]>
Commit | Line | Data |
---|---|---|
b20a3503 | 1 | /* |
14e0f9bc | 2 | * Memory Migration functionality - linux/mm/migrate.c |
b20a3503 CL |
3 | * |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <[email protected]> | |
10 | * Hirokazu Takahashi <[email protected]> | |
11 | * Dave Hansen <[email protected]> | |
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> |
42cb14b1 | 33 | #include <linux/backing-dev.h> |
bda807d4 | 34 | #include <linux/compaction.h> |
4f5ca265 | 35 | #include <linux/syscalls.h> |
290408d4 | 36 | #include <linux/hugetlb.h> |
8e6ac7fa | 37 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 38 | #include <linux/gfp.h> |
df6ad698 | 39 | #include <linux/pfn_t.h> |
a5430dda | 40 | #include <linux/memremap.h> |
8315ada7 | 41 | #include <linux/userfaultfd_k.h> |
bf6bddf1 | 42 | #include <linux/balloon_compaction.h> |
f714f4f2 | 43 | #include <linux/mmu_notifier.h> |
33c3fc71 | 44 | #include <linux/page_idle.h> |
d435edca | 45 | #include <linux/page_owner.h> |
6e84f315 | 46 | #include <linux/sched/mm.h> |
197e7e52 | 47 | #include <linux/ptrace.h> |
b20a3503 | 48 | |
0d1836c3 MN |
49 | #include <asm/tlbflush.h> |
50 | ||
7b2a2d4a MG |
51 | #define CREATE_TRACE_POINTS |
52 | #include <trace/events/migrate.h> | |
53 | ||
b20a3503 CL |
54 | #include "internal.h" |
55 | ||
b20a3503 | 56 | /* |
742755a1 | 57 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
58 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
59 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
60 | */ |
61 | int migrate_prep(void) | |
62 | { | |
b20a3503 CL |
63 | /* |
64 | * Clear the LRU lists so pages can be isolated. | |
65 | * Note that pages may be moved off the LRU after we have | |
66 | * drained them. Those pages will fail to migrate like other | |
67 | * pages that may be busy. | |
68 | */ | |
69 | lru_add_drain_all(); | |
70 | ||
71 | return 0; | |
72 | } | |
73 | ||
748446bb MG |
74 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
75 | int migrate_prep_local(void) | |
76 | { | |
77 | lru_add_drain(); | |
78 | ||
79 | return 0; | |
80 | } | |
81 | ||
9e5bcd61 | 82 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
83 | { |
84 | struct address_space *mapping; | |
85 | ||
86 | /* | |
87 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
88 | * or just got freed under us. | |
89 | * | |
90 | * In case we 'win' a race for a movable page being freed under us and | |
91 | * raise its refcount preventing __free_pages() from doing its job | |
92 | * the put_page() at the end of this block will take care of | |
93 | * release this page, thus avoiding a nasty leakage. | |
94 | */ | |
95 | if (unlikely(!get_page_unless_zero(page))) | |
96 | goto out; | |
97 | ||
98 | /* | |
99 | * Check PageMovable before holding a PG_lock because page's owner | |
100 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
101 | * so unconditionally grapping the lock ruins page's owner side. | |
102 | */ | |
103 | if (unlikely(!__PageMovable(page))) | |
104 | goto out_putpage; | |
105 | /* | |
106 | * As movable pages are not isolated from LRU lists, concurrent | |
107 | * compaction threads can race against page migration functions | |
108 | * as well as race against the releasing a page. | |
109 | * | |
110 | * In order to avoid having an already isolated movable page | |
111 | * being (wrongly) re-isolated while it is under migration, | |
112 | * or to avoid attempting to isolate pages being released, | |
113 | * lets be sure we have the page lock | |
114 | * before proceeding with the movable page isolation steps. | |
115 | */ | |
116 | if (unlikely(!trylock_page(page))) | |
117 | goto out_putpage; | |
118 | ||
119 | if (!PageMovable(page) || PageIsolated(page)) | |
120 | goto out_no_isolated; | |
121 | ||
122 | mapping = page_mapping(page); | |
123 | VM_BUG_ON_PAGE(!mapping, page); | |
124 | ||
125 | if (!mapping->a_ops->isolate_page(page, mode)) | |
126 | goto out_no_isolated; | |
127 | ||
128 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
129 | WARN_ON_ONCE(PageIsolated(page)); | |
130 | __SetPageIsolated(page); | |
131 | unlock_page(page); | |
132 | ||
9e5bcd61 | 133 | return 0; |
bda807d4 MK |
134 | |
135 | out_no_isolated: | |
136 | unlock_page(page); | |
137 | out_putpage: | |
138 | put_page(page); | |
139 | out: | |
9e5bcd61 | 140 | return -EBUSY; |
bda807d4 MK |
141 | } |
142 | ||
143 | /* It should be called on page which is PG_movable */ | |
144 | void putback_movable_page(struct page *page) | |
145 | { | |
146 | struct address_space *mapping; | |
147 | ||
148 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
149 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
150 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
151 | ||
152 | mapping = page_mapping(page); | |
153 | mapping->a_ops->putback_page(page); | |
154 | __ClearPageIsolated(page); | |
155 | } | |
156 | ||
5733c7d1 RA |
157 | /* |
158 | * Put previously isolated pages back onto the appropriate lists | |
159 | * from where they were once taken off for compaction/migration. | |
160 | * | |
59c82b70 JK |
161 | * This function shall be used whenever the isolated pageset has been |
162 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
163 | * and isolate_huge_page(). | |
5733c7d1 RA |
164 | */ |
165 | void putback_movable_pages(struct list_head *l) | |
166 | { | |
167 | struct page *page; | |
168 | struct page *page2; | |
169 | ||
b20a3503 | 170 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
171 | if (unlikely(PageHuge(page))) { |
172 | putback_active_hugepage(page); | |
173 | continue; | |
174 | } | |
e24f0b8f | 175 | list_del(&page->lru); |
bda807d4 MK |
176 | /* |
177 | * We isolated non-lru movable page so here we can use | |
178 | * __PageMovable because LRU page's mapping cannot have | |
179 | * PAGE_MAPPING_MOVABLE. | |
180 | */ | |
b1123ea6 | 181 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
182 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
183 | lock_page(page); | |
184 | if (PageMovable(page)) | |
185 | putback_movable_page(page); | |
186 | else | |
187 | __ClearPageIsolated(page); | |
188 | unlock_page(page); | |
189 | put_page(page); | |
190 | } else { | |
e8db67eb NH |
191 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
192 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
fc280fe8 | 193 | putback_lru_page(page); |
bda807d4 | 194 | } |
b20a3503 | 195 | } |
b20a3503 CL |
196 | } |
197 | ||
0697212a CL |
198 | /* |
199 | * Restore a potential migration pte to a working pte entry | |
200 | */ | |
e4b82222 | 201 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 202 | unsigned long addr, void *old) |
0697212a | 203 | { |
3fe87967 KS |
204 | struct page_vma_mapped_walk pvmw = { |
205 | .page = old, | |
206 | .vma = vma, | |
207 | .address = addr, | |
208 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
209 | }; | |
210 | struct page *new; | |
211 | pte_t pte; | |
0697212a | 212 | swp_entry_t entry; |
0697212a | 213 | |
3fe87967 KS |
214 | VM_BUG_ON_PAGE(PageTail(page), page); |
215 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
216 | if (PageKsm(page)) |
217 | new = page; | |
218 | else | |
219 | new = page - pvmw.page->index + | |
220 | linear_page_index(vma, pvmw.address); | |
0697212a | 221 | |
616b8371 ZY |
222 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
223 | /* PMD-mapped THP migration entry */ | |
224 | if (!pvmw.pte) { | |
225 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
226 | remove_migration_pmd(&pvmw, new); | |
227 | continue; | |
228 | } | |
229 | #endif | |
230 | ||
3fe87967 KS |
231 | get_page(new); |
232 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
233 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
234 | pte = pte_mksoft_dirty(pte); | |
0697212a | 235 | |
3fe87967 KS |
236 | /* |
237 | * Recheck VMA as permissions can change since migration started | |
238 | */ | |
239 | entry = pte_to_swp_entry(*pvmw.pte); | |
240 | if (is_write_migration_entry(entry)) | |
241 | pte = maybe_mkwrite(pte, vma); | |
d3cb8bf6 | 242 | |
df6ad698 JG |
243 | if (unlikely(is_zone_device_page(new))) { |
244 | if (is_device_private_page(new)) { | |
245 | entry = make_device_private_entry(new, pte_write(pte)); | |
246 | pte = swp_entry_to_pte(entry); | |
247 | } else if (is_device_public_page(new)) { | |
248 | pte = pte_mkdevmap(pte); | |
249 | flush_dcache_page(new); | |
250 | } | |
a5430dda JG |
251 | } else |
252 | flush_dcache_page(new); | |
253 | ||
3ef8fd7f | 254 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 KS |
255 | if (PageHuge(new)) { |
256 | pte = pte_mkhuge(pte); | |
257 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
383321ab | 258 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
259 | if (PageAnon(new)) |
260 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
261 | else | |
262 | page_dup_rmap(new, true); | |
383321ab AK |
263 | } else |
264 | #endif | |
265 | { | |
266 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 267 | |
383321ab AK |
268 | if (PageAnon(new)) |
269 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
270 | else | |
271 | page_add_file_rmap(new, false); | |
272 | } | |
3fe87967 KS |
273 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
274 | mlock_vma_page(new); | |
275 | ||
276 | /* No need to invalidate - it was non-present before */ | |
277 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
278 | } | |
51afb12b | 279 | |
e4b82222 | 280 | return true; |
0697212a CL |
281 | } |
282 | ||
04e62a29 CL |
283 | /* |
284 | * Get rid of all migration entries and replace them by | |
285 | * references to the indicated page. | |
286 | */ | |
e388466d | 287 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 288 | { |
051ac83a JK |
289 | struct rmap_walk_control rwc = { |
290 | .rmap_one = remove_migration_pte, | |
291 | .arg = old, | |
292 | }; | |
293 | ||
e388466d KS |
294 | if (locked) |
295 | rmap_walk_locked(new, &rwc); | |
296 | else | |
297 | rmap_walk(new, &rwc); | |
04e62a29 CL |
298 | } |
299 | ||
0697212a CL |
300 | /* |
301 | * Something used the pte of a page under migration. We need to | |
302 | * get to the page and wait until migration is finished. | |
303 | * When we return from this function the fault will be retried. | |
0697212a | 304 | */ |
e66f17ff | 305 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 306 | spinlock_t *ptl) |
0697212a | 307 | { |
30dad309 | 308 | pte_t pte; |
0697212a CL |
309 | swp_entry_t entry; |
310 | struct page *page; | |
311 | ||
30dad309 | 312 | spin_lock(ptl); |
0697212a CL |
313 | pte = *ptep; |
314 | if (!is_swap_pte(pte)) | |
315 | goto out; | |
316 | ||
317 | entry = pte_to_swp_entry(pte); | |
318 | if (!is_migration_entry(entry)) | |
319 | goto out; | |
320 | ||
321 | page = migration_entry_to_page(entry); | |
322 | ||
e286781d NP |
323 | /* |
324 | * Once radix-tree replacement of page migration started, page_count | |
325 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
326 | * against a page without get_page(). | |
327 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
328 | * will occur again. | |
329 | */ | |
330 | if (!get_page_unless_zero(page)) | |
331 | goto out; | |
0697212a CL |
332 | pte_unmap_unlock(ptep, ptl); |
333 | wait_on_page_locked(page); | |
334 | put_page(page); | |
335 | return; | |
336 | out: | |
337 | pte_unmap_unlock(ptep, ptl); | |
338 | } | |
339 | ||
30dad309 NH |
340 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
341 | unsigned long address) | |
342 | { | |
343 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
344 | pte_t *ptep = pte_offset_map(pmd, address); | |
345 | __migration_entry_wait(mm, ptep, ptl); | |
346 | } | |
347 | ||
cb900f41 KS |
348 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
349 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 350 | { |
cb900f41 | 351 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
352 | __migration_entry_wait(mm, pte, ptl); |
353 | } | |
354 | ||
616b8371 ZY |
355 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
356 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
357 | { | |
358 | spinlock_t *ptl; | |
359 | struct page *page; | |
360 | ||
361 | ptl = pmd_lock(mm, pmd); | |
362 | if (!is_pmd_migration_entry(*pmd)) | |
363 | goto unlock; | |
364 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
365 | if (!get_page_unless_zero(page)) | |
366 | goto unlock; | |
367 | spin_unlock(ptl); | |
368 | wait_on_page_locked(page); | |
369 | put_page(page); | |
370 | return; | |
371 | unlock: | |
372 | spin_unlock(ptl); | |
373 | } | |
374 | #endif | |
375 | ||
b969c4ab MG |
376 | #ifdef CONFIG_BLOCK |
377 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
378 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
379 | enum migrate_mode mode) | |
b969c4ab MG |
380 | { |
381 | struct buffer_head *bh = head; | |
382 | ||
383 | /* Simple case, sync compaction */ | |
a6bc32b8 | 384 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
385 | do { |
386 | get_bh(bh); | |
387 | lock_buffer(bh); | |
388 | bh = bh->b_this_page; | |
389 | ||
390 | } while (bh != head); | |
391 | ||
392 | return true; | |
393 | } | |
394 | ||
395 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
396 | do { | |
397 | get_bh(bh); | |
398 | if (!trylock_buffer(bh)) { | |
399 | /* | |
400 | * We failed to lock the buffer and cannot stall in | |
401 | * async migration. Release the taken locks | |
402 | */ | |
403 | struct buffer_head *failed_bh = bh; | |
404 | put_bh(failed_bh); | |
405 | bh = head; | |
406 | while (bh != failed_bh) { | |
407 | unlock_buffer(bh); | |
408 | put_bh(bh); | |
409 | bh = bh->b_this_page; | |
410 | } | |
411 | return false; | |
412 | } | |
413 | ||
