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> |
b20a3503 | 53 | |
0d1836c3 MN |
54 | #include <asm/tlbflush.h> |
55 | ||
7b2a2d4a MG |
56 | #include <trace/events/migrate.h> |
57 | ||
b20a3503 CL |
58 | #include "internal.h" |
59 | ||
9e5bcd61 | 60 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 | 61 | { |
68f2736a | 62 | const struct movable_operations *mops; |
bda807d4 MK |
63 | |
64 | /* | |
65 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
66 | * or just got freed under us. | |
67 | * | |
68 | * In case we 'win' a race for a movable page being freed under us and | |
69 | * raise its refcount preventing __free_pages() from doing its job | |
70 | * the put_page() at the end of this block will take care of | |
71 | * release this page, thus avoiding a nasty leakage. | |
72 | */ | |
73 | if (unlikely(!get_page_unless_zero(page))) | |
74 | goto out; | |
75 | ||
76 | /* | |
77 | * Check PageMovable before holding a PG_lock because page's owner | |
78 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 79 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
80 | */ |
81 | if (unlikely(!__PageMovable(page))) | |
82 | goto out_putpage; | |
83 | /* | |
84 | * As movable pages are not isolated from LRU lists, concurrent | |
85 | * compaction threads can race against page migration functions | |
86 | * as well as race against the releasing a page. | |
87 | * | |
88 | * In order to avoid having an already isolated movable page | |
89 | * being (wrongly) re-isolated while it is under migration, | |
90 | * or to avoid attempting to isolate pages being released, | |
91 | * lets be sure we have the page lock | |
92 | * before proceeding with the movable page isolation steps. | |
93 | */ | |
94 | if (unlikely(!trylock_page(page))) | |
95 | goto out_putpage; | |
96 | ||
97 | if (!PageMovable(page) || PageIsolated(page)) | |
98 | goto out_no_isolated; | |
99 | ||
68f2736a MWO |
100 | mops = page_movable_ops(page); |
101 | VM_BUG_ON_PAGE(!mops, page); | |
bda807d4 | 102 | |
68f2736a | 103 | if (!mops->isolate_page(page, mode)) |
bda807d4 MK |
104 | goto out_no_isolated; |
105 | ||
106 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
107 | WARN_ON_ONCE(PageIsolated(page)); | |
356ea386 | 108 | SetPageIsolated(page); |
bda807d4 MK |
109 | unlock_page(page); |
110 | ||
9e5bcd61 | 111 | return 0; |
bda807d4 MK |
112 | |
113 | out_no_isolated: | |
114 | unlock_page(page); | |
115 | out_putpage: | |
116 | put_page(page); | |
117 | out: | |
9e5bcd61 | 118 | return -EBUSY; |
bda807d4 MK |
119 | } |
120 | ||
606a6f71 | 121 | static void putback_movable_page(struct page *page) |
bda807d4 | 122 | { |
68f2736a | 123 | const struct movable_operations *mops = page_movable_ops(page); |
bda807d4 | 124 | |
68f2736a | 125 | mops->putback_page(page); |
356ea386 | 126 | ClearPageIsolated(page); |
bda807d4 MK |
127 | } |
128 | ||
5733c7d1 RA |
129 | /* |
130 | * Put previously isolated pages back onto the appropriate lists | |
131 | * from where they were once taken off for compaction/migration. | |
132 | * | |
59c82b70 JK |
133 | * This function shall be used whenever the isolated pageset has been |
134 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
7ce82f4c | 135 | * and isolate_hugetlb(). |
5733c7d1 RA |
136 | */ |
137 | void putback_movable_pages(struct list_head *l) | |
138 | { | |
139 | struct page *page; | |
140 | struct page *page2; | |
141 | ||
b20a3503 | 142 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
143 | if (unlikely(PageHuge(page))) { |
144 | putback_active_hugepage(page); | |
145 | continue; | |
146 | } | |
e24f0b8f | 147 | list_del(&page->lru); |
bda807d4 MK |
148 | /* |
149 | * We isolated non-lru movable page so here we can use | |
150 | * __PageMovable because LRU page's mapping cannot have | |
151 | * PAGE_MAPPING_MOVABLE. | |
152 | */ | |
b1123ea6 | 153 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
154 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
155 | lock_page(page); | |
156 | if (PageMovable(page)) | |
157 | putback_movable_page(page); | |
158 | else | |
356ea386 | 159 | ClearPageIsolated(page); |
bda807d4 MK |
160 | unlock_page(page); |
161 | put_page(page); | |
162 | } else { | |
e8db67eb | 163 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 164 | page_is_file_lru(page), -thp_nr_pages(page)); |
fc280fe8 | 165 | putback_lru_page(page); |
bda807d4 | 166 | } |
b20a3503 | 167 | } |
b20a3503 CL |
168 | } |
169 | ||
0697212a CL |
170 | /* |
171 | * Restore a potential migration pte to a working pte entry | |
172 | */ | |
2f031c6f MWO |
173 | static bool remove_migration_pte(struct folio *folio, |
174 | struct vm_area_struct *vma, unsigned long addr, void *old) | |
0697212a | 175 | { |
4eecb8b9 | 176 | DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION); |
0697212a | 177 | |
3fe87967 | 178 | while (page_vma_mapped_walk(&pvmw)) { |
6c287605 | 179 | rmap_t rmap_flags = RMAP_NONE; |
4eecb8b9 MWO |
180 | pte_t pte; |
181 | swp_entry_t entry; | |
182 | struct page *new; | |
183 | unsigned long idx = 0; | |
184 | ||
185 | /* pgoff is invalid for ksm pages, but they are never large */ | |
186 | if (folio_test_large(folio) && !folio_test_hugetlb(folio)) | |
187 | idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; | |
188 | new = folio_page(folio, idx); | |
0697212a | 189 | |
616b8371 ZY |
190 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
191 | /* PMD-mapped THP migration entry */ | |
192 | if (!pvmw.pte) { | |
4eecb8b9 MWO |
193 | VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || |
194 | !folio_test_pmd_mappable(folio), folio); | |
616b8371 ZY |
195 | remove_migration_pmd(&pvmw, new); |
196 | continue; | |
197 | } | |
198 | #endif | |
199 | ||
4eecb8b9 | 200 | folio_get(folio); |
2e346877 | 201 | pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); |
3fe87967 KS |
202 | if (pte_swp_soft_dirty(*pvmw.pte)) |
203 | pte = pte_mksoft_dirty(pte); | |
0697212a | 204 | |
3fe87967 KS |
205 | /* |
206 | * Recheck VMA as permissions can change since migration started | |
207 | */ | |
208 | entry = pte_to_swp_entry(*pvmw.pte); | |
2e346877 PX |
209 | if (!is_migration_entry_young(entry)) |
210 | pte = pte_mkold(pte); | |
211 | if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) | |
212 | pte = pte_mkdirty(pte); | |
4dd845b5 | 213 | if (is_writable_migration_entry(entry)) |
3fe87967 | 214 | pte = maybe_mkwrite(pte, vma); |
f45ec5ff PX |
215 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
216 | pte = pte_mkuffd_wp(pte); | |
d3cb8bf6 | 217 | |
6c287605 DH |
218 | if (folio_test_anon(folio) && !is_readable_migration_entry(entry)) |
219 | rmap_flags |= RMAP_EXCLUSIVE; | |
220 | ||
6128763f | 221 | if (unlikely(is_device_private_page(new))) { |
4dd845b5 AP |
222 | if (pte_write(pte)) |
223 | entry = make_writable_device_private_entry( | |
224 | page_to_pfn(new)); | |
225 | else | |
226 | entry = make_readable_device_private_entry( | |
227 | page_to_pfn(new)); | |
6128763f | 228 | pte = swp_entry_to_pte(entry); |
3d321bf8 RC |
229 | if (pte_swp_soft_dirty(*pvmw.pte)) |
230 | pte = pte_swp_mksoft_dirty(pte); | |
6128763f RC |
231 | if (pte_swp_uffd_wp(*pvmw.pte)) |
232 | pte = pte_swp_mkuffd_wp(pte); | |
d2b2c6dd | 233 | } |
a5430dda | 234 | |
3ef8fd7f | 235 | #ifdef CONFIG_HUGETLB_PAGE |
4eecb8b9 | 236 | if (folio_test_hugetlb(folio)) { |
79c1c594 CL |
237 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
238 | ||
3fe87967 | 239 | pte = pte_mkhuge(pte); |
79c1c594 | 240 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
4eecb8b9 | 241 | if (folio_test_anon(folio)) |
28c5209d | 242 | hugepage_add_anon_rmap(new, vma, pvmw.address, |
6c287605 | 243 | rmap_flags); |
3fe87967 | 244 | else |
fb3d824d | 245 | page_dup_file_rmap(new, true); |
1eba86c0 | 246 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab AK |
247 | } else |
248 | #endif | |
249 | { | |
4eecb8b9 | 250 | if (folio_test_anon(folio)) |
f1e2db12 | 251 | page_add_anon_rmap(new, vma, pvmw.address, |
6c287605 | 252 | rmap_flags); |
383321ab | 253 | else |
cea86fe2 | 254 | page_add_file_rmap(new, vma, false); |
1eba86c0 | 255 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab | 256 | } |
b7435507 | 257 | if (vma->vm_flags & VM_LOCKED) |
adb11e78 | 258 | mlock_page_drain_local(); |
e125fe40 | 259 | |
4cc79b33 AK |
260 | trace_remove_migration_pte(pvmw.address, pte_val(pte), |
261 | compound_order(new)); | |
262 | ||
3fe87967 KS |
263 | /* No need to invalidate - it was non-present before */ |
264 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
265 | } | |
51afb12b | 266 | |
e4b82222 | 267 | return true; |
0697212a CL |
268 | } |
269 | ||
04e62a29 CL |
270 | /* |
271 | * Get rid of all migration entries and replace them by | |
272 | * references to the indicated page. | |
273 | */ | |
4eecb8b9 | 274 | void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) |
04e62a29 | 275 | { |
051ac83a JK |
276 | struct rmap_walk_control rwc = { |
277 | .rmap_one = remove_migration_pte, | |
4eecb8b9 | 278 | .arg = src, |
051ac83a JK |
279 | }; |
280 | ||
e388466d | 281 | if (locked) |
2f031c6f | 282 | rmap_walk_locked(dst, &rwc); |
e388466d | 283 | else |
2f031c6f | 284 | rmap_walk(dst, &rwc); |
04e62a29 CL |
285 | } |
286 | ||
0697212a CL |
287 | /* |
288 | * Something used the pte of a page under migration. We need to | |
289 | * get to the page and wait until migration is finished. | |
290 | * When we return from this function the fault will be retried. | |
0697212a | 291 | */ |
e66f17ff | 292 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 293 | spinlock_t *ptl) |
0697212a | 294 | { |
30dad309 | 295 | pte_t pte; |
0697212a | 296 | swp_entry_t entry; |
0697212a | 297 | |
30dad309 | 298 | spin_lock(ptl); |
0697212a CL |
299 | pte = *ptep; |
300 | if (!is_swap_pte(pte)) | |
301 | goto out; | |
302 | ||
303 | entry = pte_to_swp_entry(pte); | |
304 | if (!is_migration_entry(entry)) | |
305 | goto out; | |
306 | ||
ffa65753 | 307 | migration_entry_wait_on_locked(entry, ptep, ptl); |
0697212a CL |
308 | return; |
309 | out: | |
310 | pte_unmap_unlock(ptep, ptl); | |
311 | } | |
312 | ||
30dad309 NH |
313 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
314 | unsigned long address) | |
315 | { | |
316 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
317 | pte_t *ptep = pte_offset_map(pmd, address); | |
318 | __migration_entry_wait(mm, ptep, ptl); | |
319 | } | |
320 | ||
ad1ac596 ML |
321 | #ifdef CONFIG_HUGETLB_PAGE |
322 | void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl) | |
30dad309 | 323 | { |
ad1ac596 ML |
324 | pte_t pte; |
325 | ||
326 | spin_lock(ptl); | |
327 | pte = huge_ptep_get(ptep); | |
328 | ||
329 | if (unlikely(!is_hugetlb_entry_migration(pte))) | |
330 | spin_unlock(ptl); | |
331 | else | |
332 | migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl); | |
30dad309 NH |
333 | } |
334 | ||
ad1ac596 ML |
335 | void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte) |
336 | { | |
337 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte); | |
338 | ||
339 | __migration_entry_wait_huge(pte, ptl); | |
340 | } | |
341 | #endif | |
342 | ||
616b8371 ZY |
343 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
344 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
345 | { | |
346 | spinlock_t *ptl; | |
616b8371 ZY |
347 | |
348 | ptl = pmd_lock(mm, pmd); | |
349 | if (!is_pmd_migration_entry(*pmd)) | |
350 | goto unlock; | |
ffa65753 | 351 | migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); |
616b8371 ZY |
352 | return; |
353 | unlock: | |
354 | spin_unlock(ptl); | |
355 | } | |
356 | #endif | |
357 | ||
108ca835 MWO |
358 | static int folio_expected_refs(struct address_space *mapping, |
359 | struct folio *folio) | |
0b3901b3 | 360 | { |
108ca835 MWO |
361 | int refs = 1; |
362 | if (!mapping) | |
363 | return refs; | |
0b3901b3 | 364 | |
108ca835 MWO |
365 | refs += folio_nr_pages(folio); |
366 | if (folio_test_private(folio)) | |
367 | refs++; | |
368 | ||
369 | return refs; | |
0b3901b3 JK |
370 | } |
371 | ||
b20a3503 | 372 | /* |
c3fcf8a5 | 373 | * Replace the page in the mapping. |
5b5c7120 CL |
374 | * |
375 | * The number of remaining references must be: | |
376 | * 1 for anonymous pages without a mapping | |
377 | * 2 for pages with a mapping | |
266cf658 | 378 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 379 | */ |
3417013e MWO |
380 | int folio_migrate_mapping(struct address_space *mapping, |
381 | struct folio *newfolio, struct folio *folio, int extra_count) | |
b20a3503 | 382 | { |
3417013e | 383 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
42cb14b1 HD |
384 | struct zone *oldzone, *newzone; |
385 | int dirty; | |
108ca835 | 386 | int expected_count = folio_expected_refs(mapping, folio) + extra_count; |
