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 | ||
47118af0 MN |
48 | static inline bool migrate_async_suitable(int migratetype) |
49 | { | |
50 | return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE; | |
51 | } | |
52 | ||
85aa125f MN |
53 | /* |
54 | * Isolate free pages onto a private freelist. Caller must hold zone->lock. | |
55 | * If @strict is true, will abort returning 0 on any invalid PFNs or non-free | |
56 | * pages inside of the pageblock (even though it may still end up isolating | |
57 | * some pages). | |
58 | */ | |
59 | static unsigned long isolate_freepages_block(unsigned long blockpfn, | |
60 | unsigned long end_pfn, | |
61 | struct list_head *freelist, | |
62 | bool strict) | |
748446bb | 63 | { |
b7aba698 | 64 | int nr_scanned = 0, total_isolated = 0; |
748446bb MG |
65 | struct page *cursor; |
66 | ||
748446bb MG |
67 | cursor = pfn_to_page(blockpfn); |
68 | ||
69 | /* Isolate free pages. This assumes the block is valid */ | |
70 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { | |
71 | int isolated, i; | |
72 | struct page *page = cursor; | |
73 | ||
85aa125f MN |
74 | if (!pfn_valid_within(blockpfn)) { |
75 | if (strict) | |
76 | return 0; | |
748446bb | 77 | continue; |
85aa125f | 78 | } |
b7aba698 | 79 | nr_scanned++; |
748446bb | 80 | |
85aa125f MN |
81 | if (!PageBuddy(page)) { |
82 | if (strict) | |
83 | return 0; | |
748446bb | 84 | continue; |
85aa125f | 85 | } |
748446bb MG |
86 | |
87 | /* Found a free page, break it into order-0 pages */ | |
88 | isolated = split_free_page(page); | |
85aa125f MN |
89 | if (!isolated && strict) |
90 | return 0; | |
748446bb MG |
91 | total_isolated += isolated; |
92 | for (i = 0; i < isolated; i++) { | |
93 | list_add(&page->lru, freelist); | |
94 | page++; | |
95 | } | |
96 | ||
97 | /* If a page was split, advance to the end of it */ | |
98 | if (isolated) { | |
99 | blockpfn += isolated - 1; | |
100 | cursor += isolated - 1; | |
101 | } | |
102 | } | |
103 | ||
b7aba698 | 104 | trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated); |
748446bb MG |
105 | return total_isolated; |
106 | } | |
107 | ||
85aa125f MN |
108 | /** |
109 | * isolate_freepages_range() - isolate free pages. | |
110 | * @start_pfn: The first PFN to start isolating. | |
111 | * @end_pfn: The one-past-last PFN. | |
112 | * | |
113 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
114 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
115 | * undo its actions and return zero. | |
116 | * | |
117 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
118 | * (which may be greater then end_pfn if end fell in a middle of | |
119 | * a free page). | |
120 | */ | |
ff9543fd | 121 | unsigned long |
85aa125f MN |
122 | isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn) |
123 | { | |
124 | unsigned long isolated, pfn, block_end_pfn, flags; | |
125 | struct zone *zone = NULL; | |
126 | LIST_HEAD(freelist); | |
127 | ||
128 | if (pfn_valid(start_pfn)) | |
129 | zone = page_zone(pfn_to_page(start_pfn)); | |
130 | ||
131 | for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) { | |
132 | if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn))) | |
133 | break; | |
134 | ||
135 | /* | |
136 | * On subsequent iterations ALIGN() is actually not needed, | |
137 | * but we keep it that we not to complicate the code. | |
138 | */ | |
139 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
140 | block_end_pfn = min(block_end_pfn, end_pfn); | |
141 | ||
142 | spin_lock_irqsave(&zone->lock, flags); | |
143 | isolated = isolate_freepages_block(pfn, block_end_pfn, | |
144 | &freelist, true); | |
145 | spin_unlock_irqrestore(&zone->lock, flags); | |
146 | ||
147 | /* | |
148 | * In strict mode, isolate_freepages_block() returns 0 if | |
149 | * there are any holes in the block (ie. invalid PFNs or | |
150 | * non-free pages). | |
151 | */ | |
152 | if (!isolated) | |
153 | break; | |
154 | ||
155 | /* | |
156 | * If we managed to isolate pages, it is always (1 << n) * | |
