]>
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 | ||
b7aba698 MG |
19 | #define CREATE_TRACE_POINTS |
20 | #include <trace/events/compaction.h> | |
21 | ||
748446bb MG |
22 | /* |
23 | * compact_control is used to track pages being migrated and the free pages | |
24 | * they are being migrated to during memory compaction. The free_pfn starts | |
25 | * at the end of a zone and migrate_pfn begins at the start. Movable pages | |
26 | * are moved to the end of a zone during a compaction run and the run | |
27 | * completes when free_pfn <= migrate_pfn | |
28 | */ | |
29 | struct compact_control { | |
30 | struct list_head freepages; /* List of free pages to migrate to */ | |
31 | struct list_head migratepages; /* List of pages being migrated */ | |
32 | unsigned long nr_freepages; /* Number of isolated free pages */ | |
33 | unsigned long nr_migratepages; /* Number of pages to migrate */ | |
34 | unsigned long free_pfn; /* isolate_freepages search base */ | |
35 | unsigned long migrate_pfn; /* isolate_migratepages search base */ | |
77f1fe6b | 36 | bool sync; /* Synchronous migration */ |
748446bb | 37 | |
aad6ec37 | 38 | int order; /* order a direct compactor needs */ |
56de7263 | 39 | int migratetype; /* MOVABLE, RECLAIMABLE etc */ |
748446bb MG |
40 | struct zone *zone; |
41 | }; | |
42 | ||
43 | static unsigned long release_freepages(struct list_head *freelist) | |
44 | { | |
45 | struct page *page, *next; | |
46 | unsigned long count = 0; | |
47 | ||
48 | list_for_each_entry_safe(page, next, freelist, lru) { | |
49 | list_del(&page->lru); | |
50 | __free_page(page); | |
51 | count++; | |
52 | } | |
53 | ||
54 | return count; | |
55 | } | |
56 | ||
85aa125f MN |
57 | /* |
58 | * Isolate free pages onto a private freelist. Caller must hold zone->lock. | |
59 | * If @strict is true, will abort returning 0 on any invalid PFNs or non-free | |
60 | * pages inside of the pageblock (even though it may still end up isolating | |
61 | * some pages). | |
62 | */ | |
63 | static unsigned long isolate_freepages_block(unsigned long blockpfn, | |
64 | unsigned long end_pfn, | |
65 | struct list_head *freelist, | |
66 | bool strict) | |
748446bb | 67 | { |
b7aba698 | 68 | int nr_scanned = 0, total_isolated = 0; |
748446bb MG |
69 | struct page *cursor; |
70 | ||
748446bb MG |
71 | cursor = pfn_to_page(blockpfn); |
72 | ||
73 | /* Isolate free pages. This assumes the block is valid */ | |
74 | for (; blockpfn < end_pfn; blockpfn++, cursor++) { | |
75 | int isolated, i; | |
76 | struct page *page = cursor; | |
77 | ||
85aa125f MN |
78 | if (!pfn_valid_within(blockpfn)) { |
79 | if (strict) | |
80 | return 0; | |
748446bb | 81 | continue; |
85aa125f | 82 | } |
b7aba698 | 83 | nr_scanned++; |
748446bb | 84 | |
85aa125f MN |
85 | if (!PageBuddy(page)) { |
86 | if (strict) | |
87 | return 0; | |
748446bb | 88 | continue; |
85aa125f | 89 | } |
748446bb MG |
90 | |
91 | /* Found a free page, break it into order-0 pages */ | |
92 | isolated = split_free_page(page); | |
85aa125f MN |
93 | if (!isolated && strict) |
94 | return 0; | |
748446bb MG |
95 | total_isolated += isolated; |
96 | for (i = 0; i < isolated; i++) { | |
97 | list_add(&page->lru, freelist); | |
98 | page++; | |
99 | } | |
100 | ||
101 | /* If a page was split, advance to the end of it */ | |
102 | if (isolated) { | |
103 | blockpfn += isolated - 1; | |
104 | cursor += isolated - 1; | |
105 | } | |
106 | } | |
107 | ||
b7aba698 | 108 | trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated); |
748446bb MG |
109 | return total_isolated; |
110 | } | |
111 | ||
85aa125f MN |
112 | /** |
113 | * isolate_freepages_range() - isolate free pages. | |
114 | * @start_pfn: The first PFN to start isolating. | |
115 | * @end_pfn: The one-past-last PFN. | |
116 | * | |
117 | * Non-free pages, invalid PFNs, or zone boundaries within the | |
118 | * [start_pfn, end_pfn) range are considered errors, cause function to | |
119 | * undo its actions and return zero. | |
120 | * | |
121 | * Otherwise, function returns one-past-the-last PFN of isolated page | |
122 | * (which may be greater then end_pfn if end fell in a middle of | |
123 | * a free page). | |
124 | */ | |
125 | static unsigned long | |
