 VB |
22 | #include <linux/kthread.h> |
23 | #include <linux/freezer.h> | |
83358ece | 24 | #include <linux/page_owner.h> |
eb414681 | 25 | #include <linux/psi.h> |
748446bb MG |
26 | #include "internal.h" |
27 | ||
010fc29a MK |
28 | #ifdef CONFIG_COMPACTION |
29 | static inline void count_compact_event(enum vm_event_item item) | |
30 | { | |
31 | count_vm_event(item); | |
32 | } | |
33 | ||
34 | static inline void count_compact_events(enum vm_event_item item, long delta) | |
35 | { | |
36 | count_vm_events(item, delta); | |
37 | } | |
38 | #else | |
39 | #define count_compact_event(item) do { } while (0) | |
40 | #define count_compact_events(item, delta) do { } while (0) | |
41 | #endif | |
42 | ||
ff9543fd MN |
43 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
44 | ||
b7aba698 MG |
45 | #define CREATE_TRACE_POINTS |
46 | #include <trace/events/compaction.h> | |
47 | ||
06b6640a VB |
48 | #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) |
49 | #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) | |
50 | #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order) | |
51 | #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order) | |
52 | ||
748446bb MG |
53 | static unsigned long release_freepages(struct list_head *freelist) |
54 | { | |
55 | struct page *page, *next; | |
6bace090 | 56 | unsigned long high_pfn = 0; |
748446bb MG |
57 | |
58 | list_for_each_entry_safe(page, next, freelist, lru) { | |
6bace090 | 59 | unsigned long pfn = page_to_pfn(page); |
748446bb MG |
60 | list_del(&page->lru); |
61 | __free_page(page); | |
6bace090 VB |
62 | if (pfn > high_pfn) |
63 | high_pfn = pfn; | |
748446bb MG |
64 | } |
65 | ||
6bace090 | 66 | return high_pfn; |
748446bb MG |
67 | } |
68 | ||
4469ab98 | 69 | static void split_map_pages(struct list_head *list) |
ff9543fd | 70 | { |
66c64223 JK |
71 | unsigned int i, order, nr_pages; |
72 | struct page *page, *next; | |
73 | LIST_HEAD(tmp_list); | |
74 | ||
75 | list_for_each_entry_safe(page, next, list, lru) { | |
76 | list_del(&page->lru); | |
77 | ||
78 | order = page_private(page); | |
79 | nr_pages = 1 << order; | |
66c64223 | 80 | |
46f24fd8 | 81 | post_alloc_hook(page, order, __GFP_MOVABLE); |
66c64223 JK |
82 | if (order) |
83 | split_page(page, order); | |
ff9543fd | 84 | |
66c64223 JK |
85 | for (i = 0; i < nr_pages; i++) { |
86 | list_add(&page->lru, &tmp_list); | |
87 | page++; | |
88 | } | |
ff9543fd | 89 | } |
66c64223 JK |
90 | |
91 | list_splice(&tmp_list, list); | |
ff9543fd MN |
92 | } |
93 | ||
bb13ffeb | 94 | #ifdef CONFIG_COMPACTION |
24e2716f | 95 | |
bda807d4 MK |
96 | int PageMovable(struct page *page) |
97 | { | |
98 | struct address_space *mapping; | |
99 | ||
100 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
101 | if (!__PageMovable(page)) | |
102 | return 0; | |
103 | ||
104 | mapping = page_mapping(page); | |
105 | if (mapping && mapping->a_ops && mapping->a_ops->isolate_page) | |
106 | return 1; | |
107 | ||
108 | return 0; | |
109 | } | |
110 | EXPORT_SYMBOL(PageMovable); | |
111 | ||
112 | void __SetPageMovable(struct page *page, struct address_space *mapping) | |
113 | { | |
114 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
115 | VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page); | |
116 | page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE); | |
117 | } | |
118 | EXPORT_SYMBOL(__SetPageMovable); | |
119 | ||
120 | void __ClearPageMovable(struct page *page) | |
121 | { | |
122 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
123 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
124 | /* | |
125 | * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE | |
126 | * flag so that VM can catch up released page by driver after isolation. | |
127 | * With it, VM migration doesn't try to put it back. | |
128 | */ | |
129 | page->mapping = (void *)((unsigned long)page->mapping & | |
130 | PAGE_MAPPING_MOVABLE); | |
131 | } | |
132 | EXPORT_SYMBOL(__ClearPageMovable); | |
133 | ||
24e2716f JK |
134 | /* Do not skip compaction more than 64 times */ |
135 | #define COMPACT_MAX_DEFER_SHIFT 6 | |
136 | ||
137 | /* | |
138 | * Compaction is deferred when compaction fails to result in a page | |
139 | * allocation success. 1 << compact_defer_limit compactions are skipped up | |
140 | * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT | |
141 | */ | |
142 | void defer_compaction(struct zone *zone, int order) | |
143 | { | |
144 | zone->compact_considered = 0; | |
145 | zone->compact_defer_shift++; | |
146 | ||
147 | if (order < zone->compact_order_failed) | |
148 | zone->compact_order_failed = order; | |
149 | ||
150 | if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) | |
151 | zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; | |
152 | ||
153 | trace_mm_compaction_defer_compaction(zone, order); | |
154 | } | |
155 | ||
156 | /* Returns true if compaction should be skipped this time */ | |
157 | bool compaction_deferred(struct zone *zone, int order) | |
158 | { | |
159 | unsigned long defer_limit = 1UL << zone->compact_defer_shift; | |
160 | ||
161 | if (order < zone->compact_order_failed) | |
162 | return false; | |
163 | ||
164 | /* Avoid possible overflow */ | |
165 | if (++zone->compact_considered > defer_limit) | |
166 | zone->compact_considered = defer_limit; | |
167 | ||
168 | if (zone->compact_considered >= defer_limit) | |
169 | return false; | |
170 | ||
171 | trace_mm_compaction_deferred(zone, order); | |
172 | ||
173 | return true; | |
174 | } | |
175 | ||
176 | /* | |
177 | * Update defer tracking counters after successful compaction of given order, | |
178 | * which means an allocation either succeeded (alloc_success == true) or is | |
179 | * expected to succeed. | |
180 | */ | |
181 | void compaction_defer_reset(struct zone *zone, int order, | |
182 | bool alloc_success) | |
183 | { | |
184 | if (alloc_success) { | |
185 | zone->compact_considered = 0; | |
186 | zone->compact_defer_shift = 0; | |
187 | } | |
188 | if (order >= zone->compact_order_failed) | |
189 | zone->compact_order_failed = order + 1; | |
190 | ||
191 | trace_mm_compaction_defer_reset(zone, order); | |
192 | } | |
193 | ||
194 | /* Returns true if restarting compaction after many failures */ | |
195 | bool compaction_restarting(struct zone *zone, int order) | |
196 | { | |
197 | if (order < zone->compact_order_failed) | |
198 | return false; | |
199 | ||
200 | return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && | |
201 | zone->compact_considered >= 1UL << zone->compact_defer_shift; | |
202 | } | |
203 | ||
bb13ffeb MG |
204 | /* Returns true if the pageblock should be scanned for pages to isolate. */ |
205 | static inline bool isolation_suitable(struct compact_control *cc, | |
206 | struct page *page) | |
207 | { | |
208 | if (cc->ignore_skip_hint) | |
209 | return true; | |
210 | ||
211 | return !get_pageblock_skip(page); | |
212 | } | |
213 | ||
02333641 VB |
214 | static void reset_cached_positions(struct zone *zone) |
215 | { | |
216 | zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; | |
217 | zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; | |
623446e4 | 218 | zone->compact_cached_free_pfn = |
06b6640a | 219 | pageblock_start_pfn(zone_end_pfn(zone) - 1); |
02333641 VB |
220 | } |
221 | ||
21dc7e02 | 222 | /* |
b527cfe5 VB |
223 | * Compound pages of >= pageblock_order should consistenly be skipped until |
224 | * released. It is always pointless to compact pages of such order (if they are | |
225 | * migratable), and the pageblocks they occupy cannot contain any free pages. | |
21dc7e02 | 226 | */ |
b527cfe5 | 227 | static bool pageblock_skip_persistent(struct page *page) |
21dc7e02 | 228 | { |
b527cfe5 | 229 | if (!PageCompound(page)) |
21dc7e02 | 230 | return false; |
b527cfe5 VB |
231 | |
232 | page = compound_head(page); | |
233 | ||
234 | if (compound_order(page) >= pageblock_order) | |
235 | return true; | |
236 | ||
237 | return false; | |
21dc7e02 DR |
238 | } |
239 | ||
bb13ffeb MG |
240 | /* |
241 | * This function is called to clear all cached information on pageblocks that | |
242 | * should be skipped for page isolation when the migrate and free page scanner | |
243 | * meet. | |
244 | */ | |
62997027 | 245 | static void __reset_isolation_suitable(struct zone *zone) |
bb13ffeb MG |
246 | { |
247 | unsigned long start_pfn = zone->zone_start_pfn; | |
108bcc96 | 248 | unsigned long end_pfn = zone_end_pfn(zone); |
bb13ffeb MG |
249 | unsigned long pfn; |
250 | ||
62997027 | 251 | zone->compact_blockskip_flush = false; |
bb13ffeb MG |
252 | |
253 | /* Walk the zone and mark every pageblock as suitable for isolation */ | |
254 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
255 | struct page *page; | |
256 | ||
257 | cond_resched(); | |
258 | ||
ccbe1e4d MH |
259 | page = pfn_to_online_page(pfn); |
260 | if (!page) | |
bb13ffeb | 261 | continue; |
bb13ffeb MG |
262 | if (zone != page_zone(page)) |
263 | continue; | |
b527cfe5 | 264 | if (pageblock_skip_persistent(page)) |
21dc7e02 | 265 | continue; |
bb13ffeb MG |
266 | |
267 | clear_pageblock_skip(page); | |
268 | } | |
02333641 VB |
269 | |
270 | reset_cached_positions(zone); | |
bb13ffeb MG |
271 | } |
272 | ||
62997027 MG |
273 | void reset_isolation_suitable(pg_data_t *pgdat) |
274 | { | |
275 | int zoneid; | |
276 | ||
277 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { | |
278 | struct zone *zone = &pgdat->node_zones[zoneid]; | |
279 | if (!populated_zone(zone)) | |
280 | continue; | |
281 | ||
282 | /* Only flush if a full compaction finished recently */ | |
283 | if (zone->compact_blockskip_flush) | |
284 | __reset_isolation_suitable(zone); | |
285 | } | |
286 | } | |
287 | ||
e380bebe MG |
288 | /* |
289 | * Sets the pageblock skip bit if it was clear. Note that this is a hint as | |
290 | * locks are not required for read/writers. Returns true if it was already set. | |
291 | */ | |
292 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
293 | unsigned long pfn) | |
294 | { | |
295 | bool skip; | |
296 | ||
297 | /* Do no update if skip hint is being ignored */ | |
298 | if (cc->ignore_skip_hint) | |
299 | return false; | |
300 | ||
301 | if (!IS_ALIGNED(pfn, pageblock_nr_pages)) | |
302 | return false; | |
303 | ||
304 | skip = get_pageblock_skip(page); | |
305 | if (!skip && !cc->no_set_skip_hint) | |
306 | set_pageblock_skip(page); | |
307 | ||
308 | return skip; | |
309 | } | |
310 | ||
311 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
312 | { | |
313 | struct zone *zone = cc->zone; | |
314 | ||
315 | pfn = pageblock_end_pfn(pfn); | |
316 | ||
317 | /* Set for isolation rather than compaction */ | |
318 | if (cc->no_set_skip_hint) | |
319 | return; | |
320 | ||
321 | if (pfn > zone->compact_cached_migrate_pfn[0]) | |
322 | zone->compact_cached_migrate_pfn[0] = pfn; | |
323 | if (cc->mode != MIGRATE_ASYNC && | |
324 | pfn > zone->compact_cached_migrate_pfn[1]) | |
325 | zone->compact_cached_migrate_pfn[1] = pfn; | |
326 | } | |
327 | ||
bb13ffeb MG |
328 | /* |
329 | * If no pages were isolated then mark this pageblock to be skipped in the | |
62997027 | 330 | * future. The information is later cleared by __reset_isolation_suitable(). |
bb13ffeb | 331 | */ |
c89511ab | 332 | static void update_pageblock_skip(struct compact_control *cc, |
d097a6f6 | 333 | struct page *page, unsigned long pfn) |
bb13ffeb | 334 | { |
c89511ab | 335 | struct zone *zone = cc->zone; |
6815bf3f | 336 | |
2583d671 | 337 | if (cc->no_set_skip_hint) |
6815bf3f JK |
338 | return; |
339 | ||
bb13ffeb MG |
340 | if (!page) |
341 | return; | |
342 | ||
edc2ca61 | 343 | set_pageblock_skip(page); |
c89511ab | 344 | |
35979ef3 | 345 | /* Update where async and sync compaction should restart */ |
e380bebe MG |
346 | if (pfn < zone->compact_cached_free_pfn) |
347 | zone->compact_cached_free_pfn = pfn; | |
bb13ffeb MG |
348 | } |
349 | #else | |
350 | static inline bool isolation_suitable(struct compact_control *cc, | |
351 | struct page *page) | |
352 | { | |
353 | return true; | |
354 | } | |
355 | ||
b527cfe5 | 356 | static inline bool pageblock_skip_persistent(struct page *page) |
21dc7e02 DR |
357 | { |
358 | return false; | |
359 | } | |
360 | ||
361 | static inline void update_pageblock_skip(struct compact_control *cc, | |
d097a6f6 | 362 | struct page *page, unsigned long pfn) |
bb13ffeb MG |
363 | { |
364 | } | |
e380bebe MG |
365 | |
366 | static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) | |
367 | { | |
368 | } | |
369 | ||
370 | static bool test_and_set_skip(struct compact_control *cc, struct page *page, | |
371 | unsigned long pfn) | |
372 | { | |
373 | return false; | |
374 | } | |
bb13ffeb MG |
375 | #endif /* CONFIG_COMPACTION */ |
376 | ||
8b44d279 VB |
377 | /* |
