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> |
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 | ||
173 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 174 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 175 | unlock: |
0697212a | 176 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
177 | out: |
178 | return SWAP_AGAIN; | |
0697212a CL |
179 | } |
180 | ||
04e62a29 CL |
181 | /* |
182 | * Get rid of all migration entries and replace them by | |
183 | * references to the indicated page. | |
184 | */ | |
185 | static void remove_migration_ptes(struct page *old, struct page *new) | |
186 | { | |
051ac83a JK |
187 | struct rmap_walk_control rwc = { |
188 | .rmap_one = remove_migration_pte, | |
189 | .arg = old, | |
190 | }; | |
191 | ||
192 | rmap_walk(new, &rwc); | |
04e62a29 CL |
193 | } |
194 | ||
0697212a CL |
195 | /* |
196 | * Something used the pte of a page under migration. We need to | |
197 | * get to the page and wait until migration is finished. | |
198 | * When we return from this function the fault will be retried. | |
0697212a | 199 | */ |
e66f17ff | 200 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 201 | spinlock_t *ptl) |
0697212a | 202 | { |
30dad309 | 203 | pte_t pte; |
0697212a CL |
204 | swp_entry_t entry; |
205 | struct page *page; | |
206 | ||
30dad309 | 207 | spin_lock(ptl); |
0697212a CL |
208 | pte = *ptep; |
209 | if (!is_swap_pte(pte)) | |
210 | goto out; | |
211 | ||
212 | entry = pte_to_swp_entry(pte); | |
213 | if (!is_migration_entry(entry)) | |
214 | goto out; | |
215 | ||
216 | page = migration_entry_to_page(entry); | |
217 | ||
e286781d NP |
218 | /* |
219 | * Once radix-tree replacement of page migration started, page_count | |
220 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
221 | * against a page without get_page(). | |
222 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
223 | * will occur again. | |
224 | */ | |
225 | if (!get_page_unless_zero(page)) | |
226 | goto out; | |
0697212a CL |
227 | pte_unmap_unlock(ptep, ptl); |
228 | wait_on_page_locked(page); | |
229 | put_page(page); | |
230 | return; | |
231 | out: | |
232 | pte_unmap_unlock(ptep, ptl); | |
233 | } | |
234 | ||
30dad309 NH |
235 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
236 | unsigned long address) | |
237 | { | |
238 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
239 | pte_t *ptep = pte_offset_map(pmd, address); | |
240 | __migration_entry_wait(mm, ptep, ptl); | |
241 | } | |
242 | ||
cb900f41 KS |
243 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
244 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 245 | { |
cb900f41 | 246 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
247 | __migration_entry_wait(mm, pte, ptl); |
248 | } | |
249 | ||
b969c4ab MG |
250 | #ifdef CONFIG_BLOCK |
251 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
252 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
253 | enum migrate_mode mode) | |
b969c4ab MG |
254 | { |
255 | struct buffer_head *bh = head; | |
256 | ||
257 | /* Simple case, sync compaction */ | |
a6bc32b8 | 258 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
259 | do { |
260 | get_bh(bh); | |
261 | lock_buffer(bh); | |
262 | bh = bh->b_this_page; | |
263 | ||
264 | } while (bh != head); | |
265 | ||
266 | return true; | |
267 | } | |
268 | ||
269 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
270 | do { | |
271 | get_bh(bh); | |
272 | if (!trylock_buffer(bh)) { | |
273 | /* | |
274 | * We failed to lock the buffer and cannot stall in | |
275 | * async migration. Release the taken locks | |
276 | */ | |
277 | struct buffer_head *failed_bh = bh; | |
278 | put_bh(failed_bh); | |
279 | bh = head; | |
280 | while (bh != failed_bh) { | |
281 | unlock_buffer(bh); | |
282 | put_bh(bh); | |
283 | bh = bh->b_this_page; | |
284 | } | |
285 | return false; | |
286 | } | |
287 | ||
288 | bh = bh->b_this_page; | |
289 | } while (bh != head); | |
290 | return true; | |
291 | } | |
292 | #else | |
293 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 294 | enum migrate_mode mode) |
b969c4ab MG |
295 | { |
296 | return true; | |
297 | } | |
298 | #endif /* CONFIG_BLOCK */ | |
299 | ||
b20a3503 | 300 | /* |
c3fcf8a5 | 301 | * Replace the page in the mapping. |
5b5c7120 CL |
302 | * |
303 | * The number of remaining references must be: | |
304 | * 1 for anonymous pages without a mapping | |
305 | * 2 for pages with a mapping | |
266cf658 | 306 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 307 | */ |
36bc08cc | 308 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 309 | struct page *newpage, struct page *page, |
8e321fef BL |
310 | struct buffer_head *head, enum migrate_mode mode, |
311 | int extra_count) | |
b20a3503 | 312 | { |
8e321fef | 313 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 314 | void **pslot; |
b20a3503 | 315 | |
6c5240ae | 316 | if (!mapping) { |
0e8c7d0f | 317 | /* Anonymous page without mapping */ |
8e321fef | 318 | if (page_count(page) != expected_count) |
6c5240ae | 319 | return -EAGAIN; |
78bd5209 | 320 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
321 | } |
322 | ||
19fd6231 | 323 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 324 | |
7cf9c2c7 NP |
325 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
326 | page_index(page)); | |
b20a3503 | 327 | |
8e321fef | 328 | expected_count += 1 + page_has_private(page); |
e286781d | 329 | if (page_count(page) != expected_count || |
29c1f677 | 330 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 331 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 332 | return -EAGAIN; |
b20a3503 CL |
333 | } |
334 | ||
e286781d | 335 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 336 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
337 | return -EAGAIN; |
338 | } | |
339 | ||
b969c4ab MG |
340 | /* |
341 | * In the async migration case of moving a page with buffers, lock the | |
342 | * buffers using trylock before the mapping is moved. If the mapping | |
343 | * was moved, we later failed to lock the buffers and could not move | |
344 | * the mapping back due to an elevated page count, we would have to | |
345 | * block waiting on other references to be dropped. | |
346 | */ | |
a6bc32b8 MG |
347 | if (mode == MIGRATE_ASYNC && head && |
348 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
349 | page_unfreeze_refs(page, expected_count); |
350 | spin_unlock_irq(&mapping->tree_lock); | |
351 | return -EAGAIN; | |
352 | } | |
353 | ||
b20a3503 CL |
354 | /* |
355 | * Now we know that no one else is looking at the page. | |
b20a3503 | 356 | */ |
7cf9c2c7 | 357 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
358 | if (PageSwapCache(page)) { |
359 | SetPageSwapCache(newpage); | |
360 | set_page_private(newpage, page_private(page)); | |
361 | } | |
362 | ||
7cf9c2c7 NP |
363 | radix_tree_replace_slot(pslot, newpage); |
364 | ||
365 | /* | |
937a94c9 JG |
366 | * Drop cache reference from old page by unfreezing |
367 | * to one less reference. | |
7cf9c2c7 NP |
368 | * We know this isn't the last reference. |
369 | */ | |
937a94c9 | 370 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 371 | |
0e8c7d0f CL |
372 | /* |
373 | * If moved to a different zone then also account | |
374 | * the page for that zone. Other VM counters will be | |
375 | * taken care of when we establish references to the | |
376 | * new page and drop references to the old page. | |
377 | * | |
378 | * Note that anonymous pages are accounted for | |
379 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
380 | * are mapped to swap space. | |
381 | */ | |
382 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
383 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 384 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
385 | __dec_zone_page_state(page, NR_SHMEM); |
