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> |
bf6bddf1 | 38 | #include <linux/balloon_compaction.h> |
b20a3503 | 39 | |
0d1836c3 MN |
40 | #include <asm/tlbflush.h> |
41 | ||
7b2a2d4a MG |
42 | #define CREATE_TRACE_POINTS |
43 | #include <trace/events/migrate.h> | |
44 | ||
b20a3503 CL |
45 | #include "internal.h" |
46 | ||
b20a3503 | 47 | /* |
742755a1 | 48 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
49 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
50 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
51 | */ |
52 | int migrate_prep(void) | |
53 | { | |
b20a3503 CL |
54 | /* |
55 | * Clear the LRU lists so pages can be isolated. | |
56 | * Note that pages may be moved off the LRU after we have | |
57 | * drained them. Those pages will fail to migrate like other | |
58 | * pages that may be busy. | |
59 | */ | |
60 | lru_add_drain_all(); | |
61 | ||
62 | return 0; | |
63 | } | |
64 | ||
748446bb MG |
65 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
66 | int migrate_prep_local(void) | |
67 | { | |
68 | lru_add_drain(); | |
69 | ||
70 | return 0; | |
71 | } | |
72 | ||
b20a3503 | 73 | /* |
894bc310 LS |
74 | * Add isolated pages on the list back to the LRU under page lock |
75 | * to avoid leaking evictable pages back onto unevictable list. | |
b20a3503 | 76 | */ |
e13861d8 | 77 | void putback_lru_pages(struct list_head *l) |
b20a3503 CL |
78 | { |
79 | struct page *page; | |
80 | struct page *page2; | |
b20a3503 | 81 | |
5733c7d1 RA |
82 | list_for_each_entry_safe(page, page2, l, lru) { |
83 | list_del(&page->lru); | |
84 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
85 | page_is_file_cache(page)); | |
86 | putback_lru_page(page); | |
87 | } | |
88 | } | |
89 | ||
90 | /* | |
91 | * Put previously isolated pages back onto the appropriate lists | |
92 | * from where they were once taken off for compaction/migration. | |
93 | * | |
94 | * This function shall be used instead of putback_lru_pages(), | |
95 | * whenever the isolated pageset has been built by isolate_migratepages_range() | |
96 | */ | |
97 | void putback_movable_pages(struct list_head *l) | |
98 | { | |
99 | struct page *page; | |
100 | struct page *page2; | |
101 | ||
b20a3503 | 102 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
103 | if (unlikely(PageHuge(page))) { |
104 | putback_active_hugepage(page); | |
105 | continue; | |
106 | } | |
e24f0b8f | 107 | list_del(&page->lru); |
a731286d | 108 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 109 | page_is_file_cache(page)); |
117aad1e | 110 | if (unlikely(isolated_balloon_page(page))) |
bf6bddf1 RA |
111 | balloon_page_putback(page); |
112 | else | |
113 | putback_lru_page(page); | |
b20a3503 | 114 | } |
b20a3503 CL |
115 | } |
116 | ||
0697212a CL |
117 | /* |
118 | * Restore a potential migration pte to a working pte entry | |
119 | */ | |
e9995ef9 HD |
120 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, |
121 | unsigned long addr, void *old) | |
0697212a CL |
122 | { |
123 | struct mm_struct *mm = vma->vm_mm; | |
124 | swp_entry_t entry; | |
0697212a CL |
125 | pmd_t *pmd; |
126 | pte_t *ptep, pte; | |
127 | spinlock_t *ptl; | |
128 | ||
290408d4 NH |
129 | if (unlikely(PageHuge(new))) { |
130 | ptep = huge_pte_offset(mm, addr); | |
131 | if (!ptep) | |
132 | goto out; | |
cb900f41 | 133 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep); |
290408d4 | 134 | } else { |
6219049a BL |
135 | pmd = mm_find_pmd(mm, addr); |
136 | if (!pmd) | |
290408d4 | 137 | goto out; |
500d65d4 AA |
138 | if (pmd_trans_huge(*pmd)) |
139 | goto out; | |
0697212a | 140 | |
290408d4 | 141 | ptep = pte_offset_map(pmd, addr); |
0697212a | 142 | |
486cf46f HD |
143 | /* |
144 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
145 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
146 | */ | |
290408d4 NH |
147 | |
148 | ptl = pte_lockptr(mm, pmd); | |
149 | } | |
0697212a | 150 | |
0697212a CL |
151 | spin_lock(ptl); |
152 | pte = *ptep; | |
153 | if (!is_swap_pte(pte)) | |
e9995ef9 | 154 | goto unlock; |
0697212a CL |
155 | |
156 | entry = pte_to_swp_entry(pte); | |
157 | ||
e9995ef9 HD |
158 | if (!is_migration_entry(entry) || |
159 | migration_entry_to_page(entry) != old) | |
160 | goto unlock; | |
0697212a | 161 | |
0697212a CL |
162 | get_page(new); |
163 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
c3d16e16 CG |
164 | if (pte_swp_soft_dirty(*ptep)) |
165 | pte = pte_mksoft_dirty(pte); | |
0697212a CL |
166 | if (is_write_migration_entry(entry)) |
167 | pte = pte_mkwrite(pte); | |
3ef8fd7f | 168 | #ifdef CONFIG_HUGETLB_PAGE |
be7517d6 | 169 | if (PageHuge(new)) { |
290408d4 | 170 | pte = pte_mkhuge(pte); |
be7517d6 TL |
171 | pte = arch_make_huge_pte(pte, vma, new, 0); |
172 | } | |
3ef8fd7f | 173 | #endif |
c2cc499c | 174 | flush_dcache_page(new); |
0697212a | 175 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 176 | |
290408d4 NH |
177 | if (PageHuge(new)) { |
178 | if (PageAnon(new)) | |
179 | hugepage_add_anon_rmap(new, vma, addr); | |
180 | else | |
181 | page_dup_rmap(new); | |
182 | } else if (PageAnon(new)) | |
04e62a29 CL |
183 | page_add_anon_rmap(new, vma, addr); |
184 | else | |
185 | page_add_file_rmap(new); | |
186 | ||
187 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 188 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 189 | unlock: |
0697212a | 190 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
191 | out: |
192 | return SWAP_AGAIN; | |
0697212a CL |
193 | } |
194 | ||
04e62a29 CL |
195 | /* |
196 | * Get rid of all migration entries and replace them by | |
197 | * references to the indicated page. | |
198 | */ | |
199 | static void remove_migration_ptes(struct page *old, struct page *new) | |
200 | { | |
e9995ef9 | 201 | rmap_walk(new, remove_migration_pte, old); |
04e62a29 CL |
202 | } |
203 | ||
0697212a CL |
204 | /* |
205 | * Something used the pte of a page under migration. We need to | |
206 | * get to the page and wait until migration is finished. | |
207 | * When we return from this function the fault will be retried. | |
0697212a | 208 | */ |
30dad309 NH |
209 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
210 | spinlock_t *ptl) | |
0697212a | 211 | { |
30dad309 | 212 | pte_t pte; |
0697212a CL |
213 | swp_entry_t entry; |
214 | struct page *page; | |
215 | ||
30dad309 | 216 | spin_lock(ptl); |
0697212a CL |
217 | pte = *ptep; |
218 | if (!is_swap_pte(pte)) | |
219 | goto out; | |
220 | ||
221 | entry = pte_to_swp_entry(pte); | |
222 | if (!is_migration_entry(entry)) | |
223 | goto out; | |
224 | ||
225 | page = migration_entry_to_page(entry); | |
226 | ||
e286781d NP |
227 | /* |
228 | * Once radix-tree replacement of page migration started, page_count | |
229 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
230 | * against a page without get_page(). | |
231 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
232 | * will occur again. | |
233 | */ | |
234 | if (!get_page_unless_zero(page)) | |
235 | goto out; | |
0697212a CL |
236 | pte_unmap_unlock(ptep, ptl); |
237 | wait_on_page_locked(page); | |
238 | put_page(page); | |
239 | return; | |
240 | out: | |
241 | pte_unmap_unlock(ptep, ptl); | |
242 | } | |
243 | ||
30dad309 NH |
244 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
245 | unsigned long address) | |
246 | { | |
247 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
248 | pte_t *ptep = pte_offset_map(pmd, address); | |
249 | __migration_entry_wait(mm, ptep, ptl); | |
250 | } | |
251 | ||
cb900f41 KS |
252 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
253 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 254 | { |
cb900f41 | 255 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
256 | __migration_entry_wait(mm, pte, ptl); |
257 | } | |
258 | ||
b969c4ab MG |
259 | #ifdef CONFIG_BLOCK |
260 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
261 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
262 | enum migrate_mode mode) | |
b969c4ab MG |
263 | { |
264 | struct buffer_head *bh = head; | |
265 | ||
266 | /* Simple case, sync compaction */ | |
a6bc32b8 | 267 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
268 | do { |
269 | get_bh(bh); | |
270 | lock_buffer(bh); | |
271 | bh = bh->b_this_page; | |
272 | ||
273 | } while (bh != head); | |
274 | ||
275 | return true; | |
276 | } | |
277 | ||
278 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
279 | do { | |
280 | get_bh(bh); | |
281 | if (!trylock_buffer(bh)) { | |
282 | /* | |
283 | * We failed to lock the buffer and cannot stall in | |
284 | * async migration. Release the taken locks | |
285 | */ | |
286 | struct buffer_head *failed_bh = bh; | |
287 | put_bh(failed_bh); | |
288 | bh = head; | |
289 | while (bh != failed_bh) { | |
290 | unlock_buffer(bh); | |
291 | put_bh(bh); | |
292 | bh = bh->b_this_page; | |
293 | } | |
294 | return false; | |
295 | } | |
296 | ||
297 | bh = bh->b_this_page; | |
298 | } while (bh != head); | |
299 | return true; | |
300 | } | |
301 | #else | |
302 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 303 | enum migrate_mode mode) |
b969c4ab MG |
304 | { |
305 | return true; | |
306 | } | |
307 | #endif /* CONFIG_BLOCK */ | |
308 | ||
b20a3503 | 309 | /* |
c3fcf8a5 | 310 | * Replace the page in the mapping. |
5b5c7120 CL |
311 | * |
312 | * The number of remaining references must be: | |
313 | * 1 for anonymous pages without a mapping | |
314 | * 2 for pages with a mapping | |
266cf658 | 315 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 316 | */ |
36bc08cc | 317 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 318 | struct page *newpage, struct page *page, |
a6bc32b8 | 319 | struct buffer_head *head, enum migrate_mode mode) |
b20a3503 | 320 | { |
7039e1db | 321 | int expected_count = 0; |
7cf9c2c7 | 322 | void **pslot; |
b20a3503 | 323 | |
6c5240ae | 324 | if (!mapping) { |
0e8c7d0f | 325 | /* Anonymous page without mapping */ |
6c5240ae CL |
326 | if (page_count(page) != 1) |
327 | return -EAGAIN; | |
78bd5209 | 328 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
329 | } |
330 | ||
19fd6231 | 331 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 332 | |
7cf9c2c7 NP |
333 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
334 | page_index(page)); | |
b20a3503 | 335 | |
edcf4748 | 336 | expected_count = 2 + page_has_private(page); |
e286781d | 337 | if (page_count(page) != expected_count || |
29c1f677 | 338 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 339 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 340 | return -EAGAIN; |
b20a3503 CL |
341 | } |
342 | ||
e286781d | 343 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 344 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
345 | return -EAGAIN; |
346 | } | |
347 | ||
b969c4ab MG |
348 | /* |
349 | * In the async migration case of moving a page with buffers, lock the | |
350 | * buffers using trylock before the mapping is moved. If the mapping | |
351 | * was moved, we later failed to lock the buffers and could not move | |
352 | * the mapping back due to an elevated page count, we would have to | |
353 | * block waiting on other references to be dropped. | |
354 | */ | |
a6bc32b8 MG |
355 | if (mode == MIGRATE_ASYNC && head && |
356 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
357 | page_unfreeze_refs(page, expected_count); |
358 | spin_unlock_irq(&mapping->tree_lock); | |
359 | return -EAGAIN; | |
360 | } | |
361 | ||
b20a3503 CL |
362 | /* |
363 | * Now we know that no one else is looking at the page. | |
b20a3503 | 364 | */ |
7cf9c2c7 | 365 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
366 | if (PageSwapCache(page)) { |
367 | SetPageSwapCache(newpage); | |
368 | set_page_private(newpage, page_private(page)); | |
369 | } | |
370 | ||
7cf9c2c7 NP |
371 | radix_tree_replace_slot(pslot, newpage); |
372 | ||
373 | /* | |
937a94c9 JG |
374 | * Drop cache reference from old page by unfreezing |
375 | * to one less reference. | |
7cf9c2c7 NP |
376 | * We know this isn't the last reference. |
377 | */ | |
937a94c9 | 378 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 379 | |
0e8c7d0f CL |
380 | /* |
381 | * If moved to a different zone then also account | |
382 | * the page for that zone. Other VM counters will be | |
383 | * taken care of when we establish references to the | |
384 | * new page and drop references to the old page. | |
385 | * | |
386 | * Note that anonymous pages are accounted for | |
387 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
388 | * are mapped to swap space. | |
389 | */ | |
390 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
391 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 392 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
393 | __dec_zone_page_state(page, NR_SHMEM); |
394 | __inc_zone_page_state(newpage, NR_SHMEM); | |
395 | } | |
19fd6231 | 396 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 397 | |
78bd5209 | 398 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 399 | } |
b20a3503 | 400 | |
290408d4 NH |
401 | /* |
402 | * The expected number of remaining references is the same as that | |
403 | * of migrate_page_move_mapping(). | |
404 | */ | |
405 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
406 | struct page *newpage, struct page *page) | |
407 | { | |
408 | int expected_count; | |
409 | void **pslot; | |
410 | ||
411 | if (!mapping) { | |
412 | if (page_count(page) != 1) | |
413 | return -EAGAIN; | |
78bd5209 | 414 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
415 | } |
416 | ||
417 | spin_lock_irq(&mapping->tree_lock); | |
418 | ||
419 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
420 | page_index(page)); | |
421 | ||
422 | expected_count = 2 + page_has_private(page); | |
423 | if (page_count(page) != expected_count || | |
29c1f677 | 424 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
425 | spin_unlock_irq(&mapping->tree_lock); |
426 | return -EAGAIN; | |
427 | } | |
428 | ||
429 | if (!page_freeze_refs(page, expected_count)) { | |
430 | spin_unlock_irq(&mapping->tree_lock); | |
431 | return -EAGAIN; | |
432 | } | |
433 | ||
434 | get_page(newpage); | |
435 | ||
436 | radix_tree_replace_slot(pslot, newpage); | |
437 | ||
937a94c9 | 438 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
439 | |
440 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 441 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
442 | } |
443 | ||
b20a3503 CL |
444 | /* |
445 | * Copy the page to its new location | |
446 | */ | |
290408d4 | 447 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 448 | { |
7851a45c RR |
449 | int cpupid; |
450 | ||
b32967ff | 451 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
452 | copy_huge_page(newpage, page); |
453 | else | |
454 | copy_highpage(newpage, page); | |
b20a3503 CL |
455 | |
456 | if (PageError(page)) | |
457 | SetPageError(newpage); | |
458 | if (PageReferenced(page)) | |
459 | SetPageReferenced(newpage); | |
460 | if (PageUptodate(page)) | |
461 | SetPageUptodate(newpage); | |
894bc310 LS |
462 | if (TestClearPageActive(page)) { |
463 | VM_BUG_ON(PageUnevictable(page)); | |
b20a3503 | 464 | SetPageActive(newpage); |
418b27ef LS |
465 | } else if (TestClearPageUnevictable(page)) |
466 | SetPageUnevictable(newpage); | |
b20a3503 CL |
467 | if (PageChecked(page)) |
468 | SetPageChecked(newpage); | |
469 | if (PageMappedToDisk(page)) | |
470 | SetPageMappedToDisk(newpage); | |
471 | ||
472 | if (PageDirty(page)) { | |
473 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
474 | /* |
475 | * Want to mark the page and the radix tree as dirty, and | |
476 | * redo the accounting that clear_page_dirty_for_io undid, | |
477 | * but we can't use set_page_dirty because that function | |
478 | * is actually a signal that all of the page has become dirty. | |
25985edc | 479 | * Whereas only part of our page may be dirty. |
3a902c5f | 480 | */ |
752dc185 HD |
481 | if (PageSwapBacked(page)) |
482 | SetPageDirty(newpage); | |
483 | else | |
484 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
485 | } |
486 | ||
7851a45c RR |
487 | /* |
488 | * Copy NUMA information to the new page, to prevent over-eager | |
489 | * future migrations of this same page. | |
490 | */ | |
491 | cpupid = page_cpupid_xchg_last(page, -1); | |
492 | page_cpupid_xchg_last(newpage, cpupid); | |
493 | ||
b291f000 | 494 | mlock_migrate_page(newpage, page); |
e9995ef9 | 495 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
496 | /* |
497 | * Please do not reorder this without considering how mm/ksm.c's | |
498 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
499 | */ | |
b20a3503 | 500 | ClearPageSwapCache(page); |
b20a3503 CL |
501 | ClearPagePrivate(page); |
502 | set_page_private(page, 0); | |
b20a3503 CL |
503 | |
504 | /* | |
505 | * If any waiters have accumulated on the new page then | |
506 | * wake them up. | |
507 | */ | |
508 | if (PageWriteback(newpage)) | |
509 | end_page_writeback(newpage); | |
510 | } | |
b20a3503 | 511 | |
1d8b85cc CL |
512 | /************************************************************ |
513 | * Migration functions | |
514 | ***********************************************************/ | |
515 | ||
516 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
517 | int fail_migrate_page(struct address_space *mapping, |
518 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
519 | { |
520 | return -EIO; | |
521 | } | |
522 | EXPORT_SYMBOL(fail_migrate_page); | |
523 | ||
b20a3503 CL |
524 | /* |
525 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 526 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
527 | * |
528 | * Pages are locked upon entry and exit. | |
529 | */ | |
2d1db3b1 | 530 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
531 | struct page *newpage, struct page *page, |
532 | enum migrate_mode mode) | |
b20a3503 CL |
533 | { |
534 | int rc; | |
535 | ||
536 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
537 | ||
a6bc32b8 | 538 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode); |
b20a3503 | 539 | |
78bd5209 | 540 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
541 | return rc; |
542 | ||
543 | migrate_page_copy(newpage, page); | |
78bd5209 | 544 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
545 | } |
546 | EXPORT_SYMBOL(migrate_page); | |
547 | ||
9361401e | 548 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
549 | /* |
550 | * Migration function for pages with buffers. This function can only be used | |
551 | * if the underlying filesystem guarantees that no other references to "page" | |
552 | * exist. | |
553 | */ | |
2d1db3b1 | 554 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 555 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 556 | { |
1d8b85cc CL |
557 | struct buffer_head *bh, *head; |
558 | int rc; | |
559 | ||
1d8b85cc | 560 | if (!page_has_buffers(page)) |
a6bc32b8 | 561 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
562 | |
563 | head = page_buffers(page); | |
564 | ||
a6bc32b8 | 565 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode); |
1d8b85cc | 566 | |
78bd5209 | 567 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
568 | return rc; |
569 | ||
b969c4ab MG |
570 | /* |
571 | * In the async case, migrate_page_move_mapping locked the buffers | |
572 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
573 | * need to be locked now | |
574 | */ | |
a6bc32b8 MG |
575 | if (mode != MIGRATE_ASYNC) |
576 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
577 | |
578 | ClearPagePrivate(page); | |
579 | set_page_private(newpage, page_private(page)); | |
580 | set_page_private(page, 0); | |
581 | put_page(page); | |
582 | get_page(newpage); | |
583 | ||
584 | bh = head; | |
585 | do { | |
586 | set_bh_page(bh, newpage, bh_offset(bh)); | |
587 | bh = bh->b_this_page; | |
588 | ||
589 | } while (bh != head); | |
590 | ||
591 | SetPagePrivate(newpage); | |
592 | ||
593 | migrate_page_copy(newpage, page); | |
594 | ||
595 | bh = head; | |
596 | do { | |
597 | unlock_buffer(bh); | |
598 | put_bh(bh); | |
599 | bh = bh->b_this_page; | |
600 | ||
601 | } while (bh != head); | |
602 | ||
78bd5209 | 603 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
604 | } |
605 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 606 | #endif |
1d8b85cc | 607 | |
04e62a29 CL |
608 | /* |
609 | * Writeback a page to clean the dirty state | |
610 | */ | |
611 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 612 | { |
04e62a29 CL |
613 | struct writeback_control wbc = { |
614 | .sync_mode = WB_SYNC_NONE, | |
615 | .nr_to_write = 1, | |
616 | .range_start = 0, | |
617 | .range_end = LLONG_MAX, | |
04e62a29 CL |
618 | .for_reclaim = 1 |
619 | }; | |
620 | int rc; | |
621 | ||
622 | if (!mapping->a_ops->writepage) | |
623 | /* No write method for the address space */ | |
624 | return -EINVAL; | |
625 | ||
626 | if (!clear_page_dirty_for_io(page)) | |
627 | /* Someone else already triggered a write */ | |
628 | return -EAGAIN; | |
629 | ||
8351a6e4 | 630 | /* |
04e62a29 CL |
631 | * A dirty page may imply that the underlying filesystem has |
632 | * the page on some queue. So the page must be clean for | |
633 | * migration. Writeout may mean we loose the lock and the | |
634 | * page state is no longer what we checked for earlier. | |
635 | * At this point we know that the migration attempt cannot | |
636 | * be successful. | |
8351a6e4 | 637 | */ |
04e62a29 | 638 | remove_migration_ptes(page, page); |
8351a6e4 | 639 | |
04e62a29 | 640 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 641 | |
04e62a29 CL |
642 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
643 | /* unlocked. Relock */ | |
644 | lock_page(page); | |
645 | ||
bda8550d | 646 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
647 | } |
648 | ||
649 | /* | |
650 | * Default handling if a filesystem does not provide a migration function. | |
651 | */ | |
652 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 653 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 654 | { |
b969c4ab | 655 | if (PageDirty(page)) { |
a6bc32b8 MG |
656 | /* Only writeback pages in full synchronous migration */ |
657 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 658 | return -EBUSY; |
04e62a29 | 659 | return writeout(mapping, page); |
b969c4ab | 660 | } |
8351a6e4 CL |
661 | |
662 | /* | |
663 | * Buffers may be managed in a filesystem specific way. | |
664 | * We must have no buffers or drop them. | |
665 | */ | |
266cf658 | 666 | if (page_has_private(page) && |
8351a6e4 CL |
667 | !try_to_release_page(page, GFP_KERNEL)) |
668 | return -EAGAIN; | |
669 | ||
a6bc32b8 | 670 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
671 | } |
672 | ||
e24f0b8f CL |
673 | /* |
674 | * Move a page to a newly allocated page | |
675 | * The page is locked and all ptes have been successfully removed. | |
676 | * | |
677 | * The new page will have replaced the old page if this function | |
678 | * is successful. | |
894bc310 LS |
679 | * |
680 | * Return value: | |
681 | * < 0 - error code | |
78bd5209 | 682 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 683 | */ |
3fe2011f | 684 | static int move_to_new_page(struct page *newpage, struct page *page, |
a6bc32b8 | 685 | int remap_swapcache, enum migrate_mode mode) |
e24f0b8f CL |
686 | { |
687 | struct address_space *mapping; | |
688 | int rc; | |
689 | ||
690 | /* | |
691 | * Block others from accessing the page when we get around to | |
692 | * establishing additional references. We are the only one | |
693 | * holding a reference to the new page at this point. | |
694 | */ | |
529ae9aa | 695 | if (!trylock_page(newpage)) |
e24f0b8f CL |
696 | BUG(); |
697 | ||
698 | /* Prepare mapping for the new page.*/ | |
699 | newpage->index = page->index; | |
700 | newpage->mapping = page->mapping; | |
b2e18538 RR |
701 | if (PageSwapBacked(page)) |
702 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
703 | |
704 | mapping = page_mapping(page); | |
705 | if (!mapping) | |
a6bc32b8 | 706 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 707 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 708 | /* |
b969c4ab MG |
709 | * Most pages have a mapping and most filesystems provide a |
710 | * migratepage callback. Anonymous pages are part of swap | |
711 | * space which also has its own migratepage callback. This | |
712 | * is the most common path for page migration. | |
e24f0b8f | 713 | */ |
b969c4ab | 714 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 715 | newpage, page, mode); |
b969c4ab | 716 | else |
a6bc32b8 | 717 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 718 | |
78bd5209 | 719 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 720 | newpage->mapping = NULL; |
3fe2011f MG |
721 | } else { |
722 | if (remap_swapcache) | |
723 | remove_migration_ptes(page, newpage); | |
35512eca | 724 | page->mapping = NULL; |
3fe2011f | 725 | } |
e24f0b8f CL |
726 | |
727 | unlock_page(newpage); | |
728 | ||
729 | return rc; | |
730 | } | |
731 | ||
0dabec93 | 732 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 733 | int force, enum migrate_mode mode) |
e24f0b8f | 734 | { |
0dabec93 | 735 | int rc = -EAGAIN; |
3fe2011f | 736 | int remap_swapcache = 1; |
56039efa | 737 | struct mem_cgroup *mem; |
3f6c8272 | 738 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 739 | |
529ae9aa | 740 | if (!trylock_page(page)) { |
a6bc32b8 | 741 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 742 | goto out; |
3e7d3449 MG |
743 | |
744 | /* | |
745 | * It's not safe for direct compaction to call lock_page. | |
746 | * For example, during page readahead pages are added locked | |
747 | * to the LRU. Later, when the IO completes the pages are | |
748 | * marked uptodate and unlocked. However, the queueing | |
749 | * could be merging multiple pages for one bio (e.g. | |
750 | * mpage_readpages). If an allocation happens for the | |
751 | * second or third page, the process can end up locking | |
752 | * the same page twice and deadlocking. Rather than | |
753 | * trying to be clever about what pages can be locked, | |
754 | * avoid the use of lock_page for direct compaction | |
755 | * altogether. | |
756 | */ | |
757 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 758 | goto out; |
3e7d3449 | 759 | |
e24f0b8f CL |
760 | lock_page(page); |
761 | } | |
762 | ||
01b1ae63 | 763 | /* charge against new page */ |
0030f535 | 764 | mem_cgroup_prepare_migration(page, newpage, &mem); |
01b1ae63 | 765 | |
e24f0b8f | 766 | if (PageWriteback(page)) { |
11bc82d6 | 767 | /* |
fed5b64a | 768 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
769 | * necessary to wait for PageWriteback. In the async case, |
770 | * the retry loop is too short and in the sync-light case, | |
771 | * the overhead of stalling is too much | |
11bc82d6 | 772 | */ |
a6bc32b8 | 773 | if (mode != MIGRATE_SYNC) { |
11bc82d6 AA |
774 | rc = -EBUSY; |
775 | goto uncharge; | |
776 | } | |
777 | if (!force) | |
01b1ae63 | 778 | goto uncharge; |
e24f0b8f CL |
779 | wait_on_page_writeback(page); |
780 | } | |
e24f0b8f | 781 | /* |
dc386d4d KH |
782 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
783 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 784 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 785 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
786 | * File Caches may use write_page() or lock_page() in migration, then, |
787 | * just care Anon page here. | |
dc386d4d | 788 | */ |
b79bc0a0 | 789 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 790 | /* |
4fc3f1d6 | 791 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
792 | * getting a hold on an anon_vma from outside one of its mms. |
793 | */ | |
746b18d4 | 794 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
795 | if (anon_vma) { |
796 | /* | |
746b18d4 | 797 | * Anon page |
1ce82b69 | 798 | */ |
1ce82b69 | 799 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
800 | /* |
801 | * We cannot be sure that the anon_vma of an unmapped | |
802 | * swapcache page is safe to use because we don't | |
803 | * know in advance if the VMA that this page belonged | |
804 | * to still exists. If the VMA and others sharing the | |
805 | * data have been freed, then the anon_vma could | |
806 | * already be invalid. | |
807 | * | |
808 | * To avoid this possibility, swapcache pages get | |
809 | * migrated but are not remapped when migration | |
810 | * completes | |
811 | */ | |
812 | remap_swapcache = 0; | |
813 | } else { | |
1ce82b69 | 814 | goto uncharge; |
3fe2011f | 815 | } |
989f89c5 | 816 | } |
62e1c553 | 817 | |
bf6bddf1 RA |
818 | if (unlikely(balloon_page_movable(page))) { |
819 | /* | |
820 | * A ballooned page does not need any special attention from | |
821 | * physical to virtual reverse mapping procedures. | |
822 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
823 | * in order to avoid burning cycles at rmap level, and perform | |
824 | * the page migration right away (proteced by page lock). | |
825 | */ | |
826 | rc = balloon_page_migrate(newpage, page, mode); | |
827 | goto uncharge; | |
828 | } | |
829 | ||
dc386d4d | 830 | /* |
62e1c553 SL |
831 | * Corner case handling: |
832 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
833 | * and treated as swapcache but it has no rmap yet. | |
834 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
835 | * trigger a BUG. So handle it here. | |
836 | * 2. An orphaned page (see truncate_complete_page) might have | |
837 | * fs-private metadata. The page can be picked up due to memory | |
838 | * offlining. Everywhere else except page reclaim, the page is | |
839 | * invisible to the vm, so the page can not be migrated. So try to | |
840 | * free the metadata, so the page can be freed. | |
e24f0b8f | 841 | */ |
62e1c553 | 842 | if (!page->mapping) { |
1ce82b69 HD |
843 | VM_BUG_ON(PageAnon(page)); |
844 | if (page_has_private(page)) { | |
62e1c553 | 845 | try_to_free_buffers(page); |
1ce82b69 | 846 | goto uncharge; |
62e1c553 | 847 | } |
abfc3488 | 848 | goto skip_unmap; |
62e1c553 SL |
849 | } |
850 | ||
dc386d4d | 851 | /* Establish migration ptes or remove ptes */ |
14fa31b8 | 852 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
dc386d4d | 853 | |
abfc3488 | 854 | skip_unmap: |
e6a1530d | 855 | if (!page_mapped(page)) |
a6bc32b8 | 856 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); |
e24f0b8f | 857 | |
3fe2011f | 858 | if (rc && remap_swapcache) |
e24f0b8f | 859 | remove_migration_ptes(page, page); |
3f6c8272 MG |
860 | |
861 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 862 | if (anon_vma) |
9e60109f | 863 | put_anon_vma(anon_vma); |
3f6c8272 | 864 | |
01b1ae63 | 865 | uncharge: |
bf6bddf1 RA |
866 | mem_cgroup_end_migration(mem, page, newpage, |
867 | (rc == MIGRATEPAGE_SUCCESS || | |
868 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
e24f0b8f | 869 | unlock_page(page); |
0dabec93 MK |
870 | out: |
871 | return rc; | |
872 | } | |
95a402c3 | 873 | |
0dabec93 MK |
874 | /* |
875 | * Obtain the lock on page, remove all ptes and migrate the page | |
876 | * to the newly allocated page in newpage. | |
877 | */ | |
878 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
9c620e2b | 879 | struct page *page, int force, enum migrate_mode mode) |
0dabec93 MK |
880 | { |
881 | int rc = 0; | |
882 | int *result = NULL; | |
883 | struct page *newpage = get_new_page(page, private, &result); | |
884 | ||
885 | if (!newpage) | |
886 | return -ENOMEM; | |
887 | ||
888 | if (page_count(page) == 1) { | |
889 | /* page was freed from under us. So we are done. */ | |
890 | goto out; | |
891 | } | |
892 | ||
893 | if (unlikely(PageTransHuge(page))) | |
894 | if (unlikely(split_huge_page(page))) | |
895 | goto out; | |
896 | ||
9c620e2b | 897 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 RA |
898 | |
899 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
900 | /* | |
901 | * A ballooned page has been migrated already. | |
902 | * Now, it's the time to wrap-up counters, | |
903 | * handle the page back to Buddy and return. | |
904 | */ | |
905 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
906 | page_is_file_cache(page)); | |
907 | balloon_page_free(page); | |
908 | return MIGRATEPAGE_SUCCESS; | |
909 | } | |
0dabec93 | 910 | out: |
e24f0b8f | 911 | if (rc != -EAGAIN) { |
0dabec93 MK |
912 | /* |
913 | * A page that has been migrated has all references | |
914 | * removed and will be freed. A page that has not been | |
915 | * migrated will have kepts its references and be | |
916 | * restored. | |
917 | */ | |
918 | list_del(&page->lru); | |
a731286d | 919 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 920 | page_is_file_cache(page)); |
894bc310 | 921 | putback_lru_page(page); |
e24f0b8f | 922 | } |
95a402c3 CL |
923 | /* |
924 | * Move the new page to the LRU. If migration was not successful | |
925 | * then this will free the page. | |
926 | */ | |
894bc310 | 927 | putback_lru_page(newpage); |
742755a1 CL |
928 | if (result) { |
929 | if (rc) | |
930 | *result = rc; | |
931 | else | |
932 | *result = page_to_nid(newpage); | |
933 | } | |
e24f0b8f CL |
934 | return rc; |
935 | } | |
936 | ||
290408d4 NH |
937 | /* |
938 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
939 | * | |
940 | * This function doesn't wait the completion of hugepage I/O | |
941 | * because there is no race between I/O and migration for hugepage. | |
942 | * Note that currently hugepage I/O occurs only in direct I/O | |
943 | * where no lock is held and PG_writeback is irrelevant, | |
944 | * and writeback status of all subpages are counted in the reference | |
945 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
946 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
947 | * This means that when we try to migrate hugepage whose subpages are | |
948 | * doing direct I/O, some references remain after try_to_unmap() and | |
949 | * hugepage migration fails without data corruption. | |
950 | * | |
951 | * There is also no race when direct I/O is issued on the page under migration, | |
952 | * because then pte is replaced with migration swap entry and direct I/O code | |
953 | * will wait in the page fault for migration to complete. | |
954 | */ | |
955 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
956 | unsigned long private, struct page *hpage, | |
9c620e2b | 957 | int force, enum migrate_mode mode) |
290408d4 NH |
958 | { |
959 | int rc = 0; | |
960 | int *result = NULL; | |
961 | struct page *new_hpage = get_new_page(hpage, private, &result); | |
290408d4 NH |
962 | struct anon_vma *anon_vma = NULL; |
963 | ||
83467efb NH |
964 | /* |
965 | * Movability of hugepages depends on architectures and hugepage size. | |
966 | * This check is necessary because some callers of hugepage migration | |
967 | * like soft offline and memory hotremove don't walk through page | |
968 | * tables or check whether the hugepage is pmd-based or not before | |
969 | * kicking migration. | |
970 | */ | |
971 | if (!hugepage_migration_support(page_hstate(hpage))) | |
972 | return -ENOSYS; | |
973 | ||
290408d4 NH |
974 | if (!new_hpage) |
975 | return -ENOMEM; | |
976 | ||
977 | rc = -EAGAIN; | |
978 | ||
979 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 980 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
981 | goto out; |
982 | lock_page(hpage); | |
983 | } | |
984 | ||
746b18d4 PZ |
985 | if (PageAnon(hpage)) |
986 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 NH |
987 | |
988 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
989 | ||
990 | if (!page_mapped(hpage)) | |
a6bc32b8 | 991 | rc = move_to_new_page(new_hpage, hpage, 1, mode); |
290408d4 NH |
992 | |
993 | if (rc) | |
994 | remove_migration_ptes(hpage, hpage); | |
995 | ||
fd4a4663 | 996 | if (anon_vma) |
9e60109f | 997 | put_anon_vma(anon_vma); |
8e6ac7fa AK |
998 | |
999 | if (!rc) | |
1000 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1001 | ||
290408d4 | 1002 | unlock_page(hpage); |
09761333 | 1003 | out: |
b8ec1cee NH |
1004 | if (rc != -EAGAIN) |
1005 | putback_active_hugepage(hpage); | |
290408d4 | 1006 | put_page(new_hpage); |
290408d4 NH |
1007 | if (result) { |
1008 | if (rc) | |
1009 | *result = rc; | |
1010 | else | |
1011 | *result = page_to_nid(new_hpage); | |
1012 | } | |
1013 | return rc; | |
1014 | } | |
1015 | ||
b20a3503 | 1016 | /* |
c73e5c9c SB |
1017 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1018 | * supplied as the target for the page migration | |
b20a3503 | 1019 | * |
c73e5c9c SB |
1020 | * @from: The list of pages to be migrated. |
1021 | * @get_new_page: The function used to allocate free pages to be used | |
1022 | * as the target of the page migration. | |
1023 | * @private: Private data to be passed on to get_new_page() | |
1024 | * @mode: The migration mode that specifies the constraints for | |
1025 | * page migration, if any. | |
1026 | * @reason: The reason for page migration. | |
b20a3503 | 1027 | * |
c73e5c9c SB |
1028 | * The function returns after 10 attempts or if no pages are movable any more |
1029 | * because the list has become empty or no retryable pages exist any more. | |
1030 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1031 | * or free list only if ret != 0. |
b20a3503 | 1032 | * |
c73e5c9c | 1033 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1034 | */ |
9c620e2b HD |
1035 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
1036 | unsigned long private, enum migrate_mode mode, int reason) | |
b20a3503 | 1037 | { |
e24f0b8f | 1038 | int retry = 1; |
b20a3503 | 1039 | int nr_failed = 0; |
5647bc29 | 1040 | int nr_succeeded = 0; |
b20a3503 CL |
1041 | int pass = 0; |
1042 | struct page *page; | |
1043 | struct page *page2; | |
1044 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1045 | int rc; | |
1046 | ||
1047 | if (!swapwrite) | |
1048 | current->flags |= PF_SWAPWRITE; | |
1049 | ||
e24f0b8f CL |
1050 | for(pass = 0; pass < 10 && retry; pass++) { |
1051 | retry = 0; | |
b20a3503 | 1052 | |
e24f0b8f | 1053 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1054 | cond_resched(); |
2d1db3b1 | 1055 | |
31caf665 NH |
1056 | if (PageHuge(page)) |
1057 | rc = unmap_and_move_huge_page(get_new_page, | |
1058 | private, page, pass > 2, mode); | |
1059 | else | |
1060 | rc = unmap_and_move(get_new_page, private, | |
9c620e2b | 1061 | page, pass > 2, mode); |
2d1db3b1 | 1062 | |
e24f0b8f | 1063 | switch(rc) { |
95a402c3 CL |
1064 | case -ENOMEM: |
1065 | goto out; | |
e24f0b8f | 1066 | case -EAGAIN: |
2d1db3b1 | 1067 | retry++; |
e24f0b8f | 1068 | break; |
78bd5209 | 1069 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1070 | nr_succeeded++; |
e24f0b8f CL |
1071 | break; |
1072 | default: | |
2d1db3b1 | 1073 | /* Permanent failure */ |
2d1db3b1 | 1074 | nr_failed++; |
e24f0b8f | 1075 | break; |
2d1db3b1 | 1076 | } |
b20a3503 CL |
1077 | } |
1078 | } | |
78bd5209 | 1079 | rc = nr_failed + retry; |
95a402c3 | 1080 | out: |
5647bc29 MG |
1081 | if (nr_succeeded) |
1082 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1083 | if (nr_failed) | |
1084 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1085 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1086 | ||
b20a3503 CL |
1087 | if (!swapwrite) |
1088 | current->flags &= ~PF_SWAPWRITE; | |
1089 | ||
78bd5209 | 1090 | return rc; |
b20a3503 | 1091 | } |
95a402c3 | 1092 | |
742755a1 CL |
1093 | #ifdef CONFIG_NUMA |
1094 | /* | |
1095 | * Move a list of individual pages | |
1096 | */ | |
1097 | struct page_to_node { | |
1098 | unsigned long addr; | |
1099 | struct page *page; | |
1100 | int node; | |
1101 | int status; | |
1102 | }; | |
1103 | ||
1104 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1105 | int **result) | |
1106 | { | |
1107 | struct page_to_node *pm = (struct page_to_node *)private; | |
1108 | ||
1109 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1110 | pm++; | |
1111 | ||
1112 | if (pm->node == MAX_NUMNODES) | |
1113 | return NULL; | |
1114 | ||
1115 | *result = &pm->status; | |
1116 | ||
e632a938 NH |
1117 | if (PageHuge(p)) |
1118 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1119 | pm->node); | |
1120 | else | |
1121 | return alloc_pages_exact_node(pm->node, | |
769848c0 | 1122 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); |
742755a1 CL |
1123 | } |
1124 | ||
1125 | /* | |
1126 | * Move a set of pages as indicated in the pm array. The addr | |
1127 | * field must be set to the virtual address of the page to be moved | |
1128 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1129 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1130 | */ |
5e9a0f02 BG |
1131 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1132 | struct page_to_node *pm, | |
1133 | int migrate_all) | |
742755a1 CL |
1134 | { |
1135 | int err; | |
1136 | struct page_to_node *pp; | |
1137 | LIST_HEAD(pagelist); | |
1138 | ||
1139 | down_read(&mm->mmap_sem); | |
1140 | ||
1141 | /* | |
1142 | * Build a list of pages to migrate | |
1143 | */ | |
742755a1 CL |
1144 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1145 | struct vm_area_struct *vma; | |
1146 | struct page *page; | |
1147 | ||
742755a1 CL |
1148 | err = -EFAULT; |
1149 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1150 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1151 | goto set_status; |
1152 | ||
500d65d4 | 1153 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1154 | |
1155 | err = PTR_ERR(page); | |
1156 | if (IS_ERR(page)) | |
1157 | goto set_status; | |
1158 | ||
742755a1 CL |
1159 | err = -ENOENT; |
1160 | if (!page) | |
1161 | goto set_status; | |
1162 | ||
62b61f61 | 1163 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1164 | if (PageReserved(page)) |
742755a1 CL |
1165 | goto put_and_set; |
1166 | ||
1167 | pp->page = page; | |
1168 | err = page_to_nid(page); | |
1169 | ||
1170 | if (err == pp->node) | |
1171 | /* | |
1172 | * Node already in the right place | |
1173 | */ | |
1174 | goto put_and_set; | |
1175 | ||
1176 | err = -EACCES; | |
1177 | if (page_mapcount(page) > 1 && | |
1178 | !migrate_all) | |
1179 | goto put_and_set; | |
1180 | ||
e632a938 NH |
1181 | if (PageHuge(page)) { |
1182 | isolate_huge_page(page, &pagelist); | |
1183 | goto put_and_set; | |
1184 | } | |
1185 | ||
62695a84 | 1186 | err = isolate_lru_page(page); |
6d9c285a | 1187 | if (!err) { |
62695a84 | 1188 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1189 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1190 | page_is_file_cache(page)); | |
1191 | } | |
742755a1 CL |
1192 | put_and_set: |
1193 | /* | |
1194 | * Either remove the duplicate refcount from | |
1195 | * isolate_lru_page() or drop the page ref if it was | |
1196 | * not isolated. | |
1197 | */ | |
1198 | put_page(page); | |
1199 | set_status: | |
1200 | pp->status = err; | |
1201 | } | |
1202 | ||
e78bbfa8 | 1203 | err = 0; |
cf608ac1 | 1204 | if (!list_empty(&pagelist)) { |
742755a1 | 1205 | err = migrate_pages(&pagelist, new_page_node, |
9c620e2b | 1206 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1207 | if (err) |
e632a938 | 1208 | putback_movable_pages(&pagelist); |
cf608ac1 | 1209 | } |
742755a1 CL |
1210 | |
1211 | up_read(&mm->mmap_sem); | |
1212 | return err; | |
1213 | } | |
1214 | ||
5e9a0f02 BG |
1215 | /* |
1216 | * Migrate an array of page address onto an array of nodes and fill | |
1217 | * the corresponding array of status. | |
1218 | */ | |
3268c63e | 1219 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1220 | unsigned long nr_pages, |
1221 | const void __user * __user *pages, | |
1222 | const int __user *nodes, | |
1223 | int __user *status, int flags) | |
1224 | { | |
3140a227 | 1225 | struct page_to_node *pm; |
3140a227 BG |
1226 | unsigned long chunk_nr_pages; |
1227 | unsigned long chunk_start; | |
1228 | int err; | |
5e9a0f02 | 1229 | |
3140a227 BG |
1230 | err = -ENOMEM; |
1231 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1232 | if (!pm) | |
5e9a0f02 | 1233 | goto out; |
35282a2d BG |
1234 | |
1235 | migrate_prep(); | |
1236 | ||
5e9a0f02 | 1237 | /* |
3140a227 BG |
1238 | * Store a chunk of page_to_node array in a page, |
1239 | * but keep the last one as a marker | |
5e9a0f02 | 1240 | */ |
3140a227 | 1241 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1242 | |
3140a227 BG |
1243 | for (chunk_start = 0; |
1244 | chunk_start < nr_pages; | |
1245 | chunk_start += chunk_nr_pages) { | |
1246 | int j; | |
5e9a0f02 | 1247 | |
3140a227 BG |
1248 | if (chunk_start + chunk_nr_pages > nr_pages) |
1249 | chunk_nr_pages = nr_pages - chunk_start; | |
1250 | ||
1251 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1252 | for (j = 0; j < chunk_nr_pages; j++) { | |
1253 | const void __user *p; | |
5e9a0f02 BG |
1254 | int node; |
1255 | ||
3140a227 BG |
1256 | err = -EFAULT; |
1257 | if (get_user(p, pages + j + chunk_start)) | |
1258 | goto out_pm; | |
1259 | pm[j].addr = (unsigned long) p; | |
1260 | ||
1261 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1262 | goto out_pm; |
1263 | ||
1264 | err = -ENODEV; | |
6f5a55f1 LT |
1265 | if (node < 0 || node >= MAX_NUMNODES) |
1266 | goto out_pm; | |
1267 | ||
389162c2 | 1268 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1269 | goto out_pm; |
1270 | ||
1271 | err = -EACCES; | |
1272 | if (!node_isset(node, task_nodes)) | |
1273 | goto out_pm; | |
1274 | ||
3140a227 BG |
1275 | pm[j].node = node; |
1276 | } | |
1277 | ||
1278 | /* End marker for this chunk */ | |
1279 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1280 | ||
1281 | /* Migrate this chunk */ | |
1282 | err = do_move_page_to_node_array(mm, pm, | |
1283 | flags & MPOL_MF_MOVE_ALL); | |
1284 | if (err < 0) | |
1285 | goto out_pm; | |
5e9a0f02 | 1286 | |
5e9a0f02 | 1287 | /* Return status information */ |
3140a227 BG |
1288 | for (j = 0; j < chunk_nr_pages; j++) |
1289 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1290 | err = -EFAULT; |
3140a227 BG |
1291 | goto out_pm; |
1292 | } | |
1293 | } | |
1294 | err = 0; | |
5e9a0f02 BG |
1295 | |
1296 | out_pm: | |
3140a227 | 1297 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1298 | out: |
1299 | return err; | |
1300 | } | |
1301 | ||
742755a1 | 1302 | /* |
2f007e74 | 1303 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1304 | */ |
80bba129 BG |
1305 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1306 | const void __user **pages, int *status) | |
742755a1 | 1307 | { |
2f007e74 | 1308 | unsigned long i; |
2f007e74 | 1309 | |
742755a1 CL |
1310 | down_read(&mm->mmap_sem); |
1311 | ||
2f007e74 | 1312 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1313 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1314 | struct vm_area_struct *vma; |
1315 | struct page *page; | |
c095adbc | 1316 | int err = -EFAULT; |
2f007e74 BG |
1317 | |
1318 | vma = find_vma(mm, addr); | |
70384dc6 | 1319 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1320 | goto set_status; |
1321 | ||
2f007e74 | 1322 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1323 | |
1324 | err = PTR_ERR(page); | |
1325 | if (IS_ERR(page)) | |
1326 | goto set_status; | |
1327 | ||
742755a1 CL |
1328 | err = -ENOENT; |
1329 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1330 | if (!page || PageReserved(page)) |
742755a1 CL |
1331 | goto set_status; |
1332 | ||
1333 | err = page_to_nid(page); | |
1334 | set_status: | |
80bba129 BG |
1335 | *status = err; |
1336 | ||
1337 | pages++; | |
1338 | status++; | |
1339 | } | |
1340 | ||
1341 | up_read(&mm->mmap_sem); | |
1342 | } | |
1343 | ||
1344 | /* | |
1345 | * Determine the nodes of a user array of pages and store it in | |
1346 | * a user array of status. | |
1347 | */ | |
1348 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1349 | const void __user * __user *pages, | |
1350 | int __user *status) | |
1351 | { | |
1352 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1353 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1354 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1355 | |
87b8d1ad PA |
1356 | while (nr_pages) { |
1357 | unsigned long chunk_nr; | |
80bba129 | 1358 | |
87b8d1ad PA |
1359 | chunk_nr = nr_pages; |
1360 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1361 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1362 | ||
1363 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1364 | break; | |
80bba129 BG |
1365 | |
1366 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1367 | ||
87b8d1ad PA |
1368 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1369 | break; | |
742755a1 | 1370 | |
87b8d1ad PA |
1371 | pages += chunk_nr; |
1372 | status += chunk_nr; | |
1373 | nr_pages -= chunk_nr; | |
1374 | } | |
1375 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1376 | } |
1377 | ||
1378 | /* | |
1379 | * Move a list of pages in the address space of the currently executing | |
1380 | * process. | |
1381 | */ | |
938bb9f5 HC |
1382 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1383 | const void __user * __user *, pages, | |
1384 | const int __user *, nodes, | |
1385 | int __user *, status, int, flags) | |
742755a1 | 1386 | { |
c69e8d9c | 1387 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1388 | struct task_struct *task; |
742755a1 | 1389 | struct mm_struct *mm; |
5e9a0f02 | 1390 | int err; |
3268c63e | 1391 | nodemask_t task_nodes; |
742755a1 CL |
1392 | |
1393 | /* Check flags */ | |
1394 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1395 | return -EINVAL; | |
1396 | ||
1397 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1398 | return -EPERM; | |
1399 | ||
1400 | /* Find the mm_struct */ | |
a879bf58 | 1401 | rcu_read_lock(); |
228ebcbe | 1402 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1403 | if (!task) { |
a879bf58 | 1404 | rcu_read_unlock(); |
742755a1 CL |
1405 | return -ESRCH; |
1406 | } | |
3268c63e | 1407 | get_task_struct(task); |
742755a1 CL |
1408 | |
1409 | /* | |
1410 | * Check if this process has the right to modify the specified | |
1411 | * process. The right exists if the process has administrative | |
1412 | * capabilities, superuser privileges or the same | |
1413 | * userid as the target process. | |
1414 | */ | |
c69e8d9c | 1415 | tcred = __task_cred(task); |
b38a86eb EB |
1416 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1417 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1418 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1419 | rcu_read_unlock(); |
742755a1 | 1420 | err = -EPERM; |
5e9a0f02 | 1421 | goto out; |
742755a1 | 1422 | } |
c69e8d9c | 1423 | rcu_read_unlock(); |
742755a1 | 1424 | |
86c3a764 DQ |
1425 | err = security_task_movememory(task); |
1426 | if (err) | |
5e9a0f02 | 1427 | goto out; |
86c3a764 | 1428 | |
3268c63e CL |
1429 | task_nodes = cpuset_mems_allowed(task); |
1430 | mm = get_task_mm(task); | |
1431 | put_task_struct(task); | |
1432 | ||
6e8b09ea SL |
1433 | if (!mm) |
1434 | return -EINVAL; | |
1435 | ||
1436 | if (nodes) | |
1437 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1438 | nodes, status, flags); | |
1439 | else | |
1440 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1441 | |
742755a1 CL |
1442 | mmput(mm); |
1443 | return err; | |
3268c63e CL |
1444 | |
1445 | out: | |
1446 | put_task_struct(task); | |
1447 | return err; | |
742755a1 | 1448 | } |
742755a1 | 1449 | |
7b2259b3 CL |
1450 | /* |
1451 | * Call migration functions in the vma_ops that may prepare | |
1452 | * memory in a vm for migration. migration functions may perform | |
1453 | * the migration for vmas that do not have an underlying page struct. | |
1454 | */ | |
1455 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1456 | const nodemask_t *from, unsigned long flags) | |
1457 | { | |
1458 | struct vm_area_struct *vma; | |
1459 | int err = 0; | |
1460 | ||
1001c9fb | 1461 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { |
7b2259b3 CL |
1462 | if (vma->vm_ops && vma->vm_ops->migrate) { |
1463 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1464 | if (err) | |
1465 | break; | |
1466 | } | |
1467 | } | |
1468 | return err; | |
1469 | } | |
7039e1db PZ |
1470 | |
1471 | #ifdef CONFIG_NUMA_BALANCING | |
1472 | /* | |
1473 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1474 | * pages. Currently it only checks the watermarks which crude | |
1475 | */ | |
1476 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1477 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1478 | { |
1479 | int z; | |
1480 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1481 | struct zone *zone = pgdat->node_zones + z; | |
1482 | ||
1483 | if (!populated_zone(zone)) | |
1484 | continue; | |
1485 | ||
6e543d57 | 1486 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1487 | continue; |
1488 | ||
1489 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1490 | if (!zone_watermark_ok(zone, 0, | |
1491 | high_wmark_pages(zone) + | |
1492 | nr_migrate_pages, | |
1493 | 0, 0)) | |
1494 | continue; | |
1495 | return true; | |
1496 | } | |
1497 | return false; | |
1498 | } | |
1499 | ||
1500 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1501 | unsigned long data, | |
1502 | int **result) | |
1503 | { | |
1504 | int nid = (int) data; | |
1505 | struct page *newpage; | |
1506 | ||
1507 | newpage = alloc_pages_exact_node(nid, | |
1508 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1509 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1510 | __GFP_NOWARN) & | |
1511 | ~GFP_IOFS, 0); | |
bac0382c | 1512 | if (newpage) |
90572890 | 1513 | page_cpupid_xchg_last(newpage, page_cpupid_last(page)); |
bac0382c | 1514 | |
7039e1db PZ |
1515 | return newpage; |
1516 | } | |
1517 | ||
a8f60772 MG |
1518 | /* |
1519 | * page migration rate limiting control. | |
1520 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1521 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1522 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1523 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1524 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1525 | * throttle window closed. | |
a8f60772 MG |
1526 | */ |
1527 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1528 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1529 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1530 | ||
e14808b4 MG |
1531 | /* Returns true if NUMA migration is currently rate limited */ |
1532 | bool migrate_ratelimited(int node) | |
1533 | { | |
1534 | pg_data_t *pgdat = NODE_DATA(node); | |
1535 | ||
1536 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1537 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1538 | return false; | |
1539 | ||
1540 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1541 | return false; | |
1542 | ||
1543 | return true; | |
1544 | } | |
1545 | ||
b32967ff | 1546 | /* Returns true if the node is migrate rate-limited after the update */ |
d28d4335 | 1547 | bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages) |
7039e1db | 1548 | { |
b32967ff | 1549 | bool rate_limited = false; |
7039e1db | 1550 | |
a8f60772 MG |
1551 | /* |
1552 | * Rate-limit the amount of data that is being migrated to a node. | |
1553 | * Optimal placement is no good if the memory bus is saturated and | |
1554 | * all the time is being spent migrating! | |
1555 | */ | |
1556 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1557 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1558 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1559 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1560 | msecs_to_jiffies(migrate_interval_millisecs); | |
1561 | } | |
b32967ff MG |
1562 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) |
1563 | rate_limited = true; | |
1564 | else | |
d28d4335 | 1565 | pgdat->numabalancing_migrate_nr_pages += nr_pages; |
a8f60772 | 1566 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
b32967ff MG |
1567 | |
1568 | return rate_limited; | |
1569 | } | |
1570 | ||
1571 | int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1572 | { | |
340ef390 | 1573 | int page_lru; |
a8f60772 | 1574 | |
3abef4e6 MG |
1575 | VM_BUG_ON(compound_order(page) && !PageTransHuge(page)); |
1576 | ||
7039e1db | 1577 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1578 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1579 | return 0; | |
7039e1db | 1580 | |
340ef390 HD |
1581 | if (isolate_lru_page(page)) |
1582 | return 0; | |
7039e1db | 1583 | |
340ef390 HD |
1584 | /* |
1585 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1586 | * check on page_count(), so we must do it here, now that the page | |
1587 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1588 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1589 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1590 | */ | |
1591 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1592 | putback_lru_page(page); | |
1593 | return 0; | |
7039e1db PZ |
1594 | } |
1595 | ||
340ef390 HD |
1596 | page_lru = page_is_file_cache(page); |
1597 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1598 | hpage_nr_pages(page)); | |
1599 | ||
149c33e1 | 1600 | /* |
340ef390 HD |
1601 | * Isolating the page has taken another reference, so the |
1602 | * caller's reference can be safely dropped without the page | |
1603 | * disappearing underneath us during migration. | |
149c33e1 MG |
1604 | */ |
1605 | put_page(page); | |
340ef390 | 1606 | return 1; |
b32967ff MG |
1607 | } |
1608 | ||
1609 | /* | |
1610 | * Attempt to migrate a misplaced page to the specified destination | |
1611 | * node. Caller is expected to have an elevated reference count on | |
1612 | * the page that will be dropped by this function before returning. | |
1613 | */ | |
1bc115d8 MG |
1614 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1615 | int node) | |
b32967ff MG |
1616 | { |
1617 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1618 | int isolated; |
b32967ff MG |
1619 | int nr_remaining; |
1620 | LIST_HEAD(migratepages); | |
1621 | ||
1622 | /* | |
1bc115d8 MG |
1623 | * Don't migrate file pages that are mapped in multiple processes |
1624 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1625 | */ |
1bc115d8 MG |
1626 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1627 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1628 | goto out; |
b32967ff MG |
1629 | |
1630 | /* | |
1631 | * Rate-limit the amount of data that is being migrated to a node. | |
1632 | * Optimal placement is no good if the memory bus is saturated and | |
1633 | * all the time is being spent migrating! | |
1634 | */ | |
340ef390 | 1635 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1636 | goto out; |
b32967ff MG |
1637 | |
1638 | isolated = numamigrate_isolate_page(pgdat, page); | |
1639 | if (!isolated) | |
1640 | goto out; | |
1641 | ||
1642 | list_add(&page->lru, &migratepages); | |
9c620e2b HD |
1643 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
1644 | node, MIGRATE_ASYNC, MR_NUMA_MISPLACED); | |
b32967ff MG |
1645 | if (nr_remaining) { |
1646 | putback_lru_pages(&migratepages); | |
1647 | isolated = 0; | |
1648 | } else | |
1649 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1650 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1651 | return isolated; |
340ef390 HD |
1652 | |
1653 | out: | |
1654 | put_page(page); | |
1655 | return 0; | |
7039e1db | 1656 | } |
220018d3 | 1657 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1658 | |
220018d3 | 1659 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1660 | /* |
1661 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1662 | * before returning. | |
1663 | */ | |
b32967ff MG |
1664 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1665 | struct vm_area_struct *vma, | |
1666 | pmd_t *pmd, pmd_t entry, | |
1667 | unsigned long address, | |
1668 | struct page *page, int node) | |
1669 | { | |
c4088ebd | 1670 | spinlock_t *ptl; |
b32967ff MG |
1671 | unsigned long haddr = address & HPAGE_PMD_MASK; |
1672 | pg_data_t *pgdat = NODE_DATA(node); | |
1673 | int isolated = 0; | |
1674 | struct page *new_page = NULL; | |
1675 | struct mem_cgroup *memcg = NULL; | |
1676 | int page_lru = page_is_file_cache(page); | |
1677 | ||
b32967ff MG |
1678 | /* |
1679 | * Rate-limit the amount of data that is being migrated to a node. | |
1680 | * Optimal placement is no good if the memory bus is saturated and | |
1681 | * all the time is being spent migrating! | |
1682 | */ | |
d28d4335 | 1683 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1684 | goto out_dropref; |
1685 | ||
1686 | new_page = alloc_pages_node(node, | |
1687 | (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER); | |
340ef390 HD |
1688 | if (!new_page) |
1689 | goto out_fail; | |
1690 | ||
90572890 | 1691 | page_cpupid_xchg_last(new_page, page_cpupid_last(page)); |
b32967ff MG |
1692 | |
1693 | isolated = numamigrate_isolate_page(pgdat, page); | |
340ef390 | 1694 | if (!isolated) { |
b32967ff | 1695 | put_page(new_page); |
340ef390 | 1696 | goto out_fail; |
b32967ff MG |
1697 | } |
1698 | ||
1699 | /* Prepare a page as a migration target */ | |
1700 | __set_page_locked(new_page); | |
1701 | SetPageSwapBacked(new_page); | |
1702 | ||
1703 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1704 | new_page->mapping = page->mapping; | |
1705 | new_page->index = page->index; | |
1706 | migrate_page_copy(new_page, page); | |
1707 | WARN_ON(PageLRU(new_page)); | |
1708 | ||
1709 | /* Recheck the target PMD */ | |
c4088ebd | 1710 | ptl = pmd_lock(mm, pmd); |
b32967ff | 1711 | if (unlikely(!pmd_same(*pmd, entry))) { |
c4088ebd | 1712 | spin_unlock(ptl); |
b32967ff MG |
1713 | |
1714 | /* Reverse changes made by migrate_page_copy() */ | |
1715 | if (TestClearPageActive(new_page)) | |
1716 | SetPageActive(page); | |
1717 | if (TestClearPageUnevictable(new_page)) | |
1718 | SetPageUnevictable(page); | |
1719 | mlock_migrate_page(page, new_page); | |
1720 | ||
1721 | unlock_page(new_page); | |
1722 | put_page(new_page); /* Free it */ | |
1723 | ||
a54a407f MG |
1724 | /* Retake the callers reference and putback on LRU */ |
1725 | get_page(page); | |
b32967ff | 1726 | putback_lru_page(page); |
a54a407f MG |
1727 | mod_zone_page_state(page_zone(page), |
1728 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
1729 | goto out_fail; | |
b32967ff MG |
1730 | } |
1731 | ||
1732 | /* | |
1733 | * Traditional migration needs to prepare the memcg charge | |
1734 | * transaction early to prevent the old page from being | |
1735 | * uncharged when installing migration entries. Here we can | |
1736 | * save the potential rollback and start the charge transfer | |
1737 | * only when migration is already known to end successfully. | |
1738 | */ | |
1739 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1740 | ||
1741 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
1742 | entry = pmd_mknonnuma(entry); | |
1743 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1744 | entry = pmd_mkhuge(entry); | |
1745 | ||
a54a407f | 1746 | pmdp_clear_flush(vma, haddr, pmd); |
b32967ff | 1747 | set_pmd_at(mm, haddr, pmd, entry); |
a54a407f | 1748 | page_add_new_anon_rmap(new_page, vma, haddr); |
ce4a9cc5 | 1749 | update_mmu_cache_pmd(vma, address, &entry); |
b32967ff MG |
1750 | page_remove_rmap(page); |
1751 | /* | |
1752 | * Finish the charge transaction under the page table lock to | |
1753 | * prevent split_huge_page() from dividing up the charge | |
1754 | * before it's fully transferred to the new page. | |
1755 | */ | |
1756 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
c4088ebd | 1757 | spin_unlock(ptl); |
b32967ff MG |
1758 | |
1759 | unlock_page(new_page); | |
1760 | unlock_page(page); | |
1761 | put_page(page); /* Drop the rmap reference */ | |
1762 | put_page(page); /* Drop the LRU isolation reference */ | |
1763 | ||
1764 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1765 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1766 | ||
b32967ff MG |
1767 | mod_zone_page_state(page_zone(page), |
1768 | NR_ISOLATED_ANON + page_lru, | |
1769 | -HPAGE_PMD_NR); | |
1770 | return isolated; | |
1771 | ||
340ef390 HD |
1772 | out_fail: |
1773 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1774 | out_dropref: |
a54a407f MG |
1775 | entry = pmd_mknonnuma(entry); |
1776 | set_pmd_at(mm, haddr, pmd, entry); | |
1777 | update_mmu_cache_pmd(vma, address, &entry); | |
1778 | ||
340ef390 | 1779 | unlock_page(page); |
b32967ff | 1780 | put_page(page); |
b32967ff MG |
1781 | return 0; |
1782 | } | |
7039e1db PZ |
1783 | #endif /* CONFIG_NUMA_BALANCING */ |
1784 | ||
1785 | #endif /* CONFIG_NUMA */ |