Commit | Line | Data |
---|---|---|
748446bb MG |
1 | /* |
2 | * linux/mm/compaction.c | |
3 | * | |
4 | * Memory compaction for the reduction of external fragmentation. Note that | |
5 | * this heavily depends upon page migration to do all the real heavy | |
6 | * lifting | |
7 | * | |
8 | * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> | |
9 | */ | |
10 | #include <linux/swap.h> | |
11 | #include <linux/migrate.h> | |
12 | #include <linux/compaction.h> | |
13 | #include <linux/mm_inline.h> | |
14 | #include <linux/backing-dev.h> | |
76ab0f53 | 15 | #include <linux/sysctl.h> |
ed4a6d7f | 16 | #include <linux/sysfs.h> |
748446bb MG |
17 | #include "internal.h" |
18 | ||
ff9543fd MN |
19 | #if defined CONFIG_COMPACTION || defined CONFIG_CMA |
20 | ||
b7aba698 MG |
21 | #define CREATE_TRACE_POINTS |
22 | #include <trace/events/compaction.h> | |
23 | ||
748446bb MG |
24 | static unsigned long release_freepages(struct list_head *freelist) |
25 | { | |
26 | struct page *page, *next; | |
27 | unsigned long count = 0; | |
28 | ||
29 | list_for_each_entry_safe(page, next, freelist, lru) { | |
30 | list_del(&page->lru); | |
31 | __free_page(page); | |
32 | count++; | |
33 | } | |
34 | ||
35 | return count; | |
36 | } | |
37 | ||
ff9543fd MN |
38 | static void map_pages(struct list_head *list) |
39 | { | |
40 | struct page *page; | |
41 | ||
42 | list_for_each_entry(page, list, lru) { | |
43 | arch_alloc_page(page, 0); | |
44 | kernel_map_pages(page, 1, 1); | |
45 | } | |
46 | } | |
47 | ||
85aa125f MN |
48 | /* |
49 | * Isolate free pages onto a private freelist. Caller must hold zone->lock. | |
50 | * If @strict is true, will abort returning 0 on any invalid PFNs or non-free | |
51 | * pages inside of the pageblock (even though it may still end up isolating | |
52 | * some pages). | |
53 | */ | |
54 | static unsigned long isolate_freepages_block(unsigned long blockpfn, | |
55 | unsigned long end_pfn, | |
56 | struct list_head *freelist, | |
57 | bool strict) | |
748446bb | 58 | { |
b7aba698 | 59 | int nr_scanned = 0, total_isolated = 0; |
748446bb MG |
60 | struct page *cursor; |
61 | ||
748446bb MG |
62 | cursor = pfn_to_page(blockpfn); |
63 | ||
64 | /* Isolate free pages. This assumes the block is valid */ | |
65 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { | |
66 | int isolated, i; | |
67 | struct page *page = cursor; | |
68 | ||
85aa125f MN |
69 | if (!pfn_valid_within(blockpfn)) { |
70 | if (strict) | |
71 | return 0; | |
748446bb | 72 | continue; |
85aa125f | 73 | } |
b7aba698 | 74 | nr_scanned++; |
748446bb | 75 | |
85aa125f MN |
76 | if (!PageBuddy(page)) { |
77 | if (strict) | |
78 | return 0; | |
748446bb | 79 | continue; |
85aa125f | 80 | } |
748446bb MG |
81 | |
82 | /* Found a free page, break it into order-0 pages */ | |
83 | isolated = split_free_page(page); | |
85aa125f MN |
84 | if (!isolated && strict) |
85 | return 0; | |
748446bb MG |
86 | total_isolated += isolated; |
87 | for (i = 0; i < isolated; i++) { | |
88 | list_add(&page->lru, freelist); | |
89 | page++; | |
90 | } | |
91 | ||
92 | /* If a page was split, advance to the end of it */ | |
93 | if (isolated) { | |
94 | blockpfn += isolated - 1; | |
95 | cursor += isolated - 1; | |
96 | } | |
97 | } | |
98 | ||
b7aba698 | 99 | trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated); |
748446bb MG |
100 | return total_isolated; |
101 | } | |
102 | ||
85aa125f MN |
103 | /** |
104 | * isolate_freepages_range() - isolate free pages. | |
105 | * @start_pfn: The first PFN to start isolating. | |
106 | * @end_pfn: The one-past-last PFN. | |
107 | * | |
108 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
109 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
110 | * undo its actions and return zero. | |
111 | * | |
112 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
113 | * (which may be greater then end_pfn if end fell in a middle of | |
114 | * a free page). | |
115 | */ | |
ff9543fd | 116 | unsigned long |
85aa125f MN |
117 | isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn) |
118 | { | |
119 | unsigned long isolated, pfn, block_end_pfn, flags; | |