414 | bh = bh->b_this_page; | |
415 | } while (bh != head); | |
416 | return true; | |
417 | } | |
418 | #else | |
419 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 420 | enum migrate_mode mode) |
b969c4ab MG |
421 | { |
422 | return true; | |
423 | } | |
424 | #endif /* CONFIG_BLOCK */ | |
425 | ||
b20a3503 | 426 | /* |
c3fcf8a5 | 427 | * Replace the page in the mapping. |
5b5c7120 CL |
428 | * |
429 | * The number of remaining references must be: | |
430 | * 1 for anonymous pages without a mapping | |
431 | * 2 for pages with a mapping | |
266cf658 | 432 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 433 | */ |
36bc08cc | 434 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 435 | struct page *newpage, struct page *page, |
8e321fef BL |
436 | struct buffer_head *head, enum migrate_mode mode, |
437 | int extra_count) | |
b20a3503 | 438 | { |
42cb14b1 HD |
439 | struct zone *oldzone, *newzone; |
440 | int dirty; | |
8e321fef | 441 | int expected_count = 1 + extra_count; |
7cf9c2c7 | 442 | void **pslot; |
b20a3503 | 443 | |
8763cb45 | 444 | /* |
df6ad698 JG |
445 | * Device public or private pages have an extra refcount as they are |
446 | * ZONE_DEVICE pages. | |
8763cb45 | 447 | */ |
df6ad698 JG |
448 | expected_count += is_device_private_page(page); |
449 | expected_count += is_device_public_page(page); | |
8763cb45 | 450 | |
6c5240ae | 451 | if (!mapping) { |
0e8c7d0f | 452 | /* Anonymous page without mapping */ |
8e321fef | 453 | if (page_count(page) != expected_count) |
6c5240ae | 454 | return -EAGAIN; |
cf4b769a HD |
455 | |
456 | /* No turning back from here */ | |
cf4b769a HD |
457 | newpage->index = page->index; |
458 | newpage->mapping = page->mapping; | |
459 | if (PageSwapBacked(page)) | |
fa9949da | 460 | __SetPageSwapBacked(newpage); |
cf4b769a | 461 | |
78bd5209 | 462 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
463 | } |
464 | ||
42cb14b1 HD |
465 | oldzone = page_zone(page); |
466 | newzone = page_zone(newpage); | |
467 | ||
19fd6231 | 468 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 469 | |
7cf9c2c7 NP |
470 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
471 | page_index(page)); | |
b20a3503 | 472 | |
8e321fef | 473 | expected_count += 1 + page_has_private(page); |
e286781d | 474 | if (page_count(page) != expected_count || |
29c1f677 | 475 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 476 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 477 | return -EAGAIN; |
b20a3503 CL |
478 | } |
479 | ||
fe896d18 | 480 | if (!page_ref_freeze(page, expected_count)) { |
19fd6231 | 481 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
482 | return -EAGAIN; |
483 | } | |
484 | ||
b969c4ab MG |
485 | /* |
486 | * In the async migration case of moving a page with buffers, lock the | |
487 | * buffers using trylock before the mapping is moved. If the mapping | |
488 | * was moved, we later failed to lock the buffers and could not move | |
489 | * the mapping back due to an elevated page count, we would have to | |
490 | * block waiting on other references to be dropped. | |
491 | */ | |
a6bc32b8 MG |
492 | if (mode == MIGRATE_ASYNC && head && |
493 | !buffer_migrate_lock_buffers(head, mode)) { | |
fe896d18 | 494 | page_ref_unfreeze(page, expected_count); |
b969c4ab MG |
495 | spin_unlock_irq(&mapping->tree_lock); |
496 | return -EAGAIN; | |
497 | } | |
498 | ||
b20a3503 | 499 | /* |
cf4b769a HD |
500 | * Now we know that no one else is looking at the page: |
501 | * no turning back from here. | |
b20a3503 | 502 | */ |
cf4b769a HD |
503 | newpage->index = page->index; |
504 | newpage->mapping = page->mapping; | |
7cf9c2c7 | 505 | get_page(newpage); /* add cache reference */ |
6326fec1 NP |
506 | if (PageSwapBacked(page)) { |
507 | __SetPageSwapBacked(newpage); | |
508 | if (PageSwapCache(page)) { | |
509 | SetPageSwapCache(newpage); | |
510 | set_page_private(newpage, page_private(page)); | |
511 | } | |
512 | } else { | |
513 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
514 | } |
515 | ||
42cb14b1 HD |
516 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
517 | dirty = PageDirty(page); | |
518 | if (dirty) { | |
519 | ClearPageDirty(page); | |
520 | SetPageDirty(newpage); | |
521 | } | |
522 | ||
6d75f366 | 523 | radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); |
7cf9c2c7 NP |
524 | |
525 | /* | |
937a94c9 JG |
526 | * Drop cache reference from old page by unfreezing |
527 | * to one less reference. | |
7cf9c2c7 NP |
528 | * We know this isn't the last reference. |
529 | */ | |
fe896d18 | 530 | page_ref_unfreeze(page, expected_count - 1); |
7cf9c2c7 | 531 | |
42cb14b1 HD |
532 | spin_unlock(&mapping->tree_lock); |
533 | /* Leave irq disabled to prevent preemption while updating stats */ | |
534 | ||
0e8c7d0f CL |
535 | /* |
536 | * If moved to a different zone then also account | |
537 | * the page for that zone. Other VM counters will be | |
538 | * taken care of when we establish references to the | |
539 | * new page and drop references to the old page. | |
540 | * | |
541 | * Note that anonymous pages are accounted for | |
4b9d0fab | 542 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
543 | * are mapped to swap space. |
544 | */ | |
42cb14b1 | 545 | if (newzone != oldzone) { |
11fb9989 MG |
546 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); |
547 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); | |
42cb14b1 | 548 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
11fb9989 MG |
549 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); |
550 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); | |
42cb14b1 HD |
551 | } |
552 | if (dirty && mapping_cap_account_dirty(mapping)) { | |
11fb9989 | 553 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 554 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); |
11fb9989 | 555 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 556 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); |
42cb14b1 | 557 | } |
4b02108a | 558 | } |
42cb14b1 | 559 | local_irq_enable(); |
b20a3503 | 560 | |
78bd5209 | 561 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 562 | } |
1118dce7 | 563 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 564 | |
290408d4 NH |
565 | /* |
566 | * The expected number of remaining references is the same as that | |
567 | * of migrate_page_move_mapping(). | |
568 | */ | |
569 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
570 | struct page *newpage, struct page *page) | |
571 | { | |
572 | int expected_count; | |
573 | void **pslot; | |
574 | ||
290408d4 NH |
575 | spin_lock_irq(&mapping->tree_lock); |
576 | ||
577 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
578 | page_index(page)); | |
579 | ||
580 | expected_count = 2 + page_has_private(page); | |
581 | if (page_count(page) != expected_count || | |
29c1f677 | 582 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
583 | spin_unlock_irq(&mapping->tree_lock); |
584 | return -EAGAIN; | |
585 | } | |
586 | ||
fe896d18 | 587 | if (!page_ref_freeze(page, expected_count)) { |
290408d4 NH |
588 | spin_unlock_irq(&mapping->tree_lock); |
589 | return -EAGAIN; | |
590 | } | |
591 | ||
cf4b769a HD |
592 | newpage->index = page->index; |
593 | newpage->mapping = page->mapping; | |
6a93ca8f | 594 | |
290408d4 NH |
595 | get_page(newpage); |
596 | ||
6d75f366 | 597 | radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); |
290408d4 | 598 | |
fe896d18 | 599 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 NH |
600 | |
601 | spin_unlock_irq(&mapping->tree_lock); | |
6a93ca8f | 602 | |
78bd5209 | 603 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
604 | } |
605 | ||
30b0a105 DH |
606 | /* |
607 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
608 | * arithmetic will work across the entire page. We need something more | |
609 | * specialized. | |
610 | */ | |
611 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
612 | int nr_pages) | |
613 | { | |
614 | int i; | |
615 | struct page *dst_base = dst; | |
616 | struct page *src_base = src; | |
617 | ||
618 | for (i = 0; i < nr_pages; ) { | |
619 | cond_resched(); | |
620 | copy_highpage(dst, src); | |
621 | ||
622 | i++; | |
623 | dst = mem_map_next(dst, dst_base, i); | |
624 | src = mem_map_next(src, src_base, i); | |
625 | } | |
626 | } | |
627 | ||
628 | static void copy_huge_page(struct page *dst, struct page *src) | |
629 | { | |
630 | int i; | |
631 | int nr_pages; | |
632 | ||
633 | if (PageHuge(src)) { | |
634 | /* hugetlbfs page */ | |
635 | struct hstate *h = page_hstate(src); | |
636 | nr_pages = pages_per_huge_page(h); | |
637 | ||
638 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
639 | __copy_gigantic_page(dst, src, nr_pages); | |
640 | return; | |
641 | } | |
642 | } else { | |
643 | /* thp page */ | |
644 | BUG_ON(!PageTransHuge(src)); | |
645 | nr_pages = hpage_nr_pages(src); | |
646 | } | |
647 | ||
648 | for (i = 0; i < nr_pages; i++) { | |
649 | cond_resched(); | |
650 | copy_highpage(dst + i, src + i); | |
651 | } | |
652 | } | |
653 | ||
b20a3503 CL |
654 | /* |
655 | * Copy the page to its new location | |
656 | */ | |
2916ecc0 | 657 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 658 | { |
7851a45c RR |
659 | int cpupid; |
660 | ||
b20a3503 CL |
661 | if (PageError(page)) |
662 | SetPageError(newpage); | |
663 | if (PageReferenced(page)) | |
664 | SetPageReferenced(newpage); | |
665 | if (PageUptodate(page)) | |
666 | SetPageUptodate(newpage); | |
894bc310 | 667 | if (TestClearPageActive(page)) { |
309381fe | 668 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 669 | SetPageActive(newpage); |
418b27ef LS |
670 | } else if (TestClearPageUnevictable(page)) |
671 | SetPageUnevictable(newpage); | |
b20a3503 CL |
672 | if (PageChecked(page)) |
673 | SetPageChecked(newpage); | |
674 | if (PageMappedToDisk(page)) | |
675 | SetPageMappedToDisk(newpage); | |
676 | ||
42cb14b1 HD |
677 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
678 | if (PageDirty(page)) | |
679 | SetPageDirty(newpage); | |
b20a3503 | 680 | |
33c3fc71 VD |
681 | if (page_is_young(page)) |
682 | set_page_young(newpage); | |
683 | if (page_is_idle(page)) | |
684 | set_page_idle(newpage); | |
685 | ||
7851a45c RR |
686 | /* |
687 | * Copy NUMA information to the new page, to prevent over-eager | |
688 | * future migrations of this same page. | |
689 | */ | |
690 | cpupid = page_cpupid_xchg_last(page, -1); | |
691 | page_cpupid_xchg_last(newpage, cpupid); | |
692 | ||
e9995ef9 | 693 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
694 | /* |
695 | * Please do not reorder this without considering how mm/ksm.c's | |
696 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
697 | */ | |
b3b3a99c NH |
698 | if (PageSwapCache(page)) |
699 | ClearPageSwapCache(page); | |
b20a3503 CL |
700 | ClearPagePrivate(page); |
701 | set_page_private(page, 0); | |
b20a3503 CL |
702 | |
703 | /* | |
704 | * If any waiters have accumulated on the new page then | |
705 | * wake them up. | |
706 | */ | |
707 | if (PageWriteback(newpage)) | |
708 | end_page_writeback(newpage); | |
d435edca VB |
709 | |
710 | copy_page_owner(page, newpage); | |
74485cf2 JW |
711 | |
712 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 713 | } |
2916ecc0 JG |
714 | EXPORT_SYMBOL(migrate_page_states); |
715 | ||
716 | void migrate_page_copy(struct page *newpage, struct page *page) | |
717 | { | |
718 | if (PageHuge(page) || PageTransHuge(page)) | |
719 | copy_huge_page(newpage, page); | |
720 | else | |
721 | copy_highpage(newpage, page); | |
722 | ||
723 | migrate_page_states(newpage, page); | |
724 | } | |
1118dce7 | 725 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 726 | |
1d8b85cc CL |
727 | /************************************************************ |
728 | * Migration functions | |
729 | ***********************************************************/ | |
730 | ||
b20a3503 | 731 | /* |
bda807d4 | 732 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 733 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
734 | * |
735 | * Pages are locked upon entry and exit. | |
736 | */ | |
2d1db3b1 | 737 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
738 | struct page *newpage, struct page *page, |
739 | enum migrate_mode mode) | |
b20a3503 CL |
740 | { |
741 | int rc; | |
742 | ||
743 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
744 | ||
8e321fef | 745 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 746 | |
78bd5209 | 747 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
748 | return rc; |
749 | ||
2916ecc0 JG |
750 | if (mode != MIGRATE_SYNC_NO_COPY) |
751 | migrate_page_copy(newpage, page); | |
752 | else | |
753 | migrate_page_states(newpage, page); | |
78bd5209 | 754 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
755 | } |
756 | EXPORT_SYMBOL(migrate_page); | |
757 | ||
9361401e | 758 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
759 | /* |
760 | * Migration function for pages with buffers. This function can only be used | |
761 | * if the underlying filesystem guarantees that no other references to "page" | |
762 | * exist. | |
763 | */ | |
2d1db3b1 | 764 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 765 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 766 | { |
1d8b85cc CL |
767 | struct buffer_head *bh, *head; |
768 | int rc; | |
769 | ||
1d8b85cc | 770 | if (!page_has_buffers(page)) |
a6bc32b8 | 771 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
772 | |
773 | head = page_buffers(page); | |
774 | ||
8e321fef | 775 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); |
1d8b85cc | 776 | |
78bd5209 | 777 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
778 | return rc; |
779 | ||
b969c4ab MG |
780 | /* |
