Loading...
1/*
2 * linux/mm/compaction.c
3 *
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
6 * lifting
7 *
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9 */
10#include <linux/swap.h>
11#include <linux/migrate.h>
12#include <linux/compaction.h>
13#include <linux/mm_inline.h>
14#include <linux/backing-dev.h>
15#include <linux/sysctl.h>
16#include <linux/sysfs.h>
17#include "internal.h"
18
19#define CREATE_TRACE_POINTS
20#include <trace/events/compaction.h>
21
22/*
23 * compact_control is used to track pages being migrated and the free pages
24 * they are being migrated to during memory compaction. The free_pfn starts
25 * at the end of a zone and migrate_pfn begins at the start. Movable pages
26 * are moved to the end of a zone during a compaction run and the run
27 * completes when free_pfn <= migrate_pfn
28 */
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 */
36 bool sync; /* Synchronous migration */
37
38 /* Account for isolated anon and file pages */
39 unsigned long nr_anon;
40 unsigned long nr_file;
41
42 unsigned int order; /* order a direct compactor needs */
43 int migratetype; /* MOVABLE, RECLAIMABLE etc */
44 struct zone *zone;
45};
46
47static unsigned long release_freepages(struct list_head *freelist)
48{
49 struct page *page, *next;
50 unsigned long count = 0;
51
52 list_for_each_entry_safe(page, next, freelist, lru) {
53 list_del(&page->lru);
54 __free_page(page);
55 count++;
56 }
57
58 return count;
59}
60
61/* Isolate free pages onto a private freelist. Must hold zone->lock */
62static unsigned long isolate_freepages_block(struct zone *zone,
63 unsigned long blockpfn,
64 struct list_head *freelist)
65{
66 unsigned long zone_end_pfn, end_pfn;
67 int nr_scanned = 0, total_isolated = 0;
68 struct page *cursor;
69
70 /* Get the last PFN we should scan for free pages at */
71 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
72 end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
73
74 /* Find the first usable PFN in the block to initialse page cursor */
75 for (; blockpfn < end_pfn; blockpfn++) {
76 if (pfn_valid_within(blockpfn))
77 break;
78 }
79 cursor = pfn_to_page(blockpfn);
80
81 /* Isolate free pages. This assumes the block is valid */
82 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
83 int isolated, i;
84 struct page *page = cursor;
85
86 if (!pfn_valid_within(blockpfn))
87 continue;
88 nr_scanned++;
89
90 if (!PageBuddy(page))
91 continue;
92
93 /* Found a free page, break it into order-0 pages */
94 isolated = split_free_page(page);
95 total_isolated += isolated;
96 for (i = 0; i < isolated; i++) {
97 list_add(&page->lru, freelist);
98 page++;
99 }
100
101 /* If a page was split, advance to the end of it */
102 if (isolated) {
103 blockpfn += isolated - 1;
104 cursor += isolated - 1;
105 }
106 }
107
108 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
109 return total_isolated;
110}
111
112/* Returns true if the page is within a block suitable for migration to */
113static bool suitable_migration_target(struct page *page)
114{
115
116 int migratetype = get_pageblock_migratetype(page);
117
118 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
119 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
120 return false;
121
122 /* If the page is a large free page, then allow migration */
123 if (PageBuddy(page) && page_order(page) >= pageblock_order)
124 return true;
125
126 /* If the block is MIGRATE_MOVABLE, allow migration */
127 if (migratetype == MIGRATE_MOVABLE)
128 return true;
129
130 /* Otherwise skip the block */
131 return false;
132}
133
134/*
135 * Based on information in the current compact_control, find blocks
136 * suitable for isolating free pages from and then isolate them.
137 */
138static void isolate_freepages(struct zone *zone,
139 struct compact_control *cc)
140{
141 struct page *page;
142 unsigned long high_pfn, low_pfn, pfn;
143 unsigned long flags;
144 int nr_freepages = cc->nr_freepages;
145 struct list_head *freelist = &cc->freepages;
146
147 /*
148 * Initialise the free scanner. The starting point is where we last
149 * scanned from (or the end of the zone if starting). The low point
150 * is the end of the pageblock the migration scanner is using.
151 */
152 pfn = cc->free_pfn;
153 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
154
155 /*
156 * Take care that if the migration scanner is at the end of the zone
157 * that the free scanner does not accidentally move to the next zone
158 * in the next isolation cycle.
