]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
b20a3503 | 2 | /* |
14e0f9bc | 3 | * Memory Migration functionality - linux/mm/migrate.c |
b20a3503 CL |
4 | * |
5 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
6 | * | |
7 | * Page migration was first developed in the context of the memory hotplug | |
8 | * project. The main authors of the migration code are: | |
9 | * | |
10 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
11 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
12 | * Dave Hansen <haveblue@us.ibm.com> | |
cde53535 | 13 | * Christoph Lameter |
b20a3503 CL |
14 | */ |
15 | ||
16 | #include <linux/migrate.h> | |
b95f1b31 | 17 | #include <linux/export.h> |
b20a3503 | 18 | #include <linux/swap.h> |
0697212a | 19 | #include <linux/swapops.h> |
b20a3503 | 20 | #include <linux/pagemap.h> |
e23ca00b | 21 | #include <linux/buffer_head.h> |
b20a3503 | 22 | #include <linux/mm_inline.h> |
b488893a | 23 | #include <linux/nsproxy.h> |
b20a3503 | 24 | #include <linux/pagevec.h> |
e9995ef9 | 25 | #include <linux/ksm.h> |
b20a3503 CL |
26 | #include <linux/rmap.h> |
27 | #include <linux/topology.h> | |
28 | #include <linux/cpu.h> | |
29 | #include <linux/cpuset.h> | |
04e62a29 | 30 | #include <linux/writeback.h> |
742755a1 CL |
31 | #include <linux/mempolicy.h> |
32 | #include <linux/vmalloc.h> | |
86c3a764 | 33 | #include <linux/security.h> |
42cb14b1 | 34 | #include <linux/backing-dev.h> |
bda807d4 | 35 | #include <linux/compaction.h> |
4f5ca265 | 36 | #include <linux/syscalls.h> |
7addf443 | 37 | #include <linux/compat.h> |
290408d4 | 38 | #include <linux/hugetlb.h> |
8e6ac7fa | 39 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 40 | #include <linux/gfp.h> |
a520110e | 41 | #include <linux/pagewalk.h> |
df6ad698 | 42 | #include <linux/pfn_t.h> |
a5430dda | 43 | #include <linux/memremap.h> |
8315ada7 | 44 | #include <linux/userfaultfd_k.h> |
bf6bddf1 | 45 | #include <linux/balloon_compaction.h> |
f714f4f2 | 46 | #include <linux/mmu_notifier.h> |
33c3fc71 | 47 | #include <linux/page_idle.h> |
d435edca | 48 | #include <linux/page_owner.h> |
6e84f315 | 49 | #include <linux/sched/mm.h> |
197e7e52 | 50 | #include <linux/ptrace.h> |
b20a3503 | 51 | |
0d1836c3 MN |
52 | #include <asm/tlbflush.h> |
53 | ||
7b2a2d4a MG |
54 | #define CREATE_TRACE_POINTS |
55 | #include <trace/events/migrate.h> | |
56 | ||
b20a3503 CL |
57 | #include "internal.h" |
58 | ||
b20a3503 | 59 | /* |
742755a1 | 60 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
61 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
62 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
63 | */ |
64 | int migrate_prep(void) | |
65 | { | |
b20a3503 CL |
66 | /* |
67 | * Clear the LRU lists so pages can be isolated. | |
68 | * Note that pages may be moved off the LRU after we have | |
69 | * drained them. Those pages will fail to migrate like other | |
70 | * pages that may be busy. | |
71 | */ | |
72 | lru_add_drain_all(); | |
73 | ||
74 | return 0; | |
75 | } | |
76 | ||
748446bb MG |
77 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
78 | int migrate_prep_local(void) | |
79 | { | |
80 | lru_add_drain(); | |
81 | ||
82 | return 0; | |
83 | } | |
84 | ||
9e5bcd61 | 85 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
86 | { |
87 | struct address_space *mapping; | |
88 | ||
89 | /* | |
90 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
91 | * or just got freed under us. | |
92 | * | |
93 | * In case we 'win' a race for a movable page being freed under us and | |
94 | * raise its refcount preventing __free_pages() from doing its job | |
95 | * the put_page() at the end of this block will take care of | |
96 | * release this page, thus avoiding a nasty leakage. | |
97 | */ | |
98 | if (unlikely(!get_page_unless_zero(page))) | |
99 | goto out; | |
100 | ||
101 | /* | |
102 | * Check PageMovable before holding a PG_lock because page's owner | |
103 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 104 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
105 | */ |
106 | if (unlikely(!__PageMovable(page))) | |
107 | goto out_putpage; | |
108 | /* | |
109 | * As movable pages are not isolated from LRU lists, concurrent | |
110 | * compaction threads can race against page migration functions | |
111 | * as well as race against the releasing a page. | |
112 | * | |
113 | * In order to avoid having an already isolated movable page | |
114 | * being (wrongly) re-isolated while it is under migration, | |
115 | * or to avoid attempting to isolate pages being released, | |
116 | * lets be sure we have the page lock | |
117 | * before proceeding with the movable page isolation steps. | |
118 | */ | |
119 | if (unlikely(!trylock_page(page))) | |
120 | goto out_putpage; | |
121 | ||
122 | if (!PageMovable(page) || PageIsolated(page)) | |
123 | goto out_no_isolated; | |
124 | ||
125 | mapping = page_mapping(page); | |
126 | VM_BUG_ON_PAGE(!mapping, page); | |
127 | ||
128 | if (!mapping->a_ops->isolate_page(page, mode)) | |
129 | goto out_no_isolated; | |
130 | ||
131 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
132 | WARN_ON_ONCE(PageIsolated(page)); | |
133 | __SetPageIsolated(page); | |
134 | unlock_page(page); | |
135 | ||
9e5bcd61 | 136 | return 0; |
bda807d4 MK |
137 | |
138 | out_no_isolated: | |
139 | unlock_page(page); | |
140 | out_putpage: | |
141 | put_page(page); | |
142 | out: | |
9e5bcd61 | 143 | return -EBUSY; |
bda807d4 MK |
144 | } |
145 | ||
146 | /* It should be called on page which is PG_movable */ | |
147 | void putback_movable_page(struct page *page) | |
148 | { | |
149 | struct address_space *mapping; | |
150 | ||
151 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
152 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
153 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
154 | ||
155 | mapping = page_mapping(page); | |
156 | mapping->a_ops->putback_page(page); | |
157 | __ClearPageIsolated(page); | |
158 | } | |
159 | ||
5733c7d1 RA |
160 | /* |
161 | * Put previously isolated pages back onto the appropriate lists | |
162 | * from where they were once taken off for compaction/migration. | |
163 | * | |
59c82b70 JK |
164 | * This function shall be used whenever the isolated pageset has been |
165 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
166 | * and isolate_huge_page(). | |
5733c7d1 RA |
167 | */ |
168 | void putback_movable_pages(struct list_head *l) | |
169 | { | |
170 | struct page *page; | |
171 | struct page *page2; | |
172 | ||
b20a3503 | 173 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
174 | if (unlikely(PageHuge(page))) { |
175 | putback_active_hugepage(page); | |
176 | continue; | |
177 | } | |
e24f0b8f | 178 | list_del(&page->lru); |
bda807d4 MK |
179 | /* |
180 | * We isolated non-lru movable page so here we can use | |
181 | * __PageMovable because LRU page's mapping cannot have | |
182 | * PAGE_MAPPING_MOVABLE. | |
183 | */ | |
b1123ea6 | 184 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
185 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
186 | lock_page(page); | |
187 | if (PageMovable(page)) | |
188 | putback_movable_page(page); | |
189 | else | |
190 | __ClearPageIsolated(page); | |
191 | unlock_page(page); | |
192 | put_page(page); | |
193 | } else { | |
e8db67eb NH |
194 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
195 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
fc280fe8 | 196 | putback_lru_page(page); |
bda807d4 | 197 | } |
b20a3503 | 198 | } |
b20a3503 CL |
199 | } |
200 | ||
0697212a CL |
201 | /* |
202 | * Restore a potential migration pte to a working pte entry | |
203 | */ | |
e4b82222 | 204 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 205 | unsigned long addr, void *old) |
0697212a | 206 | { |
3fe87967 KS |
207 | struct page_vma_mapped_walk pvmw = { |
208 | .page = old, | |
209 | .vma = vma, | |
210 | .address = addr, | |
211 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
212 | }; | |
213 | struct page *new; | |
214 | pte_t pte; | |
0697212a | 215 | swp_entry_t entry; |
0697212a | 216 | |
3fe87967 KS |
217 | VM_BUG_ON_PAGE(PageTail(page), page); |
218 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
219 | if (PageKsm(page)) |
220 | new = page; | |
221 | else | |
222 | new = page - pvmw.page->index + | |
223 | linear_page_index(vma, pvmw.address); | |
0697212a | 224 | |
616b8371 ZY |
225 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
226 | /* PMD-mapped THP migration entry */ | |
227 | if (!pvmw.pte) { | |
228 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
229 | remove_migration_pmd(&pvmw, new); | |
230 | continue; | |
231 | } | |
232 | #endif | |
233 | ||
3fe87967 KS |
234 | get_page(new); |
235 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
236 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
237 | pte = pte_mksoft_dirty(pte); | |
0697212a | 238 | |
3fe87967 KS |
239 | /* |
240 | * Recheck VMA as permissions can change since migration started | |
241 | */ | |
242 | entry = pte_to_swp_entry(*pvmw.pte); | |
243 | if (is_write_migration_entry(entry)) | |
244 | pte = maybe_mkwrite(pte, vma); | |
d3cb8bf6 | 245 | |
df6ad698 JG |
246 | if (unlikely(is_zone_device_page(new))) { |
247 | if (is_device_private_page(new)) { | |
248 | entry = make_device_private_entry(new, pte_write(pte)); | |
249 | pte = swp_entry_to_pte(entry); | |
df6ad698 | 250 | } |
d2b2c6dd | 251 | } |
a5430dda | 252 | |
3ef8fd7f | 253 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 KS |
254 | if (PageHuge(new)) { |
255 | pte = pte_mkhuge(pte); | |
256 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
383321ab | 257 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
258 | if (PageAnon(new)) |
259 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
260 | else | |
261 | page_dup_rmap(new, true); | |
383321ab AK |
262 | } else |
263 | #endif | |
264 | { | |
265 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 266 | |
383321ab AK |
267 | if (PageAnon(new)) |
268 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
269 | else | |
270 | page_add_file_rmap(new, false); | |
271 | } | |
3fe87967 KS |
272 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
273 | mlock_vma_page(new); | |
274 | ||
e125fe40 KS |
275 | if (PageTransHuge(page) && PageMlocked(page)) |
276 | clear_page_mlock(page); | |
277 | ||
3fe87967 KS |
278 | /* No need to invalidate - it was non-present before */ |
279 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
280 | } | |
51afb12b | 281 | |
e4b82222 | 282 | return true; |
0697212a CL |
283 | } |
284 | ||
04e62a29 CL |
285 | /* |
286 | * Get rid of all migration entries and replace them by | |
287 | * references to the indicated page. | |
288 | */ | |
e388466d | 289 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 290 | { |
051ac83a JK |
291 | struct rmap_walk_control rwc = { |
292 | .rmap_one = remove_migration_pte, | |
293 | .arg = old, | |
294 | }; | |
295 | ||
e388466d KS |
296 | if (locked) |
297 | rmap_walk_locked(new, &rwc); | |
298 | else | |
299 | rmap_walk(new, &rwc); | |
04e62a29 CL |
300 | } |
301 | ||
0697212a CL |
302 | /* |
303 | * Something used the pte of a page under migration. We need to | |
304 | * get to the page and wait until migration is finished. | |
305 | * When we return from this function the fault will be retried. | |
0697212a | 306 | */ |
e66f17ff | 307 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 308 | spinlock_t *ptl) |
0697212a | 309 | { |
30dad309 | 310 | pte_t pte; |
0697212a CL |
311 | swp_entry_t entry; |
312 | struct page *page; | |
313 | ||
30dad309 | 314 | spin_lock(ptl); |
0697212a CL |
315 | pte = *ptep; |
316 | if (!is_swap_pte(pte)) | |
317 | goto out; | |
318 | ||
319 | entry = pte_to_swp_entry(pte); | |
320 | if (!is_migration_entry(entry)) | |
321 | goto out; | |
322 | ||
323 | page = migration_entry_to_page(entry); | |
324 | ||
e286781d | 325 | /* |
89eb946a | 326 | * Once page cache replacement of page migration started, page_count |
9a1ea439 HD |
327 | * is zero; but we must not call put_and_wait_on_page_locked() without |
328 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
e286781d NP |
329 | */ |
330 | if (!get_page_unless_zero(page)) | |
331 | goto out; | |
0697212a | 332 | pte_unmap_unlock(ptep, ptl); |
9a1ea439 | 333 | put_and_wait_on_page_locked(page); |
0697212a CL |
334 | return; |
335 | out: | |
336 | pte_unmap_unlock(ptep, ptl); | |
337 | } | |
338 | ||
30dad309 NH |
339 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
340 | unsigned long address) | |
341 | { | |
342 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
343 | pte_t *ptep = pte_offset_map(pmd, address); | |
344 | __migration_entry_wait(mm, ptep, ptl); | |
345 | } | |
346 | ||
cb900f41 KS |
347 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
348 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 349 | { |
cb900f41 | 350 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
351 | __migration_entry_wait(mm, pte, ptl); |
352 | } | |
353 | ||
616b8371 ZY |
354 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
355 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
356 | { | |
357 | spinlock_t *ptl; | |
358 | struct page *page; | |
359 | ||
360 | ptl = pmd_lock(mm, pmd); | |
361 | if (!is_pmd_migration_entry(*pmd)) | |
362 | goto unlock; | |
363 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
364 | if (!get_page_unless_zero(page)) | |
365 | goto unlock; | |
366 | spin_unlock(ptl); | |
9a1ea439 | 367 | put_and_wait_on_page_locked(page); |
616b8371 ZY |
368 | return; |
369 | unlock: | |
370 | spin_unlock(ptl); | |
371 | } | |
372 | #endif | |
373 | ||
f900482d | 374 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
0b3901b3 JK |
375 | { |
376 | int expected_count = 1; | |
377 | ||
378 | /* | |
379 | * Device public or private pages have an extra refcount as they are | |
380 | * ZONE_DEVICE pages. | |