781 | * In the async case, migrate_page_move_mapping locked the buffers | |
782 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
783 | * need to be locked now | |
784 | */ | |
a6bc32b8 MG |
785 | if (mode != MIGRATE_ASYNC) |
786 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
787 | |
788 | ClearPagePrivate(page); | |
789 | set_page_private(newpage, page_private(page)); | |
790 | set_page_private(page, 0); | |
791 | put_page(page); | |
792 | get_page(newpage); | |
793 | ||
794 | bh = head; | |
795 | do { | |
796 | set_bh_page(bh, newpage, bh_offset(bh)); | |
797 | bh = bh->b_this_page; | |
798 | ||
799 | } while (bh != head); | |
800 | ||
801 | SetPagePrivate(newpage); | |
802 | ||
2916ecc0 JG |
803 | if (mode != MIGRATE_SYNC_NO_COPY) |
804 | migrate_page_copy(newpage, page); | |
805 | else | |
806 | migrate_page_states(newpage, page); | |
1d8b85cc CL |
807 | |
808 | bh = head; | |
809 | do { | |
810 | unlock_buffer(bh); | |
2916ecc0 | 811 | put_bh(bh); |
1d8b85cc CL |
812 | bh = bh->b_this_page; |
813 | ||
814 | } while (bh != head); | |
815 | ||
78bd5209 | 816 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
817 | } |
818 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 819 | #endif |
1d8b85cc | 820 | |
04e62a29 CL |
821 | /* |
822 | * Writeback a page to clean the dirty state | |
823 | */ | |
824 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 825 | { |
04e62a29 CL |
826 | struct writeback_control wbc = { |
827 | .sync_mode = WB_SYNC_NONE, | |
828 | .nr_to_write = 1, | |
829 | .range_start = 0, | |
830 | .range_end = LLONG_MAX, | |
04e62a29 CL |
831 | .for_reclaim = 1 |
832 | }; | |
833 | int rc; | |
834 | ||
835 | if (!mapping->a_ops->writepage) | |
836 | /* No write method for the address space */ | |
837 | return -EINVAL; | |
838 | ||
839 | if (!clear_page_dirty_for_io(page)) | |
840 | /* Someone else already triggered a write */ | |
841 | return -EAGAIN; | |
842 | ||
8351a6e4 | 843 | /* |
04e62a29 CL |
844 | * A dirty page may imply that the underlying filesystem has |
845 | * the page on some queue. So the page must be clean for | |
846 | * migration. Writeout may mean we loose the lock and the | |
847 | * page state is no longer what we checked for earlier. | |
848 | * At this point we know that the migration attempt cannot | |
849 | * be successful. | |
8351a6e4 | 850 | */ |
e388466d | 851 | remove_migration_ptes(page, page, false); |
8351a6e4 | 852 | |
04e62a29 | 853 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 854 | |
04e62a29 CL |
855 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
856 | /* unlocked. Relock */ | |
857 | lock_page(page); | |
858 | ||
bda8550d | 859 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
860 | } |
861 | ||
862 | /* | |
863 | * Default handling if a filesystem does not provide a migration function. | |
864 | */ | |
865 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 866 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 867 | { |
b969c4ab | 868 | if (PageDirty(page)) { |
a6bc32b8 | 869 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
870 | switch (mode) { |
871 | case MIGRATE_SYNC: | |
872 | case MIGRATE_SYNC_NO_COPY: | |
873 | break; | |
874 | default: | |
b969c4ab | 875 | return -EBUSY; |
2916ecc0 | 876 | } |
04e62a29 | 877 | return writeout(mapping, page); |
b969c4ab | 878 | } |
8351a6e4 CL |
879 | |
880 | /* | |
881 | * Buffers may be managed in a filesystem specific way. | |
882 | * We must have no buffers or drop them. | |
883 | */ | |
266cf658 | 884 | if (page_has_private(page) && |
8351a6e4 CL |
885 | !try_to_release_page(page, GFP_KERNEL)) |
886 | return -EAGAIN; | |
887 | ||
a6bc32b8 | 888 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
889 | } |
890 | ||
e24f0b8f CL |
891 | /* |
892 | * Move a page to a newly allocated page | |
893 | * The page is locked and all ptes have been successfully removed. | |
894 | * | |
895 | * The new page will have replaced the old page if this function | |
896 | * is successful. | |
894bc310 LS |
897 | * |
898 | * Return value: | |
899 | * < 0 - error code | |
78bd5209 | 900 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 901 | */ |
3fe2011f | 902 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 903 | enum migrate_mode mode) |
e24f0b8f CL |
904 | { |
905 | struct address_space *mapping; | |
bda807d4 MK |
906 | int rc = -EAGAIN; |
907 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 908 | |
7db7671f HD |
909 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
910 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 911 | |
e24f0b8f | 912 | mapping = page_mapping(page); |
bda807d4 MK |
913 | |
914 | if (likely(is_lru)) { | |
915 | if (!mapping) | |
916 | rc = migrate_page(mapping, newpage, page, mode); | |
917 | else if (mapping->a_ops->migratepage) | |
918 | /* | |
919 | * Most pages have a mapping and most filesystems | |
920 | * provide a migratepage callback. Anonymous pages | |
921 | * are part of swap space which also has its own | |
922 | * migratepage callback. This is the most common path | |
923 | * for page migration. | |
924 | */ | |
925 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
926 | page, mode); | |
927 | else | |
928 | rc = fallback_migrate_page(mapping, newpage, | |
929 | page, mode); | |
930 | } else { | |
e24f0b8f | 931 | /* |
bda807d4 MK |
932 | * In case of non-lru page, it could be released after |
933 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 934 | */ |
bda807d4 MK |
935 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
936 | if (!PageMovable(page)) { | |
937 | rc = MIGRATEPAGE_SUCCESS; | |
938 | __ClearPageIsolated(page); | |
939 | goto out; | |
940 | } | |
941 | ||
942 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
943 | page, mode); | |
944 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
945 | !PageIsolated(page)); | |
946 | } | |
e24f0b8f | 947 | |
5c3f9a67 HD |
948 | /* |
949 | * When successful, old pagecache page->mapping must be cleared before | |
950 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
951 | */ | |
952 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
953 | if (__PageMovable(page)) { |
954 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
955 | ||
956 | /* | |
957 | * We clear PG_movable under page_lock so any compactor | |
958 | * cannot try to migrate this page. | |
959 | */ | |
960 | __ClearPageIsolated(page); | |
961 | } | |
962 | ||
963 | /* | |
964 | * Anonymous and movable page->mapping will be cleard by | |
965 | * free_pages_prepare so don't reset it here for keeping | |
966 | * the type to work PageAnon, for example. | |
967 | */ | |
968 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 969 | page->mapping = NULL; |
3fe2011f | 970 | } |
bda807d4 | 971 | out: |
e24f0b8f CL |
972 | return rc; |
973 | } | |
974 | ||
0dabec93 | 975 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 976 | int force, enum migrate_mode mode) |
e24f0b8f | 977 | { |
0dabec93 | 978 | int rc = -EAGAIN; |
2ebba6b7 | 979 | int page_was_mapped = 0; |
3f6c8272 | 980 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 981 | bool is_lru = !__PageMovable(page); |
95a402c3 | 982 | |
529ae9aa | 983 | if (!trylock_page(page)) { |
a6bc32b8 | 984 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 985 | goto out; |
3e7d3449 MG |
986 | |
987 | /* | |
988 | * It's not safe for direct compaction to call lock_page. | |
989 | * For example, during page readahead pages are added locked | |
990 | * to the LRU. Later, when the IO completes the pages are | |
991 | * marked uptodate and unlocked. However, the queueing | |
992 | * could be merging multiple pages for one bio (e.g. | |
993 | * mpage_readpages). If an allocation happens for the | |
994 | * second or third page, the process can end up locking | |
995 | * the same page twice and deadlocking. Rather than | |
996 | * trying to be clever about what pages can be locked, | |
997 | * avoid the use of lock_page for direct compaction | |
998 | * altogether. | |
999 | */ | |
1000 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1001 | goto out; |
3e7d3449 | 1002 | |
e24f0b8f CL |
1003 | lock_page(page); |
1004 | } | |
1005 | ||
1006 | if (PageWriteback(page)) { | |
11bc82d6 | 1007 | /* |
fed5b64a | 1008 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1009 | * necessary to wait for PageWriteback. In the async case, |
1010 | * the retry loop is too short and in the sync-light case, | |
1011 | * the overhead of stalling is too much | |
11bc82d6 | 1012 | */ |
2916ecc0 JG |
1013 | switch (mode) { |
1014 | case MIGRATE_SYNC: | |
1015 | case MIGRATE_SYNC_NO_COPY: | |
1016 | break; | |
1017 | default: | |
11bc82d6 | 1018 | rc = -EBUSY; |
0a31bc97 | 1019 | goto out_unlock; |
11bc82d6 AA |
1020 | } |
1021 | if (!force) | |
0a31bc97 | 1022 | goto out_unlock; |
e24f0b8f CL |
1023 | wait_on_page_writeback(page); |
1024 | } | |
03f15c86 | 1025 | |
e24f0b8f | 1026 | /* |
dc386d4d KH |
1027 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1028 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 1029 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1030 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1031 | * File Caches may use write_page() or lock_page() in migration, then, |
1032 | * just care Anon page here. | |
03f15c86 HD |
1033 | * |
1034 | * Only page_get_anon_vma() understands the subtleties of | |
1035 | * getting a hold on an anon_vma from outside one of its mms. | |
1036 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1037 | * because that implies that the anon page is no longer mapped | |
1038 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1039 | */ |
03f15c86 | 1040 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1041 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1042 | |
7db7671f HD |
1043 | /* |
1044 | * Block others from accessing the new page when we get around to | |
1045 | * establishing additional references. We are usually the only one | |
1046 | * holding a reference to newpage at this point. We used to have a BUG | |
1047 | * here if trylock_page(newpage) fails, but would like to allow for | |
1048 | * cases where there might be a race with the previous use of newpage. | |
1049 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1050 | */ | |
1051 | if (unlikely(!trylock_page(newpage))) | |
1052 | goto out_unlock; | |
1053 | ||
bda807d4 MK |
1054 | if (unlikely(!is_lru)) { |
1055 | rc = move_to_new_page(newpage, page, mode); | |
1056 | goto out_unlock_both; | |
1057 | } | |
1058 | ||
dc386d4d | 1059 | /* |
62e1c553 SL |
1060 | * Corner case handling: |
1061 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1062 | * and treated as swapcache but it has no rmap yet. | |
1063 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1064 | * trigger a BUG. So handle it here. | |
1065 | * 2. An orphaned page (see truncate_complete_page) might have | |
1066 | * fs-private metadata. The page can be picked up due to memory | |
1067 | * offlining. Everywhere else except page reclaim, the page is | |
1068 | * invisible to the vm, so the page can not be migrated. So try to | |
1069 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1070 | */ |
62e1c553 | 1071 | if (!page->mapping) { |
309381fe | 1072 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1073 | if (page_has_private(page)) { |
62e1c553 | 1074 | try_to_free_buffers(page); |
7db7671f | 1075 | goto out_unlock_both; |
62e1c553 | 1076 | } |
7db7671f HD |
1077 | } else if (page_mapped(page)) { |
1078 | /* Establish migration ptes */ | |
03f15c86 HD |
1079 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1080 | page); | |
2ebba6b7 | 1081 | try_to_unmap(page, |
da1b13cc | 1082 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
1083 | page_was_mapped = 1; |
1084 | } | |
dc386d4d | 1085 | |
e6a1530d | 1086 | if (!page_mapped(page)) |
5c3f9a67 | 1087 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1088 | |
5c3f9a67 HD |
1089 | if (page_was_mapped) |
1090 | remove_migration_ptes(page, | |
e388466d | 1091 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1092 | |
7db7671f HD |
1093 | out_unlock_both: |
1094 | unlock_page(newpage); | |
1095 | out_unlock: | |
3f6c8272 | 1096 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1097 | if (anon_vma) |
9e60109f | 1098 | put_anon_vma(anon_vma); |
e24f0b8f | 1099 | unlock_page(page); |
0dabec93 | 1100 | out: |
c6c919eb MK |
1101 | /* |
1102 | * If migration is successful, decrease refcount of the newpage | |
1103 | * which will not free the page because new page owner increased | |
1104 | * refcounter. As well, if it is LRU page, add the page to LRU | |
1105 | * list in here. | |
1106 | */ | |
1107 | if (rc == MIGRATEPAGE_SUCCESS) { | |
b1123ea6 | 1108 | if (unlikely(__PageMovable(newpage))) |
c6c919eb MK |
1109 | put_page(newpage); |
1110 | else | |
1111 | putback_lru_page(newpage); | |
1112 | } | |
1113 | ||
0dabec93 MK |
1114 | return rc; |
1115 | } | |
95a402c3 | 1116 | |
ef2a5153 GU |
1117 | /* |
1118 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
1119 | * around it. | |
1120 | */ | |
1121 | #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) | |
1122 | #define ICE_noinline noinline | |
1123 | #else | |
1124 | #define ICE_noinline | |
1125 | #endif | |
1126 | ||
0dabec93 MK |
1127 | /* |
1128 | * Obtain the lock on page, remove all ptes and migrate the page | |
1129 | * to the newly allocated page in newpage. | |
1130 | */ | |
ef2a5153 GU |
1131 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
1132 | free_page_t put_new_page, | |
1133 | unsigned long private, struct page *page, | |
add05cec NH |
1134 | int force, enum migrate_mode mode, |
1135 | enum migrate_reason reason) | |
0dabec93 | 1136 | { |