159 */
160 high_pfn = min(low_pfn, pfn);
161
162 /*
163 * Isolate free pages until enough are available to migrate the
164 * pages on cc->migratepages. We stop searching if the migrate
165 * and free page scanners meet or enough free pages are isolated.
166 */
167 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
168 pfn -= pageblock_nr_pages) {
169 unsigned long isolated;
170
171 if (!pfn_valid(pfn))
172 continue;
173
174 /*
175 * Check for overlapping nodes/zones. It's possible on some
176 * configurations to have a setup like
177 * node0 node1 node0
178 * i.e. it's possible that all pages within a zones range of
179 * pages do not belong to a single zone.
180 */
181 page = pfn_to_page(pfn);
182 if (page_zone(page) != zone)
183 continue;
184
185 /* Check the block is suitable for migration */
186 if (!suitable_migration_target(page))
187 continue;
188
189 /*
190 * Found a block suitable for isolating free pages from. Now
191 * we disabled interrupts, double check things are ok and
192 * isolate the pages. This is to minimise the time IRQs
193 * are disabled
194 */
195 isolated = 0;
196 spin_lock_irqsave(&zone->lock, flags);
197 if (suitable_migration_target(page)) {
198 isolated = isolate_freepages_block(zone, pfn, freelist);
199 nr_freepages += isolated;
200 }
201 spin_unlock_irqrestore(&zone->lock, flags);
202
203 /*
204 * Record the highest PFN we isolated pages from. When next
205 * looking for free pages, the search will restart here as
206 * page migration may have returned some pages to the allocator
207 */
208 if (isolated)
209 high_pfn = max(high_pfn, pfn);
210 }
211
212 /* split_free_page does not map the pages */
213 list_for_each_entry(page, freelist, lru) {
214 arch_alloc_page(page, 0);
215 kernel_map_pages(page, 1, 1);
216 }
217
218 cc->free_pfn = high_pfn;
219 cc->nr_freepages = nr_freepages;
220}
221
222/* Update the number of anon and file isolated pages in the zone */
223static void acct_isolated(struct zone *zone, struct compact_control *cc)
224{
225 struct page *page;
226 unsigned int count[NR_LRU_LISTS] = { 0, };
227
228 list_for_each_entry(page, &cc->migratepages, lru) {
229 int lru = page_lru_base_type(page);
230 count[lru]++;
231 }
232
233 cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
234 cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
235 __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
236 __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
237}
238
239/* Similar to reclaim, but different enough that they don't share logic */
240static bool too_many_isolated(struct zone *zone)
241{
242 unsigned long active, inactive, isolated;
243
244 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
245 zone_page_state(zone, NR_INACTIVE_ANON);
246 active = zone_page_state(zone, NR_ACTIVE_FILE) +
247 zone_page_state(zone, NR_ACTIVE_ANON);
248 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
249 zone_page_state(zone, NR_ISOLATED_ANON);
250
251 return isolated > (inactive + active) / 2;
252}
253
254/* possible outcome of isolate_migratepages */
255typedef enum {
256 ISOLATE_ABORT, /* Abort compaction now */
257 ISOLATE_NONE, /* No pages isolated, continue scanning */
258 ISOLATE_SUCCESS, /* Pages isolated, migrate */
259} isolate_migrate_t;
260
261/*
262 * Isolate all pages that can be migrated from the block pointed to by
263 * the migrate scanner within compact_control.
