Commit | Line | Data |
---|---|---|
b20a3503 | 1 | /* |
14e0f9bc | 2 | * Memory Migration functionality - linux/mm/migrate.c |
b20a3503 CL |
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> |
4f5ca265 | 33 | #include <linux/syscalls.h> |
290408d4 | 34 | #include <linux/hugetlb.h> |
8e6ac7fa | 35 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 36 | #include <linux/gfp.h> |
bf6bddf1 | 37 | #include <linux/balloon_compaction.h> |
f714f4f2 | 38 | #include <linux/mmu_notifier.h> |
33c3fc71 | 39 | #include <linux/page_idle.h> |
b20a3503 | 40 | |
0d1836c3 MN |
41 | #include <asm/tlbflush.h> |
42 | ||
7b2a2d4a MG |
43 | #define CREATE_TRACE_POINTS |
44 | #include <trace/events/migrate.h> | |
45 | ||
b20a3503 CL |
46 | #include "internal.h" |
47 | ||
b20a3503 | 48 | /* |
742755a1 | 49 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
50 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
51 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
52 | */ |
53 | int migrate_prep(void) | |
54 | { | |
b20a3503 CL |
55 | /* |
56 | * Clear the LRU lists so pages can be isolated. | |
57 | * Note that pages may be moved off the LRU after we have | |
58 | * drained them. Those pages will fail to migrate like other | |
59 | * pages that may be busy. | |
60 | */ | |
61 | lru_add_drain_all(); | |
62 | ||
63 | return 0; | |
64 | } | |
65 | ||
748446bb MG |
66 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
67 | int migrate_prep_local(void) | |
68 | { | |
69 | lru_add_drain(); | |
70 | ||
71 | return 0; | |
72 | } | |
73 | ||
5733c7d1 RA |
74 | /* |
75 | * Put previously isolated pages back onto the appropriate lists | |
76 | * from where they were once taken off for compaction/migration. | |
77 | * | |
59c82b70 JK |
78 | * This function shall be used whenever the isolated pageset has been |
79 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
80 | * and isolate_huge_page(). | |
5733c7d1 RA |
81 | */ |
82 | void putback_movable_pages(struct list_head *l) | |
83 | { | |
84 | struct page *page; | |
85 | struct page *page2; | |
86 | ||
b20a3503 | 87 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
88 | if (unlikely(PageHuge(page))) { |
89 | putback_active_hugepage(page); | |
90 | continue; | |
91 | } | |
e24f0b8f | 92 | list_del(&page->lru); |
a731286d | 93 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 94 | page_is_file_cache(page)); |
117aad1e | 95 | if (unlikely(isolated_balloon_page(page))) |
bf6bddf1 RA |
96 | balloon_page_putback(page); |
97 | else | |
98 | putback_lru_page(page); | |
b20a3503 | 99 | } |
b20a3503 CL |
100 | } |
101 | ||
0697212a CL |
102 | /* |
103 | * Restore a potential migration pte to a working pte entry | |
104 | */ | |
e9995ef9 HD |
105 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, |
106 | unsigned long addr, void *old) | |
0697212a CL |
107 | { |
108 | struct mm_struct *mm = vma->vm_mm; | |
109 | swp_entry_t entry; | |
0697212a CL |
110 | pmd_t *pmd; |
111 | pte_t *ptep, pte; | |
112 | spinlock_t *ptl; | |
113 | ||
290408d4 NH |
114 | if (unlikely(PageHuge(new))) { |
115 | ptep = huge_pte_offset(mm, addr); | |
116 | if (!ptep) | |
117 | goto out; | |
cb900f41 | 118 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep); |
290408d4 | 119 | } else { |
6219049a BL |
120 | pmd = mm_find_pmd(mm, addr); |
121 | if (!pmd) | |
290408d4 | 122 | goto out; |
0697212a | 123 | |
290408d4 | 124 | ptep = pte_offset_map(pmd, addr); |
0697212a | 125 | |
486cf46f HD |
126 | /* |
127 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
128 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
129 | */ | |
290408d4 NH |
130 | |
131 | ptl = pte_lockptr(mm, pmd); | |
132 | } | |
0697212a | 133 | |
0697212a CL |
134 | spin_lock(ptl); |
135 | pte = *ptep; | |
136 | if (!is_swap_pte(pte)) | |
e9995ef9 | 137 | goto unlock; |
0697212a CL |
138 | |
139 | entry = pte_to_swp_entry(pte); | |
140 | ||
e9995ef9 HD |
141 | if (!is_migration_entry(entry) || |
142 | migration_entry_to_page(entry) != old) | |
143 | goto unlock; | |
0697212a | 144 | |
0697212a CL |
145 | get_page(new); |
146 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
c3d16e16 CG |
147 | if (pte_swp_soft_dirty(*ptep)) |
148 | pte = pte_mksoft_dirty(pte); | |
d3cb8bf6 MG |
149 | |
150 | /* Recheck VMA as permissions can change since migration started */ | |
0697212a | 151 | if (is_write_migration_entry(entry)) |
d3cb8bf6 MG |
152 | pte = maybe_mkwrite(pte, vma); |
153 | ||
3ef8fd7f | 154 | #ifdef CONFIG_HUGETLB_PAGE |
be7517d6 | 155 | if (PageHuge(new)) { |
290408d4 | 156 | pte = pte_mkhuge(pte); |
be7517d6 TL |
157 | pte = arch_make_huge_pte(pte, vma, new, 0); |
158 | } | |
3ef8fd7f | 159 | #endif |
c2cc499c | 160 | flush_dcache_page(new); |
0697212a | 161 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 162 | |
290408d4 NH |
163 | if (PageHuge(new)) { |
164 | if (PageAnon(new)) | |
165 | hugepage_add_anon_rmap(new, vma, addr); | |
166 | else | |
167 | page_dup_rmap(new); | |
168 | } else if (PageAnon(new)) | |
04e62a29 CL |
169 | page_add_anon_rmap(new, vma, addr); |
170 | else | |
171 | page_add_file_rmap(new); | |
172 | ||
51afb12b HD |
173 | if (vma->vm_flags & VM_LOCKED) |
174 | mlock_vma_page(new); | |
175 | ||
04e62a29 | 176 | /* No need to invalidate - it was non-present before */ |
4b3073e1 | 177 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 178 | unlock: |
0697212a | 179 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
180 | out: |
181 | return SWAP_AGAIN; | |
0697212a CL |
182 | } |
183 | ||
04e62a29 CL |
184 | /* |
185 | * Get rid of all migration entries and replace them by | |
186 | * references to the indicated page. | |
187 | */ | |
188 | static void remove_migration_ptes(struct page *old, struct page *new) | |
189 | { | |
051ac83a JK |
190 | struct rmap_walk_control rwc = { |
191 | .rmap_one = remove_migration_pte, | |
192 | .arg = old, | |
193 | }; | |
194 | ||
195 | rmap_walk(new, &rwc); | |
04e62a29 CL |
196 | } |
197 | ||
0697212a CL |
198 | /* |
199 | * Something used the pte of a page under migration. We need to | |
200 | * get to the page and wait until migration is finished. | |
201 | * When we return from this function the fault will be retried. | |
0697212a | 202 | */ |
e66f17ff | 203 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 204 | spinlock_t *ptl) |
0697212a | 205 | { |
30dad309 | 206 | pte_t pte; |
0697212a CL |
207 | swp_entry_t entry; |
208 | struct page *page; | |
209 | ||
30dad309 | 210 | spin_lock(ptl); |
0697212a CL |
211 | pte = *ptep; |
212 | if (!is_swap_pte(pte)) | |
213 | goto out; | |
214 | ||
215 | entry = pte_to_swp_entry(pte); | |
216 | if (!is_migration_entry(entry)) | |
217 | goto out; | |
218 | ||
219 | page = migration_entry_to_page(entry); | |
220 | ||
e286781d NP |
221 | /* |
222 | * Once radix-tree replacement of page migration started, page_count | |
223 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
224 | * against a page without get_page(). | |
225 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
226 | * will occur again. | |
227 | */ | |
228 | if (!get_page_unless_zero(page)) | |
229 | goto out; | |
0697212a CL |
230 | pte_unmap_unlock(ptep, ptl); |
231 | wait_on_page_locked(page); | |
232 | put_page(page); | |
233 | return; | |
234 | out: | |
235 | pte_unmap_unlock(ptep, ptl); | |
236 | } | |
237 | ||
30dad309 NH |
238 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
239 | unsigned long address) | |
240 | { | |
241 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
242 | pte_t *ptep = pte_offset_map(pmd, address); | |
243 | __migration_entry_wait(mm, ptep, ptl); | |
244 | } | |
245 | ||
cb900f41 KS |
246 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
247 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 248 | { |
cb900f41 | 249 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
250 | __migration_entry_wait(mm, pte, ptl); |
251 | } | |
252 | ||
b969c4ab MG |
253 | #ifdef CONFIG_BLOCK |
254 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
255 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
256 | enum migrate_mode mode) | |
b969c4ab MG |
257 | { |
258 | struct buffer_head *bh = head; | |
259 | ||
260 | /* Simple case, sync compaction */ | |
a6bc32b8 | 261 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
262 | do { |
263 | get_bh(bh); | |
264 | lock_buffer(bh); | |
265 | bh = bh->b_this_page; | |
266 | ||
267 | } while (bh != head); | |
268 | ||
269 | return true; | |
270 | } | |
271 | ||
272 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
273 | do { | |
274 | get_bh(bh); | |
