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; |
f3fd4a61 | 229 | isolate_mode_t mode = 0; |
fa9add64 | 230 | struct lruvec *lruvec; |
748446bb | 231 | |
748446bb MG |
232 | /* |
233 | * Ensure that there are not too many pages isolated from the LRU | |
234 | * list by either parallel reclaimers or compaction. If there are, | |
235 | * delay for some time until fewer pages are isolated | |
236 | */ | |
237 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b | 238 | /* async migration should just abort */ |
5ceb9ce6 | 239 | if (cc->mode != COMPACT_SYNC) |
2fe86e00 | 240 | return 0; |
f9e35b3b | 241 | |
748446bb MG |
242 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
243 | ||
244 | if (fatal_signal_pending(current)) | |
2fe86e00 | 245 | return 0; |
748446bb MG |
246 | } |
247 | ||
248 | /* Time to isolate some pages for migration */ | |
b2eef8c0 | 249 | cond_resched(); |
748446bb MG |
250 | spin_lock_irq(&zone->lru_lock); |
251 | for (; low_pfn < end_pfn; low_pfn++) { | |
252 | struct page *page; | |
b2eef8c0 AA |
253 | bool locked = true; |
254 | ||
255 | /* give a chance to irqs before checking need_resched() */ | |
256 | if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) { | |
257 | spin_unlock_irq(&zone->lru_lock); | |
258 | locked = false; | |
259 | } | |
260 | if (need_resched() || spin_is_contended(&zone->lru_lock)) { | |
261 | if (locked) | |
262 | spin_unlock_irq(&zone->lru_lock); | |
263 | cond_resched(); | |
264 | spin_lock_irq(&zone->lru_lock); | |
265 | if (fatal_signal_pending(current)) | |
266 | break; | |
267 | } else if (!locked) | |
268 | spin_lock_irq(&zone->lru_lock); | |
269 | ||
0bf380bc MG |
270 | /* |
271 | * migrate_pfn does not necessarily start aligned to a | |
272 | * pageblock. Ensure that pfn_valid is called when moving | |
273 | * into a new MAX_ORDER_NR_PAGES range in case of large | |
274 | * memory holes within the zone | |
275 | */ | |
276 | if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { | |
277 | if (!pfn_valid(low_pfn)) { | |
278 | low_pfn += MAX_ORDER_NR_PAGES - 1; | |
279 | continue; | |
280 | } | |
281 | } | |
282 | ||
748446bb MG |
283 | if (!pfn_valid_within(low_pfn)) |
284 | continue; | |
b7aba698 | 285 | nr_scanned++; |
748446bb | 286 | |
dc908600 MG |
287 | /* |
288 | * Get the page and ensure the page is within the same zone. | |
289 | * See the comment in isolate_freepages about overlapping | |
290 | * nodes. It is deliberate that the new zone lock is not taken | |
291 | * as memory compaction should not move pages between nodes. | |
292 | */ | |
748446bb | 293 | page = pfn_to_page(low_pfn); |
dc908600 MG |
294 | if (page_zone(page) != zone) |
295 | continue; | |
296 | ||
297 | /* Skip if free */ | |
748446bb MG |
298 | if (PageBuddy(page)) |
299 | continue; | |
300 | ||
9927af74 MG |
301 | /* |
302 | * For async migration, also only scan in MOVABLE blocks. Async | |
303 | * migration is optimistic to see if the minimum amount of work | |
304 | * satisfies the allocation | |
305 | */ | |
306 | pageblock_nr = low_pfn >> pageblock_order; | |
5ceb9ce6 BZ |
307 | if (cc->mode != COMPACT_SYNC && |
308 | last_pageblock_nr != pageblock_nr && | |
47118af0 | 309 | !migrate_async_suitable(get_pageblock_migratetype(page))) { |
9927af74 MG |
310 | low_pfn += pageblock_nr_pages; |
311 | low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1; | |
312 | last_pageblock_nr = pageblock_nr; | |
313 | continue; | |
314 | } | |
315 | ||
bc835011 AA |
316 | if (!PageLRU(page)) |
317 | continue; | |
318 | ||
319 | /* | |
320 | * PageLRU is set, and lru_lock excludes isolation, | |
