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 | ||
7e09e738 HD |
181 | /* |
182 | * Congratulations to trinity for discovering this bug. | |
183 | * mm/fremap.c's remap_file_pages() accepts any range within a single vma to | |
184 | * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then | |
185 | * replace the specified range by file ptes throughout (maybe populated after). | |
186 | * If page migration finds a page within that range, while it's still located | |
187 | * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem: | |
188 | * zap_pte() clears the temporary migration entry before mmap_sem is dropped. | |
189 | * But if the migrating page is in a part of the vma outside the range to be | |
190 | * remapped, then it will not be cleared, and remove_migration_ptes() needs to | |
191 | * deal with it. Fortunately, this part of the vma is of course still linear, | |
192 | * so we just need to use linear location on the nonlinear list. | |
193 | */ | |
194 | static int remove_linear_migration_ptes_from_nonlinear(struct page *page, | |
195 | struct address_space *mapping, void *arg) | |
196 | { | |
197 | struct vm_area_struct *vma; | |
198 | /* hugetlbfs does not support remap_pages, so no huge pgoff worries */ | |
199 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
200 | unsigned long addr; | |
201 | ||
202 | list_for_each_entry(vma, | |
203 | &mapping->i_mmap_nonlinear, shared.nonlinear) { | |
204 | ||
205 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
206 | if (addr >= vma->vm_start && addr < vma->vm_end) | |
207 | remove_migration_pte(page, vma, addr, arg); | |
208 | } | |
209 | return SWAP_AGAIN; | |
210 | } | |
211 | ||
04e62a29 CL |
212 | /* |
213 | * Get rid of all migration entries and replace them by | |
214 | * references to the indicated page. | |
215 | */ | |
216 | static void remove_migration_ptes(struct page *old, struct page *new) | |
217 | { | |
051ac83a JK |
218 | struct rmap_walk_control rwc = { |
219 | .rmap_one = remove_migration_pte, | |
220 | .arg = old, | |
7e09e738 | 221 | .file_nonlinear = remove_linear_migration_ptes_from_nonlinear, |
051ac83a JK |
222 | }; |
223 | ||
224 | rmap_walk(new, &rwc); | |
04e62a29 CL |
225 | } |
226 | ||
0697212a CL |
227 | /* |
228 | * Something used the pte of a page under migration. We need to | |
229 | * get to the page and wait until migration is finished. | |
230 | * When we return from this function the fault will be retried. | |
0697212a | 231 | */ |
30dad309 NH |
232 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
233 | spinlock_t *ptl) | |
0697212a | 234 | { |
30dad309 | 235 | pte_t pte; |
0697212a CL |
236 | swp_entry_t entry; |
237 | struct page *page; | |
238 | ||
30dad309 | 239 | spin_lock(ptl); |
0697212a CL |
240 | pte = *ptep; |
241 | if (!is_swap_pte(pte)) | |
242 | goto out; | |
243 | ||
244 | entry = pte_to_swp_entry(pte); | |
245 | if (!is_migration_entry(entry)) | |
246 | goto out; | |
247 | ||
248 | page = migration_entry_to_page(entry); | |
249 | ||
e286781d NP |
250 | /* |
251 | * Once radix-tree replacement of page migration started, page_count | |
252 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
253 | * against a page without get_page(). | |
254 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
255 | * will occur again. | |
256 | */ | |
257 | if (!get_page_unless_zero(page)) | |
258 | goto out; | |
0697212a CL |
259 | pte_unmap_unlock(ptep, ptl); |
260 | wait_on_page_locked(page); | |
261 | put_page(page); | |
262 | return; | |
263 | out: | |
264 | pte_unmap_unlock(ptep, ptl); | |
265 | } | |
266 | ||
30dad309 NH |
267 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
268 | unsigned long address) | |
269 | { | |
270 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
271 | pte_t *ptep = pte_offset_map(pmd, address); | |
272 | __migration_entry_wait(mm, ptep, ptl); | |
273 | } | |
274 | ||
cb900f41 KS |
275 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
276 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 277 | { |
cb900f41 | 278 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
279 | __migration_entry_wait(mm, pte, ptl); |
280 | } | |
281 | ||
b969c4ab MG |
282 | #ifdef CONFIG_BLOCK |
283 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
284 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
285 | enum migrate_mode mode) | |
b969c4ab MG |
286 | { |
287 | struct buffer_head *bh = head; | |
288 | ||
289 | /* Simple case, sync compaction */ | |
a6bc32b8 | 290 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
291 | do { |
292 | get_bh(bh); | |
293 | lock_buffer(bh); | |
294 | bh = bh->b_this_page; | |
295 | ||
296 | } while (bh != head); | |
297 | ||
298 | return true; | |
299 | } | |
300 | ||
301 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
302 | do { | |
303 | get_bh(bh); | |
304 | if (!trylock_buffer(bh)) { | |
305 | /* | |
306 | * We failed to lock the buffer and cannot stall in | |
307 | * async migration. Release the taken locks | |
308 | */ | |
309 | struct buffer_head *failed_bh = bh; | |
310 | put_bh(failed_bh); | |
311 | bh = head; | |
312 | while (bh != failed_bh) { | |
313 | unlock_buffer(bh); | |
314 | put_bh(bh); | |
315 | bh = bh->b_this_page; | |
316 | } | |
317 | return false; | |
318 | } | |
319 | ||
320 | bh = bh->b_this_page; | |
321 | } while (bh != head); | |
322 | return true; | |
323 | } | |
324 | #else | |
325 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 326 | enum migrate_mode mode) |
b969c4ab MG |
327 | { |
328 | return true; | |
329 | } | |
330 | #endif /* CONFIG_BLOCK */ | |
331 | ||
b20a3503 | 332 | /* |
c3fcf8a5 | 333 | * Replace the page in the mapping. |
5b5c7120 CL |
334 | * |
335 | * The number of remaining references must be: | |
336 | * 1 for anonymous pages without a mapping | |
337 | * 2 for pages with a mapping | |
266cf658 | 338 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 339 | */ |
36bc08cc | 340 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 341 | struct page *newpage, struct page *page, |
8e321fef BL |
342 | struct buffer_head *head, enum migrate_mode mode, |
343 | int extra_count) | |
b20a3503 | 344 | { |
8e321fef | 345 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 346 | void **pslot; |
b20a3503 | 347 | |
6c5240ae | 348 | if (!mapping) { |
0e8c7d0f | 349 | /* Anonymous page without mapping */ |
8e321fef | 350 | if (page_count(page) != expected_count) |
6c5240ae | 351 | return -EAGAIN; |
78bd5209 | 352 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
353 | } |
354 | ||
19fd6231 | 355 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 356 | |
7cf9c2c7 NP |
357 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
358 | page_index(page)); | |
b20a3503 | 359 | |
8e321fef | 360 | expected_count += 1 + page_has_private(page); |
e286781d | 361 | if (page_count(page) != expected_count || |
29c1f677 | 362 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 363 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 364 | return -EAGAIN; |
b20a3503 CL |
365 | } |
366 | ||
e286781d | 367 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 368 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
369 | return -EAGAIN; |
370 | } | |
371 | ||
b969c4ab MG |
372 | /* |
373 | * In the async migration case of moving a page with buffers, lock the | |
374 | * buffers using trylock before the mapping is moved. If the mapping | |
375 | * was moved, we later failed to lock the buffers and could not move | |
376 | * the mapping back due to an elevated page count, we would have to | |
377 | * block waiting on other references to be dropped. | |
378 | */ | |
a6bc32b8 MG |
379 | if (mode == MIGRATE_ASYNC && head && |
380 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
381 | page_unfreeze_refs(page, expected_count); |
382 | spin_unlock_irq(&mapping->tree_lock); | |
383 | return -EAGAIN; | |
384 | } | |
385 | ||
b20a3503 CL |
386 | /* |
387 | * Now we know that no one else is looking at the page. | |
b20a3503 | 388 | */ |
7cf9c2c7 | 389 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
