Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
b20a3503 | 2 | /* |
14e0f9bc | 3 | * Memory Migration functionality - linux/mm/migrate.c |
b20a3503 CL |
4 | * |
5 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
6 | * | |
7 | * Page migration was first developed in the context of the memory hotplug | |
8 | * project. The main authors of the migration code are: | |
9 | * | |
10 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
11 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
12 | * Dave Hansen <haveblue@us.ibm.com> | |
cde53535 | 13 | * Christoph Lameter |
b20a3503 CL |
14 | */ |
15 | ||
16 | #include <linux/migrate.h> | |
b95f1b31 | 17 | #include <linux/export.h> |
b20a3503 | 18 | #include <linux/swap.h> |
0697212a | 19 | #include <linux/swapops.h> |
b20a3503 | 20 | #include <linux/pagemap.h> |
e23ca00b | 21 | #include <linux/buffer_head.h> |
b20a3503 | 22 | #include <linux/mm_inline.h> |
b488893a | 23 | #include <linux/nsproxy.h> |
b20a3503 | 24 | #include <linux/pagevec.h> |
e9995ef9 | 25 | #include <linux/ksm.h> |
b20a3503 CL |
26 | #include <linux/rmap.h> |
27 | #include <linux/topology.h> | |
28 | #include <linux/cpu.h> | |
29 | #include <linux/cpuset.h> | |
04e62a29 | 30 | #include <linux/writeback.h> |
742755a1 CL |
31 | #include <linux/mempolicy.h> |
32 | #include <linux/vmalloc.h> | |
86c3a764 | 33 | #include <linux/security.h> |
42cb14b1 | 34 | #include <linux/backing-dev.h> |
bda807d4 | 35 | #include <linux/compaction.h> |
4f5ca265 | 36 | #include <linux/syscalls.h> |
7addf443 | 37 | #include <linux/compat.h> |
290408d4 | 38 | #include <linux/hugetlb.h> |
8e6ac7fa | 39 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 40 | #include <linux/gfp.h> |
a520110e | 41 | #include <linux/pagewalk.h> |
df6ad698 | 42 | #include <linux/pfn_t.h> |
a5430dda | 43 | #include <linux/memremap.h> |
8315ada7 | 44 | #include <linux/userfaultfd_k.h> |
bf6bddf1 | 45 | #include <linux/balloon_compaction.h> |
f714f4f2 | 46 | #include <linux/mmu_notifier.h> |
33c3fc71 | 47 | #include <linux/page_idle.h> |
d435edca | 48 | #include <linux/page_owner.h> |
6e84f315 | 49 | #include <linux/sched/mm.h> |
197e7e52 | 50 | #include <linux/ptrace.h> |
34290e2c | 51 | #include <linux/oom.h> |
884a6e5d | 52 | #include <linux/memory.h> |
ac16ec83 | 53 | #include <linux/random.h> |
c574bbe9 | 54 | #include <linux/sched/sysctl.h> |
b20a3503 | 55 | |
0d1836c3 MN |
56 | #include <asm/tlbflush.h> |
57 | ||
7b2a2d4a MG |
58 | #define CREATE_TRACE_POINTS |
59 | #include <trace/events/migrate.h> | |
60 | ||
b20a3503 CL |
61 | #include "internal.h" |
62 | ||
9e5bcd61 | 63 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
64 | { |
65 | struct address_space *mapping; | |
66 | ||
67 | /* | |
68 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
69 | * or just got freed under us. | |
70 | * | |
71 | * In case we 'win' a race for a movable page being freed under us and | |
72 | * raise its refcount preventing __free_pages() from doing its job | |
73 | * the put_page() at the end of this block will take care of | |
74 | * release this page, thus avoiding a nasty leakage. | |
75 | */ | |
76 | if (unlikely(!get_page_unless_zero(page))) | |
77 | goto out; | |
78 | ||
79 | /* | |
80 | * Check PageMovable before holding a PG_lock because page's owner | |
81 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 82 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
83 | */ |
84 | if (unlikely(!__PageMovable(page))) | |
85 | goto out_putpage; | |
86 | /* | |
87 | * As movable pages are not isolated from LRU lists, concurrent | |
88 | * compaction threads can race against page migration functions | |
89 | * as well as race against the releasing a page. | |
90 | * | |
91 | * In order to avoid having an already isolated movable page | |
92 | * being (wrongly) re-isolated while it is under migration, | |
93 | * or to avoid attempting to isolate pages being released, | |
94 | * lets be sure we have the page lock | |
95 | * before proceeding with the movable page isolation steps. | |
96 | */ | |
97 | if (unlikely(!trylock_page(page))) | |
98 | goto out_putpage; | |
99 | ||
100 | if (!PageMovable(page) || PageIsolated(page)) | |
101 | goto out_no_isolated; | |
102 | ||
103 | mapping = page_mapping(page); | |
104 | VM_BUG_ON_PAGE(!mapping, page); | |
105 | ||
106 | if (!mapping->a_ops->isolate_page(page, mode)) | |
107 | goto out_no_isolated; | |
108 | ||
109 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
110 | WARN_ON_ONCE(PageIsolated(page)); | |
356ea386 | 111 | SetPageIsolated(page); |
bda807d4 MK |
112 | unlock_page(page); |
113 | ||
9e5bcd61 | 114 | return 0; |
bda807d4 MK |
115 | |
116 | out_no_isolated: | |
117 | unlock_page(page); | |
118 | out_putpage: | |
119 | put_page(page); | |
120 | out: | |
9e5bcd61 | 121 | return -EBUSY; |
bda807d4 MK |
122 | } |
123 | ||
606a6f71 | 124 | static void putback_movable_page(struct page *page) |
bda807d4 MK |
125 | { |
126 | struct address_space *mapping; | |
127 | ||
bda807d4 MK |
128 | mapping = page_mapping(page); |
129 | mapping->a_ops->putback_page(page); | |
356ea386 | 130 | ClearPageIsolated(page); |
bda807d4 MK |
131 | } |
132 | ||
5733c7d1 RA |
133 | /* |
134 | * Put previously isolated pages back onto the appropriate lists | |
135 | * from where they were once taken off for compaction/migration. | |
136 | * | |
59c82b70 JK |
137 | * This function shall be used whenever the isolated pageset has been |
138 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
139 | * and isolate_huge_page(). | |
5733c7d1 RA |
140 | */ |
141 | void putback_movable_pages(struct list_head *l) | |
142 | { | |
143 | struct page *page; | |
144 | struct page *page2; | |
145 | ||
b20a3503 | 146 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
147 | if (unlikely(PageHuge(page))) { |
148 | putback_active_hugepage(page); | |
149 | continue; | |
150 | } | |
e24f0b8f | 151 | list_del(&page->lru); |
bda807d4 MK |
152 | /* |
153 | * We isolated non-lru movable page so here we can use | |
154 | * __PageMovable because LRU page's mapping cannot have | |
155 | * PAGE_MAPPING_MOVABLE. | |
156 | */ | |
b1123ea6 | 157 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
158 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
159 | lock_page(page); | |
160 | if (PageMovable(page)) | |
161 | putback_movable_page(page); | |
162 | else | |
356ea386 | 163 | ClearPageIsolated(page); |
bda807d4 MK |
164 | unlock_page(page); |
165 | put_page(page); | |
166 | } else { | |
e8db67eb | 167 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 168 | page_is_file_lru(page), -thp_nr_pages(page)); |
fc280fe8 | 169 | putback_lru_page(page); |
