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> |
df6ad698 | 41 | #include <linux/pfn_t.h> |
a5430dda | 42 | #include <linux/memremap.h> |
8315ada7 | 43 | #include <linux/userfaultfd_k.h> |
bf6bddf1 | 44 | #include <linux/balloon_compaction.h> |
f714f4f2 | 45 | #include <linux/mmu_notifier.h> |
33c3fc71 | 46 | #include <linux/page_idle.h> |
d435edca | 47 | #include <linux/page_owner.h> |
6e84f315 | 48 | #include <linux/sched/mm.h> |
197e7e52 | 49 | #include <linux/ptrace.h> |
b20a3503 | 50 | |
0d1836c3 MN |
51 | #include <asm/tlbflush.h> |
52 | ||
7b2a2d4a MG |
53 | #define CREATE_TRACE_POINTS |
54 | #include <trace/events/migrate.h> | |
55 | ||
b20a3503 CL |
56 | #include "internal.h" |
57 | ||
b20a3503 | 58 | /* |
742755a1 | 59 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
60 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
61 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
62 | */ |
63 | int migrate_prep(void) | |
64 | { | |
b20a3503 CL |
65 | /* |
66 | * Clear the LRU lists so pages can be isolated. | |
67 | * Note that pages may be moved off the LRU after we have | |
68 | * drained them. Those pages will fail to migrate like other | |
69 | * pages that may be busy. | |
70 | */ | |
71 | lru_add_drain_all(); | |
72 | ||
73 | return 0; | |
74 | } | |
75 | ||
748446bb MG |
76 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
77 | int migrate_prep_local(void) | |
78 | { | |
79 | lru_add_drain(); | |
80 | ||
81 | return 0; | |
82 | } | |
83 | ||
9e5bcd61 | 84 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
85 | { |
86 | struct address_space *mapping; | |
87 | ||
88 | /* | |
89 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
90 | * or just got freed under us. | |
91 | * | |
92 | * In case we 'win' a race for a movable page being freed under us and | |
93 | * raise its refcount preventing __free_pages() from doing its job | |
94 | * the put_page() at the end of this block will take care of | |
95 | * release this page, thus avoiding a nasty leakage. | |
96 | */ | |
97 | if (unlikely(!get_page_unless_zero(page))) | |
98 | goto out; | |
99 | ||
100 | /* | |
101 | * Check PageMovable before holding a PG_lock because page's owner | |
102 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
103 | * so unconditionally grapping the lock ruins page's owner side. | |
104 | */ | |
105 | if (unlikely(!__PageMovable(page))) | |
106 | goto out_putpage; | |
107 | /* | |
108 | * As movable pages are not isolated from LRU lists, concurrent | |
109 | * compaction threads can race against page migration functions | |
110 | * as well as race against the releasing a page. | |
111 | * | |
112 | * In order to avoid having an already isolated movable page | |
113 | * being (wrongly) re-isolated while it is under migration, | |
114 | * or to avoid attempting to isolate pages being released, | |
115 | * lets be sure we have the page lock | |
116 | * before proceeding with the movable page isolation steps. | |
117 | */ | |
118 | if (unlikely(!trylock_page(page))) | |
119 | goto out_putpage; | |
120 | ||
121 | if (!PageMovable(page) || PageIsolated(page)) | |
122 | goto out_no_isolated; | |
123 | ||
124 | mapping = page_mapping(page); | |
125 | VM_BUG_ON_PAGE(!mapping, page); | |
126 | ||
127 | if (!mapping->a_ops->isolate_page(page, mode)) | |
128 | goto out_no_isolated; | |
129 | ||
130 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
131 | WARN_ON_ONCE(PageIsolated(page)); | |
132 | __SetPageIsolated(page); | |
133 | unlock_page(page); | |
134 | ||
9e5bcd61 | 135 | return 0; |
bda807d4 MK |
136 | |
137 | out_no_isolated: | |
138 | unlock_page(page); | |
139 | out_putpage: | |
140 | put_page(page); | |
141 | out: | |
9e5bcd61 | 142 | return -EBUSY; |
bda807d4 MK |
143 | } |
144 | ||
145 | /* It should be called on page which is PG_movable */ | |
146 | void putback_movable_page(struct page *page) | |
147 | { | |
148 | struct address_space *mapping; | |
149 | ||
150 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
151 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
152 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
153 | ||
154 | mapping = page_mapping(page); | |
155 | mapping->a_ops->putback_page(page); | |
156 | __ClearPageIsolated(page); | |
157 | } | |
158 | ||
5733c7d1 RA |
159 | /* |
160 | * Put previously isolated pages back onto the appropriate lists | |
161 | * from where they were once taken off for compaction/migration. | |
162 | * | |
59c82b70 JK |
163 | * This function shall be used whenever the isolated pageset has been |
164 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
165 | * and isolate_huge_page(). | |
5733c7d1 RA |
166 | */ |
167 | void putback_movable_pages(struct list_head *l) | |
168 | { | |
169 | struct page *page; | |
170 | struct page *page2; | |
171 | ||
b20a3503 | 172 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
173 | if (unlikely(PageHuge(page))) { |
174 | putback_active_hugepage(page); | |
175 | continue; | |
176 | } | |
e24f0b8f | 177 | list_del(&page->lru); |
bda807d4 MK |
178 | /* |
179 | * We isolated non-lru movable page so here we can use | |
180 | * __PageMovable because LRU page's mapping cannot have | |
181 | * PAGE_MAPPING_MOVABLE. | |
182 | */ | |
b1123ea6 | 183 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
184 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
185 | lock_page(page); | |
186 | if (PageMovable(page)) | |
187 | putback_movable_page(page); | |
188 | else | |
189 | __ClearPageIsolated(page); | |
190 | unlock_page(page); | |
191 | put_page(page); | |
192 | } else { | |
e8db67eb NH |
193 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
194 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
fc280fe8 | 195 | putback_lru_page(page); |
bda807d4 | 196 | } |
b20a3503 | 197 | } |
b20a3503 CL |
198 | } |
199 | ||
0697212a CL |
200 | /* |
201 | * Restore a potential migration pte to a working pte entry | |
202 | */ | |
e4b82222 | 203 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 204 | unsigned long addr, void *old) |
0697212a | 205 | { |
3fe87967 KS |
206 | struct page_vma_mapped_walk pvmw = { |
207 | .page = old, | |
208 | .vma = vma, | |
209 | .address = addr, | |
210 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
211 | }; | |
212 | struct page *new; | |
213 | pte_t pte; | |
0697212a | 214 | swp_entry_t entry; |
0697212a | 215 | |
3fe87967 KS |
216 | VM_BUG_ON_PAGE(PageTail(page), page); |
217 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
218 | if (PageKsm(page)) |
219 | new = page; | |
220 | else | |
221 | new = page - pvmw.page->index + | |
222 | linear_page_index(vma, pvmw.address); | |
0697212a | 223 | |
616b8371 ZY |
224 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
225 | /* PMD-mapped THP migration entry */ | |
226 | if (!pvmw.pte) { | |
227 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
228 | remove_migration_pmd(&pvmw, new); | |
229 | continue; | |
230 | } | |
231 | #endif | |
232 | ||
3fe87967 KS |
233 | get_page(new); |
234 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
235 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
236 | pte = pte_mksoft_dirty(pte); | |
0697212a | 237 | |
3fe87967 KS |
238 | /* |
239 | * Recheck VMA as permissions can change since migration started | |
240 | */ | |
241 | entry = pte_to_swp_entry(*pvmw.pte); | |
242 | if (is_write_migration_entry(entry)) | |
243 | pte = maybe_mkwrite(pte, vma); | |
d3cb8bf6 | 244 | |
df6ad698 JG |
245 | if (unlikely(is_zone_device_page(new))) { |
246 | if (is_device_private_page(new)) { | |
247 | entry = make_device_private_entry(new, pte_write(pte)); | |
248 | pte = swp_entry_to_pte(entry); | |
249 | } else if (is_device_public_page(new)) { | |
250 | pte = pte_mkdevmap(pte); | |
251 | flush_dcache_page(new); | |
252 | } | |
a5430dda JG |
253 | } else |
254 | flush_dcache_page(new); | |
255 | ||
3ef8fd7f | 256 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 KS |
257 | if (PageHuge(new)) { |
258 | pte = pte_mkhuge(pte); | |
259 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
383321ab | 260 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
261 | if (PageAnon(new)) |
262 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
263 | else | |
264 | page_dup_rmap(new, true); | |
383321ab AK |
265 | } else |
266 | #endif | |
267 | { | |
268 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 269 | |
383321ab AK |
270 | if (PageAnon(new)) |
271 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
272 | else | |
273 | page_add_file_rmap(new, false); | |
274 | } | |
3fe87967 KS |
275 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
276 | mlock_vma_page(new); | |
277 | ||
e125fe40 KS |
278 | if (PageTransHuge(page) && PageMlocked(page)) |
279 | clear_page_mlock(page); | |
280 | ||
3fe87967 KS |
281 | /* No need to invalidate - it was non-present before */ |
282 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
283 | } | |
51afb12b | 284 | |
e4b82222 | 285 | return true; |
0697212a CL |
286 | } |
287 | ||
04e62a29 CL |
288 | /* |
289 | * Get rid of all migration entries and replace them by | |
290 | * references to the indicated page. | |
291 | */ | |
e388466d | 292 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 293 | { |
051ac83a JK |
294 | struct rmap_walk_control rwc = { |
295 | .rmap_one = remove_migration_pte, | |
296 | .arg = old, | |
297 | }; | |
298 | ||
e388466d KS |
299 | if (locked) |
300 | rmap_walk_locked(new, &rwc); | |
301 | else | |
302 | rmap_walk(new, &rwc); | |
04e62a29 CL |
303 | } |
304 | ||
0697212a CL |
305 | /* |
306 | * Something used the pte of a page under migration. We need to | |
307 | * get to the page and wait until migration is finished. | |
308 | * When we return from this function the fault will be retried. | |
0697212a | 309 | */ |
e66f17ff | 310 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 311 | spinlock_t *ptl) |
0697212a | 312 | { |
30dad309 | 313 | pte_t pte; |
0697212a CL |
314 | swp_entry_t entry; |
315 | struct page *page; | |
316 | ||
30dad309 | 317 | spin_lock(ptl); |
0697212a CL |
318 | pte = *ptep; |
319 | if (!is_swap_pte(pte)) | |
320 | goto out; | |
321 | ||
322 | entry = pte_to_swp_entry(pte); | |
323 | if (!is_migration_entry(entry)) | |
324 | goto out; | |
325 | ||
326 | page = migration_entry_to_page(entry); | |
327 | ||
e286781d | 328 | /* |
89eb946a | 329 | * Once page cache replacement of page migration started, page_count |
9a1ea439 HD |
330 | * is zero; but we must not call put_and_wait_on_page_locked() without |
331 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
e286781d NP |
332 | */ |
333 | if (!get_page_unless_zero(page)) | |
334 | goto out; | |
0697212a | 335 | pte_unmap_unlock(ptep, ptl); |
9a1ea439 | 336 | put_and_wait_on_page_locked(page); |
0697212a CL |
337 | return; |
338 | out: | |
339 | pte_unmap_unlock(ptep, ptl); | |
340 | } | |
341 | ||
30dad309 NH |
342 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
343 | unsigned long address) | |
344 | { | |
345 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
346 | pte_t *ptep = pte_offset_map(pmd, address); | |
347 | __migration_entry_wait(mm, ptep, ptl); | |
348 | } | |
349 | ||
cb900f41 KS |
350 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
351 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 352 | { |
cb900f41 | 353 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
354 | __migration_entry_wait(mm, pte, ptl); |
355 | } | |
356 | ||
616b8371 ZY |
357 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
358 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
359 | { | |
360 | spinlock_t *ptl; | |
361 | struct page *page; | |
362 | ||
363 | ptl = pmd_lock(mm, pmd); | |
364 | if (!is_pmd_migration_entry(*pmd)) | |
365 | goto unlock; | |
366 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
367 | if (!get_page_unless_zero(page)) | |
368 | goto unlock; | |
369 | spin_unlock(ptl); | |
9a1ea439 | 370 | put_and_wait_on_page_locked(page); |
616b8371 ZY |
371 | return; |
372 | unlock: | |
373 | spin_unlock(ptl); | |
374 | } | |
375 | #endif | |
376 | ||
0b3901b3 JK |
377 | static int expected_page_refs(struct page *page) |
378 | { | |
379 | int expected_count = 1; | |
380 | ||
381 | /* | |
382 | * Device public or private pages have an extra refcount as they are | |
383 | * ZONE_DEVICE pages. | |
384 | */ | |
385 | expected_count += is_device_private_page(page); | |
386 | expected_count += is_device_public_page(page); | |
387 | if (page_mapping(page)) | |
388 | expected_count += hpage_nr_pages(page) + page_has_private(page); | |
389 | ||
390 | return expected_count; | |
391 | } | |
392 | ||
b20a3503 | 393 | /* |
c3fcf8a5 | 394 | * Replace the page in the mapping. |
5b5c7120 CL |
395 | * |
396 | * The number of remaining references must be: | |
397 | * 1 for anonymous pages without a mapping | |
398 | * 2 for pages with a mapping | |
266cf658 | 399 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 400 | */ |
36bc08cc | 401 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 402 | struct page *newpage, struct page *page, |
8e321fef BL |
403 | struct buffer_head *head, enum migrate_mode mode, |
404 | int extra_count) | |
b20a3503 | 405 | { |
89eb946a | 406 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
42cb14b1 HD |
407 | struct zone *oldzone, *newzone; |
408 | int dirty; | |