3417013e | 387 | long nr = folio_nr_pages(folio); |
8763cb45 | 388 | |
6c5240ae | 389 | if (!mapping) { |
0e8c7d0f | 390 | /* Anonymous page without mapping */ |
3417013e | 391 | if (folio_ref_count(folio) != expected_count) |
6c5240ae | 392 | return -EAGAIN; |
cf4b769a HD |
393 | |
394 | /* No turning back from here */ | |
3417013e MWO |
395 | newfolio->index = folio->index; |
396 | newfolio->mapping = folio->mapping; | |
397 | if (folio_test_swapbacked(folio)) | |
398 | __folio_set_swapbacked(newfolio); | |
cf4b769a | 399 | |
78bd5209 | 400 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
401 | } |
402 | ||
3417013e MWO |
403 | oldzone = folio_zone(folio); |
404 | newzone = folio_zone(newfolio); | |
42cb14b1 | 405 | |
89eb946a | 406 | xas_lock_irq(&xas); |
3417013e | 407 | if (!folio_ref_freeze(folio, expected_count)) { |
89eb946a | 408 | xas_unlock_irq(&xas); |
e286781d NP |
409 | return -EAGAIN; |
410 | } | |
411 | ||
b20a3503 | 412 | /* |
3417013e | 413 | * Now we know that no one else is looking at the folio: |
cf4b769a | 414 | * no turning back from here. |
b20a3503 | 415 | */ |
3417013e MWO |
416 | newfolio->index = folio->index; |
417 | newfolio->mapping = folio->mapping; | |
418 | folio_ref_add(newfolio, nr); /* add cache reference */ | |
419 | if (folio_test_swapbacked(folio)) { | |
420 | __folio_set_swapbacked(newfolio); | |
421 | if (folio_test_swapcache(folio)) { | |
422 | folio_set_swapcache(newfolio); | |
423 | newfolio->private = folio_get_private(folio); | |
6326fec1 NP |
424 | } |
425 | } else { | |
3417013e | 426 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); |
b20a3503 CL |
427 | } |
428 | ||
42cb14b1 | 429 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
3417013e | 430 | dirty = folio_test_dirty(folio); |
42cb14b1 | 431 | if (dirty) { |
3417013e MWO |
432 | folio_clear_dirty(folio); |
433 | folio_set_dirty(newfolio); | |
42cb14b1 HD |
434 | } |
435 | ||
3417013e | 436 | xas_store(&xas, newfolio); |
7cf9c2c7 NP |
437 | |
438 | /* | |
937a94c9 JG |
439 | * Drop cache reference from old page by unfreezing |
440 | * to one less reference. | |
7cf9c2c7 NP |
441 | * We know this isn't the last reference. |
442 | */ | |
3417013e | 443 | folio_ref_unfreeze(folio, expected_count - nr); |
7cf9c2c7 | 444 | |
89eb946a | 445 | xas_unlock(&xas); |
42cb14b1 HD |
446 | /* Leave irq disabled to prevent preemption while updating stats */ |
447 | ||
0e8c7d0f CL |
448 | /* |
449 | * If moved to a different zone then also account | |
450 | * the page for that zone. Other VM counters will be | |
451 | * taken care of when we establish references to the | |
452 | * new page and drop references to the old page. | |
453 | * | |
454 | * Note that anonymous pages are accounted for | |
4b9d0fab | 455 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
456 | * are mapped to swap space. |
457 | */ | |
42cb14b1 | 458 | if (newzone != oldzone) { |
0d1c2072 JW |
459 | struct lruvec *old_lruvec, *new_lruvec; |
460 | struct mem_cgroup *memcg; | |
461 | ||
3417013e | 462 | memcg = folio_memcg(folio); |
0d1c2072 JW |
463 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); |
464 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
465 | ||
5c447d27 SB |
466 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
467 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); | |
3417013e | 468 | if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { |
5c447d27 SB |
469 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
470 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); | |
42cb14b1 | 471 | } |
b6038942 | 472 | #ifdef CONFIG_SWAP |
3417013e | 473 | if (folio_test_swapcache(folio)) { |
b6038942 SB |
474 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); |
475 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); | |
476 | } | |
477 | #endif | |
f56753ac | 478 | if (dirty && mapping_can_writeback(mapping)) { |
5c447d27 SB |
479 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
480 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); | |
481 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); | |
482 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); | |
42cb14b1 | 483 | } |
4b02108a | 484 | } |
42cb14b1 | 485 | local_irq_enable(); |
b20a3503 | 486 | |
78bd5209 | 487 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 488 | } |
3417013e | 489 | EXPORT_SYMBOL(folio_migrate_mapping); |
b20a3503 | 490 | |
290408d4 NH |
491 | /* |
492 | * The expected number of remaining references is the same as that | |
3417013e | 493 | * of folio_migrate_mapping(). |
290408d4 NH |
494 | */ |
495 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
b890ec2a | 496 | struct folio *dst, struct folio *src) |
290408d4 | 497 | { |
b890ec2a | 498 | XA_STATE(xas, &mapping->i_pages, folio_index(src)); |
290408d4 | 499 | int expected_count; |
290408d4 | 500 | |
89eb946a | 501 | xas_lock_irq(&xas); |
b890ec2a MWO |
502 | expected_count = 2 + folio_has_private(src); |
503 | if (!folio_ref_freeze(src, expected_count)) { | |
89eb946a | 504 | xas_unlock_irq(&xas); |
290408d4 NH |
505 | return -EAGAIN; |
506 | } | |
507 | ||
b890ec2a MWO |
508 | dst->index = src->index; |
509 | dst->mapping = src->mapping; | |
6a93ca8f | 510 | |
b890ec2a | 511 | folio_get(dst); |
290408d4 | 512 | |
b890ec2a | 513 | xas_store(&xas, dst); |
290408d4 | 514 | |
b890ec2a | 515 | folio_ref_unfreeze(src, expected_count - 1); |
290408d4 | 516 | |
89eb946a | 517 | xas_unlock_irq(&xas); |
6a93ca8f | 518 | |
78bd5209 | 519 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
520 | } |
521 | ||
b20a3503 | 522 | /* |
19138349 | 523 | * Copy the flags and some other ancillary information |
b20a3503 | 524 | */ |
19138349 | 525 | void folio_migrate_flags(struct folio *newfolio, struct folio *folio) |
b20a3503 | 526 | { |
7851a45c RR |
527 | int cpupid; |
528 | ||
19138349 MWO |
529 | if (folio_test_error(folio)) |
530 | folio_set_error(newfolio); | |
531 | if (folio_test_referenced(folio)) | |
532 | folio_set_referenced(newfolio); | |
533 | if (folio_test_uptodate(folio)) | |
534 | folio_mark_uptodate(newfolio); | |
535 | if (folio_test_clear_active(folio)) { | |
536 | VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); | |
537 | folio_set_active(newfolio); | |
538 | } else if (folio_test_clear_unevictable(folio)) | |
539 | folio_set_unevictable(newfolio); | |
540 | if (folio_test_workingset(folio)) | |
541 | folio_set_workingset(newfolio); | |
542 | if (folio_test_checked(folio)) | |
543 | folio_set_checked(newfolio); | |
6c287605 DH |
544 | /* |
545 | * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via | |
546 | * migration entries. We can still have PG_anon_exclusive set on an | |
547 | * effectively unmapped and unreferenced first sub-pages of an | |
548 | * anonymous THP: we can simply copy it here via PG_mappedtodisk. | |
549 | */ | |
19138349 MWO |
550 | if (folio_test_mappedtodisk(folio)) |
551 | folio_set_mappedtodisk(newfolio); | |
b20a3503 | 552 | |
3417013e | 553 | /* Move dirty on pages not done by folio_migrate_mapping() */ |
19138349 MWO |
554 | if (folio_test_dirty(folio)) |
555 | folio_set_dirty(newfolio); | |
b20a3503 | 556 | |
19138349 MWO |
557 | if (folio_test_young(folio)) |
558 | folio_set_young(newfolio); | |
559 | if (folio_test_idle(folio)) | |
560 | folio_set_idle(newfolio); | |
33c3fc71 | 561 | |
7851a45c RR |
562 | /* |
563 | * Copy NUMA information to the new page, to prevent over-eager | |
564 | * future migrations of this same page. | |
565 | */ | |
19138349 | 566 | cpupid = page_cpupid_xchg_last(&folio->page, -1); |
33024536 HY |
567 | /* |
568 | * For memory tiering mode, when migrate between slow and fast | |
569 | * memory node, reset cpupid, because that is used to record | |
570 | * page access time in slow memory node. | |
571 | */ | |
572 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { | |
573 | bool f_toptier = node_is_toptier(page_to_nid(&folio->page)); | |
574 | bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page)); | |
575 | ||
576 | if (f_toptier != t_toptier) | |
577 | cpupid = -1; | |
578 | } | |
19138349 | 579 | page_cpupid_xchg_last(&newfolio->page, cpupid); |
7851a45c | 580 | |
19138349 | 581 | folio_migrate_ksm(newfolio, folio); |
c8d6553b HD |
582 | /* |
583 | * Please do not reorder this without considering how mm/ksm.c's | |
584 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
585 | */ | |
19138349 MWO |
586 | if (folio_test_swapcache(folio)) |
587 | folio_clear_swapcache(folio); | |
588 | folio_clear_private(folio); | |
ad2fa371 MS |
589 | |
590 | /* page->private contains hugetlb specific flags */ | |
19138349 MWO |
591 | if (!folio_test_hugetlb(folio)) |
592 | folio->private = NULL; | |
b20a3503 CL |
593 | |
594 | /* | |
595 | * If any waiters have accumulated on the new page then | |
596 | * wake them up. | |
597 | */ | |
19138349 MWO |
598 | if (folio_test_writeback(newfolio)) |
599 | folio_end_writeback(newfolio); | |
d435edca | 600 | |
6aeff241 YS |
601 | /* |
602 | * PG_readahead shares the same bit with PG_reclaim. The above | |
603 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
604 | * bit after that. | |
605 | */ | |
19138349 MWO |
606 | if (folio_test_readahead(folio)) |
607 | folio_set_readahead(newfolio); | |
6aeff241 | 608 | |
19138349 | 609 | folio_copy_owner(newfolio, folio); |
74485cf2 | 610 | |
19138349 | 611 | if (!folio_test_hugetlb(folio)) |
d21bba2b | 612 | mem_cgroup_migrate(folio, newfolio); |
b20a3503 | 613 | } |
19138349 | 614 | EXPORT_SYMBOL(folio_migrate_flags); |
2916ecc0 | 615 | |
715cbfd6 | 616 | void folio_migrate_copy(struct folio *newfolio, struct folio *folio) |
2916ecc0 | 617 | { |
715cbfd6 MWO |
618 | folio_copy(newfolio, folio); |
619 | folio_migrate_flags(newfolio, folio); | |
2916ecc0 | 620 | } |
715cbfd6 | 621 | EXPORT_SYMBOL(folio_migrate_copy); |
b20a3503 | 622 | |
1d8b85cc CL |
623 | /************************************************************ |
624 | * Migration functions | |
625 | ***********************************************************/ | |
626 | ||
54184650 MWO |
627 | /** |
628 | * migrate_folio() - Simple folio migration. | |
629 | * @mapping: The address_space containing the folio. | |
630 | * @dst: The folio to migrate the data to. | |
631 | * @src: The folio containing the current data. | |
632 | * @mode: How to migrate the page. | |
633 | * | |
634 | * Common logic to directly migrate a single LRU folio suitable for | |
635 | * folios that do not use PagePrivate/PagePrivate2. | |
b20a3503 | 636 | * |
54184650 | 637 | * Folios are locked upon entry and exit. |
b20a3503 | 638 | */ |
54184650 MWO |
639 | int migrate_folio(struct address_space *mapping, struct folio *dst, |
640 | struct folio *src, enum migrate_mode mode) | |
b20a3503 CL |
641 | { |
642 | int rc; | |
643 | ||
54184650 | 644 | BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ |
b20a3503 | 645 | |
54184650 | 646 | rc = folio_migrate_mapping(mapping, dst, src, 0); |
b20a3503 | 647 | |
78bd5209 | 648 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
649 | return rc; |
650 | ||
2916ecc0 | 651 | if (mode != MIGRATE_SYNC_NO_COPY) |
54184650 | 652 | folio_migrate_copy(dst, src); |
2916ecc0 | 653 | else |
54184650 | 654 | folio_migrate_flags(dst, src); |
78bd5209 | 655 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 656 | } |
54184650 | 657 | EXPORT_SYMBOL(migrate_folio); |
b20a3503 | 658 | |
9361401e | 659 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
660 | /* Returns true if all buffers are successfully locked */ |
661 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
662 | enum migrate_mode mode) | |
663 | { | |
664 | struct buffer_head *bh = head; | |
665 | ||
666 | /* Simple case, sync compaction */ | |
667 | if (mode != MIGRATE_ASYNC) { | |
668 | do { | |
84ade7c1 JK |
669 | lock_buffer(bh); |
670 | bh = bh->b_this_page; | |
671 | ||
672 | } while (bh != head); | |
673 | ||
674 | return true; | |
675 | } | |
676 | ||
677 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
678 | do { | |
84ade7c1 JK |
679 | if (!trylock_buffer(bh)) { |
680 | /* | |
681 | * We failed to lock the buffer and cannot stall in | |
682 | * async migration. Release the taken locks | |
683 | */ | |
684 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
685 | bh = head; |
686 | while (bh != failed_bh) { | |
687 | unlock_buffer(bh); | |
84ade7c1 JK |
688 | bh = bh->b_this_page; |
689 | } | |
690 | return false; | |
691 | } | |
692 | ||
693 | bh = bh->b_this_page; | |
694 | } while (bh != head); | |
695 | return true; | |
696 | } | |
697 | ||
67235182 MWO |
698 | static int __buffer_migrate_folio(struct address_space *mapping, |
699 | struct folio *dst, struct folio *src, enum migrate_mode mode, | |
89cb0888 | 700 | bool check_refs) |