386 | __inc_zone_page_state(newpage, NR_SHMEM); | |
387 | } | |
19fd6231 | 388 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 389 | |
78bd5209 | 390 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 391 | } |
b20a3503 | 392 | |
290408d4 NH |
393 | /* |
394 | * The expected number of remaining references is the same as that | |
395 | * of migrate_page_move_mapping(). | |
396 | */ | |
397 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
398 | struct page *newpage, struct page *page) | |
399 | { | |
400 | int expected_count; | |
401 | void **pslot; | |
402 | ||
403 | if (!mapping) { | |
404 | if (page_count(page) != 1) | |
405 | return -EAGAIN; | |
78bd5209 | 406 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
407 | } |
408 | ||
409 | spin_lock_irq(&mapping->tree_lock); | |
410 | ||
411 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
412 | page_index(page)); | |
413 | ||
414 | expected_count = 2 + page_has_private(page); | |
415 | if (page_count(page) != expected_count || | |
29c1f677 | 416 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
417 | spin_unlock_irq(&mapping->tree_lock); |
418 | return -EAGAIN; | |
419 | } | |
420 | ||
421 | if (!page_freeze_refs(page, expected_count)) { | |
422 | spin_unlock_irq(&mapping->tree_lock); | |
423 | return -EAGAIN; | |
424 | } | |
425 | ||
426 | get_page(newpage); | |
427 | ||
428 | radix_tree_replace_slot(pslot, newpage); | |
429 | ||
937a94c9 | 430 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
431 | |
432 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 433 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
434 | } |
435 | ||
30b0a105 DH |
436 | /* |
437 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
438 | * arithmetic will work across the entire page. We need something more | |
439 | * specialized. | |
440 | */ | |
441 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
442 | int nr_pages) | |
443 | { | |
444 | int i; | |
445 | struct page *dst_base = dst; | |
446 | struct page *src_base = src; | |
447 | ||
448 | for (i = 0; i < nr_pages; ) { | |
449 | cond_resched(); | |
450 | copy_highpage(dst, src); | |
451 | ||
452 | i++; | |
453 | dst = mem_map_next(dst, dst_base, i); | |
454 | src = mem_map_next(src, src_base, i); | |
455 | } | |
456 | } | |
457 | ||
458 | static void copy_huge_page(struct page *dst, struct page *src) | |
459 | { | |
460 | int i; | |
461 | int nr_pages; | |
462 | ||
463 | if (PageHuge(src)) { | |
464 | /* hugetlbfs page */ | |
465 | struct hstate *h = page_hstate(src); | |
466 | nr_pages = pages_per_huge_page(h); | |
467 | ||
468 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
469 | __copy_gigantic_page(dst, src, nr_pages); | |
470 | return; | |
471 | } | |
472 | } else { | |
473 | /* thp page */ | |
474 | BUG_ON(!PageTransHuge(src)); | |
475 | nr_pages = hpage_nr_pages(src); | |
476 | } | |
477 | ||
478 | for (i = 0; i < nr_pages; i++) { | |
479 | cond_resched(); | |
480 | copy_highpage(dst + i, src + i); | |
481 | } | |
482 | } | |
483 | ||
b20a3503 CL |
484 | /* |
485 | * Copy the page to its new location | |
486 | */ | |
290408d4 | 487 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 488 | { |
7851a45c RR |
489 | int cpupid; |
490 | ||
b32967ff | 491 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
492 | copy_huge_page(newpage, page); |
493 | else | |
494 | copy_highpage(newpage, page); | |
b20a3503 CL |
495 | |
496 | if (PageError(page)) | |
497 | SetPageError(newpage); | |
498 | if (PageReferenced(page)) | |
499 | SetPageReferenced(newpage); | |
500 | if (PageUptodate(page)) | |
501 | SetPageUptodate(newpage); | |
894bc310 | 502 | if (TestClearPageActive(page)) { |
309381fe | 503 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 504 | SetPageActive(newpage); |
418b27ef LS |
505 | } else if (TestClearPageUnevictable(page)) |
506 | SetPageUnevictable(newpage); | |
b20a3503 CL |
507 | if (PageChecked(page)) |
508 | SetPageChecked(newpage); | |
509 | if (PageMappedToDisk(page)) | |
510 | SetPageMappedToDisk(newpage); | |
511 | ||
512 | if (PageDirty(page)) { | |
513 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
514 | /* |
515 | * Want to mark the page and the radix tree as dirty, and | |
516 | * redo the accounting that clear_page_dirty_for_io undid, | |
517 | * but we can't use set_page_dirty because that function | |
518 | * is actually a signal that all of the page has become dirty. | |
25985edc | 519 | * Whereas only part of our page may be dirty. |
3a902c5f | 520 | */ |
752dc185 HD |
521 | if (PageSwapBacked(page)) |
522 | SetPageDirty(newpage); | |
523 | else | |
524 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
525 | } |
526 | ||
7851a45c RR |
527 | /* |
528 | * Copy NUMA information to the new page, to prevent over-eager | |
529 | * future migrations of this same page. | |
530 | */ | |
531 | cpupid = page_cpupid_xchg_last(page, -1); | |
532 | page_cpupid_xchg_last(newpage, cpupid); | |
533 | ||
b291f000 | 534 | mlock_migrate_page(newpage, page); |
e9995ef9 | 535 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
536 | /* |
537 | * Please do not reorder this without considering how mm/ksm.c's | |
538 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
539 | */ | |
b3b3a99c NH |
540 | if (PageSwapCache(page)) |
541 | ClearPageSwapCache(page); | |
b20a3503 CL |
542 | ClearPagePrivate(page); |
543 | set_page_private(page, 0); | |
b20a3503 CL |
544 | |
545 | /* | |
546 | * If any waiters have accumulated on the new page then | |
547 | * wake them up. | |
548 | */ | |
549 | if (PageWriteback(newpage)) | |
550 | end_page_writeback(newpage); | |
551 | } | |
b20a3503 | 552 | |
1d8b85cc CL |
553 | /************************************************************ |
554 | * Migration functions | |
555 | ***********************************************************/ | |
556 | ||
b20a3503 CL |
557 | /* |
558 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 559 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
560 | * |
561 | * Pages are locked upon entry and exit. | |
562 | */ | |
2d1db3b1 | 563 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
564 | struct page *newpage, struct page *page, |
565 | enum migrate_mode mode) | |
b20a3503 CL |
566 | { |
567 | int rc; | |
568 | ||
569 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
570 | ||
8e321fef | 571 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 572 | |
78bd5209 | 573 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
574 | return rc; |
575 | ||
576 | migrate_page_copy(newpage, page); | |
78bd5209 | 577 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
578 | } |
579 | EXPORT_SYMBOL(migrate_page); | |
580 | ||
9361401e | 581 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
582 | /* |
583 | * Migration function for pages with buffers. This function can only be used | |
584 | * if the underlying filesystem guarantees that no other references to "page" | |
585 | * exist. | |
586 | */ | |
2d1db3b1 | 587 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 588 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 589 | { |
1d8b85cc CL |
590 | struct buffer_head *bh, *head; |
591 | int rc; | |
592 | ||
1d8b85cc | 593 | if (!page_has_buffers(page)) |
a6bc32b8 | 594 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
595 | |
596 | head = page_buffers(page); | |
597 | ||
8e321fef | 598 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 599 | |
78bd5209 | 600 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
601 | return rc; |
602 | ||
b969c4ab MG |
603 | /* |
604 | * In the async case, migrate_page_move_mapping locked the buffers | |
605 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
606 | * need to be locked now | |
607 | */ | |
a6bc32b8 MG |
608 | if (mode != MIGRATE_ASYNC) |
609 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