321 | * splitting and collapsing (collapsing has already | |
322 | * happened if PageLRU is set). | |
323 | */ | |
324 | if (PageTransHuge(page)) { | |
325 | low_pfn += (1 << compound_order(page)) - 1; | |
326 | continue; | |
327 | } | |
328 | ||
5ceb9ce6 | 329 | if (cc->mode != COMPACT_SYNC) |
c8244935 MG |
330 | mode |= ISOLATE_ASYNC_MIGRATE; |
331 | ||
fa9add64 HD |
332 | lruvec = mem_cgroup_page_lruvec(page, zone); |
333 | ||
748446bb | 334 | /* Try isolate the page */ |
f3fd4a61 | 335 | if (__isolate_lru_page(page, mode) != 0) |
748446bb MG |
336 | continue; |
337 | ||
bc835011 AA |
338 | VM_BUG_ON(PageTransCompound(page)); |
339 | ||
748446bb | 340 | /* Successfully isolated */ |
fa9add64 | 341 | del_page_from_lru_list(page, lruvec, page_lru(page)); |
748446bb | 342 | list_add(&page->lru, migratelist); |
748446bb | 343 | cc->nr_migratepages++; |
b7aba698 | 344 | nr_isolated++; |
748446bb MG |
345 | |
346 | /* Avoid isolating too much */ | |
31b8384a HD |
347 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
348 | ++low_pfn; | |
748446bb | 349 | break; |
31b8384a | 350 | } |
748446bb MG |
351 | } |
352 | ||
353 | acct_isolated(zone, cc); | |
354 | ||
355 | spin_unlock_irq(&zone->lru_lock); | |
748446bb | 356 | |
b7aba698 MG |
357 | trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated); |
358 | ||
2fe86e00 MN |
359 | return low_pfn; |
360 | } | |
361 | ||
ff9543fd MN |
362 | #endif /* CONFIG_COMPACTION || CONFIG_CMA */ |
363 | #ifdef CONFIG_COMPACTION | |
5ceb9ce6 BZ |
364 | /* |
365 | * Returns true if MIGRATE_UNMOVABLE pageblock was successfully | |
366 | * converted to MIGRATE_MOVABLE type, false otherwise. | |
367 | */ | |
368 | static bool rescue_unmovable_pageblock(struct page *page) | |
369 | { | |
370 | unsigned long pfn, start_pfn, end_pfn; | |
371 | struct page *start_page, *end_page; | |
372 | ||
373 | pfn = page_to_pfn(page); | |
374 | start_pfn = pfn & ~(pageblock_nr_pages - 1); | |
375 | end_pfn = start_pfn + pageblock_nr_pages; | |
376 | ||
377 | start_page = pfn_to_page(start_pfn); | |
378 | end_page = pfn_to_page(end_pfn); | |
379 | ||
380 | /* Do not deal with pageblocks that overlap zones */ | |
381 | if (page_zone(start_page) != page_zone(end_page)) | |
382 | return false; | |
383 | ||
384 | for (page = start_page, pfn = start_pfn; page < end_page; pfn++, | |
385 | page++) { | |
386 | if (!pfn_valid_within(pfn)) | |
387 | continue; | |
388 | ||
389 | if (PageBuddy(page)) { | |
390 | int order = page_order(page); | |
391 | ||
392 | pfn += (1 << order) - 1; | |
393 | page += (1 << order) - 1; | |
394 | ||
395 | continue; | |
396 | } else if (page_count(page) == 0 || PageLRU(page)) | |
397 | continue; | |
398 | ||
399 | return false; | |
400 | } | |
401 | ||
402 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
403 | move_freepages_block(page_zone(page), page, MIGRATE_MOVABLE); | |
404 | return true; | |
405 | } | |
ff9543fd | 406 | |
5ceb9ce6 BZ |
407 | enum smt_result { |
408 | GOOD_AS_MIGRATION_TARGET, | |
409 | FAIL_UNMOVABLE_TARGET, | |
410 | FAIL_BAD_TARGET, | |
411 | }; | |
412 | ||
413 | /* | |
414 | * Returns GOOD_AS_MIGRATION_TARGET if the page is within a block | |
415 | * suitable for migration to, FAIL_UNMOVABLE_TARGET if the page | |
416 | * is within a MIGRATE_UNMOVABLE block, FAIL_BAD_TARGET otherwise. | |
417 | */ | |
418 | static enum smt_result suitable_migration_target(struct page *page, | |
419 | struct compact_control *cc) | |
ff9543fd MN |
420 | { |
421 | ||
422 | int migratetype = get_pageblock_migratetype(page); | |
423 | ||
424 | /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */ | |
425 | if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE) | |
5ceb9ce6 | 426 | return FAIL_BAD_TARGET; |
ff9543fd MN |
427 | |
428 | /* If the page is a large free page, then allow migration */ | |
429 | if (PageBuddy(page) && page_order(page) >= pageblock_order) | |
5ceb9ce6 | 430 | return GOOD_AS_MIGRATION_TARGET; |
ff9543fd | 431 | |
47118af0 | 432 | /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ |
5ceb9ce6 BZ |
433 | if (cc->mode != COMPACT_ASYNC_UNMOVABLE && |
434 | migrate_async_suitable(migratetype)) | |
435 | return GOOD_AS_MIGRATION_TARGET; | |
436 | ||
437 | if (cc->mode == COMPACT_ASYNC_MOVABLE && | |
438 | migratetype == MIGRATE_UNMOVABLE) | |
439 | return FAIL_UNMOVABLE_TARGET; | |
440 | ||
441 | if (cc->mode != COMPACT_ASYNC_MOVABLE && | |
442 | migratetype == MIGRATE_UNMOVABLE && | |
443 | rescue_unmovable_pageblock(page)) | |
444 | return GOOD_AS_MIGRATION_TARGET; | |
ff9543fd MN |
445 | |
446 | /* Otherwise skip the block */ | |
5ceb9ce6 | 447 | return FAIL_BAD_TARGET; |
ff9543fd MN |
448 | } |
449 | ||
2fe86e00 | 450 | /* |
ff9543fd MN |
451 | * Based on information in the current compact_control, find blocks |
452 | * suitable for isolating free pages from and then isolate them. | |
2fe86e00 | 453 | */ |
ff9543fd MN |
454 | static void isolate_freepages(struct zone *zone, |
455 | struct compact_control *cc) | |
2fe86e00 | 456 | { |
ff9543fd MN |
457 | struct page *page; |
458 | unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn; | |
459 | unsigned long flags; | |
460 | int nr_freepages = cc->nr_freepages; | |
461 | struct list_head *freelist = &cc->freepages; | |
2fe86e00 | 462 | |
ff9543fd MN |
463 | /* |
464 | * Initialise the free scanner. The starting point is where we last | |
465 | * scanned from (or the end of the zone if starting). The low point | |
466 | * is the end of the pageblock the migration scanner is using. | |
467 | */ | |
468 | pfn = cc->free_pfn; | |
469 | low_pfn = cc->migrate_pfn + pageblock_nr_pages; | |
2fe86e00 | 470 | |
ff9543fd MN |
471 | /* |
472 | * Take care that if the migration scanner is at the end of the zone | |
473 | * that the free scanner does not accidentally move to the next zone | |
474 | * in the next isolation cycle. | |
475 | */ | |
476 | high_pfn = min(low_pfn, pfn); | |
2fe86e00 | 477 | |
ff9543fd | 478 | zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages; |
2fe86e00 | 479 | |
5ceb9ce6 BZ |
480 | /* |
481 | * isolate_freepages() may be called more than once during | |
482 | * compact_zone_order() run and we want only the most recent | |
483 | * count. | |
484 | */ | |
485 | cc->nr_pageblocks_skipped = 0; | |
486 | ||
ff9543fd MN |
487 | /* |
488 | * Isolate free pages until enough are available to migrate the | |
489 | * pages on cc->migratepages. We stop searching if the migrate | |
490 | * and free page scanners meet or enough free pages are isolated. | |
491 | */ | |
492 | for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages; | |
493 | pfn -= pageblock_nr_pages) { | |
494 | unsigned long isolated; | |
5ceb9ce6 | 495 | enum smt_result ret; |
2fe86e00 | 496 | |
ff9543fd MN |
497 | if (!pfn_valid(pfn)) |
498 | continue; | |
2fe86e00 | 499 | |
ff9543fd MN |
500 | /* |
501 | * Check for overlapping nodes/zones. It's possible on some | |
502 | * configurations to have a setup like | |
503 | * node0 node1 node0 | |
504 | * i.e. it's possible that all pages within a zones range of | |
505 | * pages do not belong to a single zone. | |
506 | */ | |
507 | page = pfn_to_page(pfn); | |
508 | if (page_zone(page) != zone) | |