bda807d4 | 170 | } |
b20a3503 | 171 | } |
b20a3503 CL |
172 | } |
173 | ||
0697212a CL |
174 | /* |
175 | * Restore a potential migration pte to a working pte entry | |
176 | */ | |
e4b82222 | 177 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 178 | unsigned long addr, void *old) |
0697212a | 179 | { |
3fe87967 KS |
180 | struct page_vma_mapped_walk pvmw = { |
181 | .page = old, | |
182 | .vma = vma, | |
183 | .address = addr, | |
184 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
185 | }; | |
186 | struct page *new; | |
187 | pte_t pte; | |
0697212a | 188 | swp_entry_t entry; |
0697212a | 189 | |
3fe87967 KS |
190 | VM_BUG_ON_PAGE(PageTail(page), page); |
191 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
192 | if (PageKsm(page)) |
193 | new = page; | |
194 | else | |
195 | new = page - pvmw.page->index + | |
196 | linear_page_index(vma, pvmw.address); | |
0697212a | 197 | |
616b8371 ZY |
198 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
199 | /* PMD-mapped THP migration entry */ | |
200 | if (!pvmw.pte) { | |
201 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
202 | remove_migration_pmd(&pvmw, new); | |
203 | continue; | |
204 | } | |
205 | #endif | |
206 | ||
3fe87967 KS |
207 | get_page(new); |
208 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
209 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
210 | pte = pte_mksoft_dirty(pte); | |
0697212a | 211 | |
3fe87967 KS |
212 | /* |
213 | * Recheck VMA as permissions can change since migration started | |
214 | */ | |
215 | entry = pte_to_swp_entry(*pvmw.pte); | |
4dd845b5 | 216 | if (is_writable_migration_entry(entry)) |
3fe87967 | 217 | pte = maybe_mkwrite(pte, vma); |
f45ec5ff PX |
218 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
219 | pte = pte_mkuffd_wp(pte); | |
d3cb8bf6 | 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 |
3fe87967 | 236 | if (PageHuge(new)) { |
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); |
3fe87967 KS |
241 | if (PageAnon(new)) |
242 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
243 | else | |
244 | page_dup_rmap(new, true); | |
1eba86c0 | 245 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab AK |
246 | } else |
247 | #endif | |
248 | { | |
383321ab AK |
249 | if (PageAnon(new)) |
250 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
251 | else | |
252 | page_add_file_rmap(new, false); | |
1eba86c0 | 253 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
383321ab | 254 | } |
3fe87967 KS |
255 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
256 | mlock_vma_page(new); | |
257 | ||
e125fe40 KS |
258 | if (PageTransHuge(page) && PageMlocked(page)) |
259 | clear_page_mlock(page); | |
260 | ||
3fe87967 KS |
261 | /* No need to invalidate - it was non-present before */ |
262 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
263 | } | |
51afb12b | 264 | |
e4b82222 | 265 | return true; |
0697212a CL |
266 | } |
267 | ||
04e62a29 CL |
268 | /* |
269 | * Get rid of all migration entries and replace them by | |
270 | * references to the indicated page. | |
271 | */ | |
e388466d | 272 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 273 | { |
051ac83a JK |
274 | struct rmap_walk_control rwc = { |
275 | .rmap_one = remove_migration_pte, | |
276 | .arg = old, | |
277 | }; | |
278 | ||
e388466d KS |
279 | if (locked) |
280 | rmap_walk_locked(new, &rwc); | |
281 | else | |
282 | rmap_walk(new, &rwc); | |
04e62a29 CL |
283 | } |
284 | ||
0697212a CL |
285 | /* |
286 | * Something used the pte of a page under migration. We need to | |
287 | * get to the page and wait until migration is finished. | |
288 | * When we return from this function the fault will be retried. | |
0697212a | 289 | */ |
e66f17ff | 290 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 291 | spinlock_t *ptl) |
0697212a | 292 | { |
30dad309 | 293 | pte_t pte; |
0697212a | 294 | swp_entry_t entry; |
0697212a | 295 | |
30dad309 | 296 | spin_lock(ptl); |
0697212a CL |
297 | pte = *ptep; |
298 | if (!is_swap_pte(pte)) | |
299 | goto out; | |
300 | ||
301 | entry = pte_to_swp_entry(pte); | |
302 | if (!is_migration_entry(entry)) | |
303 | goto out; | |
304 | ||
ffa65753 | 305 | migration_entry_wait_on_locked(entry, ptep, ptl); |
0697212a CL |
306 | return; |
307 | out: | |
308 | pte_unmap_unlock(ptep, ptl); | |
309 | } | |
310 | ||
30dad309 NH |
311 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
312 | unsigned long address) | |
313 | { | |
314 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
315 | pte_t *ptep = pte_offset_map(pmd, address); | |
316 | __migration_entry_wait(mm, ptep, ptl); | |
317 | } | |
318 | ||
cb900f41 KS |
319 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
320 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 321 | { |
cb900f41 | 322 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
323 | __migration_entry_wait(mm, pte, ptl); |
324 | } | |
325 | ||
616b8371 ZY |
326 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
327 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
328 | { | |
329 | spinlock_t *ptl; | |
616b8371 ZY |
330 | |
331 | ptl = pmd_lock(mm, pmd); | |
332 | if (!is_pmd_migration_entry(*pmd)) | |
333 | goto unlock; | |
ffa65753 | 334 | migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); |
616b8371 ZY |
335 | return; |
336 | unlock: | |
337 | spin_unlock(ptl); | |
338 | } | |
339 | #endif | |
340 | ||
f900482d | 341 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
0b3901b3 JK |
342 | { |
343 | int expected_count = 1; | |
344 | ||
345 | /* | |
f1f4f3ab | 346 | * Device private pages have an extra refcount as they are |
0b3901b3 JK |
347 | * ZONE_DEVICE pages. |
348 | */ | |
349 | expected_count += is_device_private_page(page); | |
f900482d | 350 | if (mapping) |
3417013e | 351 | expected_count += compound_nr(page) + page_has_private(page); |
0b3901b3 JK |
352 | |
353 | return expected_count; | |
354 | } | |
355 | ||
b20a3503 | 356 | /* |
c3fcf8a5 | 357 | * Replace the page in the mapping. |
5b5c7120 CL |
358 | * |
359 | * The number of remaining references must be: | |
360 | * 1 for anonymous pages without a mapping | |
361 | * 2 for pages with a mapping | |
266cf658 | 362 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 363 | */ |
3417013e MWO |
364 | int folio_migrate_mapping(struct address_space *mapping, |
365 | struct folio *newfolio, struct folio *folio, int extra_count) | |
b20a3503 | 366 | { |
3417013e | 367 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
42cb14b1 HD |
368 | struct zone *oldzone, *newzone; |
369 | int dirty; | |
3417013e MWO |
370 | int expected_count = expected_page_refs(mapping, &folio->page) + extra_count; |
