Root/
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 | |
24 | static 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 | |
38 | static 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 | |
48 | static 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 | */ |
59 | static 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 | */ |
121 | unsigned long |
122 | isolate_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 */ |
176 | static 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 */ |
189 | static 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 | */ |
222 | unsigned long |
223 | isolate_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 */ |
365 | static 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 | */ |
390 | static 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 | /* |
426 | * Skip ahead if another thread is compacting in the area |
427 | * simultaneously. If we wrapped around, we can only skip |
428 | * ahead if zone->compact_cached_free_pfn also wrapped to |
429 | * above our starting point. |
430 | */ |
431 | if (cc->order > 0 && (!cc->wrapped || |
432 | zone->compact_cached_free_pfn > |
433 | cc->start_free_pfn)) |
434 | pfn = min(pfn, zone->compact_cached_free_pfn); |
435 | |
436 | if (!pfn_valid(pfn)) |
437 | continue; |
438 | |
439 | /* |
440 | * Check for overlapping nodes/zones. It's possible on some |
441 | * configurations to have a setup like |
442 | * node0 node1 node0 |
443 | * i.e. it's possible that all pages within a zones range of |
444 | * pages do not belong to a single zone. |
445 | */ |
446 | page = pfn_to_page(pfn); |
447 | if (page_zone(page) != zone) |
448 | continue; |
449 | |
450 | /* Check the block is suitable for migration */ |
451 | if (!suitable_migration_target(page)) |
452 | continue; |
453 | |
454 | /* |
455 | * Found a block suitable for isolating free pages from. Now |
456 | * we disabled interrupts, double check things are ok and |
457 | * isolate the pages. This is to minimise the time IRQs |
458 | * are disabled |
459 | */ |
460 | isolated = 0; |
461 | spin_lock_irqsave(&zone->lock, flags); |
462 | if (suitable_migration_target(page)) { |
463 | end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn); |
464 | isolated = isolate_freepages_block(pfn, end_pfn, |
465 | freelist, false); |
466 | nr_freepages += isolated; |
467 | } |
468 | spin_unlock_irqrestore(&zone->lock, flags); |
469 | |
470 | /* |
471 | * Record the highest PFN we isolated pages from. When next |
472 | * looking for free pages, the search will restart here as |
473 | * page migration may have returned some pages to the allocator |
474 | */ |
475 | if (isolated) { |
476 | high_pfn = max(high_pfn, pfn); |
477 | if (cc->order > 0) |
478 | zone->compact_cached_free_pfn = high_pfn; |
479 | } |
480 | } |
481 | |
482 | /* split_free_page does not map the pages */ |
483 | map_pages(freelist); |
484 | |
485 | cc->free_pfn = high_pfn; |
486 | cc->nr_freepages = nr_freepages; |
487 | } |
488 | |
489 | /* |
490 | * This is a migrate-callback that "allocates" freepages by taking pages |
491 | * from the isolated freelists in the block we are migrating to. |
492 | */ |
493 | static struct page *compaction_alloc(struct page *migratepage, |
494 | unsigned long data, |
495 | int **result) |
496 | { |
497 | struct compact_control *cc = (struct compact_control *)data; |
498 | struct page *freepage; |
499 | |
500 | /* Isolate free pages if necessary */ |
501 | if (list_empty(&cc->freepages)) { |
502 | isolate_freepages(cc->zone, cc); |
503 | |
504 | if (list_empty(&cc->freepages)) |
505 | return NULL; |
506 | } |
507 | |
508 | freepage = list_entry(cc->freepages.next, struct page, lru); |
509 | list_del(&freepage->lru); |