509 | continue; | |
510 | ||
511 | /* Check the block is suitable for migration */ | |
5ceb9ce6 BZ |
512 | ret = suitable_migration_target(page, cc); |
513 | if (ret != GOOD_AS_MIGRATION_TARGET) { | |
514 | if (ret == FAIL_UNMOVABLE_TARGET) | |
515 | cc->nr_pageblocks_skipped++; | |
ff9543fd | 516 | continue; |
5ceb9ce6 | 517 | } |
ff9543fd MN |
518 | /* |
519 | * Found a block suitable for isolating free pages from. Now | |
520 | * we disabled interrupts, double check things are ok and | |
521 | * isolate the pages. This is to minimise the time IRQs | |
522 | * are disabled | |
523 | */ | |
524 | isolated = 0; | |
525 | spin_lock_irqsave(&zone->lock, flags); | |
5ceb9ce6 BZ |
526 | ret = suitable_migration_target(page, cc); |
527 | if (ret == GOOD_AS_MIGRATION_TARGET) { | |
ff9543fd MN |
528 | end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn); |
529 | isolated = isolate_freepages_block(pfn, end_pfn, | |
530 | freelist, false); | |
531 | nr_freepages += isolated; | |
5ceb9ce6 BZ |
532 | } else if (ret == FAIL_UNMOVABLE_TARGET) |
533 | cc->nr_pageblocks_skipped++; | |
ff9543fd MN |
534 | spin_unlock_irqrestore(&zone->lock, flags); |
535 | ||
536 | /* | |
537 | * Record the highest PFN we isolated pages from. When next | |
538 | * looking for free pages, the search will restart here as | |
539 | * page migration may have returned some pages to the allocator | |
540 | */ | |
541 | if (isolated) | |
542 | high_pfn = max(high_pfn, pfn); | |
543 | } | |
544 | ||
545 | /* split_free_page does not map the pages */ | |
546 | map_pages(freelist); | |
547 | ||
548 | cc->free_pfn = high_pfn; | |
549 | cc->nr_freepages = nr_freepages; | |
748446bb MG |
550 | } |
551 | ||
552 | /* | |
553 | * This is a migrate-callback that "allocates" freepages by taking pages | |
554 | * from the isolated freelists in the block we are migrating to. | |
555 | */ | |
556 | static struct page *compaction_alloc(struct page *migratepage, | |
557 | unsigned long data, | |
558 | int **result) | |
559 | { | |
560 | struct compact_control *cc = (struct compact_control *)data; | |
561 | struct page *freepage; | |
562 | ||
563 | /* Isolate free pages if necessary */ | |
564 | if (list_empty(&cc->freepages)) { | |
565 | isolate_freepages(cc->zone, cc); | |
566 | ||
567 | if (list_empty(&cc->freepages)) | |
568 | return NULL; | |
569 | } | |
570 | ||
571 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
572 | list_del(&freepage->lru); | |
573 | cc->nr_freepages--; | |
574 | ||
575 | return freepage; | |
576 | } | |
577 | ||
578 | /* | |
579 | * We cannot control nr_migratepages and nr_freepages fully when migration is | |
580 | * running as migrate_pages() has no knowledge of compact_control. When | |
581 | * migration is complete, we count the number of pages on the lists by hand. | |
582 | */ | |
583 | static void update_nr_listpages(struct compact_control *cc) | |
584 | { | |
585 | int nr_migratepages = 0; | |
586 | int nr_freepages = 0; | |
587 | struct page *page; | |
588 | ||
589 | list_for_each_entry(page, &cc->migratepages, lru) | |
590 | nr_migratepages++; | |
591 | list_for_each_entry(page, &cc->freepages, lru) | |
592 | nr_freepages++; | |
593 | ||
594 | cc->nr_migratepages = nr_migratepages; | |
595 | cc->nr_freepages = nr_freepages; | |
596 | } | |
597 | ||
ff9543fd MN |
598 | /* possible outcome of isolate_migratepages */ |
599 | typedef enum { | |
600 | ISOLATE_ABORT, /* Abort compaction now */ | |
601 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
602 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
603 | } isolate_migrate_t; | |
