blob: 0fbc6b73a522c1210cfc3b13b441e2762ec4e2e8 [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/*
Mel Gormanc67fe372012-08-21 16:16:17 -070054 * Compaction requires the taking of some coarse locks that are potentially
55 * very heavily contended. Check if the process needs to be scheduled or
56 * if the lock is contended. For async compaction, back out in the event
57 * if contention is severe. For sync compaction, schedule.
58 *
59 * Returns true if the lock is held.
60 * Returns false if the lock is released and compaction should abort
61 */
62static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
63 bool locked, struct compact_control *cc)
64{
65 if (need_resched() || spin_is_contended(lock)) {
66 if (locked) {
67 spin_unlock_irqrestore(lock, *flags);
68 locked = false;
69 }
70
71 /* async aborts if taking too long or contended */
72 if (!cc->sync) {
73 if (cc->contended)
74 *cc->contended = true;
75 return false;
76 }
77
78 cond_resched();
79 if (fatal_signal_pending(current))
80 return false;
81 }
82
83 if (!locked)
84 spin_lock_irqsave(lock, *flags);
85 return true;
86}
87
88static inline bool compact_trylock_irqsave(spinlock_t *lock,
89 unsigned long *flags, struct compact_control *cc)
90{
91 return compact_checklock_irqsave(lock, flags, false, cc);
92}
93
Mel Gorman1fb3f8c2012-10-08 16:29:12 -070094static void compact_capture_page(struct compact_control *cc)
95{
96 unsigned long flags;
97 int mtype, mtype_low, mtype_high;
98
99 if (!cc->page || *cc->page)
100 return;
101
102 /*
103 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
104 * regardless of the migratetype of the freelist is is captured from.
105 * This is fine because the order for a high-order MIGRATE_MOVABLE
106 * allocation is typically at least a pageblock size and overall
107 * fragmentation is not impaired. Other allocation types must
108 * capture pages from their own migratelist because otherwise they
109 * could pollute other pageblocks like MIGRATE_MOVABLE with
110 * difficult to move pages and making fragmentation worse overall.
111 */
112 if (cc->migratetype == MIGRATE_MOVABLE) {
113 mtype_low = 0;
114 mtype_high = MIGRATE_PCPTYPES;
115 } else {
116 mtype_low = cc->migratetype;
117 mtype_high = cc->migratetype + 1;
118 }
119
120 /* Speculatively examine the free lists without zone lock */
121 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
122 int order;
123 for (order = cc->order; order < MAX_ORDER; order++) {
124 struct page *page;
125 struct free_area *area;
126 area = &(cc->zone->free_area[order]);
127 if (list_empty(&area->free_list[mtype]))
128 continue;
129
130 /* Take the lock and attempt capture of the page */
131 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
132 return;
133 if (!list_empty(&area->free_list[mtype])) {
134 page = list_entry(area->free_list[mtype].next,
135 struct page, lru);
136 if (capture_free_page(page, cc->order, mtype)) {
137 spin_unlock_irqrestore(&cc->zone->lock,
138 flags);
139 *cc->page = page;
140 return;
141 }
142 }
143 spin_unlock_irqrestore(&cc->zone->lock, flags);
144 }
145 }
146}
147
Mel Gormanc67fe372012-08-21 16:16:17 -0700148/*
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100149 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
150 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
151 * pages inside of the pageblock (even though it may still end up isolating
152 * some pages).
153 */
154static unsigned long isolate_freepages_block(unsigned long blockpfn,
155 unsigned long end_pfn,
156 struct list_head *freelist,
157 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700158{
Mel Gormanb7aba692011-01-13 15:45:54 -0800159 int nr_scanned = 0, total_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700160 struct page *cursor;
161
Mel Gorman748446b2010-05-24 14:32:27 -0700162 cursor = pfn_to_page(blockpfn);
163
164 /* Isolate free pages. This assumes the block is valid */
165 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
166 int isolated, i;
167 struct page *page = cursor;
168
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100169 if (!pfn_valid_within(blockpfn)) {
170 if (strict)
171 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700172 continue;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100173 }
Mel Gormanb7aba692011-01-13 15:45:54 -0800174 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700175
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100176 if (!PageBuddy(page)) {
177 if (strict)
178 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700179 continue;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100180 }
Mel Gorman748446b2010-05-24 14:32:27 -0700181
182 /* Found a free page, break it into order-0 pages */
183 isolated = split_free_page(page);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100184 if (!isolated && strict)
185 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700186 total_isolated += isolated;
187 for (i = 0; i < isolated; i++) {
188 list_add(&page->lru, freelist);
189 page++;
190 }
191
192 /* If a page was split, advance to the end of it */
193 if (isolated) {
194 blockpfn += isolated - 1;
195 cursor += isolated - 1;
196 }
197 }
198
Mel Gormanb7aba692011-01-13 15:45:54 -0800199 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700200 return total_isolated;
201}
202
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100203/**
204 * isolate_freepages_range() - isolate free pages.