1d8b85cc | 701 | { |
1d8b85cc CL |
702 | struct buffer_head *bh, *head; |
703 | int rc; | |
cc4f11e6 | 704 | int expected_count; |
1d8b85cc | 705 | |
67235182 MWO |
706 | head = folio_buffers(src); |
707 | if (!head) | |
54184650 | 708 | return migrate_folio(mapping, dst, src, mode); |
1d8b85cc | 709 | |
cc4f11e6 | 710 | /* Check whether page does not have extra refs before we do more work */ |
108ca835 | 711 | expected_count = folio_expected_refs(mapping, src); |
67235182 | 712 | if (folio_ref_count(src) != expected_count) |
cc4f11e6 | 713 | return -EAGAIN; |
1d8b85cc | 714 | |
cc4f11e6 JK |
715 | if (!buffer_migrate_lock_buffers(head, mode)) |
716 | return -EAGAIN; | |
1d8b85cc | 717 | |
89cb0888 JK |
718 | if (check_refs) { |
719 | bool busy; | |
720 | bool invalidated = false; | |
721 | ||
722 | recheck_buffers: | |
723 | busy = false; | |
724 | spin_lock(&mapping->private_lock); | |
725 | bh = head; | |
726 | do { | |
727 | if (atomic_read(&bh->b_count)) { | |
728 | busy = true; | |
729 | break; | |
730 | } | |
731 | bh = bh->b_this_page; | |
732 | } while (bh != head); | |
89cb0888 JK |
733 | if (busy) { |
734 | if (invalidated) { | |
735 | rc = -EAGAIN; | |
736 | goto unlock_buffers; | |
737 | } | |
ebdf4de5 | 738 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
739 | invalidate_bh_lrus(); |
740 | invalidated = true; | |
741 | goto recheck_buffers; | |
742 | } | |
743 | } | |
744 | ||
67235182 | 745 | rc = folio_migrate_mapping(mapping, dst, src, 0); |
78bd5209 | 746 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 747 | goto unlock_buffers; |
1d8b85cc | 748 | |
67235182 | 749 | folio_attach_private(dst, folio_detach_private(src)); |
1d8b85cc CL |
750 | |
751 | bh = head; | |
752 | do { | |
67235182 | 753 | set_bh_page(bh, &dst->page, bh_offset(bh)); |
1d8b85cc | 754 | bh = bh->b_this_page; |
1d8b85cc CL |
755 | } while (bh != head); |
756 | ||
2916ecc0 | 757 | if (mode != MIGRATE_SYNC_NO_COPY) |
67235182 | 758 | folio_migrate_copy(dst, src); |
2916ecc0 | 759 | else |
67235182 | 760 | folio_migrate_flags(dst, src); |
1d8b85cc | 761 | |
cc4f11e6 JK |
762 | rc = MIGRATEPAGE_SUCCESS; |
763 | unlock_buffers: | |
ebdf4de5 JK |
764 | if (check_refs) |
765 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
766 | bh = head; |
767 | do { | |
768 | unlock_buffer(bh); | |
1d8b85cc | 769 | bh = bh->b_this_page; |
1d8b85cc CL |
770 | } while (bh != head); |
771 | ||
cc4f11e6 | 772 | return rc; |
1d8b85cc | 773 | } |
89cb0888 | 774 | |
67235182 MWO |
775 | /** |
776 | * buffer_migrate_folio() - Migration function for folios with buffers. | |
777 | * @mapping: The address space containing @src. | |
778 | * @dst: The folio to migrate to. | |
779 | * @src: The folio to migrate from. | |
780 | * @mode: How to migrate the folio. | |
781 | * | |
782 | * This function can only be used if the underlying filesystem guarantees | |
783 | * that no other references to @src exist. For example attached buffer | |
784 | * heads are accessed only under the folio lock. If your filesystem cannot | |
785 | * provide this guarantee, buffer_migrate_folio_norefs() may be more | |
786 | * appropriate. | |
787 | * | |
788 | * Return: 0 on success or a negative errno on failure. | |
89cb0888 | 789 | */ |
67235182 MWO |
790 | int buffer_migrate_folio(struct address_space *mapping, |
791 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
89cb0888 | 792 | { |
67235182 | 793 | return __buffer_migrate_folio(mapping, dst, src, mode, false); |
89cb0888 | 794 | } |
67235182 MWO |
795 | EXPORT_SYMBOL(buffer_migrate_folio); |
796 | ||
797 | /** | |
798 | * buffer_migrate_folio_norefs() - Migration function for folios with buffers. | |
799 | * @mapping: The address space containing @src. | |
800 | * @dst: The folio to migrate to. | |
801 | * @src: The folio to migrate from. | |
802 | * @mode: How to migrate the folio. | |
803 | * | |
804 | * Like buffer_migrate_folio() except that this variant is more careful | |
805 | * and checks that there are also no buffer head references. This function | |
806 | * is the right one for mappings where buffer heads are directly looked | |
807 | * up and referenced (such as block device mappings). | |
808 | * | |
809 | * Return: 0 on success or a negative errno on failure. | |
89cb0888 | 810 | */ |
67235182 MWO |
811 | int buffer_migrate_folio_norefs(struct address_space *mapping, |
812 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
89cb0888 | 813 | { |
67235182 | 814 | return __buffer_migrate_folio(mapping, dst, src, mode, true); |
89cb0888 | 815 | } |
9361401e | 816 | #endif |
1d8b85cc | 817 | |
2ec810d5 MWO |
818 | int filemap_migrate_folio(struct address_space *mapping, |
819 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
820 | { | |
821 | int ret; | |
822 | ||
823 | ret = folio_migrate_mapping(mapping, dst, src, 0); | |
824 | if (ret != MIGRATEPAGE_SUCCESS) | |
825 | return ret; | |
826 | ||
827 | if (folio_get_private(src)) | |
828 | folio_attach_private(dst, folio_detach_private(src)); | |
829 | ||
830 | if (mode != MIGRATE_SYNC_NO_COPY) | |
831 | folio_migrate_copy(dst, src); | |
832 | else | |
833 | folio_migrate_flags(dst, src); | |
834 | return MIGRATEPAGE_SUCCESS; | |
835 | } | |
836 | EXPORT_SYMBOL_GPL(filemap_migrate_folio); | |
837 | ||
04e62a29 | 838 | /* |
2be7fa10 | 839 | * Writeback a folio to clean the dirty state |
04e62a29 | 840 | */ |
2be7fa10 | 841 | static int writeout(struct address_space *mapping, struct folio *folio) |
8351a6e4 | 842 | { |
04e62a29 CL |
843 | struct writeback_control wbc = { |
844 | .sync_mode = WB_SYNC_NONE, | |
845 | .nr_to_write = 1, | |
846 | .range_start = 0, | |
847 | .range_end = LLONG_MAX, | |
04e62a29 CL |
848 | .for_reclaim = 1 |
849 | }; | |
850 | int rc; | |
851 | ||
852 | if (!mapping->a_ops->writepage) | |
853 | /* No write method for the address space */ | |
854 | return -EINVAL; | |
855 | ||
2be7fa10 | 856 | if (!folio_clear_dirty_for_io(folio)) |
04e62a29 CL |
857 | /* Someone else already triggered a write */ |
858 | return -EAGAIN; | |
859 | ||
8351a6e4 | 860 | /* |
2be7fa10 MWO |
861 | * A dirty folio may imply that the underlying filesystem has |
862 | * the folio on some queue. So the folio must be clean for | |
863 | * migration. Writeout may mean we lose the lock and the | |
864 | * folio state is no longer what we checked for earlier. | |
04e62a29 CL |
865 | * At this point we know that the migration attempt cannot |
866 | * be successful. | |
8351a6e4 | 867 | */ |
4eecb8b9 | 868 | remove_migration_ptes(folio, folio, false); |
8351a6e4 | 869 | |
2be7fa10 | 870 | rc = mapping->a_ops->writepage(&folio->page, &wbc); |
8351a6e4 | 871 | |
04e62a29 CL |
872 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
873 | /* unlocked. Relock */ | |
2be7fa10 | 874 | folio_lock(folio); |
04e62a29 | 875 | |
bda8550d | 876 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
877 | } |
878 | ||
879 | /* | |
880 | * Default handling if a filesystem does not provide a migration function. | |
881 | */ | |
8faa8ef5 MWO |
882 | static int fallback_migrate_folio(struct address_space *mapping, |
883 | struct folio *dst, struct folio *src, enum migrate_mode mode) | |
04e62a29 | 884 | { |
8faa8ef5 MWO |
885 | if (folio_test_dirty(src)) { |
886 | /* Only writeback folios in full synchronous migration */ | |
2916ecc0 JG |
887 | switch (mode) { |
888 | case MIGRATE_SYNC: | |
889 | case MIGRATE_SYNC_NO_COPY: | |
890 | break; | |
891 | default: | |
b969c4ab | 892 | return -EBUSY; |
2916ecc0 | 893 | } |
2be7fa10 | 894 | return writeout(mapping, src); |
b969c4ab | 895 | } |
8351a6e4 CL |
896 | |
897 | /* | |
898 | * Buffers may be managed in a filesystem specific way. | |
899 | * We must have no buffers or drop them. | |
900 | */ | |
8faa8ef5 MWO |
901 | if (folio_test_private(src) && |
902 | !filemap_release_folio(src, GFP_KERNEL)) | |
806031bb | 903 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 904 | |
54184650 | 905 | return migrate_folio(mapping, dst, src, mode); |
8351a6e4 CL |
906 | } |
907 | ||
e24f0b8f CL |
908 | /* |
909 | * Move a page to a newly allocated page | |
910 | * The page is locked and all ptes have been successfully removed. | |
911 | * | |
912 | * The new page will have replaced the old page if this function | |
913 | * is successful. | |
894bc310 LS |
914 | * |
915 | * Return value: | |
916 | * < 0 - error code | |
78bd5209 | 917 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 918 | */ |
e7e3ffeb | 919 | static int move_to_new_folio(struct folio *dst, struct folio *src, |
5c3f9a67 | 920 | enum migrate_mode mode) |
e24f0b8f | 921 | { |
bda807d4 | 922 | int rc = -EAGAIN; |
e7e3ffeb | 923 | bool is_lru = !__PageMovable(&src->page); |
e24f0b8f | 924 | |
e7e3ffeb MWO |
925 | VM_BUG_ON_FOLIO(!folio_test_locked(src), src); |
926 | VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); | |
e24f0b8f | 927 | |
bda807d4 | 928 | if (likely(is_lru)) { |
68f2736a MWO |
929 | struct address_space *mapping = folio_mapping(src); |
930 | ||
bda807d4 | 931 | if (!mapping) |
54184650 | 932 | rc = migrate_folio(mapping, dst, src, mode); |
5490da4f | 933 | else if (mapping->a_ops->migrate_folio) |
bda807d4 | 934 | /* |
5490da4f MWO |
935 | * Most folios have a mapping and most filesystems |
936 | * provide a migrate_folio callback. Anonymous folios | |
bda807d4 | 937 | * are part of swap space which also has its own |
5490da4f | 938 | * migrate_folio callback. This is the most common path |
bda807d4 MK |
939 | * for page migration. |
940 | */ | |
5490da4f MWO |
941 | rc = mapping->a_ops->migrate_folio(mapping, dst, src, |
942 | mode); | |
bda807d4 | 943 | else |
8faa8ef5 | 944 | rc = fallback_migrate_folio(mapping, dst, src, mode); |
bda807d4 | 945 | } else { |
68f2736a MWO |
946 | const struct movable_operations *mops; |
947 | ||
e24f0b8f | 948 | /* |
bda807d4 MK |
949 | * In case of non-lru page, it could be released after |
950 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 951 | */ |
e7e3ffeb MWO |
952 | VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); |
953 | if (!folio_test_movable(src)) { | |
bda807d4 | 954 | rc = MIGRATEPAGE_SUCCESS; |
e7e3ffeb | 955 | folio_clear_isolated(src); |
bda807d4 MK |
956 | goto out; |
957 | } | |
958 | ||
68f2736a MWO |
959 | mops = page_movable_ops(&src->page); |
960 | rc = mops->migrate_page(&dst->page, &src->page, mode); | |
bda807d4 | 961 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && |
e7e3ffeb | 962 | !folio_test_isolated(src)); |
bda807d4 | 963 | } |
e24f0b8f | 964 | |
5c3f9a67 | 965 | /* |
e7e3ffeb MWO |
966 | * When successful, old pagecache src->mapping must be cleared before |
967 | * src is freed; but stats require that PageAnon be left as PageAnon. | |
5c3f9a67 HD |
968 | */ |
969 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e7e3ffeb MWO |
970 | if (__PageMovable(&src->page)) { |
971 | VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); | |
bda807d4 MK |
972 | |
973 | /* | |
974 | * We clear PG_movable under page_lock so any compactor | |
975 | * cannot try to migrate this page. | |
976 | */ | |
e7e3ffeb | 977 | folio_clear_isolated(src); |
bda807d4 MK |
978 | } |
979 | ||
980 | /* | |
e7e3ffeb | 981 | * Anonymous and movable src->mapping will be cleared by |
bda807d4 MK |
982 | * free_pages_prepare so don't reset it here for keeping |
983 | * the type to work PageAnon, for example. | |
984 | */ | |
e7e3ffeb MWO |
985 | if (!folio_mapping_flags(src)) |
986 | src->mapping = NULL; | |
d2b2c6dd | 987 | |
e7e3ffeb MWO |
988 | if (likely(!folio_is_zone_device(dst))) |
989 | flush_dcache_folio(dst); | |
3fe2011f | 990 | } |
bda807d4 | 991 | out: |
e24f0b8f CL |
992 | return rc; |
993 | } | |
994 | ||
0dabec93 | 995 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 996 | int force, enum migrate_mode mode) |
e24f0b8f | 997 | { |
4b8554c5 | 998 | struct folio *folio = page_folio(page); |
4eecb8b9 | 999 | struct folio *dst = page_folio(newpage); |
0dabec93 | 1000 | int rc = -EAGAIN; |
213ecb31 | 1001 | bool page_was_mapped = false; |
3f6c8272 | 1002 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 1003 | bool is_lru = !__PageMovable(page); |
95a402c3 | 1004 | |
529ae9aa | 1005 | if (!trylock_page(page)) { |
a6bc32b8 | 1006 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 1007 | goto out; |
3e7d3449 MG |
1008 | |
1009 | /* | |
1010 | * It's not safe for direct compaction to call lock_page. | |
1011 | * For example, during page readahead pages are added locked | |
1012 | * to the LRU. Later, when the IO completes the pages are | |
1013 | * marked uptodate and unlocked. However, the queueing | |
1014 | * could be merging multiple pages for one bio (e.g. | |
d4388340 | 1015 | * mpage_readahead). If an allocation happens for the |
3e7d3449 MG |