390 | if (PageSwapCache(page)) { |
391 | SetPageSwapCache(newpage); | |
392 | set_page_private(newpage, page_private(page)); | |
393 | } | |
394 | ||
7cf9c2c7 NP |
395 | radix_tree_replace_slot(pslot, newpage); |
396 | ||
397 | /* | |
937a94c9 JG |
398 | * Drop cache reference from old page by unfreezing |
399 | * to one less reference. | |
7cf9c2c7 NP |
400 | * We know this isn't the last reference. |
401 | */ | |
937a94c9 | 402 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 403 | |
0e8c7d0f CL |
404 | /* |
405 | * If moved to a different zone then also account | |
406 | * the page for that zone. Other VM counters will be | |
407 | * taken care of when we establish references to the | |
408 | * new page and drop references to the old page. | |
409 | * | |
410 | * Note that anonymous pages are accounted for | |
411 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
412 | * are mapped to swap space. | |
413 | */ | |
414 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
415 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 416 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
417 | __dec_zone_page_state(page, NR_SHMEM); |
418 | __inc_zone_page_state(newpage, NR_SHMEM); | |
419 | } | |
19fd6231 | 420 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 421 | |
78bd5209 | 422 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 423 | } |
b20a3503 | 424 | |
290408d4 NH |
425 | /* |
426 | * The expected number of remaining references is the same as that | |
427 | * of migrate_page_move_mapping(). | |
428 | */ | |
429 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
430 | struct page *newpage, struct page *page) | |
431 | { | |
432 | int expected_count; | |
433 | void **pslot; | |
434 | ||
435 | if (!mapping) { | |
436 | if (page_count(page) != 1) | |
437 | return -EAGAIN; | |
78bd5209 | 438 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
439 | } |
440 | ||
441 | spin_lock_irq(&mapping->tree_lock); | |
442 | ||
443 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
444 | page_index(page)); | |
445 | ||
446 | expected_count = 2 + page_has_private(page); | |
447 | if (page_count(page) != expected_count || | |
29c1f677 | 448 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
449 | spin_unlock_irq(&mapping->tree_lock); |
450 | return -EAGAIN; | |
451 | } | |
452 | ||
453 | if (!page_freeze_refs(page, expected_count)) { | |
454 | spin_unlock_irq(&mapping->tree_lock); | |
455 | return -EAGAIN; | |
456 | } | |
457 | ||
458 | get_page(newpage); | |
459 | ||
460 | radix_tree_replace_slot(pslot, newpage); | |
461 | ||
937a94c9 | 462 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
463 | |
464 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 465 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
466 | } |
467 | ||
30b0a105 DH |
468 | /* |
469 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
470 | * arithmetic will work across the entire page. We need something more | |
471 | * specialized. | |
472 | */ | |
473 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
474 | int nr_pages) | |
475 | { | |
476 | int i; | |
477 | struct page *dst_base = dst; | |
478 | struct page *src_base = src; | |
479 | ||
480 | for (i = 0; i < nr_pages; ) { | |
481 | cond_resched(); | |
482 | copy_highpage(dst, src); | |
483 | ||
484 | i++; | |
485 | dst = mem_map_next(dst, dst_base, i); | |
486 | src = mem_map_next(src, src_base, i); | |
487 | } | |
488 | } | |
489 | ||
490 | static void copy_huge_page(struct page *dst, struct page *src) | |
491 | { | |
492 | int i; | |
493 | int nr_pages; | |
494 | ||
495 | if (PageHuge(src)) { | |
496 | /* hugetlbfs page */ | |
497 | struct hstate *h = page_hstate(src); | |
498 | nr_pages = pages_per_huge_page(h); | |
499 | ||
500 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
501 | __copy_gigantic_page(dst, src, nr_pages); | |
502 | return; | |
503 | } | |
504 | } else { | |
505 | /* thp page */ | |
506 | BUG_ON(!PageTransHuge(src)); | |
507 | nr_pages = hpage_nr_pages(src); | |
508 | } | |
509 | ||
510 | for (i = 0; i < nr_pages; i++) { | |
511 | cond_resched(); | |
512 | copy_highpage(dst + i, src + i); | |
513 | } | |
514 | } | |
515 | ||
b20a3503 CL |
516 | /* |
517 | * Copy the page to its new location | |
518 | */ | |
290408d4 | 519 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 520 | { |
7851a45c RR |
521 | int cpupid; |
522 | ||
b32967ff | 523 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
524 | copy_huge_page(newpage, page); |
525 | else | |
526 | copy_highpage(newpage, page); | |
b20a3503 CL |
527 | |
528 | if (PageError(page)) | |
529 | SetPageError(newpage); | |
530 | if (PageReferenced(page)) | |
531 | SetPageReferenced(newpage); | |
532 | if (PageUptodate(page)) | |
533 | SetPageUptodate(newpage); | |
894bc310 | 534 | if (TestClearPageActive(page)) { |
309381fe | 535 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 536 | SetPageActive(newpage); |
418b27ef LS |
537 | } else if (TestClearPageUnevictable(page)) |
538 | SetPageUnevictable(newpage); | |
b20a3503 CL |
539 | if (PageChecked(page)) |
540 | SetPageChecked(newpage); | |
541 | if (PageMappedToDisk(page)) | |
542 | SetPageMappedToDisk(newpage); | |
543 | ||
544 | if (PageDirty(page)) { | |
545 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
546 | /* |
547 | * Want to mark the page and the radix tree as dirty, and | |
548 | * redo the accounting that clear_page_dirty_for_io undid, | |
549 | * but we can't use set_page_dirty because that function | |
550 | * is actually a signal that all of the page has become dirty. | |
25985edc | 551 | * Whereas only part of our page may be dirty. |
3a902c5f | 552 | */ |
752dc185 HD |
553 | if (PageSwapBacked(page)) |
554 | SetPageDirty(newpage); | |
555 | else | |
556 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
557 | } |
558 | ||
7851a45c RR |
559 | /* |
560 | * Copy NUMA information to the new page, to prevent over-eager | |
561 | * future migrations of this same page. | |
562 | */ | |
563 | cpupid = page_cpupid_xchg_last(page, -1); | |
564 | page_cpupid_xchg_last(newpage, cpupid); | |
565 | ||
b291f000 | 566 | mlock_migrate_page(newpage, page); |
e9995ef9 | 567 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
568 | /* |
569 | * Please do not reorder this without considering how mm/ksm.c's | |
570 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
571 | */ | |
b20a3503 | 572 | ClearPageSwapCache(page); |
b20a3503 CL |
573 | ClearPagePrivate(page); |
574 | set_page_private(page, 0); | |
b20a3503 CL |
575 | |
576 | /* | |
577 | * If any waiters have accumulated on the new page then | |
578 | * wake them up. | |
579 | */ | |
580 | if (PageWriteback(newpage)) | |
581 | end_page_writeback(newpage); | |
582 | } | |
b20a3503 | 583 | |
1d8b85cc CL |
584 | /************************************************************ |
585 | * Migration functions | |
586 | ***********************************************************/ | |
587 | ||
b20a3503 CL |
588 | /* |
589 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 590 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
591 | * |
592 | * Pages are locked upon entry and exit. | |
593 | */ | |
2d1db3b1 | 594 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
595 | struct page *newpage, struct page *page, |
596 | enum migrate_mode mode) | |
b20a3503 CL |
597 | { |
598 | int rc; | |
599 | ||
600 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
601 | ||
8e321fef | 602 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 603 | |
78bd5209 | 604 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
605 | return rc; |
606 | ||
607 | migrate_page_copy(newpage, page); | |
78bd5209 | 608 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
609 | } |
610 | EXPORT_SYMBOL(migrate_page); | |
611 | ||
9361401e | 612 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
613 | /* |
614 | * Migration function for pages with buffers. This function can only be used | |
615 | * if the underlying filesystem guarantees that no other references to "page" | |