264 */
265static isolate_migrate_t isolate_migratepages(struct zone *zone,
266 struct compact_control *cc)
267{
268 unsigned long low_pfn, end_pfn;
269 unsigned long last_pageblock_nr = 0, pageblock_nr;
270 unsigned long nr_scanned = 0, nr_isolated = 0;
271 struct list_head *migratelist = &cc->migratepages;
272
273 /* Do not scan outside zone boundaries */
274 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
275
276 /* Only scan within a pageblock boundary */
277 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
278
279 /* Do not cross the free scanner or scan within a memory hole */
280 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
281 cc->migrate_pfn = end_pfn;
282 return ISOLATE_NONE;
283 }
284
285 /*
286 * Ensure that there are not too many pages isolated from the LRU
287 * list by either parallel reclaimers or compaction. If there are,
288 * delay for some time until fewer pages are isolated
289 */
290 while (unlikely(too_many_isolated(zone))) {
291 /* async migration should just abort */
292 if (!cc->sync)
293 return ISOLATE_ABORT;
294
295 congestion_wait(BLK_RW_ASYNC, HZ/10);
296
297 if (fatal_signal_pending(current))
298 return ISOLATE_ABORT;
299 }
300
301 /* Time to isolate some pages for migration */
302 cond_resched();
303 spin_lock_irq(&zone->lru_lock);
304 for (; low_pfn < end_pfn; low_pfn++) {
305 struct page *page;
306 bool locked = true;
307
308 /* give a chance to irqs before checking need_resched() */
309 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
310 spin_unlock_irq(&zone->lru_lock);
311 locked = false;
312 }
313 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
314 if (locked)
315 spin_unlock_irq(&zone->lru_lock);
316 cond_resched();
317 spin_lock_irq(&zone->lru_lock);
318 if (fatal_signal_pending(current))
319 break;
320 } else if (!locked)
321 spin_lock_irq(&zone->lru_lock);
322
323 if (!pfn_valid_within(low_pfn))
324 continue;
325 nr_scanned++;
326
327 /* Get the page and skip if free */
328 page = pfn_to_page(low_pfn);
329 if (PageBuddy(page))
330 continue;
331
332 /*
333 * For async migration, also only scan in MOVABLE blocks. Async
334 * migration is optimistic to see if the minimum amount of work
335 * satisfies the allocation
336 */
337 pageblock_nr = low_pfn >> pageblock_order;
338 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
339 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
340 low_pfn += pageblock_nr_pages;
341 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
342 last_pageblock_nr = pageblock_nr;
343 continue;
344 }
345
346 if (!PageLRU(page))
347 continue;
348
349 /*
350 * PageLRU is set, and lru_lock excludes isolation,
351 * splitting and collapsing (collapsing has already
352 * happened if PageLRU is set).
353 */
354 if (PageTransHuge(page)) {
355 low_pfn += (1 << compound_order(page)) - 1;
356 continue;
357 }
358
359 /* Try isolate the page */
360 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
361 continue;
362
363 VM_BUG_ON(PageTransCompound(page));
364
365 /* Successfully isolated */
366 del_page_from_lru_list(zone, page, page_lru(page));
367 list_add(&page->lru, migratelist);
368 cc->nr_migratepages++;
369 nr_isolated++;
370
371 /* Avoid isolating too much */
372 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
373 break;
374 }
375
376 acct_isolated(zone, cc);
377
378 spin_unlock_irq(&zone->lru_lock);
379 cc->migrate_pfn = low_pfn;
380
381 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
382
383 return ISOLATE_SUCCESS;
384}
385
386/*
387 * This is a migrate-callback that "allocates" freepages by taking pages
388 * from the isolated freelists in the block we are migrating to.
389 */
390static struct page *compaction_alloc(struct page *migratepage,
391 unsigned long data,
392 int **result)
393{
394 struct compact_control *cc = (struct compact_control *)data;
395 struct page *freepage;
396
397 /* Isolate free pages if necessary */
398 if (list_empty(&cc->freepages)) {
399 isolate_freepages(cc->zone, cc);
400
401 if (list_empty(&cc->freepages))
402 return NULL;
403 }
404
405 freepage = list_entry(cc->freepages.next, struct page, lru);
406 list_del(&freepage->lru);
407 cc->nr_freepages--;
408
409 return freepage;
410}
411
412/*
413 * We cannot control nr_migratepages and nr_freepages fully when migration is
414 * running as migrate_pages() has no knowledge of compact_control. When
415 * migration is complete, we count the number of pages on the lists by hand.