120 | struct zone *zone = NULL; | |
121 | LIST_HEAD(freelist); | |
122 | ||
123 | if (pfn_valid(start_pfn)) | |
124 | zone = page_zone(pfn_to_page(start_pfn)); | |
125 | ||
126 | for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) { | |
127 | if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn))) | |
128 | break; | |
129 | ||
130 | /* | |
131 | * On subsequent iterations ALIGN() is actually not needed, | |
132 | * but we keep it that we not to complicate the code. | |
133 | */ | |
134 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
135 | block_end_pfn = min(block_end_pfn, end_pfn); | |
136 | ||
137 | spin_lock_irqsave(&zone->lock, flags); | |
138 | isolated = isolate_freepages_block(pfn, block_end_pfn, | |
139 | &freelist, true); | |
140 | spin_unlock_irqrestore(&zone->lock, flags); | |
141 | ||
142 | /* | |
143 | * In strict mode, isolate_freepages_block() returns 0 if | |
144 | * there are any holes in the block (ie. invalid PFNs or | |
145 | * non-free pages). | |
146 | */ | |
147 | if (!isolated) | |
148 | break; | |
149 | ||
150 | /* | |
151 | * If we managed to isolate pages, it is always (1 << n) * | |
152 | * pageblock_nr_pages for some non-negative n. (Max order | |
153 | * page may span two pageblocks). | |
154 | */ | |
155 | } | |
156 | ||
157 | /* split_free_page does not map the pages */ | |
158 | map_pages(&freelist); | |
159 | ||
160 | if (pfn < end_pfn) { | |
161 | /* Loop terminated early, cleanup. */ | |
162 | release_freepages(&freelist); | |
163 | return 0; | |
164 | } | |
165 | ||
166 | /* We don't use freelists for anything. */ | |
167 | return pfn; | |
168 | } | |
169 | ||
748446bb MG |
170 | /* Update the number of anon and file isolated pages in the zone */ |
171 | static void acct_isolated(struct zone *zone, struct compact_control *cc) | |
172 | { | |
173 | struct page *page; | |
b9e84ac1 | 174 | unsigned int count[2] = { 0, }; |
748446bb | 175 | |
b9e84ac1 MK |
176 | list_for_each_entry(page, &cc->migratepages, lru) |
177 | count[!!page_is_file_cache(page)]++; | |
748446bb | 178 | |
b9e84ac1 MK |
179 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
180 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
181 | } |
182 | ||
183 | /* Similar to reclaim, but different enough that they don't share logic */ | |
184 | static bool too_many_isolated(struct zone *zone) | |
185 | { | |
bc693045 | 186 | unsigned long active, inactive, isolated; |
748446bb MG |
187 | |
188 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
189 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
190 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
191 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
192 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
193 | zone_page_state(zone, NR_ISOLATED_ANON); | |
194 | ||
bc693045 | 195 | return isolated > (inactive + active) / 2; |
748446bb MG |
196 | } |
197 | ||
2fe86e00 MN |
198 | /** |
199 | * isolate_migratepages_range() - isolate all migrate-able pages in range. | |
200 | * @zone: Zone pages are in. | |
201 | * @cc: Compaction control structure. | |
202 | * @low_pfn: The first PFN of the range. | |
203 | * @end_pfn: The one-past-the-last PFN of the range. | |
204 | * | |
205 | * Isolate all pages that can be migrated from the range specified by | |
206 | * [low_pfn, end_pfn). Returns zero if there is a fatal signal | |
207 | * pending), otherwise PFN of the first page that was not scanned | |
208 | * (which may be both less, equal to or more then end_pfn). | |
209 | * | |
210 | * Assumes that cc->migratepages is empty and cc->nr_migratepages is | |
211 | * zero. | |
212 | * | |
213 | * Apart from cc->migratepages and cc->nr_migratetypes this function | |
214 | * does not modify any cc's fields, in particular it does not modify | |
215 | * (or read for that matter) cc->migrate_pfn. | |
748446bb | 216 | */ |
ff9543fd | 217 | unsigned long |
2fe86e00 MN |
218 | isolate_migratepages_range(struct zone *zone, struct compact_control *cc, |
219 | unsigned long low_pfn, unsigned long end_pfn) | |
748446bb | 220 | { |
9927af74 | 221 | unsigned long last_pageblock_nr = 0, pageblock_nr; |