1016 | * second or third page, the process can end up locking |
1017 | * the same page twice and deadlocking. Rather than | |
1018 | * trying to be clever about what pages can be locked, | |
1019 | * avoid the use of lock_page for direct compaction | |
1020 | * altogether. | |
1021 | */ | |
1022 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1023 | goto out; |
3e7d3449 | 1024 | |
e24f0b8f CL |
1025 | lock_page(page); |
1026 | } | |
1027 | ||
1028 | if (PageWriteback(page)) { | |
11bc82d6 | 1029 | /* |
fed5b64a | 1030 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1031 | * necessary to wait for PageWriteback. In the async case, |
1032 | * the retry loop is too short and in the sync-light case, | |
1033 | * the overhead of stalling is too much | |
11bc82d6 | 1034 | */ |
2916ecc0 JG |
1035 | switch (mode) { |
1036 | case MIGRATE_SYNC: | |
1037 | case MIGRATE_SYNC_NO_COPY: | |
1038 | break; | |
1039 | default: | |
11bc82d6 | 1040 | rc = -EBUSY; |
0a31bc97 | 1041 | goto out_unlock; |
11bc82d6 AA |
1042 | } |
1043 | if (!force) | |
0a31bc97 | 1044 | goto out_unlock; |
e24f0b8f CL |
1045 | wait_on_page_writeback(page); |
1046 | } | |
03f15c86 | 1047 | |
e24f0b8f | 1048 | /* |
68a9843f | 1049 | * By try_to_migrate(), page->mapcount goes down to 0 here. In this case, |
dc386d4d | 1050 | * we cannot notice that anon_vma is freed while we migrates a page. |
1ce82b69 | 1051 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1052 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1053 | * File Caches may use write_page() or lock_page() in migration, then, |
1054 | * just care Anon page here. | |
03f15c86 HD |
1055 | * |
1056 | * Only page_get_anon_vma() understands the subtleties of | |
1057 | * getting a hold on an anon_vma from outside one of its mms. | |
1058 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1059 | * because that implies that the anon page is no longer mapped | |
1060 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1061 | */ |
03f15c86 | 1062 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1063 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1064 | |
7db7671f HD |
1065 | /* |
1066 | * Block others from accessing the new page when we get around to | |
1067 | * establishing additional references. We are usually the only one | |
1068 | * holding a reference to newpage at this point. We used to have a BUG | |
1069 | * here if trylock_page(newpage) fails, but would like to allow for | |
1070 | * cases where there might be a race with the previous use of newpage. | |
1071 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1072 | */ | |
1073 | if (unlikely(!trylock_page(newpage))) | |
1074 | goto out_unlock; | |
1075 | ||
bda807d4 | 1076 | if (unlikely(!is_lru)) { |
e7e3ffeb | 1077 | rc = move_to_new_folio(dst, folio, mode); |
bda807d4 MK |
1078 | goto out_unlock_both; |
1079 | } | |
1080 | ||
dc386d4d | 1081 | /* |
62e1c553 SL |
1082 | * Corner case handling: |
1083 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1084 | * and treated as swapcache but it has no rmap yet. | |
1085 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1086 | * trigger a BUG. So handle it here. | |
d12b8951 | 1087 | * 2. An orphaned page (see truncate_cleanup_page) might have |
62e1c553 SL |
1088 | * fs-private metadata. The page can be picked up due to memory |
1089 | * offlining. Everywhere else except page reclaim, the page is | |
1090 | * invisible to the vm, so the page can not be migrated. So try to | |
1091 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1092 | */ |
62e1c553 | 1093 | if (!page->mapping) { |
309381fe | 1094 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1095 | if (page_has_private(page)) { |
68189fef | 1096 | try_to_free_buffers(folio); |
7db7671f | 1097 | goto out_unlock_both; |
62e1c553 | 1098 | } |
7db7671f HD |
1099 | } else if (page_mapped(page)) { |
1100 | /* Establish migration ptes */ | |
03f15c86 HD |
1101 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1102 | page); | |
4b8554c5 | 1103 | try_to_migrate(folio, 0); |
213ecb31 | 1104 | page_was_mapped = true; |
2ebba6b7 | 1105 | } |
dc386d4d | 1106 | |
e6a1530d | 1107 | if (!page_mapped(page)) |
e7e3ffeb | 1108 | rc = move_to_new_folio(dst, folio, mode); |
e24f0b8f | 1109 | |
c3096e67 HD |
1110 | /* |
1111 | * When successful, push newpage to LRU immediately: so that if it | |
1112 | * turns out to be an mlocked page, remove_migration_ptes() will | |
1113 | * automatically build up the correct newpage->mlock_count for it. | |
1114 | * | |
1115 | * We would like to do something similar for the old page, when | |
1116 | * unsuccessful, and other cases when a page has been temporarily | |
1117 | * isolated from the unevictable LRU: but this case is the easiest. | |
1118 | */ | |
1119 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1120 | lru_cache_add(newpage); | |
1121 | if (page_was_mapped) | |
1122 | lru_add_drain(); | |
1123 | } | |
1124 | ||
5c3f9a67 | 1125 | if (page_was_mapped) |
4eecb8b9 MWO |
1126 | remove_migration_ptes(folio, |
1127 | rc == MIGRATEPAGE_SUCCESS ? dst : folio, false); | |
3f6c8272 | 1128 | |
7db7671f HD |
1129 | out_unlock_both: |
1130 | unlock_page(newpage); | |
1131 | out_unlock: | |
3f6c8272 | 1132 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1133 | if (anon_vma) |
9e60109f | 1134 | put_anon_vma(anon_vma); |
e24f0b8f | 1135 | unlock_page(page); |
0dabec93 | 1136 | out: |
c6c919eb | 1137 | /* |
c3096e67 | 1138 | * If migration is successful, decrease refcount of the newpage, |
c6c919eb | 1139 | * which will not free the page because new page owner increased |
c3096e67 | 1140 | * refcounter. |
c6c919eb | 1141 | */ |
c3096e67 HD |
1142 | if (rc == MIGRATEPAGE_SUCCESS) |
1143 | put_page(newpage); | |
c6c919eb | 1144 | |
0dabec93 MK |
1145 | return rc; |
1146 | } | |
95a402c3 | 1147 | |
0dabec93 MK |
1148 | /* |
1149 | * Obtain the lock on page, remove all ptes and migrate the page | |
1150 | * to the newly allocated page in newpage. | |
1151 | */ | |
6ec4476a | 1152 | static int unmap_and_move(new_page_t get_new_page, |
ef2a5153 GU |
1153 | free_page_t put_new_page, |
1154 | unsigned long private, struct page *page, | |
add05cec | 1155 | int force, enum migrate_mode mode, |
dd4ae78a YS |
1156 | enum migrate_reason reason, |
1157 | struct list_head *ret) | |
0dabec93 | 1158 | { |
2def7424 | 1159 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1160 | struct page *newpage = NULL; |
0dabec93 | 1161 | |
94723aaf | 1162 | if (!thp_migration_supported() && PageTransHuge(page)) |
d532e2e5 | 1163 | return -ENOSYS; |
94723aaf | 1164 | |
0dabec93 | 1165 | if (page_count(page) == 1) { |
160088b3 | 1166 | /* Page was freed from under us. So we are done. */ |
c6c919eb MK |
1167 | ClearPageActive(page); |
1168 | ClearPageUnevictable(page); | |
160088b3 | 1169 | /* free_pages_prepare() will clear PG_isolated. */ |
0dabec93 MK |
1170 | goto out; |
1171 | } | |
1172 | ||
74d4a579 YS |
1173 | newpage = get_new_page(page, private); |
1174 | if (!newpage) | |
1175 | return -ENOMEM; | |
1176 | ||
b653db77 | 1177 | newpage->private = 0; |
9c620e2b | 1178 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1179 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1180 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1181 | |
0dabec93 | 1182 | out: |
e24f0b8f | 1183 | if (rc != -EAGAIN) { |
0dabec93 MK |
1184 | /* |
1185 | * A page that has been migrated has all references | |
1186 | * removed and will be freed. A page that has not been | |
c23a0c99 | 1187 | * migrated will have kept its references and be restored. |
0dabec93 MK |
1188 | */ |
1189 | list_del(&page->lru); | |
dd4ae78a | 1190 | } |
6afcf8ef | 1191 | |
dd4ae78a YS |
1192 | /* |
1193 | * If migration is successful, releases reference grabbed during | |
1194 | * isolation. Otherwise, restore the page to right list unless | |
1195 | * we want to retry. | |
1196 | */ | |
1197 | if (rc == MIGRATEPAGE_SUCCESS) { | |
6afcf8ef ML |
1198 | /* |
1199 | * Compaction can migrate also non-LRU pages which are | |
1200 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1201 | * as __PageMovable | |
1202 | */ | |
1203 | if (likely(!__PageMovable(page))) | |
e8db67eb | 1204 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 1205 | page_is_file_lru(page), -thp_nr_pages(page)); |
c6c919eb | 1206 | |
79f5f8fa | 1207 | if (reason != MR_MEMORY_FAILURE) |
d7e69488 | 1208 | /* |
79f5f8fa | 1209 | * We release the page in page_handle_poison. |
d7e69488 | 1210 | */ |
79f5f8fa | 1211 | put_page(page); |
c6c919eb | 1212 | } else { |
dd4ae78a YS |
1213 | if (rc != -EAGAIN) |
1214 | list_add_tail(&page->lru, ret); | |
bda807d4 | 1215 | |
c6c919eb MK |
1216 | if (put_new_page) |
1217 | put_new_page(newpage, private); | |
1218 | else | |
1219 | put_page(newpage); | |
e24f0b8f | 1220 | } |
68711a74 | 1221 | |
e24f0b8f CL |
1222 | return rc; |
1223 | } | |
1224 | ||
290408d4 NH |
1225 | /* |
1226 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1227 | * | |
1228 | * This function doesn't wait the completion of hugepage I/O | |
1229 | * because there is no race between I/O and migration for hugepage. | |
1230 | * Note that currently hugepage I/O occurs only in direct I/O | |
1231 | * where no lock is held and PG_writeback is irrelevant, | |
1232 | * and writeback status of all subpages are counted in the reference | |
1233 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1234 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1235 | * This means that when we try to migrate hugepage whose subpages are | |
1236 | * doing direct I/O, some references remain after try_to_unmap() and | |
1237 | * hugepage migration fails without data corruption. | |
1238 | * | |
1239 | * There is also no race when direct I/O is issued on the page under migration, | |
1240 | * because then pte is replaced with migration swap entry and direct I/O code | |
1241 | * will wait in the page fault for migration to complete. | |
1242 | */ | |
1243 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1244 | free_page_t put_new_page, unsigned long private, |
1245 | struct page *hpage, int force, | |
dd4ae78a YS |
1246 | enum migrate_mode mode, int reason, |
1247 | struct list_head *ret) | |
290408d4 | 1248 | { |
4eecb8b9 | 1249 | struct folio *dst, *src = page_folio(hpage); |
2def7424 | 1250 | int rc = -EAGAIN; |
2ebba6b7 | 1251 | int page_was_mapped = 0; |
32665f2b | 1252 | struct page *new_hpage; |
290408d4 | 1253 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1254 | struct address_space *mapping = NULL; |
290408d4 | 1255 | |
83467efb | 1256 | /* |
7ed2c31d | 1257 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1258 | * This check is necessary because some callers of hugepage migration |
1259 | * like soft offline and memory hotremove don't walk through page | |
1260 | * tables or check whether the hugepage is pmd-based or not before | |
1261 | * kicking migration. | |
1262 | */ | |
100873d7 | 1263 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
dd4ae78a | 1264 | list_move_tail(&hpage->lru, ret); |
83467efb | 1265 | return -ENOSYS; |
32665f2b | 1266 | } |
83467efb | 1267 | |
71a64f61 MS |
1268 | if (page_count(hpage) == 1) { |
1269 | /* page was freed from under us. So we are done. */ | |
1270 | putback_active_hugepage(hpage); | |
1271 | return MIGRATEPAGE_SUCCESS; | |
1272 | } | |
1273 | ||
666feb21 | 1274 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1275 | if (!new_hpage) |
1276 | return -ENOMEM; | |
4eecb8b9 | 1277 | dst = page_folio(new_hpage); |
290408d4 | 1278 | |
290408d4 | 1279 | if (!trylock_page(hpage)) { |
2916ecc0 | 1280 | if (!force) |
290408d4 | 1281 | goto out; |
2916ecc0 JG |
1282 | switch (mode) { |
1283 | case MIGRATE_SYNC: | |
1284 | case MIGRATE_SYNC_NO_COPY: | |
1285 | break; | |
1286 | default: | |
1287 | goto out; | |
1288 | } | |
290408d4 NH |
1289 | lock_page(hpage); |
1290 | } | |
1291 | ||
cb6acd01 MK |
1292 | /* |
1293 | * Check for pages which are in the process of being freed. Without | |
1294 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1295 | * be called and we could leak usage counts for subpools. | |
1296 | */ | |
6acfb5ba | 1297 | if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) { |
cb6acd01 MK |
1298 | rc = -EBUSY; |
1299 | goto out_unlock; | |
1300 | } | |
1301 | ||
746b18d4 PZ |
1302 | if (PageAnon(hpage)) |
1303 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1304 | |
7db7671f HD |
1305 | if (unlikely(!trylock_page(new_hpage))) |
1306 | goto put_anon; | |
1307 | ||
2ebba6b7 | 1308 | if (page_mapped(hpage)) { |
a98a2f0c | 1309 | enum ttu_flags ttu = 0; |
336bf30e MK |
1310 | |
1311 | if (!PageAnon(hpage)) { | |
1312 | /* | |
1313 | * In shared mappings, try_to_unmap could potentially | |