205 * @start_pfn: The first PFN to start isolating.
206 * @end_pfn: The one-past-last PFN.
207 *
208 * Non-free pages, invalid PFNs, or zone boundaries within the
209 * [start_pfn, end_pfn) range are considered errors, cause function to
210 * undo its actions and return zero.
211 *
212 * Otherwise, function returns one-past-the-last PFN of isolated page
213 * (which may be greater then end_pfn if end fell in a middle of
214 * a free page).
215 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100216unsigned long
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100217isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
218{
219 unsigned long isolated, pfn, block_end_pfn, flags;
220 struct zone *zone = NULL;
221 LIST_HEAD(freelist);
222
223 if (pfn_valid(start_pfn))
224 zone = page_zone(pfn_to_page(start_pfn));
225
226 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
227 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
228 break;
229
230 /*
231 * On subsequent iterations ALIGN() is actually not needed,
232 * but we keep it that we not to complicate the code.
233 */
234 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
235 block_end_pfn = min(block_end_pfn, end_pfn);
236
237 spin_lock_irqsave(&zone->lock, flags);
238 isolated = isolate_freepages_block(pfn, block_end_pfn,
239 &freelist, true);
240 spin_unlock_irqrestore(&zone->lock, flags);
241
242 /*
243 * In strict mode, isolate_freepages_block() returns 0 if
244 * there are any holes in the block (ie. invalid PFNs or
245 * non-free pages).
246 */
247 if (!isolated)
248 break;
249
250 /*
251 * If we managed to isolate pages, it is always (1 << n) *
252 * pageblock_nr_pages for some non-negative n. (Max order
253 * page may span two pageblocks).
254 */
255 }
256
257 /* split_free_page does not map the pages */
258 map_pages(&freelist);
259
260 if (pfn < end_pfn) {
261 /* Loop terminated early, cleanup. */
262 release_freepages(&freelist);
263 return 0;
264 }
265
266 /* We don't use freelists for anything. */
267 return pfn;
268}
269
Mel Gorman748446b2010-05-24 14:32:27 -0700270/* Update the number of anon and file isolated pages in the zone */
Mel Gormanc67fe372012-08-21 16:16:17 -0700271static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700272{
273 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700274 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700275
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700276 list_for_each_entry(page, &cc->migratepages, lru)
277 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700278
Mel Gormanc67fe372012-08-21 16:16:17 -0700279 /* If locked we can use the interrupt unsafe versions */
280 if (locked) {
281 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
282 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
283 } else {
284 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
285 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
286 }
Mel Gorman748446b2010-05-24 14:32:27 -0700287}
288
289/* Similar to reclaim, but different enough that they don't share logic */
290static bool too_many_isolated(struct zone *zone)
291{
Minchan Kimbc693042010-09-09 16:38:00 -0700292 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700293
294 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
295 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700296 active = zone_page_state(zone, NR_ACTIVE_FILE) +
297 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700298 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
299 zone_page_state(zone, NR_ISOLATED_ANON);
300
Minchan Kimbc693042010-09-09 16:38:00 -0700301 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700302}
303
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100304/**
305 * isolate_migratepages_range() - isolate all migrate-able pages in range.
306 * @zone: Zone pages are in.
307 * @cc: Compaction control structure.
308 * @low_pfn: The first PFN of the range.
309 * @end_pfn: The one-past-the-last PFN of the range.
310 *
311 * Isolate all pages that can be migrated from the range specified by
312 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
313 * pending), otherwise PFN of the first page that was not scanned
314 * (which may be both less, equal to or more then end_pfn).
315 *
316 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
317 * zero.
