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> |
33c3fc71 | 45 | #include <linux/page_idle.h> |
d435edca | 46 | #include <linux/page_owner.h> |
6e84f315 | 47 | #include <linux/sched/mm.h> |
197e7e52 | 48 | #include <linux/ptrace.h> |
34290e2c | 49 | #include <linux/oom.h> |
884a6e5d | 50 | #include <linux/memory.h> |
ac16ec83 | 51 | #include <linux/random.h> |
c574bbe9 | 52 | #include <linux/sched/sysctl.h> |
467b171a | 53 | #include <linux/memory-tiers.h> |
b20a3503 | 54 | |
0d1836c3 MN |
55 | #include <asm/tlbflush.h> |
56 | ||
7b2a2d4a MG |
57 | #include <trace/events/migrate.h> |
58 | ||
b20a3503 CL |
59 | #include "internal.h" |
60 | ||
cd775580 | 61 | bool isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 | 62 | { |
19979497 | 63 | struct folio *folio = folio_get_nontail_page(page); |
68f2736a | 64 | const struct movable_operations *mops; |
bda807d4 MK |
65 | |
66 | /* | |
67 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
68 | * or just got freed under us. | |
69 | * | |
70 | * In case we 'win' a race for a movable page being freed under us and | |
71 | * raise its refcount preventing __free_pages() from doing its job | |
72 | * the put_page() at the end of this block will take care of | |
73 | * release this page, thus avoiding a nasty leakage. | |
74 | */ | |
19979497 | 75 | if (!folio) |
bda807d4 MK |
76 | goto out; |
77 | ||
19979497 VMO |
78 | if (unlikely(folio_test_slab(folio))) |
79 | goto out_putfolio; | |
8b881763 VB |
80 | /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */ |
81 | smp_rmb(); | |
bda807d4 | 82 | /* |
8b881763 VB |
83 | * Check movable flag before taking the page lock because |
84 | * we use non-atomic bitops on newly allocated page flags so | |
85 | * unconditionally grabbing the lock ruins page's owner side. | |
bda807d4 | 86 | */ |
19979497 VMO |
87 | if (unlikely(!__folio_test_movable(folio))) |
88 | goto out_putfolio; | |
8b881763 VB |
89 | /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */ |
90 | smp_rmb(); | |
19979497 VMO |
91 | if (unlikely(folio_test_slab(folio))) |
92 | goto out_putfolio; | |
8b881763 | 93 | |
bda807d4 MK |
94 | /* |
95 | * As movable pages are not isolated from LRU lists, concurrent | |
96 | * compaction threads can race against page migration functions | |
97 | * as well as race against the releasing a page. | |
98 | * | |
99 | * In order to avoid having an already isolated movable page | |
100 | * being (wrongly) re-isolated while it is under migration, | |
101 | * or to avoid attempting to isolate pages being released, | |
102 | * lets be sure we have the page lock | |
103 | * before proceeding with the movable page isolation steps. | |
104 | */ | |
19979497 VMO |
105 | if (unlikely(!folio_trylock(folio))) |
106 | goto out_putfolio; | |
bda807d4 | 107 | |
19979497 | 108 | if (!folio_test_movable(folio) || folio_test_isolated(folio)) |
bda807d4 MK |
109 | goto out_no_isolated; |
110 | ||
19979497 VMO |
111 | mops = folio_movable_ops(folio); |
112 | VM_BUG_ON_FOLIO(!mops, folio); | |
bda807d4 | 113 | |
19979497 | 114 | if (!mops->isolate_page(&folio->page, mode)) |
bda807d4 MK |
115 | goto out_no_isolated; |
116 | ||
117 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
19979497 VMO |
118 | WARN_ON_ONCE(folio_test_isolated(folio)); |
119 | folio_set_isolated(folio); | |
120 | folio_unlock(folio); | |
bda807d4 | 121 | |
cd775580 | 122 | return true; |
bda807d4 MK |
123 | |
124 | out_no_isolated: | |
19979497 VMO |
125 | folio_unlock(folio); |
126 | out_putfolio: | |
127 | folio_put(folio); | |
bda807d4 | 128 | out: |
cd775580 | 129 | return false; |
bda807d4 MK |
130 | } |
131 | ||
280d724a | 132 | static void putback_movable_folio(struct folio *folio) |
bda807d4 | 133 | { |
280d724a | 134 | const struct movable_operations *mops = folio_movable_ops(folio); |
bda807d4 | 135 | |
280d724a VMO |
136 | mops->putback_page(&folio->page); |
137 | folio_clear_isolated(folio); | |
bda807d4 MK |
138 | } |
139 | ||
5733c7d1 RA |
140 | /* |
141 | * Put previously isolated pages back onto the appropriate lists | |
142 | * from where they were once taken off for compaction/migration. | |
143 | * | |
59c82b70 JK |
144 | * This function shall be used whenever the isolated pageset has been |
145 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
7ce82f4c | 146 | * and isolate_hugetlb(). |
5733c7d1 RA |
147 | */ |
148 | void putback_movable_pages(struct list_head *l) | |
149 | { | |
280d724a VMO |
150 | struct folio *folio; |
151 | struct folio *folio2; | |
5733c7d1 | 152 | |
280d724a VMO |
153 | list_for_each_entry_safe(folio, folio2, l, lru) { |
154 | if (unlikely(folio_test_hugetlb(folio))) { | |
155 | folio_putback_active_hugetlb(folio); | |
31caf665 NH |
156 | continue; |
157 | } | |
280d724a | 158 | list_del(&folio->lru); |
bda807d4 | 159 | /* |
280d724a VMO |
160 | * We isolated non-lru movable folio so here we can use |
161 | * __PageMovable because LRU folio's mapping cannot have | |
bda807d4 MK |
162 | * PAGE_MAPPING_MOVABLE. |
163 | */ | |
280d724a VMO |
164 | if (unlikely(__folio_test_movable(folio))) { |
165 | VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio); | |
166 | folio_lock(folio); | |
167 | if (folio_test_movable(folio)) | |
168 | putback_movable_folio(folio); | |
bda807d4 | 169 | else |
280d724a VMO |
170 | folio_clear_isolated(folio); |
171 | folio_unlock(folio); | |
172 | folio_put(folio); | |
bda807d4 | 173 | } else { |
280d724a VMO |
174 | node_stat_mod_folio(folio, NR_ISOLATED_ANON + |
175 | folio_is_file_lru(folio), -folio_nr_pages(folio)); | |
176 | folio_putback_lru(folio); | |
bda807d4 | 177 | } |
b20a3503 | 178 | } |
b20a3503 CL |
179 | } |
180 | ||
0697212a CL |
181 | /* |
182 | * Restore a potential migration pte to a working pte entry | |
183 | */ | |
2f031c6f MWO |
184 | static bool remove_migration_pte(struct folio *folio, |
185 | struct vm_area_struct *vma, unsigned long addr, void *old) | |
0697212a | 186 | { |
4eecb8b9 | 187 | DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION); |
0697212a | 188 | |
3fe87967 | 189 | while (page_vma_mapped_walk(&pvmw)) { |
6c287605 | 190 | rmap_t rmap_flags = RMAP_NONE; |
4eecb8b9 MWO |
191 | pte_t pte; |
192 | swp_entry_t entry; | |
193 | struct page *new; | |
194 | unsigned long idx = 0; | |
195 | ||
196 | /* pgoff is invalid for ksm pages, but they are never large */ | |
197 | if (folio_test_large(folio) && !folio_test_hugetlb(folio)) | |
198 | idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; | |
199 | new = folio_page(folio, idx); | |
0697212a | 200 | |
616b8371 ZY |
201 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
202 | /* PMD-mapped THP migration entry */ | |
203 | if (!pvmw.pte) { | |
4eecb8b9 MWO |
204 | VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || |
205 | !folio_test_pmd_mappable(folio), folio); | |
616b8371 ZY |
206 | remove_migration_pmd(&pvmw, new); |
207 | continue; | |
208 | } | |
209 | #endif | |
210 | ||
4eecb8b9 | 211 | folio_get(folio); |
2e346877 | 212 | pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); |
3fe87967 KS |
213 | if (pte_swp_soft_dirty(*pvmw.pte)) |
214 | pte = pte_mksoft_dirty(pte); | |
0697212a | 215 | |
3fe87967 KS |
216 | /* |
217 | * Recheck VMA as permissions can change since migration started | |
218 | */ | |
219 | entry = pte_to_swp_entry(*pvmw.pte); | |
2e346877 PX |
220 | if (!is_migration_entry_young(entry)) |
221 | pte = pte_mkold(pte); | |
222 | if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) | |
223 | pte = pte_mkdirty(pte); | |
4dd845b5 | 224 | if (is_writable_migration_entry(entry)) |
3fe87967 | 225 | pte = maybe_mkwrite(pte, vma); |
f45ec5ff PX |
226 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
227 | pte = pte_mkuffd_wp(pte); | |
96a9c287 PX |
228 | else |
229 | pte = pte_wrprotect(pte); | |
d3cb8bf6 | 230 | |
6c287605 DH |
231 | if (folio_test_anon(folio) && !is_readable_migration_entry(entry)) |
232 | rmap_flags |= RMAP_EXCLUSIVE; | |
233 | ||
6128763f | 234 | if (unlikely(is_device_private_page(new))) { |
4dd845b5 AP |
235 | if (pte_write(pte)) |
236 | entry = make_writable_device_private_entry( | |
237 | page_to_pfn(new)); | |
238 | else | |
239 | entry = make_readable_device_private_entry( | |
240 | page_to_pfn(new)); | |
6128763f | 241 | pte = swp_entry_to_pte(entry); |
3d321bf8 RC |
242 | if (pte_swp_soft_dirty(*pvmw.pte)) |
243 | pte = pte_swp_mksoft_dirty(pte); | |
6128763f RC |
244 | if (pte_swp_uffd_wp(*pvmw.pte)) |
245 | pte = pte_swp_mkuffd_wp(pte); | |
d2b2c6dd | 246 | } |
a5430dda | 247 | |
3ef8fd7f | 248 | #ifdef CONFIG_HUGETLB_PAGE |
4eecb8b9 | 249 | if (folio_test_hugetlb(folio)) { |
79c1c594 CL |
250 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
251 | ||
3fe87967 | 252 | pte = pte_mkhuge(pte); |
79c1c594 | 253 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
4eecb8b9 | 254 | if (folio_test_anon(folio)) |
28c5209d | 255 | hugepage_add_anon_rmap(new, vma, pvmw.address, |
6c287605 | 256 | rmap_flags); |
3fe87967 | 257 | else |
fb3d824d | 258 | page_dup_file_rmap(new, true); |
1eba86c0 | 259 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab AK |
260 | } else |
261 | #endif | |
262 | { | |
4eecb8b9 | 263 | if (folio_test_anon(folio)) |
f1e2db12 | 264 | page_add_anon_rmap(new, vma, pvmw.address, |
6c287605 | 265 | rmap_flags); |
383321ab | 266 | else |
cea86fe2 | 267 | page_add_file_rmap(new, vma, false); |
1eba86c0 | 268 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab | 269 | } |
b7435507 | 270 | if (vma->vm_flags & VM_LOCKED) |
96f97c43 | 271 | mlock_drain_local(); |
e125fe40 | 272 | |
4cc79b33 AK |
273 | trace_remove_migration_pte(pvmw.address, pte_val(pte), |
274 | compound_order(new)); | |
275 | ||
3fe87967 KS |
276 | /* No need to invalidate - it was non-present before */ |
277 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
278 | } | |
51afb12b | 279 | |
e4b82222 | 280 | return true; |
0697212a CL |
281 | } |
282 | ||
04e62a29 CL |
283 | /* |
284 | * Get rid of all migration entries and replace them by | |
285 | * references to the indicated page. | |
286 | */ | |
4eecb8b9 | 287 | void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) |
04e62a29 | 288 | { |
051ac83a JK |
289 | struct rmap_walk_control rwc = { |
290 | .rmap_one = remove_migration_pte, | |
4eecb8b9 | 291 | .arg = src, |
051ac83a JK |
292 | }; |
293 | ||
e388466d | 294 | if (locked) |
2f031c6f | 295 | rmap_walk_locked(dst, &rwc); |
e388466d | 296 | else |
2f031c6f | 297 | rmap_walk(dst, &rwc); |
04e62a29 CL |
298 | } |
299 | ||
0697212a CL |
300 | /* |
301 | * Something used the pte of a page under migration. We need to | |
302 | * get to the page and wait until migration is finished. | |
303 | * When we return from this function the fault will be retried. | |
0697212a | 304 | */ |
e66f17ff | 305 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 306 | spinlock_t *ptl) |
0697212a | 307 | { |
30dad309 | 308 | pte_t pte; |
0697212a | 309 | swp_entry_t entry; |
0697212a | 310 | |
30dad309 | 311 | spin_lock(ptl); |
0697212a CL |
312 | pte = *ptep; |
313 | if (!is_swap_pte(pte)) | |
314 | goto out; | |
315 | ||
316 | entry = pte_to_swp_entry(pte); | |
317 | if (!is_migration_entry(entry)) | |
318 | goto out; | |
319 | ||
ffa65753 | 320 | migration_entry_wait_on_locked(entry, ptep, ptl); |
0697212a CL |
321 | return; |
322 | out: | |
323 | pte_unmap_unlock(ptep, ptl); | |
324 | } | |
325 | ||
30dad309 NH |
326 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
327 | unsigned long address) | |
328 | { | |
329 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
330 | pte_t *ptep = pte_offset_map(pmd, address); | |
331 | __migration_entry_wait(mm, ptep, ptl); | |
332 | } | |
333 | ||
ad1ac596 | 334 | #ifdef CONFIG_HUGETLB_PAGE |
fcd48540 PX |
335 | /* |
336 | * The vma read lock must be held upon entry. Holding that lock prevents either | |
337 | * the pte or the ptl from being freed. | |
338 | * | |
339 | * This function will release the vma lock before returning. | |
340 | */ | |
341 | void __migration_entry_wait_huge(struct vm_area_struct *vma, | |
342 | pte_t *ptep, spinlock_t *ptl) | |
30dad309 | 343 | { |
ad1ac596 ML |
344 | pte_t pte; |
345 | ||
fcd48540 | 346 | hugetlb_vma_assert_locked(vma); |
ad1ac596 ML |
347 | spin_lock(ptl); |
348 | pte = huge_ptep_get(ptep); | |
349 | ||
fcd48540 | 350 | if (unlikely(!is_hugetlb_entry_migration(pte))) { |
ad1ac596 | 351 | spin_unlock(ptl); |
fcd48540 PX |
352 | hugetlb_vma_unlock_read(vma); |
353 | } else { | |
354 | /* | |
355 | * If migration entry existed, safe to release vma lock | |
356 | * here because the pgtable page won't be freed without the | |
357 | * pgtable lock released. See comment right above pgtable | |
358 | * lock release in migration_entry_wait_on_locked(). | |
359 | */ | |
360 | hugetlb_vma_unlock_read(vma); | |
ad1ac596 | 361 | migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl); |
fcd48540 | 362 | } |
30dad309 NH |
363 | } |