b7aba698 | 222 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 223 | struct list_head *migratelist = &cc->migratepages; |
39deaf85 | 224 | isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE; |
748446bb | 225 | |
748446bb MG |
226 | /* |
227 | * Ensure that there are not too many pages isolated from the LRU | |
228 | * list by either parallel reclaimers or compaction. If there are, | |
229 | * delay for some time until fewer pages are isolated | |
230 | */ | |
231 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b MG |
232 | /* async migration should just abort */ |
233 | if (!cc->sync) | |
2fe86e00 | 234 | return 0; |
f9e35b3b | 235 | |
748446bb MG |
236 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
237 | ||
238 | if (fatal_signal_pending(current)) | |
2fe86e00 | 239 | return 0; |
748446bb MG |
240 | } |
241 | ||
242 | /* Time to isolate some pages for migration */ | |
b2eef8c0 | 243 | cond_resched(); |
748446bb MG |
244 | spin_lock_irq(&zone->lru_lock); |
245 | for (; low_pfn < end_pfn; low_pfn++) { | |
246 | struct page *page; | |
b2eef8c0 AA |
247 | bool locked = true; |
248 | ||
249 | /* give a chance to irqs before checking need_resched() */ | |
250 | if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) { | |
251 | spin_unlock_irq(&zone->lru_lock); | |
252 | locked = false; | |
253 | } | |
254 | if (need_resched() || spin_is_contended(&zone->lru_lock)) { | |
255 | if (locked) | |
256 | spin_unlock_irq(&zone->lru_lock); | |
257 | cond_resched(); | |
258 | spin_lock_irq(&zone->lru_lock); | |
259 | if (fatal_signal_pending(current)) | |
260 | break; | |
261 | } else if (!locked) | |
262 | spin_lock_irq(&zone->lru_lock); | |
263 | ||
0bf380bc MG |
264 | /* |
265 | * migrate_pfn does not necessarily start aligned to a | |
266 | * pageblock. Ensure that pfn_valid is called when moving | |
267 | * into a new MAX_ORDER_NR_PAGES range in case of large | |
268 | * memory holes within the zone | |
269 | */ | |
270 | if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { | |
271 | if (!pfn_valid(low_pfn)) { | |
272 | low_pfn += MAX_ORDER_NR_PAGES - 1; | |
273 | continue; | |
274 | } | |
275 | } | |
276 | ||
748446bb MG |
277 | if (!pfn_valid_within(low_pfn)) |
278 | continue; | |
b7aba698 | 279 | nr_scanned++; |
748446bb | 280 | |
dc908600 MG |
281 | /* |
282 | * Get the page and ensure the page is within the same zone. | |
283 | * See the comment in isolate_freepages about overlapping | |
284 | * nodes. It is deliberate that the new zone lock is not taken | |
285 | * as memory compaction should not move pages between nodes. | |
286 | */ | |
748446bb | 287 | page = pfn_to_page(low_pfn); |
dc908600 MG |
288 | if (page_zone(page) != zone) |
289 | continue; | |
290 | ||
291 | /* Skip if free */ | |
748446bb MG |
292 | if (PageBuddy(page)) |
293 | continue; | |
294 | ||
9927af74 MG |
295 | /* |
296 | * For async migration, also only scan in MOVABLE blocks. Async | |
297 | * migration is optimistic to see if the minimum amount of work | |
298 | * satisfies the allocation | |
299 | */ | |
300 | pageblock_nr = low_pfn >> pageblock_order; | |
301 | if (!cc->sync && last_pageblock_nr != pageblock_nr && | |
302 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) { | |
303 | low_pfn += pageblock_nr_pages; | |
304 | low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1; | |
305 | last_pageblock_nr = pageblock_nr; | |
306 | continue; | |
307 | } | |
308 | ||
bc835011 AA |
309 | if (!PageLRU(page)) |
310 | continue; | |
311 | ||
312 | /* | |
313 | * PageLRU is set, and lru_lock excludes isolation, | |
314 | * splitting and collapsing (collapsing has already | |
315 | * happened if PageLRU is set). | |
316 | */ | |
317 | if (PageTransHuge(page)) { | |
318 | low_pfn += (1 << compound_order(page)) - 1; | |
319 | continue; | |
320 | } | |
321 | ||
c8244935 MG |
322 | if (!cc->sync) |
323 | mode |= ISOLATE_ASYNC_MIGRATE; | |
324 | ||
748446bb | 325 | /* Try isolate the page */ |
39deaf85 | 326 | if (__isolate_lru_page(page, mode, 0) != 0) |
748446bb MG |
327 | continue; |
328 | ||
bc835011 AA |
329 | VM_BUG_ON(PageTransCompound(page)); |