275 | if (!trylock_buffer(bh)) { | |
276 | /* | |
277 | * We failed to lock the buffer and cannot stall in | |
278 | * async migration. Release the taken locks | |
279 | */ | |
280 | struct buffer_head *failed_bh = bh; | |
281 | put_bh(failed_bh); | |
282 | bh = head; | |
283 | while (bh != failed_bh) { | |
284 | unlock_buffer(bh); | |
285 | put_bh(bh); | |
286 | bh = bh->b_this_page; | |
287 | } | |
288 | return false; | |
289 | } | |
290 | ||
291 | bh = bh->b_this_page; | |
292 | } while (bh != head); | |
293 | return true; | |
294 | } | |
295 | #else | |
296 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 297 | enum migrate_mode mode) |
b969c4ab MG |
298 | { |
299 | return true; | |
300 | } | |
301 | #endif /* CONFIG_BLOCK */ | |
302 | ||
b20a3503 | 303 | /* |
c3fcf8a5 | 304 | * Replace the page in the mapping. |
5b5c7120 CL |
305 | * |
306 | * The number of remaining references must be: | |
307 | * 1 for anonymous pages without a mapping | |
308 | * 2 for pages with a mapping | |
266cf658 | 309 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 310 | */ |
36bc08cc | 311 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 312 | struct page *newpage, struct page *page, |
8e321fef BL |
313 | struct buffer_head *head, enum migrate_mode mode, |
314 | int extra_count) | |
b20a3503 | 315 | { |
8e321fef | 316 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 317 | void **pslot; |
b20a3503 | 318 | |
6c5240ae | 319 | if (!mapping) { |
0e8c7d0f | 320 | /* Anonymous page without mapping */ |
8e321fef | 321 | if (page_count(page) != expected_count) |
6c5240ae | 322 | return -EAGAIN; |
78bd5209 | 323 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
324 | } |
325 | ||
19fd6231 | 326 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 327 | |
7cf9c2c7 NP |
328 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
329 | page_index(page)); | |
b20a3503 | 330 | |
8e321fef | 331 | expected_count += 1 + page_has_private(page); |
e286781d | 332 | if (page_count(page) != expected_count || |
29c1f677 | 333 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 334 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 335 | return -EAGAIN; |
b20a3503 CL |
336 | } |
337 | ||
e286781d | 338 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 339 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
340 | return -EAGAIN; |
341 | } | |
342 | ||
b969c4ab MG |
343 | /* |
344 | * In the async migration case of moving a page with buffers, lock the | |
345 | * buffers using trylock before the mapping is moved. If the mapping | |
346 | * was moved, we later failed to lock the buffers and could not move | |
347 | * the mapping back due to an elevated page count, we would have to | |
348 | * block waiting on other references to be dropped. | |
349 | */ | |
a6bc32b8 MG |
350 | if (mode == MIGRATE_ASYNC && head && |
351 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
352 | page_unfreeze_refs(page, expected_count); |
353 | spin_unlock_irq(&mapping->tree_lock); | |
354 | return -EAGAIN; | |
355 | } | |
356 | ||
b20a3503 CL |
357 | /* |
358 | * Now we know that no one else is looking at the page. | |
b20a3503 | 359 | */ |
7cf9c2c7 | 360 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
361 | if (PageSwapCache(page)) { |
362 | SetPageSwapCache(newpage); | |
363 | set_page_private(newpage, page_private(page)); | |
364 | } | |
365 | ||
7cf9c2c7 NP |
366 | radix_tree_replace_slot(pslot, newpage); |
367 | ||
368 | /* | |
937a94c9 JG |
369 | * Drop cache reference from old page by unfreezing |
370 | * to one less reference. | |
7cf9c2c7 NP |
371 | * We know this isn't the last reference. |
372 | */ | |
937a94c9 | 373 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 374 | |
0e8c7d0f CL |
375 | /* |
376 | * If moved to a different zone then also account | |
377 | * the page for that zone. Other VM counters will be | |
378 | * taken care of when we establish references to the | |
379 | * new page and drop references to the old page. | |
380 | * | |
381 | * Note that anonymous pages are accounted for | |
382 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
383 | * are mapped to swap space. | |
384 | */ | |
385 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
386 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 387 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
388 | __dec_zone_page_state(page, NR_SHMEM); |
389 | __inc_zone_page_state(newpage, NR_SHMEM); | |
390 | } | |
19fd6231 | 391 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 392 | |
78bd5209 | 393 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 394 | } |
b20a3503 | 395 | |
290408d4 NH |
396 | /* |
397 | * The expected number of remaining references is the same as that | |
398 | * of migrate_page_move_mapping(). | |
399 | */ | |
400 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
401 | struct page *newpage, struct page *page) | |
402 | { | |
403 | int expected_count; | |
404 | void **pslot; | |
405 | ||
406 | if (!mapping) { | |
407 | if (page_count(page) != 1) | |
408 | return -EAGAIN; | |
78bd5209 | 409 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
410 | } |
411 | ||
412 | spin_lock_irq(&mapping->tree_lock); | |
413 | ||
414 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
415 | page_index(page)); | |
416 | ||
417 | expected_count = 2 + page_has_private(page); | |
418 | if (page_count(page) != expected_count || | |
29c1f677 | 419 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
420 | spin_unlock_irq(&mapping->tree_lock); |
421 | return -EAGAIN; | |
422 | } | |
423 | ||
424 | if (!page_freeze_refs(page, expected_count)) { | |
425 | spin_unlock_irq(&mapping->tree_lock); | |
426 | return -EAGAIN; | |
427 | } | |
428 | ||
429 | get_page(newpage); | |
430 | ||
431 | radix_tree_replace_slot(pslot, newpage); | |
432 | ||
937a94c9 | 433 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
434 | |
435 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 436 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
437 | } |
438 | ||
30b0a105 DH |
439 | /* |
440 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
441 | * arithmetic will work across the entire page. We need something more | |
442 | * specialized. | |
443 | */ | |
444 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
445 | int nr_pages) | |
446 | { | |
447 | int i; | |
448 | struct page *dst_base = dst; | |
449 | struct page *src_base = src; | |
450 | ||
451 | for (i = 0; i < nr_pages; ) { | |
452 | cond_resched(); | |
453 | copy_highpage(dst, src); | |
454 | ||
455 | i++; | |
456 | dst = mem_map_next(dst, dst_base, i); | |
457 | src = mem_map_next(src, src_base, i); | |
458 | } | |
459 | } | |
460 | ||
461 | static void copy_huge_page(struct page *dst, struct page *src) | |
462 | { | |
463 | int i; | |
464 | int nr_pages; | |
465 | ||
466 | if (PageHuge(src)) { | |
467 | /* hugetlbfs page */ | |
468 | struct hstate *h = page_hstate(src); | |
469 | nr_pages = pages_per_huge_page(h); | |
470 | ||
471 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
472 | __copy_gigantic_page(dst, src, nr_pages); | |
473 | return; | |
474 | } | |
475 | } else { | |
476 | /* thp page */ | |
477 | BUG_ON(!PageTransHuge(src)); | |
478 | nr_pages = hpage_nr_pages(src); | |
479 | } | |
480 | ||
481 | for (i = 0; i < nr_pages; i++) { | |
482 | cond_resched(); | |
483 | copy_highpage(dst + i, src + i); | |
484 | } | |
485 | } | |
486 | ||
b20a3503 CL |
487 | /* |
488 | * Copy the page to its new location | |
489 | */ | |
290408d4 | 490 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 491 | { |
7851a45c RR |
492 | int cpupid; |
493 | ||
b32967ff | 494 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
495 | copy_huge_page(newpage, page); |
496 | else | |
497 | copy_highpage(newpage, page); | |
b20a3503 CL |
498 | |
499 | if (PageError(page)) | |
500 | SetPageError(newpage); | |
501 | if (PageReferenced(page)) | |
502 | SetPageReferenced(newpage); | |
503 | if (PageUptodate(page)) | |
504 | SetPageUptodate(newpage); | |
894bc310 | 505 | if (TestClearPageActive(page)) { |
309381fe | 506 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 507 | SetPageActive(newpage); |
418b27ef LS |
508 | } else if (TestClearPageUnevictable(page)) |
509 | SetPageUnevictable(newpage); | |
b20a3503 CL |
510 | if (PageChecked(page)) |
511 | SetPageChecked(newpage); | |
512 | if (PageMappedToDisk(page)) | |
513 | SetPageMappedToDisk(newpage); | |
514 | ||
515 | if (PageDirty(page)) { | |
516 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
517 | /* |