364 | ||
ad1ac596 ML |
365 | void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte) |
366 | { | |
367 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte); | |
368 | ||
fcd48540 | 369 | __migration_entry_wait_huge(vma, pte, ptl); |
ad1ac596 ML |
370 | } |
371 | #endif | |
372 | ||
616b8371 ZY |
373 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
374 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
375 | { | |
376 | spinlock_t *ptl; | |
616b8371 ZY |
377 | |
378 | ptl = pmd_lock(mm, pmd); | |
379 | if (!is_pmd_migration_entry(*pmd)) | |
380 | goto unlock; | |
ffa65753 | 381 | migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); |
616b8371 ZY |
382 | return; |
383 | unlock: | |
384 | spin_unlock(ptl); | |
385 | } | |
386 | #endif | |
387 | ||
108ca835 MWO |
388 | static int folio_expected_refs(struct address_space *mapping, |
389 | struct folio *folio) | |
0b3901b3 | 390 | { |
108ca835 MWO |
391 | int refs = 1; |
392 | if (!mapping) | |
393 | return refs; | |
0b3901b3 | 394 | |
108ca835 MWO |
395 | refs += folio_nr_pages(folio); |
396 | if (folio_test_private(folio)) | |
397 | refs++; | |
398 | ||
399 | return refs; | |
0b3901b3 JK |
400 | } |
401 | ||
b20a3503 | 402 | /* |
c3fcf8a5 | 403 | * Replace the page in the mapping. |
5b5c7120 CL |
404 | * |
405 | * The number of remaining references must be: | |
406 | * 1 for anonymous pages without a mapping | |
407 | * 2 for pages with a mapping | |
266cf658 | 408 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 409 | */ |
3417013e MWO |
410 | int folio_migrate_mapping(struct address_space *mapping, |
411 | struct folio *newfolio, struct folio *folio, int extra_count) | |
b20a3503 | 412 | { |
3417013e | 413 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
42cb14b1 HD |
414 | struct zone *oldzone, *newzone; |
415 | int dirty; | |
108ca835 | 416 | int expected_count = folio_expected_refs(mapping, folio) + extra_count; |
3417013e | 417 | long nr = folio_nr_pages(folio); |
8763cb45 | 418 | |
6c5240ae | 419 | if (!mapping) { |
0e8c7d0f | 420 | /* Anonymous page without mapping */ |
3417013e | 421 | if (folio_ref_count(folio) != expected_count) |
6c5240ae | 422 | return -EAGAIN; |
cf4b769a HD |
423 | |
424 | /* No turning back from here */ | |
3417013e MWO |
425 | newfolio->index = folio->index; |
426 | newfolio->mapping = folio->mapping; | |
427 | if (folio_test_swapbacked(folio)) | |
428 | __folio_set_swapbacked(newfolio); | |
cf4b769a | 429 | |
78bd5209 | 430 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
431 | } |
432 | ||
3417013e MWO |
433 | oldzone = folio_zone(folio); |
434 | newzone = folio_zone(newfolio); | |
42cb14b1 | 435 | |
89eb946a | 436 | xas_lock_irq(&xas); |
3417013e | 437 | if (!folio_ref_freeze(folio, expected_count)) { |
89eb946a | 438 | xas_unlock_irq(&xas); |
e286781d NP |
439 | return -EAGAIN; |
440 | } | |
441 | ||
b20a3503 | 442 | /* |
3417013e | 443 | * Now we know that no one else is looking at the folio: |
cf4b769a | 444 | * no turning back from here. |
b20a3503 | 445 | */ |
3417013e MWO |
446 | newfolio->index = folio->index; |
447 | newfolio->mapping = folio->mapping; | |
448 | folio_ref_add(newfolio, nr); /* add cache reference */ | |
449 | if (folio_test_swapbacked(folio)) { | |
450 | __folio_set_swapbacked(newfolio); | |
451 | if (folio_test_swapcache(folio)) { | |
452 | folio_set_swapcache(newfolio); | |
453 | newfolio->private = folio_get_private(folio); | |
6326fec1 NP |
454 | } |
455 | } else { | |
3417013e | 456 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); |
b20a3503 CL |
457 | } |
458 | ||
42cb14b1 | 459 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
3417013e | 460 | dirty = folio_test_dirty(folio); |
42cb14b1 | 461 | if (dirty) { |
3417013e MWO |
462 | folio_clear_dirty(folio); |
463 | folio_set_dirty(newfolio); | |
42cb14b1 HD |
464 | } |
465 | ||
3417013e | 466 | xas_store(&xas, newfolio); |
7cf9c2c7 NP |
467 | |
468 | /* | |
937a94c9 JG |
469 | * Drop cache reference from old page by unfreezing |
470 | * to one less reference. | |
7cf9c2c7 NP |
471 | * We know this isn't the last reference. |
472 | */ | |
3417013e | 473 | folio_ref_unfreeze(folio, expected_count - nr); |
7cf9c2c7 | 474 | |
89eb946a | 475 | xas_unlock(&xas); |
42cb14b1 HD |
476 | /* Leave irq disabled to prevent preemption while updating stats */ |
477 | ||
0e8c7d0f CL |
478 | /* |
479 | * If moved to a different zone then also account | |
480 | * the page for that zone. Other VM counters will be | |
481 | * taken care of when we establish references to the | |
482 | * new page and drop references to the old page. | |
483 | * | |
484 | * Note that anonymous pages are accounted for | |
4b9d0fab | 485 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
486 | * are mapped to swap space. |
487 | */ | |
42cb14b1 | 488 | if (newzone != oldzone) { |
0d1c2072 JW |
489 | struct lruvec *old_lruvec, *new_lruvec; |
490 | struct mem_cgroup *memcg; | |
491 | ||
3417013e | 492 | memcg = folio_memcg(folio); |
0d1c2072 JW |
493 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); |
494 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
495 | ||
5c447d27 SB |
496 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
497 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); | |
3417013e | 498 | if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { |
5c447d27 SB |
499 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
500 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); | |
42cb14b1 | 501 | } |
b6038942 | 502 | #ifdef CONFIG_SWAP |
3417013e | 503 | if (folio_test_swapcache(folio)) { |
b6038942 SB |
504 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); |
505 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); | |
506 | } | |
507 | #endif | |
f56753ac | 508 | if (dirty && mapping_can_writeback(mapping)) { |
5c447d27 SB |
509 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
510 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); | |
511 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); | |
512 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); | |
42cb14b1 | 513 | } |
4b02108a | 514 | } |
42cb14b1 | 515 | local_irq_enable(); |
b20a3503 | 516 | |
78bd5209 | 517 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 518 | } |
3417013e | 519 | EXPORT_SYMBOL(folio_migrate_mapping); |
b20a3503 | 520 | |
290408d4 NH |
521 | /* |
522 | * The expected number of remaining references is the same as that | |
3417013e | 523 | * of folio_migrate_mapping(). |
290408d4 NH |
524 | */ |
525 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
b890ec2a | 526 | struct folio *dst, struct folio *src) |
290408d4 | 527 | { |
b890ec2a | 528 | XA_STATE(xas, &mapping->i_pages, folio_index(src)); |
290408d4 | 529 | int expected_count; |
290408d4 | 530 | |
89eb946a | 531 | xas_lock_irq(&xas); |
b890ec2a MWO |
532 | expected_count = 2 + folio_has_private(src); |
533 | if (!folio_ref_freeze(src, expected_count)) { | |
89eb946a | 534 | xas_unlock_irq(&xas); |
290408d4 NH |
535 | return -EAGAIN; |
536 | } | |
537 | ||
b890ec2a MWO |
538 | dst->index = src->index; |
539 | dst->mapping = src->mapping; | |
6a93ca8f | 540 | |
b890ec2a | 541 | folio_get(dst); |
290408d4 | 542 | |
b890ec2a | 543 | xas_store(&xas, dst); |
290408d4 | 544 | |
b890ec2a | 545 | folio_ref_unfreeze(src, expected_count - 1); |
290408d4 | 546 | |
89eb946a | 547 | xas_unlock_irq(&xas); |
6a93ca8f | 548 | |
78bd5209 | 549 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
550 | } |
551 | ||
b20a3503 | 552 | /* |
19138349 | 553 | * Copy the flags and some other ancillary information |
b20a3503 | 554 | */ |
19138349 | 555 | void folio_migrate_flags(struct folio *newfolio, struct folio *folio) |
b20a3503 | 556 | { |
7851a45c RR |
557 | int cpupid; |
558 | ||
19138349 MWO |
559 | if (folio_test_error(folio)) |
560 | folio_set_error(newfolio); | |
561 | if (folio_test_referenced(folio)) | |
562 | folio_set_referenced(newfolio); | |
563 | if (folio_test_uptodate(folio)) | |
564 | folio_mark_uptodate(newfolio); | |
565 | if (folio_test_clear_active(folio)) { | |
566 | VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); | |
567 | folio_set_active(newfolio); | |
568 | } else if (folio_test_clear_unevictable(folio)) | |
569 | folio_set_unevictable(newfolio); | |
570 | if (folio_test_workingset(folio)) | |
571 | folio_set_workingset(newfolio); | |
572 | if (folio_test_checked(folio)) | |
573 | folio_set_checked(newfolio); | |
6c287605 DH |
574 | /* |
575 | * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via | |
576 | * migration entries. We can still have PG_anon_exclusive set on an | |
577 | * effectively unmapped and unreferenced first sub-pages of an | |
578 | * anonymous THP: we can simply copy it here via PG_mappedtodisk. | |
579 | */ | |
19138349 MWO |
580 | if (folio_test_mappedtodisk(folio)) |
581 | folio_set_mappedtodisk(newfolio); | |
b20a3503 | 582 | |
3417013e | 583 | /* Move dirty on pages not done by folio_migrate_mapping() */ |
19138349 MWO |
584 | if (folio_test_dirty(folio)) |
585 | folio_set_dirty(newfolio); | |
b20a3503 | 586 | |
19138349 MWO |
587 | if (folio_test_young(folio)) |
588 | folio_set_young(newfolio); | |
589 | if (folio_test_idle(folio)) | |
590 | folio_set_idle(newfolio); | |
33c3fc71 | 591 | |
7851a45c RR |
592 | /* |
593 | * Copy NUMA information to the new page, to prevent over-eager | |
594 | * future migrations of this same page. | |
595 | */ | |
19138349 | 596 | cpupid = page_cpupid_xchg_last(&folio->page, -1); |
33024536 HY |
597 | /* |
598 | * For memory tiering mode, when migrate between slow and fast | |
599 | * memory node, reset cpupid, because that is used to record | |
600 | * page access time in slow memory node. | |
601 | */ | |
602 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { | |
603 | bool f_toptier = node_is_toptier(page_to_nid(&folio->page)); | |
604 | bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page)); | |
605 | ||
606 | if (f_toptier != t_toptier) | |
607 | cpupid = -1; | |
608 | } | |
19138349 | 609 | page_cpupid_xchg_last(&newfolio->page, cpupid); |
7851a45c | 610 | |
19138349 | 611 | folio_migrate_ksm(newfolio, folio); |
c8d6553b HD |
612 | /* |
613 | * Please do not reorder this without considering how mm/ksm.c's | |
614 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
615 | */ | |
19138349 MWO |
616 | if (folio_test_swapcache(folio)) |
617 | folio_clear_swapcache(folio); | |
618 | folio_clear_private(folio); | |
ad2fa371 MS |
619 | |
620 | /* page->private contains hugetlb specific flags */ | |
19138349 MWO |
621 | if (!folio_test_hugetlb(folio)) |
622 | folio->private = NULL; | |
b20a3503 CL |
623 | |
624 | /* | |
625 | * If any waiters have accumulated on the new page then | |
626 | * wake them up. | |
627 | */ | |
19138349 MWO |
628 | if (folio_test_writeback(newfolio)) |
629 | folio_end_writeback(newfolio); | |
d435edca | 630 | |
6aeff241 YS |
631 | /* |
632 | * PG_readahead shares the same bit with PG_reclaim. The above | |
633 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
634 | * bit after that. | |
635 | */ | |
19138349 MWO |
636 | if (folio_test_readahead(folio)) |
637 | folio_set_readahead(newfolio); | |
6aeff241 | 638 | |
19138349 | 639 | folio_copy_owner(newfolio, folio); |
74485cf2 | 640 | |
19138349 | 641 | if (!folio_test_hugetlb(folio)) |
d21bba2b | 642 | mem_cgroup_migrate(folio, newfolio); |
b20a3503 | 643 | } |
19138349 | 644 | EXPORT_SYMBOL(folio_migrate_flags); |
2916ecc0 | 645 | |
715cbfd6 | 646 | void folio_migrate_copy(struct folio *newfolio, struct folio *folio) |
2916ecc0 | 647 | { |
715cbfd6 MWO |
648 | folio_copy(newfolio, folio); |
649 | folio_migrate_flags(newfolio, folio); | |
2916ecc0 | 650 | } |
715cbfd6 | 651 | EXPORT_SYMBOL(folio_migrate_copy); |
b20a3503 | 652 | |
1d8b85cc CL |
653 | /************************************************************ |
654 | * Migration functions | |
655 | ***********************************************************/ | |
656 | ||
16ce101d AP |
657 | int migrate_folio_extra(struct address_space *mapping, struct folio *dst, |
658 | struct folio *src, enum migrate_mode mode, int extra_count) | |
659 | { | |
660 | int rc; | |
661 | ||
662 | BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ | |
663 | ||
664 | rc = folio_migrate_mapping(mapping, dst, src, extra_count); | |
665 | ||
666 | if (rc != MIGRATEPAGE_SUCCESS) | |
667 | return rc; | |
668 | ||
669 | if (mode != MIGRATE_SYNC_NO_COPY) | |
670 | folio_migrate_copy(dst, src); | |
671 | else | |
672 | folio_migrate_flags(dst, src); | |
673 | return MIGRATEPAGE_SUCCESS; | |
674 | } | |
675 | ||
54184650 MWO |
676 | /** |
677 | * migrate_folio() - Simple folio migration. | |
678 | * @mapping: The address_space containing the folio. | |
679 | * @dst: The folio to migrate the data to. | |
680 | * @src: The folio containing the current data. | |
681 | * @mode: How to migrate the page. | |
682 | * | |
