]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Memory Migration functionality - linux/mm/migration.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
12 | * Christoph Lameter | |
13 | */ | |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/export.h> | |
17 | #include <linux/swap.h> | |
18 | #include <linux/swapops.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/buffer_head.h> | |
21 | #include <linux/mm_inline.h> | |
22 | #include <linux/nsproxy.h> | |
23 | #include <linux/pagevec.h> | |
24 | #include <linux/ksm.h> | |
25 | #include <linux/rmap.h> | |
26 | #include <linux/topology.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/cpuset.h> | |
29 | #include <linux/writeback.h> | |
30 | #include <linux/mempolicy.h> | |
31 | #include <linux/vmalloc.h> | |
32 | #include <linux/security.h> | |
33 | #include <linux/memcontrol.h> | |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/hugetlb.h> | |
36 | #include <linux/hugetlb_cgroup.h> | |
37 | #include <linux/gfp.h> | |
38 | #include <linux/balloon_compaction.h> | |
39 | #include <linux/mmu_notifier.h> | |
40 | ||
41 | #include <asm/tlbflush.h> | |
42 | ||
43 | #define CREATE_TRACE_POINTS | |
44 | #include <trace/events/migrate.h> | |
45 | ||
46 | #include "internal.h" | |
47 | ||
48 | /* | |
49 | * migrate_prep() needs to be called before we start compiling a list of pages | |
50 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is | |
51 | * undesirable, use migrate_prep_local() | |
52 | */ | |
53 | int migrate_prep(void) | |
54 | { | |
55 | /* | |
56 | * Clear the LRU lists so pages can be isolated. | |
57 | * Note that pages may be moved off the LRU after we have | |
58 | * drained them. Those pages will fail to migrate like other | |
59 | * pages that may be busy. | |
60 | */ | |
61 | lru_add_drain_all(); | |
62 | ||
63 | return 0; | |
64 | } | |
65 | ||
66 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ | |
67 | int migrate_prep_local(void) | |
68 | { | |
69 | lru_add_drain(); | |
70 | ||
71 | return 0; | |
72 | } | |
73 | ||
74 | /* | |
75 | * Put previously isolated pages back onto the appropriate lists | |
76 | * from where they were once taken off for compaction/migration. | |
77 | * | |
78 | * This function shall be used whenever the isolated pageset has been | |
79 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
80 | * and isolate_huge_page(). | |
81 | */ | |
82 | void putback_movable_pages(struct list_head *l) | |
83 | { | |
84 | struct page *page; | |
85 | struct page *page2; | |
86 | ||
87 | list_for_each_entry_safe(page, page2, l, lru) { | |
88 | if (unlikely(PageHuge(page))) { | |
89 | putback_active_hugepage(page); | |
90 | continue; | |
91 | } | |
92 | list_del(&page->lru); | |
93 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
94 | page_is_file_cache(page)); | |
95 | if (unlikely(isolated_balloon_page(page))) | |
96 | balloon_page_putback(page); | |
97 | else | |
98 | putback_lru_page(page); | |
99 | } | |
100 | } | |
101 | ||
102 | /* | |
103 | * Restore a potential migration pte to a working pte entry | |
104 | */ | |
105 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, | |
106 | unsigned long addr, void *old) | |
107 | { | |
108 | struct mm_struct *mm = vma->vm_mm; | |
109 | swp_entry_t entry; | |
110 | pmd_t *pmd; | |
111 | pte_t *ptep, pte; | |
112 | spinlock_t *ptl; | |
113 | ||
114 | if (unlikely(PageHuge(new))) { | |
115 | ptep = huge_pte_offset(mm, addr); | |
116 | if (!ptep) | |
117 | goto out; | |
118 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep); | |
119 | } else { | |
120 | pmd = mm_find_pmd(mm, addr); | |
121 | if (!pmd) | |
122 | goto out; | |
123 | ||
124 | ptep = pte_offset_map(pmd, addr); | |
125 | ||
126 | /* | |
127 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
128 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
129 | */ | |
130 | ||
131 | ptl = pte_lockptr(mm, pmd); | |
132 | } | |
133 | ||
134 | spin_lock(ptl); | |
135 | pte = *ptep; | |
136 | if (!is_swap_pte(pte)) | |
137 | goto unlock; | |
138 | ||
139 | entry = pte_to_swp_entry(pte); | |
140 | ||
141 | if (!is_migration_entry(entry) || | |
142 | migration_entry_to_page(entry) != old) | |
143 | goto unlock; | |
144 | ||
145 | get_page(new); | |
146 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
147 | if (pte_swp_soft_dirty(*ptep)) | |
148 | pte = pte_mksoft_dirty(pte); | |
149 | if (is_write_migration_entry(entry)) | |
150 | pte = pte_mkwrite(pte); | |
151 | #ifdef CONFIG_HUGETLB_PAGE | |
152 | if (PageHuge(new)) { | |
153 | pte = pte_mkhuge(pte); | |
154 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
155 | } | |
156 | #endif | |
157 | flush_dcache_page(new); | |
158 | set_pte_at(mm, addr, ptep, pte); | |
159 | ||
160 | if (PageHuge(new)) { | |
161 | if (PageAnon(new)) | |
162 | hugepage_add_anon_rmap(new, vma, addr); | |
163 | else | |
164 | page_dup_rmap(new); | |
165 | } else if (PageAnon(new)) | |
166 | page_add_anon_rmap(new, vma, addr); | |
167 | else | |
168 | page_add_file_rmap(new); | |
169 | ||
170 | /* No need to invalidate - it was non-present before */ | |
171 | update_mmu_cache(vma, addr, ptep); | |
172 | unlock: | |
173 | pte_unmap_unlock(ptep, ptl); | |
174 | out: | |
175 | return SWAP_AGAIN; | |
176 | } | |
177 | ||
178 | /* | |
179 | * Congratulations to trinity for discovering this bug. | |
180 | * mm/fremap.c's remap_file_pages() accepts any range within a single vma to | |
181 | * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then | |
182 | * replace the specified range by file ptes throughout (maybe populated after). | |
183 | * If page migration finds a page within that range, while it's still located | |
184 | * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem: | |
185 | * zap_pte() clears the temporary migration entry before mmap_sem is dropped. | |
186 | * But if the migrating page is in a part of the vma outside the range to be | |
187 | * remapped, then it will not be cleared, and remove_migration_ptes() needs to | |
188 | * deal with it. Fortunately, this part of the vma is of course still linear, | |
189 | * so we just need to use linear location on the nonlinear list. | |
190 | */ | |
191 | static int remove_linear_migration_ptes_from_nonlinear(struct page *page, | |
192 | struct address_space *mapping, void *arg) | |
193 | { | |
194 | struct vm_area_struct *vma; | |
195 | /* hugetlbfs does not support remap_pages, so no huge pgoff worries */ | |
196 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
197 | unsigned long addr; | |
198 | ||
199 | list_for_each_entry(vma, | |
200 | &mapping->i_mmap_nonlinear, shared.nonlinear) { | |
201 | ||
202 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
203 | if (addr >= vma->vm_start && addr < vma->vm_end) | |
204 | remove_migration_pte(page, vma, addr, arg); | |
205 | } | |
206 | return SWAP_AGAIN; | |
207 | } | |
208 | ||
209 | /* | |
210 | * Get rid of all migration entries and replace them by | |
211 | * references to the indicated page. | |
212 | */ | |
213 | static void remove_migration_ptes(struct page *old, struct page *new) | |
214 | { | |
215 | struct rmap_walk_control rwc = { | |
216 | .rmap_one = remove_migration_pte, | |
217 | .arg = old, | |
218 | .file_nonlinear = remove_linear_migration_ptes_from_nonlinear, | |
219 | }; | |
220 | ||
221 | rmap_walk(new, &rwc); | |
222 | } | |
223 | ||
224 | /* | |
225 | * Something used the pte of a page under migration. We need to | |
226 | * get to the page and wait until migration is finished. | |
227 | * When we return from this function the fault will be retried. | |
228 | */ | |
229 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, | |
230 | spinlock_t *ptl) | |
231 | { | |
232 | pte_t pte; | |
233 | swp_entry_t entry; | |
234 | struct page *page; | |
235 | ||
236 | spin_lock(ptl); | |
237 | pte = *ptep; | |
238 | if (!is_swap_pte(pte)) | |
239 | goto out; | |
240 | ||
241 | entry = pte_to_swp_entry(pte); | |
242 | if (!is_migration_entry(entry)) | |
243 | goto out; | |
244 | ||
245 | page = migration_entry_to_page(entry); | |
246 | ||
247 | /* | |
248 | * Once radix-tree replacement of page migration started, page_count | |
249 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