330 | ||
748446bb MG |
331 | /* Successfully isolated */ |
332 | del_page_from_lru_list(zone, page, page_lru(page)); | |
333 | list_add(&page->lru, migratelist); | |
748446bb | 334 | cc->nr_migratepages++; |
b7aba698 | 335 | nr_isolated++; |
748446bb MG |
336 | |
337 | /* Avoid isolating too much */ | |
31b8384a HD |
338 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
339 | ++low_pfn; | |
748446bb | 340 | break; |
31b8384a | 341 | } |
748446bb MG |
342 | } |
343 | ||
344 | acct_isolated(zone, cc); | |
345 | ||
346 | spin_unlock_irq(&zone->lru_lock); | |
748446bb | 347 | |
b7aba698 MG |
348 | trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated); |
349 | ||
2fe86e00 MN |
350 | return low_pfn; |
351 | } | |
352 | ||
ff9543fd MN |
353 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
354 | #ifdef CONFIG_COMPACTION | |
355 | ||
356 | /* Returns true if the page is within a block suitable for migration to */ | |
357 | static bool suitable_migration_target(struct page *page) | |
358 | { | |
359 | ||
360 | int migratetype = get_pageblock_migratetype(page); | |
361 | ||
362 | /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */ | |
363 | if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE) | |
364 | return false; | |
365 | ||
366 | /* If the page is a large free page, then allow migration */ | |
367 | if (PageBuddy(page) && page_order(page) >= pageblock_order) | |
368 | return true; | |
369 | ||
370 | /* If the block is MIGRATE_MOVABLE, allow migration */ | |
371 | if (migratetype == MIGRATE_MOVABLE) | |
372 | return true; | |
373 | ||
374 | /* Otherwise skip the block */ | |
375 | return false; | |
376 | } | |
377 | ||
2fe86e00 | 378 | /* |
ff9543fd MN |
379 | * Based on information in the current compact_control, find blocks |
380 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 381 | */ |
ff9543fd MN |
382 | static void isolate_freepages(struct zone *zone, |
383 | struct compact_control *cc) | |
2fe86e00 | 384 | { |
ff9543fd MN |
385 | struct page *page; |
386 | unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn; | |
387 | unsigned long flags; | |
388 | int nr_freepages = cc->nr_freepages; | |
389 | struct list_head *freelist = &cc->freepages; | |
2fe86e00 | 390 | |
ff9543fd MN |
391 | /* |
392 | * Initialise the free scanner. The starting point is where we last | |
393 | * scanned from (or the end of the zone if starting). The low point | |
394 | * is the end of the pageblock the migration scanner is using. | |
395 | */ | |
396 | pfn = cc->free_pfn; | |
397 | low_pfn = cc->migrate_pfn + pageblock_nr_pages; | |
2fe86e00 | 398 | |
ff9543fd MN |
399 | /* |
400 | * Take care that if the migration scanner is at the end of the zone | |
401 | * that the free scanner does not accidentally move to the next zone | |
402 | * in the next isolation cycle. | |
403 | */ | |
404 | high_pfn = min(low_pfn, pfn); | |
2fe86e00 | 405 | |
ff9543fd | 406 | zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages; |
2fe86e00 | 407 | |
ff9543fd MN |
408 | /* |
409 | * Isolate free pages until enough are available to migrate the | |
410 | * pages on cc->migratepages. We stop searching if the migrate | |
411 | * and free page scanners meet or enough free pages are isolated. | |
412 | */ | |
413 | for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages; | |
414 | pfn -= pageblock_nr_pages) { | |
415 | unsigned long isolated; | |
2fe86e00 | 416 | |
ff9543fd MN |
417 | if (!pfn_valid(pfn)) |
418 | continue; | |
2fe86e00 | 419 | |
ff9543fd MN |
420 | /* |
421 | * Check for overlapping nodes/zones. It's possible on some | |
422 | * configurations to have a setup like | |
423 | * node0 node1 node0 | |
424 | * i.e. it's possible that all pages within a zones range of | |
425 | * pages do not belong to a single zone. | |
426 | */ | |
427 | page = pfn_to_page(pfn); | |
428 | if (page_zone(page) != zone) | |
429 | continue; | |
430 | ||
431 | /* Check the block is suitable for migration */ | |
432 | if (!suitable_migration_target(page)) | |
433 | continue; | |
434 | ||
435 | /* | |
436 | * Found a block suitable for isolating free pages from. Now | |