381 | */ | |
382 | expected_count += is_device_private_page(page); | |
f900482d | 383 | if (mapping) |
0b3901b3 JK |
384 | expected_count += hpage_nr_pages(page) + page_has_private(page); |
385 | ||
386 | return expected_count; | |
387 | } | |
388 | ||
b20a3503 | 389 | /* |
c3fcf8a5 | 390 | * Replace the page in the mapping. |
5b5c7120 CL |
391 | * |
392 | * The number of remaining references must be: | |
393 | * 1 for anonymous pages without a mapping | |
394 | * 2 for pages with a mapping | |
266cf658 | 395 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 396 | */ |
36bc08cc | 397 | int migrate_page_move_mapping(struct address_space *mapping, |
37109694 | 398 | struct page *newpage, struct page *page, int extra_count) |
b20a3503 | 399 | { |
89eb946a | 400 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
42cb14b1 HD |
401 | struct zone *oldzone, *newzone; |
402 | int dirty; | |
f900482d | 403 | int expected_count = expected_page_refs(mapping, page) + extra_count; |
8763cb45 | 404 | |
6c5240ae | 405 | if (!mapping) { |
0e8c7d0f | 406 | /* Anonymous page without mapping */ |
8e321fef | 407 | if (page_count(page) != expected_count) |
6c5240ae | 408 | return -EAGAIN; |
cf4b769a HD |
409 | |
410 | /* No turning back from here */ | |
cf4b769a HD |
411 | newpage->index = page->index; |
412 | newpage->mapping = page->mapping; | |
413 | if (PageSwapBacked(page)) | |
fa9949da | 414 | __SetPageSwapBacked(newpage); |
cf4b769a | 415 | |
78bd5209 | 416 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
417 | } |
418 | ||
42cb14b1 HD |
419 | oldzone = page_zone(page); |
420 | newzone = page_zone(newpage); | |
421 | ||
89eb946a | 422 | xas_lock_irq(&xas); |
89eb946a MW |
423 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
424 | xas_unlock_irq(&xas); | |
e23ca00b | 425 | return -EAGAIN; |
b20a3503 CL |
426 | } |
427 | ||
fe896d18 | 428 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 429 | xas_unlock_irq(&xas); |
e286781d NP |
430 | return -EAGAIN; |
431 | } | |
432 | ||
b20a3503 | 433 | /* |
cf4b769a HD |
434 | * Now we know that no one else is looking at the page: |
435 | * no turning back from here. | |
b20a3503 | 436 | */ |
cf4b769a HD |
437 | newpage->index = page->index; |
438 | newpage->mapping = page->mapping; | |
e71769ae | 439 | page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */ |
6326fec1 NP |
440 | if (PageSwapBacked(page)) { |
441 | __SetPageSwapBacked(newpage); | |
442 | if (PageSwapCache(page)) { | |
443 | SetPageSwapCache(newpage); | |
444 | set_page_private(newpage, page_private(page)); | |
445 | } | |
446 | } else { | |
447 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
448 | } |
449 | ||
42cb14b1 HD |
450 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
451 | dirty = PageDirty(page); | |
452 | if (dirty) { | |
453 | ClearPageDirty(page); | |
454 | SetPageDirty(newpage); | |
455 | } | |
456 | ||
89eb946a | 457 | xas_store(&xas, newpage); |
e71769ae NH |
458 | if (PageTransHuge(page)) { |
459 | int i; | |
e71769ae | 460 | |
013567be | 461 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
89eb946a | 462 | xas_next(&xas); |
4101196b | 463 | xas_store(&xas, newpage); |
e71769ae | 464 | } |
e71769ae | 465 | } |
7cf9c2c7 NP |
466 | |
467 | /* | |
937a94c9 JG |
468 | * Drop cache reference from old page by unfreezing |
469 | * to one less reference. | |
7cf9c2c7 NP |
470 | * We know this isn't the last reference. |
471 | */ | |
e71769ae | 472 | page_ref_unfreeze(page, expected_count - hpage_nr_pages(page)); |
7cf9c2c7 | 473 | |
89eb946a | 474 | xas_unlock(&xas); |
42cb14b1 HD |
475 | /* Leave irq disabled to prevent preemption while updating stats */ |
476 | ||
0e8c7d0f CL |
477 | /* |
478 | * If moved to a different zone then also account | |
479 | * the page for that zone. Other VM counters will be | |
480 | * taken care of when we establish references to the | |
481 | * new page and drop references to the old page. | |
482 | * | |
483 | * Note that anonymous pages are accounted for | |
4b9d0fab | 484 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
485 | * are mapped to swap space. |
486 | */ | |
42cb14b1 | 487 | if (newzone != oldzone) { |
11fb9989 MG |
488 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); |
489 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); | |
42cb14b1 | 490 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
11fb9989 MG |
491 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); |
492 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); | |
42cb14b1 HD |
493 | } |
494 | if (dirty && mapping_cap_account_dirty(mapping)) { | |
11fb9989 | 495 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 496 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); |
11fb9989 | 497 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 498 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); |
42cb14b1 | 499 | } |
4b02108a | 500 | } |
42cb14b1 | 501 | local_irq_enable(); |
b20a3503 | 502 | |
78bd5209 | 503 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 504 | } |
1118dce7 | 505 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 506 | |
290408d4 NH |
507 | /* |
508 | * The expected number of remaining references is the same as that | |
509 | * of migrate_page_move_mapping(). | |
510 | */ | |
511 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
512 | struct page *newpage, struct page *page) | |
513 | { | |
89eb946a | 514 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 515 | int expected_count; |
290408d4 | 516 | |
89eb946a | 517 | xas_lock_irq(&xas); |
290408d4 | 518 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
519 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
520 | xas_unlock_irq(&xas); | |
290408d4 NH |
521 | return -EAGAIN; |
522 | } | |
523 | ||
fe896d18 | 524 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 525 | xas_unlock_irq(&xas); |
290408d4 NH |
526 | return -EAGAIN; |
527 | } | |
528 | ||
cf4b769a HD |
529 | newpage->index = page->index; |
530 | newpage->mapping = page->mapping; | |
6a93ca8f | 531 | |
290408d4 NH |
532 | get_page(newpage); |
533 | ||
89eb946a | 534 | xas_store(&xas, newpage); |
290408d4 | 535 | |
fe896d18 | 536 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 537 | |
89eb946a | 538 | xas_unlock_irq(&xas); |
6a93ca8f | 539 | |
78bd5209 | 540 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
541 | } |
542 | ||
30b0a105 DH |
543 | /* |
544 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
545 | * arithmetic will work across the entire page. We need something more | |
546 | * specialized. | |
547 | */ | |
548 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
549 | int nr_pages) | |
550 | { | |
551 | int i; | |
552 | struct page *dst_base = dst; | |
553 | struct page *src_base = src; | |
554 | ||
555 | for (i = 0; i < nr_pages; ) { | |
556 | cond_resched(); | |
557 | copy_highpage(dst, src); | |
558 | ||
559 | i++; | |
560 | dst = mem_map_next(dst, dst_base, i); | |
561 | src = mem_map_next(src, src_base, i); | |
562 | } | |
563 | } | |
564 | ||
565 | static void copy_huge_page(struct page *dst, struct page *src) | |
566 | { | |
567 | int i; | |
568 | int nr_pages; | |
569 | ||
570 | if (PageHuge(src)) { | |
571 | /* hugetlbfs page */ | |
572 | struct hstate *h = page_hstate(src); | |
573 | nr_pages = pages_per_huge_page(h); | |
574 | ||
575 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
576 | __copy_gigantic_page(dst, src, nr_pages); | |
577 | return; | |
578 | } | |
579 | } else { | |
580 | /* thp page */ | |
581 | BUG_ON(!PageTransHuge(src)); | |
582 | nr_pages = hpage_nr_pages(src); | |
583 | } | |
584 | ||
585 | for (i = 0; i < nr_pages; i++) { | |
586 | cond_resched(); | |
587 | copy_highpage(dst + i, src + i); | |
588 | } | |
589 | } | |
590 | ||
b20a3503 CL |
591 | /* |
592 | * Copy the page to its new location | |
593 | */ | |
2916ecc0 | 594 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 595 | { |
7851a45c RR |
596 | int cpupid; |
597 | ||
b20a3503 CL |
598 | if (PageError(page)) |
599 | SetPageError(newpage); | |
600 | if (PageReferenced(page)) | |
601 | SetPageReferenced(newpage); | |
602 | if (PageUptodate(page)) | |
603 | SetPageUptodate(newpage); | |
894bc310 | 604 | if (TestClearPageActive(page)) { |
309381fe | 605 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 606 | SetPageActive(newpage); |
418b27ef LS |
607 | } else if (TestClearPageUnevictable(page)) |
608 | SetPageUnevictable(newpage); | |
1899ad18 JW |
609 | if (PageWorkingset(page)) |
610 | SetPageWorkingset(newpage); | |
b20a3503 CL |
611 | if (PageChecked(page)) |
612 | SetPageChecked(newpage); | |
613 | if (PageMappedToDisk(page)) | |
614 | SetPageMappedToDisk(newpage); | |
615 | ||
42cb14b1 HD |
616 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
617 | if (PageDirty(page)) | |
618 | SetPageDirty(newpage); | |
b20a3503 | 619 | |
33c3fc71 VD |
620 | if (page_is_young(page)) |
621 | set_page_young(newpage); | |
622 | if (page_is_idle(page)) | |
623 | set_page_idle(newpage); | |
624 | ||
7851a45c RR |
625 | /* |
626 | * Copy NUMA information to the new page, to prevent over-eager | |
627 | * future migrations of this same page. | |
628 | */ | |
629 | cpupid = page_cpupid_xchg_last(page, -1); | |
630 | page_cpupid_xchg_last(newpage, cpupid); | |
631 | ||
e9995ef9 | 632 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
633 | /* |
634 | * Please do not reorder this without considering how mm/ksm.c's | |
635 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
636 | */ | |
b3b3a99c NH |
637 | if (PageSwapCache(page)) |
638 | ClearPageSwapCache(page); | |
b20a3503 CL |
639 | ClearPagePrivate(page); |
640 | set_page_private(page, 0); | |
b20a3503 CL |
641 | |
642 | /* | |
643 | * If any waiters have accumulated on the new page then | |
644 | * wake them up. | |
645 | */ | |
646 | if (PageWriteback(newpage)) | |
647 | end_page_writeback(newpage); | |
d435edca VB |
648 | |
649 | copy_page_owner(page, newpage); | |
74485cf2 JW |
650 | |
651 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 652 | } |
2916ecc0 JG |
653 | EXPORT_SYMBOL(migrate_page_states); |
654 | ||
655 | void migrate_page_copy(struct page *newpage, struct page *page) | |
656 | { | |
657 | if (PageHuge(page) || PageTransHuge(page)) | |
658 | copy_huge_page(newpage, page); | |
659 | else | |
660 | copy_highpage(newpage, page); | |
661 | ||
662 | migrate_page_states(newpage, page); | |
663 | } | |
1118dce7 | 664 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 665 | |
1d8b85cc CL |
666 | /************************************************************ |
667 | * Migration functions | |
668 | ***********************************************************/ | |
669 | ||
b20a3503 | 670 | /* |
bda807d4 | 671 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 672 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
673 | * |
674 | * Pages are locked upon entry and exit. | |
675 | */ | |
2d1db3b1 | 676 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
677 | struct page *newpage, struct page *page, |
678 | enum migrate_mode mode) | |
b20a3503 CL |
679 | { |
680 | int rc; | |
681 | ||
682 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
683 | ||
37109694 | 684 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
b20a3503 | 685 | |
78bd5209 | 686 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
687 | return rc; |
688 | ||
2916ecc0 JG |
689 | if (mode != MIGRATE_SYNC_NO_COPY) |
690 | migrate_page_copy(newpage, page); | |
691 | else | |
692 | migrate_page_states(newpage, page); | |
78bd5209 | 693 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
694 | } |
695 | EXPORT_SYMBOL(migrate_page); | |
696 | ||
9361401e | 697 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
698 | /* Returns true if all buffers are successfully locked */ |
699 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
700 | enum migrate_mode mode) | |
701 | { | |
702 | struct buffer_head *bh = head; | |
703 | ||
704 | /* Simple case, sync compaction */ | |
705 | if (mode != MIGRATE_ASYNC) { | |
706 | do { | |
84ade7c1 JK |
707 | lock_buffer(bh); |
708 | bh = bh->b_this_page; | |
709 | ||
710 | } while (bh != head); | |
711 | ||
712 | return true; | |
713 | } | |
714 | ||
715 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
716 | do { | |
84ade7c1 JK |
717 | if (!trylock_buffer(bh)) { |
718 | /* | |
719 | * We failed to lock the buffer and cannot stall in | |
720 | * async migration. Release the taken locks | |
721 | */ | |
722 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
723 | bh = head; |
724 | while (bh != failed_bh) { | |
725 | unlock_buffer(bh); | |
84ade7c1 JK |
726 | bh = bh->b_this_page; |
727 | } | |
728 | return false; | |
729 | } | |
730 | ||
731 | bh = bh->b_this_page; | |
732 | } while (bh != head); | |
733 | return true; | |
734 | } | |
735 | ||
89cb0888 JK |
736 | static int __buffer_migrate_page(struct address_space *mapping, |
737 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
738 | bool check_refs) | |
1d8b85cc | 739 | { |
1d8b85cc CL |
740 | struct buffer_head *bh, *head; |
741 | int rc; | |
cc4f11e6 | 742 | int expected_count; |
1d8b85cc | 743 | |
1d8b85cc | 744 | if (!page_has_buffers(page)) |
a6bc32b8 | 745 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 746 | |
cc4f11e6 | 747 | /* Check whether page does not have extra refs before we do more work */ |
f900482d | 748 | expected_count = expected_page_refs(mapping, page); |
cc4f11e6 JK |
749 | if (page_count(page) != expected_count) |
750 | return -EAGAIN; | |
1d8b85cc | 751 | |
cc4f11e6 JK |
752 | head = page_buffers(page); |
753 | if (!buffer_migrate_lock_buffers(head, mode)) | |
754 | return -EAGAIN; | |
1d8b85cc | 755 | |
89cb0888 JK |
756 | if (check_refs) { |
757 | bool busy; | |
758 | bool invalidated = false; | |
759 | ||
760 | recheck_buffers: | |
761 | busy = false; | |
762 | spin_lock(&mapping->private_lock); | |
