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> |
34290e2c | 51 | #include <linux/oom.h> |
b20a3503 | 52 | |
0d1836c3 MN |
53 | #include <asm/tlbflush.h> |
54 | ||
7b2a2d4a MG |
55 | #define CREATE_TRACE_POINTS |
56 | #include <trace/events/migrate.h> | |
57 | ||
b20a3503 CL |
58 | #include "internal.h" |
59 | ||
b20a3503 | 60 | /* |
742755a1 | 61 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
62 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
63 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
64 | */ |
65 | int migrate_prep(void) | |
66 | { | |
b20a3503 CL |
67 | /* |
68 | * Clear the LRU lists so pages can be isolated. | |
69 | * Note that pages may be moved off the LRU after we have | |
70 | * drained them. Those pages will fail to migrate like other | |
71 | * pages that may be busy. | |
72 | */ | |
73 | lru_add_drain_all(); | |
74 | ||
75 | return 0; | |
76 | } | |
77 | ||
748446bb MG |
78 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
79 | int migrate_prep_local(void) | |
80 | { | |
81 | lru_add_drain(); | |
82 | ||
83 | return 0; | |
84 | } | |
85 | ||
9e5bcd61 | 86 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
87 | { |
88 | struct address_space *mapping; | |
89 | ||
90 | /* | |
91 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
92 | * or just got freed under us. | |
93 | * | |
94 | * In case we 'win' a race for a movable page being freed under us and | |
95 | * raise its refcount preventing __free_pages() from doing its job | |
96 | * the put_page() at the end of this block will take care of | |
97 | * release this page, thus avoiding a nasty leakage. | |
98 | */ | |
99 | if (unlikely(!get_page_unless_zero(page))) | |
100 | goto out; | |
101 | ||
102 | /* | |
103 | * Check PageMovable before holding a PG_lock because page's owner | |
104 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 105 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
106 | */ |
107 | if (unlikely(!__PageMovable(page))) | |
108 | goto out_putpage; | |
109 | /* | |
110 | * As movable pages are not isolated from LRU lists, concurrent | |
111 | * compaction threads can race against page migration functions | |
112 | * as well as race against the releasing a page. | |
113 | * | |
114 | * In order to avoid having an already isolated movable page | |
115 | * being (wrongly) re-isolated while it is under migration, | |
116 | * or to avoid attempting to isolate pages being released, | |
117 | * lets be sure we have the page lock | |
118 | * before proceeding with the movable page isolation steps. | |
119 | */ | |
120 | if (unlikely(!trylock_page(page))) | |
121 | goto out_putpage; | |
122 | ||
123 | if (!PageMovable(page) || PageIsolated(page)) | |
124 | goto out_no_isolated; | |
125 | ||
126 | mapping = page_mapping(page); | |
127 | VM_BUG_ON_PAGE(!mapping, page); | |
128 | ||
129 | if (!mapping->a_ops->isolate_page(page, mode)) | |
130 | goto out_no_isolated; | |
131 | ||
132 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
133 | WARN_ON_ONCE(PageIsolated(page)); | |
134 | __SetPageIsolated(page); | |
135 | unlock_page(page); | |
136 | ||
9e5bcd61 | 137 | return 0; |
bda807d4 MK |
138 | |
139 | out_no_isolated: | |
140 | unlock_page(page); | |
141 | out_putpage: | |
142 | put_page(page); | |
143 | out: | |
9e5bcd61 | 144 | return -EBUSY; |
bda807d4 MK |
145 | } |
146 | ||
147 | /* It should be called on page which is PG_movable */ | |
148 | void putback_movable_page(struct page *page) | |
149 | { | |
150 | struct address_space *mapping; | |
151 | ||
152 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
153 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
154 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
155 | ||
156 | mapping = page_mapping(page); | |
157 | mapping->a_ops->putback_page(page); | |
158 | __ClearPageIsolated(page); | |
159 | } | |
160 | ||
5733c7d1 RA |
161 | /* |
162 | * Put previously isolated pages back onto the appropriate lists | |
163 | * from where they were once taken off for compaction/migration. | |
164 | * | |
59c82b70 JK |
165 | * This function shall be used whenever the isolated pageset has been |
166 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
167 | * and isolate_huge_page(). | |
5733c7d1 RA |
168 | */ |
169 | void putback_movable_pages(struct list_head *l) | |
170 | { | |
171 | struct page *page; | |
172 | struct page *page2; | |
173 | ||
b20a3503 | 174 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
175 | if (unlikely(PageHuge(page))) { |
176 | putback_active_hugepage(page); | |
177 | continue; | |
178 | } | |
e24f0b8f | 179 | list_del(&page->lru); |
bda807d4 MK |
180 | /* |
181 | * We isolated non-lru movable page so here we can use | |
182 | * __PageMovable because LRU page's mapping cannot have | |
183 | * PAGE_MAPPING_MOVABLE. | |
184 | */ | |
b1123ea6 | 185 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
186 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
187 | lock_page(page); | |
188 | if (PageMovable(page)) | |
189 | putback_movable_page(page); | |
190 | else | |
191 | __ClearPageIsolated(page); | |
192 | unlock_page(page); | |
193 | put_page(page); | |
194 | } else { | |
e8db67eb NH |
195 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
196 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
fc280fe8 | 197 | putback_lru_page(page); |
bda807d4 | 198 | } |
b20a3503 | 199 | } |
b20a3503 CL |
200 | } |
201 | ||
0697212a CL |
202 | /* |
203 | * Restore a potential migration pte to a working pte entry | |
204 | */ | |
e4b82222 | 205 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 206 | unsigned long addr, void *old) |
0697212a | 207 | { |
3fe87967 KS |
208 | struct page_vma_mapped_walk pvmw = { |
209 | .page = old, | |
210 | .vma = vma, | |
211 | .address = addr, | |
212 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
213 | }; | |
214 | struct page *new; | |
215 | pte_t pte; | |
0697212a | 216 | swp_entry_t entry; |
0697212a | 217 | |
3fe87967 KS |
218 | VM_BUG_ON_PAGE(PageTail(page), page); |
219 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
220 | if (PageKsm(page)) |
221 | new = page; | |
222 | else | |
223 | new = page - pvmw.page->index + | |
224 | linear_page_index(vma, pvmw.address); | |
0697212a | 225 | |
616b8371 ZY |
226 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
227 | /* PMD-mapped THP migration entry */ | |
228 | if (!pvmw.pte) { | |
229 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
230 | remove_migration_pmd(&pvmw, new); | |
231 | continue; | |
232 | } | |
233 | #endif | |
234 | ||
3fe87967 KS |
235 | get_page(new); |
236 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
237 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
238 | pte = pte_mksoft_dirty(pte); | |
0697212a | 239 | |
3fe87967 KS |
240 | /* |
241 | * Recheck VMA as permissions can change since migration started | |
242 | */ | |
243 | entry = pte_to_swp_entry(*pvmw.pte); | |
244 | if (is_write_migration_entry(entry)) | |
245 | pte = maybe_mkwrite(pte, vma); | |
d3cb8bf6 | 246 | |
df6ad698 JG |
247 | if (unlikely(is_zone_device_page(new))) { |
248 | if (is_device_private_page(new)) { | |
249 | entry = make_device_private_entry(new, pte_write(pte)); | |
250 | pte = swp_entry_to_pte(entry); | |
df6ad698 | 251 | } |
d2b2c6dd | 252 | } |
a5430dda | 253 | |
3ef8fd7f | 254 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 KS |
255 | if (PageHuge(new)) { |
256 | pte = pte_mkhuge(pte); | |
257 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
383321ab | 258 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
259 | if (PageAnon(new)) |
260 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
261 | else | |
262 | page_dup_rmap(new, true); | |
383321ab AK |
263 | } else |
264 | #endif | |
265 | { | |
266 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 267 | |
383321ab AK |
268 | if (PageAnon(new)) |
269 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
270 | else | |
271 | page_add_file_rmap(new, false); | |
272 | } | |
3fe87967 KS |
273 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
274 | mlock_vma_page(new); | |
275 | ||
e125fe40 KS |
276 | if (PageTransHuge(page) && PageMlocked(page)) |
277 | clear_page_mlock(page); | |
278 | ||
3fe87967 KS |
279 | /* No need to invalidate - it was non-present before */ |
280 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
281 | } | |
51afb12b | 282 | |
e4b82222 | 283 | return true; |
0697212a CL |
284 | } |
285 | ||
04e62a29 CL |
286 | /* |
287 | * Get rid of all migration entries and replace them by | |
288 | * references to the indicated page. | |
289 | */ | |
e388466d | 290 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 291 | { |
051ac83a JK |
292 | struct rmap_walk_control rwc = { |
293 | .rmap_one = remove_migration_pte, | |
294 | .arg = old, | |
295 | }; | |
296 | ||
e388466d KS |
297 | if (locked) |
298 | rmap_walk_locked(new, &rwc); | |
299 | else | |
300 | rmap_walk(new, &rwc); | |
04e62a29 CL |
301 | } |
302 | ||
0697212a CL |
303 | /* |
304 | * Something used the pte of a page under migration. We need to | |
305 | * get to the page and wait until migration is finished. | |
306 | * When we return from this function the fault will be retried. | |
0697212a | 307 | */ |
e66f17ff | 308 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 309 | spinlock_t *ptl) |
0697212a | 310 | { |
30dad309 | 311 | pte_t pte; |
0697212a CL |
312 | swp_entry_t entry; |
313 | struct page *page; | |
314 | ||
30dad309 | 315 | spin_lock(ptl); |
0697212a CL |
316 | pte = *ptep; |
317 | if (!is_swap_pte(pte)) | |
318 | goto out; | |
319 | ||
320 | entry = pte_to_swp_entry(pte); | |
321 | if (!is_migration_entry(entry)) | |
322 | goto out; | |
323 | ||
324 | page = migration_entry_to_page(entry); | |
325 | ||
e286781d | 326 | /* |
89eb946a | 327 | * Once page cache replacement of page migration started, page_count |
9a1ea439 HD |
328 | * is zero; but we must not call put_and_wait_on_page_locked() without |
329 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
e286781d NP |
330 | */ |
331 | if (!get_page_unless_zero(page)) | |
332 | goto out; | |
0697212a | 333 | pte_unmap_unlock(ptep, ptl); |
9a1ea439 | 334 | put_and_wait_on_page_locked(page); |
0697212a CL |
335 | return; |
336 | out: | |
337 | pte_unmap_unlock(ptep, ptl); | |
338 | } | |
339 | ||
30dad309 NH |
340 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
341 | unsigned long address) | |
342 | { | |
343 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
344 | pte_t *ptep = pte_offset_map(pmd, address); | |
345 | __migration_entry_wait(mm, ptep, ptl); | |
346 | } | |
347 | ||
cb900f41 KS |
348 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
349 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 350 | { |
cb900f41 | 351 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
352 | __migration_entry_wait(mm, pte, ptl); |
353 | } | |
354 | ||
616b8371 ZY |
355 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
356 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
357 | { | |
358 | spinlock_t *ptl; | |
359 | struct page *page; | |
360 | ||
361 | ptl = pmd_lock(mm, pmd); | |
362 | if (!is_pmd_migration_entry(*pmd)) | |
363 | goto unlock; | |
364 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
365 | if (!get_page_unless_zero(page)) | |
366 | goto unlock; | |
367 | spin_unlock(ptl); | |
9a1ea439 | 368 | put_and_wait_on_page_locked(page); |
616b8371 ZY |
369 | return; |
370 | unlock: | |
371 | spin_unlock(ptl); | |
372 | } | |
373 | #endif | |
374 | ||
f900482d | 375 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
0b3901b3 JK |
376 | { |
377 | int expected_count = 1; | |
378 | ||
379 | /* | |
380 | * Device public or private pages have an extra refcount as they are | |
381 | * ZONE_DEVICE pages. | |
382 | */ | |
383 | expected_count += is_device_private_page(page); | |
f900482d | 384 | if (mapping) |
0b3901b3 JK |
385 | expected_count += hpage_nr_pages(page) + page_has_private(page); |
386 | ||
387 | return expected_count; | |
388 | } | |
389 | ||
b20a3503 | 390 | /* |
c3fcf8a5 | 391 | * Replace the page in the mapping. |
5b5c7120 CL |
392 | * |
393 | * The number of remaining references must be: | |
394 | * 1 for anonymous pages without a mapping | |
395 | * 2 for pages with a mapping | |
266cf658 | 396 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 397 | */ |
36bc08cc | 398 | int migrate_page_move_mapping(struct address_space *mapping, |
37109694 | 399 | struct page *newpage, struct page *page, int extra_count) |
b20a3503 | 400 | { |
89eb946a | 401 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
42cb14b1 HD |
402 | struct zone *oldzone, *newzone; |
403 | int dirty; | |
f900482d | 404 | int expected_count = expected_page_refs(mapping, page) + extra_count; |
8763cb45 | 405 | |
6c5240ae | 406 | if (!mapping) { |
0e8c7d0f | 407 | /* Anonymous page without mapping */ |
8e321fef | 408 | if (page_count(page) != expected_count) |
6c5240ae | 409 | return -EAGAIN; |
cf4b769a HD |
410 | |
411 | /* No turning back from here */ | |
cf4b769a HD |
412 | newpage->index = page->index; |
413 | newpage->mapping = page->mapping; | |
414 | if (PageSwapBacked(page)) | |
fa9949da | 415 | __SetPageSwapBacked(newpage); |
cf4b769a | 416 | |
78bd5209 | 417 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
418 | } |
419 | ||
42cb14b1 HD |
420 | oldzone = page_zone(page); |