437 | * we disabled interrupts, double check things are ok and | |
438 | * isolate the pages. This is to minimise the time IRQs | |
439 | * are disabled | |
440 | */ | |
441 | isolated = 0; | |
442 | spin_lock_irqsave(&zone->lock, flags); | |
443 | if (suitable_migration_target(page)) { | |
444 | end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn); | |
445 | isolated = isolate_freepages_block(pfn, end_pfn, | |
446 | freelist, false); | |
447 | nr_freepages += isolated; | |
448 | } | |
449 | spin_unlock_irqrestore(&zone->lock, flags); | |
450 | ||
451 | /* | |
452 | * Record the highest PFN we isolated pages from. When next | |
453 | * looking for free pages, the search will restart here as | |
454 | * page migration may have returned some pages to the allocator | |
455 | */ | |
456 | if (isolated) | |
457 | high_pfn = max(high_pfn, pfn); | |
458 | } | |
459 | ||
460 | /* split_free_page does not map the pages */ | |
461 | map_pages(freelist); | |
462 | ||
463 | cc->free_pfn = high_pfn; | |
464 | cc->nr_freepages = nr_freepages; | |
748446bb MG |
465 | } |
466 | ||
467 | /* | |
468 | * This is a migrate-callback that "allocates" freepages by taking pages | |
469 | * from the isolated freelists in the block we are migrating to. | |
470 | */ | |
471 | static struct page *compaction_alloc(struct page *migratepage, | |
472 | unsigned long data, | |
473 | int **result) | |
474 | { | |
475 | struct compact_control *cc = (struct compact_control *)data; | |
476 | struct page *freepage; | |
477 | ||
478 | /* Isolate free pages if necessary */ | |
479 | if (list_empty(&cc->freepages)) { | |
480 | isolate_freepages(cc->zone, cc); | |
481 | ||
482 | if (list_empty(&cc->freepages)) | |
483 | return NULL; | |
484 | } | |
485 | ||
486 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
487 | list_del(&freepage->lru); | |
488 | cc->nr_freepages--; | |
489 | ||
490 | return freepage; | |
491 | } | |
492 | ||
493 | /* | |
494 | * We cannot control nr_migratepages and nr_freepages fully when migration is | |
495 | * running as migrate_pages() has no knowledge of compact_control. When | |
496 | * migration is complete, we count the number of pages on the lists by hand. | |
497 | */ | |
498 | static void update_nr_listpages(struct compact_control *cc) | |
499 | { | |
500 | int nr_migratepages = 0; | |
501 | int nr_freepages = 0; | |
502 | struct page *page; | |
503 | ||
504 | list_for_each_entry(page, &cc->migratepages, lru) | |
505 | nr_migratepages++; | |
506 | list_for_each_entry(page, &cc->freepages, lru) | |
507 | nr_freepages++; | |
508 | ||
509 | cc->nr_migratepages = nr_migratepages; | |
510 | cc->nr_freepages = nr_freepages; | |
511 | } | |
512 | ||
ff9543fd MN |
513 | /* possible outcome of isolate_migratepages */ |
514 | typedef enum { | |
515 | ISOLATE_ABORT, /* Abort compaction now */ | |
516 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
517 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
518 | } isolate_migrate_t; | |
519 | ||
520 | /* | |
521 | * Isolate all pages that can be migrated from the block pointed to by | |
522 | * the migrate scanner within compact_control. | |
523 | */ | |
524 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
525 | struct compact_control *cc) | |
526 | { | |
527 | unsigned long low_pfn, end_pfn; | |
528 | ||
529 | /* Do not scan outside zone boundaries */ | |
530 | low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); | |
531 | ||
532 | /* Only scan within a pageblock boundary */ | |
533 | end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); | |
534 | ||
535 | /* Do not cross the free scanner or scan within a memory hole */ | |
536 | if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { | |
537 | cc->migrate_pfn = end_pfn; | |
538 | return ISOLATE_NONE; | |
539 | } | |
540 | ||
541 | /* Perform the isolation */ | |
542 | low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn); | |
543 | if (!low_pfn) | |
544 | return ISOLATE_ABORT; | |
545 | ||
546 | cc->migrate_pfn = low_pfn; | |
547 | ||
548 | return ISOLATE_SUCCESS; | |
549 | } | |
550 | ||
748446bb | 551 | static int compact_finished(struct zone *zone, |
