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; |
500d65d4 AA |
123 | if (pmd_trans_huge(*pmd)) |
124 | goto out; | |
0697212a | 125 | |
290408d4 | 126 | ptep = pte_offset_map(pmd, addr); |
0697212a | 127 | |
486cf46f HD |
128 | /* |
129 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
130 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
131 | */ | |
290408d4 NH |
132 | |
133 | ptl = pte_lockptr(mm, pmd); | |
134 | } | |
0697212a | 135 | |
0697212a CL |
136 | spin_lock(ptl); |
137 | pte = *ptep; | |
138 | if (!is_swap_pte(pte)) | |
e9995ef9 | 139 | goto unlock; |
0697212a CL |
140 | |
141 | entry = pte_to_swp_entry(pte); | |
142 | ||
e9995ef9 HD |
143 | if (!is_migration_entry(entry) || |
144 | migration_entry_to_page(entry) != old) | |
145 | goto unlock; | |
0697212a | 146 | |
0697212a CL |
147 | get_page(new); |
148 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
c3d16e16 CG |
149 | if (pte_swp_soft_dirty(*ptep)) |
150 | pte = pte_mksoft_dirty(pte); | |
0697212a CL |
151 | if (is_write_migration_entry(entry)) |
152 | pte = pte_mkwrite(pte); | |
3ef8fd7f | 153 | #ifdef CONFIG_HUGETLB_PAGE |
be7517d6 | 154 | if (PageHuge(new)) { |
290408d4 | 155 | pte = pte_mkhuge(pte); |
be7517d6 TL |
156 | pte = arch_make_huge_pte(pte, vma, new, 0); |
157 | } | |
3ef8fd7f | 158 | #endif |
c2cc499c | 159 | flush_dcache_page(new); |
0697212a | 160 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 161 | |
290408d4 NH |
162 | if (PageHuge(new)) { |
163 | if (PageAnon(new)) | |
164 | hugepage_add_anon_rmap(new, vma, addr); | |
165 | else | |
166 | page_dup_rmap(new); | |
167 | } else if (PageAnon(new)) | |
04e62a29 CL |
168 | page_add_anon_rmap(new, vma, addr); |
169 | else | |
170 | page_add_file_rmap(new); | |
171 | ||
172 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 173 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 174 | unlock: |
0697212a | 175 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
176 | out: |
177 | return SWAP_AGAIN; | |
0697212a CL |
178 | } |
179 | ||
04e62a29 CL |
180 | /* |
181 | * Get rid of all migration entries and replace them by | |
182 | * references to the indicated page. | |
183 | */ | |
184 | static void remove_migration_ptes(struct page *old, struct page *new) | |
185 | { | |
051ac83a JK |
186 | struct rmap_walk_control rwc = { |
187 | .rmap_one = remove_migration_pte, | |
188 | .arg = old, | |
189 | }; | |
190 | ||
191 | rmap_walk(new, &rwc); | |
04e62a29 CL |
192 | } |
193 | ||
0697212a CL |
194 | /* |
195 | * Something used the pte of a page under migration. We need to | |
196 | * get to the page and wait until migration is finished. | |
197 | * When we return from this function the fault will be retried. | |
0697212a | 198 | */ |
30dad309 NH |
199 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
200 | spinlock_t *ptl) | |
0697212a | 201 | { |
30dad309 | 202 | pte_t pte; |
0697212a CL |
203 | swp_entry_t entry; |
204 | struct page *page; | |
205 | ||
30dad309 | 206 | spin_lock(ptl); |
0697212a CL |
207 | pte = *ptep; |
208 | if (!is_swap_pte(pte)) | |
209 | goto out; | |
210 | ||
211 | entry = pte_to_swp_entry(pte); | |
212 | if (!is_migration_entry(entry)) | |
213 | goto out; | |
214 | ||
215 | page = migration_entry_to_page(entry); | |
216 | ||
e286781d NP |
217 | /* |
218 | * Once radix-tree replacement of page migration started, page_count | |
219 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
220 | * against a page without get_page(). | |
221 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
222 | * will occur again. | |
223 | */ | |
224 | if (!get_page_unless_zero(page)) | |
225 | goto out; | |
0697212a CL |
226 | pte_unmap_unlock(ptep, ptl); |
227 | wait_on_page_locked(page); | |
228 | put_page(page); | |
229 | return; | |
230 | out: | |
231 | pte_unmap_unlock(ptep, ptl); | |
232 | } | |
233 | ||
30dad309 NH |
234 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
235 | unsigned long address) | |
236 | { | |
237 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
238 | pte_t *ptep = pte_offset_map(pmd, address); | |
239 | __migration_entry_wait(mm, ptep, ptl); | |
240 | } | |
241 | ||
cb900f41 KS |
242 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
243 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 244 | { |
cb900f41 | 245 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
246 | __migration_entry_wait(mm, pte, ptl); |
247 | } | |
248 | ||
b969c4ab MG |
249 | #ifdef CONFIG_BLOCK |
250 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
251 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
252 | enum migrate_mode mode) | |
b969c4ab MG |
253 | { |
254 | struct buffer_head *bh = head; | |
255 | ||
256 | /* Simple case, sync compaction */ | |
a6bc32b8 | 257 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
258 | do { |
259 | get_bh(bh); | |
260 | lock_buffer(bh); | |
261 | bh = bh->b_this_page; | |
262 | ||
263 | } while (bh != head); | |
264 | ||
265 | return true; | |
266 | } | |
267 | ||
268 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
269 | do { | |
270 | get_bh(bh); | |
271 | if (!trylock_buffer(bh)) { | |
272 | /* | |
273 | * We failed to lock the buffer and cannot stall in | |
274 | * async migration. Release the taken locks | |
275 | */ | |
276 | struct buffer_head *failed_bh = bh; | |
277 | put_bh(failed_bh); | |
278 | bh = head; | |
279 | while (bh != failed_bh) { | |
280 | unlock_buffer(bh); | |
281 | put_bh(bh); | |
282 | bh = bh->b_this_page; | |
283 | } | |
284 | return false; | |
285 | } | |
286 | ||
287 | bh = bh->b_this_page; | |
288 | } while (bh != head); | |
289 | return true; | |
290 | } | |
291 | #else | |
292 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 293 | enum migrate_mode mode) |
b969c4ab MG |
294 | { |
295 | return true; | |
296 | } | |
297 | #endif /* CONFIG_BLOCK */ | |
298 | ||
b20a3503 | 299 | /* |
c3fcf8a5 | 300 | * Replace the page in the mapping. |
5b5c7120 CL |
301 | * |
302 | * The number of remaining references must be: | |
303 | * 1 for anonymous pages without a mapping | |
304 | * 2 for pages with a mapping | |
266cf658 | 305 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 306 | */ |
36bc08cc | 307 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 308 | struct page *newpage, struct page *page, |
8e321fef BL |
309 | struct buffer_head *head, enum migrate_mode mode, |
310 | int extra_count) | |
b20a3503 | 311 | { |
8e321fef | 312 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 313 | void **pslot; |
b20a3503 | 314 | |
6c5240ae | 315 | if (!mapping) { |
0e8c7d0f | 316 | /* Anonymous page without mapping */ |
8e321fef | 317 | if (page_count(page) != expected_count) |
6c5240ae | 318 | return -EAGAIN; |
78bd5209 | 319 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
320 | } |
321 | ||
19fd6231 | 322 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 323 | |
7cf9c2c7 NP |
324 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
325 | page_index(page)); | |
b20a3503 | 326 | |
8e321fef | 327 | expected_count += 1 + page_has_private(page); |
e286781d | 328 | if (page_count(page) != expected_count || |
29c1f677 | 329 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 330 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 331 | return -EAGAIN; |
b20a3503 CL |
332 | } |
333 | ||
e286781d | 334 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 335 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
336 | return -EAGAIN; |
337 | } | |
338 | ||
b969c4ab MG |
339 | /* |
340 | * In the async migration case of moving a page with buffers, lock the | |
341 | * buffers using trylock before the mapping is moved. If the mapping | |
342 | * was moved, we later failed to lock the buffers and could not move | |
343 | * the mapping back due to an elevated page count, we would have to | |
344 | * block waiting on other references to be dropped. | |
345 | */ | |