126 | isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn) | |
127 | { | |
128 | unsigned long isolated, pfn, block_end_pfn, flags; | |
129 | struct zone *zone = NULL; | |
130 | LIST_HEAD(freelist); | |
131 | ||
132 | if (pfn_valid(start_pfn)) | |
133 | zone = page_zone(pfn_to_page(start_pfn)); | |
134 | ||
135 | for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) { | |
136 | if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn))) | |
137 | break; | |
138 | ||
139 | /* | |
140 | * On subsequent iterations ALIGN() is actually not needed, | |
141 | * but we keep it that we not to complicate the code. | |
142 | */ | |
143 | block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); | |
144 | block_end_pfn = min(block_end_pfn, end_pfn); | |
145 | ||
146 | spin_lock_irqsave(&zone->lock, flags); | |
147 | isolated = isolate_freepages_block(pfn, block_end_pfn, | |
148 | &freelist, true); | |
149 | spin_unlock_irqrestore(&zone->lock, flags); | |
150 | ||
151 | /* | |
152 | * In strict mode, isolate_freepages_block() returns 0 if | |
153 | * there are any holes in the block (ie. invalid PFNs or | |
154 | * non-free pages). | |
155 | */ | |
156 | if (!isolated) | |
157 | break; | |
158 | ||
159 | /* | |
160 | * If we managed to isolate pages, it is always (1 << n) * | |
161 | * pageblock_nr_pages for some non-negative n. (Max order | |
162 | * page may span two pageblocks). | |
163 | */ | |
164 | } | |
165 | ||
166 | /* split_free_page does not map the pages */ | |
167 | map_pages(&freelist); | |
168 | ||
169 | if (pfn < end_pfn) { | |
170 | /* Loop terminated early, cleanup. */ | |
171 | release_freepages(&freelist); | |
172 | return 0; | |
173 | } | |
174 | ||
175 | /* We don't use freelists for anything. */ | |
176 | return pfn; | |
177 | } | |
178 | ||
748446bb MG |
179 | /* Returns true if the page is within a block suitable for migration to */ |
180 | static bool suitable_migration_target(struct page *page) | |
181 | { | |
182 | ||
183 | int migratetype = get_pageblock_migratetype(page); | |
184 | ||
185 | /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */ | |
186 | if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE) | |
187 | return false; | |
188 | ||
189 | /* If the page is a large free page, then allow migration */ | |
190 | if (PageBuddy(page) && page_order(page) >= pageblock_order) | |
191 | return true; | |
192 | ||
193 | /* If the block is MIGRATE_MOVABLE, allow migration */ | |
194 | if (migratetype == MIGRATE_MOVABLE) | |
195 | return true; | |
196 | ||
197 | /* Otherwise skip the block */ | |
198 | return false; | |
199 | } | |
200 | ||
03d44192 MN |
201 | static void map_pages(struct list_head *list) |
202 | { | |
203 | struct page *page; | |
204 | ||
205 | list_for_each_entry(page, list, lru) { | |
206 | arch_alloc_page(page, 0); | |
207 | kernel_map_pages(page, 1, 1); | |
208 | } | |
209 | } | |
210 | ||
748446bb MG |
211 | /* |
212 | * Based on information in the current compact_control, find blocks | |
213 | * suitable for isolating free pages from and then isolate them. | |
214 | */ | |
215 | static void isolate_freepages(struct zone *zone, | |
216 | struct compact_control *cc) | |
217 | { | |
218 | struct page *page; | |
85aa125f | 219 | unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn; |
748446bb MG |
220 | unsigned long flags; |
221 | int nr_freepages = cc->nr_freepages; | |
222 | struct list_head *freelist = &cc->freepages; | |
223 | ||
7454f4ba MG |
224 | /* |
225 | * Initialise the free scanner. The starting point is where we last | |
226 | * scanned from (or the end of the zone if starting). The low point | |
227 | * is the end of the pageblock the migration scanner is using. | |
228 | */ | |
748446bb MG |
229 | pfn = cc->free_pfn; |
230 | low_pfn = cc->migrate_pfn + pageblock_nr_pages; | |
7454f4ba MG |
231 | |
232 | /* | |
233 | * Take care that if the migration scanner is at the end of the zone | |
234 | * that the free scanner does not accidentally move to the next zone | |
235 | * in the next isolation cycle. | |
236 | */ | |
237 | high_pfn = min(low_pfn, pfn); | |
748446bb | 238 | |
85aa125f MN |
239 | zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages; |
240 | ||
748446bb MG |