416 */
417static void update_nr_listpages(struct compact_control *cc)
418{
419 int nr_migratepages = 0;
420 int nr_freepages = 0;
421 struct page *page;
422
423 list_for_each_entry(page, &cc->migratepages, lru)
424 nr_migratepages++;
425 list_for_each_entry(page, &cc->freepages, lru)
426 nr_freepages++;
427
428 cc->nr_migratepages = nr_migratepages;
429 cc->nr_freepages = nr_freepages;
430}
431
432static int compact_finished(struct zone *zone,
433 struct compact_control *cc)
434{
435 unsigned int order;
436 unsigned long watermark;
437
438 if (fatal_signal_pending(current))
439 return COMPACT_PARTIAL;
440
441 /* Compaction run completes if the migrate and free scanner meet */
442 if (cc->free_pfn <= cc->migrate_pfn)
443 return COMPACT_COMPLETE;
444
445 /*
446 * order == -1 is expected when compacting via
447 * /proc/sys/vm/compact_memory
448 */
449 if (cc->order == -1)
450 return COMPACT_CONTINUE;
451
452 /* Compaction run is not finished if the watermark is not met */
453 watermark = low_wmark_pages(zone);
454 watermark += (1 << cc->order);
455
456 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
457 return COMPACT_CONTINUE;
458
459 /* Direct compactor: Is a suitable page free? */
460 for (order = cc->order; order < MAX_ORDER; order++) {
461 /* Job done if page is free of the right migratetype */
462 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
463 return COMPACT_PARTIAL;
464
465 /* Job done if allocation would set block type */
466 if (order >= pageblock_order && zone->free_area[order].nr_free)
467 return COMPACT_PARTIAL;
468 }
469
470 return COMPACT_CONTINUE;
471}
472
473/*
474 * compaction_suitable: Is this suitable to run compaction on this zone now?
475 * Returns
476 * COMPACT_SKIPPED - If there are too few free pages for compaction
477 * COMPACT_PARTIAL - If the allocation would succeed without compaction
478 * COMPACT_CONTINUE - If compaction should run now
479 */
480unsigned long compaction_suitable(struct zone *zone, int order)
481{
482 int fragindex;
483 unsigned long watermark;
484
485 /*
486 * order == -1 is expected when compacting via
487 * /proc/sys/vm/compact_memory
488 */
489 if (order == -1)
490 return COMPACT_CONTINUE;
491
492 /*
493 * Watermarks for order-0 must be met for compaction. Note the 2UL.
494 * This is because during migration, copies of pages need to be
495 * allocated and for a short time, the footprint is higher
496 */
497 watermark = low_wmark_pages(zone) + (2UL << order);
498 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
499 return COMPACT_SKIPPED;
500
501 /*
502 * fragmentation index determines if allocation failures are due to
503 * low memory or external fragmentation
504 *
505 * index of -1000 implies allocations might succeed depending on
506 * watermarks
507 * index towards 0 implies failure is due to lack of memory
508 * index towards 1000 implies failure is due to fragmentation
509 *
510 * Only compact if a failure would be due to fragmentation.
511 */
512 fragindex = fragmentation_index(zone, order);
513 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
514 return COMPACT_SKIPPED;
515
516 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
517 0, 0))
518 return COMPACT_PARTIAL;
519
520 return COMPACT_CONTINUE;
521}
522
523static int compact_zone(struct zone *zone, struct compact_control *cc)
524{
525 int ret;
526
527 ret = compaction_suitable(zone, cc->order);
528 switch (ret) {
529 case COMPACT_PARTIAL:
530 case COMPACT_SKIPPED:
531 /* Compaction is likely to fail */
532 return ret;
533 case COMPACT_CONTINUE:
534 /* Fall through to compaction */
535 ;
536 }
537
538 /* Setup to move all movable pages to the end of the zone */
539 cc->migrate_pfn = zone->zone_start_pfn;
540 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
541 cc->free_pfn &= ~(pageblock_nr_pages-1);
542
543 migrate_prep_local();
544
545 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
546 unsigned long nr_migrate, nr_remaining;
547 int err;
548
549 switch (isolate_migratepages(zone, cc)) {
550 case ISOLATE_ABORT:
551 ret = COMPACT_PARTIAL;
552 goto out;
553 case ISOLATE_NONE:
554 continue;
555 case ISOLATE_SUCCESS:
556 ;
557 }
558
559 nr_migrate = cc->nr_migratepages;
560 err = migrate_pages(&cc->migratepages, compaction_alloc,
561 (unsigned long)cc, false,
562 cc->sync);
563 update_nr_listpages(cc);
564 nr_remaining = cc->nr_migratepages;
565
566 count_vm_event(COMPACTBLOCKS);
567 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
568 if (nr_remaining)
569 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
570 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
571 nr_remaining);
572
573 /* Release LRU pages not migrated */
574 if (err) {
575 putback_lru_pages(&cc->migratepages);
576 cc->nr_migratepages = 0;
577 }
578
579 }
580
581out:
582 /* Release free pages and check accounting */
583 cc->nr_freepages -= release_freepages(&cc->freepages);
584 VM_BUG_ON(cc->nr_freepages != 0);
585
586 return ret;
587}
588
589unsigned long compact_zone_order(struct zone *zone,
590 int order, gfp_t gfp_mask,
591 bool sync)
592{
593 struct compact_control cc = {
594 .nr_freepages = 0,
595 .nr_migratepages = 0,
596 .order = order,
597 .migratetype = allocflags_to_migratetype(gfp_mask),
598 .zone = zone,
599 .sync = sync,
600 };
601 INIT_LIST_HEAD(&cc.freepages);
602 INIT_LIST_HEAD(&cc.migratepages);
603
604 return compact_zone(zone, &cc);
605}
606
607int sysctl_extfrag_threshold = 500;
608
609/**
610 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
611 * @zonelist: The zonelist used for the current allocation
612 * @order: The order of the current allocation
613 * @gfp_mask: The GFP mask of the current allocation
614 * @nodemask: The allowed nodes to allocate from
615 * @sync: Whether migration is synchronous or not
616 *
617 * This is the main entry point for direct page compaction.