683 | * Common logic to directly migrate a single LRU folio suitable for | |
684 | * folios that do not use PagePrivate/PagePrivate2. | |
b20a3503 | 685 | * |
54184650 | 686 | * Folios are locked upon entry and exit. |
b20a3503 | 687 | */ |
54184650 MWO |
688 | int migrate_folio(struct address_space *mapping, struct folio *dst, |
689 | struct folio *src, enum migrate_mode mode) | |
b20a3503 | 690 | { |
16ce101d | 691 | return migrate_folio_extra(mapping, dst, src, mode, 0); |
b20a3503 | 692 | } |
54184650 | 693 | EXPORT_SYMBOL(migrate_folio); |
b20a3503 | 694 | |
9361401e | 695 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
696 | /* Returns true if all buffers are successfully locked */ |
697 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
698 | enum migrate_mode mode) | |
699 | { | |
700 | struct buffer_head *bh = head; | |
701 | ||
702 | /* Simple case, sync compaction */ | |
703 | if (mode != MIGRATE_ASYNC) { | |
704 | do { | |
84ade7c1 JK |
705 | lock_buffer(bh); |
706 | bh = bh->b_this_page; | |
707 | ||
708 | } while (bh != head); | |
709 | ||
710 | return true; | |
711 | } | |
712 | ||
713 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
714 | do { | |
84ade7c1 JK |
715 | if (!trylock_buffer(bh)) { |
716 | /* | |
717 | * We failed to lock the buffer and cannot stall in | |
718 | * async migration. Release the taken locks | |
719 | */ | |
720 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
721 | bh = head; |
722 | while (bh != failed_bh) { | |
723 | unlock_buffer(bh); | |
84ade7c1 JK |
724 | bh = bh->b_this_page; |
725 | } | |
726 | return false; | |
727 | } | |
728 | ||
729 | bh = bh->b_this_page; | |
730 | } while (bh != head); | |
731 | return true; | |
732 | } | |
733 | ||
67235182 MWO |
734 | static int __buffer_migrate_folio(struct address_space *mapping, |
735 | struct folio *dst, struct folio *src, enum migrate_mode mode, | |
89cb0888 | 736 | bool check_refs) |
1d8b85cc | 737 | { |
1d8b85cc CL |
738 | struct buffer_head *bh, *head; |
739 | int rc; | |
cc4f11e6 | 740 | int expected_count; |
1d8b85cc | 741 | |
67235182 MWO |
742 | head = folio_buffers(src); |
743 | if (!head) | |
54184650 | 744 | return migrate_folio(mapping, dst, src, mode); |
1d8b85cc | 745 | |
cc4f11e6 | 746 | /* Check whether page does not have extra refs before we do more work */ |
108ca835 | 747 | expected_count = folio_expected_refs(mapping, src); |
67235182 | 748 | if (folio_ref_count(src) != expected_count) |
cc4f11e6 | 749 | return -EAGAIN; |
1d8b85cc | 750 | |
cc4f11e6 JK |
751 | if (!buffer_migrate_lock_buffers(head, mode)) |
752 | return -EAGAIN; | |
1d8b85cc | 753 | |
89cb0888 JK |
754 | if (check_refs) { |
755 | bool busy; | |
756 | bool invalidated = false; | |
757 | ||
758 | recheck_buffers: | |
759 | busy = false; | |
760 | spin_lock(&mapping->private_lock); | |
761 | bh = head; | |
762 | do { | |
763 | if (atomic_read(&bh->b_count)) { | |
764 | busy = true; | |
765 | break; | |
766 | } | |
767 | bh = bh->b_this_page; | |
768 | } while (bh != head); | |
89cb0888 JK |
769 | if (busy) { |
770 | if (invalidated) { | |
771 | rc = -EAGAIN; | |
772 | goto unlock_buffers; | |
773 | } | |
ebdf4de5 | 774 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
775 | invalidate_bh_lrus(); |
776 | invalidated = true; | |
777 | goto recheck_buffers; | |
778 | } | |
779 | } | |
780 | ||
67235182 | 781 | rc = folio_migrate_mapping(mapping, dst, src, 0); |
78bd5209 | 782 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 783 | goto unlock_buffers; |
1d8b85cc | 784 | |
67235182 | 785 | folio_attach_private(dst, folio_detach_private(src)); |
1d8b85cc CL |
786 | |
787 | bh = head; | |
788 | do { | |
67235182 | 789 | set_bh_page(bh, &dst->page, bh_offset(bh)); |
1d8b85cc | 790 | bh = bh->b_this_page; |
1d8b85cc CL |
791 | } while (bh != head); |
792 | ||
2916ecc0 | 793 | if (mode != MIGRATE_SYNC_NO_COPY) |
67235182 | 794 | folio_migrate_copy(dst, src); |
2916ecc0 | 795 | else |
67235182 | 796 | folio_migrate_flags(dst, src); |
1d8b85cc | 797 | |
cc4f11e6 JK |
798 | rc = MIGRATEPAGE_SUCCESS; |
799 | unlock_buffers: | |
ebdf4de5 JK |
800 | if (check_refs) |
801 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
802 | bh = head; |
803 | do { | |
804 | unlock_buffer(bh); | |
1d8b85cc | 805 | bh = bh->b_this_page; |
1d8b85cc CL |
806 | } while (bh != head); |
807 | ||
cc4f11e6 | 808 | return rc; |
1d8b85cc | 809 | } |
89cb0888 | 810 | |
67235182 MWO |
811 | /** |
812 | * buffer_migrate_folio() - Migration function for folios with buffers. | |
813 | * @mapping: The address space containing @src. | |
814 | * @dst: The folio to migrate to. | |
815 | * @src: The folio to migrate from. | |
816 | * @mode: How to migrate the folio. | |
817 | * | |
818 | * This function can only be used if the underlying filesystem guarantees | |
819 | * that no other references to @src exist. For example attached buffer | |
820 | * heads are accessed only under the folio lock. If your filesystem cannot | |
821 | * provide this guarantee, buffer_migrate_folio_norefs() may be more | |
822 | * appropriate. | |
823 | * | |
824 | * Return: 0 on success or a negative errno on failure. | |
89cb0888 | 825 | */ |
67235182 MWO |
826 | int buffer_migrate_folio(struct address_space *mapping, |
827 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
89cb0888 | 828 | { |
67235182 | 829 | return __buffer_migrate_folio(mapping, dst, src, mode, false); |
89cb0888 | 830 | } |
67235182 MWO |
831 | EXPORT_SYMBOL(buffer_migrate_folio); |
832 | ||
833 | /** | |
834 | * buffer_migrate_folio_norefs() - Migration function for folios with buffers. | |
835 | * @mapping: The address space containing @src. | |
836 | * @dst: The folio to migrate to. | |
837 | * @src: The folio to migrate from. | |
838 | * @mode: How to migrate the folio. | |
839 | * | |
840 | * Like buffer_migrate_folio() except that this variant is more careful | |
841 | * and checks that there are also no buffer head references. This function | |
842 | * is the right one for mappings where buffer heads are directly looked | |
843 | * up and referenced (such as block device mappings). | |
844 | * | |
845 | * Return: 0 on success or a negative errno on failure. | |
89cb0888 | 846 | */ |
67235182 MWO |
847 | int buffer_migrate_folio_norefs(struct address_space *mapping, |
848 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
89cb0888 | 849 | { |
67235182 | 850 | return __buffer_migrate_folio(mapping, dst, src, mode, true); |
89cb0888 | 851 | } |
e26355e2 | 852 | EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); |
9361401e | 853 | #endif |
1d8b85cc | 854 | |
2ec810d5 MWO |
855 | int filemap_migrate_folio(struct address_space *mapping, |
856 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
857 | { | |
858 | int ret; | |
859 | ||
860 | ret = folio_migrate_mapping(mapping, dst, src, 0); | |
861 | if (ret != MIGRATEPAGE_SUCCESS) | |
862 | return ret; | |
863 | ||
864 | if (folio_get_private(src)) | |
865 | folio_attach_private(dst, folio_detach_private(src)); | |
866 | ||
867 | if (mode != MIGRATE_SYNC_NO_COPY) | |
868 | folio_migrate_copy(dst, src); | |
869 | else | |
870 | folio_migrate_flags(dst, src); | |
871 | return MIGRATEPAGE_SUCCESS; | |
872 | } | |
873 | EXPORT_SYMBOL_GPL(filemap_migrate_folio); | |
874 | ||
04e62a29 | 875 | /* |
2be7fa10 | 876 | * Writeback a folio to clean the dirty state |
04e62a29 | 877 | */ |
2be7fa10 | 878 | static int writeout(struct address_space *mapping, struct folio *folio) |
8351a6e4 | 879 | { |
04e62a29 CL |
880 | struct writeback_control wbc = { |
881 | .sync_mode = WB_SYNC_NONE, | |
882 | .nr_to_write = 1, | |
883 | .range_start = 0, | |
884 | .range_end = LLONG_MAX, | |
04e62a29 CL |
885 | .for_reclaim = 1 |
886 | }; | |
887 | int rc; | |
888 | ||
889 | if (!mapping->a_ops->writepage) | |
890 | /* No write method for the address space */ | |
891 | return -EINVAL; | |
892 | ||
2be7fa10 | 893 | if (!folio_clear_dirty_for_io(folio)) |
04e62a29 CL |
894 | /* Someone else already triggered a write */ |
895 | return -EAGAIN; | |
896 | ||
8351a6e4 | 897 | /* |
2be7fa10 MWO |
898 | * A dirty folio may imply that the underlying filesystem has |
899 | * the folio on some queue. So the folio must be clean for | |
900 | * migration. Writeout may mean we lose the lock and the | |
901 | * folio state is no longer what we checked for earlier. | |
04e62a29 CL |
902 | * At this point we know that the migration attempt cannot |
903 | * be successful. | |
8351a6e4 | 904 | */ |
4eecb8b9 | 905 | remove_migration_ptes(folio, folio, false); |
8351a6e4 | 906 | |
2be7fa10 | 907 | rc = mapping->a_ops->writepage(&folio->page, &wbc); |
8351a6e4 | 908 | |
04e62a29 CL |
909 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
910 | /* unlocked. Relock */ | |
2be7fa10 | 911 | folio_lock(folio); |
04e62a29 | 912 | |
bda8550d | 913 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
914 | } |
915 | ||
916 | /* | |
917 | * Default handling if a filesystem does not provide a migration function. | |
918 | */ | |
8faa8ef5 MWO |
919 | static int fallback_migrate_folio(struct address_space *mapping, |
920 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
04e62a29 | 921 | { |
8faa8ef5 MWO |
922 | if (folio_test_dirty(src)) { |
923 | /* Only writeback folios in full synchronous migration */ | |
2916ecc0 JG |
924 | switch (mode) { |
925 | case MIGRATE_SYNC: | |
926 | case MIGRATE_SYNC_NO_COPY: | |
927 | break; | |
928 | default: | |
b969c4ab | 929 | return -EBUSY; |
2916ecc0 | 930 | } |
2be7fa10 | 931 | return writeout(mapping, src); |
b969c4ab | 932 | } |
8351a6e4 CL |
933 | |
934 | /* | |
935 | * Buffers may be managed in a filesystem specific way. | |
936 | * We must have no buffers or drop them. | |
937 | */ | |
8faa8ef5 MWO |
938 | if (folio_test_private(src) && |
939 | !filemap_release_folio(src, GFP_KERNEL)) | |
806031bb | 940 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 941 | |
54184650 | 942 | return migrate_folio(mapping, dst, src, mode); |
8351a6e4 CL |
943 | } |
944 | ||
e24f0b8f CL |
945 | /* |
946 | * Move a page to a newly allocated page | |
947 | * The page is locked and all ptes have been successfully removed. | |
948 | * | |
949 | * The new page will have replaced the old page if this function | |
950 | * is successful. | |
894bc310 LS |
951 | * |
952 | * Return value: | |
953 | * < 0 - error code | |
78bd5209 | 954 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 955 | */ |
e7e3ffeb | 956 | static int move_to_new_folio(struct folio *dst, struct folio *src, |
5c3f9a67 | 957 | enum migrate_mode mode) |
e24f0b8f | 958 | { |
bda807d4 | 959 | int rc = -EAGAIN; |
e7e3ffeb | 960 | bool is_lru = !__PageMovable(&src->page); |
e24f0b8f | 961 | |
e7e3ffeb MWO |
962 | VM_BUG_ON_FOLIO(!folio_test_locked(src), src); |
963 | VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); | |
e24f0b8f | 964 | |
bda807d4 | 965 | if (likely(is_lru)) { |
68f2736a MWO |
966 | struct address_space *mapping = folio_mapping(src); |
967 | ||
bda807d4 | 968 | if (!mapping) |
54184650 | 969 | rc = migrate_folio(mapping, dst, src, mode); |
5490da4f | 970 | else if (mapping->a_ops->migrate_folio) |
bda807d4 | 971 | /* |
5490da4f MWO |
972 | * Most folios have a mapping and most filesystems |
973 | * provide a migrate_folio callback. Anonymous folios | |
bda807d4 | 974 | * are part of swap space which also has its own |
5490da4f | 975 | * migrate_folio callback. This is the most common path |
bda807d4 MK |
976 | * for page migration. |
977 | */ | |
5490da4f MWO |
978 | rc = mapping->a_ops->migrate_folio(mapping, dst, src, |
979 | mode); | |
bda807d4 | 980 | else |
8faa8ef5 | 981 | rc = fallback_migrate_folio(mapping, dst, src, mode); |
bda807d4 | 982 | } else { |
68f2736a MWO |
983 | const struct movable_operations *mops; |
984 | ||
e24f0b8f | 985 | /* |
bda807d4 MK |
986 | * In case of non-lru page, it could be released after |
987 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 988 | */ |
e7e3ffeb MWO |
989 | VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); |
990 | if (!folio_test_movable(src)) { | |
bda807d4 | 991 | rc = MIGRATEPAGE_SUCCESS; |
e7e3ffeb | 992 | folio_clear_isolated(src); |
bda807d4 MK |
993 | goto out; |
994 | } | |
995 | ||
da707a6d | 996 | mops = folio_movable_ops(src); |
68f2736a | 997 | rc = mops->migrate_page(&dst->page, &src->page, mode); |
bda807d4 | 998 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && |
e7e3ffeb | 999 | !folio_test_isolated(src)); |
bda807d4 | 1000 | } |
e24f0b8f | 1001 | |
5c3f9a67 | 1002 | /* |
e7e3ffeb MWO |
1003 | * When successful, old pagecache src->mapping must be cleared before |
1004 | * src is freed; but stats require that PageAnon be left as PageAnon. | |
5c3f9a67 HD |
1005 | */ |
1006 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e7e3ffeb MWO |
1007 | if (__PageMovable(&src->page)) { |
1008 | VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); | |
bda807d4 MK |
1009 | |
1010 | /* | |