2def7424 | 1137 | int rc = MIGRATEPAGE_SUCCESS; |
0dabec93 | 1138 | int *result = NULL; |
2def7424 | 1139 | struct page *newpage; |
0dabec93 | 1140 | |
2def7424 | 1141 | newpage = get_new_page(page, private, &result); |
0dabec93 MK |
1142 | if (!newpage) |
1143 | return -ENOMEM; | |
1144 | ||
1145 | if (page_count(page) == 1) { | |
1146 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1147 | ClearPageActive(page); |
1148 | ClearPageUnevictable(page); | |
bda807d4 MK |
1149 | if (unlikely(__PageMovable(page))) { |
1150 | lock_page(page); | |
1151 | if (!PageMovable(page)) | |
1152 | __ClearPageIsolated(page); | |
1153 | unlock_page(page); | |
1154 | } | |
c6c919eb MK |
1155 | if (put_new_page) |
1156 | put_new_page(newpage, private); | |
1157 | else | |
1158 | put_page(newpage); | |
0dabec93 MK |
1159 | goto out; |
1160 | } | |
1161 | ||
616b8371 | 1162 | if (unlikely(PageTransHuge(page) && !PageTransHuge(newpage))) { |
4d2fa965 KS |
1163 | lock_page(page); |
1164 | rc = split_huge_page(page); | |
1165 | unlock_page(page); | |
1166 | if (rc) | |
0dabec93 | 1167 | goto out; |
4d2fa965 | 1168 | } |
0dabec93 | 1169 | |
9c620e2b | 1170 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1171 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1172 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1173 | |
0dabec93 | 1174 | out: |
e24f0b8f | 1175 | if (rc != -EAGAIN) { |
0dabec93 MK |
1176 | /* |
1177 | * A page that has been migrated has all references | |
1178 | * removed and will be freed. A page that has not been | |
1179 | * migrated will have kepts its references and be | |
1180 | * restored. | |
1181 | */ | |
1182 | list_del(&page->lru); | |
6afcf8ef ML |
1183 | |
1184 | /* | |
1185 | * Compaction can migrate also non-LRU pages which are | |
1186 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1187 | * as __PageMovable | |
1188 | */ | |
1189 | if (likely(!__PageMovable(page))) | |
e8db67eb NH |
1190 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
1191 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
c6c919eb MK |
1192 | } |
1193 | ||
1194 | /* | |
1195 | * If migration is successful, releases reference grabbed during | |
1196 | * isolation. Otherwise, restore the page to right list unless | |
1197 | * we want to retry. | |
1198 | */ | |
1199 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1200 | put_page(page); | |
1201 | if (reason == MR_MEMORY_FAILURE) { | |
d7e69488 | 1202 | /* |
c6c919eb MK |
1203 | * Set PG_HWPoison on just freed page |
1204 | * intentionally. Although it's rather weird, | |
1205 | * it's how HWPoison flag works at the moment. | |
d7e69488 | 1206 | */ |
da1b13cc WL |
1207 | if (!test_set_page_hwpoison(page)) |
1208 | num_poisoned_pages_inc(); | |
c6c919eb MK |
1209 | } |
1210 | } else { | |
bda807d4 MK |
1211 | if (rc != -EAGAIN) { |
1212 | if (likely(!__PageMovable(page))) { | |
1213 | putback_lru_page(page); | |
1214 | goto put_new; | |
1215 | } | |
1216 | ||
1217 | lock_page(page); | |
1218 | if (PageMovable(page)) | |
1219 | putback_movable_page(page); | |
1220 | else | |
1221 | __ClearPageIsolated(page); | |
1222 | unlock_page(page); | |
1223 | put_page(page); | |
1224 | } | |
1225 | put_new: | |
c6c919eb MK |
1226 | if (put_new_page) |
1227 | put_new_page(newpage, private); | |
1228 | else | |
1229 | put_page(newpage); | |
e24f0b8f | 1230 | } |
68711a74 | 1231 | |
742755a1 CL |
1232 | if (result) { |
1233 | if (rc) | |
1234 | *result = rc; | |
1235 | else | |
1236 | *result = page_to_nid(newpage); | |
1237 | } | |
e24f0b8f CL |
1238 | return rc; |
1239 | } | |
1240 | ||
290408d4 NH |
1241 | /* |
1242 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1243 | * | |
1244 | * This function doesn't wait the completion of hugepage I/O | |
1245 | * because there is no race between I/O and migration for hugepage. | |
1246 | * Note that currently hugepage I/O occurs only in direct I/O | |
1247 | * where no lock is held and PG_writeback is irrelevant, | |
1248 | * and writeback status of all subpages are counted in the reference | |
1249 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1250 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1251 | * This means that when we try to migrate hugepage whose subpages are | |
1252 | * doing direct I/O, some references remain after try_to_unmap() and | |
1253 | * hugepage migration fails without data corruption. | |
1254 | * | |
1255 | * There is also no race when direct I/O is issued on the page under migration, | |
1256 | * because then pte is replaced with migration swap entry and direct I/O code | |
1257 | * will wait in the page fault for migration to complete. | |
1258 | */ | |
1259 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1260 | free_page_t put_new_page, unsigned long private, |
1261 | struct page *hpage, int force, | |
7cd12b4a | 1262 | enum migrate_mode mode, int reason) |
290408d4 | 1263 | { |
2def7424 | 1264 | int rc = -EAGAIN; |
290408d4 | 1265 | int *result = NULL; |
2ebba6b7 | 1266 | int page_was_mapped = 0; |
32665f2b | 1267 | struct page *new_hpage; |
290408d4 NH |
1268 | struct anon_vma *anon_vma = NULL; |
1269 | ||
83467efb NH |
1270 | /* |
1271 | * Movability of hugepages depends on architectures and hugepage size. | |
1272 | * This check is necessary because some callers of hugepage migration | |
1273 | * like soft offline and memory hotremove don't walk through page | |
1274 | * tables or check whether the hugepage is pmd-based or not before | |
1275 | * kicking migration. | |
1276 | */ | |
100873d7 | 1277 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1278 | putback_active_hugepage(hpage); |
83467efb | 1279 | return -ENOSYS; |
32665f2b | 1280 | } |
83467efb | 1281 | |
32665f2b | 1282 | new_hpage = get_new_page(hpage, private, &result); |
290408d4 NH |
1283 | if (!new_hpage) |
1284 | return -ENOMEM; | |
1285 | ||
290408d4 | 1286 | if (!trylock_page(hpage)) { |
2916ecc0 | 1287 | if (!force) |
290408d4 | 1288 | goto out; |
2916ecc0 JG |
1289 | switch (mode) { |
1290 | case MIGRATE_SYNC: | |
1291 | case MIGRATE_SYNC_NO_COPY: | |
1292 | break; | |
1293 | default: | |
1294 | goto out; | |
1295 | } | |
290408d4 NH |
1296 | lock_page(hpage); |
1297 | } | |
1298 | ||
746b18d4 PZ |
1299 | if (PageAnon(hpage)) |
1300 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1301 | |
7db7671f HD |
1302 | if (unlikely(!trylock_page(new_hpage))) |
1303 | goto put_anon; | |
1304 | ||
2ebba6b7 HD |
1305 | if (page_mapped(hpage)) { |
1306 | try_to_unmap(hpage, | |
1307 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1308 | page_was_mapped = 1; | |
1309 | } | |
290408d4 NH |
1310 | |
1311 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1312 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1313 | |
5c3f9a67 HD |
1314 | if (page_was_mapped) |
1315 | remove_migration_ptes(hpage, | |
e388466d | 1316 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1317 | |
7db7671f HD |
1318 | unlock_page(new_hpage); |
1319 | ||
1320 | put_anon: | |
fd4a4663 | 1321 | if (anon_vma) |
9e60109f | 1322 | put_anon_vma(anon_vma); |
8e6ac7fa | 1323 | |
2def7424 | 1324 | if (rc == MIGRATEPAGE_SUCCESS) { |
8e6ac7fa | 1325 | hugetlb_cgroup_migrate(hpage, new_hpage); |
2def7424 | 1326 | put_new_page = NULL; |
7cd12b4a | 1327 | set_page_owner_migrate_reason(new_hpage, reason); |
2def7424 | 1328 | } |
8e6ac7fa | 1329 | |
290408d4 | 1330 | unlock_page(hpage); |
09761333 | 1331 | out: |
b8ec1cee NH |
1332 | if (rc != -EAGAIN) |
1333 | putback_active_hugepage(hpage); | |
c3114a84 AK |
1334 | if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage)) |
1335 | num_poisoned_pages_inc(); | |
68711a74 DR |
1336 | |
1337 | /* | |
1338 | * If migration was not successful and there's a freeing callback, use | |
1339 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1340 | * isolation. | |
1341 | */ | |
2def7424 | 1342 | if (put_new_page) |
68711a74 DR |
1343 | put_new_page(new_hpage, private); |
1344 | else | |
3aaa76e1 | 1345 | putback_active_hugepage(new_hpage); |
68711a74 | 1346 | |
290408d4 NH |
1347 | if (result) { |
1348 | if (rc) | |
1349 | *result = rc; | |
1350 | else | |
1351 | *result = page_to_nid(new_hpage); | |
1352 | } | |
1353 | return rc; | |
1354 | } | |
1355 | ||
b20a3503 | 1356 | /* |
c73e5c9c SB |
1357 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1358 | * supplied as the target for the page migration | |
b20a3503 | 1359 | * |
c73e5c9c SB |
1360 | * @from: The list of pages to be migrated. |
1361 | * @get_new_page: The function used to allocate free pages to be used | |
1362 | * as the target of the page migration. | |
68711a74 DR |
1363 | * @put_new_page: The function used to free target pages if migration |
1364 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1365 | * @private: Private data to be passed on to get_new_page() |
1366 | * @mode: The migration mode that specifies the constraints for | |
1367 | * page migration, if any. | |
1368 | * @reason: The reason for page migration. | |
b20a3503 | 1369 | * |
c73e5c9c SB |
1370 | * The function returns after 10 attempts or if no pages are movable any more |
1371 | * because the list has become empty or no retryable pages exist any more. | |
14e0f9bc | 1372 | * The caller should call putback_movable_pages() to return pages to the LRU |
28bd6578 | 1373 | * or free list only if ret != 0. |
b20a3503 | 1374 | * |
c73e5c9c | 1375 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1376 | */ |
9c620e2b | 1377 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1378 | free_page_t put_new_page, unsigned long private, |
1379 | enum migrate_mode mode, int reason) | |
b20a3503 | 1380 | { |
e24f0b8f | 1381 | int retry = 1; |
b20a3503 | 1382 | int nr_failed = 0; |
5647bc29 | 1383 | int nr_succeeded = 0; |
b20a3503 CL |
1384 | int pass = 0; |
1385 | struct page *page; | |
1386 | struct page *page2; | |
1387 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1388 | int rc; | |
1389 | ||
1390 | if (!swapwrite) | |
1391 | current->flags |= PF_SWAPWRITE; | |
1392 | ||
e24f0b8f CL |
1393 | for(pass = 0; pass < 10 && retry; pass++) { |
1394 | retry = 0; | |
b20a3503 | 1395 | |
e24f0b8f | 1396 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1397 | cond_resched(); |
2d1db3b1 | 1398 | |
31caf665 NH |
1399 | if (PageHuge(page)) |
1400 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1401 | put_new_page, private, page, |
7cd12b4a | 1402 | pass > 2, mode, reason); |
31caf665 | 1403 | else |
68711a74 | 1404 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1405 | private, page, pass > 2, mode, |
1406 | reason); | |
2d1db3b1 | 1407 | |
e24f0b8f | 1408 | switch(rc) { |
95a402c3 | 1409 | case -ENOMEM: |
dfef2ef4 | 1410 | nr_failed++; |
95a402c3 | 1411 | goto out; |
e24f0b8f | 1412 | case -EAGAIN: |
2d1db3b1 | 1413 | retry++; |
e24f0b8f | 1414 | break; |
78bd5209 | 1415 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1416 | nr_succeeded++; |
e24f0b8f CL |
1417 | break; |
1418 | default: | |
354a3363 NH |
1419 | /* |
1420 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1421 | * unlike -EAGAIN case, the failed page is | |
1422 | * removed from migration page list and not | |
1423 | * retried in the next outer loop. | |
1424 | */ | |
2d1db3b1 | 1425 | nr_failed++; |
e24f0b8f | 1426 | break; |
2d1db3b1 | 1427 | } |
b20a3503 CL |
1428 | } |
1429 | } | |
f2f81fb2 VB |
1430 | nr_failed += retry; |
1431 | rc = nr_failed; | |
95a402c3 | 1432 | out: |
5647bc29 MG |
1433 | if (nr_succeeded) |
1434 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1435 | if (nr_failed) | |
1436 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1437 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1438 | ||
b20a3503 CL |
1439 | if (!swapwrite) |
1440 | current->flags &= ~PF_SWAPWRITE; | |
1441 | ||
78bd5209 | 1442 | return rc; |
b20a3503 | 1443 | } |
95a402c3 | 1444 | |
742755a1 CL |
1445 | #ifdef CONFIG_NUMA |
1446 | /* | |
1447 | * Move a list of individual pages | |
1448 | */ | |
1449 | struct page_to_node { | |
1450 | unsigned long addr; | |
1451 | struct page *page; | |
1452 | int node; | |
1453 | int status; | |
1454 | }; | |
1455 | ||
1456 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1457 | int **result) | |
1458 | { | |
1459 | struct page_to_node *pm = (struct page_to_node *)private; | |
1460 | ||
1461 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1462 | pm++; | |
1463 | ||
1464 | if (pm->node == MAX_NUMNODES) | |
1465 | return NULL; | |
1466 | ||
1467 | *result = &pm->status; | |
1468 | ||
e632a938 NH |
1469 | if (PageHuge(p)) |
1470 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1471 | pm->node); | |
e8db67eb NH |
1472 | else if (thp_migration_supported() && PageTransHuge(p)) { |
1473 | struct page *thp; | |
1474 | ||
1475 | thp = alloc_pages_node(pm->node, | |
1476 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM, | |
1477 | HPAGE_PMD_ORDER); | |
1478 | if (!thp) | |
1479 | return NULL; | |
1480 | prep_transhuge_page(thp); | |
1481 | return thp; | |
1482 | } else | |
96db800f | 1483 | return __alloc_pages_node(pm->node, |
e97ca8e5 | 1484 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); |
742755a1 CL |
1485 | } |
1486 | ||
1487 | /* | |
1488 | * Move a set of pages as indicated in the pm array. The addr | |
1489 | * field must be set to the virtual address of the page to be moved | |
1490 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1491 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1492 | */ |
5e9a0f02 BG |
1493 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1494 | struct page_to_node *pm, | |
1495 | int migrate_all) | |
742755a1 CL |
1496 | { |
1497 | int err; | |
1498 | struct page_to_node *pp; | |
1499 | LIST_HEAD(pagelist); | |
1500 | ||