318 *
319 * Apart from cc->migratepages and cc->nr_migratetypes this function
320 * does not modify any cc's fields, in particular it does not modify
321 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700322 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100323unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100324isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
325 unsigned long low_pfn, unsigned long end_pfn)
Mel Gorman748446b2010-05-24 14:32:27 -0700326{
Mel Gorman9927af742011-01-13 15:45:59 -0800327 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800328 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700329 struct list_head *migratelist = &cc->migratepages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700330 isolate_mode_t mode = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700331 struct lruvec *lruvec;
Mel Gormanc67fe372012-08-21 16:16:17 -0700332 unsigned long flags;
333 bool locked;
Mel Gorman748446b2010-05-24 14:32:27 -0700334
Mel Gorman748446b2010-05-24 14:32:27 -0700335 /*
336 * Ensure that there are not too many pages isolated from the LRU
337 * list by either parallel reclaimers or compaction. If there are,
338 * delay for some time until fewer pages are isolated
339 */
340 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700341 /* async migration should just abort */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700342 if (!cc->sync)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100343 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700344
Mel Gorman748446b2010-05-24 14:32:27 -0700345 congestion_wait(BLK_RW_ASYNC, HZ/10);
346
347 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100348 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700349 }
350
351 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700352 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700353 spin_lock_irqsave(&zone->lru_lock, flags);
354 locked = true;
Mel Gorman748446b2010-05-24 14:32:27 -0700355 for (; low_pfn < end_pfn; low_pfn++) {
356 struct page *page;
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700357
358 /* give a chance to irqs before checking need_resched() */
359 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
Mel Gormanc67fe372012-08-21 16:16:17 -0700360 spin_unlock_irqrestore(&zone->lru_lock, flags);
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700361 locked = false;
362 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700363
364 /* Check if it is ok to still hold the lock */
365 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
366 locked, cc);
367 if (!locked)
368 break;
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700369
Mel Gorman0bf380b2012-02-03 15:37:18 -0800370 /*
371 * migrate_pfn does not necessarily start aligned to a
372 * pageblock. Ensure that pfn_valid is called when moving
373 * into a new MAX_ORDER_NR_PAGES range in case of large
374 * memory holes within the zone
375 */
376 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
377 if (!pfn_valid(low_pfn)) {
378 low_pfn += MAX_ORDER_NR_PAGES - 1;
379 continue;
380 }
381 }
382
Mel Gorman748446b2010-05-24 14:32:27 -0700383 if (!pfn_valid_within(low_pfn))
384 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800385 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700386
Mel Gormandc908602012-02-08 17:13:38 -0800387 /*
388 * Get the page and ensure the page is within the same zone.
389 * See the comment in isolate_freepages about overlapping
390 * nodes. It is deliberate that the new zone lock is not taken
391 * as memory compaction should not move pages between nodes.
392 */
Mel Gorman748446b2010-05-24 14:32:27 -0700393 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800394 if (page_zone(page) != zone)
395 continue;
396
397 /* Skip if free */
Mel Gorman748446b2010-05-24 14:32:27 -0700398 if (PageBuddy(page))
399 continue;
400
Mel Gorman9927af742011-01-13 15:45:59 -0800401 /*
402 * For async migration, also only scan in MOVABLE blocks. Async
403 * migration is optimistic to see if the minimum amount of work
404 * satisfies the allocation
405 */
406 pageblock_nr = low_pfn >> pageblock_order;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700407 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100408 !migrate_async_suitable(get_pageblock_migratetype(page))) {
Mel Gorman9927af742011-01-13 15:45:59 -0800409 low_pfn += pageblock_nr_pages;
410 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
411 last_pageblock_nr = pageblock_nr;
412 continue;
413 }
414
Andrea Arcangelibc835012011-01-13 15:47:08 -0800415 if (!PageLRU(page))
416 continue;
417
418 /*
419 * PageLRU is set, and lru_lock excludes isolation,
420 * splitting and collapsing (collapsing has already
421 * happened if PageLRU is set).
422 */
423 if (PageTransHuge(page)) {
424 low_pfn += (1 << compound_order(page)) - 1;
425 continue;
426 }
427
Linus Torvalds68e3e922012-06-03 20:05:57 -0700428 if (!cc->sync)
Mel Gormanc8244932012-01-12 17:19:38 -0800429 mode |= ISOLATE_ASYNC_MIGRATE;
430
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700431 lruvec = mem_cgroup_page_lruvec(page, zone);
432
Mel Gorman748446b2010-05-24 14:32:27 -0700433 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700434 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700435 continue;
436
Andrea Arcangelibc835012011-01-13 15:47:08 -0800437 VM_BUG_ON(PageTransCompound(page));
438
Mel Gorman748446b2010-05-24 14:32:27 -0700439 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700440 del_page_from_lru_list(page, lruvec, page_lru(page));
Mel Gorman748446b2010-05-24 14:32:27 -0700441 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700442 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800443 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700444
445 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800446 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
447 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700448 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800449 }
Mel Gorman748446b2010-05-24 14:32:27 -0700450 }
451
Mel Gormanc67fe372012-08-21 16:16:17 -0700452 acct_isolated(zone, locked, cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700453
Mel Gormanc67fe372012-08-21 16:16:17 -0700454 if (locked)
455 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700456
Mel Gormanb7aba692011-01-13 15:45:54 -0800457 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
458
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100459 return low_pfn;
460}
461
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100462#endif /* CONFIG_COMPACTION || CONFIG_CMA */
463#ifdef CONFIG_COMPACTION
464
Linus Torvalds68e3e922012-06-03 20:05:57 -0700465/* Returns true if the page is within a block suitable for migration to */
466static bool suitable_migration_target(struct page *page)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100467{
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100468
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100469 int migratetype = get_pageblock_migratetype(page);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100470
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100471 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
472 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
Linus Torvalds68e3e922012-06-03 20:05:57 -0700473 return false;
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100474
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100475 /* If the page is a large free page, then allow migration */
476 if (PageBuddy(page) && page_order(page) >= pageblock_order)
Linus Torvalds68e3e922012-06-03 20:05:57 -0700477 return true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100478
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100479 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700480 if (migrate_async_suitable(migratetype))
481 return true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100482
483 /* Otherwise skip the block */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700484 return false;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100485}
486
487/*
Mel Gormande74f1c2012-08-21 16:16:15 -0700488 * Returns the start pfn of the last page block in a zone. This is the starting
489 * point for full compaction of a zone. Compaction searches for free pages from
490 * the end of each zone, while isolate_freepages_block scans forward inside each
491 * page block.