378 | * Compaction requires the taking of some coarse locks that are potentially | |
cb2dcaf0 MG |
379 | * very heavily contended. For async compaction, trylock and record if the |
380 | * lock is contended. The lock will still be acquired but compaction will | |
381 | * abort when the current block is finished regardless of success rate. | |
382 | * Sync compaction acquires the lock. | |
8b44d279 | 383 | * |
cb2dcaf0 | 384 | * Always returns true which makes it easier to track lock state in callers. |
8b44d279 | 385 | */ |
cb2dcaf0 | 386 | static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, |
8b44d279 | 387 | struct compact_control *cc) |
2a1402aa | 388 | { |
cb2dcaf0 MG |
389 | /* Track if the lock is contended in async mode */ |
390 | if (cc->mode == MIGRATE_ASYNC && !cc->contended) { | |
391 | if (spin_trylock_irqsave(lock, *flags)) | |
392 | return true; | |
393 | ||
394 | cc->contended = true; | |
8b44d279 | 395 | } |
1f9efdef | 396 | |
cb2dcaf0 | 397 | spin_lock_irqsave(lock, *flags); |
8b44d279 | 398 | return true; |
2a1402aa MG |
399 | } |
400 | ||
c67fe375 MG |
401 | /* |
402 | * Compaction requires the taking of some coarse locks that are potentially | |
8b44d279 VB |
403 | * very heavily contended. The lock should be periodically unlocked to avoid |
404 | * having disabled IRQs for a long time, even when there is nobody waiting on | |
405 | * the lock. It might also be that allowing the IRQs will result in | |
406 | * need_resched() becoming true. If scheduling is needed, async compaction | |
407 | * aborts. Sync compaction schedules. | |
408 | * Either compaction type will also abort if a fatal signal is pending. | |
409 | * In either case if the lock was locked, it is dropped and not regained. | |
c67fe375 | 410 | * |
8b44d279 VB |
411 | * Returns true if compaction should abort due to fatal signal pending, or |
412 | * async compaction due to need_resched() | |
413 | * Returns false when compaction can continue (sync compaction might have | |
414 | * scheduled) | |
c67fe375 | 415 | */ |
8b44d279 VB |
416 | static bool compact_unlock_should_abort(spinlock_t *lock, |
417 | unsigned long flags, bool *locked, struct compact_control *cc) | |
c67fe375 | 418 | { |
8b44d279 VB |
419 | if (*locked) { |
420 | spin_unlock_irqrestore(lock, flags); | |
421 | *locked = false; | |
422 | } | |
1f9efdef | 423 | |
8b44d279 | 424 | if (fatal_signal_pending(current)) { |
c3486f53 | 425 | cc->contended = true; |
8b44d279 VB |
426 | return true; |
427 | } | |
c67fe375 | 428 | |
cf66f070 | 429 | cond_resched(); |
be976572 VB |
430 | |
431 | return false; | |
432 | } | |
433 | ||
85aa125f | 434 | /* |
9e4be470 JM |
435 | * Isolate free pages onto a private freelist. If @strict is true, will abort |
436 | * returning 0 on any invalid PFNs or non-free pages inside of the pageblock | |
437 | * (even though it may still end up isolating some pages). | |
85aa125f | 438 | */ |
f40d1e42 | 439 | static unsigned long isolate_freepages_block(struct compact_control *cc, |
e14c720e | 440 | unsigned long *start_pfn, |
85aa125f MN |
441 | unsigned long end_pfn, |
442 | struct list_head *freelist, | |
443 | bool strict) | |
748446bb | 444 | { |
b7aba698 | 445 | int nr_scanned = 0, total_isolated = 0; |
d097a6f6 | 446 | struct page *cursor; |
b8b2d825 | 447 | unsigned long flags = 0; |
f40d1e42 | 448 | bool locked = false; |
e14c720e | 449 | unsigned long blockpfn = *start_pfn; |
66c64223 | 450 | unsigned int order; |
748446bb | 451 | |
748446bb MG |
452 | cursor = pfn_to_page(blockpfn); |
453 | ||
f40d1e42 | 454 | /* Isolate free pages. */ |
748446bb | 455 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { |
66c64223 | 456 | int isolated; |
748446bb MG |
457 | struct page *page = cursor; |
458 | ||
8b44d279 VB |
459 | /* |
460 | * Periodically drop the lock (if held) regardless of its | |
461 | * contention, to give chance to IRQs. Abort if fatal signal | |
462 | * pending or async compaction detects need_resched() | |
463 | */ | |
464 | if (!(blockpfn % SWAP_CLUSTER_MAX) | |
465 | && compact_unlock_should_abort(&cc->zone->lock, flags, | |
466 | &locked, cc)) | |
467 | break; | |
468 | ||
b7aba698 | 469 | nr_scanned++; |
f40d1e42 | 470 | if (!pfn_valid_within(blockpfn)) |
2af120bc LA |
471 | goto isolate_fail; |
472 | ||
9fcd6d2e VB |
473 | /* |
474 | * For compound pages such as THP and hugetlbfs, we can save | |
475 | * potentially a lot of iterations if we skip them at once. | |
476 | * The check is racy, but we can consider only valid values | |
477 | * and the only danger is skipping too much. | |
478 | */ | |
479 | if (PageCompound(page)) { | |
21dc7e02 DR |
480 | const unsigned int order = compound_order(page); |
481 | ||
d3c85bad | 482 | if (likely(order < MAX_ORDER)) { |
21dc7e02 DR |
483 | blockpfn += (1UL << order) - 1; |
484 | cursor += (1UL << order) - 1; | |
9fcd6d2e | 485 | } |
9fcd6d2e VB |
486 | goto isolate_fail; |
487 | } | |
488 | ||
f40d1e42 | 489 | if (!PageBuddy(page)) |
2af120bc | 490 | goto isolate_fail; |
f40d1e42 MG |
491 | |
492 | /* | |
69b7189f VB |
493 | * If we already hold the lock, we can skip some rechecking. |
494 | * Note that if we hold the lock now, checked_pageblock was | |
495 | * already set in some previous iteration (or strict is true), | |
496 | * so it is correct to skip the suitable migration target | |
497 | * recheck as well. | |
f40d1e42 | 498 | */ |
69b7189f | 499 | if (!locked) { |
cb2dcaf0 | 500 | locked = compact_lock_irqsave(&cc->zone->lock, |
8b44d279 | 501 | &flags, cc); |
f40d1e42 | 502 | |
69b7189f VB |
503 | /* Recheck this is a buddy page under lock */ |
504 | if (!PageBuddy(page)) | |
505 | goto isolate_fail; | |
506 | } | |
748446bb | 507 | |
66c64223 JK |
508 | /* Found a free page, will break it into order-0 pages */ |
509 | order = page_order(page); | |
510 | isolated = __isolate_free_page(page, order); | |
a4f04f2c DR |
511 | if (!isolated) |
512 | break; | |
66c64223 | 513 | set_page_private(page, order); |
a4f04f2c | 514 | |
748446bb | 515 | total_isolated += isolated; |
a4f04f2c | 516 | cc->nr_freepages += isolated; |
66c64223 JK |
517 | list_add_tail(&page->lru, freelist); |
518 | ||
a4f04f2c DR |
519 | if (!strict && cc->nr_migratepages <= cc->nr_freepages) { |
520 | blockpfn += isolated; | |
521 | break; | |
748446bb | 522 | } |
a4f04f2c DR |
523 | /* Advance to the end of split page */ |
524 | blockpfn += isolated - 1; | |
525 | cursor += isolated - 1; | |
526 | continue; | |
2af120bc LA |
527 | |
528 | isolate_fail: | |
529 | if (strict) | |
530 | break; | |
531 | else | |
532 | continue; | |
533 | ||
748446bb MG |
534 | } |
535 | ||
a4f04f2c DR |
536 | if (locked) |
537 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
538 | ||
9fcd6d2e VB |
539 | /* |
540 | * There is a tiny chance that we have read bogus compound_order(), | |
541 | * so be careful to not go outside of the pageblock. | |
542 | */ | |
543 | if (unlikely(blockpfn > end_pfn)) | |
544 | blockpfn = end_pfn; | |
545 | ||
e34d85f0 JK |
546 | trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, |
547 | nr_scanned, total_isolated); | |
548 | ||
e14c720e VB |
549 | /* Record how far we have got within the block */ |
550 | *start_pfn = blockpfn; | |
551 | ||
f40d1e42 MG |
552 | /* |
553 | * If strict isolation is requested by CMA then check that all the | |
554 | * pages requested were isolated. If there were any failures, 0 is | |
555 | * returned and CMA will fail. | |
556 | */ | |
2af120bc | 557 | if (strict && blockpfn < end_pfn) |
f40d1e42 MG |
558 | total_isolated = 0; |
559 | ||
7f354a54 | 560 | cc->total_free_scanned += nr_scanned; |
397487db | 561 | if (total_isolated) |
010fc29a | 562 | count_compact_events(COMPACTISOLATED, total_isolated); |
748446bb MG |
563 | return total_isolated; |
564 | } | |
565 | ||
85aa125f MN |
566 | /** |
567 | * isolate_freepages_range() - isolate free pages. | |
e8b098fc | 568 | * @cc: Compaction control structure. |
85aa125f MN |
569 | * @start_pfn: The first PFN to start isolating. |
570 | * @end_pfn: The one-past-last PFN. | |
571 | * | |
572 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
573 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
574 | * undo its actions and return zero. | |
575 | * | |
576 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
577 | * (which may be greater then end_pfn if end fell in a middle of | |
578 | * a free page). | |
579 | */ | |
ff9543fd | 580 | unsigned long |
bb13ffeb MG |
581 | isolate_freepages_range(struct compact_control *cc, |
582 | unsigned long start_pfn, unsigned long end_pfn) | |
85aa125f | 583 | { |
e1409c32 | 584 | unsigned long isolated, pfn, block_start_pfn, block_end_pfn; |
85aa125f MN |
585 | LIST_HEAD(freelist); |
586 | ||
7d49d886 | 587 | pfn = start_pfn; |
06b6640a | 588 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
589 | if (block_start_pfn < cc->zone->zone_start_pfn) |
590 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 591 | block_end_pfn = pageblock_end_pfn(pfn); |
7d49d886 VB |
592 | |
593 | for (; pfn < end_pfn; pfn += isolated, | |
e1409c32 | 594 | block_start_pfn = block_end_pfn, |
7d49d886 | 595 | block_end_pfn += pageblock_nr_pages) { |
e14c720e VB |
596 | /* Protect pfn from changing by isolate_freepages_block */ |
597 | unsigned long isolate_start_pfn = pfn; | |
85aa125f | 598 | |
85aa125f MN |
599 | block_end_pfn = min(block_end_pfn, end_pfn); |
600 | ||
58420016 JK |
601 | /* |
602 | * pfn could pass the block_end_pfn if isolated freepage | |
603 | * is more than pageblock order. In this case, we adjust | |
604 | * scanning range to right one. | |
605 | */ | |
606 | if (pfn >= block_end_pfn) { | |
06b6640a VB |
607 | block_start_pfn = pageblock_start_pfn(pfn); |
608 | block_end_pfn = pageblock_end_pfn(pfn); | |
58420016 JK |
609 | block_end_pfn = min(block_end_pfn, end_pfn); |
610 | } | |
611 | ||
e1409c32 JK |
612 | if (!pageblock_pfn_to_page(block_start_pfn, |
613 | block_end_pfn, cc->zone)) | |
7d49d886 VB |
614 | break; |
615 | ||
e14c720e VB |
616 | isolated = isolate_freepages_block(cc, &isolate_start_pfn, |
617 | block_end_pfn, &freelist, true); | |
85aa125f MN |
618 | |
619 | /* | |
620 | * In strict mode, isolate_freepages_block() returns 0 if | |
621 | * there are any holes in the block (ie. invalid PFNs or | |
622 | * non-free pages). | |
623 | */ | |
624 | if (!isolated) | |
625 | break; | |
626 | ||
627 | /* | |
628 | * If we managed to isolate pages, it is always (1 << n) * | |
629 | * pageblock_nr_pages for some non-negative n. (Max order | |
630 | * page may span two pageblocks). | |
631 | */ | |
632 | } | |
633 | ||
66c64223 | 634 | /* __isolate_free_page() does not map the pages */ |
4469ab98 | 635 | split_map_pages(&freelist); |
85aa125f MN |
636 | |
637 | if (pfn < end_pfn) { | |
638 | /* Loop terminated early, cleanup. */ | |
639 | release_freepages(&freelist); | |
640 | return 0; | |
641 | } | |
642 | ||
643 | /* We don't use freelists for anything. */ | |
644 | return pfn; | |
645 | } | |
646 | ||
748446bb MG |
647 | /* Similar to reclaim, but different enough that they don't share logic */ |
648 | static bool too_many_isolated(struct zone *zone) | |
649 | { | |
bc693045 | 650 | unsigned long active, inactive, isolated; |
748446bb | 651 | |
599d0c95 MG |
652 | inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) + |
653 | node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON); | |
654 | active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) + | |
655 | node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON); | |
656 | isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) + | |
657 | node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON); | |
748446bb | 658 | |
bc693045 | 659 | return isolated > (inactive + active) / 2; |
748446bb MG |
660 | } |
661 | ||
2fe86e00 | 662 | /** |
edc2ca61 VB |
663 | * isolate_migratepages_block() - isolate all migrate-able pages within |
664 | * a single pageblock | |
2fe86e00 | 665 | * @cc: Compaction control structure. |
edc2ca61 VB |
666 | * @low_pfn: The first PFN to isolate |
667 | * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock | |
668 | * @isolate_mode: Isolation mode to be used. | |
2fe86e00 MN |
669 | * |
670 | * Isolate all pages that can be migrated from the range specified by | |
edc2ca61 VB |
671 | * [low_pfn, end_pfn). The range is expected to be within same pageblock. |
672 | * Returns zero if there is a fatal signal pending, otherwise PFN of the | |
673 | * first page that was not scanned (which may be both less, equal to or more | |
674 | * than end_pfn). | |
2fe86e00 | 675 | * |
edc2ca61 VB |
676 | * The pages are isolated on cc->migratepages list (not required to be empty), |
677 | * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field | |
678 | * is neither read nor updated. | |