5a03b051 | 552 | struct compact_control *cc) |
748446bb | 553 | { |
56de7263 | 554 | unsigned int order; |
5a03b051 | 555 | unsigned long watermark; |
56de7263 | 556 | |
748446bb MG |
557 | if (fatal_signal_pending(current)) |
558 | return COMPACT_PARTIAL; | |
559 | ||
560 | /* Compaction run completes if the migrate and free scanner meet */ | |
561 | if (cc->free_pfn <= cc->migrate_pfn) | |
562 | return COMPACT_COMPLETE; | |
563 | ||
82478fb7 JW |
564 | /* |
565 | * order == -1 is expected when compacting via | |
566 | * /proc/sys/vm/compact_memory | |
567 | */ | |
56de7263 MG |
568 | if (cc->order == -1) |
569 | return COMPACT_CONTINUE; | |
570 | ||
3957c776 MH |
571 | /* Compaction run is not finished if the watermark is not met */ |
572 | watermark = low_wmark_pages(zone); | |
573 | watermark += (1 << cc->order); | |
574 | ||
575 | if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) | |
576 | return COMPACT_CONTINUE; | |
577 | ||
56de7263 MG |
578 | /* Direct compactor: Is a suitable page free? */ |
579 | for (order = cc->order; order < MAX_ORDER; order++) { | |
580 | /* Job done if page is free of the right migratetype */ | |
581 | if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) | |
582 | return COMPACT_PARTIAL; | |
583 | ||
584 | /* Job done if allocation would set block type */ | |
585 | if (order >= pageblock_order && zone->free_area[order].nr_free) | |
586 | return COMPACT_PARTIAL; | |
587 | } | |
588 | ||
748446bb MG |
589 | return COMPACT_CONTINUE; |
590 | } | |
591 | ||
3e7d3449 MG |
592 | /* |
593 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
594 | * Returns | |
595 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
596 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
597 | * COMPACT_CONTINUE - If compaction should run now | |
598 | */ | |
599 | unsigned long compaction_suitable(struct zone *zone, int order) | |
600 | { | |
601 | int fragindex; | |
602 | unsigned long watermark; | |
603 | ||
3957c776 MH |
604 | /* |
605 | * order == -1 is expected when compacting via | |
606 | * /proc/sys/vm/compact_memory | |
607 | */ | |
608 | if (order == -1) | |
609 | return COMPACT_CONTINUE; | |
610 | ||
3e7d3449 MG |
611 | /* |
612 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
613 | * This is because during migration, copies of pages need to be | |
614 | * allocated and for a short time, the footprint is higher | |
615 | */ | |
616 | watermark = low_wmark_pages(zone) + (2UL << order); | |
617 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
618 | return COMPACT_SKIPPED; | |
619 | ||
620 | /* | |
621 | * fragmentation index determines if allocation failures are due to | |
622 | * low memory or external fragmentation | |
623 | * | |
a582a738 SL |
624 | * index of -1000 implies allocations might succeed depending on |
625 | * watermarks | |
3e7d3449 MG |
626 | * index towards 0 implies failure is due to lack of memory |
627 | * index towards 1000 implies failure is due to fragmentation | |
628 | * | |
629 | * Only compact if a failure would be due to fragmentation. | |
630 | */ | |
631 | fragindex = fragmentation_index(zone, order); | |
632 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
633 | return COMPACT_SKIPPED; | |
634 | ||
a582a738 SL |
635 | if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark, |
636 | 0, 0)) | |
3e7d3449 MG |
637 | return COMPACT_PARTIAL; |
638 | ||
639 | return COMPACT_CONTINUE; | |
640 | } | |
641 | ||
748446bb MG |
642 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
643 | { | |
644 | int ret; | |
645 | ||
3e7d3449 MG |
646 | ret = compaction_suitable(zone, cc->order); |
647 | switch (ret) { | |
648 | case COMPACT_PARTIAL: | |
649 | case COMPACT_SKIPPED: | |
650 | /* Compaction is likely to fail */ | |
651 | return ret; | |
652 | case COMPACT_CONTINUE: | |
653 | /* Fall through to compaction */ | |
654 | ; | |
655 | } | |
656 | ||
748446bb MG |
657 | /* Setup to move all movable pages to the end of the zone */ |
658 | cc->migrate_pfn = zone->zone_start_pfn; | |
659 | cc->free_pfn = cc->migrate_pfn + zone->spanned_pages; | |
660 | cc->free_pfn &= ~(pageblock_nr_pages-1); | |