371 | long nr = folio_nr_pages(folio); | |
8763cb45 | 372 | |
6c5240ae | 373 | if (!mapping) { |
0e8c7d0f | 374 | /* Anonymous page without mapping */ |
3417013e | 375 | if (folio_ref_count(folio) != expected_count) |
6c5240ae | 376 | return -EAGAIN; |
cf4b769a HD |
377 | |
378 | /* No turning back from here */ | |
3417013e MWO |
379 | newfolio->index = folio->index; |
380 | newfolio->mapping = folio->mapping; | |
381 | if (folio_test_swapbacked(folio)) | |
382 | __folio_set_swapbacked(newfolio); | |
cf4b769a | 383 | |
78bd5209 | 384 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
385 | } |
386 | ||
3417013e MWO |
387 | oldzone = folio_zone(folio); |
388 | newzone = folio_zone(newfolio); | |
42cb14b1 | 389 | |
89eb946a | 390 | xas_lock_irq(&xas); |
3417013e | 391 | if (!folio_ref_freeze(folio, expected_count)) { |
89eb946a | 392 | xas_unlock_irq(&xas); |
e286781d NP |
393 | return -EAGAIN; |
394 | } | |
395 | ||
b20a3503 | 396 | /* |
3417013e | 397 | * Now we know that no one else is looking at the folio: |
cf4b769a | 398 | * no turning back from here. |
b20a3503 | 399 | */ |
3417013e MWO |
400 | newfolio->index = folio->index; |
401 | newfolio->mapping = folio->mapping; | |
402 | folio_ref_add(newfolio, nr); /* add cache reference */ | |
403 | if (folio_test_swapbacked(folio)) { | |
404 | __folio_set_swapbacked(newfolio); | |
405 | if (folio_test_swapcache(folio)) { | |
406 | folio_set_swapcache(newfolio); | |
407 | newfolio->private = folio_get_private(folio); | |
6326fec1 NP |
408 | } |
409 | } else { | |
3417013e | 410 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); |
b20a3503 CL |
411 | } |
412 | ||
42cb14b1 | 413 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
3417013e | 414 | dirty = folio_test_dirty(folio); |
42cb14b1 | 415 | if (dirty) { |
3417013e MWO |
416 | folio_clear_dirty(folio); |
417 | folio_set_dirty(newfolio); | |
42cb14b1 HD |
418 | } |
419 | ||
3417013e | 420 | xas_store(&xas, newfolio); |
7cf9c2c7 NP |
421 | |
422 | /* | |
937a94c9 JG |
423 | * Drop cache reference from old page by unfreezing |
424 | * to one less reference. | |
7cf9c2c7 NP |
425 | * We know this isn't the last reference. |
426 | */ | |
3417013e | 427 | folio_ref_unfreeze(folio, expected_count - nr); |
7cf9c2c7 | 428 | |
89eb946a | 429 | xas_unlock(&xas); |
42cb14b1 HD |
430 | /* Leave irq disabled to prevent preemption while updating stats */ |
431 | ||
0e8c7d0f CL |
432 | /* |
433 | * If moved to a different zone then also account | |
434 | * the page for that zone. Other VM counters will be | |
435 | * taken care of when we establish references to the | |
436 | * new page and drop references to the old page. | |
437 | * | |
438 | * Note that anonymous pages are accounted for | |
4b9d0fab | 439 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
440 | * are mapped to swap space. |
441 | */ | |
42cb14b1 | 442 | if (newzone != oldzone) { |
0d1c2072 JW |
443 | struct lruvec *old_lruvec, *new_lruvec; |
444 | struct mem_cgroup *memcg; | |
445 | ||
3417013e | 446 | memcg = folio_memcg(folio); |
0d1c2072 JW |
447 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); |
448 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
449 | ||
5c447d27 SB |
450 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
451 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); | |
3417013e | 452 | if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { |
5c447d27 SB |
453 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
454 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); | |
42cb14b1 | 455 | } |
b6038942 | 456 | #ifdef CONFIG_SWAP |
3417013e | 457 | if (folio_test_swapcache(folio)) { |
b6038942 SB |
458 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); |
459 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); | |
460 | } | |
461 | #endif | |
f56753ac | 462 | if (dirty && mapping_can_writeback(mapping)) { |
5c447d27 SB |
463 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
464 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); | |
465 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); | |
466 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); | |
42cb14b1 | 467 | } |
4b02108a | 468 | } |
42cb14b1 | 469 | local_irq_enable(); |
b20a3503 | 470 | |
78bd5209 | 471 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 472 | } |
3417013e | 473 | EXPORT_SYMBOL(folio_migrate_mapping); |
b20a3503 | 474 | |
290408d4 NH |
475 | /* |
476 | * The expected number of remaining references is the same as that | |
3417013e | 477 | * of folio_migrate_mapping(). |
290408d4 NH |
478 | */ |
479 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
480 | struct page *newpage, struct page *page) | |
481 | { | |
89eb946a | 482 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 483 | int expected_count; |
290408d4 | 484 | |
89eb946a | 485 | xas_lock_irq(&xas); |
290408d4 | 486 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
487 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
488 | xas_unlock_irq(&xas); | |
290408d4 NH |
489 | return -EAGAIN; |
490 | } | |
491 | ||
fe896d18 | 492 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 493 | xas_unlock_irq(&xas); |
290408d4 NH |
494 | return -EAGAIN; |
495 | } | |
496 | ||
cf4b769a HD |
497 | newpage->index = page->index; |
498 | newpage->mapping = page->mapping; | |
6a93ca8f | 499 | |
290408d4 NH |
500 | get_page(newpage); |
501 | ||
89eb946a | 502 | xas_store(&xas, newpage); |
290408d4 | 503 | |
fe896d18 | 504 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 505 | |
89eb946a | 506 | xas_unlock_irq(&xas); |
6a93ca8f | 507 | |
78bd5209 | 508 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
509 | } |
510 | ||
b20a3503 | 511 | /* |
19138349 | 512 | * Copy the flags and some other ancillary information |
b20a3503 | 513 | */ |
19138349 | 514 | void folio_migrate_flags(struct folio *newfolio, struct folio *folio) |
b20a3503 | 515 | { |
7851a45c RR |
516 | int cpupid; |
517 | ||
19138349 MWO |
518 | if (folio_test_error(folio)) |
519 | folio_set_error(newfolio); | |
520 | if (folio_test_referenced(folio)) | |
521 | folio_set_referenced(newfolio); | |
522 | if (folio_test_uptodate(folio)) | |
523 | folio_mark_uptodate(newfolio); | |
524 | if (folio_test_clear_active(folio)) { | |
525 | VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); | |
526 | folio_set_active(newfolio); | |
527 | } else if (folio_test_clear_unevictable(folio)) | |
528 | folio_set_unevictable(newfolio); | |
529 | if (folio_test_workingset(folio)) | |
530 | folio_set_workingset(newfolio); | |
531 | if (folio_test_checked(folio)) | |