518 | * Want to mark the page and the radix tree as dirty, and | |
519 | * redo the accounting that clear_page_dirty_for_io undid, | |
520 | * but we can't use set_page_dirty because that function | |
521 | * is actually a signal that all of the page has become dirty. | |
25985edc | 522 | * Whereas only part of our page may be dirty. |
3a902c5f | 523 | */ |
752dc185 HD |
524 | if (PageSwapBacked(page)) |
525 | SetPageDirty(newpage); | |
526 | else | |
527 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
528 | } |
529 | ||
33c3fc71 VD |
530 | if (page_is_young(page)) |
531 | set_page_young(newpage); | |
532 | if (page_is_idle(page)) | |
533 | set_page_idle(newpage); | |
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 | ||
e9995ef9 | 542 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
543 | /* |
544 | * Please do not reorder this without considering how mm/ksm.c's | |
545 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
546 | */ | |
b3b3a99c NH |
547 | if (PageSwapCache(page)) |
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 | ||
b20a3503 CL |
564 | /* |
565 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 566 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
567 | * |
568 | * Pages are locked upon entry and exit. | |
569 | */ | |
2d1db3b1 | 570 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
571 | struct page *newpage, struct page *page, |
572 | enum migrate_mode mode) | |
b20a3503 CL |
573 | { |
574 | int rc; | |
575 | ||
576 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
577 | ||
8e321fef | 578 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 579 | |
78bd5209 | 580 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
581 | return rc; |
582 | ||
583 | migrate_page_copy(newpage, page); | |
78bd5209 | 584 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
585 | } |
586 | EXPORT_SYMBOL(migrate_page); | |
587 | ||
9361401e | 588 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
589 | /* |
590 | * Migration function for pages with buffers. This function can only be used | |
591 | * if the underlying filesystem guarantees that no other references to "page" | |
592 | * exist. | |
593 | */ | |
2d1db3b1 | 594 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 595 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 596 | { |
1d8b85cc CL |
597 | struct buffer_head *bh, *head; |
598 | int rc; | |
599 | ||
1d8b85cc | 600 | if (!page_has_buffers(page)) |
a6bc32b8 | 601 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
602 | |
603 | head = page_buffers(page); | |
604 | ||
8e321fef | 605 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 606 | |
78bd5209 | 607 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
608 | return rc; |
609 | ||
b969c4ab MG |
610 | /* |
611 | * In the async case, migrate_page_move_mapping locked the buffers | |
612 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
613 | * need to be locked now | |
614 | */ | |
a6bc32b8 MG |
615 | if (mode != MIGRATE_ASYNC) |
616 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
617 | |
618 | ClearPagePrivate(page); | |
619 | set_page_private(newpage, page_private(page)); | |
620 | set_page_private(page, 0); | |
621 | put_page(page); | |
622 | get_page(newpage); | |
623 | ||
624 | bh = head; | |
625 | do { | |
626 | set_bh_page(bh, newpage, bh_offset(bh)); | |
627 | bh = bh->b_this_page; | |
628 | ||
629 | } while (bh != head); | |
630 | ||
631 | SetPagePrivate(newpage); | |
632 | ||
633 | migrate_page_copy(newpage, page); | |
634 | ||
635 | bh = head; | |
636 | do { | |
637 | unlock_buffer(bh); | |
638 | put_bh(bh); | |
639 | bh = bh->b_this_page; | |
640 | ||
641 | } while (bh != head); | |
642 | ||
78bd5209 | 643 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
644 | } |
645 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 646 | #endif |
1d8b85cc | 647 | |
04e62a29 CL |
648 | /* |
649 | * Writeback a page to clean the dirty state | |
650 | */ | |
651 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 652 | { |
04e62a29 CL |
653 | struct writeback_control wbc = { |
654 | .sync_mode = WB_SYNC_NONE, | |
655 | .nr_to_write = 1, | |
656 | .range_start = 0, | |
657 | .range_end = LLONG_MAX, | |
04e62a29 CL |
658 | .for_reclaim = 1 |
659 | }; | |
660 | int rc; | |
661 | ||
662 | if (!mapping->a_ops->writepage) | |
663 | /* No write method for the address space */ | |
664 | return -EINVAL; | |
665 | ||
666 | if (!clear_page_dirty_for_io(page)) | |
667 | /* Someone else already triggered a write */ | |
668 | return -EAGAIN; | |
669 | ||
8351a6e4 | 670 | /* |
04e62a29 CL |
671 | * A dirty page may imply that the underlying filesystem has |
672 | * the page on some queue. So the page must be clean for | |
673 | * migration. Writeout may mean we loose the lock and the | |
674 | * page state is no longer what we checked for earlier. | |
675 | * At this point we know that the migration attempt cannot | |
676 | * be successful. | |
8351a6e4 | 677 | */ |
04e62a29 | 678 | remove_migration_ptes(page, page); |
8351a6e4 | 679 | |
04e62a29 | 680 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 681 | |
04e62a29 CL |
682 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
683 | /* unlocked. Relock */ | |
684 | lock_page(page); | |
685 | ||
bda8550d | 686 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
687 | } |
688 | ||
689 | /* | |
690 | * Default handling if a filesystem does not provide a migration function. | |
691 | */ | |
692 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 693 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 694 | { |
b969c4ab | 695 | if (PageDirty(page)) { |
a6bc32b8 MG |
696 | /* Only writeback pages in full synchronous migration */ |
697 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 698 | return -EBUSY; |
04e62a29 | 699 | return writeout(mapping, page); |
b969c4ab | 700 | } |
8351a6e4 CL |
701 | |
702 | /* | |
703 | * Buffers may be managed in a filesystem specific way. | |
704 | * We must have no buffers or drop them. | |
705 | */ | |
266cf658 | 706 | if (page_has_private(page) && |
8351a6e4 CL |
707 | !try_to_release_page(page, GFP_KERNEL)) |
708 | return -EAGAIN; | |
709 | ||
a6bc32b8 | 710 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
711 | } |
712 | ||
e24f0b8f CL |
713 | /* |
714 | * Move a page to a newly allocated page | |
715 | * The page is locked and all ptes have been successfully removed. | |
716 | * | |
717 | * The new page will have replaced the old page if this function | |
718 | * is successful. | |
894bc310 LS |
719 | * |
720 | * Return value: | |
721 | * < 0 - error code | |
78bd5209 | 722 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 723 | */ |
3fe2011f | 724 | static int move_to_new_page(struct page *newpage, struct page *page, |
2ebba6b7 | 725 | int page_was_mapped, enum migrate_mode mode) |
e24f0b8f CL |
726 | { |
727 | struct address_space *mapping; | |
728 | int rc; | |
729 | ||
7db7671f HD |
730 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
731 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f CL |
732 | |
733 | /* Prepare mapping for the new page.*/ | |
734 | newpage->index = page->index; | |
735 | newpage->mapping = page->mapping; | |
b2e18538 RR |
736 | if (PageSwapBacked(page)) |
737 | SetPageSwapBacked(newpage); | |
e24f0b8f | 738 | |
0610c25d GT |
739 | /* |
740 | * Indirectly called below, migrate_page_copy() copies PG_dirty and thus | |
741 | * needs newpage's memcg set to transfer memcg dirty page accounting. | |
742 | * So perform memcg migration in two steps: | |
743 | * 1. set newpage->mem_cgroup (here) | |
744 | * 2. clear page->mem_cgroup (below) | |
745 | */ | |
746 | set_page_memcg(newpage, page_memcg(page)); | |
747 | ||
e24f0b8f CL |
748 | mapping = page_mapping(page); |
749 | if (!mapping) | |
a6bc32b8 | 750 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 751 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 752 | /* |
b969c4ab MG |
753 | * Most pages have a mapping and most filesystems provide a |
754 | * migratepage callback. Anonymous pages are part of swap | |
755 | * space which also has its own migratepage callback. This | |
756 | * is the most common path for page migration. | |
e24f0b8f | 757 | */ |
b969c4ab | 758 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 759 | newpage, page, mode); |
b969c4ab | 760 | else |
a6bc32b8 | 761 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 762 | |
78bd5209 | 763 | if (rc != MIGRATEPAGE_SUCCESS) { |
0610c25d | 764 | set_page_memcg(newpage, NULL); |
e24f0b8f | 765 | newpage->mapping = NULL; |
3fe2011f | 766 | } else { |