421 | newzone = page_zone(newpage); | |
422 | ||
89eb946a | 423 | xas_lock_irq(&xas); |
89eb946a MW |
424 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
425 | xas_unlock_irq(&xas); | |
e23ca00b | 426 | return -EAGAIN; |
b20a3503 CL |
427 | } |
428 | ||
fe896d18 | 429 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 430 | xas_unlock_irq(&xas); |
e286781d NP |
431 | return -EAGAIN; |
432 | } | |
433 | ||
b20a3503 | 434 | /* |
cf4b769a HD |
435 | * Now we know that no one else is looking at the page: |
436 | * no turning back from here. | |
b20a3503 | 437 | */ |
cf4b769a HD |
438 | newpage->index = page->index; |
439 | newpage->mapping = page->mapping; | |
e71769ae | 440 | page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */ |
6326fec1 NP |
441 | if (PageSwapBacked(page)) { |
442 | __SetPageSwapBacked(newpage); | |
443 | if (PageSwapCache(page)) { | |
444 | SetPageSwapCache(newpage); | |
445 | set_page_private(newpage, page_private(page)); | |
446 | } | |
447 | } else { | |
448 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
449 | } |
450 | ||
42cb14b1 HD |
451 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
452 | dirty = PageDirty(page); | |
453 | if (dirty) { | |
454 | ClearPageDirty(page); | |
455 | SetPageDirty(newpage); | |
456 | } | |
457 | ||
89eb946a | 458 | xas_store(&xas, newpage); |
e71769ae NH |
459 | if (PageTransHuge(page)) { |
460 | int i; | |
e71769ae | 461 | |
013567be | 462 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
89eb946a | 463 | xas_next(&xas); |
4101196b | 464 | xas_store(&xas, newpage); |
e71769ae | 465 | } |
e71769ae | 466 | } |
7cf9c2c7 NP |
467 | |
468 | /* | |
937a94c9 JG |
469 | * Drop cache reference from old page by unfreezing |
470 | * to one less reference. | |
7cf9c2c7 NP |
471 | * We know this isn't the last reference. |
472 | */ | |
e71769ae | 473 | page_ref_unfreeze(page, expected_count - hpage_nr_pages(page)); |
7cf9c2c7 | 474 | |
89eb946a | 475 | xas_unlock(&xas); |
42cb14b1 HD |
476 | /* Leave irq disabled to prevent preemption while updating stats */ |
477 | ||
0e8c7d0f CL |
478 | /* |
479 | * If moved to a different zone then also account | |
480 | * the page for that zone. Other VM counters will be | |
481 | * taken care of when we establish references to the | |
482 | * new page and drop references to the old page. | |
483 | * | |
484 | * Note that anonymous pages are accounted for | |
4b9d0fab | 485 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
486 | * are mapped to swap space. |
487 | */ | |
42cb14b1 | 488 | if (newzone != oldzone) { |
11fb9989 MG |
489 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); |
490 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); | |
42cb14b1 | 491 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
11fb9989 MG |
492 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); |
493 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); | |
42cb14b1 HD |
494 | } |
495 | if (dirty && mapping_cap_account_dirty(mapping)) { | |
11fb9989 | 496 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 497 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); |
11fb9989 | 498 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 499 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); |
42cb14b1 | 500 | } |
4b02108a | 501 | } |
42cb14b1 | 502 | local_irq_enable(); |
b20a3503 | 503 | |
78bd5209 | 504 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 505 | } |
1118dce7 | 506 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 507 | |
290408d4 NH |
508 | /* |
509 | * The expected number of remaining references is the same as that | |
510 | * of migrate_page_move_mapping(). | |
511 | */ | |
512 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
513 | struct page *newpage, struct page *page) | |
514 | { | |
89eb946a | 515 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 516 | int expected_count; |
290408d4 | 517 | |
89eb946a | 518 | xas_lock_irq(&xas); |
290408d4 | 519 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
520 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
521 | xas_unlock_irq(&xas); | |
290408d4 NH |
522 | return -EAGAIN; |
523 | } | |
524 | ||
fe896d18 | 525 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 526 | xas_unlock_irq(&xas); |
290408d4 NH |
527 | return -EAGAIN; |
528 | } | |
529 | ||
cf4b769a HD |
530 | newpage->index = page->index; |
531 | newpage->mapping = page->mapping; | |
6a93ca8f | 532 | |
290408d4 NH |
533 | get_page(newpage); |
534 | ||
89eb946a | 535 | xas_store(&xas, newpage); |
290408d4 | 536 | |
fe896d18 | 537 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 538 | |
89eb946a | 539 | xas_unlock_irq(&xas); |
6a93ca8f | 540 | |
78bd5209 | 541 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
542 | } |
543 | ||
30b0a105 DH |
544 | /* |
545 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
546 | * arithmetic will work across the entire page. We need something more | |
547 | * specialized. | |
548 | */ | |
549 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
550 | int nr_pages) | |
551 | { | |
552 | int i; | |
553 | struct page *dst_base = dst; | |
554 | struct page *src_base = src; | |
555 | ||
556 | for (i = 0; i < nr_pages; ) { | |
557 | cond_resched(); | |
558 | copy_highpage(dst, src); | |
559 | ||
560 | i++; | |
561 | dst = mem_map_next(dst, dst_base, i); | |
562 | src = mem_map_next(src, src_base, i); | |
563 | } | |
564 | } | |
565 | ||
566 | static void copy_huge_page(struct page *dst, struct page *src) | |
567 | { | |
568 | int i; | |
569 | int nr_pages; | |
570 | ||
571 | if (PageHuge(src)) { | |
572 | /* hugetlbfs page */ | |
573 | struct hstate *h = page_hstate(src); | |
574 | nr_pages = pages_per_huge_page(h); | |
575 | ||
576 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
577 | __copy_gigantic_page(dst, src, nr_pages); | |
578 | return; | |
579 | } | |
580 | } else { | |
581 | /* thp page */ | |
582 | BUG_ON(!PageTransHuge(src)); | |
583 | nr_pages = hpage_nr_pages(src); | |
584 | } | |
585 | ||
586 | for (i = 0; i < nr_pages; i++) { | |
587 | cond_resched(); | |
588 | copy_highpage(dst + i, src + i); | |
589 | } | |
590 | } | |
591 | ||
b20a3503 CL |
592 | /* |
593 | * Copy the page to its new location | |
594 | */ | |
2916ecc0 | 595 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 596 | { |
7851a45c RR |
597 | int cpupid; |
598 | ||
b20a3503 CL |
599 | if (PageError(page)) |
600 | SetPageError(newpage); | |
601 | if (PageReferenced(page)) | |
602 | SetPageReferenced(newpage); | |
603 | if (PageUptodate(page)) | |
604 | SetPageUptodate(newpage); | |
894bc310 | 605 | if (TestClearPageActive(page)) { |
309381fe | 606 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 607 | SetPageActive(newpage); |
418b27ef LS |
608 | } else if (TestClearPageUnevictable(page)) |
609 | SetPageUnevictable(newpage); | |
1899ad18 JW |
610 | if (PageWorkingset(page)) |
611 | SetPageWorkingset(newpage); | |
b20a3503 CL |
612 | if (PageChecked(page)) |
613 | SetPageChecked(newpage); | |
614 | if (PageMappedToDisk(page)) | |
615 | SetPageMappedToDisk(newpage); | |
616 | ||
42cb14b1 HD |
617 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
618 | if (PageDirty(page)) | |
619 | SetPageDirty(newpage); | |
b20a3503 | 620 | |
33c3fc71 VD |
621 | if (page_is_young(page)) |
622 | set_page_young(newpage); | |
623 | if (page_is_idle(page)) | |
624 | set_page_idle(newpage); | |
625 | ||
7851a45c RR |
626 | /* |
627 | * Copy NUMA information to the new page, to prevent over-eager | |
628 | * future migrations of this same page. | |
629 | */ | |
630 | cpupid = page_cpupid_xchg_last(page, -1); | |
631 | page_cpupid_xchg_last(newpage, cpupid); | |
632 | ||
e9995ef9 | 633 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
634 | /* |
635 | * Please do not reorder this without considering how mm/ksm.c's | |
636 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
637 | */ | |
b3b3a99c NH |
638 | if (PageSwapCache(page)) |
639 | ClearPageSwapCache(page); | |
b20a3503 CL |
640 | ClearPagePrivate(page); |
641 | set_page_private(page, 0); | |
b20a3503 CL |
642 | |
643 | /* | |
644 | * If any waiters have accumulated on the new page then | |
645 | * wake them up. | |
646 | */ | |
647 | if (PageWriteback(newpage)) | |
648 | end_page_writeback(newpage); | |
d435edca VB |
649 | |
650 | copy_page_owner(page, newpage); | |
74485cf2 JW |
651 | |
652 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 653 | } |
2916ecc0 JG |
654 | EXPORT_SYMBOL(migrate_page_states); |
655 | ||
656 | void migrate_page_copy(struct page *newpage, struct page *page) | |
657 | { | |
658 | if (PageHuge(page) || PageTransHuge(page)) | |
659 | copy_huge_page(newpage, page); | |
660 | else | |
661 | copy_highpage(newpage, page); | |
662 | ||
663 | migrate_page_states(newpage, page); | |
664 | } | |
1118dce7 | 665 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 666 | |
1d8b85cc CL |
667 | /************************************************************ |
668 | * Migration functions | |
669 | ***********************************************************/ | |
670 | ||
b20a3503 | 671 | /* |
bda807d4 | 672 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 673 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
674 | * |
675 | * Pages are locked upon entry and exit. | |
676 | */ | |
2d1db3b1 | 677 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
678 | struct page *newpage, struct page *page, |
679 | enum migrate_mode mode) | |
b20a3503 CL |
680 | { |
681 | int rc; | |
682 | ||
683 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
684 | ||
37109694 | 685 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
b20a3503 | 686 | |
78bd5209 | 687 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
688 | return rc; |
689 | ||
2916ecc0 JG |
690 | if (mode != MIGRATE_SYNC_NO_COPY) |
691 | migrate_page_copy(newpage, page); | |
692 | else | |
693 | migrate_page_states(newpage, page); | |
78bd5209 | 694 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
695 | } |
696 | EXPORT_SYMBOL(migrate_page); | |
697 | ||
9361401e | 698 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
699 | /* Returns true if all buffers are successfully locked */ |
700 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
701 | enum migrate_mode mode) | |
702 | { | |
703 | struct buffer_head *bh = head; | |
704 | ||
705 | /* Simple case, sync compaction */ | |
706 | if (mode != MIGRATE_ASYNC) { | |
707 | do { | |
84ade7c1 JK |
708 | lock_buffer(bh); |
709 | bh = bh->b_this_page; | |
710 | ||
711 | } while (bh != head); | |
712 | ||
713 | return true; | |
714 | } | |
715 | ||
716 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
717 | do { | |
84ade7c1 JK |
718 | if (!trylock_buffer(bh)) { |
719 | /* | |
720 | * We failed to lock the buffer and cannot stall in | |
721 | * async migration. Release the taken locks | |
722 | */ | |
723 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
724 | bh = head; |
725 | while (bh != failed_bh) { | |
726 | unlock_buffer(bh); | |
84ade7c1 JK |
727 | bh = bh->b_this_page; |
728 | } | |
729 | return false; | |
730 | } | |
731 | ||
732 | bh = bh->b_this_page; | |
733 | } while (bh != head); | |
734 | return true; | |
735 | } | |
736 | ||
89cb0888 JK |
737 | static int __buffer_migrate_page(struct address_space *mapping, |
738 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
739 | bool check_refs) | |
1d8b85cc | 740 | { |
1d8b85cc CL |
741 | struct buffer_head *bh, *head; |
742 | int rc; | |
cc4f11e6 | 743 | int expected_count; |
1d8b85cc | 744 | |
1d8b85cc | 745 | if (!page_has_buffers(page)) |
a6bc32b8 | 746 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 747 | |
cc4f11e6 | 748 | /* Check whether page does not have extra refs before we do more work */ |
f900482d | 749 | expected_count = expected_page_refs(mapping, page); |
cc4f11e6 JK |
750 | if (page_count(page) != expected_count) |
751 | return -EAGAIN; | |
1d8b85cc | 752 | |
cc4f11e6 JK |
753 | head = page_buffers(page); |
754 | if (!buffer_migrate_lock_buffers(head, mode)) | |
755 | return -EAGAIN; | |
1d8b85cc | 756 | |
89cb0888 JK |
757 | if (check_refs) { |
758 | bool busy; | |
759 | bool invalidated = false; | |
760 | ||
761 | recheck_buffers: | |
762 | busy = false; | |
763 | spin_lock(&mapping->private_lock); | |
764 | bh = head; | |
765 | do { | |
766 | if (atomic_read(&bh->b_count)) { | |
767 | busy = true; | |
768 | break; | |
769 | } | |
770 | bh = bh->b_this_page; | |
771 | } while (bh != head); | |
89cb0888 JK |
772 | if (busy) { |
773 | if (invalidated) { | |
774 | rc = -EAGAIN; | |
775 | goto unlock_buffers; | |
776 | } | |
ebdf4de5 | 777 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
778 | invalidate_bh_lrus(); |
779 | invalidated = true; | |
780 | goto recheck_buffers; | |
781 | } | |
782 | } | |
783 | ||
37109694 | 784 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
78bd5209 | 785 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 786 | goto unlock_buffers; |
1d8b85cc CL |
787 | |
788 | ClearPagePrivate(page); | |
789 | set_page_private(newpage, page_private(page)); | |
790 | set_page_private(page, 0); | |
791 | put_page(page); | |
792 | get_page(newpage); | |
793 | ||
794 | bh = head; | |
795 | do { | |
796 | set_bh_page(bh, newpage, bh_offset(bh)); | |
797 | bh = bh->b_this_page; | |
798 | ||
799 | } while (bh != head); | |
800 | ||
801 | SetPagePrivate(newpage); | |
802 | ||
2916ecc0 JG |