241 | /* |
242 | * Isolate free pages until enough are available to migrate the | |
243 | * pages on cc->migratepages. We stop searching if the migrate | |
244 | * and free page scanners meet or enough free pages are isolated. | |
245 | */ | |
748446bb MG |
246 | for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages; |
247 | pfn -= pageblock_nr_pages) { | |
248 | unsigned long isolated; | |
249 | ||
250 | if (!pfn_valid(pfn)) | |
251 | continue; | |
252 | ||
253 | /* | |
254 | * Check for overlapping nodes/zones. It's possible on some | |
255 | * configurations to have a setup like | |
256 | * node0 node1 node0 | |
257 | * i.e. it's possible that all pages within a zones range of | |
258 | * pages do not belong to a single zone. | |
259 | */ | |
260 | page = pfn_to_page(pfn); | |
261 | if (page_zone(page) != zone) | |
262 | continue; | |
263 | ||
264 | /* Check the block is suitable for migration */ | |
265 | if (!suitable_migration_target(page)) | |
266 | continue; | |
267 | ||
602605a4 MG |
268 | /* |
269 | * Found a block suitable for isolating free pages from. Now | |
270 | * we disabled interrupts, double check things are ok and | |
271 | * isolate the pages. This is to minimise the time IRQs | |
272 | * are disabled | |
273 | */ | |
274 | isolated = 0; | |
275 | spin_lock_irqsave(&zone->lock, flags); | |
276 | if (suitable_migration_target(page)) { | |
85aa125f MN |
277 | end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn); |
278 | isolated = isolate_freepages_block(pfn, end_pfn, | |
279 | freelist, false); | |
602605a4 MG |
280 | nr_freepages += isolated; |
281 | } | |
282 | spin_unlock_irqrestore(&zone->lock, flags); | |
748446bb MG |
283 | |
284 | /* | |
285 | * Record the highest PFN we isolated pages from. When next | |
286 | * looking for free pages, the search will restart here as | |
287 | * page migration may have returned some pages to the allocator | |
288 | */ | |
289 | if (isolated) | |
290 | high_pfn = max(high_pfn, pfn); | |
291 | } | |
748446bb MG |
292 | |
293 | /* split_free_page does not map the pages */ | |
03d44192 | 294 | map_pages(freelist); |
748446bb MG |
295 | |
296 | cc->free_pfn = high_pfn; | |
297 | cc->nr_freepages = nr_freepages; | |
298 | } | |
299 | ||
300 | /* Update the number of anon and file isolated pages in the zone */ | |
301 | static void acct_isolated(struct zone *zone, struct compact_control *cc) | |
302 | { | |
303 | struct page *page; | |
b9e84ac1 | 304 | unsigned int count[2] = { 0, }; |
748446bb | 305 | |
b9e84ac1 MK |
306 | list_for_each_entry(page, &cc->migratepages, lru) |
307 | count[!!page_is_file_cache(page)]++; | |
748446bb | 308 | |
b9e84ac1 MK |
309 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]); |
310 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]); | |
748446bb MG |
311 | } |
312 | ||
313 | /* Similar to reclaim, but different enough that they don't share logic */ | |
314 | static bool too_many_isolated(struct zone *zone) | |
315 | { | |
bc693045 | 316 | unsigned long active, inactive, isolated; |
748446bb MG |
317 | |
318 | inactive = zone_page_state(zone, NR_INACTIVE_FILE) + | |
319 | zone_page_state(zone, NR_INACTIVE_ANON); | |
bc693045 MK |
320 | active = zone_page_state(zone, NR_ACTIVE_FILE) + |
321 | zone_page_state(zone, NR_ACTIVE_ANON); | |
748446bb MG |
322 | isolated = zone_page_state(zone, NR_ISOLATED_FILE) + |
323 | zone_page_state(zone, NR_ISOLATED_ANON); | |
324 | ||
bc693045 | 325 | return isolated > (inactive + active) / 2; |
748446bb MG |
326 | } |
327 | ||
f9e35b3b MG |
328 | /* possible outcome of isolate_migratepages */ |
329 | typedef enum { | |
330 | ISOLATE_ABORT, /* Abort compaction now */ | |
331 | ISOLATE_NONE, /* No pages isolated, continue scanning */ | |
332 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ | |
333 | } isolate_migrate_t; | |
334 | ||
2fe86e00 MN |
335 | /** |
336 | * isolate_migratepages_range() - isolate all migrate-able pages in range. | |
337 | * @zone: Zone pages are in. | |
338 | * @cc: Compaction control structure. | |
339 | * @low_pfn: The first PFN of the range. | |
340 | * @end_pfn: The one-past-the-last PFN of the range. | |
341 | * | |
342 | * Isolate all pages that can be migrated from the range specified by | |