748446bb | 679 | */ |
edc2ca61 VB |
680 | static unsigned long |
681 | isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, | |
682 | unsigned long end_pfn, isolate_mode_t isolate_mode) | |
748446bb | 683 | { |
edc2ca61 | 684 | struct zone *zone = cc->zone; |
b7aba698 | 685 | unsigned long nr_scanned = 0, nr_isolated = 0; |
fa9add64 | 686 | struct lruvec *lruvec; |
b8b2d825 | 687 | unsigned long flags = 0; |
2a1402aa | 688 | bool locked = false; |
bb13ffeb | 689 | struct page *page = NULL, *valid_page = NULL; |
e34d85f0 | 690 | unsigned long start_pfn = low_pfn; |
fdd048e1 VB |
691 | bool skip_on_failure = false; |
692 | unsigned long next_skip_pfn = 0; | |
e380bebe | 693 | bool skip_updated = false; |
748446bb | 694 | |
748446bb MG |
695 | /* |
696 | * Ensure that there are not too many pages isolated from the LRU | |
697 | * list by either parallel reclaimers or compaction. If there are, | |
698 | * delay for some time until fewer pages are isolated | |
699 | */ | |
700 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 701 | /* async migration should just abort */ |
e0b9daeb | 702 | if (cc->mode == MIGRATE_ASYNC) |
2fe86e00 | 703 | return 0; |
f9e35b3b | 704 | |
748446bb MG |
705 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
706 | ||
707 | if (fatal_signal_pending(current)) | |
2fe86e00 | 708 | return 0; |
748446bb MG |
709 | } |
710 | ||
cf66f070 | 711 | cond_resched(); |
aeef4b83 | 712 | |
fdd048e1 VB |
713 | if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { |
714 | skip_on_failure = true; | |
715 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
716 | } | |
717 | ||
748446bb | 718 | /* Time to isolate some pages for migration */ |
748446bb | 719 | for (; low_pfn < end_pfn; low_pfn++) { |
29c0dde8 | 720 | |
fdd048e1 VB |
721 | if (skip_on_failure && low_pfn >= next_skip_pfn) { |
722 | /* | |
723 | * We have isolated all migration candidates in the | |
724 | * previous order-aligned block, and did not skip it due | |
725 | * to failure. We should migrate the pages now and | |
726 | * hopefully succeed compaction. | |
727 | */ | |
728 | if (nr_isolated) | |
729 | break; | |
730 | ||
731 | /* | |
732 | * We failed to isolate in the previous order-aligned | |
733 | * block. Set the new boundary to the end of the | |
734 | * current block. Note we can't simply increase | |
735 | * next_skip_pfn by 1 << order, as low_pfn might have | |
736 | * been incremented by a higher number due to skipping | |
737 | * a compound or a high-order buddy page in the | |
738 | * previous loop iteration. | |
739 | */ | |
740 | next_skip_pfn = block_end_pfn(low_pfn, cc->order); | |
741 | } | |
742 | ||
8b44d279 VB |
743 | /* |
744 | * Periodically drop the lock (if held) regardless of its | |
745 | * contention, to give chance to IRQs. Abort async compaction | |
746 | * if contended. | |
747 | */ | |
748 | if (!(low_pfn % SWAP_CLUSTER_MAX) | |
a52633d8 | 749 | && compact_unlock_should_abort(zone_lru_lock(zone), flags, |
8b44d279 VB |
750 | &locked, cc)) |
751 | break; | |
c67fe375 | 752 | |
748446bb | 753 | if (!pfn_valid_within(low_pfn)) |
fdd048e1 | 754 | goto isolate_fail; |
b7aba698 | 755 | nr_scanned++; |
748446bb | 756 | |
748446bb | 757 | page = pfn_to_page(low_pfn); |
dc908600 | 758 | |
e380bebe MG |
759 | /* |
760 | * Check if the pageblock has already been marked skipped. | |
761 | * Only the aligned PFN is checked as the caller isolates | |
762 | * COMPACT_CLUSTER_MAX at a time so the second call must | |
763 | * not falsely conclude that the block should be skipped. | |
764 | */ | |
765 | if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) { | |
766 | if (!cc->ignore_skip_hint && get_pageblock_skip(page)) { | |
767 | low_pfn = end_pfn; | |
768 | goto isolate_abort; | |
769 | } | |
bb13ffeb | 770 | valid_page = page; |
e380bebe | 771 | } |
bb13ffeb | 772 | |
6c14466c | 773 | /* |
99c0fd5e VB |
774 | * Skip if free. We read page order here without zone lock |
775 | * which is generally unsafe, but the race window is small and | |
776 | * the worst thing that can happen is that we skip some | |
777 | * potential isolation targets. | |
6c14466c | 778 | */ |
99c0fd5e VB |
779 | if (PageBuddy(page)) { |
780 | unsigned long freepage_order = page_order_unsafe(page); | |
781 | ||
782 | /* | |
783 | * Without lock, we cannot be sure that what we got is | |
784 | * a valid page order. Consider only values in the | |
785 | * valid order range to prevent low_pfn overflow. | |
786 | */ | |
787 | if (freepage_order > 0 && freepage_order < MAX_ORDER) | |
788 | low_pfn += (1UL << freepage_order) - 1; | |
748446bb | 789 | continue; |
99c0fd5e | 790 | } |
748446bb | 791 | |
bc835011 | 792 | /* |
29c0dde8 VB |
793 | * Regardless of being on LRU, compound pages such as THP and |
794 | * hugetlbfs are not to be compacted. We can potentially save | |
795 | * a lot of iterations if we skip them at once. The check is | |
796 | * racy, but we can consider only valid values and the only | |
797 | * danger is skipping too much. | |
bc835011 | 798 | */ |
29c0dde8 | 799 | if (PageCompound(page)) { |
21dc7e02 | 800 | const unsigned int order = compound_order(page); |
edc2ca61 | 801 | |
d3c85bad | 802 | if (likely(order < MAX_ORDER)) |
21dc7e02 | 803 | low_pfn += (1UL << order) - 1; |
fdd048e1 | 804 | goto isolate_fail; |
2a1402aa MG |
805 | } |
806 | ||
bda807d4 MK |
807 | /* |
808 | * Check may be lockless but that's ok as we recheck later. | |
809 | * It's possible to migrate LRU and non-lru movable pages. | |
810 | * Skip any other type of page | |
811 | */ | |
812 | if (!PageLRU(page)) { | |
bda807d4 MK |
813 | /* |
814 | * __PageMovable can return false positive so we need | |
815 | * to verify it under page_lock. | |
816 | */ | |
817 | if (unlikely(__PageMovable(page)) && | |
818 | !PageIsolated(page)) { | |
819 | if (locked) { | |
a52633d8 | 820 | spin_unlock_irqrestore(zone_lru_lock(zone), |
bda807d4 MK |
821 | flags); |
822 | locked = false; | |
823 | } | |
824 | ||
9e5bcd61 | 825 | if (!isolate_movable_page(page, isolate_mode)) |
bda807d4 MK |
826 | goto isolate_success; |
827 | } | |
828 | ||
fdd048e1 | 829 | goto isolate_fail; |
bda807d4 | 830 | } |
29c0dde8 | 831 | |
119d6d59 DR |
832 | /* |
833 | * Migration will fail if an anonymous page is pinned in memory, | |
834 | * so avoid taking lru_lock and isolating it unnecessarily in an | |
835 | * admittedly racy check. | |
836 | */ | |
837 | if (!page_mapping(page) && | |
838 | page_count(page) > page_mapcount(page)) | |
fdd048e1 | 839 | goto isolate_fail; |
119d6d59 | 840 | |
73e64c51 MH |
841 | /* |
842 | * Only allow to migrate anonymous pages in GFP_NOFS context | |
843 | * because those do not depend on fs locks. | |
844 | */ | |
845 | if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page)) | |
846 | goto isolate_fail; | |
847 | ||
69b7189f VB |
848 | /* If we already hold the lock, we can skip some rechecking */ |
849 | if (!locked) { | |
cb2dcaf0 | 850 | locked = compact_lock_irqsave(zone_lru_lock(zone), |
8b44d279 | 851 | &flags, cc); |
e380bebe | 852 | |
e380bebe MG |
853 | /* Try get exclusive access under lock */ |
854 | if (!skip_updated) { | |
855 | skip_updated = true; | |
856 | if (test_and_set_skip(cc, page, low_pfn)) | |
857 | goto isolate_abort; | |
858 | } | |
2a1402aa | 859 | |
29c0dde8 | 860 | /* Recheck PageLRU and PageCompound under lock */ |
69b7189f | 861 | if (!PageLRU(page)) |
fdd048e1 | 862 | goto isolate_fail; |
29c0dde8 VB |
863 | |
864 | /* | |
865 | * Page become compound since the non-locked check, | |
866 | * and it's on LRU. It can only be a THP so the order | |
867 | * is safe to read and it's 0 for tail pages. | |
868 | */ | |
869 | if (unlikely(PageCompound(page))) { | |
d3c85bad | 870 | low_pfn += (1UL << compound_order(page)) - 1; |
fdd048e1 | 871 | goto isolate_fail; |
69b7189f | 872 | } |
bc835011 AA |
873 | } |
874 | ||
599d0c95 | 875 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
fa9add64 | 876 | |
748446bb | 877 | /* Try isolate the page */ |
edc2ca61 | 878 | if (__isolate_lru_page(page, isolate_mode) != 0) |
fdd048e1 | 879 | goto isolate_fail; |
748446bb | 880 | |
29c0dde8 | 881 | VM_BUG_ON_PAGE(PageCompound(page), page); |
bc835011 | 882 | |
748446bb | 883 | /* Successfully isolated */ |
fa9add64 | 884 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
6afcf8ef ML |
885 | inc_node_page_state(page, |
886 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
b6c75016 JK |
887 | |
888 | isolate_success: | |
fdd048e1 | 889 | list_add(&page->lru, &cc->migratepages); |
748446bb | 890 | cc->nr_migratepages++; |
b7aba698 | 891 | nr_isolated++; |
748446bb | 892 | |
804d3121 MG |
893 | /* |
894 | * Avoid isolating too much unless this block is being | |
cb2dcaf0 MG |
895 | * rescanned (e.g. dirty/writeback pages, parallel allocation) |
896 | * or a lock is contended. For contention, isolate quickly to | |
897 | * potentially remove one source of contention. | |
804d3121 | 898 | */ |
cb2dcaf0 MG |
899 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX && |
900 | !cc->rescan && !cc->contended) { | |
31b8384a | 901 | ++low_pfn; |
748446bb | 902 | break; |
31b8384a | 903 | } |
fdd048e1 VB |
904 | |
905 | continue; | |
906 | isolate_fail: | |
907 | if (!skip_on_failure) | |
908 | continue; | |
909 | ||
910 | /* | |
911 | * We have isolated some pages, but then failed. Release them | |
912 | * instead of migrating, as we cannot form the cc->order buddy | |
913 | * page anyway. | |
914 | */ | |
915 | if (nr_isolated) { | |
916 | if (locked) { | |
a52633d8 | 917 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
fdd048e1 VB |
918 | locked = false; |
919 | } | |
fdd048e1 VB |
920 | putback_movable_pages(&cc->migratepages); |
921 | cc->nr_migratepages = 0; | |
fdd048e1 VB |
922 | nr_isolated = 0; |
923 | } | |
924 | ||
925 | if (low_pfn < next_skip_pfn) { | |
926 | low_pfn = next_skip_pfn - 1; | |
927 | /* | |
928 | * The check near the loop beginning would have updated | |
929 | * next_skip_pfn too, but this is a bit simpler. | |
930 | */ | |
931 | next_skip_pfn += 1UL << cc->order; | |
932 | } | |
748446bb MG |
933 | } |
934 | ||
99c0fd5e VB |
935 | /* |
936 | * The PageBuddy() check could have potentially brought us outside | |
937 | * the range to be scanned. | |
938 | */ | |
939 | if (unlikely(low_pfn > end_pfn)) | |
940 | low_pfn = end_pfn; | |
941 | ||
e380bebe | 942 | isolate_abort: |
c67fe375 | 943 | if (locked) |
a52633d8 | 944 | spin_unlock_irqrestore(zone_lru_lock(zone), flags); |
748446bb | 945 | |
50b5b094 | 946 | /* |
804d3121 MG |
947 | * Updated the cached scanner pfn once the pageblock has been scanned |
948 | * Pages will either be migrated in which case there is no point | |
949 | * scanning in the near future or migration failed in which case the | |
950 | * failure reason may persist. The block is marked for skipping if | |
951 | * there were no pages isolated in the block or if the block is | |
952 | * rescanned twice in a row. | |
50b5b094 | 953 | */ |
804d3121 | 954 | if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) { |
e380bebe MG |
955 | if (valid_page && !skip_updated) |
956 | set_pageblock_skip(valid_page); | |
957 | update_cached_migrate(cc, low_pfn); | |
958 | } | |
bb13ffeb | 959 | |
e34d85f0 JK |
960 | trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, |
961 | nr_scanned, nr_isolated); | |
b7aba698 | 962 | |
7f354a54 | 963 | cc->total_migrate_scanned += nr_scanned; |
397487db | 964 | if (nr_isolated) |
010fc29a | 965 | count_compact_events(COMPACTISOLATED, nr_isolated); |
397487db | 966 | |
2fe86e00 MN |
967 | return low_pfn; |
968 | } | |
969 | ||
edc2ca61 VB |
970 | /** |
971 | * isolate_migratepages_range() - isolate migrate-able pages in a PFN range | |
972 | * @cc: Compaction control structure. | |
973 | * @start_pfn: The first PFN to start isolating. | |
974 | * @end_pfn: The one-past-last PFN. | |
975 | * | |
976 | * Returns zero if isolation fails fatally due to e.g. pending signal. | |
977 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
978 | * (which may be greater than end_pfn if end fell in a middle of a THP page). | |
979 | */ | |
980 | unsigned long | |
981 | isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, | |
982 | unsigned long end_pfn) | |
983 | { | |
e1409c32 | 984 | unsigned long pfn, block_start_pfn, block_end_pfn; |
edc2ca61 VB |
985 | |
986 | /* Scan block by block. First and last block may be incomplete */ | |
987 | pfn = start_pfn; | |
06b6640a | 988 | block_start_pfn = pageblock_start_pfn(pfn); |
e1409c32 JK |
989 | if (block_start_pfn < cc->zone->zone_start_pfn) |
990 | block_start_pfn = cc->zone->zone_start_pfn; | |
06b6640a | 991 | block_end_pfn = pageblock_end_pfn(pfn); |
edc2ca61 VB |
992 | |
993 | for (; pfn < end_pfn; pfn = block_end_pfn, | |