1501 | down_read(&mm->mmap_sem); | |
1502 | ||
1503 | /* | |
1504 | * Build a list of pages to migrate | |
1505 | */ | |
742755a1 CL |
1506 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1507 | struct vm_area_struct *vma; | |
1508 | struct page *page; | |
e8db67eb NH |
1509 | struct page *head; |
1510 | unsigned int follflags; | |
742755a1 | 1511 | |
742755a1 CL |
1512 | err = -EFAULT; |
1513 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1514 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1515 | goto set_status; |
1516 | ||
d899844e | 1517 | /* FOLL_DUMP to ignore special (like zero) pages */ |
e8db67eb NH |
1518 | follflags = FOLL_GET | FOLL_DUMP; |
1519 | if (!thp_migration_supported()) | |
1520 | follflags |= FOLL_SPLIT; | |
1521 | page = follow_page(vma, pp->addr, follflags); | |
89f5b7da LT |
1522 | |
1523 | err = PTR_ERR(page); | |
1524 | if (IS_ERR(page)) | |
1525 | goto set_status; | |
1526 | ||
742755a1 CL |
1527 | err = -ENOENT; |
1528 | if (!page) | |
1529 | goto set_status; | |
1530 | ||
742755a1 CL |
1531 | err = page_to_nid(page); |
1532 | ||
1533 | if (err == pp->node) | |
1534 | /* | |
1535 | * Node already in the right place | |
1536 | */ | |
1537 | goto put_and_set; | |
1538 | ||
1539 | err = -EACCES; | |
1540 | if (page_mapcount(page) > 1 && | |
1541 | !migrate_all) | |
1542 | goto put_and_set; | |
1543 | ||
e632a938 | 1544 | if (PageHuge(page)) { |
e8db67eb | 1545 | if (PageHead(page)) { |
e66f17ff | 1546 | isolate_huge_page(page, &pagelist); |
e8db67eb NH |
1547 | err = 0; |
1548 | pp->page = page; | |
1549 | } | |
e632a938 NH |
1550 | goto put_and_set; |
1551 | } | |
1552 | ||
e8db67eb NH |
1553 | pp->page = compound_head(page); |
1554 | head = compound_head(page); | |
1555 | err = isolate_lru_page(head); | |
6d9c285a | 1556 | if (!err) { |
e8db67eb NH |
1557 | list_add_tail(&head->lru, &pagelist); |
1558 | mod_node_page_state(page_pgdat(head), | |
1559 | NR_ISOLATED_ANON + page_is_file_cache(head), | |
1560 | hpage_nr_pages(head)); | |
6d9c285a | 1561 | } |
742755a1 CL |
1562 | put_and_set: |
1563 | /* | |
1564 | * Either remove the duplicate refcount from | |
1565 | * isolate_lru_page() or drop the page ref if it was | |
1566 | * not isolated. | |
1567 | */ | |
1568 | put_page(page); | |
1569 | set_status: | |
1570 | pp->status = err; | |
1571 | } | |
1572 | ||
e78bbfa8 | 1573 | err = 0; |
cf608ac1 | 1574 | if (!list_empty(&pagelist)) { |
68711a74 | 1575 | err = migrate_pages(&pagelist, new_page_node, NULL, |
9c620e2b | 1576 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1577 | if (err) |
e632a938 | 1578 | putback_movable_pages(&pagelist); |
cf608ac1 | 1579 | } |
742755a1 CL |
1580 | |
1581 | up_read(&mm->mmap_sem); | |
1582 | return err; | |
1583 | } | |
1584 | ||
5e9a0f02 BG |
1585 | /* |
1586 | * Migrate an array of page address onto an array of nodes and fill | |
1587 | * the corresponding array of status. | |
1588 | */ | |
3268c63e | 1589 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1590 | unsigned long nr_pages, |
1591 | const void __user * __user *pages, | |
1592 | const int __user *nodes, | |
1593 | int __user *status, int flags) | |
1594 | { | |
3140a227 | 1595 | struct page_to_node *pm; |
3140a227 BG |
1596 | unsigned long chunk_nr_pages; |
1597 | unsigned long chunk_start; | |
1598 | int err; | |
5e9a0f02 | 1599 | |
3140a227 BG |
1600 | err = -ENOMEM; |
1601 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1602 | if (!pm) | |
5e9a0f02 | 1603 | goto out; |
35282a2d BG |
1604 | |
1605 | migrate_prep(); | |
1606 | ||
5e9a0f02 | 1607 | /* |
3140a227 BG |
1608 | * Store a chunk of page_to_node array in a page, |
1609 | * but keep the last one as a marker | |
5e9a0f02 | 1610 | */ |
3140a227 | 1611 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1612 | |
3140a227 BG |
1613 | for (chunk_start = 0; |
1614 | chunk_start < nr_pages; | |
1615 | chunk_start += chunk_nr_pages) { | |
1616 | int j; | |
5e9a0f02 | 1617 | |
3140a227 BG |
1618 | if (chunk_start + chunk_nr_pages > nr_pages) |
1619 | chunk_nr_pages = nr_pages - chunk_start; | |
1620 | ||
1621 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1622 | for (j = 0; j < chunk_nr_pages; j++) { | |
1623 | const void __user *p; | |
5e9a0f02 BG |
1624 | int node; |
1625 | ||
3140a227 BG |
1626 | err = -EFAULT; |
1627 | if (get_user(p, pages + j + chunk_start)) | |
1628 | goto out_pm; | |
1629 | pm[j].addr = (unsigned long) p; | |
1630 | ||
1631 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1632 | goto out_pm; |
1633 | ||
1634 | err = -ENODEV; | |
6f5a55f1 LT |
1635 | if (node < 0 || node >= MAX_NUMNODES) |
1636 | goto out_pm; | |
1637 | ||
389162c2 | 1638 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1639 | goto out_pm; |
1640 | ||
1641 | err = -EACCES; | |
1642 | if (!node_isset(node, task_nodes)) | |
1643 | goto out_pm; | |
1644 | ||
3140a227 BG |
1645 | pm[j].node = node; |
1646 | } | |
1647 | ||
1648 | /* End marker for this chunk */ | |
1649 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1650 | ||
1651 | /* Migrate this chunk */ | |
1652 | err = do_move_page_to_node_array(mm, pm, | |
1653 | flags & MPOL_MF_MOVE_ALL); | |
1654 | if (err < 0) | |
1655 | goto out_pm; | |
5e9a0f02 | 1656 | |
5e9a0f02 | 1657 | /* Return status information */ |
3140a227 BG |
1658 | for (j = 0; j < chunk_nr_pages; j++) |
1659 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1660 | err = -EFAULT; |
3140a227 BG |
1661 | goto out_pm; |
1662 | } | |
1663 | } | |
1664 | err = 0; | |
5e9a0f02 BG |
1665 | |
1666 | out_pm: | |
3140a227 | 1667 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1668 | out: |
1669 | return err; | |
1670 | } | |
1671 | ||
742755a1 | 1672 | /* |
2f007e74 | 1673 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1674 | */ |
80bba129 BG |
1675 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1676 | const void __user **pages, int *status) | |
742755a1 | 1677 | { |
2f007e74 | 1678 | unsigned long i; |
2f007e74 | 1679 | |
742755a1 CL |
1680 | down_read(&mm->mmap_sem); |
1681 | ||
2f007e74 | 1682 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1683 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1684 | struct vm_area_struct *vma; |
1685 | struct page *page; | |
c095adbc | 1686 | int err = -EFAULT; |
2f007e74 BG |
1687 | |
1688 | vma = find_vma(mm, addr); | |
70384dc6 | 1689 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1690 | goto set_status; |
1691 | ||
d899844e KS |
1692 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1693 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1694 | |
1695 | err = PTR_ERR(page); | |
1696 | if (IS_ERR(page)) | |
1697 | goto set_status; | |
1698 | ||
d899844e | 1699 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1700 | set_status: |
80bba129 BG |
1701 | *status = err; |
1702 | ||
1703 | pages++; | |
1704 | status++; | |
1705 | } | |
1706 | ||
1707 | up_read(&mm->mmap_sem); | |
1708 | } | |
1709 | ||
1710 | /* | |
1711 | * Determine the nodes of a user array of pages and store it in | |
1712 | * a user array of status. | |
1713 | */ | |
1714 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1715 | const void __user * __user *pages, | |
1716 | int __user *status) | |
1717 | { | |
1718 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1719 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1720 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1721 | |
87b8d1ad PA |
1722 | while (nr_pages) { |
1723 | unsigned long chunk_nr; | |
80bba129 | 1724 | |
87b8d1ad PA |
1725 | chunk_nr = nr_pages; |
1726 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1727 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1728 | ||
1729 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1730 | break; | |
80bba129 BG |
1731 | |
1732 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1733 | ||
87b8d1ad PA |
1734 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1735 | break; | |
742755a1 | 1736 | |
87b8d1ad PA |
1737 | pages += chunk_nr; |
1738 | status += chunk_nr; | |
1739 | nr_pages -= chunk_nr; | |
1740 | } | |
1741 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1742 | } |
1743 | ||
1744 | /* | |
1745 | * Move a list of pages in the address space of the currently executing | |
1746 | * process. | |
1747 | */ | |
938bb9f5 HC |
1748 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1749 | const void __user * __user *, pages, | |
1750 | const int __user *, nodes, | |
1751 | int __user *, status, int, flags) | |
742755a1 | 1752 | { |
742755a1 | 1753 | struct task_struct *task; |
742755a1 | 1754 | struct mm_struct *mm; |
5e9a0f02 | 1755 | int err; |
3268c63e | 1756 | nodemask_t task_nodes; |
742755a1 CL |
1757 | |
1758 | /* Check flags */ | |
1759 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1760 | return -EINVAL; | |
1761 | ||
1762 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1763 | return -EPERM; | |
1764 | ||
1765 | /* Find the mm_struct */ | |
a879bf58 | 1766 | rcu_read_lock(); |
228ebcbe | 1767 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1768 | if (!task) { |
a879bf58 | 1769 | rcu_read_unlock(); |
742755a1 CL |
1770 | return -ESRCH; |
1771 | } | |
3268c63e | 1772 | get_task_struct(task); |
742755a1 CL |
1773 | |
1774 | /* | |
1775 | * Check if this process has the right to modify the specified | |
197e7e52 | 1776 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1777 | */ |
197e7e52 | 1778 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1779 | rcu_read_unlock(); |
742755a1 | 1780 | err = -EPERM; |
5e9a0f02 | 1781 | goto out; |
742755a1 | 1782 | } |
c69e8d9c | 1783 | rcu_read_unlock(); |
742755a1 | 1784 | |
86c3a764 DQ |
1785 | err = security_task_movememory(task); |
1786 | if (err) | |
5e9a0f02 | 1787 | goto out; |
86c3a764 | 1788 | |
3268c63e CL |
1789 | task_nodes = cpuset_mems_allowed(task); |
1790 | mm = get_task_mm(task); | |
1791 | put_task_struct(task); | |
1792 | ||
6e8b09ea SL |
1793 | if (!mm) |
1794 | return -EINVAL; | |
1795 | ||
1796 | if (nodes) | |
1797 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1798 | nodes, status, flags); | |
1799 | else | |
1800 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1801 | |
742755a1 CL |
1802 | mmput(mm); |
1803 | return err; | |
3268c63e CL |
1804 | |
1805 | out: | |
1806 | put_task_struct(task); | |
1807 | return err; | |
742755a1 | 1808 | } |
742755a1 | 1809 | |
7039e1db PZ |
1810 | #ifdef CONFIG_NUMA_BALANCING |
1811 | /* | |
1812 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1813 | * pages. Currently it only checks the watermarks which crude | |
1814 | */ | |
1815 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1816 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1817 | { |
1818 | int z; | |
599d0c95 | 1819 | |
7039e1db PZ |
1820 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1821 | struct zone *zone = pgdat->node_zones + z; | |
1822 | ||
1823 | if (!populated_zone(zone)) | |
1824 | continue; | |
1825 | ||
7039e1db PZ |
1826 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1827 | if (!zone_watermark_ok(zone, 0, | |
1828 | high_wmark_pages(zone) + | |
1829 | nr_migrate_pages, | |
1830 | 0, 0)) | |
1831 | continue; | |
1832 | return true; | |
1833 | } | |
1834 | return false; | |
1835 | } | |
1836 | ||
1837 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1838 | unsigned long data, | |
1839 | int **result) | |
1840 | { | |
1841 | int nid = (int) data; | |
1842 | struct page *newpage; | |
1843 | ||
96db800f | 1844 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1845 | (GFP_HIGHUSER_MOVABLE | |
1846 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1847 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 1848 | ~__GFP_RECLAIM, 0); |
bac0382c | 1849 | |
7039e1db PZ |
1850 | return newpage; |
1851 | } | |
1852 | ||
a8f60772 MG |
1853 | /* |
1854 | * page migration rate limiting control. | |
1855 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1856 | * window of time. Default here says do not migrate more than 1280M per second. | |
1857 | */ | |
1858 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1859 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1860 | ||
b32967ff | 1861 | /* Returns true if the node is migrate rate-limited after the update */ |
1c30e017 MG |
1862 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, |
1863 | unsigned long nr_pages) | |
7039e1db | 1864 | { |
a8f60772 MG |
1865 | /* |
1866 | * Rate-limit the amount of data that is being migrated to a node. | |
1867 | * Optimal placement is no good if the memory bus is saturated and | |
1868 | * all the time is being spent migrating! | |
1869 | */ | |
a8f60772 | 1870 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { |
1c5e9c27 | 1871 | spin_lock(&pgdat->numabalancing_migrate_lock); |
a8f60772 MG |
1872 | pgdat->numabalancing_migrate_nr_pages = 0; |
1873 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1874 | msecs_to_jiffies(migrate_interval_millisecs); | |
1c5e9c27 | 1875 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
a8f60772 | 1876 | } |
af1839d7 MG |
1877 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { |
1878 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1879 | nr_pages); | |
1c5e9c27 | 1880 | return true; |
af1839d7 | 1881 | } |
1c5e9c27 MG |
1882 | |
1883 | /* | |
1884 | * This is an unlocked non-atomic update so errors are possible. | |
1885 | * The consequences are failing to migrate when we potentiall should | |
1886 | * have which is not severe enough to warrant locking. If it is ever | |
1887 | * a problem, it can be converted to a per-cpu counter. | |
1888 | */ | |
1889 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1890 | return false; | |
b32967ff MG |
1891 | } |
1892 | ||