604 | ||
605 | /* | |
606 | * Isolate all pages that can be migrated from the block pointed to by | |
607 | * the migrate scanner within compact_control. | |
608 | */ | |
609 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
610 | struct compact_control *cc) | |
611 | { | |
612 | unsigned long low_pfn, end_pfn; | |
613 | ||
614 | /* Do not scan outside zone boundaries */ | |
615 | low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); | |
616 | ||
617 | /* Only scan within a pageblock boundary */ | |
618 | end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); | |
619 | ||
620 | /* Do not cross the free scanner or scan within a memory hole */ | |
621 | if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { | |
622 | cc->migrate_pfn = end_pfn; | |
623 | return ISOLATE_NONE; | |
624 | } | |
625 | ||
626 | /* Perform the isolation */ | |
627 | low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn); | |
628 | if (!low_pfn) | |
629 | return ISOLATE_ABORT; | |
630 | ||
631 | cc->migrate_pfn = low_pfn; | |
632 | ||
633 | return ISOLATE_SUCCESS; | |
634 | } | |
635 | ||
748446bb | 636 | static int compact_finished(struct zone *zone, |
5a03b051 | 637 | struct compact_control *cc) |
748446bb | 638 | { |
56de7263 | 639 | unsigned int order; |
5a03b051 | 640 | unsigned long watermark; |
56de7263 | 641 | |
748446bb MG |
642 | if (fatal_signal_pending(current)) |
643 | return COMPACT_PARTIAL; | |
644 | ||
645 | /* Compaction run completes if the migrate and free scanner meet */ | |
646 | if (cc->free_pfn <= cc->migrate_pfn) | |
647 | return COMPACT_COMPLETE; | |
648 | ||
82478fb7 JW |
649 | /* |
650 | * order == -1 is expected when compacting via | |
651 | * /proc/sys/vm/compact_memory | |
652 | */ | |
56de7263 MG |
653 | if (cc->order == -1) |
654 | return COMPACT_CONTINUE; | |
655 | ||
3957c776 MH |
656 | /* Compaction run is not finished if the watermark is not met */ |
657 | watermark = low_wmark_pages(zone); | |
658 | watermark += (1 << cc->order); | |
659 | ||
660 | if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) | |
661 | return COMPACT_CONTINUE; | |
662 | ||
56de7263 MG |
663 | /* Direct compactor: Is a suitable page free? */ |
664 | for (order = cc->order; order < MAX_ORDER; order++) { | |
665 | /* Job done if page is free of the right migratetype */ | |
666 | if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) | |
667 | return COMPACT_PARTIAL; | |
668 | ||
669 | /* Job done if allocation would set block type */ | |
670 | if (order >= pageblock_order && zone->free_area[order].nr_free) | |
671 | return COMPACT_PARTIAL; | |
672 | } | |
673 | ||
748446bb MG |
674 | return COMPACT_CONTINUE; |
675 | } | |
676 | ||
3e7d3449 MG |
677 | /* |
678 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
679 | * Returns | |
680 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
681 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
682 | * COMPACT_CONTINUE - If compaction should run now | |
683 | */ | |
684 | unsigned long compaction_suitable(struct zone *zone, int order) | |
685 | { | |
686 | int fragindex; | |
687 | unsigned long watermark; | |
688 | ||
3957c776 MH |
689 | /* |
690 | * order == -1 is expected when compacting via | |
691 | * /proc/sys/vm/compact_memory | |
692 | */ | |
693 | if (order == -1) | |
694 | return COMPACT_CONTINUE; | |
695 | ||
3e7d3449 MG |
696 | /* |
697 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
698 | * This is because during migration, copies of pages need to be | |
699 | * allocated and for a short time, the footprint is higher | |
700 | */ | |