492 */
493static unsigned long start_free_pfn(struct zone *zone)
494{
495 unsigned long free_pfn;
496 free_pfn = zone->zone_start_pfn + zone->spanned_pages;
497 free_pfn &= ~(pageblock_nr_pages-1);
498 return free_pfn;
499}
500
501/*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100502 * Based on information in the current compact_control, find blocks
503 * suitable for isolating free pages from and then isolate them.
504 */
505static void isolate_freepages(struct zone *zone,
506 struct compact_control *cc)
507{
508 struct page *page;
509 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
510 unsigned long flags;
511 int nr_freepages = cc->nr_freepages;
512 struct list_head *freelist = &cc->freepages;
513
514 /*
515 * Initialise the free scanner. The starting point is where we last
516 * scanned from (or the end of the zone if starting). The low point
517 * is the end of the pageblock the migration scanner is using.
518 */
519 pfn = cc->free_pfn;
520 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
521
522 /*
523 * Take care that if the migration scanner is at the end of the zone
524 * that the free scanner does not accidentally move to the next zone
525 * in the next isolation cycle.
526 */
527 high_pfn = min(low_pfn, pfn);
528
529 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
530
531 /*
532 * Isolate free pages until enough are available to migrate the
533 * pages on cc->migratepages. We stop searching if the migrate
534 * and free page scanners meet or enough free pages are isolated.
535 */
536 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
537 pfn -= pageblock_nr_pages) {
538 unsigned long isolated;
539
540 if (!pfn_valid(pfn))
541 continue;
542
543 /*
544 * Check for overlapping nodes/zones. It's possible on some
545 * configurations to have a setup like
546 * node0 node1 node0
547 * i.e. it's possible that all pages within a zones range of
548 * pages do not belong to a single zone.
549 */
550 page = pfn_to_page(pfn);
551 if (page_zone(page) != zone)
552 continue;
553
554 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700555 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100556 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700557
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100558 /*
559 * Found a block suitable for isolating free pages from. Now
560 * we disabled interrupts, double check things are ok and
561 * isolate the pages. This is to minimise the time IRQs
562 * are disabled
563 */
564 isolated = 0;
Mel Gormanc67fe372012-08-21 16:16:17 -0700565
566 /*
567 * The zone lock must be held to isolate freepages. This
568 * unfortunately this is a very coarse lock and can be
569 * heavily contended if there are parallel allocations
570 * or parallel compactions. For async compaction do not
571 * spin on the lock
572 */
573 if (!compact_trylock_irqsave(&zone->lock, &flags, cc))
574 break;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700575 if (suitable_migration_target(page)) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100576 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
577 isolated = isolate_freepages_block(pfn, end_pfn,
578 freelist, false);
579 nr_freepages += isolated;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700580 }
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100581 spin_unlock_irqrestore(&zone->lock, flags);
582
583 /*
584 * Record the highest PFN we isolated pages from. When next
585 * looking for free pages, the search will restart here as
586 * page migration may have returned some pages to the allocator
587 */
Rik van Riel7db88892012-07-31 16:43:12 -0700588 if (isolated) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100589 high_pfn = max(high_pfn, pfn);
Mel Gormande74f1c2012-08-21 16:16:15 -0700590
591 /*
592 * If the free scanner has wrapped, update
593 * compact_cached_free_pfn to point to the highest
594 * pageblock with free pages. This reduces excessive
595 * scanning of full pageblocks near the end of the
596 * zone
597 */
598 if (cc->order > 0 && cc->wrapped)
Rik van Riel7db88892012-07-31 16:43:12 -0700599 zone->compact_cached_free_pfn = high_pfn;
600 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100601 }
602
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100603 /* split_free_page does not map the pages */
604 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100605
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100606 cc->free_pfn = high_pfn;
607 cc->nr_freepages = nr_freepages;
Mel Gormande74f1c2012-08-21 16:16:15 -0700608
609 /* If compact_cached_free_pfn is reset then set it now */
610 if (cc->order > 0 && !cc->wrapped &&
611 zone->compact_cached_free_pfn == start_free_pfn(zone))
612 zone->compact_cached_free_pfn = high_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700613}
614
615/*
616 * This is a migrate-callback that "allocates" freepages by taking pages
617 * from the isolated freelists in the block we are migrating to.