e1409c32 | 994 | block_start_pfn = block_end_pfn, |
edc2ca61 VB |
995 | block_end_pfn += pageblock_nr_pages) { |
996 | ||
997 | block_end_pfn = min(block_end_pfn, end_pfn); | |
998 | ||
e1409c32 JK |
999 | if (!pageblock_pfn_to_page(block_start_pfn, |
1000 | block_end_pfn, cc->zone)) | |
edc2ca61 VB |
1001 | continue; |
1002 | ||
1003 | pfn = isolate_migratepages_block(cc, pfn, block_end_pfn, | |
1004 | ISOLATE_UNEVICTABLE); | |
1005 | ||
14af4a5e | 1006 | if (!pfn) |
edc2ca61 | 1007 | break; |
6ea41c0c JK |
1008 | |
1009 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) | |
1010 | break; | |
edc2ca61 | 1011 | } |
edc2ca61 VB |
1012 | |
1013 | return pfn; | |
1014 | } | |
1015 | ||
ff9543fd MN |
1016 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
1017 | #ifdef CONFIG_COMPACTION | |
018e9a49 | 1018 | |
b682debd VB |
1019 | static bool suitable_migration_source(struct compact_control *cc, |
1020 | struct page *page) | |
1021 | { | |
282722b0 VB |
1022 | int block_mt; |
1023 | ||
9bebefd5 MG |
1024 | if (pageblock_skip_persistent(page)) |
1025 | return false; | |
1026 | ||
282722b0 | 1027 | if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) |
b682debd VB |
1028 | return true; |
1029 | ||
282722b0 VB |
1030 | block_mt = get_pageblock_migratetype(page); |
1031 | ||
1032 | if (cc->migratetype == MIGRATE_MOVABLE) | |
1033 | return is_migrate_movable(block_mt); | |
1034 | else | |
1035 | return block_mt == cc->migratetype; | |
b682debd VB |
1036 | } |
1037 | ||
018e9a49 | 1038 | /* Returns true if the page is within a block suitable for migration to */ |
9f7e3387 VB |
1039 | static bool suitable_migration_target(struct compact_control *cc, |
1040 | struct page *page) | |
018e9a49 AM |
1041 | { |
1042 | /* If the page is a large free page, then disallow migration */ | |
1043 | if (PageBuddy(page)) { | |
1044 | /* | |
1045 | * We are checking page_order without zone->lock taken. But | |
1046 | * the only small danger is that we skip a potentially suitable | |
1047 | * pageblock, so it's not worth to check order for valid range. | |
1048 | */ | |
1049 | if (page_order_unsafe(page) >= pageblock_order) | |
1050 | return false; | |
1051 | } | |
1052 | ||
1ef36db2 YX |
1053 | if (cc->ignore_block_suitable) |
1054 | return true; | |
1055 | ||
018e9a49 | 1056 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
b682debd | 1057 | if (is_migrate_movable(get_pageblock_migratetype(page))) |
018e9a49 AM |
1058 | return true; |
1059 | ||
1060 | /* Otherwise skip the block */ | |
1061 | return false; | |
1062 | } | |
1063 | ||
70b44595 MG |
1064 | static inline unsigned int |
1065 | freelist_scan_limit(struct compact_control *cc) | |
1066 | { | |
1067 | return (COMPACT_CLUSTER_MAX >> cc->fast_search_fail) + 1; | |
1068 | } | |
1069 | ||
f2849aa0 VB |
1070 | /* |
1071 | * Test whether the free scanner has reached the same or lower pageblock than | |
1072 | * the migration scanner, and compaction should thus terminate. | |
1073 | */ | |
1074 | static inline bool compact_scanners_met(struct compact_control *cc) | |
1075 | { | |
1076 | return (cc->free_pfn >> pageblock_order) | |
1077 | <= (cc->migrate_pfn >> pageblock_order); | |
1078 | } | |
1079 | ||
5a811889 MG |
1080 | /* |
1081 | * Used when scanning for a suitable migration target which scans freelists | |
1082 | * in reverse. Reorders the list such as the unscanned pages are scanned | |
1083 | * first on the next iteration of the free scanner | |
1084 | */ | |
1085 | static void | |
1086 | move_freelist_head(struct list_head *freelist, struct page *freepage) | |
1087 | { | |
1088 | LIST_HEAD(sublist); | |
1089 | ||
1090 | if (!list_is_last(freelist, &freepage->lru)) { | |
1091 | list_cut_before(&sublist, freelist, &freepage->lru); | |
1092 | if (!list_empty(&sublist)) | |
1093 | list_splice_tail(&sublist, freelist); | |
1094 | } | |
1095 | } | |
1096 | ||
1097 | /* | |
1098 | * Similar to move_freelist_head except used by the migration scanner | |
1099 | * when scanning forward. It's possible for these list operations to | |
1100 | * move against each other if they search the free list exactly in | |
1101 | * lockstep. | |
1102 | */ | |
70b44595 MG |
1103 | static void |
1104 | move_freelist_tail(struct list_head *freelist, struct page *freepage) | |
1105 | { | |
1106 | LIST_HEAD(sublist); | |
1107 | ||
1108 | if (!list_is_first(freelist, &freepage->lru)) { | |
1109 | list_cut_position(&sublist, freelist, &freepage->lru); | |
1110 | if (!list_empty(&sublist)) | |
1111 | list_splice_tail(&sublist, freelist); | |
1112 | } | |
1113 | } | |
1114 | ||
5a811889 MG |
1115 | static void |
1116 | fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated) | |
1117 | { | |
1118 | unsigned long start_pfn, end_pfn; | |
1119 | struct page *page = pfn_to_page(pfn); | |
1120 | ||
1121 | /* Do not search around if there are enough pages already */ | |
1122 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1123 | return; | |
1124 | ||
1125 | /* Minimise scanning during async compaction */ | |
1126 | if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) | |
1127 | return; | |
1128 | ||
1129 | /* Pageblock boundaries */ | |
1130 | start_pfn = pageblock_start_pfn(pfn); | |
1131 | end_pfn = min(start_pfn + pageblock_nr_pages, zone_end_pfn(cc->zone)); | |
1132 | ||
1133 | /* Scan before */ | |
1134 | if (start_pfn != pfn) { | |
1135 | isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, false); | |
1136 | if (cc->nr_freepages >= cc->nr_migratepages) | |
1137 | return; | |
1138 | } | |
1139 | ||
1140 | /* Scan after */ | |
1141 | start_pfn = pfn + nr_isolated; | |
1142 | if (start_pfn != end_pfn) | |
1143 | isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, false); | |
1144 | ||
1145 | /* Skip this pageblock in the future as it's full or nearly full */ | |
1146 | if (cc->nr_freepages < cc->nr_migratepages) | |
1147 | set_pageblock_skip(page); | |
1148 | } | |
1149 | ||
1150 | static unsigned long | |
1151 | fast_isolate_freepages(struct compact_control *cc) | |
1152 | { | |
1153 | unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1); | |
1154 | unsigned int nr_scanned = 0; | |
1155 | unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0; | |
1156 | unsigned long nr_isolated = 0; | |
1157 | unsigned long distance; | |
1158 | struct page *page = NULL; | |
1159 | bool scan_start = false; | |
1160 | int order; | |
1161 | ||
1162 | /* Full compaction passes in a negative order */ | |
1163 | if (cc->order <= 0) | |
1164 | return cc->free_pfn; | |
1165 | ||
1166 | /* | |
1167 | * If starting the scan, use a deeper search and use the highest | |
1168 | * PFN found if a suitable one is not found. | |
1169 | */ | |
1170 | if (cc->free_pfn == pageblock_start_pfn(zone_end_pfn(cc->zone) - 1)) { | |
1171 | limit = pageblock_nr_pages >> 1; | |
1172 | scan_start = true; | |
1173 | } | |
1174 | ||
1175 | /* | |
1176 | * Preferred point is in the top quarter of the scan space but take | |
1177 | * a pfn from the top half if the search is problematic. | |
1178 | */ | |
1179 | distance = (cc->free_pfn - cc->migrate_pfn); | |
1180 | low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); | |
1181 | min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); | |
1182 | ||
1183 | if (WARN_ON_ONCE(min_pfn > low_pfn)) | |
1184 | low_pfn = min_pfn; | |
1185 | ||
1186 | for (order = cc->order - 1; | |
1187 | order >= 0 && !page; | |
1188 | order--) { | |
1189 | struct free_area *area = &cc->zone->free_area[order]; | |
1190 | struct list_head *freelist; | |
1191 | struct page *freepage; | |
1192 | unsigned long flags; | |
1193 | unsigned int order_scanned = 0; | |
1194 | ||
1195 | if (!area->nr_free) | |
1196 | continue; | |
1197 | ||
1198 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1199 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1200 | list_for_each_entry_reverse(freepage, freelist, lru) { | |
1201 | unsigned long pfn; | |
1202 | ||
1203 | order_scanned++; | |
1204 | nr_scanned++; | |
1205 | pfn = page_to_pfn(freepage); | |
1206 | ||
1207 | if (pfn >= highest) | |
1208 | highest = pageblock_start_pfn(pfn); | |
1209 | ||
1210 | if (pfn >= low_pfn) { | |
1211 | cc->fast_search_fail = 0; | |
1212 | page = freepage; | |
1213 | break; | |
1214 | } | |
1215 | ||
1216 | if (pfn >= min_pfn && pfn > high_pfn) { | |
1217 | high_pfn = pfn; | |
1218 | ||
1219 | /* Shorten the scan if a candidate is found */ | |
1220 | limit >>= 1; | |
1221 | } | |
1222 | ||
1223 | if (order_scanned >= limit) | |
1224 | break; | |
1225 | } | |
1226 | ||
1227 | /* Use a minimum pfn if a preferred one was not found */ | |
1228 | if (!page && high_pfn) { | |
1229 | page = pfn_to_page(high_pfn); | |
1230 | ||
1231 | /* Update freepage for the list reorder below */ | |
1232 | freepage = page; | |
1233 | } | |
1234 | ||
1235 | /* Reorder to so a future search skips recent pages */ | |
1236 | move_freelist_head(freelist, freepage); | |
1237 | ||
1238 | /* Isolate the page if available */ | |
1239 | if (page) { | |
1240 | if (__isolate_free_page(page, order)) { | |
1241 | set_page_private(page, order); | |
1242 | nr_isolated = 1 << order; | |
1243 | cc->nr_freepages += nr_isolated; | |
1244 | list_add_tail(&page->lru, &cc->freepages); | |
1245 | count_compact_events(COMPACTISOLATED, nr_isolated); | |
1246 | } else { | |
1247 | /* If isolation fails, abort the search */ | |
1248 | order = -1; | |
1249 | page = NULL; | |
1250 | } | |
1251 | } | |
1252 | ||
1253 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1254 | ||
1255 | /* | |
1256 | * Smaller scan on next order so the total scan ig related | |
1257 | * to freelist_scan_limit. | |
1258 | */ | |
1259 | if (order_scanned >= limit) | |
1260 | limit = min(1U, limit >> 1); | |
1261 | } | |
1262 | ||
1263 | if (!page) { | |
1264 | cc->fast_search_fail++; | |
1265 | if (scan_start) { | |
1266 | /* | |
1267 | * Use the highest PFN found above min. If one was | |
1268 | * not found, be pessemistic for direct compaction | |
1269 | * and use the min mark. | |
1270 | */ | |
1271 | if (highest) { | |
1272 | page = pfn_to_page(highest); | |
1273 | cc->free_pfn = highest; | |
1274 | } else { | |
1275 | if (cc->direct_compaction) { | |
1276 | page = pfn_to_page(min_pfn); | |
1277 | cc->free_pfn = min_pfn; | |
1278 | } | |
1279 | } | |
1280 | } | |
1281 | } | |
1282 | ||
d097a6f6 MG |
1283 | if (highest && highest >= cc->zone->compact_cached_free_pfn) { |
1284 | highest -= pageblock_nr_pages; | |
5a811889 | 1285 | cc->zone->compact_cached_free_pfn = highest; |
d097a6f6 | 1286 | } |
5a811889 MG |
1287 | |
1288 | cc->total_free_scanned += nr_scanned; | |
1289 | if (!page) | |
1290 | return cc->free_pfn; | |
1291 | ||
1292 | low_pfn = page_to_pfn(page); | |
1293 | fast_isolate_around(cc, low_pfn, nr_isolated); | |
1294 | return low_pfn; | |
1295 | } | |
1296 | ||
2fe86e00 | 1297 | /* |
ff9543fd MN |
1298 | * Based on information in the current compact_control, find blocks |
1299 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 1300 | */ |
edc2ca61 | 1301 | static void isolate_freepages(struct compact_control *cc) |
2fe86e00 | 1302 | { |
edc2ca61 | 1303 | struct zone *zone = cc->zone; |
ff9543fd | 1304 | struct page *page; |
c96b9e50 | 1305 | unsigned long block_start_pfn; /* start of current pageblock */ |
e14c720e | 1306 | unsigned long isolate_start_pfn; /* exact pfn we start at */ |
c96b9e50 VB |
1307 | unsigned long block_end_pfn; /* end of current pageblock */ |
1308 | unsigned long low_pfn; /* lowest pfn scanner is able to scan */ | |
ff9543fd | 1309 | struct list_head *freelist = &cc->freepages; |
2fe86e00 | 1310 | |
5a811889 MG |
1311 | /* Try a small search of the free lists for a candidate */ |
1312 | isolate_start_pfn = fast_isolate_freepages(cc); | |
1313 | if (cc->nr_freepages) | |
1314 | goto splitmap; | |
1315 | ||
ff9543fd MN |
1316 | /* |
1317 | * Initialise the free scanner. The starting point is where we last | |
49e068f0 | 1318 | * successfully isolated from, zone-cached value, or the end of the |
e14c720e VB |
1319 | * zone when isolating for the first time. For looping we also need |
1320 | * this pfn aligned down to the pageblock boundary, because we do | |
c96b9e50 VB |
1321 | * block_start_pfn -= pageblock_nr_pages in the for loop. |
1322 | * For ending point, take care when isolating in last pageblock of a | |
1323 | * a zone which ends in the middle of a pageblock. | |
49e068f0 VB |
1324 | * The low boundary is the end of the pageblock the migration scanner |
1325 | * is using. | |
ff9543fd | 1326 | */ |
e14c720e | 1327 | isolate_start_pfn = cc->free_pfn; |
5a811889 | 1328 | block_start_pfn = pageblock_start_pfn(isolate_start_pfn); |
c96b9e50 VB |
1329 | block_end_pfn = min(block_start_pfn + pageblock_nr_pages, |
1330 | zone_end_pfn(zone)); | |