1011 | * We clear PG_movable under page_lock so any compactor | |
1012 | * cannot try to migrate this page. | |
1013 | */ | |
e7e3ffeb | 1014 | folio_clear_isolated(src); |
bda807d4 MK |
1015 | } |
1016 | ||
1017 | /* | |
e7e3ffeb | 1018 | * Anonymous and movable src->mapping will be cleared by |
bda807d4 MK |
1019 | * free_pages_prepare so don't reset it here for keeping |
1020 | * the type to work PageAnon, for example. | |
1021 | */ | |
e7e3ffeb MWO |
1022 | if (!folio_mapping_flags(src)) |
1023 | src->mapping = NULL; | |
d2b2c6dd | 1024 | |
e7e3ffeb MWO |
1025 | if (likely(!folio_is_zone_device(dst))) |
1026 | flush_dcache_folio(dst); | |
3fe2011f | 1027 | } |
bda807d4 | 1028 | out: |
e24f0b8f CL |
1029 | return rc; |
1030 | } | |
1031 | ||
64c8902e HY |
1032 | /* |
1033 | * To record some information during migration, we use some unused | |
1034 | * fields (mapping and private) of struct folio of the newly allocated | |
1035 | * destination folio. This is safe because nobody is using them | |
1036 | * except us. | |
1037 | */ | |
e77d587a LT |
1038 | union migration_ptr { |
1039 | struct anon_vma *anon_vma; | |
1040 | struct address_space *mapping; | |
1041 | }; | |
64c8902e HY |
1042 | static void __migrate_folio_record(struct folio *dst, |
1043 | unsigned long page_was_mapped, | |
1044 | struct anon_vma *anon_vma) | |
1045 | { | |
e77d587a LT |
1046 | union migration_ptr ptr = { .anon_vma = anon_vma }; |
1047 | dst->mapping = ptr.mapping; | |
64c8902e HY |
1048 | dst->private = (void *)page_was_mapped; |
1049 | } | |
1050 | ||
1051 | static void __migrate_folio_extract(struct folio *dst, | |
1052 | int *page_was_mappedp, | |
1053 | struct anon_vma **anon_vmap) | |
1054 | { | |
e77d587a LT |
1055 | union migration_ptr ptr = { .mapping = dst->mapping }; |
1056 | *anon_vmap = ptr.anon_vma; | |
64c8902e HY |
1057 | *page_was_mappedp = (unsigned long)dst->private; |
1058 | dst->mapping = NULL; | |
1059 | dst->private = NULL; | |
1060 | } | |
1061 | ||
5dfab109 HY |
1062 | /* Restore the source folio to the original state upon failure */ |
1063 | static void migrate_folio_undo_src(struct folio *src, | |
1064 | int page_was_mapped, | |
1065 | struct anon_vma *anon_vma, | |
ebe75e47 | 1066 | bool locked, |
5dfab109 HY |
1067 | struct list_head *ret) |
1068 | { | |
1069 | if (page_was_mapped) | |
1070 | remove_migration_ptes(src, src, false); | |
1071 | /* Drop an anon_vma reference if we took one */ | |
1072 | if (anon_vma) | |
1073 | put_anon_vma(anon_vma); | |
ebe75e47 HY |
1074 | if (locked) |
1075 | folio_unlock(src); | |
1076 | if (ret) | |
1077 | list_move_tail(&src->lru, ret); | |
5dfab109 HY |
1078 | } |
1079 | ||
1080 | /* Restore the destination folio to the original state upon failure */ | |
1081 | static void migrate_folio_undo_dst(struct folio *dst, | |
ebe75e47 | 1082 | bool locked, |
5dfab109 HY |
1083 | free_page_t put_new_page, |
1084 | unsigned long private) | |
1085 | { | |
ebe75e47 HY |
1086 | if (locked) |
1087 | folio_unlock(dst); | |
5dfab109 HY |
1088 | if (put_new_page) |
1089 | put_new_page(&dst->page, private); | |
1090 | else | |
1091 | folio_put(dst); | |
1092 | } | |
1093 | ||
64c8902e HY |
1094 | /* Cleanup src folio upon migration success */ |
1095 | static void migrate_folio_done(struct folio *src, | |
1096 | enum migrate_reason reason) | |
1097 | { | |
1098 | /* | |
1099 | * Compaction can migrate also non-LRU pages which are | |
1100 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1101 | * as __PageMovable | |
1102 | */ | |
1103 | if (likely(!__folio_test_movable(src))) | |
1104 | mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + | |
1105 | folio_is_file_lru(src), -folio_nr_pages(src)); | |
1106 | ||
1107 | if (reason != MR_MEMORY_FAILURE) | |
1108 | /* We release the page in page_handle_poison. */ | |
1109 | folio_put(src); | |
1110 | } | |
1111 | ||
ebe75e47 HY |
1112 | /* Obtain the lock on page, remove all ptes. */ |
1113 | static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page, | |
1114 | unsigned long private, struct folio *src, | |
2ef7dbb2 HY |
1115 | struct folio **dstp, enum migrate_mode mode, |
1116 | enum migrate_reason reason, struct list_head *ret) | |
e24f0b8f | 1117 | { |
ebe75e47 | 1118 | struct folio *dst; |
0dabec93 | 1119 | int rc = -EAGAIN; |
ebe75e47 | 1120 | struct page *newpage = NULL; |
64c8902e | 1121 | int page_was_mapped = 0; |
3f6c8272 | 1122 | struct anon_vma *anon_vma = NULL; |
682a71a1 | 1123 | bool is_lru = !__PageMovable(&src->page); |
ebe75e47 HY |
1124 | bool locked = false; |
1125 | bool dst_locked = false; | |
1126 | ||
ebe75e47 HY |
1127 | if (folio_ref_count(src) == 1) { |
1128 | /* Folio was freed from under us. So we are done. */ | |
1129 | folio_clear_active(src); | |
1130 | folio_clear_unevictable(src); | |
1131 | /* free_pages_prepare() will clear PG_isolated. */ | |
1132 | list_del(&src->lru); | |
1133 | migrate_folio_done(src, reason); | |
1134 | return MIGRATEPAGE_SUCCESS; | |
1135 | } | |
1136 | ||
1137 | newpage = get_new_page(&src->page, private); | |
1138 | if (!newpage) | |
1139 | return -ENOMEM; | |
1140 | dst = page_folio(newpage); | |
1141 | *dstp = dst; | |
1142 | ||
1143 | dst->private = NULL; | |
95a402c3 | 1144 | |
682a71a1 | 1145 | if (!folio_trylock(src)) { |
2ef7dbb2 | 1146 | if (mode == MIGRATE_ASYNC) |
0dabec93 | 1147 | goto out; |
3e7d3449 MG |
1148 | |
1149 | /* | |
1150 | * It's not safe for direct compaction to call lock_page. | |
1151 | * For example, during page readahead pages are added locked | |
1152 | * to the LRU. Later, when the IO completes the pages are | |
1153 | * marked uptodate and unlocked. However, the queueing | |
1154 | * could be merging multiple pages for one bio (e.g. | |
d4388340 | 1155 | * mpage_readahead). If an allocation happens for the |
3e7d3449 MG |
1156 | * second or third page, the process can end up locking |
1157 | * the same page twice and deadlocking. Rather than | |
1158 | * trying to be clever about what pages can be locked, | |
1159 | * avoid the use of lock_page for direct compaction | |
1160 | * altogether. | |
1161 | */ | |
1162 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1163 | goto out; |
3e7d3449 | 1164 | |
682a71a1 | 1165 | folio_lock(src); |
e24f0b8f | 1166 | } |
ebe75e47 | 1167 | locked = true; |
e24f0b8f | 1168 | |
682a71a1 | 1169 | if (folio_test_writeback(src)) { |
11bc82d6 | 1170 | /* |
fed5b64a | 1171 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1172 | * necessary to wait for PageWriteback. In the async case, |
1173 | * the retry loop is too short and in the sync-light case, | |
1174 | * the overhead of stalling is too much | |
11bc82d6 | 1175 | */ |
2916ecc0 JG |
1176 | switch (mode) { |
1177 | case MIGRATE_SYNC: | |
1178 | case MIGRATE_SYNC_NO_COPY: | |
1179 | break; | |
1180 | default: | |
11bc82d6 | 1181 | rc = -EBUSY; |
ebe75e47 | 1182 | goto out; |
11bc82d6 | 1183 | } |
682a71a1 | 1184 | folio_wait_writeback(src); |
e24f0b8f | 1185 | } |
03f15c86 | 1186 | |
e24f0b8f | 1187 | /* |
682a71a1 MWO |
1188 | * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, |
1189 | * we cannot notice that anon_vma is freed while we migrate a page. | |
1ce82b69 | 1190 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1191 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1192 | * File Caches may use write_page() or lock_page() in migration, then, |
1193 | * just care Anon page here. | |
03f15c86 | 1194 | * |
29eea9b5 | 1195 | * Only folio_get_anon_vma() understands the subtleties of |
03f15c86 HD |
1196 | * getting a hold on an anon_vma from outside one of its mms. |
1197 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1198 | * because that implies that the anon page is no longer mapped | |
1199 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1200 | */ |
682a71a1 | 1201 | if (folio_test_anon(src) && !folio_test_ksm(src)) |
29eea9b5 | 1202 | anon_vma = folio_get_anon_vma(src); |
62e1c553 | 1203 | |
7db7671f HD |
1204 | /* |
1205 | * Block others from accessing the new page when we get around to | |
1206 | * establishing additional references. We are usually the only one | |
682a71a1 MWO |
1207 | * holding a reference to dst at this point. We used to have a BUG |
1208 | * here if folio_trylock(dst) fails, but would like to allow for | |
1209 | * cases where there might be a race with the previous use of dst. | |
7db7671f HD |
1210 | * This is much like races on refcount of oldpage: just don't BUG(). |
1211 | */ | |
682a71a1 | 1212 | if (unlikely(!folio_trylock(dst))) |
ebe75e47 HY |
1213 | goto out; |
1214 | dst_locked = true; | |
7db7671f | 1215 | |
bda807d4 | 1216 | if (unlikely(!is_lru)) { |
64c8902e HY |
1217 | __migrate_folio_record(dst, page_was_mapped, anon_vma); |
1218 | return MIGRATEPAGE_UNMAP; | |
bda807d4 MK |
1219 | } |
1220 | ||
dc386d4d | 1221 | /* |
62e1c553 SL |
1222 | * Corner case handling: |
1223 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1224 | * and treated as swapcache but it has no rmap yet. | |
682a71a1 | 1225 | * Calling try_to_unmap() against a src->mapping==NULL page will |
62e1c553 | 1226 | * trigger a BUG. So handle it here. |
d12b8951 | 1227 | * 2. An orphaned page (see truncate_cleanup_page) might have |
62e1c553 SL |
1228 | * fs-private metadata. The page can be picked up due to memory |
1229 | * offlining. Everywhere else except page reclaim, the page is | |
1230 | * invisible to the vm, so the page can not be migrated. So try to | |
1231 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1232 | */ |
682a71a1 MWO |
1233 | if (!src->mapping) { |
1234 | if (folio_test_private(src)) { | |
1235 | try_to_free_buffers(src); | |
ebe75e47 | 1236 | goto out; |
62e1c553 | 1237 | } |
682a71a1 | 1238 | } else if (folio_mapped(src)) { |
7db7671f | 1239 | /* Establish migration ptes */ |
682a71a1 MWO |
1240 | VM_BUG_ON_FOLIO(folio_test_anon(src) && |
1241 | !folio_test_ksm(src) && !anon_vma, src); | |
fb3592c4 | 1242 | try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0); |
64c8902e | 1243 | page_was_mapped = 1; |
2ebba6b7 | 1244 | } |
dc386d4d | 1245 | |
64c8902e HY |
1246 | if (!folio_mapped(src)) { |
1247 | __migrate_folio_record(dst, page_was_mapped, anon_vma); | |
1248 | return MIGRATEPAGE_UNMAP; | |
1249 | } | |
1250 | ||
64c8902e | 1251 | out: |
80562ba0 HY |
1252 | /* |
1253 | * A folio that has not been unmapped will be restored to | |
1254 | * right list unless we want to retry. | |
1255 | */ | |
fb3592c4 | 1256 | if (rc == -EAGAIN) |
ebe75e47 | 1257 | ret = NULL; |
80562ba0 | 1258 | |
ebe75e47 HY |
1259 | migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret); |
1260 | migrate_folio_undo_dst(dst, dst_locked, put_new_page, private); | |
80562ba0 HY |
1261 | |
1262 | return rc; | |
1263 | } | |
1264 | ||
ebe75e47 HY |
1265 | /* Migrate the folio to the newly allocated folio in dst. */ |
1266 | static int migrate_folio_move(free_page_t put_new_page, unsigned long private, | |
1267 | struct folio *src, struct folio *dst, | |
1268 | enum migrate_mode mode, enum migrate_reason reason, | |
1269 | struct list_head *ret) | |
64c8902e HY |
1270 | { |
1271 | int rc; | |
1272 | int page_was_mapped = 0; | |
1273 | struct anon_vma *anon_vma = NULL; | |
1274 | bool is_lru = !__PageMovable(&src->page); | |
5dfab109 | 1275 | struct list_head *prev; |
64c8902e HY |
1276 | |
1277 | __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); | |
5dfab109 HY |
1278 | prev = dst->lru.prev; |
1279 | list_del(&dst->lru); | |
64c8902e HY |
1280 | |
1281 | rc = move_to_new_folio(dst, src, mode); | |
ebe75e47 HY |
1282 | if (rc) |
1283 | goto out; | |
5dfab109 | 1284 | |
64c8902e HY |
1285 | if (unlikely(!is_lru)) |
1286 | goto out_unlock_both; | |
e24f0b8f | 1287 | |
c3096e67 | 1288 | /* |
682a71a1 | 1289 | * When successful, push dst to LRU immediately: so that if it |
c3096e67 | 1290 | * turns out to be an mlocked page, remove_migration_ptes() will |
682a71a1 | 1291 | * automatically build up the correct dst->mlock_count for it. |
c3096e67 HD |
1292 | * |
1293 | * We would like to do something similar for the old page, when | |
1294 | * unsuccessful, and other cases when a page has been temporarily | |
1295 | * isolated from the unevictable LRU: but this case is the easiest. | |
1296 | */ | |
ebe75e47 HY |
1297 | folio_add_lru(dst); |
1298 | if (page_was_mapped) | |
1299 | lru_add_drain(); | |
c3096e67 | 1300 | |
5c3f9a67 | 1301 | if (page_was_mapped) |
ebe75e47 | 1302 | remove_migration_ptes(src, dst, false); |
3f6c8272 | 1303 | |
7db7671f | 1304 | out_unlock_both: |
682a71a1 | 1305 | folio_unlock(dst); |
ebe75e47 | 1306 | set_page_owner_migrate_reason(&dst->page, reason); |
c6c919eb | 1307 | /* |
682a71a1 | 1308 | * If migration is successful, decrease refcount of dst, |
c6c919eb | 1309 | * which will not free the page because new page owner increased |
c3096e67 | 1310 | * refcounter. |
c6c919eb | 1311 | */ |
ebe75e47 | 1312 | folio_put(dst); |
c6c919eb | 1313 | |