0b3901b3 | 409 | int expected_count = expected_page_refs(page) + extra_count; |
8763cb45 | 410 | |
6c5240ae | 411 | if (!mapping) { |
0e8c7d0f | 412 | /* Anonymous page without mapping */ |
8e321fef | 413 | if (page_count(page) != expected_count) |
6c5240ae | 414 | return -EAGAIN; |
cf4b769a HD |
415 | |
416 | /* No turning back from here */ | |
cf4b769a HD |
417 | newpage->index = page->index; |
418 | newpage->mapping = page->mapping; | |
419 | if (PageSwapBacked(page)) | |
fa9949da | 420 | __SetPageSwapBacked(newpage); |
cf4b769a | 421 | |
78bd5209 | 422 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
423 | } |
424 | ||
42cb14b1 HD |
425 | oldzone = page_zone(page); |
426 | newzone = page_zone(newpage); | |
427 | ||
89eb946a | 428 | xas_lock_irq(&xas); |
89eb946a MW |
429 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
430 | xas_unlock_irq(&xas); | |
e23ca00b | 431 | return -EAGAIN; |
b20a3503 CL |
432 | } |
433 | ||
fe896d18 | 434 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 435 | xas_unlock_irq(&xas); |
e286781d NP |
436 | return -EAGAIN; |
437 | } | |
438 | ||
b20a3503 | 439 | /* |
cf4b769a HD |
440 | * Now we know that no one else is looking at the page: |
441 | * no turning back from here. | |
b20a3503 | 442 | */ |
cf4b769a HD |
443 | newpage->index = page->index; |
444 | newpage->mapping = page->mapping; | |
e71769ae | 445 | page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */ |
6326fec1 NP |
446 | if (PageSwapBacked(page)) { |
447 | __SetPageSwapBacked(newpage); | |
448 | if (PageSwapCache(page)) { | |
449 | SetPageSwapCache(newpage); | |
450 | set_page_private(newpage, page_private(page)); | |
451 | } | |
452 | } else { | |
453 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
454 | } |
455 | ||
42cb14b1 HD |
456 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
457 | dirty = PageDirty(page); | |
458 | if (dirty) { | |
459 | ClearPageDirty(page); | |
460 | SetPageDirty(newpage); | |
461 | } | |
462 | ||
89eb946a | 463 | xas_store(&xas, newpage); |
e71769ae NH |
464 | if (PageTransHuge(page)) { |
465 | int i; | |
e71769ae | 466 | |
013567be | 467 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
89eb946a MW |
468 | xas_next(&xas); |
469 | xas_store(&xas, newpage + i); | |
e71769ae | 470 | } |
e71769ae | 471 | } |
7cf9c2c7 NP |
472 | |
473 | /* | |
937a94c9 JG |
474 | * Drop cache reference from old page by unfreezing |
475 | * to one less reference. | |
7cf9c2c7 NP |
476 | * We know this isn't the last reference. |
477 | */ | |
e71769ae | 478 | page_ref_unfreeze(page, expected_count - hpage_nr_pages(page)); |
7cf9c2c7 | 479 | |
89eb946a | 480 | xas_unlock(&xas); |
42cb14b1 HD |
481 | /* Leave irq disabled to prevent preemption while updating stats */ |
482 | ||
0e8c7d0f CL |
483 | /* |
484 | * If moved to a different zone then also account | |
485 | * the page for that zone. Other VM counters will be | |
486 | * taken care of when we establish references to the | |
487 | * new page and drop references to the old page. | |
488 | * | |
489 | * Note that anonymous pages are accounted for | |
4b9d0fab | 490 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
491 | * are mapped to swap space. |
492 | */ | |
42cb14b1 | 493 | if (newzone != oldzone) { |
11fb9989 MG |
494 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); |
495 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); | |
42cb14b1 | 496 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
11fb9989 MG |
497 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); |
498 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); | |
42cb14b1 HD |
499 | } |
500 | if (dirty && mapping_cap_account_dirty(mapping)) { | |
11fb9989 | 501 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 502 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); |
11fb9989 | 503 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); |
5a1c84b4 | 504 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); |
42cb14b1 | 505 | } |
4b02108a | 506 | } |
42cb14b1 | 507 | local_irq_enable(); |
b20a3503 | 508 | |
78bd5209 | 509 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 510 | } |
1118dce7 | 511 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 512 | |
290408d4 NH |
513 | /* |
514 | * The expected number of remaining references is the same as that | |
515 | * of migrate_page_move_mapping(). | |
516 | */ | |
517 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
518 | struct page *newpage, struct page *page) | |
519 | { | |
89eb946a | 520 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 521 | int expected_count; |
290408d4 | 522 | |
89eb946a | 523 | xas_lock_irq(&xas); |
290408d4 | 524 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
525 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
526 | xas_unlock_irq(&xas); | |
290408d4 NH |
527 | return -EAGAIN; |
528 | } | |
529 | ||
fe896d18 | 530 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 531 | xas_unlock_irq(&xas); |
290408d4 NH |
532 | return -EAGAIN; |
533 | } | |
534 | ||
cf4b769a HD |
535 | newpage->index = page->index; |
536 | newpage->mapping = page->mapping; | |
6a93ca8f | 537 | |
290408d4 NH |
538 | get_page(newpage); |
539 | ||
89eb946a | 540 | xas_store(&xas, newpage); |
290408d4 | 541 | |
fe896d18 | 542 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 543 | |
89eb946a | 544 | xas_unlock_irq(&xas); |
6a93ca8f | 545 | |
78bd5209 | 546 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
547 | } |
548 | ||
30b0a105 DH |
549 | /* |
550 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
551 | * arithmetic will work across the entire page. We need something more | |
552 | * specialized. | |
553 | */ | |
554 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
555 | int nr_pages) | |
556 | { | |
557 | int i; | |
558 | struct page *dst_base = dst; | |
559 | struct page *src_base = src; | |
560 | ||
561 | for (i = 0; i < nr_pages; ) { | |
562 | cond_resched(); | |
563 | copy_highpage(dst, src); | |
564 | ||
565 | i++; | |
566 | dst = mem_map_next(dst, dst_base, i); | |
567 | src = mem_map_next(src, src_base, i); | |
568 | } | |
569 | } | |
570 | ||
571 | static void copy_huge_page(struct page *dst, struct page *src) | |
572 | { | |
573 | int i; | |
574 | int nr_pages; | |
575 | ||
576 | if (PageHuge(src)) { | |
577 | /* hugetlbfs page */ | |
578 | struct hstate *h = page_hstate(src); | |
579 | nr_pages = pages_per_huge_page(h); | |
580 | ||
581 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
582 | __copy_gigantic_page(dst, src, nr_pages); | |
583 | return; | |
584 | } | |
585 | } else { | |
586 | /* thp page */ | |
587 | BUG_ON(!PageTransHuge(src)); | |
588 | nr_pages = hpage_nr_pages(src); | |
589 | } | |
590 | ||
591 | for (i = 0; i < nr_pages; i++) { | |
592 | cond_resched(); | |
593 | copy_highpage(dst + i, src + i); | |
594 | } | |
595 | } | |
596 | ||
b20a3503 CL |
597 | /* |
598 | * Copy the page to its new location | |
599 | */ | |
2916ecc0 | 600 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 601 | { |
7851a45c RR |
602 | int cpupid; |
603 | ||
b20a3503 CL |
604 | if (PageError(page)) |
605 | SetPageError(newpage); | |
606 | if (PageReferenced(page)) | |
607 | SetPageReferenced(newpage); | |
608 | if (PageUptodate(page)) | |
609 | SetPageUptodate(newpage); | |
894bc310 | 610 | if (TestClearPageActive(page)) { |
309381fe | 611 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 612 | SetPageActive(newpage); |
418b27ef LS |
613 | } else if (TestClearPageUnevictable(page)) |
614 | SetPageUnevictable(newpage); | |
1899ad18 JW |
615 | if (PageWorkingset(page)) |
616 | SetPageWorkingset(newpage); | |
b20a3503 CL |
617 | if (PageChecked(page)) |
618 | SetPageChecked(newpage); | |
619 | if (PageMappedToDisk(page)) | |
620 | SetPageMappedToDisk(newpage); | |
621 | ||
42cb14b1 HD |
622 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
623 | if (PageDirty(page)) | |
624 | SetPageDirty(newpage); | |
b20a3503 | 625 | |
33c3fc71 VD |
626 | if (page_is_young(page)) |
627 | set_page_young(newpage); | |
628 | if (page_is_idle(page)) | |
629 | set_page_idle(newpage); | |
630 | ||
7851a45c RR |
631 | /* |
632 | * Copy NUMA information to the new page, to prevent over-eager | |
633 | * future migrations of this same page. | |
634 | */ | |
635 | cpupid = page_cpupid_xchg_last(page, -1); | |
636 | page_cpupid_xchg_last(newpage, cpupid); | |
637 | ||
e9995ef9 | 638 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
639 | /* |
640 | * Please do not reorder this without considering how mm/ksm.c's | |
641 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
642 | */ | |
b3b3a99c NH |
643 | if (PageSwapCache(page)) |
644 | ClearPageSwapCache(page); | |
b20a3503 CL |
645 | ClearPagePrivate(page); |
646 | set_page_private(page, 0); | |
b20a3503 CL |
647 | |
648 | /* | |
649 | * If any waiters have accumulated on the new page then | |
650 | * wake them up. | |
651 | */ | |
652 | if (PageWriteback(newpage)) | |
653 | end_page_writeback(newpage); | |
d435edca VB |
654 | |
655 | copy_page_owner(page, newpage); | |
74485cf2 JW |
656 | |
657 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 658 | } |
2916ecc0 JG |
659 | EXPORT_SYMBOL(migrate_page_states); |
660 | ||
661 | void migrate_page_copy(struct page *newpage, struct page *page) | |
662 | { | |
663 | if (PageHuge(page) || PageTransHuge(page)) | |
664 | copy_huge_page(newpage, page); | |
665 | else | |
666 | copy_highpage(newpage, page); | |
667 | ||
668 | migrate_page_states(newpage, page); | |
669 | } | |
1118dce7 | 670 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 671 | |
1d8b85cc CL |
672 | /************************************************************ |
673 | * Migration functions | |
674 | ***********************************************************/ | |
675 | ||
b20a3503 | 676 | /* |
bda807d4 | 677 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 678 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
679 | * |
680 | * Pages are locked upon entry and exit. | |
681 | */ | |
2d1db3b1 | 682 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
683 | struct page *newpage, struct page *page, |
684 | enum migrate_mode mode) | |
b20a3503 CL |
685 | { |
686 | int rc; | |
687 | ||
688 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
689 | ||
8e321fef | 690 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
b20a3503 | 691 | |
78bd5209 | 692 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
693 | return rc; |
694 | ||
2916ecc0 JG |
695 | if (mode != MIGRATE_SYNC_NO_COPY) |
696 | migrate_page_copy(newpage, page); | |
697 | else | |
698 | migrate_page_states(newpage, page); | |
78bd5209 | 699 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
700 | } |
701 | EXPORT_SYMBOL(migrate_page); | |
702 | ||
9361401e | 703 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
704 | /* Returns true if all buffers are successfully locked */ |
705 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
706 | enum migrate_mode mode) | |
707 | { | |
708 | struct buffer_head *bh = head; | |
709 | ||
710 | /* Simple case, sync compaction */ | |
711 | if (mode != MIGRATE_ASYNC) { | |
712 | do { | |
713 | get_bh(bh); | |
714 | lock_buffer(bh); | |
715 | bh = bh->b_this_page; | |
716 | ||
717 | } while (bh != head); | |
718 | ||
719 | return true; | |
720 | } | |
721 | ||
722 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
723 | do { | |
724 | get_bh(bh); | |
725 | if (!trylock_buffer(bh)) { | |
726 | /* | |
727 | * We failed to lock the buffer and cannot stall in | |
728 | * async migration. Release the taken locks | |
729 | */ | |
730 | struct buffer_head *failed_bh = bh; | |
731 | put_bh(failed_bh); | |
732 | bh = head; | |
733 | while (bh != failed_bh) { | |
734 | unlock_buffer(bh); | |
735 | put_bh(bh); | |
736 | bh = bh->b_this_page; | |
737 | } | |
738 | return false; | |
739 | } | |
740 | ||
741 | bh = bh->b_this_page; | |
742 | } while (bh != head); | |
743 | return true; | |
744 | } | |
745 | ||
1d8b85cc CL |
746 | /* |
747 | * Migration function for pages with buffers. This function can only be used | |
748 | * if the underlying filesystem guarantees that no other references to "page" | |
749 | * exist. | |
750 | */ | |
2d1db3b1 | 751 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 752 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 753 | { |
1d8b85cc CL |
754 | struct buffer_head *bh, *head; |
755 | int rc; | |
cc4f11e6 | 756 | int expected_count; |
1d8b85cc | 757 | |
1d8b85cc | 758 | if (!page_has_buffers(page)) |
a6bc32b8 | 759 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 760 | |
cc4f11e6 JK |
761 | /* Check whether page does not have extra refs before we do more work */ |
762 | expected_count = expected_page_refs(page); | |
763 | if (page_count(page) != expected_count) | |
764 | return -EAGAIN; | |
1d8b85cc | 765 | |
cc4f11e6 JK |
766 | head = page_buffers(page); |
767 | if (!buffer_migrate_lock_buffers(head, mode)) | |
768 | return -EAGAIN; | |
1d8b85cc | 769 | |
cc4f11e6 | 770 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); |
78bd5209 | 771 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 772 | goto unlock_buffers; |
1d8b85cc CL |
773 | |
774 | ClearPagePrivate(page); | |
775 | set_page_private(newpage, page_private(page)); | |
776 | set_page_private(page, 0); | |
777 | put_page(page); | |
778 | get_page(newpage); | |
779 | ||
780 | bh = head; | |
781 | do { | |
