blob: 71a58f67f4817720a4eca34678b2353fa96863e6 [file] [log] [blame]
Mel Gorman748446b2010-05-24 14:32:27 -07001/*
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>
Mel Gorman76ab0f52010-05-24 14:32:28 -070015#include <linux/sysctl.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070016#include <linux/sysfs.h>
Mel Gorman748446b2010-05-24 14:32:27 -070017#include "internal.h"
18
Mel Gormanb7aba692011-01-13 15:45:54 -080019#define CREATE_TRACE_POINTS
20#include <trace/events/compaction.h>
21
Mel Gorman748446b2010-05-24 14:32:27 -070022/*
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 */
29struct 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 */
Mel Gorman77f1fe62011-01-13 15:45:57 -080036 bool sync; /* Synchronous migration */
Mel Gorman748446b2010-05-24 14:32:27 -070037
Mel Gorman56de7262010-05-24 14:32:30 -070038 unsigned int order; /* order a direct compactor needs */
39 int migratetype; /* MOVABLE, RECLAIMABLE etc */
Mel Gorman748446b2010-05-24 14:32:27 -070040 struct zone *zone;
41};
42
43static 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 */
58static 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;
Mel Gormanb7aba692011-01-13 15:45:54 -080063 int nr_scanned = 0, total_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -070064 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;
Mel Gormanb7aba692011-01-13 15:45:54 -080084 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -070085
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
Mel Gormanb7aba692011-01-13 15:45:54 -0800104 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700105 return total_isolated;
106}
107
108/* Returns true if the page is within a block suitable for migration to */
109static 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
130/*
131 * Based on information in the current compact_control, find blocks
132 * suitable for isolating free pages from and then isolate them.
133 */
134static void isolate_freepages(struct zone *zone,
135 struct compact_control *cc)
136{
137 struct page *page;
138 unsigned long high_pfn, low_pfn, pfn;
139 unsigned long flags;
140 int nr_freepages = cc->nr_freepages;
141 struct list_head *freelist = &cc->freepages;
142
Mel Gorman7454f4b2011-06-15 15:08:50 -0700143 /*
144 * Initialise the free scanner. The starting point is where we last
145 * scanned from (or the end of the zone if starting). The low point
146 * is the end of the pageblock the migration scanner is using.
147 */
Mel Gorman748446b2010-05-24 14:32:27 -0700148 pfn = cc->free_pfn;
149 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
Mel Gorman7454f4b2011-06-15 15:08:50 -0700150
151 /*
152 * Take care that if the migration scanner is at the end of the zone
153 * that the free scanner does not accidentally move to the next zone
154 * in the next isolation cycle.
155 */
156 high_pfn = min(low_pfn, pfn);
Mel Gorman748446b2010-05-24 14:32:27 -0700157
158 /*
159 * Isolate free pages until enough are available to migrate the
160 * pages on cc->migratepages. We stop searching if the migrate
161 * and free page scanners meet or enough free pages are isolated.
162 */
Mel Gorman748446b2010-05-24 14:32:27 -0700163 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
164 pfn -= pageblock_nr_pages) {
165 unsigned long isolated;
166
167 if (!pfn_valid(pfn))
168 continue;
169
170 /*
171 * Check for overlapping nodes/zones. It's possible on some
172 * configurations to have a setup like
173 * node0 node1 node0
174 * i.e. it's possible that all pages within a zones range of
175 * pages do not belong to a single zone.