763 | bh = head; | |
764 | do { | |
765 | if (atomic_read(&bh->b_count)) { | |
766 | busy = true; | |
767 | break; | |
768 | } | |
769 | bh = bh->b_this_page; | |
770 | } while (bh != head); | |
89cb0888 JK |
771 | if (busy) { |
772 | if (invalidated) { | |
773 | rc = -EAGAIN; | |
774 | goto unlock_buffers; | |
775 | } | |
ebdf4de5 | 776 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
777 | invalidate_bh_lrus(); |
778 | invalidated = true; | |
779 | goto recheck_buffers; | |
780 | } | |
781 | } | |
782 | ||
37109694 | 783 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
78bd5209 | 784 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 785 | goto unlock_buffers; |
1d8b85cc CL |
786 | |
787 | ClearPagePrivate(page); | |
788 | set_page_private(newpage, page_private(page)); | |
789 | set_page_private(page, 0); | |
790 | put_page(page); | |
791 | get_page(newpage); | |
792 | ||
793 | bh = head; | |
794 | do { | |
795 | set_bh_page(bh, newpage, bh_offset(bh)); | |
796 | bh = bh->b_this_page; | |
797 | ||
798 | } while (bh != head); | |
799 | ||
800 | SetPagePrivate(newpage); | |
801 | ||
2916ecc0 JG |
802 | if (mode != MIGRATE_SYNC_NO_COPY) |
803 | migrate_page_copy(newpage, page); | |
804 | else | |
805 | migrate_page_states(newpage, page); | |
1d8b85cc | 806 | |
cc4f11e6 JK |
807 | rc = MIGRATEPAGE_SUCCESS; |
808 | unlock_buffers: | |
ebdf4de5 JK |
809 | if (check_refs) |
810 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
811 | bh = head; |
812 | do { | |
813 | unlock_buffer(bh); | |
1d8b85cc CL |
814 | bh = bh->b_this_page; |
815 | ||
816 | } while (bh != head); | |
817 | ||
cc4f11e6 | 818 | return rc; |
1d8b85cc | 819 | } |
89cb0888 JK |
820 | |
821 | /* | |
822 | * Migration function for pages with buffers. This function can only be used | |
823 | * if the underlying filesystem guarantees that no other references to "page" | |
824 | * exist. For example attached buffer heads are accessed only under page lock. | |
825 | */ | |
826 | int buffer_migrate_page(struct address_space *mapping, | |
827 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
828 | { | |
829 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
830 | } | |
1d8b85cc | 831 | EXPORT_SYMBOL(buffer_migrate_page); |
89cb0888 JK |
832 | |
833 | /* | |
834 | * Same as above except that this variant is more careful and checks that there | |
835 | * are also no buffer head references. This function is the right one for | |
836 | * mappings where buffer heads are directly looked up and referenced (such as | |
837 | * block device mappings). | |
838 | */ | |
839 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
840 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
841 | { | |
842 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
843 | } | |
9361401e | 844 | #endif |
1d8b85cc | 845 | |
04e62a29 CL |
846 | /* |
847 | * Writeback a page to clean the dirty state | |
848 | */ | |
849 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 850 | { |
04e62a29 CL |
851 | struct writeback_control wbc = { |
852 | .sync_mode = WB_SYNC_NONE, | |
853 | .nr_to_write = 1, | |
854 | .range_start = 0, | |
855 | .range_end = LLONG_MAX, | |
04e62a29 CL |
856 | .for_reclaim = 1 |
857 | }; | |
858 | int rc; | |
859 | ||
860 | if (!mapping->a_ops->writepage) | |
861 | /* No write method for the address space */ | |
862 | return -EINVAL; | |
863 | ||
864 | if (!clear_page_dirty_for_io(page)) | |
865 | /* Someone else already triggered a write */ | |
866 | return -EAGAIN; | |
867 | ||
8351a6e4 | 868 | /* |
04e62a29 CL |
869 | * A dirty page may imply that the underlying filesystem has |
870 | * the page on some queue. So the page must be clean for | |
871 | * migration. Writeout may mean we loose the lock and the | |
872 | * page state is no longer what we checked for earlier. | |
873 | * At this point we know that the migration attempt cannot | |
874 | * be successful. | |
8351a6e4 | 875 | */ |
e388466d | 876 | remove_migration_ptes(page, page, false); |
8351a6e4 | 877 | |
04e62a29 | 878 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 879 | |
04e62a29 CL |
880 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
881 | /* unlocked. Relock */ | |
882 | lock_page(page); | |
883 | ||
bda8550d | 884 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
885 | } |
886 | ||
887 | /* | |
888 | * Default handling if a filesystem does not provide a migration function. | |
889 | */ | |
890 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 891 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 892 | { |
b969c4ab | 893 | if (PageDirty(page)) { |
a6bc32b8 | 894 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
895 | switch (mode) { |
896 | case MIGRATE_SYNC: | |
897 | case MIGRATE_SYNC_NO_COPY: | |
898 | break; | |
899 | default: | |
b969c4ab | 900 | return -EBUSY; |
2916ecc0 | 901 | } |
04e62a29 | 902 | return writeout(mapping, page); |
b969c4ab | 903 | } |
8351a6e4 CL |
904 | |
905 | /* | |
906 | * Buffers may be managed in a filesystem specific way. | |
907 | * We must have no buffers or drop them. | |
908 | */ | |
266cf658 | 909 | if (page_has_private(page) && |
8351a6e4 | 910 | !try_to_release_page(page, GFP_KERNEL)) |
806031bb | 911 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 912 | |
a6bc32b8 | 913 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
914 | } |
915 | ||
e24f0b8f CL |
916 | /* |
917 | * Move a page to a newly allocated page | |
918 | * The page is locked and all ptes have been successfully removed. | |
919 | * | |
920 | * The new page will have replaced the old page if this function | |
921 | * is successful. | |
894bc310 LS |
922 | * |
923 | * Return value: | |
924 | * < 0 - error code | |
78bd5209 | 925 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 926 | */ |
3fe2011f | 927 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 928 | enum migrate_mode mode) |
e24f0b8f CL |
929 | { |
930 | struct address_space *mapping; | |
bda807d4 MK |
931 | int rc = -EAGAIN; |
932 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 933 | |
7db7671f HD |
934 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
935 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 936 | |
e24f0b8f | 937 | mapping = page_mapping(page); |
bda807d4 MK |
938 | |
939 | if (likely(is_lru)) { | |
940 | if (!mapping) | |
941 | rc = migrate_page(mapping, newpage, page, mode); | |
942 | else if (mapping->a_ops->migratepage) | |
943 | /* | |
944 | * Most pages have a mapping and most filesystems | |
945 | * provide a migratepage callback. Anonymous pages | |
946 | * are part of swap space which also has its own | |
947 | * migratepage callback. This is the most common path | |
948 | * for page migration. | |
949 | */ | |
950 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
951 | page, mode); | |
952 | else | |
953 | rc = fallback_migrate_page(mapping, newpage, | |
954 | page, mode); | |
955 | } else { | |
e24f0b8f | 956 | /* |
bda807d4 MK |
957 | * In case of non-lru page, it could be released after |
958 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 959 | */ |
bda807d4 MK |
960 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
961 | if (!PageMovable(page)) { | |
962 | rc = MIGRATEPAGE_SUCCESS; | |
963 | __ClearPageIsolated(page); | |
964 | goto out; | |
965 | } | |
966 | ||
967 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
968 | page, mode); | |
969 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
970 | !PageIsolated(page)); | |
971 | } | |
e24f0b8f | 972 | |
5c3f9a67 HD |
973 | /* |
974 | * When successful, old pagecache page->mapping must be cleared before | |
975 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
976 | */ | |
977 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
978 | if (__PageMovable(page)) { |
979 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
980 | ||
981 | /* | |
982 | * We clear PG_movable under page_lock so any compactor | |
983 | * cannot try to migrate this page. | |
984 | */ | |
985 | __ClearPageIsolated(page); | |
986 | } | |
987 | ||
988 | /* | |
989 | * Anonymous and movable page->mapping will be cleard by | |
990 | * free_pages_prepare so don't reset it here for keeping | |
991 | * the type to work PageAnon, for example. | |
992 | */ | |
993 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 994 | page->mapping = NULL; |
d2b2c6dd | 995 | |
25b2995a | 996 | if (likely(!is_zone_device_page(newpage))) |
d2b2c6dd LP |
997 | flush_dcache_page(newpage); |
998 | ||
3fe2011f | 999 | } |
bda807d4 | 1000 | out: |
e24f0b8f CL |
1001 | return rc; |
1002 | } | |
1003 | ||
0dabec93 | 1004 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 1005 | int force, enum migrate_mode mode) |
e24f0b8f | 1006 | { |
0dabec93 | 1007 | int rc = -EAGAIN; |
2ebba6b7 | 1008 | int page_was_mapped = 0; |
3f6c8272 | 1009 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 1010 | bool is_lru = !__PageMovable(page); |
95a402c3 | 1011 | |
529ae9aa | 1012 | if (!trylock_page(page)) { |
a6bc32b8 | 1013 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 1014 | goto out; |
3e7d3449 MG |
1015 | |
1016 | /* | |
1017 | * It's not safe for direct compaction to call lock_page. | |
1018 | * For example, during page readahead pages are added locked | |
1019 | * to the LRU. Later, when the IO completes the pages are | |
1020 | * marked uptodate and unlocked. However, the queueing | |
1021 | * could be merging multiple pages for one bio (e.g. | |
1022 | * mpage_readpages). If an allocation happens for the | |
1023 | * second or third page, the process can end up locking | |
1024 | * the same page twice and deadlocking. Rather than | |
1025 | * trying to be clever about what pages can be locked, | |
1026 | * avoid the use of lock_page for direct compaction | |
1027 | * altogether. | |
1028 | */ | |
1029 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1030 | goto out; |
3e7d3449 | 1031 | |
e24f0b8f CL |
1032 | lock_page(page); |
1033 | } | |
1034 | ||
1035 | if (PageWriteback(page)) { | |
11bc82d6 | 1036 | /* |
fed5b64a | 1037 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1038 | * necessary to wait for PageWriteback. In the async case, |
1039 | * the retry loop is too short and in the sync-light case, | |
1040 | * the overhead of stalling is too much | |
11bc82d6 | 1041 | */ |
2916ecc0 JG |
1042 | switch (mode) { |
1043 | case MIGRATE_SYNC: | |
1044 | case MIGRATE_SYNC_NO_COPY: | |
1045 | break; | |
1046 | default: | |
11bc82d6 | 1047 | rc = -EBUSY; |
0a31bc97 | 1048 | goto out_unlock; |
11bc82d6 AA |
1049 | } |
1050 | if (!force) | |
0a31bc97 | 1051 | goto out_unlock; |
e24f0b8f CL |
1052 | wait_on_page_writeback(page); |
1053 | } | |
03f15c86 | 1054 | |
e24f0b8f | 1055 | /* |
dc386d4d KH |
1056 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1057 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 1058 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1059 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1060 | * File Caches may use write_page() or lock_page() in migration, then, |
1061 | * just care Anon page here. | |
03f15c86 HD |
1062 | * |
1063 | * Only page_get_anon_vma() understands the subtleties of | |
1064 | * getting a hold on an anon_vma from outside one of its mms. | |
1065 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1066 | * because that implies that the anon page is no longer mapped | |
1067 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1068 | */ |
03f15c86 | 1069 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1070 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1071 | |
7db7671f HD |
1072 | /* |
1073 | * Block others from accessing the new page when we get around to | |
1074 | * establishing additional references. We are usually the only one | |
1075 | * holding a reference to newpage at this point. We used to have a BUG | |
1076 | * here if trylock_page(newpage) fails, but would like to allow for | |
1077 | * cases where there might be a race with the previous use of newpage. | |
1078 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1079 | */ | |
1080 | if (unlikely(!trylock_page(newpage))) | |
1081 | goto out_unlock; | |
1082 | ||
bda807d4 MK |
1083 | if (unlikely(!is_lru)) { |
1084 | rc = move_to_new_page(newpage, page, mode); | |
1085 | goto out_unlock_both; | |
1086 | } | |
1087 | ||
dc386d4d | 1088 | /* |
62e1c553 SL |
1089 | * Corner case handling: |
1090 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1091 | * and treated as swapcache but it has no rmap yet. | |
1092 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1093 | * trigger a BUG. So handle it here. | |
1094 | * 2. An orphaned page (see truncate_complete_page) might have | |
1095 | * fs-private metadata. The page can be picked up due to memory | |
1096 | * offlining. Everywhere else except page reclaim, the page is | |
1097 | * invisible to the vm, so the page can not be migrated. So try to | |
1098 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1099 | */ |
62e1c553 | 1100 | if (!page->mapping) { |
309381fe | 1101 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1102 | if (page_has_private(page)) { |
62e1c553 | 1103 | try_to_free_buffers(page); |
7db7671f | 1104 | goto out_unlock_both; |
62e1c553 | 1105 | } |
7db7671f HD |
1106 | } else if (page_mapped(page)) { |
1107 | /* Establish migration ptes */ | |
03f15c86 HD |
1108 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1109 | page); | |
2ebba6b7 | 1110 | try_to_unmap(page, |
da1b13cc | 1111 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
1112 | page_was_mapped = 1; |
1113 | } | |
dc386d4d | 1114 | |
e6a1530d | 1115 | if (!page_mapped(page)) |
5c3f9a67 | 1116 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1117 | |
5c3f9a67 HD |
1118 | if (page_was_mapped) |
1119 | remove_migration_ptes(page, | |
e388466d | 1120 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1121 | |
7db7671f HD |
