blob: 7ea259d82a998c0397e976216ac861576e638100 [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
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010019#if defined CONFIG_COMPACTION || defined CONFIG_CMA
20
Mel Gormanb7aba692011-01-13 15:45:54 -080021#define CREATE_TRACE_POINTS
22#include <trace/events/compaction.h>
23
Mel Gorman748446b2010-05-24 14:32:27 -070024static unsigned long release_freepages(struct list_head *freelist)
25{
26 struct page *page, *next;
27 unsigned long count = 0;
28
29 list_for_each_entry_safe(page, next, freelist, lru) {
30 list_del(&page->lru);
31 __free_page(page);
32 count++;
33 }
34
35 return count;
36}
37
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010038static void map_pages(struct list_head *list)
39{
40 struct page *page;
41
42 list_for_each_entry(page, list, lru) {
43 arch_alloc_page(page, 0);
44 kernel_map_pages(page, 1, 1);
45 }
46}
47
Michal Nazarewicz47118af2011-12-29 13:09:50 +010048static inline bool migrate_async_suitable(int migratetype)
49{
50 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
51}
52
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010053/*
54 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
55 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
56 * pages inside of the pageblock (even though it may still end up isolating
57 * some pages).
58 */
59static unsigned long isolate_freepages_block(unsigned long blockpfn,
60 unsigned long end_pfn,
61 struct list_head *freelist,
62 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -070063{
Mel Gormanb7aba692011-01-13 15:45:54 -080064 int nr_scanned = 0, total_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -070065 struct page *cursor;
66
Mel Gorman748446b2010-05-24 14:32:27 -070067 cursor = pfn_to_page(blockpfn);
68
69 /* Isolate free pages. This assumes the block is valid */
70 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
71 int isolated, i;
72 struct page *page = cursor;
73
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010074 if (!pfn_valid_within(blockpfn)) {
75 if (strict)
76 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -070077 continue;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010078 }
Mel Gormanb7aba692011-01-13 15:45:54 -080079 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -070080
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010081 if (!PageBuddy(page)) {
82 if (strict)
83 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -070084 continue;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010085 }
Mel Gorman748446b2010-05-24 14:32:27 -070086
87 /* Found a free page, break it into order-0 pages */
88 isolated = split_free_page(page);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +010089 if (!isolated && strict)
90 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -070091 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
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100108/**
109 * isolate_freepages_range() - isolate free pages.
110 * @start_pfn: The first PFN to start isolating.
111 * @end_pfn: The one-past-last PFN.
112 *
113 * Non-free pages, invalid PFNs, or zone boundaries within the
114 * [start_pfn, end_pfn) range are considered errors, cause function to
115 * undo its actions and return zero.
116 *
117 * Otherwise, function returns one-past-the-last PFN of isolated page
118 * (which may be greater then end_pfn if end fell in a middle of
119 * a free page).
120 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100121unsigned long
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100122isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
123{
124 unsigned long isolated, pfn, block_end_pfn, flags;
125 struct zone *zone = NULL;
126 LIST_HEAD(freelist);
127
128 if (pfn_valid(start_pfn))
129 zone = page_zone(pfn_to_page(start_pfn));
130
131 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
132 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
133 break;
134
135 /*
136 * On subsequent iterations ALIGN() is actually not needed,
137 * but we keep it that we not to complicate the code.
138 */
139 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
140 block_end_pfn = min(block_end_pfn, end_pfn);
141
142 spin_lock_irqsave(&zone->lock, flags);
143 isolated = isolate_freepages_block(pfn, block_end_pfn,
144 &freelist, true);
145 spin_unlock_irqrestore(&zone->lock, flags);
146
147 /*
148 * In strict mode, isolate_freepages_block() returns 0 if
149 * there are any holes in the block (ie. invalid PFNs or
150 * non-free pages).
151 */
152 if (!isolated)
153 break;
154
155 /*
156 * If we managed to isolate pages, it is always (1 << n) *
157 * pageblock_nr_pages for some non-negative n. (Max order
158 * page may span two pageblocks).