176 */
177 page = pfn_to_page(pfn);
178 if (page_zone(page) != zone)
179 continue;
180
181 /* Check the block is suitable for migration */
182 if (!suitable_migration_target(page))
183 continue;
184
Mel Gorman602605a2011-03-22 16:33:08 -0700185 /*
186 * Found a block suitable for isolating free pages from. Now
187 * we disabled interrupts, double check things are ok and
188 * isolate the pages. This is to minimise the time IRQs
189 * are disabled
190 */
191 isolated = 0;
192 spin_lock_irqsave(&zone->lock, flags);
193 if (suitable_migration_target(page)) {
194 isolated = isolate_freepages_block(zone, pfn, freelist);
195 nr_freepages += isolated;
196 }
197 spin_unlock_irqrestore(&zone->lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700198
199 /*
200 * Record the highest PFN we isolated pages from. When next
201 * looking for free pages, the search will restart here as
202 * page migration may have returned some pages to the allocator
203 */
204 if (isolated)
205 high_pfn = max(high_pfn, pfn);
206 }
Mel Gorman748446b2010-05-24 14:32:27 -0700207
208 /* split_free_page does not map the pages */
209 list_for_each_entry(page, freelist, lru) {
210 arch_alloc_page(page, 0);
211 kernel_map_pages(page, 1, 1);
212 }
213
214 cc->free_pfn = high_pfn;
215 cc->nr_freepages = nr_freepages;
216}
217
218/* Update the number of anon and file isolated pages in the zone */
219static void acct_isolated(struct zone *zone, struct compact_control *cc)
220{
221 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700222 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700223
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700224 list_for_each_entry(page, &cc->migratepages, lru)
225 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700226
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700227 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
228 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
Mel Gorman748446b2010-05-24 14:32:27 -0700229}
230
231/* Similar to reclaim, but different enough that they don't share logic */
232static bool too_many_isolated(struct zone *zone)
233{
Minchan Kimbc693042010-09-09 16:38:00 -0700234 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700235
236 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
237 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700238 active = zone_page_state(zone, NR_ACTIVE_FILE) +
239 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700240 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
241 zone_page_state(zone, NR_ISOLATED_ANON);
242
Minchan Kimbc693042010-09-09 16:38:00 -0700243 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700244}
245
Mel Gormanf9e35b32011-06-15 15:08:52 -0700246/* possible outcome of isolate_migratepages */
247typedef enum {
248 ISOLATE_ABORT, /* Abort compaction now */
249 ISOLATE_NONE, /* No pages isolated, continue scanning */
250 ISOLATE_SUCCESS, /* Pages isolated, migrate */
251} isolate_migrate_t;
252
Mel Gorman748446b2010-05-24 14:32:27 -0700253/*
254 * Isolate all pages that can be migrated from the block pointed to by
255 * the migrate scanner within compact_control.
256 */
Mel Gormanf9e35b32011-06-15 15:08:52 -0700257static isolate_migrate_t isolate_migratepages(struct zone *zone,
Mel Gorman748446b2010-05-24 14:32:27 -0700258 struct compact_control *cc)
259{
260 unsigned long low_pfn, end_pfn;
Mel Gorman9927af742011-01-13 15:45:59 -0800261 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800262 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700263 struct list_head *migratelist = &cc->migratepages;
Minchan Kim39deaf82011-10-31 17:06:51 -0700264 isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE;
Mel Gorman748446b2010-05-24 14:32:27 -0700265
266 /* Do not scan outside zone boundaries */
267 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
268
269 /* Only scan within a pageblock boundary */
270 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
271
272 /* Do not cross the free scanner or scan within a memory hole */
273 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
274 cc->migrate_pfn = end_pfn;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700275 return ISOLATE_NONE;
Mel Gorman748446b2010-05-24 14:32:27 -0700276 }
277
278 /*
279 * Ensure that there are not too many pages isolated from the LRU
280 * list by either parallel reclaimers or compaction. If there are,
281 * delay for some time until fewer pages are isolated
282 */
283 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700284 /* async migration should just abort */
285 if (!cc->sync)