343 | * [low_pfn, end_pfn). Returns zero if there is a fatal signal | |
344 | * pending), otherwise PFN of the first page that was not scanned | |
345 | * (which may be both less, equal to or more then end_pfn). | |
346 | * | |
347 | * Assumes that cc->migratepages is empty and cc->nr_migratepages is | |
348 | * zero. | |
349 | * | |
350 | * Apart from cc->migratepages and cc->nr_migratetypes this function | |
351 | * does not modify any cc's fields, in particular it does not modify | |
352 | * (or read for that matter) cc->migrate_pfn. | |
748446bb | 353 | */ |
2fe86e00 MN |
354 | static unsigned long |
355 | isolate_migratepages_range(struct zone *zone, struct compact_control *cc, | |
356 | unsigned long low_pfn, unsigned long end_pfn) | |
748446bb | 357 | { |
9927af74 | 358 | unsigned long last_pageblock_nr = 0, pageblock_nr; |
b7aba698 | 359 | unsigned long nr_scanned = 0, nr_isolated = 0; |
748446bb | 360 | struct list_head *migratelist = &cc->migratepages; |
39deaf85 | 361 | isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE; |
748446bb | 362 | |
748446bb MG |
363 | /* |
364 | * Ensure that there are not too many pages isolated from the LRU | |
365 | * list by either parallel reclaimers or compaction. If there are, | |
366 | * delay for some time until fewer pages are isolated | |
367 | */ | |
368 | while (unlikely(too_many_isolated(zone))) { | |
f9e35b3b MG |
369 | /* async migration should just abort */ |
370 | if (!cc->sync) | |
2fe86e00 | 371 | return 0; |
f9e35b3b | 372 | |
748446bb MG |
373 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
374 | ||
375 | if (fatal_signal_pending(current)) | |
2fe86e00 | 376 | return 0; |
748446bb MG |
377 | } |
378 | ||
379 | /* Time to isolate some pages for migration */ | |
b2eef8c0 | 380 | cond_resched(); |
748446bb MG |
381 | spin_lock_irq(&zone->lru_lock); |
382 | for (; low_pfn < end_pfn; low_pfn++) { | |
383 | struct page *page; | |
b2eef8c0 AA |
384 | bool locked = true; |
385 | ||
386 | /* give a chance to irqs before checking need_resched() */ | |
387 | if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) { | |
388 | spin_unlock_irq(&zone->lru_lock); | |
389 | locked = false; | |
390 | } | |
391 | if (need_resched() || spin_is_contended(&zone->lru_lock)) { | |
392 | if (locked) | |
393 | spin_unlock_irq(&zone->lru_lock); | |
394 | cond_resched(); | |
395 | spin_lock_irq(&zone->lru_lock); | |
396 | if (fatal_signal_pending(current)) | |
397 | break; | |
398 | } else if (!locked) | |
399 | spin_lock_irq(&zone->lru_lock); | |
400 | ||
0bf380bc MG |
401 | /* |
402 | * migrate_pfn does not necessarily start aligned to a | |
403 | * pageblock. Ensure that pfn_valid is called when moving | |
404 | * into a new MAX_ORDER_NR_PAGES range in case of large | |
405 | * memory holes within the zone | |
406 | */ | |
407 | if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { | |
408 | if (!pfn_valid(low_pfn)) { | |
409 | low_pfn += MAX_ORDER_NR_PAGES - 1; | |
410 | continue; | |
411 | } | |
412 | } | |
413 | ||
748446bb MG |
414 | if (!pfn_valid_within(low_pfn)) |
415 | continue; | |
b7aba698 | 416 | nr_scanned++; |
748446bb | 417 | |
dc908600 MG |
418 | /* |
419 | * Get the page and ensure the page is within the same zone. | |
420 | * See the comment in isolate_freepages about overlapping | |
421 | * nodes. It is deliberate that the new zone lock is not taken | |
422 | * as memory compaction should not move pages between nodes. | |
423 | */ | |
748446bb | 424 | page = pfn_to_page(low_pfn); |
dc908600 MG |
425 | if (page_zone(page) != zone) |
426 | continue; | |
427 | ||
428 | /* Skip if free */ | |
748446bb MG |
429 | if (PageBuddy(page)) |
430 | continue; | |
431 | ||
9927af74 MG |
432 | /* |
433 | * For async migration, also only scan in MOVABLE blocks. Async | |
434 | * migration is optimistic to see if the minimum amount of work | |
435 | * satisfies the allocation | |
436 | */ | |
437 | pageblock_nr = low_pfn >> pageblock_order; | |
438 | if (!cc->sync && last_pageblock_nr != pageblock_nr && | |
439 | get_pageblock_migratetype(page) != MIGRATE_MOVABLE) { | |
440 | low_pfn += pageblock_nr_pages; | |
441 | low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1; | |