610 | |
611 | ClearPagePrivate(page); | |
612 | set_page_private(newpage, page_private(page)); | |
613 | set_page_private(page, 0); | |
614 | put_page(page); | |
615 | get_page(newpage); | |
616 | ||
617 | bh = head; | |
618 | do { | |
619 | set_bh_page(bh, newpage, bh_offset(bh)); | |
620 | bh = bh->b_this_page; | |
621 | ||
622 | } while (bh != head); | |
623 | ||
624 | SetPagePrivate(newpage); | |
625 | ||
626 | migrate_page_copy(newpage, page); | |
627 | ||
628 | bh = head; | |
629 | do { | |
630 | unlock_buffer(bh); | |
631 | put_bh(bh); | |
632 | bh = bh->b_this_page; | |
633 | ||
634 | } while (bh != head); | |
635 | ||
78bd5209 | 636 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
637 | } |
638 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 639 | #endif |
1d8b85cc | 640 | |
04e62a29 CL |
641 | /* |
642 | * Writeback a page to clean the dirty state | |
643 | */ | |
644 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 645 | { |
04e62a29 CL |
646 | struct writeback_control wbc = { |
647 | .sync_mode = WB_SYNC_NONE, | |
648 | .nr_to_write = 1, | |
649 | .range_start = 0, | |
650 | .range_end = LLONG_MAX, | |
04e62a29 CL |
651 | .for_reclaim = 1 |
652 | }; | |
653 | int rc; | |
654 | ||
655 | if (!mapping->a_ops->writepage) | |
656 | /* No write method for the address space */ | |
657 | return -EINVAL; | |
658 | ||
659 | if (!clear_page_dirty_for_io(page)) | |
660 | /* Someone else already triggered a write */ | |
661 | return -EAGAIN; | |
662 | ||
8351a6e4 | 663 | /* |
04e62a29 CL |
664 | * A dirty page may imply that the underlying filesystem has |
665 | * the page on some queue. So the page must be clean for | |
666 | * migration. Writeout may mean we loose the lock and the | |
667 | * page state is no longer what we checked for earlier. | |
668 | * At this point we know that the migration attempt cannot | |
669 | * be successful. | |
8351a6e4 | 670 | */ |
04e62a29 | 671 | remove_migration_ptes(page, page); |
8351a6e4 | 672 | |
04e62a29 | 673 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 674 | |
04e62a29 CL |
675 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
676 | /* unlocked. Relock */ | |
677 | lock_page(page); | |
678 | ||
bda8550d | 679 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
680 | } |
681 | ||
682 | /* | |
683 | * Default handling if a filesystem does not provide a migration function. | |
684 | */ | |
685 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 686 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 687 | { |
b969c4ab | 688 | if (PageDirty(page)) { |
a6bc32b8 MG |
689 | /* Only writeback pages in full synchronous migration */ |
690 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 691 | return -EBUSY; |
04e62a29 | 692 | return writeout(mapping, page); |
b969c4ab | 693 | } |
8351a6e4 CL |
694 | |
695 | /* | |
696 | * Buffers may be managed in a filesystem specific way. | |
697 | * We must have no buffers or drop them. | |
698 | */ | |
266cf658 | 699 | if (page_has_private(page) && |
8351a6e4 CL |
700 | !try_to_release_page(page, GFP_KERNEL)) |
701 | return -EAGAIN; | |
702 | ||
a6bc32b8 | 703 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
704 | } |
705 | ||
e24f0b8f CL |
706 | /* |
707 | * Move a page to a newly allocated page | |
708 | * The page is locked and all ptes have been successfully removed. | |
709 | * | |
710 | * The new page will have replaced the old page if this function | |
711 | * is successful. | |
894bc310 LS |
712 | * |
713 | * Return value: | |
714 | * < 0 - error code | |
78bd5209 | 715 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 716 | */ |
3fe2011f | 717 | static int move_to_new_page(struct page *newpage, struct page *page, |
2ebba6b7 | 718 | int page_was_mapped, enum migrate_mode mode) |
e24f0b8f CL |
719 | { |
720 | struct address_space *mapping; | |
721 | int rc; | |
722 | ||
723 | /* | |
724 | * Block others from accessing the page when we get around to | |
725 | * establishing additional references. We are the only one | |
726 | * holding a reference to the new page at this point. | |
727 | */ | |
529ae9aa | 728 | if (!trylock_page(newpage)) |
e24f0b8f CL |
729 | BUG(); |
730 | ||
731 | /* Prepare mapping for the new page.*/ | |
732 | newpage->index = page->index; | |
733 | newpage->mapping = page->mapping; | |
b2e18538 RR |
734 | if (PageSwapBacked(page)) |
735 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
736 | |
737 | mapping = page_mapping(page); | |
738 | if (!mapping) | |
a6bc32b8 | 739 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 740 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 741 | /* |
b969c4ab MG |
742 | * Most pages have a mapping and most filesystems provide a |
743 | * migratepage callback. Anonymous pages are part of swap | |
744 | * space which also has its own migratepage callback. This | |
745 | * is the most common path for page migration. | |
e24f0b8f | 746 | */ |
b969c4ab | 747 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 748 | newpage, page, mode); |
b969c4ab | 749 | else |
a6bc32b8 | 750 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 751 | |
78bd5209 | 752 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 753 | newpage->mapping = NULL; |
3fe2011f | 754 | } else { |
0a31bc97 | 755 | mem_cgroup_migrate(page, newpage, false); |
2ebba6b7 | 756 | if (page_was_mapped) |
3fe2011f | 757 | remove_migration_ptes(page, newpage); |
35512eca | 758 | page->mapping = NULL; |
3fe2011f | 759 | } |
e24f0b8f CL |
760 | |
761 | unlock_page(newpage); | |
762 | ||
763 | return rc; | |
764 | } | |
765 | ||
0dabec93 | 766 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 767 | int force, enum migrate_mode mode) |
e24f0b8f | 768 | { |
0dabec93 | 769 | int rc = -EAGAIN; |
2ebba6b7 | 770 | int page_was_mapped = 0; |
3f6c8272 | 771 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 772 | |
529ae9aa | 773 | if (!trylock_page(page)) { |
a6bc32b8 | 774 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 775 | goto out; |
3e7d3449 MG |
776 | |
777 | /* | |
778 | * It's not safe for direct compaction to call lock_page. | |
779 | * For example, during page readahead pages are added locked | |
780 | * to the LRU. Later, when the IO completes the pages are | |
781 | * marked uptodate and unlocked. However, the queueing | |
782 | * could be merging multiple pages for one bio (e.g. | |
783 | * mpage_readpages). If an allocation happens for the | |
784 | * second or third page, the process can end up locking | |
785 | * the same page twice and deadlocking. Rather than | |
786 | * trying to be clever about what pages can be locked, | |
787 | * avoid the use of lock_page for direct compaction | |
788 | * altogether. | |
789 | */ | |
790 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 791 | goto out; |
3e7d3449 | 792 | |
e24f0b8f CL |
793 | lock_page(page); |
794 | } | |
795 | ||
796 | if (PageWriteback(page)) { | |
11bc82d6 | 797 | /* |
fed5b64a | 798 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
799 | * necessary to wait for PageWriteback. In the async case, |
800 | * the retry loop is too short and in the sync-light case, | |
801 | * the overhead of stalling is too much | |
11bc82d6 | 802 | */ |
a6bc32b8 | 803 | if (mode != MIGRATE_SYNC) { |
11bc82d6 | 804 | rc = -EBUSY; |
0a31bc97 | 805 | goto out_unlock; |
11bc82d6 AA |
806 | } |
807 | if (!force) | |
0a31bc97 | 808 | goto out_unlock; |
e24f0b8f CL |
809 | wait_on_page_writeback(page); |
810 | } | |
e24f0b8f | 811 | /* |
dc386d4d KH |
812 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
813 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 814 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 815 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
816 | * File Caches may use write_page() or lock_page() in migration, then, |