1122 | out_unlock_both: |
1123 | unlock_page(newpage); | |
1124 | out_unlock: | |
3f6c8272 | 1125 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1126 | if (anon_vma) |
9e60109f | 1127 | put_anon_vma(anon_vma); |
e24f0b8f | 1128 | unlock_page(page); |
0dabec93 | 1129 | out: |
c6c919eb MK |
1130 | /* |
1131 | * If migration is successful, decrease refcount of the newpage | |
1132 | * which will not free the page because new page owner increased | |
1133 | * refcounter. As well, if it is LRU page, add the page to LRU | |
e0a352fa DH |
1134 | * list in here. Use the old state of the isolated source page to |
1135 | * determine if we migrated a LRU page. newpage was already unlocked | |
1136 | * and possibly modified by its owner - don't rely on the page | |
1137 | * state. | |
c6c919eb MK |
1138 | */ |
1139 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e0a352fa | 1140 | if (unlikely(!is_lru)) |
c6c919eb MK |
1141 | put_page(newpage); |
1142 | else | |
1143 | putback_lru_page(newpage); | |
1144 | } | |
1145 | ||
0dabec93 MK |
1146 | return rc; |
1147 | } | |
95a402c3 | 1148 | |
ef2a5153 GU |
1149 | /* |
1150 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
1151 | * around it. | |
1152 | */ | |
815f0ddb ND |
1153 | #if defined(CONFIG_ARM) && \ |
1154 | defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700 | |
ef2a5153 GU |
1155 | #define ICE_noinline noinline |
1156 | #else | |
1157 | #define ICE_noinline | |
1158 | #endif | |
1159 | ||
0dabec93 MK |
1160 | /* |
1161 | * Obtain the lock on page, remove all ptes and migrate the page | |
1162 | * to the newly allocated page in newpage. | |
1163 | */ | |
ef2a5153 GU |
1164 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
1165 | free_page_t put_new_page, | |
1166 | unsigned long private, struct page *page, | |
add05cec NH |
1167 | int force, enum migrate_mode mode, |
1168 | enum migrate_reason reason) | |
0dabec93 | 1169 | { |
2def7424 | 1170 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1171 | struct page *newpage = NULL; |
0dabec93 | 1172 | |
94723aaf MH |
1173 | if (!thp_migration_supported() && PageTransHuge(page)) |
1174 | return -ENOMEM; | |
1175 | ||
0dabec93 MK |
1176 | if (page_count(page) == 1) { |
1177 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1178 | ClearPageActive(page); |
1179 | ClearPageUnevictable(page); | |
bda807d4 MK |
1180 | if (unlikely(__PageMovable(page))) { |
1181 | lock_page(page); | |
1182 | if (!PageMovable(page)) | |
1183 | __ClearPageIsolated(page); | |
1184 | unlock_page(page); | |
1185 | } | |
0dabec93 MK |
1186 | goto out; |
1187 | } | |
1188 | ||
74d4a579 YS |
1189 | newpage = get_new_page(page, private); |
1190 | if (!newpage) | |
1191 | return -ENOMEM; | |
1192 | ||
9c620e2b | 1193 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1194 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1195 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1196 | |
0dabec93 | 1197 | out: |
e24f0b8f | 1198 | if (rc != -EAGAIN) { |
0dabec93 MK |
1199 | /* |
1200 | * A page that has been migrated has all references | |
1201 | * removed and will be freed. A page that has not been | |
1202 | * migrated will have kepts its references and be | |
1203 | * restored. | |
1204 | */ | |
1205 | list_del(&page->lru); | |
6afcf8ef ML |
1206 | |
1207 | /* | |
1208 | * Compaction can migrate also non-LRU pages which are | |
1209 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1210 | * as __PageMovable | |
1211 | */ | |
1212 | if (likely(!__PageMovable(page))) | |
e8db67eb NH |
1213 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
1214 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
c6c919eb MK |
1215 | } |
1216 | ||
1217 | /* | |
1218 | * If migration is successful, releases reference grabbed during | |
1219 | * isolation. Otherwise, restore the page to right list unless | |
1220 | * we want to retry. | |
1221 | */ | |
1222 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1223 | put_page(page); | |
1224 | if (reason == MR_MEMORY_FAILURE) { | |
d7e69488 | 1225 | /* |
c6c919eb MK |
1226 | * Set PG_HWPoison on just freed page |
1227 | * intentionally. Although it's rather weird, | |
1228 | * it's how HWPoison flag works at the moment. | |
d7e69488 | 1229 | */ |
d4ae9916 | 1230 | if (set_hwpoison_free_buddy_page(page)) |
da1b13cc | 1231 | num_poisoned_pages_inc(); |
c6c919eb MK |
1232 | } |
1233 | } else { | |
bda807d4 MK |
1234 | if (rc != -EAGAIN) { |
1235 | if (likely(!__PageMovable(page))) { | |
1236 | putback_lru_page(page); | |
1237 | goto put_new; | |
1238 | } | |
1239 | ||
1240 | lock_page(page); | |
1241 | if (PageMovable(page)) | |
1242 | putback_movable_page(page); | |
1243 | else | |
1244 | __ClearPageIsolated(page); | |
1245 | unlock_page(page); | |
1246 | put_page(page); | |
1247 | } | |
1248 | put_new: | |
c6c919eb MK |
1249 | if (put_new_page) |
1250 | put_new_page(newpage, private); | |
1251 | else | |
1252 | put_page(newpage); | |
e24f0b8f | 1253 | } |
68711a74 | 1254 | |
e24f0b8f CL |
1255 | return rc; |
1256 | } | |
1257 | ||
290408d4 NH |
1258 | /* |
1259 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1260 | * | |
1261 | * This function doesn't wait the completion of hugepage I/O | |
1262 | * because there is no race between I/O and migration for hugepage. | |
1263 | * Note that currently hugepage I/O occurs only in direct I/O | |
1264 | * where no lock is held and PG_writeback is irrelevant, | |
1265 | * and writeback status of all subpages are counted in the reference | |
1266 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1267 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1268 | * This means that when we try to migrate hugepage whose subpages are | |
1269 | * doing direct I/O, some references remain after try_to_unmap() and | |
1270 | * hugepage migration fails without data corruption. | |
1271 | * | |
1272 | * There is also no race when direct I/O is issued on the page under migration, | |
1273 | * because then pte is replaced with migration swap entry and direct I/O code | |
1274 | * will wait in the page fault for migration to complete. | |
1275 | */ | |
1276 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1277 | free_page_t put_new_page, unsigned long private, |
1278 | struct page *hpage, int force, | |
7cd12b4a | 1279 | enum migrate_mode mode, int reason) |
290408d4 | 1280 | { |
2def7424 | 1281 | int rc = -EAGAIN; |
2ebba6b7 | 1282 | int page_was_mapped = 0; |
32665f2b | 1283 | struct page *new_hpage; |
290408d4 NH |
1284 | struct anon_vma *anon_vma = NULL; |
1285 | ||
83467efb | 1286 | /* |
7ed2c31d | 1287 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1288 | * This check is necessary because some callers of hugepage migration |
1289 | * like soft offline and memory hotremove don't walk through page | |
1290 | * tables or check whether the hugepage is pmd-based or not before | |
1291 | * kicking migration. | |
1292 | */ | |
100873d7 | 1293 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1294 | putback_active_hugepage(hpage); |
83467efb | 1295 | return -ENOSYS; |
32665f2b | 1296 | } |
83467efb | 1297 | |
666feb21 | 1298 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1299 | if (!new_hpage) |
1300 | return -ENOMEM; | |
1301 | ||
290408d4 | 1302 | if (!trylock_page(hpage)) { |
2916ecc0 | 1303 | if (!force) |
290408d4 | 1304 | goto out; |
2916ecc0 JG |
1305 | switch (mode) { |
1306 | case MIGRATE_SYNC: | |
1307 | case MIGRATE_SYNC_NO_COPY: | |
1308 | break; | |
1309 | default: | |
1310 | goto out; | |
1311 | } | |
290408d4 NH |
1312 | lock_page(hpage); |
1313 | } | |
1314 | ||
cb6acd01 MK |
1315 | /* |
1316 | * Check for pages which are in the process of being freed. Without | |
1317 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1318 | * be called and we could leak usage counts for subpools. | |
1319 | */ | |
1320 | if (page_private(hpage) && !page_mapping(hpage)) { | |
1321 | rc = -EBUSY; | |
1322 | goto out_unlock; | |
1323 | } | |
1324 | ||
746b18d4 PZ |
1325 | if (PageAnon(hpage)) |
1326 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1327 | |
7db7671f HD |
1328 | if (unlikely(!trylock_page(new_hpage))) |
1329 | goto put_anon; | |
1330 | ||
2ebba6b7 HD |
1331 | if (page_mapped(hpage)) { |
1332 | try_to_unmap(hpage, | |
ddeaab32 | 1333 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
1334 | page_was_mapped = 1; |
1335 | } | |
290408d4 NH |
1336 | |
1337 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1338 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1339 | |
5c3f9a67 HD |
1340 | if (page_was_mapped) |
1341 | remove_migration_ptes(hpage, | |
e388466d | 1342 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1343 | |
7db7671f HD |
1344 | unlock_page(new_hpage); |
1345 | ||
1346 | put_anon: | |
fd4a4663 | 1347 | if (anon_vma) |
9e60109f | 1348 | put_anon_vma(anon_vma); |
8e6ac7fa | 1349 | |
2def7424 | 1350 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1351 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1352 | put_new_page = NULL; |
1353 | } | |
8e6ac7fa | 1354 | |
cb6acd01 | 1355 | out_unlock: |
290408d4 | 1356 | unlock_page(hpage); |
09761333 | 1357 | out: |
b8ec1cee NH |
1358 | if (rc != -EAGAIN) |
1359 | putback_active_hugepage(hpage); | |
68711a74 DR |
1360 | |
1361 | /* | |
1362 | * If migration was not successful and there's a freeing callback, use | |
1363 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1364 | * isolation. | |
1365 | */ | |
2def7424 | 1366 | if (put_new_page) |
68711a74 DR |
1367 | put_new_page(new_hpage, private); |
1368 | else | |
3aaa76e1 | 1369 | putback_active_hugepage(new_hpage); |
68711a74 | 1370 | |
290408d4 NH |
1371 | return rc; |
1372 | } | |
1373 | ||
b20a3503 | 1374 | /* |
c73e5c9c SB |
1375 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1376 | * supplied as the target for the page migration | |
b20a3503 | 1377 | * |
c73e5c9c SB |
1378 | * @from: The list of pages to be migrated. |
1379 | * @get_new_page: The function used to allocate free pages to be used | |
1380 | * as the target of the page migration. | |
68711a74 DR |
1381 | * @put_new_page: The function used to free target pages if migration |
1382 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1383 | * @private: Private data to be passed on to get_new_page() |
1384 | * @mode: The migration mode that specifies the constraints for | |
1385 | * page migration, if any. | |
1386 | * @reason: The reason for page migration. | |
b20a3503 | 1387 | * |
c73e5c9c SB |
1388 | * The function returns after 10 attempts or if no pages are movable any more |
1389 | * because the list has become empty or no retryable pages exist any more. | |
14e0f9bc | 1390 | * The caller should call putback_movable_pages() to return pages to the LRU |
28bd6578 | 1391 | * or free list only if ret != 0. |
b20a3503 | 1392 | * |
c73e5c9c | 1393 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1394 | */ |
9c620e2b | 1395 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1396 | free_page_t put_new_page, unsigned long private, |
1397 | enum migrate_mode mode, int reason) | |
b20a3503 | 1398 | { |
e24f0b8f | 1399 | int retry = 1; |
b20a3503 | 1400 | int nr_failed = 0; |
5647bc29 | 1401 | int nr_succeeded = 0; |
b20a3503 CL |
1402 | int pass = 0; |
1403 | struct page *page; | |
1404 | struct page *page2; | |
1405 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1406 | int rc; | |
1407 | ||
1408 | if (!swapwrite) | |
1409 | current->flags |= PF_SWAPWRITE; | |
1410 | ||
e24f0b8f CL |
1411 | for(pass = 0; pass < 10 && retry; pass++) { |
1412 | retry = 0; | |
b20a3503 | 1413 | |
e24f0b8f | 1414 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1415 | retry: |
e24f0b8f | 1416 | cond_resched(); |
2d1db3b1 | 1417 | |
31caf665 NH |
1418 | if (PageHuge(page)) |
1419 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1420 | put_new_page, private, page, |
7cd12b4a | 1421 | pass > 2, mode, reason); |
31caf665 | 1422 | else |
68711a74 | 1423 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1424 | private, page, pass > 2, mode, |
1425 | reason); | |
2d1db3b1 | 1426 | |
e24f0b8f | 1427 | switch(rc) { |
95a402c3 | 1428 | case -ENOMEM: |
94723aaf MH |
1429 | /* |
1430 | * THP migration might be unsupported or the | |
1431 | * allocation could've failed so we should | |
1432 | * retry on the same page with the THP split | |
1433 | * to base pages. | |
1434 | * | |
1435 | * Head page is retried immediately and tail | |
1436 | * pages are added to the tail of the list so | |
1437 | * we encounter them after the rest of the list | |
1438 | * is processed. | |
1439 | */ | |
e6112fc3 | 1440 | if (PageTransHuge(page) && !PageHuge(page)) { |
94723aaf MH |
1441 | lock_page(page); |
1442 | rc = split_huge_page_to_list(page, from); | |
1443 | unlock_page(page); | |
1444 | if (!rc) { | |
1445 | list_safe_reset_next(page, page2, lru); | |
1446 | goto retry; | |
1447 | } | |
1448 | } | |
dfef2ef4 | 1449 | nr_failed++; |
95a402c3 | 1450 | goto out; |
e24f0b8f | 1451 | case -EAGAIN: |
2d1db3b1 | 1452 | retry++; |
e24f0b8f | 1453 | break; |
78bd5209 | 1454 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1455 | nr_succeeded++; |
e24f0b8f CL |
1456 | break; |
1457 | default: | |
354a3363 NH |
1458 | /* |
1459 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1460 | * unlike -EAGAIN case, the failed page is | |
1461 | * removed from migration page list and not | |
1462 | * retried in the next outer loop. | |
1463 | */ | |
2d1db3b1 | 1464 | nr_failed++; |
e24f0b8f | 1465 | break; |
2d1db3b1 | 1466 | } |
b20a3503 CL |
1467 | } |
1468 | } | |
f2f81fb2 VB |
1469 | nr_failed += retry; |
1470 | rc = nr_failed; | |
95a402c3 | 1471 | out: |
5647bc29 MG |
1472 | if (nr_succeeded) |
1473 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1474 | if (nr_failed) | |
1475 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1476 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1477 | ||