532 | folio_set_checked(newfolio); | |
533 | if (folio_test_mappedtodisk(folio)) | |
534 | folio_set_mappedtodisk(newfolio); | |
b20a3503 | 535 | |
3417013e | 536 | /* Move dirty on pages not done by folio_migrate_mapping() */ |
19138349 MWO |
537 | if (folio_test_dirty(folio)) |
538 | folio_set_dirty(newfolio); | |
b20a3503 | 539 | |
19138349 MWO |
540 | if (folio_test_young(folio)) |
541 | folio_set_young(newfolio); | |
542 | if (folio_test_idle(folio)) | |
543 | folio_set_idle(newfolio); | |
33c3fc71 | 544 | |
7851a45c RR |
545 | /* |
546 | * Copy NUMA information to the new page, to prevent over-eager | |
547 | * future migrations of this same page. | |
548 | */ | |
19138349 MWO |
549 | cpupid = page_cpupid_xchg_last(&folio->page, -1); |
550 | page_cpupid_xchg_last(&newfolio->page, cpupid); | |
7851a45c | 551 | |
19138349 | 552 | folio_migrate_ksm(newfolio, folio); |
c8d6553b HD |
553 | /* |
554 | * Please do not reorder this without considering how mm/ksm.c's | |
555 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
556 | */ | |
19138349 MWO |
557 | if (folio_test_swapcache(folio)) |
558 | folio_clear_swapcache(folio); | |
559 | folio_clear_private(folio); | |
ad2fa371 MS |
560 | |
561 | /* page->private contains hugetlb specific flags */ | |
19138349 MWO |
562 | if (!folio_test_hugetlb(folio)) |
563 | folio->private = NULL; | |
b20a3503 CL |
564 | |
565 | /* | |
566 | * If any waiters have accumulated on the new page then | |
567 | * wake them up. | |
568 | */ | |
19138349 MWO |
569 | if (folio_test_writeback(newfolio)) |
570 | folio_end_writeback(newfolio); | |
d435edca | 571 | |
6aeff241 YS |
572 | /* |
573 | * PG_readahead shares the same bit with PG_reclaim. The above | |
574 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
575 | * bit after that. | |
576 | */ | |
19138349 MWO |
577 | if (folio_test_readahead(folio)) |
578 | folio_set_readahead(newfolio); | |
6aeff241 | 579 | |
19138349 | 580 | folio_copy_owner(newfolio, folio); |
74485cf2 | 581 | |
19138349 | 582 | if (!folio_test_hugetlb(folio)) |
d21bba2b | 583 | mem_cgroup_migrate(folio, newfolio); |
b20a3503 | 584 | } |
19138349 | 585 | EXPORT_SYMBOL(folio_migrate_flags); |
2916ecc0 | 586 | |
715cbfd6 | 587 | void folio_migrate_copy(struct folio *newfolio, struct folio *folio) |
2916ecc0 | 588 | { |
715cbfd6 MWO |
589 | folio_copy(newfolio, folio); |
590 | folio_migrate_flags(newfolio, folio); | |
2916ecc0 | 591 | } |
715cbfd6 | 592 | EXPORT_SYMBOL(folio_migrate_copy); |
b20a3503 | 593 | |
1d8b85cc CL |
594 | /************************************************************ |
595 | * Migration functions | |
596 | ***********************************************************/ | |
597 | ||
b20a3503 | 598 | /* |
bda807d4 | 599 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 600 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
601 | * |
602 | * Pages are locked upon entry and exit. | |
603 | */ | |
2d1db3b1 | 604 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
605 | struct page *newpage, struct page *page, |
606 | enum migrate_mode mode) | |
b20a3503 | 607 | { |
3417013e MWO |
608 | struct folio *newfolio = page_folio(newpage); |
609 | struct folio *folio = page_folio(page); | |
b20a3503 CL |
610 | int rc; |
611 | ||
3417013e | 612 | BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */ |
b20a3503 | 613 | |
3417013e | 614 | rc = folio_migrate_mapping(mapping, newfolio, folio, 0); |
b20a3503 | 615 | |
78bd5209 | 616 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
617 | return rc; |
618 | ||
2916ecc0 | 619 | if (mode != MIGRATE_SYNC_NO_COPY) |
715cbfd6 | 620 | folio_migrate_copy(newfolio, folio); |
2916ecc0 | 621 | else |
19138349 | 622 | folio_migrate_flags(newfolio, folio); |
78bd5209 | 623 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
624 | } |
625 | EXPORT_SYMBOL(migrate_page); | |
626 | ||
9361401e | 627 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
628 | /* Returns true if all buffers are successfully locked */ |
629 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
630 | enum migrate_mode mode) | |
631 | { | |
632 | struct buffer_head *bh = head; | |
633 | ||
634 | /* Simple case, sync compaction */ | |
635 | if (mode != MIGRATE_ASYNC) { | |
636 | do { | |
84ade7c1 JK |
637 | lock_buffer(bh); |
638 | bh = bh->b_this_page; | |
639 | ||
640 | } while (bh != head); | |
641 | ||
642 | return true; | |
643 | } | |
644 | ||
645 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
646 | do { | |
84ade7c1 JK |
647 | if (!trylock_buffer(bh)) { |
648 | /* | |
649 | * We failed to lock the buffer and cannot stall in | |
650 | * async migration. Release the taken locks | |
651 | */ | |
652 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
653 | bh = head; |
654 | while (bh != failed_bh) { | |
655 | unlock_buffer(bh); | |
84ade7c1 JK |
656 | bh = bh->b_this_page; |
657 | } | |
658 | return false; | |
659 | } | |
660 | ||
661 | bh = bh->b_this_page; | |
662 | } while (bh != head); | |
663 | return true; | |
664 | } | |
665 | ||
89cb0888 JK |
666 | static int __buffer_migrate_page(struct address_space *mapping, |
667 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
668 | bool check_refs) | |
1d8b85cc | 669 | { |
1d8b85cc CL |
670 | struct buffer_head *bh, *head; |
671 | int rc; | |
cc4f11e6 | 672 | int expected_count; |
1d8b85cc | 673 | |
1d8b85cc | 674 | if (!page_has_buffers(page)) |
a6bc32b8 | 675 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 676 | |
cc4f11e6 | 677 | /* Check whether page does not have extra refs before we do more work */ |
f900482d | 678 | expected_count = expected_page_refs(mapping, page); |
cc4f11e6 JK |
679 | if (page_count(page) != expected_count) |
680 | return -EAGAIN; | |
1d8b85cc | 681 | |
cc4f11e6 JK |
682 | head = page_buffers(page); |
683 | if (!buffer_migrate_lock_buffers(head, mode)) | |
684 | return -EAGAIN; | |
1d8b85cc | 685 | |
89cb0888 JK |
686 | if (check_refs) { |
687 | bool busy; | |
688 | bool invalidated = false; | |
689 | ||
690 | recheck_buffers: | |
691 | busy = false; | |
692 | spin_lock(&mapping->private_lock); | |
693 | bh = head; | |
694 | do { | |
695 | if (atomic_read(&bh->b_count)) { | |
696 | busy = true; | |
697 | break; | |
698 | } | |
699 | bh = bh->b_this_page; | |
700 | } while (bh != head); | |
89cb0888 JK |
701 | if (busy) { |
702 | if (invalidated) { | |
703 | rc = -EAGAIN; | |
704 | goto unlock_buffers; | |
705 | } | |
ebdf4de5 | 706 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
707 | invalidate_bh_lrus(); |
708 | invalidated = true; | |
709 | goto recheck_buffers; | |
710 | } | |
711 | } | |
712 | ||
37109694 | 713 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