a6bc32b8 MG |
346 | if (mode == MIGRATE_ASYNC && head && |
347 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
348 | page_unfreeze_refs(page, expected_count); |
349 | spin_unlock_irq(&mapping->tree_lock); | |
350 | return -EAGAIN; | |
351 | } | |
352 | ||
b20a3503 CL |
353 | /* |
354 | * Now we know that no one else is looking at the page. | |
b20a3503 | 355 | */ |
7cf9c2c7 | 356 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
357 | if (PageSwapCache(page)) { |
358 | SetPageSwapCache(newpage); | |
359 | set_page_private(newpage, page_private(page)); | |
360 | } | |
361 | ||
7cf9c2c7 NP |
362 | radix_tree_replace_slot(pslot, newpage); |
363 | ||
364 | /* | |
937a94c9 JG |
365 | * Drop cache reference from old page by unfreezing |
366 | * to one less reference. | |
7cf9c2c7 NP |
367 | * We know this isn't the last reference. |
368 | */ | |
937a94c9 | 369 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 370 | |
0e8c7d0f CL |
371 | /* |
372 | * If moved to a different zone then also account | |
373 | * the page for that zone. Other VM counters will be | |
374 | * taken care of when we establish references to the | |
375 | * new page and drop references to the old page. | |
376 | * | |
377 | * Note that anonymous pages are accounted for | |
378 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
379 | * are mapped to swap space. | |
380 | */ | |
381 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
382 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 383 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
384 | __dec_zone_page_state(page, NR_SHMEM); |
385 | __inc_zone_page_state(newpage, NR_SHMEM); | |
386 | } | |
19fd6231 | 387 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 388 | |
78bd5209 | 389 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 390 | } |
b20a3503 | 391 | |
290408d4 NH |
392 | /* |
393 | * The expected number of remaining references is the same as that | |
394 | * of migrate_page_move_mapping(). | |
395 | */ | |
396 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
397 | struct page *newpage, struct page *page) | |
398 | { | |
399 | int expected_count; | |
400 | void **pslot; | |
401 | ||
402 | if (!mapping) { | |
403 | if (page_count(page) != 1) | |
404 | return -EAGAIN; | |
78bd5209 | 405 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
406 | } |
407 | ||
408 | spin_lock_irq(&mapping->tree_lock); | |
409 | ||
410 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
411 | page_index(page)); | |
412 | ||
413 | expected_count = 2 + page_has_private(page); | |
414 | if (page_count(page) != expected_count || | |
29c1f677 | 415 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
416 | spin_unlock_irq(&mapping->tree_lock); |
417 | return -EAGAIN; | |
418 | } | |
419 | ||
420 | if (!page_freeze_refs(page, expected_count)) { | |
421 | spin_unlock_irq(&mapping->tree_lock); | |
422 | return -EAGAIN; | |
423 | } | |
424 | ||
425 | get_page(newpage); | |
426 | ||
427 | radix_tree_replace_slot(pslot, newpage); | |
428 | ||
937a94c9 | 429 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
430 | |
431 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 432 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
433 | } |
434 | ||
30b0a105 DH |
435 | /* |
436 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
437 | * arithmetic will work across the entire page. We need something more | |
438 | * specialized. | |
439 | */ | |
440 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
441 | int nr_pages) | |
442 | { | |
443 | int i; | |
444 | struct page *dst_base = dst; | |
445 | struct page *src_base = src; | |
446 | ||
447 | for (i = 0; i < nr_pages; ) { | |
448 | cond_resched(); | |
449 | copy_highpage(dst, src); | |
450 | ||
451 | i++; | |
452 | dst = mem_map_next(dst, dst_base, i); | |
453 | src = mem_map_next(src, src_base, i); | |
454 | } | |
455 | } | |
456 | ||
457 | static void copy_huge_page(struct page *dst, struct page *src) | |
458 | { | |
459 | int i; | |
460 | int nr_pages; | |
461 | ||
462 | if (PageHuge(src)) { | |
463 | /* hugetlbfs page */ | |
464 | struct hstate *h = page_hstate(src); | |
465 | nr_pages = pages_per_huge_page(h); | |
466 | ||
467 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
468 | __copy_gigantic_page(dst, src, nr_pages); | |
469 | return; | |
470 | } | |
471 | } else { | |
472 | /* thp page */ | |
473 | BUG_ON(!PageTransHuge(src)); | |
474 | nr_pages = hpage_nr_pages(src); | |
475 | } | |
476 | ||
477 | for (i = 0; i < nr_pages; i++) { | |
478 | cond_resched(); | |
479 | copy_highpage(dst + i, src + i); | |
480 | } | |
481 | } | |
482 | ||
b20a3503 CL |
483 | /* |
484 | * Copy the page to its new location | |
485 | */ | |
290408d4 | 486 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 487 | { |
7851a45c RR |
488 | int cpupid; |
489 | ||
b32967ff | 490 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
491 | copy_huge_page(newpage, page); |
492 | else | |
493 | copy_highpage(newpage, page); | |
b20a3503 CL |
494 | |
495 | if (PageError(page)) | |
496 | SetPageError(newpage); | |
497 | if (PageReferenced(page)) | |
498 | SetPageReferenced(newpage); | |
499 | if (PageUptodate(page)) | |
500 | SetPageUptodate(newpage); | |
894bc310 LS |
501 | if (TestClearPageActive(page)) { |
502 | VM_BUG_ON(PageUnevictable(page)); | |
b20a3503 | 503 | SetPageActive(newpage); |
418b27ef LS |
504 | } else if (TestClearPageUnevictable(page)) |
505 | SetPageUnevictable(newpage); | |
b20a3503 CL |
506 | if (PageChecked(page)) |
507 | SetPageChecked(newpage); | |
508 | if (PageMappedToDisk(page)) | |
509 | SetPageMappedToDisk(newpage); | |
510 | ||
511 | if (PageDirty(page)) { | |
512 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
513 | /* |
514 | * Want to mark the page and the radix tree as dirty, and | |
515 | * redo the accounting that clear_page_dirty_for_io undid, | |
516 | * but we can't use set_page_dirty because that function | |
517 | * is actually a signal that all of the page has become dirty. | |
25985edc | 518 | * Whereas only part of our page may be dirty. |
3a902c5f | 519 | */ |
752dc185 HD |
520 | if (PageSwapBacked(page)) |
521 | SetPageDirty(newpage); | |
522 | else | |
523 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
524 | } |
525 | ||
7851a45c RR |
526 | /* |
527 | * Copy NUMA information to the new page, to prevent over-eager | |
528 | * future migrations of this same page. | |
529 | */ | |
530 | cpupid = page_cpupid_xchg_last(page, -1); | |
531 | page_cpupid_xchg_last(newpage, cpupid); | |
532 | ||
b291f000 | 533 | mlock_migrate_page(newpage, page); |
e9995ef9 | 534 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
535 | /* |
536 | * Please do not reorder this without considering how mm/ksm.c's | |
537 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
538 | */ | |
b20a3503 | 539 | ClearPageSwapCache(page); |
b20a3503 CL |
540 | ClearPagePrivate(page); |
541 | set_page_private(page, 0); | |
b20a3503 CL |
542 | |
543 | /* | |
544 | * If any waiters have accumulated on the new page then | |
545 | * wake them up. | |
546 | */ | |
547 | if (PageWriteback(newpage)) | |
548 | end_page_writeback(newpage); | |
549 | } | |
b20a3503 | 550 | |
1d8b85cc CL |
551 | /************************************************************ |
552 | * Migration functions | |
553 | ***********************************************************/ | |
554 | ||
b20a3503 CL |
555 | /* |
556 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 557 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
558 | * |
559 | * Pages are locked upon entry and exit. | |
560 | */ | |
2d1db3b1 | 561 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
562 | struct page *newpage, struct page *page, |
563 | enum migrate_mode mode) | |
b20a3503 CL |
564 | { |
565 | int rc; | |
566 | ||
567 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
568 | ||