286 return ISOLATE_ABORT;
287
Mel Gorman748446b2010-05-24 14:32:27 -0700288 congestion_wait(BLK_RW_ASYNC, HZ/10);
289
290 if (fatal_signal_pending(current))
Mel Gormanf9e35b32011-06-15 15:08:52 -0700291 return ISOLATE_ABORT;
Mel Gorman748446b2010-05-24 14:32:27 -0700292 }
293
294 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700295 cond_resched();
Mel Gorman748446b2010-05-24 14:32:27 -0700296 spin_lock_irq(&zone->lru_lock);
297 for (; low_pfn < end_pfn; low_pfn++) {
298 struct page *page;
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700299 bool locked = true;
300
301 /* give a chance to irqs before checking need_resched() */
302 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
303 spin_unlock_irq(&zone->lru_lock);
304 locked = false;
305 }
306 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
307 if (locked)
308 spin_unlock_irq(&zone->lru_lock);
309 cond_resched();
310 spin_lock_irq(&zone->lru_lock);
311 if (fatal_signal_pending(current))
312 break;
313 } else if (!locked)
314 spin_lock_irq(&zone->lru_lock);
315
Mel Gorman748446b2010-05-24 14:32:27 -0700316 if (!pfn_valid_within(low_pfn))
317 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800318 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700319
320 /* Get the page and skip if free */
321 page = pfn_to_page(low_pfn);
322 if (PageBuddy(page))
323 continue;
324
Mel Gorman9927af742011-01-13 15:45:59 -0800325 /*
326 * For async migration, also only scan in MOVABLE blocks. Async
327 * migration is optimistic to see if the minimum amount of work
328 * satisfies the allocation
329 */
330 pageblock_nr = low_pfn >> pageblock_order;
331 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
332 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
333 low_pfn += pageblock_nr_pages;
334 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
335 last_pageblock_nr = pageblock_nr;
336 continue;
337 }
338
Andrea Arcangelibc835012011-01-13 15:47:08 -0800339 if (!PageLRU(page))
340 continue;
341
342 /*
343 * PageLRU is set, and lru_lock excludes isolation,
344 * splitting and collapsing (collapsing has already
345 * happened if PageLRU is set).
346 */
347 if (PageTransHuge(page)) {
348 low_pfn += (1 << compound_order(page)) - 1;
349 continue;
350 }
351
Mel Gormanc8244932012-01-12 17:19:38 -0800352 if (!cc->sync)
353 mode |= ISOLATE_ASYNC_MIGRATE;
354
Mel Gorman748446b2010-05-24 14:32:27 -0700355 /* Try isolate the page */
Minchan Kim39deaf82011-10-31 17:06:51 -0700356 if (__isolate_lru_page(page, mode, 0) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700357 continue;
358
Andrea Arcangelibc835012011-01-13 15:47:08 -0800359 VM_BUG_ON(PageTransCompound(page));
360
Mel Gorman748446b2010-05-24 14:32:27 -0700361 /* Successfully isolated */
362 del_page_from_lru_list(zone, page, page_lru(page));
363 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700364 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800365 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700366
367 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800368 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
369 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700370 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800371 }
Mel Gorman748446b2010-05-24 14:32:27 -0700372 }
373
374 acct_isolated(zone, cc);
375
376 spin_unlock_irq(&zone->lru_lock);
377 cc->migrate_pfn = low_pfn;
378
Mel Gormanb7aba692011-01-13 15:45:54 -0800379 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
380
Mel Gormanf9e35b32011-06-15 15:08:52 -0700381 return ISOLATE_SUCCESS;
Mel Gorman748446b2010-05-24 14:32:27 -0700382}
383
384/*
385 * This is a migrate-callback that "allocates" freepages by taking pages
386 * from the isolated freelists in the block we are migrating to.
387 */
388static struct page *compaction_alloc(struct page *migratepage,
389 unsigned long data,
390 int **result)
391{
392 struct compact_control *cc = (struct compact_control *)data;
393 struct page *freepage;
394
395 /* Isolate free pages if necessary */
396 if (list_empty(&cc->freepages)) {
397 isolate_freepages(cc->zone, cc);
398
399 if (list_empty(&cc->freepages))
400 return NULL;
401 }
402
403 freepage = list_entry(cc->freepages.next, struct page, lru);
404 list_del(&freepage->lru);
405 cc->nr_freepages--;
406
407 return freepage;
408}
409
410/*
411 * We cannot control nr_migratepages and nr_freepages fully when migration is
412 * running as migrate_pages() has no knowledge of compact_control. When
413 * migration is complete, we count the number of pages on the lists by hand.