b20a3503 CL |
1478 | if (!swapwrite) |
1479 | current->flags &= ~PF_SWAPWRITE; | |
1480 | ||
78bd5209 | 1481 | return rc; |
b20a3503 | 1482 | } |
95a402c3 | 1483 | |
742755a1 | 1484 | #ifdef CONFIG_NUMA |
742755a1 | 1485 | |
a49bd4d7 | 1486 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1487 | { |
a49bd4d7 MH |
1488 | while (nr-- > 0) { |
1489 | if (put_user(value, status + start)) | |
1490 | return -EFAULT; | |
1491 | start++; | |
1492 | } | |
1493 | ||
1494 | return 0; | |
1495 | } | |
1496 | ||
1497 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1498 | struct list_head *pagelist, int node) | |
1499 | { | |
1500 | int err; | |
1501 | ||
1502 | if (list_empty(pagelist)) | |
1503 | return 0; | |
1504 | ||
1505 | err = migrate_pages(pagelist, alloc_new_node_page, NULL, node, | |
1506 | MIGRATE_SYNC, MR_SYSCALL); | |
1507 | if (err) | |
1508 | putback_movable_pages(pagelist); | |
1509 | return err; | |
742755a1 CL |
1510 | } |
1511 | ||
1512 | /* | |
a49bd4d7 MH |
1513 | * Resolves the given address to a struct page, isolates it from the LRU and |
1514 | * puts it to the given pagelist. | |
e0153fc2 YS |
1515 | * Returns: |
1516 | * errno - if the page cannot be found/isolated | |
1517 | * 0 - when it doesn't have to be migrated because it is already on the | |
1518 | * target node | |
1519 | * 1 - when it has been queued | |
742755a1 | 1520 | */ |
a49bd4d7 MH |
1521 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1522 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1523 | { |
a49bd4d7 MH |
1524 | struct vm_area_struct *vma; |
1525 | struct page *page; | |
1526 | unsigned int follflags; | |
742755a1 | 1527 | int err; |
742755a1 CL |
1528 | |
1529 | down_read(&mm->mmap_sem); | |
a49bd4d7 MH |
1530 | err = -EFAULT; |
1531 | vma = find_vma(mm, addr); | |
1532 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1533 | goto out; | |
742755a1 | 1534 | |
a49bd4d7 MH |
1535 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1536 | follflags = FOLL_GET | FOLL_DUMP; | |
a49bd4d7 | 1537 | page = follow_page(vma, addr, follflags); |
89f5b7da | 1538 | |
a49bd4d7 MH |
1539 | err = PTR_ERR(page); |
1540 | if (IS_ERR(page)) | |
1541 | goto out; | |
89f5b7da | 1542 | |
a49bd4d7 MH |
1543 | err = -ENOENT; |
1544 | if (!page) | |
1545 | goto out; | |
742755a1 | 1546 | |
a49bd4d7 MH |
1547 | err = 0; |
1548 | if (page_to_nid(page) == node) | |
1549 | goto out_putpage; | |
742755a1 | 1550 | |
a49bd4d7 MH |
1551 | err = -EACCES; |
1552 | if (page_mapcount(page) > 1 && !migrate_all) | |
1553 | goto out_putpage; | |
742755a1 | 1554 | |
a49bd4d7 MH |
1555 | if (PageHuge(page)) { |
1556 | if (PageHead(page)) { | |
1557 | isolate_huge_page(page, pagelist); | |
e0153fc2 | 1558 | err = 1; |
e632a938 | 1559 | } |
a49bd4d7 MH |
1560 | } else { |
1561 | struct page *head; | |
e632a938 | 1562 | |
e8db67eb NH |
1563 | head = compound_head(page); |
1564 | err = isolate_lru_page(head); | |
cf608ac1 | 1565 | if (err) |
a49bd4d7 | 1566 | goto out_putpage; |
742755a1 | 1567 | |
e0153fc2 | 1568 | err = 1; |
a49bd4d7 MH |
1569 | list_add_tail(&head->lru, pagelist); |
1570 | mod_node_page_state(page_pgdat(head), | |
1571 | NR_ISOLATED_ANON + page_is_file_cache(head), | |
1572 | hpage_nr_pages(head)); | |
1573 | } | |
1574 | out_putpage: | |
1575 | /* | |
1576 | * Either remove the duplicate refcount from | |
1577 | * isolate_lru_page() or drop the page ref if it was | |
1578 | * not isolated. | |
1579 | */ | |
1580 | put_page(page); | |
1581 | out: | |
742755a1 CL |
1582 | up_read(&mm->mmap_sem); |
1583 | return err; | |
1584 | } | |
1585 | ||
5e9a0f02 BG |
1586 | /* |
1587 | * Migrate an array of page address onto an array of nodes and fill | |
1588 | * the corresponding array of status. | |
1589 | */ | |
3268c63e | 1590 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1591 | unsigned long nr_pages, |
1592 | const void __user * __user *pages, | |
1593 | const int __user *nodes, | |
1594 | int __user *status, int flags) | |
1595 | { | |
a49bd4d7 MH |
1596 | int current_node = NUMA_NO_NODE; |
1597 | LIST_HEAD(pagelist); | |
1598 | int start, i; | |
1599 | int err = 0, err1; | |
35282a2d BG |
1600 | |
1601 | migrate_prep(); | |
1602 | ||
a49bd4d7 MH |
1603 | for (i = start = 0; i < nr_pages; i++) { |
1604 | const void __user *p; | |
1605 | unsigned long addr; | |
1606 | int node; | |
3140a227 | 1607 | |
a49bd4d7 MH |
1608 | err = -EFAULT; |
1609 | if (get_user(p, pages + i)) | |
1610 | goto out_flush; | |
1611 | if (get_user(node, nodes + i)) | |
1612 | goto out_flush; | |
057d3389 | 1613 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1614 | |
1615 | err = -ENODEV; | |
1616 | if (node < 0 || node >= MAX_NUMNODES) | |
1617 | goto out_flush; | |
1618 | if (!node_state(node, N_MEMORY)) | |
1619 | goto out_flush; | |
5e9a0f02 | 1620 | |
a49bd4d7 MH |
1621 | err = -EACCES; |
1622 | if (!node_isset(node, task_nodes)) | |
1623 | goto out_flush; | |
1624 | ||
1625 | if (current_node == NUMA_NO_NODE) { | |
1626 | current_node = node; | |
1627 | start = i; | |
1628 | } else if (node != current_node) { | |
1629 | err = do_move_pages_to_node(mm, &pagelist, current_node); | |
5984fabb YS |
1630 | if (err) { |
1631 | /* | |
1632 | * Positive err means the number of failed | |
1633 | * pages to migrate. Since we are going to | |
1634 | * abort and return the number of non-migrated | |
1635 | * pages, so need to incude the rest of the | |
1636 | * nr_pages that have not been attempted as | |
1637 | * well. | |
1638 | */ | |
1639 | if (err > 0) | |
1640 | err += nr_pages - i - 1; | |
a49bd4d7 | 1641 | goto out; |
5984fabb | 1642 | } |
a49bd4d7 MH |
1643 | err = store_status(status, start, current_node, i - start); |
1644 | if (err) | |
1645 | goto out; | |
1646 | start = i; | |
1647 | current_node = node; | |
3140a227 BG |
1648 | } |
1649 | ||
a49bd4d7 MH |
1650 | /* |
1651 | * Errors in the page lookup or isolation are not fatal and we simply | |
1652 | * report them via status | |
1653 | */ | |
1654 | err = add_page_for_migration(mm, addr, current_node, | |
1655 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 YS |
1656 | |
1657 | if (!err) { | |
1658 | /* The page is already on the target node */ | |
1659 | err = store_status(status, i, current_node, 1); | |
1660 | if (err) | |
1661 | goto out_flush; | |
a49bd4d7 | 1662 | continue; |
e0153fc2 YS |
1663 | } else if (err > 0) { |
1664 | /* The page is successfully queued for migration */ | |
1665 | continue; | |
1666 | } | |
3140a227 | 1667 | |
a49bd4d7 MH |
1668 | err = store_status(status, i, err, 1); |
1669 | if (err) | |
1670 | goto out_flush; | |
5e9a0f02 | 1671 | |
a49bd4d7 | 1672 | err = do_move_pages_to_node(mm, &pagelist, current_node); |
5984fabb YS |
1673 | if (err) { |
1674 | if (err > 0) | |
1675 | err += nr_pages - i - 1; | |
a49bd4d7 | 1676 | goto out; |
5984fabb | 1677 | } |
a49bd4d7 MH |
1678 | if (i > start) { |
1679 | err = store_status(status, start, current_node, i - start); | |
1680 | if (err) | |
1681 | goto out; | |
1682 | } | |
1683 | current_node = NUMA_NO_NODE; | |
3140a227 | 1684 | } |
a49bd4d7 | 1685 | out_flush: |
8f175cf5 MH |
1686 | if (list_empty(&pagelist)) |
1687 | return err; | |
1688 | ||
a49bd4d7 MH |
1689 | /* Make sure we do not overwrite the existing error */ |
1690 | err1 = do_move_pages_to_node(mm, &pagelist, current_node); | |
5984fabb YS |
1691 | /* |
1692 | * Don't have to report non-attempted pages here since: | |
1693 | * - If the above loop is done gracefully all pages have been | |
1694 | * attempted. | |
1695 | * - If the above loop is aborted it means a fatal error | |
1696 | * happened, should return ret. | |
1697 | */ | |
a49bd4d7 MH |
1698 | if (!err1) |
1699 | err1 = store_status(status, start, current_node, i - start); | |
dfe9aa23 | 1700 | if (err >= 0) |
a49bd4d7 | 1701 | err = err1; |
5e9a0f02 BG |
1702 | out: |
1703 | return err; | |
1704 | } | |
1705 | ||
742755a1 | 1706 | /* |
2f007e74 | 1707 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1708 | */ |
80bba129 BG |
1709 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1710 | const void __user **pages, int *status) | |
742755a1 | 1711 | { |
2f007e74 | 1712 | unsigned long i; |
2f007e74 | 1713 | |
742755a1 CL |
1714 | down_read(&mm->mmap_sem); |
1715 | ||
2f007e74 | 1716 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1717 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1718 | struct vm_area_struct *vma; |
1719 | struct page *page; | |
c095adbc | 1720 | int err = -EFAULT; |
2f007e74 BG |
1721 | |
1722 | vma = find_vma(mm, addr); | |
70384dc6 | 1723 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1724 | goto set_status; |
1725 | ||
d899844e KS |
1726 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1727 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1728 | |
1729 | err = PTR_ERR(page); | |
1730 | if (IS_ERR(page)) | |
1731 | goto set_status; | |
1732 | ||
d899844e | 1733 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1734 | set_status: |
80bba129 BG |
1735 | *status = err; |
1736 | ||
1737 | pages++; | |
1738 | status++; | |
1739 | } | |
1740 | ||
1741 | up_read(&mm->mmap_sem); | |
1742 | } | |
1743 | ||
1744 | /* | |
1745 | * Determine the nodes of a user array of pages and store it in | |
1746 | * a user array of status. | |
1747 | */ | |
1748 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1749 | const void __user * __user *pages, | |
1750 | int __user *status) | |
1751 | { | |
1752 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1753 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1754 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1755 | |
87b8d1ad PA |
1756 | while (nr_pages) { |
1757 | unsigned long chunk_nr; | |
80bba129 | 1758 | |
87b8d1ad PA |
1759 | chunk_nr = nr_pages; |
1760 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1761 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1762 | ||
1763 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1764 | break; | |
80bba129 BG |
1765 | |
1766 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1767 | ||
87b8d1ad PA |
1768 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1769 | break; | |
742755a1 | 1770 | |
87b8d1ad PA |
1771 | pages += chunk_nr; |
1772 | status += chunk_nr; | |
1773 | nr_pages -= chunk_nr; | |
1774 | } | |
1775 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1776 | } |
1777 | ||
1778 | /* | |
1779 | * Move a list of pages in the address space of the currently executing | |
1780 | * process. | |
1781 | */ | |
7addf443 DB |
1782 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, |
1783 | const void __user * __user *pages, | |
1784 | const int __user *nodes, | |
1785 | int __user *status, int flags) | |
742755a1 | 1786 | { |
742755a1 | 1787 | struct task_struct *task; |
742755a1 | 1788 | struct mm_struct *mm; |
5e9a0f02 | 1789 | int err; |
3268c63e | 1790 | nodemask_t task_nodes; |
742755a1 CL |
1791 | |
1792 | /* Check flags */ | |
1793 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1794 | return -EINVAL; | |
1795 | ||
1796 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1797 | return -EPERM; | |
1798 | ||
1799 | /* Find the mm_struct */ | |
a879bf58 | 1800 | rcu_read_lock(); |
228ebcbe | 1801 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1802 | if (!task) { |
a879bf58 | 1803 | rcu_read_unlock(); |
742755a1 CL |
1804 | return -ESRCH; |
1805 | } | |
3268c63e | 1806 | get_task_struct(task); |
742755a1 CL |
1807 | |
1808 | /* | |
1809 | * Check if this process has the right to modify the specified | |
197e7e52 | 1810 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1811 | */ |
197e7e52 | 1812 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1813 | rcu_read_unlock(); |
742755a1 | 1814 | err = -EPERM; |
5e9a0f02 | 1815 | goto out; |
742755a1 | 1816 | } |
c69e8d9c | 1817 | rcu_read_unlock(); |
742755a1 | 1818 | |
86c3a764 DQ |
1819 | err = security_task_movememory(task); |
1820 | if (err) | |
5e9a0f02 | 1821 | goto out; |
86c3a764 | 1822 | |
3268c63e CL |
1823 | task_nodes = cpuset_mems_allowed(task); |
1824 | mm = get_task_mm(task); | |
1825 | put_task_struct(task); | |
1826 | ||
6e8b09ea SL |
1827 | if (!mm) |
1828 | return -EINVAL; | |
1829 | ||
1830 | if (nodes) | |
1831 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1832 | nodes, status, flags); | |
1833 | else | |
1834 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1835 | |
742755a1 CL |
1836 | mmput(mm); |
1837 | return err; | |
3268c63e CL |
1838 | |
1839 | out: | |
1840 | put_task_struct(task); | |
1841 | return err; | |
742755a1 | 1842 | } |
742755a1 | 1843 | |
7addf443 DB |
1844 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1845 | const void __user * __user *, pages, | |
1846 | const int __user *, nodes, | |
1847 | int __user *, status, int, flags) | |
1848 | { | |
1849 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1850 | } | |
1851 | ||
1852 | #ifdef CONFIG_COMPAT | |
1853 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, | |
1854 | compat_uptr_t __user *, pages32, | |
1855 | const int __user *, nodes, | |
1856 | int __user *, status, | |
1857 | int, flags) | |
1858 | { | |
1859 | const void __user * __user *pages; | |
1860 | int i; | |
1861 | ||
1862 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); | |
1863 | for (i = 0; i < nr_pages; i++) { | |
1864 | compat_uptr_t p; | |
1865 | ||
1866 | if (get_user(p, pages32 + i) || | |
1867 | put_user(compat_ptr(p), pages + i)) | |
1868 | return -EFAULT; | |
1869 | } | |
1870 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1871 | } | |
1872 | #endif /* CONFIG_COMPAT */ | |
1873 | ||
7039e1db PZ |
1874 | #ifdef CONFIG_NUMA_BALANCING |
1875 | /* | |
1876 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1877 | * pages. Currently it only checks the watermarks which crude | |
1878 | */ | |