78bd5209 | 714 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 715 | goto unlock_buffers; |
1d8b85cc | 716 | |
cd0f3715 | 717 | attach_page_private(newpage, detach_page_private(page)); |
1d8b85cc CL |
718 | |
719 | bh = head; | |
720 | do { | |
721 | set_bh_page(bh, newpage, bh_offset(bh)); | |
722 | bh = bh->b_this_page; | |
723 | ||
724 | } while (bh != head); | |
725 | ||
2916ecc0 JG |
726 | if (mode != MIGRATE_SYNC_NO_COPY) |
727 | migrate_page_copy(newpage, page); | |
728 | else | |
729 | migrate_page_states(newpage, page); | |
1d8b85cc | 730 | |
cc4f11e6 JK |
731 | rc = MIGRATEPAGE_SUCCESS; |
732 | unlock_buffers: | |
ebdf4de5 JK |
733 | if (check_refs) |
734 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
735 | bh = head; |
736 | do { | |
737 | unlock_buffer(bh); | |
1d8b85cc CL |
738 | bh = bh->b_this_page; |
739 | ||
740 | } while (bh != head); | |
741 | ||
cc4f11e6 | 742 | return rc; |
1d8b85cc | 743 | } |
89cb0888 JK |
744 | |
745 | /* | |
746 | * Migration function for pages with buffers. This function can only be used | |
747 | * if the underlying filesystem guarantees that no other references to "page" | |
748 | * exist. For example attached buffer heads are accessed only under page lock. | |
749 | */ | |
750 | int buffer_migrate_page(struct address_space *mapping, | |
751 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
752 | { | |
753 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
754 | } | |
1d8b85cc | 755 | EXPORT_SYMBOL(buffer_migrate_page); |
89cb0888 JK |
756 | |
757 | /* | |
758 | * Same as above except that this variant is more careful and checks that there | |
759 | * are also no buffer head references. This function is the right one for | |
760 | * mappings where buffer heads are directly looked up and referenced (such as | |
761 | * block device mappings). | |
762 | */ | |
763 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
764 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
765 | { | |
766 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
767 | } | |
9361401e | 768 | #endif |
1d8b85cc | 769 | |
04e62a29 CL |
770 | /* |
771 | * Writeback a page to clean the dirty state | |
772 | */ | |
773 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 774 | { |
04e62a29 CL |
775 | struct writeback_control wbc = { |
776 | .sync_mode = WB_SYNC_NONE, | |
777 | .nr_to_write = 1, | |
778 | .range_start = 0, | |
779 | .range_end = LLONG_MAX, | |
04e62a29 CL |
780 | .for_reclaim = 1 |
781 | }; | |
782 | int rc; | |
783 | ||
784 | if (!mapping->a_ops->writepage) | |
785 | /* No write method for the address space */ | |
786 | return -EINVAL; | |
787 | ||
788 | if (!clear_page_dirty_for_io(page)) | |
789 | /* Someone else already triggered a write */ | |
790 | return -EAGAIN; | |
791 | ||
8351a6e4 | 792 | /* |
04e62a29 CL |
793 | * A dirty page may imply that the underlying filesystem has |
794 | * the page on some queue. So the page must be clean for | |
795 | * migration. Writeout may mean we loose the lock and the | |
796 | * page state is no longer what we checked for earlier. | |
797 | * At this point we know that the migration attempt cannot | |
798 | * be successful. | |
8351a6e4 | 799 | */ |
e388466d | 800 | remove_migration_ptes(page, page, false); |
8351a6e4 | 801 | |
04e62a29 | 802 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 803 | |
04e62a29 CL |
804 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
805 | /* unlocked. Relock */ | |
806 | lock_page(page); | |
807 | ||
bda8550d | 808 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
809 | } |
810 | ||
811 | /* | |
812 | * Default handling if a filesystem does not provide a migration function. | |
813 | */ | |
814 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 815 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 816 | { |
b969c4ab | 817 | if (PageDirty(page)) { |
a6bc32b8 | 818 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
819 | switch (mode) { |
820 | case MIGRATE_SYNC: | |
821 | case MIGRATE_SYNC_NO_COPY: | |
822 | break; | |
823 | default: | |
b969c4ab | 824 | return -EBUSY; |
2916ecc0 | 825 | } |
04e62a29 | 826 | return writeout(mapping, page); |
b969c4ab | 827 | } |
8351a6e4 CL |
828 | |
829 | /* | |
830 | * Buffers may be managed in a filesystem specific way. | |
831 | * We must have no buffers or drop them. | |
832 | */ | |
266cf658 | 833 | if (page_has_private(page) && |
8351a6e4 | 834 | !try_to_release_page(page, GFP_KERNEL)) |
806031bb | 835 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 836 | |
a6bc32b8 | 837 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
838 | } |
839 | ||
e24f0b8f CL |
840 | /* |
841 | * Move a page to a newly allocated page | |
842 | * The page is locked and all ptes have been successfully removed. | |
843 | * | |
844 | * The new page will have replaced the old page if this function | |
845 | * is successful. | |
894bc310 LS |
846 | * |
847 | * Return value: | |
848 | * < 0 - error code | |
78bd5209 | 849 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 850 | */ |
3fe2011f | 851 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 852 | enum migrate_mode mode) |
e24f0b8f CL |
853 | { |
854 | struct address_space *mapping; | |
bda807d4 MK |
855 | int rc = -EAGAIN; |
856 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 857 | |
7db7671f HD |
858 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
859 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 860 | |
e24f0b8f | 861 | mapping = page_mapping(page); |
bda807d4 MK |
862 | |
863 | if (likely(is_lru)) { | |
864 | if (!mapping) | |
865 | rc = migrate_page(mapping, newpage, page, mode); | |
866 | else if (mapping->a_ops->migratepage) | |
867 | /* | |
868 | * Most pages have a mapping and most filesystems | |
869 | * provide a migratepage callback. Anonymous pages | |
870 | * are part of swap space which also has its own | |
871 | * migratepage callback. This is the most common path | |
872 | * for page migration. | |
873 | */ | |
874 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
875 | page, mode); | |
876 | else | |
877 | rc = fallback_migrate_page(mapping, newpage, | |
878 | page, mode); | |
879 | } else { | |
e24f0b8f | 880 | /* |
bda807d4 MK |
881 | * In case of non-lru page, it could be released after |
882 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 883 | */ |
bda807d4 MK |
884 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
885 | if (!PageMovable(page)) { | |
886 | rc = MIGRATEPAGE_SUCCESS; | |
356ea386 | 887 | ClearPageIsolated(page); |
bda807d4 MK |
888 | goto out; |
889 | } | |
890 | ||
891 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
