]>
Commit | Line | Data |
---|---|---|
b20a3503 CL |
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> | |
cde53535 | 12 | * Christoph Lameter |
b20a3503 CL |
13 | */ |
14 | ||
15 | #include <linux/migrate.h> | |
b95f1b31 | 16 | #include <linux/export.h> |
b20a3503 | 17 | #include <linux/swap.h> |
0697212a | 18 | #include <linux/swapops.h> |
b20a3503 | 19 | #include <linux/pagemap.h> |
e23ca00b | 20 | #include <linux/buffer_head.h> |
b20a3503 | 21 | #include <linux/mm_inline.h> |
b488893a | 22 | #include <linux/nsproxy.h> |
b20a3503 | 23 | #include <linux/pagevec.h> |
e9995ef9 | 24 | #include <linux/ksm.h> |
b20a3503 CL |
25 | #include <linux/rmap.h> |
26 | #include <linux/topology.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/cpuset.h> | |
04e62a29 | 29 | #include <linux/writeback.h> |
742755a1 CL |
30 | #include <linux/mempolicy.h> |
31 | #include <linux/vmalloc.h> | |
86c3a764 | 32 | #include <linux/security.h> |
8a9f3ccd | 33 | #include <linux/memcontrol.h> |
4f5ca265 | 34 | #include <linux/syscalls.h> |
290408d4 | 35 | #include <linux/hugetlb.h> |
8e6ac7fa | 36 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 37 | #include <linux/gfp.h> |
b20a3503 | 38 | |
0d1836c3 MN |
39 | #include <asm/tlbflush.h> |
40 | ||
7b2a2d4a MG |
41 | #define CREATE_TRACE_POINTS |
42 | #include <trace/events/migrate.h> | |
43 | ||
b20a3503 CL |
44 | #include "internal.h" |
45 | ||
b20a3503 | 46 | /* |
742755a1 | 47 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
48 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
49 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
50 | */ |
51 | int migrate_prep(void) | |
52 | { | |
b20a3503 CL |
53 | /* |
54 | * Clear the LRU lists so pages can be isolated. | |
55 | * Note that pages may be moved off the LRU after we have | |
56 | * drained them. Those pages will fail to migrate like other | |
57 | * pages that may be busy. | |
58 | */ | |
59 | lru_add_drain_all(); | |
60 | ||
61 | return 0; | |
62 | } | |
63 | ||
748446bb MG |
64 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
65 | int migrate_prep_local(void) | |
66 | { | |
67 | lru_add_drain(); | |
68 | ||
69 | return 0; | |
70 | } | |
71 | ||
b20a3503 | 72 | /* |
894bc310 LS |
73 | * Add isolated pages on the list back to the LRU under page lock |
74 | * to avoid leaking evictable pages back onto unevictable list. | |
b20a3503 | 75 | */ |
e13861d8 | 76 | void putback_lru_pages(struct list_head *l) |
b20a3503 CL |
77 | { |
78 | struct page *page; | |
79 | struct page *page2; | |
b20a3503 CL |
80 | |
81 | list_for_each_entry_safe(page, page2, l, lru) { | |
e24f0b8f | 82 | list_del(&page->lru); |
a731286d | 83 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 84 | page_is_file_cache(page)); |
894bc310 | 85 | putback_lru_page(page); |
b20a3503 | 86 | } |
b20a3503 CL |
87 | } |
88 | ||
0697212a CL |
89 | /* |
90 | * Restore a potential migration pte to a working pte entry | |
91 | */ | |
e9995ef9 HD |
92 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, |
93 | unsigned long addr, void *old) | |
0697212a CL |
94 | { |
95 | struct mm_struct *mm = vma->vm_mm; | |
96 | swp_entry_t entry; | |
97 | pgd_t *pgd; | |
98 | pud_t *pud; | |
99 | pmd_t *pmd; | |
100 | pte_t *ptep, pte; | |
101 | spinlock_t *ptl; | |
102 | ||
290408d4 NH |
103 | if (unlikely(PageHuge(new))) { |
104 | ptep = huge_pte_offset(mm, addr); | |
105 | if (!ptep) | |
106 | goto out; | |
107 | ptl = &mm->page_table_lock; | |
108 | } else { | |
109 | pgd = pgd_offset(mm, addr); | |
110 | if (!pgd_present(*pgd)) | |
111 | goto out; | |
0697212a | 112 | |
290408d4 NH |
113 | pud = pud_offset(pgd, addr); |
114 | if (!pud_present(*pud)) | |
115 | goto out; | |
0697212a | 116 | |
290408d4 | 117 | pmd = pmd_offset(pud, addr); |
500d65d4 AA |
118 | if (pmd_trans_huge(*pmd)) |
119 | goto out; | |
290408d4 NH |
120 | if (!pmd_present(*pmd)) |
121 | goto out; | |
0697212a | 122 | |
290408d4 | 123 | ptep = pte_offset_map(pmd, addr); |
0697212a | 124 | |
486cf46f HD |
125 | /* |
126 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
127 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
128 | */ | |
290408d4 NH |
129 | |
130 | ptl = pte_lockptr(mm, pmd); | |
131 | } | |
0697212a | 132 | |
0697212a CL |
133 | spin_lock(ptl); |
134 | pte = *ptep; | |
135 | if (!is_swap_pte(pte)) | |
e9995ef9 | 136 | goto unlock; |
0697212a CL |
137 | |
138 | entry = pte_to_swp_entry(pte); | |
139 | ||
e9995ef9 HD |
140 | if (!is_migration_entry(entry) || |
141 | migration_entry_to_page(entry) != old) | |
142 | goto unlock; | |
0697212a | 143 | |
0697212a CL |
144 | get_page(new); |
145 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
146 | if (is_write_migration_entry(entry)) | |
147 | pte = pte_mkwrite(pte); | |
3ef8fd7f | 148 | #ifdef CONFIG_HUGETLB_PAGE |
290408d4 NH |
149 | if (PageHuge(new)) |
150 | pte = pte_mkhuge(pte); | |
3ef8fd7f | 151 | #endif |
97ee0524 | 152 | flush_cache_page(vma, addr, pte_pfn(pte)); |
0697212a | 153 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 154 | |
290408d4 NH |
155 | if (PageHuge(new)) { |
156 | if (PageAnon(new)) | |
157 | hugepage_add_anon_rmap(new, vma, addr); | |
158 | else | |
159 | page_dup_rmap(new); | |
160 | } else if (PageAnon(new)) | |
04e62a29 CL |
161 | page_add_anon_rmap(new, vma, addr); |
162 | else | |
163 | page_add_file_rmap(new); | |
164 | ||
165 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 166 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 167 | unlock: |
0697212a | 168 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
169 | out: |
170 | return SWAP_AGAIN; | |
0697212a CL |
171 | } |
172 | ||
04e62a29 CL |
173 | /* |
174 | * Get rid of all migration entries and replace them by | |
175 | * references to the indicated page. | |
176 | */ | |
177 | static void remove_migration_ptes(struct page *old, struct page *new) | |
178 | { | |
e9995ef9 | 179 | rmap_walk(new, remove_migration_pte, old); |
04e62a29 CL |
180 | } |
181 | ||
0697212a CL |
182 | /* |
183 | * Something used the pte of a page under migration. We need to | |
184 | * get to the page and wait until migration is finished. | |
185 | * When we return from this function the fault will be retried. | |
0697212a CL |
186 | */ |
187 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
188 | unsigned long address) | |
189 | { | |
190 | pte_t *ptep, pte; | |
191 | spinlock_t *ptl; | |
192 | swp_entry_t entry; | |
193 | struct page *page; | |
194 | ||
195 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
196 | pte = *ptep; | |
197 | if (!is_swap_pte(pte)) | |
198 | goto out; | |
199 | ||
200 | entry = pte_to_swp_entry(pte); | |
201 | if (!is_migration_entry(entry)) | |
202 | goto out; | |
203 | ||
204 | page = migration_entry_to_page(entry); | |
205 | ||
e286781d NP |
206 | /* |
207 | * Once radix-tree replacement of page migration started, page_count | |
208 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
209 | * against a page without get_page(). | |
210 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
211 | * will occur again. | |
212 | */ | |
213 | if (!get_page_unless_zero(page)) | |
214 | goto out; | |
0697212a CL |
215 | pte_unmap_unlock(ptep, ptl); |
216 | wait_on_page_locked(page); | |
217 | put_page(page); | |
218 | return; | |
219 | out: | |
220 | pte_unmap_unlock(ptep, ptl); | |
221 | } | |
222 | ||
b969c4ab MG |
223 | #ifdef CONFIG_BLOCK |
224 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
225 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
226 | enum migrate_mode mode) | |
b969c4ab MG |
227 | { |
228 | struct buffer_head *bh = head; | |
229 | ||
230 | /* Simple case, sync compaction */ | |
a6bc32b8 | 231 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
232 | do { |
233 | get_bh(bh); | |
234 | lock_buffer(bh); | |
235 | bh = bh->b_this_page; | |
236 | ||
237 | } while (bh != head); | |
238 | ||
239 | return true; | |
240 | } | |
241 | ||
242 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
243 | do { | |
244 | get_bh(bh); | |
245 | if (!trylock_buffer(bh)) { | |
246 | /* | |
247 | * We failed to lock the buffer and cannot stall in | |
248 | * async migration. Release the taken locks | |
249 | */ | |
250 | struct buffer_head *failed_bh = bh; | |
251 | put_bh(failed_bh); | |
252 | bh = head; | |
253 | while (bh != failed_bh) { | |
254 | unlock_buffer(bh); | |
255 | put_bh(bh); | |
256 | bh = bh->b_this_page; | |
257 | } | |
258 | return false; | |
259 | } | |
260 | ||
261 | bh = bh->b_this_page; | |
262 | } while (bh != head); | |
263 | return true; | |
264 | } | |
265 | #else | |
266 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 267 | enum migrate_mode mode) |
b969c4ab MG |
268 | { |
269 | return true; | |
270 | } | |
271 | #endif /* CONFIG_BLOCK */ | |
272 | ||
b20a3503 | 273 | /* |
c3fcf8a5 | 274 | * Replace the page in the mapping. |
5b5c7120 CL |
275 | * |
276 | * The number of remaining references must be: | |
277 | * 1 for anonymous pages without a mapping | |
278 | * 2 for pages with a mapping | |
266cf658 | 279 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 280 | */ |
2d1db3b1 | 281 | static int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 282 | struct page *newpage, struct page *page, |
a6bc32b8 | 283 | struct buffer_head *head, enum migrate_mode mode) |
b20a3503 | 284 | { |
7039e1db | 285 | int expected_count = 0; |
7cf9c2c7 | 286 | void **pslot; |
b20a3503 | 287 | |
6c5240ae | 288 | if (!mapping) { |
0e8c7d0f | 289 | /* Anonymous page without mapping */ |
6c5240ae CL |
290 | if (page_count(page) != 1) |
291 | return -EAGAIN; | |
292 | return 0; | |
293 | } | |
294 | ||
19fd6231 | 295 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 296 | |
7cf9c2c7 NP |
297 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
298 | page_index(page)); | |
b20a3503 | 299 | |
edcf4748 | 300 | expected_count = 2 + page_has_private(page); |
e286781d | 301 | if (page_count(page) != expected_count || |
29c1f677 | 302 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 303 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 304 | return -EAGAIN; |
b20a3503 CL |
305 | } |
306 | ||
e286781d | 307 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 308 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
309 | return -EAGAIN; |
310 | } | |
311 | ||
b969c4ab MG |
312 | /* |
313 | * In the async migration case of moving a page with buffers, lock the | |
314 | * buffers using trylock before the mapping is moved. If the mapping | |
315 | * was moved, we later failed to lock the buffers and could not move | |
316 | * the mapping back due to an elevated page count, we would have to | |
317 | * block waiting on other references to be dropped. | |
318 | */ | |
a6bc32b8 MG |
319 | if (mode == MIGRATE_ASYNC && head && |
320 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
321 | page_unfreeze_refs(page, expected_count); |
322 | spin_unlock_irq(&mapping->tree_lock); | |
323 | return -EAGAIN; | |
324 | } | |
325 | ||
b20a3503 CL |
326 | /* |
327 | * Now we know that no one else is looking at the page. | |
b20a3503 | 328 | */ |
7cf9c2c7 | 329 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
330 | if (PageSwapCache(page)) { |
331 | SetPageSwapCache(newpage); | |
332 | set_page_private(newpage, page_private(page)); | |
333 | } | |
334 | ||
7cf9c2c7 NP |
335 | radix_tree_replace_slot(pslot, newpage); |
336 | ||
337 | /* | |
937a94c9 JG |
338 | * Drop cache reference from old page by unfreezing |
339 | * to one less reference. | |
7cf9c2c7 NP |
340 | * We know this isn't the last reference. |
341 | */ | |
937a94c9 | 342 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 343 | |
0e8c7d0f CL |
344 | /* |
345 | * If moved to a different zone then also account | |
346 | * the page for that zone. Other VM counters will be | |
347 | * taken care of when we establish references to the | |
348 | * new page and drop references to the old page. | |
349 | * | |
350 | * Note that anonymous pages are accounted for | |
351 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
352 | * are mapped to swap space. | |
353 | */ | |
354 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
355 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 356 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
357 | __dec_zone_page_state(page, NR_SHMEM); |
358 | __inc_zone_page_state(newpage, NR_SHMEM); | |
359 | } | |
19fd6231 | 360 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 CL |
361 | |
362 | return 0; | |
363 | } | |
b20a3503 | 364 | |
290408d4 NH |
365 | /* |
366 | * The expected number of remaining references is the same as that | |
367 | * of migrate_page_move_mapping(). | |
368 | */ | |
369 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
370 | struct page *newpage, struct page *page) | |
371 | { | |
372 | int expected_count; | |
373 | void **pslot; | |
374 | ||
375 | if (!mapping) { | |
376 | if (page_count(page) != 1) | |
377 | return -EAGAIN; | |
378 | return 0; | |
379 | } | |
380 | ||
381 | spin_lock_irq(&mapping->tree_lock); | |
382 | ||
383 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
384 | page_index(page)); | |
385 | ||
386 | expected_count = 2 + page_has_private(page); | |
387 | if (page_count(page) != expected_count || | |
29c1f677 | 388 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
389 | spin_unlock_irq(&mapping->tree_lock); |
390 | return -EAGAIN; | |
391 | } | |
392 | ||
393 | if (!page_freeze_refs(page, expected_count)) { | |
394 | spin_unlock_irq(&mapping->tree_lock); | |
395 | return -EAGAIN; | |
396 | } | |
397 | ||
398 | get_page(newpage); | |
399 | ||
400 | radix_tree_replace_slot(pslot, newpage); | |
401 | ||
937a94c9 | 402 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
403 | |
404 | spin_unlock_irq(&mapping->tree_lock); | |
405 | return 0; | |
406 | } | |
407 | ||
b20a3503 CL |
408 | /* |
409 | * Copy the page to its new location | |
410 | */ | |
290408d4 | 411 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 412 | { |
290408d4 NH |
413 | if (PageHuge(page)) |
414 | copy_huge_page(newpage, page); | |
415 | else | |
416 | copy_highpage(newpage, page); | |
b20a3503 CL |
417 | |
418 | if (PageError(page)) | |
419 | SetPageError(newpage); | |
420 | if (PageReferenced(page)) | |
421 | SetPageReferenced(newpage); | |
422 | if (PageUptodate(page)) | |
423 | SetPageUptodate(newpage); | |
894bc310 LS |
424 | if (TestClearPageActive(page)) { |
425 | VM_BUG_ON(PageUnevictable(page)); | |
b20a3503 | 426 | SetPageActive(newpage); |
418b27ef LS |
427 | } else if (TestClearPageUnevictable(page)) |
428 | SetPageUnevictable(newpage); | |
b20a3503 CL |
429 | if (PageChecked(page)) |
430 | SetPageChecked(newpage); | |
431 | if (PageMappedToDisk(page)) | |
432 | SetPageMappedToDisk(newpage); | |
433 | ||
434 | if (PageDirty(page)) { | |
435 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
436 | /* |
437 | * Want to mark the page and the radix tree as dirty, and | |
438 | * redo the accounting that clear_page_dirty_for_io undid, | |
439 | * but we can't use set_page_dirty because that function | |
440 | * is actually a signal that all of the page has become dirty. | |
25985edc | 441 | * Whereas only part of our page may be dirty. |
3a902c5f | 442 | */ |
752dc185 HD |
443 | if (PageSwapBacked(page)) |
444 | SetPageDirty(newpage); | |
445 | else | |
446 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
447 | } |
448 | ||
b291f000 | 449 | mlock_migrate_page(newpage, page); |
e9995ef9 | 450 | ksm_migrate_page(newpage, page); |
b291f000 | 451 | |
b20a3503 | 452 | ClearPageSwapCache(page); |
b20a3503 CL |
453 | ClearPagePrivate(page); |
454 | set_page_private(page, 0); | |
b20a3503 CL |
455 | |
456 | /* | |
457 | * If any waiters have accumulated on the new page then | |
458 | * wake them up. | |
459 | */ | |
460 | if (PageWriteback(newpage)) | |
461 | end_page_writeback(newpage); | |
462 | } | |
b20a3503 | 463 | |
1d8b85cc CL |
464 | /************************************************************ |
465 | * Migration functions | |
466 | ***********************************************************/ | |
467 | ||
468 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
469 | int fail_migrate_page(struct address_space *mapping, |
470 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
471 | { |
472 | return -EIO; | |
473 | } | |
474 | EXPORT_SYMBOL(fail_migrate_page); | |
475 | ||
b20a3503 CL |
476 | /* |
477 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 478 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
479 | * |
480 | * Pages are locked upon entry and exit. | |
481 | */ | |
2d1db3b1 | 482 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
483 | struct page *newpage, struct page *page, |
484 | enum migrate_mode mode) | |
b20a3503 CL |
485 | { |
486 | int rc; | |
487 | ||
488 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
489 | ||
a6bc32b8 | 490 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode); |
b20a3503 CL |
491 | |
492 | if (rc) | |
493 | return rc; | |
494 | ||
495 | migrate_page_copy(newpage, page); | |
b20a3503 CL |
496 | return 0; |
497 | } | |
498 | EXPORT_SYMBOL(migrate_page); | |
499 | ||
9361401e | 500 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
501 | /* |
502 | * Migration function for pages with buffers. This function can only be used | |
503 | * if the underlying filesystem guarantees that no other references to "page" | |
504 | * exist. | |
505 | */ | |
2d1db3b1 | 506 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 507 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 508 | { |
1d8b85cc CL |
509 | struct buffer_head *bh, *head; |
510 | int rc; | |
511 | ||
1d8b85cc | 512 | if (!page_has_buffers(page)) |
a6bc32b8 | 513 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
514 | |
515 | head = page_buffers(page); | |
516 | ||
a6bc32b8 | 517 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode); |
1d8b85cc CL |
518 | |
519 | if (rc) | |
520 | return rc; | |
521 | ||
b969c4ab MG |
522 | /* |
523 | * In the async case, migrate_page_move_mapping locked the buffers | |
524 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
525 | * need to be locked now | |
526 | */ | |
a6bc32b8 MG |
527 | if (mode != MIGRATE_ASYNC) |
528 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
529 | |
530 | ClearPagePrivate(page); | |
531 | set_page_private(newpage, page_private(page)); | |
532 | set_page_private(page, 0); | |
533 | put_page(page); | |
534 | get_page(newpage); | |
535 | ||
536 | bh = head; | |
537 | do { | |
538 | set_bh_page(bh, newpage, bh_offset(bh)); | |
539 | bh = bh->b_this_page; | |
540 | ||
541 | } while (bh != head); | |
542 | ||
543 | SetPagePrivate(newpage); | |
544 | ||
545 | migrate_page_copy(newpage, page); | |
546 | ||
547 | bh = head; | |
548 | do { | |
549 | unlock_buffer(bh); | |
550 | put_bh(bh); | |
551 | bh = bh->b_this_page; | |
552 | ||
553 | } while (bh != head); | |
554 | ||
555 | return 0; | |
556 | } | |
557 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 558 | #endif |
1d8b85cc | 559 | |
04e62a29 CL |
560 | /* |
561 | * Writeback a page to clean the dirty state | |
562 | */ | |
563 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 564 | { |
04e62a29 CL |
565 | struct writeback_control wbc = { |
566 | .sync_mode = WB_SYNC_NONE, | |
567 | .nr_to_write = 1, | |
568 | .range_start = 0, | |
569 | .range_end = LLONG_MAX, | |
04e62a29 CL |
570 | .for_reclaim = 1 |
571 | }; | |
572 | int rc; | |
573 | ||
574 | if (!mapping->a_ops->writepage) | |
575 | /* No write method for the address space */ | |
576 | return -EINVAL; | |
577 | ||
578 | if (!clear_page_dirty_for_io(page)) | |
579 | /* Someone else already triggered a write */ | |
580 | return -EAGAIN; | |
581 | ||
8351a6e4 | 582 | /* |
04e62a29 CL |
583 | * A dirty page may imply that the underlying filesystem has |
584 | * the page on some queue. So the page must be clean for | |
585 | * migration. Writeout may mean we loose the lock and the | |
586 | * page state is no longer what we checked for earlier. | |
587 | * At this point we know that the migration attempt cannot | |
588 | * be successful. | |
8351a6e4 | 589 | */ |
04e62a29 | 590 | remove_migration_ptes(page, page); |
8351a6e4 | 591 | |
04e62a29 | 592 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 593 | |
04e62a29 CL |
594 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
595 | /* unlocked. Relock */ | |
596 | lock_page(page); | |
597 | ||
bda8550d | 598 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
599 | } |
600 | ||
601 | /* | |
602 | * Default handling if a filesystem does not provide a migration function. | |
603 | */ | |
604 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 605 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 606 | { |
b969c4ab | 607 | if (PageDirty(page)) { |
a6bc32b8 MG |
608 | /* Only writeback pages in full synchronous migration */ |
609 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 610 | return -EBUSY; |
04e62a29 | 611 | return writeout(mapping, page); |
b969c4ab | 612 | } |
8351a6e4 CL |
613 | |
614 | /* | |
615 | * Buffers may be managed in a filesystem specific way. | |
616 | * We must have no buffers or drop them. | |
617 | */ | |
266cf658 | 618 | if (page_has_private(page) && |
8351a6e4 CL |
619 | !try_to_release_page(page, GFP_KERNEL)) |
620 | return -EAGAIN; | |
621 | ||
a6bc32b8 | 622 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
623 | } |
624 | ||
e24f0b8f CL |
625 | /* |
626 | * Move a page to a newly allocated page | |
627 | * The page is locked and all ptes have been successfully removed. | |
628 | * | |
629 | * The new page will have replaced the old page if this function | |
630 | * is successful. | |
894bc310 LS |
631 | * |
632 | * Return value: | |
633 | * < 0 - error code | |
634 | * == 0 - success | |
e24f0b8f | 635 | */ |
3fe2011f | 636 | static int move_to_new_page(struct page *newpage, struct page *page, |
a6bc32b8 | 637 | int remap_swapcache, enum migrate_mode mode) |
e24f0b8f CL |
638 | { |
639 | struct address_space *mapping; | |
640 | int rc; | |
641 | ||
642 | /* | |
643 | * Block others from accessing the page when we get around to | |
644 | * establishing additional references. We are the only one | |
645 | * holding a reference to the new page at this point. | |
646 | */ | |
529ae9aa | 647 | if (!trylock_page(newpage)) |
e24f0b8f CL |
648 | BUG(); |
649 | ||
650 | /* Prepare mapping for the new page.*/ | |
651 | newpage->index = page->index; | |
652 | newpage->mapping = page->mapping; | |
b2e18538 RR |
653 | if (PageSwapBacked(page)) |
654 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
655 | |
656 | mapping = page_mapping(page); | |
657 | if (!mapping) | |
a6bc32b8 | 658 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 659 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 660 | /* |
b969c4ab MG |
661 | * Most pages have a mapping and most filesystems provide a |
662 | * migratepage callback. Anonymous pages are part of swap | |
663 | * space which also has its own migratepage callback. This | |
664 | * is the most common path for page migration. | |
e24f0b8f | 665 | */ |
b969c4ab | 666 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 667 | newpage, page, mode); |
b969c4ab | 668 | else |
a6bc32b8 | 669 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 670 | |
3fe2011f | 671 | if (rc) { |
e24f0b8f | 672 | newpage->mapping = NULL; |
3fe2011f MG |
673 | } else { |
674 | if (remap_swapcache) | |
675 | remove_migration_ptes(page, newpage); | |
35512eca | 676 | page->mapping = NULL; |
3fe2011f | 677 | } |
e24f0b8f CL |
678 | |
679 | unlock_page(newpage); | |
680 | ||
681 | return rc; | |
682 | } | |
683 | ||
0dabec93 | 684 | static int __unmap_and_move(struct page *page, struct page *newpage, |
a6bc32b8 | 685 | int force, bool offlining, enum migrate_mode mode) |
e24f0b8f | 686 | { |
0dabec93 | 687 | int rc = -EAGAIN; |
3fe2011f | 688 | int remap_swapcache = 1; |
56039efa | 689 | struct mem_cgroup *mem; |
3f6c8272 | 690 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 691 | |
529ae9aa | 692 | if (!trylock_page(page)) { |
a6bc32b8 | 693 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 694 | goto out; |
3e7d3449 MG |
695 | |
696 | /* | |
697 | * It's not safe for direct compaction to call lock_page. | |
698 | * For example, during page readahead pages are added locked | |
699 | * to the LRU. Later, when the IO completes the pages are | |
700 | * marked uptodate and unlocked. However, the queueing | |
701 | * could be merging multiple pages for one bio (e.g. | |
702 | * mpage_readpages). If an allocation happens for the | |
703 | * second or third page, the process can end up locking | |
704 | * the same page twice and deadlocking. Rather than | |
705 | * trying to be clever about what pages can be locked, | |
706 | * avoid the use of lock_page for direct compaction | |
707 | * altogether. | |
708 | */ | |
709 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 710 | goto out; |
3e7d3449 | 711 | |
e24f0b8f CL |
712 | lock_page(page); |
713 | } | |
714 | ||
62b61f61 HD |
715 | /* |
716 | * Only memory hotplug's offline_pages() caller has locked out KSM, | |
717 | * and can safely migrate a KSM page. The other cases have skipped | |
718 | * PageKsm along with PageReserved - but it is only now when we have | |
719 | * the page lock that we can be certain it will not go KSM beneath us | |
720 | * (KSM will not upgrade a page from PageAnon to PageKsm when it sees | |
721 | * its pagecount raised, but only here do we take the page lock which | |
722 | * serializes that). | |
723 | */ | |
724 | if (PageKsm(page) && !offlining) { | |
725 | rc = -EBUSY; | |
726 | goto unlock; | |
727 | } | |
728 | ||
01b1ae63 | 729 | /* charge against new page */ |
0030f535 | 730 | mem_cgroup_prepare_migration(page, newpage, &mem); |
01b1ae63 | 731 | |
e24f0b8f | 732 | if (PageWriteback(page)) { |
11bc82d6 | 733 | /* |
a6bc32b8 MG |
734 | * Only in the case of a full syncronous migration is it |
735 | * necessary to wait for PageWriteback. In the async case, | |
736 | * the retry loop is too short and in the sync-light case, | |
737 | * the overhead of stalling is too much | |
11bc82d6 | 738 | */ |
a6bc32b8 | 739 | if (mode != MIGRATE_SYNC) { |
11bc82d6 AA |
740 | rc = -EBUSY; |
741 | goto uncharge; | |
742 | } | |
743 | if (!force) | |
01b1ae63 | 744 | goto uncharge; |
e24f0b8f CL |
745 | wait_on_page_writeback(page); |
746 | } | |
e24f0b8f | 747 | /* |
dc386d4d KH |
748 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
749 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 750 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 751 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
752 | * File Caches may use write_page() or lock_page() in migration, then, |
753 | * just care Anon page here. | |
dc386d4d | 754 | */ |
989f89c5 | 755 | if (PageAnon(page)) { |
1ce82b69 HD |
756 | /* |
757 | * Only page_lock_anon_vma() understands the subtleties of | |
758 | * getting a hold on an anon_vma from outside one of its mms. | |
759 | */ | |
746b18d4 | 760 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
761 | if (anon_vma) { |
762 | /* | |
746b18d4 | 763 | * Anon page |
1ce82b69 | 764 | */ |
1ce82b69 | 765 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
766 | /* |
767 | * We cannot be sure that the anon_vma of an unmapped | |
768 | * swapcache page is safe to use because we don't | |
769 | * know in advance if the VMA that this page belonged | |
770 | * to still exists. If the VMA and others sharing the | |
771 | * data have been freed, then the anon_vma could | |
772 | * already be invalid. | |
773 | * | |
774 | * To avoid this possibility, swapcache pages get | |
775 | * migrated but are not remapped when migration | |
776 | * completes | |
777 | */ | |
778 | remap_swapcache = 0; | |
779 | } else { | |
1ce82b69 | 780 | goto uncharge; |
3fe2011f | 781 | } |
989f89c5 | 782 | } |
62e1c553 | 783 | |
dc386d4d | 784 | /* |
62e1c553 SL |
785 | * Corner case handling: |
786 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
787 | * and treated as swapcache but it has no rmap yet. | |
788 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
789 | * trigger a BUG. So handle it here. | |
790 | * 2. An orphaned page (see truncate_complete_page) might have | |
791 | * fs-private metadata. The page can be picked up due to memory | |
792 | * offlining. Everywhere else except page reclaim, the page is | |
793 | * invisible to the vm, so the page can not be migrated. So try to | |
794 | * free the metadata, so the page can be freed. | |
e24f0b8f | 795 | */ |
62e1c553 | 796 | if (!page->mapping) { |
1ce82b69 HD |
797 | VM_BUG_ON(PageAnon(page)); |
798 | if (page_has_private(page)) { | |
62e1c553 | 799 | try_to_free_buffers(page); |
1ce82b69 | 800 | goto uncharge; |
62e1c553 | 801 | } |
abfc3488 | 802 | goto skip_unmap; |
62e1c553 SL |
803 | } |
804 | ||
dc386d4d | 805 | /* Establish migration ptes or remove ptes */ |
14fa31b8 | 806 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
dc386d4d | 807 | |
abfc3488 | 808 | skip_unmap: |
e6a1530d | 809 | if (!page_mapped(page)) |
a6bc32b8 | 810 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); |
e24f0b8f | 811 | |
3fe2011f | 812 | if (rc && remap_swapcache) |
e24f0b8f | 813 | remove_migration_ptes(page, page); |
3f6c8272 MG |
814 | |
815 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 816 | if (anon_vma) |
9e60109f | 817 | put_anon_vma(anon_vma); |
3f6c8272 | 818 | |
01b1ae63 | 819 | uncharge: |
0030f535 | 820 | mem_cgroup_end_migration(mem, page, newpage, rc == 0); |
e24f0b8f CL |
821 | unlock: |
822 | unlock_page(page); | |
0dabec93 MK |
823 | out: |
824 | return rc; | |
825 | } | |
95a402c3 | 826 | |
0dabec93 MK |
827 | /* |
828 | * Obtain the lock on page, remove all ptes and migrate the page | |
829 | * to the newly allocated page in newpage. | |
830 | */ | |
831 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
a6bc32b8 MG |
832 | struct page *page, int force, bool offlining, |
833 | enum migrate_mode mode) | |
0dabec93 MK |
834 | { |
835 | int rc = 0; | |
836 | int *result = NULL; | |
837 | struct page *newpage = get_new_page(page, private, &result); | |
838 | ||
839 | if (!newpage) | |
840 | return -ENOMEM; | |
841 | ||
842 | if (page_count(page) == 1) { | |
843 | /* page was freed from under us. So we are done. */ | |
844 | goto out; | |
845 | } | |
846 | ||
847 | if (unlikely(PageTransHuge(page))) | |
848 | if (unlikely(split_huge_page(page))) | |
849 | goto out; | |
850 | ||
a6bc32b8 | 851 | rc = __unmap_and_move(page, newpage, force, offlining, mode); |
0dabec93 | 852 | out: |
e24f0b8f | 853 | if (rc != -EAGAIN) { |
0dabec93 MK |
854 | /* |
855 | * A page that has been migrated has all references | |
856 | * removed and will be freed. A page that has not been | |
857 | * migrated will have kepts its references and be | |
858 | * restored. | |
859 | */ | |
860 | list_del(&page->lru); | |
a731286d | 861 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 862 | page_is_file_cache(page)); |
894bc310 | 863 | putback_lru_page(page); |
e24f0b8f | 864 | } |
95a402c3 CL |
865 | /* |
866 | * Move the new page to the LRU. If migration was not successful | |
867 | * then this will free the page. | |
868 | */ | |
894bc310 | 869 | putback_lru_page(newpage); |
742755a1 CL |
870 | if (result) { |
871 | if (rc) | |
872 | *result = rc; | |
873 | else | |
874 | *result = page_to_nid(newpage); | |
875 | } | |
e24f0b8f CL |
876 | return rc; |
877 | } | |
878 | ||
290408d4 NH |
879 | /* |
880 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
881 | * | |
882 | * This function doesn't wait the completion of hugepage I/O | |
883 | * because there is no race between I/O and migration for hugepage. | |
884 | * Note that currently hugepage I/O occurs only in direct I/O | |
885 | * where no lock is held and PG_writeback is irrelevant, | |
886 | * and writeback status of all subpages are counted in the reference | |
887 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
888 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
889 | * This means that when we try to migrate hugepage whose subpages are | |
890 | * doing direct I/O, some references remain after try_to_unmap() and | |
891 | * hugepage migration fails without data corruption. | |
892 | * | |
893 | * There is also no race when direct I/O is issued on the page under migration, | |
894 | * because then pte is replaced with migration swap entry and direct I/O code | |
895 | * will wait in the page fault for migration to complete. | |
896 | */ | |
897 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
898 | unsigned long private, struct page *hpage, | |
a6bc32b8 MG |
899 | int force, bool offlining, |
900 | enum migrate_mode mode) | |
290408d4 NH |
901 | { |
902 | int rc = 0; | |
903 | int *result = NULL; | |
904 | struct page *new_hpage = get_new_page(hpage, private, &result); | |
290408d4 NH |
905 | struct anon_vma *anon_vma = NULL; |
906 | ||
907 | if (!new_hpage) | |
908 | return -ENOMEM; | |
909 | ||
910 | rc = -EAGAIN; | |
911 | ||
912 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 913 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
914 | goto out; |
915 | lock_page(hpage); | |
916 | } | |
917 | ||
746b18d4 PZ |
918 | if (PageAnon(hpage)) |
919 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 NH |
920 | |
921 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
922 | ||
923 | if (!page_mapped(hpage)) | |
a6bc32b8 | 924 | rc = move_to_new_page(new_hpage, hpage, 1, mode); |
290408d4 NH |
925 | |
926 | if (rc) | |
927 | remove_migration_ptes(hpage, hpage); | |
928 | ||
fd4a4663 | 929 | if (anon_vma) |
9e60109f | 930 | put_anon_vma(anon_vma); |
8e6ac7fa AK |
931 | |
932 | if (!rc) | |
933 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
934 | ||
290408d4 | 935 | unlock_page(hpage); |
09761333 | 936 | out: |
290408d4 | 937 | put_page(new_hpage); |
290408d4 NH |
938 | if (result) { |
939 | if (rc) | |
940 | *result = rc; | |
941 | else | |
942 | *result = page_to_nid(new_hpage); | |
943 | } | |
944 | return rc; | |
945 | } | |
946 | ||
b20a3503 CL |
947 | /* |
948 | * migrate_pages | |
949 | * | |
95a402c3 CL |
950 | * The function takes one list of pages to migrate and a function |
951 | * that determines from the page to be migrated and the private data | |
952 | * the target of the move and allocates the page. | |
b20a3503 CL |
953 | * |
954 | * The function returns after 10 attempts or if no pages | |
955 | * are movable anymore because to has become empty | |
cf608ac1 MK |
956 | * or no retryable pages exist anymore. |
957 | * Caller should call putback_lru_pages to return pages to the LRU | |
28bd6578 | 958 | * or free list only if ret != 0. |
b20a3503 | 959 | * |
95a402c3 | 960 | * Return: Number of pages not migrated or error code. |
b20a3503 | 961 | */ |
95a402c3 | 962 | int migrate_pages(struct list_head *from, |
7f0f2496 | 963 | new_page_t get_new_page, unsigned long private, bool offlining, |
7b2a2d4a | 964 | enum migrate_mode mode, int reason) |
b20a3503 | 965 | { |
e24f0b8f | 966 | int retry = 1; |
b20a3503 | 967 | int nr_failed = 0; |
5647bc29 | 968 | int nr_succeeded = 0; |
b20a3503 CL |
969 | int pass = 0; |
970 | struct page *page; | |
971 | struct page *page2; | |
972 | int swapwrite = current->flags & PF_SWAPWRITE; | |
973 | int rc; | |
974 | ||
975 | if (!swapwrite) | |
976 | current->flags |= PF_SWAPWRITE; | |
977 | ||
e24f0b8f CL |
978 | for(pass = 0; pass < 10 && retry; pass++) { |
979 | retry = 0; | |
b20a3503 | 980 | |
e24f0b8f | 981 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 982 | cond_resched(); |
2d1db3b1 | 983 | |
95a402c3 | 984 | rc = unmap_and_move(get_new_page, private, |
77f1fe6b | 985 | page, pass > 2, offlining, |
a6bc32b8 | 986 | mode); |
2d1db3b1 | 987 | |
e24f0b8f | 988 | switch(rc) { |
95a402c3 CL |
989 | case -ENOMEM: |
990 | goto out; | |
e24f0b8f | 991 | case -EAGAIN: |
2d1db3b1 | 992 | retry++; |
e24f0b8f CL |
993 | break; |
994 | case 0: | |
5647bc29 | 995 | nr_succeeded++; |
e24f0b8f CL |
996 | break; |
997 | default: | |
2d1db3b1 | 998 | /* Permanent failure */ |
2d1db3b1 | 999 | nr_failed++; |
e24f0b8f | 1000 | break; |
2d1db3b1 | 1001 | } |
b20a3503 CL |
1002 | } |
1003 | } | |
95a402c3 CL |
1004 | rc = 0; |
1005 | out: | |
5647bc29 MG |
1006 | if (nr_succeeded) |
1007 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1008 | if (nr_failed) | |
1009 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1010 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1011 | ||
b20a3503 CL |
1012 | if (!swapwrite) |
1013 | current->flags &= ~PF_SWAPWRITE; | |
1014 | ||
95a402c3 CL |
1015 | if (rc) |
1016 | return rc; | |
b20a3503 | 1017 | |
95a402c3 | 1018 | return nr_failed + retry; |
b20a3503 | 1019 | } |
95a402c3 | 1020 | |
189ebff2 AK |
1021 | int migrate_huge_page(struct page *hpage, new_page_t get_new_page, |
1022 | unsigned long private, bool offlining, | |
1023 | enum migrate_mode mode) | |
290408d4 | 1024 | { |
189ebff2 AK |
1025 | int pass, rc; |
1026 | ||
1027 | for (pass = 0; pass < 10; pass++) { | |
1028 | rc = unmap_and_move_huge_page(get_new_page, | |
1029 | private, hpage, pass > 2, offlining, | |
1030 | mode); | |
1031 | switch (rc) { | |
1032 | case -ENOMEM: | |
1033 | goto out; | |
1034 | case -EAGAIN: | |
1035 | /* try again */ | |
290408d4 | 1036 | cond_resched(); |
189ebff2 AK |
1037 | break; |
1038 | case 0: | |
1039 | goto out; | |
1040 | default: | |
1041 | rc = -EIO; | |
1042 | goto out; | |
290408d4 NH |
1043 | } |
1044 | } | |
290408d4 | 1045 | out: |
189ebff2 | 1046 | return rc; |
290408d4 NH |
1047 | } |
1048 | ||
742755a1 CL |
1049 | #ifdef CONFIG_NUMA |
1050 | /* | |
1051 | * Move a list of individual pages | |
1052 | */ | |
1053 | struct page_to_node { | |
1054 | unsigned long addr; | |
1055 | struct page *page; | |
1056 | int node; | |
1057 | int status; | |
1058 | }; | |
1059 | ||
1060 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1061 | int **result) | |
1062 | { | |
1063 | struct page_to_node *pm = (struct page_to_node *)private; | |
1064 | ||
1065 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1066 | pm++; | |
1067 | ||
1068 | if (pm->node == MAX_NUMNODES) | |
1069 | return NULL; | |
1070 | ||
1071 | *result = &pm->status; | |
1072 | ||
6484eb3e | 1073 | return alloc_pages_exact_node(pm->node, |
769848c0 | 1074 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); |
742755a1 CL |
1075 | } |
1076 | ||
1077 | /* | |
1078 | * Move a set of pages as indicated in the pm array. The addr | |
1079 | * field must be set to the virtual address of the page to be moved | |
1080 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1081 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1082 | */ |
5e9a0f02 BG |
1083 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1084 | struct page_to_node *pm, | |
1085 | int migrate_all) | |
742755a1 CL |
1086 | { |
1087 | int err; | |
1088 | struct page_to_node *pp; | |
1089 | LIST_HEAD(pagelist); | |
1090 | ||
1091 | down_read(&mm->mmap_sem); | |
1092 | ||
1093 | /* | |
1094 | * Build a list of pages to migrate | |
1095 | */ | |
742755a1 CL |
1096 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1097 | struct vm_area_struct *vma; | |
1098 | struct page *page; | |
1099 | ||
742755a1 CL |
1100 | err = -EFAULT; |
1101 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1102 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1103 | goto set_status; |
1104 | ||
500d65d4 | 1105 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1106 | |
1107 | err = PTR_ERR(page); | |
1108 | if (IS_ERR(page)) | |
1109 | goto set_status; | |
1110 | ||
742755a1 CL |
1111 | err = -ENOENT; |
1112 | if (!page) | |
1113 | goto set_status; | |
1114 | ||
62b61f61 HD |
1115 | /* Use PageReserved to check for zero page */ |
1116 | if (PageReserved(page) || PageKsm(page)) | |
742755a1 CL |
1117 | goto put_and_set; |
1118 | ||
1119 | pp->page = page; | |
1120 | err = page_to_nid(page); | |
1121 | ||
1122 | if (err == pp->node) | |
1123 | /* | |
1124 | * Node already in the right place | |
1125 | */ | |
1126 | goto put_and_set; | |
1127 | ||
1128 | err = -EACCES; | |
1129 | if (page_mapcount(page) > 1 && | |
1130 | !migrate_all) | |
1131 | goto put_and_set; | |
1132 | ||
62695a84 | 1133 | err = isolate_lru_page(page); |
6d9c285a | 1134 | if (!err) { |
62695a84 | 1135 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1136 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1137 | page_is_file_cache(page)); | |
1138 | } | |
742755a1 CL |
1139 | put_and_set: |
1140 | /* | |
1141 | * Either remove the duplicate refcount from | |
1142 | * isolate_lru_page() or drop the page ref if it was | |
1143 | * not isolated. | |
1144 | */ | |
1145 | put_page(page); | |
1146 | set_status: | |
1147 | pp->status = err; | |
1148 | } | |
1149 | ||
e78bbfa8 | 1150 | err = 0; |
cf608ac1 | 1151 | if (!list_empty(&pagelist)) { |
742755a1 | 1152 | err = migrate_pages(&pagelist, new_page_node, |
7b2a2d4a MG |
1153 | (unsigned long)pm, 0, MIGRATE_SYNC, |
1154 | MR_SYSCALL); | |
cf608ac1 MK |
1155 | if (err) |
1156 | putback_lru_pages(&pagelist); | |
1157 | } | |
742755a1 CL |
1158 | |
1159 | up_read(&mm->mmap_sem); | |
1160 | return err; | |
1161 | } | |
1162 | ||
5e9a0f02 BG |
1163 | /* |
1164 | * Migrate an array of page address onto an array of nodes and fill | |
1165 | * the corresponding array of status. | |
1166 | */ | |
3268c63e | 1167 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1168 | unsigned long nr_pages, |
1169 | const void __user * __user *pages, | |
1170 | const int __user *nodes, | |
1171 | int __user *status, int flags) | |
1172 | { | |
3140a227 | 1173 | struct page_to_node *pm; |
3140a227 BG |
1174 | unsigned long chunk_nr_pages; |
1175 | unsigned long chunk_start; | |
1176 | int err; | |
5e9a0f02 | 1177 | |
3140a227 BG |
1178 | err = -ENOMEM; |
1179 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1180 | if (!pm) | |
5e9a0f02 | 1181 | goto out; |
35282a2d BG |
1182 | |
1183 | migrate_prep(); | |
1184 | ||
5e9a0f02 | 1185 | /* |
3140a227 BG |
1186 | * Store a chunk of page_to_node array in a page, |
1187 | * but keep the last one as a marker | |
5e9a0f02 | 1188 | */ |
3140a227 | 1189 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1190 | |
3140a227 BG |
1191 | for (chunk_start = 0; |
1192 | chunk_start < nr_pages; | |
1193 | chunk_start += chunk_nr_pages) { | |
1194 | int j; | |
5e9a0f02 | 1195 | |
3140a227 BG |
1196 | if (chunk_start + chunk_nr_pages > nr_pages) |
1197 | chunk_nr_pages = nr_pages - chunk_start; | |
1198 | ||
1199 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1200 | for (j = 0; j < chunk_nr_pages; j++) { | |
1201 | const void __user *p; | |
5e9a0f02 BG |
1202 | int node; |
1203 | ||
3140a227 BG |
1204 | err = -EFAULT; |
1205 | if (get_user(p, pages + j + chunk_start)) | |
1206 | goto out_pm; | |
1207 | pm[j].addr = (unsigned long) p; | |
1208 | ||
1209 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1210 | goto out_pm; |
1211 | ||
1212 | err = -ENODEV; | |
6f5a55f1 LT |
1213 | if (node < 0 || node >= MAX_NUMNODES) |
1214 | goto out_pm; | |
1215 | ||
5e9a0f02 BG |
1216 | if (!node_state(node, N_HIGH_MEMORY)) |
1217 | goto out_pm; | |
1218 | ||
1219 | err = -EACCES; | |
1220 | if (!node_isset(node, task_nodes)) | |
1221 | goto out_pm; | |
1222 | ||
3140a227 BG |
1223 | pm[j].node = node; |
1224 | } | |
1225 | ||
1226 | /* End marker for this chunk */ | |
1227 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1228 | ||
1229 | /* Migrate this chunk */ | |
1230 | err = do_move_page_to_node_array(mm, pm, | |
1231 | flags & MPOL_MF_MOVE_ALL); | |
1232 | if (err < 0) | |
1233 | goto out_pm; | |
5e9a0f02 | 1234 | |
5e9a0f02 | 1235 | /* Return status information */ |
3140a227 BG |
1236 | for (j = 0; j < chunk_nr_pages; j++) |
1237 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1238 | err = -EFAULT; |
3140a227 BG |
1239 | goto out_pm; |
1240 | } | |
1241 | } | |
1242 | err = 0; | |
5e9a0f02 BG |
1243 | |
1244 | out_pm: | |
3140a227 | 1245 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1246 | out: |
1247 | return err; | |
1248 | } | |
1249 | ||
742755a1 | 1250 | /* |
2f007e74 | 1251 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1252 | */ |
80bba129 BG |
1253 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1254 | const void __user **pages, int *status) | |
742755a1 | 1255 | { |
2f007e74 | 1256 | unsigned long i; |
2f007e74 | 1257 | |
742755a1 CL |
1258 | down_read(&mm->mmap_sem); |
1259 | ||
2f007e74 | 1260 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1261 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1262 | struct vm_area_struct *vma; |
1263 | struct page *page; | |
c095adbc | 1264 | int err = -EFAULT; |
2f007e74 BG |
1265 | |
1266 | vma = find_vma(mm, addr); | |
70384dc6 | 1267 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1268 | goto set_status; |
1269 | ||
2f007e74 | 1270 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1271 | |
1272 | err = PTR_ERR(page); | |
1273 | if (IS_ERR(page)) | |
1274 | goto set_status; | |
1275 | ||
742755a1 CL |
1276 | err = -ENOENT; |
1277 | /* Use PageReserved to check for zero page */ | |
62b61f61 | 1278 | if (!page || PageReserved(page) || PageKsm(page)) |
742755a1 CL |
1279 | goto set_status; |
1280 | ||
1281 | err = page_to_nid(page); | |
1282 | set_status: | |
80bba129 BG |
1283 | *status = err; |
1284 | ||
1285 | pages++; | |
1286 | status++; | |
1287 | } | |
1288 | ||
1289 | up_read(&mm->mmap_sem); | |
1290 | } | |
1291 | ||
1292 | /* | |
1293 | * Determine the nodes of a user array of pages and store it in | |
1294 | * a user array of status. | |
1295 | */ | |
1296 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1297 | const void __user * __user *pages, | |
1298 | int __user *status) | |
1299 | { | |
1300 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1301 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1302 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1303 | |
87b8d1ad PA |
1304 | while (nr_pages) { |
1305 | unsigned long chunk_nr; | |
80bba129 | 1306 | |
87b8d1ad PA |
1307 | chunk_nr = nr_pages; |
1308 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1309 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1310 | ||
1311 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1312 | break; | |
80bba129 BG |
1313 | |
1314 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1315 | ||
87b8d1ad PA |
1316 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1317 | break; | |
742755a1 | 1318 | |
87b8d1ad PA |
1319 | pages += chunk_nr; |
1320 | status += chunk_nr; | |
1321 | nr_pages -= chunk_nr; | |
1322 | } | |
1323 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1324 | } |
1325 | ||
1326 | /* | |
1327 | * Move a list of pages in the address space of the currently executing | |
1328 | * process. | |
1329 | */ | |
938bb9f5 HC |
1330 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1331 | const void __user * __user *, pages, | |
1332 | const int __user *, nodes, | |
1333 | int __user *, status, int, flags) | |
742755a1 | 1334 | { |
c69e8d9c | 1335 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1336 | struct task_struct *task; |
742755a1 | 1337 | struct mm_struct *mm; |
5e9a0f02 | 1338 | int err; |
3268c63e | 1339 | nodemask_t task_nodes; |
742755a1 CL |
1340 | |
1341 | /* Check flags */ | |
1342 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1343 | return -EINVAL; | |
1344 | ||
1345 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1346 | return -EPERM; | |
1347 | ||
1348 | /* Find the mm_struct */ | |
a879bf58 | 1349 | rcu_read_lock(); |
228ebcbe | 1350 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1351 | if (!task) { |
a879bf58 | 1352 | rcu_read_unlock(); |
742755a1 CL |
1353 | return -ESRCH; |
1354 | } | |
3268c63e | 1355 | get_task_struct(task); |
742755a1 CL |
1356 | |
1357 | /* | |
1358 | * Check if this process has the right to modify the specified | |
1359 | * process. The right exists if the process has administrative | |
1360 | * capabilities, superuser privileges or the same | |
1361 | * userid as the target process. | |
1362 | */ | |
c69e8d9c | 1363 | tcred = __task_cred(task); |
b38a86eb EB |
1364 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1365 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1366 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1367 | rcu_read_unlock(); |
742755a1 | 1368 | err = -EPERM; |
5e9a0f02 | 1369 | goto out; |
742755a1 | 1370 | } |
c69e8d9c | 1371 | rcu_read_unlock(); |
742755a1 | 1372 | |
86c3a764 DQ |
1373 | err = security_task_movememory(task); |
1374 | if (err) | |
5e9a0f02 | 1375 | goto out; |
86c3a764 | 1376 | |
3268c63e CL |
1377 | task_nodes = cpuset_mems_allowed(task); |
1378 | mm = get_task_mm(task); | |
1379 | put_task_struct(task); | |
1380 | ||
6e8b09ea SL |
1381 | if (!mm) |
1382 | return -EINVAL; | |
1383 | ||
1384 | if (nodes) | |
1385 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1386 | nodes, status, flags); | |
1387 | else | |
1388 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1389 | |
742755a1 CL |
1390 | mmput(mm); |
1391 | return err; | |
3268c63e CL |
1392 | |
1393 | out: | |
1394 | put_task_struct(task); | |
1395 | return err; | |
742755a1 | 1396 | } |
742755a1 | 1397 | |
7b2259b3 CL |
1398 | /* |
1399 | * Call migration functions in the vma_ops that may prepare | |
1400 | * memory in a vm for migration. migration functions may perform | |
1401 | * the migration for vmas that do not have an underlying page struct. | |
1402 | */ | |
1403 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1404 | const nodemask_t *from, unsigned long flags) | |
1405 | { | |
1406 | struct vm_area_struct *vma; | |
1407 | int err = 0; | |
1408 | ||
1001c9fb | 1409 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { |
7b2259b3 CL |
1410 | if (vma->vm_ops && vma->vm_ops->migrate) { |
1411 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1412 | if (err) | |
1413 | break; | |
1414 | } | |
1415 | } | |
1416 | return err; | |
1417 | } | |
7039e1db PZ |
1418 | |
1419 | #ifdef CONFIG_NUMA_BALANCING | |
1420 | /* | |
1421 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1422 | * pages. Currently it only checks the watermarks which crude | |
1423 | */ | |
1424 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
1425 | int nr_migrate_pages) | |
1426 | { | |
1427 | int z; | |
1428 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1429 | struct zone *zone = pgdat->node_zones + z; | |
1430 | ||
1431 | if (!populated_zone(zone)) | |
1432 | continue; | |
1433 | ||
1434 | if (zone->all_unreclaimable) | |
1435 | continue; | |
1436 | ||
1437 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1438 | if (!zone_watermark_ok(zone, 0, | |
1439 | high_wmark_pages(zone) + | |
1440 | nr_migrate_pages, | |
1441 | 0, 0)) | |
1442 | continue; | |
1443 | return true; | |
1444 | } | |
1445 | return false; | |
1446 | } | |
1447 | ||
1448 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1449 | unsigned long data, | |
1450 | int **result) | |
1451 | { | |
1452 | int nid = (int) data; | |
1453 | struct page *newpage; | |
1454 | ||
1455 | newpage = alloc_pages_exact_node(nid, | |
1456 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1457 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1458 | __GFP_NOWARN) & | |
1459 | ~GFP_IOFS, 0); | |
1460 | return newpage; | |
1461 | } | |
1462 | ||
a8f60772 MG |
1463 | /* |
1464 | * page migration rate limiting control. | |
1465 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1466 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1467 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1468 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1469 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1470 | * throttle window closed. | |
a8f60772 MG |
1471 | */ |
1472 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1473 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1474 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1475 | ||
e14808b4 MG |
1476 | /* Returns true if NUMA migration is currently rate limited */ |
1477 | bool migrate_ratelimited(int node) | |
1478 | { | |
1479 | pg_data_t *pgdat = NODE_DATA(node); | |
1480 | ||
1481 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1482 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1483 | return false; | |
1484 | ||
1485 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1486 | return false; | |
1487 | ||
1488 | return true; | |
1489 | } | |
1490 | ||
7039e1db PZ |
1491 | /* |
1492 | * Attempt to migrate a misplaced page to the specified destination | |
1493 | * node. Caller is expected to have an elevated reference count on | |
1494 | * the page that will be dropped by this function before returning. | |
1495 | */ | |
1496 | int migrate_misplaced_page(struct page *page, int node) | |
1497 | { | |
a8f60772 | 1498 | pg_data_t *pgdat = NODE_DATA(node); |
7039e1db PZ |
1499 | int isolated = 0; |
1500 | LIST_HEAD(migratepages); | |
1501 | ||
1502 | /* | |
1503 | * Don't migrate pages that are mapped in multiple processes. | |
1504 | * TODO: Handle false sharing detection instead of this hammer | |
1505 | */ | |
1506 | if (page_mapcount(page) != 1) { | |
1507 | put_page(page); | |
1508 | goto out; | |
1509 | } | |
1510 | ||
a8f60772 MG |
1511 | /* |
1512 | * Rate-limit the amount of data that is being migrated to a node. | |
1513 | * Optimal placement is no good if the memory bus is saturated and | |
1514 | * all the time is being spent migrating! | |
1515 | */ | |
1516 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1517 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1518 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1519 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1520 | msecs_to_jiffies(migrate_interval_millisecs); | |
1521 | } | |
1522 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { | |
1523 | spin_unlock(&pgdat->numabalancing_migrate_lock); | |
1524 | put_page(page); | |
1525 | goto out; | |
1526 | } | |
1527 | pgdat->numabalancing_migrate_nr_pages++; | |
1528 | spin_unlock(&pgdat->numabalancing_migrate_lock); | |
1529 | ||
7039e1db | 1530 | /* Avoid migrating to a node that is nearly full */ |
a8f60772 | 1531 | if (migrate_balanced_pgdat(pgdat, 1)) { |
7039e1db PZ |
1532 | int page_lru; |
1533 | ||
1534 | if (isolate_lru_page(page)) { | |
1535 | put_page(page); | |
1536 | goto out; | |
1537 | } | |
1538 | isolated = 1; | |
1539 | ||
7039e1db PZ |
1540 | page_lru = page_is_file_cache(page); |
1541 | inc_zone_page_state(page, NR_ISOLATED_ANON + page_lru); | |
1542 | list_add(&page->lru, &migratepages); | |
1543 | } | |
1544 | ||
149c33e1 MG |
1545 | /* |
1546 | * Page is either isolated or there is not enough space on the target | |
1547 | * node. If isolated, then it has taken a reference count and the | |
1548 | * callers reference can be safely dropped without the page | |
1549 | * disappearing underneath us during migration. Otherwise the page is | |
1550 | * not to be migrated but the callers reference should still be | |
1551 | * dropped so it does not leak. | |
1552 | */ | |
1553 | put_page(page); | |
1554 | ||
7039e1db PZ |
1555 | if (isolated) { |
1556 | int nr_remaining; | |
1557 | ||
1558 | nr_remaining = migrate_pages(&migratepages, | |
1559 | alloc_misplaced_dst_page, | |
1560 | node, false, MIGRATE_ASYNC, | |
1561 | MR_NUMA_MISPLACED); | |
1562 | if (nr_remaining) { | |
1563 | putback_lru_pages(&migratepages); | |
1564 | isolated = 0; | |
03c5a6e1 MG |
1565 | } else |
1566 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db PZ |
1567 | } |
1568 | BUG_ON(!list_empty(&migratepages)); | |
1569 | out: | |
1570 | return isolated; | |
1571 | } | |
1572 | #endif /* CONFIG_NUMA_BALANCING */ | |
1573 | ||
1574 | #endif /* CONFIG_NUMA */ |