442 | last_pageblock_nr = pageblock_nr; | |
443 | continue; | |
444 | } | |
445 | ||
bc835011 AA |
446 | if (!PageLRU(page)) |
447 | continue; | |
448 | ||
449 | /* | |
450 | * PageLRU is set, and lru_lock excludes isolation, | |
451 | * splitting and collapsing (collapsing has already | |
452 | * happened if PageLRU is set). | |
453 | */ | |
454 | if (PageTransHuge(page)) { | |
455 | low_pfn += (1 << compound_order(page)) - 1; | |
456 | continue; | |
457 | } | |
458 | ||
c8244935 MG |
459 | if (!cc->sync) |
460 | mode |= ISOLATE_ASYNC_MIGRATE; | |
461 | ||
748446bb | 462 | /* Try isolate the page */ |
39deaf85 | 463 | if (__isolate_lru_page(page, mode, 0) != 0) |
748446bb MG |
464 | continue; |
465 | ||
bc835011 AA |
466 | VM_BUG_ON(PageTransCompound(page)); |
467 | ||
748446bb MG |
468 | /* Successfully isolated */ |
469 | del_page_from_lru_list(zone, page, page_lru(page)); | |
470 | list_add(&page->lru, migratelist); | |
748446bb | 471 | cc->nr_migratepages++; |
b7aba698 | 472 | nr_isolated++; |
748446bb MG |
473 | |
474 | /* Avoid isolating too much */ | |
31b8384a HD |
475 | if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) { |
476 | ++low_pfn; | |
748446bb | 477 | break; |
31b8384a | 478 | } |
748446bb MG |
479 | } |
480 | ||
481 | acct_isolated(zone, cc); | |
482 | ||
483 | spin_unlock_irq(&zone->lru_lock); | |
748446bb | 484 | |
b7aba698 MG |
485 | trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated); |
486 | ||
2fe86e00 MN |
487 | return low_pfn; |
488 | } | |
489 | ||
490 | /* | |
491 | * Isolate all pages that can be migrated from the block pointed to by | |
492 | * the migrate scanner within compact_control. | |
493 | */ | |
494 | static isolate_migrate_t isolate_migratepages(struct zone *zone, | |
495 | struct compact_control *cc) | |
496 | { | |
497 | unsigned long low_pfn, end_pfn; | |
498 | ||
499 | /* Do not scan outside zone boundaries */ | |
500 | low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); | |
501 | ||
502 | /* Only scan within a pageblock boundary */ | |
503 | end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); | |
504 | ||
505 | /* Do not cross the free scanner or scan within a memory hole */ | |
506 | if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { | |
507 | cc->migrate_pfn = end_pfn; | |
508 | return ISOLATE_NONE; | |
509 | } | |
510 | ||
511 | /* Perform the isolation */ | |
512 | low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn); | |
513 | if (!low_pfn) | |
514 | return ISOLATE_ABORT; | |
515 | ||
516 | cc->migrate_pfn = low_pfn; | |
517 | ||
f9e35b3b | 518 | return ISOLATE_SUCCESS; |
748446bb MG |
519 | } |
520 | ||
521 | /* | |
522 | * This is a migrate-callback that "allocates" freepages by taking pages | |
523 | * from the isolated freelists in the block we are migrating to. | |
524 | */ | |
525 | static struct page *compaction_alloc(struct page *migratepage, | |
526 | unsigned long data, | |
527 | int **result) | |
528 | { | |
529 | struct compact_control *cc = (struct compact_control *)data; | |
530 | struct page *freepage; | |
531 | ||
532 | /* Isolate free pages if necessary */ | |
533 | if (list_empty(&cc->freepages)) { | |
534 | isolate_freepages(cc->zone, cc); | |
535 | ||
536 | if (list_empty(&cc->freepages)) | |
537 | return NULL; | |
538 | } | |
539 | ||
540 | freepage = list_entry(cc->freepages.next, struct page, lru); | |
541 | list_del(&freepage->lru); | |
542 | cc->nr_freepages--; | |
543 | ||
544 | return freepage; | |
545 | } | |
546 | ||
547 | /* | |
548 | * We cannot control nr_migratepages and nr_freepages fully when migration is | |
549 | * running as migrate_pages() has no knowledge of compact_control. When | |
550 | * migration is complete, we count the number of pages on the lists by hand. | |
551 | */ | |
552 | static void update_nr_listpages(struct compact_control *cc) | |
553 | { | |
554 | int nr_migratepages = 0; | |
555 | int nr_freepages = 0; | |
556 | struct page *page; | |
557 | ||
558 | list_for_each_entry(page, &cc->migratepages, lru) | |
559 | nr_migratepages++; | |
560 | list_for_each_entry(page, &cc->freepages, lru) | |
561 | nr_freepages++; | |
562 | ||
563 | cc->nr_migratepages = nr_migratepages; | |