618 */
619unsigned long try_to_compact_pages(struct zonelist *zonelist,
620 int order, gfp_t gfp_mask, nodemask_t *nodemask,
621 bool sync)
622{
623 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
624 int may_enter_fs = gfp_mask & __GFP_FS;
625 int may_perform_io = gfp_mask & __GFP_IO;
626 struct zoneref *z;
627 struct zone *zone;
628 int rc = COMPACT_SKIPPED;
629
630 /*
631 * Check whether it is worth even starting compaction. The order check is
632 * made because an assumption is made that the page allocator can satisfy
633 * the "cheaper" orders without taking special steps
634 */
635 if (!order || !may_enter_fs || !may_perform_io)
636 return rc;
637
638 count_vm_event(COMPACTSTALL);
639
640 /* Compact each zone in the list */
641 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
642 nodemask) {
643 int status;
644
645 status = compact_zone_order(zone, order, gfp_mask, sync);
646 rc = max(status, rc);
647
648 /* If a normal allocation would succeed, stop compacting */
649 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
650 break;
651 }
652
653 return rc;
654}
655
656
657/* Compact all zones within a node */
658static int compact_node(int nid)
659{
660 int zoneid;
661 pg_data_t *pgdat;
662 struct zone *zone;
663
664 if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
665 return -EINVAL;
666 pgdat = NODE_DATA(nid);
667
668 /* Flush pending updates to the LRU lists */
669 lru_add_drain_all();
670
671 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
672 struct compact_control cc = {
673 .nr_freepages = 0,
674 .nr_migratepages = 0,
675 .order = -1,
676 };
677
678 zone = &pgdat->node_zones[zoneid];
679 if (!populated_zone(zone))
680 continue;
681
682 cc.zone = zone;
683 INIT_LIST_HEAD(&cc.freepages);
684 INIT_LIST_HEAD(&cc.migratepages);
685
686 compact_zone(zone, &cc);
687
688 VM_BUG_ON(!list_empty(&cc.freepages));
689 VM_BUG_ON(!list_empty(&cc.migratepages));
690 }
691
692 return 0;
693}
694
695/* Compact all nodes in the system */
696static int compact_nodes(void)
697{
698 int nid;
699
700 for_each_online_node(nid)
701 compact_node(nid);
702
703 return COMPACT_COMPLETE;
704}
705
706/* The written value is actually unused, all memory is compacted */
707int sysctl_compact_memory;
708
709/* This is the entry point for compacting all nodes via /proc/sys/vm */
710int sysctl_compaction_handler(struct ctl_table *table, int write,
711 void __user *buffer, size_t *length, loff_t *ppos)
712{
713 if (write)
714 return compact_nodes();
715
716 return 0;
717}
718
719int sysctl_extfrag_handler(struct ctl_table *table, int write,
720 void __user *buffer, size_t *length, loff_t *ppos)
721{
722 proc_dointvec_minmax(table, write, buffer, length, ppos);
723
724 return 0;
725}
726
727#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
728ssize_t sysfs_compact_node(struct sys_device *dev,
729 struct sysdev_attribute *attr,
730 const char *buf, size_t count)
731{
732 compact_node(dev->id);
733
734 return count;
735}
736static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
737
738int compaction_register_node(struct node *node)
739{
740 return sysdev_create_file(&node->sysdev, &attr_compact);
741}
742
743void compaction_unregister_node(struct node *node)
744{
745 return sysdev_remove_file(&node->sysdev, &attr_compact);
746}
747#endif /* CONFIG_SYSFS && CONFIG_NUMA */
1/*
2 * linux/mm/compaction.c
3 *
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
6 * lifting
7 *
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9 */
10#include <linux/swap.h>
11#include <linux/migrate.h>
12#include <linux/compaction.h>
13#include <linux/mm_inline.h>
14#include <linux/backing-dev.h>
15#include <linux/sysctl.h>
16#include <linux/sysfs.h>
17#include "internal.h"
18
19#if defined CONFIG_COMPACTION || defined CONFIG_CMA
20
21#define CREATE_TRACE_POINTS
22#include <trace/events/compaction.h>
23
24static 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
38static 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
48static inline bool migrate_async_suitable(int migratetype)
49{
50 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
51}
52
53/*
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)
63{