1c30e017 | 1893 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1894 | { |
340ef390 | 1895 | int page_lru; |
a8f60772 | 1896 | |
309381fe | 1897 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1898 | |
7039e1db | 1899 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1900 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1901 | return 0; | |
7039e1db | 1902 | |
340ef390 HD |
1903 | if (isolate_lru_page(page)) |
1904 | return 0; | |
7039e1db | 1905 | |
340ef390 HD |
1906 | /* |
1907 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1908 | * check on page_count(), so we must do it here, now that the page | |
1909 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1910 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1911 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1912 | */ | |
1913 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1914 | putback_lru_page(page); | |
1915 | return 0; | |
7039e1db PZ |
1916 | } |
1917 | ||
340ef390 | 1918 | page_lru = page_is_file_cache(page); |
599d0c95 | 1919 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
340ef390 HD |
1920 | hpage_nr_pages(page)); |
1921 | ||
149c33e1 | 1922 | /* |
340ef390 HD |
1923 | * Isolating the page has taken another reference, so the |
1924 | * caller's reference can be safely dropped without the page | |
1925 | * disappearing underneath us during migration. | |
149c33e1 MG |
1926 | */ |
1927 | put_page(page); | |
340ef390 | 1928 | return 1; |
b32967ff MG |
1929 | } |
1930 | ||
de466bd6 MG |
1931 | bool pmd_trans_migrating(pmd_t pmd) |
1932 | { | |
1933 | struct page *page = pmd_page(pmd); | |
1934 | return PageLocked(page); | |
1935 | } | |
1936 | ||
b32967ff MG |
1937 | /* |
1938 | * Attempt to migrate a misplaced page to the specified destination | |
1939 | * node. Caller is expected to have an elevated reference count on | |
1940 | * the page that will be dropped by this function before returning. | |
1941 | */ | |
1bc115d8 MG |
1942 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1943 | int node) | |
b32967ff MG |
1944 | { |
1945 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1946 | int isolated; |
b32967ff MG |
1947 | int nr_remaining; |
1948 | LIST_HEAD(migratepages); | |
1949 | ||
1950 | /* | |
1bc115d8 MG |
1951 | * Don't migrate file pages that are mapped in multiple processes |
1952 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1953 | */ |
1bc115d8 MG |
1954 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1955 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1956 | goto out; |
b32967ff MG |
1957 | |
1958 | /* | |
1959 | * Rate-limit the amount of data that is being migrated to a node. | |
1960 | * Optimal placement is no good if the memory bus is saturated and | |
1961 | * all the time is being spent migrating! | |
1962 | */ | |
340ef390 | 1963 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1964 | goto out; |
b32967ff MG |
1965 | |
1966 | isolated = numamigrate_isolate_page(pgdat, page); | |
1967 | if (!isolated) | |
1968 | goto out; | |
1969 | ||
1970 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1971 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1972 | NULL, node, MIGRATE_ASYNC, |
1973 | MR_NUMA_MISPLACED); | |
b32967ff | 1974 | if (nr_remaining) { |
59c82b70 JK |
1975 | if (!list_empty(&migratepages)) { |
1976 | list_del(&page->lru); | |
599d0c95 | 1977 | dec_node_page_state(page, NR_ISOLATED_ANON + |
59c82b70 JK |
1978 | page_is_file_cache(page)); |
1979 | putback_lru_page(page); | |
1980 | } | |
b32967ff MG |
1981 | isolated = 0; |
1982 | } else | |
1983 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1984 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1985 | return isolated; |
340ef390 HD |
1986 | |
1987 | out: | |
1988 | put_page(page); | |
1989 | return 0; | |
7039e1db | 1990 | } |
220018d3 | 1991 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1992 | |
220018d3 | 1993 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1994 | /* |
1995 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1996 | * before returning. | |
1997 | */ | |
b32967ff MG |
1998 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1999 | struct vm_area_struct *vma, | |
2000 | pmd_t *pmd, pmd_t entry, | |
2001 | unsigned long address, | |
2002 | struct page *page, int node) | |
2003 | { | |
c4088ebd | 2004 | spinlock_t *ptl; |
b32967ff MG |
2005 | pg_data_t *pgdat = NODE_DATA(node); |
2006 | int isolated = 0; | |
2007 | struct page *new_page = NULL; | |
b32967ff | 2008 | int page_lru = page_is_file_cache(page); |
f714f4f2 MG |
2009 | unsigned long mmun_start = address & HPAGE_PMD_MASK; |
2010 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
b32967ff | 2011 | |
b32967ff MG |
2012 | /* |
2013 | * Rate-limit the amount of data that is being migrated to a node. | |
2014 | * Optimal placement is no good if the memory bus is saturated and | |
2015 | * all the time is being spent migrating! | |
2016 | */ | |
d28d4335 | 2017 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
2018 | goto out_dropref; |
2019 | ||
2020 | new_page = alloc_pages_node(node, | |
25160354 | 2021 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
e97ca8e5 | 2022 | HPAGE_PMD_ORDER); |
340ef390 HD |
2023 | if (!new_page) |
2024 | goto out_fail; | |
9a982250 | 2025 | prep_transhuge_page(new_page); |
340ef390 | 2026 | |
b32967ff | 2027 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 2028 | if (!isolated) { |
b32967ff | 2029 | put_page(new_page); |
340ef390 | 2030 | goto out_fail; |
b32967ff | 2031 | } |
b0943d61 | 2032 | |
b32967ff | 2033 | /* Prepare a page as a migration target */ |
48c935ad | 2034 | __SetPageLocked(new_page); |
d44d363f SL |
2035 | if (PageSwapBacked(page)) |
2036 | __SetPageSwapBacked(new_page); | |
b32967ff MG |
2037 | |
2038 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2039 | new_page->mapping = page->mapping; | |
2040 | new_page->index = page->index; | |
2041 | migrate_page_copy(new_page, page); | |
2042 | WARN_ON(PageLRU(new_page)); | |
2043 | ||
2044 | /* Recheck the target PMD */ | |
f714f4f2 | 2045 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 2046 | ptl = pmd_lock(mm, pmd); |
f4e177d1 | 2047 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
c4088ebd | 2048 | spin_unlock(ptl); |
f714f4f2 | 2049 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff MG |
2050 | |
2051 | /* Reverse changes made by migrate_page_copy() */ | |
2052 | if (TestClearPageActive(new_page)) | |
2053 | SetPageActive(page); | |
2054 | if (TestClearPageUnevictable(new_page)) | |
2055 | SetPageUnevictable(page); | |
b32967ff MG |
2056 | |
2057 | unlock_page(new_page); | |
2058 | put_page(new_page); /* Free it */ | |
2059 | ||
a54a407f MG |
2060 | /* Retake the callers reference and putback on LRU */ |
2061 | get_page(page); | |
b32967ff | 2062 | putback_lru_page(page); |
599d0c95 | 2063 | mod_node_page_state(page_pgdat(page), |
a54a407f | 2064 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
eb4489f6 MG |
2065 | |
2066 | goto out_unlock; | |
b32967ff MG |
2067 | } |
2068 | ||
10102459 | 2069 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
2b4847e7 | 2070 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 2071 | |
2b4847e7 MG |
2072 | /* |
2073 | * Clear the old entry under pagetable lock and establish the new PTE. | |
2074 | * Any parallel GUP will either observe the old page blocking on the | |
2075 | * page lock, block on the page table lock or observe the new page. | |
2076 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
2077 | * guarantee the copy is visible before the pagetable update. | |
2078 | */ | |
f714f4f2 | 2079 | flush_cache_range(vma, mmun_start, mmun_end); |
d281ee61 | 2080 | page_add_anon_rmap(new_page, vma, mmun_start, true); |
8809aa2d | 2081 | pmdp_huge_clear_flush_notify(vma, mmun_start, pmd); |
f714f4f2 | 2082 | set_pmd_at(mm, mmun_start, pmd, entry); |
ce4a9cc5 | 2083 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 | 2084 | |
f4e177d1 | 2085 | page_ref_unfreeze(page, 2); |
51afb12b | 2086 | mlock_migrate_page(new_page, page); |
d281ee61 | 2087 | page_remove_rmap(page, true); |
7cd12b4a | 2088 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2b4847e7 | 2089 | |
c4088ebd | 2090 | spin_unlock(ptl); |
f714f4f2 | 2091 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b32967ff | 2092 | |
11de9927 MG |
2093 | /* Take an "isolate" reference and put new page on the LRU. */ |
2094 | get_page(new_page); | |
2095 | putback_lru_page(new_page); | |
2096 | ||
b32967ff MG |
2097 | unlock_page(new_page); |
2098 | unlock_page(page); | |
2099 | put_page(page); /* Drop the rmap reference */ | |
2100 | put_page(page); /* Drop the LRU isolation reference */ | |
2101 | ||
2102 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2103 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2104 | ||
599d0c95 | 2105 | mod_node_page_state(page_pgdat(page), |
b32967ff MG |
2106 | NR_ISOLATED_ANON + page_lru, |
2107 | -HPAGE_PMD_NR); | |
2108 | return isolated; | |
2109 | ||
340ef390 HD |
2110 | out_fail: |
2111 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 2112 | out_dropref: |
2b4847e7 MG |
2113 | ptl = pmd_lock(mm, pmd); |
2114 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 2115 | entry = pmd_modify(entry, vma->vm_page_prot); |
f714f4f2 | 2116 | set_pmd_at(mm, mmun_start, pmd, entry); |
2b4847e7 MG |
2117 | update_mmu_cache_pmd(vma, address, &entry); |
2118 | } | |
2119 | spin_unlock(ptl); | |
a54a407f | 2120 | |
eb4489f6 | 2121 | out_unlock: |
340ef390 | 2122 | unlock_page(page); |
b32967ff | 2123 | put_page(page); |
b32967ff MG |
2124 | return 0; |
2125 | } | |
7039e1db PZ |
2126 | #endif /* CONFIG_NUMA_BALANCING */ |
2127 | ||
2128 | #endif /* CONFIG_NUMA */ | |
8763cb45 | 2129 | |
6b368cd4 | 2130 | #if defined(CONFIG_MIGRATE_VMA_HELPER) |
8763cb45 JG |
2131 | struct migrate_vma { |
2132 | struct vm_area_struct *vma; | |
2133 | unsigned long *dst; | |
2134 | unsigned long *src; | |
2135 | unsigned long cpages; | |
2136 | unsigned long npages; | |
2137 | unsigned long start; | |
2138 | unsigned long end; | |
2139 | }; | |
2140 | ||
2141 | static int migrate_vma_collect_hole(unsigned long start, | |
2142 | unsigned long end, | |
2143 | struct mm_walk *walk) | |
2144 | { | |
2145 | struct migrate_vma *migrate = walk->private; | |
2146 | unsigned long addr; | |
2147 | ||
8315ada7 | 2148 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
e20d103b | 2149 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8315ada7 | 2150 | migrate->dst[migrate->npages] = 0; |
e20d103b | 2151 | migrate->npages++; |
8315ada7 JG |
2152 | migrate->cpages++; |
2153 | } | |
2154 | ||
2155 | return 0; | |
2156 | } | |
2157 | ||
2158 | static int migrate_vma_collect_skip(unsigned long start, | |
2159 | unsigned long end, | |
2160 | struct mm_walk *walk) | |
2161 | { | |
2162 | struct migrate_vma *migrate = walk->private; | |
2163 | unsigned long addr; | |
2164 | ||
8763cb45 JG |
2165 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
2166 | migrate->dst[migrate->npages] = 0; | |
2167 | migrate->src[migrate->npages++] = 0; | |
2168 | } | |
2169 | ||
2170 | return 0; | |
2171 | } | |
2172 | ||
2173 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2174 | unsigned long start, | |
2175 | unsigned long end, | |
2176 | struct mm_walk *walk) | |
2177 | { | |
2178 | struct migrate_vma *migrate = walk->private; | |
2179 | struct vm_area_struct *vma = walk->vma; | |
2180 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2181 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2182 | spinlock_t *ptl; |
2183 | pte_t *ptep; | |
2184 | ||
2185 | again: | |
2186 | if (pmd_none(*pmdp)) | |
2187 | return migrate_vma_collect_hole(start, end, walk); | |
2188 | ||
2189 | if (pmd_trans_huge(*pmdp)) { | |
2190 | struct page *page; | |
2191 | ||
2192 | ptl = pmd_lock(mm, pmdp); | |
2193 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2194 | spin_unlock(ptl); | |
2195 | goto again; | |
2196 | } | |
2197 | ||
2198 | page = pmd_page(*pmdp); | |
2199 | if (is_huge_zero_page(page)) { | |
2200 | spin_unlock(ptl); | |
2201 | split_huge_pmd(vma, pmdp, addr); | |
2202 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2203 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2204 | walk); |
2205 | } else { | |
2206 | int ret; | |
2207 | ||
2208 | get_page(page); | |
2209 | spin_unlock(ptl); | |
2210 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2211 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2212 | walk); |
2213 | ret = split_huge_page(page); | |
2214 | unlock_page(page); | |
2215 | put_page(page); | |
8315ada7 JG |
2216 | if (ret) |
2217 | return migrate_vma_collect_skip(start, end, | |
2218 | walk); | |
2219 | if (pmd_none(*pmdp)) | |
8763cb45 JG |
2220 | return migrate_vma_collect_hole(start, end, |
2221 | walk); | |
2222 | } | |
2223 | } | |
2224 | ||
2225 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2226 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2227 | |
2228 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2229 | arch_enter_lazy_mmu_mode(); |
2230 | ||
8763cb45 JG |
2231 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
2232 | unsigned long mpfn, pfn; | |
2233 | struct page *page; | |
8c3328f1 | 2234 | swp_entry_t entry; |
8763cb45 JG |
2235 | pte_t pte; |
2236 | ||
2237 | pte = *ptep; | |
2238 | pfn = pte_pfn(pte); | |
2239 | ||
a5430dda | 2240 | if (pte_none(pte)) { |
8315ada7 JG |
2241 | mpfn = MIGRATE_PFN_MIGRATE; |
2242 | migrate->cpages++; | |
2243 | pfn = 0; | |
8763cb45 JG |
2244 | goto next; |
2245 | } | |
2246 | ||
a5430dda JG |
2247 | if (!pte_present(pte)) { |
2248 | mpfn = pfn = 0; | |
2249 | ||
2250 | /* | |
2251 | * Only care about unaddressable device page special | |