817 | * just care Anon page here. | |
dc386d4d | 818 | */ |
b79bc0a0 | 819 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 820 | /* |
4fc3f1d6 | 821 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
822 | * getting a hold on an anon_vma from outside one of its mms. |
823 | */ | |
746b18d4 | 824 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
825 | if (anon_vma) { |
826 | /* | |
746b18d4 | 827 | * Anon page |
1ce82b69 | 828 | */ |
1ce82b69 | 829 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
830 | /* |
831 | * We cannot be sure that the anon_vma of an unmapped | |
832 | * swapcache page is safe to use because we don't | |
833 | * know in advance if the VMA that this page belonged | |
834 | * to still exists. If the VMA and others sharing the | |
835 | * data have been freed, then the anon_vma could | |
836 | * already be invalid. | |
837 | * | |
838 | * To avoid this possibility, swapcache pages get | |
839 | * migrated but are not remapped when migration | |
840 | * completes | |
841 | */ | |
3fe2011f | 842 | } else { |
0a31bc97 | 843 | goto out_unlock; |
3fe2011f | 844 | } |
989f89c5 | 845 | } |
62e1c553 | 846 | |
d6d86c0a | 847 | if (unlikely(isolated_balloon_page(page))) { |
bf6bddf1 RA |
848 | /* |
849 | * A ballooned page does not need any special attention from | |
850 | * physical to virtual reverse mapping procedures. | |
851 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
852 | * in order to avoid burning cycles at rmap level, and perform | |
853 | * the page migration right away (proteced by page lock). | |
854 | */ | |
855 | rc = balloon_page_migrate(newpage, page, mode); | |
0a31bc97 | 856 | goto out_unlock; |
bf6bddf1 RA |
857 | } |
858 | ||
dc386d4d | 859 | /* |
62e1c553 SL |
860 | * Corner case handling: |
861 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
862 | * and treated as swapcache but it has no rmap yet. | |
863 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
864 | * trigger a BUG. So handle it here. | |
865 | * 2. An orphaned page (see truncate_complete_page) might have | |
866 | * fs-private metadata. The page can be picked up due to memory | |
867 | * offlining. Everywhere else except page reclaim, the page is | |
868 | * invisible to the vm, so the page can not be migrated. So try to | |
869 | * free the metadata, so the page can be freed. | |
e24f0b8f | 870 | */ |
62e1c553 | 871 | if (!page->mapping) { |
309381fe | 872 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 873 | if (page_has_private(page)) { |
62e1c553 | 874 | try_to_free_buffers(page); |
0a31bc97 | 875 | goto out_unlock; |
62e1c553 | 876 | } |
abfc3488 | 877 | goto skip_unmap; |
62e1c553 SL |
878 | } |
879 | ||
dc386d4d | 880 | /* Establish migration ptes or remove ptes */ |
2ebba6b7 HD |
881 | if (page_mapped(page)) { |
882 | try_to_unmap(page, | |
883 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
884 | page_was_mapped = 1; | |
885 | } | |
dc386d4d | 886 | |
abfc3488 | 887 | skip_unmap: |
e6a1530d | 888 | if (!page_mapped(page)) |
2ebba6b7 | 889 | rc = move_to_new_page(newpage, page, page_was_mapped, mode); |
e24f0b8f | 890 | |
2ebba6b7 | 891 | if (rc && page_was_mapped) |
e24f0b8f | 892 | remove_migration_ptes(page, page); |
3f6c8272 MG |
893 | |
894 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 895 | if (anon_vma) |
9e60109f | 896 | put_anon_vma(anon_vma); |
3f6c8272 | 897 | |
0a31bc97 | 898 | out_unlock: |
e24f0b8f | 899 | unlock_page(page); |
0dabec93 MK |
900 | out: |
901 | return rc; | |
902 | } | |
95a402c3 | 903 | |
ef2a5153 GU |
904 | /* |
905 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
906 | * around it. | |
907 | */ | |
908 | #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) | |
909 | #define ICE_noinline noinline | |
910 | #else | |
911 | #define ICE_noinline | |
912 | #endif | |
913 | ||
0dabec93 MK |
914 | /* |
915 | * Obtain the lock on page, remove all ptes and migrate the page | |
916 | * to the newly allocated page in newpage. | |
917 | */ | |
ef2a5153 GU |
918 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
919 | free_page_t put_new_page, | |
920 | unsigned long private, struct page *page, | |
add05cec NH |
921 | int force, enum migrate_mode mode, |
922 | enum migrate_reason reason) | |
0dabec93 MK |
923 | { |
924 | int rc = 0; | |
925 | int *result = NULL; | |
926 | struct page *newpage = get_new_page(page, private, &result); | |
927 | ||
928 | if (!newpage) | |
929 | return -ENOMEM; | |
930 | ||
931 | if (page_count(page) == 1) { | |
932 | /* page was freed from under us. So we are done. */ | |
933 | goto out; | |
934 | } | |
935 | ||
936 | if (unlikely(PageTransHuge(page))) | |
937 | if (unlikely(split_huge_page(page))) | |
938 | goto out; | |
939 | ||
9c620e2b | 940 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 | 941 | |
0dabec93 | 942 | out: |
e24f0b8f | 943 | if (rc != -EAGAIN) { |
0dabec93 MK |
944 | /* |
945 | * A page that has been migrated has all references | |
946 | * removed and will be freed. A page that has not been | |
947 | * migrated will have kepts its references and be | |
948 | * restored. | |
949 | */ | |
950 | list_del(&page->lru); | |
a731286d | 951 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 952 | page_is_file_cache(page)); |
add05cec NH |
953 | if (reason != MR_MEMORY_FAILURE) |
954 | putback_lru_page(page); | |
e24f0b8f | 955 | } |
68711a74 | 956 | |
95a402c3 | 957 | /* |
68711a74 DR |
958 | * If migration was not successful and there's a freeing callback, use |
959 | * it. Otherwise, putback_lru_page() will drop the reference grabbed | |
960 | * during isolation. | |
95a402c3 | 961 | */ |
8bdd6380 HD |
962 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) { |
963 | ClearPageSwapBacked(newpage); | |
68711a74 | 964 | put_new_page(newpage, private); |
d6d86c0a KK |
965 | } else if (unlikely(__is_movable_balloon_page(newpage))) { |
966 | /* drop our reference, page already in the balloon */ | |
967 | put_page(newpage); | |
8bdd6380 | 968 | } else |
68711a74 DR |
969 | putback_lru_page(newpage); |
970 | ||
742755a1 CL |
971 | if (result) { |
972 | if (rc) | |
973 | *result = rc; | |
974 | else | |
975 | *result = page_to_nid(newpage); | |
976 | } | |
e24f0b8f CL |
977 | return rc; |
978 | } | |
979 | ||
290408d4 NH |
980 | /* |
981 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
982 | * | |
983 | * This function doesn't wait the completion of hugepage I/O | |
984 | * because there is no race between I/O and migration for hugepage. | |
985 | * Note that currently hugepage I/O occurs only in direct I/O | |
986 | * where no lock is held and PG_writeback is irrelevant, | |
987 | * and writeback status of all subpages are counted in the reference | |
988 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
989 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
990 | * This means that when we try to migrate hugepage whose subpages are | |
991 | * doing direct I/O, some references remain after try_to_unmap() and | |
992 | * hugepage migration fails without data corruption. | |
993 | * | |
994 | * There is also no race when direct I/O is issued on the page under migration, | |
995 | * because then pte is replaced with migration swap entry and direct I/O code | |
996 | * will wait in the page fault for migration to complete. | |
997 | */ | |
998 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
999 | free_page_t put_new_page, unsigned long private, |
1000 | struct page *hpage, int force, | |
1001 | enum migrate_mode mode) | |
290408d4 NH |
1002 | { |
1003 | int rc = 0; | |
1004 | int *result = NULL; | |
2ebba6b7 | 1005 | int page_was_mapped = 0; |
32665f2b | 1006 | struct page *new_hpage; |
290408d4 NH |
1007 | struct anon_vma *anon_vma = NULL; |
1008 | ||
83467efb NH |
1009 | /* |
1010 | * Movability of hugepages depends on architectures and hugepage size. | |