64 int nr_scanned = 0, total_isolated = 0;
65 struct page *cursor;
66
67 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
74 if (!pfn_valid_within(blockpfn)) {
75 if (strict)
76 return 0;
77 continue;
78 }
79 nr_scanned++;
80
81 if (!PageBuddy(page)) {
82 if (strict)
83 return 0;
84 continue;
85 }
86
87 /* Found a free page, break it into order-0 pages */
88 isolated = split_free_page(page);
89 if (!isolated && strict)
90 return 0;
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
104 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
105 return total_isolated;
106}
107
108/**
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 */
121unsigned long
122isolate_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
175/* 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;
179 unsigned int count[2] = { 0, };
180
181 list_for_each_entry(page, &cc->migratepages, lru)
182 count[!!page_is_file_cache(page)]++;
183
184 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
185 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
186}
187
188/* Similar to reclaim, but different enough that they don't share logic */
189static bool too_many_isolated(struct zone *zone)
190{
191 unsigned long active, inactive, isolated;
192
193 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
194 zone_page_state(zone, NR_INACTIVE_ANON);
195 active = zone_page_state(zone, NR_ACTIVE_FILE) +
196 zone_page_state(zone, NR_ACTIVE_ANON);
197 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
198 zone_page_state(zone, NR_ISOLATED_ANON);
199
200 return isolated > (inactive + active) / 2;
201}
202
203/**
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.
221 */
222unsigned long
223isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
224 unsigned long low_pfn, unsigned long end_pfn)
225{
226 unsigned long last_pageblock_nr = 0, pageblock_nr;
227 unsigned long nr_scanned = 0, nr_isolated = 0;
228 struct list_head *migratelist = &cc->migratepages;
229 isolate_mode_t mode = 0;
230 struct lruvec *lruvec;
231
232 /*
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))) {
238 /* async migration should just abort */
239 if (!cc->sync)
240 return 0;
241
242 congestion_wait(BLK_RW_ASYNC, HZ/10);
243
244 if (fatal_signal_pending(current))
245 return 0;
246 }
247
248 /* Time to isolate some pages for migration */
249 cond_resched();
250 spin_lock_irq(&zone->lru_lock);
251 for (; low_pfn < end_pfn; low_pfn++) {
252 struct page *page;
253 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
270 /*
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
283 if (!pfn_valid_within(low_pfn))
284 continue;
285 nr_scanned++;
286
287 /*
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 */
293 page = pfn_to_page(low_pfn);
294 if (page_zone(page) != zone)
295 continue;
296
297 /* Skip if free */
298 if (PageBuddy(page))
299 continue;
300
301 /*
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;
307 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
308 !migrate_async_suitable(get_pageblock_migratetype(page))) {
309 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
315 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
328 if (!cc->sync)
329 mode |= ISOLATE_ASYNC_MIGRATE;
330
331 lruvec = mem_cgroup_page_lruvec(page, zone);
332
333 /* Try isolate the page */
334 if (__isolate_lru_page(page, mode) != 0)
335 continue;
336
337 VM_BUG_ON(PageTransCompound(page));
338
339 /* Successfully isolated */
340 del_page_from_lru_list(page, lruvec, page_lru(page));
341 list_add(&page->lru, migratelist);
342 cc->nr_migratepages++;
343 nr_isolated++;
344
345 /* Avoid isolating too much */
346 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
347 ++low_pfn;
348 break;
349 }
350 }
351
352 acct_isolated(zone, cc);
353
354 spin_unlock_irq(&zone->lru_lock);
355
356 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
357
358 return low_pfn;
359}
360
361#endif /* CONFIG_COMPACTION || CONFIG_CMA */
362#ifdef CONFIG_COMPACTION
363
364/* Returns true if the page is within a block suitable for migration to */
365static bool suitable_migration_target(struct page *page)
366{
367
368 int migratetype = get_pageblock_migratetype(page);
369