616 | * exist. | |
617 | */ | |
2d1db3b1 | 618 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 619 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 620 | { |
1d8b85cc CL |
621 | struct buffer_head *bh, *head; |
622 | int rc; | |
623 | ||
1d8b85cc | 624 | if (!page_has_buffers(page)) |
a6bc32b8 | 625 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
626 | |
627 | head = page_buffers(page); | |
628 | ||
8e321fef | 629 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 630 | |
78bd5209 | 631 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
632 | return rc; |
633 | ||
b969c4ab MG |
634 | /* |
635 | * In the async case, migrate_page_move_mapping locked the buffers | |
636 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
637 | * need to be locked now | |
638 | */ | |
a6bc32b8 MG |
639 | if (mode != MIGRATE_ASYNC) |
640 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
641 | |
642 | ClearPagePrivate(page); | |
643 | set_page_private(newpage, page_private(page)); | |
644 | set_page_private(page, 0); | |
645 | put_page(page); | |
646 | get_page(newpage); | |
647 | ||
648 | bh = head; | |
649 | do { | |
650 | set_bh_page(bh, newpage, bh_offset(bh)); | |
651 | bh = bh->b_this_page; | |
652 | ||
653 | } while (bh != head); | |
654 | ||
655 | SetPagePrivate(newpage); | |
656 | ||
657 | migrate_page_copy(newpage, page); | |
658 | ||
659 | bh = head; | |
660 | do { | |
661 | unlock_buffer(bh); | |
662 | put_bh(bh); | |
663 | bh = bh->b_this_page; | |
664 | ||
665 | } while (bh != head); | |
666 | ||
78bd5209 | 667 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
668 | } |
669 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 670 | #endif |
1d8b85cc | 671 | |
04e62a29 CL |
672 | /* |
673 | * Writeback a page to clean the dirty state | |
674 | */ | |
675 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 676 | { |
04e62a29 CL |
677 | struct writeback_control wbc = { |
678 | .sync_mode = WB_SYNC_NONE, | |
679 | .nr_to_write = 1, | |
680 | .range_start = 0, | |
681 | .range_end = LLONG_MAX, | |
04e62a29 CL |
682 | .for_reclaim = 1 |
683 | }; | |
684 | int rc; | |
685 | ||
686 | if (!mapping->a_ops->writepage) | |
687 | /* No write method for the address space */ | |
688 | return -EINVAL; | |
689 | ||
690 | if (!clear_page_dirty_for_io(page)) | |
691 | /* Someone else already triggered a write */ | |
692 | return -EAGAIN; | |
693 | ||
8351a6e4 | 694 | /* |
04e62a29 CL |
695 | * A dirty page may imply that the underlying filesystem has |
696 | * the page on some queue. So the page must be clean for | |
697 | * migration. Writeout may mean we loose the lock and the | |
698 | * page state is no longer what we checked for earlier. | |
699 | * At this point we know that the migration attempt cannot | |
700 | * be successful. | |
8351a6e4 | 701 | */ |
04e62a29 | 702 | remove_migration_ptes(page, page); |
8351a6e4 | 703 | |
04e62a29 | 704 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 705 | |
04e62a29 CL |
706 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
707 | /* unlocked. Relock */ | |
708 | lock_page(page); | |
709 | ||
bda8550d | 710 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
711 | } |
712 | ||
713 | /* | |
714 | * Default handling if a filesystem does not provide a migration function. | |
715 | */ | |
716 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 717 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 718 | { |
b969c4ab | 719 | if (PageDirty(page)) { |
a6bc32b8 MG |
720 | /* Only writeback pages in full synchronous migration */ |
721 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 722 | return -EBUSY; |
04e62a29 | 723 | return writeout(mapping, page); |
b969c4ab | 724 | } |
8351a6e4 CL |
725 | |
726 | /* | |
727 | * Buffers may be managed in a filesystem specific way. | |
728 | * We must have no buffers or drop them. | |
729 | */ | |
266cf658 | 730 | if (page_has_private(page) && |
8351a6e4 CL |
731 | !try_to_release_page(page, GFP_KERNEL)) |
732 | return -EAGAIN; | |
733 | ||
a6bc32b8 | 734 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
735 | } |
736 | ||
e24f0b8f CL |
737 | /* |
738 | * Move a page to a newly allocated page | |
739 | * The page is locked and all ptes have been successfully removed. | |
740 | * | |
741 | * The new page will have replaced the old page if this function | |
742 | * is successful. | |
894bc310 LS |
743 | * |
744 | * Return value: | |
745 | * < 0 - error code | |
78bd5209 | 746 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 747 | */ |
3fe2011f | 748 | static int move_to_new_page(struct page *newpage, struct page *page, |
2ebba6b7 | 749 | int page_was_mapped, enum migrate_mode mode) |
e24f0b8f CL |
750 | { |
751 | struct address_space *mapping; | |
752 | int rc; | |
753 | ||
754 | /* | |
755 | * Block others from accessing the page when we get around to | |
756 | * establishing additional references. We are the only one | |
757 | * holding a reference to the new page at this point. | |
758 | */ | |
529ae9aa | 759 | if (!trylock_page(newpage)) |
e24f0b8f CL |
760 | BUG(); |
761 | ||
762 | /* Prepare mapping for the new page.*/ | |
763 | newpage->index = page->index; | |
764 | newpage->mapping = page->mapping; | |
b2e18538 RR |
765 | if (PageSwapBacked(page)) |
766 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
767 | |
768 | mapping = page_mapping(page); | |
769 | if (!mapping) | |
a6bc32b8 | 770 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 771 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 772 | /* |
b969c4ab MG |
773 | * Most pages have a mapping and most filesystems provide a |
774 | * migratepage callback. Anonymous pages are part of swap | |
775 | * space which also has its own migratepage callback. This | |
776 | * is the most common path for page migration. | |
e24f0b8f | 777 | */ |
b969c4ab | 778 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 779 | newpage, page, mode); |
b969c4ab | 780 | else |
a6bc32b8 | 781 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 782 | |
78bd5209 | 783 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 784 | newpage->mapping = NULL; |
3fe2011f | 785 | } else { |
0a31bc97 | 786 | mem_cgroup_migrate(page, newpage, false); |
2ebba6b7 | 787 | if (page_was_mapped) |
3fe2011f | 788 | remove_migration_ptes(page, newpage); |
35512eca | 789 | page->mapping = NULL; |
3fe2011f | 790 | } |
e24f0b8f CL |
791 | |
792 | unlock_page(newpage); | |
793 | ||
794 | return rc; | |
795 | } | |
796 | ||
0dabec93 | 797 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 798 | int force, enum migrate_mode mode) |
e24f0b8f | 799 | { |
0dabec93 | 800 | int rc = -EAGAIN; |
2ebba6b7 | 801 | int page_was_mapped = 0; |
3f6c8272 | 802 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 803 | |
529ae9aa | 804 | if (!trylock_page(page)) { |
a6bc32b8 | 805 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 806 | goto out; |
3e7d3449 MG |
807 | |
808 | /* | |
809 | * It's not safe for direct compaction to call lock_page. | |
810 | * For example, during page readahead pages are added locked | |
811 | * to the LRU. Later, when the IO completes the pages are | |
812 | * marked uptodate and unlocked. However, the queueing | |
813 | * could be merging multiple pages for one bio (e.g. | |
814 | * mpage_readpages). If an allocation happens for the | |
815 | * second or third page, the process can end up locking | |
816 | * the same page twice and deadlocking. Rather than | |
817 | * trying to be clever about what pages can be locked, | |
818 | * avoid the use of lock_page for direct compaction | |
819 | * altogether. | |
820 | */ | |
821 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 822 | goto out; |
3e7d3449 | 823 | |
e24f0b8f CL |
824 | lock_page(page); |
825 | } | |
826 | ||
827 | if (PageWriteback(page)) { | |
11bc82d6 | 828 | /* |
fed5b64a | 829 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
830 | * necessary to wait for PageWriteback. In the async case, |
831 | * the retry loop is too short and in the sync-light case, | |