dd4ae78a | 1314 | /* |
ebe75e47 HY |
1315 | * A folio that has been migrated has all references removed |
1316 | * and will be freed. | |
dd4ae78a | 1317 | */ |
ebe75e47 HY |
1318 | list_del(&src->lru); |
1319 | /* Drop an anon_vma reference if we took one */ | |
1320 | if (anon_vma) | |
1321 | put_anon_vma(anon_vma); | |
1322 | folio_unlock(src); | |
1323 | migrate_folio_done(src, reason); | |
bf6bddf1 | 1324 | |
ebe75e47 | 1325 | return rc; |
0dabec93 | 1326 | out: |
dd4ae78a | 1327 | /* |
ebe75e47 HY |
1328 | * A folio that has not been migrated will be restored to |
1329 | * right list unless we want to retry. | |
dd4ae78a | 1330 | */ |
ebe75e47 HY |
1331 | if (rc == -EAGAIN) { |
1332 | list_add(&dst->lru, prev); | |
1333 | __migrate_folio_record(dst, page_was_mapped, anon_vma); | |
1334 | return rc; | |
e24f0b8f | 1335 | } |
68711a74 | 1336 | |
ebe75e47 HY |
1337 | migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret); |
1338 | migrate_folio_undo_dst(dst, true, put_new_page, private); | |
1339 | ||
e24f0b8f CL |
1340 | return rc; |
1341 | } | |
1342 | ||
290408d4 NH |
1343 | /* |
1344 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1345 | * | |
1346 | * This function doesn't wait the completion of hugepage I/O | |
1347 | * because there is no race between I/O and migration for hugepage. | |
1348 | * Note that currently hugepage I/O occurs only in direct I/O | |
1349 | * where no lock is held and PG_writeback is irrelevant, | |
1350 | * and writeback status of all subpages are counted in the reference | |
1351 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1352 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1353 | * This means that when we try to migrate hugepage whose subpages are | |
1354 | * doing direct I/O, some references remain after try_to_unmap() and | |
1355 | * hugepage migration fails without data corruption. | |
1356 | * | |
1357 | * There is also no race when direct I/O is issued on the page under migration, | |
1358 | * because then pte is replaced with migration swap entry and direct I/O code | |
1359 | * will wait in the page fault for migration to complete. | |
1360 | */ | |
1361 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1362 | free_page_t put_new_page, unsigned long private, |
1363 | struct page *hpage, int force, | |
dd4ae78a YS |
1364 | enum migrate_mode mode, int reason, |
1365 | struct list_head *ret) | |
290408d4 | 1366 | { |
4eecb8b9 | 1367 | struct folio *dst, *src = page_folio(hpage); |
2def7424 | 1368 | int rc = -EAGAIN; |
2ebba6b7 | 1369 | int page_was_mapped = 0; |
32665f2b | 1370 | struct page *new_hpage; |
290408d4 | 1371 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1372 | struct address_space *mapping = NULL; |
290408d4 | 1373 | |
c33db292 | 1374 | if (folio_ref_count(src) == 1) { |
71a64f61 | 1375 | /* page was freed from under us. So we are done. */ |
ea8e72f4 | 1376 | folio_putback_active_hugetlb(src); |
71a64f61 MS |
1377 | return MIGRATEPAGE_SUCCESS; |
1378 | } | |
1379 | ||
666feb21 | 1380 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1381 | if (!new_hpage) |
1382 | return -ENOMEM; | |
4eecb8b9 | 1383 | dst = page_folio(new_hpage); |
290408d4 | 1384 | |
c33db292 | 1385 | if (!folio_trylock(src)) { |
2916ecc0 | 1386 | if (!force) |
290408d4 | 1387 | goto out; |
2916ecc0 JG |
1388 | switch (mode) { |
1389 | case MIGRATE_SYNC: | |
1390 | case MIGRATE_SYNC_NO_COPY: | |
1391 | break; | |
1392 | default: | |
1393 | goto out; | |
1394 | } | |
c33db292 | 1395 | folio_lock(src); |
290408d4 NH |
1396 | } |
1397 | ||
cb6acd01 MK |
1398 | /* |
1399 | * Check for pages which are in the process of being freed. Without | |
c33db292 | 1400 | * folio_mapping() set, hugetlbfs specific move page routine will not |
cb6acd01 MK |
1401 | * be called and we could leak usage counts for subpools. |
1402 | */ | |
345c62d1 | 1403 | if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { |
cb6acd01 MK |
1404 | rc = -EBUSY; |
1405 | goto out_unlock; | |
1406 | } | |
1407 | ||
c33db292 | 1408 | if (folio_test_anon(src)) |
29eea9b5 | 1409 | anon_vma = folio_get_anon_vma(src); |
290408d4 | 1410 | |
c33db292 | 1411 | if (unlikely(!folio_trylock(dst))) |
7db7671f HD |
1412 | goto put_anon; |
1413 | ||
c33db292 | 1414 | if (folio_mapped(src)) { |
a98a2f0c | 1415 | enum ttu_flags ttu = 0; |
336bf30e | 1416 | |
c33db292 | 1417 | if (!folio_test_anon(src)) { |
336bf30e MK |
1418 | /* |
1419 | * In shared mappings, try_to_unmap could potentially | |
1420 | * call huge_pmd_unshare. Because of this, take | |
1421 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1422 | * to let lower levels know we have taken the lock. | |
1423 | */ | |
1424 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1425 | if (unlikely(!mapping)) | |
1426 | goto unlock_put_anon; | |
1427 | ||
5202978b | 1428 | ttu = TTU_RMAP_LOCKED; |
336bf30e | 1429 | } |
c0d0381a | 1430 | |
4b8554c5 | 1431 | try_to_migrate(src, ttu); |
2ebba6b7 | 1432 | page_was_mapped = 1; |
336bf30e | 1433 | |
5202978b | 1434 | if (ttu & TTU_RMAP_LOCKED) |
336bf30e | 1435 | i_mmap_unlock_write(mapping); |
2ebba6b7 | 1436 | } |
290408d4 | 1437 | |
c33db292 | 1438 | if (!folio_mapped(src)) |
e7e3ffeb | 1439 | rc = move_to_new_folio(dst, src, mode); |
290408d4 | 1440 | |
336bf30e | 1441 | if (page_was_mapped) |
4eecb8b9 MWO |
1442 | remove_migration_ptes(src, |
1443 | rc == MIGRATEPAGE_SUCCESS ? dst : src, false); | |
290408d4 | 1444 | |
c0d0381a | 1445 | unlock_put_anon: |
c33db292 | 1446 | folio_unlock(dst); |
7db7671f HD |
1447 | |
1448 | put_anon: | |
fd4a4663 | 1449 | if (anon_vma) |
9e60109f | 1450 | put_anon_vma(anon_vma); |
8e6ac7fa | 1451 | |
2def7424 | 1452 | if (rc == MIGRATEPAGE_SUCCESS) { |
345c62d1 | 1453 | move_hugetlb_state(src, dst, reason); |
2def7424 HD |
1454 | put_new_page = NULL; |
1455 | } | |
8e6ac7fa | 1456 | |
cb6acd01 | 1457 | out_unlock: |
c33db292 | 1458 | folio_unlock(src); |
09761333 | 1459 | out: |
dd4ae78a | 1460 | if (rc == MIGRATEPAGE_SUCCESS) |
ea8e72f4 | 1461 | folio_putback_active_hugetlb(src); |
a04840c6 | 1462 | else if (rc != -EAGAIN) |
c33db292 | 1463 | list_move_tail(&src->lru, ret); |
68711a74 DR |
1464 | |
1465 | /* | |
1466 | * If migration was not successful and there's a freeing callback, use | |
1467 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1468 | * isolation. | |
1469 | */ | |
2def7424 | 1470 | if (put_new_page) |
68711a74 DR |
1471 | put_new_page(new_hpage, private); |
1472 | else | |
ea8e72f4 | 1473 | folio_putback_active_hugetlb(dst); |
68711a74 | 1474 | |
290408d4 NH |
1475 | return rc; |
1476 | } | |
1477 | ||
eaec4e63 | 1478 | static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) |
d532e2e5 | 1479 | { |
9c62ff00 | 1480 | int rc; |
d532e2e5 | 1481 | |
eaec4e63 HY |
1482 | folio_lock(folio); |
1483 | rc = split_folio_to_list(folio, split_folios); | |
1484 | folio_unlock(folio); | |
e6fa8a79 | 1485 | if (!rc) |
eaec4e63 | 1486 | list_move_tail(&folio->lru, split_folios); |
d532e2e5 YS |
1487 | |
1488 | return rc; | |
1489 | } | |
1490 | ||
42012e04 HY |
1491 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1492 | #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR | |
1493 | #else | |
1494 | #define NR_MAX_BATCHED_MIGRATION 512 | |
1495 | #endif | |
e5bfff8b | 1496 | #define NR_MAX_MIGRATE_PAGES_RETRY 10 |
2ef7dbb2 HY |
1497 | #define NR_MAX_MIGRATE_ASYNC_RETRY 3 |
1498 | #define NR_MAX_MIGRATE_SYNC_RETRY \ | |
1499 | (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY) | |
e5bfff8b | 1500 | |
5b855937 HY |
1501 | struct migrate_pages_stats { |
1502 | int nr_succeeded; /* Normal and large folios migrated successfully, in | |
1503 | units of base pages */ | |
1504 | int nr_failed_pages; /* Normal and large folios failed to be migrated, in | |
1505 | units of base pages. Untried folios aren't counted */ | |
1506 | int nr_thp_succeeded; /* THP migrated successfully */ | |
1507 | int nr_thp_failed; /* THP failed to be migrated */ | |
1508 | int nr_thp_split; /* THP split before migrating */ | |
1509 | }; | |
1510 | ||
b20a3503 | 1511 | /* |
e5bfff8b HY |
1512 | * Returns the number of hugetlb folios that were not migrated, or an error code |
1513 | * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable | |
1514 | * any more because the list has become empty or no retryable hugetlb folios | |
1515 | * exist any more. It is caller's responsibility to call putback_movable_pages() | |
1516 | * only if ret != 0. | |
b20a3503 | 1517 | */ |
e5bfff8b HY |
1518 | static int migrate_hugetlbs(struct list_head *from, new_page_t get_new_page, |
1519 | free_page_t put_new_page, unsigned long private, | |
1520 | enum migrate_mode mode, int reason, | |
1521 | struct migrate_pages_stats *stats, | |
1522 | struct list_head *ret_folios) | |
b20a3503 | 1523 | { |
e24f0b8f | 1524 | int retry = 1; |
e5bfff8b HY |
1525 | int nr_failed = 0; |
1526 | int nr_retry_pages = 0; | |
1527 | int pass = 0; | |
1528 | struct folio *folio, *folio2; | |
1529 | int rc, nr_pages; | |
1530 | ||
1531 | for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) { | |
1532 | retry = 0; | |
1533 | nr_retry_pages = 0; | |
1534 | ||
1535 | list_for_each_entry_safe(folio, folio2, from, lru) { | |
1536 | if (!folio_test_hugetlb(folio)) | |
1537 | continue; | |
1538 | ||
1539 | nr_pages = folio_nr_pages(folio); | |
1540 | ||
1541 | cond_resched(); | |
1542 | ||
6f7d760e HY |
1543 | /* |
1544 | * Migratability of hugepages depends on architectures and | |
1545 | * their size. This check is necessary because some callers | |
1546 | * of hugepage migration like soft offline and memory | |
1547 | * hotremove don't walk through page tables or check whether | |
1548 | * the hugepage is pmd-based or not before kicking migration. | |
1549 | */ | |
1550 | if (!hugepage_migration_supported(folio_hstate(folio))) { | |
1551 | nr_failed++; | |
1552 | stats->nr_failed_pages += nr_pages; | |
1553 | list_move_tail(&folio->lru, ret_folios); | |
1554 | continue; | |
1555 | } | |
1556 | ||
e5bfff8b HY |
1557 | rc = unmap_and_move_huge_page(get_new_page, |
1558 | put_new_page, private, | |
1559 | &folio->page, pass > 2, mode, | |
1560 | reason, ret_folios); | |
1561 | /* | |
1562 | * The rules are: | |
1563 | * Success: hugetlb folio will be put back | |
1564 | * -EAGAIN: stay on the from list | |
1565 | * -ENOMEM: stay on the from list | |
e5bfff8b HY |
1566 | * Other errno: put on ret_folios list |
1567 | */ | |
1568 | switch(rc) { | |
e5bfff8b HY |
1569 | case -ENOMEM: |
1570 | /* | |
1571 | * When memory is low, don't bother to try to migrate | |
1572 | * other folios, just exit. | |
1573 | */ | |
1574 | stats->nr_failed_pages += nr_pages + nr_retry_pages; | |
1575 | return -ENOMEM; | |
1576 | case -EAGAIN: | |
1577 | retry++; | |
1578 | nr_retry_pages += nr_pages; | |
1579 | break; | |
1580 | case MIGRATEPAGE_SUCCESS: | |
1581 | stats->nr_succeeded += nr_pages; | |
1582 | break; | |
1583 | default: | |
1584 | /* | |
1585 | * Permanent failure (-EBUSY, etc.): | |
1586 | * unlike -EAGAIN case, the failed folio is | |
1587 | * removed from migration folio list and not | |
1588 | * retried in the next outer loop. | |
1589 | */ | |
1590 | nr_failed++; | |
1591 | stats->nr_failed_pages += nr_pages; | |
1592 | break; | |
1593 | } | |
1594 | } | |
1595 | } | |
1596 | /* | |
1597 | * nr_failed is number of hugetlb folios failed to be migrated. After | |
1598 | * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb | |
1599 | * folios as failed. | |
1600 | */ | |
1601 | nr_failed += retry; | |
1602 | stats->nr_failed_pages += nr_retry_pages; | |
1603 | ||
1604 | return nr_failed; | |
1605 | } | |
1606 | ||
5dfab109 HY |
1607 | /* |
1608 | * migrate_pages_batch() first unmaps folios in the from list as many as | |
1609 | * possible, then move the unmapped folios. | |
fb3592c4 HY |
1610 | * |
1611 | * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a | |
1612 | * lock or bit when we have locked more than one folio. Which may cause | |
1613 | * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the | |
1614 | * length of the from list must be <= 1. | |
5dfab109 | 1615 | */ |
42012e04 | 1616 | static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page, |
68711a74 | 1617 | free_page_t put_new_page, unsigned long private, |
42012e04 | 1618 | enum migrate_mode mode, int reason, struct list_head *ret_folios, |
a21d2133 HY |
1619 | struct list_head *split_folios, struct migrate_pages_stats *stats, |
1620 | int nr_pass) | |
b20a3503 | 1621 | { |
a21d2133 | 1622 | int retry = 1; |
eaec4e63 | 1623 | int large_retry = 1; |
1a5bae25 | 1624 | int thp_retry = 1; |
b20a3503 | 1625 | int nr_failed = 0; |
077309bc | 1626 | int nr_retry_pages = 0; |
eaec4e63 | 1627 | int nr_large_failed = 0; |
b20a3503 | 1628 | int pass = 0; |
eaec4e63 | 1629 | bool is_large = false; |
1a5bae25 | 1630 | bool is_thp = false; |
5dfab109 | 1631 | struct folio *folio, *folio2, *dst = NULL, *dst2; |