250 | * against a page without get_page(). | |
251 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
252 | * will occur again. | |
253 | */ | |
254 | if (!get_page_unless_zero(page)) | |
255 | goto out; | |
256 | pte_unmap_unlock(ptep, ptl); | |
257 | wait_on_page_locked(page); | |
258 | put_page(page); | |
259 | return; | |
260 | out: | |
261 | pte_unmap_unlock(ptep, ptl); | |
262 | } | |
263 | ||
264 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
265 | unsigned long address) | |
266 | { | |
267 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
268 | pte_t *ptep = pte_offset_map(pmd, address); | |
269 | __migration_entry_wait(mm, ptep, ptl); | |
270 | } | |
271 | ||
272 | void migration_entry_wait_huge(struct vm_area_struct *vma, | |
273 | struct mm_struct *mm, pte_t *pte) | |
274 | { | |
275 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); | |
276 | __migration_entry_wait(mm, pte, ptl); | |
277 | } | |
278 | ||
279 | #ifdef CONFIG_BLOCK | |
280 | /* Returns true if all buffers are successfully locked */ | |
281 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
282 | enum migrate_mode mode) | |
283 | { | |
284 | struct buffer_head *bh = head; | |
285 | ||
286 | /* Simple case, sync compaction */ | |
287 | if (mode != MIGRATE_ASYNC) { | |
288 | do { | |
289 | get_bh(bh); | |
290 | lock_buffer(bh); | |
291 | bh = bh->b_this_page; | |
292 | ||
293 | } while (bh != head); | |
294 | ||
295 | return true; | |
296 | } | |
297 | ||
298 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
299 | do { | |
300 | get_bh(bh); | |
301 | if (!trylock_buffer(bh)) { | |
302 | /* | |
303 | * We failed to lock the buffer and cannot stall in | |
304 | * async migration. Release the taken locks | |
305 | */ | |
306 | struct buffer_head *failed_bh = bh; | |
307 | put_bh(failed_bh); | |
308 | bh = head; | |
309 | while (bh != failed_bh) { | |
310 | unlock_buffer(bh); | |
311 | put_bh(bh); | |
312 | bh = bh->b_this_page; | |
313 | } | |
314 | return false; | |
315 | } | |
316 | ||
317 | bh = bh->b_this_page; | |
318 | } while (bh != head); | |
319 | return true; | |
320 | } | |
321 | #else | |
322 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
323 | enum migrate_mode mode) | |
324 | { | |
325 | return true; | |
326 | } | |
327 | #endif /* CONFIG_BLOCK */ | |
328 | ||
329 | /* | |
330 | * Replace the page in the mapping. | |
331 | * | |
332 | * The number of remaining references must be: | |
333 | * 1 for anonymous pages without a mapping | |
334 | * 2 for pages with a mapping | |
335 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. | |
336 | */ | |
337 | int migrate_page_move_mapping(struct address_space *mapping, | |
338 | struct page *newpage, struct page *page, | |
339 | struct buffer_head *head, enum migrate_mode mode, | |
340 | int extra_count) | |
341 | { | |
342 | int expected_count = 1 + extra_count; | |
343 | void **pslot; | |
344 | ||
345 | if (!mapping) { | |
346 | /* Anonymous page without mapping */ | |
347 | if (page_count(page) != expected_count) | |
348 | return -EAGAIN; | |
349 | return MIGRATEPAGE_SUCCESS; | |
350 | } | |
351 | ||
352 | spin_lock_irq(&mapping->tree_lock); | |
353 | ||
354 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
355 | page_index(page)); | |
356 | ||
357 | expected_count += 1 + page_has_private(page); | |
358 | if (page_count(page) != expected_count || | |
359 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
360 | spin_unlock_irq(&mapping->tree_lock); | |
361 | return -EAGAIN; | |
362 | } | |
363 | ||
364 | if (!page_freeze_refs(page, expected_count)) { | |
365 | spin_unlock_irq(&mapping->tree_lock); | |
366 | return -EAGAIN; | |
367 | } | |
368 | ||
369 | /* | |
370 | * In the async migration case of moving a page with buffers, lock the | |
371 | * buffers using trylock before the mapping is moved. If the mapping | |
372 | * was moved, we later failed to lock the buffers and could not move | |
373 | * the mapping back due to an elevated page count, we would have to | |
374 | * block waiting on other references to be dropped. | |
375 | */ | |
376 | if (mode == MIGRATE_ASYNC && head && | |
377 | !buffer_migrate_lock_buffers(head, mode)) { | |
378 | page_unfreeze_refs(page, expected_count); | |
379 | spin_unlock_irq(&mapping->tree_lock); | |
380 | return -EAGAIN; | |
381 | } | |
382 | ||
383 | /* | |
384 | * Now we know that no one else is looking at the page. | |
385 | */ | |
386 | get_page(newpage); /* add cache reference */ | |
387 | if (PageSwapCache(page)) { | |
388 | SetPageSwapCache(newpage); | |
389 | set_page_private(newpage, page_private(page)); | |
390 | } | |
391 | ||
392 | radix_tree_replace_slot(pslot, newpage); | |
393 | ||
394 | /* | |
395 | * Drop cache reference from old page by unfreezing | |
396 | * to one less reference. | |
397 | * We know this isn't the last reference. | |
398 | */ | |
399 | page_unfreeze_refs(page, expected_count - 1); | |
400 | ||
401 | /* | |
402 | * If moved to a different zone then also account | |
403 | * the page for that zone. Other VM counters will be | |
404 | * taken care of when we establish references to the | |
405 | * new page and drop references to the old page. | |
406 | * | |
407 | * Note that anonymous pages are accounted for | |
408 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
409 | * are mapped to swap space. | |
410 | */ | |
411 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
412 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
413 | if (!PageSwapCache(page) && PageSwapBacked(page)) { | |
414 | __dec_zone_page_state(page, NR_SHMEM); | |
415 | __inc_zone_page_state(newpage, NR_SHMEM); | |
416 | } | |
417 | spin_unlock_irq(&mapping->tree_lock); | |
418 | ||
419 | return MIGRATEPAGE_SUCCESS; | |
420 | } | |
421 | ||
422 | /* | |
423 | * The expected number of remaining references is the same as that | |
424 | * of migrate_page_move_mapping(). | |
425 | */ | |
426 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
427 | struct page *newpage, struct page *page) | |
428 | { | |
429 | int expected_count; | |
430 | void **pslot; | |
431 | ||
432 | if (!mapping) { | |
433 | if (page_count(page) != 1) | |
434 | return -EAGAIN; | |
435 | return MIGRATEPAGE_SUCCESS; | |
436 | } | |
437 | ||
438 | spin_lock_irq(&mapping->tree_lock); | |
439 | ||
440 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
441 | page_index(page)); | |
442 | ||
443 | expected_count = 2 + page_has_private(page); | |
444 | if (page_count(page) != expected_count || | |
445 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
446 | spin_unlock_irq(&mapping->tree_lock); | |
447 | return -EAGAIN; | |
448 | } | |
449 | ||
450 | if (!page_freeze_refs(page, expected_count)) { | |
451 | spin_unlock_irq(&mapping->tree_lock); | |
452 | return -EAGAIN; | |
453 | } | |
454 | ||
455 | get_page(newpage); | |
456 | ||
457 | radix_tree_replace_slot(pslot, newpage); | |
458 | ||
459 | page_unfreeze_refs(page, expected_count - 1); | |
460 | ||
461 | spin_unlock_irq(&mapping->tree_lock); | |
462 | return MIGRATEPAGE_SUCCESS; | |
463 | } | |
464 | ||
465 | /* | |
466 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
467 | * arithmetic will work across the entire page. We need something more | |
468 | * specialized. | |
469 | */ | |
470 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
471 | int nr_pages) | |
472 | { | |
473 | int i; | |
474 | struct page *dst_base = dst; | |
475 | struct page *src_base = src; | |
476 | ||
477 | for (i = 0; i < nr_pages; ) { | |
478 | cond_resched(); | |
479 | copy_highpage(dst, src); | |
480 | ||
481 | i++; | |
482 | dst = mem_map_next(dst, dst_base, i); | |
483 | src = mem_map_next(src, src_base, i); | |
484 | } | |
485 | } | |
486 | ||
487 | static void copy_huge_page(struct page *dst, struct page *src) | |
488 | { | |
489 | int i; | |
490 | int nr_pages; | |
491 | ||
492 | if (PageHuge(src)) { | |
493 | /* hugetlbfs page */ | |
494 | struct hstate *h = page_hstate(src); | |
495 | nr_pages = pages_per_huge_page(h); | |
496 | ||
497 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
498 | __copy_gigantic_page(dst, src, nr_pages); | |
499 | return; | |
500 | } | |
501 | } else { | |
502 | /* thp page */ | |
503 | BUG_ON(!PageTransHuge(src)); | |