832 | * the overhead of stalling is too much | |
11bc82d6 | 833 | */ |
a6bc32b8 | 834 | if (mode != MIGRATE_SYNC) { |
11bc82d6 | 835 | rc = -EBUSY; |
0a31bc97 | 836 | goto out_unlock; |
11bc82d6 AA |
837 | } |
838 | if (!force) | |
0a31bc97 | 839 | goto out_unlock; |
e24f0b8f CL |
840 | wait_on_page_writeback(page); |
841 | } | |
e24f0b8f | 842 | /* |
dc386d4d KH |
843 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
844 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 845 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 846 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
847 | * File Caches may use write_page() or lock_page() in migration, then, |
848 | * just care Anon page here. | |
dc386d4d | 849 | */ |
b79bc0a0 | 850 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 851 | /* |
4fc3f1d6 | 852 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
853 | * getting a hold on an anon_vma from outside one of its mms. |
854 | */ | |
746b18d4 | 855 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
856 | if (anon_vma) { |
857 | /* | |
746b18d4 | 858 | * Anon page |
1ce82b69 | 859 | */ |
1ce82b69 | 860 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
861 | /* |
862 | * We cannot be sure that the anon_vma of an unmapped | |
863 | * swapcache page is safe to use because we don't | |
864 | * know in advance if the VMA that this page belonged | |
865 | * to still exists. If the VMA and others sharing the | |
866 | * data have been freed, then the anon_vma could | |
867 | * already be invalid. | |
868 | * | |
869 | * To avoid this possibility, swapcache pages get | |
870 | * migrated but are not remapped when migration | |
871 | * completes | |
872 | */ | |
3fe2011f | 873 | } else { |
0a31bc97 | 874 | goto out_unlock; |
3fe2011f | 875 | } |
989f89c5 | 876 | } |
62e1c553 | 877 | |
d6d86c0a | 878 | if (unlikely(isolated_balloon_page(page))) { |
bf6bddf1 RA |
879 | /* |
880 | * A ballooned page does not need any special attention from | |
881 | * physical to virtual reverse mapping procedures. | |
882 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
883 | * in order to avoid burning cycles at rmap level, and perform | |
884 | * the page migration right away (proteced by page lock). | |
885 | */ | |
886 | rc = balloon_page_migrate(newpage, page, mode); | |
0a31bc97 | 887 | goto out_unlock; |
bf6bddf1 RA |
888 | } |
889 | ||
dc386d4d | 890 | /* |
62e1c553 SL |
891 | * Corner case handling: |
892 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
893 | * and treated as swapcache but it has no rmap yet. | |
894 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
895 | * trigger a BUG. So handle it here. | |
896 | * 2. An orphaned page (see truncate_complete_page) might have | |
897 | * fs-private metadata. The page can be picked up due to memory | |
898 | * offlining. Everywhere else except page reclaim, the page is | |
899 | * invisible to the vm, so the page can not be migrated. So try to | |
900 | * free the metadata, so the page can be freed. | |
e24f0b8f | 901 | */ |
62e1c553 | 902 | if (!page->mapping) { |
309381fe | 903 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 904 | if (page_has_private(page)) { |
62e1c553 | 905 | try_to_free_buffers(page); |
0a31bc97 | 906 | goto out_unlock; |
62e1c553 | 907 | } |
abfc3488 | 908 | goto skip_unmap; |
62e1c553 SL |
909 | } |
910 | ||
dc386d4d | 911 | /* Establish migration ptes or remove ptes */ |
2ebba6b7 HD |
912 | if (page_mapped(page)) { |
913 | try_to_unmap(page, | |
914 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
915 | page_was_mapped = 1; | |
916 | } | |
dc386d4d | 917 | |
abfc3488 | 918 | skip_unmap: |
e6a1530d | 919 | if (!page_mapped(page)) |
2ebba6b7 | 920 | rc = move_to_new_page(newpage, page, page_was_mapped, mode); |
e24f0b8f | 921 | |
2ebba6b7 | 922 | if (rc && page_was_mapped) |
e24f0b8f | 923 | remove_migration_ptes(page, page); |
3f6c8272 MG |
924 | |
925 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 926 | if (anon_vma) |
9e60109f | 927 | put_anon_vma(anon_vma); |
3f6c8272 | 928 | |
0a31bc97 | 929 | out_unlock: |
e24f0b8f | 930 | unlock_page(page); |
0dabec93 MK |
931 | out: |
932 | return rc; | |
933 | } | |
95a402c3 | 934 | |
0dabec93 MK |
935 | /* |
936 | * Obtain the lock on page, remove all ptes and migrate the page | |
937 | * to the newly allocated page in newpage. | |
938 | */ | |
68711a74 DR |
939 | static int unmap_and_move(new_page_t get_new_page, free_page_t put_new_page, |
940 | unsigned long private, struct page *page, int force, | |
941 | enum migrate_mode mode) | |
0dabec93 MK |
942 | { |
943 | int rc = 0; | |
944 | int *result = NULL; | |
945 | struct page *newpage = get_new_page(page, private, &result); | |
946 | ||
947 | if (!newpage) | |
948 | return -ENOMEM; | |
949 | ||
950 | if (page_count(page) == 1) { | |
951 | /* page was freed from under us. So we are done. */ | |
952 | goto out; | |
953 | } | |
954 | ||
955 | if (unlikely(PageTransHuge(page))) | |
956 | if (unlikely(split_huge_page(page))) | |
957 | goto out; | |
958 | ||
9c620e2b | 959 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 | 960 | |
0dabec93 | 961 | out: |
e24f0b8f | 962 | if (rc != -EAGAIN) { |
0dabec93 MK |
963 | /* |
964 | * A page that has been migrated has all references | |
965 | * removed and will be freed. A page that has not been | |
966 | * migrated will have kepts its references and be | |
967 | * restored. | |
968 | */ | |
969 | list_del(&page->lru); | |
a731286d | 970 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 971 | page_is_file_cache(page)); |
894bc310 | 972 | putback_lru_page(page); |
e24f0b8f | 973 | } |
68711a74 | 974 | |
95a402c3 | 975 | /* |
68711a74 DR |
976 | * If migration was not successful and there's a freeing callback, use |
977 | * it. Otherwise, putback_lru_page() will drop the reference grabbed | |
978 | * during isolation. | |
95a402c3 | 979 | */ |
8bdd6380 HD |
980 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) { |
981 | ClearPageSwapBacked(newpage); | |
68711a74 | 982 | put_new_page(newpage, private); |
d6d86c0a KK |
983 | } else if (unlikely(__is_movable_balloon_page(newpage))) { |
984 | /* drop our reference, page already in the balloon */ | |
985 | put_page(newpage); | |
8bdd6380 | 986 | } else |
68711a74 DR |
987 | putback_lru_page(newpage); |
988 | ||
742755a1 CL |
989 | if (result) { |
990 | if (rc) | |
991 | *result = rc; | |
992 | else | |
993 | *result = page_to_nid(newpage); | |
994 | } | |
e24f0b8f CL |
995 | return rc; |
996 | } | |
997 | ||
290408d4 NH |
998 | /* |
999 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1000 | * | |
1001 | * This function doesn't wait the completion of hugepage I/O | |
1002 | * because there is no race between I/O and migration for hugepage. | |
1003 | * Note that currently hugepage I/O occurs only in direct I/O | |
1004 | * where no lock is held and PG_writeback is irrelevant, | |
1005 | * and writeback status of all subpages are counted in the reference | |
1006 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1007 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1008 | * This means that when we try to migrate hugepage whose subpages are | |
1009 | * doing direct I/O, some references remain after try_to_unmap() and | |
1010 | * hugepage migration fails without data corruption. | |
1011 | * | |
1012 | * There is also no race when direct I/O is issued on the page under migration, | |
1013 | * because then pte is replaced with migration swap entry and direct I/O code | |
1014 | * will wait in the page fault for migration to complete. | |
1015 | */ | |
1016 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1017 | free_page_t put_new_page, unsigned long private, |
1018 | struct page *hpage, int force, | |
1019 | enum migrate_mode mode) | |
290408d4 NH |
1020 | { |
1021 | int rc = 0; | |
1022 | int *result = NULL; | |
2ebba6b7 | 1023 | int page_was_mapped = 0; |
32665f2b | 1024 | struct page *new_hpage; |
290408d4 NH |
1025 | struct anon_vma *anon_vma = NULL; |
1026 | ||
83467efb NH |