159 */
160 }
161
162 /* split_free_page does not map the pages */
163 map_pages(&freelist);
164
165 if (pfn < end_pfn) {
166 /* Loop terminated early, cleanup. */
167 release_freepages(&freelist);
168 return 0;
169 }
170
171 /* We don't use freelists for anything. */
172 return pfn;
173}
174
Mel Gorman748446b2010-05-24 14:32:27 -0700175/* Update the number of anon and file isolated pages in the zone */
176static void acct_isolated(struct zone *zone, struct compact_control *cc)
177{
178 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700179 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700180
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700181 list_for_each_entry(page, &cc->migratepages, lru)
182 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700183
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700184 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
185 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
Mel Gorman748446b2010-05-24 14:32:27 -0700186}
187
188/* Similar to reclaim, but different enough that they don't share logic */
189static bool too_many_isolated(struct zone *zone)
190{
Minchan Kimbc693042010-09-09 16:38:00 -0700191 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700192
193 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
194 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700195 active = zone_page_state(zone, NR_ACTIVE_FILE) +
196 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700197 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
198 zone_page_state(zone, NR_ISOLATED_ANON);
199
Minchan Kimbc693042010-09-09 16:38:00 -0700200 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700201}
202
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100203/**
204 * isolate_migratepages_range() - isolate all migrate-able pages in range.
205 * @zone: Zone pages are in.
206 * @cc: Compaction control structure.
207 * @low_pfn: The first PFN of the range.
208 * @end_pfn: The one-past-the-last PFN of the range.
209 *
210 * Isolate all pages that can be migrated from the range specified by
211 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
212 * pending), otherwise PFN of the first page that was not scanned
213 * (which may be both less, equal to or more then end_pfn).
214 *
215 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
216 * zero.
217 *
218 * Apart from cc->migratepages and cc->nr_migratetypes this function
219 * does not modify any cc's fields, in particular it does not modify
220 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700221 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100222unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100223isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
224 unsigned long low_pfn, unsigned long end_pfn)
Mel Gorman748446b2010-05-24 14:32:27 -0700225{
Mel Gorman9927af742011-01-13 15:45:59 -0800226 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800227 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700228 struct list_head *migratelist = &cc->migratepages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700229 isolate_mode_t mode = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700230 struct lruvec *lruvec;
Mel Gorman748446b2010-05-24 14:32:27 -0700231
Mel Gorman748446b2010-05-24 14:32:27 -0700232 /*
233 * Ensure that there are not too many pages isolated from the LRU
234 * list by either parallel reclaimers or compaction. If there are,
235 * delay for some time until fewer pages are isolated
236 */
237 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700238 /* async migration should just abort */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700239 if (!cc->sync)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100240 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700241
Mel Gorman748446b2010-05-24 14:32:27 -0700242 congestion_wait(BLK_RW_ASYNC, HZ/10);
243
244 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100245 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700246 }
247
248 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700249 cond_resched();
Mel Gorman748446b2010-05-24 14:32:27 -0700250 spin_lock_irq(&zone->lru_lock);
251 for (; low_pfn < end_pfn; low_pfn++) {
252 struct page *page;
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700253 bool locked = true;
254
255 /* give a chance to irqs before checking need_resched() */
256 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
257 spin_unlock_irq(&zone->lru_lock);
258 locked = false;
259 }
260 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
261 if (locked)
262 spin_unlock_irq(&zone->lru_lock);
263 cond_resched();
264 spin_lock_irq(&zone->lru_lock);
265 if (fatal_signal_pending(current))
266 break;
267 } else if (!locked)
268 spin_lock_irq(&zone->lru_lock);
269
Mel Gorman0bf380b2012-02-03 15:37:18 -0800270 /*
271 * migrate_pfn does not necessarily start aligned to a
272 * pageblock. Ensure that pfn_valid is called when moving
273 * into a new MAX_ORDER_NR_PAGES range in case of large
274 * memory holes within the zone
275 */
276 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
277 if (!pfn_valid(low_pfn)) {
278 low_pfn += MAX_ORDER_NR_PAGES - 1;
279 continue;
280 }
281 }
282
Mel Gorman748446b2010-05-24 14:32:27 -0700283 if (!pfn_valid_within(low_pfn))
284 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800285 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700286
Mel Gormandc908602012-02-08 17:13:38 -0800287 /*
288 * Get the page and ensure the page is within the same zone.
289 * See the comment in isolate_freepages about overlapping
290 * nodes. It is deliberate that the new zone lock is not taken
291 * as memory compaction should not move pages between nodes.