1314 | * call huge_pmd_unshare. Because of this, take | |
1315 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1316 | * to let lower levels know we have taken the lock. | |
1317 | */ | |
1318 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1319 | if (unlikely(!mapping)) | |
1320 | goto unlock_put_anon; | |
1321 | ||
5202978b | 1322 | ttu = TTU_RMAP_LOCKED; |
336bf30e | 1323 | } |
c0d0381a | 1324 | |
4b8554c5 | 1325 | try_to_migrate(src, ttu); |
2ebba6b7 | 1326 | page_was_mapped = 1; |
336bf30e | 1327 | |
5202978b | 1328 | if (ttu & TTU_RMAP_LOCKED) |
336bf30e | 1329 | i_mmap_unlock_write(mapping); |
2ebba6b7 | 1330 | } |
290408d4 NH |
1331 | |
1332 | if (!page_mapped(hpage)) | |
e7e3ffeb | 1333 | rc = move_to_new_folio(dst, src, mode); |
290408d4 | 1334 | |
336bf30e | 1335 | if (page_was_mapped) |
4eecb8b9 MWO |
1336 | remove_migration_ptes(src, |
1337 | rc == MIGRATEPAGE_SUCCESS ? dst : src, false); | |
290408d4 | 1338 | |
c0d0381a | 1339 | unlock_put_anon: |
7db7671f HD |
1340 | unlock_page(new_hpage); |
1341 | ||
1342 | put_anon: | |
fd4a4663 | 1343 | if (anon_vma) |
9e60109f | 1344 | put_anon_vma(anon_vma); |
8e6ac7fa | 1345 | |
2def7424 | 1346 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1347 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1348 | put_new_page = NULL; |
1349 | } | |
8e6ac7fa | 1350 | |
cb6acd01 | 1351 | out_unlock: |
290408d4 | 1352 | unlock_page(hpage); |
09761333 | 1353 | out: |
dd4ae78a | 1354 | if (rc == MIGRATEPAGE_SUCCESS) |
b8ec1cee | 1355 | putback_active_hugepage(hpage); |
a04840c6 | 1356 | else if (rc != -EAGAIN) |
dd4ae78a | 1357 | list_move_tail(&hpage->lru, ret); |
68711a74 DR |
1358 | |
1359 | /* | |
1360 | * If migration was not successful and there's a freeing callback, use | |
1361 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1362 | * isolation. | |
1363 | */ | |
2def7424 | 1364 | if (put_new_page) |
68711a74 DR |
1365 | put_new_page(new_hpage, private); |
1366 | else | |
3aaa76e1 | 1367 | putback_active_hugepage(new_hpage); |
68711a74 | 1368 | |
290408d4 NH |
1369 | return rc; |
1370 | } | |
1371 | ||
d532e2e5 YS |
1372 | static inline int try_split_thp(struct page *page, struct page **page2, |
1373 | struct list_head *from) | |
1374 | { | |
1375 | int rc = 0; | |
1376 | ||
1377 | lock_page(page); | |
1378 | rc = split_huge_page_to_list(page, from); | |
1379 | unlock_page(page); | |
1380 | if (!rc) | |
1381 | list_safe_reset_next(page, *page2, lru); | |
1382 | ||
1383 | return rc; | |
1384 | } | |
1385 | ||
b20a3503 | 1386 | /* |
c73e5c9c SB |
1387 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1388 | * supplied as the target for the page migration | |
b20a3503 | 1389 | * |
c73e5c9c SB |
1390 | * @from: The list of pages to be migrated. |
1391 | * @get_new_page: The function used to allocate free pages to be used | |
1392 | * as the target of the page migration. | |
68711a74 DR |
1393 | * @put_new_page: The function used to free target pages if migration |
1394 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1395 | * @private: Private data to be passed on to get_new_page() |
1396 | * @mode: The migration mode that specifies the constraints for | |
1397 | * page migration, if any. | |
1398 | * @reason: The reason for page migration. | |
b5bade97 | 1399 | * @ret_succeeded: Set to the number of normal pages migrated successfully if |
5ac95884 | 1400 | * the caller passes a non-NULL pointer. |
b20a3503 | 1401 | * |
c73e5c9c SB |
1402 | * The function returns after 10 attempts or if no pages are movable any more |
1403 | * because the list has become empty or no retryable pages exist any more. | |
dd4ae78a YS |
1404 | * It is caller's responsibility to call putback_movable_pages() to return pages |
1405 | * to the LRU or free list only if ret != 0. | |
b20a3503 | 1406 | * |
5d39a7eb BW |
1407 | * Returns the number of {normal page, THP, hugetlb} that were not migrated, or |
1408 | * an error code. The number of THP splits will be considered as the number of | |
1409 | * non-migrated THP, no matter how many subpages of the THP are migrated successfully. | |
b20a3503 | 1410 | */ |
9c620e2b | 1411 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 | 1412 | free_page_t put_new_page, unsigned long private, |
5ac95884 | 1413 | enum migrate_mode mode, int reason, unsigned int *ret_succeeded) |
b20a3503 | 1414 | { |
e24f0b8f | 1415 | int retry = 1; |
1a5bae25 | 1416 | int thp_retry = 1; |
b20a3503 | 1417 | int nr_failed = 0; |
b5bade97 | 1418 | int nr_failed_pages = 0; |
5647bc29 | 1419 | int nr_succeeded = 0; |
1a5bae25 AK |
1420 | int nr_thp_succeeded = 0; |
1421 | int nr_thp_failed = 0; | |
1422 | int nr_thp_split = 0; | |
b20a3503 | 1423 | int pass = 0; |
1a5bae25 | 1424 | bool is_thp = false; |
b20a3503 CL |
1425 | struct page *page; |
1426 | struct page *page2; | |
1a5bae25 | 1427 | int rc, nr_subpages; |
dd4ae78a | 1428 | LIST_HEAD(ret_pages); |
b5bade97 | 1429 | LIST_HEAD(thp_split_pages); |
b0b515bf | 1430 | bool nosplit = (reason == MR_NUMA_MISPLACED); |
b5bade97 | 1431 | bool no_subpage_counting = false; |
b20a3503 | 1432 | |
7bc1aec5 LM |
1433 | trace_mm_migrate_pages_start(mode, reason); |
1434 | ||
b5bade97 | 1435 | thp_subpage_migration: |
1a5bae25 | 1436 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { |
e24f0b8f | 1437 | retry = 0; |
1a5bae25 | 1438 | thp_retry = 0; |
b20a3503 | 1439 | |
e24f0b8f | 1440 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1441 | retry: |
1a5bae25 AK |
1442 | /* |
1443 | * THP statistics is based on the source huge page. | |
1444 | * Capture required information that might get lost | |
1445 | * during migration. | |
1446 | */ | |
6c5c7b9f | 1447 | is_thp = PageTransHuge(page) && !PageHuge(page); |
5d39a7eb | 1448 | nr_subpages = compound_nr(page); |
e24f0b8f | 1449 | cond_resched(); |
2d1db3b1 | 1450 | |
31caf665 NH |
1451 | if (PageHuge(page)) |
1452 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1453 | put_new_page, private, page, |
dd4ae78a YS |
1454 | pass > 2, mode, reason, |
1455 | &ret_pages); | |
31caf665 | 1456 | else |
68711a74 | 1457 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec | 1458 | private, page, pass > 2, mode, |
dd4ae78a YS |
1459 | reason, &ret_pages); |
1460 | /* | |
1461 | * The rules are: | |
1462 | * Success: non hugetlb page will be freed, hugetlb | |
1463 | * page will be put back | |
1464 | * -EAGAIN: stay on the from list | |
1465 | * -ENOMEM: stay on the from list | |
1466 | * Other errno: put on ret_pages list then splice to | |
1467 | * from list | |
1468 | */ | |
e24f0b8f | 1469 | switch(rc) { |
d532e2e5 YS |
1470 | /* |
1471 | * THP migration might be unsupported or the | |
1472 | * allocation could've failed so we should | |
1473 | * retry on the same page with the THP split | |
1474 | * to base pages. | |
1475 | * | |
1476 | * Head page is retried immediately and tail | |
1477 | * pages are added to the tail of the list so | |
1478 | * we encounter them after the rest of the list | |
1479 | * is processed. | |
1480 | */ | |
1481 | case -ENOSYS: | |
1482 | /* THP migration is unsupported */ | |
1483 | if (is_thp) { | |
b5bade97 BW |
1484 | nr_thp_failed++; |
1485 | if (!try_split_thp(page, &page2, &thp_split_pages)) { | |
d532e2e5 YS |
1486 | nr_thp_split++; |
1487 | goto retry; | |
1488 | } | |
d532e2e5 | 1489 | /* Hugetlb migration is unsupported */ |
f430893b | 1490 | } else if (!no_subpage_counting) { |
b5bade97 | 1491 | nr_failed++; |
f430893b ML |
1492 | } |
1493 | ||
5d39a7eb | 1494 | nr_failed_pages += nr_subpages; |
d532e2e5 | 1495 | break; |
95a402c3 | 1496 | case -ENOMEM: |
94723aaf | 1497 | /* |
d532e2e5 YS |
1498 | * When memory is low, don't bother to try to migrate |
1499 | * other pages, just exit. | |
b0b515bf | 1500 | * THP NUMA faulting doesn't split THP to retry. |
94723aaf | 1501 | */ |
b0b515bf | 1502 | if (is_thp && !nosplit) { |
b5bade97 BW |
1503 | nr_thp_failed++; |
1504 | if (!try_split_thp(page, &page2, &thp_split_pages)) { | |
1a5bae25 | 1505 | nr_thp_split++; |
94723aaf MH |
1506 | goto retry; |
1507 | } | |
f430893b ML |
1508 | } else if (!no_subpage_counting) { |
1509 | nr_failed++; | |
1a5bae25 | 1510 | } |
b5bade97 | 1511 | |
5d39a7eb | 1512 | nr_failed_pages += nr_subpages; |
69a041ff ML |
1513 | /* |
1514 | * There might be some subpages of fail-to-migrate THPs | |
1515 | * left in thp_split_pages list. Move them back to migration | |
1516 | * list so that they could be put back to the right list by | |
1517 | * the caller otherwise the page refcnt will be leaked. | |
1518 | */ | |
1519 | list_splice_init(&thp_split_pages, from); | |
1520 | nr_thp_failed += thp_retry; | |
95a402c3 | 1521 | goto out; |
e24f0b8f | 1522 | case -EAGAIN: |
f430893b | 1523 | if (is_thp) |
1a5bae25 | 1524 | thp_retry++; |
f430893b ML |
1525 | else |
1526 | retry++; | |
e24f0b8f | 1527 | break; |
78bd5209 | 1528 | case MIGRATEPAGE_SUCCESS: |
5d39a7eb | 1529 | nr_succeeded += nr_subpages; |
f430893b | 1530 | if (is_thp) |
1a5bae25 | 1531 | nr_thp_succeeded++; |
e24f0b8f CL |
1532 | break; |
1533 | default: | |
354a3363 | 1534 | /* |
d532e2e5 | 1535 | * Permanent failure (-EBUSY, etc.): |
354a3363 NH |
1536 | * unlike -EAGAIN case, the failed page is |
1537 | * removed from migration page list and not | |
1538 | * retried in the next outer loop. | |
1539 | */ | |
f430893b | 1540 | if (is_thp) |
1a5bae25 | 1541 | nr_thp_failed++; |
f430893b | 1542 | else if (!no_subpage_counting) |
b5bade97 | 1543 | nr_failed++; |
f430893b | 1544 | |
5d39a7eb | 1545 | nr_failed_pages += nr_subpages; |
e24f0b8f | 1546 | break; |
2d1db3b1 | 1547 | } |
b20a3503 CL |
1548 | } |
1549 | } | |
b5bade97 | 1550 | nr_failed += retry; |
1a5bae25 | 1551 | nr_thp_failed += thp_retry; |
b5bade97 BW |
1552 | /* |
1553 | * Try to migrate subpages of fail-to-migrate THPs, no nr_failed | |
1554 | * counting in this round, since all subpages of a THP is counted | |
1555 | * as 1 failure in the first round. | |
1556 | */ | |
1557 | if (!list_empty(&thp_split_pages)) { | |
1558 | /* | |
1559 | * Move non-migrated pages (after 10 retries) to ret_pages | |
1560 | * to avoid migrating them again. | |
1561 | */ | |
1562 | list_splice_init(from, &ret_pages); | |
1563 | list_splice_init(&thp_split_pages, from); | |
1564 | no_subpage_counting = true; | |
1565 | retry = 1; | |
1566 | goto thp_subpage_migration; | |
1567 | } | |
1568 | ||
1569 | rc = nr_failed + nr_thp_failed; | |
95a402c3 | 1570 | out: |
dd4ae78a YS |
1571 | /* |
1572 | * Put the permanent failure page back to migration list, they | |
1573 | * will be put back to the right list by the caller. | |
1574 | */ | |
1575 | list_splice(&ret_pages, from); | |
1576 | ||
1a5bae25 | 1577 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
b5bade97 | 1578 | count_vm_events(PGMIGRATE_FAIL, nr_failed_pages); |
1a5bae25 AK |
1579 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); |
1580 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); | |
1581 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); | |
b5bade97 | 1582 | trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded, |
1a5bae25 | 1583 | nr_thp_failed, nr_thp_split, mode, reason); |
7b2a2d4a | 1584 | |
5ac95884 YS |
1585 | if (ret_succeeded) |
1586 | *ret_succeeded = nr_succeeded; | |
1587 | ||
78bd5209 | 1588 | return rc; |
b20a3503 | 1589 | } |
95a402c3 | 1590 | |
19fc7bed | 1591 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
b4b38223 | 1592 | { |
ffe06786 | 1593 | struct folio *folio = page_folio(page); |
19fc7bed JK |
1594 | struct migration_target_control *mtc; |
1595 | gfp_t gfp_mask; | |
b4b38223 | 1596 | unsigned int order = 0; |
ffe06786 | 1597 | struct folio *new_folio = NULL; |
19fc7bed JK |
1598 | int nid; |
1599 | int zidx; | |
1600 | ||
1601 | mtc = (struct migration_target_control *)private; | |
1602 | gfp_mask = mtc->gfp_mask; | |
1603 | nid = mtc->nid; | |
1604 | if (nid == NUMA_NO_NODE) | |
ffe06786 | 1605 | nid = folio_nid(folio); |
b4b38223 | 1606 | |
ffe06786 MWO |
1607 | if (folio_test_hugetlb(folio)) { |
1608 | struct hstate *h = page_hstate(&folio->page); | |
d92bbc27 | 1609 | |
19fc7bed JK |
1610 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
1611 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); | |
d92bbc27 | 1612 | } |
b4b38223 | 1613 | |
ffe06786 | 1614 | if (folio_test_large(folio)) { |
9933a0c8 JK |
1615 | /* |
1616 | * clear __GFP_RECLAIM to make the migration callback | |
1617 | * consistent with regular THP allocations. | |
1618 | */ | |
1619 | gfp_mask &= ~__GFP_RECLAIM; | |
b4b38223 | 1620 | gfp_mask |= GFP_TRANSHUGE; |
ffe06786 | 1621 | order = folio_order(folio); |
b4b38223 | 1622 | } |
ffe06786 | 1623 | zidx = zone_idx(folio_zone(folio)); |
19fc7bed | 1624 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) |