510 | cc->nr_freepages--; |
511 | |
512 | return freepage; |
513 | } |
514 | |
515 | /* |
516 | * We cannot control nr_migratepages and nr_freepages fully when migration is |
517 | * running as migrate_pages() has no knowledge of compact_control. When |
518 | * migration is complete, we count the number of pages on the lists by hand. |
519 | */ |
520 | static void update_nr_listpages(struct compact_control *cc) |
521 | { |
522 | int nr_migratepages = 0; |
523 | int nr_freepages = 0; |
524 | struct page *page; |
525 | |
526 | list_for_each_entry(page, &cc->migratepages, lru) |
527 | nr_migratepages++; |
528 | list_for_each_entry(page, &cc->freepages, lru) |
529 | nr_freepages++; |
530 | |
531 | cc->nr_migratepages = nr_migratepages; |
532 | cc->nr_freepages = nr_freepages; |
533 | } |
534 | |
535 | /* possible outcome of isolate_migratepages */ |
536 | typedef enum { |
537 | ISOLATE_ABORT, /* Abort compaction now */ |
538 | ISOLATE_NONE, /* No pages isolated, continue scanning */ |
539 | ISOLATE_SUCCESS, /* Pages isolated, migrate */ |
540 | } isolate_migrate_t; |
541 | |
542 | /* |
543 | * Isolate all pages that can be migrated from the block pointed to by |
544 | * the migrate scanner within compact_control. |
545 | */ |
546 | static isolate_migrate_t isolate_migratepages(struct zone *zone, |
547 | struct compact_control *cc) |
548 | { |
549 | unsigned long low_pfn, end_pfn; |
550 | |
551 | /* Do not scan outside zone boundaries */ |
552 | low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); |
553 | |
554 | /* Only scan within a pageblock boundary */ |
555 | end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); |
556 | |
557 | /* Do not cross the free scanner or scan within a memory hole */ |
558 | if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { |
559 | cc->migrate_pfn = end_pfn; |
560 | return ISOLATE_NONE; |
561 | } |
562 | |
563 | /* Perform the isolation */ |
564 | low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn); |
565 | if (!low_pfn) |
566 | return ISOLATE_ABORT; |
567 | |
568 | cc->migrate_pfn = low_pfn; |
569 | |
570 | return ISOLATE_SUCCESS; |
571 | } |
572 | |
573 | /* |
574 | * Returns the start pfn of the last page block in a zone. This is the starting |
575 | * point for full compaction of a zone. Compaction searches for free pages from |
576 | * the end of each zone, while isolate_freepages_block scans forward inside each |
577 | * page block. |
578 | */ |
579 | static unsigned long start_free_pfn(struct zone *zone) |
580 | { |
581 | unsigned long free_pfn; |
582 | free_pfn = zone->zone_start_pfn + zone->spanned_pages; |
583 | free_pfn &= ~(pageblock_nr_pages-1); |
584 | return free_pfn; |
585 | } |
586 | |
587 | static int compact_finished(struct zone *zone, |
588 | struct compact_control *cc) |
589 | { |
590 | unsigned int order; |
591 | unsigned long watermark; |
592 | |
593 | if (fatal_signal_pending(current)) |
594 | return COMPACT_PARTIAL; |
595 | |
596 | /* |
597 | * A full (order == -1) compaction run starts at the beginning and |
598 | * end of a zone; it completes when the migrate and free scanner meet. |
599 | * A partial (order > 0) compaction can start with the free scanner |
600 | * at a random point in the zone, and may have to restart. |
601 | */ |
602 | if (cc->free_pfn <= cc->migrate_pfn) { |
603 | if (cc->order > 0 && !cc->wrapped) { |
604 | /* We started partway through; restart at the end. */ |
605 | unsigned long free_pfn = start_free_pfn(zone); |
606 | zone->compact_cached_free_pfn = free_pfn; |
607 | cc->free_pfn = free_pfn; |
608 | cc->wrapped = 1; |
609 | return COMPACT_CONTINUE; |
610 | } |
611 | return COMPACT_COMPLETE; |
612 | } |
613 | |
614 | /* We wrapped around and ended up where we started. */ |