292 */
Mel Gorman748446b2010-05-24 14:32:27 -0700293 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800294 if (page_zone(page) != zone)
295 continue;
296
297 /* Skip if free */
Mel Gorman748446b2010-05-24 14:32:27 -0700298 if (PageBuddy(page))
299 continue;
300
Mel Gorman9927af742011-01-13 15:45:59 -0800301 /*
302 * For async migration, also only scan in MOVABLE blocks. Async
303 * migration is optimistic to see if the minimum amount of work
304 * satisfies the allocation
305 */
306 pageblock_nr = low_pfn >> pageblock_order;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700307 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100308 !migrate_async_suitable(get_pageblock_migratetype(page))) {
Mel Gorman9927af742011-01-13 15:45:59 -0800309 low_pfn += pageblock_nr_pages;
310 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
311 last_pageblock_nr = pageblock_nr;
312 continue;
313 }
314
Andrea Arcangelibc835012011-01-13 15:47:08 -0800315 if (!PageLRU(page))
316 continue;
317
318 /*
319 * PageLRU is set, and lru_lock excludes isolation,
320 * splitting and collapsing (collapsing has already
321 * happened if PageLRU is set).
322 */
323 if (PageTransHuge(page)) {
324 low_pfn += (1 << compound_order(page)) - 1;
325 continue;
326 }
327
Linus Torvalds68e3e922012-06-03 20:05:57 -0700328 if (!cc->sync)
Mel Gormanc8244932012-01-12 17:19:38 -0800329 mode |= ISOLATE_ASYNC_MIGRATE;
330
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700331 lruvec = mem_cgroup_page_lruvec(page, zone);
332
Mel Gorman748446b2010-05-24 14:32:27 -0700333 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700334 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700335 continue;
336
Andrea Arcangelibc835012011-01-13 15:47:08 -0800337 VM_BUG_ON(PageTransCompound(page));
338
Mel Gorman748446b2010-05-24 14:32:27 -0700339 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700340 del_page_from_lru_list(page, lruvec, page_lru(page));
Mel Gorman748446b2010-05-24 14:32:27 -0700341 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700342 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800343 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700344
345 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800346 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
347 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700348 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800349 }
Mel Gorman748446b2010-05-24 14:32:27 -0700350 }
351
352 acct_isolated(zone, cc);
353
354 spin_unlock_irq(&zone->lru_lock);
Mel Gorman748446b2010-05-24 14:32:27 -0700355
Mel Gormanb7aba692011-01-13 15:45:54 -0800356 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
357
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100358 return low_pfn;
359}
360
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100361#endif /* CONFIG_COMPACTION || CONFIG_CMA */
362#ifdef CONFIG_COMPACTION
363
Linus Torvalds68e3e922012-06-03 20:05:57 -0700364/* Returns true if the page is within a block suitable for migration to */
365static bool suitable_migration_target(struct page *page)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100366{
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100367
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100368 int migratetype = get_pageblock_migratetype(page);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100369
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100370 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
371 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
Linus Torvalds68e3e922012-06-03 20:05:57 -0700372 return false;
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100373
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100374 /* If the page is a large free page, then allow migration */
375 if (PageBuddy(page) && page_order(page) >= pageblock_order)
Linus Torvalds68e3e922012-06-03 20:05:57 -0700376 return true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100377
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100378 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700379 if (migrate_async_suitable(migratetype))
380 return true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100381
382 /* Otherwise skip the block */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700383 return false;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100384}
385
386/*
387 * Based on information in the current compact_control, find blocks
388 * suitable for isolating free pages from and then isolate them.
389 */
390static void isolate_freepages(struct zone *zone,
391 struct compact_control *cc)
392{
393 struct page *page;
394 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
395 unsigned long flags;
396 int nr_freepages = cc->nr_freepages;
397 struct list_head *freelist = &cc->freepages;
398
399 /*
400 * Initialise the free scanner. The starting point is where we last
401 * scanned from (or the end of the zone if starting). The low point
402 * is the end of the pageblock the migration scanner is using.
403 */
404 pfn = cc->free_pfn;
405 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
406
407 /*
408 * Take care that if the migration scanner is at the end of the zone
409 * that the free scanner does not accidentally move to the next zone
410 * in the next isolation cycle.
411 */
412 high_pfn = min(low_pfn, pfn);
413
414 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
415
416 /*
417 * Isolate free pages until enough are available to migrate the
418 * pages on cc->migratepages. We stop searching if the migrate
419 * and free page scanners meet or enough free pages are isolated.
