]>
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> |
f714f4f2 | 39 | #include <linux/mmu_notifier.h> |
33c3fc71 | 40 | #include <linux/page_idle.h> |
b20a3503 | 41 | |
0d1836c3 MN |
42 | #include <asm/tlbflush.h> |
43 | ||
7b2a2d4a MG |
44 | #define CREATE_TRACE_POINTS |
45 | #include <trace/events/migrate.h> | |
46 | ||
b20a3503 CL |
47 | #include "internal.h" |
48 | ||
b20a3503 | 49 | /* |
742755a1 | 50 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
51 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
52 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
53 | */ |
54 | int migrate_prep(void) | |
55 | { | |
b20a3503 CL |
56 | /* |
57 | * Clear the LRU lists so pages can be isolated. | |
58 | * Note that pages may be moved off the LRU after we have | |
59 | * drained them. Those pages will fail to migrate like other | |
60 | * pages that may be busy. | |
61 | */ | |
62 | lru_add_drain_all(); | |
63 | ||
64 | return 0; | |
65 | } | |
66 | ||
748446bb MG |
67 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
68 | int migrate_prep_local(void) | |
69 | { | |
70 | lru_add_drain(); | |
71 | ||
72 | return 0; | |
73 | } | |
74 | ||
5733c7d1 RA |
75 | /* |
76 | * Put previously isolated pages back onto the appropriate lists | |
77 | * from where they were once taken off for compaction/migration. | |
78 | * | |
59c82b70 JK |
79 | * This function shall be used whenever the isolated pageset has been |
80 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
81 | * and isolate_huge_page(). | |
5733c7d1 RA |
82 | */ |
83 | void putback_movable_pages(struct list_head *l) | |
84 | { | |
85 | struct page *page; | |
86 | struct page *page2; | |
87 | ||
b20a3503 | 88 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
89 | if (unlikely(PageHuge(page))) { |
90 | putback_active_hugepage(page); | |
91 | continue; | |
92 | } | |
e24f0b8f | 93 | list_del(&page->lru); |
a731286d | 94 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 95 | page_is_file_cache(page)); |
117aad1e | 96 | if (unlikely(isolated_balloon_page(page))) |
bf6bddf1 RA |
97 | balloon_page_putback(page); |
98 | else | |
99 | putback_lru_page(page); | |
b20a3503 | 100 | } |
b20a3503 CL |
101 | } |
102 | ||
0697212a CL |
103 | /* |
104 | * Restore a potential migration pte to a working pte entry | |
105 | */ | |
e9995ef9 HD |
106 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, |
107 | unsigned long addr, void *old) | |
0697212a CL |
108 | { |
109 | struct mm_struct *mm = vma->vm_mm; | |
110 | swp_entry_t entry; | |
0697212a CL |
111 | pmd_t *pmd; |
112 | pte_t *ptep, pte; | |
113 | spinlock_t *ptl; | |
114 | ||
290408d4 NH |
115 | if (unlikely(PageHuge(new))) { |
116 | ptep = huge_pte_offset(mm, addr); | |
117 | if (!ptep) | |
118 | goto out; | |
cb900f41 | 119 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep); |
290408d4 | 120 | } else { |
6219049a BL |
121 | pmd = mm_find_pmd(mm, addr); |
122 | if (!pmd) | |
290408d4 | 123 | goto out; |
0697212a | 124 | |
290408d4 | 125 | ptep = pte_offset_map(pmd, addr); |
0697212a | 126 | |
486cf46f HD |
127 | /* |
128 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
129 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
130 | */ | |
290408d4 NH |
131 | |
132 | ptl = pte_lockptr(mm, pmd); | |
133 | } | |
0697212a | 134 | |
0697212a CL |
135 | spin_lock(ptl); |
136 | pte = *ptep; | |
137 | if (!is_swap_pte(pte)) | |
e9995ef9 | 138 | goto unlock; |
0697212a CL |
139 | |
140 | entry = pte_to_swp_entry(pte); | |
141 | ||
e9995ef9 HD |
142 | if (!is_migration_entry(entry) || |
143 | migration_entry_to_page(entry) != old) | |
144 | goto unlock; | |
0697212a | 145 | |
0697212a CL |
146 | get_page(new); |
147 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
c3d16e16 CG |
148 | if (pte_swp_soft_dirty(*ptep)) |
149 | pte = pte_mksoft_dirty(pte); | |
d3cb8bf6 MG |
150 | |
151 | /* Recheck VMA as permissions can change since migration started */ | |
0697212a | 152 | if (is_write_migration_entry(entry)) |
d3cb8bf6 MG |
153 | pte = maybe_mkwrite(pte, vma); |
154 | ||
3ef8fd7f | 155 | #ifdef CONFIG_HUGETLB_PAGE |
be7517d6 | 156 | if (PageHuge(new)) { |
290408d4 | 157 | pte = pte_mkhuge(pte); |
be7517d6 TL |
158 | pte = arch_make_huge_pte(pte, vma, new, 0); |
159 | } | |
3ef8fd7f | 160 | #endif |
c2cc499c | 161 | flush_dcache_page(new); |
0697212a | 162 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 163 | |
290408d4 NH |
164 | if (PageHuge(new)) { |
165 | if (PageAnon(new)) | |
166 | hugepage_add_anon_rmap(new, vma, addr); | |
167 | else | |
168 | page_dup_rmap(new); | |
169 | } else if (PageAnon(new)) | |
04e62a29 CL |
170 | page_add_anon_rmap(new, vma, addr); |
171 | else | |
172 | page_add_file_rmap(new); | |
173 | ||
174 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 175 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 176 | unlock: |
0697212a | 177 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
178 | out: |
179 | return SWAP_AGAIN; | |
0697212a CL |
180 | } |
181 | ||
04e62a29 CL |
182 | /* |
183 | * Get rid of all migration entries and replace them by | |
184 | * references to the indicated page. | |
185 | */ | |
186 | static void remove_migration_ptes(struct page *old, struct page *new) | |
187 | { | |
051ac83a JK |
188 | struct rmap_walk_control rwc = { |
189 | .rmap_one = remove_migration_pte, | |
190 | .arg = old, | |
191 | }; | |
192 | ||
193 | rmap_walk(new, &rwc); | |
04e62a29 CL |
194 | } |
195 | ||
0697212a CL |
196 | /* |
197 | * Something used the pte of a page under migration. We need to | |
198 | * get to the page and wait until migration is finished. | |
199 | * When we return from this function the fault will be retried. | |
0697212a | 200 | */ |
e66f17ff | 201 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 202 | spinlock_t *ptl) |
0697212a | 203 | { |
30dad309 | 204 | pte_t pte; |
0697212a CL |
205 | swp_entry_t entry; |
206 | struct page *page; | |
207 | ||
30dad309 | 208 | spin_lock(ptl); |
0697212a CL |
209 | pte = *ptep; |
210 | if (!is_swap_pte(pte)) | |
211 | goto out; | |
212 | ||
213 | entry = pte_to_swp_entry(pte); | |
214 | if (!is_migration_entry(entry)) | |
215 | goto out; | |
216 | ||
217 | page = migration_entry_to_page(entry); | |
218 | ||
e286781d NP |
219 | /* |
220 | * Once radix-tree replacement of page migration started, page_count | |
221 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
222 | * against a page without get_page(). | |
223 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
224 | * will occur again. | |
225 | */ | |
226 | if (!get_page_unless_zero(page)) | |
227 | goto out; | |
0697212a CL |
228 | pte_unmap_unlock(ptep, ptl); |
229 | wait_on_page_locked(page); | |
230 | put_page(page); | |
231 | return; | |
232 | out: | |
233 | pte_unmap_unlock(ptep, ptl); | |
234 | } | |
235 | ||
30dad309 NH |
236 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
237 | unsigned long address) | |
238 | { | |
239 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
240 | pte_t *ptep = pte_offset_map(pmd, address); | |
241 | __migration_entry_wait(mm, ptep, ptl); | |
242 | } | |
243 | ||
cb900f41 KS |
244 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
245 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 246 | { |
cb900f41 | 247 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
248 | __migration_entry_wait(mm, pte, ptl); |
249 | } | |
250 | ||
b969c4ab MG |
251 | #ifdef CONFIG_BLOCK |
252 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
253 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
254 | enum migrate_mode mode) | |
b969c4ab MG |
255 | { |
256 | struct buffer_head *bh = head; | |
257 | ||
258 | /* Simple case, sync compaction */ | |
a6bc32b8 | 259 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
260 | do { |
261 | get_bh(bh); | |
262 | lock_buffer(bh); | |
263 | bh = bh->b_this_page; | |
264 | ||
265 | } while (bh != head); | |
266 | ||
267 | return true; | |
268 | } | |
269 | ||
270 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
271 | do { | |
272 | get_bh(bh); | |
273 | if (!trylock_buffer(bh)) { | |
274 | /* | |
275 | * We failed to lock the buffer and cannot stall in | |
276 | * async migration. Release the taken locks | |
277 | */ | |
278 | struct buffer_head *failed_bh = bh; | |
279 | put_bh(failed_bh); | |
280 | bh = head; | |