615 | if (cc->wrapped && cc->free_pfn <= cc->start_free_pfn) |
616 | return COMPACT_COMPLETE; |
617 | |
618 | /* |
619 | * order == -1 is expected when compacting via |
620 | * /proc/sys/vm/compact_memory |
621 | */ |
622 | if (cc->order == -1) |
623 | return COMPACT_CONTINUE; |
624 | |
625 | /* Compaction run is not finished if the watermark is not met */ |
626 | watermark = low_wmark_pages(zone); |
627 | watermark += (1 << cc->order); |
628 | |
629 | if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) |
630 | return COMPACT_CONTINUE; |
631 | |
632 | /* Direct compactor: Is a suitable page free? */ |
633 | for (order = cc->order; order < MAX_ORDER; order++) { |
634 | /* Job done if page is free of the right migratetype */ |
635 | if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) |
636 | return COMPACT_PARTIAL; |
637 | |
638 | /* Job done if allocation would set block type */ |
639 | if (order >= pageblock_order && zone->free_area[order].nr_free) |
640 | return COMPACT_PARTIAL; |
641 | } |
642 | |
643 | return COMPACT_CONTINUE; |
644 | } |
645 | |
646 | /* |
647 | * compaction_suitable: Is this suitable to run compaction on this zone now? |
648 | * Returns |
649 | * COMPACT_SKIPPED - If there are too few free pages for compaction |
650 | * COMPACT_PARTIAL - If the allocation would succeed without compaction |
651 | * COMPACT_CONTINUE - If compaction should run now |
652 | */ |
653 | unsigned long compaction_suitable(struct zone *zone, int order) |
654 | { |
655 | int fragindex; |
656 | unsigned long watermark; |
657 | |
658 | /* |
659 | * order == -1 is expected when compacting via |
660 | * /proc/sys/vm/compact_memory |
661 | */ |
662 | if (order == -1) |
663 | return COMPACT_CONTINUE; |
664 | |
665 | /* |
666 | * Watermarks for order-0 must be met for compaction. Note the 2UL. |
667 | * This is because during migration, copies of pages need to be |
668 | * allocated and for a short time, the footprint is higher |
669 | */ |
670 | watermark = low_wmark_pages(zone) + (2UL << order); |
671 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) |
672 | return COMPACT_SKIPPED; |
673 | |
674 | /* |
675 | * fragmentation index determines if allocation failures are due to |
676 | * low memory or external fragmentation |
677 | * |
678 | * index of -1000 implies allocations might succeed depending on |
679 | * watermarks |
680 | * index towards 0 implies failure is due to lack of memory |
681 | * index towards 1000 implies failure is due to fragmentation |
682 | * |
683 | * Only compact if a failure would be due to fragmentation. |
684 | */ |
685 | fragindex = fragmentation_index(zone, order); |
686 | if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) |
687 | return COMPACT_SKIPPED; |
688 | |
689 | if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark, |
690 | 0, 0)) |
691 | return COMPACT_PARTIAL; |
692 | |
693 | return COMPACT_CONTINUE; |
694 | } |
695 | |
696 | static int compact_zone(struct zone *zone, struct compact_control *cc) |
697 | { |
698 | int ret; |
699 | |
700 | ret = compaction_suitable(zone, cc->order); |
701 | switch (ret) { |
702 | case COMPACT_PARTIAL: |
703 | case COMPACT_SKIPPED: |
704 | /* Compaction is likely to fail */ |
705 | return ret; |
706 | case COMPACT_CONTINUE: |
707 | /* Fall through to compaction */ |
708 | ; |
709 | } |
710 | |
711 | /* Setup to move all movable pages to the end of the zone */ |
712 | cc->migrate_pfn = zone->zone_start_pfn; |
713 | |
714 | if (cc->order > 0) { |
715 | /* Incremental compaction. Start where the last one stopped. */ |
716 | cc->free_pfn = zone->compact_cached_free_pfn; |
717 | cc->start_free_pfn = cc->free_pfn; |
718 | } else { |
719 | /* Order == -1 starts at the end of the zone. */ |