420 */
421 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
422 pfn -= pageblock_nr_pages) {
423 unsigned long isolated;
424
425 if (!pfn_valid(pfn))
426 continue;
427
428 /*
429 * Check for overlapping nodes/zones. It's possible on some
430 * configurations to have a setup like
431 * node0 node1 node0
432 * i.e. it's possible that all pages within a zones range of
433 * pages do not belong to a single zone.
434 */
435 page = pfn_to_page(pfn);
436 if (page_zone(page) != zone)
437 continue;
438
439 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700440 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100441 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700442
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100443 /*
444 * Found a block suitable for isolating free pages from. Now
445 * we disabled interrupts, double check things are ok and
446 * isolate the pages. This is to minimise the time IRQs
447 * are disabled
448 */
449 isolated = 0;
450 spin_lock_irqsave(&zone->lock, flags);
Linus Torvalds68e3e922012-06-03 20:05:57 -0700451 if (suitable_migration_target(page)) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100452 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
453 isolated = isolate_freepages_block(pfn, end_pfn,
454 freelist, false);
455 nr_freepages += isolated;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700456 }
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100457 spin_unlock_irqrestore(&zone->lock, flags);
458
459 /*
460 * Record the highest PFN we isolated pages from. When next
461 * looking for free pages, the search will restart here as
462 * page migration may have returned some pages to the allocator
463 */
464 if (isolated)
465 high_pfn = max(high_pfn, pfn);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100466 }
467
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100468 /* split_free_page does not map the pages */
469 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100470
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100471 cc->free_pfn = high_pfn;
472 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700473}
474
475/*
476 * This is a migrate-callback that "allocates" freepages by taking pages
477 * from the isolated freelists in the block we are migrating to.
478 */
479static struct page *compaction_alloc(struct page *migratepage,
480 unsigned long data,
481 int **result)
482{
483 struct compact_control *cc = (struct compact_control *)data;
484 struct page *freepage;
485
486 /* Isolate free pages if necessary */
487 if (list_empty(&cc->freepages)) {
488 isolate_freepages(cc->zone, cc);
489
490 if (list_empty(&cc->freepages))
491 return NULL;
492 }
493
494 freepage = list_entry(cc->freepages.next, struct page, lru);
495 list_del(&freepage->lru);
496 cc->nr_freepages--;
497
498 return freepage;
499}
500
501/*
502 * We cannot control nr_migratepages and nr_freepages fully when migration is
503 * running as migrate_pages() has no knowledge of compact_control. When
504 * migration is complete, we count the number of pages on the lists by hand.
505 */
506static void update_nr_listpages(struct compact_control *cc)
507{
508 int nr_migratepages = 0;
509 int nr_freepages = 0;
510 struct page *page;
511
512 list_for_each_entry(page, &cc->migratepages, lru)
513 nr_migratepages++;
514 list_for_each_entry(page, &cc->freepages, lru)
515 nr_freepages++;
516
517 cc->nr_migratepages = nr_migratepages;
518 cc->nr_freepages = nr_freepages;
519}
520
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100521/* possible outcome of isolate_migratepages */
522typedef enum {
523 ISOLATE_ABORT, /* Abort compaction now */
524 ISOLATE_NONE, /* No pages isolated, continue scanning */
525 ISOLATE_SUCCESS, /* Pages isolated, migrate */
526} isolate_migrate_t;
527
528/*
529 * Isolate all pages that can be migrated from the block pointed to by
530 * the migrate scanner within compact_control.