281 | while (bh != failed_bh) { | |
282 | unlock_buffer(bh); | |
283 | put_bh(bh); | |
284 | bh = bh->b_this_page; | |
285 | } | |
286 | return false; | |
287 | } | |
288 | ||
289 | bh = bh->b_this_page; | |
290 | } while (bh != head); | |
291 | return true; | |
292 | } | |
293 | #else | |
294 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 295 | enum migrate_mode mode) |
b969c4ab MG |
296 | { |
297 | return true; | |
298 | } | |
299 | #endif /* CONFIG_BLOCK */ | |
300 | ||
b20a3503 | 301 | /* |
c3fcf8a5 | 302 | * Replace the page in the mapping. |
5b5c7120 CL |
303 | * |
304 | * The number of remaining references must be: | |
305 | * 1 for anonymous pages without a mapping | |
306 | * 2 for pages with a mapping | |
266cf658 | 307 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 308 | */ |
36bc08cc | 309 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 310 | struct page *newpage, struct page *page, |
8e321fef BL |
311 | struct buffer_head *head, enum migrate_mode mode, |
312 | int extra_count) | |
b20a3503 | 313 | { |
8e321fef | 314 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 315 | void **pslot; |
b20a3503 | 316 | |
6c5240ae | 317 | if (!mapping) { |
0e8c7d0f | 318 | /* Anonymous page without mapping */ |
8e321fef | 319 | if (page_count(page) != expected_count) |
6c5240ae | 320 | return -EAGAIN; |
78bd5209 | 321 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
322 | } |
323 | ||
19fd6231 | 324 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 325 | |
7cf9c2c7 NP |
326 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
327 | page_index(page)); | |
b20a3503 | 328 | |
8e321fef | 329 | expected_count += 1 + page_has_private(page); |
e286781d | 330 | if (page_count(page) != expected_count || |
29c1f677 | 331 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 332 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 333 | return -EAGAIN; |
b20a3503 CL |
334 | } |
335 | ||
e286781d | 336 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 337 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
338 | return -EAGAIN; |
339 | } | |
340 | ||
b969c4ab MG |
341 | /* |
342 | * In the async migration case of moving a page with buffers, lock the | |
343 | * buffers using trylock before the mapping is moved. If the mapping | |
344 | * was moved, we later failed to lock the buffers and could not move | |
345 | * the mapping back due to an elevated page count, we would have to | |
346 | * block waiting on other references to be dropped. | |
347 | */ | |
a6bc32b8 MG |
348 | if (mode == MIGRATE_ASYNC && head && |
349 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
350 | page_unfreeze_refs(page, expected_count); |
351 | spin_unlock_irq(&mapping->tree_lock); | |
352 | return -EAGAIN; | |
353 | } | |
354 | ||
b20a3503 CL |
355 | /* |
356 | * Now we know that no one else is looking at the page. | |
b20a3503 | 357 | */ |
7cf9c2c7 | 358 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
359 | if (PageSwapCache(page)) { |
360 | SetPageSwapCache(newpage); | |
361 | set_page_private(newpage, page_private(page)); | |
362 | } | |
363 | ||
7cf9c2c7 NP |
364 | radix_tree_replace_slot(pslot, newpage); |
365 | ||
366 | /* | |
937a94c9 JG |
367 | * Drop cache reference from old page by unfreezing |
368 | * to one less reference. | |
7cf9c2c7 NP |
369 | * We know this isn't the last reference. |
370 | */ | |
937a94c9 | 371 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 372 | |
0e8c7d0f CL |
373 | /* |
374 | * If moved to a different zone then also account | |
375 | * the page for that zone. Other VM counters will be | |
376 | * taken care of when we establish references to the | |
377 | * new page and drop references to the old page. | |
378 | * | |
379 | * Note that anonymous pages are accounted for | |
380 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
381 | * are mapped to swap space. | |
382 | */ | |
383 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
384 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 385 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
386 | __dec_zone_page_state(page, NR_SHMEM); |
387 | __inc_zone_page_state(newpage, NR_SHMEM); | |
388 | } | |
19fd6231 | 389 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 390 | |
78bd5209 | 391 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 392 | } |
b20a3503 | 393 | |
290408d4 NH |
394 | /* |
395 | * The expected number of remaining references is the same as that | |
396 | * of migrate_page_move_mapping(). | |
397 | */ | |
398 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
399 | struct page *newpage, struct page *page) | |
400 | { | |
401 | int expected_count; | |
402 | void **pslot; | |
403 | ||
404 | if (!mapping) { | |
405 | if (page_count(page) != 1) | |
406 | return -EAGAIN; | |
78bd5209 | 407 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
408 | } |
409 | ||
410 | spin_lock_irq(&mapping->tree_lock); | |
411 | ||
412 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
413 | page_index(page)); | |
414 | ||
415 | expected_count = 2 + page_has_private(page); | |
416 | if (page_count(page) != expected_count || | |
29c1f677 | 417 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
418 | spin_unlock_irq(&mapping->tree_lock); |
419 | return -EAGAIN; | |
420 | } | |
421 | ||
422 | if (!page_freeze_refs(page, expected_count)) { | |
423 | spin_unlock_irq(&mapping->tree_lock); | |
424 | return -EAGAIN; | |
425 | } | |
426 | ||
427 | get_page(newpage); | |
428 | ||
429 | radix_tree_replace_slot(pslot, newpage); | |
430 | ||
937a94c9 | 431 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
432 | |
433 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 434 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
435 | } |
436 | ||
30b0a105 DH |
437 | /* |
438 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
439 | * arithmetic will work across the entire page. We need something more | |
440 | * specialized. | |
441 | */ | |
442 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
443 | int nr_pages) | |
444 | { | |
445 | int i; | |
446 | struct page *dst_base = dst; | |
447 | struct page *src_base = src; | |
448 | ||
449 | for (i = 0; i < nr_pages; ) { | |
450 | cond_resched(); | |
451 | copy_highpage(dst, src); | |
452 | ||
453 | i++; | |
454 | dst = mem_map_next(dst, dst_base, i); | |
455 | src = mem_map_next(src, src_base, i); | |
456 | } | |
457 | } | |
458 | ||
459 | static void copy_huge_page(struct page *dst, struct page *src) | |
460 | { | |
461 | int i; | |
462 | int nr_pages; | |
463 | ||
464 | if (PageHuge(src)) { | |
465 | /* hugetlbfs page */ | |
466 | struct hstate *h = page_hstate(src); | |
467 | nr_pages = pages_per_huge_page(h); | |
468 | ||
469 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
470 | __copy_gigantic_page(dst, src, nr_pages); | |
471 | return; | |
472 | } | |
473 | } else { | |
474 | /* thp page */ | |
475 | BUG_ON(!PageTransHuge(src)); | |
476 | nr_pages = hpage_nr_pages(src); | |
477 | } | |
478 | ||
479 | for (i = 0; i < nr_pages; i++) { | |
480 | cond_resched(); | |
481 | copy_highpage(dst + i, src + i); | |
482 | } | |
483 | } | |
484 | ||
b20a3503 CL |
485 | /* |
486 | * Copy the page to its new location | |
487 | */ | |
290408d4 | 488 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 489 | { |
7851a45c RR |
490 | int cpupid; |
491 | ||
b32967ff | 492 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
493 | copy_huge_page(newpage, page); |
494 | else | |
495 | copy_highpage(newpage, page); | |
b20a3503 CL |
496 | |
497 | if (PageError(page)) | |
498 | SetPageError(newpage); | |
499 | if (PageReferenced(page)) | |
500 | SetPageReferenced(newpage); | |
501 | if (PageUptodate(page)) | |
502 | SetPageUptodate(newpage); | |
894bc310 | 503 | if (TestClearPageActive(page)) { |
309381fe | 504 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 505 | SetPageActive(newpage); |
418b27ef LS |
506 | } else if (TestClearPageUnevictable(page)) |
507 | SetPageUnevictable(newpage); | |
b20a3503 CL |
508 | if (PageChecked(page)) |
509 | SetPageChecked(newpage); | |
510 | if (PageMappedToDisk(page)) | |
511 | SetPageMappedToDisk(newpage); | |
512 | ||
513 | if (PageDirty(page)) { | |
514 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
515 | /* |
516 | * Want to mark the page and the radix tree as dirty, and | |