782 | set_bh_page(bh, newpage, bh_offset(bh)); | |
783 | bh = bh->b_this_page; | |
784 | ||
785 | } while (bh != head); | |
786 | ||
787 | SetPagePrivate(newpage); | |
788 | ||
2916ecc0 JG |
789 | if (mode != MIGRATE_SYNC_NO_COPY) |
790 | migrate_page_copy(newpage, page); | |
791 | else | |
792 | migrate_page_states(newpage, page); | |
1d8b85cc | 793 | |
cc4f11e6 JK |
794 | rc = MIGRATEPAGE_SUCCESS; |
795 | unlock_buffers: | |
1d8b85cc CL |
796 | bh = head; |
797 | do { | |
798 | unlock_buffer(bh); | |
2916ecc0 | 799 | put_bh(bh); |
1d8b85cc CL |
800 | bh = bh->b_this_page; |
801 | ||
802 | } while (bh != head); | |
803 | ||
cc4f11e6 | 804 | return rc; |
1d8b85cc CL |
805 | } |
806 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 807 | #endif |
1d8b85cc | 808 | |
04e62a29 CL |
809 | /* |
810 | * Writeback a page to clean the dirty state | |
811 | */ | |
812 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 813 | { |
04e62a29 CL |
814 | struct writeback_control wbc = { |
815 | .sync_mode = WB_SYNC_NONE, | |
816 | .nr_to_write = 1, | |
817 | .range_start = 0, | |
818 | .range_end = LLONG_MAX, | |
04e62a29 CL |
819 | .for_reclaim = 1 |
820 | }; | |
821 | int rc; | |
822 | ||
823 | if (!mapping->a_ops->writepage) | |
824 | /* No write method for the address space */ | |
825 | return -EINVAL; | |
826 | ||
827 | if (!clear_page_dirty_for_io(page)) | |
828 | /* Someone else already triggered a write */ | |
829 | return -EAGAIN; | |
830 | ||
8351a6e4 | 831 | /* |
04e62a29 CL |
832 | * A dirty page may imply that the underlying filesystem has |
833 | * the page on some queue. So the page must be clean for | |
834 | * migration. Writeout may mean we loose the lock and the | |
835 | * page state is no longer what we checked for earlier. | |
836 | * At this point we know that the migration attempt cannot | |
837 | * be successful. | |
8351a6e4 | 838 | */ |
e388466d | 839 | remove_migration_ptes(page, page, false); |
8351a6e4 | 840 | |
04e62a29 | 841 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 842 | |
04e62a29 CL |
843 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
844 | /* unlocked. Relock */ | |
845 | lock_page(page); | |
846 | ||
bda8550d | 847 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
848 | } |
849 | ||
850 | /* | |
851 | * Default handling if a filesystem does not provide a migration function. | |
852 | */ | |
853 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 854 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 855 | { |
b969c4ab | 856 | if (PageDirty(page)) { |
a6bc32b8 | 857 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
858 | switch (mode) { |
859 | case MIGRATE_SYNC: | |
860 | case MIGRATE_SYNC_NO_COPY: | |
861 | break; | |
862 | default: | |
b969c4ab | 863 | return -EBUSY; |
2916ecc0 | 864 | } |
04e62a29 | 865 | return writeout(mapping, page); |
b969c4ab | 866 | } |
8351a6e4 CL |
867 | |
868 | /* | |
869 | * Buffers may be managed in a filesystem specific way. | |
870 | * We must have no buffers or drop them. | |
871 | */ | |
266cf658 | 872 | if (page_has_private(page) && |
8351a6e4 CL |
873 | !try_to_release_page(page, GFP_KERNEL)) |
874 | return -EAGAIN; | |
875 | ||
a6bc32b8 | 876 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
877 | } |
878 | ||
e24f0b8f CL |
879 | /* |
880 | * Move a page to a newly allocated page | |
881 | * The page is locked and all ptes have been successfully removed. | |
882 | * | |
883 | * The new page will have replaced the old page if this function | |
884 | * is successful. | |
894bc310 LS |
885 | * |
886 | * Return value: | |
887 | * < 0 - error code | |
78bd5209 | 888 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 889 | */ |
3fe2011f | 890 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 891 | enum migrate_mode mode) |
e24f0b8f CL |
892 | { |
893 | struct address_space *mapping; | |
bda807d4 MK |
894 | int rc = -EAGAIN; |
895 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 896 | |
7db7671f HD |
897 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
898 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 899 | |
e24f0b8f | 900 | mapping = page_mapping(page); |
bda807d4 MK |
901 | |
902 | if (likely(is_lru)) { | |
903 | if (!mapping) | |
904 | rc = migrate_page(mapping, newpage, page, mode); | |
905 | else if (mapping->a_ops->migratepage) | |
906 | /* | |
907 | * Most pages have a mapping and most filesystems | |
908 | * provide a migratepage callback. Anonymous pages | |
909 | * are part of swap space which also has its own | |
910 | * migratepage callback. This is the most common path | |
911 | * for page migration. | |
912 | */ | |
913 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
914 | page, mode); | |
915 | else | |
916 | rc = fallback_migrate_page(mapping, newpage, | |
917 | page, mode); | |
918 | } else { | |
e24f0b8f | 919 | /* |
bda807d4 MK |
920 | * In case of non-lru page, it could be released after |
921 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 922 | */ |
bda807d4 MK |
923 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
924 | if (!PageMovable(page)) { | |
925 | rc = MIGRATEPAGE_SUCCESS; | |
926 | __ClearPageIsolated(page); | |
927 | goto out; | |
928 | } | |
929 | ||
930 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
931 | page, mode); | |
932 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
933 | !PageIsolated(page)); | |
934 | } | |
e24f0b8f | 935 | |
5c3f9a67 HD |
936 | /* |
937 | * When successful, old pagecache page->mapping must be cleared before | |
938 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
939 | */ | |
940 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
941 | if (__PageMovable(page)) { |
942 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
943 | ||
944 | /* | |
945 | * We clear PG_movable under page_lock so any compactor | |
946 | * cannot try to migrate this page. | |
947 | */ | |
948 | __ClearPageIsolated(page); | |
949 | } | |
950 | ||
951 | /* | |
952 | * Anonymous and movable page->mapping will be cleard by | |
953 | * free_pages_prepare so don't reset it here for keeping | |
954 | * the type to work PageAnon, for example. | |
955 | */ | |
956 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 957 | page->mapping = NULL; |
3fe2011f | 958 | } |
bda807d4 | 959 | out: |
e24f0b8f CL |
960 | return rc; |
961 | } | |
962 | ||
0dabec93 | 963 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 964 | int force, enum migrate_mode mode) |
e24f0b8f | 965 | { |
0dabec93 | 966 | int rc = -EAGAIN; |
2ebba6b7 | 967 | int page_was_mapped = 0; |
3f6c8272 | 968 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 969 | bool is_lru = !__PageMovable(page); |
95a402c3 | 970 | |
529ae9aa | 971 | if (!trylock_page(page)) { |
a6bc32b8 | 972 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 973 | goto out; |
3e7d3449 MG |
974 | |
975 | /* | |
976 | * It's not safe for direct compaction to call lock_page. | |
977 | * For example, during page readahead pages are added locked | |
978 | * to the LRU. Later, when the IO completes the pages are | |
979 | * marked uptodate and unlocked. However, the queueing | |
980 | * could be merging multiple pages for one bio (e.g. | |
981 | * mpage_readpages). If an allocation happens for the | |
982 | * second or third page, the process can end up locking | |
983 | * the same page twice and deadlocking. Rather than | |
984 | * trying to be clever about what pages can be locked, | |
985 | * avoid the use of lock_page for direct compaction | |
986 | * altogether. | |
987 | */ | |
988 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 989 | goto out; |
3e7d3449 | 990 | |
e24f0b8f CL |
991 | lock_page(page); |
992 | } | |
993 | ||
994 | if (PageWriteback(page)) { | |
11bc82d6 | 995 | /* |
fed5b64a | 996 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
997 | * necessary to wait for PageWriteback. In the async case, |
998 | * the retry loop is too short and in the sync-light case, | |
999 | * the overhead of stalling is too much | |
11bc82d6 | 1000 | */ |
2916ecc0 JG |
1001 | switch (mode) { |
1002 | case MIGRATE_SYNC: | |
1003 | case MIGRATE_SYNC_NO_COPY: | |
1004 | break; | |
1005 | default: | |
11bc82d6 | 1006 | rc = -EBUSY; |
0a31bc97 | 1007 | goto out_unlock; |
11bc82d6 AA |
1008 | } |
1009 | if (!force) | |
0a31bc97 | 1010 | goto out_unlock; |
e24f0b8f CL |
1011 | wait_on_page_writeback(page); |
1012 | } | |
03f15c86 | 1013 | |
e24f0b8f | 1014 | /* |
dc386d4d KH |
1015 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1016 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 1017 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1018 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1019 | * File Caches may use write_page() or lock_page() in migration, then, |
1020 | * just care Anon page here. | |
03f15c86 HD |
1021 | * |
1022 | * Only page_get_anon_vma() understands the subtleties of | |
1023 | * getting a hold on an anon_vma from outside one of its mms. | |
1024 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1025 | * because that implies that the anon page is no longer mapped | |
1026 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1027 | */ |
03f15c86 | 1028 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1029 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1030 | |
7db7671f HD |
1031 | /* |
1032 | * Block others from accessing the new page when we get around to | |
1033 | * establishing additional references. We are usually the only one | |
1034 | * holding a reference to newpage at this point. We used to have a BUG | |
1035 | * here if trylock_page(newpage) fails, but would like to allow for | |
1036 | * cases where there might be a race with the previous use of newpage. | |
1037 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1038 | */ | |
1039 | if (unlikely(!trylock_page(newpage))) | |
1040 | goto out_unlock; | |
1041 | ||
bda807d4 MK |
1042 | if (unlikely(!is_lru)) { |
1043 | rc = move_to_new_page(newpage, page, mode); | |
1044 | goto out_unlock_both; | |
1045 | } | |
1046 | ||
dc386d4d | 1047 | /* |
62e1c553 SL |
1048 | * Corner case handling: |
1049 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1050 | * and treated as swapcache but it has no rmap yet. | |
1051 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1052 | * trigger a BUG. So handle it here. | |
1053 | * 2. An orphaned page (see truncate_complete_page) might have | |
1054 | * fs-private metadata. The page can be picked up due to memory | |
1055 | * offlining. Everywhere else except page reclaim, the page is | |
1056 | * invisible to the vm, so the page can not be migrated. So try to | |
1057 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1058 | */ |
62e1c553 | 1059 | if (!page->mapping) { |
309381fe | 1060 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1061 | if (page_has_private(page)) { |
62e1c553 | 1062 | try_to_free_buffers(page); |
7db7671f | 1063 | goto out_unlock_both; |
62e1c553 | 1064 | } |
7db7671f HD |
1065 | } else if (page_mapped(page)) { |
1066 | /* Establish migration ptes */ | |
03f15c86 HD |
1067 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1068 | page); | |
2ebba6b7 | 1069 | try_to_unmap(page, |
da1b13cc | 1070 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
2ebba6b7 HD |
1071 | page_was_mapped = 1; |
1072 | } | |
dc386d4d | 1073 | |
e6a1530d | 1074 | if (!page_mapped(page)) |
5c3f9a67 | 1075 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1076 | |
5c3f9a67 HD |
1077 | if (page_was_mapped) |
1078 | remove_migration_ptes(page, | |
e388466d | 1079 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1080 | |
7db7671f HD |
1081 | out_unlock_both: |
1082 | unlock_page(newpage); | |
1083 | out_unlock: | |
3f6c8272 | 1084 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1085 | if (anon_vma) |
9e60109f | 1086 | put_anon_vma(anon_vma); |
e24f0b8f | 1087 | unlock_page(page); |
0dabec93 | 1088 | out: |
c6c919eb MK |
1089 | /* |
1090 | * If migration is successful, decrease refcount of the newpage | |
1091 | * which will not free the page because new page owner increased | |
1092 | * refcounter. As well, if it is LRU page, add the page to LRU | |
1093 | * list in here. | |
1094 | */ | |
1095 | if (rc == MIGRATEPAGE_SUCCESS) { | |
b1123ea6 | 1096 | if (unlikely(__PageMovable(newpage))) |
c6c919eb MK |
1097 | put_page(newpage); |
1098 | else | |
1099 | putback_lru_page(newpage); | |
1100 | } | |
1101 | ||
0dabec93 MK |
1102 | return rc; |
1103 | } | |
95a402c3 | 1104 | |
ef2a5153 GU |
1105 | /* |
1106 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
1107 | * around it. | |
1108 | */ | |
815f0ddb ND |
1109 | #if defined(CONFIG_ARM) && \ |
1110 | defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700 | |
ef2a5153 GU |
1111 | #define ICE_noinline noinline |
1112 | #else | |
1113 | #define ICE_noinline | |
1114 | #endif | |
1115 | ||
0dabec93 MK |
1116 | /* |
1117 | * Obtain the lock on page, remove all ptes and migrate the page | |
1118 | * to the newly allocated page in newpage. | |
1119 | */ | |
ef2a5153 GU |
1120 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
1121 | free_page_t put_new_page, | |
1122 | unsigned long private, struct page *page, | |
add05cec NH |
1123 | int force, enum migrate_mode mode, |
1124 | enum migrate_reason reason) | |
0dabec93 | 1125 | { |
2def7424 | 1126 | int rc = MIGRATEPAGE_SUCCESS; |