1879 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1880 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1881 | { |
1882 | int z; | |
599d0c95 | 1883 | |
7039e1db PZ |
1884 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1885 | struct zone *zone = pgdat->node_zones + z; | |
1886 | ||
1887 | if (!populated_zone(zone)) | |
1888 | continue; | |
1889 | ||
7039e1db PZ |
1890 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1891 | if (!zone_watermark_ok(zone, 0, | |
1892 | high_wmark_pages(zone) + | |
1893 | nr_migrate_pages, | |
bfe9d006 | 1894 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
1895 | continue; |
1896 | return true; | |
1897 | } | |
1898 | return false; | |
1899 | } | |
1900 | ||
1901 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 1902 | unsigned long data) |
7039e1db PZ |
1903 | { |
1904 | int nid = (int) data; | |
1905 | struct page *newpage; | |
1906 | ||
96db800f | 1907 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1908 | (GFP_HIGHUSER_MOVABLE | |
1909 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1910 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 1911 | ~__GFP_RECLAIM, 0); |
bac0382c | 1912 | |
7039e1db PZ |
1913 | return newpage; |
1914 | } | |
1915 | ||
1c30e017 | 1916 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1917 | { |
340ef390 | 1918 | int page_lru; |
a8f60772 | 1919 | |
309381fe | 1920 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1921 | |
7039e1db | 1922 | /* Avoid migrating to a node that is nearly full */ |
d8c6546b | 1923 | if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) |
340ef390 | 1924 | return 0; |
7039e1db | 1925 | |
340ef390 HD |
1926 | if (isolate_lru_page(page)) |
1927 | return 0; | |
7039e1db | 1928 | |
340ef390 HD |
1929 | /* |
1930 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1931 | * check on page_count(), so we must do it here, now that the page | |
1932 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1933 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1934 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1935 | */ | |
1936 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1937 | putback_lru_page(page); | |
1938 | return 0; | |
7039e1db PZ |
1939 | } |
1940 | ||
340ef390 | 1941 | page_lru = page_is_file_cache(page); |
599d0c95 | 1942 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
340ef390 HD |
1943 | hpage_nr_pages(page)); |
1944 | ||
149c33e1 | 1945 | /* |
340ef390 HD |
1946 | * Isolating the page has taken another reference, so the |
1947 | * caller's reference can be safely dropped without the page | |
1948 | * disappearing underneath us during migration. | |
149c33e1 MG |
1949 | */ |
1950 | put_page(page); | |
340ef390 | 1951 | return 1; |
b32967ff MG |
1952 | } |
1953 | ||
de466bd6 MG |
1954 | bool pmd_trans_migrating(pmd_t pmd) |
1955 | { | |
1956 | struct page *page = pmd_page(pmd); | |
1957 | return PageLocked(page); | |
1958 | } | |
1959 | ||
b32967ff MG |
1960 | /* |
1961 | * Attempt to migrate a misplaced page to the specified destination | |
1962 | * node. Caller is expected to have an elevated reference count on | |
1963 | * the page that will be dropped by this function before returning. | |
1964 | */ | |
1bc115d8 MG |
1965 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1966 | int node) | |
b32967ff MG |
1967 | { |
1968 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1969 | int isolated; |
b32967ff MG |
1970 | int nr_remaining; |
1971 | LIST_HEAD(migratepages); | |
1972 | ||
1973 | /* | |
1bc115d8 MG |
1974 | * Don't migrate file pages that are mapped in multiple processes |
1975 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1976 | */ |
1bc115d8 MG |
1977 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1978 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1979 | goto out; |
b32967ff | 1980 | |
09a913a7 MG |
1981 | /* |
1982 | * Also do not migrate dirty pages as not all filesystems can move | |
1983 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
1984 | */ | |
1985 | if (page_is_file_cache(page) && PageDirty(page)) | |
1986 | goto out; | |
1987 | ||
b32967ff MG |
1988 | isolated = numamigrate_isolate_page(pgdat, page); |
1989 | if (!isolated) | |
1990 | goto out; | |
1991 | ||
1992 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1993 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1994 | NULL, node, MIGRATE_ASYNC, |
1995 | MR_NUMA_MISPLACED); | |
b32967ff | 1996 | if (nr_remaining) { |
59c82b70 JK |
1997 | if (!list_empty(&migratepages)) { |
1998 | list_del(&page->lru); | |
599d0c95 | 1999 | dec_node_page_state(page, NR_ISOLATED_ANON + |
59c82b70 JK |
2000 | page_is_file_cache(page)); |
2001 | putback_lru_page(page); | |
2002 | } | |
b32967ff MG |
2003 | isolated = 0; |
2004 | } else | |
2005 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 2006 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2007 | return isolated; |
340ef390 HD |
2008 | |
2009 | out: | |
2010 | put_page(page); | |
2011 | return 0; | |
7039e1db | 2012 | } |
220018d3 | 2013 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 2014 | |
220018d3 | 2015 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
2016 | /* |
2017 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
2018 | * before returning. | |
2019 | */ | |
b32967ff MG |
2020 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
2021 | struct vm_area_struct *vma, | |
2022 | pmd_t *pmd, pmd_t entry, | |
2023 | unsigned long address, | |
2024 | struct page *page, int node) | |
2025 | { | |
c4088ebd | 2026 | spinlock_t *ptl; |
b32967ff MG |
2027 | pg_data_t *pgdat = NODE_DATA(node); |
2028 | int isolated = 0; | |
2029 | struct page *new_page = NULL; | |
b32967ff | 2030 | int page_lru = page_is_file_cache(page); |
7066f0f9 | 2031 | unsigned long start = address & HPAGE_PMD_MASK; |
b32967ff | 2032 | |
b32967ff | 2033 | new_page = alloc_pages_node(node, |
25160354 | 2034 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
e97ca8e5 | 2035 | HPAGE_PMD_ORDER); |
340ef390 HD |
2036 | if (!new_page) |
2037 | goto out_fail; | |
9a982250 | 2038 | prep_transhuge_page(new_page); |
340ef390 | 2039 | |
b32967ff | 2040 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 2041 | if (!isolated) { |
b32967ff | 2042 | put_page(new_page); |
340ef390 | 2043 | goto out_fail; |
b32967ff | 2044 | } |
b0943d61 | 2045 | |
b32967ff | 2046 | /* Prepare a page as a migration target */ |
48c935ad | 2047 | __SetPageLocked(new_page); |
d44d363f SL |
2048 | if (PageSwapBacked(page)) |
2049 | __SetPageSwapBacked(new_page); | |
b32967ff MG |
2050 | |
2051 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2052 | new_page->mapping = page->mapping; | |
2053 | new_page->index = page->index; | |
7eef5f97 AA |
2054 | /* flush the cache before copying using the kernel virtual address */ |
2055 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); | |
b32967ff MG |
2056 | migrate_page_copy(new_page, page); |
2057 | WARN_ON(PageLRU(new_page)); | |
2058 | ||
2059 | /* Recheck the target PMD */ | |
c4088ebd | 2060 | ptl = pmd_lock(mm, pmd); |
f4e177d1 | 2061 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
c4088ebd | 2062 | spin_unlock(ptl); |
b32967ff MG |
2063 | |
2064 | /* Reverse changes made by migrate_page_copy() */ | |
2065 | if (TestClearPageActive(new_page)) | |
2066 | SetPageActive(page); | |
2067 | if (TestClearPageUnevictable(new_page)) | |
2068 | SetPageUnevictable(page); | |
b32967ff MG |
2069 | |
2070 | unlock_page(new_page); | |
2071 | put_page(new_page); /* Free it */ | |
2072 | ||
a54a407f MG |
2073 | /* Retake the callers reference and putback on LRU */ |
2074 | get_page(page); | |
b32967ff | 2075 | putback_lru_page(page); |
599d0c95 | 2076 | mod_node_page_state(page_pgdat(page), |
a54a407f | 2077 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
eb4489f6 MG |
2078 | |
2079 | goto out_unlock; | |
b32967ff MG |
2080 | } |
2081 | ||
10102459 | 2082 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
f55e1014 | 2083 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 2084 | |
2b4847e7 | 2085 | /* |
d7c33934 AA |
2086 | * Overwrite the old entry under pagetable lock and establish |
2087 | * the new PTE. Any parallel GUP will either observe the old | |
2088 | * page blocking on the page lock, block on the page table | |
2089 | * lock or observe the new page. The SetPageUptodate on the | |
2090 | * new page and page_add_new_anon_rmap guarantee the copy is | |
2091 | * visible before the pagetable update. | |
2b4847e7 | 2092 | */ |
7066f0f9 | 2093 | page_add_anon_rmap(new_page, vma, start, true); |
d7c33934 AA |
2094 | /* |
2095 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB | |
2096 | * has already been flushed globally. So no TLB can be currently | |
2097 | * caching this non present pmd mapping. There's no need to clear the | |
2098 | * pmd before doing set_pmd_at(), nor to flush the TLB after | |
2099 | * set_pmd_at(). Clearing the pmd here would introduce a race | |
2100 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the | |
2101 | * mmap_sem for reading. If the pmd is set to NULL at any given time, | |
2102 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this | |
2103 | * pmd. | |
2104 | */ | |
7066f0f9 | 2105 | set_pmd_at(mm, start, pmd, entry); |
ce4a9cc5 | 2106 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 | 2107 | |
f4e177d1 | 2108 | page_ref_unfreeze(page, 2); |
51afb12b | 2109 | mlock_migrate_page(new_page, page); |
d281ee61 | 2110 | page_remove_rmap(page, true); |
7cd12b4a | 2111 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2b4847e7 | 2112 | |
c4088ebd | 2113 | spin_unlock(ptl); |
b32967ff | 2114 | |
11de9927 MG |
2115 | /* Take an "isolate" reference and put new page on the LRU. */ |
2116 | get_page(new_page); | |
2117 | putback_lru_page(new_page); | |
2118 | ||
b32967ff MG |
2119 | unlock_page(new_page); |
2120 | unlock_page(page); | |
2121 | put_page(page); /* Drop the rmap reference */ | |
2122 | put_page(page); /* Drop the LRU isolation reference */ | |
2123 | ||
2124 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2125 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2126 | ||
599d0c95 | 2127 | mod_node_page_state(page_pgdat(page), |
b32967ff MG |
2128 | NR_ISOLATED_ANON + page_lru, |
2129 | -HPAGE_PMD_NR); | |
2130 | return isolated; | |
2131 | ||
340ef390 HD |
2132 | out_fail: |
2133 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2b4847e7 MG |
2134 | ptl = pmd_lock(mm, pmd); |
2135 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 2136 | entry = pmd_modify(entry, vma->vm_page_prot); |
7066f0f9 | 2137 | set_pmd_at(mm, start, pmd, entry); |
2b4847e7 MG |
2138 | update_mmu_cache_pmd(vma, address, &entry); |
2139 | } | |
2140 | spin_unlock(ptl); | |
a54a407f | 2141 | |
eb4489f6 | 2142 | out_unlock: |
340ef390 | 2143 | unlock_page(page); |
b32967ff | 2144 | put_page(page); |
b32967ff MG |
2145 | return 0; |
2146 | } | |
7039e1db PZ |
2147 | #endif /* CONFIG_NUMA_BALANCING */ |
2148 | ||
2149 | #endif /* CONFIG_NUMA */ | |
8763cb45 | 2150 | |
9b2ed9cb | 2151 | #ifdef CONFIG_DEVICE_PRIVATE |
8763cb45 JG |
2152 | static int migrate_vma_collect_hole(unsigned long start, |
2153 | unsigned long end, | |
2154 | struct mm_walk *walk) | |
2155 | { | |
2156 | struct migrate_vma *migrate = walk->private; | |
2157 | unsigned long addr; | |
2158 | ||
872ea707 | 2159 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
e20d103b | 2160 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8315ada7 | 2161 | migrate->dst[migrate->npages] = 0; |
e20d103b | 2162 | migrate->npages++; |
8315ada7 JG |
2163 | migrate->cpages++; |
2164 | } | |
2165 | ||
2166 | return 0; | |
2167 | } | |
2168 | ||
2169 | static int migrate_vma_collect_skip(unsigned long start, | |
2170 | unsigned long end, | |
2171 | struct mm_walk *walk) | |
2172 | { | |
2173 | struct migrate_vma *migrate = walk->private; | |
2174 | unsigned long addr; | |
2175 | ||
872ea707 | 2176 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
8763cb45 JG |
2177 | migrate->dst[migrate->npages] = 0; |
2178 | migrate->src[migrate->npages++] = 0; | |
2179 | } | |
2180 | ||
2181 | return 0; | |
2182 | } | |
2183 | ||
2184 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2185 | unsigned long start, | |
2186 | unsigned long end, | |
2187 | struct mm_walk *walk) | |
2188 | { | |
2189 | struct migrate_vma *migrate = walk->private; | |
2190 | struct vm_area_struct *vma = walk->vma; | |
2191 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2192 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2193 | spinlock_t *ptl; |
2194 | pte_t *ptep; | |
2195 | ||
2196 | again: | |
2197 | if (pmd_none(*pmdp)) | |
2198 | return migrate_vma_collect_hole(start, end, walk); | |
2199 | ||
2200 | if (pmd_trans_huge(*pmdp)) { | |
2201 | struct page *page; | |
2202 | ||
2203 | ptl = pmd_lock(mm, pmdp); | |
2204 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2205 | spin_unlock(ptl); | |
2206 | goto again; | |
2207 | } | |
2208 | ||
2209 | page = pmd_page(*pmdp); | |
2210 | if (is_huge_zero_page(page)) { | |
2211 | spin_unlock(ptl); | |
2212 | split_huge_pmd(vma, pmdp, addr); | |
2213 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2214 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2215 | walk); |
2216 | } else { | |
2217 | int ret; | |
2218 | ||
2219 | get_page(page); | |
2220 | spin_unlock(ptl); | |
2221 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2222 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2223 | walk); |
2224 | ret = split_huge_page(page); | |
2225 | unlock_page(page); | |
2226 | put_page(page); | |
8315ada7 JG |
2227 | if (ret) |
2228 | return migrate_vma_collect_skip(start, end, | |
2229 | walk); | |
2230 | if (pmd_none(*pmdp)) | |
8763cb45 JG |
2231 | return migrate_vma_collect_hole(start, end, |
2232 | walk); | |
2233 | } | |
2234 | } | |
2235 | ||