564 | cc->nr_freepages = nr_freepages; | |
565 | } | |
566 | ||
567 | static int compact_finished(struct zone *zone, | |
5a03b051 | 568 | struct compact_control *cc) |
748446bb | 569 | { |
56de7263 | 570 | unsigned int order; |
5a03b051 | 571 | unsigned long watermark; |
56de7263 | 572 | |
748446bb MG |
573 | if (fatal_signal_pending(current)) |
574 | return COMPACT_PARTIAL; | |
575 | ||
576 | /* Compaction run completes if the migrate and free scanner meet */ | |
577 | if (cc->free_pfn <= cc->migrate_pfn) | |
578 | return COMPACT_COMPLETE; | |
579 | ||
82478fb7 JW |
580 | /* |
581 | * order == -1 is expected when compacting via | |
582 | * /proc/sys/vm/compact_memory | |
583 | */ | |
56de7263 MG |
584 | if (cc->order == -1) |
585 | return COMPACT_CONTINUE; | |
586 | ||
3957c776 MH |
587 | /* Compaction run is not finished if the watermark is not met */ |
588 | watermark = low_wmark_pages(zone); | |
589 | watermark += (1 << cc->order); | |
590 | ||
591 | if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) | |
592 | return COMPACT_CONTINUE; | |
593 | ||
56de7263 MG |
594 | /* Direct compactor: Is a suitable page free? */ |
595 | for (order = cc->order; order < MAX_ORDER; order++) { | |
596 | /* Job done if page is free of the right migratetype */ | |
597 | if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) | |
598 | return COMPACT_PARTIAL; | |
599 | ||
600 | /* Job done if allocation would set block type */ | |
601 | if (order >= pageblock_order && zone->free_area[order].nr_free) | |
602 | return COMPACT_PARTIAL; | |
603 | } | |
604 | ||
748446bb MG |
605 | return COMPACT_CONTINUE; |
606 | } | |
607 | ||
3e7d3449 MG |
608 | /* |
609 | * compaction_suitable: Is this suitable to run compaction on this zone now? | |
610 | * Returns | |
611 | * COMPACT_SKIPPED - If there are too few free pages for compaction | |
612 | * COMPACT_PARTIAL - If the allocation would succeed without compaction | |
613 | * COMPACT_CONTINUE - If compaction should run now | |
614 | */ | |
615 | unsigned long compaction_suitable(struct zone *zone, int order) | |
616 | { | |
617 | int fragindex; | |
618 | unsigned long watermark; | |
619 | ||
3957c776 MH |
620 | /* |
621 | * order == -1 is expected when compacting via | |
622 | * /proc/sys/vm/compact_memory | |
623 | */ | |
624 | if (order == -1) | |
625 | return COMPACT_CONTINUE; | |
626 | ||
3e7d3449 MG |
627 | /* |
628 | * Watermarks for order-0 must be met for compaction. Note the 2UL. | |
629 | * This is because during migration, copies of pages need to be | |
630 | * allocated and for a short time, the footprint is higher | |
631 | */ | |
632 | watermark = low_wmark_pages(zone) + (2UL << order); | |
633 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
634 | return COMPACT_SKIPPED; | |
635 | ||
636 | /* | |
637 | * fragmentation index determines if allocation failures are due to | |
638 | * low memory or external fragmentation | |
639 | * | |
a582a738 SL |
640 | * index of -1000 implies allocations might succeed depending on |
641 | * watermarks | |
3e7d3449 MG |
642 | * index towards 0 implies failure is due to lack of memory |
643 | * index towards 1000 implies failure is due to fragmentation | |
644 | * | |
645 | * Only compact if a failure would be due to fragmentation. | |
646 | */ | |
647 | fragindex = fragmentation_index(zone, order); | |
648 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) | |
649 | return COMPACT_SKIPPED; | |
650 | ||
a582a738 SL |
651 | if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark, |
652 | 0, 0)) | |
3e7d3449 MG |
653 | return COMPACT_PARTIAL; |
654 | ||
655 | return COMPACT_CONTINUE; | |
656 | } | |
657 | ||
748446bb MG |
658 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
659 | { | |
660 | int ret; | |
661 | ||
3e7d3449 MG |
662 | ret = compaction_suitable(zone, cc->order); |
663 | switch (ret) { | |
664 | case COMPACT_PARTIAL: | |
665 | case COMPACT_SKIPPED: | |
666 | /* Compaction is likely to fail */ | |
667 | return ret; | |
668 | case COMPACT_CONTINUE: | |
669 | /* Fall through to compaction */ | |
670 | ; | |
671 | } | |
672 | ||
748446bb MG |
673 | /* Setup to move all movable pages to the end of the zone */ |
674 | cc->migrate_pfn = zone->zone_start_pfn; | |