2def7424 | 1127 | struct page *newpage; |
0dabec93 | 1128 | |
94723aaf MH |
1129 | if (!thp_migration_supported() && PageTransHuge(page)) |
1130 | return -ENOMEM; | |
1131 | ||
666feb21 | 1132 | newpage = get_new_page(page, private); |
0dabec93 MK |
1133 | if (!newpage) |
1134 | return -ENOMEM; | |
1135 | ||
1136 | if (page_count(page) == 1) { | |
1137 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1138 | ClearPageActive(page); |
1139 | ClearPageUnevictable(page); | |
bda807d4 MK |
1140 | if (unlikely(__PageMovable(page))) { |
1141 | lock_page(page); | |
1142 | if (!PageMovable(page)) | |
1143 | __ClearPageIsolated(page); | |
1144 | unlock_page(page); | |
1145 | } | |
c6c919eb MK |
1146 | if (put_new_page) |
1147 | put_new_page(newpage, private); | |
1148 | else | |
1149 | put_page(newpage); | |
0dabec93 MK |
1150 | goto out; |
1151 | } | |
1152 | ||
9c620e2b | 1153 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1154 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1155 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1156 | |
0dabec93 | 1157 | out: |
e24f0b8f | 1158 | if (rc != -EAGAIN) { |
0dabec93 MK |
1159 | /* |
1160 | * A page that has been migrated has all references | |
1161 | * removed and will be freed. A page that has not been | |
1162 | * migrated will have kepts its references and be | |
1163 | * restored. | |
1164 | */ | |
1165 | list_del(&page->lru); | |
6afcf8ef ML |
1166 | |
1167 | /* | |
1168 | * Compaction can migrate also non-LRU pages which are | |
1169 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1170 | * as __PageMovable | |
1171 | */ | |
1172 | if (likely(!__PageMovable(page))) | |
e8db67eb NH |
1173 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
1174 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
c6c919eb MK |
1175 | } |
1176 | ||
1177 | /* | |
1178 | * If migration is successful, releases reference grabbed during | |
1179 | * isolation. Otherwise, restore the page to right list unless | |
1180 | * we want to retry. | |
1181 | */ | |
1182 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1183 | put_page(page); | |
1184 | if (reason == MR_MEMORY_FAILURE) { | |
d7e69488 | 1185 | /* |
c6c919eb MK |
1186 | * Set PG_HWPoison on just freed page |
1187 | * intentionally. Although it's rather weird, | |
1188 | * it's how HWPoison flag works at the moment. | |
d7e69488 | 1189 | */ |
d4ae9916 | 1190 | if (set_hwpoison_free_buddy_page(page)) |
da1b13cc | 1191 | num_poisoned_pages_inc(); |
c6c919eb MK |
1192 | } |
1193 | } else { | |
bda807d4 MK |
1194 | if (rc != -EAGAIN) { |
1195 | if (likely(!__PageMovable(page))) { | |
1196 | putback_lru_page(page); | |
1197 | goto put_new; | |
1198 | } | |
1199 | ||
1200 | lock_page(page); | |
1201 | if (PageMovable(page)) | |
1202 | putback_movable_page(page); | |
1203 | else | |
1204 | __ClearPageIsolated(page); | |
1205 | unlock_page(page); | |
1206 | put_page(page); | |
1207 | } | |
1208 | put_new: | |
c6c919eb MK |
1209 | if (put_new_page) |
1210 | put_new_page(newpage, private); | |
1211 | else | |
1212 | put_page(newpage); | |
e24f0b8f | 1213 | } |
68711a74 | 1214 | |
e24f0b8f CL |
1215 | return rc; |
1216 | } | |
1217 | ||
290408d4 NH |
1218 | /* |
1219 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1220 | * | |
1221 | * This function doesn't wait the completion of hugepage I/O | |
1222 | * because there is no race between I/O and migration for hugepage. | |
1223 | * Note that currently hugepage I/O occurs only in direct I/O | |
1224 | * where no lock is held and PG_writeback is irrelevant, | |
1225 | * and writeback status of all subpages are counted in the reference | |
1226 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1227 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1228 | * This means that when we try to migrate hugepage whose subpages are | |
1229 | * doing direct I/O, some references remain after try_to_unmap() and | |
1230 | * hugepage migration fails without data corruption. | |
1231 | * | |
1232 | * There is also no race when direct I/O is issued on the page under migration, | |
1233 | * because then pte is replaced with migration swap entry and direct I/O code | |
1234 | * will wait in the page fault for migration to complete. | |
1235 | */ | |
1236 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1237 | free_page_t put_new_page, unsigned long private, |
1238 | struct page *hpage, int force, | |
7cd12b4a | 1239 | enum migrate_mode mode, int reason) |
290408d4 | 1240 | { |
2def7424 | 1241 | int rc = -EAGAIN; |
2ebba6b7 | 1242 | int page_was_mapped = 0; |
32665f2b | 1243 | struct page *new_hpage; |
290408d4 NH |
1244 | struct anon_vma *anon_vma = NULL; |
1245 | ||
83467efb NH |
1246 | /* |
1247 | * Movability of hugepages depends on architectures and hugepage size. | |
1248 | * This check is necessary because some callers of hugepage migration | |
1249 | * like soft offline and memory hotremove don't walk through page | |
1250 | * tables or check whether the hugepage is pmd-based or not before | |
1251 | * kicking migration. | |
1252 | */ | |
100873d7 | 1253 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
32665f2b | 1254 | putback_active_hugepage(hpage); |
83467efb | 1255 | return -ENOSYS; |
32665f2b | 1256 | } |
83467efb | 1257 | |
666feb21 | 1258 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1259 | if (!new_hpage) |
1260 | return -ENOMEM; | |
1261 | ||
290408d4 | 1262 | if (!trylock_page(hpage)) { |
2916ecc0 | 1263 | if (!force) |
290408d4 | 1264 | goto out; |
2916ecc0 JG |
1265 | switch (mode) { |
1266 | case MIGRATE_SYNC: | |
1267 | case MIGRATE_SYNC_NO_COPY: | |
1268 | break; | |
1269 | default: | |
1270 | goto out; | |
1271 | } | |
290408d4 NH |
1272 | lock_page(hpage); |
1273 | } | |
1274 | ||
746b18d4 PZ |
1275 | if (PageAnon(hpage)) |
1276 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1277 | |
7db7671f HD |
1278 | if (unlikely(!trylock_page(new_hpage))) |
1279 | goto put_anon; | |
1280 | ||
2ebba6b7 HD |
1281 | if (page_mapped(hpage)) { |
1282 | try_to_unmap(hpage, | |
1283 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1284 | page_was_mapped = 1; | |
1285 | } | |
290408d4 NH |
1286 | |
1287 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1288 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1289 | |
5c3f9a67 HD |
1290 | if (page_was_mapped) |
1291 | remove_migration_ptes(hpage, | |
e388466d | 1292 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1293 | |
7db7671f HD |
1294 | unlock_page(new_hpage); |
1295 | ||
1296 | put_anon: | |
fd4a4663 | 1297 | if (anon_vma) |
9e60109f | 1298 | put_anon_vma(anon_vma); |
8e6ac7fa | 1299 | |
2def7424 | 1300 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1301 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1302 | put_new_page = NULL; |
1303 | } | |
8e6ac7fa | 1304 | |
290408d4 | 1305 | unlock_page(hpage); |
09761333 | 1306 | out: |
b8ec1cee NH |
1307 | if (rc != -EAGAIN) |
1308 | putback_active_hugepage(hpage); | |
68711a74 DR |
1309 | |
1310 | /* | |
1311 | * If migration was not successful and there's a freeing callback, use | |
1312 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1313 | * isolation. | |
1314 | */ | |
2def7424 | 1315 | if (put_new_page) |
68711a74 DR |
1316 | put_new_page(new_hpage, private); |
1317 | else | |
3aaa76e1 | 1318 | putback_active_hugepage(new_hpage); |
68711a74 | 1319 | |
290408d4 NH |
1320 | return rc; |
1321 | } | |
1322 | ||
b20a3503 | 1323 | /* |
c73e5c9c SB |
1324 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1325 | * supplied as the target for the page migration | |
b20a3503 | 1326 | * |
c73e5c9c SB |
1327 | * @from: The list of pages to be migrated. |
1328 | * @get_new_page: The function used to allocate free pages to be used | |
1329 | * as the target of the page migration. | |
68711a74 DR |
1330 | * @put_new_page: The function used to free target pages if migration |
1331 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1332 | * @private: Private data to be passed on to get_new_page() |
1333 | * @mode: The migration mode that specifies the constraints for | |
1334 | * page migration, if any. | |
1335 | * @reason: The reason for page migration. | |
b20a3503 | 1336 | * |
c73e5c9c SB |
1337 | * The function returns after 10 attempts or if no pages are movable any more |
1338 | * because the list has become empty or no retryable pages exist any more. | |
14e0f9bc | 1339 | * The caller should call putback_movable_pages() to return pages to the LRU |
28bd6578 | 1340 | * or free list only if ret != 0. |
b20a3503 | 1341 | * |
c73e5c9c | 1342 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1343 | */ |
9c620e2b | 1344 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1345 | free_page_t put_new_page, unsigned long private, |
1346 | enum migrate_mode mode, int reason) | |
b20a3503 | 1347 | { |
e24f0b8f | 1348 | int retry = 1; |
b20a3503 | 1349 | int nr_failed = 0; |
5647bc29 | 1350 | int nr_succeeded = 0; |
b20a3503 CL |
1351 | int pass = 0; |
1352 | struct page *page; | |
1353 | struct page *page2; | |
1354 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1355 | int rc; | |
1356 | ||
1357 | if (!swapwrite) | |
1358 | current->flags |= PF_SWAPWRITE; | |
1359 | ||
e24f0b8f CL |
1360 | for(pass = 0; pass < 10 && retry; pass++) { |
1361 | retry = 0; | |
b20a3503 | 1362 | |
e24f0b8f | 1363 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1364 | retry: |
e24f0b8f | 1365 | cond_resched(); |
2d1db3b1 | 1366 | |
31caf665 NH |
1367 | if (PageHuge(page)) |
1368 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1369 | put_new_page, private, page, |
7cd12b4a | 1370 | pass > 2, mode, reason); |
31caf665 | 1371 | else |
68711a74 | 1372 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec NH |
1373 | private, page, pass > 2, mode, |
1374 | reason); | |
2d1db3b1 | 1375 | |
e24f0b8f | 1376 | switch(rc) { |
95a402c3 | 1377 | case -ENOMEM: |
94723aaf MH |
1378 | /* |
1379 | * THP migration might be unsupported or the | |
1380 | * allocation could've failed so we should | |
1381 | * retry on the same page with the THP split | |
1382 | * to base pages. | |
1383 | * | |
1384 | * Head page is retried immediately and tail | |
1385 | * pages are added to the tail of the list so | |
1386 | * we encounter them after the rest of the list | |
1387 | * is processed. | |
1388 | */ | |
e6112fc3 | 1389 | if (PageTransHuge(page) && !PageHuge(page)) { |
94723aaf MH |
1390 | lock_page(page); |
1391 | rc = split_huge_page_to_list(page, from); | |
1392 | unlock_page(page); | |
1393 | if (!rc) { | |
1394 | list_safe_reset_next(page, page2, lru); | |
1395 | goto retry; | |
1396 | } | |
1397 | } | |
dfef2ef4 | 1398 | nr_failed++; |
95a402c3 | 1399 | goto out; |
e24f0b8f | 1400 | case -EAGAIN: |
2d1db3b1 | 1401 | retry++; |
e24f0b8f | 1402 | break; |
78bd5209 | 1403 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1404 | nr_succeeded++; |
e24f0b8f CL |
1405 | break; |
1406 | default: | |
354a3363 NH |
1407 | /* |
1408 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1409 | * unlike -EAGAIN case, the failed page is | |
1410 | * removed from migration page list and not | |
1411 | * retried in the next outer loop. | |
1412 | */ | |
2d1db3b1 | 1413 | nr_failed++; |
e24f0b8f | 1414 | break; |
2d1db3b1 | 1415 | } |
b20a3503 CL |
1416 | } |
1417 | } | |
f2f81fb2 VB |
1418 | nr_failed += retry; |
1419 | rc = nr_failed; | |
95a402c3 | 1420 | out: |
5647bc29 MG |
1421 | if (nr_succeeded) |
1422 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1423 | if (nr_failed) | |
1424 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1425 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1426 | ||
b20a3503 CL |
1427 | if (!swapwrite) |
1428 | current->flags &= ~PF_SWAPWRITE; | |
1429 | ||
78bd5209 | 1430 | return rc; |
b20a3503 | 1431 | } |
95a402c3 | 1432 | |
742755a1 | 1433 | #ifdef CONFIG_NUMA |
742755a1 | 1434 | |
a49bd4d7 | 1435 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1436 | { |
a49bd4d7 MH |
1437 | while (nr-- > 0) { |
1438 | if (put_user(value, status + start)) | |
1439 | return -EFAULT; | |
1440 | start++; | |
1441 | } | |
1442 | ||
1443 | return 0; | |
1444 | } | |
1445 | ||
1446 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1447 | struct list_head *pagelist, int node) | |
1448 | { | |
1449 | int err; | |
1450 | ||
1451 | if (list_empty(pagelist)) | |
1452 | return 0; | |
1453 | ||
1454 | err = migrate_pages(pagelist, alloc_new_node_page, NULL, node, | |
1455 | MIGRATE_SYNC, MR_SYSCALL); | |
1456 | if (err) | |
1457 | putback_movable_pages(pagelist); | |
1458 | return err; | |
742755a1 CL |
1459 | } |
1460 | ||
1461 | /* | |
a49bd4d7 MH |
1462 | * Resolves the given address to a struct page, isolates it from the LRU and |
1463 | * puts it to the given pagelist. | |
1464 | * Returns -errno if the page cannot be found/isolated or 0 when it has been | |
1465 | * queued or the page doesn't need to be migrated because it is already on | |
1466 | * the target node | |
742755a1 | 1467 | */ |
a49bd4d7 MH |
1468 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1469 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1470 | { |
a49bd4d7 MH |
1471 | struct vm_area_struct *vma; |
1472 | struct page *page; | |
1473 | unsigned int follflags; | |
742755a1 | 1474 | int err; |
742755a1 CL |
1475 | |
1476 | down_read(&mm->mmap_sem); | |
a49bd4d7 MH |
1477 | err = -EFAULT; |