531 */
532static isolate_migrate_t isolate_migratepages(struct zone *zone,
533 struct compact_control *cc)
534{
535 unsigned long low_pfn, end_pfn;
536
537 /* Do not scan outside zone boundaries */
538 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
539
540 /* Only scan within a pageblock boundary */
541 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
542
543 /* Do not cross the free scanner or scan within a memory hole */
544 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
545 cc->migrate_pfn = end_pfn;
546 return ISOLATE_NONE;
547 }
548
549 /* Perform the isolation */
550 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
551 if (!low_pfn)
552 return ISOLATE_ABORT;
553
554 cc->migrate_pfn = low_pfn;
555
556 return ISOLATE_SUCCESS;
557}
558
Mel Gorman748446b2010-05-24 14:32:27 -0700559static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800560 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700561{
Mel Gorman56de7262010-05-24 14:32:30 -0700562 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800563 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700564
Mel Gorman748446b2010-05-24 14:32:27 -0700565 if (fatal_signal_pending(current))
566 return COMPACT_PARTIAL;
567
568 /* Compaction run completes if the migrate and free scanner meet */
569 if (cc->free_pfn <= cc->migrate_pfn)
570 return COMPACT_COMPLETE;
571
Johannes Weiner82478fb2011-01-20 14:44:21 -0800572 /*
573 * order == -1 is expected when compacting via
574 * /proc/sys/vm/compact_memory
575 */
Mel Gorman56de7262010-05-24 14:32:30 -0700576 if (cc->order == -1)
577 return COMPACT_CONTINUE;
578
Michal Hocko3957c772011-06-15 15:08:25 -0700579 /* Compaction run is not finished if the watermark is not met */
580 watermark = low_wmark_pages(zone);
581 watermark += (1 << cc->order);
582
583 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
584 return COMPACT_CONTINUE;
585
Mel Gorman56de7262010-05-24 14:32:30 -0700586 /* Direct compactor: Is a suitable page free? */
587 for (order = cc->order; order < MAX_ORDER; order++) {
588 /* Job done if page is free of the right migratetype */
589 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
590 return COMPACT_PARTIAL;
591
592 /* Job done if allocation would set block type */
593 if (order >= pageblock_order && zone->free_area[order].nr_free)
594 return COMPACT_PARTIAL;
595 }
596
Mel Gorman748446b2010-05-24 14:32:27 -0700597 return COMPACT_CONTINUE;
598}
599
Mel Gorman3e7d3442011-01-13 15:45:56 -0800600/*
601 * compaction_suitable: Is this suitable to run compaction on this zone now?
602 * Returns
603 * COMPACT_SKIPPED - If there are too few free pages for compaction
604 * COMPACT_PARTIAL - If the allocation would succeed without compaction
605 * COMPACT_CONTINUE - If compaction should run now
606 */
607unsigned long compaction_suitable(struct zone *zone, int order)
608{
609 int fragindex;
610 unsigned long watermark;
611
612 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700613 * order == -1 is expected when compacting via
614 * /proc/sys/vm/compact_memory
615 */
616 if (order == -1)
617 return COMPACT_CONTINUE;
618
619 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800620 * Watermarks for order-0 must be met for compaction. Note the 2UL.
621 * This is because during migration, copies of pages need to be
622 * allocated and for a short time, the footprint is higher
623 */
624 watermark = low_wmark_pages(zone) + (2UL << order);
625 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
626 return COMPACT_SKIPPED;
627
628 /*
629 * fragmentation index determines if allocation failures are due to
630 * low memory or external fragmentation
631 *
Shaohua Lia582a732011-06-15 15:08:49 -0700632 * index of -1000 implies allocations might succeed depending on
633 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800634 * index towards 0 implies failure is due to lack of memory
635 * index towards 1000 implies failure is due to fragmentation
636 *
637 * Only compact if a failure would be due to fragmentation.
638 */
639 fragindex = fragmentation_index(zone, order);
640 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
641 return COMPACT_SKIPPED;
642
Shaohua Lia582a732011-06-15 15:08:49 -0700643 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
644 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800645 return COMPACT_PARTIAL;
646
647 return COMPACT_CONTINUE;
648}
649
Mel Gorman748446b2010-05-24 14:32:27 -0700650static int compact_zone(struct zone *zone, struct compact_control *cc)
651{
652 int ret;
653
Mel Gorman3e7d3442011-01-13 15:45:56 -0800654 ret = compaction_suitable(zone, cc->order);
655 switch (ret) {
656 case COMPACT_PARTIAL:
657 case COMPACT_SKIPPED:
658 /* Compaction is likely to fail */