618 */
619static struct page *compaction_alloc(struct page *migratepage,
620 unsigned long data,
621 int **result)
622{
623 struct compact_control *cc = (struct compact_control *)data;
624 struct page *freepage;
625
626 /* Isolate free pages if necessary */
627 if (list_empty(&cc->freepages)) {
628 isolate_freepages(cc->zone, cc);
629
630 if (list_empty(&cc->freepages))
631 return NULL;
632 }
633
634 freepage = list_entry(cc->freepages.next, struct page, lru);
635 list_del(&freepage->lru);
636 cc->nr_freepages--;
637
638 return freepage;
639}
640
641/*
642 * We cannot control nr_migratepages and nr_freepages fully when migration is
643 * running as migrate_pages() has no knowledge of compact_control. When
644 * migration is complete, we count the number of pages on the lists by hand.
645 */
646static void update_nr_listpages(struct compact_control *cc)
647{
648 int nr_migratepages = 0;
649 int nr_freepages = 0;
650 struct page *page;
651
652 list_for_each_entry(page, &cc->migratepages, lru)
653 nr_migratepages++;
654 list_for_each_entry(page, &cc->freepages, lru)
655 nr_freepages++;
656
657 cc->nr_migratepages = nr_migratepages;
658 cc->nr_freepages = nr_freepages;
659}
660
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100661/* possible outcome of isolate_migratepages */
662typedef enum {
663 ISOLATE_ABORT, /* Abort compaction now */
664 ISOLATE_NONE, /* No pages isolated, continue scanning */
665 ISOLATE_SUCCESS, /* Pages isolated, migrate */
666} isolate_migrate_t;
667
668/*
669 * Isolate all pages that can be migrated from the block pointed to by
670 * the migrate scanner within compact_control.
671 */
672static isolate_migrate_t isolate_migratepages(struct zone *zone,
673 struct compact_control *cc)
674{
675 unsigned long low_pfn, end_pfn;
676
677 /* Do not scan outside zone boundaries */
678 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
679
680 /* Only scan within a pageblock boundary */
681 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
682
683 /* Do not cross the free scanner or scan within a memory hole */
684 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
685 cc->migrate_pfn = end_pfn;
686 return ISOLATE_NONE;
687 }
688
689 /* Perform the isolation */
690 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
691 if (!low_pfn)
692 return ISOLATE_ABORT;
693
694 cc->migrate_pfn = low_pfn;
695
696 return ISOLATE_SUCCESS;
697}
698
Mel Gorman748446b2010-05-24 14:32:27 -0700699static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800700 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700701{
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800702 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700703
Mel Gorman748446b2010-05-24 14:32:27 -0700704 if (fatal_signal_pending(current))
705 return COMPACT_PARTIAL;
706
Rik van Riel7db88892012-07-31 16:43:12 -0700707 /*
708 * A full (order == -1) compaction run starts at the beginning and
709 * end of a zone; it completes when the migrate and free scanner meet.
710 * A partial (order > 0) compaction can start with the free scanner
711 * at a random point in the zone, and may have to restart.