1478 | vma = find_vma(mm, addr); | |
1479 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1480 | goto out; | |
742755a1 | 1481 | |
a49bd4d7 MH |
1482 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1483 | follflags = FOLL_GET | FOLL_DUMP; | |
a49bd4d7 | 1484 | page = follow_page(vma, addr, follflags); |
89f5b7da | 1485 | |
a49bd4d7 MH |
1486 | err = PTR_ERR(page); |
1487 | if (IS_ERR(page)) | |
1488 | goto out; | |
89f5b7da | 1489 | |
a49bd4d7 MH |
1490 | err = -ENOENT; |
1491 | if (!page) | |
1492 | goto out; | |
742755a1 | 1493 | |
a49bd4d7 MH |
1494 | err = 0; |
1495 | if (page_to_nid(page) == node) | |
1496 | goto out_putpage; | |
742755a1 | 1497 | |
a49bd4d7 MH |
1498 | err = -EACCES; |
1499 | if (page_mapcount(page) > 1 && !migrate_all) | |
1500 | goto out_putpage; | |
742755a1 | 1501 | |
a49bd4d7 MH |
1502 | if (PageHuge(page)) { |
1503 | if (PageHead(page)) { | |
1504 | isolate_huge_page(page, pagelist); | |
1505 | err = 0; | |
e632a938 | 1506 | } |
a49bd4d7 MH |
1507 | } else { |
1508 | struct page *head; | |
e632a938 | 1509 | |
e8db67eb NH |
1510 | head = compound_head(page); |
1511 | err = isolate_lru_page(head); | |
cf608ac1 | 1512 | if (err) |
a49bd4d7 | 1513 | goto out_putpage; |
742755a1 | 1514 | |
a49bd4d7 MH |
1515 | err = 0; |
1516 | list_add_tail(&head->lru, pagelist); | |
1517 | mod_node_page_state(page_pgdat(head), | |
1518 | NR_ISOLATED_ANON + page_is_file_cache(head), | |
1519 | hpage_nr_pages(head)); | |
1520 | } | |
1521 | out_putpage: | |
1522 | /* | |
1523 | * Either remove the duplicate refcount from | |
1524 | * isolate_lru_page() or drop the page ref if it was | |
1525 | * not isolated. | |
1526 | */ | |
1527 | put_page(page); | |
1528 | out: | |
742755a1 CL |
1529 | up_read(&mm->mmap_sem); |
1530 | return err; | |
1531 | } | |
1532 | ||
5e9a0f02 BG |
1533 | /* |
1534 | * Migrate an array of page address onto an array of nodes and fill | |
1535 | * the corresponding array of status. | |
1536 | */ | |
3268c63e | 1537 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1538 | unsigned long nr_pages, |
1539 | const void __user * __user *pages, | |
1540 | const int __user *nodes, | |
1541 | int __user *status, int flags) | |
1542 | { | |
a49bd4d7 MH |
1543 | int current_node = NUMA_NO_NODE; |
1544 | LIST_HEAD(pagelist); | |
1545 | int start, i; | |
1546 | int err = 0, err1; | |
35282a2d BG |
1547 | |
1548 | migrate_prep(); | |
1549 | ||
a49bd4d7 MH |
1550 | for (i = start = 0; i < nr_pages; i++) { |
1551 | const void __user *p; | |
1552 | unsigned long addr; | |
1553 | int node; | |
3140a227 | 1554 | |
a49bd4d7 MH |
1555 | err = -EFAULT; |
1556 | if (get_user(p, pages + i)) | |
1557 | goto out_flush; | |
1558 | if (get_user(node, nodes + i)) | |
1559 | goto out_flush; | |
1560 | addr = (unsigned long)p; | |
1561 | ||
1562 | err = -ENODEV; | |
1563 | if (node < 0 || node >= MAX_NUMNODES) | |
1564 | goto out_flush; | |
1565 | if (!node_state(node, N_MEMORY)) | |
1566 | goto out_flush; | |
5e9a0f02 | 1567 | |
a49bd4d7 MH |
1568 | err = -EACCES; |
1569 | if (!node_isset(node, task_nodes)) | |
1570 | goto out_flush; | |
1571 | ||
1572 | if (current_node == NUMA_NO_NODE) { | |
1573 | current_node = node; | |
1574 | start = i; | |
1575 | } else if (node != current_node) { | |
1576 | err = do_move_pages_to_node(mm, &pagelist, current_node); | |
1577 | if (err) | |
1578 | goto out; | |
1579 | err = store_status(status, start, current_node, i - start); | |
1580 | if (err) | |
1581 | goto out; | |
1582 | start = i; | |
1583 | current_node = node; | |
3140a227 BG |
1584 | } |
1585 | ||
a49bd4d7 MH |
1586 | /* |
1587 | * Errors in the page lookup or isolation are not fatal and we simply | |
1588 | * report them via status | |
1589 | */ | |
1590 | err = add_page_for_migration(mm, addr, current_node, | |
1591 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
1592 | if (!err) | |
1593 | continue; | |
3140a227 | 1594 | |
a49bd4d7 MH |
1595 | err = store_status(status, i, err, 1); |
1596 | if (err) | |
1597 | goto out_flush; | |
5e9a0f02 | 1598 | |
a49bd4d7 MH |
1599 | err = do_move_pages_to_node(mm, &pagelist, current_node); |
1600 | if (err) | |
1601 | goto out; | |
1602 | if (i > start) { | |
1603 | err = store_status(status, start, current_node, i - start); | |
1604 | if (err) | |
1605 | goto out; | |
1606 | } | |
1607 | current_node = NUMA_NO_NODE; | |
3140a227 | 1608 | } |
a49bd4d7 | 1609 | out_flush: |
8f175cf5 MH |
1610 | if (list_empty(&pagelist)) |
1611 | return err; | |
1612 | ||
a49bd4d7 MH |
1613 | /* Make sure we do not overwrite the existing error */ |
1614 | err1 = do_move_pages_to_node(mm, &pagelist, current_node); | |
1615 | if (!err1) | |
1616 | err1 = store_status(status, start, current_node, i - start); | |
1617 | if (!err) | |
1618 | err = err1; | |
5e9a0f02 BG |
1619 | out: |
1620 | return err; | |
1621 | } | |
1622 | ||
742755a1 | 1623 | /* |
2f007e74 | 1624 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1625 | */ |
80bba129 BG |
1626 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1627 | const void __user **pages, int *status) | |
742755a1 | 1628 | { |
2f007e74 | 1629 | unsigned long i; |
2f007e74 | 1630 | |
742755a1 CL |
1631 | down_read(&mm->mmap_sem); |
1632 | ||
2f007e74 | 1633 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1634 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1635 | struct vm_area_struct *vma; |
1636 | struct page *page; | |
c095adbc | 1637 | int err = -EFAULT; |
2f007e74 BG |
1638 | |
1639 | vma = find_vma(mm, addr); | |
70384dc6 | 1640 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1641 | goto set_status; |
1642 | ||
d899844e KS |
1643 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1644 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1645 | |
1646 | err = PTR_ERR(page); | |
1647 | if (IS_ERR(page)) | |
1648 | goto set_status; | |
1649 | ||
d899844e | 1650 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1651 | set_status: |
80bba129 BG |
1652 | *status = err; |
1653 | ||
1654 | pages++; | |
1655 | status++; | |
1656 | } | |
1657 | ||
1658 | up_read(&mm->mmap_sem); | |
1659 | } | |
1660 | ||
1661 | /* | |
1662 | * Determine the nodes of a user array of pages and store it in | |
1663 | * a user array of status. | |
1664 | */ | |
1665 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1666 | const void __user * __user *pages, | |
1667 | int __user *status) | |
1668 | { | |
1669 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1670 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1671 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1672 | |
87b8d1ad PA |
1673 | while (nr_pages) { |
1674 | unsigned long chunk_nr; | |
80bba129 | 1675 | |
87b8d1ad PA |
1676 | chunk_nr = nr_pages; |
1677 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1678 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1679 | ||
1680 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1681 | break; | |
80bba129 BG |
1682 | |
1683 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1684 | ||
87b8d1ad PA |
1685 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1686 | break; | |
742755a1 | 1687 | |
87b8d1ad PA |
1688 | pages += chunk_nr; |
1689 | status += chunk_nr; | |
1690 | nr_pages -= chunk_nr; | |
1691 | } | |
1692 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1693 | } |
1694 | ||
1695 | /* | |
1696 | * Move a list of pages in the address space of the currently executing | |
1697 | * process. | |
1698 | */ | |
7addf443 DB |
1699 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, |
1700 | const void __user * __user *pages, | |
1701 | const int __user *nodes, | |
1702 | int __user *status, int flags) | |
742755a1 | 1703 | { |
742755a1 | 1704 | struct task_struct *task; |
742755a1 | 1705 | struct mm_struct *mm; |
5e9a0f02 | 1706 | int err; |
3268c63e | 1707 | nodemask_t task_nodes; |
742755a1 CL |
1708 | |
1709 | /* Check flags */ | |
1710 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1711 | return -EINVAL; | |
1712 | ||
1713 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1714 | return -EPERM; | |
1715 | ||
1716 | /* Find the mm_struct */ | |
a879bf58 | 1717 | rcu_read_lock(); |
228ebcbe | 1718 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1719 | if (!task) { |
a879bf58 | 1720 | rcu_read_unlock(); |
742755a1 CL |
1721 | return -ESRCH; |
1722 | } | |
3268c63e | 1723 | get_task_struct(task); |
742755a1 CL |
1724 | |
1725 | /* | |
1726 | * Check if this process has the right to modify the specified | |
197e7e52 | 1727 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1728 | */ |
197e7e52 | 1729 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1730 | rcu_read_unlock(); |
742755a1 | 1731 | err = -EPERM; |
5e9a0f02 | 1732 | goto out; |
742755a1 | 1733 | } |
c69e8d9c | 1734 | rcu_read_unlock(); |
742755a1 | 1735 | |
86c3a764 DQ |
1736 | err = security_task_movememory(task); |
1737 | if (err) | |
5e9a0f02 | 1738 | goto out; |
86c3a764 | 1739 | |
3268c63e CL |
1740 | task_nodes = cpuset_mems_allowed(task); |
1741 | mm = get_task_mm(task); | |
1742 | put_task_struct(task); | |
1743 | ||
6e8b09ea SL |
1744 | if (!mm) |
1745 | return -EINVAL; | |
1746 | ||
1747 | if (nodes) | |
1748 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1749 | nodes, status, flags); | |
1750 | else | |
1751 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1752 | |
742755a1 CL |
1753 | mmput(mm); |
1754 | return err; | |
3268c63e CL |
1755 | |
1756 | out: | |
1757 | put_task_struct(task); | |
1758 | return err; | |
742755a1 | 1759 | } |
742755a1 | 1760 | |
7addf443 DB |
1761 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1762 | const void __user * __user *, pages, | |
1763 | const int __user *, nodes, | |
1764 | int __user *, status, int, flags) | |
1765 | { | |
1766 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1767 | } | |
1768 | ||
1769 | #ifdef CONFIG_COMPAT | |
1770 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, | |
1771 | compat_uptr_t __user *, pages32, | |
1772 | const int __user *, nodes, | |
1773 | int __user *, status, | |
1774 | int, flags) | |
1775 | { | |
1776 | const void __user * __user *pages; | |
1777 | int i; | |
1778 | ||
1779 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); | |
1780 | for (i = 0; i < nr_pages; i++) { | |
1781 | compat_uptr_t p; | |
1782 | ||
1783 | if (get_user(p, pages32 + i) || | |
1784 | put_user(compat_ptr(p), pages + i)) | |
1785 | return -EFAULT; | |
1786 | } | |
1787 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1788 | } | |
1789 | #endif /* CONFIG_COMPAT */ | |
1790 | ||
7039e1db PZ |
1791 | #ifdef CONFIG_NUMA_BALANCING |
1792 | /* | |
1793 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1794 | * pages. Currently it only checks the watermarks which crude | |
1795 | */ | |
1796 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1797 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1798 | { |
1799 | int z; | |
599d0c95 | 1800 | |
7039e1db PZ |
1801 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1802 | struct zone *zone = pgdat->node_zones + z; | |
1803 | ||
1804 | if (!populated_zone(zone)) | |
1805 | continue; | |
1806 | ||
7039e1db PZ |
1807 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1808 | if (!zone_watermark_ok(zone, 0, | |
1809 | high_wmark_pages(zone) + | |
1810 | nr_migrate_pages, | |
1811 | 0, 0)) | |
1812 | continue; | |
1813 | return true; | |
1814 | } | |
1815 | return false; | |
1816 | } | |
1817 | ||
1818 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 1819 | unsigned long data) |
7039e1db PZ |
1820 | { |
1821 | int nid = (int) data; | |
1822 | struct page *newpage; | |
1823 | ||
96db800f | 1824 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
1825 | (GFP_HIGHUSER_MOVABLE | |
1826 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1827 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 1828 | ~__GFP_RECLAIM, 0); |
bac0382c | 1829 | |
7039e1db PZ |
1830 | return newpage; |
1831 | } | |
1832 | ||
1c30e017 | 1833 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 1834 | { |
340ef390 | 1835 | int page_lru; |
a8f60772 | 1836 | |
309381fe | 1837 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 1838 | |
7039e1db | 1839 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1840 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1841 | return 0; | |
7039e1db | 1842 | |
340ef390 HD |
1843 | if (isolate_lru_page(page)) |
1844 | return 0; | |
7039e1db | 1845 | |
340ef390 HD |
1846 | /* |
1847 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1848 | * check on page_count(), so we must do it here, now that the page | |
1849 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1850 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1851 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1852 | */ | |
1853 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1854 | putback_lru_page(page); | |
1855 | return 0; | |
7039e1db PZ |
1856 | } |
1857 | ||
340ef390 | 1858 | page_lru = page_is_file_cache(page); |
599d0c95 | 1859 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
340ef390 HD |
1860 | hpage_nr_pages(page)); |
1861 | ||
149c33e1 | 1862 | /* |
340ef390 HD |
1863 | * Isolating the page has taken another reference, so the |
1864 | * caller's reference can be safely dropped without the page | |
1865 | * disappearing underneath us during migration. | |
149c33e1 MG |
1866 | */ |