8e321fef | 569 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 570 | |
78bd5209 | 571 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
572 | return rc; |
573 | ||
574 | migrate_page_copy(newpage, page); | |
78bd5209 | 575 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
576 | } |
577 | EXPORT_SYMBOL(migrate_page); | |
578 | ||
9361401e | 579 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
580 | /* |
581 | * Migration function for pages with buffers. This function can only be used | |
582 | * if the underlying filesystem guarantees that no other references to "page" | |
583 | * exist. | |
584 | */ | |
2d1db3b1 | 585 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 586 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 587 | { |
1d8b85cc CL |
588 | struct buffer_head *bh, *head; |
589 | int rc; | |
590 | ||
1d8b85cc | 591 | if (!page_has_buffers(page)) |
a6bc32b8 | 592 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
593 | |
594 | head = page_buffers(page); | |
595 | ||
8e321fef | 596 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 597 | |
78bd5209 | 598 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
599 | return rc; |
600 | ||
b969c4ab MG |
601 | /* |
602 | * In the async case, migrate_page_move_mapping locked the buffers | |
603 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
604 | * need to be locked now | |
605 | */ | |
a6bc32b8 MG |
606 | if (mode != MIGRATE_ASYNC) |
607 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
608 | |
609 | ClearPagePrivate(page); | |
610 | set_page_private(newpage, page_private(page)); | |
611 | set_page_private(page, 0); | |
612 | put_page(page); | |
613 | get_page(newpage); | |
614 | ||
615 | bh = head; | |
616 | do { | |
617 | set_bh_page(bh, newpage, bh_offset(bh)); | |
618 | bh = bh->b_this_page; | |
619 | ||
620 | } while (bh != head); | |
621 | ||
622 | SetPagePrivate(newpage); | |
623 | ||
624 | migrate_page_copy(newpage, page); | |
625 | ||
626 | bh = head; | |
627 | do { | |
628 | unlock_buffer(bh); | |
629 | put_bh(bh); | |
630 | bh = bh->b_this_page; | |
631 | ||
632 | } while (bh != head); | |
633 | ||
78bd5209 | 634 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
635 | } |
636 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 637 | #endif |
1d8b85cc | 638 | |
04e62a29 CL |
639 | /* |
640 | * Writeback a page to clean the dirty state | |
641 | */ | |
642 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 643 | { |
04e62a29 CL |
644 | struct writeback_control wbc = { |
645 | .sync_mode = WB_SYNC_NONE, | |
646 | .nr_to_write = 1, | |
647 | .range_start = 0, | |
648 | .range_end = LLONG_MAX, | |
04e62a29 CL |
649 | .for_reclaim = 1 |
650 | }; | |
651 | int rc; | |
652 | ||
653 | if (!mapping->a_ops->writepage) | |
654 | /* No write method for the address space */ | |
655 | return -EINVAL; | |
656 | ||
657 | if (!clear_page_dirty_for_io(page)) | |
658 | /* Someone else already triggered a write */ | |
659 | return -EAGAIN; | |
660 | ||
8351a6e4 | 661 | /* |
04e62a29 CL |
662 | * A dirty page may imply that the underlying filesystem has |
663 | * the page on some queue. So the page must be clean for | |
664 | * migration. Writeout may mean we loose the lock and the | |
665 | * page state is no longer what we checked for earlier. | |
666 | * At this point we know that the migration attempt cannot | |
667 | * be successful. | |
8351a6e4 | 668 | */ |
04e62a29 | 669 | remove_migration_ptes(page, page); |
8351a6e4 | 670 | |
04e62a29 | 671 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 672 | |
04e62a29 CL |
673 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
674 | /* unlocked. Relock */ | |
675 | lock_page(page); | |
676 | ||
bda8550d | 677 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
678 | } |
679 | ||
680 | /* | |
681 | * Default handling if a filesystem does not provide a migration function. | |
682 | */ | |
683 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 684 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 685 | { |
b969c4ab | 686 | if (PageDirty(page)) { |
a6bc32b8 MG |
687 | /* Only writeback pages in full synchronous migration */ |
688 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 689 | return -EBUSY; |
04e62a29 | 690 | return writeout(mapping, page); |
b969c4ab | 691 | } |
8351a6e4 CL |
692 | |
693 | /* | |
694 | * Buffers may be managed in a filesystem specific way. | |
695 | * We must have no buffers or drop them. | |
696 | */ | |
266cf658 | 697 | if (page_has_private(page) && |
8351a6e4 CL |
698 | !try_to_release_page(page, GFP_KERNEL)) |
699 | return -EAGAIN; | |
700 | ||
a6bc32b8 | 701 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
702 | } |
703 | ||
e24f0b8f CL |
704 | /* |
705 | * Move a page to a newly allocated page | |
706 | * The page is locked and all ptes have been successfully removed. | |
707 | * | |
708 | * The new page will have replaced the old page if this function | |
709 | * is successful. | |
894bc310 LS |
710 | * |
711 | * Return value: | |
712 | * < 0 - error code | |
78bd5209 | 713 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 714 | */ |
3fe2011f | 715 | static int move_to_new_page(struct page *newpage, struct page *page, |
a6bc32b8 | 716 | int remap_swapcache, enum migrate_mode mode) |
e24f0b8f CL |
717 | { |
718 | struct address_space *mapping; | |
719 | int rc; | |
720 | ||
721 | /* | |
722 | * Block others from accessing the page when we get around to | |
723 | * establishing additional references. We are the only one | |
724 | * holding a reference to the new page at this point. | |
725 | */ | |
529ae9aa | 726 | if (!trylock_page(newpage)) |
e24f0b8f CL |
727 | BUG(); |
728 | ||
729 | /* Prepare mapping for the new page.*/ | |
730 | newpage->index = page->index; | |
731 | newpage->mapping = page->mapping; | |
b2e18538 RR |
732 | if (PageSwapBacked(page)) |
733 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
734 | |
735 | mapping = page_mapping(page); | |
736 | if (!mapping) | |
a6bc32b8 | 737 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 738 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 739 | /* |
b969c4ab MG |
740 | * Most pages have a mapping and most filesystems provide a |
741 | * migratepage callback. Anonymous pages are part of swap | |
742 | * space which also has its own migratepage callback. This | |
743 | * is the most common path for page migration. | |
e24f0b8f | 744 | */ |
b969c4ab | 745 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 746 | newpage, page, mode); |
b969c4ab | 747 | else |
a6bc32b8 | 748 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 749 | |
78bd5209 | 750 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 751 | newpage->mapping = NULL; |
3fe2011f MG |
752 | } else { |
753 | if (remap_swapcache) | |
754 | remove_migration_ptes(page, newpage); | |
35512eca | 755 | page->mapping = NULL; |
3fe2011f | 756 | } |
e24f0b8f CL |
757 | |
758 | unlock_page(newpage); | |
759 | ||
760 | return rc; | |
761 | } | |
762 | ||
0dabec93 | 763 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 764 | int force, enum migrate_mode mode) |
e24f0b8f | 765 | { |
0dabec93 | 766 | int rc = -EAGAIN; |
3fe2011f | 767 | int remap_swapcache = 1; |
56039efa | 768 | struct mem_cgroup *mem; |
3f6c8272 | 769 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 770 | |
529ae9aa | 771 | if (!trylock_page(page)) { |
a6bc32b8 | 772 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 773 | goto out; |
3e7d3449 MG |
774 | |
775 | /* | |
776 | * It's not safe for direct compaction to call lock_page. | |
777 | * For example, during page readahead pages are added locked | |
778 | * to the LRU. Later, when the IO completes the pages are | |
779 | * marked uptodate and unlocked. However, the queueing | |
780 | * could be merging multiple pages for one bio (e.g. | |
781 | * mpage_readpages). If an allocation happens for the | |