659 return ret;
660 case COMPACT_CONTINUE:
661 /* Fall through to compaction */
662 ;
663 }
664
Mel Gorman748446b2010-05-24 14:32:27 -0700665 /* Setup to move all movable pages to the end of the zone */
666 cc->migrate_pfn = zone->zone_start_pfn;
667 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
668 cc->free_pfn &= ~(pageblock_nr_pages-1);
669
670 migrate_prep_local();
671
672 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
673 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -0700674 int err;
Mel Gorman748446b2010-05-24 14:32:27 -0700675
Mel Gormanf9e35b32011-06-15 15:08:52 -0700676 switch (isolate_migratepages(zone, cc)) {
677 case ISOLATE_ABORT:
678 ret = COMPACT_PARTIAL;
679 goto out;
680 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -0700681 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700682 case ISOLATE_SUCCESS:
683 ;
684 }
Mel Gorman748446b2010-05-24 14:32:27 -0700685
686 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -0700687 err = migrate_pages(&cc->migratepages, compaction_alloc,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700688 (unsigned long)cc, false,
689 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
Mel Gorman748446b2010-05-24 14:32:27 -0700690 update_nr_listpages(cc);
691 nr_remaining = cc->nr_migratepages;
692
693 count_vm_event(COMPACTBLOCKS);
694 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
695 if (nr_remaining)
696 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
Mel Gormanb7aba692011-01-13 15:45:54 -0800697 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
698 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -0700699
700 /* Release LRU pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -0700701 if (err) {
Mel Gorman748446b2010-05-24 14:32:27 -0700702 putback_lru_pages(&cc->migratepages);
703 cc->nr_migratepages = 0;
704 }
705
706 }
707
Mel Gormanf9e35b32011-06-15 15:08:52 -0700708out:
Mel Gorman748446b2010-05-24 14:32:27 -0700709 /* Release free pages and check accounting */
710 cc->nr_freepages -= release_freepages(&cc->freepages);
711 VM_BUG_ON(cc->nr_freepages != 0);
712
713 return ret;
714}
Mel Gorman76ab0f52010-05-24 14:32:28 -0700715
Kyungmin Parkd43a87e2011-10-31 17:09:08 -0700716static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800717 int order, gfp_t gfp_mask,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700718 bool sync)
Mel Gorman56de7262010-05-24 14:32:30 -0700719{
720 struct compact_control cc = {
721 .nr_freepages = 0,
722 .nr_migratepages = 0,
723 .order = order,
724 .migratetype = allocflags_to_migratetype(gfp_mask),
725 .zone = zone,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700726 .sync = sync,
Mel Gorman56de7262010-05-24 14:32:30 -0700727 };
728 INIT_LIST_HEAD(&cc.freepages);
729 INIT_LIST_HEAD(&cc.migratepages);
730
Linus Torvalds68e3e922012-06-03 20:05:57 -0700731 return compact_zone(zone, &cc);
Mel Gorman56de7262010-05-24 14:32:30 -0700732}
733
Mel Gorman5e771902010-05-24 14:32:31 -0700734int sysctl_extfrag_threshold = 500;
735
Mel Gorman56de7262010-05-24 14:32:30 -0700736/**
737 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
738 * @zonelist: The zonelist used for the current allocation
739 * @order: The order of the current allocation
740 * @gfp_mask: The GFP mask of the current allocation
741 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -0800742 * @sync: Whether migration is synchronous or not
Mel Gorman56de7262010-05-24 14:32:30 -0700743 *
744 * This is the main entry point for direct page compaction.
745 */
746unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800747 int order, gfp_t gfp_mask, nodemask_t *nodemask,
748 bool sync)
Mel Gorman56de7262010-05-24 14:32:30 -0700749{
750 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
751 int may_enter_fs = gfp_mask & __GFP_FS;
752 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -0700753 struct zoneref *z;
754 struct zone *zone;
755 int rc = COMPACT_SKIPPED;
756
757 /*
758 * Check whether it is worth even starting compaction. The order check is
759 * made because an assumption is made that the page allocator can satisfy
760 * the "cheaper" orders without taking special steps
761 */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -0800762 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -0700763 return rc;
764
765 count_vm_event(COMPACTSTALL);
766
767 /* Compact each zone in the list */
768 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
769 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -0700770 int status;
771
Linus Torvalds68e3e922012-06-03 20:05:57 -0700772 status = compact_zone_order(zone, order, gfp_mask, sync);
Mel Gorman56de7262010-05-24 14:32:30 -0700773 rc = max(status, rc);
774