661 | ||
662 | migrate_prep_local(); | |
663 | ||
664 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { | |
665 | unsigned long nr_migrate, nr_remaining; | |
9d502c1c | 666 | int err; |
748446bb | 667 | |
f9e35b3b MG |
668 | switch (isolate_migratepages(zone, cc)) { |
669 | case ISOLATE_ABORT: | |
670 | ret = COMPACT_PARTIAL; | |
671 | goto out; | |
672 | case ISOLATE_NONE: | |
748446bb | 673 | continue; |
f9e35b3b MG |
674 | case ISOLATE_SUCCESS: |
675 | ; | |
676 | } | |
748446bb MG |
677 | |
678 | nr_migrate = cc->nr_migratepages; | |
9d502c1c | 679 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
7f0f2496 | 680 | (unsigned long)cc, false, |
a6bc32b8 | 681 | cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC); |
748446bb MG |
682 | update_nr_listpages(cc); |
683 | nr_remaining = cc->nr_migratepages; | |
684 | ||
685 | count_vm_event(COMPACTBLOCKS); | |
686 | count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); | |
687 | if (nr_remaining) | |
688 | count_vm_events(COMPACTPAGEFAILED, nr_remaining); | |
b7aba698 MG |
689 | trace_mm_compaction_migratepages(nr_migrate - nr_remaining, |
690 | nr_remaining); | |
748446bb MG |
691 | |
692 | /* Release LRU pages not migrated */ | |
9d502c1c | 693 | if (err) { |
748446bb MG |
694 | putback_lru_pages(&cc->migratepages); |
695 | cc->nr_migratepages = 0; | |
696 | } | |
697 | ||
698 | } | |
699 | ||
f9e35b3b | 700 | out: |
748446bb MG |
701 | /* Release free pages and check accounting */ |
702 | cc->nr_freepages -= release_freepages(&cc->freepages); | |
703 | VM_BUG_ON(cc->nr_freepages != 0); | |
704 | ||
705 | return ret; | |
706 | } | |
76ab0f53 | 707 | |
d43a87e6 | 708 | static unsigned long compact_zone_order(struct zone *zone, |
5a03b051 | 709 | int order, gfp_t gfp_mask, |
d527caf2 | 710 | bool sync) |
56de7263 MG |
711 | { |
712 | struct compact_control cc = { | |
713 | .nr_freepages = 0, | |
714 | .nr_migratepages = 0, | |
715 | .order = order, | |
716 | .migratetype = allocflags_to_migratetype(gfp_mask), | |
717 | .zone = zone, | |
77f1fe6b | 718 | .sync = sync, |
56de7263 MG |
719 | }; |
720 | INIT_LIST_HEAD(&cc.freepages); | |
721 | INIT_LIST_HEAD(&cc.migratepages); | |
722 | ||
723 | return compact_zone(zone, &cc); | |
724 | } | |
725 | ||
5e771905 MG |
726 | int sysctl_extfrag_threshold = 500; |
727 | ||
56de7263 MG |
728 | /** |
729 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
730 | * @zonelist: The zonelist used for the current allocation | |
731 | * @order: The order of the current allocation | |
732 | * @gfp_mask: The GFP mask of the current allocation | |
733 | * @nodemask: The allowed nodes to allocate from | |
77f1fe6b | 734 | * @sync: Whether migration is synchronous or not |
56de7263 MG |
735 | * |
736 | * This is the main entry point for direct page compaction. | |
737 | */ | |
738 | unsigned long try_to_compact_pages(struct zonelist *zonelist, | |
77f1fe6b MG |
739 | int order, gfp_t gfp_mask, nodemask_t *nodemask, |
740 | bool sync) | |
56de7263 MG |
741 | { |
742 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
743 | int may_enter_fs = gfp_mask & __GFP_FS; | |
744 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
745 | struct zoneref *z; |
746 | struct zone *zone; | |
747 | int rc = COMPACT_SKIPPED; | |
748 | ||
749 | /* | |
750 | * Check whether it is worth even starting compaction. The order check is | |
751 | * made because an assumption is made that the page allocator can satisfy | |
752 | * the "cheaper" orders without taking special steps | |
753 | */ | |
c5a73c3d | 754 | if (!order || !may_enter_fs || !may_perform_io) |
56de7263 MG |
755 | return rc; |
756 | ||
757 | count_vm_event(COMPACTSTALL); | |
758 | ||
759 | /* Compact each zone in the list */ | |
760 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, | |
761 | nodemask) { | |
56de7263 MG |
762 | int status; |
763 | ||
d527caf2 | 764 | status = compact_zone_order(zone, order, gfp_mask, sync); |
56de7263 MG |
765 | rc = max(status, rc); |
766 | ||
3e7d3449 MG |
767 | /* If a normal allocation would succeed, stop compacting */ |