782 | * second or third page, the process can end up locking | |
783 | * the same page twice and deadlocking. Rather than | |
784 | * trying to be clever about what pages can be locked, | |
785 | * avoid the use of lock_page for direct compaction | |
786 | * altogether. | |
787 | */ | |
788 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 789 | goto out; |
3e7d3449 | 790 | |
e24f0b8f CL |
791 | lock_page(page); |
792 | } | |
793 | ||
01b1ae63 | 794 | /* charge against new page */ |
0030f535 | 795 | mem_cgroup_prepare_migration(page, newpage, &mem); |
01b1ae63 | 796 | |
e24f0b8f | 797 | if (PageWriteback(page)) { |
11bc82d6 | 798 | /* |
fed5b64a | 799 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
800 | * necessary to wait for PageWriteback. In the async case, |
801 | * the retry loop is too short and in the sync-light case, | |
802 | * the overhead of stalling is too much | |
11bc82d6 | 803 | */ |
a6bc32b8 | 804 | if (mode != MIGRATE_SYNC) { |
11bc82d6 AA |
805 | rc = -EBUSY; |
806 | goto uncharge; | |
807 | } | |
808 | if (!force) | |
01b1ae63 | 809 | goto uncharge; |
e24f0b8f CL |
810 | wait_on_page_writeback(page); |
811 | } | |
e24f0b8f | 812 | /* |
dc386d4d KH |
813 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
814 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 815 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 816 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
817 | * File Caches may use write_page() or lock_page() in migration, then, |
818 | * just care Anon page here. | |
dc386d4d | 819 | */ |
b79bc0a0 | 820 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 821 | /* |
4fc3f1d6 | 822 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
823 | * getting a hold on an anon_vma from outside one of its mms. |
824 | */ | |
746b18d4 | 825 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
826 | if (anon_vma) { |
827 | /* | |
746b18d4 | 828 | * Anon page |
1ce82b69 | 829 | */ |
1ce82b69 | 830 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
831 | /* |
832 | * We cannot be sure that the anon_vma of an unmapped | |
833 | * swapcache page is safe to use because we don't | |
834 | * know in advance if the VMA that this page belonged | |
835 | * to still exists. If the VMA and others sharing the | |
836 | * data have been freed, then the anon_vma could | |
837 | * already be invalid. | |
838 | * | |
839 | * To avoid this possibility, swapcache pages get | |
840 | * migrated but are not remapped when migration | |
841 | * completes | |
842 | */ | |
843 | remap_swapcache = 0; | |
844 | } else { | |
1ce82b69 | 845 | goto uncharge; |
3fe2011f | 846 | } |
989f89c5 | 847 | } |
62e1c553 | 848 | |
bf6bddf1 RA |
849 | if (unlikely(balloon_page_movable(page))) { |
850 | /* | |
851 | * A ballooned page does not need any special attention from | |
852 | * physical to virtual reverse mapping procedures. | |
853 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
854 | * in order to avoid burning cycles at rmap level, and perform | |
855 | * the page migration right away (proteced by page lock). | |
856 | */ | |
857 | rc = balloon_page_migrate(newpage, page, mode); | |
858 | goto uncharge; | |
859 | } | |
860 | ||
dc386d4d | 861 | /* |
62e1c553 SL |
862 | * Corner case handling: |
863 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
864 | * and treated as swapcache but it has no rmap yet. | |
865 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
866 | * trigger a BUG. So handle it here. | |
867 | * 2. An orphaned page (see truncate_complete_page) might have | |
868 | * fs-private metadata. The page can be picked up due to memory | |
869 | * offlining. Everywhere else except page reclaim, the page is | |
870 | * invisible to the vm, so the page can not be migrated. So try to | |
871 | * free the metadata, so the page can be freed. | |
e24f0b8f | 872 | */ |
62e1c553 | 873 | if (!page->mapping) { |
1ce82b69 HD |
874 | VM_BUG_ON(PageAnon(page)); |
875 | if (page_has_private(page)) { | |
62e1c553 | 876 | try_to_free_buffers(page); |
1ce82b69 | 877 | goto uncharge; |
62e1c553 | 878 | } |
abfc3488 | 879 | goto skip_unmap; |
62e1c553 SL |
880 | } |
881 | ||
dc386d4d | 882 | /* Establish migration ptes or remove ptes */ |
14fa31b8 | 883 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
dc386d4d | 884 | |
abfc3488 | 885 | skip_unmap: |
e6a1530d | 886 | if (!page_mapped(page)) |
a6bc32b8 | 887 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); |
e24f0b8f | 888 | |
3fe2011f | 889 | if (rc && remap_swapcache) |
e24f0b8f | 890 | remove_migration_ptes(page, page); |
3f6c8272 MG |
891 | |
892 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 893 | if (anon_vma) |
9e60109f | 894 | put_anon_vma(anon_vma); |
3f6c8272 | 895 | |
01b1ae63 | 896 | uncharge: |
bf6bddf1 RA |
897 | mem_cgroup_end_migration(mem, page, newpage, |
898 | (rc == MIGRATEPAGE_SUCCESS || | |
899 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
e24f0b8f | 900 | unlock_page(page); |
0dabec93 MK |
901 | out: |
902 | return rc; | |
903 | } | |
95a402c3 | 904 | |
0dabec93 MK |
905 | /* |
906 | * Obtain the lock on page, remove all ptes and migrate the page | |
907 | * to the newly allocated page in newpage. | |
908 | */ | |
909 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
9c620e2b | 910 | struct page *page, int force, enum migrate_mode mode) |
0dabec93 MK |
911 | { |
912 | int rc = 0; | |
913 | int *result = NULL; | |
914 | struct page *newpage = get_new_page(page, private, &result); | |
915 | ||
916 | if (!newpage) | |
917 | return -ENOMEM; | |
918 | ||
919 | if (page_count(page) == 1) { | |
920 | /* page was freed from under us. So we are done. */ | |
921 | goto out; | |
922 | } | |
923 | ||
924 | if (unlikely(PageTransHuge(page))) | |
925 | if (unlikely(split_huge_page(page))) | |
926 | goto out; | |
927 | ||
9c620e2b | 928 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 RA |
929 | |
930 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
931 | /* | |
932 | * A ballooned page has been migrated already. | |
933 | * Now, it's the time to wrap-up counters, | |
934 | * handle the page back to Buddy and return. | |
935 | */ | |
936 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
937 | page_is_file_cache(page)); | |
938 | balloon_page_free(page); | |
939 | return MIGRATEPAGE_SUCCESS; | |
940 | } | |
0dabec93 | 941 | out: |
e24f0b8f | 942 | if (rc != -EAGAIN) { |
0dabec93 MK |
943 | /* |
944 | * A page that has been migrated has all references | |
945 | * removed and will be freed. A page that has not been | |
946 | * migrated will have kepts its references and be | |
947 | * restored. | |
948 | */ | |
949 | list_del(&page->lru); | |
a731286d | 950 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 951 | page_is_file_cache(page)); |
894bc310 | 952 | putback_lru_page(page); |
e24f0b8f | 953 | } |
95a402c3 CL |
954 | /* |
955 | * Move the new page to the LRU. If migration was not successful | |
956 | * then this will free the page. | |
957 | */ | |
894bc310 | 958 | putback_lru_page(newpage); |
742755a1 CL |
959 | if (result) { |
960 | if (rc) | |
961 | *result = rc; | |
962 | else | |
963 | *result = page_to_nid(newpage); | |
964 | } | |
e24f0b8f CL |
965 | return rc; |
966 | } | |
967 | ||
290408d4 NH |
968 | /* |
969 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
970 | * | |
971 | * This function doesn't wait the completion of hugepage I/O | |
972 | * because there is no race between I/O and migration for hugepage. | |
973 | * Note that currently hugepage I/O occurs only in direct I/O | |
974 | * where no lock is held and PG_writeback is irrelevant, | |
975 | * and writeback status of all subpages are counted in the reference | |
976 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
977 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
978 | * This means that when we try to migrate hugepage whose subpages are | |