504 | nr_pages = hpage_nr_pages(src); | |
505 | } | |
506 | ||
507 | for (i = 0; i < nr_pages; i++) { | |
508 | cond_resched(); | |
509 | copy_highpage(dst + i, src + i); | |
510 | } | |
511 | } | |
512 | ||
513 | /* | |
514 | * Copy the page to its new location | |
515 | */ | |
516 | void migrate_page_copy(struct page *newpage, struct page *page) | |
517 | { | |
518 | int cpupid; | |
519 | ||
520 | if (PageHuge(page) || PageTransHuge(page)) | |
521 | copy_huge_page(newpage, page); | |
522 | else | |
523 | copy_highpage(newpage, page); | |
524 | ||
525 | if (PageError(page)) | |
526 | SetPageError(newpage); | |
527 | if (PageReferenced(page)) | |
528 | SetPageReferenced(newpage); | |
529 | if (PageUptodate(page)) | |
530 | SetPageUptodate(newpage); | |
531 | if (TestClearPageActive(page)) { | |
532 | VM_BUG_ON_PAGE(PageUnevictable(page), page); | |
533 | SetPageActive(newpage); | |
534 | } else if (TestClearPageUnevictable(page)) | |
535 | SetPageUnevictable(newpage); | |
536 | if (PageChecked(page)) | |
537 | SetPageChecked(newpage); | |
538 | if (PageMappedToDisk(page)) | |
539 | SetPageMappedToDisk(newpage); | |
540 | ||
541 | if (PageDirty(page)) { | |
542 | clear_page_dirty_for_io(page); | |
543 | /* | |
544 | * Want to mark the page and the radix tree as dirty, and | |
545 | * redo the accounting that clear_page_dirty_for_io undid, | |
546 | * but we can't use set_page_dirty because that function | |
547 | * is actually a signal that all of the page has become dirty. | |
548 | * Whereas only part of our page may be dirty. | |
549 | */ | |
550 | if (PageSwapBacked(page)) | |
551 | SetPageDirty(newpage); | |
552 | else | |
553 | __set_page_dirty_nobuffers(newpage); | |
554 | } | |
555 | ||
556 | /* | |
557 | * Copy NUMA information to the new page, to prevent over-eager | |
558 | * future migrations of this same page. | |
559 | */ | |
560 | cpupid = page_cpupid_xchg_last(page, -1); | |
561 | page_cpupid_xchg_last(newpage, cpupid); | |
562 | ||
563 | mlock_migrate_page(newpage, page); | |
564 | ksm_migrate_page(newpage, page); | |
565 | /* | |
566 | * Please do not reorder this without considering how mm/ksm.c's | |
567 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
568 | */ | |
569 | ClearPageSwapCache(page); | |
570 | ClearPagePrivate(page); | |
571 | set_page_private(page, 0); | |
572 | ||
573 | /* | |
574 | * If any waiters have accumulated on the new page then | |
575 | * wake them up. | |
576 | */ | |
577 | if (PageWriteback(newpage)) | |
578 | end_page_writeback(newpage); | |
579 | } | |
580 | ||
581 | /************************************************************ | |
582 | * Migration functions | |
583 | ***********************************************************/ | |
584 | ||
585 | /* | |
586 | * Common logic to directly migrate a single page suitable for | |
587 | * pages that do not use PagePrivate/PagePrivate2. | |
588 | * | |
589 | * Pages are locked upon entry and exit. | |
590 | */ | |
591 | int migrate_page(struct address_space *mapping, | |
592 | struct page *newpage, struct page *page, | |
593 | enum migrate_mode mode) | |
594 | { | |
595 | int rc; | |
596 | ||
597 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
598 | ||
599 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); | |
600 | ||
601 | if (rc != MIGRATEPAGE_SUCCESS) | |
602 | return rc; | |
603 | ||
604 | migrate_page_copy(newpage, page); | |
605 | return MIGRATEPAGE_SUCCESS; | |
606 | } | |
607 | EXPORT_SYMBOL(migrate_page); | |
608 | ||
609 | #ifdef CONFIG_BLOCK | |
610 | /* | |
611 | * Migration function for pages with buffers. This function can only be used | |
612 | * if the underlying filesystem guarantees that no other references to "page" | |
613 | * exist. | |
614 | */ | |
615 | int buffer_migrate_page(struct address_space *mapping, | |
616 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
617 | { | |
618 | struct buffer_head *bh, *head; | |
619 | int rc; | |
620 | ||
621 | if (!page_has_buffers(page)) | |
622 | return migrate_page(mapping, newpage, page, mode); | |
623 | ||
624 | head = page_buffers(page); | |
625 | ||
626 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); | |
627 | ||
628 | if (rc != MIGRATEPAGE_SUCCESS) | |
629 | return rc; | |
630 | ||
631 | /* | |
632 | * In the async case, migrate_page_move_mapping locked the buffers | |
633 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
634 | * need to be locked now | |
635 | */ | |
636 | if (mode != MIGRATE_ASYNC) | |
637 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
638 | ||
639 | ClearPagePrivate(page); | |
640 | set_page_private(newpage, page_private(page)); | |
641 | set_page_private(page, 0); | |
642 | put_page(page); | |
643 | get_page(newpage); | |
644 | ||
645 | bh = head; | |
646 | do { | |
647 | set_bh_page(bh, newpage, bh_offset(bh)); | |
648 | bh = bh->b_this_page; | |
649 | ||
650 | } while (bh != head); | |
651 | ||
652 | SetPagePrivate(newpage); | |
653 | ||
654 | migrate_page_copy(newpage, page); | |
655 | ||
656 | bh = head; | |
657 | do { | |
658 | unlock_buffer(bh); | |
659 | put_bh(bh); | |
660 | bh = bh->b_this_page; | |
661 | ||
662 | } while (bh != head); | |
663 | ||
664 | return MIGRATEPAGE_SUCCESS; | |
665 | } | |
666 | EXPORT_SYMBOL(buffer_migrate_page); | |
667 | #endif | |
668 | ||
669 | /* | |
670 | * Writeback a page to clean the dirty state | |
671 | */ | |
672 | static int writeout(struct address_space *mapping, struct page *page) | |
673 | { | |
674 | struct writeback_control wbc = { | |
675 | .sync_mode = WB_SYNC_NONE, | |
676 | .nr_to_write = 1, | |
677 | .range_start = 0, | |
678 | .range_end = LLONG_MAX, | |
679 | .for_reclaim = 1 | |
680 | }; | |
681 | int rc; | |
682 | ||
683 | if (!mapping->a_ops->writepage) | |
684 | /* No write method for the address space */ | |
685 | return -EINVAL; | |
686 | ||
687 | if (!clear_page_dirty_for_io(page)) | |
688 | /* Someone else already triggered a write */ | |
689 | return -EAGAIN; | |
690 | ||
691 | /* | |
692 | * A dirty page may imply that the underlying filesystem has | |
693 | * the page on some queue. So the page must be clean for | |
694 | * migration. Writeout may mean we loose the lock and the | |
695 | * page state is no longer what we checked for earlier. | |
696 | * At this point we know that the migration attempt cannot | |
697 | * be successful. | |
698 | */ | |
699 | remove_migration_ptes(page, page); | |
700 | ||
701 | rc = mapping->a_ops->writepage(page, &wbc); | |
702 | ||
703 | if (rc != AOP_WRITEPAGE_ACTIVATE) | |
704 | /* unlocked. Relock */ | |
705 | lock_page(page); | |
706 | ||
707 | return (rc < 0) ? -EIO : -EAGAIN; | |
708 | } | |
709 | ||
710 | /* | |
711 | * Default handling if a filesystem does not provide a migration function. | |
712 | */ | |
713 | static int fallback_migrate_page(struct address_space *mapping, | |
714 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
715 | { | |
716 | if (PageDirty(page)) { | |
717 | /* Only writeback pages in full synchronous migration */ | |
718 | if (mode != MIGRATE_SYNC) | |
719 | return -EBUSY; | |
720 | return writeout(mapping, page); | |
721 | } | |
722 | ||
723 | /* | |
724 | * Buffers may be managed in a filesystem specific way. | |
725 | * We must have no buffers or drop them. | |
726 | */ | |
727 | if (page_has_private(page) && | |
728 | !try_to_release_page(page, GFP_KERNEL)) | |
729 | return -EAGAIN; | |
730 | ||
731 | return migrate_page(mapping, newpage, page, mode); | |
732 | } | |
733 | ||
734 | /* | |
735 | * Move a page to a newly allocated page | |
736 | * The page is locked and all ptes have been successfully removed. | |
737 | * | |
738 | * The new page will have replaced the old page if this function | |
739 | * is successful. | |
740 | * | |
741 | * Return value: | |
742 | * < 0 - error code | |
743 | * MIGRATEPAGE_SUCCESS - success | |
744 | */ | |
745 | static int move_to_new_page(struct page *newpage, struct page *page, | |
746 | int remap_swapcache, enum migrate_mode mode) | |
747 | { | |
748 | struct address_space *mapping; | |
749 | int rc; | |
750 | ||
751 | /* | |
752 | * Block others from accessing the page when we get around to | |
753 | * establishing additional references. We are the only one | |