1011 | * This check is necessary because some callers of hugepage migration | |
1012 | * like soft offline and memory hotremove don't walk through page | |
1013 | * tables or check whether the hugepage is pmd-based or not before | |
1014 | * kicking migration. | |
1015 | */ | |
100873d7 | 1016 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1017 | putback_active_hugepage(hpage); |
83467efb | 1018 | return -ENOSYS; |
32665f2b | 1019 | } |
83467efb | 1020 | |
32665f2b | 1021 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1022 | if (!new_hpage) |
1023 | return -ENOMEM; | |
1024 | ||
1025 | rc = -EAGAIN; | |
1026 | ||
1027 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 1028 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1029 | goto out; |
1030 | lock_page(hpage); | |
1031 | } | |
1032 | ||
746b18d4 PZ |
1033 | if (PageAnon(hpage)) |
1034 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1035 | |
2ebba6b7 HD |
1036 | if (page_mapped(hpage)) { |
1037 | try_to_unmap(hpage, | |
1038 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1039 | page_was_mapped = 1; | |
1040 | } | |
290408d4 NH |
1041 | |
1042 | if (!page_mapped(hpage)) | |
2ebba6b7 | 1043 | rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode); |
290408d4 | 1044 | |
2ebba6b7 | 1045 | if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped) |
290408d4 NH |
1046 | remove_migration_ptes(hpage, hpage); |
1047 | ||
fd4a4663 | 1048 | if (anon_vma) |
9e60109f | 1049 | put_anon_vma(anon_vma); |
8e6ac7fa | 1050 | |
68711a74 | 1051 | if (rc == MIGRATEPAGE_SUCCESS) |
8e6ac7fa AK |
1052 | hugetlb_cgroup_migrate(hpage, new_hpage); |
1053 | ||
290408d4 | 1054 | unlock_page(hpage); |
09761333 | 1055 | out: |
b8ec1cee NH |
1056 | if (rc != -EAGAIN) |
1057 | putback_active_hugepage(hpage); | |
68711a74 DR |
1058 | |
1059 | /* | |
1060 | * If migration was not successful and there's a freeing callback, use | |
1061 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1062 | * isolation. | |
1063 | */ | |
1064 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) | |
1065 | put_new_page(new_hpage, private); | |
1066 | else | |
1067 | put_page(new_hpage); | |
1068 | ||
290408d4 NH |
1069 | if (result) { |
1070 | if (rc) | |
1071 | *result = rc; | |
1072 | else | |
1073 | *result = page_to_nid(new_hpage); | |
1074 | } | |
1075 | return rc; | |
1076 | } | |
1077 | ||
b20a3503 | 1078 | /* |
c73e5c9c SB |
1079 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1080 | * supplied as the target for the page migration | |
b20a3503 | 1081 | * |
c73e5c9c SB |
1082 | * @from: The list of pages to be migrated. |
1083 | * @get_new_page: The function used to allocate free pages to be used | |
1084 | * as the target of the page migration. | |
68711a74 DR |
1085 | * @put_new_page: The function used to free target pages if migration |
1086 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1087 | * @private: Private data to be passed on to get_new_page() |
1088 | * @mode: The migration mode that specifies the constraints for | |
1089 | * page migration, if any. | |
1090 | * @reason: The reason for page migration. | |
b20a3503 | 1091 | * |
c73e5c9c SB |
1092 | * The function returns after 10 attempts or if no pages are movable any more |
1093 | * because the list has become empty or no retryable pages exist any more. | |
1094 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1095 | * or free list only if ret != 0. |
b20a3503 | 1096 | * |
c73e5c9c | 1097 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1098 | */ |
9c620e2b | 1099 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1100 | free_page_t put_new_page, unsigned long private, |
1101 | enum migrate_mode mode, int reason) | |
b20a3503 | 1102 | { |
e24f0b8f | 1103 | int retry = 1; |
b20a3503 | 1104 | int nr_failed = 0; |
5647bc29 | 1105 | int nr_succeeded = 0; |
b20a3503 CL |
1106 | int pass = 0; |
1107 | struct page *page; | |
1108 | struct page *page2; | |
1109 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1110 | int rc; | |
1111 | ||
1112 | if (!swapwrite) | |
1113 | current->flags |= PF_SWAPWRITE; | |
1114 | ||
e24f0b8f CL |
1115 | for(pass = 0; pass < 10 && retry; pass++) { |
1116 | retry = 0; | |
b20a3503 | 1117 | |
e24f0b8f | 1118 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1119 | cond_resched(); |
2d1db3b1 | 1120 | |
31caf665 NH |
1121 | if (PageHuge(page)) |
1122 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 DR |
1123 | put_new_page, private, page, |
1124 | pass > 2, mode); | |
31caf665 | 1125 | else |
68711a74 | 1126 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1127 | private, page, pass > 2, mode, |
1128 | reason); | |
2d1db3b1 | 1129 | |
e24f0b8f | 1130 | switch(rc) { |
95a402c3 CL |
1131 | case -ENOMEM: |
1132 | goto out; | |
e24f0b8f | 1133 | case -EAGAIN: |
2d1db3b1 | 1134 | retry++; |
e24f0b8f | 1135 | break; |
78bd5209 | 1136 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1137 | nr_succeeded++; |
e24f0b8f CL |
1138 | break; |
1139 | default: | |
354a3363 NH |
1140 | /* |
1141 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1142 | * unlike -EAGAIN case, the failed page is | |
1143 | * removed from migration page list and not | |
1144 | * retried in the next outer loop. | |
1145 | */ | |
2d1db3b1 | 1146 | nr_failed++; |
e24f0b8f | 1147 | break; |
2d1db3b1 | 1148 | } |
b20a3503 CL |
1149 | } |
1150 | } | |
78bd5209 | 1151 | rc = nr_failed + retry; |
95a402c3 | 1152 | out: |
5647bc29 MG |
1153 | if (nr_succeeded) |
1154 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1155 | if (nr_failed) | |
1156 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1157 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1158 | ||
b20a3503 CL |
1159 | if (!swapwrite) |
1160 | current->flags &= ~PF_SWAPWRITE; | |
1161 | ||
78bd5209 | 1162 | return rc; |
b20a3503 | 1163 | } |
95a402c3 | 1164 | |
742755a1 CL |
1165 | #ifdef CONFIG_NUMA |
1166 | /* | |
1167 | * Move a list of individual pages | |
1168 | */ | |
1169 | struct page_to_node { | |
1170 | unsigned long addr; | |
1171 | struct page *page; | |
1172 | int node; | |
1173 | int status; | |
1174 | }; | |
1175 | ||
1176 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1177 | int **result) | |
1178 | { | |
1179 | struct page_to_node *pm = (struct page_to_node *)private; | |
1180 | ||
1181 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1182 | pm++; | |
1183 | ||
1184 | if (pm->node == MAX_NUMNODES) | |
1185 | return NULL; | |
1186 | ||
1187 | *result = &pm->status; | |
1188 | ||
e632a938 NH |
1189 | if (PageHuge(p)) |
1190 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1191 | pm->node); | |
1192 | else | |
1193 | return alloc_pages_exact_node(pm->node, | |
e97ca8e5 | 1194 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); |
742755a1 CL |
1195 | } |
1196 | ||
1197 | /* | |
1198 | * Move a set of pages as indicated in the pm array. The addr | |
1199 | * field must be set to the virtual address of the page to be moved | |
1200 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1201 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1202 | */ |
5e9a0f02 BG |
1203 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1204 | struct page_to_node *pm, | |
1205 | int migrate_all) | |
742755a1 CL |
1206 | { |
1207 | int err; | |
1208 | struct page_to_node *pp; | |
1209 | LIST_HEAD(pagelist); | |
1210 | ||
1211 | down_read(&mm->mmap_sem); | |
1212 | ||
1213 | /* | |
1214 | * Build a list of pages to migrate | |
1215 | */ | |
742755a1 CL |
1216 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1217 | struct vm_area_struct *vma; | |
1218 | struct page *page; | |
1219 | ||
742755a1 CL |
1220 | err = -EFAULT; |
1221 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1222 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1223 | goto set_status; |