517 | * redo the accounting that clear_page_dirty_for_io undid, | |
518 | * but we can't use set_page_dirty because that function | |
519 | * is actually a signal that all of the page has become dirty. | |
25985edc | 520 | * Whereas only part of our page may be dirty. |
3a902c5f | 521 | */ |
752dc185 HD |
522 | if (PageSwapBacked(page)) |
523 | SetPageDirty(newpage); | |
524 | else | |
525 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
526 | } |
527 | ||
33c3fc71 VD |
528 | if (page_is_young(page)) |
529 | set_page_young(newpage); | |
530 | if (page_is_idle(page)) | |
531 | set_page_idle(newpage); | |
532 | ||
7851a45c RR |
533 | /* |
534 | * Copy NUMA information to the new page, to prevent over-eager | |
535 | * future migrations of this same page. | |
536 | */ | |
537 | cpupid = page_cpupid_xchg_last(page, -1); | |
538 | page_cpupid_xchg_last(newpage, cpupid); | |
539 | ||
b291f000 | 540 | mlock_migrate_page(newpage, page); |
e9995ef9 | 541 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
542 | /* |
543 | * Please do not reorder this without considering how mm/ksm.c's | |
544 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
545 | */ | |
b3b3a99c NH |
546 | if (PageSwapCache(page)) |
547 | ClearPageSwapCache(page); | |
b20a3503 CL |
548 | ClearPagePrivate(page); |
549 | set_page_private(page, 0); | |
b20a3503 CL |
550 | |
551 | /* | |
552 | * If any waiters have accumulated on the new page then | |
553 | * wake them up. | |
554 | */ | |
555 | if (PageWriteback(newpage)) | |
556 | end_page_writeback(newpage); | |
557 | } | |
b20a3503 | 558 | |
1d8b85cc CL |
559 | /************************************************************ |
560 | * Migration functions | |
561 | ***********************************************************/ | |
562 | ||
b20a3503 CL |
563 | /* |
564 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 565 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
566 | * |
567 | * Pages are locked upon entry and exit. | |
568 | */ | |
2d1db3b1 | 569 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
570 | struct page *newpage, struct page *page, |
571 | enum migrate_mode mode) | |
b20a3503 CL |
572 | { |
573 | int rc; | |
574 | ||
575 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
576 | ||
8e321fef | 577 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 578 | |
78bd5209 | 579 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
580 | return rc; |
581 | ||
582 | migrate_page_copy(newpage, page); | |
78bd5209 | 583 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
584 | } |
585 | EXPORT_SYMBOL(migrate_page); | |
586 | ||
9361401e | 587 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
588 | /* |
589 | * Migration function for pages with buffers. This function can only be used | |
590 | * if the underlying filesystem guarantees that no other references to "page" | |
591 | * exist. | |
592 | */ | |
2d1db3b1 | 593 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 594 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 595 | { |
1d8b85cc CL |
596 | struct buffer_head *bh, *head; |
597 | int rc; | |
598 | ||
1d8b85cc | 599 | if (!page_has_buffers(page)) |
a6bc32b8 | 600 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
601 | |
602 | head = page_buffers(page); | |
603 | ||
8e321fef | 604 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 605 | |
78bd5209 | 606 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
607 | return rc; |
608 | ||
b969c4ab MG |
609 | /* |
610 | * In the async case, migrate_page_move_mapping locked the buffers | |
611 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
612 | * need to be locked now | |
613 | */ | |
a6bc32b8 MG |
614 | if (mode != MIGRATE_ASYNC) |
615 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
616 | |
617 | ClearPagePrivate(page); | |
618 | set_page_private(newpage, page_private(page)); | |
619 | set_page_private(page, 0); | |
620 | put_page(page); | |
621 | get_page(newpage); | |
622 | ||
623 | bh = head; | |
624 | do { | |
625 | set_bh_page(bh, newpage, bh_offset(bh)); | |
626 | bh = bh->b_this_page; | |
627 | ||
628 | } while (bh != head); | |
629 | ||
630 | SetPagePrivate(newpage); | |
631 | ||
632 | migrate_page_copy(newpage, page); | |
633 | ||
634 | bh = head; | |
635 | do { | |
636 | unlock_buffer(bh); | |
637 | put_bh(bh); | |
638 | bh = bh->b_this_page; | |
639 | ||
640 | } while (bh != head); | |
641 | ||
78bd5209 | 642 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
643 | } |
644 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 645 | #endif |
1d8b85cc | 646 | |
04e62a29 CL |
647 | /* |
648 | * Writeback a page to clean the dirty state | |
649 | */ | |
650 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 651 | { |
04e62a29 CL |
652 | struct writeback_control wbc = { |
653 | .sync_mode = WB_SYNC_NONE, | |
654 | .nr_to_write = 1, | |
655 | .range_start = 0, | |
656 | .range_end = LLONG_MAX, | |
04e62a29 CL |
657 | .for_reclaim = 1 |
658 | }; | |
659 | int rc; | |
660 | ||
661 | if (!mapping->a_ops->writepage) | |
662 | /* No write method for the address space */ | |
663 | return -EINVAL; | |
664 | ||
665 | if (!clear_page_dirty_for_io(page)) | |
666 | /* Someone else already triggered a write */ | |
667 | return -EAGAIN; | |
668 | ||
8351a6e4 | 669 | /* |
04e62a29 CL |
670 | * A dirty page may imply that the underlying filesystem has |
671 | * the page on some queue. So the page must be clean for | |
672 | * migration. Writeout may mean we loose the lock and the | |
673 | * page state is no longer what we checked for earlier. | |
674 | * At this point we know that the migration attempt cannot | |
675 | * be successful. | |
8351a6e4 | 676 | */ |
04e62a29 | 677 | remove_migration_ptes(page, page); |
8351a6e4 | 678 | |
04e62a29 | 679 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 680 | |
04e62a29 CL |
681 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
682 | /* unlocked. Relock */ | |
683 | lock_page(page); | |
684 | ||
bda8550d | 685 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
686 | } |
687 | ||
688 | /* | |
689 | * Default handling if a filesystem does not provide a migration function. | |
690 | */ | |
691 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 692 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 693 | { |
b969c4ab | 694 | if (PageDirty(page)) { |
a6bc32b8 MG |
695 | /* Only writeback pages in full synchronous migration */ |
696 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 697 | return -EBUSY; |
04e62a29 | 698 | return writeout(mapping, page); |
b969c4ab | 699 | } |
8351a6e4 CL |
700 | |
701 | /* | |
702 | * Buffers may be managed in a filesystem specific way. | |
703 | * We must have no buffers or drop them. | |
704 | */ | |
266cf658 | 705 | if (page_has_private(page) && |
8351a6e4 CL |
706 | !try_to_release_page(page, GFP_KERNEL)) |
707 | return -EAGAIN; | |
708 | ||
a6bc32b8 | 709 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
710 | } |
711 | ||
e24f0b8f CL |
712 | /* |
713 | * Move a page to a newly allocated page | |
714 | * The page is locked and all ptes have been successfully removed. | |
715 | * | |
716 | * The new page will have replaced the old page if this function | |
717 | * is successful. | |
894bc310 LS |
718 | * |
719 | * Return value: | |
720 | * < 0 - error code | |
78bd5209 | 721 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 722 | */ |
3fe2011f | 723 | static int move_to_new_page(struct page *newpage, struct page *page, |
2ebba6b7 | 724 | int page_was_mapped, enum migrate_mode mode) |
e24f0b8f CL |
725 | { |
726 | struct address_space *mapping; | |
727 | int rc; | |
728 | ||
729 | /* | |
730 | * Block others from accessing the page when we get around to | |
731 | * establishing additional references. We are the only one | |
732 | * holding a reference to the new page at this point. | |
733 | */ | |
529ae9aa | 734 | if (!trylock_page(newpage)) |
e24f0b8f CL |
735 | BUG(); |
736 | ||
737 | /* Prepare mapping for the new page.*/ | |
738 | newpage->index = page->index; | |
739 | newpage->mapping = page->mapping; | |
b2e18538 RR |
740 | if (PageSwapBacked(page)) |
741 | SetPageSwapBacked(newpage); | |
e24f0b8f | 742 | |
0610c25d GT |
743 | /* |
744 | * Indirectly called below, migrate_page_copy() copies PG_dirty and thus | |
745 | * needs newpage's memcg set to transfer memcg dirty page accounting. | |