b4b38223 JK |
1625 | gfp_mask |= __GFP_HIGHMEM; |
1626 | ||
ffe06786 | 1627 | new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask); |
b4b38223 | 1628 | |
ffe06786 | 1629 | return &new_folio->page; |
b4b38223 JK |
1630 | } |
1631 | ||
742755a1 | 1632 | #ifdef CONFIG_NUMA |
742755a1 | 1633 | |
a49bd4d7 | 1634 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1635 | { |
a49bd4d7 MH |
1636 | while (nr-- > 0) { |
1637 | if (put_user(value, status + start)) | |
1638 | return -EFAULT; | |
1639 | start++; | |
1640 | } | |
1641 | ||
1642 | return 0; | |
1643 | } | |
1644 | ||
1645 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1646 | struct list_head *pagelist, int node) | |
1647 | { | |
1648 | int err; | |
a0976311 JK |
1649 | struct migration_target_control mtc = { |
1650 | .nid = node, | |
1651 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
1652 | }; | |
a49bd4d7 | 1653 | |
a0976311 | 1654 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
5ac95884 | 1655 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); |
a49bd4d7 MH |
1656 | if (err) |
1657 | putback_movable_pages(pagelist); | |
1658 | return err; | |
742755a1 CL |
1659 | } |
1660 | ||
1661 | /* | |
a49bd4d7 MH |
1662 | * Resolves the given address to a struct page, isolates it from the LRU and |
1663 | * puts it to the given pagelist. | |
e0153fc2 YS |
1664 | * Returns: |
1665 | * errno - if the page cannot be found/isolated | |
1666 | * 0 - when it doesn't have to be migrated because it is already on the | |
1667 | * target node | |
1668 | * 1 - when it has been queued | |
742755a1 | 1669 | */ |
a49bd4d7 MH |
1670 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1671 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1672 | { |
a49bd4d7 MH |
1673 | struct vm_area_struct *vma; |
1674 | struct page *page; | |
742755a1 | 1675 | int err; |
742755a1 | 1676 | |
d8ed45c5 | 1677 | mmap_read_lock(mm); |
a49bd4d7 | 1678 | err = -EFAULT; |
cb1c37b1 ML |
1679 | vma = vma_lookup(mm, addr); |
1680 | if (!vma || !vma_migratable(vma)) | |
a49bd4d7 | 1681 | goto out; |
742755a1 | 1682 | |
a49bd4d7 | 1683 | /* FOLL_DUMP to ignore special (like zero) pages */ |
87d2762e | 1684 | page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); |
89f5b7da | 1685 | |
a49bd4d7 MH |
1686 | err = PTR_ERR(page); |
1687 | if (IS_ERR(page)) | |
1688 | goto out; | |
89f5b7da | 1689 | |
a49bd4d7 | 1690 | err = -ENOENT; |
3218f871 | 1691 | if (!page || is_zone_device_page(page)) |
a49bd4d7 | 1692 | goto out; |
742755a1 | 1693 | |
a49bd4d7 MH |
1694 | err = 0; |
1695 | if (page_to_nid(page) == node) | |
1696 | goto out_putpage; | |
742755a1 | 1697 | |
a49bd4d7 MH |
1698 | err = -EACCES; |
1699 | if (page_mapcount(page) > 1 && !migrate_all) | |
1700 | goto out_putpage; | |
742755a1 | 1701 | |
a49bd4d7 MH |
1702 | if (PageHuge(page)) { |
1703 | if (PageHead(page)) { | |
7ce82f4c ML |
1704 | err = isolate_hugetlb(page, pagelist); |
1705 | if (!err) | |
1706 | err = 1; | |
e632a938 | 1707 | } |
a49bd4d7 MH |
1708 | } else { |
1709 | struct page *head; | |
e632a938 | 1710 | |
e8db67eb NH |
1711 | head = compound_head(page); |
1712 | err = isolate_lru_page(head); | |
cf608ac1 | 1713 | if (err) |
a49bd4d7 | 1714 | goto out_putpage; |
742755a1 | 1715 | |
e0153fc2 | 1716 | err = 1; |
a49bd4d7 MH |
1717 | list_add_tail(&head->lru, pagelist); |
1718 | mod_node_page_state(page_pgdat(head), | |
9de4f22a | 1719 | NR_ISOLATED_ANON + page_is_file_lru(head), |
6c357848 | 1720 | thp_nr_pages(head)); |
a49bd4d7 MH |
1721 | } |
1722 | out_putpage: | |
1723 | /* | |
1724 | * Either remove the duplicate refcount from | |
1725 | * isolate_lru_page() or drop the page ref if it was | |
1726 | * not isolated. | |
1727 | */ | |
1728 | put_page(page); | |
1729 | out: | |
d8ed45c5 | 1730 | mmap_read_unlock(mm); |
742755a1 CL |
1731 | return err; |
1732 | } | |
1733 | ||
7ca8783a WY |
1734 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
1735 | struct list_head *pagelist, int __user *status, | |
1736 | int start, int i, unsigned long nr_pages) | |
1737 | { | |
1738 | int err; | |
1739 | ||
5d7ae891 WY |
1740 | if (list_empty(pagelist)) |
1741 | return 0; | |
1742 | ||
7ca8783a WY |
1743 | err = do_move_pages_to_node(mm, pagelist, node); |
1744 | if (err) { | |
1745 | /* | |
1746 | * Positive err means the number of failed | |
1747 | * pages to migrate. Since we are going to | |
1748 | * abort and return the number of non-migrated | |
ab9dd4f8 | 1749 | * pages, so need to include the rest of the |
7ca8783a WY |
1750 | * nr_pages that have not been attempted as |
1751 | * well. | |
1752 | */ | |
1753 | if (err > 0) | |
a7504ed1 | 1754 | err += nr_pages - i; |
7ca8783a WY |
1755 | return err; |
1756 | } | |
1757 | return store_status(status, start, node, i - start); | |
1758 | } | |
1759 | ||
5e9a0f02 BG |
1760 | /* |
1761 | * Migrate an array of page address onto an array of nodes and fill | |
1762 | * the corresponding array of status. | |
1763 | */ | |
3268c63e | 1764 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1765 | unsigned long nr_pages, |
1766 | const void __user * __user *pages, | |
1767 | const int __user *nodes, | |
1768 | int __user *status, int flags) | |
1769 | { | |
a49bd4d7 MH |
1770 | int current_node = NUMA_NO_NODE; |
1771 | LIST_HEAD(pagelist); | |
1772 | int start, i; | |
1773 | int err = 0, err1; | |
35282a2d | 1774 | |
361a2a22 | 1775 | lru_cache_disable(); |
35282a2d | 1776 | |
a49bd4d7 MH |
1777 | for (i = start = 0; i < nr_pages; i++) { |
1778 | const void __user *p; | |
1779 | unsigned long addr; | |
1780 | int node; | |
3140a227 | 1781 | |
a49bd4d7 MH |
1782 | err = -EFAULT; |
1783 | if (get_user(p, pages + i)) | |
1784 | goto out_flush; | |
1785 | if (get_user(node, nodes + i)) | |
1786 | goto out_flush; | |
057d3389 | 1787 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1788 | |
1789 | err = -ENODEV; | |
1790 | if (node < 0 || node >= MAX_NUMNODES) | |
1791 | goto out_flush; | |
1792 | if (!node_state(node, N_MEMORY)) | |
1793 | goto out_flush; | |
5e9a0f02 | 1794 | |
a49bd4d7 MH |
1795 | err = -EACCES; |
1796 | if (!node_isset(node, task_nodes)) | |
1797 | goto out_flush; | |
1798 | ||
1799 | if (current_node == NUMA_NO_NODE) { | |
1800 | current_node = node; | |
1801 | start = i; | |
1802 | } else if (node != current_node) { | |
7ca8783a WY |
1803 | err = move_pages_and_store_status(mm, current_node, |
1804 | &pagelist, status, start, i, nr_pages); | |
a49bd4d7 MH |
1805 | if (err) |
1806 | goto out; | |
1807 | start = i; | |
1808 | current_node = node; | |
3140a227 BG |
1809 | } |
1810 | ||
a49bd4d7 MH |
1811 | /* |
1812 | * Errors in the page lookup or isolation are not fatal and we simply | |
1813 | * report them via status | |
1814 | */ | |
1815 | err = add_page_for_migration(mm, addr, current_node, | |
1816 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 | 1817 | |
d08221a0 | 1818 | if (err > 0) { |
e0153fc2 YS |
1819 | /* The page is successfully queued for migration */ |
1820 | continue; | |
1821 | } | |
3140a227 | 1822 | |
65462462 JH |
1823 | /* |
1824 | * The move_pages() man page does not have an -EEXIST choice, so | |
1825 | * use -EFAULT instead. | |
1826 | */ | |
1827 | if (err == -EEXIST) | |
1828 | err = -EFAULT; | |
1829 | ||
d08221a0 WY |
1830 | /* |
1831 | * If the page is already on the target node (!err), store the | |
1832 | * node, otherwise, store the err. | |
1833 | */ | |
1834 | err = store_status(status, i, err ? : current_node, 1); | |
a49bd4d7 MH |
1835 | if (err) |
1836 | goto out_flush; | |
5e9a0f02 | 1837 | |
7ca8783a WY |
1838 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
1839 | status, start, i, nr_pages); | |
a7504ed1 HY |
1840 | if (err) { |
1841 | /* We have accounted for page i */ | |
1842 | if (err > 0) | |
1843 | err--; | |
4afdacec | 1844 | goto out; |
a7504ed1 | 1845 | } |
a49bd4d7 | 1846 | current_node = NUMA_NO_NODE; |
3140a227 | 1847 | } |
a49bd4d7 MH |
1848 | out_flush: |
1849 | /* Make sure we do not overwrite the existing error */ | |
7ca8783a WY |
1850 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
1851 | status, start, i, nr_pages); | |
dfe9aa23 | 1852 | if (err >= 0) |
a49bd4d7 | 1853 | err = err1; |
5e9a0f02 | 1854 | out: |
361a2a22 | 1855 | lru_cache_enable(); |
5e9a0f02 BG |
1856 | return err; |
1857 | } | |
1858 | ||
742755a1 | 1859 | /* |
2f007e74 | 1860 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1861 | */ |
80bba129 BG |
1862 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1863 | const void __user **pages, int *status) | |
742755a1 | 1864 | { |
2f007e74 | 1865 | unsigned long i; |
2f007e74 | 1866 | |
d8ed45c5 | 1867 | mmap_read_lock(mm); |
742755a1 | 1868 | |
2f007e74 | 1869 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1870 | unsigned long addr = (unsigned long)(*pages); |
83156821 | 1871 | unsigned int foll_flags = FOLL_DUMP; |
742755a1 CL |
1872 | struct vm_area_struct *vma; |
1873 | struct page *page; | |
c095adbc | 1874 | int err = -EFAULT; |
2f007e74 | 1875 | |
059b8b48 LH |
1876 | vma = vma_lookup(mm, addr); |
1877 | if (!vma) | |
742755a1 CL |
1878 | goto set_status; |
1879 | ||
83156821 HW |
1880 | /* Not all huge page follow APIs support 'FOLL_GET' */ |
1881 | if (!is_vm_hugetlb_page(vma)) | |
1882 | foll_flags |= FOLL_GET; | |
1883 | ||
d899844e | 1884 | /* FOLL_DUMP to ignore special (like zero) pages */ |
83156821 | 1885 | page = follow_page(vma, addr, foll_flags); |
89f5b7da LT |
1886 | |
1887 | err = PTR_ERR(page); | |
1888 | if (IS_ERR(page)) | |
1889 | goto set_status; | |
1890 | ||
3218f871 | 1891 | if (page && !is_zone_device_page(page)) { |
4cd61484 | 1892 | err = page_to_nid(page); |
83156821 HW |
1893 | if (foll_flags & FOLL_GET) |
1894 | put_page(page); | |
4cd61484 ML |
1895 | } else { |
1896 | err = -ENOENT; | |
1897 | } | |
742755a1 | 1898 | set_status: |
80bba129 BG |
1899 | *status = err; |
1900 | ||
1901 | pages++; | |
1902 | status++; | |
1903 | } | |
1904 | ||
d8ed45c5 | 1905 | mmap_read_unlock(mm); |
80bba129 BG |
1906 | } |
1907 | ||
5b1b561b AB |
1908 | static int get_compat_pages_array(const void __user *chunk_pages[], |
1909 | const void __user * __user *pages, | |
1910 | unsigned long chunk_nr) | |
1911 | { | |
1912 | compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; | |
1913 | compat_uptr_t p; | |
1914 | int i; | |
1915 | ||
1916 | for (i = 0; i < chunk_nr; i++) { | |
1917 | if (get_user(p, pages32 + i)) | |
1918 | return -EFAULT; | |
1919 | chunk_pages[i] = compat_ptr(p); | |
1920 | } | |
1921 | ||
1922 | return 0; | |
1923 | } | |
1924 | ||
80bba129 BG |
1925 | /* |
1926 | * Determine the nodes of a user array of pages and store it in | |
1927 | * a user array of status. | |
1928 | */ | |
1929 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1930 | const void __user * __user *pages, | |
1931 | int __user *status) | |
1932 | { | |
3eefb826 | 1933 | #define DO_PAGES_STAT_CHUNK_NR 16UL |
80bba129 BG |
1934 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; |
1935 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1936 | |
87b8d1ad | 1937 | while (nr_pages) { |
3eefb826 | 1938 | unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); |
87b8d1ad | 1939 | |
5b1b561b AB |
1940 | if (in_compat_syscall()) { |
1941 | if (get_compat_pages_array(chunk_pages, pages, | |
1942 | chunk_nr)) | |
1943 | break; | |
1944 | } else { | |
1945 | if (copy_from_user(chunk_pages, pages, | |
1946 | chunk_nr * sizeof(*chunk_pages))) | |
1947 | break; | |
1948 | } | |
80bba129 BG |
1949 | |
1950 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1951 | ||
87b8d1ad PA |
1952 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1953 | break; | |
742755a1 | 1954 | |
87b8d1ad PA |
1955 | pages += chunk_nr; |
1956 | status += chunk_nr; | |
1957 | nr_pages -= chunk_nr; | |
1958 | } | |
1959 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1960 | } |
1961 | ||
4dc200ce | 1962 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
742755a1 | 1963 | { |
742755a1 | 1964 | struct task_struct *task; |
742755a1 | 1965 | struct mm_struct *mm; |
742755a1 | 1966 | |
4dc200ce ML |
1967 | /* |
1968 | * There is no need to check if current process has the right to modify | |
1969 | * the specified process when they are same. | |
1970 | */ | |
1971 | if (!pid) { | |
1972 | mmget(current->mm); | |
1973 | *mem_nodes = cpuset_mems_allowed(current); | |
1974 | return current->mm; | |
1975 | } | |
742755a1 CL |
1976 | |
1977 | /* Find the mm_struct */ | |
a879bf58 | 1978 | rcu_read_lock(); |
4dc200ce | 1979 | task = find_task_by_vpid(pid); |
742755a1 | 1980 | if (!task) { |
a879bf58 | 1981 | rcu_read_unlock(); |
4dc200ce | 1982 | return ERR_PTR(-ESRCH); |
742755a1 | 1983 | } |
3268c63e | 1984 | get_task_struct(task); |
742755a1 CL |
1985 | |