1224 | ||
500d65d4 | 1225 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1226 | |
1227 | err = PTR_ERR(page); | |
1228 | if (IS_ERR(page)) | |
1229 | goto set_status; | |
1230 | ||
742755a1 CL |
1231 | err = -ENOENT; |
1232 | if (!page) | |
1233 | goto set_status; | |
1234 | ||
62b61f61 | 1235 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1236 | if (PageReserved(page)) |
742755a1 CL |
1237 | goto put_and_set; |
1238 | ||
1239 | pp->page = page; | |
1240 | err = page_to_nid(page); | |
1241 | ||
1242 | if (err == pp->node) | |
1243 | /* | |
1244 | * Node already in the right place | |
1245 | */ | |
1246 | goto put_and_set; | |
1247 | ||
1248 | err = -EACCES; | |
1249 | if (page_mapcount(page) > 1 && | |
1250 | !migrate_all) | |
1251 | goto put_and_set; | |
1252 | ||
e632a938 | 1253 | if (PageHuge(page)) { |
e66f17ff NH |
1254 | if (PageHead(page)) |
1255 | isolate_huge_page(page, &pagelist); | |
e632a938 NH |
1256 | goto put_and_set; |
1257 | } | |
1258 | ||
62695a84 | 1259 | err = isolate_lru_page(page); |
6d9c285a | 1260 | if (!err) { |
62695a84 | 1261 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1262 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1263 | page_is_file_cache(page)); | |
1264 | } | |
742755a1 CL |
1265 | put_and_set: |
1266 | /* | |
1267 | * Either remove the duplicate refcount from | |
1268 | * isolate_lru_page() or drop the page ref if it was | |
1269 | * not isolated. | |
1270 | */ | |
1271 | put_page(page); | |
1272 | set_status: | |
1273 | pp->status = err; | |
1274 | } | |
1275 | ||
e78bbfa8 | 1276 | err = 0; |
cf608ac1 | 1277 | if (!list_empty(&pagelist)) { |
68711a74 | 1278 | err = migrate_pages(&pagelist, new_page_node, NULL, |
9c620e2b | 1279 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1280 | if (err) |
e632a938 | 1281 | putback_movable_pages(&pagelist); |
cf608ac1 | 1282 | } |
742755a1 CL |
1283 | |
1284 | up_read(&mm->mmap_sem); | |
1285 | return err; | |
1286 | } | |
1287 | ||
5e9a0f02 BG |
1288 | /* |
1289 | * Migrate an array of page address onto an array of nodes and fill | |
1290 | * the corresponding array of status. | |
1291 | */ | |
3268c63e | 1292 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1293 | unsigned long nr_pages, |
1294 | const void __user * __user *pages, | |
1295 | const int __user *nodes, | |
1296 | int __user *status, int flags) | |
1297 | { | |
3140a227 | 1298 | struct page_to_node *pm; |
3140a227 BG |
1299 | unsigned long chunk_nr_pages; |
1300 | unsigned long chunk_start; | |
1301 | int err; | |
5e9a0f02 | 1302 | |
3140a227 BG |
1303 | err = -ENOMEM; |
1304 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1305 | if (!pm) | |
5e9a0f02 | 1306 | goto out; |
35282a2d BG |
1307 | |
1308 | migrate_prep(); | |
1309 | ||
5e9a0f02 | 1310 | /* |
3140a227 BG |
1311 | * Store a chunk of page_to_node array in a page, |
1312 | * but keep the last one as a marker | |
5e9a0f02 | 1313 | */ |
3140a227 | 1314 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1315 | |
3140a227 BG |
1316 | for (chunk_start = 0; |
1317 | chunk_start < nr_pages; | |
1318 | chunk_start += chunk_nr_pages) { | |
1319 | int j; | |
5e9a0f02 | 1320 | |
3140a227 BG |
1321 | if (chunk_start + chunk_nr_pages > nr_pages) |
1322 | chunk_nr_pages = nr_pages - chunk_start; | |
1323 | ||
1324 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1325 | for (j = 0; j < chunk_nr_pages; j++) { | |
1326 | const void __user *p; | |
5e9a0f02 BG |
1327 | int node; |
1328 | ||
3140a227 BG |
1329 | err = -EFAULT; |
1330 | if (get_user(p, pages + j + chunk_start)) | |
1331 | goto out_pm; | |
1332 | pm[j].addr = (unsigned long) p; | |
1333 | ||
1334 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1335 | goto out_pm; |
1336 | ||
1337 | err = -ENODEV; | |
6f5a55f1 LT |
1338 | if (node < 0 || node >= MAX_NUMNODES) |
1339 | goto out_pm; | |
1340 | ||
389162c2 | 1341 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1342 | goto out_pm; |
1343 | ||
1344 | err = -EACCES; | |
1345 | if (!node_isset(node, task_nodes)) | |
1346 | goto out_pm; | |
1347 | ||
3140a227 BG |
1348 | pm[j].node = node; |
1349 | } | |
1350 | ||
1351 | /* End marker for this chunk */ | |
1352 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1353 | ||
1354 | /* Migrate this chunk */ | |
1355 | err = do_move_page_to_node_array(mm, pm, | |
1356 | flags & MPOL_MF_MOVE_ALL); | |
1357 | if (err < 0) | |
1358 | goto out_pm; | |
5e9a0f02 | 1359 | |
5e9a0f02 | 1360 | /* Return status information */ |
3140a227 BG |
1361 | for (j = 0; j < chunk_nr_pages; j++) |
1362 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1363 | err = -EFAULT; |
3140a227 BG |
1364 | goto out_pm; |
1365 | } | |
1366 | } | |
1367 | err = 0; | |
5e9a0f02 BG |
1368 | |
1369 | out_pm: | |
3140a227 | 1370 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1371 | out: |
1372 | return err; | |
1373 | } | |
1374 | ||
742755a1 | 1375 | /* |
2f007e74 | 1376 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1377 | */ |
80bba129 BG |
1378 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1379 | const void __user **pages, int *status) | |
742755a1 | 1380 | { |
2f007e74 | 1381 | unsigned long i; |
2f007e74 | 1382 | |
742755a1 CL |
1383 | down_read(&mm->mmap_sem); |
1384 | ||
2f007e74 | 1385 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1386 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1387 | struct vm_area_struct *vma; |
1388 | struct page *page; | |
c095adbc | 1389 | int err = -EFAULT; |
2f007e74 BG |
1390 | |
1391 | vma = find_vma(mm, addr); | |
70384dc6 | 1392 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1393 | goto set_status; |
1394 | ||
2f007e74 | 1395 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1396 | |
1397 | err = PTR_ERR(page); | |
1398 | if (IS_ERR(page)) | |
1399 | goto set_status; | |
1400 | ||
742755a1 CL |
1401 | err = -ENOENT; |
1402 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1403 | if (!page || PageReserved(page)) |
742755a1 CL |
1404 | goto set_status; |
1405 | ||
1406 | err = page_to_nid(page); | |
1407 | set_status: | |
80bba129 BG |
1408 | *status = err; |
1409 | ||
1410 | pages++; | |
1411 | status++; | |
1412 | } | |
1413 | ||
1414 | up_read(&mm->mmap_sem); | |
1415 | } | |
1416 | ||
1417 | /* | |
1418 | * Determine the nodes of a user array of pages and store it in | |
1419 | * a user array of status. | |
1420 | */ | |
1421 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1422 | const void __user * __user *pages, | |
1423 | int __user *status) | |
1424 | { | |
1425 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1426 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1427 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1428 | |
87b8d1ad PA |
1429 | while (nr_pages) { |
1430 | unsigned long chunk_nr; | |
80bba129 | 1431 | |
87b8d1ad PA |
1432 | chunk_nr = nr_pages; |
1433 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1434 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1435 | ||
1436 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1437 | break; | |
80bba129 BG |
1438 | |
1439 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1440 | ||
87b8d1ad PA |
1441 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1442 | break; | |
742755a1 | 1443 | |
87b8d1ad PA |
1444 | pages += chunk_nr; |
1445 | status += chunk_nr; | |
1446 | nr_pages -= chunk_nr; | |
1447 | } | |
1448 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1449 | } |
1450 | ||
1451 | /* | |
1452 | * Move a list of pages in the address space of the currently executing | |
1453 | * process. | |
1454 | */ | |
938bb9f5 HC |
1455 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1456 | const void __user * __user *, pages, | |
1457 | const int __user *, nodes, | |
1458 | int __user *, status, int, flags) | |