414 */
415static void update_nr_listpages(struct compact_control *cc)
416{
417 int nr_migratepages = 0;
418 int nr_freepages = 0;
419 struct page *page;
420
421 list_for_each_entry(page, &cc->migratepages, lru)
422 nr_migratepages++;
423 list_for_each_entry(page, &cc->freepages, lru)
424 nr_freepages++;
425
426 cc->nr_migratepages = nr_migratepages;
427 cc->nr_freepages = nr_freepages;
428}
429
430static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800431 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700432{
Mel Gorman56de7262010-05-24 14:32:30 -0700433 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800434 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700435
Mel Gorman748446b2010-05-24 14:32:27 -0700436 if (fatal_signal_pending(current))
437 return COMPACT_PARTIAL;
438
439 /* Compaction run completes if the migrate and free scanner meet */
440 if (cc->free_pfn <= cc->migrate_pfn)
441 return COMPACT_COMPLETE;
442
Johannes Weiner82478fb2011-01-20 14:44:21 -0800443 /*
444 * order == -1 is expected when compacting via
445 * /proc/sys/vm/compact_memory
446 */
Mel Gorman56de7262010-05-24 14:32:30 -0700447 if (cc->order == -1)
448 return COMPACT_CONTINUE;
449
Michal Hocko3957c772011-06-15 15:08:25 -0700450 /* Compaction run is not finished if the watermark is not met */
451 watermark = low_wmark_pages(zone);
452 watermark += (1 << cc->order);
453
454 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
455 return COMPACT_CONTINUE;
456
Mel Gorman56de7262010-05-24 14:32:30 -0700457 /* Direct compactor: Is a suitable page free? */
458 for (order = cc->order; order < MAX_ORDER; order++) {
459 /* Job done if page is free of the right migratetype */
460 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
461 return COMPACT_PARTIAL;
462
463 /* Job done if allocation would set block type */
464 if (order >= pageblock_order && zone->free_area[order].nr_free)
465 return COMPACT_PARTIAL;
466 }
467
Mel Gorman748446b2010-05-24 14:32:27 -0700468 return COMPACT_CONTINUE;
469}
470
Mel Gorman3e7d3442011-01-13 15:45:56 -0800471/*
472 * compaction_suitable: Is this suitable to run compaction on this zone now?
473 * Returns
474 * COMPACT_SKIPPED - If there are too few free pages for compaction
475 * COMPACT_PARTIAL - If the allocation would succeed without compaction
476 * COMPACT_CONTINUE - If compaction should run now
477 */
478unsigned long compaction_suitable(struct zone *zone, int order)
479{
480 int fragindex;
481 unsigned long watermark;
482
483 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700484 * order == -1 is expected when compacting via
485 * /proc/sys/vm/compact_memory
486 */
487 if (order == -1)
488 return COMPACT_CONTINUE;
489
490 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800491 * Watermarks for order-0 must be met for compaction. Note the 2UL.
492 * This is because during migration, copies of pages need to be
493 * allocated and for a short time, the footprint is higher
494 */
495 watermark = low_wmark_pages(zone) + (2UL << order);
496 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
497 return COMPACT_SKIPPED;
498
499 /*
500 * fragmentation index determines if allocation failures are due to
501 * low memory or external fragmentation
502 *
Shaohua Lia582a732011-06-15 15:08:49 -0700503 * index of -1000 implies allocations might succeed depending on
504 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800505 * index towards 0 implies failure is due to lack of memory
506 * index towards 1000 implies failure is due to fragmentation
507 *
508 * Only compact if a failure would be due to fragmentation.