06b6640a | 1331 | low_pfn = pageblock_end_pfn(cc->migrate_pfn); |
2fe86e00 | 1332 | |
ff9543fd MN |
1333 | /* |
1334 | * Isolate free pages until enough are available to migrate the | |
1335 | * pages on cc->migratepages. We stop searching if the migrate | |
1336 | * and free page scanners meet or enough free pages are isolated. | |
1337 | */ | |
f5f61a32 | 1338 | for (; block_start_pfn >= low_pfn; |
c96b9e50 | 1339 | block_end_pfn = block_start_pfn, |
e14c720e VB |
1340 | block_start_pfn -= pageblock_nr_pages, |
1341 | isolate_start_pfn = block_start_pfn) { | |
f6ea3adb DR |
1342 | /* |
1343 | * This can iterate a massively long zone without finding any | |
cb810ad2 | 1344 | * suitable migration targets, so periodically check resched. |
f6ea3adb | 1345 | */ |
cb810ad2 | 1346 | if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1347 | cond_resched(); |
f6ea3adb | 1348 | |
7d49d886 VB |
1349 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1350 | zone); | |
1351 | if (!page) | |
ff9543fd MN |
1352 | continue; |
1353 | ||
1354 | /* Check the block is suitable for migration */ | |
9f7e3387 | 1355 | if (!suitable_migration_target(cc, page)) |
ff9543fd | 1356 | continue; |
68e3e926 | 1357 | |
bb13ffeb MG |
1358 | /* If isolation recently failed, do not retry */ |
1359 | if (!isolation_suitable(cc, page)) | |
1360 | continue; | |
1361 | ||
e14c720e | 1362 | /* Found a block suitable for isolating free pages from. */ |
a46cbf3b DR |
1363 | isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn, |
1364 | freelist, false); | |
ff9543fd | 1365 | |
d097a6f6 MG |
1366 | /* Update the skip hint if the full pageblock was scanned */ |
1367 | if (isolate_start_pfn == block_end_pfn) | |
1368 | update_pageblock_skip(cc, page, block_start_pfn); | |
1369 | ||
cb2dcaf0 MG |
1370 | /* Are enough freepages isolated? */ |
1371 | if (cc->nr_freepages >= cc->nr_migratepages) { | |
a46cbf3b DR |
1372 | if (isolate_start_pfn >= block_end_pfn) { |
1373 | /* | |
1374 | * Restart at previous pageblock if more | |
1375 | * freepages can be isolated next time. | |
1376 | */ | |
f5f61a32 VB |
1377 | isolate_start_pfn = |
1378 | block_start_pfn - pageblock_nr_pages; | |
a46cbf3b | 1379 | } |
be976572 | 1380 | break; |
a46cbf3b | 1381 | } else if (isolate_start_pfn < block_end_pfn) { |
f5f61a32 | 1382 | /* |
a46cbf3b DR |
1383 | * If isolation failed early, do not continue |
1384 | * needlessly. | |
f5f61a32 | 1385 | */ |
a46cbf3b | 1386 | break; |
f5f61a32 | 1387 | } |
ff9543fd MN |
1388 | } |
1389 | ||
7ed695e0 | 1390 | /* |
f5f61a32 VB |
1391 | * Record where the free scanner will restart next time. Either we |
1392 | * broke from the loop and set isolate_start_pfn based on the last | |
1393 | * call to isolate_freepages_block(), or we met the migration scanner | |
1394 | * and the loop terminated due to isolate_start_pfn < low_pfn | |
7ed695e0 | 1395 | */ |
f5f61a32 | 1396 | cc->free_pfn = isolate_start_pfn; |
5a811889 MG |
1397 | |
1398 | splitmap: | |
1399 | /* __isolate_free_page() does not map the pages */ | |
1400 | split_map_pages(freelist); | |
748446bb MG |
1401 | } |
1402 | ||
1403 | /* | |
1404 | * This is a migrate-callback that "allocates" freepages by taking pages | |
1405 | * from the isolated freelists in the block we are migrating to. | |
1406 | */ | |
1407 | static struct page *compaction_alloc(struct page *migratepage, | |
666feb21 | 1408 | unsigned long data) |
748446bb MG |
1409 | { |
1410 | struct compact_control *cc = (struct compact_control *)data; | |
1411 | struct page *freepage; | |
1412 | ||
748446bb | 1413 | if (list_empty(&cc->freepages)) { |
cb2dcaf0 | 1414 | isolate_freepages(cc); |
748446bb MG |
1415 | |
1416 | if (list_empty(&cc->freepages)) | |
1417 | return NULL; | |
1418 | } | |
1419 | ||
1420 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
1421 | list_del(&freepage->lru); | |
1422 | cc->nr_freepages--; | |
1423 | ||
1424 | return freepage; | |
1425 | } | |
1426 | ||
1427 | /* | |
d53aea3d DR |
1428 | * This is a migrate-callback that "frees" freepages back to the isolated |
1429 | * freelist. All pages on the freelist are from the same zone, so there is no | |
1430 | * special handling needed for NUMA. | |
1431 | */ | |
1432 | static void compaction_free(struct page *page, unsigned long data) | |
1433 | { | |
1434 | struct compact_control *cc = (struct compact_control *)data; | |
1435 | ||
1436 | list_add(&page->lru, &cc->freepages); | |
1437 | cc->nr_freepages++; | |
1438 | } | |
1439 | ||
ff9543fd MN |
1440 | /* possible outcome of isolate_migratepages */ |
1441 | typedef enum { | |
1442 | ISOLATE_ABORT, /* Abort compaction now */ | |
1443 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
1444 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
1445 | } isolate_migrate_t; | |
1446 | ||
5bbe3547 EM |
1447 | /* |
1448 | * Allow userspace to control policy on scanning the unevictable LRU for | |
1449 | * compactable pages. | |
1450 | */ | |
1451 | int sysctl_compact_unevictable_allowed __read_mostly = 1; | |
1452 | ||
70b44595 MG |
1453 | static inline void |
1454 | update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) | |
1455 | { | |
1456 | if (cc->fast_start_pfn == ULONG_MAX) | |
1457 | return; | |
1458 | ||
1459 | if (!cc->fast_start_pfn) | |
1460 | cc->fast_start_pfn = pfn; | |
1461 | ||
1462 | cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); | |
1463 | } | |
1464 | ||
1465 | static inline unsigned long | |
1466 | reinit_migrate_pfn(struct compact_control *cc) | |
1467 | { | |
1468 | if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) | |
1469 | return cc->migrate_pfn; | |
1470 | ||
1471 | cc->migrate_pfn = cc->fast_start_pfn; | |
1472 | cc->fast_start_pfn = ULONG_MAX; | |
1473 | ||
1474 | return cc->migrate_pfn; | |
1475 | } | |
1476 | ||
1477 | /* | |
1478 | * Briefly search the free lists for a migration source that already has | |
1479 | * some free pages to reduce the number of pages that need migration | |
1480 | * before a pageblock is free. | |
1481 | */ | |
1482 | static unsigned long fast_find_migrateblock(struct compact_control *cc) | |
1483 | { | |
1484 | unsigned int limit = freelist_scan_limit(cc); | |
1485 | unsigned int nr_scanned = 0; | |
1486 | unsigned long distance; | |
1487 | unsigned long pfn = cc->migrate_pfn; | |
1488 | unsigned long high_pfn; | |
1489 | int order; | |
1490 | ||
1491 | /* Skip hints are relied on to avoid repeats on the fast search */ | |
1492 | if (cc->ignore_skip_hint) | |
1493 | return pfn; | |
1494 | ||
1495 | /* | |
1496 | * If the migrate_pfn is not at the start of a zone or the start | |
1497 | * of a pageblock then assume this is a continuation of a previous | |
1498 | * scan restarted due to COMPACT_CLUSTER_MAX. | |
1499 | */ | |
1500 | if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) | |
1501 | return pfn; | |
1502 | ||
1503 | /* | |
1504 | * For smaller orders, just linearly scan as the number of pages | |
1505 | * to migrate should be relatively small and does not necessarily | |
1506 | * justify freeing up a large block for a small allocation. | |
1507 | */ | |
1508 | if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1509 | return pfn; | |
1510 | ||
1511 | /* | |
1512 | * Only allow kcompactd and direct requests for movable pages to | |
1513 | * quickly clear out a MOVABLE pageblock for allocation. This | |
1514 | * reduces the risk that a large movable pageblock is freed for | |
1515 | * an unmovable/reclaimable small allocation. | |
1516 | */ | |
1517 | if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) | |
1518 | return pfn; | |
1519 | ||
1520 | /* | |
1521 | * When starting the migration scanner, pick any pageblock within the | |
1522 | * first half of the search space. Otherwise try and pick a pageblock | |
1523 | * within the first eighth to reduce the chances that a migration | |
1524 | * target later becomes a source. | |
1525 | */ | |
1526 | distance = (cc->free_pfn - cc->migrate_pfn) >> 1; | |
1527 | if (cc->migrate_pfn != cc->zone->zone_start_pfn) | |
1528 | distance >>= 2; | |
1529 | high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); | |
1530 | ||
1531 | for (order = cc->order - 1; | |
1532 | order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; | |
1533 | order--) { | |
1534 | struct free_area *area = &cc->zone->free_area[order]; | |
1535 | struct list_head *freelist; | |
1536 | unsigned long flags; | |
1537 | struct page *freepage; | |
1538 | ||
1539 | if (!area->nr_free) | |
1540 | continue; | |
1541 | ||
1542 | spin_lock_irqsave(&cc->zone->lock, flags); | |
1543 | freelist = &area->free_list[MIGRATE_MOVABLE]; | |
1544 | list_for_each_entry(freepage, freelist, lru) { | |
1545 | unsigned long free_pfn; | |
1546 | ||
1547 | nr_scanned++; | |
1548 | free_pfn = page_to_pfn(freepage); | |
1549 | if (free_pfn < high_pfn) { | |
70b44595 MG |
1550 | /* |
1551 | * Avoid if skipped recently. Ideally it would | |
1552 | * move to the tail but even safe iteration of | |
1553 | * the list assumes an entry is deleted, not | |
1554 | * reordered. | |
1555 | */ | |
1556 | if (get_pageblock_skip(freepage)) { | |
1557 | if (list_is_last(freelist, &freepage->lru)) | |
1558 | break; | |
1559 | ||
1560 | continue; | |
1561 | } | |
1562 | ||
1563 | /* Reorder to so a future search skips recent pages */ | |
1564 | move_freelist_tail(freelist, freepage); | |
1565 | ||
e380bebe | 1566 | update_fast_start_pfn(cc, free_pfn); |
70b44595 MG |
1567 | pfn = pageblock_start_pfn(free_pfn); |
1568 | cc->fast_search_fail = 0; | |
1569 | set_pageblock_skip(freepage); | |
1570 | break; | |
1571 | } | |
1572 | ||
1573 | if (nr_scanned >= limit) { | |
1574 | cc->fast_search_fail++; | |
1575 | move_freelist_tail(freelist, freepage); | |
1576 | break; | |
1577 | } | |
1578 | } | |
1579 | spin_unlock_irqrestore(&cc->zone->lock, flags); | |
1580 | } | |
1581 | ||
1582 | cc->total_migrate_scanned += nr_scanned; | |
1583 | ||
1584 | /* | |
1585 | * If fast scanning failed then use a cached entry for a page block | |
1586 | * that had free pages as the basis for starting a linear scan. | |
1587 | */ | |
1588 | if (pfn == cc->migrate_pfn) | |
1589 | pfn = reinit_migrate_pfn(cc); | |
1590 | ||
1591 | return pfn; | |
1592 | } | |
1593 | ||
ff9543fd | 1594 | /* |
edc2ca61 VB |
1595 | * Isolate all pages that can be migrated from the first suitable block, |
1596 | * starting at the block pointed to by the migrate scanner pfn within | |
1597 | * compact_control. | |
ff9543fd MN |
1598 | */ |
1599 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
1600 | struct compact_control *cc) | |
1601 | { | |
e1409c32 JK |
1602 | unsigned long block_start_pfn; |
1603 | unsigned long block_end_pfn; | |
1604 | unsigned long low_pfn; | |
edc2ca61 VB |
1605 | struct page *page; |
1606 | const isolate_mode_t isolate_mode = | |
5bbe3547 | 1607 | (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | |
1d2047fe | 1608 | (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); |
70b44595 | 1609 | bool fast_find_block; |
ff9543fd | 1610 | |
edc2ca61 VB |
1611 | /* |
1612 | * Start at where we last stopped, or beginning of the zone as | |
70b44595 MG |
1613 | * initialized by compact_zone(). The first failure will use |
1614 | * the lowest PFN as the starting point for linear scanning. | |
edc2ca61 | 1615 | */ |
70b44595 | 1616 | low_pfn = fast_find_migrateblock(cc); |
06b6640a | 1617 | block_start_pfn = pageblock_start_pfn(low_pfn); |
e1409c32 JK |
1618 | if (block_start_pfn < zone->zone_start_pfn) |
1619 | block_start_pfn = zone->zone_start_pfn; | |
ff9543fd | 1620 | |
70b44595 MG |
1621 | /* |
1622 | * fast_find_migrateblock marks a pageblock skipped so to avoid | |
1623 | * the isolation_suitable check below, check whether the fast | |
1624 | * search was successful. | |
1625 | */ | |
1626 | fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; | |
1627 | ||
ff9543fd | 1628 | /* Only scan within a pageblock boundary */ |
06b6640a | 1629 | block_end_pfn = pageblock_end_pfn(low_pfn); |
ff9543fd | 1630 | |
edc2ca61 VB |
1631 | /* |
1632 | * Iterate over whole pageblocks until we find the first suitable. | |
1633 | * Do not cross the free scanner. | |
1634 | */ | |
e1409c32 | 1635 | for (; block_end_pfn <= cc->free_pfn; |
70b44595 | 1636 | fast_find_block = false, |
e1409c32 JK |
1637 | low_pfn = block_end_pfn, |
1638 | block_start_pfn = block_end_pfn, | |
1639 | block_end_pfn += pageblock_nr_pages) { | |
ff9543fd | 1640 | |
edc2ca61 VB |
1641 | /* |
1642 | * This can potentially iterate a massively long zone with | |
1643 | * many pageblocks unsuitable, so periodically check if we | |
cb810ad2 | 1644 | * need to schedule. |
edc2ca61 | 1645 | */ |
cb810ad2 | 1646 | if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))) |