157 | * pageblock_nr_pages for some non-negative n. (Max order | |
158 | * page may span two pageblocks). | |
159 | */ | |
160 | } | |
161 | ||
162 | /* split_free_page does not map the pages */ | |
163 | map_pages(&freelist); | |
164 | ||
165 | if (pfn < end_pfn) { | |
166 | /* Loop terminated early, cleanup. */ | |
167 | release_freepages(&freelist); | |
168 | return 0; | |
169 | } | |
170 | ||
171 | /* We don't use freelists for anything. */ | |
172 | return pfn; | |
173 | } | |
174 | ||
748446bb MG |
175 | /* Update the number of anon and file isolated pages in the zone */ |
176 | static void acct_isolated(struct zone *zone, struct compact_control *cc) | |
177 | { | |
178 | struct page *page; | |
b9e84ac1 | 179 | unsigned int count[2] = { 0, }; |
748446bb | 180 | |
b9e84ac1 MK |
181 | list_for_each_entry(page, &cc->migratepages, lru) |
182 | count[!!page_is_file_cache(page)]++; | |
748446bb | 183 | |
b9e84ac1 MK |
184 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
185 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
186 | } |
187 | ||
188 | /* Similar to reclaim, but different enough that they don't share logic */ | |
189 | static bool too_many_isolated(struct zone *zone) | |
190 | { | |
bc693045 | 191 | unsigned long active, inactive, isolated; |
748446bb MG |
192 | |
193 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
194 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
195 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
196 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
197 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
198 | zone_page_state(zone, NR_ISOLATED_ANON); | |
199 | ||
bc693045 | 200 | return isolated > (inactive + active) / 2; |
748446bb MG |
201 | } |
202 | ||
2fe86e00 MN |
203 | /** |
204 | * isolate_migratepages_range() - isolate all migrate-able pages in range. | |
205 | * @zone: Zone pages are in. | |
206 | * @cc: Compaction control structure. | |
207 | * @low_pfn: The first PFN of the range. | |
208 | * @end_pfn: The one-past-the-last PFN of the range. | |
209 | * | |
210 | * Isolate all pages that can be migrated from the range specified by | |
211 | * [low_pfn, end_pfn). Returns zero if there is a fatal signal | |
212 | * pending), otherwise PFN of the first page that was not scanned | |
213 | * (which may be both less, equal to or more then end_pfn). | |
214 | * | |
215 | * Assumes that cc->migratepages is empty and cc->nr_migratepages is | |
216 | * zero. | |
217 | * | |
218 | * Apart from cc->migratepages and cc->nr_migratetypes this function | |
219 | * does not modify any cc's fields, in particular it does not modify | |
220 | * (or read for that matter) cc->migrate_pfn. | |
748446bb | 221 | */ |
ff9543fd | 222 | unsigned long |
2fe86e00 MN |
223 | isolate_migratepages_range(struct zone *zone, struct compact_control *cc, |
224 | unsigned long low_pfn, unsigned long end_pfn) | |
748446bb | 225 | { |
9927af74 | 226 | unsigned long last_pageblock_nr = 0, pageblock_nr; |
b7aba698 | 227 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 228 | struct list_head *migratelist = &cc->migratepages; |
39deaf85 | 229 | isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE; |
748446bb | 230 | |
748446bb MG |
231 | /* |
232 | * Ensure that there are not too many pages isolated from the LRU | |
233 | * list by either parallel reclaimers or compaction. If there are, | |
234 | * delay for some time until fewer pages are isolated | |
235 | */ | |
236 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b MG |
237 | /* async migration should just abort */ |
238 | if (!cc->sync) | |
2fe86e00 | 239 | return 0; |
f9e35b3b | 240 | |
748446bb MG |
241 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
242 | ||
243 | if (fatal_signal_pending(current)) | |
2fe86e00 | 244 | return 0; |
748446bb MG |
245 | } |
246 | ||
247 | /* Time to isolate some pages for migration */ | |
b2eef8c0 | 248 | cond_resched(); |
748446bb MG |