0610c25d | 767 | set_page_memcg(page, NULL); |
2ebba6b7 | 768 | if (page_was_mapped) |
3fe2011f | 769 | remove_migration_ptes(page, newpage); |
35512eca | 770 | page->mapping = NULL; |
3fe2011f | 771 | } |
e24f0b8f CL |
772 | return rc; |
773 | } | |
774 | ||
0dabec93 | 775 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 776 | int force, enum migrate_mode mode) |
e24f0b8f | 777 | { |
0dabec93 | 778 | int rc = -EAGAIN; |
2ebba6b7 | 779 | int page_was_mapped = 0; |
3f6c8272 | 780 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 781 | |
529ae9aa | 782 | if (!trylock_page(page)) { |
a6bc32b8 | 783 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 784 | goto out; |
3e7d3449 MG |
785 | |
786 | /* | |
787 | * It's not safe for direct compaction to call lock_page. | |
788 | * For example, during page readahead pages are added locked | |
789 | * to the LRU. Later, when the IO completes the pages are | |
790 | * marked uptodate and unlocked. However, the queueing | |
791 | * could be merging multiple pages for one bio (e.g. | |
792 | * mpage_readpages). If an allocation happens for the | |
793 | * second or third page, the process can end up locking | |
794 | * the same page twice and deadlocking. Rather than | |
795 | * trying to be clever about what pages can be locked, | |
796 | * avoid the use of lock_page for direct compaction | |
797 | * altogether. | |
798 | */ | |
799 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 800 | goto out; |
3e7d3449 | 801 | |
e24f0b8f CL |
802 | lock_page(page); |
803 | } | |
804 | ||
805 | if (PageWriteback(page)) { | |
11bc82d6 | 806 | /* |
fed5b64a | 807 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
808 | * necessary to wait for PageWriteback. In the async case, |
809 | * the retry loop is too short and in the sync-light case, | |
810 | * the overhead of stalling is too much | |
11bc82d6 | 811 | */ |
a6bc32b8 | 812 | if (mode != MIGRATE_SYNC) { |
11bc82d6 | 813 | rc = -EBUSY; |
0a31bc97 | 814 | goto out_unlock; |
11bc82d6 AA |
815 | } |
816 | if (!force) | |
0a31bc97 | 817 | goto out_unlock; |
e24f0b8f CL |
818 | wait_on_page_writeback(page); |
819 | } | |
e24f0b8f | 820 | /* |
dc386d4d KH |
821 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
822 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 823 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 824 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
825 | * File Caches may use write_page() or lock_page() in migration, then, |
826 | * just care Anon page here. | |
dc386d4d | 827 | */ |
b79bc0a0 | 828 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 829 | /* |
4fc3f1d6 | 830 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
831 | * getting a hold on an anon_vma from outside one of its mms. |
832 | */ | |
746b18d4 | 833 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
834 | if (anon_vma) { |
835 | /* | |
746b18d4 | 836 | * Anon page |
1ce82b69 | 837 | */ |
1ce82b69 | 838 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
839 | /* |
840 | * We cannot be sure that the anon_vma of an unmapped | |
841 | * swapcache page is safe to use because we don't | |
842 | * know in advance if the VMA that this page belonged | |
843 | * to still exists. If the VMA and others sharing the | |
844 | * data have been freed, then the anon_vma could | |
845 | * already be invalid. | |
846 | * | |
847 | * To avoid this possibility, swapcache pages get | |
848 | * migrated but are not remapped when migration | |
849 | * completes | |
850 | */ | |
3fe2011f | 851 | } else { |
0a31bc97 | 852 | goto out_unlock; |
3fe2011f | 853 | } |
989f89c5 | 854 | } |
62e1c553 | 855 | |
7db7671f HD |
856 | /* |
857 | * Block others from accessing the new page when we get around to | |
858 | * establishing additional references. We are usually the only one | |
859 | * holding a reference to newpage at this point. We used to have a BUG | |
860 | * here if trylock_page(newpage) fails, but would like to allow for | |
861 | * cases where there might be a race with the previous use of newpage. | |
862 | * This is much like races on refcount of oldpage: just don't BUG(). | |
863 | */ | |
864 | if (unlikely(!trylock_page(newpage))) | |
865 | goto out_unlock; | |
866 | ||
d6d86c0a | 867 | if (unlikely(isolated_balloon_page(page))) { |
bf6bddf1 RA |
868 | /* |
869 | * A ballooned page does not need any special attention from | |
870 | * physical to virtual reverse mapping procedures. | |
871 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
872 | * in order to avoid burning cycles at rmap level, and perform | |
873 | * the page migration right away (proteced by page lock). | |
874 | */ | |
875 | rc = balloon_page_migrate(newpage, page, mode); | |
7db7671f | 876 | goto out_unlock_both; |
bf6bddf1 RA |
877 | } |
878 | ||
dc386d4d | 879 | /* |
62e1c553 SL |
880 | * Corner case handling: |
881 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
882 | * and treated as swapcache but it has no rmap yet. | |
883 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
884 | * trigger a BUG. So handle it here. | |
885 | * 2. An orphaned page (see truncate_complete_page) might have | |
886 | * fs-private metadata. The page can be picked up due to memory | |
887 | * offlining. Everywhere else except page reclaim, the page is | |
888 | * invisible to the vm, so the page can not be migrated. So try to | |
889 | * free the metadata, so the page can be freed. | |
e24f0b8f | 890 | */ |
62e1c553 | 891 | if (!page->mapping) { |
309381fe | 892 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 893 | if (page_has_private(page)) { |
62e1c553 | 894 | try_to_free_buffers(page); |
7db7671f | 895 | goto out_unlock_both; |
62e1c553 | 896 | } |
7db7671f HD |
897 | } else if (page_mapped(page)) { |
898 | /* Establish migration ptes */ | |
2ebba6b7 | 899 | try_to_unmap(page, |
da1b13cc | 900 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
901 | page_was_mapped = 1; |
902 | } | |
dc386d4d | 903 | |
e6a1530d | 904 | if (!page_mapped(page)) |
2ebba6b7 | 905 | rc = move_to_new_page(newpage, page, page_was_mapped, mode); |
e24f0b8f | 906 | |
2ebba6b7 | 907 | if (rc && page_was_mapped) |
e24f0b8f | 908 | remove_migration_ptes(page, page); |
3f6c8272 | 909 | |
7db7671f HD |
910 | out_unlock_both: |
911 | unlock_page(newpage); | |
912 | out_unlock: | |
3f6c8272 | 913 | /* Drop an anon_vma reference if we took one */ |
76545066 | 914 | if (anon_vma) |
9e60109f | 915 | put_anon_vma(anon_vma); |
e24f0b8f | 916 | unlock_page(page); |
0dabec93 MK |
917 | out: |
918 | return rc; | |
919 | } | |
95a402c3 | 920 | |
ef2a5153 GU |
921 | /* |
922 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
923 | * around it. | |
924 | */ | |
925 | #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) | |
926 | #define ICE_noinline noinline | |
927 | #else | |
928 | #define ICE_noinline | |
929 | #endif | |
930 | ||
0dabec93 MK |
931 | /* |
932 | * Obtain the lock on page, remove all ptes and migrate the page | |
933 | * to the newly allocated page in newpage. | |
934 | */ | |
ef2a5153 GU |
935 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
936 | free_page_t put_new_page, | |
937 | unsigned long private, struct page *page, | |
add05cec NH |
938 | int force, enum migrate_mode mode, |
939 | enum migrate_reason reason) | |
0dabec93 | 940 | { |
2def7424 | 941 | int rc = MIGRATEPAGE_SUCCESS; |
0dabec93 | 942 | int *result = NULL; |
2def7424 | 943 | struct page *newpage; |
0dabec93 | 944 | |
2def7424 | 945 | newpage = get_new_page(page, private, &result); |
0dabec93 MK |
946 | if (!newpage) |
947 | return -ENOMEM; | |
948 | ||
949 | if (page_count(page) == 1) { | |
950 | /* page was freed from under us. So we are done. */ | |
951 | goto out; | |
952 | } | |
953 | ||
954 | if (unlikely(PageTransHuge(page))) | |
955 | if (unlikely(split_huge_page(page))) | |
956 | goto out; | |
957 | ||
9c620e2b | 958 | rc = __unmap_and_move(page, newpage, force, mode); |
2def7424 HD |
959 | if (rc == MIGRATEPAGE_SUCCESS) |
960 | put_new_page = NULL; | |
bf6bddf1 | 961 | |
0dabec93 | 962 | out: |
e24f0b8f | 963 | if (rc != -EAGAIN) { |
0dabec93 MK |
964 | /* |
965 | * A page that has been migrated has all references | |
966 | * removed and will be freed. A page that has not been | |
967 | * migrated will have kepts its references and be | |
968 | * restored. | |
969 | */ | |
970 | list_del(&page->lru); | |
a731286d | 971 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 972 | page_is_file_cache(page)); |
f4c18e6f | 973 | /* Soft-offlined page shouldn't go through lru cache list */ |