1027 | /* |
1028 | * Movability of hugepages depends on architectures and hugepage size. | |
1029 | * This check is necessary because some callers of hugepage migration | |
1030 | * like soft offline and memory hotremove don't walk through page | |
1031 | * tables or check whether the hugepage is pmd-based or not before | |
1032 | * kicking migration. | |
1033 | */ | |
100873d7 | 1034 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1035 | putback_active_hugepage(hpage); |
83467efb | 1036 | return -ENOSYS; |
32665f2b | 1037 | } |
83467efb | 1038 | |
32665f2b | 1039 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1040 | if (!new_hpage) |
1041 | return -ENOMEM; | |
1042 | ||
1043 | rc = -EAGAIN; | |
1044 | ||
1045 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 1046 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
1047 | goto out; |
1048 | lock_page(hpage); | |
1049 | } | |
1050 | ||
746b18d4 PZ |
1051 | if (PageAnon(hpage)) |
1052 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1053 | |
2ebba6b7 HD |
1054 | if (page_mapped(hpage)) { |
1055 | try_to_unmap(hpage, | |
1056 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1057 | page_was_mapped = 1; | |
1058 | } | |
290408d4 NH |
1059 | |
1060 | if (!page_mapped(hpage)) | |
2ebba6b7 | 1061 | rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode); |
290408d4 | 1062 | |
2ebba6b7 | 1063 | if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped) |
290408d4 NH |
1064 | remove_migration_ptes(hpage, hpage); |
1065 | ||
fd4a4663 | 1066 | if (anon_vma) |
9e60109f | 1067 | put_anon_vma(anon_vma); |
8e6ac7fa | 1068 | |
68711a74 | 1069 | if (rc == MIGRATEPAGE_SUCCESS) |
8e6ac7fa AK |
1070 | hugetlb_cgroup_migrate(hpage, new_hpage); |
1071 | ||
290408d4 | 1072 | unlock_page(hpage); |
09761333 | 1073 | out: |
b8ec1cee NH |
1074 | if (rc != -EAGAIN) |
1075 | putback_active_hugepage(hpage); | |
68711a74 DR |
1076 | |
1077 | /* | |
1078 | * If migration was not successful and there's a freeing callback, use | |
1079 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1080 | * isolation. | |
1081 | */ | |
1082 | if (rc != MIGRATEPAGE_SUCCESS && put_new_page) | |
1083 | put_new_page(new_hpage, private); | |
1084 | else | |
1085 | put_page(new_hpage); | |
1086 | ||
290408d4 NH |
1087 | if (result) { |
1088 | if (rc) | |
1089 | *result = rc; | |
1090 | else | |
1091 | *result = page_to_nid(new_hpage); | |
1092 | } | |
1093 | return rc; | |
1094 | } | |
1095 | ||
b20a3503 | 1096 | /* |
c73e5c9c SB |
1097 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1098 | * supplied as the target for the page migration | |
b20a3503 | 1099 | * |
c73e5c9c SB |
1100 | * @from: The list of pages to be migrated. |
1101 | * @get_new_page: The function used to allocate free pages to be used | |
1102 | * as the target of the page migration. | |
68711a74 DR |
1103 | * @put_new_page: The function used to free target pages if migration |
1104 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1105 | * @private: Private data to be passed on to get_new_page() |
1106 | * @mode: The migration mode that specifies the constraints for | |
1107 | * page migration, if any. | |
1108 | * @reason: The reason for page migration. | |
b20a3503 | 1109 | * |
c73e5c9c SB |
1110 | * The function returns after 10 attempts or if no pages are movable any more |
1111 | * because the list has become empty or no retryable pages exist any more. | |
1112 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1113 | * or free list only if ret != 0. |
b20a3503 | 1114 | * |
c73e5c9c | 1115 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1116 | */ |
9c620e2b | 1117 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1118 | free_page_t put_new_page, unsigned long private, |
1119 | enum migrate_mode mode, int reason) | |
b20a3503 | 1120 | { |
e24f0b8f | 1121 | int retry = 1; |
b20a3503 | 1122 | int nr_failed = 0; |
5647bc29 | 1123 | int nr_succeeded = 0; |
b20a3503 CL |
1124 | int pass = 0; |
1125 | struct page *page; | |
1126 | struct page *page2; | |
1127 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1128 | int rc; | |
1129 | ||
1130 | if (!swapwrite) | |
1131 | current->flags |= PF_SWAPWRITE; | |
1132 | ||
e24f0b8f CL |
1133 | for(pass = 0; pass < 10 && retry; pass++) { |
1134 | retry = 0; | |
b20a3503 | 1135 | |
e24f0b8f | 1136 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1137 | cond_resched(); |
2d1db3b1 | 1138 | |
31caf665 NH |
1139 | if (PageHuge(page)) |
1140 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 DR |
1141 | put_new_page, private, page, |
1142 | pass > 2, mode); | |
31caf665 | 1143 | else |
68711a74 DR |
1144 | rc = unmap_and_move(get_new_page, put_new_page, |
1145 | private, page, pass > 2, mode); | |
2d1db3b1 | 1146 | |
e24f0b8f | 1147 | switch(rc) { |
95a402c3 CL |
1148 | case -ENOMEM: |
1149 | goto out; | |
e24f0b8f | 1150 | case -EAGAIN: |
2d1db3b1 | 1151 | retry++; |
e24f0b8f | 1152 | break; |
78bd5209 | 1153 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1154 | nr_succeeded++; |
e24f0b8f CL |
1155 | break; |
1156 | default: | |
354a3363 NH |
1157 | /* |
1158 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1159 | * unlike -EAGAIN case, the failed page is | |
1160 | * removed from migration page list and not | |
1161 | * retried in the next outer loop. | |
1162 | */ | |
2d1db3b1 | 1163 | nr_failed++; |
e24f0b8f | 1164 | break; |
2d1db3b1 | 1165 | } |
b20a3503 CL |
1166 | } |
1167 | } | |
78bd5209 | 1168 | rc = nr_failed + retry; |
95a402c3 | 1169 | out: |
5647bc29 MG |
1170 | if (nr_succeeded) |
1171 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1172 | if (nr_failed) | |
1173 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1174 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1175 | ||
b20a3503 CL |
1176 | if (!swapwrite) |
1177 | current->flags &= ~PF_SWAPWRITE; | |
1178 | ||
78bd5209 | 1179 | return rc; |
b20a3503 | 1180 | } |
95a402c3 | 1181 | |
742755a1 CL |
1182 | #ifdef CONFIG_NUMA |
1183 | /* | |
1184 | * Move a list of individual pages | |
1185 | */ | |
1186 | struct page_to_node { | |
1187 | unsigned long addr; | |
1188 | struct page *page; | |
1189 | int node; | |
1190 | int status; | |
1191 | }; | |
1192 | ||
1193 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1194 | int **result) | |
1195 | { | |
1196 | struct page_to_node *pm = (struct page_to_node *)private; | |
1197 | ||
1198 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1199 | pm++; | |
1200 | ||
1201 | if (pm->node == MAX_NUMNODES) | |
1202 | return NULL; | |
1203 | ||
1204 | *result = &pm->status; | |
1205 | ||
e632a938 NH |
1206 | if (PageHuge(p)) |
1207 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1208 | pm->node); | |
1209 | else | |
1210 | return alloc_pages_exact_node(pm->node, | |
e97ca8e5 | 1211 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); |
742755a1 CL |
1212 | } |
1213 | ||
1214 | /* | |
1215 | * Move a set of pages as indicated in the pm array. The addr | |
1216 | * field must be set to the virtual address of the page to be moved | |
1217 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1218 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1219 | */ |
5e9a0f02 BG |
1220 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1221 | struct page_to_node *pm, | |
1222 | int migrate_all) | |
742755a1 CL |
1223 | { |
1224 | int err; | |
1225 | struct page_to_node *pp; | |
1226 | LIST_HEAD(pagelist); | |
1227 | ||
1228 | down_read(&mm->mmap_sem); | |
1229 | ||
1230 | /* | |
1231 | * Build a list of pages to migrate | |
1232 | */ | |
742755a1 CL |
1233 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1234 | struct vm_area_struct *vma; | |
1235 | struct page *page; | |
1236 | ||
742755a1 CL |
1237 | err = -EFAULT; |
1238 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1239 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1240 | goto set_status; |
1241 | ||
500d65d4 | 1242 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1243 | |
1244 | err = PTR_ERR(page); | |
1245 | if (IS_ERR(page)) | |