746 | * So perform memcg migration in two steps: | |
747 | * 1. set newpage->mem_cgroup (here) | |
748 | * 2. clear page->mem_cgroup (below) | |
749 | */ | |
750 | set_page_memcg(newpage, page_memcg(page)); | |
751 | ||
e24f0b8f CL |
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) { |
0610c25d | 768 | set_page_memcg(newpage, NULL); |
e24f0b8f | 769 | newpage->mapping = NULL; |
3fe2011f | 770 | } else { |
0610c25d | 771 | set_page_memcg(page, NULL); |
2ebba6b7 | 772 | if (page_was_mapped) |
3fe2011f | 773 | remove_migration_ptes(page, newpage); |
35512eca | 774 | page->mapping = NULL; |
3fe2011f | 775 | } |
e24f0b8f CL |
776 | |
777 | unlock_page(newpage); | |
778 | ||
779 | return rc; | |
780 | } | |
781 | ||
0dabec93 | 782 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 783 | int force, enum migrate_mode mode) |
e24f0b8f | 784 | { |
0dabec93 | 785 | int rc = -EAGAIN; |
2ebba6b7 | 786 | int page_was_mapped = 0; |
3f6c8272 | 787 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 788 | |
529ae9aa | 789 | if (!trylock_page(page)) { |
a6bc32b8 | 790 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 791 | goto out; |
3e7d3449 MG |
792 | |
793 | /* | |
794 | * It's not safe for direct compaction to call lock_page. | |
795 | * For example, during page readahead pages are added locked | |
796 | * to the LRU. Later, when the IO completes the pages are | |
797 | * marked uptodate and unlocked. However, the queueing | |
798 | * could be merging multiple pages for one bio (e.g. | |
799 | * mpage_readpages). If an allocation happens for the | |
800 | * second or third page, the process can end up locking | |
801 | * the same page twice and deadlocking. Rather than | |
802 | * trying to be clever about what pages can be locked, | |
803 | * avoid the use of lock_page for direct compaction | |
804 | * altogether. | |
805 | */ | |
806 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 807 | goto out; |
3e7d3449 | 808 | |
e24f0b8f CL |
809 | lock_page(page); |
810 | } | |
811 | ||
812 | if (PageWriteback(page)) { | |
11bc82d6 | 813 | /* |
fed5b64a | 814 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
815 | * necessary to wait for PageWriteback. In the async case, |
816 | * the retry loop is too short and in the sync-light case, | |
817 | * the overhead of stalling is too much | |
11bc82d6 | 818 | */ |
a6bc32b8 | 819 | if (mode != MIGRATE_SYNC) { |
11bc82d6 | 820 | rc = -EBUSY; |
0a31bc97 | 821 | goto out_unlock; |
11bc82d6 AA |
822 | } |
823 | if (!force) | |
0a31bc97 | 824 | goto out_unlock; |
e24f0b8f CL |
825 | wait_on_page_writeback(page); |
826 | } | |
e24f0b8f | 827 | /* |
dc386d4d KH |
828 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
829 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 830 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 831 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
832 | * File Caches may use write_page() or lock_page() in migration, then, |
833 | * just care Anon page here. | |
dc386d4d | 834 | */ |
b79bc0a0 | 835 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 836 | /* |
4fc3f1d6 | 837 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
838 | * getting a hold on an anon_vma from outside one of its mms. |
839 | */ | |
746b18d4 | 840 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
841 | if (anon_vma) { |
842 | /* | |
746b18d4 | 843 | * Anon page |
1ce82b69 | 844 | */ |
1ce82b69 | 845 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
846 | /* |
847 | * We cannot be sure that the anon_vma of an unmapped | |
848 | * swapcache page is safe to use because we don't | |
849 | * know in advance if the VMA that this page belonged | |
850 | * to still exists. If the VMA and others sharing the | |
851 | * data have been freed, then the anon_vma could | |
852 | * already be invalid. | |
853 | * | |
854 | * To avoid this possibility, swapcache pages get | |
855 | * migrated but are not remapped when migration | |
856 | * completes | |
857 | */ | |
3fe2011f | 858 | } else { |
0a31bc97 | 859 | goto out_unlock; |
3fe2011f | 860 | } |
989f89c5 | 861 | } |
62e1c553 | 862 | |
d6d86c0a | 863 | if (unlikely(isolated_balloon_page(page))) { |
bf6bddf1 RA |
864 | /* |
865 | * A ballooned page does not need any special attention from | |
866 | * physical to virtual reverse mapping procedures. | |
867 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
868 | * in order to avoid burning cycles at rmap level, and perform | |
869 | * the page migration right away (proteced by page lock). | |
870 | */ | |
871 | rc = balloon_page_migrate(newpage, page, mode); | |
0a31bc97 | 872 | goto out_unlock; |
bf6bddf1 RA |
873 | } |
874 | ||
dc386d4d | 875 | /* |
62e1c553 SL |
876 | * Corner case handling: |
877 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
878 | * and treated as swapcache but it has no rmap yet. | |
879 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
880 | * trigger a BUG. So handle it here. | |
881 | * 2. An orphaned page (see truncate_complete_page) might have | |
882 | * fs-private metadata. The page can be picked up due to memory | |
883 | * offlining. Everywhere else except page reclaim, the page is | |
884 | * invisible to the vm, so the page can not be migrated. So try to | |
885 | * free the metadata, so the page can be freed. | |
e24f0b8f | 886 | */ |
62e1c553 | 887 | if (!page->mapping) { |
309381fe | 888 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 889 | if (page_has_private(page)) { |
62e1c553 | 890 | try_to_free_buffers(page); |
0a31bc97 | 891 | goto out_unlock; |
62e1c553 | 892 | } |
abfc3488 | 893 | goto skip_unmap; |
62e1c553 SL |
894 | } |
895 | ||
dc386d4d | 896 | /* Establish migration ptes or remove ptes */ |
2ebba6b7 HD |
897 | if (page_mapped(page)) { |
898 | try_to_unmap(page, | |
da1b13cc | 899 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
900 | page_was_mapped = 1; |
901 | } | |
dc386d4d | 902 | |
abfc3488 | 903 | skip_unmap: |
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 MG |
909 | |
910 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 911 | if (anon_vma) |
9e60109f | 912 | put_anon_vma(anon_vma); |
3f6c8272 | 913 | |
0a31bc97 | 914 | out_unlock: |
e24f0b8f | 915 | unlock_page(page); |
0dabec93 MK |
916 | out: |
917 | return rc; | |
918 | } | |
95a402c3 | 919 | |
ef2a5153 GU |
920 | /* |
921 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
922 | * around it. | |
923 | */ | |
924 | #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) | |
925 | #define ICE_noinline noinline | |
926 | #else | |
927 | #define ICE_noinline | |
928 | #endif | |
929 | ||
0dabec93 MK |
930 | /* |
931 | * Obtain the lock on page, remove all ptes and migrate the page | |
932 | * to the newly allocated page in newpage. | |
933 | */ | |
ef2a5153 GU |
934 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
935 | free_page_t put_new_page, | |
936 | unsigned long private, struct page *page, | |
add05cec NH |
937 | int force, enum migrate_mode mode, |
938 | enum migrate_reason reason) | |
0dabec93 MK |
939 | { |
940 | int rc = 0; | |
941 | int *result = NULL; | |
942 | struct page *newpage = get_new_page(page, private, &result); | |
943 | ||
944 | if (!newpage) | |
945 | return -ENOMEM; | |
946 | ||
947 | if (page_count(page) == 1) { | |
948 | /* page was freed from under us. So we are done. */ | |
949 | goto out; | |
950 | } | |
951 | ||
952 | if (unlikely(PageTransHuge(page))) | |
953 | if (unlikely(split_huge_page(page))) | |
954 | goto out; | |
955 | ||
9c620e2b | 956 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 | 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)); |
f4c18e6f | 969 | /* Soft-offlined page shouldn't go through lru cache list */ |
da1b13cc | 970 | if (reason == MR_MEMORY_FAILURE) { |
f4c18e6f | 971 | put_page(page); |
da1b13cc WL |
972 | if (!test_set_page_hwpoison(page)) |
973 | num_poisoned_pages_inc(); | |
974 | } else | |
add05cec | 975 | putback_lru_page(page); |
e24f0b8f | 976 | } |
68711a74 | 977 | |
95a402c3 | 978 | /* |
68711a74 DR |
979 | * If migration was not successful and there's a freeing callback, use |
980 | * it. Otherwise, putback_lru_page() will drop the reference grabbed | |
981 | * during isolation. | |
95a402c3 | 982 | */ |
8bdd6380 HD |
983 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) { |