892 | page, mode); | |
893 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
894 | !PageIsolated(page)); | |
895 | } | |
e24f0b8f | 896 | |
5c3f9a67 HD |
897 | /* |
898 | * When successful, old pagecache page->mapping must be cleared before | |
899 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
900 | */ | |
901 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
902 | if (__PageMovable(page)) { |
903 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
904 | ||
905 | /* | |
906 | * We clear PG_movable under page_lock so any compactor | |
907 | * cannot try to migrate this page. | |
908 | */ | |
356ea386 | 909 | ClearPageIsolated(page); |
bda807d4 MK |
910 | } |
911 | ||
912 | /* | |
c23a0c99 | 913 | * Anonymous and movable page->mapping will be cleared by |
bda807d4 MK |
914 | * free_pages_prepare so don't reset it here for keeping |
915 | * the type to work PageAnon, for example. | |
916 | */ | |
917 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 918 | page->mapping = NULL; |
d2b2c6dd | 919 | |
3150be8f MS |
920 | if (likely(!is_zone_device_page(newpage))) |
921 | flush_dcache_folio(page_folio(newpage)); | |
3fe2011f | 922 | } |
bda807d4 | 923 | out: |
e24f0b8f CL |
924 | return rc; |
925 | } | |
926 | ||
0dabec93 | 927 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 928 | int force, enum migrate_mode mode) |
e24f0b8f | 929 | { |
0dabec93 | 930 | int rc = -EAGAIN; |
213ecb31 | 931 | bool page_was_mapped = false; |
3f6c8272 | 932 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 933 | bool is_lru = !__PageMovable(page); |
95a402c3 | 934 | |
529ae9aa | 935 | if (!trylock_page(page)) { |
a6bc32b8 | 936 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 937 | goto out; |
3e7d3449 MG |
938 | |
939 | /* | |
940 | * It's not safe for direct compaction to call lock_page. | |
941 | * For example, during page readahead pages are added locked | |
942 | * to the LRU. Later, when the IO completes the pages are | |
943 | * marked uptodate and unlocked. However, the queueing | |
944 | * could be merging multiple pages for one bio (e.g. | |
d4388340 | 945 | * mpage_readahead). If an allocation happens for the |
3e7d3449 MG |
946 | * second or third page, the process can end up locking |
947 | * the same page twice and deadlocking. Rather than | |
948 | * trying to be clever about what pages can be locked, | |
949 | * avoid the use of lock_page for direct compaction | |
950 | * altogether. | |
951 | */ | |
952 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 953 | goto out; |
3e7d3449 | 954 | |
e24f0b8f CL |
955 | lock_page(page); |
956 | } | |
957 | ||
958 | if (PageWriteback(page)) { | |
11bc82d6 | 959 | /* |
fed5b64a | 960 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
961 | * necessary to wait for PageWriteback. In the async case, |
962 | * the retry loop is too short and in the sync-light case, | |
963 | * the overhead of stalling is too much | |
11bc82d6 | 964 | */ |
2916ecc0 JG |
965 | switch (mode) { |
966 | case MIGRATE_SYNC: | |
967 | case MIGRATE_SYNC_NO_COPY: | |
968 | break; | |
969 | default: | |
11bc82d6 | 970 | rc = -EBUSY; |
0a31bc97 | 971 | goto out_unlock; |
11bc82d6 AA |
972 | } |
973 | if (!force) | |
0a31bc97 | 974 | goto out_unlock; |
e24f0b8f CL |
975 | wait_on_page_writeback(page); |
976 | } | |
03f15c86 | 977 | |
e24f0b8f | 978 | /* |
68a9843f | 979 | * By try_to_migrate(), page->mapcount goes down to 0 here. In this case, |
dc386d4d | 980 | * we cannot notice that anon_vma is freed while we migrates a page. |
1ce82b69 | 981 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 982 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
983 | * File Caches may use write_page() or lock_page() in migration, then, |
984 | * just care Anon page here. | |
03f15c86 HD |
985 | * |
986 | * Only page_get_anon_vma() understands the subtleties of | |
987 | * getting a hold on an anon_vma from outside one of its mms. | |
988 | * But if we cannot get anon_vma, then we won't need it anyway, | |
989 | * because that implies that the anon page is no longer mapped | |
990 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 991 | */ |
03f15c86 | 992 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 993 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 994 | |
7db7671f HD |
995 | /* |
996 | * Block others from accessing the new page when we get around to | |
997 | * establishing additional references. We are usually the only one | |
998 | * holding a reference to newpage at this point. We used to have a BUG | |
999 | * here if trylock_page(newpage) fails, but would like to allow for | |
1000 | * cases where there might be a race with the previous use of newpage. | |
1001 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1002 | */ | |
1003 | if (unlikely(!trylock_page(newpage))) | |
1004 | goto out_unlock; | |
1005 | ||
bda807d4 MK |
1006 | if (unlikely(!is_lru)) { |
1007 | rc = move_to_new_page(newpage, page, mode); | |
1008 | goto out_unlock_both; | |
1009 | } | |
1010 | ||
dc386d4d | 1011 | /* |
62e1c553 SL |
1012 | * Corner case handling: |
1013 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1014 | * and treated as swapcache but it has no rmap yet. | |
1015 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1016 | * trigger a BUG. So handle it here. | |
d12b8951 | 1017 | * 2. An orphaned page (see truncate_cleanup_page) might have |
62e1c553 SL |
1018 | * fs-private metadata. The page can be picked up due to memory |
1019 | * offlining. Everywhere else except page reclaim, the page is | |
1020 | * invisible to the vm, so the page can not be migrated. So try to | |
1021 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1022 | */ |
62e1c553 | 1023 | if (!page->mapping) { |
309381fe | 1024 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1025 | if (page_has_private(page)) { |
62e1c553 | 1026 | try_to_free_buffers(page); |
7db7671f | 1027 | goto out_unlock_both; |
62e1c553 | 1028 | } |
7db7671f HD |
1029 | } else if (page_mapped(page)) { |
1030 | /* Establish migration ptes */ | |
03f15c86 HD |
1031 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1032 | page); | |
a98a2f0c | 1033 | try_to_migrate(page, 0); |
213ecb31 | 1034 | page_was_mapped = true; |
2ebba6b7 | 1035 | } |
dc386d4d | 1036 | |
e6a1530d | 1037 | if (!page_mapped(page)) |
5c3f9a67 | 1038 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1039 | |
5c3f9a67 HD |
1040 | if (page_was_mapped) |
1041 | remove_migration_ptes(page, | |
e388466d | 1042 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1043 | |
7db7671f HD |
1044 | out_unlock_both: |
1045 | unlock_page(newpage); | |