979 | * doing direct I/O, some references remain after try_to_unmap() and | |
980 | * hugepage migration fails without data corruption. | |
981 | * | |
982 | * There is also no race when direct I/O is issued on the page under migration, | |
983 | * because then pte is replaced with migration swap entry and direct I/O code | |
984 | * will wait in the page fault for migration to complete. | |
985 | */ | |
986 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
987 | unsigned long private, struct page *hpage, | |
9c620e2b | 988 | int force, enum migrate_mode mode) |
290408d4 NH |
989 | { |
990 | int rc = 0; | |
991 | int *result = NULL; | |
32665f2b | 992 | struct page *new_hpage; |
290408d4 NH |
993 | struct anon_vma *anon_vma = NULL; |
994 | ||
83467efb NH |
995 | /* |
996 | * Movability of hugepages depends on architectures and hugepage size. | |
997 | * This check is necessary because some callers of hugepage migration | |
998 | * like soft offline and memory hotremove don't walk through page | |
999 | * tables or check whether the hugepage is pmd-based or not before | |
1000 | * kicking migration. | |
1001 | */ | |
32665f2b JK |
1002 | if (!hugepage_migration_support(page_hstate(hpage))) { |
1003 | putback_active_hugepage(hpage); | |
83467efb | 1004 | return -ENOSYS; |
32665f2b | 1005 | } |
83467efb | 1006 | |
32665f2b | 1007 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1008 | if (!new_hpage) |
1009 | return -ENOMEM; | |
1010 | ||
1011 | rc = -EAGAIN; | |
1012 | ||
1013 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 1014 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1015 | goto out; |
1016 | lock_page(hpage); | |
1017 | } | |
1018 | ||
746b18d4 PZ |
1019 | if (PageAnon(hpage)) |
1020 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 NH |
1021 | |
1022 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1023 | ||
1024 | if (!page_mapped(hpage)) | |
a6bc32b8 | 1025 | rc = move_to_new_page(new_hpage, hpage, 1, mode); |
290408d4 NH |
1026 | |
1027 | if (rc) | |
1028 | remove_migration_ptes(hpage, hpage); | |
1029 | ||
fd4a4663 | 1030 | if (anon_vma) |
9e60109f | 1031 | put_anon_vma(anon_vma); |
8e6ac7fa AK |
1032 | |
1033 | if (!rc) | |
1034 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1035 | ||
290408d4 | 1036 | unlock_page(hpage); |
09761333 | 1037 | out: |
b8ec1cee NH |
1038 | if (rc != -EAGAIN) |
1039 | putback_active_hugepage(hpage); | |
290408d4 | 1040 | put_page(new_hpage); |
290408d4 NH |
1041 | if (result) { |
1042 | if (rc) | |
1043 | *result = rc; | |
1044 | else | |
1045 | *result = page_to_nid(new_hpage); | |
1046 | } | |
1047 | return rc; | |
1048 | } | |
1049 | ||
b20a3503 | 1050 | /* |
c73e5c9c SB |
1051 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1052 | * supplied as the target for the page migration | |
b20a3503 | 1053 | * |
c73e5c9c SB |
1054 | * @from: The list of pages to be migrated. |
1055 | * @get_new_page: The function used to allocate free pages to be used | |
1056 | * as the target of the page migration. | |
1057 | * @private: Private data to be passed on to get_new_page() | |
1058 | * @mode: The migration mode that specifies the constraints for | |
1059 | * page migration, if any. | |
1060 | * @reason: The reason for page migration. | |
b20a3503 | 1061 | * |
c73e5c9c SB |
1062 | * The function returns after 10 attempts or if no pages are movable any more |
1063 | * because the list has become empty or no retryable pages exist any more. | |
1064 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1065 | * or free list only if ret != 0. |
b20a3503 | 1066 | * |
c73e5c9c | 1067 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1068 | */ |
9c620e2b HD |
1069 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
1070 | unsigned long private, enum migrate_mode mode, int reason) | |
b20a3503 | 1071 | { |
e24f0b8f | 1072 | int retry = 1; |
b20a3503 | 1073 | int nr_failed = 0; |
5647bc29 | 1074 | int nr_succeeded = 0; |
b20a3503 CL |
1075 | int pass = 0; |
1076 | struct page *page; | |
1077 | struct page *page2; | |
1078 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1079 | int rc; | |
1080 | ||
1081 | if (!swapwrite) | |
1082 | current->flags |= PF_SWAPWRITE; | |
1083 | ||
e24f0b8f CL |
1084 | for(pass = 0; pass < 10 && retry; pass++) { |
1085 | retry = 0; | |
b20a3503 | 1086 | |
e24f0b8f | 1087 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1088 | cond_resched(); |
2d1db3b1 | 1089 | |
31caf665 NH |
1090 | if (PageHuge(page)) |
1091 | rc = unmap_and_move_huge_page(get_new_page, | |
1092 | private, page, pass > 2, mode); | |
1093 | else | |
1094 | rc = unmap_and_move(get_new_page, private, | |
9c620e2b | 1095 | page, pass > 2, mode); |
2d1db3b1 | 1096 | |
e24f0b8f | 1097 | switch(rc) { |
95a402c3 CL |
1098 | case -ENOMEM: |
1099 | goto out; | |
e24f0b8f | 1100 | case -EAGAIN: |
2d1db3b1 | 1101 | retry++; |
e24f0b8f | 1102 | break; |
78bd5209 | 1103 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1104 | nr_succeeded++; |
e24f0b8f CL |
1105 | break; |
1106 | default: | |
354a3363 NH |
1107 | /* |
1108 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1109 | * unlike -EAGAIN case, the failed page is | |
1110 | * removed from migration page list and not | |
1111 | * retried in the next outer loop. | |
1112 | */ | |
2d1db3b1 | 1113 | nr_failed++; |
e24f0b8f | 1114 | break; |
2d1db3b1 | 1115 | } |
b20a3503 CL |
1116 | } |
1117 | } | |
78bd5209 | 1118 | rc = nr_failed + retry; |
95a402c3 | 1119 | out: |
5647bc29 MG |
1120 | if (nr_succeeded) |
1121 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1122 | if (nr_failed) | |
1123 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1124 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1125 | ||
b20a3503 CL |
1126 | if (!swapwrite) |
1127 | current->flags &= ~PF_SWAPWRITE; | |
1128 | ||
78bd5209 | 1129 | return rc; |
b20a3503 | 1130 | } |
95a402c3 | 1131 | |
742755a1 CL |
1132 | #ifdef CONFIG_NUMA |
1133 | /* | |
1134 | * Move a list of individual pages | |
1135 | */ | |
1136 | struct page_to_node { | |
1137 | unsigned long addr; | |
1138 | struct page *page; | |
1139 | int node; | |
1140 | int status; | |
1141 | }; | |
1142 | ||
1143 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1144 | int **result) | |
1145 | { | |
1146 | struct page_to_node *pm = (struct page_to_node *)private; | |
1147 | ||
1148 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1149 | pm++; | |
1150 | ||
1151 | if (pm->node == MAX_NUMNODES) | |
1152 | return NULL; | |
1153 | ||
1154 | *result = &pm->status; | |
1155 | ||
e632a938 NH |
1156 | if (PageHuge(p)) |
1157 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1158 | pm->node); | |
1159 | else | |
1160 | return alloc_pages_exact_node(pm->node, | |
769848c0 | 1161 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); |
742755a1 CL |
1162 | } |
1163 | ||
1164 | /* | |
1165 | * Move a set of pages as indicated in the pm array. The addr | |
1166 | * field must be set to the virtual address of the page to be moved | |
1167 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1168 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1169 | */ |
5e9a0f02 BG |
1170 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1171 | struct page_to_node *pm, | |
1172 | int migrate_all) | |
742755a1 CL |
1173 | { |
1174 | int err; | |
1175 | struct page_to_node *pp; | |
1176 | LIST_HEAD(pagelist); | |
1177 | ||
1178 | down_read(&mm->mmap_sem); | |
1179 | ||
1180 | /* | |
1181 | * Build a list of pages to migrate | |
1182 | */ | |
742755a1 CL |
1183 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1184 | struct vm_area_struct *vma; | |
1185 | struct page *page; | |
1186 | ||
742755a1 CL |
1187 | err = -EFAULT; |
1188 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1189 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1190 | goto set_status; |