754 | * holding a reference to the new page at this point. | |
755 | */ | |
756 | if (!trylock_page(newpage)) | |
757 | BUG(); | |
758 | ||
759 | /* Prepare mapping for the new page.*/ | |
760 | newpage->index = page->index; | |
761 | newpage->mapping = page->mapping; | |
762 | if (PageSwapBacked(page)) | |
763 | SetPageSwapBacked(newpage); | |
764 | ||
765 | mapping = page_mapping(page); | |
766 | if (!mapping) | |
767 | rc = migrate_page(mapping, newpage, page, mode); | |
768 | else if (mapping->a_ops->migratepage) | |
769 | /* | |
770 | * Most pages have a mapping and most filesystems provide a | |
771 | * migratepage callback. Anonymous pages are part of swap | |
772 | * space which also has its own migratepage callback. This | |
773 | * is the most common path for page migration. | |
774 | */ | |
775 | rc = mapping->a_ops->migratepage(mapping, | |
776 | newpage, page, mode); | |
777 | else | |
778 | rc = fallback_migrate_page(mapping, newpage, page, mode); | |
779 | ||
780 | if (rc != MIGRATEPAGE_SUCCESS) { | |
781 | newpage->mapping = NULL; | |
782 | } else { | |
783 | if (remap_swapcache) | |
784 | remove_migration_ptes(page, newpage); | |
785 | page->mapping = NULL; | |
786 | } | |
787 | ||
788 | unlock_page(newpage); | |
789 | ||
790 | return rc; | |
791 | } | |
792 | ||
793 | static int __unmap_and_move(struct page *page, struct page *newpage, | |
794 | int force, enum migrate_mode mode) | |
795 | { | |
796 | int rc = -EAGAIN; | |
797 | int remap_swapcache = 1; | |
798 | struct mem_cgroup *mem; | |
799 | struct anon_vma *anon_vma = NULL; | |
800 | ||
801 | if (!trylock_page(page)) { | |
802 | if (!force || mode == MIGRATE_ASYNC) | |
803 | goto out; | |
804 | ||
805 | /* | |
806 | * It's not safe for direct compaction to call lock_page. | |
807 | * For example, during page readahead pages are added locked | |
808 | * to the LRU. Later, when the IO completes the pages are | |
809 | * marked uptodate and unlocked. However, the queueing | |
810 | * could be merging multiple pages for one bio (e.g. | |
811 | * mpage_readpages). If an allocation happens for the | |
812 | * second or third page, the process can end up locking | |
813 | * the same page twice and deadlocking. Rather than | |
814 | * trying to be clever about what pages can be locked, | |
815 | * avoid the use of lock_page for direct compaction | |
816 | * altogether. | |
817 | */ | |
818 | if (current->flags & PF_MEMALLOC) | |
819 | goto out; | |
820 | ||
821 | lock_page(page); | |
822 | } | |
823 | ||
824 | /* charge against new page */ | |
825 | mem_cgroup_prepare_migration(page, newpage, &mem); | |
826 | ||
827 | if (PageWriteback(page)) { | |
828 | /* | |
829 | * Only in the case of a full synchronous migration is it | |
830 | * necessary to wait for PageWriteback. In the async case, | |
831 | * the retry loop is too short and in the sync-light case, | |
832 | * the overhead of stalling is too much | |
833 | */ | |
834 | if (mode != MIGRATE_SYNC) { | |
835 | rc = -EBUSY; | |
836 | goto uncharge; | |
837 | } | |
838 | if (!force) | |
839 | goto uncharge; | |
840 | wait_on_page_writeback(page); | |
841 | } | |
842 | /* | |
843 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, | |
844 | * we cannot notice that anon_vma is freed while we migrates a page. | |
845 | * This get_anon_vma() delays freeing anon_vma pointer until the end | |
846 | * of migration. File cache pages are no problem because of page_lock() | |
847 | * File Caches may use write_page() or lock_page() in migration, then, | |
848 | * just care Anon page here. | |
849 | */ | |
850 | if (PageAnon(page) && !PageKsm(page)) { | |
851 | /* | |
852 | * Only page_lock_anon_vma_read() understands the subtleties of | |
853 | * getting a hold on an anon_vma from outside one of its mms. | |
854 | */ | |
855 | anon_vma = page_get_anon_vma(page); | |
856 | if (anon_vma) { | |
857 | /* | |
858 | * Anon page | |
859 | */ | |
860 | } else if (PageSwapCache(page)) { | |
861 | /* | |
862 | * We cannot be sure that the anon_vma of an unmapped | |
863 | * swapcache page is safe to use because we don't | |
864 | * know in advance if the VMA that this page belonged | |
865 | * to still exists. If the VMA and others sharing the | |
866 | * data have been freed, then the anon_vma could | |
867 | * already be invalid. | |
868 | * | |
869 | * To avoid this possibility, swapcache pages get | |
870 | * migrated but are not remapped when migration | |
871 | * completes | |
872 | */ | |
873 | remap_swapcache = 0; | |
874 | } else { | |
875 | goto uncharge; | |
876 | } | |
877 | } | |
878 | ||
879 | if (unlikely(balloon_page_movable(page))) { | |
880 | /* | |
881 | * A ballooned page does not need any special attention from | |
882 | * physical to virtual reverse mapping procedures. | |
883 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
884 | * in order to avoid burning cycles at rmap level, and perform | |
885 | * the page migration right away (proteced by page lock). | |
886 | */ | |
887 | rc = balloon_page_migrate(newpage, page, mode); | |
888 | goto uncharge; | |
889 | } | |
890 | ||
891 | /* | |
892 | * Corner case handling: | |
893 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
894 | * and treated as swapcache but it has no rmap yet. | |
895 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
896 | * trigger a BUG. So handle it here. | |
897 | * 2. An orphaned page (see truncate_complete_page) might have | |
898 | * fs-private metadata. The page can be picked up due to memory | |
899 | * offlining. Everywhere else except page reclaim, the page is | |
900 | * invisible to the vm, so the page can not be migrated. So try to | |
901 | * free the metadata, so the page can be freed. | |
902 | */ | |
903 | if (!page->mapping) { | |
904 | VM_BUG_ON_PAGE(PageAnon(page), page); | |
905 | if (page_has_private(page)) { | |
906 | try_to_free_buffers(page); | |
907 | goto uncharge; | |
908 | } | |
909 | goto skip_unmap; | |
910 | } | |
911 | ||
912 | /* Establish migration ptes or remove ptes */ | |
913 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
914 | ||
915 | skip_unmap: | |
916 | if (!page_mapped(page)) | |
917 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); | |
918 | ||
919 | if (rc && remap_swapcache) | |
920 | remove_migration_ptes(page, page); | |
921 | ||
922 | /* Drop an anon_vma reference if we took one */ | |
923 | if (anon_vma) | |
924 | put_anon_vma(anon_vma); | |
925 | ||
926 | uncharge: | |
927 | mem_cgroup_end_migration(mem, page, newpage, | |
928 | (rc == MIGRATEPAGE_SUCCESS || | |
929 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
930 | unlock_page(page); | |
931 | out: | |
932 | return rc; | |
933 | } | |
934 | ||
935 | /* | |
936 | * Obtain the lock on page, remove all ptes and migrate the page | |
937 | * to the newly allocated page in newpage. | |
938 | */ | |
939 | static int unmap_and_move(new_page_t get_new_page, free_page_t put_new_page, | |
940 | unsigned long private, struct page *page, int force, | |
941 | enum migrate_mode mode) | |
942 | { | |
943 | int rc = 0; | |
944 | int *result = NULL; | |
945 | struct page *newpage = get_new_page(page, private, &result); | |
946 | ||
947 | if (!newpage) | |
948 | return -ENOMEM; | |
949 | ||
950 | if (page_count(page) == 1) { | |
951 | /* page was freed from under us. So we are done. */ | |
952 | goto out; | |
953 | } | |
954 | ||
955 | if (unlikely(PageTransHuge(page))) | |
956 | if (unlikely(split_huge_page(page))) | |
957 | goto out; | |
958 | ||
959 | rc = __unmap_and_move(page, newpage, force, mode); | |
960 | ||
961 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
962 | /* | |
963 | * A ballooned page has been migrated already. | |
964 | * Now, it's the time to wrap-up counters, | |
965 | * handle the page back to Buddy and return. | |
966 | */ | |
967 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
968 | page_is_file_cache(page)); | |
969 | balloon_page_free(page); | |
970 | return MIGRATEPAGE_SUCCESS; | |
971 | } | |
972 | out: | |
973 | if (rc != -EAGAIN) { | |
974 | /* | |
975 | * A page that has been migrated has all references | |
976 | * removed and will be freed. A page that has not been | |
977 | * migrated will have kepts its references and be | |