370 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
371 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
372 return false;
373
374 /* If the page is a large free page, then allow migration */
375 if (PageBuddy(page) && page_order(page) >= pageblock_order)
376 return true;
377
378 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
379 if (migrate_async_suitable(migratetype))
380 return true;
381
382 /* Otherwise skip the block */
383 return false;
384}
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 */
440 if (!suitable_migration_target(page))
441 continue;
442
443 /*
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);
451 if (suitable_migration_target(page)) {
452 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;
456 }
457 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);
466 }
467
468 /* split_free_page does not map the pages */
469 map_pages(freelist);
470
471 cc->free_pfn = high_pfn;
472 cc->nr_freepages = nr_freepages;
473}
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
521/* 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
559static int compact_finished(struct zone *zone,
560 struct compact_control *cc)
561{
562 unsigned int order;
563 unsigned long watermark;
564
565 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
572 /*
573 * order == -1 is expected when compacting via
574 * /proc/sys/vm/compact_memory
575 */
576 if (cc->order == -1)
577 return COMPACT_CONTINUE;
578
579 /* 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
586 /* 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
597 return COMPACT_CONTINUE;
598}
599
600/*
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 /*
613 * order == -1 is expected when compacting via
614 * /proc/sys/vm/compact_memory
615 */
616 if (order == -1)
617 return COMPACT_CONTINUE;
618
619 /*
620 * 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 *
632 * index of -1000 implies allocations might succeed depending on
633 * watermarks
634 * 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
643 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
644 0, 0))
645 return COMPACT_PARTIAL;
646
647 return COMPACT_CONTINUE;
648}
649
650static int compact_zone(struct zone *zone, struct compact_control *cc)
651{
652 int ret;
653
654 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
665 /* 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;
674 int err;
675
676 switch (isolate_migratepages(zone, cc)) {
677 case ISOLATE_ABORT:
678 ret = COMPACT_PARTIAL;
679 goto out;
680 case ISOLATE_NONE:
681 continue;
682 case ISOLATE_SUCCESS:
683 ;
684 }
685
686 nr_migrate = cc->nr_migratepages;
687 err = migrate_pages(&cc->migratepages, compaction_alloc,
688 (unsigned long)cc, false,
689 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
690 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);
697 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
698 nr_remaining);
699
700 /* Release LRU pages not migrated */
701 if (err) {
702 putback_lru_pages(&cc->migratepages);
703 cc->nr_migratepages = 0;
704 if (err == -ENOMEM) {
705 ret = COMPACT_PARTIAL;
706 goto out;
707 }
708 }
709 }
710
711out:
712 /* Release free pages and check accounting */
713 cc->nr_freepages -= release_freepages(&cc->freepages);
714 VM_BUG_ON(cc->nr_freepages != 0);
715
716 return ret;
717}
718
719static unsigned long compact_zone_order(struct zone *zone,
720 int order, gfp_t gfp_mask,
721 bool sync)
722{
723 struct compact_control cc = {
724 .nr_freepages = 0,
725 .nr_migratepages = 0,
726 .order = order,
727 .migratetype = allocflags_to_migratetype(gfp_mask),
728 .zone = zone,
729 .sync = sync,
730 };
731 INIT_LIST_HEAD(&cc.freepages);
732 INIT_LIST_HEAD(&cc.migratepages);
733
734 return compact_zone(zone, &cc);
735}
736
737int sysctl_extfrag_threshold = 500;
738
739/**
740 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
741 * @zonelist: The zonelist used for the current allocation
742 * @order: The order of the current allocation
743 * @gfp_mask: The GFP mask of the current allocation
744 * @nodemask: The allowed nodes to allocate from
745 * @sync: Whether migration is synchronous or not
746 *
747 * This is the main entry point for direct page compaction.
