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