984 | ClearPageSwapBacked(newpage); | |
68711a74 | 985 | put_new_page(newpage, private); |
d6d86c0a KK |
986 | } else if (unlikely(__is_movable_balloon_page(newpage))) { |
987 | /* drop our reference, page already in the balloon */ | |
988 | put_page(newpage); | |
8bdd6380 | 989 | } else |
68711a74 DR |
990 | putback_lru_page(newpage); |
991 | ||
742755a1 CL |
992 | if (result) { |
993 | if (rc) | |
994 | *result = rc; | |
995 | else | |
996 | *result = page_to_nid(newpage); | |
997 | } | |
e24f0b8f CL |
998 | return rc; |
999 | } | |
1000 | ||
290408d4 NH |
1001 | /* |
1002 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1003 | * | |
1004 | * This function doesn't wait the completion of hugepage I/O | |
1005 | * because there is no race between I/O and migration for hugepage. | |
1006 | * Note that currently hugepage I/O occurs only in direct I/O | |
1007 | * where no lock is held and PG_writeback is irrelevant, | |
1008 | * and writeback status of all subpages are counted in the reference | |
1009 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1010 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1011 | * This means that when we try to migrate hugepage whose subpages are | |
1012 | * doing direct I/O, some references remain after try_to_unmap() and | |
1013 | * hugepage migration fails without data corruption. | |
1014 | * | |
1015 | * There is also no race when direct I/O is issued on the page under migration, | |
1016 | * because then pte is replaced with migration swap entry and direct I/O code | |
1017 | * will wait in the page fault for migration to complete. | |
1018 | */ | |
1019 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1020 | free_page_t put_new_page, unsigned long private, |
1021 | struct page *hpage, int force, | |
1022 | enum migrate_mode mode) | |
290408d4 NH |
1023 | { |
1024 | int rc = 0; | |
1025 | int *result = NULL; | |
2ebba6b7 | 1026 | int page_was_mapped = 0; |
32665f2b | 1027 | struct page *new_hpage; |
290408d4 NH |
1028 | struct anon_vma *anon_vma = NULL; |
1029 | ||
83467efb NH |
1030 | /* |
1031 | * Movability of hugepages depends on architectures and hugepage size. | |
1032 | * This check is necessary because some callers of hugepage migration | |
1033 | * like soft offline and memory hotremove don't walk through page | |
1034 | * tables or check whether the hugepage is pmd-based or not before | |
1035 | * kicking migration. | |
1036 | */ | |
100873d7 | 1037 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1038 | putback_active_hugepage(hpage); |
83467efb | 1039 | return -ENOSYS; |
32665f2b | 1040 | } |
83467efb | 1041 | |
32665f2b | 1042 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1043 | if (!new_hpage) |
1044 | return -ENOMEM; | |
1045 | ||
1046 | rc = -EAGAIN; | |
1047 | ||
1048 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 1049 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1050 | goto out; |
1051 | lock_page(hpage); | |
1052 | } | |
1053 | ||
746b18d4 PZ |
1054 | if (PageAnon(hpage)) |
1055 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1056 | |
2ebba6b7 HD |
1057 | if (page_mapped(hpage)) { |
1058 | try_to_unmap(hpage, | |
1059 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1060 | page_was_mapped = 1; | |
1061 | } | |
290408d4 NH |
1062 | |
1063 | if (!page_mapped(hpage)) | |
2ebba6b7 | 1064 | rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode); |
290408d4 | 1065 | |
2ebba6b7 | 1066 | if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped) |
290408d4 NH |
1067 | remove_migration_ptes(hpage, hpage); |
1068 | ||
fd4a4663 | 1069 | if (anon_vma) |
9e60109f | 1070 | put_anon_vma(anon_vma); |
8e6ac7fa | 1071 | |
68711a74 | 1072 | if (rc == MIGRATEPAGE_SUCCESS) |
8e6ac7fa AK |
1073 | hugetlb_cgroup_migrate(hpage, new_hpage); |
1074 | ||
290408d4 | 1075 | unlock_page(hpage); |
09761333 | 1076 | out: |
b8ec1cee NH |
1077 | if (rc != -EAGAIN) |
1078 | putback_active_hugepage(hpage); | |
68711a74 DR |
1079 | |
1080 | /* | |
1081 | * If migration was not successful and there's a freeing callback, use | |
1082 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1083 | * isolation. | |
1084 | */ | |
1085 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) | |
1086 | put_new_page(new_hpage, private); | |
1087 | else | |
3aaa76e1 | 1088 | putback_active_hugepage(new_hpage); |
68711a74 | 1089 | |
290408d4 NH |
1090 | if (result) { |
1091 | if (rc) | |
1092 | *result = rc; | |
1093 | else | |
1094 | *result = page_to_nid(new_hpage); | |
1095 | } | |
1096 | return rc; | |
1097 | } | |
1098 | ||
b20a3503 | 1099 | /* |
c73e5c9c SB |
1100 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1101 | * supplied as the target for the page migration | |
b20a3503 | 1102 | * |
c73e5c9c SB |
1103 | * @from: The list of pages to be migrated. |
1104 | * @get_new_page: The function used to allocate free pages to be used | |
1105 | * as the target of the page migration. | |
68711a74 DR |
1106 | * @put_new_page: The function used to free target pages if migration |
1107 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1108 | * @private: Private data to be passed on to get_new_page() |
1109 | * @mode: The migration mode that specifies the constraints for | |
1110 | * page migration, if any. | |
1111 | * @reason: The reason for page migration. | |
b20a3503 | 1112 | * |
c73e5c9c SB |
1113 | * The function returns after 10 attempts or if no pages are movable any more |
1114 | * because the list has become empty or no retryable pages exist any more. | |
1115 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1116 | * or free list only if ret != 0. |
b20a3503 | 1117 | * |
c73e5c9c | 1118 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1119 | */ |
9c620e2b | 1120 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1121 | free_page_t put_new_page, unsigned long private, |
1122 | enum migrate_mode mode, int reason) | |
b20a3503 | 1123 | { |
e24f0b8f | 1124 | int retry = 1; |
b20a3503 | 1125 | int nr_failed = 0; |
5647bc29 | 1126 | int nr_succeeded = 0; |
b20a3503 CL |
1127 | int pass = 0; |
1128 | struct page *page; | |
1129 | struct page *page2; | |
1130 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1131 | int rc; | |
1132 | ||
1133 | if (!swapwrite) | |
1134 | current->flags |= PF_SWAPWRITE; | |
1135 | ||
e24f0b8f CL |
1136 | for(pass = 0; pass < 10 && retry; pass++) { |
1137 | retry = 0; | |
b20a3503 | 1138 | |
e24f0b8f | 1139 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1140 | cond_resched(); |
2d1db3b1 | 1141 | |
31caf665 NH |
1142 | if (PageHuge(page)) |
1143 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 DR |
1144 | put_new_page, private, page, |
1145 | pass > 2, mode); | |
31caf665 | 1146 | else |
68711a74 | 1147 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1148 | private, page, pass > 2, mode, |
1149 | reason); | |
2d1db3b1 | 1150 | |
e24f0b8f | 1151 | switch(rc) { |
95a402c3 CL |
1152 | case -ENOMEM: |
1153 | goto out; | |
e24f0b8f | 1154 | case -EAGAIN: |
2d1db3b1 | 1155 | retry++; |
e24f0b8f | 1156 | break; |
78bd5209 | 1157 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1158 | nr_succeeded++; |
e24f0b8f CL |
1159 | break; |
1160 | default: | |
354a3363 NH |
1161 | /* |
1162 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1163 | * unlike -EAGAIN case, the failed page is | |
1164 | * removed from migration page list and not | |
1165 | * retried in the next outer loop. | |
1166 | */ | |
2d1db3b1 | 1167 | nr_failed++; |
e24f0b8f | 1168 | break; |
2d1db3b1 | 1169 | } |
b20a3503 CL |
1170 | } |
1171 | } | |
78bd5209 | 1172 | rc = nr_failed + retry; |
95a402c3 | 1173 | out: |
5647bc29 MG |
1174 | if (nr_succeeded) |
1175 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1176 | if (nr_failed) | |
1177 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1178 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1179 | ||
b20a3503 CL |
1180 | if (!swapwrite) |
1181 | current->flags &= ~PF_SWAPWRITE; | |
1182 | ||
78bd5209 | 1183 | return rc; |
b20a3503 | 1184 | } |
95a402c3 | 1185 | |
742755a1 CL |
1186 | #ifdef CONFIG_NUMA |
1187 | /* | |
1188 | * Move a list of individual pages | |
1189 | */ | |
1190 | struct page_to_node { | |
1191 | unsigned long addr; | |
1192 | struct page *page; | |
1193 | int node; | |
1194 | int status; | |