720 | cc->free_pfn = start_free_pfn(zone); |
721 | } |
722 | |
723 | migrate_prep_local(); |
724 | |
725 | while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { |
726 | unsigned long nr_migrate, nr_remaining; |
727 | int err; |
728 | |
729 | switch (isolate_migratepages(zone, cc)) { |
730 | case ISOLATE_ABORT: |
731 | ret = COMPACT_PARTIAL; |
732 | goto out; |
733 | case ISOLATE_NONE: |
734 | continue; |
735 | case ISOLATE_SUCCESS: |
736 | ; |
737 | } |
738 | |
739 | nr_migrate = cc->nr_migratepages; |
740 | err = migrate_pages(&cc->migratepages, compaction_alloc, |
741 | (unsigned long)cc, false, |
742 | cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC); |
743 | update_nr_listpages(cc); |
744 | nr_remaining = cc->nr_migratepages; |
745 | |
746 | count_vm_event(COMPACTBLOCKS); |
747 | count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); |
748 | if (nr_remaining) |
749 | count_vm_events(COMPACTPAGEFAILED, nr_remaining); |
750 | trace_mm_compaction_migratepages(nr_migrate - nr_remaining, |
751 | nr_remaining); |
752 | |
753 | /* Release LRU pages not migrated */ |
754 | if (err) { |
755 | putback_lru_pages(&cc->migratepages); |
756 | cc->nr_migratepages = 0; |
757 | if (err == -ENOMEM) { |
758 | ret = COMPACT_PARTIAL; |
759 | goto out; |
760 | } |
761 | } |
762 | } |
763 | |
764 | out: |
765 | /* Release free pages and check accounting */ |
766 | cc->nr_freepages -= release_freepages(&cc->freepages); |
767 | VM_BUG_ON(cc->nr_freepages != 0); |
768 | |
769 | return ret; |
770 | } |
771 | |
772 | static unsigned long compact_zone_order(struct zone *zone, |
773 | int order, gfp_t gfp_mask, |
774 | bool sync) |
775 | { |
776 | struct compact_control cc = { |
777 | .nr_freepages = 0, |
778 | .nr_migratepages = 0, |
779 | .order = order, |
780 | .migratetype = allocflags_to_migratetype(gfp_mask), |
781 | .zone = zone, |
782 | .sync = sync, |
783 | }; |
784 | INIT_LIST_HEAD(&cc.freepages); |
785 | INIT_LIST_HEAD(&cc.migratepages); |
786 | |
787 | return compact_zone(zone, &cc); |
788 | } |
789 | |
790 | int sysctl_extfrag_threshold = 500; |
791 | |
792 | /** |
793 | * try_to_compact_pages - Direct compact to satisfy a high-order allocation |
794 | * @zonelist: The zonelist used for the current allocation |
795 | * @order: The order of the current allocation |
796 | * @gfp_mask: The GFP mask of the current allocation |
797 | * @nodemask: The allowed nodes to allocate from |
798 | * @sync: Whether migration is synchronous or not |
799 | * |
800 | * This is the main entry point for direct page compaction. |
801 | */ |
802 | unsigned long try_to_compact_pages(struct zonelist *zonelist, |
803 | int order, gfp_t gfp_mask, nodemask_t *nodemask, |
804 | bool sync) |
805 | { |
806 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); |
807 | int may_enter_fs = gfp_mask & __GFP_FS; |
808 | int may_perform_io = gfp_mask & __GFP_IO; |
809 | struct zoneref *z; |
810 | struct zone *zone; |
811 | int rc = COMPACT_SKIPPED; |
812 | |
813 | /* |
814 | * Check whether it is worth even starting compaction. The order check is |
815 | * made because an assumption is made that the page allocator can satisfy |
816 | * the "cheaper" orders without taking special steps |
817 | */ |
818 | if (!order || !may_enter_fs || !may_perform_io) |
819 | return rc; |
820 | |
821 | count_vm_event(COMPACTSTALL); |
822 | |
823 | /* Compact each zone in the list */ |
824 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, |
825 | nodemask) { |
826 | int status; |
827 | |
828 | status = compact_zone_order(zone, order, gfp_mask, sync); |
829 | rc = max(status, rc); |
830 | |
831 | /* If a normal allocation would succeed, stop compacting */ |
832 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) |