1867 | put_page(page); | |
340ef390 | 1868 | return 1; |
b32967ff MG |
1869 | } |
1870 | ||
de466bd6 MG |
1871 | bool pmd_trans_migrating(pmd_t pmd) |
1872 | { | |
1873 | struct page *page = pmd_page(pmd); | |
1874 | return PageLocked(page); | |
1875 | } | |
1876 | ||
b32967ff MG |
1877 | /* |
1878 | * Attempt to migrate a misplaced page to the specified destination | |
1879 | * node. Caller is expected to have an elevated reference count on | |
1880 | * the page that will be dropped by this function before returning. | |
1881 | */ | |
1bc115d8 MG |
1882 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1883 | int node) | |
b32967ff MG |
1884 | { |
1885 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1886 | int isolated; |
b32967ff MG |
1887 | int nr_remaining; |
1888 | LIST_HEAD(migratepages); | |
1889 | ||
1890 | /* | |
1bc115d8 MG |
1891 | * Don't migrate file pages that are mapped in multiple processes |
1892 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1893 | */ |
1bc115d8 MG |
1894 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1895 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1896 | goto out; |
b32967ff | 1897 | |
09a913a7 MG |
1898 | /* |
1899 | * Also do not migrate dirty pages as not all filesystems can move | |
1900 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
1901 | */ | |
1902 | if (page_is_file_cache(page) && PageDirty(page)) | |
1903 | goto out; | |
1904 | ||
b32967ff MG |
1905 | isolated = numamigrate_isolate_page(pgdat, page); |
1906 | if (!isolated) | |
1907 | goto out; | |
1908 | ||
1909 | list_add(&page->lru, &migratepages); | |
9c620e2b | 1910 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
1911 | NULL, node, MIGRATE_ASYNC, |
1912 | MR_NUMA_MISPLACED); | |
b32967ff | 1913 | if (nr_remaining) { |
59c82b70 JK |
1914 | if (!list_empty(&migratepages)) { |
1915 | list_del(&page->lru); | |
599d0c95 | 1916 | dec_node_page_state(page, NR_ISOLATED_ANON + |
59c82b70 JK |
1917 | page_is_file_cache(page)); |
1918 | putback_lru_page(page); | |
1919 | } | |
b32967ff MG |
1920 | isolated = 0; |
1921 | } else | |
1922 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1923 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1924 | return isolated; |
340ef390 HD |
1925 | |
1926 | out: | |
1927 | put_page(page); | |
1928 | return 0; | |
7039e1db | 1929 | } |
220018d3 | 1930 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1931 | |
220018d3 | 1932 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1933 | /* |
1934 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1935 | * before returning. | |
1936 | */ | |
b32967ff MG |
1937 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1938 | struct vm_area_struct *vma, | |
1939 | pmd_t *pmd, pmd_t entry, | |
1940 | unsigned long address, | |
1941 | struct page *page, int node) | |
1942 | { | |
c4088ebd | 1943 | spinlock_t *ptl; |
b32967ff MG |
1944 | pg_data_t *pgdat = NODE_DATA(node); |
1945 | int isolated = 0; | |
1946 | struct page *new_page = NULL; | |
b32967ff | 1947 | int page_lru = page_is_file_cache(page); |
7066f0f9 | 1948 | unsigned long start = address & HPAGE_PMD_MASK; |
b32967ff | 1949 | |
b32967ff | 1950 | new_page = alloc_pages_node(node, |
25160354 | 1951 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
e97ca8e5 | 1952 | HPAGE_PMD_ORDER); |
340ef390 HD |
1953 | if (!new_page) |
1954 | goto out_fail; | |
9a982250 | 1955 | prep_transhuge_page(new_page); |
340ef390 | 1956 | |
b32967ff | 1957 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 1958 | if (!isolated) { |
b32967ff | 1959 | put_page(new_page); |
340ef390 | 1960 | goto out_fail; |
b32967ff | 1961 | } |
b0943d61 | 1962 | |
b32967ff | 1963 | /* Prepare a page as a migration target */ |
48c935ad | 1964 | __SetPageLocked(new_page); |
d44d363f SL |
1965 | if (PageSwapBacked(page)) |
1966 | __SetPageSwapBacked(new_page); | |
b32967ff MG |
1967 | |
1968 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1969 | new_page->mapping = page->mapping; | |
1970 | new_page->index = page->index; | |
7eef5f97 AA |
1971 | /* flush the cache before copying using the kernel virtual address */ |
1972 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); | |
b32967ff MG |
1973 | migrate_page_copy(new_page, page); |
1974 | WARN_ON(PageLRU(new_page)); | |
1975 | ||
1976 | /* Recheck the target PMD */ | |
c4088ebd | 1977 | ptl = pmd_lock(mm, pmd); |
f4e177d1 | 1978 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
c4088ebd | 1979 | spin_unlock(ptl); |
b32967ff MG |
1980 | |
1981 | /* Reverse changes made by migrate_page_copy() */ | |
1982 | if (TestClearPageActive(new_page)) | |
1983 | SetPageActive(page); | |
1984 | if (TestClearPageUnevictable(new_page)) | |
1985 | SetPageUnevictable(page); | |
b32967ff MG |
1986 | |
1987 | unlock_page(new_page); | |
1988 | put_page(new_page); /* Free it */ | |
1989 | ||
a54a407f MG |
1990 | /* Retake the callers reference and putback on LRU */ |
1991 | get_page(page); | |
b32967ff | 1992 | putback_lru_page(page); |
599d0c95 | 1993 | mod_node_page_state(page_pgdat(page), |
a54a407f | 1994 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
eb4489f6 MG |
1995 | |
1996 | goto out_unlock; | |
b32967ff MG |
1997 | } |
1998 | ||
10102459 | 1999 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
f55e1014 | 2000 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 2001 | |
2b4847e7 | 2002 | /* |
d7c33934 AA |
2003 | * Overwrite the old entry under pagetable lock and establish |
2004 | * the new PTE. Any parallel GUP will either observe the old | |
2005 | * page blocking on the page lock, block on the page table | |
2006 | * lock or observe the new page. The SetPageUptodate on the | |
2007 | * new page and page_add_new_anon_rmap guarantee the copy is | |
2008 | * visible before the pagetable update. | |
2b4847e7 | 2009 | */ |
7066f0f9 | 2010 | page_add_anon_rmap(new_page, vma, start, true); |
d7c33934 AA |
2011 | /* |
2012 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB | |
2013 | * has already been flushed globally. So no TLB can be currently | |
2014 | * caching this non present pmd mapping. There's no need to clear the | |
2015 | * pmd before doing set_pmd_at(), nor to flush the TLB after | |
2016 | * set_pmd_at(). Clearing the pmd here would introduce a race | |
2017 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the | |
2018 | * mmap_sem for reading. If the pmd is set to NULL at any given time, | |
2019 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this | |
2020 | * pmd. | |
2021 | */ | |
7066f0f9 | 2022 | set_pmd_at(mm, start, pmd, entry); |
ce4a9cc5 | 2023 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 | 2024 | |
f4e177d1 | 2025 | page_ref_unfreeze(page, 2); |
51afb12b | 2026 | mlock_migrate_page(new_page, page); |
d281ee61 | 2027 | page_remove_rmap(page, true); |
7cd12b4a | 2028 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2b4847e7 | 2029 | |
c4088ebd | 2030 | spin_unlock(ptl); |
b32967ff | 2031 | |
11de9927 MG |
2032 | /* Take an "isolate" reference and put new page on the LRU. */ |
2033 | get_page(new_page); | |
2034 | putback_lru_page(new_page); | |
2035 | ||
b32967ff MG |
2036 | unlock_page(new_page); |
2037 | unlock_page(page); | |
2038 | put_page(page); /* Drop the rmap reference */ | |
2039 | put_page(page); /* Drop the LRU isolation reference */ | |
2040 | ||
2041 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2042 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2043 | ||
599d0c95 | 2044 | mod_node_page_state(page_pgdat(page), |
b32967ff MG |
2045 | NR_ISOLATED_ANON + page_lru, |
2046 | -HPAGE_PMD_NR); | |
2047 | return isolated; | |
2048 | ||
340ef390 HD |
2049 | out_fail: |
2050 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2b4847e7 MG |
2051 | ptl = pmd_lock(mm, pmd); |
2052 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 2053 | entry = pmd_modify(entry, vma->vm_page_prot); |
7066f0f9 | 2054 | set_pmd_at(mm, start, pmd, entry); |
2b4847e7 MG |
2055 | update_mmu_cache_pmd(vma, address, &entry); |
2056 | } | |
2057 | spin_unlock(ptl); | |
a54a407f | 2058 | |
eb4489f6 | 2059 | out_unlock: |
340ef390 | 2060 | unlock_page(page); |
b32967ff | 2061 | put_page(page); |
b32967ff MG |
2062 | return 0; |
2063 | } | |
7039e1db PZ |
2064 | #endif /* CONFIG_NUMA_BALANCING */ |
2065 | ||
2066 | #endif /* CONFIG_NUMA */ | |
8763cb45 | 2067 | |
6b368cd4 | 2068 | #if defined(CONFIG_MIGRATE_VMA_HELPER) |
8763cb45 JG |
2069 | struct migrate_vma { |
2070 | struct vm_area_struct *vma; | |
2071 | unsigned long *dst; | |
2072 | unsigned long *src; | |
2073 | unsigned long cpages; | |
2074 | unsigned long npages; | |
2075 | unsigned long start; | |
2076 | unsigned long end; | |
2077 | }; | |
2078 | ||
2079 | static int migrate_vma_collect_hole(unsigned long start, | |
2080 | unsigned long end, | |
2081 | struct mm_walk *walk) | |
2082 | { | |
2083 | struct migrate_vma *migrate = walk->private; | |
2084 | unsigned long addr; | |
2085 | ||
8315ada7 | 2086 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
e20d103b | 2087 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8315ada7 | 2088 | migrate->dst[migrate->npages] = 0; |
e20d103b | 2089 | migrate->npages++; |
8315ada7 JG |
2090 | migrate->cpages++; |
2091 | } | |
2092 | ||
2093 | return 0; | |
2094 | } | |
2095 | ||
2096 | static int migrate_vma_collect_skip(unsigned long start, | |
2097 | unsigned long end, | |
2098 | struct mm_walk *walk) | |
2099 | { | |
2100 | struct migrate_vma *migrate = walk->private; | |
2101 | unsigned long addr; | |
2102 | ||
8763cb45 JG |
2103 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
2104 | migrate->dst[migrate->npages] = 0; | |
2105 | migrate->src[migrate->npages++] = 0; | |
2106 | } | |
2107 | ||
2108 | return 0; | |
2109 | } | |
2110 | ||
2111 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2112 | unsigned long start, | |
2113 | unsigned long end, | |
2114 | struct mm_walk *walk) | |
2115 | { | |
2116 | struct migrate_vma *migrate = walk->private; | |
2117 | struct vm_area_struct *vma = walk->vma; | |
2118 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2119 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2120 | spinlock_t *ptl; |
2121 | pte_t *ptep; | |
2122 | ||
2123 | again: | |
2124 | if (pmd_none(*pmdp)) | |
2125 | return migrate_vma_collect_hole(start, end, walk); | |
2126 | ||
2127 | if (pmd_trans_huge(*pmdp)) { | |
2128 | struct page *page; | |
2129 | ||
2130 | ptl = pmd_lock(mm, pmdp); | |
2131 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2132 | spin_unlock(ptl); | |
2133 | goto again; | |
2134 | } | |
2135 | ||
2136 | page = pmd_page(*pmdp); | |
2137 | if (is_huge_zero_page(page)) { | |
2138 | spin_unlock(ptl); | |
2139 | split_huge_pmd(vma, pmdp, addr); | |
2140 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2141 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2142 | walk); |
2143 | } else { | |
2144 | int ret; | |
2145 | ||
2146 | get_page(page); | |
2147 | spin_unlock(ptl); | |
2148 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2149 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2150 | walk); |
2151 | ret = split_huge_page(page); | |
2152 | unlock_page(page); | |
2153 | put_page(page); | |
8315ada7 JG |
2154 | if (ret) |
2155 | return migrate_vma_collect_skip(start, end, | |
2156 | walk); | |
2157 | if (pmd_none(*pmdp)) | |
8763cb45 JG |
2158 | return migrate_vma_collect_hole(start, end, |
2159 | walk); | |
2160 | } | |
2161 | } | |
2162 | ||
2163 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2164 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2165 | |
2166 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2167 | arch_enter_lazy_mmu_mode(); |
2168 | ||
8763cb45 JG |
2169 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
2170 | unsigned long mpfn, pfn; | |
2171 | struct page *page; | |
8c3328f1 | 2172 | swp_entry_t entry; |
8763cb45 JG |
2173 | pte_t pte; |
2174 | ||
2175 | pte = *ptep; | |
2176 | pfn = pte_pfn(pte); | |
2177 | ||
a5430dda | 2178 | if (pte_none(pte)) { |
8315ada7 JG |
2179 | mpfn = MIGRATE_PFN_MIGRATE; |
2180 | migrate->cpages++; | |
2181 | pfn = 0; | |
8763cb45 JG |
2182 | goto next; |
2183 | } | |
2184 | ||
a5430dda JG |
2185 | if (!pte_present(pte)) { |
2186 | mpfn = pfn = 0; | |
2187 | ||
2188 | /* | |
2189 | * Only care about unaddressable device page special | |
2190 | * page table entry. Other special swap entries are not | |
2191 | * migratable, and we ignore regular swapped page. | |
2192 | */ | |
2193 | entry = pte_to_swp_entry(pte); | |
2194 | if (!is_device_private_entry(entry)) | |
2195 | goto next; | |
2196 | ||
2197 | page = device_private_entry_to_page(entry); | |
2198 | mpfn = migrate_pfn(page_to_pfn(page))| | |
2199 | MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE; | |
2200 | if (is_write_device_private_entry(entry)) | |
2201 | mpfn |= MIGRATE_PFN_WRITE; | |
2202 | } else { | |
8315ada7 JG |
2203 | if (is_zero_pfn(pfn)) { |
2204 | mpfn = MIGRATE_PFN_MIGRATE; | |
2205 | migrate->cpages++; | |
2206 | pfn = 0; | |
2207 | goto next; | |
2208 | } | |
df6ad698 | 2209 | page = _vm_normal_page(migrate->vma, addr, pte, true); |
a5430dda JG |
2210 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2211 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2212 | } | |
2213 | ||
8763cb45 | 2214 | /* FIXME support THP */ |
8763cb45 JG |