cf66f070 | 1647 | cond_resched(); |
ff9543fd | 1648 | |
e1409c32 JK |
1649 | page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, |
1650 | zone); | |
7d49d886 | 1651 | if (!page) |
edc2ca61 VB |
1652 | continue; |
1653 | ||
e380bebe MG |
1654 | /* |
1655 | * If isolation recently failed, do not retry. Only check the | |
1656 | * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock | |
1657 | * to be visited multiple times. Assume skip was checked | |
1658 | * before making it "skip" so other compaction instances do | |
1659 | * not scan the same block. | |
1660 | */ | |
1661 | if (IS_ALIGNED(low_pfn, pageblock_nr_pages) && | |
1662 | !fast_find_block && !isolation_suitable(cc, page)) | |
edc2ca61 VB |
1663 | continue; |
1664 | ||
1665 | /* | |
9bebefd5 MG |
1666 | * For async compaction, also only scan in MOVABLE blocks |
1667 | * without huge pages. Async compaction is optimistic to see | |
1668 | * if the minimum amount of work satisfies the allocation. | |
1669 | * The cached PFN is updated as it's possible that all | |
1670 | * remaining blocks between source and target are unsuitable | |
1671 | * and the compaction scanners fail to meet. | |
edc2ca61 | 1672 | */ |
9bebefd5 MG |
1673 | if (!suitable_migration_source(cc, page)) { |
1674 | update_cached_migrate(cc, block_end_pfn); | |
edc2ca61 | 1675 | continue; |
9bebefd5 | 1676 | } |
edc2ca61 VB |
1677 | |
1678 | /* Perform the isolation */ | |
e1409c32 JK |
1679 | low_pfn = isolate_migratepages_block(cc, low_pfn, |
1680 | block_end_pfn, isolate_mode); | |
edc2ca61 | 1681 | |
cb2dcaf0 | 1682 | if (!low_pfn) |
edc2ca61 VB |
1683 | return ISOLATE_ABORT; |
1684 | ||
1685 | /* | |
1686 | * Either we isolated something and proceed with migration. Or | |
1687 | * we failed and compact_zone should decide if we should | |
1688 | * continue or not. | |
1689 | */ | |
1690 | break; | |
1691 | } | |
1692 | ||
f2849aa0 VB |
1693 | /* Record where migration scanner will be restarted. */ |
1694 | cc->migrate_pfn = low_pfn; | |
ff9543fd | 1695 | |
edc2ca61 | 1696 | return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; |
ff9543fd MN |
1697 | } |
1698 | ||
21c527a3 YB |
1699 | /* |
1700 | * order == -1 is expected when compacting via | |
1701 | * /proc/sys/vm/compact_memory | |
1702 | */ | |
1703 | static inline bool is_via_compact_memory(int order) | |
1704 | { | |
1705 | return order == -1; | |
1706 | } | |
1707 | ||
40cacbcb | 1708 | static enum compact_result __compact_finished(struct compact_control *cc) |
748446bb | 1709 | { |
8fb74b9f | 1710 | unsigned int order; |
d39773a0 | 1711 | const int migratetype = cc->migratetype; |
cb2dcaf0 | 1712 | int ret; |
748446bb | 1713 | |
753341a4 | 1714 | /* Compaction run completes if the migrate and free scanner meet */ |
f2849aa0 | 1715 | if (compact_scanners_met(cc)) { |
55b7c4c9 | 1716 | /* Let the next compaction start anew. */ |
40cacbcb | 1717 | reset_cached_positions(cc->zone); |
55b7c4c9 | 1718 | |
62997027 MG |
1719 | /* |
1720 | * Mark that the PG_migrate_skip information should be cleared | |
accf6242 | 1721 | * by kswapd when it goes to sleep. kcompactd does not set the |
62997027 MG |
1722 | * flag itself as the decision to be clear should be directly |
1723 | * based on an allocation request. | |
1724 | */ | |
accf6242 | 1725 | if (cc->direct_compaction) |
40cacbcb | 1726 | cc->zone->compact_blockskip_flush = true; |
62997027 | 1727 | |
c8f7de0b MH |
1728 | if (cc->whole_zone) |
1729 | return COMPACT_COMPLETE; | |
1730 | else | |
1731 | return COMPACT_PARTIAL_SKIPPED; | |
bb13ffeb | 1732 | } |
748446bb | 1733 | |
21c527a3 | 1734 | if (is_via_compact_memory(cc->order)) |
56de7263 MG |
1735 | return COMPACT_CONTINUE; |
1736 | ||
efe771c7 MG |
1737 | /* |
1738 | * Always finish scanning a pageblock to reduce the possibility of | |
1739 | * fallbacks in the future. This is particularly important when | |
1740 | * migration source is unmovable/reclaimable but it's not worth | |
1741 | * special casing. | |
1742 | */ | |
1743 | if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages)) | |
1744 | return COMPACT_CONTINUE; | |
baf6a9a1 | 1745 | |
56de7263 | 1746 | /* Direct compactor: Is a suitable page free? */ |
cb2dcaf0 | 1747 | ret = COMPACT_NO_SUITABLE_PAGE; |
8fb74b9f | 1748 | for (order = cc->order; order < MAX_ORDER; order++) { |
40cacbcb | 1749 | struct free_area *area = &cc->zone->free_area[order]; |
2149cdae | 1750 | bool can_steal; |
8fb74b9f MG |
1751 | |
1752 | /* Job done if page is free of the right migratetype */ | |
6d7ce559 | 1753 | if (!list_empty(&area->free_list[migratetype])) |
cf378319 | 1754 | return COMPACT_SUCCESS; |
8fb74b9f | 1755 | |
2149cdae JK |
1756 | #ifdef CONFIG_CMA |
1757 | /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ | |
1758 | if (migratetype == MIGRATE_MOVABLE && | |
1759 | !list_empty(&area->free_list[MIGRATE_CMA])) | |
cf378319 | 1760 | return COMPACT_SUCCESS; |
2149cdae JK |
1761 | #endif |
1762 | /* | |
1763 | * Job done if allocation would steal freepages from | |
1764 | * other migratetype buddy lists. | |
1765 | */ | |
1766 | if (find_suitable_fallback(area, order, migratetype, | |
baf6a9a1 VB |
1767 | true, &can_steal) != -1) { |
1768 | ||
1769 | /* movable pages are OK in any pageblock */ | |
1770 | if (migratetype == MIGRATE_MOVABLE) | |
1771 | return COMPACT_SUCCESS; | |
1772 | ||
1773 | /* | |
1774 | * We are stealing for a non-movable allocation. Make | |
1775 | * sure we finish compacting the current pageblock | |
1776 | * first so it is as free as possible and we won't | |
1777 | * have to steal another one soon. This only applies | |
1778 | * to sync compaction, as async compaction operates | |
1779 | * on pageblocks of the same migratetype. | |
1780 | */ | |
1781 | if (cc->mode == MIGRATE_ASYNC || | |
1782 | IS_ALIGNED(cc->migrate_pfn, | |
1783 | pageblock_nr_pages)) { | |
1784 | return COMPACT_SUCCESS; | |
1785 | } | |
1786 | ||
cb2dcaf0 MG |
1787 | ret = COMPACT_CONTINUE; |
1788 | break; | |
baf6a9a1 | 1789 | } |
56de7263 MG |
1790 | } |
1791 | ||
cb2dcaf0 MG |
1792 | if (cc->contended || fatal_signal_pending(current)) |
1793 | ret = COMPACT_CONTENDED; | |
1794 | ||
1795 | return ret; | |
837d026d JK |
1796 | } |
1797 | ||
40cacbcb | 1798 | static enum compact_result compact_finished(struct compact_control *cc) |
837d026d JK |
1799 | { |
1800 | int ret; | |
1801 | ||
40cacbcb MG |
1802 | ret = __compact_finished(cc); |
1803 | trace_mm_compaction_finished(cc->zone, cc->order, ret); | |
837d026d JK |
1804 | if (ret == COMPACT_NO_SUITABLE_PAGE) |
1805 | ret = COMPACT_CONTINUE; | |
1806 | ||
1807 | return ret; | |
748446bb MG |
1808 | } |
1809 | ||
3e7d3449 MG |
1810 | /* |
1811 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
1812 | * Returns | |
1813 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
cf378319 | 1814 | * COMPACT_SUCCESS - If the allocation would succeed without compaction |
3e7d3449 MG |
1815 | * COMPACT_CONTINUE - If compaction should run now |
1816 | */ | |
ea7ab982 | 1817 | static enum compact_result __compaction_suitable(struct zone *zone, int order, |
c603844b | 1818 | unsigned int alloc_flags, |
86a294a8 MH |
1819 | int classzone_idx, |
1820 | unsigned long wmark_target) | |
3e7d3449 | 1821 | { |
3e7d3449 MG |
1822 | unsigned long watermark; |
1823 | ||
21c527a3 | 1824 | if (is_via_compact_memory(order)) |
3957c776 MH |
1825 | return COMPACT_CONTINUE; |
1826 | ||
a9214443 | 1827 | watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
ebff3980 VB |
1828 | /* |
1829 | * If watermarks for high-order allocation are already met, there | |
1830 | * should be no need for compaction at all. | |
1831 | */ | |
1832 | if (zone_watermark_ok(zone, order, watermark, classzone_idx, | |
1833 | alloc_flags)) | |
cf378319 | 1834 | return COMPACT_SUCCESS; |
ebff3980 | 1835 | |
3e7d3449 | 1836 | /* |
9861a62c | 1837 | * Watermarks for order-0 must be met for compaction to be able to |
984fdba6 VB |
1838 | * isolate free pages for migration targets. This means that the |
1839 | * watermark and alloc_flags have to match, or be more pessimistic than | |
1840 | * the check in __isolate_free_page(). We don't use the direct | |
1841 | * compactor's alloc_flags, as they are not relevant for freepage | |
1842 | * isolation. We however do use the direct compactor's classzone_idx to | |
1843 | * skip over zones where lowmem reserves would prevent allocation even | |
1844 | * if compaction succeeds. | |
8348faf9 VB |
1845 | * For costly orders, we require low watermark instead of min for |
1846 | * compaction to proceed to increase its chances. | |
d883c6cf JK |
1847 | * ALLOC_CMA is used, as pages in CMA pageblocks are considered |
1848 | * suitable migration targets | |
3e7d3449 | 1849 | */ |
8348faf9 VB |
1850 | watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
1851 | low_wmark_pages(zone) : min_wmark_pages(zone); | |
1852 | watermark += compact_gap(order); | |
86a294a8 | 1853 | if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx, |
d883c6cf | 1854 | ALLOC_CMA, wmark_target)) |
3e7d3449 MG |
1855 | return COMPACT_SKIPPED; |
1856 | ||
cc5c9f09 VB |
1857 | return COMPACT_CONTINUE; |
1858 | } | |
1859 | ||
1860 | enum compact_result compaction_suitable(struct zone *zone, int order, | |
1861 | unsigned int alloc_flags, | |
1862 | int classzone_idx) | |
1863 | { | |
1864 | enum compact_result ret; | |
1865 | int fragindex; | |
1866 | ||
1867 | ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx, | |
1868 | zone_page_state(zone, NR_FREE_PAGES)); | |
3e7d3449 MG |
1869 | /* |
1870 | * fragmentation index determines if allocation failures are due to | |
1871 | * low memory or external fragmentation | |
1872 | * | |
ebff3980 VB |
1873 | * index of -1000 would imply allocations might succeed depending on |
1874 | * watermarks, but we already failed the high-order watermark check | |
3e7d3449 MG |
1875 | * index towards 0 implies failure is due to lack of memory |
1876 | * index towards 1000 implies failure is due to fragmentation | |
1877 | * | |
20311420 VB |
1878 | * Only compact if a failure would be due to fragmentation. Also |
1879 | * ignore fragindex for non-costly orders where the alternative to | |
1880 | * a successful reclaim/compaction is OOM. Fragindex and the | |
1881 | * vm.extfrag_threshold sysctl is meant as a heuristic to prevent | |
1882 | * excessive compaction for costly orders, but it should not be at the | |
1883 | * expense of system stability. | |
3e7d3449 | 1884 | */ |
20311420 | 1885 | if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) { |
cc5c9f09 VB |
1886 | fragindex = fragmentation_index(zone, order); |
1887 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
1888 | ret = COMPACT_NOT_SUITABLE_ZONE; | |
1889 | } | |
837d026d | 1890 | |
837d026d JK |
1891 | trace_mm_compaction_suitable(zone, order, ret); |
1892 | if (ret == COMPACT_NOT_SUITABLE_ZONE) | |
1893 | ret = COMPACT_SKIPPED; | |
1894 | ||
1895 | return ret; | |
1896 | } | |
1897 | ||
86a294a8 MH |
1898 | bool compaction_zonelist_suitable(struct alloc_context *ac, int order, |
1899 | int alloc_flags) | |
1900 | { | |
1901 | struct zone *zone; | |
1902 | struct zoneref *z; | |
1903 | ||
1904 | /* | |
1905 | * Make sure at least one zone would pass __compaction_suitable if we continue | |
1906 | * retrying the reclaim. | |
1907 | */ | |
1908 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, | |
1909 | ac->nodemask) { | |
1910 | unsigned long available; | |
1911 | enum compact_result compact_result; | |
1912 | ||
1913 | /* | |
1914 | * Do not consider all the reclaimable memory because we do not | |
1915 | * want to trash just for a single high order allocation which | |
1916 | * is even not guaranteed to appear even if __compaction_suitable | |
1917 | * is happy about the watermark check. | |
1918 | */ | |
5a1c84b4 | 1919 | available = zone_reclaimable_pages(zone) / order; |
86a294a8 MH |
1920 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
1921 | compact_result = __compaction_suitable(zone, order, alloc_flags, | |
1922 | ac_classzone_idx(ac), available); | |
cc5c9f09 | 1923 | if (compact_result != COMPACT_SKIPPED) |
86a294a8 MH |
1924 | return true; |
1925 | } | |
1926 | ||
1927 | return false; | |
1928 | } | |
1929 | ||
40cacbcb | 1930 | static enum compact_result compact_zone(struct compact_control *cc) |
748446bb | 1931 | { |
ea7ab982 | 1932 | enum compact_result ret; |
40cacbcb MG |
1933 | unsigned long start_pfn = cc->zone->zone_start_pfn; |
1934 | unsigned long end_pfn = zone_end_pfn(cc->zone); | |
566e54e1 | 1935 | unsigned long last_migrated_pfn; |
e0b9daeb | 1936 | const bool sync = cc->mode != MIGRATE_ASYNC; |
8854c55f | 1937 | bool update_cached; |
748446bb | 1938 | |
d39773a0 | 1939 | cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask); |