712 */
713 if (cc->free_pfn <= cc->migrate_pfn) {
714 if (cc->order > 0 && !cc->wrapped) {
715 /* We started partway through; restart at the end. */
716 unsigned long free_pfn = start_free_pfn(zone);
717 zone->compact_cached_free_pfn = free_pfn;
718 cc->free_pfn = free_pfn;
719 cc->wrapped = 1;
720 return COMPACT_CONTINUE;
721 }
722 return COMPACT_COMPLETE;
723 }
724
725 /* We wrapped around and ended up where we started. */
726 if (cc->wrapped && cc->free_pfn <= cc->start_free_pfn)
Mel Gorman748446b2010-05-24 14:32:27 -0700727 return COMPACT_COMPLETE;
728
Johannes Weiner82478fb2011-01-20 14:44:21 -0800729 /*
730 * order == -1 is expected when compacting via
731 * /proc/sys/vm/compact_memory
732 */
Mel Gorman56de7262010-05-24 14:32:30 -0700733 if (cc->order == -1)
734 return COMPACT_CONTINUE;
735
Michal Hocko3957c772011-06-15 15:08:25 -0700736 /* Compaction run is not finished if the watermark is not met */
737 watermark = low_wmark_pages(zone);
738 watermark += (1 << cc->order);
739
740 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
741 return COMPACT_CONTINUE;
742
Mel Gorman56de7262010-05-24 14:32:30 -0700743 /* Direct compactor: Is a suitable page free? */
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700744 if (cc->page) {
745 /* Was a suitable page captured? */
746 if (*cc->page)
Mel Gorman56de7262010-05-24 14:32:30 -0700747 return COMPACT_PARTIAL;
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700748 } else {
749 unsigned int order;
750 for (order = cc->order; order < MAX_ORDER; order++) {
751 struct free_area *area = &zone->free_area[cc->order];
752 /* Job done if page is free of the right migratetype */
753 if (!list_empty(&area->free_list[cc->migratetype]))
754 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -0700755
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700756 /* Job done if allocation would set block type */
757 if (cc->order >= pageblock_order && area->nr_free)
758 return COMPACT_PARTIAL;
759 }
Mel Gorman56de7262010-05-24 14:32:30 -0700760 }
761
Mel Gorman748446b2010-05-24 14:32:27 -0700762 return COMPACT_CONTINUE;
763}
764
Mel Gorman3e7d3442011-01-13 15:45:56 -0800765/*
766 * compaction_suitable: Is this suitable to run compaction on this zone now?
767 * Returns
768 * COMPACT_SKIPPED - If there are too few free pages for compaction
769 * COMPACT_PARTIAL - If the allocation would succeed without compaction
770 * COMPACT_CONTINUE - If compaction should run now
771 */
772unsigned long compaction_suitable(struct zone *zone, int order)
773{
774 int fragindex;
775 unsigned long watermark;
776
777 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700778 * order == -1 is expected when compacting via
779 * /proc/sys/vm/compact_memory
780 */
781 if (order == -1)
782 return COMPACT_CONTINUE;
783
784 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800785 * Watermarks for order-0 must be met for compaction. Note the 2UL.
786 * This is because during migration, copies of pages need to be
787 * allocated and for a short time, the footprint is higher
788 */
789 watermark = low_wmark_pages(zone) + (2UL << order);
790 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
791 return COMPACT_SKIPPED;
792
793 /*
794 * fragmentation index determines if allocation failures are due to
795 * low memory or external fragmentation
796 *
Shaohua Lia582a732011-06-15 15:08:49 -0700797 * index of -1000 implies allocations might succeed depending on
798 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800799 * index towards 0 implies failure is due to lack of memory
800 * index towards 1000 implies failure is due to fragmentation
801 *
802 * Only compact if a failure would be due to fragmentation.
803 */
804 fragindex = fragmentation_index(zone, order);
805 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
806 return COMPACT_SKIPPED;
807
Shaohua Lia582a732011-06-15 15:08:49 -0700808 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
809 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800810 return COMPACT_PARTIAL;
811
812 return COMPACT_CONTINUE;
813}
814
Mel Gorman748446b2010-05-24 14:32:27 -0700815static int compact_zone(struct zone *zone, struct compact_control *cc)
816{
817 int ret;
818
Mel Gorman3e7d3442011-01-13 15:45:56 -0800819 ret = compaction_suitable(zone, cc->order);
820 switch (ret) {
821 case COMPACT_PARTIAL:
822 case COMPACT_SKIPPED:
823 /* Compaction is likely to fail */
824 return ret;
825 case COMPACT_CONTINUE:
826 /* Fall through to compaction */
827 ;
828 }
829
Mel Gorman748446b2010-05-24 14:32:27 -0700830 /* Setup to move all movable pages to the end of the zone */
831 cc->migrate_pfn = zone->zone_start_pfn;
Rik van Riel7db88892012-07-31 16:43:12 -0700832
833 if (cc->order > 0) {
834 /* Incremental compaction. Start where the last one stopped. */
835 cc->free_pfn = zone->compact_cached_free_pfn;
836 cc->start_free_pfn = cc->free_pfn;
837 } else {
838 /* Order == -1 starts at the end of the zone. */
839 cc->free_pfn = start_free_pfn(zone);
840 }
Mel Gorman748446b2010-05-24 14:32:27 -0700841
842 migrate_prep_local();
843
844 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
845 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -0700846 int err;
Mel Gorman748446b2010-05-24 14:32:27 -0700847
Mel Gormanf9e35b32011-06-15 15:08:52 -0700848 switch (isolate_migratepages(zone, cc)) {
849 case ISOLATE_ABORT:
850 ret = COMPACT_PARTIAL;
851 goto out;
852 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -0700853 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700854 case ISOLATE_SUCCESS:
855 ;
856 }
Mel Gorman748446b2010-05-24 14:32:27 -0700857
858 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -0700859 err = migrate_pages(&cc->migratepages, compaction_alloc,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700860 (unsigned long)cc, false,
861 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
Mel Gorman748446b2010-05-24 14:32:27 -0700862 update_nr_listpages(cc);
863 nr_remaining = cc->nr_migratepages;
864
865 count_vm_event(COMPACTBLOCKS);
866 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
867 if (nr_remaining)
868 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
Mel Gormanb7aba692011-01-13 15:45:54 -0800869 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
870 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -0700871
872 /* Release LRU pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -0700873 if (err) {
Mel Gorman748446b2010-05-24 14:32:27 -0700874 putback_lru_pages(&cc->migratepages);
875 cc->nr_migratepages = 0;
David Rientjes4bf2bba2012-07-11 14:02:13 -0700876 if (err == -ENOMEM) {
877 ret = COMPACT_PARTIAL;
878 goto out;
879 }
Mel Gorman748446b2010-05-24 14:32:27 -0700880 }
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700881
882 /* Capture a page now if it is a suitable size */
883 compact_capture_page(cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700884 }
885
Mel Gormanf9e35b32011-06-15 15:08:52 -0700886out:
Mel Gorman748446b2010-05-24 14:32:27 -0700887 /* Release free pages and check accounting */
888 cc->nr_freepages -= release_freepages(&cc->freepages);
889 VM_BUG_ON(cc->nr_freepages != 0);
890
891 return ret;
892}
Mel Gorman76ab0f52010-05-24 14:32:28 -0700893
Kyungmin Parkd43a87e2011-10-31 17:09:08 -0700894static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800895 int order, gfp_t gfp_mask,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700896 bool sync, bool *contended,
897 struct page **page)
Mel Gorman56de7262010-05-24 14:32:30 -0700898{
899 struct compact_control cc = {
900 .nr_freepages = 0,
901 .nr_migratepages = 0,
902 .order = order,
903 .migratetype = allocflags_to_migratetype(gfp_mask),
904 .zone = zone,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700905 .sync = sync,
Mel Gormanc67fe372012-08-21 16:16:17 -0700906 .contended = contended,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700907 .page = page,
Mel Gorman56de7262010-05-24 14:32:30 -0700908 };
909 INIT_LIST_HEAD(&cc.freepages);
910 INIT_LIST_HEAD(&cc.migratepages);
911
Linus Torvalds68e3e922012-06-03 20:05:57 -0700912 return compact_zone(zone, &cc);
Mel Gorman56de7262010-05-24 14:32:30 -0700913}
914
Mel Gorman5e771902010-05-24 14:32:31 -0700915int sysctl_extfrag_threshold = 500;
916
Mel Gorman56de7262010-05-24 14:32:30 -0700917/**
918 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
919 * @zonelist: The zonelist used for the current allocation
920 * @order: The order of the current allocation
921 * @gfp_mask: The GFP mask of the current allocation
922 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -0800923 * @sync: Whether migration is synchronous or not
Mel Gorman56de7262010-05-24 14:32:30 -0700924 *
925 * This is the main entry point for direct page compaction.