1195 | }; | |
1196 | ||
1197 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1198 | int **result) | |
1199 | { | |
1200 | struct page_to_node *pm = (struct page_to_node *)private; | |
1201 | ||
1202 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1203 | pm++; | |
1204 | ||
1205 | if (pm->node == MAX_NUMNODES) | |
1206 | return NULL; | |
1207 | ||
1208 | *result = &pm->status; | |
1209 | ||
e632a938 NH |
1210 | if (PageHuge(p)) |
1211 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1212 | pm->node); | |
1213 | else | |
96db800f | 1214 | return __alloc_pages_node(pm->node, |
e97ca8e5 | 1215 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); |
742755a1 CL |
1216 | } |
1217 | ||
1218 | /* | |
1219 | * Move a set of pages as indicated in the pm array. The addr | |
1220 | * field must be set to the virtual address of the page to be moved | |
1221 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1222 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1223 | */ |
5e9a0f02 BG |
1224 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1225 | struct page_to_node *pm, | |
1226 | int migrate_all) | |
742755a1 CL |
1227 | { |
1228 | int err; | |
1229 | struct page_to_node *pp; | |
1230 | LIST_HEAD(pagelist); | |
1231 | ||
1232 | down_read(&mm->mmap_sem); | |
1233 | ||
1234 | /* | |
1235 | * Build a list of pages to migrate | |
1236 | */ | |
742755a1 CL |
1237 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1238 | struct vm_area_struct *vma; | |
1239 | struct page *page; | |
1240 | ||
742755a1 CL |
1241 | err = -EFAULT; |
1242 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1243 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1244 | goto set_status; |
1245 | ||
d899844e KS |
1246 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1247 | page = follow_page(vma, pp->addr, | |
1248 | FOLL_GET | FOLL_SPLIT | FOLL_DUMP); | |
89f5b7da LT |
1249 | |
1250 | err = PTR_ERR(page); | |
1251 | if (IS_ERR(page)) | |
1252 | goto set_status; | |
1253 | ||
742755a1 CL |
1254 | err = -ENOENT; |
1255 | if (!page) | |
1256 | goto set_status; | |
1257 | ||
742755a1 CL |
1258 | pp->page = page; |
1259 | err = page_to_nid(page); | |
1260 | ||
1261 | if (err == pp->node) | |
1262 | /* | |
1263 | * Node already in the right place | |
1264 | */ | |
1265 | goto put_and_set; | |
1266 | ||
1267 | err = -EACCES; | |
1268 | if (page_mapcount(page) > 1 && | |
1269 | !migrate_all) | |
1270 | goto put_and_set; | |
1271 | ||
e632a938 | 1272 | if (PageHuge(page)) { |
e66f17ff NH |
1273 | if (PageHead(page)) |
1274 | isolate_huge_page(page, &pagelist); | |
e632a938 NH |
1275 | goto put_and_set; |
1276 | } | |
1277 | ||
62695a84 | 1278 | err = isolate_lru_page(page); |
6d9c285a | 1279 | if (!err) { |
62695a84 | 1280 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1281 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1282 | page_is_file_cache(page)); | |
1283 | } | |
742755a1 CL |
1284 | put_and_set: |
1285 | /* | |
1286 | * Either remove the duplicate refcount from | |
1287 | * isolate_lru_page() or drop the page ref if it was | |
1288 | * not isolated. | |
1289 | */ | |
1290 | put_page(page); | |
1291 | set_status: | |
1292 | pp->status = err; | |
1293 | } | |
1294 | ||
e78bbfa8 | 1295 | err = 0; |
cf608ac1 | 1296 | if (!list_empty(&pagelist)) { |
68711a74 | 1297 | err = migrate_pages(&pagelist, new_page_node, NULL, |
9c620e2b | 1298 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1299 | if (err) |
e632a938 | 1300 | putback_movable_pages(&pagelist); |
cf608ac1 | 1301 | } |
742755a1 CL |
1302 | |
1303 | up_read(&mm->mmap_sem); | |
1304 | return err; | |
1305 | } | |
1306 | ||
5e9a0f02 BG |
1307 | /* |
1308 | * Migrate an array of page address onto an array of nodes and fill | |
1309 | * the corresponding array of status. | |
1310 | */ | |
3268c63e | 1311 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1312 | unsigned long nr_pages, |
1313 | const void __user * __user *pages, | |
1314 | const int __user *nodes, | |
1315 | int __user *status, int flags) | |
1316 | { | |
3140a227 | 1317 | struct page_to_node *pm; |
3140a227 BG |
1318 | unsigned long chunk_nr_pages; |
1319 | unsigned long chunk_start; | |
1320 | int err; | |
5e9a0f02 | 1321 | |
3140a227 BG |
1322 | err = -ENOMEM; |
1323 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1324 | if (!pm) | |
5e9a0f02 | 1325 | goto out; |
35282a2d BG |
1326 | |
1327 | migrate_prep(); | |
1328 | ||
5e9a0f02 | 1329 | /* |
3140a227 BG |
1330 | * Store a chunk of page_to_node array in a page, |
1331 | * but keep the last one as a marker | |
5e9a0f02 | 1332 | */ |
3140a227 | 1333 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1334 | |
3140a227 BG |
1335 | for (chunk_start = 0; |
1336 | chunk_start < nr_pages; | |
1337 | chunk_start += chunk_nr_pages) { | |
1338 | int j; | |
5e9a0f02 | 1339 | |
3140a227 BG |
1340 | if (chunk_start + chunk_nr_pages > nr_pages) |
1341 | chunk_nr_pages = nr_pages - chunk_start; | |
1342 | ||
1343 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1344 | for (j = 0; j < chunk_nr_pages; j++) { | |
1345 | const void __user *p; | |
5e9a0f02 BG |
1346 | int node; |
1347 | ||
3140a227 BG |
1348 | err = -EFAULT; |
1349 | if (get_user(p, pages + j + chunk_start)) | |
1350 | goto out_pm; | |
1351 | pm[j].addr = (unsigned long) p; | |
1352 | ||
1353 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1354 | goto out_pm; |
1355 | ||
1356 | err = -ENODEV; | |
6f5a55f1 LT |
1357 | if (node < 0 || node >= MAX_NUMNODES) |
1358 | goto out_pm; | |
1359 | ||
389162c2 | 1360 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1361 | goto out_pm; |
1362 | ||
1363 | err = -EACCES; | |
1364 | if (!node_isset(node, task_nodes)) | |
1365 | goto out_pm; | |
1366 | ||
3140a227 BG |
1367 | pm[j].node = node; |
1368 | } | |
1369 | ||
1370 | /* End marker for this chunk */ | |
1371 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1372 | ||
1373 | /* Migrate this chunk */ | |
1374 | err = do_move_page_to_node_array(mm, pm, | |
1375 | flags & MPOL_MF_MOVE_ALL); | |
1376 | if (err < 0) | |
1377 | goto out_pm; | |
5e9a0f02 | 1378 | |
5e9a0f02 | 1379 | /* Return status information */ |
3140a227 BG |
1380 | for (j = 0; j < chunk_nr_pages; j++) |
1381 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1382 | err = -EFAULT; |
3140a227 BG |
1383 | goto out_pm; |
1384 | } | |
1385 | } | |
1386 | err = 0; | |
5e9a0f02 BG |
1387 | |
1388 | out_pm: | |
3140a227 | 1389 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1390 | out: |
1391 | return err; | |
1392 | } | |
1393 | ||
742755a1 | 1394 | /* |
2f007e74 | 1395 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1396 | */ |
80bba129 BG |
1397 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1398 | const void __user **pages, int *status) | |
742755a1 | 1399 | { |
2f007e74 | 1400 | unsigned long i; |
2f007e74 | 1401 | |
742755a1 CL |
1402 | down_read(&mm->mmap_sem); |
1403 | ||
2f007e74 | 1404 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1405 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1406 | struct vm_area_struct *vma; |
1407 | struct page *page; | |
c095adbc | 1408 | int err = -EFAULT; |
2f007e74 BG |
1409 | |
1410 | vma = find_vma(mm, addr); | |
70384dc6 | 1411 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1412 | goto set_status; |
1413 | ||
d899844e KS |
1414 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1415 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1416 | |
1417 | err = PTR_ERR(page); | |
1418 | if (IS_ERR(page)) | |
1419 | goto set_status; | |
1420 | ||
d899844e | 1421 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1422 | set_status: |
80bba129 BG |
1423 | *status = err; |
1424 | ||
1425 | pages++; | |
1426 | status++; | |
1427 | } | |
1428 | ||
1429 | up_read(&mm->mmap_sem); | |
1430 | } | |
1431 | ||
1432 | /* | |
1433 | * Determine the nodes of a user array of pages and store it in | |
1434 | * a user array of status. | |
1435 | */ | |
1436 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1437 | const void __user * __user *pages, | |
1438 | int __user *status) | |
1439 | { | |
1440 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1441 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1442 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1443 | |
87b8d1ad PA |
1444 | while (nr_pages) { |