2215 | if (!page || !page->mapping || PageTransCompound(page)) { |
2216 | mpfn = pfn = 0; | |
2217 | goto next; | |
2218 | } | |
a5430dda | 2219 | pfn = page_to_pfn(page); |
8763cb45 JG |
2220 | |
2221 | /* | |
2222 | * By getting a reference on the page we pin it and that blocks | |
2223 | * any kind of migration. Side effect is that it "freezes" the | |
2224 | * pte. | |
2225 | * | |
2226 | * We drop this reference after isolating the page from the lru | |
2227 | * for non device page (device page are not on the lru and thus | |
2228 | * can't be dropped from it). | |
2229 | */ | |
2230 | get_page(page); | |
2231 | migrate->cpages++; | |
8763cb45 | 2232 | |
8c3328f1 JG |
2233 | /* |
2234 | * Optimize for the common case where page is only mapped once | |
2235 | * in one process. If we can lock the page, then we can safely | |
2236 | * set up a special migration page table entry now. | |
2237 | */ | |
2238 | if (trylock_page(page)) { | |
2239 | pte_t swp_pte; | |
2240 | ||
2241 | mpfn |= MIGRATE_PFN_LOCKED; | |
2242 | ptep_get_and_clear(mm, addr, ptep); | |
2243 | ||
2244 | /* Setup special migration page table entry */ | |
07707125 RC |
2245 | entry = make_migration_entry(page, mpfn & |
2246 | MIGRATE_PFN_WRITE); | |
8c3328f1 JG |
2247 | swp_pte = swp_entry_to_pte(entry); |
2248 | if (pte_soft_dirty(pte)) | |
2249 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2250 | set_pte_at(mm, addr, ptep, swp_pte); | |
2251 | ||
2252 | /* | |
2253 | * This is like regular unmap: we remove the rmap and | |
2254 | * drop page refcount. Page won't be freed, as we took | |
2255 | * a reference just above. | |
2256 | */ | |
2257 | page_remove_rmap(page, false); | |
2258 | put_page(page); | |
a5430dda JG |
2259 | |
2260 | if (pte_present(pte)) | |
2261 | unmapped++; | |
8c3328f1 JG |
2262 | } |
2263 | ||
8763cb45 | 2264 | next: |
a5430dda | 2265 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2266 | migrate->src[migrate->npages++] = mpfn; |
2267 | } | |
8c3328f1 | 2268 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2269 | pte_unmap_unlock(ptep - 1, ptl); |
2270 | ||
8c3328f1 JG |
2271 | /* Only flush the TLB if we actually modified any entries */ |
2272 | if (unmapped) | |
2273 | flush_tlb_range(walk->vma, start, end); | |
2274 | ||
8763cb45 JG |
2275 | return 0; |
2276 | } | |
2277 | ||
2278 | /* | |
2279 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2280 | * @migrate: migrate struct containing all migration information | |
2281 | * | |
2282 | * This will walk the CPU page table. For each virtual address backed by a | |
2283 | * valid page, it updates the src array and takes a reference on the page, in | |
2284 | * order to pin the page until we lock it and unmap it. | |
2285 | */ | |
2286 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2287 | { | |
ac46d4f3 | 2288 | struct mmu_notifier_range range; |
8763cb45 JG |
2289 | struct mm_walk mm_walk; |
2290 | ||
2291 | mm_walk.pmd_entry = migrate_vma_collect_pmd; | |
2292 | mm_walk.pte_entry = NULL; | |
2293 | mm_walk.pte_hole = migrate_vma_collect_hole; | |
2294 | mm_walk.hugetlb_entry = NULL; | |
2295 | mm_walk.test_walk = NULL; | |
2296 | mm_walk.vma = migrate->vma; | |
2297 | mm_walk.mm = migrate->vma->vm_mm; | |
2298 | mm_walk.private = migrate; | |
2299 | ||
ac46d4f3 JG |
2300 | mmu_notifier_range_init(&range, mm_walk.mm, migrate->start, |
2301 | migrate->end); | |
2302 | mmu_notifier_invalidate_range_start(&range); | |
8763cb45 | 2303 | walk_page_range(migrate->start, migrate->end, &mm_walk); |
ac46d4f3 | 2304 | mmu_notifier_invalidate_range_end(&range); |
8763cb45 JG |
2305 | |
2306 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
2307 | } | |
2308 | ||
2309 | /* | |
2310 | * migrate_vma_check_page() - check if page is pinned or not | |
2311 | * @page: struct page to check | |
2312 | * | |
2313 | * Pinned pages cannot be migrated. This is the same test as in | |
2314 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2315 | * ZONE_DEVICE page. | |
2316 | */ | |
2317 | static bool migrate_vma_check_page(struct page *page) | |
2318 | { | |
2319 | /* | |
2320 | * One extra ref because caller holds an extra reference, either from | |
2321 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2322 | * a device page. | |
2323 | */ | |
2324 | int extra = 1; | |
2325 | ||
2326 | /* | |
2327 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2328 | * check them than regular pages, because they can be mapped with a pmd | |
2329 | * or with a pte (split pte mapping). | |
2330 | */ | |
2331 | if (PageCompound(page)) | |
2332 | return false; | |
2333 | ||
a5430dda JG |
2334 | /* Page from ZONE_DEVICE have one extra reference */ |
2335 | if (is_zone_device_page(page)) { | |
2336 | /* | |
2337 | * Private page can never be pin as they have no valid pte and | |
2338 | * GUP will fail for those. Yet if there is a pending migration | |
2339 | * a thread might try to wait on the pte migration entry and | |
2340 | * will bump the page reference count. Sadly there is no way to | |
2341 | * differentiate a regular pin from migration wait. Hence to | |
2342 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
2343 | * infinite loop (one stoping migration because the other is | |
2344 | * waiting on pte migration entry). We always return true here. | |
2345 | * | |
2346 | * FIXME proper solution is to rework migration_entry_wait() so | |
2347 | * it does not need to take a reference on page. | |
2348 | */ | |
2349 | if (is_device_private_page(page)) | |
2350 | return true; | |
2351 | ||
df6ad698 JG |
2352 | /* |
2353 | * Only allow device public page to be migrated and account for | |
2354 | * the extra reference count imply by ZONE_DEVICE pages. | |
2355 | */ | |
2356 | if (!is_device_public_page(page)) | |
2357 | return false; | |
2358 | extra++; | |
a5430dda JG |
2359 | } |
2360 | ||
df6ad698 JG |
2361 | /* For file back page */ |
2362 | if (page_mapping(page)) | |
2363 | extra += 1 + page_has_private(page); | |
2364 | ||
8763cb45 JG |
2365 | if ((page_count(page) - extra) > page_mapcount(page)) |
2366 | return false; | |
2367 | ||
2368 | return true; | |
2369 | } | |
2370 | ||
2371 | /* | |
2372 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2373 | * @migrate: migrate struct containing all migration information | |
2374 | * | |
2375 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2376 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2377 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2378 | * migrated by concurrent kernel threads. | |
2379 | */ | |
2380 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2381 | { | |
2382 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2383 | const unsigned long start = migrate->start; |
2384 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2385 | bool allow_drain = true; |
8763cb45 JG |
2386 | |
2387 | lru_add_drain(); | |
2388 | ||
2389 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2390 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2391 | bool remap = true; |
8763cb45 JG |
2392 | |
2393 | if (!page) | |
2394 | continue; | |
2395 | ||
8c3328f1 JG |
2396 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2397 | /* | |
2398 | * Because we are migrating several pages there can be | |
2399 | * a deadlock between 2 concurrent migration where each | |
2400 | * are waiting on each other page lock. | |
2401 | * | |
2402 | * Make migrate_vma() a best effort thing and backoff | |
2403 | * for any page we can not lock right away. | |
2404 | */ | |
2405 | if (!trylock_page(page)) { | |
2406 | migrate->src[i] = 0; | |
2407 | migrate->cpages--; | |
2408 | put_page(page); | |
2409 | continue; | |
2410 | } | |
2411 | remap = false; | |
2412 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2413 | } |
8763cb45 | 2414 | |
a5430dda JG |
2415 | /* ZONE_DEVICE pages are not on LRU */ |
2416 | if (!is_zone_device_page(page)) { | |
2417 | if (!PageLRU(page) && allow_drain) { | |
2418 | /* Drain CPU's pagevec */ | |
2419 | lru_add_drain_all(); | |
2420 | allow_drain = false; | |
2421 | } | |
8763cb45 | 2422 | |
a5430dda JG |
2423 | if (isolate_lru_page(page)) { |
2424 | if (remap) { | |
2425 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2426 | migrate->cpages--; | |
2427 | restore++; | |
2428 | } else { | |
2429 | migrate->src[i] = 0; | |
2430 | unlock_page(page); | |
2431 | migrate->cpages--; | |
2432 | put_page(page); | |
2433 | } | |
2434 | continue; | |
8c3328f1 | 2435 | } |
a5430dda JG |
2436 | |
2437 | /* Drop the reference we took in collect */ | |
2438 | put_page(page); | |
8763cb45 JG |
2439 | } |
2440 | ||
2441 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2442 | if (remap) { |
2443 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2444 | migrate->cpages--; | |
2445 | restore++; | |
8763cb45 | 2446 | |
a5430dda JG |
2447 | if (!is_zone_device_page(page)) { |
2448 | get_page(page); | |
2449 | putback_lru_page(page); | |
2450 | } | |
8c3328f1 JG |
2451 | } else { |
2452 | migrate->src[i] = 0; | |
2453 | unlock_page(page); | |
2454 | migrate->cpages--; | |
2455 | ||
a5430dda JG |
2456 | if (!is_zone_device_page(page)) |
2457 | putback_lru_page(page); | |
2458 | else | |
2459 | put_page(page); | |
8c3328f1 | 2460 | } |
8763cb45 JG |
2461 | } |
2462 | } | |
8c3328f1 JG |
2463 | |
2464 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2465 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2466 | ||
2467 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2468 | continue; | |
2469 | ||
2470 | remove_migration_pte(page, migrate->vma, addr, page); | |
2471 | ||
2472 | migrate->src[i] = 0; | |
2473 | unlock_page(page); | |
2474 | put_page(page); | |
2475 | restore--; | |
2476 | } | |
8763cb45 JG |
2477 | } |
2478 | ||
2479 | /* | |
2480 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2481 | * @migrate: migrate struct containing all migration information | |
2482 | * | |
2483 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2484 | * and check again if it has been pinned. Pinned pages are restored because we | |
2485 | * cannot migrate them. | |
2486 | * | |
2487 | * This is the last step before we call the device driver callback to allocate | |
2488 | * destination memory and copy contents of original page over to new page. | |
2489 | */ | |
2490 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2491 | { | |
2492 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
2493 | const unsigned long npages = migrate->npages; | |
2494 | const unsigned long start = migrate->start; | |
2495 | unsigned long addr, i, restore = 0; | |
2496 | ||
2497 | for (i = 0; i < npages; i++) { | |
2498 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2499 | ||
2500 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2501 | continue; | |
2502 | ||
8c3328f1 JG |
2503 | if (page_mapped(page)) { |
2504 | try_to_unmap(page, flags); | |
2505 | if (page_mapped(page)) | |
2506 | goto restore; | |
8763cb45 | 2507 | } |
8c3328f1 JG |
2508 | |
2509 | if (migrate_vma_check_page(page)) | |
2510 | continue; | |
2511 | ||
2512 | restore: | |
2513 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2514 | migrate->cpages--; | |
2515 | restore++; | |
8763cb45 JG |
2516 | } |
2517 | ||
2518 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2519 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2520 | ||
2521 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2522 | continue; | |
2523 | ||
2524 | remove_migration_ptes(page, page, false); | |
2525 | ||
2526 | migrate->src[i] = 0; | |
2527 | unlock_page(page); | |
2528 | restore--; | |
2529 | ||
a5430dda JG |
2530 | if (is_zone_device_page(page)) |
2531 | put_page(page); | |
2532 | else | |
2533 | putback_lru_page(page); | |
8763cb45 JG |
2534 | } |
2535 | } | |
2536 | ||
8315ada7 JG |
2537 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2538 | unsigned long addr, | |
2539 | struct page *page, | |
2540 | unsigned long *src, | |
2541 | unsigned long *dst) | |
2542 | { | |
2543 | struct vm_area_struct *vma = migrate->vma; | |
2544 | struct mm_struct *mm = vma->vm_mm; | |
2545 | struct mem_cgroup *memcg; | |
2546 | bool flush = false; | |
2547 | spinlock_t *ptl; | |
2548 | pte_t entry; | |
2549 | pgd_t *pgdp; | |
2550 | p4d_t *p4dp; | |
2551 | pud_t *pudp; | |
2552 | pmd_t *pmdp; | |
2553 | pte_t *ptep; | |
2554 | ||
2555 | /* Only allow populating anonymous memory */ | |
2556 | if (!vma_is_anonymous(vma)) | |
2557 | goto abort; | |
2558 | ||
2559 | pgdp = pgd_offset(mm, addr); | |
2560 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2561 | if (!p4dp) | |
2562 | goto abort; | |
2563 | pudp = pud_alloc(mm, p4dp, addr); | |
2564 | if (!pudp) | |
2565 | goto abort; | |
2566 | pmdp = pmd_alloc(mm, pudp, addr); | |
2567 | if (!pmdp) | |
2568 | goto abort; | |
2569 | ||
2570 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2571 | goto abort; | |
2572 | ||
2573 | /* | |
2574 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2575 | * pte_offset_map() on pmds where a huge pmd might be created | |
2576 | * from a different thread. | |
2577 | * | |
2578 | * pte_alloc_map() is safe to use under down_write(mmap_sem) or when | |
2579 | * parallel threads are excluded by other means. | |
2580 | * | |
2581 | * Here we only have down_read(mmap_sem). | |
2582 | */ | |
2583 | if (pte_alloc(mm, pmdp, addr)) | |
2584 | goto abort; | |
2585 | ||
2586 | /* See the comment in pte_alloc_one_map() */ | |