803 | if (mode != MIGRATE_SYNC_NO_COPY) |
804 | migrate_page_copy(newpage, page); | |
805 | else | |
806 | migrate_page_states(newpage, page); | |
1d8b85cc | 807 | |
cc4f11e6 JK |
808 | rc = MIGRATEPAGE_SUCCESS; |
809 | unlock_buffers: | |
ebdf4de5 JK |
810 | if (check_refs) |
811 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
812 | bh = head; |
813 | do { | |
814 | unlock_buffer(bh); | |
1d8b85cc CL |
815 | bh = bh->b_this_page; |
816 | ||
817 | } while (bh != head); | |
818 | ||
cc4f11e6 | 819 | return rc; |
1d8b85cc | 820 | } |
89cb0888 JK |
821 | |
822 | /* | |
823 | * Migration function for pages with buffers. This function can only be used | |
824 | * if the underlying filesystem guarantees that no other references to "page" | |
825 | * exist. For example attached buffer heads are accessed only under page lock. | |
826 | */ | |
827 | int buffer_migrate_page(struct address_space *mapping, | |
828 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
829 | { | |
830 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
831 | } | |
1d8b85cc | 832 | EXPORT_SYMBOL(buffer_migrate_page); |
89cb0888 JK |
833 | |
834 | /* | |
835 | * Same as above except that this variant is more careful and checks that there | |
836 | * are also no buffer head references. This function is the right one for | |
837 | * mappings where buffer heads are directly looked up and referenced (such as | |
838 | * block device mappings). | |
839 | */ | |
840 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
841 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
842 | { | |
843 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
844 | } | |
9361401e | 845 | #endif |
1d8b85cc | 846 | |
04e62a29 CL |
847 | /* |
848 | * Writeback a page to clean the dirty state | |
849 | */ | |
850 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 851 | { |
04e62a29 CL |
852 | struct writeback_control wbc = { |
853 | .sync_mode = WB_SYNC_NONE, | |
854 | .nr_to_write = 1, | |
855 | .range_start = 0, | |
856 | .range_end = LLONG_MAX, | |
04e62a29 CL |
857 | .for_reclaim = 1 |
858 | }; | |
859 | int rc; | |
860 | ||
861 | if (!mapping->a_ops->writepage) | |
862 | /* No write method for the address space */ | |
863 | return -EINVAL; | |
864 | ||
865 | if (!clear_page_dirty_for_io(page)) | |
866 | /* Someone else already triggered a write */ | |
867 | return -EAGAIN; | |
868 | ||
8351a6e4 | 869 | /* |
04e62a29 CL |
870 | * A dirty page may imply that the underlying filesystem has |
871 | * the page on some queue. So the page must be clean for | |
872 | * migration. Writeout may mean we loose the lock and the | |
873 | * page state is no longer what we checked for earlier. | |
874 | * At this point we know that the migration attempt cannot | |
875 | * be successful. | |
8351a6e4 | 876 | */ |
e388466d | 877 | remove_migration_ptes(page, page, false); |
8351a6e4 | 878 | |
04e62a29 | 879 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 880 | |
04e62a29 CL |
881 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
882 | /* unlocked. Relock */ | |
883 | lock_page(page); | |
884 | ||
bda8550d | 885 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
886 | } |
887 | ||
888 | /* | |
889 | * Default handling if a filesystem does not provide a migration function. | |
890 | */ | |
891 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 892 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 893 | { |
b969c4ab | 894 | if (PageDirty(page)) { |
a6bc32b8 | 895 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
896 | switch (mode) { |
897 | case MIGRATE_SYNC: | |
898 | case MIGRATE_SYNC_NO_COPY: | |
899 | break; | |
900 | default: | |
b969c4ab | 901 | return -EBUSY; |
2916ecc0 | 902 | } |
04e62a29 | 903 | return writeout(mapping, page); |
b969c4ab | 904 | } |
8351a6e4 CL |
905 | |
906 | /* | |
907 | * Buffers may be managed in a filesystem specific way. | |
908 | * We must have no buffers or drop them. | |
909 | */ | |
266cf658 | 910 | if (page_has_private(page) && |
8351a6e4 | 911 | !try_to_release_page(page, GFP_KERNEL)) |
806031bb | 912 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 913 | |
a6bc32b8 | 914 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
915 | } |
916 | ||
e24f0b8f CL |
917 | /* |
918 | * Move a page to a newly allocated page | |
919 | * The page is locked and all ptes have been successfully removed. | |
920 | * | |
921 | * The new page will have replaced the old page if this function | |
922 | * is successful. | |
894bc310 LS |
923 | * |
924 | * Return value: | |
925 | * < 0 - error code | |
78bd5209 | 926 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 927 | */ |
3fe2011f | 928 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 929 | enum migrate_mode mode) |
e24f0b8f CL |
930 | { |
931 | struct address_space *mapping; | |
bda807d4 MK |
932 | int rc = -EAGAIN; |
933 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 934 | |
7db7671f HD |
935 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
936 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 937 | |
e24f0b8f | 938 | mapping = page_mapping(page); |
bda807d4 MK |
939 | |
940 | if (likely(is_lru)) { | |
941 | if (!mapping) | |
942 | rc = migrate_page(mapping, newpage, page, mode); | |
943 | else if (mapping->a_ops->migratepage) | |
944 | /* | |
945 | * Most pages have a mapping and most filesystems | |
946 | * provide a migratepage callback. Anonymous pages | |
947 | * are part of swap space which also has its own | |
948 | * migratepage callback. This is the most common path | |
949 | * for page migration. | |
950 | */ | |
951 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
952 | page, mode); | |
953 | else | |
954 | rc = fallback_migrate_page(mapping, newpage, | |
955 | page, mode); | |
956 | } else { | |
e24f0b8f | 957 | /* |
bda807d4 MK |
958 | * In case of non-lru page, it could be released after |
959 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 960 | */ |
bda807d4 MK |
961 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
962 | if (!PageMovable(page)) { | |
963 | rc = MIGRATEPAGE_SUCCESS; | |
964 | __ClearPageIsolated(page); | |
965 | goto out; | |
966 | } | |
967 | ||
968 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
969 | page, mode); | |
970 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
971 | !PageIsolated(page)); | |
972 | } | |
e24f0b8f | 973 | |
5c3f9a67 HD |
974 | /* |
975 | * When successful, old pagecache page->mapping must be cleared before | |
976 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
977 | */ | |
978 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
979 | if (__PageMovable(page)) { |
980 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
981 | ||
982 | /* | |
983 | * We clear PG_movable under page_lock so any compactor | |
984 | * cannot try to migrate this page. | |
985 | */ | |
986 | __ClearPageIsolated(page); | |
987 | } | |
988 | ||
989 | /* | |
c23a0c99 | 990 | * Anonymous and movable page->mapping will be cleared by |
bda807d4 MK |
991 | * free_pages_prepare so don't reset it here for keeping |
992 | * the type to work PageAnon, for example. | |
993 | */ | |
994 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 995 | page->mapping = NULL; |
d2b2c6dd | 996 | |
25b2995a | 997 | if (likely(!is_zone_device_page(newpage))) |
d2b2c6dd LP |
998 | flush_dcache_page(newpage); |
999 | ||
3fe2011f | 1000 | } |
bda807d4 | 1001 | out: |
e24f0b8f CL |
1002 | return rc; |
1003 | } | |
1004 | ||
0dabec93 | 1005 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 1006 | int force, enum migrate_mode mode) |
e24f0b8f | 1007 | { |
0dabec93 | 1008 | int rc = -EAGAIN; |
2ebba6b7 | 1009 | int page_was_mapped = 0; |
3f6c8272 | 1010 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 1011 | bool is_lru = !__PageMovable(page); |
95a402c3 | 1012 | |
529ae9aa | 1013 | if (!trylock_page(page)) { |
a6bc32b8 | 1014 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 1015 | goto out; |
3e7d3449 MG |
1016 | |
1017 | /* | |
1018 | * It's not safe for direct compaction to call lock_page. | |
1019 | * For example, during page readahead pages are added locked | |
1020 | * to the LRU. Later, when the IO completes the pages are | |
1021 | * marked uptodate and unlocked. However, the queueing | |
1022 | * could be merging multiple pages for one bio (e.g. | |
1023 | * mpage_readpages). If an allocation happens for the | |
1024 | * second or third page, the process can end up locking | |
1025 | * the same page twice and deadlocking. Rather than | |
1026 | * trying to be clever about what pages can be locked, | |
1027 | * avoid the use of lock_page for direct compaction | |
1028 | * altogether. | |
1029 | */ | |
1030 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1031 | goto out; |
3e7d3449 | 1032 | |
e24f0b8f CL |
1033 | lock_page(page); |
1034 | } | |
1035 | ||
1036 | if (PageWriteback(page)) { | |
11bc82d6 | 1037 | /* |
fed5b64a | 1038 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1039 | * necessary to wait for PageWriteback. In the async case, |
1040 | * the retry loop is too short and in the sync-light case, | |
1041 | * the overhead of stalling is too much | |
11bc82d6 | 1042 | */ |
2916ecc0 JG |
1043 | switch (mode) { |
1044 | case MIGRATE_SYNC: | |
1045 | case MIGRATE_SYNC_NO_COPY: | |
1046 | break; | |
1047 | default: | |
11bc82d6 | 1048 | rc = -EBUSY; |
0a31bc97 | 1049 | goto out_unlock; |
11bc82d6 AA |
1050 | } |
1051 | if (!force) | |
0a31bc97 | 1052 | goto out_unlock; |
e24f0b8f CL |
1053 | wait_on_page_writeback(page); |
1054 | } | |
03f15c86 | 1055 | |
e24f0b8f | 1056 | /* |
dc386d4d KH |
1057 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1058 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 1059 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1060 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1061 | * File Caches may use write_page() or lock_page() in migration, then, |
1062 | * just care Anon page here. | |
03f15c86 HD |
1063 | * |
1064 | * Only page_get_anon_vma() understands the subtleties of | |
1065 | * getting a hold on an anon_vma from outside one of its mms. | |
1066 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1067 | * because that implies that the anon page is no longer mapped | |
1068 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1069 | */ |
03f15c86 | 1070 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1071 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1072 | |
7db7671f HD |
1073 | /* |
1074 | * Block others from accessing the new page when we get around to | |
1075 | * establishing additional references. We are usually the only one | |
1076 | * holding a reference to newpage at this point. We used to have a BUG | |
1077 | * here if trylock_page(newpage) fails, but would like to allow for | |
1078 | * cases where there might be a race with the previous use of newpage. | |
1079 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1080 | */ | |
1081 | if (unlikely(!trylock_page(newpage))) | |
1082 | goto out_unlock; | |
1083 | ||
bda807d4 MK |
1084 | if (unlikely(!is_lru)) { |
1085 | rc = move_to_new_page(newpage, page, mode); | |
1086 | goto out_unlock_both; | |
1087 | } | |
1088 | ||
dc386d4d | 1089 | /* |
62e1c553 SL |
1090 | * Corner case handling: |
1091 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1092 | * and treated as swapcache but it has no rmap yet. | |
1093 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1094 | * trigger a BUG. So handle it here. | |
1095 | * 2. An orphaned page (see truncate_complete_page) might have | |
1096 | * fs-private metadata. The page can be picked up due to memory | |
1097 | * offlining. Everywhere else except page reclaim, the page is | |
1098 | * invisible to the vm, so the page can not be migrated. So try to | |
1099 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1100 | */ |
62e1c553 | 1101 | if (!page->mapping) { |
309381fe | 1102 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1103 | if (page_has_private(page)) { |
62e1c553 | 1104 | try_to_free_buffers(page); |
7db7671f | 1105 | goto out_unlock_both; |
62e1c553 | 1106 | } |
7db7671f HD |
1107 | } else if (page_mapped(page)) { |
1108 | /* Establish migration ptes */ | |
03f15c86 HD |
1109 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1110 | page); | |
2ebba6b7 | 1111 | try_to_unmap(page, |
da1b13cc | 1112 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
1113 | page_was_mapped = 1; |
1114 | } | |
dc386d4d | 1115 | |
e6a1530d | 1116 | if (!page_mapped(page)) |
5c3f9a67 | 1117 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1118 | |
5c3f9a67 HD |
1119 | if (page_was_mapped) |
1120 | remove_migration_ptes(page, | |
e388466d | 1121 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1122 | |
7db7671f HD |
1123 | out_unlock_both: |
1124 | unlock_page(newpage); | |
1125 | out_unlock: | |
3f6c8272 | 1126 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1127 | if (anon_vma) |
9e60109f | 1128 | put_anon_vma(anon_vma); |
e24f0b8f | 1129 | unlock_page(page); |
0dabec93 | 1130 | out: |
c6c919eb MK |
1131 | /* |
1132 | * If migration is successful, decrease refcount of the newpage | |
1133 | * which will not free the page because new page owner increased | |
1134 | * refcounter. As well, if it is LRU page, add the page to LRU | |
e0a352fa DH |
1135 | * list in here. Use the old state of the isolated source page to |