249 | spin_lock_irq(&zone->lru_lock); |
250 | for (; low_pfn < end_pfn; low_pfn++) { | |
251 | struct page *page; | |
b2eef8c0 AA |
252 | bool locked = true; |
253 | ||
254 | /* give a chance to irqs before checking need_resched() */ | |
255 | if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) { | |
256 | spin_unlock_irq(&zone->lru_lock); | |
257 | locked = false; | |
258 | } | |
259 | if (need_resched() || spin_is_contended(&zone->lru_lock)) { | |
260 | if (locked) | |
261 | spin_unlock_irq(&zone->lru_lock); | |
262 | cond_resched(); | |
263 | spin_lock_irq(&zone->lru_lock); | |
264 | if (fatal_signal_pending(current)) | |
265 | break; | |
266 | } else if (!locked) | |
267 | spin_lock_irq(&zone->lru_lock); | |
268 | ||
0bf380bc MG |
269 | /* |
270 | * migrate_pfn does not necessarily start aligned to a | |
271 | * pageblock. Ensure that pfn_valid is called when moving | |
272 | * into a new MAX_ORDER_NR_PAGES range in case of large | |
273 | * memory holes within the zone | |
274 | */ | |
275 | if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { | |
276 | if (!pfn_valid(low_pfn)) { | |
277 | low_pfn += MAX_ORDER_NR_PAGES - 1; | |
278 | continue; | |
279 | } | |
280 | } | |
281 | ||
748446bb MG |
282 | if (!pfn_valid_within(low_pfn)) |
283 | continue; | |
b7aba698 | 284 | nr_scanned++; |
748446bb | 285 | |
dc908600 MG |
286 | /* |
287 | * Get the page and ensure the page is within the same zone. | |
288 | * See the comment in isolate_freepages about overlapping | |
289 | * nodes. It is deliberate that the new zone lock is not taken | |
290 | * as memory compaction should not move pages between nodes. | |
291 | */ | |
748446bb | 292 | page = pfn_to_page(low_pfn); |
dc908600 MG |
293 | if (page_zone(page) != zone) |
294 | continue; | |
295 | ||
296 | /* Skip if free */ | |
748446bb MG |
297 | if (PageBuddy(page)) |
298 | continue; | |
299 | ||
9927af74 MG |
300 | /* |
301 | * For async migration, also only scan in MOVABLE blocks. Async | |
302 | * migration is optimistic to see if the minimum amount of work | |
303 | * satisfies the allocation | |
304 | */ | |
305 | pageblock_nr = low_pfn >> pageblock_order; | |
306 | if (!cc->sync && last_pageblock_nr != pageblock_nr && | |
47118af0 | 307 | !migrate_async_suitable(get_pageblock_migratetype(page))) { |
9927af74 MG |
308 | low_pfn += pageblock_nr_pages; |
309 | low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1; | |
310 | last_pageblock_nr = pageblock_nr; | |
311 | continue; | |
312 | } | |
313 | ||
bc835011 AA |
314 | if (!PageLRU(page)) |
315 | continue; | |
316 | ||
317 | /* | |
318 | * PageLRU is set, and lru_lock excludes isolation, | |
319 | * splitting and collapsing (collapsing has already | |
320 | * happened if PageLRU is set). | |
321 | */ | |
322 | if (PageTransHuge(page)) { | |
323 | low_pfn += (1 << compound_order(page)) - 1; | |
324 | continue; | |
325 | } | |
326 | ||
c8244935 MG |
327 | if (!cc->sync) |
328 | mode |= ISOLATE_ASYNC_MIGRATE; | |
329 | ||
748446bb | 330 | /* Try isolate the page */ |
39deaf85 | 331 | if (__isolate_lru_page(page, mode, 0) != 0) |
748446bb MG |
332 | continue; |
333 | ||
bc835011 AA |
334 | VM_BUG_ON(PageTransCompound(page)); |
335 | ||
748446bb MG |
336 | /* Successfully isolated */ |
337 | del_page_from_lru_list(zone, page, page_lru(page)); | |
338 | list_add(&page->lru, migratelist); | |
748446bb | 339 | cc->nr_migratepages++; |
b7aba698 | 340 | nr_isolated++; |
748446bb MG |
341 | |
342 | /* Avoid isolating too much */ | |
31b8384a HD |
343 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
344 | ++low_pfn; | |
748446bb | 345 | break; |
31b8384a | 346 | } |
748446bb MG |
347 | } |
348 | ||
349 | acct_isolated(zone, cc); | |
350 | ||
351 | spin_unlock_irq(&zone->lru_lock); | |
748446bb | 352 | |
b7aba698 MG |
353 | trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated); |
354 | ||
2fe86e00 MN |
355 | return low_pfn; |
356 | } | |
357 | ||
ff9543fd MN |
358 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