1191 | ||
500d65d4 | 1192 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1193 | |
1194 | err = PTR_ERR(page); | |
1195 | if (IS_ERR(page)) | |
1196 | goto set_status; | |
1197 | ||
742755a1 CL |
1198 | err = -ENOENT; |
1199 | if (!page) | |
1200 | goto set_status; | |
1201 | ||
62b61f61 | 1202 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1203 | if (PageReserved(page)) |
742755a1 CL |
1204 | goto put_and_set; |
1205 | ||
1206 | pp->page = page; | |
1207 | err = page_to_nid(page); | |
1208 | ||
1209 | if (err == pp->node) | |
1210 | /* | |
1211 | * Node already in the right place | |
1212 | */ | |
1213 | goto put_and_set; | |
1214 | ||
1215 | err = -EACCES; | |
1216 | if (page_mapcount(page) > 1 && | |
1217 | !migrate_all) | |
1218 | goto put_and_set; | |
1219 | ||
e632a938 NH |
1220 | if (PageHuge(page)) { |
1221 | isolate_huge_page(page, &pagelist); | |
1222 | goto put_and_set; | |
1223 | } | |
1224 | ||
62695a84 | 1225 | err = isolate_lru_page(page); |
6d9c285a | 1226 | if (!err) { |
62695a84 | 1227 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1228 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1229 | page_is_file_cache(page)); | |
1230 | } | |
742755a1 CL |
1231 | put_and_set: |
1232 | /* | |
1233 | * Either remove the duplicate refcount from | |
1234 | * isolate_lru_page() or drop the page ref if it was | |
1235 | * not isolated. | |
1236 | */ | |
1237 | put_page(page); | |
1238 | set_status: | |
1239 | pp->status = err; | |
1240 | } | |
1241 | ||
e78bbfa8 | 1242 | err = 0; |
cf608ac1 | 1243 | if (!list_empty(&pagelist)) { |
742755a1 | 1244 | err = migrate_pages(&pagelist, new_page_node, |
9c620e2b | 1245 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1246 | if (err) |
e632a938 | 1247 | putback_movable_pages(&pagelist); |
cf608ac1 | 1248 | } |
742755a1 CL |
1249 | |
1250 | up_read(&mm->mmap_sem); | |
1251 | return err; | |
1252 | } | |
1253 | ||
5e9a0f02 BG |
1254 | /* |
1255 | * Migrate an array of page address onto an array of nodes and fill | |
1256 | * the corresponding array of status. | |
1257 | */ | |
3268c63e | 1258 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1259 | unsigned long nr_pages, |
1260 | const void __user * __user *pages, | |
1261 | const int __user *nodes, | |
1262 | int __user *status, int flags) | |
1263 | { | |
3140a227 | 1264 | struct page_to_node *pm; |
3140a227 BG |
1265 | unsigned long chunk_nr_pages; |
1266 | unsigned long chunk_start; | |
1267 | int err; | |
5e9a0f02 | 1268 | |
3140a227 BG |
1269 | err = -ENOMEM; |
1270 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1271 | if (!pm) | |
5e9a0f02 | 1272 | goto out; |
35282a2d BG |
1273 | |
1274 | migrate_prep(); | |
1275 | ||
5e9a0f02 | 1276 | /* |
3140a227 BG |
1277 | * Store a chunk of page_to_node array in a page, |
1278 | * but keep the last one as a marker | |
5e9a0f02 | 1279 | */ |
3140a227 | 1280 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1281 | |
3140a227 BG |
1282 | for (chunk_start = 0; |
1283 | chunk_start < nr_pages; | |
1284 | chunk_start += chunk_nr_pages) { | |
1285 | int j; | |
5e9a0f02 | 1286 | |
3140a227 BG |
1287 | if (chunk_start + chunk_nr_pages > nr_pages) |
1288 | chunk_nr_pages = nr_pages - chunk_start; | |
1289 | ||
1290 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1291 | for (j = 0; j < chunk_nr_pages; j++) { | |
1292 | const void __user *p; | |
5e9a0f02 BG |
1293 | int node; |
1294 | ||
3140a227 BG |
1295 | err = -EFAULT; |
1296 | if (get_user(p, pages + j + chunk_start)) | |
1297 | goto out_pm; | |
1298 | pm[j].addr = (unsigned long) p; | |
1299 | ||
1300 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1301 | goto out_pm; |
1302 | ||
1303 | err = -ENODEV; | |
6f5a55f1 LT |
1304 | if (node < 0 || node >= MAX_NUMNODES) |
1305 | goto out_pm; | |
1306 | ||
389162c2 | 1307 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1308 | goto out_pm; |
1309 | ||
1310 | err = -EACCES; | |
1311 | if (!node_isset(node, task_nodes)) | |
1312 | goto out_pm; | |
1313 | ||
3140a227 BG |
1314 | pm[j].node = node; |
1315 | } | |
1316 | ||
1317 | /* End marker for this chunk */ | |
1318 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1319 | ||
1320 | /* Migrate this chunk */ | |
1321 | err = do_move_page_to_node_array(mm, pm, | |
1322 | flags & MPOL_MF_MOVE_ALL); | |
1323 | if (err < 0) | |
1324 | goto out_pm; | |
5e9a0f02 | 1325 | |
5e9a0f02 | 1326 | /* Return status information */ |
3140a227 BG |
1327 | for (j = 0; j < chunk_nr_pages; j++) |
1328 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1329 | err = -EFAULT; |
3140a227 BG |
1330 | goto out_pm; |
1331 | } | |
1332 | } | |
1333 | err = 0; | |
5e9a0f02 BG |
1334 | |
1335 | out_pm: | |
3140a227 | 1336 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1337 | out: |
1338 | return err; | |
1339 | } | |
1340 | ||
742755a1 | 1341 | /* |
2f007e74 | 1342 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1343 | */ |
80bba129 BG |
1344 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1345 | const void __user **pages, int *status) | |
742755a1 | 1346 | { |
2f007e74 | 1347 | unsigned long i; |
2f007e74 | 1348 | |
742755a1 CL |
1349 | down_read(&mm->mmap_sem); |
1350 | ||
2f007e74 | 1351 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1352 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1353 | struct vm_area_struct *vma; |
1354 | struct page *page; | |
c095adbc | 1355 | int err = -EFAULT; |
2f007e74 BG |
1356 | |
1357 | vma = find_vma(mm, addr); | |
70384dc6 | 1358 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1359 | goto set_status; |
1360 | ||
2f007e74 | 1361 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1362 | |
1363 | err = PTR_ERR(page); | |
1364 | if (IS_ERR(page)) | |
1365 | goto set_status; | |
1366 | ||
742755a1 CL |
1367 | err = -ENOENT; |
1368 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1369 | if (!page || PageReserved(page)) |
742755a1 CL |
1370 | goto set_status; |
1371 | ||
1372 | err = page_to_nid(page); | |
1373 | set_status: | |
80bba129 BG |
1374 | *status = err; |
1375 | ||
1376 | pages++; | |
1377 | status++; | |
1378 | } | |
1379 | ||
1380 | up_read(&mm->mmap_sem); | |
1381 | } | |
1382 | ||
1383 | /* | |
1384 | * Determine the nodes of a user array of pages and store it in | |
1385 | * a user array of status. | |
1386 | */ | |
1387 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1388 | const void __user * __user *pages, | |
1389 | int __user *status) | |
1390 | { | |
1391 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1392 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1393 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1394 | |
87b8d1ad PA |
1395 | while (nr_pages) { |
1396 | unsigned long chunk_nr; | |
80bba129 | 1397 | |
87b8d1ad PA |
1398 | chunk_nr = nr_pages; |
1399 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1400 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1401 | ||
1402 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1403 | break; | |
80bba129 BG |
1404 | |
1405 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1406 | ||
87b8d1ad PA |
1407 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1408 | break; | |
742755a1 | 1409 | |
87b8d1ad PA |
1410 | pages += chunk_nr; |
1411 | status += chunk_nr; | |
1412 | nr_pages -= chunk_nr; | |
1413 | } | |
1414 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1415 | } |
1416 | ||
1417 | /* | |
1418 | * Move a list of pages in the address space of the currently executing | |
1419 | * process. | |
1420 | */ | |
938bb9f5 HC |
1421 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1422 | const void __user * __user *, pages, | |
1423 | const int __user *, nodes, | |
1424 | int __user *, status, int, flags) | |
742755a1 | 1425 | { |
c69e8d9c | 1426 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1427 | struct task_struct *task; |
742755a1 | 1428 | struct mm_struct *mm; |