1136 | * determine if we migrated a LRU page. newpage was already unlocked | |
1137 | * and possibly modified by its owner - don't rely on the page | |
1138 | * state. | |
c6c919eb MK |
1139 | */ |
1140 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e0a352fa | 1141 | if (unlikely(!is_lru)) |
c6c919eb MK |
1142 | put_page(newpage); |
1143 | else | |
1144 | putback_lru_page(newpage); | |
1145 | } | |
1146 | ||
0dabec93 MK |
1147 | return rc; |
1148 | } | |
95a402c3 | 1149 | |
ef2a5153 GU |
1150 | /* |
1151 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
1152 | * around it. | |
1153 | */ | |
815f0ddb ND |
1154 | #if defined(CONFIG_ARM) && \ |
1155 | defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700 | |
ef2a5153 GU |
1156 | #define ICE_noinline noinline |
1157 | #else | |
1158 | #define ICE_noinline | |
1159 | #endif | |
1160 | ||
0dabec93 MK |
1161 | /* |
1162 | * Obtain the lock on page, remove all ptes and migrate the page | |
1163 | * to the newly allocated page in newpage. | |
1164 | */ | |
ef2a5153 GU |
1165 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
1166 | free_page_t put_new_page, | |
1167 | unsigned long private, struct page *page, | |
add05cec NH |
1168 | int force, enum migrate_mode mode, |
1169 | enum migrate_reason reason) | |
0dabec93 | 1170 | { |
2def7424 | 1171 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1172 | struct page *newpage = NULL; |
0dabec93 | 1173 | |
94723aaf MH |
1174 | if (!thp_migration_supported() && PageTransHuge(page)) |
1175 | return -ENOMEM; | |
1176 | ||
0dabec93 MK |
1177 | if (page_count(page) == 1) { |
1178 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1179 | ClearPageActive(page); |
1180 | ClearPageUnevictable(page); | |
bda807d4 MK |
1181 | if (unlikely(__PageMovable(page))) { |
1182 | lock_page(page); | |
1183 | if (!PageMovable(page)) | |
1184 | __ClearPageIsolated(page); | |
1185 | unlock_page(page); | |
1186 | } | |
0dabec93 MK |
1187 | goto out; |
1188 | } | |
1189 | ||
74d4a579 YS |
1190 | newpage = get_new_page(page, private); |
1191 | if (!newpage) | |
1192 | return -ENOMEM; | |
1193 | ||
9c620e2b | 1194 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1195 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1196 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1197 | |
0dabec93 | 1198 | out: |
e24f0b8f | 1199 | if (rc != -EAGAIN) { |
0dabec93 MK |
1200 | /* |
1201 | * A page that has been migrated has all references | |
1202 | * removed and will be freed. A page that has not been | |
c23a0c99 | 1203 | * migrated will have kept its references and be restored. |
0dabec93 MK |
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 | 1284 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1285 | struct address_space *mapping = NULL; |
290408d4 | 1286 | |
83467efb | 1287 | /* |
7ed2c31d | 1288 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1289 | * This check is necessary because some callers of hugepage migration |
1290 | * like soft offline and memory hotremove don't walk through page | |
1291 | * tables or check whether the hugepage is pmd-based or not before | |
1292 | * kicking migration. | |
1293 | */ | |
100873d7 | 1294 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1295 | putback_active_hugepage(hpage); |
83467efb | 1296 | return -ENOSYS; |
32665f2b | 1297 | } |
83467efb | 1298 | |
666feb21 | 1299 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1300 | if (!new_hpage) |
1301 | return -ENOMEM; | |
1302 | ||
290408d4 | 1303 | if (!trylock_page(hpage)) { |
2916ecc0 | 1304 | if (!force) |
290408d4 | 1305 | goto out; |
2916ecc0 JG |
1306 | switch (mode) { |
1307 | case MIGRATE_SYNC: | |
1308 | case MIGRATE_SYNC_NO_COPY: | |
1309 | break; | |
1310 | default: | |
1311 | goto out; | |
1312 | } | |
290408d4 NH |
1313 | lock_page(hpage); |
1314 | } | |
1315 | ||
cb6acd01 MK |
1316 | /* |
1317 | * Check for pages which are in the process of being freed. Without | |
1318 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1319 | * be called and we could leak usage counts for subpools. | |
1320 | */ | |
1321 | if (page_private(hpage) && !page_mapping(hpage)) { | |
1322 | rc = -EBUSY; | |
1323 | goto out_unlock; | |
1324 | } | |
1325 | ||
746b18d4 PZ |
1326 | if (PageAnon(hpage)) |
1327 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1328 | |
7db7671f HD |
1329 | if (unlikely(!trylock_page(new_hpage))) |
1330 | goto put_anon; | |
1331 | ||
2ebba6b7 | 1332 | if (page_mapped(hpage)) { |
c0d0381a MK |
1333 | /* |
1334 | * try_to_unmap could potentially call huge_pmd_unshare. | |
1335 | * Because of this, take semaphore in write mode here and | |
1336 | * set TTU_RMAP_LOCKED to let lower levels know we have | |
1337 | * taken the lock. | |
1338 | */ | |
1339 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1340 | if (unlikely(!mapping)) | |
1341 | goto unlock_put_anon; | |
1342 | ||
2ebba6b7 | 1343 | try_to_unmap(hpage, |
c0d0381a MK |
1344 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS| |
1345 | TTU_RMAP_LOCKED); | |
2ebba6b7 | 1346 | page_was_mapped = 1; |
c0d0381a MK |
1347 | /* |
1348 | * Leave mapping locked until after subsequent call to | |
1349 | * remove_migration_ptes() | |
1350 | */ | |
2ebba6b7 | 1351 | } |
290408d4 NH |
1352 | |
1353 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1354 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1355 | |
c0d0381a | 1356 | if (page_was_mapped) { |
5c3f9a67 | 1357 | remove_migration_ptes(hpage, |
c0d0381a MK |
1358 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, true); |
1359 | i_mmap_unlock_write(mapping); | |
1360 | } | |
290408d4 | 1361 | |
c0d0381a | 1362 | unlock_put_anon: |
7db7671f HD |
1363 | unlock_page(new_hpage); |
1364 | ||
1365 | put_anon: | |
fd4a4663 | 1366 | if (anon_vma) |
9e60109f | 1367 | put_anon_vma(anon_vma); |
8e6ac7fa | 1368 | |
2def7424 | 1369 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1370 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1371 | put_new_page = NULL; |
1372 | } | |
8e6ac7fa | 1373 | |
cb6acd01 | 1374 | out_unlock: |
290408d4 | 1375 | unlock_page(hpage); |
09761333 | 1376 | out: |
b8ec1cee NH |
1377 | if (rc != -EAGAIN) |
1378 | putback_active_hugepage(hpage); | |
68711a74 DR |
1379 | |
1380 | /* | |
1381 | * If migration was not successful and there's a freeing callback, use | |
1382 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1383 | * isolation. | |
1384 | */ | |
2def7424 | 1385 | if (put_new_page) |
68711a74 DR |
1386 | put_new_page(new_hpage, private); |
1387 | else | |
3aaa76e1 | 1388 | putback_active_hugepage(new_hpage); |
68711a74 | 1389 | |
290408d4 NH |
1390 | return rc; |
1391 | } | |
1392 | ||
b20a3503 | 1393 | /* |
c73e5c9c SB |
1394 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1395 | * supplied as the target for the page migration | |
b20a3503 | 1396 | * |
c73e5c9c SB |
1397 | * @from: The list of pages to be migrated. |
1398 | * @get_new_page: The function used to allocate free pages to be used | |
1399 | * as the target of the page migration. | |
68711a74 DR |
1400 | * @put_new_page: The function used to free target pages if migration |
1401 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1402 | * @private: Private data to be passed on to get_new_page() |
1403 | * @mode: The migration mode that specifies the constraints for | |
1404 | * page migration, if any. | |
1405 | * @reason: The reason for page migration. | |
b20a3503 | 1406 | * |
c73e5c9c SB |
1407 | * The function returns after 10 attempts or if no pages are movable any more |
1408 | * because the list has become empty or no retryable pages exist any more. | |
14e0f9bc | 1409 | * The caller should call putback_movable_pages() to return pages to the LRU |
28bd6578 | 1410 | * or free list only if ret != 0. |
b20a3503 | 1411 | * |
c73e5c9c | 1412 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1413 | */ |
9c620e2b | 1414 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1415 | free_page_t put_new_page, unsigned long private, |
1416 | enum migrate_mode mode, int reason) | |
b20a3503 | 1417 | { |
e24f0b8f | 1418 | int retry = 1; |
b20a3503 | 1419 | int nr_failed = 0; |
5647bc29 | 1420 | int nr_succeeded = 0; |
b20a3503 CL |
1421 | int pass = 0; |
1422 | struct page *page; | |
1423 | struct page *page2; | |
1424 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1425 | int rc; | |
1426 | ||
1427 | if (!swapwrite) | |
1428 | current->flags |= PF_SWAPWRITE; | |
1429 | ||
e24f0b8f CL |
1430 | for(pass = 0; pass < 10 && retry; pass++) { |
1431 | retry = 0; | |
b20a3503 | 1432 | |
e24f0b8f | 1433 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1434 | retry: |
e24f0b8f | 1435 | cond_resched(); |
2d1db3b1 | 1436 | |
31caf665 NH |
1437 | if (PageHuge(page)) |
1438 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1439 | put_new_page, private, page, |
7cd12b4a | 1440 | pass > 2, mode, reason); |
31caf665 | 1441 | else |
68711a74 | 1442 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1443 | private, page, pass > 2, mode, |
1444 | reason); | |
2d1db3b1 | 1445 | |
e24f0b8f | 1446 | switch(rc) { |
95a402c3 | 1447 | case -ENOMEM: |
94723aaf MH |
1448 | /* |
1449 | * THP migration might be unsupported or the | |
1450 | * allocation could've failed so we should | |
1451 | * retry on the same page with the THP split | |
1452 | * to base pages. | |
1453 | * | |
1454 | * Head page is retried immediately and tail | |
1455 | * pages are added to the tail of the list so | |
1456 | * we encounter them after the rest of the list | |
1457 | * is processed. | |
1458 | */ | |
e6112fc3 | 1459 | if (PageTransHuge(page) && !PageHuge(page)) { |
94723aaf MH |
1460 | lock_page(page); |
1461 | rc = split_huge_page_to_list(page, from); | |
1462 | unlock_page(page); | |
1463 | if (!rc) { | |
1464 | list_safe_reset_next(page, page2, lru); | |
1465 | goto retry; | |
1466 | } | |
1467 | } | |
dfef2ef4 | 1468 | nr_failed++; |
95a402c3 | 1469 | goto out; |
e24f0b8f | 1470 | case -EAGAIN: |
2d1db3b1 | 1471 | retry++; |
e24f0b8f | 1472 | break; |
78bd5209 | 1473 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1474 | nr_succeeded++; |
e24f0b8f CL |
1475 | break; |
1476 | default: | |
354a3363 NH |
1477 | /* |
1478 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1479 | * unlike -EAGAIN case, the failed page is | |
1480 | * removed from migration page list and not | |
1481 | * retried in the next outer loop. | |
1482 | */ | |
2d1db3b1 | 1483 | nr_failed++; |
e24f0b8f | 1484 | break; |
2d1db3b1 | 1485 | } |
b20a3503 CL |
1486 | } |
1487 | } | |
f2f81fb2 VB |
1488 | nr_failed += retry; |
1489 | rc = nr_failed; | |
95a402c3 | 1490 | out: |
5647bc29 MG |
1491 | if (nr_succeeded) |
1492 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1493 | if (nr_failed) | |
1494 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1495 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1496 | ||
b20a3503 CL |
1497 | if (!swapwrite) |
1498 | current->flags &= ~PF_SWAPWRITE; | |
1499 | ||
78bd5209 | 1500 | return rc; |
b20a3503 | 1501 | } |
95a402c3 | 1502 | |
742755a1 | 1503 | #ifdef CONFIG_NUMA |
742755a1 | 1504 | |
a49bd4d7 | 1505 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1506 | { |
a49bd4d7 MH |
1507 | while (nr-- > 0) { |
1508 | if (put_user(value, status + start)) | |
1509 | return -EFAULT; | |
1510 | start++; | |
1511 | } | |
1512 | ||
1513 | return 0; | |
1514 | } | |
1515 | ||
1516 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1517 | struct list_head *pagelist, int node) | |
1518 | { | |
1519 | int err; | |
1520 | ||
1521 | if (list_empty(pagelist)) | |
1522 | return 0; | |
1523 | ||
1524 | err = migrate_pages(pagelist, alloc_new_node_page, NULL, node, | |
1525 | MIGRATE_SYNC, MR_SYSCALL); | |
1526 | if (err) | |
1527 | putback_movable_pages(pagelist); | |
1528 | return err; | |
742755a1 CL |
1529 | } |
1530 | ||
1531 | /* | |
a49bd4d7 MH |
1532 | * Resolves the given address to a struct page, isolates it from the LRU and |
1533 | * puts it to the given pagelist. | |
e0153fc2 YS |
1534 | * Returns: |
1535 | * errno - if the page cannot be found/isolated | |
1536 | * 0 - when it doesn't have to be migrated because it is already on the | |
1537 | * target node | |
1538 | * 1 - when it has been queued | |
742755a1 | 1539 | */ |
a49bd4d7 MH |
1540 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1541 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1542 | { |
a49bd4d7 MH |
1543 | struct vm_area_struct *vma; |
1544 | struct page *page; | |
1545 | unsigned int follflags; | |
742755a1 | 1546 | int err; |
742755a1 CL |
1547 | |
1548 | down_read(&mm->mmap_sem); | |
a49bd4d7 MH |
1549 | err = -EFAULT; |
1550 | vma = find_vma(mm, addr); | |
1551 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1552 | goto out; | |
742755a1 | 1553 | |
a49bd4d7 MH |
1554 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1555 | follflags = FOLL_GET | FOLL_DUMP; | |
a49bd4d7 | 1556 | page = follow_page(vma, addr, follflags); |
89f5b7da | 1557 | |
a49bd4d7 MH |
1558 | err = PTR_ERR(page); |
1559 | if (IS_ERR(page)) | |
1560 | goto out; | |
89f5b7da | 1561 | |
a49bd4d7 MH |
1562 | err = -ENOENT; |
1563 | if (!page) | |
1564 | goto out; | |
742755a1 | 1565 | |
a49bd4d7 MH |
1566 | err = 0; |
1567 | if (page_to_nid(page) == node) | |
1568 | goto out_putpage; | |
742755a1 | 1569 | |
a49bd4d7 MH |
1570 | err = -EACCES; |