da1b13cc | 974 | if (reason == MR_MEMORY_FAILURE) { |
f4c18e6f | 975 | put_page(page); |
da1b13cc WL |
976 | if (!test_set_page_hwpoison(page)) |
977 | num_poisoned_pages_inc(); | |
978 | } else | |
add05cec | 979 | putback_lru_page(page); |
e24f0b8f | 980 | } |
68711a74 | 981 | |
95a402c3 | 982 | /* |
68711a74 DR |
983 | * If migration was not successful and there's a freeing callback, use |
984 | * it. Otherwise, putback_lru_page() will drop the reference grabbed | |
985 | * during isolation. | |
95a402c3 | 986 | */ |
2def7424 | 987 | if (put_new_page) { |
8bdd6380 | 988 | ClearPageSwapBacked(newpage); |
68711a74 | 989 | put_new_page(newpage, private); |
d6d86c0a KK |
990 | } else if (unlikely(__is_movable_balloon_page(newpage))) { |
991 | /* drop our reference, page already in the balloon */ | |
992 | put_page(newpage); | |
8bdd6380 | 993 | } else |
68711a74 DR |
994 | putback_lru_page(newpage); |
995 | ||
742755a1 CL |
996 | if (result) { |
997 | if (rc) | |
998 | *result = rc; | |
999 | else | |
1000 | *result = page_to_nid(newpage); | |
1001 | } | |
e24f0b8f CL |
1002 | return rc; |
1003 | } | |
1004 | ||
290408d4 NH |
1005 | /* |
1006 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1007 | * | |
1008 | * This function doesn't wait the completion of hugepage I/O | |
1009 | * because there is no race between I/O and migration for hugepage. | |
1010 | * Note that currently hugepage I/O occurs only in direct I/O | |
1011 | * where no lock is held and PG_writeback is irrelevant, | |
1012 | * and writeback status of all subpages are counted in the reference | |
1013 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1014 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1015 | * This means that when we try to migrate hugepage whose subpages are | |
1016 | * doing direct I/O, some references remain after try_to_unmap() and | |
1017 | * hugepage migration fails without data corruption. | |
1018 | * | |
1019 | * There is also no race when direct I/O is issued on the page under migration, | |
1020 | * because then pte is replaced with migration swap entry and direct I/O code | |
1021 | * will wait in the page fault for migration to complete. | |
1022 | */ | |
1023 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1024 | free_page_t put_new_page, unsigned long private, |
1025 | struct page *hpage, int force, | |
1026 | enum migrate_mode mode) | |
290408d4 | 1027 | { |
2def7424 | 1028 | int rc = -EAGAIN; |
290408d4 | 1029 | int *result = NULL; |
2ebba6b7 | 1030 | int page_was_mapped = 0; |
32665f2b | 1031 | struct page *new_hpage; |
290408d4 NH |
1032 | struct anon_vma *anon_vma = NULL; |
1033 | ||
83467efb NH |
1034 | /* |
1035 | * Movability of hugepages depends on architectures and hugepage size. | |
1036 | * This check is necessary because some callers of hugepage migration | |
1037 | * like soft offline and memory hotremove don't walk through page | |
1038 | * tables or check whether the hugepage is pmd-based or not before | |
1039 | * kicking migration. | |
1040 | */ | |
100873d7 | 1041 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1042 | putback_active_hugepage(hpage); |
83467efb | 1043 | return -ENOSYS; |
32665f2b | 1044 | } |
83467efb | 1045 | |
32665f2b | 1046 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1047 | if (!new_hpage) |
1048 | return -ENOMEM; | |
1049 | ||
290408d4 | 1050 | if (!trylock_page(hpage)) { |
a6bc32b8 | 1051 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1052 | goto out; |
1053 | lock_page(hpage); | |
1054 | } | |
1055 | ||
746b18d4 PZ |
1056 | if (PageAnon(hpage)) |
1057 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1058 | |
7db7671f HD |
1059 | if (unlikely(!trylock_page(new_hpage))) |
1060 | goto put_anon; | |
1061 | ||
2ebba6b7 HD |
1062 | if (page_mapped(hpage)) { |
1063 | try_to_unmap(hpage, | |
1064 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1065 | page_was_mapped = 1; | |
1066 | } | |
290408d4 NH |
1067 | |
1068 | if (!page_mapped(hpage)) | |
2ebba6b7 | 1069 | rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode); |
290408d4 | 1070 | |
2ebba6b7 | 1071 | if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped) |
290408d4 NH |
1072 | remove_migration_ptes(hpage, hpage); |
1073 | ||
7db7671f HD |
1074 | unlock_page(new_hpage); |
1075 | ||
1076 | put_anon: | |
fd4a4663 | 1077 | if (anon_vma) |
9e60109f | 1078 | put_anon_vma(anon_vma); |
8e6ac7fa | 1079 | |
2def7424 | 1080 | if (rc == MIGRATEPAGE_SUCCESS) { |
8e6ac7fa | 1081 | hugetlb_cgroup_migrate(hpage, new_hpage); |
2def7424 HD |
1082 | put_new_page = NULL; |
1083 | } | |
8e6ac7fa | 1084 | |
290408d4 | 1085 | unlock_page(hpage); |
09761333 | 1086 | out: |
b8ec1cee NH |
1087 | if (rc != -EAGAIN) |
1088 | putback_active_hugepage(hpage); | |
68711a74 DR |
1089 | |
1090 | /* | |
1091 | * If migration was not successful and there's a freeing callback, use | |
1092 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1093 | * isolation. | |
1094 | */ | |
2def7424 | 1095 | if (put_new_page) |
68711a74 DR |
1096 | put_new_page(new_hpage, private); |
1097 | else | |
3aaa76e1 | 1098 | putback_active_hugepage(new_hpage); |
68711a74 | 1099 | |
290408d4 NH |
1100 | if (result) { |
1101 | if (rc) | |
1102 | *result = rc; | |
1103 | else | |
1104 | *result = page_to_nid(new_hpage); | |
1105 | } | |
1106 | return rc; | |
1107 | } | |
1108 | ||
b20a3503 | 1109 | /* |
c73e5c9c SB |
1110 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1111 | * supplied as the target for the page migration | |
b20a3503 | 1112 | * |
c73e5c9c SB |
1113 | * @from: The list of pages to be migrated. |
1114 | * @get_new_page: The function used to allocate free pages to be used | |
1115 | * as the target of the page migration. | |
68711a74 DR |
1116 | * @put_new_page: The function used to free target pages if migration |
1117 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1118 | * @private: Private data to be passed on to get_new_page() |
1119 | * @mode: The migration mode that specifies the constraints for | |
1120 | * page migration, if any. | |
1121 | * @reason: The reason for page migration. | |
b20a3503 | 1122 | * |
c73e5c9c SB |
1123 | * The function returns after 10 attempts or if no pages are movable any more |
1124 | * because the list has become empty or no retryable pages exist any more. | |
14e0f9bc | 1125 | * The caller should call putback_movable_pages() to return pages to the LRU |
28bd6578 | 1126 | * or free list only if ret != 0. |
b20a3503 | 1127 | * |
c73e5c9c | 1128 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1129 | */ |
9c620e2b | 1130 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1131 | free_page_t put_new_page, unsigned long private, |
1132 | enum migrate_mode mode, int reason) | |
b20a3503 | 1133 | { |
e24f0b8f | 1134 | int retry = 1; |
b20a3503 | 1135 | int nr_failed = 0; |
5647bc29 | 1136 | int nr_succeeded = 0; |
b20a3503 CL |
1137 | int pass = 0; |
1138 | struct page *page; | |
1139 | struct page *page2; | |
1140 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1141 | int rc; | |
1142 | ||
1143 | if (!swapwrite) | |
1144 | current->flags |= PF_SWAPWRITE; | |
1145 | ||
e24f0b8f CL |
1146 | for(pass = 0; pass < 10 && retry; pass++) { |
1147 | retry = 0; | |
b20a3503 | 1148 | |
e24f0b8f | 1149 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1150 | cond_resched(); |
2d1db3b1 | 1151 | |
31caf665 NH |
1152 | if (PageHuge(page)) |
1153 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 DR |
1154 | put_new_page, private, page, |
1155 | pass > 2, mode); | |
31caf665 | 1156 | else |
68711a74 | 1157 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1158 | private, page, pass > 2, mode, |
1159 | reason); | |
2d1db3b1 | 1160 | |
e24f0b8f | 1161 | switch(rc) { |
95a402c3 CL |
1162 | case -ENOMEM: |
1163 | goto out; | |
e24f0b8f | 1164 | case -EAGAIN: |
2d1db3b1 | 1165 | retry++; |
e24f0b8f | 1166 | break; |
78bd5209 | 1167 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1168 | nr_succeeded++; |
e24f0b8f CL |
1169 | break; |
1170 | default: | |
354a3363 NH |
1171 | /* |
1172 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1173 | * unlike -EAGAIN case, the failed page is | |
1174 | * removed from migration page list and not | |
1175 | * retried in the next outer loop. | |
1176 | */ | |
2d1db3b1 | 1177 | nr_failed++; |
e24f0b8f | 1178 | break; |
2d1db3b1 | 1179 | } |
b20a3503 CL |
1180 | } |
1181 | } | |
f2f81fb2 VB |
1182 | nr_failed += retry; |
1183 | rc = nr_failed; | |
95a402c3 | 1184 | out: |
5647bc29 MG |
1185 | if (nr_succeeded) |