5e9a0f02 | 1429 | int err; |
3268c63e | 1430 | nodemask_t task_nodes; |
742755a1 CL |
1431 | |
1432 | /* Check flags */ | |
1433 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1434 | return -EINVAL; | |
1435 | ||
1436 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1437 | return -EPERM; | |
1438 | ||
1439 | /* Find the mm_struct */ | |
a879bf58 | 1440 | rcu_read_lock(); |
228ebcbe | 1441 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1442 | if (!task) { |
a879bf58 | 1443 | rcu_read_unlock(); |
742755a1 CL |
1444 | return -ESRCH; |
1445 | } | |
3268c63e | 1446 | get_task_struct(task); |
742755a1 CL |
1447 | |
1448 | /* | |
1449 | * Check if this process has the right to modify the specified | |
1450 | * process. The right exists if the process has administrative | |
1451 | * capabilities, superuser privileges or the same | |
1452 | * userid as the target process. | |
1453 | */ | |
c69e8d9c | 1454 | tcred = __task_cred(task); |
b38a86eb EB |
1455 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1456 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1457 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1458 | rcu_read_unlock(); |
742755a1 | 1459 | err = -EPERM; |
5e9a0f02 | 1460 | goto out; |
742755a1 | 1461 | } |
c69e8d9c | 1462 | rcu_read_unlock(); |
742755a1 | 1463 | |
86c3a764 DQ |
1464 | err = security_task_movememory(task); |
1465 | if (err) | |
5e9a0f02 | 1466 | goto out; |
86c3a764 | 1467 | |
3268c63e CL |
1468 | task_nodes = cpuset_mems_allowed(task); |
1469 | mm = get_task_mm(task); | |
1470 | put_task_struct(task); | |
1471 | ||
6e8b09ea SL |
1472 | if (!mm) |
1473 | return -EINVAL; | |
1474 | ||
1475 | if (nodes) | |
1476 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1477 | nodes, status, flags); | |
1478 | else | |
1479 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1480 | |
742755a1 CL |
1481 | mmput(mm); |
1482 | return err; | |
3268c63e CL |
1483 | |
1484 | out: | |
1485 | put_task_struct(task); | |
1486 | return err; | |
742755a1 | 1487 | } |
742755a1 | 1488 | |
7b2259b3 CL |
1489 | /* |
1490 | * Call migration functions in the vma_ops that may prepare | |
1491 | * memory in a vm for migration. migration functions may perform | |
1492 | * the migration for vmas that do not have an underlying page struct. | |
1493 | */ | |
1494 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1495 | const nodemask_t *from, unsigned long flags) | |
1496 | { | |
1497 | struct vm_area_struct *vma; | |
1498 | int err = 0; | |
1499 | ||
1001c9fb | 1500 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { |
7b2259b3 CL |
1501 | if (vma->vm_ops && vma->vm_ops->migrate) { |
1502 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1503 | if (err) | |
1504 | break; | |
1505 | } | |
1506 | } | |
1507 | return err; | |
1508 | } | |
7039e1db PZ |
1509 | |
1510 | #ifdef CONFIG_NUMA_BALANCING | |
1511 | /* | |
1512 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1513 | * pages. Currently it only checks the watermarks which crude | |
1514 | */ | |
1515 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1516 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1517 | { |
1518 | int z; | |
1519 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1520 | struct zone *zone = pgdat->node_zones + z; | |
1521 | ||
1522 | if (!populated_zone(zone)) | |
1523 | continue; | |
1524 | ||
6e543d57 | 1525 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1526 | continue; |
1527 | ||
1528 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1529 | if (!zone_watermark_ok(zone, 0, | |
1530 | high_wmark_pages(zone) + | |
1531 | nr_migrate_pages, | |
1532 | 0, 0)) | |
1533 | continue; | |
1534 | return true; | |
1535 | } | |
1536 | return false; | |
1537 | } | |
1538 | ||
1539 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1540 | unsigned long data, | |
1541 | int **result) | |
1542 | { | |
1543 | int nid = (int) data; | |
1544 | struct page *newpage; | |
1545 | ||
1546 | newpage = alloc_pages_exact_node(nid, | |
1547 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1548 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1549 | __GFP_NOWARN) & | |
1550 | ~GFP_IOFS, 0); | |
bac0382c | 1551 | if (newpage) |
90572890 | 1552 | page_cpupid_xchg_last(newpage, page_cpupid_last(page)); |
bac0382c | 1553 | |
7039e1db PZ |
1554 | return newpage; |
1555 | } | |
1556 | ||
a8f60772 MG |
1557 | /* |
1558 | * page migration rate limiting control. | |
1559 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1560 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1561 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1562 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1563 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1564 | * throttle window closed. | |
a8f60772 MG |
1565 | */ |
1566 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1567 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1568 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1569 | ||
e14808b4 MG |
1570 | /* Returns true if NUMA migration is currently rate limited */ |
1571 | bool migrate_ratelimited(int node) | |
1572 | { | |
1573 | pg_data_t *pgdat = NODE_DATA(node); | |
1574 | ||
1575 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1576 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1577 | return false; | |
1578 | ||
1579 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1580 | return false; | |
1581 | ||
1582 | return true; | |
1583 | } | |
1584 | ||
b32967ff | 1585 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1586 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1587 | unsigned long nr_pages) | |
7039e1db | 1588 | { |
a8f60772 MG |
1589 | /* |
1590 | * Rate-limit the amount of data that is being migrated to a node. | |
1591 | * Optimal placement is no good if the memory bus is saturated and | |
1592 | * all the time is being spent migrating! | |
1593 | */ | |
a8f60772 | 1594 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1595 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1596 | pgdat->numabalancing_migrate_nr_pages = 0; |
1597 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1598 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1599 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1600 | } |
af1839d7 MG |
1601 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1602 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1603 | nr_pages); | |
1c5e9c27 | 1604 | return true; |
af1839d7 | 1605 | } |
1c5e9c27 MG |
1606 | |
1607 | /* | |
1608 | * This is an unlocked non-atomic update so errors are possible. | |
1609 | * The consequences are failing to migrate when we potentiall should | |
1610 | * have which is not severe enough to warrant locking. If it is ever | |
1611 | * a problem, it can be converted to a per-cpu counter. | |
1612 | */ | |
1613 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1614 | return false; | |
b32967ff MG |
1615 | } |
1616 | ||
1c30e017 | 1617 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1618 | { |
340ef390 | 1619 | int page_lru; |
a8f60772 | 1620 | |
3abef4e6 MG |
1621 | VM_BUG_ON(compound_order(page) && !PageTransHuge(page)); |
1622 | ||
7039e1db | 1623 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1624 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1625 | return 0; | |
7039e1db | 1626 | |
340ef390 HD |
1627 | if (isolate_lru_page(page)) |
1628 | return 0; | |
7039e1db | 1629 | |
340ef390 HD |
1630 | /* |
1631 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1632 | * check on page_count(), so we must do it here, now that the page | |
1633 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1634 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1635 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1636 | */ | |
1637 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1638 | putback_lru_page(page); | |