1571 | if (page_mapcount(page) > 1 && !migrate_all) | |
1572 | goto out_putpage; | |
742755a1 | 1573 | |
a49bd4d7 MH |
1574 | if (PageHuge(page)) { |
1575 | if (PageHead(page)) { | |
1576 | isolate_huge_page(page, pagelist); | |
e0153fc2 | 1577 | err = 1; |
e632a938 | 1578 | } |
a49bd4d7 MH |
1579 | } else { |
1580 | struct page *head; | |
e632a938 | 1581 | |
e8db67eb NH |
1582 | head = compound_head(page); |
1583 | err = isolate_lru_page(head); | |
cf608ac1 | 1584 | if (err) |
a49bd4d7 | 1585 | goto out_putpage; |
742755a1 | 1586 | |
e0153fc2 | 1587 | err = 1; |
a49bd4d7 MH |
1588 | list_add_tail(&head->lru, pagelist); |
1589 | mod_node_page_state(page_pgdat(head), | |
1590 | NR_ISOLATED_ANON + page_is_file_cache(head), | |
1591 | hpage_nr_pages(head)); | |
1592 | } | |
1593 | out_putpage: | |
1594 | /* | |
1595 | * Either remove the duplicate refcount from | |
1596 | * isolate_lru_page() or drop the page ref if it was | |
1597 | * not isolated. | |
1598 | */ | |
1599 | put_page(page); | |
1600 | out: | |
742755a1 CL |
1601 | up_read(&mm->mmap_sem); |
1602 | return err; | |
1603 | } | |
1604 | ||
5e9a0f02 BG |
1605 | /* |
1606 | * Migrate an array of page address onto an array of nodes and fill | |
1607 | * the corresponding array of status. | |
1608 | */ | |
3268c63e | 1609 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1610 | unsigned long nr_pages, |
1611 | const void __user * __user *pages, | |
1612 | const int __user *nodes, | |
1613 | int __user *status, int flags) | |
1614 | { | |
a49bd4d7 MH |
1615 | int current_node = NUMA_NO_NODE; |
1616 | LIST_HEAD(pagelist); | |
1617 | int start, i; | |
1618 | int err = 0, err1; | |
35282a2d BG |
1619 | |
1620 | migrate_prep(); | |
1621 | ||
a49bd4d7 MH |
1622 | for (i = start = 0; i < nr_pages; i++) { |
1623 | const void __user *p; | |
1624 | unsigned long addr; | |
1625 | int node; | |
3140a227 | 1626 | |
a49bd4d7 MH |
1627 | err = -EFAULT; |
1628 | if (get_user(p, pages + i)) | |
1629 | goto out_flush; | |
1630 | if (get_user(node, nodes + i)) | |
1631 | goto out_flush; | |
057d3389 | 1632 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1633 | |
1634 | err = -ENODEV; | |
1635 | if (node < 0 || node >= MAX_NUMNODES) | |
1636 | goto out_flush; | |
1637 | if (!node_state(node, N_MEMORY)) | |
1638 | goto out_flush; | |
5e9a0f02 | 1639 | |
a49bd4d7 MH |
1640 | err = -EACCES; |
1641 | if (!node_isset(node, task_nodes)) | |
1642 | goto out_flush; | |
1643 | ||
1644 | if (current_node == NUMA_NO_NODE) { | |
1645 | current_node = node; | |
1646 | start = i; | |
1647 | } else if (node != current_node) { | |
1648 | err = do_move_pages_to_node(mm, &pagelist, current_node); | |
5984fabb YS |
1649 | if (err) { |
1650 | /* | |
1651 | * Positive err means the number of failed | |
1652 | * pages to migrate. Since we are going to | |
1653 | * abort and return the number of non-migrated | |
1654 | * pages, so need to incude the rest of the | |
1655 | * nr_pages that have not been attempted as | |
1656 | * well. | |
1657 | */ | |
1658 | if (err > 0) | |
1659 | err += nr_pages - i - 1; | |
a49bd4d7 | 1660 | goto out; |
5984fabb | 1661 | } |
a49bd4d7 MH |
1662 | err = store_status(status, start, current_node, i - start); |
1663 | if (err) | |
1664 | goto out; | |
1665 | start = i; | |
1666 | current_node = node; | |
3140a227 BG |
1667 | } |
1668 | ||
a49bd4d7 MH |
1669 | /* |
1670 | * Errors in the page lookup or isolation are not fatal and we simply | |
1671 | * report them via status | |
1672 | */ | |
1673 | err = add_page_for_migration(mm, addr, current_node, | |
1674 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 YS |
1675 | |
1676 | if (!err) { | |
1677 | /* The page is already on the target node */ | |
1678 | err = store_status(status, i, current_node, 1); | |
1679 | if (err) | |
1680 | goto out_flush; | |
a49bd4d7 | 1681 | continue; |
e0153fc2 YS |
1682 | } else if (err > 0) { |
1683 | /* The page is successfully queued for migration */ | |
1684 | continue; | |
1685 | } | |
3140a227 | 1686 | |
a49bd4d7 MH |
1687 | err = store_status(status, i, err, 1); |
1688 | if (err) | |
1689 | goto out_flush; | |
5e9a0f02 | 1690 | |
a49bd4d7 | 1691 | err = do_move_pages_to_node(mm, &pagelist, current_node); |
5984fabb YS |
1692 | if (err) { |
1693 | if (err > 0) | |
1694 | err += nr_pages - i - 1; | |
a49bd4d7 | 1695 | goto out; |
5984fabb | 1696 | } |
a49bd4d7 MH |
1697 | if (i > start) { |
1698 | err = store_status(status, start, current_node, i - start); | |
1699 | if (err) | |
1700 | goto out; | |
1701 | } | |
1702 | current_node = NUMA_NO_NODE; | |
3140a227 | 1703 | } |
a49bd4d7 | 1704 | out_flush: |
8f175cf5 MH |
1705 | if (list_empty(&pagelist)) |
1706 | return err; | |
1707 | ||
a49bd4d7 MH |
1708 | /* Make sure we do not overwrite the existing error */ |
1709 | err1 = do_move_pages_to_node(mm, &pagelist, current_node); | |
5984fabb YS |
1710 | /* |
1711 | * Don't have to report non-attempted pages here since: | |
1712 | * - If the above loop is done gracefully all pages have been | |
1713 | * attempted. | |
1714 | * - If the above loop is aborted it means a fatal error | |
1715 | * happened, should return ret. | |
1716 | */ | |
a49bd4d7 MH |
1717 | if (!err1) |
1718 | err1 = store_status(status, start, current_node, i - start); | |
dfe9aa23 | 1719 | if (err >= 0) |
a49bd4d7 | 1720 | err = err1; |
5e9a0f02 BG |
1721 | out: |
1722 | return err; | |
1723 | } | |
1724 | ||
742755a1 | 1725 | /* |
2f007e74 | 1726 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1727 | */ |
80bba129 BG |
1728 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1729 | const void __user **pages, int *status) | |
742755a1 | 1730 | { |
2f007e74 | 1731 | unsigned long i; |
2f007e74 | 1732 | |
742755a1 CL |
1733 | down_read(&mm->mmap_sem); |
1734 | ||
2f007e74 | 1735 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1736 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1737 | struct vm_area_struct *vma; |
1738 | struct page *page; | |
c095adbc | 1739 | int err = -EFAULT; |
2f007e74 BG |
1740 | |
1741 | vma = find_vma(mm, addr); | |
70384dc6 | 1742 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1743 | goto set_status; |
1744 | ||
d899844e KS |
1745 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1746 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1747 | |
1748 | err = PTR_ERR(page); | |
1749 | if (IS_ERR(page)) | |
1750 | goto set_status; | |
1751 | ||
d899844e | 1752 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1753 | set_status: |
80bba129 BG |
1754 | *status = err; |
1755 | ||
1756 | pages++; | |
1757 | status++; | |
1758 | } | |
1759 | ||
1760 | up_read(&mm->mmap_sem); | |
1761 | } | |
1762 | ||
1763 | /* | |
1764 | * Determine the nodes of a user array of pages and store it in | |
1765 | * a user array of status. | |
1766 | */ | |
1767 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1768 | const void __user * __user *pages, | |
1769 | int __user *status) | |
1770 | { | |
1771 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1772 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1773 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1774 | |
87b8d1ad PA |
1775 | while (nr_pages) { |
1776 | unsigned long chunk_nr; | |
80bba129 | 1777 | |
87b8d1ad PA |
1778 | chunk_nr = nr_pages; |
1779 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1780 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1781 | ||
1782 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1783 | break; | |
80bba129 BG |
1784 | |
1785 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1786 | ||
87b8d1ad PA |
1787 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1788 | break; | |
742755a1 | 1789 | |
87b8d1ad PA |
1790 | pages += chunk_nr; |
1791 | status += chunk_nr; | |
1792 | nr_pages -= chunk_nr; | |
1793 | } | |
1794 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1795 | } |
1796 | ||
1797 | /* | |
1798 | * Move a list of pages in the address space of the currently executing | |
1799 | * process. | |
1800 | */ | |
7addf443 DB |
1801 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, |
1802 | const void __user * __user *pages, | |
1803 | const int __user *nodes, | |
1804 | int __user *status, int flags) | |
742755a1 | 1805 | { |
742755a1 | 1806 | struct task_struct *task; |
742755a1 | 1807 | struct mm_struct *mm; |
5e9a0f02 | 1808 | int err; |
3268c63e | 1809 | nodemask_t task_nodes; |
742755a1 CL |
1810 | |
1811 | /* Check flags */ | |
1812 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1813 | return -EINVAL; | |
1814 | ||
1815 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1816 | return -EPERM; | |
1817 | ||
1818 | /* Find the mm_struct */ | |
a879bf58 | 1819 | rcu_read_lock(); |
228ebcbe | 1820 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1821 | if (!task) { |
a879bf58 | 1822 | rcu_read_unlock(); |
742755a1 CL |
1823 | return -ESRCH; |
1824 | } | |
3268c63e | 1825 | get_task_struct(task); |
742755a1 CL |
1826 | |
1827 | /* | |
1828 | * Check if this process has the right to modify the specified | |
197e7e52 | 1829 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1830 | */ |
197e7e52 | 1831 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1832 | rcu_read_unlock(); |
742755a1 | 1833 | err = -EPERM; |
5e9a0f02 | 1834 | goto out; |
742755a1 | 1835 | } |
c69e8d9c | 1836 | rcu_read_unlock(); |
742755a1 | 1837 | |
86c3a764 DQ |
1838 | err = security_task_movememory(task); |
1839 | if (err) | |
5e9a0f02 | 1840 | goto out; |
86c3a764 | 1841 | |
3268c63e CL |
1842 | task_nodes = cpuset_mems_allowed(task); |
1843 | mm = get_task_mm(task); | |
1844 | put_task_struct(task); | |
1845 | ||
6e8b09ea SL |
1846 | if (!mm) |
1847 | return -EINVAL; | |
1848 | ||
1849 | if (nodes) | |
1850 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1851 | nodes, status, flags); | |
1852 | else | |
1853 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1854 | |
742755a1 CL |
1855 | mmput(mm); |
1856 | return err; | |
3268c63e CL |
1857 | |
1858 | out: | |
1859 | put_task_struct(task); | |
1860 | return err; | |
742755a1 | 1861 | } |
742755a1 | 1862 | |
7addf443 DB |
1863 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1864 | const void __user * __user *, pages, | |
1865 | const int __user *, nodes, | |
1866 | int __user *, status, int, flags) | |
1867 | { | |
1868 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1869 | } | |
1870 | ||
1871 | #ifdef CONFIG_COMPAT | |
1872 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, | |
1873 | compat_uptr_t __user *, pages32, | |
1874 | const int __user *, nodes, | |
1875 | int __user *, status, | |
1876 | int, flags) | |
1877 | { | |
1878 | const void __user * __user *pages; | |
1879 | int i; | |
1880 | ||
1881 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); | |
1882 | for (i = 0; i < nr_pages; i++) { | |
1883 | compat_uptr_t p; | |
1884 | ||
1885 | if (get_user(p, pages32 + i) || | |
1886 | put_user(compat_ptr(p), pages + i)) | |
1887 | return -EFAULT; | |
1888 | } | |
1889 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1890 | } | |
1891 | #endif /* CONFIG_COMPAT */ | |
1892 | ||
7039e1db PZ |
1893 | #ifdef CONFIG_NUMA_BALANCING |
1894 | /* | |
1895 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1896 | * pages. Currently it only checks the watermarks which crude | |
1897 | */ | |
1898 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1899 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1900 | { |
1901 | int z; | |
599d0c95 | 1902 | |
7039e1db PZ |
1903 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1904 | struct zone *zone = pgdat->node_zones + z; | |
1905 | ||
1906 | if (!populated_zone(zone)) | |
1907 | continue; | |
1908 | ||
7039e1db PZ |
1909 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1910 | if (!zone_watermark_ok(zone, 0, | |
1911 | high_wmark_pages(zone) + | |
1912 | nr_migrate_pages, | |
bfe9d006 | 1913 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
1914 | continue; |
1915 | return true; | |
1916 | } | |
1917 | return false; | |
1918 | } | |
1919 | ||
1920 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 1921 | unsigned long data) |
7039e1db PZ |
1922 | { |
1923 | int nid = (int) data; | |
1924 | struct page *newpage; | |
1925 | ||
96db800f | 1926 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1927 | (GFP_HIGHUSER_MOVABLE | |
1928 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1929 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 1930 | ~__GFP_RECLAIM, 0); |
bac0382c | 1931 | |
7039e1db PZ |
1932 | return newpage; |
1933 | } | |
1934 | ||
1c30e017 | 1935 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1936 | { |
340ef390 | 1937 | int page_lru; |
a8f60772 | 1938 | |
309381fe | 1939 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1940 | |
7039e1db | 1941 | /* Avoid migrating to a node that is nearly full */ |
d8c6546b | 1942 | if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) |
340ef390 | 1943 | return 0; |
7039e1db | 1944 | |
340ef390 HD |
1945 | if (isolate_lru_page(page)) |
1946 | return 0; | |
7039e1db | 1947 | |
340ef390 HD |
1948 | /* |
1949 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1950 | * check on page_count(), so we must do it here, now that the page | |
1951 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1952 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1953 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1954 | */ | |