1246 | goto set_status; | |
1247 | ||
742755a1 CL |
1248 | err = -ENOENT; |
1249 | if (!page) | |
1250 | goto set_status; | |
1251 | ||
62b61f61 | 1252 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1253 | if (PageReserved(page)) |
742755a1 CL |
1254 | goto put_and_set; |
1255 | ||
1256 | pp->page = page; | |
1257 | err = page_to_nid(page); | |
1258 | ||
1259 | if (err == pp->node) | |
1260 | /* | |
1261 | * Node already in the right place | |
1262 | */ | |
1263 | goto put_and_set; | |
1264 | ||
1265 | err = -EACCES; | |
1266 | if (page_mapcount(page) > 1 && | |
1267 | !migrate_all) | |
1268 | goto put_and_set; | |
1269 | ||
e632a938 NH |
1270 | if (PageHuge(page)) { |
1271 | isolate_huge_page(page, &pagelist); | |
1272 | goto put_and_set; | |
1273 | } | |
1274 | ||
62695a84 | 1275 | err = isolate_lru_page(page); |
6d9c285a | 1276 | if (!err) { |
62695a84 | 1277 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1278 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1279 | page_is_file_cache(page)); | |
1280 | } | |
742755a1 CL |
1281 | put_and_set: |
1282 | /* | |
1283 | * Either remove the duplicate refcount from | |
1284 | * isolate_lru_page() or drop the page ref if it was | |
1285 | * not isolated. | |
1286 | */ | |
1287 | put_page(page); | |
1288 | set_status: | |
1289 | pp->status = err; | |
1290 | } | |
1291 | ||
e78bbfa8 | 1292 | err = 0; |
cf608ac1 | 1293 | if (!list_empty(&pagelist)) { |
68711a74 | 1294 | err = migrate_pages(&pagelist, new_page_node, NULL, |
9c620e2b | 1295 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1296 | if (err) |
e632a938 | 1297 | putback_movable_pages(&pagelist); |
cf608ac1 | 1298 | } |
742755a1 CL |
1299 | |
1300 | up_read(&mm->mmap_sem); | |
1301 | return err; | |
1302 | } | |
1303 | ||
5e9a0f02 BG |
1304 | /* |
1305 | * Migrate an array of page address onto an array of nodes and fill | |
1306 | * the corresponding array of status. | |
1307 | */ | |
3268c63e | 1308 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1309 | unsigned long nr_pages, |
1310 | const void __user * __user *pages, | |
1311 | const int __user *nodes, | |
1312 | int __user *status, int flags) | |
1313 | { | |
3140a227 | 1314 | struct page_to_node *pm; |
3140a227 BG |
1315 | unsigned long chunk_nr_pages; |
1316 | unsigned long chunk_start; | |
1317 | int err; | |
5e9a0f02 | 1318 | |
3140a227 BG |
1319 | err = -ENOMEM; |
1320 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1321 | if (!pm) | |
5e9a0f02 | 1322 | goto out; |
35282a2d BG |
1323 | |
1324 | migrate_prep(); | |
1325 | ||
5e9a0f02 | 1326 | /* |
3140a227 BG |
1327 | * Store a chunk of page_to_node array in a page, |
1328 | * but keep the last one as a marker | |
5e9a0f02 | 1329 | */ |
3140a227 | 1330 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1331 | |
3140a227 BG |
1332 | for (chunk_start = 0; |
1333 | chunk_start < nr_pages; | |
1334 | chunk_start += chunk_nr_pages) { | |
1335 | int j; | |
5e9a0f02 | 1336 | |
3140a227 BG |
1337 | if (chunk_start + chunk_nr_pages > nr_pages) |
1338 | chunk_nr_pages = nr_pages - chunk_start; | |
1339 | ||
1340 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1341 | for (j = 0; j < chunk_nr_pages; j++) { | |
1342 | const void __user *p; | |
5e9a0f02 BG |
1343 | int node; |
1344 | ||
3140a227 BG |
1345 | err = -EFAULT; |
1346 | if (get_user(p, pages + j + chunk_start)) | |
1347 | goto out_pm; | |
1348 | pm[j].addr = (unsigned long) p; | |
1349 | ||
1350 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1351 | goto out_pm; |
1352 | ||
1353 | err = -ENODEV; | |
6f5a55f1 LT |
1354 | if (node < 0 || node >= MAX_NUMNODES) |
1355 | goto out_pm; | |
1356 | ||
389162c2 | 1357 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1358 | goto out_pm; |
1359 | ||
1360 | err = -EACCES; | |
1361 | if (!node_isset(node, task_nodes)) | |
1362 | goto out_pm; | |
1363 | ||
3140a227 BG |
1364 | pm[j].node = node; |
1365 | } | |
1366 | ||
1367 | /* End marker for this chunk */ | |
1368 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1369 | ||
1370 | /* Migrate this chunk */ | |
1371 | err = do_move_page_to_node_array(mm, pm, | |
1372 | flags & MPOL_MF_MOVE_ALL); | |
1373 | if (err < 0) | |
1374 | goto out_pm; | |
5e9a0f02 | 1375 | |
5e9a0f02 | 1376 | /* Return status information */ |
3140a227 BG |
1377 | for (j = 0; j < chunk_nr_pages; j++) |
1378 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1379 | err = -EFAULT; |
3140a227 BG |
1380 | goto out_pm; |
1381 | } | |
1382 | } | |
1383 | err = 0; | |
5e9a0f02 BG |
1384 | |
1385 | out_pm: | |
3140a227 | 1386 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1387 | out: |
1388 | return err; | |
1389 | } | |
1390 | ||
742755a1 | 1391 | /* |
2f007e74 | 1392 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1393 | */ |
80bba129 BG |
1394 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1395 | const void __user **pages, int *status) | |
742755a1 | 1396 | { |
2f007e74 | 1397 | unsigned long i; |
2f007e74 | 1398 | |
742755a1 CL |
1399 | down_read(&mm->mmap_sem); |
1400 | ||
2f007e74 | 1401 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1402 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1403 | struct vm_area_struct *vma; |
1404 | struct page *page; | |
c095adbc | 1405 | int err = -EFAULT; |
2f007e74 BG |
1406 | |
1407 | vma = find_vma(mm, addr); | |
70384dc6 | 1408 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1409 | goto set_status; |
1410 | ||
2f007e74 | 1411 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1412 | |
1413 | err = PTR_ERR(page); | |
1414 | if (IS_ERR(page)) | |
1415 | goto set_status; | |
1416 | ||
742755a1 CL |
1417 | err = -ENOENT; |
1418 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1419 | if (!page || PageReserved(page)) |
742755a1 CL |
1420 | goto set_status; |
1421 | ||
1422 | err = page_to_nid(page); | |
1423 | set_status: | |
80bba129 BG |
1424 | *status = err; |
1425 | ||
1426 | pages++; | |
1427 | status++; | |
1428 | } | |
1429 | ||
1430 | up_read(&mm->mmap_sem); | |
1431 | } | |
1432 | ||
1433 | /* | |
1434 | * Determine the nodes of a user array of pages and store it in | |
1435 | * a user array of status. | |
1436 | */ | |
1437 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1438 | const void __user * __user *pages, | |
1439 | int __user *status) | |
1440 | { | |
1441 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1442 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1443 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1444 | |
87b8d1ad PA |
1445 | while (nr_pages) { |
1446 | unsigned long chunk_nr; | |
80bba129 | 1447 | |
87b8d1ad PA |
1448 | chunk_nr = nr_pages; |
1449 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1450 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1451 | ||
1452 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1453 | break; | |
80bba129 BG |
1454 | |
1455 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1456 | ||
87b8d1ad PA |
1457 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1458 | break; | |
742755a1 | 1459 | |
87b8d1ad PA |
1460 | pages += chunk_nr; |
1461 | status += chunk_nr; | |
1462 | nr_pages -= chunk_nr; | |
1463 | } | |
1464 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1465 | } |
1466 | ||
1467 | /* | |
1468 | * Move a list of pages in the address space of the currently executing | |
1469 | * process. | |
1470 | */ | |
938bb9f5 HC |
1471 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1472 | const void __user * __user *, pages, | |
1473 | const int __user *, nodes, | |
1474 | int __user *, status, int, flags) | |
742755a1 | 1475 | { |
c69e8d9c | 1476 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1477 | struct task_struct *task; |
742755a1 | 1478 | struct mm_struct *mm; |
5e9a0f02 | 1479 | int err; |
3268c63e | 1480 | nodemask_t task_nodes; |
742755a1 CL |
1481 | |
1482 | /* Check flags */ | |
1483 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1484 | return -EINVAL; | |
1485 | ||
1486 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1487 | return -EPERM; | |