748 */
749unsigned long try_to_compact_pages(struct zonelist *zonelist,
750 int order, gfp_t gfp_mask, nodemask_t *nodemask,
751 bool sync)
752{
753 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
754 int may_enter_fs = gfp_mask & __GFP_FS;
755 int may_perform_io = gfp_mask & __GFP_IO;
756 struct zoneref *z;
757 struct zone *zone;
758 int rc = COMPACT_SKIPPED;
759
760 /*
761 * Check whether it is worth even starting compaction. The order check is
762 * made because an assumption is made that the page allocator can satisfy
763 * the "cheaper" orders without taking special steps
764 */
765 if (!order || !may_enter_fs || !may_perform_io)
766 return rc;
767
768 count_vm_event(COMPACTSTALL);
769
770 /* Compact each zone in the list */
771 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
772 nodemask) {
773 int status;
774
775 status = compact_zone_order(zone, order, gfp_mask, sync);
776 rc = max(status, rc);
777
778 /* If a normal allocation would succeed, stop compacting */
779 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
780 break;
781 }
782
783 return rc;
784}
785
786
787/* Compact all zones within a node */
788static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
789{
790 int zoneid;
791 struct zone *zone;
792
793 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
794
795 zone = &pgdat->node_zones[zoneid];
796 if (!populated_zone(zone))
797 continue;
798
799 cc->nr_freepages = 0;
800 cc->nr_migratepages = 0;
801 cc->zone = zone;
802 INIT_LIST_HEAD(&cc->freepages);
803 INIT_LIST_HEAD(&cc->migratepages);
804
805 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
806 compact_zone(zone, cc);
807
808 if (cc->order > 0) {
809 int ok = zone_watermark_ok(zone, cc->order,
810 low_wmark_pages(zone), 0, 0);
811 if (ok && cc->order > zone->compact_order_failed)
812 zone->compact_order_failed = cc->order + 1;
813 /* Currently async compaction is never deferred. */
814 else if (!ok && cc->sync)
815 defer_compaction(zone, cc->order);
816 }
817
818 VM_BUG_ON(!list_empty(&cc->freepages));
819 VM_BUG_ON(!list_empty(&cc->migratepages));
820 }
821
822 return 0;
823}
824
825int compact_pgdat(pg_data_t *pgdat, int order)
826{
827 struct compact_control cc = {
828 .order = order,
829 .sync = false,
830 };
831
832 return __compact_pgdat(pgdat, &cc);
833}
834
835static int compact_node(int nid)
836{
837 struct compact_control cc = {
838 .order = -1,
839 .sync = true,
840 };
841
842 return __compact_pgdat(NODE_DATA(nid), &cc);
843}
844
845/* Compact all nodes in the system */
846static int compact_nodes(void)
847{
848 int nid;
849
850 /* Flush pending updates to the LRU lists */
851 lru_add_drain_all();
852
853 for_each_online_node(nid)
854 compact_node(nid);
855
856 return COMPACT_COMPLETE;
857}
858
859/* The written value is actually unused, all memory is compacted */
860int sysctl_compact_memory;
861
862/* This is the entry point for compacting all nodes via /proc/sys/vm */
863int sysctl_compaction_handler(struct ctl_table *table, int write,
864 void __user *buffer, size_t *length, loff_t *ppos)
865{
866 if (write)
867 return compact_nodes();
868
869 return 0;
870}
871
872int sysctl_extfrag_handler(struct ctl_table *table, int write,
873 void __user *buffer, size_t *length, loff_t *ppos)
874{
875 proc_dointvec_minmax(table, write, buffer, length, ppos);
876
877 return 0;
878}
879
880#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
881ssize_t sysfs_compact_node(struct device *dev,
882 struct device_attribute *attr,
883 const char *buf, size_t count)
884{
885 int nid = dev->id;
886
887 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
888 /* Flush pending updates to the LRU lists */
889 lru_add_drain_all();
890
891 compact_node(nid);
892 }
893
894 return count;
895}
896static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
897
898int compaction_register_node(struct node *node)
899{
900 return device_create_file(&node->dev, &dev_attr_compact);
901}
902
903void compaction_unregister_node(struct node *node)
904{
905 return device_remove_file(&node->dev, &dev_attr_compact);
906}
907#endif /* CONFIG_SYSFS && CONFIG_NUMA */
908
909#endif /* CONFIG_COMPACTION */