359 | #ifdef CONFIG_COMPACTION | |
360 | ||
361 | /* Returns true if the page is within a block suitable for migration to */ | |
362 | static bool suitable_migration_target(struct page *page) | |
363 | { | |
364 | ||
365 | int migratetype = get_pageblock_migratetype(page); | |
366 | ||
367 | /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */ | |
368 | if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE) | |
369 | return false; | |
370 | ||
371 | /* If the page is a large free page, then allow migration */ | |
372 | if (PageBuddy(page) && page_order(page) >= pageblock_order) | |
373 | return true; | |
374 | ||
47118af0 MN |
375 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
376 | if (migrate_async_suitable(migratetype)) | |
ff9543fd MN |
377 | return true; |
378 | ||
379 | /* Otherwise skip the block */ | |
380 | return false; | |
381 | } | |
382 | ||
2fe86e00 | 383 | /* |
ff9543fd MN |
384 | * Based on information in the current compact_control, find blocks |
385 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 386 | */ |
ff9543fd MN |
387 | static void isolate_freepages(struct zone *zone, |
388 | struct compact_control *cc) | |
2fe86e00 | 389 | { |
ff9543fd MN |
390 | struct page *page; |
391 | unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn; | |
392 | unsigned long flags; | |
393 | int nr_freepages = cc->nr_freepages; | |
394 | struct list_head *freelist = &cc->freepages; | |
2fe86e00 | 395 | |
ff9543fd MN |
396 | /* |
397 | * Initialise the free scanner. The starting point is where we last | |
398 | * scanned from (or the end of the zone if starting). The low point | |
399 | * is the end of the pageblock the migration scanner is using. | |
400 | */ | |
401 | pfn = cc->free_pfn; | |
402 | low_pfn = cc->migrate_pfn + pageblock_nr_pages; | |
2fe86e00 | 403 | |
ff9543fd MN |
404 | /* |
405 | * Take care that if the migration scanner is at the end of the zone | |
406 | * that the free scanner does not accidentally move to the next zone | |
407 | * in the next isolation cycle. | |
408 | */ | |
409 | high_pfn = min(low_pfn, pfn); | |
2fe86e00 | 410 | |
ff9543fd | 411 | zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages; |
2fe86e00 | 412 | |
ff9543fd MN |
413 | /* |
414 | * Isolate free pages until enough are available to migrate the | |
415 | * pages on cc->migratepages. We stop searching if the migrate | |
416 | * and free page scanners meet or enough free pages are isolated. | |
417 | */ | |
418 | for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages; | |
419 | pfn -= pageblock_nr_pages) { | |
420 | unsigned long isolated; | |
2fe86e00 | 421 | |
ff9543fd MN |
422 | if (!pfn_valid(pfn)) |
423 | continue; | |
2fe86e00 | 424 | |
ff9543fd MN |
425 | /* |
426 | * Check for overlapping nodes/zones. It's possible on some | |
427 | * configurations to have a setup like | |
428 | * node0 node1 node0 | |
429 | * i.e. it's possible that all pages within a zones range of | |
430 | * pages do not belong to a single zone. | |
431 | */ | |
432 | page = pfn_to_page(pfn); | |
433 | if (page_zone(page) != zone) | |
434 | continue; | |
435 | ||
436 | /* Check the block is suitable for migration */ | |
437 | if (!suitable_migration_target(page)) | |
438 | continue; | |
439 | ||
440 | /* | |
441 | * Found a block suitable for isolating free pages from. Now | |
442 | * we disabled interrupts, double check things are ok and | |
443 | * isolate the pages. This is to minimise the time IRQs | |
444 | * are disabled | |
445 | */ | |
446 | isolated = 0; | |
447 | spin_lock_irqsave(&zone->lock, flags); | |
448 | if (suitable_migration_target(page)) { | |
449 | end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn); | |
450 | isolated = isolate_freepages_block(pfn, end_pfn, | |
451 | freelist, false); | |
452 | nr_freepages += isolated; | |
453 | } | |
454 | spin_unlock_irqrestore(&zone->lock, flags); | |
455 | ||
456 | /* | |
457 | * Record the highest PFN we isolated pages from. When next | |