1445 | unsigned long chunk_nr; | |
80bba129 | 1446 | |
87b8d1ad PA |
1447 | chunk_nr = nr_pages; |
1448 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1449 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1450 | ||
1451 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1452 | break; | |
80bba129 BG |
1453 | |
1454 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1455 | ||
87b8d1ad PA |
1456 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1457 | break; | |
742755a1 | 1458 | |
87b8d1ad PA |
1459 | pages += chunk_nr; |
1460 | status += chunk_nr; | |
1461 | nr_pages -= chunk_nr; | |
1462 | } | |
1463 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1464 | } |
1465 | ||
1466 | /* | |
1467 | * Move a list of pages in the address space of the currently executing | |
1468 | * process. | |
1469 | */ | |
938bb9f5 HC |
1470 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1471 | const void __user * __user *, pages, | |
1472 | const int __user *, nodes, | |
1473 | int __user *, status, int, flags) | |
742755a1 | 1474 | { |
c69e8d9c | 1475 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1476 | struct task_struct *task; |
742755a1 | 1477 | struct mm_struct *mm; |
5e9a0f02 | 1478 | int err; |
3268c63e | 1479 | nodemask_t task_nodes; |
742755a1 CL |
1480 | |
1481 | /* Check flags */ | |
1482 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1483 | return -EINVAL; | |
1484 | ||
1485 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1486 | return -EPERM; | |
1487 | ||
1488 | /* Find the mm_struct */ | |
a879bf58 | 1489 | rcu_read_lock(); |
228ebcbe | 1490 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1491 | if (!task) { |
a879bf58 | 1492 | rcu_read_unlock(); |
742755a1 CL |
1493 | return -ESRCH; |
1494 | } | |
3268c63e | 1495 | get_task_struct(task); |
742755a1 CL |
1496 | |
1497 | /* | |
1498 | * Check if this process has the right to modify the specified | |
1499 | * process. The right exists if the process has administrative | |
1500 | * capabilities, superuser privileges or the same | |
1501 | * userid as the target process. | |
1502 | */ | |
c69e8d9c | 1503 | tcred = __task_cred(task); |
b38a86eb EB |
1504 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1505 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1506 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1507 | rcu_read_unlock(); |
742755a1 | 1508 | err = -EPERM; |
5e9a0f02 | 1509 | goto out; |
742755a1 | 1510 | } |
c69e8d9c | 1511 | rcu_read_unlock(); |
742755a1 | 1512 | |
86c3a764 DQ |
1513 | err = security_task_movememory(task); |
1514 | if (err) | |
5e9a0f02 | 1515 | goto out; |
86c3a764 | 1516 | |
3268c63e CL |
1517 | task_nodes = cpuset_mems_allowed(task); |
1518 | mm = get_task_mm(task); | |
1519 | put_task_struct(task); | |
1520 | ||
6e8b09ea SL |
1521 | if (!mm) |
1522 | return -EINVAL; | |
1523 | ||
1524 | if (nodes) | |
1525 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1526 | nodes, status, flags); | |
1527 | else | |
1528 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1529 | |
742755a1 CL |
1530 | mmput(mm); |
1531 | return err; | |
3268c63e CL |
1532 | |
1533 | out: | |
1534 | put_task_struct(task); | |
1535 | return err; | |
742755a1 | 1536 | } |
742755a1 | 1537 | |
7039e1db PZ |
1538 | #ifdef CONFIG_NUMA_BALANCING |
1539 | /* | |
1540 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1541 | * pages. Currently it only checks the watermarks which crude | |
1542 | */ | |
1543 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1544 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1545 | { |
1546 | int z; | |
1547 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1548 | struct zone *zone = pgdat->node_zones + z; | |
1549 | ||
1550 | if (!populated_zone(zone)) | |
1551 | continue; | |
1552 | ||
6e543d57 | 1553 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1554 | continue; |
1555 | ||
1556 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1557 | if (!zone_watermark_ok(zone, 0, | |
1558 | high_wmark_pages(zone) + | |
1559 | nr_migrate_pages, | |
1560 | 0, 0)) | |
1561 | continue; | |
1562 | return true; | |
1563 | } | |
1564 | return false; | |
1565 | } | |
1566 | ||
1567 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1568 | unsigned long data, | |
1569 | int **result) | |
1570 | { | |
1571 | int nid = (int) data; | |
1572 | struct page *newpage; | |
1573 | ||
96db800f | 1574 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1575 | (GFP_HIGHUSER_MOVABLE | |
1576 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1577 | __GFP_NORETRY | __GFP_NOWARN) & | |
7039e1db | 1578 | ~GFP_IOFS, 0); |
bac0382c | 1579 | |
7039e1db PZ |
1580 | return newpage; |
1581 | } | |
1582 | ||
a8f60772 MG |
1583 | /* |
1584 | * page migration rate limiting control. | |
1585 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1586 | * window of time. Default here says do not migrate more than 1280M per second. | |
1587 | */ | |
1588 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1589 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1590 | ||
b32967ff | 1591 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1592 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1593 | unsigned long nr_pages) | |
7039e1db | 1594 | { |
a8f60772 MG |
1595 | /* |
1596 | * Rate-limit the amount of data that is being migrated to a node. | |
1597 | * Optimal placement is no good if the memory bus is saturated and | |
1598 | * all the time is being spent migrating! | |
1599 | */ | |
a8f60772 | 1600 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1601 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1602 | pgdat->numabalancing_migrate_nr_pages = 0; |
1603 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1604 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1605 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1606 | } |
af1839d7 MG |
1607 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1608 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1609 | nr_pages); | |
1c5e9c27 | 1610 | return true; |
af1839d7 | 1611 | } |
1c5e9c27 MG |
1612 | |
1613 | /* | |
1614 | * This is an unlocked non-atomic update so errors are possible. | |
1615 | * The consequences are failing to migrate when we potentiall should | |
1616 | * have which is not severe enough to warrant locking. If it is ever | |
1617 | * a problem, it can be converted to a per-cpu counter. | |
1618 | */ | |
1619 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1620 | return false; | |
b32967ff MG |
1621 | } |
1622 | ||
1c30e017 | 1623 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1624 | { |
340ef390 | 1625 | int page_lru; |
a8f60772 | 1626 | |
309381fe | 1627 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1628 | |
7039e1db | 1629 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1630 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1631 | return 0; | |
7039e1db | 1632 | |
340ef390 HD |
1633 | if (isolate_lru_page(page)) |
1634 | return 0; | |
7039e1db | 1635 | |
340ef390 HD |
1636 | /* |
1637 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1638 | * check on page_count(), so we must do it here, now that the page | |
1639 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1640 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1641 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1642 | */ | |
1643 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1644 | putback_lru_page(page); | |
1645 | return 0; | |
7039e1db PZ |
1646 | } |
1647 | ||
340ef390 HD |
1648 | page_lru = page_is_file_cache(page); |
1649 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1650 | hpage_nr_pages(page)); | |
1651 | ||
149c33e1 | 1652 | /* |
340ef390 HD |
1653 | * Isolating the page has taken another reference, so the |
1654 | * caller's reference can be safely dropped without the page | |
1655 | * disappearing underneath us during migration. | |
149c33e1 MG |
1656 | */ |
1657 | put_page(page); | |
340ef390 | 1658 | return 1; |
b32967ff MG |
1659 | } |
1660 | ||
de466bd6 MG |
1661 | bool pmd_trans_migrating(pmd_t pmd) |
1662 | { | |
1663 | struct page *page = pmd_page(pmd); | |
1664 | return PageLocked(page); | |
1665 | } | |
1666 | ||
b32967ff MG |
1667 | /* |
1668 | * Attempt to migrate a misplaced page to the specified destination | |
1669 | * node. Caller is expected to have an elevated reference count on | |