2587 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2588 | goto abort; | |
2589 | ||
2590 | if (unlikely(anon_vma_prepare(vma))) | |
2591 | goto abort; | |
2592 | if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false)) | |
2593 | goto abort; | |
2594 | ||
2595 | /* | |
2596 | * The memory barrier inside __SetPageUptodate makes sure that | |
2597 | * preceding stores to the page contents become visible before | |
2598 | * the set_pte_at() write. | |
2599 | */ | |
2600 | __SetPageUptodate(page); | |
2601 | ||
df6ad698 JG |
2602 | if (is_zone_device_page(page)) { |
2603 | if (is_device_private_page(page)) { | |
2604 | swp_entry_t swp_entry; | |
2605 | ||
2606 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2607 | entry = swp_entry_to_pte(swp_entry); | |
2608 | } else if (is_device_public_page(page)) { | |
2609 | entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); | |
2610 | if (vma->vm_flags & VM_WRITE) | |
2611 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2612 | entry = pte_mkdevmap(entry); | |
2613 | } | |
8315ada7 JG |
2614 | } else { |
2615 | entry = mk_pte(page, vma->vm_page_prot); | |
2616 | if (vma->vm_flags & VM_WRITE) | |
2617 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2618 | } | |
2619 | ||
2620 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2621 | ||
2622 | if (pte_present(*ptep)) { | |
2623 | unsigned long pfn = pte_pfn(*ptep); | |
2624 | ||
2625 | if (!is_zero_pfn(pfn)) { | |
2626 | pte_unmap_unlock(ptep, ptl); | |
2627 | mem_cgroup_cancel_charge(page, memcg, false); | |
2628 | goto abort; | |
2629 | } | |
2630 | flush = true; | |
2631 | } else if (!pte_none(*ptep)) { | |
2632 | pte_unmap_unlock(ptep, ptl); | |
2633 | mem_cgroup_cancel_charge(page, memcg, false); | |
2634 | goto abort; | |
2635 | } | |
2636 | ||
2637 | /* | |
2638 | * Check for usefaultfd but do not deliver the fault. Instead, | |
2639 | * just back off. | |
2640 | */ | |
2641 | if (userfaultfd_missing(vma)) { | |
2642 | pte_unmap_unlock(ptep, ptl); | |
2643 | mem_cgroup_cancel_charge(page, memcg, false); | |
2644 | goto abort; | |
2645 | } | |
2646 | ||
2647 | inc_mm_counter(mm, MM_ANONPAGES); | |
2648 | page_add_new_anon_rmap(page, vma, addr, false); | |
2649 | mem_cgroup_commit_charge(page, memcg, false, false); | |
2650 | if (!is_zone_device_page(page)) | |
2651 | lru_cache_add_active_or_unevictable(page, vma); | |
2652 | get_page(page); | |
2653 | ||
2654 | if (flush) { | |
2655 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2656 | ptep_clear_flush_notify(vma, addr, ptep); | |
2657 | set_pte_at_notify(mm, addr, ptep, entry); | |
2658 | update_mmu_cache(vma, addr, ptep); | |
2659 | } else { | |
2660 | /* No need to invalidate - it was non-present before */ | |
2661 | set_pte_at(mm, addr, ptep, entry); | |
2662 | update_mmu_cache(vma, addr, ptep); | |
2663 | } | |
2664 | ||
2665 | pte_unmap_unlock(ptep, ptl); | |
2666 | *src = MIGRATE_PFN_MIGRATE; | |
2667 | return; | |
2668 | ||
2669 | abort: | |
2670 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2671 | } | |
2672 | ||
8763cb45 JG |
2673 | /* |
2674 | * migrate_vma_pages() - migrate meta-data from src page to dst page | |
2675 | * @migrate: migrate struct containing all migration information | |
2676 | * | |
2677 | * This migrates struct page meta-data from source struct page to destination | |
2678 | * struct page. This effectively finishes the migration from source page to the | |
2679 | * destination page. | |
2680 | */ | |
2681 | static void migrate_vma_pages(struct migrate_vma *migrate) | |
2682 | { | |
2683 | const unsigned long npages = migrate->npages; | |
2684 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
2685 | struct mmu_notifier_range range; |
2686 | unsigned long addr, i; | |
8315ada7 | 2687 | bool notified = false; |
8763cb45 JG |
2688 | |
2689 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2690 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2691 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2692 | struct address_space *mapping; | |
2693 | int r; | |
2694 | ||
8315ada7 JG |
2695 | if (!newpage) { |
2696 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2697 | continue; |
8315ada7 JG |
2698 | } |
2699 | ||
2700 | if (!page) { | |
2701 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) { | |
2702 | continue; | |
2703 | } | |
2704 | if (!notified) { | |
8315ada7 | 2705 | notified = true; |
ac46d4f3 JG |
2706 | |
2707 | mmu_notifier_range_init(&range, | |
2708 | migrate->vma->vm_mm, | |
2709 | addr, migrate->end); | |
2710 | mmu_notifier_invalidate_range_start(&range); | |
8315ada7 JG |
2711 | } |
2712 | migrate_vma_insert_page(migrate, addr, newpage, | |
2713 | &migrate->src[i], | |
2714 | &migrate->dst[i]); | |
8763cb45 | 2715 | continue; |
8315ada7 | 2716 | } |
8763cb45 JG |
2717 | |
2718 | mapping = page_mapping(page); | |
2719 | ||
a5430dda JG |
2720 | if (is_zone_device_page(newpage)) { |
2721 | if (is_device_private_page(newpage)) { | |
2722 | /* | |
2723 | * For now only support private anonymous when | |
2724 | * migrating to un-addressable device memory. | |
2725 | */ | |
2726 | if (mapping) { | |
2727 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2728 | continue; | |
2729 | } | |
df6ad698 | 2730 | } else if (!is_device_public_page(newpage)) { |
a5430dda JG |
2731 | /* |
2732 | * Other types of ZONE_DEVICE page are not | |
2733 | * supported. | |
2734 | */ | |
2735 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2736 | continue; | |
2737 | } | |
2738 | } | |
2739 | ||
8763cb45 JG |
2740 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2741 | if (r != MIGRATEPAGE_SUCCESS) | |
2742 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2743 | } | |
8315ada7 | 2744 | |
4645b9fe JG |
2745 | /* |
2746 | * No need to double call mmu_notifier->invalidate_range() callback as | |
2747 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
2748 | * did already call it. | |
2749 | */ | |
8315ada7 | 2750 | if (notified) |
ac46d4f3 | 2751 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 JG |
2752 | } |
2753 | ||
2754 | /* | |
2755 | * migrate_vma_finalize() - restore CPU page table entry | |
2756 | * @migrate: migrate struct containing all migration information | |
2757 | * | |
2758 | * This replaces the special migration pte entry with either a mapping to the | |
2759 | * new page if migration was successful for that page, or to the original page | |
2760 | * otherwise. | |
2761 | * | |
2762 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2763 | * refcount, for device pages. | |
2764 | */ | |
2765 | static void migrate_vma_finalize(struct migrate_vma *migrate) | |
2766 | { | |
2767 | const unsigned long npages = migrate->npages; | |
2768 | unsigned long i; | |
2769 | ||
2770 | for (i = 0; i < npages; i++) { | |
2771 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2772 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2773 | ||
8315ada7 JG |
2774 | if (!page) { |
2775 | if (newpage) { | |
2776 | unlock_page(newpage); | |
2777 | put_page(newpage); | |
2778 | } | |
8763cb45 | 2779 | continue; |
8315ada7 JG |
2780 | } |
2781 | ||
8763cb45 JG |
2782 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2783 | if (newpage) { | |
2784 | unlock_page(newpage); | |
2785 | put_page(newpage); | |
2786 | } | |
2787 | newpage = page; | |
2788 | } | |
2789 | ||
2790 | remove_migration_ptes(page, newpage, false); | |
2791 | unlock_page(page); | |
2792 | migrate->cpages--; | |
2793 | ||
a5430dda JG |
2794 | if (is_zone_device_page(page)) |
2795 | put_page(page); | |
2796 | else | |
2797 | putback_lru_page(page); | |
8763cb45 JG |
2798 | |
2799 | if (newpage != page) { | |
2800 | unlock_page(newpage); | |
a5430dda JG |
2801 | if (is_zone_device_page(newpage)) |
2802 | put_page(newpage); | |
2803 | else | |
2804 | putback_lru_page(newpage); | |
8763cb45 JG |
2805 | } |
2806 | } | |
2807 | } | |
2808 | ||
2809 | /* | |
2810 | * migrate_vma() - migrate a range of memory inside vma | |
2811 | * | |
2812 | * @ops: migration callback for allocating destination memory and copying | |
2813 | * @vma: virtual memory area containing the range to be migrated | |
2814 | * @start: start address of the range to migrate (inclusive) | |
2815 | * @end: end address of the range to migrate (exclusive) | |
2816 | * @src: array of hmm_pfn_t containing source pfns | |
2817 | * @dst: array of hmm_pfn_t containing destination pfns | |
2818 | * @private: pointer passed back to each of the callback | |
2819 | * Returns: 0 on success, error code otherwise | |
2820 | * | |
2821 | * This function tries to migrate a range of memory virtual address range, using | |
2822 | * callbacks to allocate and copy memory from source to destination. First it | |
2823 | * collects all the pages backing each virtual address in the range, saving this | |
2824 | * inside the src array. Then it locks those pages and unmaps them. Once the pages | |
2825 | * are locked and unmapped, it checks whether each page is pinned or not. Pages | |
2826 | * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) | |
2827 | * in the corresponding src array entry. It then restores any pages that are | |
2828 | * pinned, by remapping and unlocking those pages. | |
2829 | * | |
2830 | * At this point it calls the alloc_and_copy() callback. For documentation on | |
2831 | * what is expected from that callback, see struct migrate_vma_ops comments in | |
2832 | * include/linux/migrate.h | |
2833 | * | |
2834 | * After the alloc_and_copy() callback, this function goes over each entry in | |
2835 | * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2836 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2837 | * then the function tries to migrate struct page information from the source | |
2838 | * struct page to the destination struct page. If it fails to migrate the struct | |
2839 | * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src | |
2840 | * array. | |
2841 | * | |
2842 | * At this point all successfully migrated pages have an entry in the src | |
2843 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2844 | * array entry with MIGRATE_PFN_VALID flag set. | |
2845 | * | |
2846 | * It then calls the finalize_and_map() callback. See comments for "struct | |
2847 | * migrate_vma_ops", in include/linux/migrate.h for details about | |
2848 | * finalize_and_map() behavior. | |
2849 | * | |
2850 | * After the finalize_and_map() callback, for successfully migrated pages, this | |
2851 | * function updates the CPU page table to point to new pages, otherwise it | |
2852 | * restores the CPU page table to point to the original source pages. | |
2853 | * | |
2854 | * Function returns 0 after the above steps, even if no pages were migrated | |
2855 | * (The function only returns an error if any of the arguments are invalid.) | |
2856 | * | |
2857 | * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT | |
2858 | * unsigned long entries. | |
2859 | */ | |
2860 | int migrate_vma(const struct migrate_vma_ops *ops, | |
2861 | struct vm_area_struct *vma, | |
2862 | unsigned long start, | |
2863 | unsigned long end, | |
2864 | unsigned long *src, | |
2865 | unsigned long *dst, | |
2866 | void *private) | |
2867 | { | |
2868 | struct migrate_vma migrate; | |
2869 | ||
2870 | /* Sanity check the arguments */ | |
2871 | start &= PAGE_MASK; | |
2872 | end &= PAGE_MASK; | |
e1fb4a08 DJ |
2873 | if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) || |
2874 | vma_is_dax(vma)) | |
8763cb45 JG |
2875 | return -EINVAL; |
2876 | if (start < vma->vm_start || start >= vma->vm_end) | |
2877 | return -EINVAL; | |
2878 | if (end <= vma->vm_start || end > vma->vm_end) | |
2879 | return -EINVAL; | |
2880 | if (!ops || !src || !dst || start >= end) | |
2881 | return -EINVAL; | |
2882 | ||
2883 | memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT)); | |
2884 | migrate.src = src; | |
2885 | migrate.dst = dst; | |
2886 | migrate.start = start; | |
2887 | migrate.npages = 0; | |
2888 | migrate.cpages = 0; | |
2889 | migrate.end = end; | |
2890 | migrate.vma = vma; | |
2891 | ||
2892 | /* Collect, and try to unmap source pages */ | |
2893 | migrate_vma_collect(&migrate); | |
2894 | if (!migrate.cpages) | |
2895 | return 0; | |
2896 | ||
2897 | /* Lock and isolate page */ | |
2898 | migrate_vma_prepare(&migrate); | |
2899 | if (!migrate.cpages) | |
2900 | return 0; | |
2901 | ||
2902 | /* Unmap pages */ | |
2903 | migrate_vma_unmap(&migrate); | |
2904 | if (!migrate.cpages) | |
2905 | return 0; | |
2906 | ||
2907 | /* | |
2908 | * At this point pages are locked and unmapped, and thus they have | |
2909 | * stable content and can safely be copied to destination memory that | |
2910 | * is allocated by the callback. | |
2911 | * | |
2912 | * Note that migration can fail in migrate_vma_struct_page() for each | |
2913 | * individual page. | |
2914 | */ | |
2915 | ops->alloc_and_copy(vma, src, dst, start, end, private); | |
2916 | ||
2917 | /* This does the real migration of struct page */ | |
2918 | migrate_vma_pages(&migrate); | |
2919 | ||
2920 | ops->finalize_and_map(vma, src, dst, start, end, private); | |
2921 | ||
2922 | /* Unlock and remap pages */ | |
2923 | migrate_vma_finalize(&migrate); | |
2924 | ||
2925 | return 0; | |
2926 | } | |
2927 | EXPORT_SYMBOL(migrate_vma); | |
6b368cd4 | 2928 | #endif /* defined(MIGRATE_VMA_HELPER) */ |