458 | * looking for free pages, the search will restart here as | |
459 | * page migration may have returned some pages to the allocator | |
460 | */ | |
461 | if (isolated) | |
462 | high_pfn = max(high_pfn, pfn); | |
463 | } | |
464 | ||
465 | /* split_free_page does not map the pages */ | |
466 | map_pages(freelist); | |
467 | ||
468 | cc->free_pfn = high_pfn; | |
469 | cc->nr_freepages = nr_freepages; | |
748446bb MG |
470 | } |
471 | ||
472 | /* | |
473 | * This is a migrate-callback that "allocates" freepages by taking pages | |
474 | * from the isolated freelists in the block we are migrating to. | |
475 | */ | |
476 | static struct page *compaction_alloc(struct page *migratepage, | |
477 | unsigned long data, | |
478 | int **result) | |
479 | { | |
480 | struct compact_control *cc = (struct compact_control *)data; | |
481 | struct page *freepage; | |
482 | ||
483 | /* Isolate free pages if necessary */ | |
484 | if (list_empty(&cc->freepages)) { | |
485 | isolate_freepages(cc->zone, cc); | |
486 | ||
487 | if (list_empty(&cc->freepages)) | |
488 | return NULL; | |
489 | } | |
490 | ||
491 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
492 | list_del(&freepage->lru); | |
493 | cc->nr_freepages--; | |
494 | ||
495 | return freepage; | |
496 | } | |
497 | ||
498 | /* | |
499 | * We cannot control nr_migratepages and nr_freepages fully when migration is | |
500 | * running as migrate_pages() has no knowledge of compact_control. When | |
501 | * migration is complete, we count the number of pages on the lists by hand. | |
502 | */ | |
503 | static void update_nr_listpages(struct compact_control *cc) | |
504 | { | |
505 | int nr_migratepages = 0; | |
506 | int nr_freepages = 0; | |
507 | struct page *page; | |
508 | ||
509 | list_for_each_entry(page, &cc->migratepages, lru) | |
510 | nr_migratepages++; | |
511 | list_for_each_entry(page, &cc->freepages, lru) | |
512 | nr_freepages++; | |
513 | ||
514 | cc->nr_migratepages = nr_migratepages; | |
515 | cc->nr_freepages = nr_freepages; | |
516 | } | |
517 | ||
ff9543fd MN |
518 | /* possible outcome of isolate_migratepages */ |
519 | typedef enum { | |
520 | ISOLATE_ABORT, /* Abort compaction now */ | |
521 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
522 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
523 | } isolate_migrate_t; | |
524 | ||
525 | /* | |
526 | * Isolate all pages that can be migrated from the block pointed to by | |
527 | * the migrate scanner within compact_control. | |
528 | */ | |
529 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
530 | struct compact_control *cc) | |
531 | { | |
532 | unsigned long low_pfn, end_pfn; | |
533 | ||
534 | /* Do not scan outside zone boundaries */ | |
535 | low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); | |
536 | ||
537 | /* Only scan within a pageblock boundary */ | |
538 | end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); | |
539 | ||
540 | /* Do not cross the free scanner or scan within a memory hole */ | |
541 | if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { | |
542 | cc->migrate_pfn = end_pfn; | |
543 | return ISOLATE_NONE; | |
544 | } | |
545 | ||
546 | /* Perform the isolation */ | |
547 | low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn); | |
548 | if (!low_pfn) | |
549 | return ISOLATE_ABORT; | |
550 | ||
551 | cc->migrate_pfn = low_pfn; | |
552 | ||
553 | return ISOLATE_SUCCESS; | |
554 | } | |
555 | ||
748446bb | 556 | static int compact_finished(struct zone *zone, |
5a03b051 | 557 | struct compact_control *cc) |
748446bb | 558 | { |
56de7263 | 559 | unsigned int order; |
5a03b051 | 560 | unsigned long watermark; |
56de7263 | 561 | |
748446bb MG |
562 | if (fatal_signal_pending(current)) |
563 | return COMPACT_PARTIAL; | |
564 | ||
565 | /* Compaction run completes if the migrate and free scanner meet */ | |
566 | if (cc->free_pfn <= cc->migrate_pfn) | |
567 | return COMPACT_COMPLETE; | |
568 | ||