2236 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2237 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2238 | |
2239 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2240 | arch_enter_lazy_mmu_mode(); |
2241 | ||
8763cb45 JG |
2242 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
2243 | unsigned long mpfn, pfn; | |
2244 | struct page *page; | |
8c3328f1 | 2245 | swp_entry_t entry; |
8763cb45 JG |
2246 | pte_t pte; |
2247 | ||
2248 | pte = *ptep; | |
8763cb45 | 2249 | |
a5430dda | 2250 | if (pte_none(pte)) { |
8315ada7 JG |
2251 | mpfn = MIGRATE_PFN_MIGRATE; |
2252 | migrate->cpages++; | |
8763cb45 JG |
2253 | goto next; |
2254 | } | |
2255 | ||
a5430dda | 2256 | if (!pte_present(pte)) { |
276f756d | 2257 | mpfn = 0; |
a5430dda JG |
2258 | |
2259 | /* | |
2260 | * Only care about unaddressable device page special | |
2261 | * page table entry. Other special swap entries are not | |
2262 | * migratable, and we ignore regular swapped page. | |
2263 | */ | |
2264 | entry = pte_to_swp_entry(pte); | |
2265 | if (!is_device_private_entry(entry)) | |
2266 | goto next; | |
2267 | ||
2268 | page = device_private_entry_to_page(entry); | |
06d462be CH |
2269 | mpfn = migrate_pfn(page_to_pfn(page)) | |
2270 | MIGRATE_PFN_MIGRATE; | |
a5430dda JG |
2271 | if (is_write_device_private_entry(entry)) |
2272 | mpfn |= MIGRATE_PFN_WRITE; | |
2273 | } else { | |
276f756d | 2274 | pfn = pte_pfn(pte); |
8315ada7 JG |
2275 | if (is_zero_pfn(pfn)) { |
2276 | mpfn = MIGRATE_PFN_MIGRATE; | |
2277 | migrate->cpages++; | |
8315ada7 JG |
2278 | goto next; |
2279 | } | |
25b2995a | 2280 | page = vm_normal_page(migrate->vma, addr, pte); |
a5430dda JG |
2281 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2282 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2283 | } | |
2284 | ||
8763cb45 | 2285 | /* FIXME support THP */ |
8763cb45 | 2286 | if (!page || !page->mapping || PageTransCompound(page)) { |
276f756d | 2287 | mpfn = 0; |
8763cb45 JG |
2288 | goto next; |
2289 | } | |
2290 | ||
2291 | /* | |
2292 | * By getting a reference on the page we pin it and that blocks | |
2293 | * any kind of migration. Side effect is that it "freezes" the | |
2294 | * pte. | |
2295 | * | |
2296 | * We drop this reference after isolating the page from the lru | |
2297 | * for non device page (device page are not on the lru and thus | |
2298 | * can't be dropped from it). | |
2299 | */ | |
2300 | get_page(page); | |
2301 | migrate->cpages++; | |
8763cb45 | 2302 | |
8c3328f1 JG |
2303 | /* |
2304 | * Optimize for the common case where page is only mapped once | |
2305 | * in one process. If we can lock the page, then we can safely | |
2306 | * set up a special migration page table entry now. | |
2307 | */ | |
2308 | if (trylock_page(page)) { | |
2309 | pte_t swp_pte; | |
2310 | ||
2311 | mpfn |= MIGRATE_PFN_LOCKED; | |
2312 | ptep_get_and_clear(mm, addr, ptep); | |
2313 | ||
2314 | /* Setup special migration page table entry */ | |
07707125 RC |
2315 | entry = make_migration_entry(page, mpfn & |
2316 | MIGRATE_PFN_WRITE); | |
8c3328f1 JG |
2317 | swp_pte = swp_entry_to_pte(entry); |
2318 | if (pte_soft_dirty(pte)) | |
2319 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2320 | set_pte_at(mm, addr, ptep, swp_pte); | |
2321 | ||
2322 | /* | |
2323 | * This is like regular unmap: we remove the rmap and | |
2324 | * drop page refcount. Page won't be freed, as we took | |
2325 | * a reference just above. | |
2326 | */ | |
2327 | page_remove_rmap(page, false); | |
2328 | put_page(page); | |
a5430dda JG |
2329 | |
2330 | if (pte_present(pte)) | |
2331 | unmapped++; | |
8c3328f1 JG |
2332 | } |
2333 | ||
8763cb45 | 2334 | next: |
a5430dda | 2335 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2336 | migrate->src[migrate->npages++] = mpfn; |
2337 | } | |
8c3328f1 | 2338 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2339 | pte_unmap_unlock(ptep - 1, ptl); |
2340 | ||
8c3328f1 JG |
2341 | /* Only flush the TLB if we actually modified any entries */ |
2342 | if (unmapped) | |
2343 | flush_tlb_range(walk->vma, start, end); | |
2344 | ||
8763cb45 JG |
2345 | return 0; |
2346 | } | |
2347 | ||
7b86ac33 CH |
2348 | static const struct mm_walk_ops migrate_vma_walk_ops = { |
2349 | .pmd_entry = migrate_vma_collect_pmd, | |
2350 | .pte_hole = migrate_vma_collect_hole, | |
2351 | }; | |
2352 | ||
8763cb45 JG |
2353 | /* |
2354 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2355 | * @migrate: migrate struct containing all migration information | |
2356 | * | |
2357 | * This will walk the CPU page table. For each virtual address backed by a | |
2358 | * valid page, it updates the src array and takes a reference on the page, in | |
2359 | * order to pin the page until we lock it and unmap it. | |
2360 | */ | |
2361 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2362 | { | |
ac46d4f3 | 2363 | struct mmu_notifier_range range; |
8763cb45 | 2364 | |
7b86ac33 CH |
2365 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, |
2366 | migrate->vma->vm_mm, migrate->start, migrate->end); | |
ac46d4f3 | 2367 | mmu_notifier_invalidate_range_start(&range); |
8763cb45 | 2368 | |
7b86ac33 CH |
2369 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, |
2370 | &migrate_vma_walk_ops, migrate); | |
2371 | ||
2372 | mmu_notifier_invalidate_range_end(&range); | |
8763cb45 JG |
2373 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2374 | } | |
2375 | ||
2376 | /* | |
2377 | * migrate_vma_check_page() - check if page is pinned or not | |
2378 | * @page: struct page to check | |
2379 | * | |
2380 | * Pinned pages cannot be migrated. This is the same test as in | |
2381 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2382 | * ZONE_DEVICE page. | |
2383 | */ | |
2384 | static bool migrate_vma_check_page(struct page *page) | |
2385 | { | |
2386 | /* | |
2387 | * One extra ref because caller holds an extra reference, either from | |
2388 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2389 | * a device page. | |
2390 | */ | |
2391 | int extra = 1; | |
2392 | ||
2393 | /* | |
2394 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2395 | * check them than regular pages, because they can be mapped with a pmd | |
2396 | * or with a pte (split pte mapping). | |
2397 | */ | |
2398 | if (PageCompound(page)) | |
2399 | return false; | |
2400 | ||
a5430dda JG |
2401 | /* Page from ZONE_DEVICE have one extra reference */ |
2402 | if (is_zone_device_page(page)) { | |
2403 | /* | |
2404 | * Private page can never be pin as they have no valid pte and | |
2405 | * GUP will fail for those. Yet if there is a pending migration | |
2406 | * a thread might try to wait on the pte migration entry and | |
2407 | * will bump the page reference count. Sadly there is no way to | |
2408 | * differentiate a regular pin from migration wait. Hence to | |
2409 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
2410 | * infinite loop (one stoping migration because the other is | |
2411 | * waiting on pte migration entry). We always return true here. | |
2412 | * | |
2413 | * FIXME proper solution is to rework migration_entry_wait() so | |
2414 | * it does not need to take a reference on page. | |
2415 | */ | |
25b2995a | 2416 | return is_device_private_page(page); |
a5430dda JG |
2417 | } |
2418 | ||
df6ad698 JG |
2419 | /* For file back page */ |
2420 | if (page_mapping(page)) | |
2421 | extra += 1 + page_has_private(page); | |
2422 | ||
8763cb45 JG |
2423 | if ((page_count(page) - extra) > page_mapcount(page)) |
2424 | return false; | |
2425 | ||
2426 | return true; | |
2427 | } | |
2428 | ||
2429 | /* | |
2430 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2431 | * @migrate: migrate struct containing all migration information | |
2432 | * | |
2433 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2434 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2435 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2436 | * migrated by concurrent kernel threads. | |
2437 | */ | |
2438 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2439 | { | |
2440 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2441 | const unsigned long start = migrate->start; |
2442 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2443 | bool allow_drain = true; |
8763cb45 JG |
2444 | |
2445 | lru_add_drain(); | |
2446 | ||
2447 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2448 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2449 | bool remap = true; |
8763cb45 JG |
2450 | |
2451 | if (!page) | |
2452 | continue; | |
2453 | ||
8c3328f1 JG |
2454 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2455 | /* | |
2456 | * Because we are migrating several pages there can be | |
2457 | * a deadlock between 2 concurrent migration where each | |
2458 | * are waiting on each other page lock. | |
2459 | * | |
2460 | * Make migrate_vma() a best effort thing and backoff | |
2461 | * for any page we can not lock right away. | |
2462 | */ | |
2463 | if (!trylock_page(page)) { | |
2464 | migrate->src[i] = 0; | |
2465 | migrate->cpages--; | |
2466 | put_page(page); | |
2467 | continue; | |
2468 | } | |
2469 | remap = false; | |
2470 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2471 | } |
8763cb45 | 2472 | |
a5430dda JG |
2473 | /* ZONE_DEVICE pages are not on LRU */ |
2474 | if (!is_zone_device_page(page)) { | |
2475 | if (!PageLRU(page) && allow_drain) { | |
2476 | /* Drain CPU's pagevec */ | |
2477 | lru_add_drain_all(); | |
2478 | allow_drain = false; | |
2479 | } | |
8763cb45 | 2480 | |
a5430dda JG |
2481 | if (isolate_lru_page(page)) { |
2482 | if (remap) { | |
2483 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2484 | migrate->cpages--; | |
2485 | restore++; | |
2486 | } else { | |
2487 | migrate->src[i] = 0; | |
2488 | unlock_page(page); | |
2489 | migrate->cpages--; | |
2490 | put_page(page); | |
2491 | } | |
2492 | continue; | |
8c3328f1 | 2493 | } |
a5430dda JG |
2494 | |
2495 | /* Drop the reference we took in collect */ | |
2496 | put_page(page); | |
8763cb45 JG |
2497 | } |
2498 | ||
2499 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2500 | if (remap) { |
2501 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2502 | migrate->cpages--; | |
2503 | restore++; | |
8763cb45 | 2504 | |
a5430dda JG |
2505 | if (!is_zone_device_page(page)) { |
2506 | get_page(page); | |
2507 | putback_lru_page(page); | |
2508 | } | |
8c3328f1 JG |
2509 | } else { |
2510 | migrate->src[i] = 0; | |
2511 | unlock_page(page); | |
2512 | migrate->cpages--; | |
2513 | ||
a5430dda JG |
2514 | if (!is_zone_device_page(page)) |
2515 | putback_lru_page(page); | |
2516 | else | |
2517 | put_page(page); | |
8c3328f1 | 2518 | } |
8763cb45 JG |
2519 | } |
2520 | } | |
8c3328f1 JG |
2521 | |
2522 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2523 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2524 | ||
2525 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2526 | continue; | |
2527 | ||
2528 | remove_migration_pte(page, migrate->vma, addr, page); | |
2529 | ||
2530 | migrate->src[i] = 0; | |
2531 | unlock_page(page); | |
2532 | put_page(page); | |
2533 | restore--; | |
2534 | } | |
8763cb45 JG |
2535 | } |
2536 | ||
2537 | /* | |
2538 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2539 | * @migrate: migrate struct containing all migration information | |
2540 | * | |
2541 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2542 | * and check again if it has been pinned. Pinned pages are restored because we | |
2543 | * cannot migrate them. | |
2544 | * | |
2545 | * This is the last step before we call the device driver callback to allocate | |
2546 | * destination memory and copy contents of original page over to new page. | |
2547 | */ | |
2548 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2549 | { | |
2550 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
2551 | const unsigned long npages = migrate->npages; | |
2552 | const unsigned long start = migrate->start; | |
2553 | unsigned long addr, i, restore = 0; | |
2554 | ||
2555 | for (i = 0; i < npages; i++) { | |
2556 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2557 | ||
2558 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2559 | continue; | |
2560 | ||
8c3328f1 JG |
2561 | if (page_mapped(page)) { |
2562 | try_to_unmap(page, flags); | |
2563 | if (page_mapped(page)) | |
2564 | goto restore; | |
8763cb45 | 2565 | } |
8c3328f1 JG |
2566 | |
2567 | if (migrate_vma_check_page(page)) | |
2568 | continue; | |
2569 | ||
2570 | restore: | |
2571 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2572 | migrate->cpages--; | |
2573 | restore++; | |
8763cb45 JG |
2574 | } |
2575 | ||
2576 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2577 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2578 | ||
2579 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2580 | continue; | |
2581 | ||
2582 | remove_migration_ptes(page, page, false); | |
2583 | ||
2584 | migrate->src[i] = 0; | |
2585 | unlock_page(page); | |
2586 | restore--; | |
2587 | ||
a5430dda JG |
2588 | if (is_zone_device_page(page)) |
2589 | put_page(page); | |
2590 | else | |
2591 | putback_lru_page(page); | |
8763cb45 JG |
2592 | } |
2593 | } | |
2594 | ||
a7d1f22b CH |
2595 | /** |
2596 | * migrate_vma_setup() - prepare to migrate a range of memory | |
2597 | * @args: contains the vma, start, and and pfns arrays for the migration | |
2598 | * | |
2599 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2600 | * without an error. | |
2601 | * | |
2602 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2603 | * the pages backing each virtual address in the range, saving them inside the | |
2604 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2605 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2606 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2607 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2608 | * remapping and unlocking those pages. | |
2609 | * | |