40cacbcb | 1940 | ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags, |
ebff3980 | 1941 | cc->classzone_idx); |
c46649de | 1942 | /* Compaction is likely to fail */ |
cf378319 | 1943 | if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED) |
3e7d3449 | 1944 | return ret; |
c46649de MH |
1945 | |
1946 | /* huh, compaction_suitable is returning something unexpected */ | |
1947 | VM_BUG_ON(ret != COMPACT_CONTINUE); | |
3e7d3449 | 1948 | |
d3132e4b VB |
1949 | /* |
1950 | * Clear pageblock skip if there were failures recently and compaction | |
accf6242 | 1951 | * is about to be retried after being deferred. |
d3132e4b | 1952 | */ |
40cacbcb MG |
1953 | if (compaction_restarting(cc->zone, cc->order)) |
1954 | __reset_isolation_suitable(cc->zone); | |
d3132e4b | 1955 | |
c89511ab MG |
1956 | /* |
1957 | * Setup to move all movable pages to the end of the zone. Used cached | |
06ed2998 VB |
1958 | * information on where the scanners should start (unless we explicitly |
1959 | * want to compact the whole zone), but check that it is initialised | |
1960 | * by ensuring the values are within zone boundaries. | |
c89511ab | 1961 | */ |
70b44595 | 1962 | cc->fast_start_pfn = 0; |
06ed2998 | 1963 | if (cc->whole_zone) { |
c89511ab | 1964 | cc->migrate_pfn = start_pfn; |
06ed2998 VB |
1965 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); |
1966 | } else { | |
40cacbcb MG |
1967 | cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; |
1968 | cc->free_pfn = cc->zone->compact_cached_free_pfn; | |
06ed2998 VB |
1969 | if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { |
1970 | cc->free_pfn = pageblock_start_pfn(end_pfn - 1); | |
40cacbcb | 1971 | cc->zone->compact_cached_free_pfn = cc->free_pfn; |
06ed2998 VB |
1972 | } |
1973 | if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { | |
1974 | cc->migrate_pfn = start_pfn; | |
40cacbcb MG |
1975 | cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; |
1976 | cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; | |
06ed2998 | 1977 | } |
c8f7de0b | 1978 | |
06ed2998 VB |
1979 | if (cc->migrate_pfn == start_pfn) |
1980 | cc->whole_zone = true; | |
1981 | } | |
c8f7de0b | 1982 | |
566e54e1 | 1983 | last_migrated_pfn = 0; |
748446bb | 1984 | |
8854c55f MG |
1985 | /* |
1986 | * Migrate has separate cached PFNs for ASYNC and SYNC* migration on | |
1987 | * the basis that some migrations will fail in ASYNC mode. However, | |
1988 | * if the cached PFNs match and pageblocks are skipped due to having | |
1989 | * no isolation candidates, then the sync state does not matter. | |
1990 | * Until a pageblock with isolation candidates is found, keep the | |
1991 | * cached PFNs in sync to avoid revisiting the same blocks. | |
1992 | */ | |
1993 | update_cached = !sync && | |
1994 | cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; | |
1995 | ||
16c4a097 JK |
1996 | trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, |
1997 | cc->free_pfn, end_pfn, sync); | |
0eb927c0 | 1998 | |
748446bb MG |
1999 | migrate_prep_local(); |
2000 | ||
40cacbcb | 2001 | while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { |
9d502c1c | 2002 | int err; |
566e54e1 | 2003 | unsigned long start_pfn = cc->migrate_pfn; |
748446bb | 2004 | |
804d3121 MG |
2005 | /* |
2006 | * Avoid multiple rescans which can happen if a page cannot be | |
2007 | * isolated (dirty/writeback in async mode) or if the migrated | |
2008 | * pages are being allocated before the pageblock is cleared. | |
2009 | * The first rescan will capture the entire pageblock for | |
2010 | * migration. If it fails, it'll be marked skip and scanning | |
2011 | * will proceed as normal. | |
2012 | */ | |
2013 | cc->rescan = false; | |
2014 | if (pageblock_start_pfn(last_migrated_pfn) == | |
2015 | pageblock_start_pfn(start_pfn)) { | |
2016 | cc->rescan = true; | |
2017 | } | |
2018 | ||
40cacbcb | 2019 | switch (isolate_migratepages(cc->zone, cc)) { |
f9e35b3b | 2020 | case ISOLATE_ABORT: |
2d1e1041 | 2021 | ret = COMPACT_CONTENDED; |
5733c7d1 | 2022 | putback_movable_pages(&cc->migratepages); |
e64c5237 | 2023 | cc->nr_migratepages = 0; |
566e54e1 | 2024 | last_migrated_pfn = 0; |
f9e35b3b MG |
2025 | goto out; |
2026 | case ISOLATE_NONE: | |
8854c55f MG |
2027 | if (update_cached) { |
2028 | cc->zone->compact_cached_migrate_pfn[1] = | |
2029 | cc->zone->compact_cached_migrate_pfn[0]; | |
2030 | } | |
2031 | ||
fdaf7f5c VB |
2032 | /* |
2033 | * We haven't isolated and migrated anything, but | |
2034 | * there might still be unflushed migrations from | |
2035 | * previous cc->order aligned block. | |
2036 | */ | |
2037 | goto check_drain; | |
f9e35b3b | 2038 | case ISOLATE_SUCCESS: |
8854c55f | 2039 | update_cached = false; |
566e54e1 | 2040 | last_migrated_pfn = start_pfn; |
f9e35b3b MG |
2041 | ; |
2042 | } | |
748446bb | 2043 | |
d53aea3d | 2044 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
e0b9daeb | 2045 | compaction_free, (unsigned long)cc, cc->mode, |
7b2a2d4a | 2046 | MR_COMPACTION); |
748446bb | 2047 | |
f8c9301f VB |
2048 | trace_mm_compaction_migratepages(cc->nr_migratepages, err, |
2049 | &cc->migratepages); | |
748446bb | 2050 | |
f8c9301f VB |
2051 | /* All pages were either migrated or will be released */ |
2052 | cc->nr_migratepages = 0; | |
9d502c1c | 2053 | if (err) { |
5733c7d1 | 2054 | putback_movable_pages(&cc->migratepages); |
7ed695e0 VB |
2055 | /* |
2056 | * migrate_pages() may return -ENOMEM when scanners meet | |
2057 | * and we want compact_finished() to detect it | |
2058 | */ | |
f2849aa0 | 2059 | if (err == -ENOMEM && !compact_scanners_met(cc)) { |
2d1e1041 | 2060 | ret = COMPACT_CONTENDED; |
4bf2bba3 DR |
2061 | goto out; |
2062 | } | |
fdd048e1 VB |
2063 | /* |
2064 | * We failed to migrate at least one page in the current | |
2065 | * order-aligned block, so skip the rest of it. | |
2066 | */ | |
2067 | if (cc->direct_compaction && | |
2068 | (cc->mode == MIGRATE_ASYNC)) { | |
2069 | cc->migrate_pfn = block_end_pfn( | |
2070 | cc->migrate_pfn - 1, cc->order); | |
2071 | /* Draining pcplists is useless in this case */ | |
566e54e1 | 2072 | last_migrated_pfn = 0; |
fdd048e1 | 2073 | } |
748446bb | 2074 | } |
fdaf7f5c | 2075 | |
fdaf7f5c VB |
2076 | check_drain: |
2077 | /* | |
2078 | * Has the migration scanner moved away from the previous | |
2079 | * cc->order aligned block where we migrated from? If yes, | |
2080 | * flush the pages that were freed, so that they can merge and | |
2081 | * compact_finished() can detect immediately if allocation | |
2082 | * would succeed. | |
2083 | */ | |
566e54e1 | 2084 | if (cc->order > 0 && last_migrated_pfn) { |
fdaf7f5c VB |
2085 | int cpu; |
2086 | unsigned long current_block_start = | |
06b6640a | 2087 | block_start_pfn(cc->migrate_pfn, cc->order); |
fdaf7f5c | 2088 | |
566e54e1 | 2089 | if (last_migrated_pfn < current_block_start) { |
fdaf7f5c VB |
2090 | cpu = get_cpu(); |
2091 | lru_add_drain_cpu(cpu); | |
40cacbcb | 2092 | drain_local_pages(cc->zone); |
fdaf7f5c VB |
2093 | put_cpu(); |
2094 | /* No more flushing until we migrate again */ | |
566e54e1 | 2095 | last_migrated_pfn = 0; |
fdaf7f5c VB |
2096 | } |
2097 | } | |
2098 | ||
748446bb MG |
2099 | } |
2100 | ||
f9e35b3b | 2101 | out: |
6bace090 VB |
2102 | /* |
2103 | * Release free pages and update where the free scanner should restart, | |
2104 | * so we don't leave any returned pages behind in the next attempt. | |
2105 | */ | |
2106 | if (cc->nr_freepages > 0) { | |
2107 | unsigned long free_pfn = release_freepages(&cc->freepages); | |
2108 | ||
2109 | cc->nr_freepages = 0; | |
2110 | VM_BUG_ON(free_pfn == 0); | |
2111 | /* The cached pfn is always the first in a pageblock */ | |
06b6640a | 2112 | free_pfn = pageblock_start_pfn(free_pfn); |
6bace090 VB |
2113 | /* |
2114 | * Only go back, not forward. The cached pfn might have been | |
2115 | * already reset to zone end in compact_finished() | |
2116 | */ | |
40cacbcb MG |
2117 | if (free_pfn > cc->zone->compact_cached_free_pfn) |
2118 | cc->zone->compact_cached_free_pfn = free_pfn; | |
6bace090 | 2119 | } |
748446bb | 2120 | |
7f354a54 DR |
2121 | count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); |
2122 | count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); | |
2123 | ||
16c4a097 JK |
2124 | trace_mm_compaction_end(start_pfn, cc->migrate_pfn, |
2125 | cc->free_pfn, end_pfn, sync, ret); | |
0eb927c0 | 2126 | |
748446bb MG |
2127 | return ret; |
2128 | } | |
76ab0f53 | 2129 | |
ea7ab982 | 2130 | static enum compact_result compact_zone_order(struct zone *zone, int order, |
c3486f53 | 2131 | gfp_t gfp_mask, enum compact_priority prio, |
c603844b | 2132 | unsigned int alloc_flags, int classzone_idx) |
56de7263 | 2133 | { |
ea7ab982 | 2134 | enum compact_result ret; |
56de7263 MG |
2135 | struct compact_control cc = { |
2136 | .nr_freepages = 0, | |
2137 | .nr_migratepages = 0, | |
7f354a54 DR |
2138 | .total_migrate_scanned = 0, |
2139 | .total_free_scanned = 0, | |
56de7263 | 2140 | .order = order, |
6d7ce559 | 2141 | .gfp_mask = gfp_mask, |
56de7263 | 2142 | .zone = zone, |
a5508cd8 VB |
2143 | .mode = (prio == COMPACT_PRIO_ASYNC) ? |
2144 | MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, | |
ebff3980 VB |
2145 | .alloc_flags = alloc_flags, |
2146 | .classzone_idx = classzone_idx, | |
accf6242 | 2147 | .direct_compaction = true, |
a8e025e5 | 2148 | .whole_zone = (prio == MIN_COMPACT_PRIORITY), |
9f7e3387 VB |
2149 | .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), |
2150 | .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) | |
56de7263 MG |
2151 | }; |
2152 | INIT_LIST_HEAD(&cc.freepages); | |
2153 | INIT_LIST_HEAD(&cc.migratepages); | |
2154 | ||
40cacbcb | 2155 | ret = compact_zone(&cc); |
e64c5237 SL |
2156 | |
2157 | VM_BUG_ON(!list_empty(&cc.freepages)); | |
2158 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2159 | ||
e64c5237 | 2160 | return ret; |
56de7263 MG |
2161 | } |
2162 | ||
5e771905 MG |
2163 | int sysctl_extfrag_threshold = 500; |
2164 | ||
56de7263 MG |
2165 | /** |
2166 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
56de7263 | 2167 | * @gfp_mask: The GFP mask of the current allocation |
1a6d53a1 VB |
2168 | * @order: The order of the current allocation |
2169 | * @alloc_flags: The allocation flags of the current allocation | |
2170 | * @ac: The context of current allocation | |
112d2d29 | 2171 | * @prio: Determines how hard direct compaction should try to succeed |
56de7263 MG |
2172 | * |
2173 | * This is the main entry point for direct page compaction. | |
2174 | */ | |
ea7ab982 | 2175 | enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, |
c603844b | 2176 | unsigned int alloc_flags, const struct alloc_context *ac, |
c3486f53 | 2177 | enum compact_priority prio) |
56de7263 | 2178 | { |
56de7263 | 2179 | int may_perform_io = gfp_mask & __GFP_IO; |
56de7263 MG |
2180 | struct zoneref *z; |
2181 | struct zone *zone; | |
1d4746d3 | 2182 | enum compact_result rc = COMPACT_SKIPPED; |
56de7263 | 2183 | |
73e64c51 MH |
2184 | /* |
2185 | * Check if the GFP flags allow compaction - GFP_NOIO is really | |
2186 | * tricky context because the migration might require IO | |
2187 | */ | |
2188 | if (!may_perform_io) | |
53853e2d | 2189 | return COMPACT_SKIPPED; |
56de7263 | 2190 | |
a5508cd8 | 2191 | trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); |
837d026d | 2192 | |
56de7263 | 2193 | /* Compact each zone in the list */ |
1a6d53a1 VB |
2194 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, |
2195 | ac->nodemask) { | |
ea7ab982 | 2196 | enum compact_result status; |
56de7263 | 2197 | |
a8e025e5 VB |
2198 | if (prio > MIN_COMPACT_PRIORITY |
2199 | && compaction_deferred(zone, order)) { | |
1d4746d3 | 2200 | rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); |
53853e2d | 2201 | continue; |
1d4746d3 | 2202 | } |
53853e2d | 2203 | |
a5508cd8 | 2204 | status = compact_zone_order(zone, order, gfp_mask, prio, |
c3486f53 | 2205 | alloc_flags, ac_classzone_idx(ac)); |
56de7263 MG |
2206 | rc = max(status, rc); |
2207 | ||
7ceb009a VB |
2208 | /* The allocation should succeed, stop compacting */ |
2209 | if (status == COMPACT_SUCCESS) { | |
53853e2d VB |
2210 | /* |
2211 | * We think the allocation will succeed in this zone, | |
2212 | * but it is not certain, hence the false. The caller | |
2213 | * will repeat this with true if allocation indeed | |
2214 | * succeeds in this zone. | |
2215 | */ | |
2216 | compaction_defer_reset(zone, order, false); | |
1f9efdef | 2217 | |
c3486f53 | 2218 | break; |
1f9efdef VB |
2219 | } |
2220 | ||
a5508cd8 | 2221 | if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || |
c3486f53 | 2222 | status == COMPACT_PARTIAL_SKIPPED)) |
53853e2d VB |
2223 | /* |
2224 | * We think that allocation won't succeed in this zone | |