82478fb7 JW |
569 | /* |
570 | * order == -1 is expected when compacting via | |
571 | * /proc/sys/vm/compact_memory | |
572 | */ | |
56de7263 MG |
573 | if (cc->order == -1) |
574 | return COMPACT_CONTINUE; | |
575 | ||
3957c776 MH |
576 | /* Compaction run is not finished if the watermark is not met */ |
577 | watermark = low_wmark_pages(zone); | |
578 | watermark += (1 << cc->order); | |
579 | ||
580 | if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) | |
581 | return COMPACT_CONTINUE; | |
582 | ||
56de7263 MG |
583 | /* Direct compactor: Is a suitable page free? */ |
584 | for (order = cc->order; order < MAX_ORDER; order++) { | |
585 | /* Job done if page is free of the right migratetype */ | |
586 | if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) | |
587 | return COMPACT_PARTIAL; | |
588 | ||
589 | /* Job done if allocation would set block type */ | |
590 | if (order >= pageblock_order && zone->free_area[order].nr_free) | |
591 | return COMPACT_PARTIAL; | |
592 | } | |
593 | ||
748446bb MG |
594 | return COMPACT_CONTINUE; |
595 | } | |
596 | ||
3e7d3449 MG |
597 | /* |
598 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
599 | * Returns | |
600 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
601 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
602 | * COMPACT_CONTINUE - If compaction should run now | |
603 | */ | |
604 | unsigned long compaction_suitable(struct zone *zone, int order) | |
605 | { | |
606 | int fragindex; | |
607 | unsigned long watermark; | |
608 | ||
3957c776 MH |
609 | /* |
610 | * order == -1 is expected when compacting via | |
611 | * /proc/sys/vm/compact_memory | |
612 | */ | |
613 | if (order == -1) | |
614 | return COMPACT_CONTINUE; | |
615 | ||
3e7d3449 MG |
616 | /* |
617 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
618 | * This is because during migration, copies of pages need to be | |
619 | * allocated and for a short time, the footprint is higher | |
620 | */ | |
621 | watermark = low_wmark_pages(zone) + (2UL << order); | |
622 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
623 | return COMPACT_SKIPPED; | |
624 | ||
625 | /* | |
626 | * fragmentation index determines if allocation failures are due to | |
627 | * low memory or external fragmentation | |
628 | * | |
a582a738 SL |
629 | * index of -1000 implies allocations might succeed depending on |
630 | * watermarks | |
3e7d3449 MG |
631 | * index towards 0 implies failure is due to lack of memory |
632 | * index towards 1000 implies failure is due to fragmentation | |
633 | * | |
634 | * Only compact if a failure would be due to fragmentation. | |
635 | */ | |
636 | fragindex = fragmentation_index(zone, order); | |
637 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
638 | return COMPACT_SKIPPED; | |
639 | ||
a582a738 SL |
640 | if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark, |
641 | 0, 0)) | |
3e7d3449 MG |
642 | return COMPACT_PARTIAL; |
643 | ||
644 | return COMPACT_CONTINUE; | |
645 | } | |
646 | ||
748446bb MG |
647 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
648 | { | |
649 | int ret; | |
650 | ||
3e7d3449 MG |
651 | ret = compaction_suitable(zone, cc->order); |
652 | switch (ret) { | |
653 | case COMPACT_PARTIAL: | |
654 | case COMPACT_SKIPPED: | |
655 | /* Compaction is likely to fail */ | |
656 | return ret; | |
657 | case COMPACT_CONTINUE: | |
658 | /* Fall through to compaction */ | |
659 | ; | |
660 | } | |
661 | ||
748446bb MG |
662 | /* Setup to move all movable pages to the end of the zone */ |
663 | cc->migrate_pfn = zone->zone_start_pfn; | |
664 | cc->free_pfn = cc->migrate_pfn + zone->spanned_pages; | |
665 | cc->free_pfn &= ~(pageblock_nr_pages-1); | |
666 | ||
667 | migrate_prep_local(); | |
668 | ||
669 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { | |
670 | unsigned long nr_migrate, nr_remaining; | |
9d502c1c | 671 | int err; |
748446bb | 672 | |
f9e35b3b MG |
673 | switch (isolate_migratepages(zone, cc)) { |