978 | * restored. | |
979 | */ | |
980 | list_del(&page->lru); | |
981 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
982 | page_is_file_cache(page)); | |
983 | putback_lru_page(page); | |
984 | } | |
985 | ||
986 | /* | |
987 | * If migration was not successful and there's a freeing callback, use | |
988 | * it. Otherwise, putback_lru_page() will drop the reference grabbed | |
989 | * during isolation. | |
990 | */ | |
991 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) { | |
992 | ClearPageSwapBacked(newpage); | |
993 | put_new_page(newpage, private); | |
994 | } else | |
995 | putback_lru_page(newpage); | |
996 | ||
997 | if (result) { | |
998 | if (rc) | |
999 | *result = rc; | |
1000 | else | |
1001 | *result = page_to_nid(newpage); | |
1002 | } | |
1003 | return rc; | |
1004 | } | |
1005 | ||
1006 | /* | |
1007 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1008 | * | |
1009 | * This function doesn't wait the completion of hugepage I/O | |
1010 | * because there is no race between I/O and migration for hugepage. | |
1011 | * Note that currently hugepage I/O occurs only in direct I/O | |
1012 | * where no lock is held and PG_writeback is irrelevant, | |
1013 | * and writeback status of all subpages are counted in the reference | |
1014 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1015 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1016 | * This means that when we try to migrate hugepage whose subpages are | |
1017 | * doing direct I/O, some references remain after try_to_unmap() and | |
1018 | * hugepage migration fails without data corruption. | |
1019 | * | |
1020 | * There is also no race when direct I/O is issued on the page under migration, | |
1021 | * because then pte is replaced with migration swap entry and direct I/O code | |
1022 | * will wait in the page fault for migration to complete. | |
1023 | */ | |
1024 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
1025 | free_page_t put_new_page, unsigned long private, | |
1026 | struct page *hpage, int force, | |
1027 | enum migrate_mode mode) | |
1028 | { | |
1029 | int rc = 0; | |
1030 | int *result = NULL; | |
1031 | struct page *new_hpage; | |
1032 | struct anon_vma *anon_vma = NULL; | |
1033 | ||
1034 | /* | |
1035 | * Movability of hugepages depends on architectures and hugepage size. | |
1036 | * This check is necessary because some callers of hugepage migration | |
1037 | * like soft offline and memory hotremove don't walk through page | |
1038 | * tables or check whether the hugepage is pmd-based or not before | |
1039 | * kicking migration. | |
1040 | */ | |
1041 | if (!hugepage_migration_supported(page_hstate(hpage))) { | |
1042 | putback_active_hugepage(hpage); | |
1043 | return -ENOSYS; | |
1044 | } | |
1045 | ||
1046 | new_hpage = get_new_page(hpage, private, &result); | |
1047 | if (!new_hpage) | |
1048 | return -ENOMEM; | |
1049 | ||
1050 | rc = -EAGAIN; | |
1051 | ||
1052 | if (!trylock_page(hpage)) { | |
1053 | if (!force || mode != MIGRATE_SYNC) | |
1054 | goto out; | |
1055 | lock_page(hpage); | |
1056 | } | |
1057 | ||
1058 | if (PageAnon(hpage)) | |
1059 | anon_vma = page_get_anon_vma(hpage); | |
1060 | ||
1061 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1062 | ||
1063 | if (!page_mapped(hpage)) | |
1064 | rc = move_to_new_page(new_hpage, hpage, 1, mode); | |
1065 | ||
1066 | if (rc != MIGRATEPAGE_SUCCESS) | |
1067 | remove_migration_ptes(hpage, hpage); | |
1068 | ||
1069 | if (anon_vma) | |
1070 | put_anon_vma(anon_vma); | |
1071 | ||
1072 | if (rc == MIGRATEPAGE_SUCCESS) | |
1073 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1074 | ||
1075 | unlock_page(hpage); | |
1076 | out: | |
1077 | if (rc != -EAGAIN) | |
1078 | putback_active_hugepage(hpage); | |
1079 | ||
1080 | /* | |
1081 | * If migration was not successful and there's a freeing callback, use | |
1082 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1083 | * isolation. | |
1084 | */ | |
1085 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) | |
1086 | put_new_page(new_hpage, private); | |
1087 | else | |
1088 | put_page(new_hpage); | |
1089 | ||
1090 | if (result) { | |
1091 | if (rc) | |
1092 | *result = rc; | |
1093 | else | |
1094 | *result = page_to_nid(new_hpage); | |
1095 | } | |
1096 | return rc; | |
1097 | } | |
1098 | ||
1099 | /* | |
1100 | * migrate_pages - migrate the pages specified in a list, to the free pages | |
1101 | * supplied as the target for the page migration | |
1102 | * | |
1103 | * @from: The list of pages to be migrated. | |
1104 | * @get_new_page: The function used to allocate free pages to be used | |
1105 | * as the target of the page migration. | |
1106 | * @put_new_page: The function used to free target pages if migration | |
1107 | * fails, or NULL if no special handling is necessary. | |
1108 | * @private: Private data to be passed on to get_new_page() | |
1109 | * @mode: The migration mode that specifies the constraints for | |
1110 | * page migration, if any. | |
1111 | * @reason: The reason for page migration. | |
1112 | * | |
1113 | * The function returns after 10 attempts or if no pages are movable any more | |
1114 | * because the list has become empty or no retryable pages exist any more. | |
1115 | * The caller should call putback_lru_pages() to return pages to the LRU | |
1116 | * or free list only if ret != 0. | |
1117 | * | |
1118 | * Returns the number of pages that were not migrated, or an error code. | |
1119 | */ | |
1120 | int migrate_pages(struct list_head *from, new_page_t get_new_page, | |
1121 | free_page_t put_new_page, unsigned long private, | |
1122 | enum migrate_mode mode, int reason) | |
1123 | { | |
1124 | int retry = 1; | |
1125 | int nr_failed = 0; | |
1126 | int nr_succeeded = 0; | |
1127 | int pass = 0; | |
1128 | struct page *page; | |
1129 | struct page *page2; | |
1130 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1131 | int rc; | |
1132 | ||
1133 | if (!swapwrite) | |
1134 | current->flags |= PF_SWAPWRITE; | |
1135 | ||
1136 | for(pass = 0; pass < 10 && retry; pass++) { | |
1137 | retry = 0; | |
1138 | ||
1139 | list_for_each_entry_safe(page, page2, from, lru) { | |
1140 | cond_resched(); | |
1141 | ||
1142 | if (PageHuge(page)) | |
1143 | rc = unmap_and_move_huge_page(get_new_page, | |
1144 | put_new_page, private, page, | |
1145 | pass > 2, mode); | |
1146 | else | |
1147 | rc = unmap_and_move(get_new_page, put_new_page, | |
1148 | private, page, pass > 2, mode); | |
1149 | ||
1150 | switch(rc) { | |
1151 | case -ENOMEM: | |
1152 | goto out; | |
1153 | case -EAGAIN: | |
1154 | retry++; | |
1155 | break; | |
1156 | case MIGRATEPAGE_SUCCESS: | |
1157 | nr_succeeded++; | |
1158 | break; | |
1159 | default: | |
1160 | /* | |
1161 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1162 | * unlike -EAGAIN case, the failed page is | |
1163 | * removed from migration page list and not | |
1164 | * retried in the next outer loop. | |
1165 | */ | |
1166 | nr_failed++; | |
1167 | break; | |
1168 | } | |
1169 | } | |
1170 | } | |
1171 | rc = nr_failed + retry; | |
1172 | out: | |
1173 | if (nr_succeeded) | |
1174 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1175 | if (nr_failed) | |
1176 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1177 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); | |
1178 | ||
1179 | if (!swapwrite) | |
1180 | current->flags &= ~PF_SWAPWRITE; | |
1181 | ||
1182 | return rc; | |
1183 | } | |
1184 | ||
1185 | #ifdef CONFIG_NUMA | |
1186 | /* | |
1187 | * Move a list of individual pages | |
1188 | */ | |
1189 | struct page_to_node { | |
1190 | unsigned long addr; | |
1191 | struct page *page; | |
1192 | int node; | |
1193 | int status; | |
1194 | }; | |
1195 | ||
1196 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1197 | int **result) | |
1198 | { | |
1199 | struct page_to_node *pm = (struct page_to_node *)private; | |
1200 | ||
1201 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1202 | pm++; | |
1203 | ||
1204 | if (pm->node == MAX_NUMNODES) | |
1205 | return NULL; | |
1206 | ||
1207 | *result = &pm->status; | |
1208 | ||
1209 | if (PageHuge(p)) | |
1210 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1211 | pm->node); | |
1212 | else | |