1639 | return 0; | |
7039e1db PZ |
1640 | } |
1641 | ||
340ef390 HD |
1642 | page_lru = page_is_file_cache(page); |
1643 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1644 | hpage_nr_pages(page)); | |
1645 | ||
149c33e1 | 1646 | /* |
340ef390 HD |
1647 | * Isolating the page has taken another reference, so the |
1648 | * caller's reference can be safely dropped without the page | |
1649 | * disappearing underneath us during migration. | |
149c33e1 MG |
1650 | */ |
1651 | put_page(page); | |
340ef390 | 1652 | return 1; |
b32967ff MG |
1653 | } |
1654 | ||
de466bd6 MG |
1655 | bool pmd_trans_migrating(pmd_t pmd) |
1656 | { | |
1657 | struct page *page = pmd_page(pmd); | |
1658 | return PageLocked(page); | |
1659 | } | |
1660 | ||
1661 | void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd) | |
1662 | { | |
1663 | struct page *page = pmd_page(*pmd); | |
1664 | wait_on_page_locked(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 HD |
1701 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
1702 | node, MIGRATE_ASYNC, MR_NUMA_MISPLACED); | |
b32967ff | 1703 | if (nr_remaining) { |
59c82b70 JK |
1704 | if (!list_empty(&migratepages)) { |
1705 | list_del(&page->lru); | |
1706 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1707 | page_is_file_cache(page)); | |
1708 | putback_lru_page(page); | |
1709 | } | |
b32967ff MG |
1710 | isolated = 0; |
1711 | } else | |
1712 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1713 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1714 | return isolated; |
340ef390 HD |
1715 | |
1716 | out: | |
1717 | put_page(page); | |
1718 | return 0; | |
7039e1db | 1719 | } |
220018d3 | 1720 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1721 | |
220018d3 | 1722 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1723 | /* |
1724 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1725 | * before returning. | |
1726 | */ | |
b32967ff MG |
1727 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1728 | struct vm_area_struct *vma, | |
1729 | pmd_t *pmd, pmd_t entry, | |
1730 | unsigned long address, | |
1731 | struct page *page, int node) | |
1732 | { | |
c4088ebd | 1733 | spinlock_t *ptl; |
b32967ff MG |
1734 | pg_data_t *pgdat = NODE_DATA(node); |
1735 | int isolated = 0; | |
1736 | struct page *new_page = NULL; | |
1737 | struct mem_cgroup *memcg = NULL; | |
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, | |
1752 | (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER); | |
340ef390 HD |
1753 | if (!new_page) |
1754 | goto out_fail; | |
1755 | ||
90572890 | 1756 | page_cpupid_xchg_last(new_page, page_cpupid_last(page)); |
b32967ff MG |
1757 | |
1758 | isolated = numamigrate_isolate_page(pgdat, page); | |
340ef390 | 1759 | if (!isolated) { |
b32967ff | 1760 | put_page(new_page); |
340ef390 | 1761 | goto out_fail; |
b32967ff MG |
1762 | } |
1763 | ||
b0943d61 MG |
1764 | if (mm_tlb_flush_pending(mm)) |
1765 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1766 | ||
b32967ff MG |
1767 | /* Prepare a page as a migration target */ |
1768 | __set_page_locked(new_page); | |
1769 | SetPageSwapBacked(new_page); | |
1770 | ||
1771 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1772 | new_page->mapping = page->mapping; | |
1773 | new_page->index = page->index; | |
1774 | migrate_page_copy(new_page, page); | |
1775 | WARN_ON(PageLRU(new_page)); | |
1776 | ||
1777 | /* Recheck the target PMD */ | |
f714f4f2 | 1778 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 1779 | ptl = pmd_lock(mm, pmd); |
2b4847e7 MG |
1780 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { |
1781 | fail_putback: | |
c4088ebd | 1782 | spin_unlock(ptl); |
f714f4f2 | 1783 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1784 | |
1785 | /* Reverse changes made by migrate_page_copy() */ | |
1786 | if (TestClearPageActive(new_page)) | |
1787 | SetPageActive(page); | |
1788 | if (TestClearPageUnevictable(new_page)) | |
1789 | SetPageUnevictable(page); | |
1790 | mlock_migrate_page(page, new_page); | |
1791 | ||
1792 | unlock_page(new_page); | |
1793 | put_page(new_page); /* Free it */ | |
1794 | ||
a54a407f MG |
1795 | /* Retake the callers reference and putback on LRU */ |
1796 | get_page(page); | |
b32967ff | 1797 | putback_lru_page(page); |
a54a407f MG |
1798 | mod_zone_page_state(page_zone(page), |
1799 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
eb4489f6 MG |
1800 | |
1801 | goto out_unlock; | |
b32967ff MG |
1802 | } |
1803 | ||
1804 | /* | |
1805 | * Traditional migration needs to prepare the memcg charge | |
1806 | * transaction early to prevent the old page from being | |
1807 | * uncharged when installing migration entries. Here we can | |
1808 | * save the potential rollback and start the charge transfer | |
1809 | * only when migration is already known to end successfully. | |
1810 | */ | |
1811 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1812 | ||
2b4847e7 | 1813 | orig_entry = *pmd; |
b32967ff | 1814 | entry = mk_pmd(new_page, vma->vm_page_prot); |
b32967ff | 1815 | entry = pmd_mkhuge(entry); |
2b4847e7 | 1816 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 1817 | |
2b4847e7 MG |
1818 | /* |
1819 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1820 | * Any parallel GUP will either observe the old page blocking on the | |
1821 | * page lock, block on the page table lock or observe the new page. | |
1822 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1823 | * guarantee the copy is visible before the pagetable update. | |
1824 | */ | |
f714f4f2 MG |
1825 | flush_cache_range(vma, mmun_start, mmun_end); |
1826 | page_add_new_anon_rmap(new_page, vma, mmun_start); | |
1827 | pmdp_clear_flush(vma, mmun_start, pmd); | |
1828 | set_pmd_at(mm, mmun_start, pmd, entry); | |
1829 | flush_tlb_range(vma, mmun_start, mmun_end); | |
ce4a9cc5 | 1830 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 MG |
1831 | |
1832 | if (page_count(page) != 2) { | |
f714f4f2 MG |
1833 | set_pmd_at(mm, mmun_start, pmd, orig_entry); |
1834 | flush_tlb_range(vma, mmun_start, mmun_end); | |
2b4847e7 MG |
1835 | update_mmu_cache_pmd(vma, address, &entry); |
1836 | page_remove_rmap(new_page); | |
1837 | goto fail_putback; | |
1838 | } | |
1839 | ||
b32967ff | 1840 | page_remove_rmap(page); |
2b4847e7 | 1841 | |
b32967ff MG |
1842 | /* |
1843 | * Finish the charge transaction under the page table lock to | |
1844 | * prevent split_huge_page() from dividing up the charge | |
1845 | * before it's fully transferred to the new page. | |
1846 | */ | |
1847 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
c4088ebd | 1848 | spin_unlock(ptl); |
f714f4f2 | 1849 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1850 | |
1851 | unlock_page(new_page); | |
1852 | unlock_page(page); | |
1853 | put_page(page); /* Drop the rmap reference */ | |
1854 | put_page(page); /* Drop the LRU isolation reference */ | |
1855 | ||
1856 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1857 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1858 | ||
b32967ff MG |
1859 | mod_zone_page_state(page_zone(page), |
1860 | NR_ISOLATED_ANON + page_lru, | |
1861 | -HPAGE_PMD_NR); | |
1862 | return isolated; | |
1863 | ||
340ef390 HD |
1864 | out_fail: |
1865 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1866 | out_dropref: |
2b4847e7 MG |
1867 | ptl = pmd_lock(mm, pmd); |
1868 | if (pmd_same(*pmd, entry)) { | |
1869 | entry = pmd_mknonnuma(entry); | |
f714f4f2 | 1870 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
1871 | update_mmu_cache_pmd(vma, address, &entry); |
1872 | } | |
1873 | spin_unlock(ptl); | |
a54a407f | 1874 | |
eb4489f6 | 1875 | out_unlock: |
340ef390 | 1876 | unlock_page(page); |
b32967ff | 1877 | put_page(page); |
b32967ff MG |
1878 | return 0; |
1879 | } | |
7039e1db PZ |
1880 | #endif /* CONFIG_NUMA_BALANCING */ |
1881 | ||
1882 | #endif /* CONFIG_NUMA */ |