926 */
927unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800928 int order, gfp_t gfp_mask, nodemask_t *nodemask,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700929 bool sync, bool *contended, struct page **page)
Mel Gorman56de7262010-05-24 14:32:30 -0700930{
931 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
932 int may_enter_fs = gfp_mask & __GFP_FS;
933 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -0700934 struct zoneref *z;
935 struct zone *zone;
936 int rc = COMPACT_SKIPPED;
937
Mel Gorman4ffb6332012-10-08 16:29:09 -0700938 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -0800939 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -0700940 return rc;
941
942 count_vm_event(COMPACTSTALL);
943
944 /* Compact each zone in the list */
945 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
946 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -0700947 int status;
948
Mel Gormanc67fe372012-08-21 16:16:17 -0700949 status = compact_zone_order(zone, order, gfp_mask, sync,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -0700950 contended, page);
Mel Gorman56de7262010-05-24 14:32:30 -0700951 rc = max(status, rc);
952
Mel Gorman3e7d3442011-01-13 15:45:56 -0800953 /* If a normal allocation would succeed, stop compacting */
954 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Mel Gorman56de7262010-05-24 14:32:30 -0700955 break;
956 }
957
958 return rc;
959}
960
961
Mel Gorman76ab0f52010-05-24 14:32:28 -0700962/* Compact all zones within a node */
Rik van Riel7be62de2012-03-21 16:33:52 -0700963static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -0700964{
965 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -0700966 struct zone *zone;
967
Mel Gorman76ab0f52010-05-24 14:32:28 -0700968 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -0700969
970 zone = &pgdat->node_zones[zoneid];
971 if (!populated_zone(zone))
972 continue;
973
Rik van Riel7be62de2012-03-21 16:33:52 -0700974 cc->nr_freepages = 0;
975 cc->nr_migratepages = 0;
976 cc->zone = zone;
977 INIT_LIST_HEAD(&cc->freepages);
978 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -0700979
Dan Carpenteraad6ec32012-03-21 16:33:54 -0700980 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -0700981 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -0700982
Rik van Rielaff62242012-03-21 16:33:52 -0700983 if (cc->order > 0) {
984 int ok = zone_watermark_ok(zone, cc->order,
985 low_wmark_pages(zone), 0, 0);
Minchan Kimc81758f2012-08-21 16:16:03 -0700986 if (ok && cc->order >= zone->compact_order_failed)
Rik van Rielaff62242012-03-21 16:33:52 -0700987 zone->compact_order_failed = cc->order + 1;
988 /* Currently async compaction is never deferred. */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700989 else if (!ok && cc->sync)
Rik van Rielaff62242012-03-21 16:33:52 -0700990 defer_compaction(zone, cc->order);
991 }
992
Rik van Riel7be62de2012-03-21 16:33:52 -0700993 VM_BUG_ON(!list_empty(&cc->freepages));
994 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -0700995 }
996
997 return 0;
998}
999
Rik van Riel7be62de2012-03-21 16:33:52 -07001000int compact_pgdat(pg_data_t *pgdat, int order)
1001{
1002 struct compact_control cc = {
1003 .order = order,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001004 .sync = false,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001005 .page = NULL,
Rik van Riel7be62de2012-03-21 16:33:52 -07001006 };
1007
1008 return __compact_pgdat(pgdat, &cc);
1009}
1010
1011static int compact_node(int nid)
1012{
Rik van Riel7be62de2012-03-21 16:33:52 -07001013 struct compact_control cc = {
1014 .order = -1,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001015 .sync = true,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001016 .page = NULL,
Rik van Riel7be62de2012-03-21 16:33:52 -07001017 };
1018
Hugh Dickins8575ec22012-03-21 16:33:53 -07001019 return __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001020}
1021
Mel Gorman76ab0f52010-05-24 14:32:28 -07001022/* Compact all nodes in the system */
1023static int compact_nodes(void)
1024{
1025 int nid;
1026
Hugh Dickins8575ec22012-03-21 16:33:53 -07001027 /* Flush pending updates to the LRU lists */
1028 lru_add_drain_all();
1029
Mel Gorman76ab0f52010-05-24 14:32:28 -07001030 for_each_online_node(nid)
1031 compact_node(nid);
1032
1033 return COMPACT_COMPLETE;
1034}
1035
1036/* The written value is actually unused, all memory is compacted */
1037int sysctl_compact_memory;
1038
1039/* This is the entry point for compacting all nodes via /proc/sys/vm */
1040int sysctl_compaction_handler(struct ctl_table *table, int write,
1041 void __user *buffer, size_t *length, loff_t *ppos)
1042{
1043 if (write)
1044 return compact_nodes();
1045
1046 return 0;
1047}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001048
Mel Gorman5e771902010-05-24 14:32:31 -07001049int sysctl_extfrag_handler(struct ctl_table *table, int write,
1050 void __user *buffer, size_t *length, loff_t *ppos)
1051{
1052 proc_dointvec_minmax(table, write, buffer, length, ppos);
1053
1054 return 0;
1055}
1056
Mel Gormaned4a6d72010-05-24 14:32:29 -07001057#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Kay Sievers10fbcf42011-12-21 14:48:43 -08001058ssize_t sysfs_compact_node(struct device *dev,
1059 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001060 const char *buf, size_t count)
1061{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001062 int nid = dev->id;
1063
1064 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1065 /* Flush pending updates to the LRU lists */
1066 lru_add_drain_all();
1067
1068 compact_node(nid);
1069 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001070
1071 return count;
1072}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001073static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001074
1075int compaction_register_node(struct node *node)
1076{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001077 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001078}
1079
1080void compaction_unregister_node(struct node *node)
1081{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001082 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001083}
1084#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001085
1086#endif /* CONFIG_COMPACTION */