1046 | out_unlock: | |
3f6c8272 | 1047 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1048 | if (anon_vma) |
9e60109f | 1049 | put_anon_vma(anon_vma); |
e24f0b8f | 1050 | unlock_page(page); |
0dabec93 | 1051 | out: |
c6c919eb MK |
1052 | /* |
1053 | * If migration is successful, decrease refcount of the newpage | |
1054 | * which will not free the page because new page owner increased | |
1055 | * refcounter. As well, if it is LRU page, add the page to LRU | |
e0a352fa DH |
1056 | * list in here. Use the old state of the isolated source page to |
1057 | * determine if we migrated a LRU page. newpage was already unlocked | |
1058 | * and possibly modified by its owner - don't rely on the page | |
1059 | * state. | |
c6c919eb MK |
1060 | */ |
1061 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e0a352fa | 1062 | if (unlikely(!is_lru)) |
c6c919eb MK |
1063 | put_page(newpage); |
1064 | else | |
1065 | putback_lru_page(newpage); | |
1066 | } | |
1067 | ||
0dabec93 MK |
1068 | return rc; |
1069 | } | |
95a402c3 | 1070 | |
0dabec93 MK |
1071 | /* |
1072 | * Obtain the lock on page, remove all ptes and migrate the page | |
1073 | * to the newly allocated page in newpage. | |
1074 | */ | |
6ec4476a | 1075 | static int unmap_and_move(new_page_t get_new_page, |
ef2a5153 GU |
1076 | free_page_t put_new_page, |
1077 | unsigned long private, struct page *page, | |
add05cec | 1078 | int force, enum migrate_mode mode, |
dd4ae78a YS |
1079 | enum migrate_reason reason, |
1080 | struct list_head *ret) | |
0dabec93 | 1081 | { |
2def7424 | 1082 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1083 | struct page *newpage = NULL; |
0dabec93 | 1084 | |
94723aaf | 1085 | if (!thp_migration_supported() && PageTransHuge(page)) |
d532e2e5 | 1086 | return -ENOSYS; |
94723aaf | 1087 | |
0dabec93 MK |
1088 | if (page_count(page) == 1) { |
1089 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1090 | ClearPageActive(page); |
1091 | ClearPageUnevictable(page); | |
bda807d4 MK |
1092 | if (unlikely(__PageMovable(page))) { |
1093 | lock_page(page); | |
1094 | if (!PageMovable(page)) | |
356ea386 | 1095 | ClearPageIsolated(page); |
bda807d4 MK |
1096 | unlock_page(page); |
1097 | } | |
0dabec93 MK |
1098 | goto out; |
1099 | } | |
1100 | ||
74d4a579 YS |
1101 | newpage = get_new_page(page, private); |
1102 | if (!newpage) | |
1103 | return -ENOMEM; | |
1104 | ||
9c620e2b | 1105 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1106 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1107 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1108 | |
0dabec93 | 1109 | out: |
e24f0b8f | 1110 | if (rc != -EAGAIN) { |
0dabec93 MK |
1111 | /* |
1112 | * A page that has been migrated has all references | |
1113 | * removed and will be freed. A page that has not been | |
c23a0c99 | 1114 | * migrated will have kept its references and be restored. |
0dabec93 MK |
1115 | */ |
1116 | list_del(&page->lru); | |
dd4ae78a | 1117 | } |
6afcf8ef | 1118 | |
dd4ae78a YS |
1119 | /* |
1120 | * If migration is successful, releases reference grabbed during | |
1121 | * isolation. Otherwise, restore the page to right list unless | |
1122 | * we want to retry. | |
1123 | */ | |
1124 | if (rc == MIGRATEPAGE_SUCCESS) { | |
6afcf8ef ML |
1125 | /* |
1126 | * Compaction can migrate also non-LRU pages which are | |
1127 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1128 | * as __PageMovable | |
1129 | */ | |
1130 | if (likely(!__PageMovable(page))) | |
e8db67eb | 1131 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 1132 | page_is_file_lru(page), -thp_nr_pages(page)); |
c6c919eb | 1133 | |
79f5f8fa | 1134 | if (reason != MR_MEMORY_FAILURE) |
d7e69488 | 1135 | /* |
79f5f8fa | 1136 | * We release the page in page_handle_poison. |
d7e69488 | 1137 | */ |
79f5f8fa | 1138 | put_page(page); |
c6c919eb | 1139 | } else { |
dd4ae78a YS |
1140 | if (rc != -EAGAIN) |
1141 | list_add_tail(&page->lru, ret); | |
bda807d4 | 1142 | |
c6c919eb MK |
1143 | if (put_new_page) |
1144 | put_new_page(newpage, private); | |
1145 | else | |
1146 | put_page(newpage); | |
e24f0b8f | 1147 | } |
68711a74 | 1148 | |
e24f0b8f CL |
1149 | return rc; |
1150 | } | |
1151 | ||
290408d4 NH |
1152 | /* |
1153 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1154 | * | |
1155 | * This function doesn't wait the completion of hugepage I/O | |
1156 | * because there is no race between I/O and migration for hugepage. | |
1157 | * Note that currently hugepage I/O occurs only in direct I/O | |
1158 | * where no lock is held and PG_writeback is irrelevant, | |
1159 | * and writeback status of all subpages are counted in the reference | |
1160 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1161 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1162 | * This means that when we try to migrate hugepage whose subpages are | |
1163 | * doing direct I/O, some references remain after try_to_unmap() and | |
1164 | * hugepage migration fails without data corruption. | |
1165 | * | |
1166 | * There is also no race when direct I/O is issued on the page under migration, | |
1167 | * because then pte is replaced with migration swap entry and direct I/O code | |
1168 | * will wait in the page fault for migration to complete. | |
1169 | */ | |
1170 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1171 | free_page_t put_new_page, unsigned long private, |
1172 | struct page *hpage, int force, | |
dd4ae78a YS |
1173 | enum migrate_mode mode, int reason, |
1174 | struct list_head *ret) | |
290408d4 | 1175 | { |
2def7424 | 1176 | int rc = -EAGAIN; |
2ebba6b7 | 1177 | int page_was_mapped = 0; |
32665f2b | 1178 | struct page *new_hpage; |
290408d4 | 1179 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1180 | struct address_space *mapping = NULL; |
290408d4 | 1181 | |
83467efb | 1182 | /* |
7ed2c31d | 1183 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1184 | * This check is necessary because some callers of hugepage migration |
1185 | * like soft offline and memory hotremove don't walk through page | |
1186 | * tables or check whether the hugepage is pmd-based or not before | |
1187 | * kicking migration. | |
1188 | */ | |
100873d7 | 1189 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
dd4ae78a | 1190 | list_move_tail(&hpage->lru, ret); |
83467efb | 1191 | return -ENOSYS; |
32665f2b | 1192 | } |
83467efb | 1193 | |
71a64f61 MS |
1194 | if (page_count(hpage) == 1) { |
1195 | /* page was freed from under us. So we are done. */ | |
1196 | putback_active_hugepage(hpage); | |
1197 | return MIGRATEPAGE_SUCCESS; | |
1198 | } | |
1199 | ||
666feb21 | 1200 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1201 | if (!new_hpage) |