1186 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1187 | if (nr_failed) | |
1188 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1189 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1190 | ||
b20a3503 CL |
1191 | if (!swapwrite) |
1192 | current->flags &= ~PF_SWAPWRITE; | |
1193 | ||
78bd5209 | 1194 | return rc; |
b20a3503 | 1195 | } |
95a402c3 | 1196 | |
742755a1 CL |
1197 | #ifdef CONFIG_NUMA |
1198 | /* | |
1199 | * Move a list of individual pages | |
1200 | */ | |
1201 | struct page_to_node { | |
1202 | unsigned long addr; | |
1203 | struct page *page; | |
1204 | int node; | |
1205 | int status; | |
1206 | }; | |
1207 | ||
1208 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1209 | int **result) | |
1210 | { | |
1211 | struct page_to_node *pm = (struct page_to_node *)private; | |
1212 | ||
1213 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1214 | pm++; | |
1215 | ||
1216 | if (pm->node == MAX_NUMNODES) | |
1217 | return NULL; | |
1218 | ||
1219 | *result = &pm->status; | |
1220 | ||
e632a938 NH |
1221 | if (PageHuge(p)) |
1222 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1223 | pm->node); | |
1224 | else | |
96db800f | 1225 | return __alloc_pages_node(pm->node, |
e97ca8e5 | 1226 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); |
742755a1 CL |
1227 | } |
1228 | ||
1229 | /* | |
1230 | * Move a set of pages as indicated in the pm array. The addr | |
1231 | * field must be set to the virtual address of the page to be moved | |
1232 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1233 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1234 | */ |
5e9a0f02 BG |
1235 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1236 | struct page_to_node *pm, | |
1237 | int migrate_all) | |
742755a1 CL |
1238 | { |
1239 | int err; | |
1240 | struct page_to_node *pp; | |
1241 | LIST_HEAD(pagelist); | |
1242 | ||
1243 | down_read(&mm->mmap_sem); | |
1244 | ||
1245 | /* | |
1246 | * Build a list of pages to migrate | |
1247 | */ | |
742755a1 CL |
1248 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1249 | struct vm_area_struct *vma; | |
1250 | struct page *page; | |
1251 | ||
742755a1 CL |
1252 | err = -EFAULT; |
1253 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1254 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1255 | goto set_status; |
1256 | ||
d899844e KS |
1257 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1258 | page = follow_page(vma, pp->addr, | |
1259 | FOLL_GET | FOLL_SPLIT | FOLL_DUMP); | |
89f5b7da LT |
1260 | |
1261 | err = PTR_ERR(page); | |
1262 | if (IS_ERR(page)) | |
1263 | goto set_status; | |
1264 | ||
742755a1 CL |
1265 | err = -ENOENT; |
1266 | if (!page) | |
1267 | goto set_status; | |
1268 | ||
742755a1 CL |
1269 | pp->page = page; |
1270 | err = page_to_nid(page); | |
1271 | ||
1272 | if (err == pp->node) | |
1273 | /* | |
1274 | * Node already in the right place | |
1275 | */ | |
1276 | goto put_and_set; | |
1277 | ||
1278 | err = -EACCES; | |
1279 | if (page_mapcount(page) > 1 && | |
1280 | !migrate_all) | |
1281 | goto put_and_set; | |
1282 | ||
e632a938 | 1283 | if (PageHuge(page)) { |
e66f17ff NH |
1284 | if (PageHead(page)) |
1285 | isolate_huge_page(page, &pagelist); | |
e632a938 NH |
1286 | goto put_and_set; |
1287 | } | |
1288 | ||
62695a84 | 1289 | err = isolate_lru_page(page); |
6d9c285a | 1290 | if (!err) { |
62695a84 | 1291 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1292 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1293 | page_is_file_cache(page)); | |
1294 | } | |
742755a1 CL |
1295 | put_and_set: |
1296 | /* | |
1297 | * Either remove the duplicate refcount from | |
1298 | * isolate_lru_page() or drop the page ref if it was | |
1299 | * not isolated. | |
1300 | */ | |
1301 | put_page(page); | |
1302 | set_status: | |
1303 | pp->status = err; | |
1304 | } | |
1305 | ||
e78bbfa8 | 1306 | err = 0; |
cf608ac1 | 1307 | if (!list_empty(&pagelist)) { |
68711a74 | 1308 | err = migrate_pages(&pagelist, new_page_node, NULL, |
9c620e2b | 1309 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1310 | if (err) |
e632a938 | 1311 | putback_movable_pages(&pagelist); |
cf608ac1 | 1312 | } |
742755a1 CL |
1313 | |
1314 | up_read(&mm->mmap_sem); | |
1315 | return err; | |
1316 | } | |
1317 | ||
5e9a0f02 BG |
1318 | /* |
1319 | * Migrate an array of page address onto an array of nodes and fill | |
1320 | * the corresponding array of status. | |
1321 | */ | |
3268c63e | 1322 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1323 | unsigned long nr_pages, |
1324 | const void __user * __user *pages, | |
1325 | const int __user *nodes, | |
1326 | int __user *status, int flags) | |
1327 | { | |
3140a227 | 1328 | struct page_to_node *pm; |
3140a227 BG |
1329 | unsigned long chunk_nr_pages; |
1330 | unsigned long chunk_start; | |
1331 | int err; | |
5e9a0f02 | 1332 | |
3140a227 BG |
1333 | err = -ENOMEM; |
1334 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1335 | if (!pm) | |
5e9a0f02 | 1336 | goto out; |
35282a2d BG |
1337 | |
1338 | migrate_prep(); | |
1339 | ||
5e9a0f02 | 1340 | /* |
3140a227 BG |
1341 | * Store a chunk of page_to_node array in a page, |
1342 | * but keep the last one as a marker | |
5e9a0f02 | 1343 | */ |
3140a227 | 1344 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1345 | |
3140a227 BG |
1346 | for (chunk_start = 0; |
1347 | chunk_start < nr_pages; | |
1348 | chunk_start += chunk_nr_pages) { | |
1349 | int j; | |
5e9a0f02 | 1350 | |
3140a227 BG |
1351 | if (chunk_start + chunk_nr_pages > nr_pages) |
1352 | chunk_nr_pages = nr_pages - chunk_start; | |
1353 | ||
1354 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1355 | for (j = 0; j < chunk_nr_pages; j++) { | |
1356 | const void __user *p; | |
5e9a0f02 BG |
1357 | int node; |
1358 | ||
3140a227 BG |
1359 | err = -EFAULT; |
1360 | if (get_user(p, pages + j + chunk_start)) | |
1361 | goto out_pm; | |
1362 | pm[j].addr = (unsigned long) p; | |
1363 | ||
1364 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1365 | goto out_pm; |
1366 | ||
1367 | err = -ENODEV; | |
6f5a55f1 LT |
1368 | if (node < 0 || node >= MAX_NUMNODES) |
1369 | goto out_pm; | |
1370 | ||
389162c2 | 1371 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1372 | goto out_pm; |
1373 | ||
1374 | err = -EACCES; | |
1375 | if (!node_isset(node, task_nodes)) | |
1376 | goto out_pm; | |
1377 | ||
3140a227 BG |
1378 | pm[j].node = node; |
1379 | } | |
1380 | ||
1381 | /* End marker for this chunk */ | |
1382 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1383 | ||
1384 | /* Migrate this chunk */ | |
1385 | err = do_move_page_to_node_array(mm, pm, | |
1386 | flags & MPOL_MF_MOVE_ALL); | |
1387 | if (err < 0) | |
1388 | goto out_pm; | |
5e9a0f02 | 1389 | |
5e9a0f02 | 1390 | /* Return status information */ |
3140a227 BG |
1391 | for (j = 0; j < chunk_nr_pages; j++) |
1392 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1393 | err = -EFAULT; |
3140a227 BG |
1394 | goto out_pm; |
1395 | } | |
1396 | } | |
1397 | err = 0; | |
5e9a0f02 BG |
1398 | |
1399 | out_pm: | |
3140a227 | 1400 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1401 | out: |
1402 | return err; | |
1403 | } | |
1404 | ||
742755a1 | 1405 | /* |
2f007e74 | 1406 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1407 | */ |
80bba129 BG |
1408 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1409 | const void __user **pages, int *status) | |
742755a1 | 1410 | { |
2f007e74 | 1411 | unsigned long i; |
2f007e74 | 1412 | |
742755a1 CL |
1413 | down_read(&mm->mmap_sem); |
1414 | ||
2f007e74 | 1415 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1416 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1417 | struct vm_area_struct *vma; |
1418 | struct page *page; | |
c095adbc | 1419 | int err = -EFAULT; |
2f007e74 BG |
1420 | |
1421 | vma = find_vma(mm, addr); | |
70384dc6 | 1422 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1423 | goto set_status; |
1424 | ||
d899844e KS |
1425 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1426 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1427 | |
1428 | err = PTR_ERR(page); | |
1429 | if (IS_ERR(page)) | |
1430 | goto set_status; | |
1431 | ||
d899844e | 1432 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1433 | set_status: |
80bba129 BG |
1434 | *status = err; |
1435 | ||
1436 | pages++; | |
1437 | status++; | |
1438 | } | |
1439 | ||
1440 | up_read(&mm->mmap_sem); | |
1441 | } | |
1442 | ||
1443 | /* | |
1444 | * Determine the nodes of a user array of pages and store it in | |
1445 | * a user array of status. | |