1213 | return alloc_pages_exact_node(pm->node, | |
1214 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); | |
1215 | } | |
1216 | ||
1217 | /* | |
1218 | * Move a set of pages as indicated in the pm array. The addr | |
1219 | * field must be set to the virtual address of the page to be moved | |
1220 | * and the node number must contain a valid target node. | |
1221 | * The pm array ends with node = MAX_NUMNODES. | |
1222 | */ | |
1223 | static int do_move_page_to_node_array(struct mm_struct *mm, | |
1224 | struct page_to_node *pm, | |
1225 | int migrate_all) | |
1226 | { | |
1227 | int err; | |
1228 | struct page_to_node *pp; | |
1229 | LIST_HEAD(pagelist); | |
1230 | ||
1231 | down_read(&mm->mmap_sem); | |
1232 | ||
1233 | /* | |
1234 | * Build a list of pages to migrate | |
1235 | */ | |
1236 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
1237 | struct vm_area_struct *vma; | |
1238 | struct page *page; | |
1239 | ||
1240 | err = -EFAULT; | |
1241 | vma = find_vma(mm, pp->addr); | |
1242 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) | |
1243 | goto set_status; | |
1244 | ||
1245 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); | |
1246 | ||
1247 | err = PTR_ERR(page); | |
1248 | if (IS_ERR(page)) | |
1249 | goto set_status; | |
1250 | ||
1251 | err = -ENOENT; | |
1252 | if (!page) | |
1253 | goto set_status; | |
1254 | ||
1255 | /* Use PageReserved to check for zero page */ | |
1256 | if (PageReserved(page)) | |
1257 | goto put_and_set; | |
1258 | ||
1259 | pp->page = page; | |
1260 | err = page_to_nid(page); | |
1261 | ||
1262 | if (err == pp->node) | |
1263 | /* | |
1264 | * Node already in the right place | |
1265 | */ | |
1266 | goto put_and_set; | |
1267 | ||
1268 | err = -EACCES; | |
1269 | if (page_mapcount(page) > 1 && | |
1270 | !migrate_all) | |
1271 | goto put_and_set; | |
1272 | ||
1273 | if (PageHuge(page)) { | |
1274 | isolate_huge_page(page, &pagelist); | |
1275 | goto put_and_set; | |
1276 | } | |
1277 | ||
1278 | err = isolate_lru_page(page); | |
1279 | if (!err) { | |
1280 | list_add_tail(&page->lru, &pagelist); | |
1281 | inc_zone_page_state(page, NR_ISOLATED_ANON + | |
1282 | page_is_file_cache(page)); | |
1283 | } | |
1284 | put_and_set: | |
1285 | /* | |
1286 | * Either remove the duplicate refcount from | |
1287 | * isolate_lru_page() or drop the page ref if it was | |
1288 | * not isolated. | |
1289 | */ | |
1290 | put_page(page); | |
1291 | set_status: | |
1292 | pp->status = err; | |
1293 | } | |
1294 | ||
1295 | err = 0; | |
1296 | if (!list_empty(&pagelist)) { | |
1297 | err = migrate_pages(&pagelist, new_page_node, NULL, | |
1298 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); | |
1299 | if (err) | |
1300 | putback_movable_pages(&pagelist); | |
1301 | } | |
1302 | ||
1303 | up_read(&mm->mmap_sem); | |
1304 | return err; | |
1305 | } | |
1306 | ||
1307 | /* | |
1308 | * Migrate an array of page address onto an array of nodes and fill | |
1309 | * the corresponding array of status. | |
1310 | */ | |
1311 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, | |
1312 | unsigned long nr_pages, | |
1313 | const void __user * __user *pages, | |
1314 | const int __user *nodes, | |
1315 | int __user *status, int flags) | |
1316 | { | |
1317 | struct page_to_node *pm; | |
1318 | unsigned long chunk_nr_pages; | |
1319 | unsigned long chunk_start; | |
1320 | int err; | |
1321 | ||
1322 | err = -ENOMEM; | |
1323 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1324 | if (!pm) | |
1325 | goto out; | |
1326 | ||
1327 | migrate_prep(); | |
1328 | ||
1329 | /* | |
1330 | * Store a chunk of page_to_node array in a page, | |
1331 | * but keep the last one as a marker | |
1332 | */ | |
1333 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; | |
1334 | ||
1335 | for (chunk_start = 0; | |
1336 | chunk_start < nr_pages; | |
1337 | chunk_start += chunk_nr_pages) { | |
1338 | int j; | |
1339 | ||
1340 | if (chunk_start + chunk_nr_pages > nr_pages) | |
1341 | chunk_nr_pages = nr_pages - chunk_start; | |
1342 | ||
1343 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1344 | for (j = 0; j < chunk_nr_pages; j++) { | |
1345 | const void __user *p; | |
1346 | int node; | |
1347 | ||
1348 | err = -EFAULT; | |
1349 | if (get_user(p, pages + j + chunk_start)) | |
1350 | goto out_pm; | |
1351 | pm[j].addr = (unsigned long) p; | |
1352 | ||
1353 | if (get_user(node, nodes + j + chunk_start)) | |
1354 | goto out_pm; | |
1355 | ||
1356 | err = -ENODEV; | |
1357 | if (node < 0 || node >= MAX_NUMNODES) | |
1358 | goto out_pm; | |
1359 | ||
1360 | if (!node_state(node, N_MEMORY)) | |
1361 | goto out_pm; | |
1362 | ||
1363 | err = -EACCES; | |
1364 | if (!node_isset(node, task_nodes)) | |
1365 | goto out_pm; | |
1366 | ||
1367 | pm[j].node = node; | |
1368 | } | |
1369 | ||
1370 | /* End marker for this chunk */ | |
1371 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1372 | ||
1373 | /* Migrate this chunk */ | |
1374 | err = do_move_page_to_node_array(mm, pm, | |
1375 | flags & MPOL_MF_MOVE_ALL); | |
1376 | if (err < 0) | |
1377 | goto out_pm; | |
1378 | ||
1379 | /* Return status information */ | |
1380 | for (j = 0; j < chunk_nr_pages; j++) | |
1381 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
1382 | err = -EFAULT; | |
1383 | goto out_pm; | |
1384 | } | |
1385 | } | |
1386 | err = 0; | |
1387 | ||
1388 | out_pm: | |
1389 | free_page((unsigned long)pm); | |
1390 | out: | |
1391 | return err; | |
1392 | } | |
1393 | ||
1394 | /* | |
1395 | * Determine the nodes of an array of pages and store it in an array of status. | |
1396 | */ | |
1397 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, | |
1398 | const void __user **pages, int *status) | |
1399 | { | |
1400 | unsigned long i; | |
1401 | ||
1402 | down_read(&mm->mmap_sem); | |
1403 | ||
1404 | for (i = 0; i < nr_pages; i++) { | |
1405 | unsigned long addr = (unsigned long)(*pages); | |
1406 | struct vm_area_struct *vma; | |
1407 | struct page *page; | |
1408 | int err = -EFAULT; | |
1409 | ||
1410 | vma = find_vma(mm, addr); | |
1411 | if (!vma || addr < vma->vm_start) | |
1412 | goto set_status; | |
1413 | ||
1414 | page = follow_page(vma, addr, 0); | |
1415 | ||
1416 | err = PTR_ERR(page); | |
1417 | if (IS_ERR(page)) | |
1418 | goto set_status; | |
1419 | ||
1420 | err = -ENOENT; | |
1421 | /* Use PageReserved to check for zero page */ | |
1422 | if (!page || PageReserved(page)) | |
1423 | goto set_status; | |
1424 | ||
1425 | err = page_to_nid(page); | |
1426 | set_status: | |
1427 | *status = err; | |
1428 | ||
1429 | pages++; | |
1430 | status++; | |
1431 | } | |
1432 | ||
1433 | up_read(&mm->mmap_sem); | |
1434 | } | |
1435 | ||
1436 | /* | |
1437 | * Determine the nodes of a user array of pages and store it in | |
1438 | * a user array of status. | |
1439 | */ | |
1440 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1441 | const void __user * __user *pages, | |
1442 | int __user *status) | |
1443 | { | |
1444 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1445 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1446 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
1447 | ||
1448 | while (nr_pages) { | |
1449 | unsigned long chunk_nr; | |
1450 | ||
1451 | chunk_nr = nr_pages; | |
1452 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1453 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1454 | ||
1455 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1456 | break; | |
1457 | ||
1458 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1459 | ||
1460 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) | |
1461 | break; | |
1462 | ||
1463 | pages += chunk_nr; | |
1464 | status += chunk_nr; | |
1465 | nr_pages -= chunk_nr; | |
1466 | } | |
1467 | return nr_pages ? -EFAULT : 0; | |
1468 | } | |
1469 | ||
1470 | /* | |
1471 | * Move a list of pages in the address space of the currently executing | |
1472 | * process. | |
1473 | */ | |
1474 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, | |
1475 | const void __user * __user *, pages, | |
1476 | const int __user *, nodes, | |
1477 | int __user *, status, int, flags) | |
1478 | { | |
1479 | const struct cred *cred = current_cred(), *tcred; | |
1480 | struct task_struct *task; | |