509 */
510 fragindex = fragmentation_index(zone, order);
511 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
512 return COMPACT_SKIPPED;
513
Shaohua Lia582a732011-06-15 15:08:49 -0700514 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
515 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800516 return COMPACT_PARTIAL;
517
518 return COMPACT_CONTINUE;
519}
520
Mel Gorman748446b2010-05-24 14:32:27 -0700521static int compact_zone(struct zone *zone, struct compact_control *cc)
522{
523 int ret;
524
Mel Gorman3e7d3442011-01-13 15:45:56 -0800525 ret = compaction_suitable(zone, cc->order);
526 switch (ret) {
527 case COMPACT_PARTIAL:
528 case COMPACT_SKIPPED:
529 /* Compaction is likely to fail */
530 return ret;
531 case COMPACT_CONTINUE:
532 /* Fall through to compaction */
533 ;
534 }
535
Mel Gorman748446b2010-05-24 14:32:27 -0700536 /* Setup to move all movable pages to the end of the zone */
537 cc->migrate_pfn = zone->zone_start_pfn;
538 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
539 cc->free_pfn &= ~(pageblock_nr_pages-1);
540
541 migrate_prep_local();
542
543 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
544 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -0700545 int err;
Mel Gorman748446b2010-05-24 14:32:27 -0700546
Mel Gormanf9e35b32011-06-15 15:08:52 -0700547 switch (isolate_migratepages(zone, cc)) {
548 case ISOLATE_ABORT:
549 ret = COMPACT_PARTIAL;
550 goto out;
551 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -0700552 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700553 case ISOLATE_SUCCESS:
554 ;
555 }
Mel Gorman748446b2010-05-24 14:32:27 -0700556
557 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -0700558 err = migrate_pages(&cc->migratepages, compaction_alloc,
Mel Gorman7f0f2492011-01-13 15:45:58 -0800559 (unsigned long)cc, false,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800560 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
Mel Gorman748446b2010-05-24 14:32:27 -0700561 update_nr_listpages(cc);
562 nr_remaining = cc->nr_migratepages;
563
564 count_vm_event(COMPACTBLOCKS);
565 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
566 if (nr_remaining)
567 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
Mel Gormanb7aba692011-01-13 15:45:54 -0800568 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
569 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -0700570
571 /* Release LRU pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -0700572 if (err) {
Mel Gorman748446b2010-05-24 14:32:27 -0700573 putback_lru_pages(&cc->migratepages);
574 cc->nr_migratepages = 0;
575 }
576
577 }
578
Mel Gormanf9e35b32011-06-15 15:08:52 -0700579out:
Mel Gorman748446b2010-05-24 14:32:27 -0700580 /* Release free pages and check accounting */
581 cc->nr_freepages -= release_freepages(&cc->freepages);
582 VM_BUG_ON(cc->nr_freepages != 0);
583
584 return ret;
585}
Mel Gorman76ab0f52010-05-24 14:32:28 -0700586
Kyungmin Parkd43a87e2011-10-31 17:09:08 -0700587static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800588 int order, gfp_t gfp_mask,
Andrea Arcangelid527caf2011-03-22 16:30:38 -0700589 bool sync)
Mel Gorman56de7262010-05-24 14:32:30 -0700590{
591 struct compact_control cc = {
592 .nr_freepages = 0,
593 .nr_migratepages = 0,
594 .order = order,
595 .migratetype = allocflags_to_migratetype(gfp_mask),
596 .zone = zone,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800597 .sync = sync,
Mel Gorman56de7262010-05-24 14:32:30 -0700598 };
599 INIT_LIST_HEAD(&cc.freepages);
600 INIT_LIST_HEAD(&cc.migratepages);
601
602 return compact_zone(zone, &cc);
603}
604
Mel Gorman5e771902010-05-24 14:32:31 -0700605int sysctl_extfrag_threshold = 500;
606
Mel Gorman56de7262010-05-24 14:32:30 -0700607/**
608 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
609 * @zonelist: The zonelist used for the current allocation
610 * @order: The order of the current allocation
611 * @gfp_mask: The GFP mask of the current allocation
612 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -0800613 * @sync: Whether migration is synchronous or not
Mel Gorman56de7262010-05-24 14:32:30 -0700614 *
615 * This is the main entry point for direct page compaction.