768 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) | |
56de7263 MG |
769 | break; |
770 | } | |
771 | ||
772 | return rc; | |
773 | } | |
774 | ||
775 | ||
76ab0f53 | 776 | /* Compact all zones within a node */ |
7be62de9 | 777 | static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
778 | { |
779 | int zoneid; | |
76ab0f53 MG |
780 | struct zone *zone; |
781 | ||
76ab0f53 | 782 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
783 | |
784 | zone = &pgdat->node_zones[zoneid]; | |
785 | if (!populated_zone(zone)) | |
786 | continue; | |
787 | ||
7be62de9 RR |
788 | cc->nr_freepages = 0; |
789 | cc->nr_migratepages = 0; | |
790 | cc->zone = zone; | |
791 | INIT_LIST_HEAD(&cc->freepages); | |
792 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 793 | |
aad6ec37 | 794 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
7be62de9 | 795 | compact_zone(zone, cc); |
76ab0f53 | 796 | |
aff62249 RR |
797 | if (cc->order > 0) { |
798 | int ok = zone_watermark_ok(zone, cc->order, | |
799 | low_wmark_pages(zone), 0, 0); | |
800 | if (ok && cc->order > zone->compact_order_failed) | |
801 | zone->compact_order_failed = cc->order + 1; | |
802 | /* Currently async compaction is never deferred. */ | |
803 | else if (!ok && cc->sync) | |
804 | defer_compaction(zone, cc->order); | |
805 | } | |
806 | ||
7be62de9 RR |
807 | VM_BUG_ON(!list_empty(&cc->freepages)); |
808 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 MG |
809 | } |
810 | ||
811 | return 0; | |
812 | } | |
813 | ||
7be62de9 RR |
814 | int compact_pgdat(pg_data_t *pgdat, int order) |
815 | { | |
816 | struct compact_control cc = { | |
817 | .order = order, | |
818 | .sync = false, | |
819 | }; | |
820 | ||
821 | return __compact_pgdat(pgdat, &cc); | |
822 | } | |
823 | ||
824 | static int compact_node(int nid) | |
825 | { | |
7be62de9 RR |
826 | struct compact_control cc = { |
827 | .order = -1, | |
828 | .sync = true, | |
829 | }; | |
830 | ||
8575ec29 | 831 | return __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
832 | } |
833 | ||
76ab0f53 MG |
834 | /* Compact all nodes in the system */ |
835 | static int compact_nodes(void) | |
836 | { | |
837 | int nid; | |
838 | ||
8575ec29 HD |
839 | /* Flush pending updates to the LRU lists */ |
840 | lru_add_drain_all(); | |
841 | ||
76ab0f53 MG |
842 | for_each_online_node(nid) |
843 | compact_node(nid); | |
844 | ||
845 | return COMPACT_COMPLETE; | |
846 | } | |
847 | ||
848 | /* The written value is actually unused, all memory is compacted */ | |
849 | int sysctl_compact_memory; | |
850 | ||
851 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
852 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
853 | void __user *buffer, size_t *length, loff_t *ppos) | |
854 | { | |
855 | if (write) | |
856 | return compact_nodes(); | |
857 | ||
858 | return 0; | |
859 | } | |
ed4a6d7f | 860 | |
5e771905 MG |
861 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
862 | void __user *buffer, size_t *length, loff_t *ppos) | |
863 | { | |
864 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
865 | ||
866 | return 0; | |
867 | } | |
868 | ||
ed4a6d7f | 869 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
10fbcf4c KS |
870 | ssize_t sysfs_compact_node(struct device *dev, |
871 | struct device_attribute *attr, | |
ed4a6d7f MG |
872 | const char *buf, size_t count) |
873 | { | |
8575ec29 HD |
874 | int nid = dev->id; |
875 | ||
876 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
877 | /* Flush pending updates to the LRU lists */ | |
878 | lru_add_drain_all(); | |
879 | ||
880 | compact_node(nid); | |
881 | } | |
ed4a6d7f MG |
882 | |
883 | return count; | |
884 | } | |
10fbcf4c | 885 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
886 | |
887 | int compaction_register_node(struct node *node) | |
888 | { | |
10fbcf4c | 889 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
890 | } |
891 | ||
892 | void compaction_unregister_node(struct node *node) | |
893 | { | |
10fbcf4c | 894 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
895 | } |
896 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd MN |
897 | |
898 | #endif /* CONFIG_COMPACTION */ |