1446 | */ | |
1447 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1448 | const void __user * __user *pages, | |
1449 | int __user *status) | |
1450 | { | |
1451 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1452 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1453 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1454 | |
87b8d1ad PA |
1455 | while (nr_pages) { |
1456 | unsigned long chunk_nr; | |
80bba129 | 1457 | |
87b8d1ad PA |
1458 | chunk_nr = nr_pages; |
1459 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1460 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1461 | ||
1462 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1463 | break; | |
80bba129 BG |
1464 | |
1465 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1466 | ||
87b8d1ad PA |
1467 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1468 | break; | |
742755a1 | 1469 | |
87b8d1ad PA |
1470 | pages += chunk_nr; |
1471 | status += chunk_nr; | |
1472 | nr_pages -= chunk_nr; | |
1473 | } | |
1474 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1475 | } |
1476 | ||
1477 | /* | |
1478 | * Move a list of pages in the address space of the currently executing | |
1479 | * process. | |
1480 | */ | |
938bb9f5 HC |
1481 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1482 | const void __user * __user *, pages, | |
1483 | const int __user *, nodes, | |
1484 | int __user *, status, int, flags) | |
742755a1 | 1485 | { |
c69e8d9c | 1486 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1487 | struct task_struct *task; |
742755a1 | 1488 | struct mm_struct *mm; |
5e9a0f02 | 1489 | int err; |
3268c63e | 1490 | nodemask_t task_nodes; |
742755a1 CL |
1491 | |
1492 | /* Check flags */ | |
1493 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1494 | return -EINVAL; | |
1495 | ||
1496 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1497 | return -EPERM; | |
1498 | ||
1499 | /* Find the mm_struct */ | |
a879bf58 | 1500 | rcu_read_lock(); |
228ebcbe | 1501 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1502 | if (!task) { |
a879bf58 | 1503 | rcu_read_unlock(); |
742755a1 CL |
1504 | return -ESRCH; |
1505 | } | |
3268c63e | 1506 | get_task_struct(task); |
742755a1 CL |
1507 | |
1508 | /* | |
1509 | * Check if this process has the right to modify the specified | |
1510 | * process. The right exists if the process has administrative | |
1511 | * capabilities, superuser privileges or the same | |
1512 | * userid as the target process. | |
1513 | */ | |
c69e8d9c | 1514 | tcred = __task_cred(task); |
b38a86eb EB |
1515 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1516 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1517 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1518 | rcu_read_unlock(); |
742755a1 | 1519 | err = -EPERM; |
5e9a0f02 | 1520 | goto out; |
742755a1 | 1521 | } |
c69e8d9c | 1522 | rcu_read_unlock(); |
742755a1 | 1523 | |
86c3a764 DQ |
1524 | err = security_task_movememory(task); |
1525 | if (err) | |
5e9a0f02 | 1526 | goto out; |
86c3a764 | 1527 | |
3268c63e CL |
1528 | task_nodes = cpuset_mems_allowed(task); |
1529 | mm = get_task_mm(task); | |
1530 | put_task_struct(task); | |
1531 | ||
6e8b09ea SL |
1532 | if (!mm) |
1533 | return -EINVAL; | |
1534 | ||
1535 | if (nodes) | |
1536 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1537 | nodes, status, flags); | |
1538 | else | |
1539 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1540 | |
742755a1 CL |
1541 | mmput(mm); |
1542 | return err; | |
3268c63e CL |
1543 | |
1544 | out: | |
1545 | put_task_struct(task); | |
1546 | return err; | |
742755a1 | 1547 | } |
742755a1 | 1548 | |
7039e1db PZ |
1549 | #ifdef CONFIG_NUMA_BALANCING |
1550 | /* | |
1551 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1552 | * pages. Currently it only checks the watermarks which crude | |
1553 | */ | |
1554 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1555 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1556 | { |
1557 | int z; | |
1558 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1559 | struct zone *zone = pgdat->node_zones + z; | |
1560 | ||
1561 | if (!populated_zone(zone)) | |
1562 | continue; | |
1563 | ||
6e543d57 | 1564 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1565 | continue; |
1566 | ||
1567 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1568 | if (!zone_watermark_ok(zone, 0, | |
1569 | high_wmark_pages(zone) + | |
1570 | nr_migrate_pages, | |
1571 | 0, 0)) | |
1572 | continue; | |
1573 | return true; | |
1574 | } | |
1575 | return false; | |
1576 | } | |
1577 | ||
1578 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1579 | unsigned long data, | |
1580 | int **result) | |
1581 | { | |
1582 | int nid = (int) data; | |
1583 | struct page *newpage; | |
1584 | ||
96db800f | 1585 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1586 | (GFP_HIGHUSER_MOVABLE | |
1587 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1588 | __GFP_NORETRY | __GFP_NOWARN) & | |
7039e1db | 1589 | ~GFP_IOFS, 0); |
bac0382c | 1590 | |
7039e1db PZ |
1591 | return newpage; |
1592 | } | |
1593 | ||
a8f60772 MG |
1594 | /* |
1595 | * page migration rate limiting control. | |
1596 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1597 | * window of time. Default here says do not migrate more than 1280M per second. | |
1598 | */ | |
1599 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1600 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1601 | ||
b32967ff | 1602 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1603 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1604 | unsigned long nr_pages) | |
7039e1db | 1605 | { |
a8f60772 MG |
1606 | /* |
1607 | * Rate-limit the amount of data that is being migrated to a node. | |
1608 | * Optimal placement is no good if the memory bus is saturated and | |
1609 | * all the time is being spent migrating! | |
1610 | */ | |
a8f60772 | 1611 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1612 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1613 | pgdat->numabalancing_migrate_nr_pages = 0; |
1614 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1615 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1616 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1617 | } |
af1839d7 MG |
1618 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1619 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1620 | nr_pages); | |
1c5e9c27 | 1621 | return true; |
af1839d7 | 1622 | } |
1c5e9c27 MG |
1623 | |
1624 | /* | |
1625 | * This is an unlocked non-atomic update so errors are possible. | |
1626 | * The consequences are failing to migrate when we potentiall should | |
1627 | * have which is not severe enough to warrant locking. If it is ever | |
1628 | * a problem, it can be converted to a per-cpu counter. | |
1629 | */ | |
1630 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1631 | return false; | |
b32967ff MG |
1632 | } |
1633 | ||
1c30e017 | 1634 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1635 | { |
340ef390 | 1636 | int page_lru; |
a8f60772 | 1637 | |
309381fe | 1638 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1639 | |
7039e1db | 1640 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1641 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1642 | return 0; | |
7039e1db | 1643 | |
340ef390 HD |
1644 | if (isolate_lru_page(page)) |
1645 | return 0; | |
7039e1db | 1646 | |
340ef390 HD |
1647 | /* |
1648 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1649 | * check on page_count(), so we must do it here, now that the page | |
1650 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1651 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1652 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1653 | */ | |
1654 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1655 | putback_lru_page(page); | |
1656 | return 0; | |
7039e1db PZ |
1657 | } |
1658 | ||
340ef390 HD |
1659 | page_lru = page_is_file_cache(page); |
1660 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1661 | hpage_nr_pages(page)); | |
1662 | ||
149c33e1 | 1663 | /* |
340ef390 HD |
1664 | * Isolating the page has taken another reference, so the |
1665 | * caller's reference can be safely dropped without the page | |
1666 | * disappearing underneath us during migration. | |
149c33e1 MG |
1667 | */ |
1668 | put_page(page); | |
340ef390 | 1669 | return 1; |
b32967ff MG |
1670 | } |
1671 | ||
de466bd6 MG |
1672 | bool pmd_trans_migrating(pmd_t pmd) |
1673 | { | |
1674 | struct page *page = pmd_page(pmd); | |
1675 | return PageLocked(page); | |