1481 | struct mm_struct *mm; | |
1482 | int err; | |
1483 | nodemask_t task_nodes; | |
1484 | ||
1485 | /* Check flags */ | |
1486 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1487 | return -EINVAL; | |
1488 | ||
1489 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1490 | return -EPERM; | |
1491 | ||
1492 | /* Find the mm_struct */ | |
1493 | rcu_read_lock(); | |
1494 | task = pid ? find_task_by_vpid(pid) : current; | |
1495 | if (!task) { | |
1496 | rcu_read_unlock(); | |
1497 | return -ESRCH; | |
1498 | } | |
1499 | get_task_struct(task); | |
1500 | ||
1501 | /* | |
1502 | * Check if this process has the right to modify the specified | |
1503 | * process. The right exists if the process has administrative | |
1504 | * capabilities, superuser privileges or the same | |
1505 | * userid as the target process. | |
1506 | */ | |
1507 | tcred = __task_cred(task); | |
1508 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && | |
1509 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
1510 | !capable(CAP_SYS_NICE)) { | |
1511 | rcu_read_unlock(); | |
1512 | err = -EPERM; | |
1513 | goto out; | |
1514 | } | |
1515 | rcu_read_unlock(); | |
1516 | ||
1517 | err = security_task_movememory(task); | |
1518 | if (err) | |
1519 | goto out; | |
1520 | ||
1521 | task_nodes = cpuset_mems_allowed(task); | |
1522 | mm = get_task_mm(task); | |
1523 | put_task_struct(task); | |
1524 | ||
1525 | if (!mm) | |
1526 | return -EINVAL; | |
1527 | ||
1528 | if (nodes) | |
1529 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1530 | nodes, status, flags); | |
1531 | else | |
1532 | err = do_pages_stat(mm, nr_pages, pages, status); | |
1533 | ||
1534 | mmput(mm); | |
1535 | return err; | |
1536 | ||
1537 | out: | |
1538 | put_task_struct(task); | |
1539 | return err; | |
1540 | } | |
1541 | ||
1542 | /* | |
1543 | * Call migration functions in the vma_ops that may prepare | |
1544 | * memory in a vm for migration. migration functions may perform | |
1545 | * the migration for vmas that do not have an underlying page struct. | |
1546 | */ | |
1547 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1548 | const nodemask_t *from, unsigned long flags) | |
1549 | { | |
1550 | struct vm_area_struct *vma; | |
1551 | int err = 0; | |
1552 | ||
1553 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { | |
1554 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1555 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1556 | if (err) | |
1557 | break; | |
1558 | } | |
1559 | } | |
1560 | return err; | |
1561 | } | |
1562 | ||
1563 | #ifdef CONFIG_NUMA_BALANCING | |
1564 | /* | |
1565 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1566 | * pages. Currently it only checks the watermarks which crude | |
1567 | */ | |
1568 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
1569 | unsigned long nr_migrate_pages) | |
1570 | { | |
1571 | int z; | |
1572 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1573 | struct zone *zone = pgdat->node_zones + z; | |
1574 | ||
1575 | if (!populated_zone(zone)) | |
1576 | continue; | |
1577 | ||
1578 | if (!zone_reclaimable(zone)) | |
1579 | continue; | |
1580 | ||
1581 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1582 | if (!zone_watermark_ok(zone, 0, | |
1583 | high_wmark_pages(zone) + | |
1584 | nr_migrate_pages, | |
1585 | 0, 0)) | |
1586 | continue; | |
1587 | return true; | |
1588 | } | |
1589 | return false; | |
1590 | } | |
1591 | ||
1592 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1593 | unsigned long data, | |
1594 | int **result) | |
1595 | { | |
1596 | int nid = (int) data; | |
1597 | struct page *newpage; | |
1598 | ||
1599 | newpage = alloc_pages_exact_node(nid, | |
1600 | (GFP_HIGHUSER_MOVABLE | | |
1601 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1602 | __GFP_NORETRY | __GFP_NOWARN) & | |
1603 | ~GFP_IOFS, 0); | |
1604 | ||
1605 | return newpage; | |
1606 | } | |
1607 | ||
1608 | /* | |
1609 | * page migration rate limiting control. | |
1610 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1611 | * window of time. Default here says do not migrate more than 1280M per second. | |
1612 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However | |
1613 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1614 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1615 | * throttle window closed. | |
1616 | */ | |
1617 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1618 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; | |
1619 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1620 | ||
1621 | /* Returns true if NUMA migration is currently rate limited */ | |
1622 | bool migrate_ratelimited(int node) | |
1623 | { | |
1624 | pg_data_t *pgdat = NODE_DATA(node); | |
1625 | ||
1626 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1627 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1628 | return false; | |
1629 | ||
1630 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1631 | return false; | |
1632 | ||
1633 | return true; | |
1634 | } | |
1635 | ||
1636 | /* Returns true if the node is migrate rate-limited after the update */ | |
1637 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, | |
1638 | unsigned long nr_pages) | |
1639 | { | |
1640 | /* | |
1641 | * Rate-limit the amount of data that is being migrated to a node. | |
1642 | * Optimal placement is no good if the memory bus is saturated and | |
1643 | * all the time is being spent migrating! | |
1644 | */ | |
1645 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1646 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1647 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1648 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1649 | msecs_to_jiffies(migrate_interval_millisecs); | |
1650 | spin_unlock(&pgdat->numabalancing_migrate_lock); | |
1651 | } | |
1652 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { | |
1653 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1654 | nr_pages); | |
1655 | return true; | |
1656 | } | |
1657 | ||
1658 | /* | |
1659 | * This is an unlocked non-atomic update so errors are possible. | |
1660 | * The consequences are failing to migrate when we potentiall should | |
1661 | * have which is not severe enough to warrant locking. If it is ever | |
1662 | * a problem, it can be converted to a per-cpu counter. | |
1663 | */ | |
1664 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1665 | return false; | |
1666 | } | |
1667 | ||
1668 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1669 | { | |
1670 | int page_lru; | |
1671 | ||
1672 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); | |
1673 | ||
1674 | /* Avoid migrating to a node that is nearly full */ | |
1675 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) | |
1676 | return 0; | |
1677 | ||
1678 | if (isolate_lru_page(page)) | |
1679 | return 0; | |
1680 | ||
1681 | /* | |
1682 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1683 | * check on page_count(), so we must do it here, now that the page | |
1684 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1685 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1686 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1687 | */ | |
1688 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1689 | putback_lru_page(page); | |
1690 | return 0; | |
1691 | } | |
1692 | ||
1693 | page_lru = page_is_file_cache(page); | |
1694 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1695 | hpage_nr_pages(page)); | |
1696 | ||
1697 | /* | |
1698 | * Isolating the page has taken another reference, so the | |
1699 | * caller's reference can be safely dropped without the page | |
1700 | * disappearing underneath us during migration. | |
1701 | */ | |
1702 | put_page(page); | |
1703 | return 1; | |
1704 | } | |
1705 | ||
1706 | bool pmd_trans_migrating(pmd_t pmd) | |
1707 | { | |
1708 | struct page *page = pmd_page(pmd); | |
1709 | return PageLocked(page); | |
1710 | } | |
1711 | ||
1712 | void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd) | |
1713 | { | |
1714 | struct page *page = pmd_page(*pmd); | |
1715 | wait_on_page_locked(page); | |
1716 | } | |
1717 | ||
1718 | /* | |
1719 | * Attempt to migrate a misplaced page to the specified destination | |
1720 | * node. Caller is expected to have an elevated reference count on | |
1721 | * the page that will be dropped by this function before returning. | |
1722 | */ | |
1723 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, | |
1724 | int node) | |
1725 | { | |
1726 | pg_data_t *pgdat = NODE_DATA(node); | |
1727 | int isolated; | |
1728 | int nr_remaining; | |
1729 | LIST_HEAD(migratepages); | |
1730 | ||
1731 | /* | |
1732 | * Don't migrate file pages that are mapped in multiple processes | |
1733 | * with execute permissions as they are probably shared libraries. | |
1734 | */ | |
1735 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && | |
1736 | (vma->vm_flags & VM_EXEC)) | |
1737 | goto out; | |
1738 | ||
1739 | /* | |
1740 | * Rate-limit the amount of data that is being migrated to a node. | |
1741 | * Optimal placement is no good if the memory bus is saturated and | |
1742 | * all the time is being spent migrating! | |
1743 | */ | |
1744 | if (numamigrate_update_ratelimit(pgdat, 1)) | |
1745 | goto out; | |
1746 | ||
1747 | isolated = numamigrate_isolate_page(pgdat, page); | |
1748 | if (!isolated) | |
1749 | goto out; | |
1750 | ||
1751 | list_add(&page->lru, &migratepages); | |
1752 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, | |
1753 | NULL, node, MIGRATE_ASYNC, | |
1754 | MR_NUMA_MISPLACED); | |
1755 | if (nr_remaining) { | |
1756 | if (!list_empty(&migratepages)) { | |
1757 | list_del(&page->lru); | |
1758 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1759 | page_is_file_cache(page)); | |
1760 | putback_lru_page(page); | |
1761 | } | |
1762 | isolated = 0; | |
1763 | } else | |
1764 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
1765 | BUG_ON(!list_empty(&migratepages)); | |
1766 | return isolated; | |
1767 | ||
1768 | out: | |
1769 | put_page(page); | |
1770 | return 0; | |
1771 | } | |
1772 | #endif /* CONFIG_NUMA_BALANCING */ | |
1773 | ||
1774 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) | |
1775 | /* | |
1776 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1777 | * before returning. | |
1778 | */ | |
1779 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, | |
1780 | struct vm_area_struct *vma, | |
1781 | pmd_t *pmd, pmd_t entry, | |
1782 | unsigned long address, | |
1783 | struct page *page, int node) | |
1784 | { | |
1785 | spinlock_t *ptl; | |
1786 | pg_data_t *pgdat = NODE_DATA(node); | |
1787 | int isolated = 0; | |
1788 | struct page *new_page = NULL; | |
1789 | struct mem_cgroup *memcg = NULL; | |
1790 | int page_lru = page_is_file_cache(page); | |
1791 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
1792 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
1793 | pmd_t orig_entry; | |
1794 | ||
1795 | /* | |
1796 | * Rate-limit the amount of data that is being migrated to a node. | |
1797 | * Optimal placement is no good if the memory bus is saturated and | |
1798 | * all the time is being spent migrating! | |
1799 | */ | |
1800 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) | |
1801 | goto out_dropref; | |
1802 | ||
1803 | new_page = alloc_pages_node(node, | |
1804 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT, | |
1805 | HPAGE_PMD_ORDER); | |
1806 | if (!new_page) | |
1807 | goto out_fail; | |
1808 | ||
1809 | isolated = numamigrate_isolate_page(pgdat, page); | |
1810 | if (!isolated) { | |
1811 | put_page(new_page); | |
1812 | goto out_fail; | |
1813 | } | |
1814 | ||
1815 | if (mm_tlb_flush_pending(mm)) | |
1816 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1817 | ||
1818 | /* Prepare a page as a migration target */ | |
1819 | __set_page_locked(new_page); | |
1820 | SetPageSwapBacked(new_page); | |
1821 | ||
1822 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1823 | new_page->mapping = page->mapping; | |
1824 | new_page->index = page->index; | |
1825 | migrate_page_copy(new_page, page); | |
1826 | WARN_ON(PageLRU(new_page)); | |
1827 | ||
1828 | /* Recheck the target PMD */ | |
1829 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1830 | ptl = pmd_lock(mm, pmd); | |
1831 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { | |
1832 | fail_putback: | |
1833 | spin_unlock(ptl); | |
1834 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
1835 | ||
1836 | /* Reverse changes made by migrate_page_copy() */ | |
1837 | if (TestClearPageActive(new_page)) | |
1838 | SetPageActive(page); | |
1839 | if (TestClearPageUnevictable(new_page)) | |
1840 | SetPageUnevictable(page); | |
1841 | mlock_migrate_page(page, new_page); | |
1842 | ||
1843 | unlock_page(new_page); | |
1844 | put_page(new_page); /* Free it */ | |
1845 | ||
1846 | /* Retake the callers reference and putback on LRU */ | |
1847 | get_page(page); | |
1848 | putback_lru_page(page); | |
1849 | mod_zone_page_state(page_zone(page), | |
1850 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
1851 | ||
1852 | goto out_unlock; | |
1853 | } | |
1854 | ||
1855 | /* | |
1856 | * Traditional migration needs to prepare the memcg charge | |
1857 | * transaction early to prevent the old page from being | |
1858 | * uncharged when installing migration entries. Here we can | |
1859 | * save the potential rollback and start the charge transfer | |
1860 | * only when migration is already known to end successfully. | |
1861 | */ | |
1862 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1863 | ||
1864 | orig_entry = *pmd; | |
1865 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
1866 | entry = pmd_mkhuge(entry); | |
1867 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1868 | ||
1869 | /* | |
1870 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1871 | * Any parallel GUP will either observe the old page blocking on the | |
1872 | * page lock, block on the page table lock or observe the new page. | |
1873 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1874 | * guarantee the copy is visible before the pagetable update. | |
1875 | */ | |
1876 | flush_cache_range(vma, mmun_start, mmun_end); | |
1877 | page_add_anon_rmap(new_page, vma, mmun_start); | |
1878 | pmdp_clear_flush(vma, mmun_start, pmd); | |
1879 | set_pmd_at(mm, mmun_start, pmd, entry); | |
1880 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1881 | update_mmu_cache_pmd(vma, address, &entry); | |
1882 | ||
1883 | if (page_count(page) != 2) { | |
1884 | set_pmd_at(mm, mmun_start, pmd, orig_entry); | |
1885 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1886 | update_mmu_cache_pmd(vma, address, &entry); | |
1887 | page_remove_rmap(new_page); | |
1888 | goto fail_putback; | |
1889 | } | |
1890 | ||
1891 | page_remove_rmap(page); | |
1892 | ||
1893 | /* | |
1894 | * Finish the charge transaction under the page table lock to | |
1895 | * prevent split_huge_page() from dividing up the charge | |
1896 | * before it's fully transferred to the new page. | |
1897 | */ | |
1898 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
1899 | spin_unlock(ptl); | |
1900 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
1901 | ||
1902 | /* Take an "isolate" reference and put new page on the LRU. */ | |
1903 | get_page(new_page); | |
1904 | putback_lru_page(new_page); | |
1905 | ||
1906 | unlock_page(new_page); | |
1907 | unlock_page(page); | |
1908 | put_page(page); /* Drop the rmap reference */ | |
1909 | put_page(page); /* Drop the LRU isolation reference */ | |
1910 | ||
1911 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1912 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1913 | ||
1914 | mod_zone_page_state(page_zone(page), | |
1915 | NR_ISOLATED_ANON + page_lru, | |
1916 | -HPAGE_PMD_NR); | |
1917 | return isolated; | |
1918 | ||
1919 | out_fail: | |
1920 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
1921 | out_dropref: | |
1922 | ptl = pmd_lock(mm, pmd); | |
1923 | if (pmd_same(*pmd, entry)) { | |
1924 | entry = pmd_mknonnuma(entry); | |
1925 | set_pmd_at(mm, mmun_start, pmd, entry); | |
1926 | update_mmu_cache_pmd(vma, address, &entry); | |
1927 | } | |
1928 | spin_unlock(ptl); | |
1929 | ||
1930 | out_unlock: | |
1931 | unlock_page(page); | |
1932 | put_page(page); | |
1933 | return 0; | |
1934 | } | |
1935 | #endif /* CONFIG_NUMA_BALANCING */ | |
1936 | ||
1937 | #endif /* CONFIG_NUMA */ |