1986 | /* | |
1987 | * Check if this process has the right to modify the specified | |
197e7e52 | 1988 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1989 | */ |
197e7e52 | 1990 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1991 | rcu_read_unlock(); |
4dc200ce | 1992 | mm = ERR_PTR(-EPERM); |
5e9a0f02 | 1993 | goto out; |
742755a1 | 1994 | } |
c69e8d9c | 1995 | rcu_read_unlock(); |
742755a1 | 1996 | |
4dc200ce ML |
1997 | mm = ERR_PTR(security_task_movememory(task)); |
1998 | if (IS_ERR(mm)) | |
5e9a0f02 | 1999 | goto out; |
4dc200ce | 2000 | *mem_nodes = cpuset_mems_allowed(task); |
3268c63e | 2001 | mm = get_task_mm(task); |
4dc200ce | 2002 | out: |
3268c63e | 2003 | put_task_struct(task); |
6e8b09ea | 2004 | if (!mm) |
4dc200ce ML |
2005 | mm = ERR_PTR(-EINVAL); |
2006 | return mm; | |
2007 | } | |
2008 | ||
2009 | /* | |
2010 | * Move a list of pages in the address space of the currently executing | |
2011 | * process. | |
2012 | */ | |
2013 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
2014 | const void __user * __user *pages, | |
2015 | const int __user *nodes, | |
2016 | int __user *status, int flags) | |
2017 | { | |
2018 | struct mm_struct *mm; | |
2019 | int err; | |
2020 | nodemask_t task_nodes; | |
2021 | ||
2022 | /* Check flags */ | |
2023 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
6e8b09ea SL |
2024 | return -EINVAL; |
2025 | ||
4dc200ce ML |
2026 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
2027 | return -EPERM; | |
2028 | ||
2029 | mm = find_mm_struct(pid, &task_nodes); | |
2030 | if (IS_ERR(mm)) | |
2031 | return PTR_ERR(mm); | |
2032 | ||
6e8b09ea SL |
2033 | if (nodes) |
2034 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
2035 | nodes, status, flags); | |
2036 | else | |
2037 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 2038 | |
742755a1 CL |
2039 | mmput(mm); |
2040 | return err; | |
2041 | } | |
742755a1 | 2042 | |
7addf443 DB |
2043 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
2044 | const void __user * __user *, pages, | |
2045 | const int __user *, nodes, | |
2046 | int __user *, status, int, flags) | |
2047 | { | |
2048 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
2049 | } | |
2050 | ||
7039e1db PZ |
2051 | #ifdef CONFIG_NUMA_BALANCING |
2052 | /* | |
2053 | * Returns true if this is a safe migration target node for misplaced NUMA | |
bc53008e | 2054 | * pages. Currently it only checks the watermarks which is crude. |
7039e1db PZ |
2055 | */ |
2056 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 2057 | unsigned long nr_migrate_pages) |
7039e1db PZ |
2058 | { |
2059 | int z; | |
599d0c95 | 2060 | |
7039e1db PZ |
2061 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
2062 | struct zone *zone = pgdat->node_zones + z; | |
2063 | ||
bc53008e | 2064 | if (!managed_zone(zone)) |
7039e1db PZ |
2065 | continue; |
2066 | ||
7039e1db PZ |
2067 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
2068 | if (!zone_watermark_ok(zone, 0, | |
2069 | high_wmark_pages(zone) + | |
2070 | nr_migrate_pages, | |
bfe9d006 | 2071 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
2072 | continue; |
2073 | return true; | |
2074 | } | |
2075 | return false; | |
2076 | } | |
2077 | ||
2078 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 2079 | unsigned long data) |
7039e1db PZ |
2080 | { |
2081 | int nid = (int) data; | |
c185e494 MWO |
2082 | int order = compound_order(page); |
2083 | gfp_t gfp = __GFP_THISNODE; | |
2084 | struct folio *new; | |
2085 | ||
2086 | if (order > 0) | |
2087 | gfp |= GFP_TRANSHUGE_LIGHT; | |
2088 | else { | |
2089 | gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
2090 | __GFP_NOWARN; | |
2091 | gfp &= ~__GFP_RECLAIM; | |
2092 | } | |
2093 | new = __folio_alloc_node(gfp, order, nid); | |
c5b5a3dd | 2094 | |
c185e494 | 2095 | return &new->page; |
c5b5a3dd YS |
2096 | } |
2097 | ||
1c30e017 | 2098 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 2099 | { |
2b9b624f | 2100 | int nr_pages = thp_nr_pages(page); |
c574bbe9 | 2101 | int order = compound_order(page); |
a8f60772 | 2102 | |
c574bbe9 | 2103 | VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); |
3abef4e6 | 2104 | |
662aeea7 YS |
2105 | /* Do not migrate THP mapped by multiple processes */ |
2106 | if (PageTransHuge(page) && total_mapcount(page) > 1) | |
2107 | return 0; | |
2108 | ||
7039e1db | 2109 | /* Avoid migrating to a node that is nearly full */ |
c574bbe9 HY |
2110 | if (!migrate_balanced_pgdat(pgdat, nr_pages)) { |
2111 | int z; | |
2112 | ||
2113 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) | |
2114 | return 0; | |
2115 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
bc53008e | 2116 | if (managed_zone(pgdat->node_zones + z)) |
c574bbe9 HY |
2117 | break; |
2118 | } | |
2119 | wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); | |
340ef390 | 2120 | return 0; |
c574bbe9 | 2121 | } |
7039e1db | 2122 | |
340ef390 HD |
2123 | if (isolate_lru_page(page)) |
2124 | return 0; | |
7039e1db | 2125 | |
b75454e1 | 2126 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page), |
2b9b624f | 2127 | nr_pages); |
340ef390 | 2128 | |
149c33e1 | 2129 | /* |
340ef390 HD |
2130 | * Isolating the page has taken another reference, so the |
2131 | * caller's reference can be safely dropped without the page | |
2132 | * disappearing underneath us during migration. | |
149c33e1 MG |
2133 | */ |
2134 | put_page(page); | |
340ef390 | 2135 | return 1; |
b32967ff MG |
2136 | } |
2137 | ||
2138 | /* | |
2139 | * Attempt to migrate a misplaced page to the specified destination | |
2140 | * node. Caller is expected to have an elevated reference count on | |
2141 | * the page that will be dropped by this function before returning. | |
2142 | */ | |
1bc115d8 MG |
2143 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
2144 | int node) | |
b32967ff MG |
2145 | { |
2146 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 2147 | int isolated; |
b32967ff | 2148 | int nr_remaining; |
e39bb6be | 2149 | unsigned int nr_succeeded; |
b32967ff | 2150 | LIST_HEAD(migratepages); |
b5916c02 | 2151 | int nr_pages = thp_nr_pages(page); |
c5b5a3dd | 2152 | |
b32967ff | 2153 | /* |
1bc115d8 MG |
2154 | * Don't migrate file pages that are mapped in multiple processes |
2155 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 2156 | */ |
7ee820ee ML |
2157 | if (page_mapcount(page) != 1 && page_is_file_lru(page) && |
2158 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 2159 | goto out; |
b32967ff | 2160 | |
09a913a7 MG |
2161 | /* |
2162 | * Also do not migrate dirty pages as not all filesystems can move | |
2163 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2164 | */ | |
9de4f22a | 2165 | if (page_is_file_lru(page) && PageDirty(page)) |
09a913a7 MG |
2166 | goto out; |
2167 | ||
b32967ff MG |
2168 | isolated = numamigrate_isolate_page(pgdat, page); |
2169 | if (!isolated) | |
2170 | goto out; | |
2171 | ||
2172 | list_add(&page->lru, &migratepages); | |
c185e494 MWO |
2173 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
2174 | NULL, node, MIGRATE_ASYNC, | |
2175 | MR_NUMA_MISPLACED, &nr_succeeded); | |
b32967ff | 2176 | if (nr_remaining) { |
59c82b70 JK |
2177 | if (!list_empty(&migratepages)) { |
2178 | list_del(&page->lru); | |
c5fc5c3a YS |
2179 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
2180 | page_is_file_lru(page), -nr_pages); | |
59c82b70 JK |
2181 | putback_lru_page(page); |
2182 | } | |
b32967ff | 2183 | isolated = 0; |
e39bb6be HY |
2184 | } |
2185 | if (nr_succeeded) { | |
2186 | count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); | |
2187 | if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) | |
2188 | mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, | |
2189 | nr_succeeded); | |
2190 | } | |
7039e1db | 2191 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2192 | return isolated; |
340ef390 HD |
2193 | |
2194 | out: | |
2195 | put_page(page); | |
2196 | return 0; | |
7039e1db | 2197 | } |
220018d3 | 2198 | #endif /* CONFIG_NUMA_BALANCING */ |
8763cb45 | 2199 | |
dcee9bf5 HY |
2200 | /* |
2201 | * node_demotion[] example: | |
2202 | * | |
2203 | * Consider a system with two sockets. Each socket has | |
2204 | * three classes of memory attached: fast, medium and slow. | |
2205 | * Each memory class is placed in its own NUMA node. The | |
2206 | * CPUs are placed in the node with the "fast" memory. The | |
2207 | * 6 NUMA nodes (0-5) might be split among the sockets like | |
2208 | * this: | |
2209 | * | |
2210 | * Socket A: 0, 1, 2 | |
2211 | * Socket B: 3, 4, 5 | |
2212 | * | |
2213 | * When Node 0 fills up, its memory should be migrated to | |
2214 | * Node 1. When Node 1 fills up, it should be migrated to | |
2215 | * Node 2. The migration path start on the nodes with the | |
2216 | * processors (since allocations default to this node) and | |
2217 | * fast memory, progress through medium and end with the | |
2218 | * slow memory: | |
2219 | * | |
2220 | * 0 -> 1 -> 2 -> stop | |
2221 | * 3 -> 4 -> 5 -> stop | |
2222 | * | |
2223 | * This is represented in the node_demotion[] like this: | |
2224 | * | |
2225 | * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1 | |
2226 | * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2 | |
2227 | * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate | |
2228 | * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4 | |
2229 | * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5 | |
2230 | * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate | |
2231 | * | |
2232 | * Moreover some systems may have multiple slow memory nodes. | |
2233 | * Suppose a system has one socket with 3 memory nodes, node 0 | |
2234 | * is fast memory type, and node 1/2 both are slow memory | |
2235 | * type, and the distance between fast memory node and slow | |
2236 | * memory node is same. So the migration path should be: | |
2237 | * | |
2238 | * 0 -> 1/2 -> stop | |
2239 | * | |
2240 | * This is represented in the node_demotion[] like this: | |
2241 | * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2 | |
2242 | * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate | |
2243 | * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate | |
2244 | */ | |
2245 | ||
2246 | /* | |
2247 | * Writes to this array occur without locking. Cycles are | |
2248 | * not allowed: Node X demotes to Y which demotes to X... | |
2249 | * | |
2250 | * If multiple reads are performed, a single rcu_read_lock() | |
2251 | * must be held over all reads to ensure that no cycles are | |
2252 | * observed. | |
2253 | */ | |
2254 | #define DEFAULT_DEMOTION_TARGET_NODES 15 | |
2255 | ||
2256 | #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES | |
2257 | #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1) | |
2258 | #else | |
2259 | #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES | |
2260 | #endif | |
2261 | ||
2262 | struct demotion_nodes { | |
2263 | unsigned short nr; | |
2264 | short nodes[DEMOTION_TARGET_NODES]; | |
2265 | }; | |
2266 | ||
2267 | static struct demotion_nodes *node_demotion __read_mostly; | |
2268 | ||
2269 | /** | |
2270 | * next_demotion_node() - Get the next node in the demotion path | |
2271 | * @node: The starting node to lookup the next node | |
2272 | * | |
2273 | * Return: node id for next memory node in the demotion path hierarchy | |
2274 | * from @node; NUMA_NO_NODE if @node is terminal. This does not keep | |
2275 | * @node online or guarantee that it *continues* to be the next demotion | |
2276 | * target. | |
2277 | */ | |
2278 | int next_demotion_node(int node) | |
2279 | { | |
2280 | struct demotion_nodes *nd; | |
2281 | unsigned short target_nr, index; | |
2282 | int target; | |
2283 | ||
2284 | if (!node_demotion) | |
2285 | return NUMA_NO_NODE; | |
2286 | ||
2287 | nd = &node_demotion[node]; | |
2288 | ||
2289 | /* | |
2290 | * node_demotion[] is updated without excluding this | |
2291 | * function from running. RCU doesn't provide any | |
2292 | * compiler barriers, so the READ_ONCE() is required | |
2293 | * to avoid compiler reordering or read merging. | |
2294 | * | |
2295 | * Make sure to use RCU over entire code blocks if | |
2296 | * node_demotion[] reads need to be consistent. | |
2297 | */ | |
2298 | rcu_read_lock(); | |
2299 | target_nr = READ_ONCE(nd->nr); | |
2300 | ||
2301 | switch (target_nr) { | |
2302 | case 0: | |
2303 | target = NUMA_NO_NODE; | |
2304 | goto out; | |
2305 | case 1: | |
2306 | index = 0; | |
2307 | break; | |
2308 | default: | |
2309 | /* | |
2310 | * If there are multiple target nodes, just select one | |
2311 | * target node randomly. | |
2312 | * | |
2313 | * In addition, we can also use round-robin to select | |
2314 | * target node, but we should introduce another variable | |