616 */
617unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800618 int order, gfp_t gfp_mask, nodemask_t *nodemask,
619 bool sync)
Mel Gorman56de7262010-05-24 14:32:30 -0700620{
621 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
622 int may_enter_fs = gfp_mask & __GFP_FS;
623 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -0700624 struct zoneref *z;
625 struct zone *zone;
626 int rc = COMPACT_SKIPPED;
627
628 /*
629 * Check whether it is worth even starting compaction. The order check is
630 * made because an assumption is made that the page allocator can satisfy
631 * the "cheaper" orders without taking special steps
632 */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -0800633 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -0700634 return rc;
635
636 count_vm_event(COMPACTSTALL);
637
638 /* Compact each zone in the list */
639 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
640 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -0700641 int status;
642
Andrea Arcangelid527caf2011-03-22 16:30:38 -0700643 status = compact_zone_order(zone, order, gfp_mask, sync);
Mel Gorman56de7262010-05-24 14:32:30 -0700644 rc = max(status, rc);
645
Mel Gorman3e7d3442011-01-13 15:45:56 -0800646 /* If a normal allocation would succeed, stop compacting */
647 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Mel Gorman56de7262010-05-24 14:32:30 -0700648 break;
649 }
650
651 return rc;
652}
653
654
Mel Gorman76ab0f52010-05-24 14:32:28 -0700655/* Compact all zones within a node */
656static int compact_node(int nid)
657{
658 int zoneid;
659 pg_data_t *pgdat;
660 struct zone *zone;
661
662 if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
663 return -EINVAL;
664 pgdat = NODE_DATA(nid);
665
666 /* Flush pending updates to the LRU lists */
667 lru_add_drain_all();
668
669 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
670 struct compact_control cc = {
671 .nr_freepages = 0,
672 .nr_migratepages = 0,
Mel Gorman56de7262010-05-24 14:32:30 -0700673 .order = -1,
Mel Gormanb16d3d52012-01-12 17:19:26 -0800674 .sync = true,
Mel Gorman76ab0f52010-05-24 14:32:28 -0700675 };
676
677 zone = &pgdat->node_zones[zoneid];
678 if (!populated_zone(zone))
679 continue;
680
681 cc.zone = zone;
682 INIT_LIST_HEAD(&cc.freepages);
683 INIT_LIST_HEAD(&cc.migratepages);
684
685 compact_zone(zone, &cc);
686
687 VM_BUG_ON(!list_empty(&cc.freepages));
688 VM_BUG_ON(!list_empty(&cc.migratepages));
689 }
690
691 return 0;
692}
693
694/* Compact all nodes in the system */
695static int compact_nodes(void)
696{
697 int nid;
698
699 for_each_online_node(nid)
700 compact_node(nid);
701
702 return COMPACT_COMPLETE;
703}
704
705/* The written value is actually unused, all memory is compacted */
706int sysctl_compact_memory;
707
708/* This is the entry point for compacting all nodes via /proc/sys/vm */
709int sysctl_compaction_handler(struct ctl_table *table, int write,
710 void __user *buffer, size_t *length, loff_t *ppos)
711{
712 if (write)
713 return compact_nodes();
714
715 return 0;
716}
Mel Gormaned4a6d72010-05-24 14:32:29 -0700717
Mel Gorman5e771902010-05-24 14:32:31 -0700718int sysctl_extfrag_handler(struct ctl_table *table, int write,
719 void __user *buffer, size_t *length, loff_t *ppos)
720{
721 proc_dointvec_minmax(table, write, buffer, length, ppos);
722
723 return 0;
724}
725
Mel Gormaned4a6d72010-05-24 14:32:29 -0700726#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Kay Sievers10fbcf42011-12-21 14:48:43 -0800727ssize_t sysfs_compact_node(struct device *dev,
728 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -0700729 const char *buf, size_t count)
730{
731 compact_node(dev->id);
732
733 return count;
734}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800735static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700736
737int compaction_register_node(struct node *node)
738{
Kay Sievers10fbcf42011-12-21 14:48:43 -0800739 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700740}
741
742void compaction_unregister_node(struct node *node)
743{
Kay Sievers10fbcf42011-12-21 14:48:43 -0800744 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700745}
746#endif /* CONFIG_SYSFS && CONFIG_NUMA */