833 | break; |
834 | } |
835 | |
836 | return rc; |
837 | } |
838 | |
839 | |
840 | /* Compact all zones within a node */ |
841 | static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc) |
842 | { |
843 | int zoneid; |
844 | struct zone *zone; |
845 | |
846 | for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { |
847 | |
848 | zone = &pgdat->node_zones[zoneid]; |
849 | if (!populated_zone(zone)) |
850 | continue; |
851 | |
852 | cc->nr_freepages = 0; |
853 | cc->nr_migratepages = 0; |
854 | cc->zone = zone; |
855 | INIT_LIST_HEAD(&cc->freepages); |
856 | INIT_LIST_HEAD(&cc->migratepages); |
857 | |
858 | if (cc->order == -1 || !compaction_deferred(zone, cc->order)) |
859 | compact_zone(zone, cc); |
860 | |
861 | if (cc->order > 0) { |
862 | int ok = zone_watermark_ok(zone, cc->order, |
863 | low_wmark_pages(zone), 0, 0); |
864 | if (ok && cc->order > zone->compact_order_failed) |
865 | zone->compact_order_failed = cc->order + 1; |
866 | /* Currently async compaction is never deferred. */ |
867 | else if (!ok && cc->sync) |
868 | defer_compaction(zone, cc->order); |
869 | } |
870 | |
871 | VM_BUG_ON(!list_empty(&cc->freepages)); |
872 | VM_BUG_ON(!list_empty(&cc->migratepages)); |
873 | } |
874 | |
875 | return 0; |
876 | } |
877 | |
878 | int compact_pgdat(pg_data_t *pgdat, int order) |
879 | { |
880 | struct compact_control cc = { |
881 | .order = order, |
882 | .sync = false, |
883 | }; |
884 | |
885 | return __compact_pgdat(pgdat, &cc); |
886 | } |
887 | |
888 | static int compact_node(int nid) |
889 | { |
890 | struct compact_control cc = { |
891 | .order = -1, |
892 | .sync = true, |
893 | }; |
894 | |
895 | return __compact_pgdat(NODE_DATA(nid), &cc); |
896 | } |
897 | |
898 | /* Compact all nodes in the system */ |
899 | static int compact_nodes(void) |
900 | { |
901 | int nid; |
902 | |
903 | /* Flush pending updates to the LRU lists */ |
904 | lru_add_drain_all(); |
905 | |
906 | for_each_online_node(nid) |
907 | compact_node(nid); |
908 | |
909 | return COMPACT_COMPLETE; |
910 | } |
911 | |
912 | /* The written value is actually unused, all memory is compacted */ |
913 | int sysctl_compact_memory; |
914 | |
915 | /* This is the entry point for compacting all nodes via /proc/sys/vm */ |
916 | int sysctl_compaction_handler(struct ctl_table *table, int write, |
917 | void __user *buffer, size_t *length, loff_t *ppos) |
918 | { |
919 | if (write) |
920 | return compact_nodes(); |
921 | |
922 | return 0; |
923 | } |
924 | |
925 | int sysctl_extfrag_handler(struct ctl_table *table, int write, |
926 | void __user *buffer, size_t *length, loff_t *ppos) |
927 | { |
928 | proc_dointvec_minmax(table, write, buffer, length, ppos); |
929 | |
930 | return 0; |
931 | } |
932 | |
933 | #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) |
934 | ssize_t sysfs_compact_node(struct device *dev, |
935 | struct device_attribute *attr, |
936 | const char *buf, size_t count) |
937 | { |
938 | int nid = dev->id; |
939 | |
940 | if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { |
941 | /* Flush pending updates to the LRU lists */ |
942 | lru_add_drain_all(); |
943 | |
944 | compact_node(nid); |
945 | } |
946 | |
947 | return count; |
948 | } |
949 | static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); |
950 | |
951 | int compaction_register_node(struct node *node) |
952 | { |
953 | return device_create_file(&node->dev, &dev_attr_compact); |
954 | } |
955 | |
956 | void compaction_unregister_node(struct node *node) |
957 | { |
958 | return device_remove_file(&node->dev, &dev_attr_compact); |
959 | } |
960 | #endif /* CONFIG_SYSFS && CONFIG_NUMA */ |
961 | |
962 | #endif /* CONFIG_COMPACTION */ |
963 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9