1955 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1956 | putback_lru_page(page); | |
1957 | return 0; | |
7039e1db PZ |
1958 | } |
1959 | ||
340ef390 | 1960 | page_lru = page_is_file_cache(page); |
599d0c95 | 1961 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
340ef390 HD |
1962 | hpage_nr_pages(page)); |
1963 | ||
149c33e1 | 1964 | /* |
340ef390 HD |
1965 | * Isolating the page has taken another reference, so the |
1966 | * caller's reference can be safely dropped without the page | |
1967 | * disappearing underneath us during migration. | |
149c33e1 MG |
1968 | */ |
1969 | put_page(page); | |
340ef390 | 1970 | return 1; |
b32967ff MG |
1971 | } |
1972 | ||
de466bd6 MG |
1973 | bool pmd_trans_migrating(pmd_t pmd) |
1974 | { | |
1975 | struct page *page = pmd_page(pmd); | |
1976 | return PageLocked(page); | |
1977 | } | |
1978 | ||
b32967ff MG |
1979 | /* |
1980 | * Attempt to migrate a misplaced page to the specified destination | |
1981 | * node. Caller is expected to have an elevated reference count on | |
1982 | * the page that will be dropped by this function before returning. | |
1983 | */ | |
1bc115d8 MG |
1984 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1985 | int node) | |
b32967ff MG |
1986 | { |
1987 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1988 | int isolated; |
b32967ff MG |
1989 | int nr_remaining; |
1990 | LIST_HEAD(migratepages); | |
1991 | ||
1992 | /* | |
1bc115d8 MG |
1993 | * Don't migrate file pages that are mapped in multiple processes |
1994 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1995 | */ |
1bc115d8 MG |
1996 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1997 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1998 | goto out; |
b32967ff | 1999 | |
09a913a7 MG |
2000 | /* |
2001 | * Also do not migrate dirty pages as not all filesystems can move | |
2002 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2003 | */ | |
2004 | if (page_is_file_cache(page) && PageDirty(page)) | |
2005 | goto out; | |
2006 | ||
b32967ff MG |
2007 | isolated = numamigrate_isolate_page(pgdat, page); |
2008 | if (!isolated) | |
2009 | goto out; | |
2010 | ||
2011 | list_add(&page->lru, &migratepages); | |
9c620e2b | 2012 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
2013 | NULL, node, MIGRATE_ASYNC, |
2014 | MR_NUMA_MISPLACED); | |
b32967ff | 2015 | if (nr_remaining) { |
59c82b70 JK |
2016 | if (!list_empty(&migratepages)) { |
2017 | list_del(&page->lru); | |
599d0c95 | 2018 | dec_node_page_state(page, NR_ISOLATED_ANON + |
59c82b70 JK |
2019 | page_is_file_cache(page)); |
2020 | putback_lru_page(page); | |
2021 | } | |
b32967ff MG |
2022 | isolated = 0; |
2023 | } else | |
2024 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 2025 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2026 | return isolated; |
340ef390 HD |
2027 | |
2028 | out: | |
2029 | put_page(page); | |
2030 | return 0; | |
7039e1db | 2031 | } |
220018d3 | 2032 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 2033 | |
220018d3 | 2034 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
2035 | /* |
2036 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
2037 | * before returning. | |
2038 | */ | |
b32967ff MG |
2039 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
2040 | struct vm_area_struct *vma, | |
2041 | pmd_t *pmd, pmd_t entry, | |
2042 | unsigned long address, | |
2043 | struct page *page, int node) | |
2044 | { | |
c4088ebd | 2045 | spinlock_t *ptl; |
b32967ff MG |
2046 | pg_data_t *pgdat = NODE_DATA(node); |
2047 | int isolated = 0; | |
2048 | struct page *new_page = NULL; | |
b32967ff | 2049 | int page_lru = page_is_file_cache(page); |
7066f0f9 | 2050 | unsigned long start = address & HPAGE_PMD_MASK; |
b32967ff | 2051 | |
b32967ff | 2052 | new_page = alloc_pages_node(node, |
25160354 | 2053 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
e97ca8e5 | 2054 | HPAGE_PMD_ORDER); |
340ef390 HD |
2055 | if (!new_page) |
2056 | goto out_fail; | |
9a982250 | 2057 | prep_transhuge_page(new_page); |
340ef390 | 2058 | |
b32967ff | 2059 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 2060 | if (!isolated) { |
b32967ff | 2061 | put_page(new_page); |
340ef390 | 2062 | goto out_fail; |
b32967ff | 2063 | } |
b0943d61 | 2064 | |
b32967ff | 2065 | /* Prepare a page as a migration target */ |
48c935ad | 2066 | __SetPageLocked(new_page); |
d44d363f SL |
2067 | if (PageSwapBacked(page)) |
2068 | __SetPageSwapBacked(new_page); | |
b32967ff MG |
2069 | |
2070 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2071 | new_page->mapping = page->mapping; | |
2072 | new_page->index = page->index; | |
7eef5f97 AA |
2073 | /* flush the cache before copying using the kernel virtual address */ |
2074 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); | |
b32967ff MG |
2075 | migrate_page_copy(new_page, page); |
2076 | WARN_ON(PageLRU(new_page)); | |
2077 | ||
2078 | /* Recheck the target PMD */ | |
c4088ebd | 2079 | ptl = pmd_lock(mm, pmd); |
f4e177d1 | 2080 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
c4088ebd | 2081 | spin_unlock(ptl); |
b32967ff MG |
2082 | |
2083 | /* Reverse changes made by migrate_page_copy() */ | |
2084 | if (TestClearPageActive(new_page)) | |
2085 | SetPageActive(page); | |
2086 | if (TestClearPageUnevictable(new_page)) | |
2087 | SetPageUnevictable(page); | |
b32967ff MG |
2088 | |
2089 | unlock_page(new_page); | |
2090 | put_page(new_page); /* Free it */ | |
2091 | ||
a54a407f MG |
2092 | /* Retake the callers reference and putback on LRU */ |
2093 | get_page(page); | |
b32967ff | 2094 | putback_lru_page(page); |
599d0c95 | 2095 | mod_node_page_state(page_pgdat(page), |
a54a407f | 2096 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
eb4489f6 MG |
2097 | |
2098 | goto out_unlock; | |
b32967ff MG |
2099 | } |
2100 | ||
10102459 | 2101 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
f55e1014 | 2102 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 2103 | |
2b4847e7 | 2104 | /* |
d7c33934 AA |
2105 | * Overwrite the old entry under pagetable lock and establish |
2106 | * the new PTE. Any parallel GUP will either observe the old | |
2107 | * page blocking on the page lock, block on the page table | |
2108 | * lock or observe the new page. The SetPageUptodate on the | |
2109 | * new page and page_add_new_anon_rmap guarantee the copy is | |
2110 | * visible before the pagetable update. | |
2b4847e7 | 2111 | */ |
7066f0f9 | 2112 | page_add_anon_rmap(new_page, vma, start, true); |
d7c33934 AA |
2113 | /* |
2114 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB | |
2115 | * has already been flushed globally. So no TLB can be currently | |
2116 | * caching this non present pmd mapping. There's no need to clear the | |
2117 | * pmd before doing set_pmd_at(), nor to flush the TLB after | |
2118 | * set_pmd_at(). Clearing the pmd here would introduce a race | |
2119 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the | |
2120 | * mmap_sem for reading. If the pmd is set to NULL at any given time, | |
2121 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this | |
2122 | * pmd. | |
2123 | */ | |
7066f0f9 | 2124 | set_pmd_at(mm, start, pmd, entry); |
ce4a9cc5 | 2125 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 | 2126 | |
f4e177d1 | 2127 | page_ref_unfreeze(page, 2); |
51afb12b | 2128 | mlock_migrate_page(new_page, page); |
d281ee61 | 2129 | page_remove_rmap(page, true); |
7cd12b4a | 2130 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2b4847e7 | 2131 | |
c4088ebd | 2132 | spin_unlock(ptl); |
b32967ff | 2133 | |
11de9927 MG |
2134 | /* Take an "isolate" reference and put new page on the LRU. */ |
2135 | get_page(new_page); | |
2136 | putback_lru_page(new_page); | |
2137 | ||
b32967ff MG |
2138 | unlock_page(new_page); |
2139 | unlock_page(page); | |
2140 | put_page(page); /* Drop the rmap reference */ | |
2141 | put_page(page); /* Drop the LRU isolation reference */ | |
2142 | ||
2143 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2144 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2145 | ||
599d0c95 | 2146 | mod_node_page_state(page_pgdat(page), |
b32967ff MG |
2147 | NR_ISOLATED_ANON + page_lru, |
2148 | -HPAGE_PMD_NR); | |
2149 | return isolated; | |
2150 | ||
340ef390 HD |
2151 | out_fail: |
2152 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2b4847e7 MG |
2153 | ptl = pmd_lock(mm, pmd); |
2154 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 2155 | entry = pmd_modify(entry, vma->vm_page_prot); |
7066f0f9 | 2156 | set_pmd_at(mm, start, pmd, entry); |
2b4847e7 MG |
2157 | update_mmu_cache_pmd(vma, address, &entry); |
2158 | } | |
2159 | spin_unlock(ptl); | |
a54a407f | 2160 | |
eb4489f6 | 2161 | out_unlock: |
340ef390 | 2162 | unlock_page(page); |
b32967ff | 2163 | put_page(page); |
b32967ff MG |
2164 | return 0; |
2165 | } | |
7039e1db PZ |
2166 | #endif /* CONFIG_NUMA_BALANCING */ |
2167 | ||
2168 | #endif /* CONFIG_NUMA */ | |
8763cb45 | 2169 | |
9b2ed9cb | 2170 | #ifdef CONFIG_DEVICE_PRIVATE |
8763cb45 JG |
2171 | static int migrate_vma_collect_hole(unsigned long start, |
2172 | unsigned long end, | |
b7a16c7a | 2173 | __always_unused int depth, |
8763cb45 JG |
2174 | struct mm_walk *walk) |
2175 | { | |
2176 | struct migrate_vma *migrate = walk->private; | |
2177 | unsigned long addr; | |
2178 | ||
872ea707 | 2179 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
e20d103b | 2180 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8315ada7 | 2181 | migrate->dst[migrate->npages] = 0; |
e20d103b | 2182 | migrate->npages++; |
8315ada7 JG |
2183 | migrate->cpages++; |
2184 | } | |
2185 | ||
2186 | return 0; | |
2187 | } | |
2188 | ||
2189 | static int migrate_vma_collect_skip(unsigned long start, | |
2190 | unsigned long end, | |
2191 | struct mm_walk *walk) | |
2192 | { | |
2193 | struct migrate_vma *migrate = walk->private; | |
2194 | unsigned long addr; | |
2195 | ||
872ea707 | 2196 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
8763cb45 JG |
2197 | migrate->dst[migrate->npages] = 0; |
2198 | migrate->src[migrate->npages++] = 0; | |
2199 | } | |
2200 | ||
2201 | return 0; | |
2202 | } | |
2203 | ||
2204 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2205 | unsigned long start, | |
2206 | unsigned long end, | |
2207 | struct mm_walk *walk) | |
2208 | { | |
2209 | struct migrate_vma *migrate = walk->private; | |
2210 | struct vm_area_struct *vma = walk->vma; | |
2211 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2212 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2213 | spinlock_t *ptl; |
2214 | pte_t *ptep; | |
2215 | ||
2216 | again: | |
2217 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2218 | return migrate_vma_collect_hole(start, end, -1, walk); |
8763cb45 JG |
2219 | |
2220 | if (pmd_trans_huge(*pmdp)) { | |
2221 | struct page *page; | |
2222 | ||
2223 | ptl = pmd_lock(mm, pmdp); | |
2224 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2225 | spin_unlock(ptl); | |
2226 | goto again; | |
2227 | } | |
2228 | ||
2229 | page = pmd_page(*pmdp); | |
2230 | if (is_huge_zero_page(page)) { | |
2231 | spin_unlock(ptl); | |
2232 | split_huge_pmd(vma, pmdp, addr); | |
2233 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2234 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2235 | walk); |
2236 | } else { | |
2237 | int ret; | |
2238 | ||
2239 | get_page(page); | |
2240 | spin_unlock(ptl); | |
2241 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2242 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2243 | walk); |
2244 | ret = split_huge_page(page); | |
2245 | unlock_page(page); | |
2246 | put_page(page); | |
8315ada7 JG |
2247 | if (ret) |
2248 | return migrate_vma_collect_skip(start, end, | |
2249 | walk); | |
2250 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2251 | return migrate_vma_collect_hole(start, end, -1, |
8763cb45 JG |
2252 | walk); |
2253 | } | |
2254 | } | |
2255 | ||
2256 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2257 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2258 | |
2259 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2260 | arch_enter_lazy_mmu_mode(); |
2261 | ||
8763cb45 | 2262 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
800bb1c8 | 2263 | unsigned long mpfn = 0, pfn; |
8763cb45 | 2264 | struct page *page; |
8c3328f1 | 2265 | swp_entry_t entry; |
8763cb45 JG |
2266 | pte_t pte; |
2267 | ||
2268 | pte = *ptep; | |
8763cb45 | 2269 | |
a5430dda | 2270 | if (pte_none(pte)) { |
8315ada7 JG |
2271 | mpfn = MIGRATE_PFN_MIGRATE; |
2272 | migrate->cpages++; | |
8763cb45 JG |
2273 | goto next; |
2274 | } | |
2275 | ||
a5430dda | 2276 | if (!pte_present(pte)) { |
a5430dda JG |
2277 | /* |
2278 | * Only care about unaddressable device page special | |
2279 | * page table entry. Other special swap entries are not | |
2280 | * migratable, and we ignore regular swapped page. | |
2281 | */ | |
2282 | entry = pte_to_swp_entry(pte); | |
2283 | if (!is_device_private_entry(entry)) | |
2284 | goto next; | |
2285 | ||
2286 | page = device_private_entry_to_page(entry); | |
800bb1c8 CH |
2287 | if (page->pgmap->owner != migrate->src_owner) |
2288 | goto next; | |
2289 | ||
06d462be CH |
2290 | mpfn = migrate_pfn(page_to_pfn(page)) | |
2291 | MIGRATE_PFN_MIGRATE; | |
a5430dda JG |
2292 | if (is_write_device_private_entry(entry)) |
2293 | mpfn |= MIGRATE_PFN_WRITE; | |
2294 | } else { | |
800bb1c8 CH |
2295 | if (migrate->src_owner) |