2610 | * The caller should then allocate destination memory and copy source memory to | |
2611 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2612 | * flag set). Once these are allocated and copied, the caller must update each | |
2613 | * corresponding entry in the dst array with the pfn value of the destination | |
2614 | * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set | |
2615 | * (destination pages must have their struct pages locked, via lock_page()). | |
2616 | * | |
2617 | * Note that the caller does not have to migrate all the pages that are marked | |
2618 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2619 | * device memory to system memory. If the caller cannot migrate a device page | |
2620 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2621 | * consequences for the userspace process, so it must be avoided if at all | |
2622 | * possible. | |
2623 | * | |
2624 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2625 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
2626 | * allowing the caller to allocate device memory for those unback virtual | |
2627 | * address. For this the caller simply has to allocate device memory and | |
2628 | * properly set the destination entry like for regular migration. Note that | |
2629 | * this can still fails and thus inside the device driver must check if the | |
2630 | * migration was successful for those entries after calling migrate_vma_pages() | |
2631 | * just like for regular migration. | |
2632 | * | |
2633 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2634 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2635 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2636 | * then migrate_vma_pages() to migrate struct page information from the source | |
2637 | * struct page to the destination struct page. If it fails to migrate the | |
2638 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2639 | * src array. | |
2640 | * | |
2641 | * At this point all successfully migrated pages have an entry in the src | |
2642 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2643 | * array entry with MIGRATE_PFN_VALID flag set. | |
2644 | * | |
2645 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2646 | * successfully migrated, and which were not. Successfully migrated pages will | |
2647 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2648 | * | |
2649 | * It is safe to update device page table after migrate_vma_pages() because | |
2650 | * both destination and source page are still locked, and the mmap_sem is held | |
2651 | * in read mode (hence no one can unmap the range being migrated). | |
2652 | * | |
2653 | * Once the caller is done cleaning up things and updating its page table (if it | |
2654 | * chose to do so, this is not an obligation) it finally calls | |
2655 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2656 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2657 | * point to the original source pages. | |
2658 | */ | |
2659 | int migrate_vma_setup(struct migrate_vma *args) | |
2660 | { | |
2661 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2662 | ||
2663 | args->start &= PAGE_MASK; | |
2664 | args->end &= PAGE_MASK; | |
2665 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2666 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2667 | return -EINVAL; | |
2668 | if (nr_pages <= 0) | |
2669 | return -EINVAL; | |
2670 | if (args->start < args->vma->vm_start || | |
2671 | args->start >= args->vma->vm_end) | |
2672 | return -EINVAL; | |
2673 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2674 | return -EINVAL; | |
2675 | if (!args->src || !args->dst) | |
2676 | return -EINVAL; | |
2677 | ||
2678 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2679 | args->cpages = 0; | |
2680 | args->npages = 0; | |
2681 | ||
2682 | migrate_vma_collect(args); | |
2683 | ||
2684 | if (args->cpages) | |
2685 | migrate_vma_prepare(args); | |
2686 | if (args->cpages) | |
2687 | migrate_vma_unmap(args); | |
2688 | ||
2689 | /* | |
2690 | * At this point pages are locked and unmapped, and thus they have | |
2691 | * stable content and can safely be copied to destination memory that | |
2692 | * is allocated by the drivers. | |
2693 | */ | |
2694 | return 0; | |
2695 | ||
2696 | } | |
2697 | EXPORT_SYMBOL(migrate_vma_setup); | |
2698 | ||
8315ada7 JG |
2699 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2700 | unsigned long addr, | |
2701 | struct page *page, | |
2702 | unsigned long *src, | |
2703 | unsigned long *dst) | |
2704 | { | |
2705 | struct vm_area_struct *vma = migrate->vma; | |
2706 | struct mm_struct *mm = vma->vm_mm; | |
2707 | struct mem_cgroup *memcg; | |
2708 | bool flush = false; | |
2709 | spinlock_t *ptl; | |
2710 | pte_t entry; | |
2711 | pgd_t *pgdp; | |
2712 | p4d_t *p4dp; | |
2713 | pud_t *pudp; | |
2714 | pmd_t *pmdp; | |
2715 | pte_t *ptep; | |
2716 | ||
2717 | /* Only allow populating anonymous memory */ | |
2718 | if (!vma_is_anonymous(vma)) | |
2719 | goto abort; | |
2720 | ||
2721 | pgdp = pgd_offset(mm, addr); | |
2722 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2723 | if (!p4dp) | |
2724 | goto abort; | |
2725 | pudp = pud_alloc(mm, p4dp, addr); | |
2726 | if (!pudp) | |
2727 | goto abort; | |
2728 | pmdp = pmd_alloc(mm, pudp, addr); | |
2729 | if (!pmdp) | |
2730 | goto abort; | |
2731 | ||
2732 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2733 | goto abort; | |
2734 | ||
2735 | /* | |
2736 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2737 | * pte_offset_map() on pmds where a huge pmd might be created | |
2738 | * from a different thread. | |
2739 | * | |
2740 | * pte_alloc_map() is safe to use under down_write(mmap_sem) or when | |
2741 | * parallel threads are excluded by other means. | |
2742 | * | |
2743 | * Here we only have down_read(mmap_sem). | |
2744 | */ | |
4cf58924 | 2745 | if (pte_alloc(mm, pmdp)) |
8315ada7 JG |
2746 | goto abort; |
2747 | ||
2748 | /* See the comment in pte_alloc_one_map() */ | |
2749 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2750 | goto abort; | |
2751 | ||
2752 | if (unlikely(anon_vma_prepare(vma))) | |
2753 | goto abort; | |
2754 | if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false)) | |
2755 | goto abort; | |
2756 | ||
2757 | /* | |
2758 | * The memory barrier inside __SetPageUptodate makes sure that | |
2759 | * preceding stores to the page contents become visible before | |
2760 | * the set_pte_at() write. | |
2761 | */ | |
2762 | __SetPageUptodate(page); | |
2763 | ||
df6ad698 JG |
2764 | if (is_zone_device_page(page)) { |
2765 | if (is_device_private_page(page)) { | |
2766 | swp_entry_t swp_entry; | |
2767 | ||
2768 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2769 | entry = swp_entry_to_pte(swp_entry); | |
df6ad698 | 2770 | } |
8315ada7 JG |
2771 | } else { |
2772 | entry = mk_pte(page, vma->vm_page_prot); | |
2773 | if (vma->vm_flags & VM_WRITE) | |
2774 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2775 | } | |
2776 | ||
2777 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2778 | ||
2779 | if (pte_present(*ptep)) { | |
2780 | unsigned long pfn = pte_pfn(*ptep); | |
2781 | ||
2782 | if (!is_zero_pfn(pfn)) { | |
2783 | pte_unmap_unlock(ptep, ptl); | |
2784 | mem_cgroup_cancel_charge(page, memcg, false); | |
2785 | goto abort; | |
2786 | } | |
2787 | flush = true; | |
2788 | } else if (!pte_none(*ptep)) { | |
2789 | pte_unmap_unlock(ptep, ptl); | |
2790 | mem_cgroup_cancel_charge(page, memcg, false); | |
2791 | goto abort; | |
2792 | } | |
2793 | ||
2794 | /* | |
2795 | * Check for usefaultfd but do not deliver the fault. Instead, | |
2796 | * just back off. | |
2797 | */ | |
2798 | if (userfaultfd_missing(vma)) { | |
2799 | pte_unmap_unlock(ptep, ptl); | |
2800 | mem_cgroup_cancel_charge(page, memcg, false); | |
2801 | goto abort; | |
2802 | } | |
2803 | ||
2804 | inc_mm_counter(mm, MM_ANONPAGES); | |
2805 | page_add_new_anon_rmap(page, vma, addr, false); | |
2806 | mem_cgroup_commit_charge(page, memcg, false, false); | |
2807 | if (!is_zone_device_page(page)) | |
2808 | lru_cache_add_active_or_unevictable(page, vma); | |
2809 | get_page(page); | |
2810 | ||
2811 | if (flush) { | |
2812 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2813 | ptep_clear_flush_notify(vma, addr, ptep); | |
2814 | set_pte_at_notify(mm, addr, ptep, entry); | |
2815 | update_mmu_cache(vma, addr, ptep); | |
2816 | } else { | |
2817 | /* No need to invalidate - it was non-present before */ | |
2818 | set_pte_at(mm, addr, ptep, entry); | |
2819 | update_mmu_cache(vma, addr, ptep); | |
2820 | } | |
2821 | ||
2822 | pte_unmap_unlock(ptep, ptl); | |
2823 | *src = MIGRATE_PFN_MIGRATE; | |
2824 | return; | |
2825 | ||
2826 | abort: | |
2827 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2828 | } | |
2829 | ||
a7d1f22b | 2830 | /** |
8763cb45 JG |
2831 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
2832 | * @migrate: migrate struct containing all migration information | |
2833 | * | |
2834 | * This migrates struct page meta-data from source struct page to destination | |
2835 | * struct page. This effectively finishes the migration from source page to the | |
2836 | * destination page. | |
2837 | */ | |
a7d1f22b | 2838 | void migrate_vma_pages(struct migrate_vma *migrate) |
8763cb45 JG |
2839 | { |
2840 | const unsigned long npages = migrate->npages; | |
2841 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
2842 | struct mmu_notifier_range range; |
2843 | unsigned long addr, i; | |
8315ada7 | 2844 | bool notified = false; |
8763cb45 JG |
2845 | |
2846 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2847 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2848 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2849 | struct address_space *mapping; | |
2850 | int r; | |
2851 | ||
8315ada7 JG |
2852 | if (!newpage) { |
2853 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2854 | continue; |
8315ada7 JG |
2855 | } |
2856 | ||
2857 | if (!page) { | |
2858 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) { | |
2859 | continue; | |
2860 | } | |
2861 | if (!notified) { | |
8315ada7 | 2862 | notified = true; |
ac46d4f3 JG |
2863 | |
2864 | mmu_notifier_range_init(&range, | |
7269f999 | 2865 | MMU_NOTIFY_CLEAR, 0, |
6f4f13e8 | 2866 | NULL, |
ac46d4f3 JG |
2867 | migrate->vma->vm_mm, |
2868 | addr, migrate->end); | |
2869 | mmu_notifier_invalidate_range_start(&range); | |
8315ada7 JG |
2870 | } |
2871 | migrate_vma_insert_page(migrate, addr, newpage, | |
2872 | &migrate->src[i], | |
2873 | &migrate->dst[i]); | |
8763cb45 | 2874 | continue; |
8315ada7 | 2875 | } |
8763cb45 JG |
2876 | |
2877 | mapping = page_mapping(page); | |
2878 | ||
a5430dda JG |
2879 | if (is_zone_device_page(newpage)) { |
2880 | if (is_device_private_page(newpage)) { | |
2881 | /* | |
2882 | * For now only support private anonymous when | |
2883 | * migrating to un-addressable device memory. | |
2884 | */ | |
2885 | if (mapping) { | |
2886 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2887 | continue; | |
2888 | } | |
25b2995a | 2889 | } else { |
a5430dda JG |
2890 | /* |
2891 | * Other types of ZONE_DEVICE page are not | |
2892 | * supported. | |
2893 | */ | |
2894 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2895 | continue; | |
2896 | } | |
2897 | } | |
2898 | ||
8763cb45 JG |
2899 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2900 | if (r != MIGRATEPAGE_SUCCESS) | |
2901 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2902 | } | |
8315ada7 | 2903 | |
4645b9fe JG |
2904 | /* |
2905 | * No need to double call mmu_notifier->invalidate_range() callback as | |
2906 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
2907 | * did already call it. | |
2908 | */ | |
8315ada7 | 2909 | if (notified) |
ac46d4f3 | 2910 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 | 2911 | } |
a7d1f22b | 2912 | EXPORT_SYMBOL(migrate_vma_pages); |
8763cb45 | 2913 | |
a7d1f22b | 2914 | /** |
8763cb45 JG |
2915 | * migrate_vma_finalize() - restore CPU page table entry |
2916 | * @migrate: migrate struct containing all migration information | |
2917 | * | |
2918 | * This replaces the special migration pte entry with either a mapping to the | |
2919 | * new page if migration was successful for that page, or to the original page | |
2920 | * otherwise. | |
2921 | * | |
2922 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2923 | * refcount, for device pages. | |
2924 | */ | |
a7d1f22b | 2925 | void migrate_vma_finalize(struct migrate_vma *migrate) |
8763cb45 JG |
2926 | { |
2927 | const unsigned long npages = migrate->npages; | |
2928 | unsigned long i; | |
2929 | ||
2930 | for (i = 0; i < npages; i++) { | |
2931 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2932 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2933 | ||
8315ada7 JG |
2934 | if (!page) { |
2935 | if (newpage) { | |
2936 | unlock_page(newpage); | |
2937 | put_page(newpage); | |
2938 | } | |
8763cb45 | 2939 | continue; |
8315ada7 JG |
2940 | } |
2941 | ||
8763cb45 JG |
2942 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2943 | if (newpage) { | |
2944 | unlock_page(newpage); | |
2945 | put_page(newpage); | |
2946 | } | |
2947 | newpage = page; | |
2948 | } | |
2949 | ||
2950 | remove_migration_ptes(page, newpage, false); | |
2951 | unlock_page(page); | |
2952 | migrate->cpages--; | |
2953 | ||
a5430dda JG |
2954 | if (is_zone_device_page(page)) |
2955 | put_page(page); | |
2956 | else | |
2957 | putback_lru_page(page); | |
8763cb45 JG |
2958 | |
2959 | if (newpage != page) { | |
2960 | unlock_page(newpage); | |
a5430dda JG |
2961 | if (is_zone_device_page(newpage)) |
2962 | put_page(newpage); | |
2963 | else | |
2964 | putback_lru_page(newpage); | |
8763cb45 JG |
2965 | } |
2966 | } | |
2967 | } | |
a7d1f22b | 2968 | EXPORT_SYMBOL(migrate_vma_finalize); |
9b2ed9cb | 2969 | #endif /* CONFIG_DEVICE_PRIVATE */ |