701 | watermark = low_wmark_pages(zone) + (2UL << order); | |
702 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
703 | return COMPACT_SKIPPED; | |
704 | ||
705 | /* | |
706 | * fragmentation index determines if allocation failures are due to | |
707 | * low memory or external fragmentation | |
708 | * | |
a582a738 SL |
709 | * index of -1000 implies allocations might succeed depending on |
710 | * watermarks | |
3e7d3449 MG |
711 | * index towards 0 implies failure is due to lack of memory |
712 | * index towards 1000 implies failure is due to fragmentation | |
713 | * | |
714 | * Only compact if a failure would be due to fragmentation. | |
715 | */ | |
716 | fragindex = fragmentation_index(zone, order); | |
717 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
718 | return COMPACT_SKIPPED; | |
719 | ||
a582a738 SL |
720 | if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark, |
721 | 0, 0)) | |
3e7d3449 MG |
722 | return COMPACT_PARTIAL; |
723 | ||
724 | return COMPACT_CONTINUE; | |
725 | } | |
726 | ||
748446bb MG |
727 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
728 | { | |
729 | int ret; | |
730 | ||
3e7d3449 MG |
731 | ret = compaction_suitable(zone, cc->order); |
732 | switch (ret) { | |
733 | case COMPACT_PARTIAL: | |
734 | case COMPACT_SKIPPED: | |
735 | /* Compaction is likely to fail */ | |
736 | return ret; | |
737 | case COMPACT_CONTINUE: | |
738 | /* Fall through to compaction */ | |
739 | ; | |
740 | } | |
741 | ||
748446bb MG |
742 | /* Setup to move all movable pages to the end of the zone */ |
743 | cc->migrate_pfn = zone->zone_start_pfn; | |
744 | cc->free_pfn = cc->migrate_pfn + zone->spanned_pages; | |
745 | cc->free_pfn &= ~(pageblock_nr_pages-1); | |
746 | ||
747 | migrate_prep_local(); | |
748 | ||
749 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { | |
750 | unsigned long nr_migrate, nr_remaining; | |
9d502c1c | 751 | int err; |
748446bb | 752 | |
f9e35b3b MG |
753 | switch (isolate_migratepages(zone, cc)) { |
754 | case ISOLATE_ABORT: | |
755 | ret = COMPACT_PARTIAL; | |
756 | goto out; | |
757 | case ISOLATE_NONE: | |
748446bb | 758 | continue; |
f9e35b3b MG |
759 | case ISOLATE_SUCCESS: |
760 | ; | |
761 | } | |
748446bb MG |
762 | |
763 | nr_migrate = cc->nr_migratepages; | |
9d502c1c | 764 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
5ceb9ce6 BZ |
765 | (unsigned long)&cc->freepages, false, |
766 | (cc->mode == COMPACT_SYNC) ? MIGRATE_SYNC_LIGHT | |
767 | : MIGRATE_ASYNC); | |
748446bb MG |
768 | update_nr_listpages(cc); |
769 | nr_remaining = cc->nr_migratepages; | |
770 | ||
771 | count_vm_event(COMPACTBLOCKS); | |
772 | count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); | |
773 | if (nr_remaining) | |
774 | count_vm_events(COMPACTPAGEFAILED, nr_remaining); | |
b7aba698 MG |
775 | trace_mm_compaction_migratepages(nr_migrate - nr_remaining, |
776 | nr_remaining); | |
748446bb MG |
777 | |
778 | /* Release LRU pages not migrated */ | |
9d502c1c | 779 | if (err) { |
748446bb MG |
780 | putback_lru_pages(&cc->migratepages); |
781 | cc->nr_migratepages = 0; | |
782 | } | |
783 | ||
784 | } | |
785 | ||
f9e35b3b | 786 | out: |
748446bb MG |
787 | /* Release free pages and check accounting */ |
788 | cc->nr_freepages -= release_freepages(&cc->freepages); | |
789 | VM_BUG_ON(cc->nr_freepages != 0); | |
790 | ||
791 | return ret; | |
792 | } | |
76ab0f53 | 793 | |
d43a87e6 | 794 | static unsigned long compact_zone_order(struct zone *zone, |
5a03b051 | 795 | int order, gfp_t gfp_mask, |
5ceb9ce6 BZ |
796 | enum compact_mode mode, |
797 | unsigned long *nr_pageblocks_skipped) | |