2296 | goto next; | |
276f756d | 2297 | pfn = pte_pfn(pte); |
8315ada7 JG |
2298 | if (is_zero_pfn(pfn)) { |
2299 | mpfn = MIGRATE_PFN_MIGRATE; | |
2300 | migrate->cpages++; | |
8315ada7 JG |
2301 | goto next; |
2302 | } | |
25b2995a | 2303 | page = vm_normal_page(migrate->vma, addr, pte); |
a5430dda JG |
2304 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2305 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2306 | } | |
2307 | ||
8763cb45 | 2308 | /* FIXME support THP */ |
8763cb45 | 2309 | if (!page || !page->mapping || PageTransCompound(page)) { |
276f756d | 2310 | mpfn = 0; |
8763cb45 JG |
2311 | goto next; |
2312 | } | |
2313 | ||
2314 | /* | |
2315 | * By getting a reference on the page we pin it and that blocks | |
2316 | * any kind of migration. Side effect is that it "freezes" the | |
2317 | * pte. | |
2318 | * | |
2319 | * We drop this reference after isolating the page from the lru | |
2320 | * for non device page (device page are not on the lru and thus | |
2321 | * can't be dropped from it). | |
2322 | */ | |
2323 | get_page(page); | |
2324 | migrate->cpages++; | |
8763cb45 | 2325 | |
8c3328f1 JG |
2326 | /* |
2327 | * Optimize for the common case where page is only mapped once | |
2328 | * in one process. If we can lock the page, then we can safely | |
2329 | * set up a special migration page table entry now. | |
2330 | */ | |
2331 | if (trylock_page(page)) { | |
2332 | pte_t swp_pte; | |
2333 | ||
2334 | mpfn |= MIGRATE_PFN_LOCKED; | |
2335 | ptep_get_and_clear(mm, addr, ptep); | |
2336 | ||
2337 | /* Setup special migration page table entry */ | |
07707125 RC |
2338 | entry = make_migration_entry(page, mpfn & |
2339 | MIGRATE_PFN_WRITE); | |
8c3328f1 JG |
2340 | swp_pte = swp_entry_to_pte(entry); |
2341 | if (pte_soft_dirty(pte)) | |
2342 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2343 | set_pte_at(mm, addr, ptep, swp_pte); | |
2344 | ||
2345 | /* | |
2346 | * This is like regular unmap: we remove the rmap and | |
2347 | * drop page refcount. Page won't be freed, as we took | |
2348 | * a reference just above. | |
2349 | */ | |
2350 | page_remove_rmap(page, false); | |
2351 | put_page(page); | |
a5430dda JG |
2352 | |
2353 | if (pte_present(pte)) | |
2354 | unmapped++; | |
8c3328f1 JG |
2355 | } |
2356 | ||
8763cb45 | 2357 | next: |
a5430dda | 2358 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2359 | migrate->src[migrate->npages++] = mpfn; |
2360 | } | |
8c3328f1 | 2361 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2362 | pte_unmap_unlock(ptep - 1, ptl); |
2363 | ||
8c3328f1 JG |
2364 | /* Only flush the TLB if we actually modified any entries */ |
2365 | if (unmapped) | |
2366 | flush_tlb_range(walk->vma, start, end); | |
2367 | ||
8763cb45 JG |
2368 | return 0; |
2369 | } | |
2370 | ||
7b86ac33 CH |
2371 | static const struct mm_walk_ops migrate_vma_walk_ops = { |
2372 | .pmd_entry = migrate_vma_collect_pmd, | |
2373 | .pte_hole = migrate_vma_collect_hole, | |
2374 | }; | |
2375 | ||
8763cb45 JG |
2376 | /* |
2377 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2378 | * @migrate: migrate struct containing all migration information | |
2379 | * | |
2380 | * This will walk the CPU page table. For each virtual address backed by a | |
2381 | * valid page, it updates the src array and takes a reference on the page, in | |
2382 | * order to pin the page until we lock it and unmap it. | |
2383 | */ | |
2384 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2385 | { | |
ac46d4f3 | 2386 | struct mmu_notifier_range range; |
8763cb45 | 2387 | |
7b86ac33 CH |
2388 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, |
2389 | migrate->vma->vm_mm, migrate->start, migrate->end); | |
ac46d4f3 | 2390 | mmu_notifier_invalidate_range_start(&range); |
8763cb45 | 2391 | |
7b86ac33 CH |
2392 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, |
2393 | &migrate_vma_walk_ops, migrate); | |
2394 | ||
2395 | mmu_notifier_invalidate_range_end(&range); | |
8763cb45 JG |
2396 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2397 | } | |
2398 | ||
2399 | /* | |
2400 | * migrate_vma_check_page() - check if page is pinned or not | |
2401 | * @page: struct page to check | |
2402 | * | |
2403 | * Pinned pages cannot be migrated. This is the same test as in | |
2404 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2405 | * ZONE_DEVICE page. | |
2406 | */ | |
2407 | static bool migrate_vma_check_page(struct page *page) | |
2408 | { | |
2409 | /* | |
2410 | * One extra ref because caller holds an extra reference, either from | |
2411 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2412 | * a device page. | |
2413 | */ | |
2414 | int extra = 1; | |
2415 | ||
2416 | /* | |
2417 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2418 | * check them than regular pages, because they can be mapped with a pmd | |
2419 | * or with a pte (split pte mapping). | |
2420 | */ | |
2421 | if (PageCompound(page)) | |
2422 | return false; | |
2423 | ||
a5430dda JG |
2424 | /* Page from ZONE_DEVICE have one extra reference */ |
2425 | if (is_zone_device_page(page)) { | |
2426 | /* | |
2427 | * Private page can never be pin as they have no valid pte and | |
2428 | * GUP will fail for those. Yet if there is a pending migration | |
2429 | * a thread might try to wait on the pte migration entry and | |
2430 | * will bump the page reference count. Sadly there is no way to | |
2431 | * differentiate a regular pin from migration wait. Hence to | |
2432 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
2433 | * infinite loop (one stoping migration because the other is | |
2434 | * waiting on pte migration entry). We always return true here. | |
2435 | * | |
2436 | * FIXME proper solution is to rework migration_entry_wait() so | |
2437 | * it does not need to take a reference on page. | |
2438 | */ | |
25b2995a | 2439 | return is_device_private_page(page); |
a5430dda JG |
2440 | } |
2441 | ||
df6ad698 JG |
2442 | /* For file back page */ |
2443 | if (page_mapping(page)) | |
2444 | extra += 1 + page_has_private(page); | |
2445 | ||
8763cb45 JG |
2446 | if ((page_count(page) - extra) > page_mapcount(page)) |
2447 | return false; | |
2448 | ||
2449 | return true; | |
2450 | } | |
2451 | ||
2452 | /* | |
2453 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2454 | * @migrate: migrate struct containing all migration information | |
2455 | * | |
2456 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2457 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2458 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2459 | * migrated by concurrent kernel threads. | |
2460 | */ | |
2461 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2462 | { | |
2463 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2464 | const unsigned long start = migrate->start; |
2465 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2466 | bool allow_drain = true; |
8763cb45 JG |
2467 | |
2468 | lru_add_drain(); | |
2469 | ||
2470 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2471 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2472 | bool remap = true; |
8763cb45 JG |
2473 | |
2474 | if (!page) | |
2475 | continue; | |
2476 | ||
8c3328f1 JG |
2477 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2478 | /* | |
2479 | * Because we are migrating several pages there can be | |
2480 | * a deadlock between 2 concurrent migration where each | |
2481 | * are waiting on each other page lock. | |
2482 | * | |
2483 | * Make migrate_vma() a best effort thing and backoff | |
2484 | * for any page we can not lock right away. | |
2485 | */ | |
2486 | if (!trylock_page(page)) { | |
2487 | migrate->src[i] = 0; | |
2488 | migrate->cpages--; | |
2489 | put_page(page); | |
2490 | continue; | |
2491 | } | |
2492 | remap = false; | |
2493 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2494 | } |
8763cb45 | 2495 | |
a5430dda JG |
2496 | /* ZONE_DEVICE pages are not on LRU */ |
2497 | if (!is_zone_device_page(page)) { | |
2498 | if (!PageLRU(page) && allow_drain) { | |
2499 | /* Drain CPU's pagevec */ | |
2500 | lru_add_drain_all(); | |
2501 | allow_drain = false; | |
2502 | } | |
8763cb45 | 2503 | |
a5430dda JG |
2504 | if (isolate_lru_page(page)) { |
2505 | if (remap) { | |
2506 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2507 | migrate->cpages--; | |
2508 | restore++; | |
2509 | } else { | |
2510 | migrate->src[i] = 0; | |
2511 | unlock_page(page); | |
2512 | migrate->cpages--; | |
2513 | put_page(page); | |
2514 | } | |
2515 | continue; | |
8c3328f1 | 2516 | } |
a5430dda JG |
2517 | |
2518 | /* Drop the reference we took in collect */ | |
2519 | put_page(page); | |
8763cb45 JG |
2520 | } |
2521 | ||
2522 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2523 | if (remap) { |
2524 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2525 | migrate->cpages--; | |
2526 | restore++; | |
8763cb45 | 2527 | |
a5430dda JG |
2528 | if (!is_zone_device_page(page)) { |
2529 | get_page(page); | |
2530 | putback_lru_page(page); | |
2531 | } | |
8c3328f1 JG |
2532 | } else { |
2533 | migrate->src[i] = 0; | |
2534 | unlock_page(page); | |
2535 | migrate->cpages--; | |
2536 | ||
a5430dda JG |
2537 | if (!is_zone_device_page(page)) |
2538 | putback_lru_page(page); | |
2539 | else | |
2540 | put_page(page); | |
8c3328f1 | 2541 | } |
8763cb45 JG |
2542 | } |
2543 | } | |
8c3328f1 JG |
2544 | |
2545 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2546 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2547 | ||
2548 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2549 | continue; | |
2550 | ||
2551 | remove_migration_pte(page, migrate->vma, addr, page); | |
2552 | ||
2553 | migrate->src[i] = 0; | |
2554 | unlock_page(page); | |
2555 | put_page(page); | |
2556 | restore--; | |
2557 | } | |
8763cb45 JG |
2558 | } |
2559 | ||
2560 | /* | |
2561 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2562 | * @migrate: migrate struct containing all migration information | |
2563 | * | |
2564 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2565 | * and check again if it has been pinned. Pinned pages are restored because we | |
2566 | * cannot migrate them. | |
2567 | * | |
2568 | * This is the last step before we call the device driver callback to allocate | |
2569 | * destination memory and copy contents of original page over to new page. | |
2570 | */ | |
2571 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2572 | { | |
2573 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
2574 | const unsigned long npages = migrate->npages; | |
2575 | const unsigned long start = migrate->start; | |
2576 | unsigned long addr, i, restore = 0; | |
2577 | ||
2578 | for (i = 0; i < npages; i++) { | |
2579 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2580 | ||
2581 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2582 | continue; | |
2583 | ||
8c3328f1 JG |
2584 | if (page_mapped(page)) { |
2585 | try_to_unmap(page, flags); | |
2586 | if (page_mapped(page)) | |
2587 | goto restore; | |
8763cb45 | 2588 | } |
8c3328f1 JG |
2589 | |
2590 | if (migrate_vma_check_page(page)) | |
2591 | continue; | |
2592 | ||
2593 | restore: | |
2594 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2595 | migrate->cpages--; | |
2596 | restore++; | |
8763cb45 JG |
2597 | } |
2598 | ||
2599 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2600 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2601 | ||
2602 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2603 | continue; | |
2604 | ||
2605 | remove_migration_ptes(page, page, false); | |
2606 | ||
2607 | migrate->src[i] = 0; | |
2608 | unlock_page(page); | |
2609 | restore--; | |
2610 | ||
a5430dda JG |
2611 | if (is_zone_device_page(page)) |
2612 | put_page(page); | |
2613 | else | |
2614 | putback_lru_page(page); | |
8763cb45 JG |
2615 | } |
2616 | } | |
2617 | ||
a7d1f22b CH |
2618 | /** |
2619 | * migrate_vma_setup() - prepare to migrate a range of memory | |
2620 | * @args: contains the vma, start, and and pfns arrays for the migration | |
2621 | * | |
2622 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2623 | * without an error. | |
2624 | * | |
2625 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2626 | * the pages backing each virtual address in the range, saving them inside the | |
2627 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2628 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2629 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2630 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2631 | * remapping and unlocking those pages. | |
2632 | * | |
2633 | * The caller should then allocate destination memory and copy source memory to | |
2634 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2635 | * flag set). Once these are allocated and copied, the caller must update each | |
2636 | * corresponding entry in the dst array with the pfn value of the destination | |
2637 | * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set | |
2638 | * (destination pages must have their struct pages locked, via lock_page()). | |
2639 | * | |
2640 | * Note that the caller does not have to migrate all the pages that are marked | |
2641 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2642 | * device memory to system memory. If the caller cannot migrate a device page | |
2643 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2644 | * consequences for the userspace process, so it must be avoided if at all | |