742755a1 | 1459 | { |
c69e8d9c | 1460 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1461 | struct task_struct *task; |
742755a1 | 1462 | struct mm_struct *mm; |
5e9a0f02 | 1463 | int err; |
3268c63e | 1464 | nodemask_t task_nodes; |
742755a1 CL |
1465 | |
1466 | /* Check flags */ | |
1467 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1468 | return -EINVAL; | |
1469 | ||
1470 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1471 | return -EPERM; | |
1472 | ||
1473 | /* Find the mm_struct */ | |
a879bf58 | 1474 | rcu_read_lock(); |
228ebcbe | 1475 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1476 | if (!task) { |
a879bf58 | 1477 | rcu_read_unlock(); |
742755a1 CL |
1478 | return -ESRCH; |
1479 | } | |
3268c63e | 1480 | get_task_struct(task); |
742755a1 CL |
1481 | |
1482 | /* | |
1483 | * Check if this process has the right to modify the specified | |
1484 | * process. The right exists if the process has administrative | |
1485 | * capabilities, superuser privileges or the same | |
1486 | * userid as the target process. | |
1487 | */ | |
c69e8d9c | 1488 | tcred = __task_cred(task); |
b38a86eb EB |
1489 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1490 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1491 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1492 | rcu_read_unlock(); |
742755a1 | 1493 | err = -EPERM; |
5e9a0f02 | 1494 | goto out; |
742755a1 | 1495 | } |
c69e8d9c | 1496 | rcu_read_unlock(); |
742755a1 | 1497 | |
86c3a764 DQ |
1498 | err = security_task_movememory(task); |
1499 | if (err) | |
5e9a0f02 | 1500 | goto out; |
86c3a764 | 1501 | |
3268c63e CL |
1502 | task_nodes = cpuset_mems_allowed(task); |
1503 | mm = get_task_mm(task); | |
1504 | put_task_struct(task); | |
1505 | ||
6e8b09ea SL |
1506 | if (!mm) |
1507 | return -EINVAL; | |
1508 | ||
1509 | if (nodes) | |
1510 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1511 | nodes, status, flags); | |
1512 | else | |
1513 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1514 | |
742755a1 CL |
1515 | mmput(mm); |
1516 | return err; | |
3268c63e CL |
1517 | |
1518 | out: | |
1519 | put_task_struct(task); | |
1520 | return err; | |
742755a1 | 1521 | } |
742755a1 | 1522 | |
7039e1db PZ |
1523 | #ifdef CONFIG_NUMA_BALANCING |
1524 | /* | |
1525 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1526 | * pages. Currently it only checks the watermarks which crude | |
1527 | */ | |
1528 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1529 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1530 | { |
1531 | int z; | |
1532 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1533 | struct zone *zone = pgdat->node_zones + z; | |
1534 | ||
1535 | if (!populated_zone(zone)) | |
1536 | continue; | |
1537 | ||
6e543d57 | 1538 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1539 | continue; |
1540 | ||
1541 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1542 | if (!zone_watermark_ok(zone, 0, | |
1543 | high_wmark_pages(zone) + | |
1544 | nr_migrate_pages, | |
1545 | 0, 0)) | |
1546 | continue; | |
1547 | return true; | |
1548 | } | |
1549 | return false; | |
1550 | } | |
1551 | ||
1552 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1553 | unsigned long data, | |
1554 | int **result) | |
1555 | { | |
1556 | int nid = (int) data; | |
1557 | struct page *newpage; | |
1558 | ||
1559 | newpage = alloc_pages_exact_node(nid, | |
e97ca8e5 JW |
1560 | (GFP_HIGHUSER_MOVABLE | |
1561 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1562 | __GFP_NORETRY | __GFP_NOWARN) & | |
7039e1db | 1563 | ~GFP_IOFS, 0); |
bac0382c | 1564 | |
7039e1db PZ |
1565 | return newpage; |
1566 | } | |
1567 | ||
a8f60772 MG |
1568 | /* |
1569 | * page migration rate limiting control. | |
1570 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1571 | * window of time. Default here says do not migrate more than 1280M per second. | |
1572 | */ | |
1573 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1574 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1575 | ||
b32967ff | 1576 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1577 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1578 | unsigned long nr_pages) | |
7039e1db | 1579 | { |
a8f60772 MG |
1580 | /* |
1581 | * Rate-limit the amount of data that is being migrated to a node. | |
1582 | * Optimal placement is no good if the memory bus is saturated and | |
1583 | * all the time is being spent migrating! | |
1584 | */ | |
a8f60772 | 1585 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1586 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1587 | pgdat->numabalancing_migrate_nr_pages = 0; |
1588 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1589 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1590 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1591 | } |
af1839d7 MG |
1592 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1593 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1594 | nr_pages); | |
1c5e9c27 | 1595 | return true; |
af1839d7 | 1596 | } |
1c5e9c27 MG |
1597 | |
1598 | /* | |
1599 | * This is an unlocked non-atomic update so errors are possible. | |
1600 | * The consequences are failing to migrate when we potentiall should | |
1601 | * have which is not severe enough to warrant locking. If it is ever | |
1602 | * a problem, it can be converted to a per-cpu counter. | |
1603 | */ | |
1604 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1605 | return false; | |
b32967ff MG |
1606 | } |
1607 | ||
1c30e017 | 1608 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1609 | { |
340ef390 | 1610 | int page_lru; |
a8f60772 | 1611 | |
309381fe | 1612 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1613 | |
7039e1db | 1614 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1615 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1616 | return 0; | |
7039e1db | 1617 | |
340ef390 HD |
1618 | if (isolate_lru_page(page)) |
1619 | return 0; | |
7039e1db | 1620 | |
340ef390 HD |
1621 | /* |
1622 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1623 | * check on page_count(), so we must do it here, now that the page | |
1624 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1625 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1626 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1627 | */ | |
1628 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1629 | putback_lru_page(page); | |
1630 | return 0; | |
7039e1db PZ |
1631 | } |
1632 | ||
340ef390 HD |
1633 | page_lru = page_is_file_cache(page); |
1634 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1635 | hpage_nr_pages(page)); | |
1636 | ||
149c33e1 | 1637 | /* |
340ef390 HD |
1638 | * Isolating the page has taken another reference, so the |
1639 | * caller's reference can be safely dropped without the page | |
1640 | * disappearing underneath us during migration. | |
149c33e1 MG |
1641 | */ |
1642 | put_page(page); | |
340ef390 | 1643 | return 1; |
b32967ff MG |
1644 | } |
1645 | ||
de466bd6 MG |
1646 | bool pmd_trans_migrating(pmd_t pmd) |
1647 | { | |
1648 | struct page *page = pmd_page(pmd); | |
1649 | return PageLocked(page); | |
1650 | } | |
1651 | ||
b32967ff MG |
1652 | /* |
1653 | * Attempt to migrate a misplaced page to the specified destination | |
1654 | * node. Caller is expected to have an elevated reference count on | |
1655 | * the page that will be dropped by this function before returning. | |
1656 | */ | |
1bc115d8 MG |
1657 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1658 | int node) | |
b32967ff MG |
1659 | { |
1660 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1661 | int isolated; |
b32967ff MG |
1662 | int nr_remaining; |
1663 | LIST_HEAD(migratepages); | |
1664 | ||
1665 | /* | |
1bc115d8 MG |
1666 | * Don't migrate file pages that are mapped in multiple processes |