674 | case ISOLATE_ABORT: | |
675 | ret = COMPACT_PARTIAL; | |
676 | goto out; | |
677 | case ISOLATE_NONE: | |
748446bb | 678 | continue; |
f9e35b3b MG |
679 | case ISOLATE_SUCCESS: |
680 | ; | |
681 | } | |
748446bb MG |
682 | |
683 | nr_migrate = cc->nr_migratepages; | |
9d502c1c | 684 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
7f0f2496 | 685 | (unsigned long)cc, false, |
a6bc32b8 | 686 | cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC); |
748446bb MG |
687 | update_nr_listpages(cc); |
688 | nr_remaining = cc->nr_migratepages; | |
689 | ||
690 | count_vm_event(COMPACTBLOCKS); | |
691 | count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); | |
692 | if (nr_remaining) | |
693 | count_vm_events(COMPACTPAGEFAILED, nr_remaining); | |
b7aba698 MG |
694 | trace_mm_compaction_migratepages(nr_migrate - nr_remaining, |
695 | nr_remaining); | |
748446bb MG |
696 | |
697 | /* Release LRU pages not migrated */ | |
9d502c1c | 698 | if (err) { |
748446bb MG |
699 | putback_lru_pages(&cc->migratepages); |
700 | cc->nr_migratepages = 0; | |
701 | } | |
702 | ||
703 | } | |
704 | ||
f9e35b3b | 705 | out: |
748446bb MG |
706 | /* Release free pages and check accounting */ |
707 | cc->nr_freepages -= release_freepages(&cc->freepages); | |
708 | VM_BUG_ON(cc->nr_freepages != 0); | |
709 | ||
710 | return ret; | |
711 | } | |
76ab0f53 | 712 | |
d43a87e6 | 713 | static unsigned long compact_zone_order(struct zone *zone, |
5a03b051 | 714 | int order, gfp_t gfp_mask, |
d527caf2 | 715 | bool sync) |
56de7263 MG |
716 | { |
717 | struct compact_control cc = { | |
718 | .nr_freepages = 0, | |
719 | .nr_migratepages = 0, | |
720 | .order = order, | |
721 | .migratetype = allocflags_to_migratetype(gfp_mask), | |
722 | .zone = zone, | |
77f1fe6b | 723 | .sync = sync, |
56de7263 MG |
724 | }; |
725 | INIT_LIST_HEAD(&cc.freepages); | |
726 | INIT_LIST_HEAD(&cc.migratepages); | |
727 | ||
728 | return compact_zone(zone, &cc); | |
729 | } | |
730 | ||
5e771905 MG |
731 | int sysctl_extfrag_threshold = 500; |
732 | ||
56de7263 MG |
733 | /** |
734 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
735 | * @zonelist: The zonelist used for the current allocation | |
736 | * @order: The order of the current allocation | |
737 | * @gfp_mask: The GFP mask of the current allocation | |
738 | * @nodemask: The allowed nodes to allocate from | |
77f1fe6b | 739 | * @sync: Whether migration is synchronous or not |
56de7263 MG |
740 | * |
741 | * This is the main entry point for direct page compaction. | |
742 | */ | |
743 | unsigned long try_to_compact_pages(struct zonelist *zonelist, | |
77f1fe6b MG |
744 | int order, gfp_t gfp_mask, nodemask_t *nodemask, |
745 | bool sync) | |
56de7263 MG |
746 | { |
747 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
748 | int may_enter_fs = gfp_mask & __GFP_FS; | |
749 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
750 | struct zoneref *z; |
751 | struct zone *zone; | |
752 | int rc = COMPACT_SKIPPED; | |
753 | ||
754 | /* | |
755 | * Check whether it is worth even starting compaction. The order check is | |
756 | * made because an assumption is made that the page allocator can satisfy | |
757 | * the "cheaper" orders without taking special steps | |
758 | */ | |
c5a73c3d | 759 | if (!order || !may_enter_fs || !may_perform_io) |
56de7263 MG |
760 | return rc; |
761 | ||
762 | count_vm_event(COMPACTSTALL); | |
763 | ||
764 | /* Compact each zone in the list */ | |
765 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, | |
766 | nodemask) { | |
56de7263 MG |
767 | int status; |
768 | ||
d527caf2 | 769 | status = compact_zone_order(zone, order, gfp_mask, sync); |
56de7263 MG |
770 | rc = max(status, rc); |
771 | ||
3e7d3449 MG |
772 | /* If a normal allocation would succeed, stop compacting */ |
773 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) | |
56de7263 MG |
774 | break; |
775 | } | |
776 | ||
777 | return rc; | |
778 | } | |
779 | ||
780 | ||