2645 | * possible. | |
2646 | * | |
2647 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2648 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
2649 | * allowing the caller to allocate device memory for those unback virtual | |
2650 | * address. For this the caller simply has to allocate device memory and | |
2651 | * properly set the destination entry like for regular migration. Note that | |
2652 | * this can still fails and thus inside the device driver must check if the | |
2653 | * migration was successful for those entries after calling migrate_vma_pages() | |
2654 | * just like for regular migration. | |
2655 | * | |
2656 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2657 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2658 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2659 | * then migrate_vma_pages() to migrate struct page information from the source | |
2660 | * struct page to the destination struct page. If it fails to migrate the | |
2661 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2662 | * src array. | |
2663 | * | |
2664 | * At this point all successfully migrated pages have an entry in the src | |
2665 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2666 | * array entry with MIGRATE_PFN_VALID flag set. | |
2667 | * | |
2668 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2669 | * successfully migrated, and which were not. Successfully migrated pages will | |
2670 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2671 | * | |
2672 | * It is safe to update device page table after migrate_vma_pages() because | |
2673 | * both destination and source page are still locked, and the mmap_sem is held | |
2674 | * in read mode (hence no one can unmap the range being migrated). | |
2675 | * | |
2676 | * Once the caller is done cleaning up things and updating its page table (if it | |
2677 | * chose to do so, this is not an obligation) it finally calls | |
2678 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2679 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2680 | * point to the original source pages. | |
2681 | */ | |
2682 | int migrate_vma_setup(struct migrate_vma *args) | |
2683 | { | |
2684 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2685 | ||
2686 | args->start &= PAGE_MASK; | |
2687 | args->end &= PAGE_MASK; | |
2688 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2689 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2690 | return -EINVAL; | |
2691 | if (nr_pages <= 0) | |
2692 | return -EINVAL; | |
2693 | if (args->start < args->vma->vm_start || | |
2694 | args->start >= args->vma->vm_end) | |
2695 | return -EINVAL; | |
2696 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2697 | return -EINVAL; | |
2698 | if (!args->src || !args->dst) | |
2699 | return -EINVAL; | |
2700 | ||
2701 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2702 | args->cpages = 0; | |
2703 | args->npages = 0; | |
2704 | ||
2705 | migrate_vma_collect(args); | |
2706 | ||
2707 | if (args->cpages) | |
2708 | migrate_vma_prepare(args); | |
2709 | if (args->cpages) | |
2710 | migrate_vma_unmap(args); | |
2711 | ||
2712 | /* | |
2713 | * At this point pages are locked and unmapped, and thus they have | |
2714 | * stable content and can safely be copied to destination memory that | |
2715 | * is allocated by the drivers. | |
2716 | */ | |
2717 | return 0; | |
2718 | ||
2719 | } | |
2720 | EXPORT_SYMBOL(migrate_vma_setup); | |
2721 | ||
34290e2c RC |
2722 | /* |
2723 | * This code closely matches the code in: | |
2724 | * __handle_mm_fault() | |
2725 | * handle_pte_fault() | |
2726 | * do_anonymous_page() | |
2727 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
2728 | * private page. | |
2729 | */ | |
8315ada7 JG |
2730 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2731 | unsigned long addr, | |
2732 | struct page *page, | |
2733 | unsigned long *src, | |
2734 | unsigned long *dst) | |
2735 | { | |
2736 | struct vm_area_struct *vma = migrate->vma; | |
2737 | struct mm_struct *mm = vma->vm_mm; | |
2738 | struct mem_cgroup *memcg; | |
2739 | bool flush = false; | |
2740 | spinlock_t *ptl; | |
2741 | pte_t entry; | |
2742 | pgd_t *pgdp; | |
2743 | p4d_t *p4dp; | |
2744 | pud_t *pudp; | |
2745 | pmd_t *pmdp; | |
2746 | pte_t *ptep; | |
2747 | ||
2748 | /* Only allow populating anonymous memory */ | |
2749 | if (!vma_is_anonymous(vma)) | |
2750 | goto abort; | |
2751 | ||
2752 | pgdp = pgd_offset(mm, addr); | |
2753 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2754 | if (!p4dp) | |
2755 | goto abort; | |
2756 | pudp = pud_alloc(mm, p4dp, addr); | |
2757 | if (!pudp) | |
2758 | goto abort; | |
2759 | pmdp = pmd_alloc(mm, pudp, addr); | |
2760 | if (!pmdp) | |
2761 | goto abort; | |
2762 | ||
2763 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2764 | goto abort; | |
2765 | ||
2766 | /* | |
2767 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2768 | * pte_offset_map() on pmds where a huge pmd might be created | |
2769 | * from a different thread. | |
2770 | * | |
2771 | * pte_alloc_map() is safe to use under down_write(mmap_sem) or when | |
2772 | * parallel threads are excluded by other means. | |
2773 | * | |
2774 | * Here we only have down_read(mmap_sem). | |
2775 | */ | |
4cf58924 | 2776 | if (pte_alloc(mm, pmdp)) |
8315ada7 JG |
2777 | goto abort; |
2778 | ||
2779 | /* See the comment in pte_alloc_one_map() */ | |
2780 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2781 | goto abort; | |
2782 | ||
2783 | if (unlikely(anon_vma_prepare(vma))) | |
2784 | goto abort; | |
2785 | if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false)) | |
2786 | goto abort; | |
2787 | ||
2788 | /* | |
2789 | * The memory barrier inside __SetPageUptodate makes sure that | |
2790 | * preceding stores to the page contents become visible before | |
2791 | * the set_pte_at() write. | |
2792 | */ | |
2793 | __SetPageUptodate(page); | |
2794 | ||
df6ad698 JG |
2795 | if (is_zone_device_page(page)) { |
2796 | if (is_device_private_page(page)) { | |
2797 | swp_entry_t swp_entry; | |
2798 | ||
2799 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2800 | entry = swp_entry_to_pte(swp_entry); | |
df6ad698 | 2801 | } |
8315ada7 JG |
2802 | } else { |
2803 | entry = mk_pte(page, vma->vm_page_prot); | |
2804 | if (vma->vm_flags & VM_WRITE) | |
2805 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2806 | } | |
2807 | ||
2808 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2809 | ||
34290e2c RC |
2810 | if (check_stable_address_space(mm)) |
2811 | goto unlock_abort; | |
2812 | ||
8315ada7 JG |
2813 | if (pte_present(*ptep)) { |
2814 | unsigned long pfn = pte_pfn(*ptep); | |
2815 | ||
c23a0c99 RC |
2816 | if (!is_zero_pfn(pfn)) |
2817 | goto unlock_abort; | |
8315ada7 | 2818 | flush = true; |
c23a0c99 RC |
2819 | } else if (!pte_none(*ptep)) |
2820 | goto unlock_abort; | |
8315ada7 JG |
2821 | |
2822 | /* | |
c23a0c99 | 2823 | * Check for userfaultfd but do not deliver the fault. Instead, |
8315ada7 JG |
2824 | * just back off. |
2825 | */ | |
c23a0c99 RC |
2826 | if (userfaultfd_missing(vma)) |
2827 | goto unlock_abort; | |
8315ada7 JG |
2828 | |
2829 | inc_mm_counter(mm, MM_ANONPAGES); | |
2830 | page_add_new_anon_rmap(page, vma, addr, false); | |
2831 | mem_cgroup_commit_charge(page, memcg, false, false); | |
2832 | if (!is_zone_device_page(page)) | |
2833 | lru_cache_add_active_or_unevictable(page, vma); | |
2834 | get_page(page); | |
2835 | ||
2836 | if (flush) { | |
2837 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2838 | ptep_clear_flush_notify(vma, addr, ptep); | |
2839 | set_pte_at_notify(mm, addr, ptep, entry); | |
2840 | update_mmu_cache(vma, addr, ptep); | |
2841 | } else { | |
2842 | /* No need to invalidate - it was non-present before */ | |
2843 | set_pte_at(mm, addr, ptep, entry); | |
2844 | update_mmu_cache(vma, addr, ptep); | |
2845 | } | |
2846 | ||
2847 | pte_unmap_unlock(ptep, ptl); | |
2848 | *src = MIGRATE_PFN_MIGRATE; | |
2849 | return; | |
2850 | ||
c23a0c99 RC |
2851 | unlock_abort: |
2852 | pte_unmap_unlock(ptep, ptl); | |
2853 | mem_cgroup_cancel_charge(page, memcg, false); | |
8315ada7 JG |
2854 | abort: |
2855 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2856 | } | |
2857 | ||
a7d1f22b | 2858 | /** |
8763cb45 JG |
2859 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
2860 | * @migrate: migrate struct containing all migration information | |
2861 | * | |
2862 | * This migrates struct page meta-data from source struct page to destination | |
2863 | * struct page. This effectively finishes the migration from source page to the | |
2864 | * destination page. | |
2865 | */ | |
a7d1f22b | 2866 | void migrate_vma_pages(struct migrate_vma *migrate) |
8763cb45 JG |
2867 | { |
2868 | const unsigned long npages = migrate->npages; | |
2869 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
2870 | struct mmu_notifier_range range; |
2871 | unsigned long addr, i; | |
8315ada7 | 2872 | bool notified = false; |
8763cb45 JG |
2873 | |
2874 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2875 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2876 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2877 | struct address_space *mapping; | |
2878 | int r; | |
2879 | ||
8315ada7 JG |
2880 | if (!newpage) { |
2881 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2882 | continue; |
8315ada7 JG |
2883 | } |
2884 | ||
2885 | if (!page) { | |
c23a0c99 | 2886 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
8315ada7 | 2887 | continue; |
8315ada7 | 2888 | if (!notified) { |
8315ada7 | 2889 | notified = true; |
ac46d4f3 JG |
2890 | |
2891 | mmu_notifier_range_init(&range, | |
7269f999 | 2892 | MMU_NOTIFY_CLEAR, 0, |
6f4f13e8 | 2893 | NULL, |
ac46d4f3 JG |
2894 | migrate->vma->vm_mm, |
2895 | addr, migrate->end); | |
2896 | mmu_notifier_invalidate_range_start(&range); | |
8315ada7 JG |
2897 | } |
2898 | migrate_vma_insert_page(migrate, addr, newpage, | |
2899 | &migrate->src[i], | |
2900 | &migrate->dst[i]); | |
8763cb45 | 2901 | continue; |
8315ada7 | 2902 | } |
8763cb45 JG |
2903 | |
2904 | mapping = page_mapping(page); | |
2905 | ||
a5430dda JG |
2906 | if (is_zone_device_page(newpage)) { |
2907 | if (is_device_private_page(newpage)) { | |
2908 | /* | |
2909 | * For now only support private anonymous when | |
2910 | * migrating to un-addressable device memory. | |
2911 | */ | |
2912 | if (mapping) { | |
2913 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2914 | continue; | |
2915 | } | |
25b2995a | 2916 | } else { |
a5430dda JG |
2917 | /* |
2918 | * Other types of ZONE_DEVICE page are not | |
2919 | * supported. | |
2920 | */ | |
2921 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2922 | continue; | |
2923 | } | |
2924 | } | |
2925 | ||
8763cb45 JG |
2926 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2927 | if (r != MIGRATEPAGE_SUCCESS) | |
2928 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2929 | } | |
8315ada7 | 2930 | |
4645b9fe JG |
2931 | /* |
2932 | * No need to double call mmu_notifier->invalidate_range() callback as | |
2933 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
2934 | * did already call it. | |
2935 | */ | |
8315ada7 | 2936 | if (notified) |
ac46d4f3 | 2937 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 | 2938 | } |
a7d1f22b | 2939 | EXPORT_SYMBOL(migrate_vma_pages); |
8763cb45 | 2940 | |
a7d1f22b | 2941 | /** |
8763cb45 JG |
2942 | * migrate_vma_finalize() - restore CPU page table entry |
2943 | * @migrate: migrate struct containing all migration information | |
2944 | * | |
2945 | * This replaces the special migration pte entry with either a mapping to the | |
2946 | * new page if migration was successful for that page, or to the original page | |
2947 | * otherwise. | |
2948 | * | |
2949 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2950 | * refcount, for device pages. | |
2951 | */ | |
a7d1f22b | 2952 | void migrate_vma_finalize(struct migrate_vma *migrate) |
8763cb45 JG |
2953 | { |
2954 | const unsigned long npages = migrate->npages; | |
2955 | unsigned long i; | |
2956 | ||
2957 | for (i = 0; i < npages; i++) { | |
2958 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2959 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2960 | ||
8315ada7 JG |
2961 | if (!page) { |
2962 | if (newpage) { | |
2963 | unlock_page(newpage); | |
2964 | put_page(newpage); | |
2965 | } | |
8763cb45 | 2966 | continue; |
8315ada7 JG |
2967 | } |
2968 | ||
8763cb45 JG |
2969 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2970 | if (newpage) { | |
2971 | unlock_page(newpage); | |
2972 | put_page(newpage); | |
2973 | } | |
2974 | newpage = page; | |
2975 | } | |
2976 | ||
2977 | remove_migration_ptes(page, newpage, false); | |
2978 | unlock_page(page); | |
2979 | migrate->cpages--; | |
2980 | ||
a5430dda JG |
2981 | if (is_zone_device_page(page)) |
2982 | put_page(page); | |
2983 | else | |
2984 | putback_lru_page(page); | |
8763cb45 JG |
2985 | |
2986 | if (newpage != page) { | |
2987 | unlock_page(newpage); | |
a5430dda JG |
2988 | if (is_zone_device_page(newpage)) |
2989 | put_page(newpage); | |
2990 | else | |
2991 | putback_lru_page(newpage); | |
8763cb45 JG |
2992 | } |
2993 | } | |
2994 | } | |
a7d1f22b | 2995 | EXPORT_SYMBOL(migrate_vma_finalize); |
9b2ed9cb | 2996 | #endif /* CONFIG_DEVICE_PRIVATE */ |