1488 | ||
1489 | /* Find the mm_struct */ | |
a879bf58 | 1490 | rcu_read_lock(); |
228ebcbe | 1491 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1492 | if (!task) { |
a879bf58 | 1493 | rcu_read_unlock(); |
742755a1 CL |
1494 | return -ESRCH; |
1495 | } | |
3268c63e | 1496 | get_task_struct(task); |
742755a1 CL |
1497 | |
1498 | /* | |
1499 | * Check if this process has the right to modify the specified | |
1500 | * process. The right exists if the process has administrative | |
1501 | * capabilities, superuser privileges or the same | |
1502 | * userid as the target process. | |
1503 | */ | |
c69e8d9c | 1504 | tcred = __task_cred(task); |
b38a86eb EB |
1505 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1506 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1507 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1508 | rcu_read_unlock(); |
742755a1 | 1509 | err = -EPERM; |
5e9a0f02 | 1510 | goto out; |
742755a1 | 1511 | } |
c69e8d9c | 1512 | rcu_read_unlock(); |
742755a1 | 1513 | |
86c3a764 DQ |
1514 | err = security_task_movememory(task); |
1515 | if (err) | |
5e9a0f02 | 1516 | goto out; |
86c3a764 | 1517 | |
3268c63e CL |
1518 | task_nodes = cpuset_mems_allowed(task); |
1519 | mm = get_task_mm(task); | |
1520 | put_task_struct(task); | |
1521 | ||
6e8b09ea SL |
1522 | if (!mm) |
1523 | return -EINVAL; | |
1524 | ||
1525 | if (nodes) | |
1526 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1527 | nodes, status, flags); | |
1528 | else | |
1529 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1530 | |
742755a1 CL |
1531 | mmput(mm); |
1532 | return err; | |
3268c63e CL |
1533 | |
1534 | out: | |
1535 | put_task_struct(task); | |
1536 | return err; | |
742755a1 | 1537 | } |
742755a1 | 1538 | |
7039e1db PZ |
1539 | #ifdef CONFIG_NUMA_BALANCING |
1540 | /* | |
1541 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1542 | * pages. Currently it only checks the watermarks which crude | |
1543 | */ | |
1544 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1545 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1546 | { |
1547 | int z; | |
1548 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1549 | struct zone *zone = pgdat->node_zones + z; | |
1550 | ||
1551 | if (!populated_zone(zone)) | |
1552 | continue; | |
1553 | ||
6e543d57 | 1554 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1555 | continue; |
1556 | ||
1557 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1558 | if (!zone_watermark_ok(zone, 0, | |
1559 | high_wmark_pages(zone) + | |
1560 | nr_migrate_pages, | |
1561 | 0, 0)) | |
1562 | continue; | |
1563 | return true; | |
1564 | } | |
1565 | return false; | |
1566 | } | |
1567 | ||
1568 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1569 | unsigned long data, | |
1570 | int **result) | |
1571 | { | |
1572 | int nid = (int) data; | |
1573 | struct page *newpage; | |
1574 | ||
1575 | newpage = alloc_pages_exact_node(nid, | |
e97ca8e5 JW |
1576 | (GFP_HIGHUSER_MOVABLE | |
1577 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1578 | __GFP_NORETRY | __GFP_NOWARN) & | |
7039e1db | 1579 | ~GFP_IOFS, 0); |
bac0382c | 1580 | |
7039e1db PZ |
1581 | return newpage; |
1582 | } | |
1583 | ||
a8f60772 MG |
1584 | /* |
1585 | * page migration rate limiting control. | |
1586 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1587 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1588 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1589 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1590 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1591 | * throttle window closed. | |
a8f60772 MG |
1592 | */ |
1593 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1594 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1595 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1596 | ||
e14808b4 MG |
1597 | /* Returns true if NUMA migration is currently rate limited */ |
1598 | bool migrate_ratelimited(int node) | |
1599 | { | |
1600 | pg_data_t *pgdat = NODE_DATA(node); | |
1601 | ||
1602 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1603 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1604 | return false; | |
1605 | ||
1606 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1607 | return false; | |
1608 | ||
1609 | return true; | |
1610 | } | |
1611 | ||
b32967ff | 1612 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1613 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1614 | unsigned long nr_pages) | |
7039e1db | 1615 | { |
a8f60772 MG |
1616 | /* |
1617 | * Rate-limit the amount of data that is being migrated to a node. | |
1618 | * Optimal placement is no good if the memory bus is saturated and | |
1619 | * all the time is being spent migrating! | |
1620 | */ | |
a8f60772 | 1621 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1622 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1623 | pgdat->numabalancing_migrate_nr_pages = 0; |
1624 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1625 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1626 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1627 | } |
af1839d7 MG |
1628 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1629 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1630 | nr_pages); | |
1c5e9c27 | 1631 | return true; |
af1839d7 | 1632 | } |
1c5e9c27 MG |
1633 | |
1634 | /* | |
1635 | * This is an unlocked non-atomic update so errors are possible. | |
1636 | * The consequences are failing to migrate when we potentiall should | |
1637 | * have which is not severe enough to warrant locking. If it is ever | |
1638 | * a problem, it can be converted to a per-cpu counter. | |
1639 | */ | |
1640 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1641 | return false; | |
b32967ff MG |
1642 | } |
1643 | ||
1c30e017 | 1644 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1645 | { |
340ef390 | 1646 | int page_lru; |
a8f60772 | 1647 | |
309381fe | 1648 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1649 | |
7039e1db | 1650 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1651 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1652 | return 0; | |
7039e1db | 1653 | |
340ef390 HD |
1654 | if (isolate_lru_page(page)) |
1655 | return 0; | |
7039e1db | 1656 | |
340ef390 HD |
1657 | /* |
1658 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1659 | * check on page_count(), so we must do it here, now that the page | |
1660 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1661 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1662 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1663 | */ | |
1664 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1665 | putback_lru_page(page); | |
1666 | return 0; | |
7039e1db PZ |
1667 | } |
1668 | ||
340ef390 HD |
1669 | page_lru = page_is_file_cache(page); |
1670 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1671 | hpage_nr_pages(page)); | |
1672 | ||
149c33e1 | 1673 | /* |
340ef390 HD |
1674 | * Isolating the page has taken another reference, so the |
1675 | * caller's reference can be safely dropped without the page | |
1676 | * disappearing underneath us during migration. | |
149c33e1 MG |
1677 | */ |
1678 | put_page(page); | |
340ef390 | 1679 | return 1; |
b32967ff MG |
1680 | } |
1681 | ||
de466bd6 MG |
1682 | bool pmd_trans_migrating(pmd_t pmd) |
1683 | { | |
1684 | struct page *page = pmd_page(pmd); | |
1685 | return PageLocked(page); | |
1686 | } | |
1687 | ||
1688 | void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd) | |
1689 | { | |
1690 | struct page *page = pmd_page(*pmd); | |
1691 | wait_on_page_locked(page); | |
1692 | } | |
1693 | ||
b32967ff MG |
1694 | /* |
1695 | * Attempt to migrate a misplaced page to the specified destination | |
1696 | * node. Caller is expected to have an elevated reference count on | |
1697 | * the page that will be dropped by this function before returning. | |
1698 | */ | |
1bc115d8 MG |
1699 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1700 | int node) | |
b32967ff MG |
1701 | { |