56de7263 MG |
798 | { |
799 | struct compact_control cc = { | |
800 | .nr_freepages = 0, | |
801 | .nr_migratepages = 0, | |
802 | .order = order, | |
803 | .migratetype = allocflags_to_migratetype(gfp_mask), | |
804 | .zone = zone, | |
5ceb9ce6 | 805 | .mode = mode, |
56de7263 | 806 | }; |
5ceb9ce6 BZ |
807 | unsigned long rc; |
808 | ||
56de7263 MG |
809 | INIT_LIST_HEAD(&cc.freepages); |
810 | INIT_LIST_HEAD(&cc.migratepages); | |
811 | ||
5ceb9ce6 BZ |
812 | rc = compact_zone(zone, &cc); |
813 | *nr_pageblocks_skipped = cc.nr_pageblocks_skipped; | |
814 | ||
815 | return rc; | |
56de7263 MG |
816 | } |
817 | ||
5e771905 MG |
818 | int sysctl_extfrag_threshold = 500; |
819 | ||
56de7263 MG |
820 | /** |
821 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
822 | * @zonelist: The zonelist used for the current allocation | |
823 | * @order: The order of the current allocation | |
824 | * @gfp_mask: The GFP mask of the current allocation | |
825 | * @nodemask: The allowed nodes to allocate from | |
77f1fe6b | 826 | * @sync: Whether migration is synchronous or not |
56de7263 MG |
827 | * |
828 | * This is the main entry point for direct page compaction. | |
829 | */ | |
830 | unsigned long try_to_compact_pages(struct zonelist *zonelist, | |
77f1fe6b MG |
831 | int order, gfp_t gfp_mask, nodemask_t *nodemask, |
832 | bool sync) | |
56de7263 MG |
833 | { |
834 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
835 | int may_enter_fs = gfp_mask & __GFP_FS; | |
836 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
837 | struct zoneref *z; |
838 | struct zone *zone; | |
839 | int rc = COMPACT_SKIPPED; | |
5ceb9ce6 BZ |
840 | unsigned long nr_pageblocks_skipped; |
841 | enum compact_mode mode; | |
56de7263 MG |
842 | |
843 | /* | |
844 | * Check whether it is worth even starting compaction. The order check is | |
845 | * made because an assumption is made that the page allocator can satisfy | |
846 | * the "cheaper" orders without taking special steps | |
847 | */ | |
c5a73c3d | 848 | if (!order || !may_enter_fs || !may_perform_io) |
56de7263 MG |
849 | return rc; |
850 | ||
851 | count_vm_event(COMPACTSTALL); | |
852 | ||
853 | /* Compact each zone in the list */ | |
854 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, | |
855 | nodemask) { | |
56de7263 MG |
856 | int status; |
857 | ||
5ceb9ce6 BZ |
858 | mode = sync ? COMPACT_SYNC : COMPACT_ASYNC_MOVABLE; |
859 | retry: | |
860 | status = compact_zone_order(zone, order, gfp_mask, mode, | |
861 | &nr_pageblocks_skipped); | |
56de7263 MG |
862 | rc = max(status, rc); |
863 | ||
3e7d3449 MG |
864 | /* If a normal allocation would succeed, stop compacting */ |
865 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) | |
56de7263 | 866 | break; |
5ceb9ce6 BZ |
867 | |
868 | if (rc == COMPACT_COMPLETE && mode == COMPACT_ASYNC_MOVABLE) { | |
869 | if (nr_pageblocks_skipped) { | |
870 | mode = COMPACT_ASYNC_UNMOVABLE; | |
871 | goto retry; | |
872 | } | |
873 | } | |
56de7263 MG |
874 | } |
875 | ||
876 | return rc; | |
877 | } | |
878 | ||
879 | ||
76ab0f53 | 880 | /* Compact all zones within a node */ |
7be62de9 | 881 | static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
882 | { |
883 | int zoneid; | |
76ab0f53 MG |
884 | struct zone *zone; |
885 | ||
76ab0f53 | 886 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
887 | |
888 | zone = &pgdat->node_zones[zoneid]; | |
889 | if (!populated_zone(zone)) | |
890 | continue; | |
891 | ||
7be62de9 RR |
892 | cc->nr_freepages = 0; |
893 | cc->nr_migratepages = 0; | |
894 | cc->zone = zone; | |