2252 | * page table entry. Other special swap entries are not | |
2253 | * migratable, and we ignore regular swapped page. | |
2254 | */ | |
2255 | entry = pte_to_swp_entry(pte); | |
2256 | if (!is_device_private_entry(entry)) | |
2257 | goto next; | |
2258 | ||
2259 | page = device_private_entry_to_page(entry); | |
2260 | mpfn = migrate_pfn(page_to_pfn(page))| | |
2261 | MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE; | |
2262 | if (is_write_device_private_entry(entry)) | |
2263 | mpfn |= MIGRATE_PFN_WRITE; | |
2264 | } else { | |
8315ada7 JG |
2265 | if (is_zero_pfn(pfn)) { |
2266 | mpfn = MIGRATE_PFN_MIGRATE; | |
2267 | migrate->cpages++; | |
2268 | pfn = 0; | |
2269 | goto next; | |
2270 | } | |
df6ad698 | 2271 | page = _vm_normal_page(migrate->vma, addr, pte, true); |
a5430dda JG |
2272 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2273 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2274 | } | |
2275 | ||
8763cb45 | 2276 | /* FIXME support THP */ |
8763cb45 JG |
2277 | if (!page || !page->mapping || PageTransCompound(page)) { |
2278 | mpfn = pfn = 0; | |
2279 | goto next; | |
2280 | } | |
a5430dda | 2281 | pfn = page_to_pfn(page); |
8763cb45 JG |
2282 | |
2283 | /* | |
2284 | * By getting a reference on the page we pin it and that blocks | |
2285 | * any kind of migration. Side effect is that it "freezes" the | |
2286 | * pte. | |
2287 | * | |
2288 | * We drop this reference after isolating the page from the lru | |
2289 | * for non device page (device page are not on the lru and thus | |
2290 | * can't be dropped from it). | |
2291 | */ | |
2292 | get_page(page); | |
2293 | migrate->cpages++; | |
8763cb45 | 2294 | |
8c3328f1 JG |
2295 | /* |
2296 | * Optimize for the common case where page is only mapped once | |
2297 | * in one process. If we can lock the page, then we can safely | |
2298 | * set up a special migration page table entry now. | |
2299 | */ | |
2300 | if (trylock_page(page)) { | |
2301 | pte_t swp_pte; | |
2302 | ||
2303 | mpfn |= MIGRATE_PFN_LOCKED; | |
2304 | ptep_get_and_clear(mm, addr, ptep); | |
2305 | ||
2306 | /* Setup special migration page table entry */ | |
2307 | entry = make_migration_entry(page, pte_write(pte)); | |
2308 | swp_pte = swp_entry_to_pte(entry); | |
2309 | if (pte_soft_dirty(pte)) | |
2310 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2311 | set_pte_at(mm, addr, ptep, swp_pte); | |
2312 | ||
2313 | /* | |
2314 | * This is like regular unmap: we remove the rmap and | |
2315 | * drop page refcount. Page won't be freed, as we took | |
2316 | * a reference just above. | |
2317 | */ | |
2318 | page_remove_rmap(page, false); | |
2319 | put_page(page); | |
a5430dda JG |
2320 | |
2321 | if (pte_present(pte)) | |
2322 | unmapped++; | |
8c3328f1 JG |
2323 | } |
2324 | ||
8763cb45 | 2325 | next: |
a5430dda | 2326 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2327 | migrate->src[migrate->npages++] = mpfn; |
2328 | } | |
8c3328f1 | 2329 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2330 | pte_unmap_unlock(ptep - 1, ptl); |
2331 | ||
8c3328f1 JG |
2332 | /* Only flush the TLB if we actually modified any entries */ |
2333 | if (unmapped) | |
2334 | flush_tlb_range(walk->vma, start, end); | |
2335 | ||
8763cb45 JG |
2336 | return 0; |
2337 | } | |
2338 | ||
2339 | /* | |
2340 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2341 | * @migrate: migrate struct containing all migration information | |
2342 | * | |
2343 | * This will walk the CPU page table. For each virtual address backed by a | |
2344 | * valid page, it updates the src array and takes a reference on the page, in | |
2345 | * order to pin the page until we lock it and unmap it. | |
2346 | */ | |
2347 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2348 | { | |
2349 | struct mm_walk mm_walk; | |
2350 | ||
2351 | mm_walk.pmd_entry = migrate_vma_collect_pmd; | |
2352 | mm_walk.pte_entry = NULL; | |
2353 | mm_walk.pte_hole = migrate_vma_collect_hole; | |
2354 | mm_walk.hugetlb_entry = NULL; | |
2355 | mm_walk.test_walk = NULL; | |
2356 | mm_walk.vma = migrate->vma; | |
2357 | mm_walk.mm = migrate->vma->vm_mm; | |
2358 | mm_walk.private = migrate; | |
2359 | ||
8c3328f1 JG |
2360 | mmu_notifier_invalidate_range_start(mm_walk.mm, |
2361 | migrate->start, | |
2362 | migrate->end); | |
8763cb45 | 2363 | walk_page_range(migrate->start, migrate->end, &mm_walk); |
8c3328f1 JG |
2364 | mmu_notifier_invalidate_range_end(mm_walk.mm, |
2365 | migrate->start, | |
2366 | migrate->end); | |
8763cb45 JG |
2367 | |
2368 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
2369 | } | |
2370 | ||
2371 | /* | |
2372 | * migrate_vma_check_page() - check if page is pinned or not | |
2373 | * @page: struct page to check | |
2374 | * | |
2375 | * Pinned pages cannot be migrated. This is the same test as in | |
2376 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2377 | * ZONE_DEVICE page. | |
2378 | */ | |
2379 | static bool migrate_vma_check_page(struct page *page) | |
2380 | { | |
2381 | /* | |
2382 | * One extra ref because caller holds an extra reference, either from | |
2383 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2384 | * a device page. | |
2385 | */ | |
2386 | int extra = 1; | |
2387 | ||
2388 | /* | |
2389 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2390 | * check them than regular pages, because they can be mapped with a pmd | |
2391 | * or with a pte (split pte mapping). | |
2392 | */ | |
2393 | if (PageCompound(page)) | |
2394 | return false; | |
2395 | ||
a5430dda JG |
2396 | /* Page from ZONE_DEVICE have one extra reference */ |
2397 | if (is_zone_device_page(page)) { | |
2398 | /* | |
2399 | * Private page can never be pin as they have no valid pte and | |
2400 | * GUP will fail for those. Yet if there is a pending migration | |
2401 | * a thread might try to wait on the pte migration entry and | |
2402 | * will bump the page reference count. Sadly there is no way to | |
2403 | * differentiate a regular pin from migration wait. Hence to | |
2404 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
2405 | * infinite loop (one stoping migration because the other is | |
2406 | * waiting on pte migration entry). We always return true here. | |
2407 | * | |
2408 | * FIXME proper solution is to rework migration_entry_wait() so | |
2409 | * it does not need to take a reference on page. | |
2410 | */ | |
2411 | if (is_device_private_page(page)) | |
2412 | return true; | |
2413 | ||
df6ad698 JG |
2414 | /* |
2415 | * Only allow device public page to be migrated and account for | |
2416 | * the extra reference count imply by ZONE_DEVICE pages. | |
2417 | */ | |
2418 | if (!is_device_public_page(page)) | |
2419 | return false; | |
2420 | extra++; | |
a5430dda JG |
2421 | } |
2422 | ||
df6ad698 JG |
2423 | /* For file back page */ |
2424 | if (page_mapping(page)) | |
2425 | extra += 1 + page_has_private(page); | |
2426 | ||
8763cb45 JG |
2427 | if ((page_count(page) - extra) > page_mapcount(page)) |
2428 | return false; | |
2429 | ||
2430 | return true; | |
2431 | } | |
2432 | ||
2433 | /* | |
2434 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2435 | * @migrate: migrate struct containing all migration information | |
2436 | * | |
2437 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2438 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2439 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2440 | * migrated by concurrent kernel threads. | |
2441 | */ | |
2442 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2443 | { | |
2444 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2445 | const unsigned long start = migrate->start; |
2446 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2447 | bool allow_drain = true; |
8763cb45 JG |
2448 | |
2449 | lru_add_drain(); | |
2450 | ||
2451 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2452 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2453 | bool remap = true; |
8763cb45 JG |
2454 | |
2455 | if (!page) | |
2456 | continue; | |
2457 | ||
8c3328f1 JG |
2458 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2459 | /* | |
2460 | * Because we are migrating several pages there can be | |
2461 | * a deadlock between 2 concurrent migration where each | |
2462 | * are waiting on each other page lock. | |
2463 | * | |
2464 | * Make migrate_vma() a best effort thing and backoff | |
2465 | * for any page we can not lock right away. | |
2466 | */ | |
2467 | if (!trylock_page(page)) { | |
2468 | migrate->src[i] = 0; | |
2469 | migrate->cpages--; | |
2470 | put_page(page); | |
2471 | continue; | |
2472 | } | |
2473 | remap = false; | |
2474 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2475 | } |
8763cb45 | 2476 | |
a5430dda JG |
2477 | /* ZONE_DEVICE pages are not on LRU */ |
2478 | if (!is_zone_device_page(page)) { | |
2479 | if (!PageLRU(page) && allow_drain) { | |
2480 | /* Drain CPU's pagevec */ | |
2481 | lru_add_drain_all(); | |
2482 | allow_drain = false; | |
2483 | } | |
8763cb45 | 2484 | |
a5430dda JG |
2485 | if (isolate_lru_page(page)) { |
2486 | if (remap) { | |
2487 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2488 | migrate->cpages--; | |
2489 | restore++; | |
2490 | } else { | |
2491 | migrate->src[i] = 0; | |
2492 | unlock_page(page); | |
2493 | migrate->cpages--; | |
2494 | put_page(page); | |
2495 | } | |
2496 | continue; | |
8c3328f1 | 2497 | } |
a5430dda JG |
2498 | |
2499 | /* Drop the reference we took in collect */ | |
2500 | put_page(page); | |
8763cb45 JG |
2501 | } |
2502 | ||
2503 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2504 | if (remap) { |
2505 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2506 | migrate->cpages--; | |
2507 | restore++; | |
8763cb45 | 2508 | |
a5430dda JG |
2509 | if (!is_zone_device_page(page)) { |
2510 | get_page(page); | |
2511 | putback_lru_page(page); | |
2512 | } | |
8c3328f1 JG |
2513 | } else { |
2514 | migrate->src[i] = 0; | |
2515 | unlock_page(page); | |
2516 | migrate->cpages--; | |
2517 | ||
a5430dda JG |
2518 | if (!is_zone_device_page(page)) |
2519 | putback_lru_page(page); | |
2520 | else | |
2521 | put_page(page); | |
8c3328f1 | 2522 | } |
8763cb45 JG |
2523 | } |
2524 | } | |
8c3328f1 JG |
2525 | |
2526 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2527 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2528 | ||
2529 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2530 | continue; | |
2531 | ||
2532 | remove_migration_pte(page, migrate->vma, addr, page); | |
2533 | ||
2534 | migrate->src[i] = 0; | |
2535 | unlock_page(page); | |
2536 | put_page(page); | |
2537 | restore--; | |
2538 | } | |
8763cb45 JG |
2539 | } |
2540 | ||
2541 | /* | |
2542 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2543 | * @migrate: migrate struct containing all migration information | |
2544 | * | |
2545 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2546 | * and check again if it has been pinned. Pinned pages are restored because we | |
2547 | * cannot migrate them. | |
2548 | * | |
2549 | * This is the last step before we call the device driver callback to allocate | |
2550 | * destination memory and copy contents of original page over to new page. | |
2551 | */ | |
2552 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2553 | { | |
2554 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
2555 | const unsigned long npages = migrate->npages; | |
2556 | const unsigned long start = migrate->start; | |
2557 | unsigned long addr, i, restore = 0; | |
2558 | ||
2559 | for (i = 0; i < npages; i++) { | |
2560 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2561 | ||
2562 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2563 | continue; | |
2564 | ||
8c3328f1 JG |
2565 | if (page_mapped(page)) { |
2566 | try_to_unmap(page, flags); | |
2567 | if (page_mapped(page)) | |
2568 | goto restore; | |
8763cb45 | 2569 | } |
8c3328f1 JG |
2570 | |
2571 | if (migrate_vma_check_page(page)) | |
2572 | continue; | |
2573 | ||
2574 | restore: | |
2575 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2576 | migrate->cpages--; | |
2577 | restore++; | |
8763cb45 JG |
2578 | } |
2579 | ||
2580 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2581 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2582 | ||
2583 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2584 | continue; | |
2585 | ||
2586 | remove_migration_ptes(page, page, false); | |
2587 | ||
2588 | migrate->src[i] = 0; | |
2589 | unlock_page(page); | |
2590 | restore--; | |
2591 | ||
a5430dda JG |
2592 | if (is_zone_device_page(page)) |
2593 | put_page(page); | |
2594 | else | |
2595 | putback_lru_page(page); | |
8763cb45 JG |
2596 | } |
2597 | } | |
2598 | ||
8315ada7 JG |
2599 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2600 | unsigned long addr, | |
2601 | struct page *page, | |
2602 | unsigned long *src, | |
2603 | unsigned long *dst) | |
2604 | { | |
2605 | struct vm_area_struct *vma = migrate->vma; | |
2606 | struct mm_struct *mm = vma->vm_mm; | |
2607 | struct mem_cgroup *memcg; | |
2608 | bool flush = false; | |
2609 | spinlock_t *ptl; | |
2610 | pte_t entry; | |
2611 | pgd_t *pgdp; | |
2612 | p4d_t *p4dp; | |
2613 | pud_t *pudp; | |
2614 | pmd_t *pmdp; | |
2615 | pte_t *ptep; | |
2616 | ||
2617 | /* Only allow populating anonymous memory */ | |
2618 | if (!vma_is_anonymous(vma)) | |
2619 | goto abort; | |
2620 | ||
2621 | pgdp = pgd_offset(mm, addr); | |
2622 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2623 | if (!p4dp) | |
2624 | goto abort; | |
2625 | pudp = pud_alloc(mm, p4dp, addr); | |
2626 | if (!pudp) | |
2627 | goto abort; | |
2628 | pmdp = pmd_alloc(mm, pudp, addr); | |
2629 | if (!pmdp) | |
2630 | goto abort; | |
2631 | ||