76ab0f53 | 781 | /* Compact all zones within a node */ |
7be62de9 | 782 | static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
783 | { |
784 | int zoneid; | |
76ab0f53 MG |
785 | struct zone *zone; |
786 | ||
76ab0f53 | 787 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
788 | |
789 | zone = &pgdat->node_zones[zoneid]; | |
790 | if (!populated_zone(zone)) | |
791 | continue; | |
792 | ||
7be62de9 RR |
793 | cc->nr_freepages = 0; |
794 | cc->nr_migratepages = 0; | |
795 | cc->zone = zone; | |
796 | INIT_LIST_HEAD(&cc->freepages); | |
797 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 798 | |
aad6ec37 | 799 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
7be62de9 | 800 | compact_zone(zone, cc); |
76ab0f53 | 801 | |
aff62249 RR |
802 | if (cc->order > 0) { |
803 | int ok = zone_watermark_ok(zone, cc->order, | |
804 | low_wmark_pages(zone), 0, 0); | |
805 | if (ok && cc->order > zone->compact_order_failed) | |
806 | zone->compact_order_failed = cc->order + 1; | |
807 | /* Currently async compaction is never deferred. */ | |
808 | else if (!ok && cc->sync) | |
809 | defer_compaction(zone, cc->order); | |
810 | } | |
811 | ||
7be62de9 RR |
812 | VM_BUG_ON(!list_empty(&cc->freepages)); |
813 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 MG |
814 | } |
815 | ||
816 | return 0; | |
817 | } | |
818 | ||
7be62de9 RR |
819 | int compact_pgdat(pg_data_t *pgdat, int order) |
820 | { | |
821 | struct compact_control cc = { | |
822 | .order = order, | |
823 | .sync = false, | |
824 | }; | |
825 | ||
826 | return __compact_pgdat(pgdat, &cc); | |
827 | } | |
828 | ||
829 | static int compact_node(int nid) | |
830 | { | |
7be62de9 RR |
831 | struct compact_control cc = { |
832 | .order = -1, | |
833 | .sync = true, | |
834 | }; | |
835 | ||
8575ec29 | 836 | return __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
837 | } |
838 | ||
76ab0f53 MG |
839 | /* Compact all nodes in the system */ |
840 | static int compact_nodes(void) | |
841 | { | |
842 | int nid; | |
843 | ||
8575ec29 HD |
844 | /* Flush pending updates to the LRU lists */ |
845 | lru_add_drain_all(); | |
846 | ||
76ab0f53 MG |
847 | for_each_online_node(nid) |
848 | compact_node(nid); | |
849 | ||
850 | return COMPACT_COMPLETE; | |
851 | } | |
852 | ||
853 | /* The written value is actually unused, all memory is compacted */ | |
854 | int sysctl_compact_memory; | |
855 | ||
856 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
857 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
858 | void __user *buffer, size_t *length, loff_t *ppos) | |
859 | { | |
860 | if (write) | |
861 | return compact_nodes(); | |
862 | ||
863 | return 0; | |
864 | } | |
ed4a6d7f | 865 | |
5e771905 MG |
866 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
867 | void __user *buffer, size_t *length, loff_t *ppos) | |
868 | { | |
869 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
870 | ||
871 | return 0; | |
872 | } | |
873 | ||
ed4a6d7f | 874 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
10fbcf4c KS |
875 | ssize_t sysfs_compact_node(struct device *dev, |
876 | struct device_attribute *attr, | |
ed4a6d7f MG |
877 | const char *buf, size_t count) |
878 | { | |
8575ec29 HD |
879 | int nid = dev->id; |
880 | ||
881 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
882 | /* Flush pending updates to the LRU lists */ | |
883 | lru_add_drain_all(); | |
884 | ||
885 | compact_node(nid); | |
886 | } | |
ed4a6d7f MG |
887 | |
888 | return count; | |
889 | } | |
10fbcf4c | 890 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
891 | |
892 | int compaction_register_node(struct node *node) | |
893 | { | |
10fbcf4c | 894 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
895 | } |
896 | ||
897 | void compaction_unregister_node(struct node *node) | |
898 | { | |
10fbcf4c | 899 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
900 | } |
901 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd MN |
902 | |
903 | #endif /* CONFIG_COMPACTION */ |