a21d2133 | 1632 | int rc, rc_saved = 0, nr_pages; |
5dfab109 HY |
1633 | LIST_HEAD(unmap_folios); |
1634 | LIST_HEAD(dst_folios); | |
b0b515bf | 1635 | bool nosplit = (reason == MR_NUMA_MISPLACED); |
e5bfff8b | 1636 | |
fb3592c4 HY |
1637 | VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC && |
1638 | !list_empty(from) && !list_is_singular(from)); | |
a21d2133 HY |
1639 | |
1640 | for (pass = 0; pass < nr_pass && (retry || large_retry); pass++) { | |
e24f0b8f | 1641 | retry = 0; |
eaec4e63 | 1642 | large_retry = 0; |
1a5bae25 | 1643 | thp_retry = 0; |
077309bc | 1644 | nr_retry_pages = 0; |
b20a3503 | 1645 | |
eaec4e63 | 1646 | list_for_each_entry_safe(folio, folio2, from, lru) { |
1a5bae25 | 1647 | /* |
eaec4e63 HY |
1648 | * Large folio statistics is based on the source large |
1649 | * folio. Capture required information that might get | |
1650 | * lost during migration. | |
1a5bae25 | 1651 | */ |
e5bfff8b | 1652 | is_large = folio_test_large(folio); |
eaec4e63 HY |
1653 | is_thp = is_large && folio_test_pmd_mappable(folio); |
1654 | nr_pages = folio_nr_pages(folio); | |
e5bfff8b | 1655 | |
e24f0b8f | 1656 | cond_resched(); |
2d1db3b1 | 1657 | |
d532e2e5 | 1658 | /* |
eaec4e63 | 1659 | * Large folio migration might be unsupported or |
6f7d760e | 1660 | * the allocation might be failed so we should retry |
eaec4e63 HY |
1661 | * on the same folio with the large folio split |
1662 | * to normal folios. | |
d532e2e5 | 1663 | * |
eaec4e63 | 1664 | * Split folios are put in split_folios, and |
e6fa8a79 HY |
1665 | * we will migrate them after the rest of the |
1666 | * list is processed. | |
d532e2e5 | 1667 | */ |
6f7d760e HY |
1668 | if (!thp_migration_supported() && is_thp) { |
1669 | nr_large_failed++; | |
1670 | stats->nr_thp_failed++; | |
a21d2133 | 1671 | if (!try_split_folio(folio, split_folios)) { |
6f7d760e HY |
1672 | stats->nr_thp_split++; |
1673 | continue; | |
f430893b | 1674 | } |
6f7d760e HY |
1675 | stats->nr_failed_pages += nr_pages; |
1676 | list_move_tail(&folio->lru, ret_folios); | |
1677 | continue; | |
1678 | } | |
f430893b | 1679 | |
64c8902e | 1680 | rc = migrate_folio_unmap(get_new_page, put_new_page, private, |
2ef7dbb2 | 1681 | folio, &dst, mode, reason, ret_folios); |
dd4ae78a YS |
1682 | /* |
1683 | * The rules are: | |
e5bfff8b | 1684 | * Success: folio will be freed |
5dfab109 HY |
1685 | * Unmap: folio will be put on unmap_folios list, |
1686 | * dst folio put on dst_folios list | |
dd4ae78a YS |
1687 | * -EAGAIN: stay on the from list |
1688 | * -ENOMEM: stay on the from list | |
42012e04 | 1689 | * Other errno: put on ret_folios list |
dd4ae78a | 1690 | */ |
e24f0b8f | 1691 | switch(rc) { |
95a402c3 | 1692 | case -ENOMEM: |
94723aaf | 1693 | /* |
d532e2e5 | 1694 | * When memory is low, don't bother to try to migrate |
5dfab109 | 1695 | * other folios, move unmapped folios, then exit. |
94723aaf | 1696 | */ |
eaec4e63 HY |
1697 | if (is_large) { |
1698 | nr_large_failed++; | |
42012e04 | 1699 | stats->nr_thp_failed += is_thp; |
eaec4e63 | 1700 | /* Large folio NUMA faulting doesn't split to retry. */ |
fd4a7ac3 | 1701 | if (!nosplit) { |
a21d2133 | 1702 | int ret = try_split_folio(folio, split_folios); |
fd4a7ac3 BW |
1703 | |
1704 | if (!ret) { | |
42012e04 | 1705 | stats->nr_thp_split += is_thp; |
fd4a7ac3 BW |
1706 | break; |
1707 | } else if (reason == MR_LONGTERM_PIN && | |
1708 | ret == -EAGAIN) { | |
1709 | /* | |
eaec4e63 HY |
1710 | * Try again to split large folio to |
1711 | * mitigate the failure of longterm pinning. | |
fd4a7ac3 | 1712 | */ |
eaec4e63 HY |
1713 | large_retry++; |
1714 | thp_retry += is_thp; | |
1715 | nr_retry_pages += nr_pages; | |
fd4a7ac3 BW |
1716 | break; |
1717 | } | |
94723aaf | 1718 | } |
a21d2133 | 1719 | } else { |
f430893b | 1720 | nr_failed++; |
1a5bae25 | 1721 | } |
b5bade97 | 1722 | |
42012e04 | 1723 | stats->nr_failed_pages += nr_pages + nr_retry_pages; |
fbed53b4 | 1724 | /* nr_failed isn't updated for not used */ |
eaec4e63 | 1725 | nr_large_failed += large_retry; |
42012e04 | 1726 | stats->nr_thp_failed += thp_retry; |
5dfab109 HY |
1727 | rc_saved = rc; |
1728 | if (list_empty(&unmap_folios)) | |
1729 | goto out; | |
1730 | else | |
1731 | goto move; | |
e24f0b8f | 1732 | case -EAGAIN: |
eaec4e63 HY |
1733 | if (is_large) { |
1734 | large_retry++; | |
1735 | thp_retry += is_thp; | |
a21d2133 | 1736 | } else { |
f430893b | 1737 | retry++; |
eaec4e63 HY |
1738 | } |
1739 | nr_retry_pages += nr_pages; | |
e24f0b8f | 1740 | break; |
78bd5209 | 1741 | case MIGRATEPAGE_SUCCESS: |
42012e04 HY |
1742 | stats->nr_succeeded += nr_pages; |
1743 | stats->nr_thp_succeeded += is_thp; | |
e24f0b8f | 1744 | break; |
5dfab109 | 1745 | case MIGRATEPAGE_UNMAP: |
5dfab109 HY |
1746 | list_move_tail(&folio->lru, &unmap_folios); |
1747 | list_add_tail(&dst->lru, &dst_folios); | |
e24f0b8f CL |
1748 | break; |
1749 | default: | |
354a3363 | 1750 | /* |
d532e2e5 | 1751 | * Permanent failure (-EBUSY, etc.): |
eaec4e63 HY |
1752 | * unlike -EAGAIN case, the failed folio is |
1753 | * removed from migration folio list and not | |
354a3363 NH |
1754 | * retried in the next outer loop. |
1755 | */ | |
eaec4e63 HY |
1756 | if (is_large) { |
1757 | nr_large_failed++; | |
42012e04 | 1758 | stats->nr_thp_failed += is_thp; |
a21d2133 | 1759 | } else { |
b5bade97 | 1760 | nr_failed++; |
eaec4e63 | 1761 | } |
f430893b | 1762 | |
42012e04 | 1763 | stats->nr_failed_pages += nr_pages; |
e24f0b8f | 1764 | break; |
2d1db3b1 | 1765 | } |
b20a3503 CL |
1766 | } |
1767 | } | |
7047b5a4 | 1768 | nr_failed += retry; |
eaec4e63 | 1769 | nr_large_failed += large_retry; |
42012e04 HY |
1770 | stats->nr_thp_failed += thp_retry; |
1771 | stats->nr_failed_pages += nr_retry_pages; | |
5dfab109 | 1772 | move: |
7e12beb8 HY |
1773 | /* Flush TLBs for all unmapped folios */ |
1774 | try_to_unmap_flush(); | |
1775 | ||
5dfab109 | 1776 | retry = 1; |
a21d2133 | 1777 | for (pass = 0; pass < nr_pass && (retry || large_retry); pass++) { |
5dfab109 HY |
1778 | retry = 0; |
1779 | large_retry = 0; | |
1780 | thp_retry = 0; | |
1781 | nr_retry_pages = 0; | |
1782 | ||
1783 | dst = list_first_entry(&dst_folios, struct folio, lru); | |
1784 | dst2 = list_next_entry(dst, lru); | |
1785 | list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { | |
1786 | is_large = folio_test_large(folio); | |
1787 | is_thp = is_large && folio_test_pmd_mappable(folio); | |
1788 | nr_pages = folio_nr_pages(folio); | |
1789 | ||
1790 | cond_resched(); | |
1791 | ||
1792 | rc = migrate_folio_move(put_new_page, private, | |
1793 | folio, dst, mode, | |
1794 | reason, ret_folios); | |
1795 | /* | |
1796 | * The rules are: | |
1797 | * Success: folio will be freed | |
1798 | * -EAGAIN: stay on the unmap_folios list | |
1799 | * Other errno: put on ret_folios list | |
1800 | */ | |
1801 | switch(rc) { | |
1802 | case -EAGAIN: | |
1803 | if (is_large) { | |
1804 | large_retry++; | |
1805 | thp_retry += is_thp; | |
a21d2133 | 1806 | } else { |
5dfab109 HY |
1807 | retry++; |
1808 | } | |
1809 | nr_retry_pages += nr_pages; | |
1810 | break; | |
1811 | case MIGRATEPAGE_SUCCESS: | |
1812 | stats->nr_succeeded += nr_pages; | |
1813 | stats->nr_thp_succeeded += is_thp; | |
1814 | break; | |
1815 | default: | |
1816 | if (is_large) { | |
1817 | nr_large_failed++; | |
1818 | stats->nr_thp_failed += is_thp; | |
a21d2133 | 1819 | } else { |
b5bade97 | 1820 | nr_failed++; |
eaec4e63 | 1821 | } |
f430893b | 1822 | |
5dfab109 | 1823 | stats->nr_failed_pages += nr_pages; |
e24f0b8f | 1824 | break; |
2d1db3b1 | 1825 | } |
5dfab109 HY |
1826 | dst = dst2; |
1827 | dst2 = list_next_entry(dst, lru); | |
b20a3503 CL |
1828 | } |
1829 | } | |
7047b5a4 | 1830 | nr_failed += retry; |
eaec4e63 | 1831 | nr_large_failed += large_retry; |
5dfab109 HY |
1832 | stats->nr_thp_failed += thp_retry; |
1833 | stats->nr_failed_pages += nr_retry_pages; | |
1834 | ||
1835 | if (rc_saved) | |
1836 | rc = rc_saved; | |
1837 | else | |
1838 | rc = nr_failed + nr_large_failed; | |
1839 | out: | |
1840 | /* Cleanup remaining folios */ | |
1841 | dst = list_first_entry(&dst_folios, struct folio, lru); | |
1842 | dst2 = list_next_entry(dst, lru); | |
1843 | list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { | |
1844 | int page_was_mapped = 0; | |
1845 | struct anon_vma *anon_vma = NULL; | |
1846 | ||
1847 | __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); | |
1848 | migrate_folio_undo_src(folio, page_was_mapped, anon_vma, | |
ebe75e47 | 1849 | true, ret_folios); |
5dfab109 | 1850 | list_del(&dst->lru); |
ebe75e47 | 1851 | migrate_folio_undo_dst(dst, true, put_new_page, private); |
5dfab109 HY |
1852 | dst = dst2; |
1853 | dst2 = list_next_entry(dst, lru); | |
1854 | } | |
1855 | ||
42012e04 HY |
1856 | return rc; |
1857 | } | |
1858 | ||
2ef7dbb2 HY |
1859 | static int migrate_pages_sync(struct list_head *from, new_page_t get_new_page, |
1860 | free_page_t put_new_page, unsigned long private, | |
1861 | enum migrate_mode mode, int reason, struct list_head *ret_folios, | |
1862 | struct list_head *split_folios, struct migrate_pages_stats *stats) | |
1863 | { | |
1864 | int rc, nr_failed = 0; | |
1865 | LIST_HEAD(folios); | |
1866 | struct migrate_pages_stats astats; | |
1867 | ||
1868 | memset(&astats, 0, sizeof(astats)); | |
1869 | /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */ | |
1870 | rc = migrate_pages_batch(from, get_new_page, put_new_page, private, MIGRATE_ASYNC, | |
1871 | reason, &folios, split_folios, &astats, | |
1872 | NR_MAX_MIGRATE_ASYNC_RETRY); | |
1873 | stats->nr_succeeded += astats.nr_succeeded; | |
1874 | stats->nr_thp_succeeded += astats.nr_thp_succeeded; | |
1875 | stats->nr_thp_split += astats.nr_thp_split; | |
1876 | if (rc < 0) { | |
1877 | stats->nr_failed_pages += astats.nr_failed_pages; | |
1878 | stats->nr_thp_failed += astats.nr_thp_failed; | |
1879 | list_splice_tail(&folios, ret_folios); | |
1880 | return rc; | |
1881 | } | |
1882 | stats->nr_thp_failed += astats.nr_thp_split; | |
1883 | nr_failed += astats.nr_thp_split; | |
1884 | /* | |
1885 | * Fall back to migrate all failed folios one by one synchronously. All | |
1886 | * failed folios except split THPs will be retried, so their failure | |
1887 | * isn't counted | |
1888 | */ | |
1889 | list_splice_tail_init(&folios, from); | |
1890 | while (!list_empty(from)) { | |
1891 | list_move(from->next, &folios); | |
1892 | rc = migrate_pages_batch(&folios, get_new_page, put_new_page, | |
1893 | private, mode, reason, ret_folios, | |
1894 | split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY); | |
1895 | list_splice_tail_init(&folios, ret_folios); | |
1896 | if (rc < 0) | |
1897 | return rc; | |
1898 | nr_failed += rc; | |
1899 | } | |
1900 | ||
1901 | return nr_failed; | |
1902 | } | |
1903 | ||
42012e04 HY |
1904 | /* |
1905 | * migrate_pages - migrate the folios specified in a list, to the free folios | |
1906 | * supplied as the target for the page migration | |
1907 | * | |
1908 | * @from: The list of folios to be migrated. | |
1909 | * @get_new_page: The function used to allocate free folios to be used | |
1910 | * as the target of the folio migration. | |
1911 | * @put_new_page: The function used to free target folios if migration | |
1912 | * fails, or NULL if no special handling is necessary. | |
1913 | * @private: Private data to be passed on to get_new_page() | |
1914 | * @mode: The migration mode that specifies the constraints for | |
1915 | * folio migration, if any. | |
1916 | * @reason: The reason for folio migration. | |
1917 | * @ret_succeeded: Set to the number of folios migrated successfully if | |
1918 | * the caller passes a non-NULL pointer. | |
1919 | * | |
1920 | * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios | |
1921 | * are movable any more because the list has become empty or no retryable folios | |
1922 | * exist any more. It is caller's responsibility to call putback_movable_pages() | |
1923 | * only if ret != 0. | |
1924 | * | |
1925 | * Returns the number of {normal folio, large folio, hugetlb} that were not | |
1926 | * migrated, or an error code. The number of large folio splits will be | |
1927 | * considered as the number of non-migrated large folio, no matter how many | |
1928 | * split folios of the large folio are migrated successfully. | |
1929 | */ | |
1930 | int migrate_pages(struct list_head *from, new_page_t get_new_page, | |
1931 | free_page_t put_new_page, unsigned long private, | |
1932 | enum migrate_mode mode, int reason, unsigned int *ret_succeeded) | |
1933 | { | |
1934 | int rc, rc_gather; | |
2ef7dbb2 | 1935 | int nr_pages; |
42012e04 HY |
1936 | struct folio *folio, *folio2; |
1937 | LIST_HEAD(folios); | |
1938 | LIST_HEAD(ret_folios); | |
a21d2133 | 1939 | LIST_HEAD(split_folios); |
42012e04 HY |
1940 | struct migrate_pages_stats stats; |
1941 | ||
1942 | trace_mm_migrate_pages_start(mode, reason); | |
1943 | ||
1944 | memset(&stats, 0, sizeof(stats)); | |