Mel Gorman3e7d3442011-01-13 15:45:56 -0800775 /* If a normal allocation would succeed, stop compacting */
776 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Mel Gorman56de7262010-05-24 14:32:30 -0700777 break;
778 }
779
780 return rc;
781}
782
783
Mel Gorman76ab0f52010-05-24 14:32:28 -0700784/* Compact all zones within a node */
Rik van Riel7be62de2012-03-21 16:33:52 -0700785static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -0700786{
787 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -0700788 struct zone *zone;
789
Mel Gorman76ab0f52010-05-24 14:32:28 -0700790 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -0700791
792 zone = &pgdat->node_zones[zoneid];
793 if (!populated_zone(zone))
794 continue;
795
Rik van Riel7be62de2012-03-21 16:33:52 -0700796 cc->nr_freepages = 0;
797 cc->nr_migratepages = 0;
798 cc->zone = zone;
799 INIT_LIST_HEAD(&cc->freepages);
800 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -0700801
Dan Carpenteraad6ec32012-03-21 16:33:54 -0700802 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -0700803 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -0700804
Rik van Rielaff62242012-03-21 16:33:52 -0700805 if (cc->order > 0) {
806 int ok = zone_watermark_ok(zone, cc->order,
807 low_wmark_pages(zone), 0, 0);
808 if (ok && cc->order > zone->compact_order_failed)
809 zone->compact_order_failed = cc->order + 1;
810 /* Currently async compaction is never deferred. */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700811 else if (!ok && cc->sync)
Rik van Rielaff62242012-03-21 16:33:52 -0700812 defer_compaction(zone, cc->order);
813 }
814
Rik van Riel7be62de2012-03-21 16:33:52 -0700815 VM_BUG_ON(!list_empty(&cc->freepages));
816 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -0700817 }
818
819 return 0;
820}
821
Rik van Riel7be62de2012-03-21 16:33:52 -0700822int compact_pgdat(pg_data_t *pgdat, int order)
823{
824 struct compact_control cc = {
825 .order = order,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700826 .sync = false,
Rik van Riel7be62de2012-03-21 16:33:52 -0700827 };
828
829 return __compact_pgdat(pgdat, &cc);
830}
831
832static int compact_node(int nid)
833{
Rik van Riel7be62de2012-03-21 16:33:52 -0700834 struct compact_control cc = {
835 .order = -1,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700836 .sync = true,
Rik van Riel7be62de2012-03-21 16:33:52 -0700837 };
838
Hugh Dickins8575ec22012-03-21 16:33:53 -0700839 return __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -0700840}
841
Mel Gorman76ab0f52010-05-24 14:32:28 -0700842/* Compact all nodes in the system */
843static int compact_nodes(void)
844{
845 int nid;
846
Hugh Dickins8575ec22012-03-21 16:33:53 -0700847 /* Flush pending updates to the LRU lists */
848 lru_add_drain_all();
849
Mel Gorman76ab0f52010-05-24 14:32:28 -0700850 for_each_online_node(nid)
851 compact_node(nid);
852
853 return COMPACT_COMPLETE;
854}
855
856/* The written value is actually unused, all memory is compacted */
857int sysctl_compact_memory;
858
859/* This is the entry point for compacting all nodes via /proc/sys/vm */
860int sysctl_compaction_handler(struct ctl_table *table, int write,
861 void __user *buffer, size_t *length, loff_t *ppos)
862{
863 if (write)
864 return compact_nodes();
865
866 return 0;
867}
Mel Gormaned4a6d72010-05-24 14:32:29 -0700868
Mel Gorman5e771902010-05-24 14:32:31 -0700869int sysctl_extfrag_handler(struct ctl_table *table, int write,
870 void __user *buffer, size_t *length, loff_t *ppos)
871{
872 proc_dointvec_minmax(table, write, buffer, length, ppos);
873
874 return 0;
875}
876
Mel Gormaned4a6d72010-05-24 14:32:29 -0700877#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Kay Sievers10fbcf42011-12-21 14:48:43 -0800878ssize_t sysfs_compact_node(struct device *dev,
879 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -0700880 const char *buf, size_t count)
881{
Hugh Dickins8575ec22012-03-21 16:33:53 -0700882 int nid = dev->id;
883
884 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
885 /* Flush pending updates to the LRU lists */
886 lru_add_drain_all();
887
888 compact_node(nid);
889 }
Mel Gormaned4a6d72010-05-24 14:32:29 -0700890
891 return count;
892}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800893static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700894
895int compaction_register_node(struct node *node)
896{
Kay Sievers10fbcf42011-12-21 14:48:43 -0800897 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700898}
899
900void compaction_unregister_node(struct node *node)
901{
Kay Sievers10fbcf42011-12-21 14:48:43 -0800902 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700903}
904#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100905
906#endif /* CONFIG_COMPACTION */