1667 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1668 | */ |
1bc115d8 MG |
1669 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1670 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1671 | goto out; |
b32967ff MG |
1672 | |
1673 | /* | |
1674 | * Rate-limit the amount of data that is being migrated to a node. | |
1675 | * Optimal placement is no good if the memory bus is saturated and | |
1676 | * all the time is being spent migrating! | |
1677 | */ | |
340ef390 | 1678 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1679 | goto out; |
b32967ff MG |
1680 | |
1681 | isolated = numamigrate_isolate_page(pgdat, page); | |
1682 | if (!isolated) | |
1683 | goto out; | |
1684 | ||
1685 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1686 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1687 | NULL, node, MIGRATE_ASYNC, |
1688 | MR_NUMA_MISPLACED); | |
b32967ff | 1689 | if (nr_remaining) { |
59c82b70 JK |
1690 | if (!list_empty(&migratepages)) { |
1691 | list_del(&page->lru); | |
1692 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1693 | page_is_file_cache(page)); | |
1694 | putback_lru_page(page); | |
1695 | } | |
b32967ff MG |
1696 | isolated = 0; |
1697 | } else | |
1698 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1699 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1700 | return isolated; |
340ef390 HD |
1701 | |
1702 | out: | |
1703 | put_page(page); | |
1704 | return 0; | |
7039e1db | 1705 | } |
220018d3 | 1706 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1707 | |
220018d3 | 1708 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1709 | /* |
1710 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1711 | * before returning. | |
1712 | */ | |
b32967ff MG |
1713 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1714 | struct vm_area_struct *vma, | |
1715 | pmd_t *pmd, pmd_t entry, | |
1716 | unsigned long address, | |
1717 | struct page *page, int node) | |
1718 | { | |
c4088ebd | 1719 | spinlock_t *ptl; |
b32967ff MG |
1720 | pg_data_t *pgdat = NODE_DATA(node); |
1721 | int isolated = 0; | |
1722 | struct page *new_page = NULL; | |
b32967ff | 1723 | int page_lru = page_is_file_cache(page); |
f714f4f2 MG |
1724 | unsigned long mmun_start = address & HPAGE_PMD_MASK; |
1725 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
2b4847e7 | 1726 | pmd_t orig_entry; |
b32967ff | 1727 | |
b32967ff MG |
1728 | /* |
1729 | * Rate-limit the amount of data that is being migrated to a node. | |
1730 | * Optimal placement is no good if the memory bus is saturated and | |
1731 | * all the time is being spent migrating! | |
1732 | */ | |
d28d4335 | 1733 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1734 | goto out_dropref; |
1735 | ||
1736 | new_page = alloc_pages_node(node, | |
e97ca8e5 JW |
1737 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT, |
1738 | HPAGE_PMD_ORDER); | |
340ef390 HD |
1739 | if (!new_page) |
1740 | goto out_fail; | |
1741 | ||
b32967ff | 1742 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 1743 | if (!isolated) { |
b32967ff | 1744 | put_page(new_page); |
340ef390 | 1745 | goto out_fail; |
b32967ff MG |
1746 | } |
1747 | ||
b0943d61 MG |
1748 | if (mm_tlb_flush_pending(mm)) |
1749 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1750 | ||
b32967ff MG |
1751 | /* Prepare a page as a migration target */ |
1752 | __set_page_locked(new_page); | |
1753 | SetPageSwapBacked(new_page); | |
1754 | ||
1755 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1756 | new_page->mapping = page->mapping; | |
1757 | new_page->index = page->index; | |
1758 | migrate_page_copy(new_page, page); | |
1759 | WARN_ON(PageLRU(new_page)); | |
1760 | ||
1761 | /* Recheck the target PMD */ | |
f714f4f2 | 1762 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 1763 | ptl = pmd_lock(mm, pmd); |
2b4847e7 MG |
1764 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { |
1765 | fail_putback: | |
c4088ebd | 1766 | spin_unlock(ptl); |
f714f4f2 | 1767 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1768 | |
1769 | /* Reverse changes made by migrate_page_copy() */ | |
1770 | if (TestClearPageActive(new_page)) | |
1771 | SetPageActive(page); | |
1772 | if (TestClearPageUnevictable(new_page)) | |
1773 | SetPageUnevictable(page); | |
1774 | mlock_migrate_page(page, new_page); | |
1775 | ||
1776 | unlock_page(new_page); | |
1777 | put_page(new_page); /* Free it */ | |
1778 | ||
a54a407f MG |
1779 | /* Retake the callers reference and putback on LRU */ |
1780 | get_page(page); | |
b32967ff | 1781 | putback_lru_page(page); |
a54a407f MG |
1782 | mod_zone_page_state(page_zone(page), |
1783 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
eb4489f6 MG |
1784 | |
1785 | goto out_unlock; | |
b32967ff MG |
1786 | } |
1787 | ||
2b4847e7 | 1788 | orig_entry = *pmd; |
b32967ff | 1789 | entry = mk_pmd(new_page, vma->vm_page_prot); |
b32967ff | 1790 | entry = pmd_mkhuge(entry); |
2b4847e7 | 1791 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 1792 | |
2b4847e7 MG |
1793 | /* |
1794 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1795 | * Any parallel GUP will either observe the old page blocking on the | |
1796 | * page lock, block on the page table lock or observe the new page. | |
1797 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1798 | * guarantee the copy is visible before the pagetable update. | |
1799 | */ | |
f714f4f2 | 1800 | flush_cache_range(vma, mmun_start, mmun_end); |
11de9927 | 1801 | page_add_anon_rmap(new_page, vma, mmun_start); |
34ee645e | 1802 | pmdp_clear_flush_notify(vma, mmun_start, pmd); |
f714f4f2 MG |
1803 | set_pmd_at(mm, mmun_start, pmd, entry); |
1804 | flush_tlb_range(vma, mmun_start, mmun_end); | |
ce4a9cc5 | 1805 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 MG |
1806 | |
1807 | if (page_count(page) != 2) { | |
f714f4f2 MG |
1808 | set_pmd_at(mm, mmun_start, pmd, orig_entry); |
1809 | flush_tlb_range(vma, mmun_start, mmun_end); | |
34ee645e | 1810 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
2b4847e7 MG |
1811 | update_mmu_cache_pmd(vma, address, &entry); |
1812 | page_remove_rmap(new_page); | |
1813 | goto fail_putback; | |
1814 | } | |
1815 | ||
0a31bc97 JW |
1816 | mem_cgroup_migrate(page, new_page, false); |
1817 | ||
b32967ff | 1818 | page_remove_rmap(page); |
2b4847e7 | 1819 | |
c4088ebd | 1820 | spin_unlock(ptl); |
f714f4f2 | 1821 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff | 1822 | |
11de9927 MG |
1823 | /* Take an "isolate" reference and put new page on the LRU. */ |
1824 | get_page(new_page); | |
1825 | putback_lru_page(new_page); | |
1826 | ||
b32967ff MG |
1827 | unlock_page(new_page); |
1828 | unlock_page(page); | |
1829 | put_page(page); /* Drop the rmap reference */ | |
1830 | put_page(page); /* Drop the LRU isolation reference */ | |
1831 | ||
1832 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1833 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1834 | ||
b32967ff MG |
1835 | mod_zone_page_state(page_zone(page), |
1836 | NR_ISOLATED_ANON + page_lru, | |
1837 | -HPAGE_PMD_NR); | |
1838 | return isolated; | |
1839 | ||
340ef390 HD |
1840 | out_fail: |
1841 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1842 | out_dropref: |
2b4847e7 MG |
1843 | ptl = pmd_lock(mm, pmd); |
1844 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 1845 | entry = pmd_modify(entry, vma->vm_page_prot); |
f714f4f2 | 1846 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
1847 | update_mmu_cache_pmd(vma, address, &entry); |
1848 | } | |
1849 | spin_unlock(ptl); | |
a54a407f | 1850 | |
eb4489f6 | 1851 | out_unlock: |
340ef390 | 1852 | unlock_page(page); |
b32967ff | 1853 | put_page(page); |
b32967ff MG |
1854 | return 0; |
1855 | } | |
7039e1db PZ |
1856 | #endif /* CONFIG_NUMA_BALANCING */ |
1857 | ||
1858 | #endif /* CONFIG_NUMA */ |