1945 | ||
1946 | rc_gather = migrate_hugetlbs(from, get_new_page, put_new_page, private, | |
1947 | mode, reason, &stats, &ret_folios); | |
1948 | if (rc_gather < 0) | |
1949 | goto out; | |
fb3592c4 | 1950 | |
42012e04 HY |
1951 | again: |
1952 | nr_pages = 0; | |
1953 | list_for_each_entry_safe(folio, folio2, from, lru) { | |
1954 | /* Retried hugetlb folios will be kept in list */ | |
1955 | if (folio_test_hugetlb(folio)) { | |
1956 | list_move_tail(&folio->lru, &ret_folios); | |
1957 | continue; | |
1958 | } | |
1959 | ||
1960 | nr_pages += folio_nr_pages(folio); | |
2ef7dbb2 | 1961 | if (nr_pages >= NR_MAX_BATCHED_MIGRATION) |
42012e04 HY |
1962 | break; |
1963 | } | |
2ef7dbb2 | 1964 | if (nr_pages >= NR_MAX_BATCHED_MIGRATION) |
fb3592c4 | 1965 | list_cut_before(&folios, from, &folio2->lru); |
42012e04 HY |
1966 | else |
1967 | list_splice_init(from, &folios); | |
2ef7dbb2 HY |
1968 | if (mode == MIGRATE_ASYNC) |
1969 | rc = migrate_pages_batch(&folios, get_new_page, put_new_page, private, | |
1970 | mode, reason, &ret_folios, &split_folios, &stats, | |
1971 | NR_MAX_MIGRATE_PAGES_RETRY); | |
1972 | else | |
1973 | rc = migrate_pages_sync(&folios, get_new_page, put_new_page, private, | |
1974 | mode, reason, &ret_folios, &split_folios, &stats); | |
42012e04 HY |
1975 | list_splice_tail_init(&folios, &ret_folios); |
1976 | if (rc < 0) { | |
1977 | rc_gather = rc; | |
a21d2133 | 1978 | list_splice_tail(&split_folios, &ret_folios); |
42012e04 HY |
1979 | goto out; |
1980 | } | |
a21d2133 HY |
1981 | if (!list_empty(&split_folios)) { |
1982 | /* | |
1983 | * Failure isn't counted since all split folios of a large folio | |
1984 | * is counted as 1 failure already. And, we only try to migrate | |
1985 | * with minimal effort, force MIGRATE_ASYNC mode and retry once. | |
1986 | */ | |
1987 | migrate_pages_batch(&split_folios, get_new_page, put_new_page, private, | |
1988 | MIGRATE_ASYNC, reason, &ret_folios, NULL, &stats, 1); | |
1989 | list_splice_tail_init(&split_folios, &ret_folios); | |
1990 | } | |
42012e04 HY |
1991 | rc_gather += rc; |
1992 | if (!list_empty(from)) | |
1993 | goto again; | |
95a402c3 | 1994 | out: |
dd4ae78a | 1995 | /* |
eaec4e63 | 1996 | * Put the permanent failure folio back to migration list, they |
dd4ae78a YS |
1997 | * will be put back to the right list by the caller. |
1998 | */ | |
eaec4e63 | 1999 | list_splice(&ret_folios, from); |
dd4ae78a | 2000 | |
03e5f82e | 2001 | /* |
eaec4e63 HY |
2002 | * Return 0 in case all split folios of fail-to-migrate large folios |
2003 | * are migrated successfully. | |
03e5f82e BW |
2004 | */ |
2005 | if (list_empty(from)) | |
42012e04 | 2006 | rc_gather = 0; |
03e5f82e | 2007 | |
5b855937 HY |
2008 | count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded); |
2009 | count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages); | |
2010 | count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded); | |
2011 | count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed); | |
2012 | count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split); | |
2013 | trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages, | |
2014 | stats.nr_thp_succeeded, stats.nr_thp_failed, | |
2015 | stats.nr_thp_split, mode, reason); | |
7b2a2d4a | 2016 | |
5ac95884 | 2017 | if (ret_succeeded) |
5b855937 | 2018 | *ret_succeeded = stats.nr_succeeded; |
5ac95884 | 2019 | |
42012e04 | 2020 | return rc_gather; |
b20a3503 | 2021 | } |
95a402c3 | 2022 | |
19fc7bed | 2023 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
b4b38223 | 2024 | { |
ffe06786 | 2025 | struct folio *folio = page_folio(page); |
19fc7bed JK |
2026 | struct migration_target_control *mtc; |
2027 | gfp_t gfp_mask; | |
b4b38223 | 2028 | unsigned int order = 0; |
e37d3e83 | 2029 | struct folio *hugetlb_folio = NULL; |
ffe06786 | 2030 | struct folio *new_folio = NULL; |
19fc7bed JK |
2031 | int nid; |
2032 | int zidx; | |
2033 | ||
2034 | mtc = (struct migration_target_control *)private; | |
2035 | gfp_mask = mtc->gfp_mask; | |
2036 | nid = mtc->nid; | |
2037 | if (nid == NUMA_NO_NODE) | |
ffe06786 | 2038 | nid = folio_nid(folio); |
b4b38223 | 2039 | |
ffe06786 | 2040 | if (folio_test_hugetlb(folio)) { |
e51da3a9 | 2041 | struct hstate *h = folio_hstate(folio); |
d92bbc27 | 2042 | |
19fc7bed | 2043 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
e37d3e83 SK |
2044 | hugetlb_folio = alloc_hugetlb_folio_nodemask(h, nid, |
2045 | mtc->nmask, gfp_mask); | |
2046 | return &hugetlb_folio->page; | |
d92bbc27 | 2047 | } |
b4b38223 | 2048 | |
ffe06786 | 2049 | if (folio_test_large(folio)) { |
9933a0c8 JK |
2050 | /* |
2051 | * clear __GFP_RECLAIM to make the migration callback | |
2052 | * consistent with regular THP allocations. | |
2053 | */ | |
2054 | gfp_mask &= ~__GFP_RECLAIM; | |
b4b38223 | 2055 | gfp_mask |= GFP_TRANSHUGE; |
ffe06786 | 2056 | order = folio_order(folio); |
b4b38223 | 2057 | } |
ffe06786 | 2058 | zidx = zone_idx(folio_zone(folio)); |
19fc7bed | 2059 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) |
b4b38223 JK |
2060 | gfp_mask |= __GFP_HIGHMEM; |
2061 | ||
ffe06786 | 2062 | new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask); |
b4b38223 | 2063 | |
ffe06786 | 2064 | return &new_folio->page; |
b4b38223 JK |
2065 | } |
2066 | ||
742755a1 | 2067 | #ifdef CONFIG_NUMA |
742755a1 | 2068 | |
a49bd4d7 | 2069 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 2070 | { |
a49bd4d7 MH |
2071 | while (nr-- > 0) { |
2072 | if (put_user(value, status + start)) | |
2073 | return -EFAULT; | |
2074 | start++; | |
2075 | } | |
2076 | ||
2077 | return 0; | |
2078 | } | |
2079 | ||
2080 | static int do_move_pages_to_node(struct mm_struct *mm, | |
2081 | struct list_head *pagelist, int node) | |
2082 | { | |
2083 | int err; | |
a0976311 JK |
2084 | struct migration_target_control mtc = { |
2085 | .nid = node, | |
2086 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
2087 | }; | |
a49bd4d7 | 2088 | |
a0976311 | 2089 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
5ac95884 | 2090 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); |
a49bd4d7 MH |
2091 | if (err) |
2092 | putback_movable_pages(pagelist); | |
2093 | return err; | |
742755a1 CL |
2094 | } |
2095 | ||
2096 | /* | |
a49bd4d7 MH |
2097 | * Resolves the given address to a struct page, isolates it from the LRU and |
2098 | * puts it to the given pagelist. | |
e0153fc2 YS |
2099 | * Returns: |
2100 | * errno - if the page cannot be found/isolated | |
2101 | * 0 - when it doesn't have to be migrated because it is already on the | |
2102 | * target node | |
2103 | * 1 - when it has been queued | |
742755a1 | 2104 | */ |
a49bd4d7 MH |
2105 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
2106 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 2107 | { |
a49bd4d7 MH |
2108 | struct vm_area_struct *vma; |
2109 | struct page *page; | |
742755a1 | 2110 | int err; |
9747b9e9 | 2111 | bool isolated; |
742755a1 | 2112 | |
d8ed45c5 | 2113 | mmap_read_lock(mm); |
a49bd4d7 | 2114 | err = -EFAULT; |
cb1c37b1 ML |
2115 | vma = vma_lookup(mm, addr); |
2116 | if (!vma || !vma_migratable(vma)) | |
a49bd4d7 | 2117 | goto out; |
742755a1 | 2118 | |
a49bd4d7 | 2119 | /* FOLL_DUMP to ignore special (like zero) pages */ |
87d2762e | 2120 | page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); |
89f5b7da | 2121 | |
a49bd4d7 MH |
2122 | err = PTR_ERR(page); |
2123 | if (IS_ERR(page)) | |
2124 | goto out; | |
89f5b7da | 2125 | |
a49bd4d7 | 2126 | err = -ENOENT; |
f7091ed6 | 2127 | if (!page) |
a49bd4d7 | 2128 | goto out; |
742755a1 | 2129 | |
f7091ed6 HW |
2130 | if (is_zone_device_page(page)) |
2131 | goto out_putpage; | |
2132 | ||
a49bd4d7 MH |
2133 | err = 0; |
2134 | if (page_to_nid(page) == node) | |
2135 | goto out_putpage; | |
742755a1 | 2136 | |
a49bd4d7 MH |
2137 | err = -EACCES; |
2138 | if (page_mapcount(page) > 1 && !migrate_all) | |
2139 | goto out_putpage; | |
742755a1 | 2140 | |
a49bd4d7 MH |
2141 | if (PageHuge(page)) { |
2142 | if (PageHead(page)) { | |
9747b9e9 BW |
2143 | isolated = isolate_hugetlb(page_folio(page), pagelist); |
2144 | err = isolated ? 1 : -EBUSY; | |
e632a938 | 2145 | } |
a49bd4d7 MH |
2146 | } else { |
2147 | struct page *head; | |
e632a938 | 2148 | |
e8db67eb | 2149 | head = compound_head(page); |
f7f9c00d BW |
2150 | isolated = isolate_lru_page(head); |
2151 | if (!isolated) { | |
2152 | err = -EBUSY; | |
a49bd4d7 | 2153 | goto out_putpage; |
f7f9c00d | 2154 | } |
742755a1 | 2155 | |
e0153fc2 | 2156 | err = 1; |
a49bd4d7 MH |
2157 | list_add_tail(&head->lru, pagelist); |
2158 | mod_node_page_state(page_pgdat(head), | |
9de4f22a | 2159 | NR_ISOLATED_ANON + page_is_file_lru(head), |
6c357848 | 2160 | thp_nr_pages(head)); |
a49bd4d7 MH |
2161 | } |
2162 | out_putpage: | |
2163 | /* | |
2164 | * Either remove the duplicate refcount from | |
2165 | * isolate_lru_page() or drop the page ref if it was | |
2166 | * not isolated. | |
2167 | */ | |
2168 | put_page(page); | |
2169 | out: | |
d8ed45c5 | 2170 | mmap_read_unlock(mm); |
742755a1 CL |
2171 | return err; |
2172 | } | |
2173 | ||
7ca8783a WY |
2174 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
2175 | struct list_head *pagelist, int __user *status, | |
2176 | int start, int i, unsigned long nr_pages) | |
2177 | { | |
2178 | int err; | |
2179 | ||
5d7ae891 WY |
2180 | if (list_empty(pagelist)) |
2181 | return 0; | |
2182 | ||
7ca8783a WY |
2183 | err = do_move_pages_to_node(mm, pagelist, node); |
2184 | if (err) { | |
2185 | /* | |
2186 | * Positive err means the number of failed | |
2187 | * pages to migrate. Since we are going to | |
2188 | * abort and return the number of non-migrated | |
ab9dd4f8 | 2189 | * pages, so need to include the rest of the |
7ca8783a WY |
2190 | * nr_pages that have not been attempted as |
2191 | * well. | |
2192 | */ | |
2193 | if (err > 0) | |
a7504ed1 | 2194 | err += nr_pages - i; |
7ca8783a WY |
2195 | return err; |
2196 | } | |
2197 | return store_status(status, start, node, i - start); | |
2198 | } | |
2199 | ||
5e9a0f02 BG |
2200 | /* |
2201 | * Migrate an array of page address onto an array of nodes and fill | |
2202 | * the corresponding array of status. | |
2203 | */ | |
3268c63e | 2204 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
2205 | unsigned long nr_pages, |
2206 | const void __user * __user *pages, | |
2207 | const int __user *nodes, | |
2208 | int __user *status, int flags) | |
2209 | { | |
a49bd4d7 MH |
2210 | int current_node = NUMA_NO_NODE; |
2211 | LIST_HEAD(pagelist); | |
2212 | int start, i; | |
2213 | int err = 0, err1; | |
35282a2d | 2214 | |
361a2a22 | 2215 | lru_cache_disable(); |
35282a2d | 2216 | |
a49bd4d7 MH |
2217 | for (i = start = 0; i < nr_pages; i++) { |
2218 | const void __user *p; | |
2219 | unsigned long addr; | |
2220 | int node; | |
3140a227 | 2221 | |
a49bd4d7 MH |
2222 | err = -EFAULT; |
2223 | if (get_user(p, pages + i)) | |
2224 | goto out_flush; | |
2225 | if (get_user(node, nodes + i)) | |
2226 | goto out_flush; | |
057d3389 | 2227 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
2228 | |
2229 | err = -ENODEV; | |
2230 | if (node < 0 || node >= MAX_NUMNODES) | |
2231 | goto out_flush; | |
2232 | if (!node_state(node, N_MEMORY)) | |
2233 | goto out_flush; | |
5e9a0f02 | 2234 | |
a49bd4d7 MH |
2235 | err = -EACCES; |
2236 | if (!node_isset(node, task_nodes)) | |
2237 | goto out_flush; | |
2238 | ||
2239 | if (current_node == NUMA_NO_NODE) { | |
2240 | current_node = node; | |
2241 | start = i; | |
2242 | } else if (node != current_node) { | |
7ca8783a WY |
2243 | err = move_pages_and_store_status(mm, current_node, |
2244 | &pagelist, status, start, i, nr_pages); | |
a49bd4d7 MH |
2245 | if (err) |
2246 | goto out; | |
2247 | start = i; | |
2248 | current_node = node; | |
3140a227 BG |
2249 | } |
2250 | ||
a49bd4d7 MH |
2251 | /* |
2252 | * Errors in the page lookup or isolation are not fatal and we simply | |
2253 | * report them via status | |
2254 | */ | |
2255 | err = add_page_for_migration(mm, addr, current_node, | |
2256 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 | 2257 | |
d08221a0 | 2258 | if (err > 0) { |
e0153fc2 YS |
2259 | /* The page is successfully queued for migration */ |
2260 | continue; | |
2261 | } | |
3140a227 | 2262 | |
65462462 JH |
2263 | /* |
2264 | * The move_pages() man page does not have an -EEXIST choice, so | |
2265 | * use -EFAULT instead. | |
2266 | */ | |
2267 | if (err == -EEXIST) | |
2268 | err = -EFAULT; | |
2269 | ||
d08221a0 WY |
2270 | /* |
2271 | * If the page is already on the target node (!err), store the | |
2272 | * node, otherwise, store the err. | |
2273 | */ | |
2274 | err = store_status(status, i, err ? : current_node, 1); | |
a49bd4d7 MH |
2275 | if (err) |
2276 | goto out_flush; | |
5e9a0f02 | 2277 | |
7ca8783a WY |
2278 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
2279 | status, start, i, nr_pages); | |
a7504ed1 HY |
2280 | if (err) { |