1676 | } | |
1677 | ||
b32967ff MG |
1678 | /* |
1679 | * Attempt to migrate a misplaced page to the specified destination | |
1680 | * node. Caller is expected to have an elevated reference count on | |
1681 | * the page that will be dropped by this function before returning. | |
1682 | */ | |
1bc115d8 MG |
1683 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1684 | int node) | |
b32967ff MG |
1685 | { |
1686 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1687 | int isolated; |
b32967ff MG |
1688 | int nr_remaining; |
1689 | LIST_HEAD(migratepages); | |
1690 | ||
1691 | /* | |
1bc115d8 MG |
1692 | * Don't migrate file pages that are mapped in multiple processes |
1693 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1694 | */ |
1bc115d8 MG |
1695 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1696 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1697 | goto out; |
b32967ff MG |
1698 | |
1699 | /* | |
1700 | * Rate-limit the amount of data that is being migrated to a node. | |
1701 | * Optimal placement is no good if the memory bus is saturated and | |
1702 | * all the time is being spent migrating! | |
1703 | */ | |
340ef390 | 1704 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1705 | goto out; |
b32967ff MG |
1706 | |
1707 | isolated = numamigrate_isolate_page(pgdat, page); | |
1708 | if (!isolated) | |
1709 | goto out; | |
1710 | ||
1711 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1712 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1713 | NULL, node, MIGRATE_ASYNC, |
1714 | MR_NUMA_MISPLACED); | |
b32967ff | 1715 | if (nr_remaining) { |
59c82b70 JK |
1716 | if (!list_empty(&migratepages)) { |
1717 | list_del(&page->lru); | |
1718 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1719 | page_is_file_cache(page)); | |
1720 | putback_lru_page(page); | |
1721 | } | |
b32967ff MG |
1722 | isolated = 0; |
1723 | } else | |
1724 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1725 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1726 | return isolated; |
340ef390 HD |
1727 | |
1728 | out: | |
1729 | put_page(page); | |
1730 | return 0; | |
7039e1db | 1731 | } |
220018d3 | 1732 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1733 | |
220018d3 | 1734 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1735 | /* |
1736 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1737 | * before returning. | |
1738 | */ | |
b32967ff MG |
1739 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1740 | struct vm_area_struct *vma, | |
1741 | pmd_t *pmd, pmd_t entry, | |
1742 | unsigned long address, | |
1743 | struct page *page, int node) | |
1744 | { | |
c4088ebd | 1745 | spinlock_t *ptl; |
b32967ff MG |
1746 | pg_data_t *pgdat = NODE_DATA(node); |
1747 | int isolated = 0; | |
1748 | struct page *new_page = NULL; | |
b32967ff | 1749 | int page_lru = page_is_file_cache(page); |
f714f4f2 MG |
1750 | unsigned long mmun_start = address & HPAGE_PMD_MASK; |
1751 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
2b4847e7 | 1752 | pmd_t orig_entry; |
b32967ff | 1753 | |
b32967ff MG |
1754 | /* |
1755 | * Rate-limit the amount of data that is being migrated to a node. | |
1756 | * Optimal placement is no good if the memory bus is saturated and | |
1757 | * all the time is being spent migrating! | |
1758 | */ | |
d28d4335 | 1759 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1760 | goto out_dropref; |
1761 | ||
1762 | new_page = alloc_pages_node(node, | |
e97ca8e5 JW |
1763 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT, |
1764 | HPAGE_PMD_ORDER); | |
340ef390 HD |
1765 | if (!new_page) |
1766 | goto out_fail; | |
1767 | ||
b32967ff | 1768 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 1769 | if (!isolated) { |
b32967ff | 1770 | put_page(new_page); |
340ef390 | 1771 | goto out_fail; |
b32967ff MG |
1772 | } |
1773 | ||
b0943d61 MG |
1774 | if (mm_tlb_flush_pending(mm)) |
1775 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1776 | ||
b32967ff MG |
1777 | /* Prepare a page as a migration target */ |
1778 | __set_page_locked(new_page); | |
1779 | SetPageSwapBacked(new_page); | |
1780 | ||
1781 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1782 | new_page->mapping = page->mapping; | |
1783 | new_page->index = page->index; | |
1784 | migrate_page_copy(new_page, page); | |
1785 | WARN_ON(PageLRU(new_page)); | |
1786 | ||
1787 | /* Recheck the target PMD */ | |
f714f4f2 | 1788 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 1789 | ptl = pmd_lock(mm, pmd); |
2b4847e7 MG |
1790 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { |
1791 | fail_putback: | |
c4088ebd | 1792 | spin_unlock(ptl); |
f714f4f2 | 1793 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1794 | |
1795 | /* Reverse changes made by migrate_page_copy() */ | |
1796 | if (TestClearPageActive(new_page)) | |
1797 | SetPageActive(page); | |
1798 | if (TestClearPageUnevictable(new_page)) | |
1799 | SetPageUnevictable(page); | |
b32967ff MG |
1800 | |
1801 | unlock_page(new_page); | |
1802 | put_page(new_page); /* Free it */ | |
1803 | ||
a54a407f MG |
1804 | /* Retake the callers reference and putback on LRU */ |
1805 | get_page(page); | |
b32967ff | 1806 | putback_lru_page(page); |
a54a407f MG |
1807 | mod_zone_page_state(page_zone(page), |
1808 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
eb4489f6 MG |
1809 | |
1810 | goto out_unlock; | |
b32967ff MG |
1811 | } |
1812 | ||
2b4847e7 | 1813 | orig_entry = *pmd; |
b32967ff | 1814 | entry = mk_pmd(new_page, vma->vm_page_prot); |
b32967ff | 1815 | entry = pmd_mkhuge(entry); |
2b4847e7 | 1816 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 1817 | |
2b4847e7 MG |
1818 | /* |
1819 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1820 | * Any parallel GUP will either observe the old page blocking on the | |
1821 | * page lock, block on the page table lock or observe the new page. | |
1822 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1823 | * guarantee the copy is visible before the pagetable update. | |
1824 | */ | |
f714f4f2 | 1825 | flush_cache_range(vma, mmun_start, mmun_end); |
11de9927 | 1826 | page_add_anon_rmap(new_page, vma, mmun_start); |
8809aa2d | 1827 | pmdp_huge_clear_flush_notify(vma, mmun_start, pmd); |
f714f4f2 MG |
1828 | set_pmd_at(mm, mmun_start, pmd, entry); |
1829 | flush_tlb_range(vma, mmun_start, mmun_end); | |
ce4a9cc5 | 1830 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 MG |
1831 | |
1832 | if (page_count(page) != 2) { | |
f714f4f2 MG |
1833 | set_pmd_at(mm, mmun_start, pmd, orig_entry); |
1834 | flush_tlb_range(vma, mmun_start, mmun_end); | |
34ee645e | 1835 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
2b4847e7 MG |
1836 | update_mmu_cache_pmd(vma, address, &entry); |
1837 | page_remove_rmap(new_page); | |
1838 | goto fail_putback; | |
1839 | } | |
1840 | ||
51afb12b | 1841 | mlock_migrate_page(new_page, page); |
45637bab HD |
1842 | set_page_memcg(new_page, page_memcg(page)); |
1843 | set_page_memcg(page, NULL); | |
b32967ff | 1844 | page_remove_rmap(page); |
2b4847e7 | 1845 | |
c4088ebd | 1846 | spin_unlock(ptl); |
f714f4f2 | 1847 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff | 1848 | |
11de9927 MG |
1849 | /* Take an "isolate" reference and put new page on the LRU. */ |
1850 | get_page(new_page); | |
1851 | putback_lru_page(new_page); | |
1852 | ||
b32967ff MG |
1853 | unlock_page(new_page); |
1854 | unlock_page(page); | |
1855 | put_page(page); /* Drop the rmap reference */ | |
1856 | put_page(page); /* Drop the LRU isolation reference */ | |
1857 | ||
1858 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1859 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1860 | ||
b32967ff MG |
1861 | mod_zone_page_state(page_zone(page), |
1862 | NR_ISOLATED_ANON + page_lru, | |
1863 | -HPAGE_PMD_NR); | |
1864 | return isolated; | |
1865 | ||
340ef390 HD |
1866 | out_fail: |
1867 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1868 | out_dropref: |
2b4847e7 MG |
1869 | ptl = pmd_lock(mm, pmd); |
1870 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 1871 | entry = pmd_modify(entry, vma->vm_page_prot); |
f714f4f2 | 1872 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
1873 | update_mmu_cache_pmd(vma, address, &entry); |
1874 | } | |
1875 | spin_unlock(ptl); | |
a54a407f | 1876 | |
eb4489f6 | 1877 | out_unlock: |
340ef390 | 1878 | unlock_page(page); |
b32967ff | 1879 | put_page(page); |
b32967ff MG |
1880 | return 0; |
1881 | } | |
7039e1db PZ |
1882 | #endif /* CONFIG_NUMA_BALANCING */ |
1883 | ||
1884 | #endif /* CONFIG_NUMA */ |