2315 | * for node_demotion[] to record last selected target node, | |
2316 | * that may cause cache ping-pong due to the changing of | |
2317 | * last target node. Or introducing per-cpu data to avoid | |
2318 | * caching issue, which seems more complicated. So selecting | |
2319 | * target node randomly seems better until now. | |
2320 | */ | |
2321 | index = get_random_int() % target_nr; | |
2322 | break; | |
2323 | } | |
2324 | ||
2325 | target = READ_ONCE(nd->nodes[index]); | |
2326 | ||
2327 | out: | |
2328 | rcu_read_unlock(); | |
2329 | return target; | |
2330 | } | |
2331 | ||
79c28a41 DH |
2332 | /* Disable reclaim-based migration. */ |
2333 | static void __disable_all_migrate_targets(void) | |
2334 | { | |
ac16ec83 | 2335 | int node, i; |
79c28a41 | 2336 | |
ac16ec83 BW |
2337 | if (!node_demotion) |
2338 | return; | |
79c28a41 | 2339 | |
ac16ec83 BW |
2340 | for_each_online_node(node) { |
2341 | node_demotion[node].nr = 0; | |
2342 | for (i = 0; i < DEMOTION_TARGET_NODES; i++) | |
2343 | node_demotion[node].nodes[i] = NUMA_NO_NODE; | |
2344 | } | |
79c28a41 DH |
2345 | } |
2346 | ||
2347 | static void disable_all_migrate_targets(void) | |
2348 | { | |
2349 | __disable_all_migrate_targets(); | |
2350 | ||
2351 | /* | |
2352 | * Ensure that the "disable" is visible across the system. | |
2353 | * Readers will see either a combination of before+disable | |
2354 | * state or disable+after. They will never see before and | |
2355 | * after state together. | |
2356 | * | |
2357 | * The before+after state together might have cycles and | |
2358 | * could cause readers to do things like loop until this | |
2359 | * function finishes. This ensures they can only see a | |
2360 | * single "bad" read and would, for instance, only loop | |
2361 | * once. | |
2362 | */ | |
2363 | synchronize_rcu(); | |
2364 | } | |
2365 | ||
2366 | /* | |
2367 | * Find an automatic demotion target for 'node'. | |
2368 | * Failing here is OK. It might just indicate | |
2369 | * being at the end of a chain. | |
2370 | */ | |
ac16ec83 BW |
2371 | static int establish_migrate_target(int node, nodemask_t *used, |
2372 | int best_distance) | |
79c28a41 | 2373 | { |
ac16ec83 BW |
2374 | int migration_target, index, val; |
2375 | struct demotion_nodes *nd; | |
79c28a41 | 2376 | |
ac16ec83 | 2377 | if (!node_demotion) |
79c28a41 DH |
2378 | return NUMA_NO_NODE; |
2379 | ||
ac16ec83 BW |
2380 | nd = &node_demotion[node]; |
2381 | ||
79c28a41 DH |
2382 | migration_target = find_next_best_node(node, used); |
2383 | if (migration_target == NUMA_NO_NODE) | |
2384 | return NUMA_NO_NODE; | |
2385 | ||
ac16ec83 BW |
2386 | /* |
2387 | * If the node has been set a migration target node before, | |
2388 | * which means it's the best distance between them. Still | |
2389 | * check if this node can be demoted to other target nodes | |
2390 | * if they have a same best distance. | |
2391 | */ | |
2392 | if (best_distance != -1) { | |
2393 | val = node_distance(node, migration_target); | |
2394 | if (val > best_distance) | |
fc89213a | 2395 | goto out_clear; |
ac16ec83 BW |
2396 | } |
2397 | ||
2398 | index = nd->nr; | |
2399 | if (WARN_ONCE(index >= DEMOTION_TARGET_NODES, | |
2400 | "Exceeds maximum demotion target nodes\n")) | |
fc89213a | 2401 | goto out_clear; |
ac16ec83 BW |
2402 | |
2403 | nd->nodes[index] = migration_target; | |
2404 | nd->nr++; | |
79c28a41 DH |
2405 | |
2406 | return migration_target; | |
fc89213a HY |
2407 | out_clear: |
2408 | node_clear(migration_target, *used); | |
2409 | return NUMA_NO_NODE; | |
79c28a41 DH |
2410 | } |
2411 | ||
2412 | /* | |
2413 | * When memory fills up on a node, memory contents can be | |
2414 | * automatically migrated to another node instead of | |
2415 | * discarded at reclaim. | |
2416 | * | |
2417 | * Establish a "migration path" which will start at nodes | |
2418 | * with CPUs and will follow the priorities used to build the | |
2419 | * page allocator zonelists. | |
2420 | * | |
2421 | * The difference here is that cycles must be avoided. If | |
2422 | * node0 migrates to node1, then neither node1, nor anything | |
ac16ec83 BW |
2423 | * node1 migrates to can migrate to node0. Also one node can |
2424 | * be migrated to multiple nodes if the target nodes all have | |
2425 | * a same best-distance against the source node. | |
79c28a41 DH |
2426 | * |
2427 | * This function can run simultaneously with readers of | |
2428 | * node_demotion[]. However, it can not run simultaneously | |
2429 | * with itself. Exclusion is provided by memory hotplug events | |
2430 | * being single-threaded. | |
2431 | */ | |
2432 | static void __set_migration_target_nodes(void) | |
2433 | { | |
91925ab8 ML |
2434 | nodemask_t next_pass; |
2435 | nodemask_t this_pass; | |
79c28a41 | 2436 | nodemask_t used_targets = NODE_MASK_NONE; |
ac16ec83 | 2437 | int node, best_distance; |
79c28a41 DH |
2438 | |
2439 | /* | |
2440 | * Avoid any oddities like cycles that could occur | |
2441 | * from changes in the topology. This will leave | |
2442 | * a momentary gap when migration is disabled. | |
2443 | */ | |
2444 | disable_all_migrate_targets(); | |
2445 | ||
2446 | /* | |
2447 | * Allocations go close to CPUs, first. Assume that | |
2448 | * the migration path starts at the nodes with CPUs. | |
2449 | */ | |
2450 | next_pass = node_states[N_CPU]; | |
2451 | again: | |
2452 | this_pass = next_pass; | |
2453 | next_pass = NODE_MASK_NONE; | |
2454 | /* | |
2455 | * To avoid cycles in the migration "graph", ensure | |
2456 | * that migration sources are not future targets by | |
2457 | * setting them in 'used_targets'. Do this only | |
2458 | * once per pass so that multiple source nodes can | |
2459 | * share a target node. | |
2460 | * | |
2461 | * 'used_targets' will become unavailable in future | |
2462 | * passes. This limits some opportunities for | |
2463 | * multiple source nodes to share a destination. | |
2464 | */ | |
2465 | nodes_or(used_targets, used_targets, this_pass); | |
79c28a41 | 2466 | |
ac16ec83 BW |
2467 | for_each_node_mask(node, this_pass) { |
2468 | best_distance = -1; | |
79c28a41 DH |
2469 | |
2470 | /* | |
ac16ec83 BW |
2471 | * Try to set up the migration path for the node, and the target |
2472 | * migration nodes can be multiple, so doing a loop to find all | |
2473 | * the target nodes if they all have a best node distance. | |
79c28a41 | 2474 | */ |
ac16ec83 BW |
2475 | do { |
2476 | int target_node = | |
2477 | establish_migrate_target(node, &used_targets, | |
2478 | best_distance); | |
2479 | ||
2480 | if (target_node == NUMA_NO_NODE) | |
2481 | break; | |
2482 | ||
2483 | if (best_distance == -1) | |
2484 | best_distance = node_distance(node, target_node); | |
2485 | ||
2486 | /* | |
2487 | * Visit targets from this pass in the next pass. | |
2488 | * Eventually, every node will have been part of | |
2489 | * a pass, and will become set in 'used_targets'. | |
2490 | */ | |
2491 | node_set(target_node, next_pass); | |
2492 | } while (1); | |
79c28a41 DH |
2493 | } |
2494 | /* | |
2495 | * 'next_pass' contains nodes which became migration | |
2496 | * targets in this pass. Make additional passes until | |
2497 | * no more migrations targets are available. | |
2498 | */ | |
2499 | if (!nodes_empty(next_pass)) | |
2500 | goto again; | |
2501 | } | |
2502 | ||
2503 | /* | |
2504 | * For callers that do not hold get_online_mems() already. | |
2505 | */ | |
734c1570 | 2506 | void set_migration_target_nodes(void) |
79c28a41 DH |
2507 | { |
2508 | get_online_mems(); | |
2509 | __set_migration_target_nodes(); | |
2510 | put_online_mems(); | |
2511 | } | |
884a6e5d | 2512 | |
884a6e5d DH |
2513 | /* |
2514 | * This leaves migrate-on-reclaim transiently disabled between | |
2515 | * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs | |
2516 | * whether reclaim-based migration is enabled or not, which | |
2517 | * ensures that the user can turn reclaim-based migration at | |
2518 | * any time without needing to recalculate migration targets. | |
2519 | * | |
2520 | * These callbacks already hold get_online_mems(). That is why | |
2521 | * __set_migration_target_nodes() can be used as opposed to | |
2522 | * set_migration_target_nodes(). | |
2523 | */ | |
7d6e2d96 | 2524 | #ifdef CONFIG_MEMORY_HOTPLUG |
884a6e5d | 2525 | static int __meminit migrate_on_reclaim_callback(struct notifier_block *self, |
295be91f | 2526 | unsigned long action, void *_arg) |
884a6e5d | 2527 | { |
295be91f DH |
2528 | struct memory_notify *arg = _arg; |
2529 | ||
2530 | /* | |
2531 | * Only update the node migration order when a node is | |
2532 | * changing status, like online->offline. This avoids | |
2533 | * the overhead of synchronize_rcu() in most cases. | |
2534 | */ | |
2535 | if (arg->status_change_nid < 0) | |
2536 | return notifier_from_errno(0); | |
2537 | ||
884a6e5d DH |
2538 | switch (action) { |
2539 | case MEM_GOING_OFFLINE: | |
2540 | /* | |
2541 | * Make sure there are not transient states where | |
2542 | * an offline node is a migration target. This | |
2543 | * will leave migration disabled until the offline | |
2544 | * completes and the MEM_OFFLINE case below runs. | |
2545 | */ | |
2546 | disable_all_migrate_targets(); | |
2547 | break; | |
2548 | case MEM_OFFLINE: | |
2549 | case MEM_ONLINE: | |
2550 | /* | |
2551 | * Recalculate the target nodes once the node | |
2552 | * reaches its final state (online or offline). | |
2553 | */ | |
2554 | __set_migration_target_nodes(); | |
2555 | break; | |
2556 | case MEM_CANCEL_OFFLINE: | |
2557 | /* | |
2558 | * MEM_GOING_OFFLINE disabled all the migration | |
2559 | * targets. Reenable them. | |
2560 | */ | |
2561 | __set_migration_target_nodes(); | |
2562 | break; | |
2563 | case MEM_GOING_ONLINE: | |
2564 | case MEM_CANCEL_ONLINE: | |
2565 | break; | |
2566 | } | |
2567 | ||
2568 | return notifier_from_errno(0); | |
2569 | } | |
7d6e2d96 | 2570 | #endif |
884a6e5d | 2571 | |
734c1570 | 2572 | void __init migrate_on_reclaim_init(void) |
76af6a05 | 2573 | { |
3f26c88b ML |
2574 | node_demotion = kcalloc(nr_node_ids, |
2575 | sizeof(struct demotion_nodes), | |
2576 | GFP_KERNEL); | |
ac16ec83 | 2577 | WARN_ON(!node_demotion); |
7d6e2d96 | 2578 | #ifdef CONFIG_MEMORY_HOTPLUG |
734c1570 | 2579 | hotplug_memory_notifier(migrate_on_reclaim_callback, 100); |
7d6e2d96 | 2580 | #endif |
884a6e5d | 2581 | /* |
734c1570 OS |
2582 | * At this point, all numa nodes with memory/CPus have their state |
2583 | * properly set, so we can build the demotion order now. | |
2584 | * Let us hold the cpu_hotplug lock just, as we could possibily have | |
2585 | * CPU hotplug events during boot. | |
884a6e5d | 2586 | */ |
734c1570 OS |
2587 | cpus_read_lock(); |
2588 | set_migration_target_nodes(); | |
2589 | cpus_read_unlock(); | |
884a6e5d | 2590 | } |
20f9ba4f YS |
2591 | |
2592 | bool numa_demotion_enabled = false; | |
2593 | ||
2594 | #ifdef CONFIG_SYSFS | |
2595 | static ssize_t numa_demotion_enabled_show(struct kobject *kobj, | |
2596 | struct kobj_attribute *attr, char *buf) | |
2597 | { | |
2598 | return sysfs_emit(buf, "%s\n", | |
2599 | numa_demotion_enabled ? "true" : "false"); | |
2600 | } | |
2601 | ||
2602 | static ssize_t numa_demotion_enabled_store(struct kobject *kobj, | |
2603 | struct kobj_attribute *attr, | |
2604 | const char *buf, size_t count) | |
2605 | { | |
717aeab4 JG |
2606 | ssize_t ret; |
2607 | ||
2608 | ret = kstrtobool(buf, &numa_demotion_enabled); | |
2609 | if (ret) | |
2610 | return ret; | |
20f9ba4f YS |
2611 | |
2612 | return count; | |
2613 | } | |
2614 | ||
2615 | static struct kobj_attribute numa_demotion_enabled_attr = | |
2616 | __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show, | |
2617 | numa_demotion_enabled_store); | |
2618 | ||
2619 | static struct attribute *numa_attrs[] = { | |
2620 | &numa_demotion_enabled_attr.attr, | |
2621 | NULL, | |
2622 | }; | |
2623 | ||
2624 | static const struct attribute_group numa_attr_group = { | |
2625 | .attrs = numa_attrs, | |
2626 | }; | |
2627 | ||
2628 | static int __init numa_init_sysfs(void) | |
2629 | { | |
2630 | int err; | |
2631 | struct kobject *numa_kobj; | |
2632 | ||
2633 | numa_kobj = kobject_create_and_add("numa", mm_kobj); | |
2634 | if (!numa_kobj) { | |
2635 | pr_err("failed to create numa kobject\n"); | |
2636 | return -ENOMEM; | |
2637 | } | |
2638 | err = sysfs_create_group(numa_kobj, &numa_attr_group); | |
2639 | if (err) { | |
2640 | pr_err("failed to register numa group\n"); | |
2641 | goto delete_obj; | |
2642 | } | |
2643 | return 0; | |
2644 | ||
2645 | delete_obj: | |
2646 | kobject_put(numa_kobj); | |
2647 | return err; | |
2648 | } | |
2649 | subsys_initcall(numa_init_sysfs); | |
7d6e2d96 OS |
2650 | #endif /* CONFIG_SYSFS */ |
2651 | #endif /* CONFIG_NUMA */ |