2281 | /* We have accounted for page i */ | |
2282 | if (err > 0) | |
2283 | err--; | |
4afdacec | 2284 | goto out; |
a7504ed1 | 2285 | } |
a49bd4d7 | 2286 | current_node = NUMA_NO_NODE; |
3140a227 | 2287 | } |
a49bd4d7 MH |
2288 | out_flush: |
2289 | /* Make sure we do not overwrite the existing error */ | |
7ca8783a WY |
2290 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
2291 | status, start, i, nr_pages); | |
dfe9aa23 | 2292 | if (err >= 0) |
a49bd4d7 | 2293 | err = err1; |
5e9a0f02 | 2294 | out: |
361a2a22 | 2295 | lru_cache_enable(); |
5e9a0f02 BG |
2296 | return err; |
2297 | } | |
2298 | ||
742755a1 | 2299 | /* |
2f007e74 | 2300 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 2301 | */ |
80bba129 BG |
2302 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
2303 | const void __user **pages, int *status) | |
742755a1 | 2304 | { |
2f007e74 | 2305 | unsigned long i; |
2f007e74 | 2306 | |
d8ed45c5 | 2307 | mmap_read_lock(mm); |
742755a1 | 2308 | |
2f007e74 | 2309 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 2310 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
2311 | struct vm_area_struct *vma; |
2312 | struct page *page; | |
c095adbc | 2313 | int err = -EFAULT; |
2f007e74 | 2314 | |
059b8b48 LH |
2315 | vma = vma_lookup(mm, addr); |
2316 | if (!vma) | |
742755a1 CL |
2317 | goto set_status; |
2318 | ||
d899844e | 2319 | /* FOLL_DUMP to ignore special (like zero) pages */ |
16fd6b31 | 2320 | page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); |
89f5b7da LT |
2321 | |
2322 | err = PTR_ERR(page); | |
2323 | if (IS_ERR(page)) | |
2324 | goto set_status; | |
2325 | ||
f7091ed6 HW |
2326 | err = -ENOENT; |
2327 | if (!page) | |
2328 | goto set_status; | |
2329 | ||
2330 | if (!is_zone_device_page(page)) | |
4cd61484 | 2331 | err = page_to_nid(page); |
f7091ed6 | 2332 | |
16fd6b31 | 2333 | put_page(page); |
742755a1 | 2334 | set_status: |
80bba129 BG |
2335 | *status = err; |
2336 | ||
2337 | pages++; | |
2338 | status++; | |
2339 | } | |
2340 | ||
d8ed45c5 | 2341 | mmap_read_unlock(mm); |
80bba129 BG |
2342 | } |
2343 | ||
5b1b561b AB |
2344 | static int get_compat_pages_array(const void __user *chunk_pages[], |
2345 | const void __user * __user *pages, | |
2346 | unsigned long chunk_nr) | |
2347 | { | |
2348 | compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; | |
2349 | compat_uptr_t p; | |
2350 | int i; | |
2351 | ||
2352 | for (i = 0; i < chunk_nr; i++) { | |
2353 | if (get_user(p, pages32 + i)) | |
2354 | return -EFAULT; | |
2355 | chunk_pages[i] = compat_ptr(p); | |
2356 | } | |
2357 | ||
2358 | return 0; | |
2359 | } | |
2360 | ||
80bba129 BG |
2361 | /* |
2362 | * Determine the nodes of a user array of pages and store it in | |
2363 | * a user array of status. | |
2364 | */ | |
2365 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
2366 | const void __user * __user *pages, | |
2367 | int __user *status) | |
2368 | { | |
3eefb826 | 2369 | #define DO_PAGES_STAT_CHUNK_NR 16UL |
80bba129 BG |
2370 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; |
2371 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 2372 | |
87b8d1ad | 2373 | while (nr_pages) { |
3eefb826 | 2374 | unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); |
87b8d1ad | 2375 | |
5b1b561b AB |
2376 | if (in_compat_syscall()) { |
2377 | if (get_compat_pages_array(chunk_pages, pages, | |
2378 | chunk_nr)) | |
2379 | break; | |
2380 | } else { | |
2381 | if (copy_from_user(chunk_pages, pages, | |
2382 | chunk_nr * sizeof(*chunk_pages))) | |
2383 | break; | |
2384 | } | |
80bba129 BG |
2385 | |
2386 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
2387 | ||
87b8d1ad PA |
2388 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
2389 | break; | |
742755a1 | 2390 | |
87b8d1ad PA |
2391 | pages += chunk_nr; |
2392 | status += chunk_nr; | |
2393 | nr_pages -= chunk_nr; | |
2394 | } | |
2395 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
2396 | } |
2397 | ||
4dc200ce | 2398 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
742755a1 | 2399 | { |
742755a1 | 2400 | struct task_struct *task; |
742755a1 | 2401 | struct mm_struct *mm; |
742755a1 | 2402 | |
4dc200ce ML |
2403 | /* |
2404 | * There is no need to check if current process has the right to modify | |
2405 | * the specified process when they are same. | |
2406 | */ | |
2407 | if (!pid) { | |
2408 | mmget(current->mm); | |
2409 | *mem_nodes = cpuset_mems_allowed(current); | |
2410 | return current->mm; | |
2411 | } | |
742755a1 CL |
2412 | |
2413 | /* Find the mm_struct */ | |
a879bf58 | 2414 | rcu_read_lock(); |
4dc200ce | 2415 | task = find_task_by_vpid(pid); |
742755a1 | 2416 | if (!task) { |
a879bf58 | 2417 | rcu_read_unlock(); |
4dc200ce | 2418 | return ERR_PTR(-ESRCH); |
742755a1 | 2419 | } |
3268c63e | 2420 | get_task_struct(task); |
742755a1 CL |
2421 | |
2422 | /* | |
2423 | * Check if this process has the right to modify the specified | |
197e7e52 | 2424 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 2425 | */ |
197e7e52 | 2426 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 2427 | rcu_read_unlock(); |
4dc200ce | 2428 | mm = ERR_PTR(-EPERM); |
5e9a0f02 | 2429 | goto out; |
742755a1 | 2430 | } |
c69e8d9c | 2431 | rcu_read_unlock(); |
742755a1 | 2432 | |
4dc200ce ML |
2433 | mm = ERR_PTR(security_task_movememory(task)); |
2434 | if (IS_ERR(mm)) | |
5e9a0f02 | 2435 | goto out; |
4dc200ce | 2436 | *mem_nodes = cpuset_mems_allowed(task); |
3268c63e | 2437 | mm = get_task_mm(task); |
4dc200ce | 2438 | out: |
3268c63e | 2439 | put_task_struct(task); |
6e8b09ea | 2440 | if (!mm) |
4dc200ce ML |
2441 | mm = ERR_PTR(-EINVAL); |
2442 | return mm; | |
2443 | } | |
2444 | ||
2445 | /* | |
2446 | * Move a list of pages in the address space of the currently executing | |
2447 | * process. | |
2448 | */ | |
2449 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
2450 | const void __user * __user *pages, | |
2451 | const int __user *nodes, | |
2452 | int __user *status, int flags) | |
2453 | { | |
2454 | struct mm_struct *mm; | |
2455 | int err; | |
2456 | nodemask_t task_nodes; | |
2457 | ||
2458 | /* Check flags */ | |
2459 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
6e8b09ea SL |
2460 | return -EINVAL; |
2461 | ||
4dc200ce ML |
2462 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
2463 | return -EPERM; | |
2464 | ||
2465 | mm = find_mm_struct(pid, &task_nodes); | |
2466 | if (IS_ERR(mm)) | |
2467 | return PTR_ERR(mm); | |
2468 | ||
6e8b09ea SL |
2469 | if (nodes) |
2470 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
2471 | nodes, status, flags); | |
2472 | else | |
2473 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 2474 | |
742755a1 CL |
2475 | mmput(mm); |
2476 | return err; | |
2477 | } | |
742755a1 | 2478 | |
7addf443 DB |
2479 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
2480 | const void __user * __user *, pages, | |
2481 | const int __user *, nodes, | |
2482 | int __user *, status, int, flags) | |
2483 | { | |
2484 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
2485 | } | |
2486 | ||
7039e1db PZ |
2487 | #ifdef CONFIG_NUMA_BALANCING |
2488 | /* | |
2489 | * Returns true if this is a safe migration target node for misplaced NUMA | |
bc53008e | 2490 | * pages. Currently it only checks the watermarks which is crude. |
7039e1db PZ |
2491 | */ |
2492 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 2493 | unsigned long nr_migrate_pages) |
7039e1db PZ |
2494 | { |
2495 | int z; | |
599d0c95 | 2496 | |
7039e1db PZ |
2497 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
2498 | struct zone *zone = pgdat->node_zones + z; | |
2499 | ||
bc53008e | 2500 | if (!managed_zone(zone)) |
7039e1db PZ |
2501 | continue; |
2502 | ||
7039e1db PZ |
2503 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
2504 | if (!zone_watermark_ok(zone, 0, | |
2505 | high_wmark_pages(zone) + | |
2506 | nr_migrate_pages, | |
bfe9d006 | 2507 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
2508 | continue; |
2509 | return true; | |
2510 | } | |
2511 | return false; | |
2512 | } | |
2513 | ||
2514 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 2515 | unsigned long data) |
7039e1db PZ |
2516 | { |
2517 | int nid = (int) data; | |
c185e494 MWO |
2518 | int order = compound_order(page); |
2519 | gfp_t gfp = __GFP_THISNODE; | |
2520 | struct folio *new; | |
2521 | ||
2522 | if (order > 0) | |
2523 | gfp |= GFP_TRANSHUGE_LIGHT; | |
2524 | else { | |
2525 | gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
2526 | __GFP_NOWARN; | |
2527 | gfp &= ~__GFP_RECLAIM; | |
2528 | } | |
2529 | new = __folio_alloc_node(gfp, order, nid); | |
c5b5a3dd | 2530 | |
c185e494 | 2531 | return &new->page; |
c5b5a3dd YS |
2532 | } |
2533 | ||
1c30e017 | 2534 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 2535 | { |
2b9b624f | 2536 | int nr_pages = thp_nr_pages(page); |
c574bbe9 | 2537 | int order = compound_order(page); |
a8f60772 | 2538 | |
c574bbe9 | 2539 | VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); |
3abef4e6 | 2540 | |
662aeea7 YS |
2541 | /* Do not migrate THP mapped by multiple processes */ |
2542 | if (PageTransHuge(page) && total_mapcount(page) > 1) | |
2543 | return 0; | |
2544 | ||
7039e1db | 2545 | /* Avoid migrating to a node that is nearly full */ |
c574bbe9 HY |
2546 | if (!migrate_balanced_pgdat(pgdat, nr_pages)) { |
2547 | int z; | |
2548 | ||
2549 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) | |
2550 | return 0; | |
2551 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
bc53008e | 2552 | if (managed_zone(pgdat->node_zones + z)) |
c574bbe9 HY |
2553 | break; |
2554 | } | |
2555 | wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); | |
340ef390 | 2556 | return 0; |
c574bbe9 | 2557 | } |
7039e1db | 2558 | |
f7f9c00d | 2559 | if (!isolate_lru_page(page)) |
340ef390 | 2560 | return 0; |
7039e1db | 2561 | |
b75454e1 | 2562 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page), |
2b9b624f | 2563 | nr_pages); |
340ef390 | 2564 | |
149c33e1 | 2565 | /* |
340ef390 HD |
2566 | * Isolating the page has taken another reference, so the |
2567 | * caller's reference can be safely dropped without the page | |
2568 | * disappearing underneath us during migration. | |
149c33e1 MG |
2569 | */ |
2570 | put_page(page); | |
340ef390 | 2571 | return 1; |
b32967ff MG |
2572 | } |
2573 | ||
2574 | /* | |
2575 | * Attempt to migrate a misplaced page to the specified destination | |
2576 | * node. Caller is expected to have an elevated reference count on | |
2577 | * the page that will be dropped by this function before returning. | |
2578 | */ | |
1bc115d8 MG |
2579 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
2580 | int node) | |
b32967ff MG |
2581 | { |
2582 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 2583 | int isolated; |
b32967ff | 2584 | int nr_remaining; |
e39bb6be | 2585 | unsigned int nr_succeeded; |
b32967ff | 2586 | LIST_HEAD(migratepages); |
b5916c02 | 2587 | int nr_pages = thp_nr_pages(page); |
c5b5a3dd | 2588 | |
b32967ff | 2589 | /* |
1bc115d8 MG |
2590 | * Don't migrate file pages that are mapped in multiple processes |
2591 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 2592 | */ |
7ee820ee ML |
2593 | if (page_mapcount(page) != 1 && page_is_file_lru(page) && |
2594 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 2595 | goto out; |
b32967ff | 2596 | |
09a913a7 MG |
2597 | /* |
2598 | * Also do not migrate dirty pages as not all filesystems can move | |
2599 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2600 | */ | |
9de4f22a | 2601 | if (page_is_file_lru(page) && PageDirty(page)) |
09a913a7 MG |
2602 | goto out; |
2603 | ||
b32967ff MG |
2604 | isolated = numamigrate_isolate_page(pgdat, page); |
2605 | if (!isolated) | |
2606 | goto out; | |
2607 | ||
2608 | list_add(&page->lru, &migratepages); | |
c185e494 MWO |
2609 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
2610 | NULL, node, MIGRATE_ASYNC, | |
2611 | MR_NUMA_MISPLACED, &nr_succeeded); | |
b32967ff | 2612 | if (nr_remaining) { |
59c82b70 JK |
2613 | if (!list_empty(&migratepages)) { |
2614 | list_del(&page->lru); | |
c5fc5c3a YS |
2615 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
2616 | page_is_file_lru(page), -nr_pages); | |
59c82b70 JK |
2617 | putback_lru_page(page); |
2618 | } | |
b32967ff | 2619 | isolated = 0; |
e39bb6be HY |
2620 | } |
2621 | if (nr_succeeded) { | |
2622 | count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); | |
2623 | if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) | |
2624 | mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, | |
2625 | nr_succeeded); | |
2626 | } | |
7039e1db | 2627 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2628 | return isolated; |
340ef390 HD |
2629 | |
2630 | out: | |
2631 | put_page(page); | |
2632 | return 0; | |
7039e1db | 2633 | } |
220018d3 | 2634 | #endif /* CONFIG_NUMA_BALANCING */ |
91952440 | 2635 | #endif /* CONFIG_NUMA */ |