2632 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2633 | goto abort; | |
2634 | ||
2635 | /* | |
2636 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2637 | * pte_offset_map() on pmds where a huge pmd might be created | |
2638 | * from a different thread. | |
2639 | * | |
2640 | * pte_alloc_map() is safe to use under down_write(mmap_sem) or when | |
2641 | * parallel threads are excluded by other means. | |
2642 | * | |
2643 | * Here we only have down_read(mmap_sem). | |
2644 | */ | |
2645 | if (pte_alloc(mm, pmdp, addr)) | |
2646 | goto abort; | |
2647 | ||
2648 | /* See the comment in pte_alloc_one_map() */ | |
2649 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2650 | goto abort; | |
2651 | ||
2652 | if (unlikely(anon_vma_prepare(vma))) | |
2653 | goto abort; | |
2654 | if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false)) | |
2655 | goto abort; | |
2656 | ||
2657 | /* | |
2658 | * The memory barrier inside __SetPageUptodate makes sure that | |
2659 | * preceding stores to the page contents become visible before | |
2660 | * the set_pte_at() write. | |
2661 | */ | |
2662 | __SetPageUptodate(page); | |
2663 | ||
df6ad698 JG |
2664 | if (is_zone_device_page(page)) { |
2665 | if (is_device_private_page(page)) { | |
2666 | swp_entry_t swp_entry; | |
2667 | ||
2668 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2669 | entry = swp_entry_to_pte(swp_entry); | |
2670 | } else if (is_device_public_page(page)) { | |
2671 | entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); | |
2672 | if (vma->vm_flags & VM_WRITE) | |
2673 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2674 | entry = pte_mkdevmap(entry); | |
2675 | } | |
8315ada7 JG |
2676 | } else { |
2677 | entry = mk_pte(page, vma->vm_page_prot); | |
2678 | if (vma->vm_flags & VM_WRITE) | |
2679 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2680 | } | |
2681 | ||
2682 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2683 | ||
2684 | if (pte_present(*ptep)) { | |
2685 | unsigned long pfn = pte_pfn(*ptep); | |
2686 | ||
2687 | if (!is_zero_pfn(pfn)) { | |
2688 | pte_unmap_unlock(ptep, ptl); | |
2689 | mem_cgroup_cancel_charge(page, memcg, false); | |
2690 | goto abort; | |
2691 | } | |
2692 | flush = true; | |
2693 | } else if (!pte_none(*ptep)) { | |
2694 | pte_unmap_unlock(ptep, ptl); | |
2695 | mem_cgroup_cancel_charge(page, memcg, false); | |
2696 | goto abort; | |
2697 | } | |
2698 | ||
2699 | /* | |
2700 | * Check for usefaultfd but do not deliver the fault. Instead, | |
2701 | * just back off. | |
2702 | */ | |
2703 | if (userfaultfd_missing(vma)) { | |
2704 | pte_unmap_unlock(ptep, ptl); | |
2705 | mem_cgroup_cancel_charge(page, memcg, false); | |
2706 | goto abort; | |
2707 | } | |
2708 | ||
2709 | inc_mm_counter(mm, MM_ANONPAGES); | |
2710 | page_add_new_anon_rmap(page, vma, addr, false); | |
2711 | mem_cgroup_commit_charge(page, memcg, false, false); | |
2712 | if (!is_zone_device_page(page)) | |
2713 | lru_cache_add_active_or_unevictable(page, vma); | |
2714 | get_page(page); | |
2715 | ||
2716 | if (flush) { | |
2717 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2718 | ptep_clear_flush_notify(vma, addr, ptep); | |
2719 | set_pte_at_notify(mm, addr, ptep, entry); | |
2720 | update_mmu_cache(vma, addr, ptep); | |
2721 | } else { | |
2722 | /* No need to invalidate - it was non-present before */ | |
2723 | set_pte_at(mm, addr, ptep, entry); | |
2724 | update_mmu_cache(vma, addr, ptep); | |
2725 | } | |
2726 | ||
2727 | pte_unmap_unlock(ptep, ptl); | |
2728 | *src = MIGRATE_PFN_MIGRATE; | |
2729 | return; | |
2730 | ||
2731 | abort: | |
2732 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2733 | } | |
2734 | ||
8763cb45 JG |
2735 | /* |
2736 | * migrate_vma_pages() - migrate meta-data from src page to dst page | |
2737 | * @migrate: migrate struct containing all migration information | |
2738 | * | |
2739 | * This migrates struct page meta-data from source struct page to destination | |
2740 | * struct page. This effectively finishes the migration from source page to the | |
2741 | * destination page. | |
2742 | */ | |
2743 | static void migrate_vma_pages(struct migrate_vma *migrate) | |
2744 | { | |
2745 | const unsigned long npages = migrate->npages; | |
2746 | const unsigned long start = migrate->start; | |
8315ada7 JG |
2747 | struct vm_area_struct *vma = migrate->vma; |
2748 | struct mm_struct *mm = vma->vm_mm; | |
2749 | unsigned long addr, i, mmu_start; | |
2750 | bool notified = false; | |
8763cb45 JG |
2751 | |
2752 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2753 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2754 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2755 | struct address_space *mapping; | |
2756 | int r; | |
2757 | ||
8315ada7 JG |
2758 | if (!newpage) { |
2759 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2760 | continue; |
8315ada7 JG |
2761 | } |
2762 | ||
2763 | if (!page) { | |
2764 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) { | |
2765 | continue; | |
2766 | } | |
2767 | if (!notified) { | |
2768 | mmu_start = addr; | |
2769 | notified = true; | |
2770 | mmu_notifier_invalidate_range_start(mm, | |
2771 | mmu_start, | |
2772 | migrate->end); | |
2773 | } | |
2774 | migrate_vma_insert_page(migrate, addr, newpage, | |
2775 | &migrate->src[i], | |
2776 | &migrate->dst[i]); | |
8763cb45 | 2777 | continue; |
8315ada7 | 2778 | } |
8763cb45 JG |
2779 | |
2780 | mapping = page_mapping(page); | |
2781 | ||
a5430dda JG |
2782 | if (is_zone_device_page(newpage)) { |
2783 | if (is_device_private_page(newpage)) { | |
2784 | /* | |
2785 | * For now only support private anonymous when | |
2786 | * migrating to un-addressable device memory. | |
2787 | */ | |
2788 | if (mapping) { | |
2789 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2790 | continue; | |
2791 | } | |
df6ad698 | 2792 | } else if (!is_device_public_page(newpage)) { |
a5430dda JG |
2793 | /* |
2794 | * Other types of ZONE_DEVICE page are not | |
2795 | * supported. | |
2796 | */ | |
2797 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2798 | continue; | |
2799 | } | |
2800 | } | |
2801 | ||
8763cb45 JG |
2802 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2803 | if (r != MIGRATEPAGE_SUCCESS) | |
2804 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2805 | } | |
8315ada7 JG |
2806 | |
2807 | if (notified) | |
2808 | mmu_notifier_invalidate_range_end(mm, mmu_start, | |
2809 | migrate->end); | |
8763cb45 JG |
2810 | } |
2811 | ||
2812 | /* | |
2813 | * migrate_vma_finalize() - restore CPU page table entry | |
2814 | * @migrate: migrate struct containing all migration information | |
2815 | * | |
2816 | * This replaces the special migration pte entry with either a mapping to the | |
2817 | * new page if migration was successful for that page, or to the original page | |
2818 | * otherwise. | |
2819 | * | |
2820 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2821 | * refcount, for device pages. | |
2822 | */ | |
2823 | static void migrate_vma_finalize(struct migrate_vma *migrate) | |
2824 | { | |
2825 | const unsigned long npages = migrate->npages; | |
2826 | unsigned long i; | |
2827 | ||
2828 | for (i = 0; i < npages; i++) { | |
2829 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2830 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2831 | ||
8315ada7 JG |
2832 | if (!page) { |
2833 | if (newpage) { | |
2834 | unlock_page(newpage); | |
2835 | put_page(newpage); | |
2836 | } | |
8763cb45 | 2837 | continue; |
8315ada7 JG |
2838 | } |
2839 | ||
8763cb45 JG |
2840 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2841 | if (newpage) { | |
2842 | unlock_page(newpage); | |
2843 | put_page(newpage); | |
2844 | } | |
2845 | newpage = page; | |
2846 | } | |
2847 | ||
2848 | remove_migration_ptes(page, newpage, false); | |
2849 | unlock_page(page); | |
2850 | migrate->cpages--; | |
2851 | ||
a5430dda JG |
2852 | if (is_zone_device_page(page)) |
2853 | put_page(page); | |
2854 | else | |
2855 | putback_lru_page(page); | |
8763cb45 JG |
2856 | |
2857 | if (newpage != page) { | |
2858 | unlock_page(newpage); | |
a5430dda JG |
2859 | if (is_zone_device_page(newpage)) |
2860 | put_page(newpage); | |
2861 | else | |
2862 | putback_lru_page(newpage); | |
8763cb45 JG |
2863 | } |
2864 | } | |
2865 | } | |
2866 | ||
2867 | /* | |
2868 | * migrate_vma() - migrate a range of memory inside vma | |
2869 | * | |
2870 | * @ops: migration callback for allocating destination memory and copying | |
2871 | * @vma: virtual memory area containing the range to be migrated | |
2872 | * @start: start address of the range to migrate (inclusive) | |
2873 | * @end: end address of the range to migrate (exclusive) | |
2874 | * @src: array of hmm_pfn_t containing source pfns | |
2875 | * @dst: array of hmm_pfn_t containing destination pfns | |
2876 | * @private: pointer passed back to each of the callback | |
2877 | * Returns: 0 on success, error code otherwise | |
2878 | * | |
2879 | * This function tries to migrate a range of memory virtual address range, using | |
2880 | * callbacks to allocate and copy memory from source to destination. First it | |
2881 | * collects all the pages backing each virtual address in the range, saving this | |
2882 | * inside the src array. Then it locks those pages and unmaps them. Once the pages | |
2883 | * are locked and unmapped, it checks whether each page is pinned or not. Pages | |
2884 | * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) | |
2885 | * in the corresponding src array entry. It then restores any pages that are | |
2886 | * pinned, by remapping and unlocking those pages. | |
2887 | * | |
2888 | * At this point it calls the alloc_and_copy() callback. For documentation on | |
2889 | * what is expected from that callback, see struct migrate_vma_ops comments in | |
2890 | * include/linux/migrate.h | |
2891 | * | |
2892 | * After the alloc_and_copy() callback, this function goes over each entry in | |
2893 | * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2894 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2895 | * then the function tries to migrate struct page information from the source | |
2896 | * struct page to the destination struct page. If it fails to migrate the struct | |
2897 | * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src | |
2898 | * array. | |
2899 | * | |
2900 | * At this point all successfully migrated pages have an entry in the src | |
2901 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2902 | * array entry with MIGRATE_PFN_VALID flag set. | |
2903 | * | |
2904 | * It then calls the finalize_and_map() callback. See comments for "struct | |
2905 | * migrate_vma_ops", in include/linux/migrate.h for details about | |
2906 | * finalize_and_map() behavior. | |
2907 | * | |
2908 | * After the finalize_and_map() callback, for successfully migrated pages, this | |
2909 | * function updates the CPU page table to point to new pages, otherwise it | |
2910 | * restores the CPU page table to point to the original source pages. | |
2911 | * | |
2912 | * Function returns 0 after the above steps, even if no pages were migrated | |
2913 | * (The function only returns an error if any of the arguments are invalid.) | |
2914 | * | |
2915 | * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT | |
2916 | * unsigned long entries. | |
2917 | */ | |
2918 | int migrate_vma(const struct migrate_vma_ops *ops, | |
2919 | struct vm_area_struct *vma, | |
2920 | unsigned long start, | |
2921 | unsigned long end, | |
2922 | unsigned long *src, | |
2923 | unsigned long *dst, | |
2924 | void *private) | |
2925 | { | |
2926 | struct migrate_vma migrate; | |
2927 | ||
2928 | /* Sanity check the arguments */ | |
2929 | start &= PAGE_MASK; | |
2930 | end &= PAGE_MASK; | |
2931 | if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) | |
2932 | return -EINVAL; | |
2933 | if (start < vma->vm_start || start >= vma->vm_end) | |
2934 | return -EINVAL; | |
2935 | if (end <= vma->vm_start || end > vma->vm_end) | |
2936 | return -EINVAL; | |
2937 | if (!ops || !src || !dst || start >= end) | |
2938 | return -EINVAL; | |
2939 | ||
2940 | memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT)); | |
2941 | migrate.src = src; | |
2942 | migrate.dst = dst; | |
2943 | migrate.start = start; | |
2944 | migrate.npages = 0; | |
2945 | migrate.cpages = 0; | |
2946 | migrate.end = end; | |
2947 | migrate.vma = vma; | |
2948 | ||
2949 | /* Collect, and try to unmap source pages */ | |
2950 | migrate_vma_collect(&migrate); | |
2951 | if (!migrate.cpages) | |
2952 | return 0; | |
2953 | ||
2954 | /* Lock and isolate page */ | |
2955 | migrate_vma_prepare(&migrate); | |
2956 | if (!migrate.cpages) | |
2957 | return 0; | |
2958 | ||
2959 | /* Unmap pages */ | |
2960 | migrate_vma_unmap(&migrate); | |
2961 | if (!migrate.cpages) | |
2962 | return 0; | |
2963 | ||
2964 | /* | |
2965 | * At this point pages are locked and unmapped, and thus they have | |
2966 | * stable content and can safely be copied to destination memory that | |
2967 | * is allocated by the callback. | |
2968 | * | |
2969 | * Note that migration can fail in migrate_vma_struct_page() for each | |
2970 | * individual page. | |
2971 | */ | |
2972 | ops->alloc_and_copy(vma, src, dst, start, end, private); | |
2973 | ||
2974 | /* This does the real migration of struct page */ | |
2975 | migrate_vma_pages(&migrate); | |
2976 | ||
2977 | ops->finalize_and_map(vma, src, dst, start, end, private); | |
2978 | ||
2979 | /* Unlock and remap pages */ | |
2980 | migrate_vma_finalize(&migrate); | |
2981 | ||
2982 | return 0; | |
2983 | } | |
2984 | EXPORT_SYMBOL(migrate_vma); | |
6b368cd4 | 2985 | #endif /* defined(MIGRATE_VMA_HELPER) */ |