1702 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1703 | int isolated; |
b32967ff MG |
1704 | int nr_remaining; |
1705 | LIST_HEAD(migratepages); | |
1706 | ||
1707 | /* | |
1bc115d8 MG |
1708 | * Don't migrate file pages that are mapped in multiple processes |
1709 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1710 | */ |
1bc115d8 MG |
1711 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1712 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1713 | goto out; |
b32967ff MG |
1714 | |
1715 | /* | |
1716 | * Rate-limit the amount of data that is being migrated to a node. | |
1717 | * Optimal placement is no good if the memory bus is saturated and | |
1718 | * all the time is being spent migrating! | |
1719 | */ | |
340ef390 | 1720 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1721 | goto out; |
b32967ff MG |
1722 | |
1723 | isolated = numamigrate_isolate_page(pgdat, page); | |
1724 | if (!isolated) | |
1725 | goto out; | |
1726 | ||
1727 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1728 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1729 | NULL, node, MIGRATE_ASYNC, |
1730 | MR_NUMA_MISPLACED); | |
b32967ff | 1731 | if (nr_remaining) { |
59c82b70 JK |
1732 | if (!list_empty(&migratepages)) { |
1733 | list_del(&page->lru); | |
1734 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1735 | page_is_file_cache(page)); | |
1736 | putback_lru_page(page); | |
1737 | } | |
b32967ff MG |
1738 | isolated = 0; |
1739 | } else | |
1740 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1741 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1742 | return isolated; |
340ef390 HD |
1743 | |
1744 | out: | |
1745 | put_page(page); | |
1746 | return 0; | |
7039e1db | 1747 | } |
220018d3 | 1748 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1749 | |
220018d3 | 1750 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1751 | /* |
1752 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1753 | * before returning. | |
1754 | */ | |
b32967ff MG |
1755 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1756 | struct vm_area_struct *vma, | |
1757 | pmd_t *pmd, pmd_t entry, | |
1758 | unsigned long address, | |
1759 | struct page *page, int node) | |
1760 | { | |
c4088ebd | 1761 | spinlock_t *ptl; |
b32967ff MG |
1762 | pg_data_t *pgdat = NODE_DATA(node); |
1763 | int isolated = 0; | |
1764 | struct page *new_page = NULL; | |
b32967ff | 1765 | int page_lru = page_is_file_cache(page); |
f714f4f2 MG |
1766 | unsigned long mmun_start = address & HPAGE_PMD_MASK; |
1767 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
2b4847e7 | 1768 | pmd_t orig_entry; |
b32967ff | 1769 | |
b32967ff MG |
1770 | /* |
1771 | * Rate-limit the amount of data that is being migrated to a node. | |
1772 | * Optimal placement is no good if the memory bus is saturated and | |
1773 | * all the time is being spent migrating! | |
1774 | */ | |
d28d4335 | 1775 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1776 | goto out_dropref; |
1777 | ||
1778 | new_page = alloc_pages_node(node, | |
e97ca8e5 JW |
1779 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT, |
1780 | HPAGE_PMD_ORDER); | |
340ef390 HD |
1781 | if (!new_page) |
1782 | goto out_fail; | |
1783 | ||
b32967ff | 1784 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 1785 | if (!isolated) { |
b32967ff | 1786 | put_page(new_page); |
340ef390 | 1787 | goto out_fail; |
b32967ff MG |
1788 | } |
1789 | ||
b0943d61 MG |
1790 | if (mm_tlb_flush_pending(mm)) |
1791 | flush_tlb_range(vma, mmun_start, mmun_end); | |
1792 | ||
b32967ff MG |
1793 | /* Prepare a page as a migration target */ |
1794 | __set_page_locked(new_page); | |
1795 | SetPageSwapBacked(new_page); | |
1796 | ||
1797 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1798 | new_page->mapping = page->mapping; | |
1799 | new_page->index = page->index; | |
1800 | migrate_page_copy(new_page, page); | |
1801 | WARN_ON(PageLRU(new_page)); | |
1802 | ||
1803 | /* Recheck the target PMD */ | |
f714f4f2 | 1804 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 1805 | ptl = pmd_lock(mm, pmd); |
2b4847e7 MG |
1806 | if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { |
1807 | fail_putback: | |
c4088ebd | 1808 | spin_unlock(ptl); |
f714f4f2 | 1809 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
1810 | |
1811 | /* Reverse changes made by migrate_page_copy() */ | |
1812 | if (TestClearPageActive(new_page)) | |
1813 | SetPageActive(page); | |
1814 | if (TestClearPageUnevictable(new_page)) | |
1815 | SetPageUnevictable(page); | |
1816 | mlock_migrate_page(page, new_page); | |
1817 | ||
1818 | unlock_page(new_page); | |
1819 | put_page(new_page); /* Free it */ | |
1820 | ||
a54a407f MG |
1821 | /* Retake the callers reference and putback on LRU */ |
1822 | get_page(page); | |
b32967ff | 1823 | putback_lru_page(page); |
a54a407f MG |
1824 | mod_zone_page_state(page_zone(page), |
1825 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
eb4489f6 MG |
1826 | |
1827 | goto out_unlock; | |
b32967ff MG |
1828 | } |
1829 | ||
2b4847e7 | 1830 | orig_entry = *pmd; |
b32967ff | 1831 | entry = mk_pmd(new_page, vma->vm_page_prot); |
b32967ff | 1832 | entry = pmd_mkhuge(entry); |
2b4847e7 | 1833 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 1834 | |
2b4847e7 MG |
1835 | /* |
1836 | * Clear the old entry under pagetable lock and establish the new PTE. | |
1837 | * Any parallel GUP will either observe the old page blocking on the | |
1838 | * page lock, block on the page table lock or observe the new page. | |
1839 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
1840 | * guarantee the copy is visible before the pagetable update. | |
1841 | */ | |
f714f4f2 | 1842 | flush_cache_range(vma, mmun_start, mmun_end); |
11de9927 | 1843 | page_add_anon_rmap(new_page, vma, mmun_start); |
34ee645e | 1844 | pmdp_clear_flush_notify(vma, mmun_start, pmd); |
f714f4f2 MG |
1845 | set_pmd_at(mm, mmun_start, pmd, entry); |
1846 | flush_tlb_range(vma, mmun_start, mmun_end); | |
ce4a9cc5 | 1847 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 MG |
1848 | |
1849 | if (page_count(page) != 2) { | |
f714f4f2 MG |
1850 | set_pmd_at(mm, mmun_start, pmd, orig_entry); |
1851 | flush_tlb_range(vma, mmun_start, mmun_end); | |
34ee645e | 1852 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
2b4847e7 MG |
1853 | update_mmu_cache_pmd(vma, address, &entry); |
1854 | page_remove_rmap(new_page); | |
1855 | goto fail_putback; | |
1856 | } | |
1857 | ||
0a31bc97 JW |
1858 | mem_cgroup_migrate(page, new_page, false); |
1859 | ||
b32967ff | 1860 | page_remove_rmap(page); |
2b4847e7 | 1861 | |
c4088ebd | 1862 | spin_unlock(ptl); |
f714f4f2 | 1863 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff | 1864 | |
11de9927 MG |
1865 | /* Take an "isolate" reference and put new page on the LRU. */ |
1866 | get_page(new_page); | |
1867 | putback_lru_page(new_page); | |
1868 | ||
b32967ff MG |
1869 | unlock_page(new_page); |
1870 | unlock_page(page); | |
1871 | put_page(page); /* Drop the rmap reference */ | |
1872 | put_page(page); /* Drop the LRU isolation reference */ | |
1873 | ||
1874 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1875 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1876 | ||
b32967ff MG |
1877 | mod_zone_page_state(page_zone(page), |
1878 | NR_ISOLATED_ANON + page_lru, | |
1879 | -HPAGE_PMD_NR); | |
1880 | return isolated; | |
1881 | ||
340ef390 HD |
1882 | out_fail: |
1883 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1884 | out_dropref: |
2b4847e7 MG |
1885 | ptl = pmd_lock(mm, pmd); |
1886 | if (pmd_same(*pmd, entry)) { | |
1887 | entry = pmd_mknonnuma(entry); | |
f714f4f2 | 1888 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
1889 | update_mmu_cache_pmd(vma, address, &entry); |
1890 | } | |
1891 | spin_unlock(ptl); | |
a54a407f | 1892 | |
eb4489f6 | 1893 | out_unlock: |
340ef390 | 1894 | unlock_page(page); |
b32967ff | 1895 | put_page(page); |
b32967ff MG |
1896 | return 0; |
1897 | } | |
7039e1db PZ |
1898 | #endif /* CONFIG_NUMA_BALANCING */ |
1899 | ||
1900 | #endif /* CONFIG_NUMA */ |