2225 | * so we defer compaction there. If it ends up | |
2226 | * succeeding after all, it will be reset. | |
2227 | */ | |
2228 | defer_compaction(zone, order); | |
1f9efdef VB |
2229 | |
2230 | /* | |
2231 | * We might have stopped compacting due to need_resched() in | |
2232 | * async compaction, or due to a fatal signal detected. In that | |
c3486f53 | 2233 | * case do not try further zones |
1f9efdef | 2234 | */ |
c3486f53 VB |
2235 | if ((prio == COMPACT_PRIO_ASYNC && need_resched()) |
2236 | || fatal_signal_pending(current)) | |
2237 | break; | |
56de7263 MG |
2238 | } |
2239 | ||
2240 | return rc; | |
2241 | } | |
2242 | ||
2243 | ||
76ab0f53 | 2244 | /* Compact all zones within a node */ |
791cae96 | 2245 | static void compact_node(int nid) |
76ab0f53 | 2246 | { |
791cae96 | 2247 | pg_data_t *pgdat = NODE_DATA(nid); |
76ab0f53 | 2248 | int zoneid; |
76ab0f53 | 2249 | struct zone *zone; |
791cae96 VB |
2250 | struct compact_control cc = { |
2251 | .order = -1, | |
7f354a54 DR |
2252 | .total_migrate_scanned = 0, |
2253 | .total_free_scanned = 0, | |
791cae96 VB |
2254 | .mode = MIGRATE_SYNC, |
2255 | .ignore_skip_hint = true, | |
2256 | .whole_zone = true, | |
73e64c51 | 2257 | .gfp_mask = GFP_KERNEL, |
791cae96 VB |
2258 | }; |
2259 | ||
76ab0f53 | 2260 | |
76ab0f53 | 2261 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
2262 | |
2263 | zone = &pgdat->node_zones[zoneid]; | |
2264 | if (!populated_zone(zone)) | |
2265 | continue; | |
2266 | ||
791cae96 VB |
2267 | cc.nr_freepages = 0; |
2268 | cc.nr_migratepages = 0; | |
2269 | cc.zone = zone; | |
2270 | INIT_LIST_HEAD(&cc.freepages); | |
2271 | INIT_LIST_HEAD(&cc.migratepages); | |
76ab0f53 | 2272 | |
40cacbcb | 2273 | compact_zone(&cc); |
75469345 | 2274 | |
791cae96 VB |
2275 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2276 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
76ab0f53 | 2277 | } |
76ab0f53 MG |
2278 | } |
2279 | ||
2280 | /* Compact all nodes in the system */ | |
7964c06d | 2281 | static void compact_nodes(void) |
76ab0f53 MG |
2282 | { |
2283 | int nid; | |
2284 | ||
8575ec29 HD |
2285 | /* Flush pending updates to the LRU lists */ |
2286 | lru_add_drain_all(); | |
2287 | ||
76ab0f53 MG |
2288 | for_each_online_node(nid) |
2289 | compact_node(nid); | |
76ab0f53 MG |
2290 | } |
2291 | ||
2292 | /* The written value is actually unused, all memory is compacted */ | |
2293 | int sysctl_compact_memory; | |
2294 | ||
fec4eb2c YB |
2295 | /* |
2296 | * This is the entry point for compacting all nodes via | |
2297 | * /proc/sys/vm/compact_memory | |
2298 | */ | |
76ab0f53 MG |
2299 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
2300 | void __user *buffer, size_t *length, loff_t *ppos) | |
2301 | { | |
2302 | if (write) | |
7964c06d | 2303 | compact_nodes(); |
76ab0f53 MG |
2304 | |
2305 | return 0; | |
2306 | } | |
ed4a6d7f MG |
2307 | |
2308 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) | |
74e77fb9 | 2309 | static ssize_t sysfs_compact_node(struct device *dev, |
10fbcf4c | 2310 | struct device_attribute *attr, |
ed4a6d7f MG |
2311 | const char *buf, size_t count) |
2312 | { | |
8575ec29 HD |
2313 | int nid = dev->id; |
2314 | ||
2315 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
2316 | /* Flush pending updates to the LRU lists */ | |
2317 | lru_add_drain_all(); | |
2318 | ||
2319 | compact_node(nid); | |
2320 | } | |
ed4a6d7f MG |
2321 | |
2322 | return count; | |
2323 | } | |
0825a6f9 | 2324 | static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node); |
ed4a6d7f MG |
2325 | |
2326 | int compaction_register_node(struct node *node) | |
2327 | { | |
10fbcf4c | 2328 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2329 | } |
2330 | ||
2331 | void compaction_unregister_node(struct node *node) | |
2332 | { | |
10fbcf4c | 2333 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
2334 | } |
2335 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd | 2336 | |
698b1b30 VB |
2337 | static inline bool kcompactd_work_requested(pg_data_t *pgdat) |
2338 | { | |
172400c6 | 2339 | return pgdat->kcompactd_max_order > 0 || kthread_should_stop(); |
698b1b30 VB |
2340 | } |
2341 | ||
2342 | static bool kcompactd_node_suitable(pg_data_t *pgdat) | |
2343 | { | |
2344 | int zoneid; | |
2345 | struct zone *zone; | |
2346 | enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx; | |
2347 | ||
6cd9dc3e | 2348 | for (zoneid = 0; zoneid <= classzone_idx; zoneid++) { |
698b1b30 VB |
2349 | zone = &pgdat->node_zones[zoneid]; |
2350 | ||
2351 | if (!populated_zone(zone)) | |
2352 | continue; | |
2353 | ||
2354 | if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0, | |
2355 | classzone_idx) == COMPACT_CONTINUE) | |
2356 | return true; | |
2357 | } | |
2358 | ||
2359 | return false; | |
2360 | } | |
2361 | ||
2362 | static void kcompactd_do_work(pg_data_t *pgdat) | |
2363 | { | |
2364 | /* | |
2365 | * With no special task, compact all zones so that a page of requested | |
2366 | * order is allocatable. | |
2367 | */ | |
2368 | int zoneid; | |
2369 | struct zone *zone; | |
2370 | struct compact_control cc = { | |
2371 | .order = pgdat->kcompactd_max_order, | |
7f354a54 DR |
2372 | .total_migrate_scanned = 0, |
2373 | .total_free_scanned = 0, | |
698b1b30 VB |
2374 | .classzone_idx = pgdat->kcompactd_classzone_idx, |
2375 | .mode = MIGRATE_SYNC_LIGHT, | |
a0647dc9 | 2376 | .ignore_skip_hint = false, |
73e64c51 | 2377 | .gfp_mask = GFP_KERNEL, |
698b1b30 | 2378 | }; |
698b1b30 VB |
2379 | trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, |
2380 | cc.classzone_idx); | |
7f354a54 | 2381 | count_compact_event(KCOMPACTD_WAKE); |
698b1b30 | 2382 | |
6cd9dc3e | 2383 | for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) { |
698b1b30 VB |
2384 | int status; |
2385 | ||
2386 | zone = &pgdat->node_zones[zoneid]; | |
2387 | if (!populated_zone(zone)) | |
2388 | continue; | |
2389 | ||
2390 | if (compaction_deferred(zone, cc.order)) | |
2391 | continue; | |
2392 | ||
2393 | if (compaction_suitable(zone, cc.order, 0, zoneid) != | |
2394 | COMPACT_CONTINUE) | |
2395 | continue; | |
2396 | ||
2397 | cc.nr_freepages = 0; | |
2398 | cc.nr_migratepages = 0; | |
7f354a54 DR |
2399 | cc.total_migrate_scanned = 0; |
2400 | cc.total_free_scanned = 0; | |
698b1b30 VB |
2401 | cc.zone = zone; |
2402 | INIT_LIST_HEAD(&cc.freepages); | |
2403 | INIT_LIST_HEAD(&cc.migratepages); | |
2404 | ||
172400c6 VB |
2405 | if (kthread_should_stop()) |
2406 | return; | |
40cacbcb | 2407 | status = compact_zone(&cc); |
698b1b30 | 2408 | |
7ceb009a | 2409 | if (status == COMPACT_SUCCESS) { |
698b1b30 | 2410 | compaction_defer_reset(zone, cc.order, false); |
c8f7de0b | 2411 | } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { |
bc3106b2 DR |
2412 | /* |
2413 | * Buddy pages may become stranded on pcps that could | |
2414 | * otherwise coalesce on the zone's free area for | |
2415 | * order >= cc.order. This is ratelimited by the | |
2416 | * upcoming deferral. | |
2417 | */ | |
2418 | drain_all_pages(zone); | |
2419 | ||
698b1b30 VB |
2420 | /* |
2421 | * We use sync migration mode here, so we defer like | |
2422 | * sync direct compaction does. | |
2423 | */ | |
2424 | defer_compaction(zone, cc.order); | |
2425 | } | |
2426 | ||
7f354a54 DR |
2427 | count_compact_events(KCOMPACTD_MIGRATE_SCANNED, |
2428 | cc.total_migrate_scanned); | |
2429 | count_compact_events(KCOMPACTD_FREE_SCANNED, | |
2430 | cc.total_free_scanned); | |
2431 | ||
698b1b30 VB |
2432 | VM_BUG_ON(!list_empty(&cc.freepages)); |
2433 | VM_BUG_ON(!list_empty(&cc.migratepages)); | |
2434 | } | |
2435 | ||
2436 | /* | |
2437 | * Regardless of success, we are done until woken up next. But remember | |
2438 | * the requested order/classzone_idx in case it was higher/tighter than | |
2439 | * our current ones | |
2440 | */ | |
2441 | if (pgdat->kcompactd_max_order <= cc.order) | |
2442 | pgdat->kcompactd_max_order = 0; | |
2443 | if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx) | |
2444 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2445 | } | |
2446 | ||
2447 | void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx) | |
2448 | { | |
2449 | if (!order) | |
2450 | return; | |
2451 | ||
2452 | if (pgdat->kcompactd_max_order < order) | |
2453 | pgdat->kcompactd_max_order = order; | |
2454 | ||
2455 | if (pgdat->kcompactd_classzone_idx > classzone_idx) | |
2456 | pgdat->kcompactd_classzone_idx = classzone_idx; | |
2457 | ||
6818600f DB |
2458 | /* |
2459 | * Pairs with implicit barrier in wait_event_freezable() | |
2460 | * such that wakeups are not missed. | |
2461 | */ | |
2462 | if (!wq_has_sleeper(&pgdat->kcompactd_wait)) | |
698b1b30 VB |
2463 | return; |
2464 | ||
2465 | if (!kcompactd_node_suitable(pgdat)) | |
2466 | return; | |
2467 | ||
2468 | trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, | |
2469 | classzone_idx); | |
2470 | wake_up_interruptible(&pgdat->kcompactd_wait); | |
2471 | } | |
2472 | ||
2473 | /* | |
2474 | * The background compaction daemon, started as a kernel thread | |
2475 | * from the init process. | |
2476 | */ | |
2477 | static int kcompactd(void *p) | |
2478 | { | |
2479 | pg_data_t *pgdat = (pg_data_t*)p; | |
2480 | struct task_struct *tsk = current; | |
2481 | ||
2482 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2483 | ||
2484 | if (!cpumask_empty(cpumask)) | |
2485 | set_cpus_allowed_ptr(tsk, cpumask); | |
2486 | ||
2487 | set_freezable(); | |
2488 | ||
2489 | pgdat->kcompactd_max_order = 0; | |
2490 | pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1; | |
2491 | ||
2492 | while (!kthread_should_stop()) { | |
eb414681 JW |
2493 | unsigned long pflags; |
2494 | ||
698b1b30 VB |
2495 | trace_mm_compaction_kcompactd_sleep(pgdat->node_id); |
2496 | wait_event_freezable(pgdat->kcompactd_wait, | |
2497 | kcompactd_work_requested(pgdat)); | |
2498 | ||
eb414681 | 2499 | psi_memstall_enter(&pflags); |
698b1b30 | 2500 | kcompactd_do_work(pgdat); |
eb414681 | 2501 | psi_memstall_leave(&pflags); |
698b1b30 VB |
2502 | } |
2503 | ||
2504 | return 0; | |
2505 | } | |
2506 | ||
2507 | /* | |
2508 | * This kcompactd start function will be called by init and node-hot-add. | |
2509 | * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. | |
2510 | */ | |
2511 | int kcompactd_run(int nid) | |
2512 | { | |
2513 | pg_data_t *pgdat = NODE_DATA(nid); | |
2514 | int ret = 0; | |
2515 | ||
2516 | if (pgdat->kcompactd) | |
2517 | return 0; | |
2518 | ||
2519 | pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); | |
2520 | if (IS_ERR(pgdat->kcompactd)) { | |
2521 | pr_err("Failed to start kcompactd on node %d\n", nid); | |
2522 | ret = PTR_ERR(pgdat->kcompactd); | |
2523 | pgdat->kcompactd = NULL; | |
2524 | } | |
2525 | return ret; | |
2526 | } | |
2527 | ||
2528 | /* | |
2529 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
2530 | * hold mem_hotplug_begin/end(). | |
2531 | */ | |
2532 | void kcompactd_stop(int nid) | |
2533 | { | |
2534 | struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; | |
2535 | ||
2536 | if (kcompactd) { | |
2537 | kthread_stop(kcompactd); | |
2538 | NODE_DATA(nid)->kcompactd = NULL; | |
2539 | } | |
2540 | } | |
2541 | ||
2542 | /* | |
2543 | * It's optimal to keep kcompactd on the same CPUs as their memory, but | |
2544 | * not required for correctness. So if the last cpu in a node goes | |
2545 | * away, we get changed to run anywhere: as the first one comes back, | |
2546 | * restore their cpu bindings. | |
2547 | */ | |
e46b1db2 | 2548 | static int kcompactd_cpu_online(unsigned int cpu) |
698b1b30 VB |
2549 | { |
2550 | int nid; | |
2551 | ||
e46b1db2 AMG |
2552 | for_each_node_state(nid, N_MEMORY) { |
2553 | pg_data_t *pgdat = NODE_DATA(nid); | |
2554 | const struct cpumask *mask; | |
698b1b30 | 2555 | |
e46b1db2 | 2556 | mask = cpumask_of_node(pgdat->node_id); |
698b1b30 | 2557 | |
e46b1db2 AMG |
2558 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
2559 | /* One of our CPUs online: restore mask */ | |
2560 | set_cpus_allowed_ptr(pgdat->kcompactd, mask); | |
698b1b30 | 2561 | } |
e46b1db2 | 2562 | return 0; |
698b1b30 VB |
2563 | } |
2564 | ||
2565 | static int __init kcompactd_init(void) | |
2566 | { | |
2567 | int nid; | |
e46b1db2 AMG |
2568 | int ret; |
2569 | ||
2570 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
2571 | "mm/compaction:online", | |
2572 | kcompactd_cpu_online, NULL); | |
2573 | if (ret < 0) { | |
2574 | pr_err("kcompactd: failed to register hotplug callbacks.\n"); | |
2575 | return ret; | |
2576 | } | |
698b1b30 VB |
2577 | |
2578 | for_each_node_state(nid, N_MEMORY) | |
2579 | kcompactd_run(nid); | |
698b1b30 VB |
2580 | return 0; |
2581 | } | |
2582 | subsys_initcall(kcompactd_init) | |
2583 | ||
ff9543fd | 2584 | #endif /* CONFIG_COMPACTION */ |