675 | cc->free_pfn = cc->migrate_pfn + zone->spanned_pages; | |
676 | cc->free_pfn &= ~(pageblock_nr_pages-1); | |
677 | ||
678 | migrate_prep_local(); | |
679 | ||
680 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { | |
681 | unsigned long nr_migrate, nr_remaining; | |
9d502c1c | 682 | int err; |
748446bb | 683 | |
f9e35b3b MG |
684 | switch (isolate_migratepages(zone, cc)) { |
685 | case ISOLATE_ABORT: | |
686 | ret = COMPACT_PARTIAL; | |
687 | goto out; | |
688 | case ISOLATE_NONE: | |
748446bb | 689 | continue; |
f9e35b3b MG |
690 | case ISOLATE_SUCCESS: |
691 | ; | |
692 | } | |
748446bb MG |
693 | |
694 | nr_migrate = cc->nr_migratepages; | |
9d502c1c | 695 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
7f0f2496 | 696 | (unsigned long)cc, false, |
a6bc32b8 | 697 | cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC); |
748446bb MG |
698 | update_nr_listpages(cc); |
699 | nr_remaining = cc->nr_migratepages; | |
700 | ||
701 | count_vm_event(COMPACTBLOCKS); | |
702 | count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); | |
703 | if (nr_remaining) | |
704 | count_vm_events(COMPACTPAGEFAILED, nr_remaining); | |
b7aba698 MG |
705 | trace_mm_compaction_migratepages(nr_migrate - nr_remaining, |
706 | nr_remaining); | |
748446bb MG |
707 | |
708 | /* Release LRU pages not migrated */ | |
9d502c1c | 709 | if (err) { |
748446bb MG |
710 | putback_lru_pages(&cc->migratepages); |
711 | cc->nr_migratepages = 0; | |
712 | } | |
713 | ||
714 | } | |
715 | ||
f9e35b3b | 716 | out: |
748446bb MG |
717 | /* Release free pages and check accounting */ |
718 | cc->nr_freepages -= release_freepages(&cc->freepages); | |
719 | VM_BUG_ON(cc->nr_freepages != 0); | |
720 | ||
721 | return ret; | |
722 | } | |
76ab0f53 | 723 | |
d43a87e6 | 724 | static unsigned long compact_zone_order(struct zone *zone, |
5a03b051 | 725 | int order, gfp_t gfp_mask, |
d527caf2 | 726 | bool sync) |
56de7263 MG |
727 | { |
728 | struct compact_control cc = { | |
729 | .nr_freepages = 0, | |
730 | .nr_migratepages = 0, | |
731 | .order = order, | |
732 | .migratetype = allocflags_to_migratetype(gfp_mask), | |
733 | .zone = zone, | |
77f1fe6b | 734 | .sync = sync, |
56de7263 MG |
735 | }; |
736 | INIT_LIST_HEAD(&cc.freepages); | |
737 | INIT_LIST_HEAD(&cc.migratepages); | |
738 | ||
739 | return compact_zone(zone, &cc); | |
740 | } | |
741 | ||
5e771905 MG |
742 | int sysctl_extfrag_threshold = 500; |
743 | ||
56de7263 MG |
744 | /** |
745 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation | |
746 | * @zonelist: The zonelist used for the current allocation | |
747 | * @order: The order of the current allocation | |
748 | * @gfp_mask: The GFP mask of the current allocation | |
749 | * @nodemask: The allowed nodes to allocate from | |
77f1fe6b | 750 | * @sync: Whether migration is synchronous or not |
56de7263 MG |
751 | * |
752 | * This is the main entry point for direct page compaction. | |
753 | */ | |
754 | unsigned long try_to_compact_pages(struct zonelist *zonelist, | |
77f1fe6b MG |
755 | int order, gfp_t gfp_mask, nodemask_t *nodemask, |
756 | bool sync) | |
56de7263 MG |
757 | { |
758 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
759 | int may_enter_fs = gfp_mask & __GFP_FS; | |
760 | int may_perform_io = gfp_mask & __GFP_IO; | |
56de7263 MG |
761 | struct zoneref *z; |
762 | struct zone *zone; | |
763 | int rc = COMPACT_SKIPPED; | |
764 | ||
765 | /* | |
766 | * Check whether it is worth even starting compaction. The order check is | |
767 | * made because an assumption is made that the page allocator can satisfy | |
768 | * the "cheaper" orders without taking special steps | |
769 | */ | |
c5a73c3d | 770 | if (!order || !may_enter_fs || !may_perform_io) |
56de7263 MG |
771 | return rc; |
772 | ||
773 | count_vm_event(COMPACTSTALL); | |
774 | ||
775 | /* Compact each zone in the list */ | |
776 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, | |
777 | nodemask) { | |
56de7263 MG |
778 | int status; |
779 | ||
d527caf2 | 780 | status = compact_zone_order(zone, order, gfp_mask, sync); |
56de7263 MG |
781 | rc = max(status, rc); |
782 | ||
3e7d3449 MG |
783 | /* If a normal allocation would succeed, stop compacting */ |