895 | INIT_LIST_HEAD(&cc->freepages); | |
896 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 897 | |
aad6ec37 | 898 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
7be62de9 | 899 | compact_zone(zone, cc); |
76ab0f53 | 900 | |
aff62249 RR |
901 | if (cc->order > 0) { |
902 | int ok = zone_watermark_ok(zone, cc->order, | |
903 | low_wmark_pages(zone), 0, 0); | |
904 | if (ok && cc->order > zone->compact_order_failed) | |
905 | zone->compact_order_failed = cc->order + 1; | |
906 | /* Currently async compaction is never deferred. */ | |
5ceb9ce6 | 907 | else if (!ok && cc->mode == COMPACT_SYNC) |
aff62249 RR |
908 | defer_compaction(zone, cc->order); |
909 | } | |
910 | ||
7be62de9 RR |
911 | VM_BUG_ON(!list_empty(&cc->freepages)); |
912 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 MG |
913 | } |
914 | ||
915 | return 0; | |
916 | } | |
917 | ||
7be62de9 RR |
918 | int compact_pgdat(pg_data_t *pgdat, int order) |
919 | { | |
920 | struct compact_control cc = { | |
921 | .order = order, | |
5ceb9ce6 | 922 | .mode = COMPACT_ASYNC_MOVABLE, |
7be62de9 RR |
923 | }; |
924 | ||
925 | return __compact_pgdat(pgdat, &cc); | |
926 | } | |
927 | ||
928 | static int compact_node(int nid) | |
929 | { | |
7be62de9 RR |
930 | struct compact_control cc = { |
931 | .order = -1, | |
5ceb9ce6 | 932 | .mode = COMPACT_SYNC, |
7be62de9 RR |
933 | }; |
934 | ||
8575ec29 | 935 | return __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
936 | } |
937 | ||
76ab0f53 MG |
938 | /* Compact all nodes in the system */ |
939 | static int compact_nodes(void) | |
940 | { | |
941 | int nid; | |
942 | ||
8575ec29 HD |
943 | /* Flush pending updates to the LRU lists */ |
944 | lru_add_drain_all(); | |
945 | ||
76ab0f53 MG |
946 | for_each_online_node(nid) |
947 | compact_node(nid); | |
948 | ||
949 | return COMPACT_COMPLETE; | |
950 | } | |
951 | ||
952 | /* The written value is actually unused, all memory is compacted */ | |
953 | int sysctl_compact_memory; | |
954 | ||
955 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
956 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
957 | void __user *buffer, size_t *length, loff_t *ppos) | |
958 | { | |
959 | if (write) | |
960 | return compact_nodes(); | |
961 | ||
962 | return 0; | |
963 | } | |
ed4a6d7f | 964 | |
5e771905 MG |
965 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
966 | void __user *buffer, size_t *length, loff_t *ppos) | |
967 | { | |
968 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
969 | ||
970 | return 0; | |
971 | } | |
972 | ||
ed4a6d7f | 973 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
10fbcf4c KS |
974 | ssize_t sysfs_compact_node(struct device *dev, |
975 | struct device_attribute *attr, | |
ed4a6d7f MG |
976 | const char *buf, size_t count) |
977 | { | |
8575ec29 HD |
978 | int nid = dev->id; |
979 | ||
980 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
981 | /* Flush pending updates to the LRU lists */ | |
982 | lru_add_drain_all(); | |
983 | ||
984 | compact_node(nid); | |
985 | } | |
ed4a6d7f MG |
986 | |
987 | return count; | |
988 | } | |
10fbcf4c | 989 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
990 | |
991 | int compaction_register_node(struct node *node) | |
992 | { | |
10fbcf4c | 993 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
994 | } |
995 | ||
996 | void compaction_unregister_node(struct node *node) | |
997 | { | |
10fbcf4c | 998 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
999 | } |
1000 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ | |
ff9543fd MN |
1001 | |
1002 | #endif /* CONFIG_COMPACTION */ |