1670 | * the page that will be dropped by this function before returning. | |
1671 | */ | |
1bc115d8 MG |
1672 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1673 | int node) | |
b32967ff MG |
1674 | { |
1675 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1676 | int isolated; |
b32967ff MG |
1677 | int nr_remaining; |
1678 | LIST_HEAD(migratepages); | |
1679 | ||
1680 | /* | |
1bc115d8 MG |
1681 | * Don't migrate file pages that are mapped in multiple processes |
1682 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1683 | */ |
1bc115d8 MG |
1684 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1685 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1686 | goto out; |
b32967ff MG |
1687 | |
1688 | /* | |
1689 | * Rate-limit the amount of data that is being migrated to a node. | |
1690 | * Optimal placement is no good if the memory bus is saturated and | |
1691 | * all the time is being spent migrating! | |
1692 | */ | |
340ef390 | 1693 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1694 | goto out; |
b32967ff MG |
1695 | |
1696 | isolated = numamigrate_isolate_page(pgdat, page); | |
1697 | if (!isolated) | |
1698 | goto out; | |
1699 | ||
1700 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1701 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1702 | NULL, node, MIGRATE_ASYNC, |
1703 | MR_NUMA_MISPLACED); | |
b32967ff | 1704 | if (nr_remaining) { |
59c82b70 JK |
1705 | if (!list_empty(&migratepages)) { |
1706 | list_del(&page->lru); | |
1707 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1708 | page_is_file_cache(page)); | |
1709 | putback_lru_page(page); | |
1710 | } | |
b32967ff MG |
1711 | isolated = 0; |
1712 | } else | |
1713 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1714 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1715 | return isolated; |
340ef390 HD |
1716 | |
1717 | out: | |
1718 | put_page(page); | |
1719 | return 0; | |
7039e1db | 1720 | } |
220018d3 | 1721 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1722 | |
220018d3 | 1723 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1724 | /* |
1725 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1726 | * before returning. | |
1727 | */ | |
b32967ff MG |
1728 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1729 | struct vm_area_struct *vma, | |
1730 | pmd_t *pmd, pmd_t entry, | |
1731 | unsigned long address, | |
1732 | struct page *page, int node) | |
1733 | { | |
c4088ebd | 1734 | spinlock_t *ptl; |
b32967ff MG |
1735 | pg_data_t *pgdat = NODE_DATA(node); |
1736 | int isolated = 0; | |
1737 | struct page *new_page = NULL; | |
b32967ff | 1738 | int page_lru = page_is_file_cache(page); |
f714f4f2 MG |
1739 | unsigned long mmun_start = address & HPAGE_PMD_MASK; |
1740 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
2b4847e7 | 1741 | pmd_t orig_entry; |
b32967ff | 1742 | |
b32967ff MG |
1743 | /* |
1744 | * Rate-limit the amount of data that is being migrated to a node. | |
1745 | * Optimal placement is no good if the memory bus is saturated and | |
1746 | * all the time is being spent migrating! | |
1747 | */ | |
d28d4335 | 1748 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1749 | goto out_dropref; |
1750 | ||
1751 | new_page = alloc_pages_node(node, | |
e97ca8e5 JW |
1752 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT, |
1753 | HPAGE_PMD_ORDER); | |
340ef390 HD |
1754 | if (!new_page) |
1755 | goto out_fail; | |
1756 | ||
b32967ff | 1757 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 1758 | if (!isolated) { |
b32967ff | 1759 | put_page(new_page); |
340ef390 | 1760 | goto out_fail; |
b32967ff MG |
1761 | } |
1762 | ||
b0943d61 MG |
1763 | if (mm_tlb_flush_pending(mm)) |
1764 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1765 | ||
b32967ff MG |
1766 | /* Prepare a page as a migration target */ |
1767 | __set_page_locked(new_page); | |
1768 | SetPageSwapBacked(new_page); | |
1769 | ||
1770 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1771 | new_page->mapping = page->mapping; | |
1772 | new_page->index = page->index; | |
1773 | migrate_page_copy(new_page, page); | |
1774 | WARN_ON(PageLRU(new_page)); | |
1775 | ||
1776 | /* Recheck the target PMD */ | |
f714f4f2 | 1777 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 1778 | ptl = pmd_lock(mm, pmd); |
2b4847e7 MG |
1779 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { |
1780 | fail_putback: | |
c4088ebd | 1781 | spin_unlock(ptl); |
f714f4f2 | 1782 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1783 | |
1784 | /* Reverse changes made by migrate_page_copy() */ | |
1785 | if (TestClearPageActive(new_page)) | |
1786 | SetPageActive(page); | |
1787 | if (TestClearPageUnevictable(new_page)) | |
1788 | SetPageUnevictable(page); | |
1789 | mlock_migrate_page(page, new_page); | |
1790 | ||
1791 | unlock_page(new_page); | |
1792 | put_page(new_page); /* Free it */ | |
1793 | ||
a54a407f MG |
1794 | /* Retake the callers reference and putback on LRU */ |
1795 | get_page(page); | |
b32967ff | 1796 | putback_lru_page(page); |
a54a407f MG |
1797 | mod_zone_page_state(page_zone(page), |
1798 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
eb4489f6 MG |
1799 | |
1800 | goto out_unlock; | |
b32967ff MG |
1801 | } |
1802 | ||
2b4847e7 | 1803 | orig_entry = *pmd; |
b32967ff | 1804 | entry = mk_pmd(new_page, vma->vm_page_prot); |
b32967ff | 1805 | entry = pmd_mkhuge(entry); |
2b4847e7 | 1806 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 1807 | |
2b4847e7 MG |
1808 | /* |
1809 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1810 | * Any parallel GUP will either observe the old page blocking on the | |
1811 | * page lock, block on the page table lock or observe the new page. | |
1812 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1813 | * guarantee the copy is visible before the pagetable update. | |
1814 | */ | |
f714f4f2 | 1815 | flush_cache_range(vma, mmun_start, mmun_end); |
11de9927 | 1816 | page_add_anon_rmap(new_page, vma, mmun_start); |
8809aa2d | 1817 | pmdp_huge_clear_flush_notify(vma, mmun_start, pmd); |
f714f4f2 MG |
1818 | set_pmd_at(mm, mmun_start, pmd, entry); |
1819 | flush_tlb_range(vma, mmun_start, mmun_end); | |
ce4a9cc5 | 1820 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 MG |
1821 | |
1822 | if (page_count(page) != 2) { | |
f714f4f2 MG |
1823 | set_pmd_at(mm, mmun_start, pmd, orig_entry); |
1824 | flush_tlb_range(vma, mmun_start, mmun_end); | |
34ee645e | 1825 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
2b4847e7 MG |
1826 | update_mmu_cache_pmd(vma, address, &entry); |
1827 | page_remove_rmap(new_page); | |
1828 | goto fail_putback; | |
1829 | } | |
1830 | ||
0a31bc97 JW |
1831 | mem_cgroup_migrate(page, new_page, false); |
1832 | ||
b32967ff | 1833 | page_remove_rmap(page); |
2b4847e7 | 1834 | |
c4088ebd | 1835 | spin_unlock(ptl); |
f714f4f2 | 1836 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff | 1837 | |
11de9927 MG |
1838 | /* Take an "isolate" reference and put new page on the LRU. */ |
1839 | get_page(new_page); | |
1840 | putback_lru_page(new_page); | |
1841 | ||
b32967ff MG |
1842 | unlock_page(new_page); |
1843 | unlock_page(page); | |
1844 | put_page(page); /* Drop the rmap reference */ | |
1845 | put_page(page); /* Drop the LRU isolation reference */ | |
1846 | ||
1847 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1848 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1849 | ||
b32967ff MG |
1850 | mod_zone_page_state(page_zone(page), |
1851 | NR_ISOLATED_ANON + page_lru, | |
1852 | -HPAGE_PMD_NR); | |
1853 | return isolated; | |
1854 | ||
340ef390 HD |
1855 | out_fail: |
1856 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1857 | out_dropref: |
2b4847e7 MG |
1858 | ptl = pmd_lock(mm, pmd); |
1859 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 1860 | entry = pmd_modify(entry, vma->vm_page_prot); |
f714f4f2 | 1861 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
1862 | update_mmu_cache_pmd(vma, address, &entry); |
1863 | } | |
1864 | spin_unlock(ptl); | |
a54a407f | 1865 | |
eb4489f6 | 1866 | out_unlock: |
340ef390 | 1867 | unlock_page(page); |
b32967ff | 1868 | put_page(page); |
b32967ff MG |
1869 | return 0; |
1870 | } | |
7039e1db PZ |
1871 | #endif /* CONFIG_NUMA_BALANCING */ |
1872 | ||
1873 | #endif /* CONFIG_NUMA */ |