784 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) | |
56de7263 MG |
785 | break; |
786 | } | |
787 | ||
788 | return rc; | |
789 | } | |
790 | ||
791 | ||
76ab0f53 | 792 | /* Compact all zones within a node */ |
7be62de9 | 793 | static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
76ab0f53 MG |
794 | { |
795 | int zoneid; | |
76ab0f53 MG |
796 | struct zone *zone; |
797 | ||
76ab0f53 | 798 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
76ab0f53 MG |
799 | |
800 | zone = &pgdat->node_zones[zoneid]; | |
801 | if (!populated_zone(zone)) | |
802 | continue; | |
803 | ||
7be62de9 RR |
804 | cc->nr_freepages = 0; |
805 | cc->nr_migratepages = 0; | |
806 | cc->zone = zone; | |
807 | INIT_LIST_HEAD(&cc->freepages); | |
808 | INIT_LIST_HEAD(&cc->migratepages); | |
76ab0f53 | 809 | |
aad6ec37 | 810 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
7be62de9 | 811 | compact_zone(zone, cc); |
76ab0f53 | 812 | |
aff62249 RR |
813 | if (cc->order > 0) { |
814 | int ok = zone_watermark_ok(zone, cc->order, | |
815 | low_wmark_pages(zone), 0, 0); | |
816 | if (ok && cc->order > zone->compact_order_failed) | |
817 | zone->compact_order_failed = cc->order + 1; | |
818 | /* Currently async compaction is never deferred. */ | |
819 | else if (!ok && cc->sync) | |
820 | defer_compaction(zone, cc->order); | |
821 | } | |
822 | ||
7be62de9 RR |
823 | VM_BUG_ON(!list_empty(&cc->freepages)); |
824 | VM_BUG_ON(!list_empty(&cc->migratepages)); | |
76ab0f53 MG |
825 | } |
826 | ||
827 | return 0; | |
828 | } | |
829 | ||
7be62de9 RR |
830 | int compact_pgdat(pg_data_t *pgdat, int order) |
831 | { | |
832 | struct compact_control cc = { | |
833 | .order = order, | |
834 | .sync = false, | |
835 | }; | |
836 | ||
837 | return __compact_pgdat(pgdat, &cc); | |
838 | } | |
839 | ||
840 | static int compact_node(int nid) | |
841 | { | |
7be62de9 RR |
842 | struct compact_control cc = { |
843 | .order = -1, | |
844 | .sync = true, | |
845 | }; | |
846 | ||
8575ec29 | 847 | return __compact_pgdat(NODE_DATA(nid), &cc); |
7be62de9 RR |
848 | } |
849 | ||
76ab0f53 MG |
850 | /* Compact all nodes in the system */ |
851 | static int compact_nodes(void) | |
852 | { | |
853 | int nid; | |
854 | ||
8575ec29 HD |
855 | /* Flush pending updates to the LRU lists */ |
856 | lru_add_drain_all(); | |
857 | ||
76ab0f53 MG |
858 | for_each_online_node(nid) |
859 | compact_node(nid); | |
860 | ||
861 | return COMPACT_COMPLETE; | |
862 | } | |
863 | ||
864 | /* The written value is actually unused, all memory is compacted */ | |
865 | int sysctl_compact_memory; | |
866 | ||
867 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ | |
868 | int sysctl_compaction_handler(struct ctl_table *table, int write, | |
869 | void __user *buffer, size_t *length, loff_t *ppos) | |
870 | { | |
871 | if (write) | |
872 | return compact_nodes(); | |
873 | ||
874 | return 0; | |
875 | } | |
ed4a6d7f | 876 | |
5e771905 MG |
877 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
878 | void __user *buffer, size_t *length, loff_t *ppos) | |
879 | { | |
880 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
881 | ||
882 | return 0; | |
883 | } | |
884 | ||
ed4a6d7f | 885 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
10fbcf4c KS |
886 | ssize_t sysfs_compact_node(struct device *dev, |
887 | struct device_attribute *attr, | |
ed4a6d7f MG |
888 | const char *buf, size_t count) |
889 | { | |
8575ec29 HD |
890 | int nid = dev->id; |
891 | ||
892 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { | |
893 | /* Flush pending updates to the LRU lists */ | |
894 | lru_add_drain_all(); | |
895 | ||
896 | compact_node(nid); | |
897 | } | |
ed4a6d7f MG |
898 | |
899 | return count; | |
900 | } | |
10fbcf4c | 901 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
ed4a6d7f MG |
902 | |
903 | int compaction_register_node(struct node *node) | |
904 | { | |
10fbcf4c | 905 | return device_create_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
906 | } |
907 | ||
908 | void compaction_unregister_node(struct node *node) | |
909 | { | |
10fbcf4c | 910 | return device_remove_file(&node->dev, &dev_attr_compact); |
ed4a6d7f MG |
911 | } |
912 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ |