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 | ||
30b0a105 DH |
444 | /* |
445 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
446 | * arithmetic will work across the entire page. We need something more | |
447 | * specialized. | |
448 | */ | |
449 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
450 | int nr_pages) | |
451 | { | |
452 | int i; | |
453 | struct page *dst_base = dst; | |
454 | struct page *src_base = src; | |
455 | ||
456 | for (i = 0; i < nr_pages; ) { | |
457 | cond_resched(); | |
458 | copy_highpage(dst, src); | |
459 | ||
460 | i++; | |
461 | dst = mem_map_next(dst, dst_base, i); | |
462 | src = mem_map_next(src, src_base, i); | |
463 | } | |
464 | } | |
465 | ||
466 | static void copy_huge_page(struct page *dst, struct page *src) | |
467 | { | |
468 | int i; | |
469 | int nr_pages; | |
470 | ||
471 | if (PageHuge(src)) { | |
472 | /* hugetlbfs page */ | |
473 | struct hstate *h = page_hstate(src); | |
474 | nr_pages = pages_per_huge_page(h); | |
475 | ||
476 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
477 | __copy_gigantic_page(dst, src, nr_pages); | |
478 | return; | |
479 | } | |
480 | } else { | |
481 | /* thp page */ | |
482 | BUG_ON(!PageTransHuge(src)); | |
483 | nr_pages = hpage_nr_pages(src); | |
484 | } | |
485 | ||
486 | for (i = 0; i < nr_pages; i++) { | |
487 | cond_resched(); | |
488 | copy_highpage(dst + i, src + i); | |
489 | } | |
490 | } | |
491 | ||
b20a3503 CL |
492 | /* |
493 | * Copy the page to its new location | |
494 | */ | |
290408d4 | 495 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 496 | { |
7851a45c RR |
497 | int cpupid; |
498 | ||
b32967ff | 499 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
500 | copy_huge_page(newpage, page); |
501 | else | |
502 | copy_highpage(newpage, page); | |
b20a3503 CL |
503 | |
504 | if (PageError(page)) | |
505 | SetPageError(newpage); | |
506 | if (PageReferenced(page)) | |
507 | SetPageReferenced(newpage); | |
508 | if (PageUptodate(page)) | |
509 | SetPageUptodate(newpage); | |
894bc310 LS |
510 | if (TestClearPageActive(page)) { |
511 | VM_BUG_ON(PageUnevictable(page)); | |
b20a3503 | 512 | SetPageActive(newpage); |
418b27ef LS |
513 | } else if (TestClearPageUnevictable(page)) |
514 | SetPageUnevictable(newpage); | |
b20a3503 CL |
515 | if (PageChecked(page)) |
516 | SetPageChecked(newpage); | |
517 | if (PageMappedToDisk(page)) | |
518 | SetPageMappedToDisk(newpage); | |
519 | ||
520 | if (PageDirty(page)) { | |
521 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
522 | /* |
523 | * Want to mark the page and the radix tree as dirty, and | |
524 | * redo the accounting that clear_page_dirty_for_io undid, | |
525 | * but we can't use set_page_dirty because that function | |
526 | * is actually a signal that all of the page has become dirty. | |
25985edc | 527 | * Whereas only part of our page may be dirty. |
3a902c5f | 528 | */ |
752dc185 HD |
529 | if (PageSwapBacked(page)) |
530 | SetPageDirty(newpage); | |
531 | else | |
532 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
533 | } |
534 | ||
7851a45c RR |
535 | /* |
536 | * Copy NUMA information to the new page, to prevent over-eager | |
537 | * future migrations of this same page. | |
538 | */ | |
539 | cpupid = page_cpupid_xchg_last(page, -1); | |
540 | page_cpupid_xchg_last(newpage, cpupid); | |
541 | ||
b291f000 | 542 | mlock_migrate_page(newpage, page); |
e9995ef9 | 543 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
544 | /* |
545 | * Please do not reorder this without considering how mm/ksm.c's | |
546 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
547 | */ | |
b20a3503 | 548 | ClearPageSwapCache(page); |
b20a3503 CL |
549 | ClearPagePrivate(page); |
550 | set_page_private(page, 0); | |
b20a3503 CL |
551 | |
552 | /* | |
553 | * If any waiters have accumulated on the new page then | |
554 | * wake them up. | |
555 | */ | |
556 | if (PageWriteback(newpage)) | |
557 | end_page_writeback(newpage); | |
558 | } | |
b20a3503 | 559 | |
1d8b85cc CL |
560 | /************************************************************ |
561 | * Migration functions | |
562 | ***********************************************************/ | |
563 | ||
564 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
565 | int fail_migrate_page(struct address_space *mapping, |
566 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
567 | { |
568 | return -EIO; | |
569 | } | |
570 | EXPORT_SYMBOL(fail_migrate_page); | |
571 | ||
b20a3503 CL |
572 | /* |
573 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 574 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
575 | * |
576 | * Pages are locked upon entry and exit. | |
577 | */ | |
2d1db3b1 | 578 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
579 | struct page *newpage, struct page *page, |
580 | enum migrate_mode mode) | |
b20a3503 CL |
581 | { |
582 | int rc; | |
583 | ||
584 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
585 | ||
a6bc32b8 | 586 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode); |
b20a3503 | 587 | |
78bd5209 | 588 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
589 | return rc; |
590 | ||
591 | migrate_page_copy(newpage, page); | |
78bd5209 | 592 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
593 | } |
594 | EXPORT_SYMBOL(migrate_page); | |
595 | ||
9361401e | 596 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
597 | /* |
598 | * Migration function for pages with buffers. This function can only be used | |
599 | * if the underlying filesystem guarantees that no other references to "page" | |
600 | * exist. | |
601 | */ | |
2d1db3b1 | 602 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 603 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 604 | { |
1d8b85cc CL |
605 | struct buffer_head *bh, *head; |
606 | int rc; | |
607 | ||
1d8b85cc | 608 | if (!page_has_buffers(page)) |
a6bc32b8 | 609 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
610 | |
611 | head = page_buffers(page); | |
612 | ||
a6bc32b8 | 613 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode); |
1d8b85cc | 614 | |
78bd5209 | 615 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
616 | return rc; |
617 | ||
b969c4ab MG |
618 | /* |
619 | * In the async case, migrate_page_move_mapping locked the buffers | |
620 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
621 | * need to be locked now | |
622 | */ | |
a6bc32b8 MG |
623 | if (mode != MIGRATE_ASYNC) |
624 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
625 | |
626 | ClearPagePrivate(page); | |
627 | set_page_private(newpage, page_private(page)); | |
628 | set_page_private(page, 0); | |
629 | put_page(page); | |
630 | get_page(newpage); | |
631 | ||
632 | bh = head; | |
633 | do { | |
634 | set_bh_page(bh, newpage, bh_offset(bh)); | |
635 | bh = bh->b_this_page; | |
636 | ||
637 | } while (bh != head); | |
638 | ||
639 | SetPagePrivate(newpage); | |
640 | ||
641 | migrate_page_copy(newpage, page); | |
642 | ||
643 | bh = head; | |
644 | do { | |
645 | unlock_buffer(bh); | |
646 | put_bh(bh); | |
647 | bh = bh->b_this_page; | |
648 | ||
649 | } while (bh != head); | |
650 | ||
78bd5209 | 651 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
652 | } |
653 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 654 | #endif |
1d8b85cc | 655 | |
04e62a29 CL |
656 | /* |
657 | * Writeback a page to clean the dirty state | |
658 | */ | |
659 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 660 | { |
04e62a29 CL |
661 | struct writeback_control wbc = { |
662 | .sync_mode = WB_SYNC_NONE, | |
663 | .nr_to_write = 1, | |
664 | .range_start = 0, | |
665 | .range_end = LLONG_MAX, | |
04e62a29 CL |
666 | .for_reclaim = 1 |
667 | }; | |
668 | int rc; | |
669 | ||
670 | if (!mapping->a_ops->writepage) | |
671 | /* No write method for the address space */ | |
672 | return -EINVAL; | |
673 | ||
674 | if (!clear_page_dirty_for_io(page)) | |
675 | /* Someone else already triggered a write */ | |
676 | return -EAGAIN; | |
677 | ||
8351a6e4 | 678 | /* |
04e62a29 CL |
679 | * A dirty page may imply that the underlying filesystem has |
680 | * the page on some queue. So the page must be clean for | |
681 | * migration. Writeout may mean we loose the lock and the | |
682 | * page state is no longer what we checked for earlier. | |
683 | * At this point we know that the migration attempt cannot | |
684 | * be successful. | |
8351a6e4 | 685 | */ |
04e62a29 | 686 | remove_migration_ptes(page, page); |
8351a6e4 | 687 | |
04e62a29 | 688 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 689 | |
04e62a29 CL |
690 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
691 | /* unlocked. Relock */ | |
692 | lock_page(page); | |
693 | ||
bda8550d | 694 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
695 | } |
696 | ||
697 | /* | |
698 | * Default handling if a filesystem does not provide a migration function. | |
699 | */ | |
700 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 701 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 702 | { |
b969c4ab | 703 | if (PageDirty(page)) { |
a6bc32b8 MG |
704 | /* Only writeback pages in full synchronous migration */ |
705 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 706 | return -EBUSY; |
04e62a29 | 707 | return writeout(mapping, page); |
b969c4ab | 708 | } |
8351a6e4 CL |
709 | |
710 | /* | |
711 | * Buffers may be managed in a filesystem specific way. | |
712 | * We must have no buffers or drop them. | |
713 | */ | |
266cf658 | 714 | if (page_has_private(page) && |
8351a6e4 CL |
715 | !try_to_release_page(page, GFP_KERNEL)) |
716 | return -EAGAIN; | |
717 | ||
a6bc32b8 | 718 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
719 | } |
720 | ||
e24f0b8f CL |
721 | /* |
722 | * Move a page to a newly allocated page | |
723 | * The page is locked and all ptes have been successfully removed. | |
724 | * | |
725 | * The new page will have replaced the old page if this function | |
726 | * is successful. | |
894bc310 LS |
727 | * |
728 | * Return value: | |
729 | * < 0 - error code | |
78bd5209 | 730 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 731 | */ |
3fe2011f | 732 | static int move_to_new_page(struct page *newpage, struct page *page, |
a6bc32b8 | 733 | int remap_swapcache, enum migrate_mode mode) |
e24f0b8f CL |
734 | { |
735 | struct address_space *mapping; | |
736 | int rc; | |
737 | ||
738 | /* | |
739 | * Block others from accessing the page when we get around to | |
740 | * establishing additional references. We are the only one | |
741 | * holding a reference to the new page at this point. | |
742 | */ | |
529ae9aa | 743 | if (!trylock_page(newpage)) |
e24f0b8f CL |
744 | BUG(); |
745 | ||
746 | /* Prepare mapping for the new page.*/ | |
747 | newpage->index = page->index; | |
748 | newpage->mapping = page->mapping; | |
b2e18538 RR |
749 | if (PageSwapBacked(page)) |
750 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
751 | |
752 | mapping = page_mapping(page); | |
753 | if (!mapping) | |
a6bc32b8 | 754 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 755 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 756 | /* |
b969c4ab MG |
757 | * Most pages have a mapping and most filesystems provide a |
758 | * migratepage callback. Anonymous pages are part of swap | |
759 | * space which also has its own migratepage callback. This | |
760 | * is the most common path for page migration. | |
e24f0b8f | 761 | */ |
b969c4ab | 762 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 763 | newpage, page, mode); |
b969c4ab | 764 | else |
a6bc32b8 | 765 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 766 | |
78bd5209 | 767 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 768 | newpage->mapping = NULL; |
3fe2011f MG |
769 | } else { |
770 | if (remap_swapcache) | |
771 | remove_migration_ptes(page, newpage); | |
35512eca | 772 | page->mapping = NULL; |
3fe2011f | 773 | } |
e24f0b8f CL |
774 | |
775 | unlock_page(newpage); | |
776 | ||
777 | return rc; | |
778 | } | |
779 | ||
0dabec93 | 780 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 781 | int force, enum migrate_mode mode) |
e24f0b8f | 782 | { |
0dabec93 | 783 | int rc = -EAGAIN; |
3fe2011f | 784 | int remap_swapcache = 1; |
56039efa | 785 | struct mem_cgroup *mem; |
3f6c8272 | 786 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 787 | |
529ae9aa | 788 | if (!trylock_page(page)) { |
a6bc32b8 | 789 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 790 | goto out; |
3e7d3449 MG |
791 | |
792 | /* | |
793 | * It's not safe for direct compaction to call lock_page. | |
794 | * For example, during page readahead pages are added locked | |
795 | * to the LRU. Later, when the IO completes the pages are | |
796 | * marked uptodate and unlocked. However, the queueing | |
797 | * could be merging multiple pages for one bio (e.g. | |
798 | * mpage_readpages). If an allocation happens for the | |
799 | * second or third page, the process can end up locking | |
800 | * the same page twice and deadlocking. Rather than | |
801 | * trying to be clever about what pages can be locked, | |
802 | * avoid the use of lock_page for direct compaction | |
803 | * altogether. | |
804 | */ | |
805 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 806 | goto out; |
3e7d3449 | 807 | |
e24f0b8f CL |
808 | lock_page(page); |
809 | } | |
810 | ||
01b1ae63 | 811 | /* charge against new page */ |
0030f535 | 812 | mem_cgroup_prepare_migration(page, newpage, &mem); |
01b1ae63 | 813 | |
e24f0b8f | 814 | if (PageWriteback(page)) { |
11bc82d6 | 815 | /* |
fed5b64a | 816 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
817 | * necessary to wait for PageWriteback. In the async case, |
818 | * the retry loop is too short and in the sync-light case, | |
819 | * the overhead of stalling is too much | |
11bc82d6 | 820 | */ |
a6bc32b8 | 821 | if (mode != MIGRATE_SYNC) { |
11bc82d6 AA |
822 | rc = -EBUSY; |
823 | goto uncharge; | |
824 | } | |
825 | if (!force) | |
01b1ae63 | 826 | goto uncharge; |
e24f0b8f CL |
827 | wait_on_page_writeback(page); |
828 | } | |
e24f0b8f | 829 | /* |
dc386d4d KH |
830 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
831 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 832 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 833 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
834 | * File Caches may use write_page() or lock_page() in migration, then, |
835 | * just care Anon page here. | |
dc386d4d | 836 | */ |
b79bc0a0 | 837 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 838 | /* |
4fc3f1d6 | 839 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
840 | * getting a hold on an anon_vma from outside one of its mms. |
841 | */ | |
746b18d4 | 842 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
843 | if (anon_vma) { |
844 | /* | |
746b18d4 | 845 | * Anon page |
1ce82b69 | 846 | */ |
1ce82b69 | 847 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
848 | /* |
849 | * We cannot be sure that the anon_vma of an unmapped | |
850 | * swapcache page is safe to use because we don't | |
851 | * know in advance if the VMA that this page belonged | |
852 | * to still exists. If the VMA and others sharing the | |
853 | * data have been freed, then the anon_vma could | |
854 | * already be invalid. | |
855 | * | |
856 | * To avoid this possibility, swapcache pages get | |
857 | * migrated but are not remapped when migration | |
858 | * completes | |
859 | */ | |
860 | remap_swapcache = 0; | |
861 | } else { | |
1ce82b69 | 862 | goto uncharge; |
3fe2011f | 863 | } |
989f89c5 | 864 | } |
62e1c553 | 865 | |
bf6bddf1 RA |
866 | if (unlikely(balloon_page_movable(page))) { |
867 | /* | |
868 | * A ballooned page does not need any special attention from | |
869 | * physical to virtual reverse mapping procedures. | |
870 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
871 | * in order to avoid burning cycles at rmap level, and perform | |
872 | * the page migration right away (proteced by page lock). | |
873 | */ | |
874 | rc = balloon_page_migrate(newpage, page, mode); | |
875 | goto uncharge; | |
876 | } | |
877 | ||
dc386d4d | 878 | /* |
62e1c553 SL |
879 | * Corner case handling: |
880 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
881 | * and treated as swapcache but it has no rmap yet. | |
882 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
883 | * trigger a BUG. So handle it here. | |
884 | * 2. An orphaned page (see truncate_complete_page) might have | |
885 | * fs-private metadata. The page can be picked up due to memory | |
886 | * offlining. Everywhere else except page reclaim, the page is | |
887 | * invisible to the vm, so the page can not be migrated. So try to | |
888 | * free the metadata, so the page can be freed. | |
e24f0b8f | 889 | */ |
62e1c553 | 890 | if (!page->mapping) { |
1ce82b69 HD |
891 | VM_BUG_ON(PageAnon(page)); |
892 | if (page_has_private(page)) { | |
62e1c553 | 893 | try_to_free_buffers(page); |
1ce82b69 | 894 | goto uncharge; |
62e1c553 | 895 | } |
abfc3488 | 896 | goto skip_unmap; |
62e1c553 SL |
897 | } |
898 | ||
dc386d4d | 899 | /* Establish migration ptes or remove ptes */ |
14fa31b8 | 900 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
dc386d4d | 901 | |
abfc3488 | 902 | skip_unmap: |
e6a1530d | 903 | if (!page_mapped(page)) |
a6bc32b8 | 904 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); |
e24f0b8f | 905 | |
3fe2011f | 906 | if (rc && remap_swapcache) |
e24f0b8f | 907 | remove_migration_ptes(page, page); |
3f6c8272 MG |
908 | |
909 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 910 | if (anon_vma) |
9e60109f | 911 | put_anon_vma(anon_vma); |
3f6c8272 | 912 | |
01b1ae63 | 913 | uncharge: |
bf6bddf1 RA |
914 | mem_cgroup_end_migration(mem, page, newpage, |
915 | (rc == MIGRATEPAGE_SUCCESS || | |
916 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
e24f0b8f | 917 | unlock_page(page); |
0dabec93 MK |
918 | out: |
919 | return rc; | |
920 | } | |
95a402c3 | 921 | |
0dabec93 MK |
922 | /* |
923 | * Obtain the lock on page, remove all ptes and migrate the page | |
924 | * to the newly allocated page in newpage. | |
925 | */ | |
926 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
9c620e2b | 927 | struct page *page, int force, enum migrate_mode mode) |
0dabec93 MK |
928 | { |
929 | int rc = 0; | |
930 | int *result = NULL; | |
931 | struct page *newpage = get_new_page(page, private, &result); | |
932 | ||
933 | if (!newpage) | |
934 | return -ENOMEM; | |
935 | ||
936 | if (page_count(page) == 1) { | |
937 | /* page was freed from under us. So we are done. */ | |
938 | goto out; | |
939 | } | |
940 | ||
941 | if (unlikely(PageTransHuge(page))) | |
942 | if (unlikely(split_huge_page(page))) | |
943 | goto out; | |
944 | ||
9c620e2b | 945 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 RA |
946 | |
947 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
948 | /* | |
949 | * A ballooned page has been migrated already. | |
950 | * Now, it's the time to wrap-up counters, | |
951 | * handle the page back to Buddy and return. | |
952 | */ | |
953 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
954 | page_is_file_cache(page)); | |
955 | balloon_page_free(page); | |
956 | return MIGRATEPAGE_SUCCESS; | |
957 | } | |
0dabec93 | 958 | out: |
e24f0b8f | 959 | if (rc != -EAGAIN) { |
0dabec93 MK |
960 | /* |
961 | * A page that has been migrated has all references | |
962 | * removed and will be freed. A page that has not been | |
963 | * migrated will have kepts its references and be | |
964 | * restored. | |
965 | */ | |
966 | list_del(&page->lru); | |
a731286d | 967 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 968 | page_is_file_cache(page)); |
894bc310 | 969 | putback_lru_page(page); |
e24f0b8f | 970 | } |
95a402c3 CL |
971 | /* |
972 | * Move the new page to the LRU. If migration was not successful | |
973 | * then this will free the page. | |
974 | */ | |
894bc310 | 975 | putback_lru_page(newpage); |
742755a1 CL |
976 | if (result) { |
977 | if (rc) | |
978 | *result = rc; | |
979 | else | |
980 | *result = page_to_nid(newpage); | |
981 | } | |
e24f0b8f CL |
982 | return rc; |
983 | } | |
984 | ||
290408d4 NH |
985 | /* |
986 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
987 | * | |
988 | * This function doesn't wait the completion of hugepage I/O | |
989 | * because there is no race between I/O and migration for hugepage. | |
990 | * Note that currently hugepage I/O occurs only in direct I/O | |
991 | * where no lock is held and PG_writeback is irrelevant, | |
992 | * and writeback status of all subpages are counted in the reference | |
993 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
994 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
995 | * This means that when we try to migrate hugepage whose subpages are | |
996 | * doing direct I/O, some references remain after try_to_unmap() and | |
997 | * hugepage migration fails without data corruption. | |
998 | * | |
999 | * There is also no race when direct I/O is issued on the page under migration, | |
1000 | * because then pte is replaced with migration swap entry and direct I/O code | |
1001 | * will wait in the page fault for migration to complete. | |
1002 | */ | |
1003 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
1004 | unsigned long private, struct page *hpage, | |
9c620e2b | 1005 | int force, enum migrate_mode mode) |
290408d4 NH |
1006 | { |
1007 | int rc = 0; | |
1008 | int *result = NULL; | |
1009 | struct page *new_hpage = get_new_page(hpage, private, &result); | |
290408d4 NH |
1010 | struct anon_vma *anon_vma = NULL; |
1011 | ||
83467efb NH |
1012 | /* |
1013 | * Movability of hugepages depends on architectures and hugepage size. | |
1014 | * This check is necessary because some callers of hugepage migration | |
1015 | * like soft offline and memory hotremove don't walk through page | |
1016 | * tables or check whether the hugepage is pmd-based or not before | |
1017 | * kicking migration. | |
1018 | */ | |
1019 | if (!hugepage_migration_support(page_hstate(hpage))) | |
1020 | return -ENOSYS; | |
1021 | ||
290408d4 NH |
1022 | if (!new_hpage) |
1023 | return -ENOMEM; | |
1024 | ||
1025 | rc = -EAGAIN; | |
1026 | ||
1027 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 1028 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1029 | goto out; |
1030 | lock_page(hpage); | |
1031 | } | |
1032 | ||
746b18d4 PZ |
1033 | if (PageAnon(hpage)) |
1034 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 NH |
1035 | |
1036 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1037 | ||
1038 | if (!page_mapped(hpage)) | |
a6bc32b8 | 1039 | rc = move_to_new_page(new_hpage, hpage, 1, mode); |
290408d4 NH |
1040 | |
1041 | if (rc) | |
1042 | remove_migration_ptes(hpage, hpage); | |
1043 | ||
fd4a4663 | 1044 | if (anon_vma) |
9e60109f | 1045 | put_anon_vma(anon_vma); |
8e6ac7fa AK |
1046 | |
1047 | if (!rc) | |
1048 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1049 | ||
290408d4 | 1050 | unlock_page(hpage); |
09761333 | 1051 | out: |
b8ec1cee NH |
1052 | if (rc != -EAGAIN) |
1053 | putback_active_hugepage(hpage); | |
290408d4 | 1054 | put_page(new_hpage); |
290408d4 NH |
1055 | if (result) { |
1056 | if (rc) | |
1057 | *result = rc; | |
1058 | else | |
1059 | *result = page_to_nid(new_hpage); | |
1060 | } | |
1061 | return rc; | |
1062 | } | |
1063 | ||
b20a3503 | 1064 | /* |
c73e5c9c SB |
1065 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1066 | * supplied as the target for the page migration | |
b20a3503 | 1067 | * |
c73e5c9c SB |
1068 | * @from: The list of pages to be migrated. |
1069 | * @get_new_page: The function used to allocate free pages to be used | |
1070 | * as the target of the page migration. | |
1071 | * @private: Private data to be passed on to get_new_page() | |
1072 | * @mode: The migration mode that specifies the constraints for | |
1073 | * page migration, if any. | |
1074 | * @reason: The reason for page migration. | |
b20a3503 | 1075 | * |
c73e5c9c SB |
1076 | * The function returns after 10 attempts or if no pages are movable any more |
1077 | * because the list has become empty or no retryable pages exist any more. | |
1078 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1079 | * or free list only if ret != 0. |
b20a3503 | 1080 | * |
c73e5c9c | 1081 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1082 | */ |
9c620e2b HD |
1083 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
1084 | unsigned long private, enum migrate_mode mode, int reason) | |
b20a3503 | 1085 | { |
e24f0b8f | 1086 | int retry = 1; |
b20a3503 | 1087 | int nr_failed = 0; |
5647bc29 | 1088 | int nr_succeeded = 0; |
b20a3503 CL |
1089 | int pass = 0; |
1090 | struct page *page; | |
1091 | struct page *page2; | |
1092 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1093 | int rc; | |
1094 | ||
1095 | if (!swapwrite) | |
1096 | current->flags |= PF_SWAPWRITE; | |
1097 | ||
e24f0b8f CL |
1098 | for(pass = 0; pass < 10 && retry; pass++) { |
1099 | retry = 0; | |
b20a3503 | 1100 | |
e24f0b8f | 1101 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1102 | cond_resched(); |
2d1db3b1 | 1103 | |
31caf665 NH |
1104 | if (PageHuge(page)) |
1105 | rc = unmap_and_move_huge_page(get_new_page, | |
1106 | private, page, pass > 2, mode); | |
1107 | else | |
1108 | rc = unmap_and_move(get_new_page, private, | |
9c620e2b | 1109 | page, pass > 2, mode); |
2d1db3b1 | 1110 | |
e24f0b8f | 1111 | switch(rc) { |
95a402c3 CL |
1112 | case -ENOMEM: |
1113 | goto out; | |
e24f0b8f | 1114 | case -EAGAIN: |
2d1db3b1 | 1115 | retry++; |
e24f0b8f | 1116 | break; |
78bd5209 | 1117 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1118 | nr_succeeded++; |
e24f0b8f CL |
1119 | break; |
1120 | default: | |
2d1db3b1 | 1121 | /* Permanent failure */ |
2d1db3b1 | 1122 | nr_failed++; |
e24f0b8f | 1123 | break; |
2d1db3b1 | 1124 | } |
b20a3503 CL |
1125 | } |
1126 | } | |
78bd5209 | 1127 | rc = nr_failed + retry; |
95a402c3 | 1128 | out: |
5647bc29 MG |
1129 | if (nr_succeeded) |
1130 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1131 | if (nr_failed) | |
1132 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1133 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1134 | ||
b20a3503 CL |
1135 | if (!swapwrite) |
1136 | current->flags &= ~PF_SWAPWRITE; | |
1137 | ||
78bd5209 | 1138 | return rc; |
b20a3503 | 1139 | } |
95a402c3 | 1140 | |
742755a1 CL |
1141 | #ifdef CONFIG_NUMA |
1142 | /* | |
1143 | * Move a list of individual pages | |
1144 | */ | |
1145 | struct page_to_node { | |
1146 | unsigned long addr; | |
1147 | struct page *page; | |
1148 | int node; | |
1149 | int status; | |
1150 | }; | |
1151 | ||
1152 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1153 | int **result) | |
1154 | { | |
1155 | struct page_to_node *pm = (struct page_to_node *)private; | |
1156 | ||
1157 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1158 | pm++; | |
1159 | ||
1160 | if (pm->node == MAX_NUMNODES) | |
1161 | return NULL; | |
1162 | ||
1163 | *result = &pm->status; | |
1164 | ||
e632a938 NH |
1165 | if (PageHuge(p)) |
1166 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1167 | pm->node); | |
1168 | else | |
1169 | return alloc_pages_exact_node(pm->node, | |
769848c0 | 1170 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); |
742755a1 CL |
1171 | } |
1172 | ||
1173 | /* | |
1174 | * Move a set of pages as indicated in the pm array. The addr | |
1175 | * field must be set to the virtual address of the page to be moved | |
1176 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1177 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1178 | */ |
5e9a0f02 BG |
1179 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1180 | struct page_to_node *pm, | |
1181 | int migrate_all) | |
742755a1 CL |
1182 | { |
1183 | int err; | |
1184 | struct page_to_node *pp; | |
1185 | LIST_HEAD(pagelist); | |
1186 | ||
1187 | down_read(&mm->mmap_sem); | |
1188 | ||
1189 | /* | |
1190 | * Build a list of pages to migrate | |
1191 | */ | |
742755a1 CL |
1192 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1193 | struct vm_area_struct *vma; | |
1194 | struct page *page; | |
1195 | ||
742755a1 CL |
1196 | err = -EFAULT; |
1197 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1198 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1199 | goto set_status; |
1200 | ||
500d65d4 | 1201 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1202 | |
1203 | err = PTR_ERR(page); | |
1204 | if (IS_ERR(page)) | |
1205 | goto set_status; | |
1206 | ||
742755a1 CL |
1207 | err = -ENOENT; |
1208 | if (!page) | |
1209 | goto set_status; | |
1210 | ||
62b61f61 | 1211 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1212 | if (PageReserved(page)) |
742755a1 CL |
1213 | goto put_and_set; |
1214 | ||
1215 | pp->page = page; | |
1216 | err = page_to_nid(page); | |
1217 | ||
1218 | if (err == pp->node) | |
1219 | /* | |
1220 | * Node already in the right place | |
1221 | */ | |
1222 | goto put_and_set; | |
1223 | ||
1224 | err = -EACCES; | |
1225 | if (page_mapcount(page) > 1 && | |
1226 | !migrate_all) | |
1227 | goto put_and_set; | |
1228 | ||
e632a938 NH |
1229 | if (PageHuge(page)) { |
1230 | isolate_huge_page(page, &pagelist); | |
1231 | goto put_and_set; | |
1232 | } | |
1233 | ||
62695a84 | 1234 | err = isolate_lru_page(page); |
6d9c285a | 1235 | if (!err) { |
62695a84 | 1236 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1237 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1238 | page_is_file_cache(page)); | |
1239 | } | |
742755a1 CL |
1240 | put_and_set: |
1241 | /* | |
1242 | * Either remove the duplicate refcount from | |
1243 | * isolate_lru_page() or drop the page ref if it was | |
1244 | * not isolated. | |
1245 | */ | |
1246 | put_page(page); | |
1247 | set_status: | |
1248 | pp->status = err; | |
1249 | } | |
1250 | ||
e78bbfa8 | 1251 | err = 0; |
cf608ac1 | 1252 | if (!list_empty(&pagelist)) { |
742755a1 | 1253 | err = migrate_pages(&pagelist, new_page_node, |
9c620e2b | 1254 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1255 | if (err) |
e632a938 | 1256 | putback_movable_pages(&pagelist); |
cf608ac1 | 1257 | } |
742755a1 CL |
1258 | |
1259 | up_read(&mm->mmap_sem); | |
1260 | return err; | |
1261 | } | |
1262 | ||
5e9a0f02 BG |
1263 | /* |
1264 | * Migrate an array of page address onto an array of nodes and fill | |
1265 | * the corresponding array of status. | |
1266 | */ | |
3268c63e | 1267 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1268 | unsigned long nr_pages, |
1269 | const void __user * __user *pages, | |
1270 | const int __user *nodes, | |
1271 | int __user *status, int flags) | |
1272 | { | |
3140a227 | 1273 | struct page_to_node *pm; |
3140a227 BG |
1274 | unsigned long chunk_nr_pages; |
1275 | unsigned long chunk_start; | |
1276 | int err; | |
5e9a0f02 | 1277 | |
3140a227 BG |
1278 | err = -ENOMEM; |
1279 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1280 | if (!pm) | |
5e9a0f02 | 1281 | goto out; |
35282a2d BG |
1282 | |
1283 | migrate_prep(); | |
1284 | ||
5e9a0f02 | 1285 | /* |
3140a227 BG |
1286 | * Store a chunk of page_to_node array in a page, |
1287 | * but keep the last one as a marker | |
5e9a0f02 | 1288 | */ |
3140a227 | 1289 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1290 | |
3140a227 BG |
1291 | for (chunk_start = 0; |
1292 | chunk_start < nr_pages; | |
1293 | chunk_start += chunk_nr_pages) { | |
1294 | int j; | |
5e9a0f02 | 1295 | |
3140a227 BG |
1296 | if (chunk_start + chunk_nr_pages > nr_pages) |
1297 | chunk_nr_pages = nr_pages - chunk_start; | |
1298 | ||
1299 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1300 | for (j = 0; j < chunk_nr_pages; j++) { | |
1301 | const void __user *p; | |
5e9a0f02 BG |
1302 | int node; |
1303 | ||
3140a227 BG |
1304 | err = -EFAULT; |
1305 | if (get_user(p, pages + j + chunk_start)) | |
1306 | goto out_pm; | |
1307 | pm[j].addr = (unsigned long) p; | |
1308 | ||
1309 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1310 | goto out_pm; |
1311 | ||
1312 | err = -ENODEV; | |
6f5a55f1 LT |
1313 | if (node < 0 || node >= MAX_NUMNODES) |
1314 | goto out_pm; | |
1315 | ||
389162c2 | 1316 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1317 | goto out_pm; |
1318 | ||
1319 | err = -EACCES; | |
1320 | if (!node_isset(node, task_nodes)) | |
1321 | goto out_pm; | |
1322 | ||
3140a227 BG |
1323 | pm[j].node = node; |
1324 | } | |
1325 | ||
1326 | /* End marker for this chunk */ | |
1327 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1328 | ||
1329 | /* Migrate this chunk */ | |
1330 | err = do_move_page_to_node_array(mm, pm, | |
1331 | flags & MPOL_MF_MOVE_ALL); | |
1332 | if (err < 0) | |
1333 | goto out_pm; | |
5e9a0f02 | 1334 | |
5e9a0f02 | 1335 | /* Return status information */ |
3140a227 BG |
1336 | for (j = 0; j < chunk_nr_pages; j++) |
1337 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1338 | err = -EFAULT; |
3140a227 BG |
1339 | goto out_pm; |
1340 | } | |
1341 | } | |
1342 | err = 0; | |
5e9a0f02 BG |
1343 | |
1344 | out_pm: | |
3140a227 | 1345 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1346 | out: |
1347 | return err; | |
1348 | } | |
1349 | ||
742755a1 | 1350 | /* |
2f007e74 | 1351 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1352 | */ |
80bba129 BG |
1353 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1354 | const void __user **pages, int *status) | |
742755a1 | 1355 | { |
2f007e74 | 1356 | unsigned long i; |
2f007e74 | 1357 | |
742755a1 CL |
1358 | down_read(&mm->mmap_sem); |
1359 | ||
2f007e74 | 1360 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1361 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1362 | struct vm_area_struct *vma; |
1363 | struct page *page; | |
c095adbc | 1364 | int err = -EFAULT; |
2f007e74 BG |
1365 | |
1366 | vma = find_vma(mm, addr); | |
70384dc6 | 1367 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1368 | goto set_status; |
1369 | ||
2f007e74 | 1370 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1371 | |
1372 | err = PTR_ERR(page); | |
1373 | if (IS_ERR(page)) | |
1374 | goto set_status; | |
1375 | ||
742755a1 CL |
1376 | err = -ENOENT; |
1377 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1378 | if (!page || PageReserved(page)) |
742755a1 CL |
1379 | goto set_status; |
1380 | ||
1381 | err = page_to_nid(page); | |
1382 | set_status: | |
80bba129 BG |
1383 | *status = err; |
1384 | ||
1385 | pages++; | |
1386 | status++; | |
1387 | } | |
1388 | ||
1389 | up_read(&mm->mmap_sem); | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * Determine the nodes of a user array of pages and store it in | |
1394 | * a user array of status. | |
1395 | */ | |
1396 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1397 | const void __user * __user *pages, | |
1398 | int __user *status) | |
1399 | { | |
1400 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1401 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1402 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1403 | |
87b8d1ad PA |
1404 | while (nr_pages) { |
1405 | unsigned long chunk_nr; | |
80bba129 | 1406 | |
87b8d1ad PA |
1407 | chunk_nr = nr_pages; |
1408 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1409 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1410 | ||
1411 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1412 | break; | |
80bba129 BG |
1413 | |
1414 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1415 | ||
87b8d1ad PA |
1416 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1417 | break; | |
742755a1 | 1418 | |
87b8d1ad PA |
1419 | pages += chunk_nr; |
1420 | status += chunk_nr; | |
1421 | nr_pages -= chunk_nr; | |
1422 | } | |
1423 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1424 | } |
1425 | ||
1426 | /* | |
1427 | * Move a list of pages in the address space of the currently executing | |
1428 | * process. | |
1429 | */ | |
938bb9f5 HC |
1430 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1431 | const void __user * __user *, pages, | |
1432 | const int __user *, nodes, | |
1433 | int __user *, status, int, flags) | |
742755a1 | 1434 | { |
c69e8d9c | 1435 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1436 | struct task_struct *task; |
742755a1 | 1437 | struct mm_struct *mm; |
5e9a0f02 | 1438 | int err; |
3268c63e | 1439 | nodemask_t task_nodes; |
742755a1 CL |
1440 | |
1441 | /* Check flags */ | |
1442 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1443 | return -EINVAL; | |
1444 | ||
1445 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1446 | return -EPERM; | |
1447 | ||
1448 | /* Find the mm_struct */ | |
a879bf58 | 1449 | rcu_read_lock(); |
228ebcbe | 1450 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1451 | if (!task) { |
a879bf58 | 1452 | rcu_read_unlock(); |
742755a1 CL |
1453 | return -ESRCH; |
1454 | } | |
3268c63e | 1455 | get_task_struct(task); |
742755a1 CL |
1456 | |
1457 | /* | |
1458 | * Check if this process has the right to modify the specified | |
1459 | * process. The right exists if the process has administrative | |
1460 | * capabilities, superuser privileges or the same | |
1461 | * userid as the target process. | |
1462 | */ | |
c69e8d9c | 1463 | tcred = __task_cred(task); |
b38a86eb EB |
1464 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1465 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1466 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1467 | rcu_read_unlock(); |
742755a1 | 1468 | err = -EPERM; |
5e9a0f02 | 1469 | goto out; |
742755a1 | 1470 | } |
c69e8d9c | 1471 | rcu_read_unlock(); |
742755a1 | 1472 | |
86c3a764 DQ |
1473 | err = security_task_movememory(task); |
1474 | if (err) | |
5e9a0f02 | 1475 | goto out; |
86c3a764 | 1476 | |
3268c63e CL |
1477 | task_nodes = cpuset_mems_allowed(task); |
1478 | mm = get_task_mm(task); | |
1479 | put_task_struct(task); | |
1480 | ||
6e8b09ea SL |
1481 | if (!mm) |
1482 | return -EINVAL; | |
1483 | ||
1484 | if (nodes) | |
1485 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1486 | nodes, status, flags); | |
1487 | else | |
1488 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1489 | |
742755a1 CL |
1490 | mmput(mm); |
1491 | return err; | |
3268c63e CL |
1492 | |
1493 | out: | |
1494 | put_task_struct(task); | |
1495 | return err; | |
742755a1 | 1496 | } |
742755a1 | 1497 | |
7b2259b3 CL |
1498 | /* |
1499 | * Call migration functions in the vma_ops that may prepare | |
1500 | * memory in a vm for migration. migration functions may perform | |
1501 | * the migration for vmas that do not have an underlying page struct. | |
1502 | */ | |
1503 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1504 | const nodemask_t *from, unsigned long flags) | |
1505 | { | |
1506 | struct vm_area_struct *vma; | |
1507 | int err = 0; | |
1508 | ||
1001c9fb | 1509 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { |
7b2259b3 CL |
1510 | if (vma->vm_ops && vma->vm_ops->migrate) { |
1511 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1512 | if (err) | |
1513 | break; | |
1514 | } | |
1515 | } | |
1516 | return err; | |
1517 | } | |
7039e1db PZ |
1518 | |
1519 | #ifdef CONFIG_NUMA_BALANCING | |
1520 | /* | |
1521 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1522 | * pages. Currently it only checks the watermarks which crude | |
1523 | */ | |
1524 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1525 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1526 | { |
1527 | int z; | |
1528 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1529 | struct zone *zone = pgdat->node_zones + z; | |
1530 | ||
1531 | if (!populated_zone(zone)) | |
1532 | continue; | |
1533 | ||
6e543d57 | 1534 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1535 | continue; |
1536 | ||
1537 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1538 | if (!zone_watermark_ok(zone, 0, | |
1539 | high_wmark_pages(zone) + | |
1540 | nr_migrate_pages, | |
1541 | 0, 0)) | |
1542 | continue; | |
1543 | return true; | |
1544 | } | |
1545 | return false; | |
1546 | } | |
1547 | ||
1548 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1549 | unsigned long data, | |
1550 | int **result) | |
1551 | { | |
1552 | int nid = (int) data; | |
1553 | struct page *newpage; | |
1554 | ||
1555 | newpage = alloc_pages_exact_node(nid, | |
1556 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1557 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1558 | __GFP_NOWARN) & | |
1559 | ~GFP_IOFS, 0); | |
bac0382c | 1560 | if (newpage) |
90572890 | 1561 | page_cpupid_xchg_last(newpage, page_cpupid_last(page)); |
bac0382c | 1562 | |
7039e1db PZ |
1563 | return newpage; |
1564 | } | |
1565 | ||
a8f60772 MG |
1566 | /* |
1567 | * page migration rate limiting control. | |
1568 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1569 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1570 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1571 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1572 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1573 | * throttle window closed. | |
a8f60772 MG |
1574 | */ |
1575 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1576 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1577 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1578 | ||
e14808b4 MG |
1579 | /* Returns true if NUMA migration is currently rate limited */ |
1580 | bool migrate_ratelimited(int node) | |
1581 | { | |
1582 | pg_data_t *pgdat = NODE_DATA(node); | |
1583 | ||
1584 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1585 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1586 | return false; | |
1587 | ||
1588 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1589 | return false; | |
1590 | ||
1591 | return true; | |
1592 | } | |
1593 | ||
b32967ff | 1594 | /* Returns true if the node is migrate rate-limited after the update */ |
d28d4335 | 1595 | bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages) |
7039e1db | 1596 | { |
b32967ff | 1597 | bool rate_limited = false; |
7039e1db | 1598 | |
a8f60772 MG |
1599 | /* |
1600 | * Rate-limit the amount of data that is being migrated to a node. | |
1601 | * Optimal placement is no good if the memory bus is saturated and | |
1602 | * all the time is being spent migrating! | |
1603 | */ | |
1604 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1605 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1606 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1607 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1608 | msecs_to_jiffies(migrate_interval_millisecs); | |
1609 | } | |
b32967ff MG |
1610 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) |
1611 | rate_limited = true; | |
1612 | else | |
d28d4335 | 1613 | pgdat->numabalancing_migrate_nr_pages += nr_pages; |
a8f60772 | 1614 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
b32967ff MG |
1615 | |
1616 | return rate_limited; | |
1617 | } | |
1618 | ||
1619 | int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1620 | { | |
340ef390 | 1621 | int page_lru; |
a8f60772 | 1622 | |
3abef4e6 MG |
1623 | VM_BUG_ON(compound_order(page) && !PageTransHuge(page)); |
1624 | ||
7039e1db | 1625 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1626 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1627 | return 0; | |
7039e1db | 1628 | |
340ef390 HD |
1629 | if (isolate_lru_page(page)) |
1630 | return 0; | |
7039e1db | 1631 | |
340ef390 HD |
1632 | /* |
1633 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1634 | * check on page_count(), so we must do it here, now that the page | |
1635 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1636 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1637 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1638 | */ | |
1639 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1640 | putback_lru_page(page); | |
1641 | return 0; | |
7039e1db PZ |
1642 | } |
1643 | ||
340ef390 HD |
1644 | page_lru = page_is_file_cache(page); |
1645 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1646 | hpage_nr_pages(page)); | |
1647 | ||
149c33e1 | 1648 | /* |
340ef390 HD |
1649 | * Isolating the page has taken another reference, so the |
1650 | * caller's reference can be safely dropped without the page | |
1651 | * disappearing underneath us during migration. | |
149c33e1 MG |
1652 | */ |
1653 | put_page(page); | |
340ef390 | 1654 | return 1; |
b32967ff MG |
1655 | } |
1656 | ||
1657 | /* | |
1658 | * Attempt to migrate a misplaced page to the specified destination | |
1659 | * node. Caller is expected to have an elevated reference count on | |
1660 | * the page that will be dropped by this function before returning. | |
1661 | */ | |
1bc115d8 MG |
1662 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1663 | int node) | |
b32967ff MG |
1664 | { |
1665 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1666 | int isolated; |
b32967ff MG |
1667 | int nr_remaining; |
1668 | LIST_HEAD(migratepages); | |
1669 | ||
1670 | /* | |
1bc115d8 MG |
1671 | * Don't migrate file pages that are mapped in multiple processes |
1672 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1673 | */ |
1bc115d8 MG |
1674 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1675 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1676 | goto out; |
b32967ff MG |
1677 | |
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 | */ | |
340ef390 | 1683 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1684 | goto out; |
b32967ff MG |
1685 | |
1686 | isolated = numamigrate_isolate_page(pgdat, page); | |
1687 | if (!isolated) | |
1688 | goto out; | |
1689 | ||
1690 | list_add(&page->lru, &migratepages); | |
9c620e2b HD |
1691 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
1692 | node, MIGRATE_ASYNC, MR_NUMA_MISPLACED); | |
b32967ff MG |
1693 | if (nr_remaining) { |
1694 | putback_lru_pages(&migratepages); | |
1695 | isolated = 0; | |
1696 | } else | |
1697 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1698 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1699 | return isolated; |
340ef390 HD |
1700 | |
1701 | out: | |
1702 | put_page(page); | |
1703 | return 0; | |
7039e1db | 1704 | } |
220018d3 | 1705 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1706 | |
220018d3 | 1707 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1708 | /* |
1709 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1710 | * before returning. | |
1711 | */ | |
b32967ff MG |
1712 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1713 | struct vm_area_struct *vma, | |
1714 | pmd_t *pmd, pmd_t entry, | |
1715 | unsigned long address, | |
1716 | struct page *page, int node) | |
1717 | { | |
c4088ebd | 1718 | spinlock_t *ptl; |
b32967ff MG |
1719 | unsigned long haddr = address & HPAGE_PMD_MASK; |
1720 | pg_data_t *pgdat = NODE_DATA(node); | |
1721 | int isolated = 0; | |
1722 | struct page *new_page = NULL; | |
1723 | struct mem_cgroup *memcg = NULL; | |
1724 | int page_lru = page_is_file_cache(page); | |
1725 | ||
b32967ff MG |
1726 | /* |
1727 | * Rate-limit the amount of data that is being migrated to a node. | |
1728 | * Optimal placement is no good if the memory bus is saturated and | |
1729 | * all the time is being spent migrating! | |
1730 | */ | |
d28d4335 | 1731 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1732 | goto out_dropref; |
1733 | ||
1734 | new_page = alloc_pages_node(node, | |
1735 | (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER); | |
340ef390 HD |
1736 | if (!new_page) |
1737 | goto out_fail; | |
1738 | ||
90572890 | 1739 | page_cpupid_xchg_last(new_page, page_cpupid_last(page)); |
b32967ff MG |
1740 | |
1741 | isolated = numamigrate_isolate_page(pgdat, page); | |
340ef390 | 1742 | if (!isolated) { |
b32967ff | 1743 | put_page(new_page); |
340ef390 | 1744 | goto out_fail; |
b32967ff MG |
1745 | } |
1746 | ||
1747 | /* Prepare a page as a migration target */ | |
1748 | __set_page_locked(new_page); | |
1749 | SetPageSwapBacked(new_page); | |
1750 | ||
1751 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1752 | new_page->mapping = page->mapping; | |
1753 | new_page->index = page->index; | |
1754 | migrate_page_copy(new_page, page); | |
1755 | WARN_ON(PageLRU(new_page)); | |
1756 | ||
1757 | /* Recheck the target PMD */ | |
c4088ebd | 1758 | ptl = pmd_lock(mm, pmd); |
b32967ff | 1759 | if (unlikely(!pmd_same(*pmd, entry))) { |
c4088ebd | 1760 | spin_unlock(ptl); |
b32967ff MG |
1761 | |
1762 | /* Reverse changes made by migrate_page_copy() */ | |
1763 | if (TestClearPageActive(new_page)) | |
1764 | SetPageActive(page); | |
1765 | if (TestClearPageUnevictable(new_page)) | |
1766 | SetPageUnevictable(page); | |
1767 | mlock_migrate_page(page, new_page); | |
1768 | ||
1769 | unlock_page(new_page); | |
1770 | put_page(new_page); /* Free it */ | |
1771 | ||
a54a407f MG |
1772 | /* Retake the callers reference and putback on LRU */ |
1773 | get_page(page); | |
b32967ff | 1774 | putback_lru_page(page); |
a54a407f MG |
1775 | mod_zone_page_state(page_zone(page), |
1776 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
1777 | goto out_fail; | |
b32967ff MG |
1778 | } |
1779 | ||
1780 | /* | |
1781 | * Traditional migration needs to prepare the memcg charge | |
1782 | * transaction early to prevent the old page from being | |
1783 | * uncharged when installing migration entries. Here we can | |
1784 | * save the potential rollback and start the charge transfer | |
1785 | * only when migration is already known to end successfully. | |
1786 | */ | |
1787 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1788 | ||
1789 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
1790 | entry = pmd_mknonnuma(entry); | |
1791 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1792 | entry = pmd_mkhuge(entry); | |
1793 | ||
a54a407f | 1794 | pmdp_clear_flush(vma, haddr, pmd); |
b32967ff | 1795 | set_pmd_at(mm, haddr, pmd, entry); |
a54a407f | 1796 | page_add_new_anon_rmap(new_page, vma, haddr); |
ce4a9cc5 | 1797 | update_mmu_cache_pmd(vma, address, &entry); |
b32967ff MG |
1798 | page_remove_rmap(page); |
1799 | /* | |
1800 | * Finish the charge transaction under the page table lock to | |
1801 | * prevent split_huge_page() from dividing up the charge | |
1802 | * before it's fully transferred to the new page. | |
1803 | */ | |
1804 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
c4088ebd | 1805 | spin_unlock(ptl); |
b32967ff MG |
1806 | |
1807 | unlock_page(new_page); | |
1808 | unlock_page(page); | |
1809 | put_page(page); /* Drop the rmap reference */ | |
1810 | put_page(page); /* Drop the LRU isolation reference */ | |
1811 | ||
1812 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1813 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1814 | ||
b32967ff MG |
1815 | mod_zone_page_state(page_zone(page), |
1816 | NR_ISOLATED_ANON + page_lru, | |
1817 | -HPAGE_PMD_NR); | |
1818 | return isolated; | |
1819 | ||
340ef390 HD |
1820 | out_fail: |
1821 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1822 | out_dropref: |
a54a407f MG |
1823 | entry = pmd_mknonnuma(entry); |
1824 | set_pmd_at(mm, haddr, pmd, entry); | |
1825 | update_mmu_cache_pmd(vma, address, &entry); | |
1826 | ||
340ef390 | 1827 | unlock_page(page); |
b32967ff | 1828 | put_page(page); |
b32967ff MG |
1829 | return 0; |
1830 | } | |
7039e1db PZ |
1831 | #endif /* CONFIG_NUMA_BALANCING */ |
1832 | ||
1833 | #endif /* CONFIG_NUMA */ |