1202 | return -ENOMEM; | |
1203 | ||
290408d4 | 1204 | if (!trylock_page(hpage)) { |
2916ecc0 | 1205 | if (!force) |
290408d4 | 1206 | goto out; |
2916ecc0 JG |
1207 | switch (mode) { |
1208 | case MIGRATE_SYNC: | |
1209 | case MIGRATE_SYNC_NO_COPY: | |
1210 | break; | |
1211 | default: | |
1212 | goto out; | |
1213 | } | |
290408d4 NH |
1214 | lock_page(hpage); |
1215 | } | |
1216 | ||
cb6acd01 MK |
1217 | /* |
1218 | * Check for pages which are in the process of being freed. Without | |
1219 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1220 | * be called and we could leak usage counts for subpools. | |
1221 | */ | |
6acfb5ba | 1222 | if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) { |
cb6acd01 MK |
1223 | rc = -EBUSY; |
1224 | goto out_unlock; | |
1225 | } | |
1226 | ||
746b18d4 PZ |
1227 | if (PageAnon(hpage)) |
1228 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1229 | |
7db7671f HD |
1230 | if (unlikely(!trylock_page(new_hpage))) |
1231 | goto put_anon; | |
1232 | ||
2ebba6b7 | 1233 | if (page_mapped(hpage)) { |
336bf30e | 1234 | bool mapping_locked = false; |
a98a2f0c | 1235 | enum ttu_flags ttu = 0; |
336bf30e MK |
1236 | |
1237 | if (!PageAnon(hpage)) { | |
1238 | /* | |
1239 | * In shared mappings, try_to_unmap could potentially | |
1240 | * call huge_pmd_unshare. Because of this, take | |
1241 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1242 | * to let lower levels know we have taken the lock. | |
1243 | */ | |
1244 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1245 | if (unlikely(!mapping)) | |
1246 | goto unlock_put_anon; | |
1247 | ||
1248 | mapping_locked = true; | |
1249 | ttu |= TTU_RMAP_LOCKED; | |
1250 | } | |
c0d0381a | 1251 | |
a98a2f0c | 1252 | try_to_migrate(hpage, ttu); |
2ebba6b7 | 1253 | page_was_mapped = 1; |
336bf30e MK |
1254 | |
1255 | if (mapping_locked) | |
1256 | i_mmap_unlock_write(mapping); | |
2ebba6b7 | 1257 | } |
290408d4 NH |
1258 | |
1259 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1260 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1261 | |
336bf30e | 1262 | if (page_was_mapped) |
5c3f9a67 | 1263 | remove_migration_ptes(hpage, |
336bf30e | 1264 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1265 | |
c0d0381a | 1266 | unlock_put_anon: |
7db7671f HD |
1267 | unlock_page(new_hpage); |
1268 | ||
1269 | put_anon: | |
fd4a4663 | 1270 | if (anon_vma) |
9e60109f | 1271 | put_anon_vma(anon_vma); |
8e6ac7fa | 1272 | |
2def7424 | 1273 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1274 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1275 | put_new_page = NULL; |
1276 | } | |
8e6ac7fa | 1277 | |
cb6acd01 | 1278 | out_unlock: |
290408d4 | 1279 | unlock_page(hpage); |
09761333 | 1280 | out: |
dd4ae78a | 1281 | if (rc == MIGRATEPAGE_SUCCESS) |
b8ec1cee | 1282 | putback_active_hugepage(hpage); |
a04840c6 | 1283 | else if (rc != -EAGAIN) |
dd4ae78a | 1284 | list_move_tail(&hpage->lru, ret); |
68711a74 DR |
1285 | |
1286 | /* | |
1287 | * If migration was not successful and there's a freeing callback, use | |
1288 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1289 | * isolation. | |
1290 | */ | |
2def7424 | 1291 | if (put_new_page) |
68711a74 DR |
1292 | put_new_page(new_hpage, private); |
1293 | else | |
3aaa76e1 | 1294 | putback_active_hugepage(new_hpage); |
68711a74 | 1295 | |
290408d4 NH |
1296 | return rc; |
1297 | } | |
1298 | ||
d532e2e5 YS |
1299 | static inline int try_split_thp(struct page *page, struct page **page2, |
1300 | struct list_head *from) | |
1301 | { | |
1302 | int rc = 0; | |
1303 | ||
1304 | lock_page(page); | |
1305 | rc = split_huge_page_to_list(page, from); | |
1306 | unlock_page(page); | |
1307 | if (!rc) | |
1308 | list_safe_reset_next(page, *page2, lru); | |
1309 | ||
1310 | return rc; | |
1311 | } | |
1312 | ||
b20a3503 | 1313 | /* |
c73e5c9c SB |
1314 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1315 | * supplied as the target for the page migration | |
b20a3503 | 1316 | * |
c73e5c9c SB |
1317 | * @from: The list of pages to be migrated. |
1318 | * @get_new_page: The function used to allocate free pages to be used | |
1319 | * as the target of the page migration. | |
68711a74 DR |
1320 | * @put_new_page: The function used to free target pages if migration |
1321 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1322 | * @private: Private data to be passed on to get_new_page() |
1323 | * @mode: The migration mode that specifies the constraints for | |
1324 | * page migration, if any. | |
1325 | * @reason: The reason for page migration. | |
b5bade97 | 1326 | * @ret_succeeded: Set to the number of normal pages migrated successfully if |
5ac95884 | 1327 | * the caller passes a non-NULL pointer. |
b20a3503 | 1328 | * |
c73e5c9c SB |
1329 | * The function returns after 10 attempts or if no pages are movable any more |
1330 | * because the list has become empty or no retryable pages exist any more. | |
dd4ae78a YS |
1331 | * It is caller's responsibility to call putback_movable_pages() to return pages |
1332 | * to the LRU or free list only if ret != 0. | |
b20a3503 | 1333 | * |
5d39a7eb BW |
1334 | * Returns the number of {normal page, THP, hugetlb} that were not migrated, or |
1335 | * an error code. The number of THP splits will be considered as the number of | |
1336 | * non-migrated THP, no matter how many subpages of the THP are migrated successfully. | |
b20a3503 | 1337 | */ |
9c620e2b | 1338 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 | 1339 | free_page_t put_new_page, unsigned long private, |
5ac95884 | 1340 | enum migrate_mode mode, int reason, unsigned int *ret_succeeded) |
b20a3503 | 1341 | { |
e24f0b8f | 1342 | int retry = 1; |
1a5bae25 | 1343 | int thp_retry = 1; |
b20a3503 | 1344 | int nr_failed = 0; |
b5bade97 | 1345 | int nr_failed_pages = 0; |
5647bc29 | 1346 | int nr_succeeded = 0; |
1a5bae25 AK |
1347 | int nr_thp_succeeded = 0; |
1348 | int nr_thp_failed = 0; | |
1349 | int nr_thp_split = 0; | |
b20a3503 | 1350 | int pass = 0; |
1a5bae25 | 1351 | bool is_thp = false; |
b20a3503 CL |
1352 | struct page *page; |
1353 | struct page *page2; | |
1a5bae25 | 1354 | int rc, nr_subpages; |
dd4ae78a | 1355 | LIST_HEAD(ret_pages); |
b5bade97 | 1356 | LIST_HEAD(thp_split_pages); |
b0b515bf | 1357 | bool nosplit = (reason == MR_NUMA_MISPLACED); |
b5bade97 | 1358 | bool no_subpage_counting = false; |
b20a3503 | 1359 | |
7bc1aec5 LM |
1360 | trace_mm_migrate_pages_start(mode, reason); |
1361 | ||
b5bade97 | 1362 | thp_subpage_migration: |
1a5bae25 | 1363 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { |
e24f0b8f | 1364 | retry = 0; |
1a5bae25 | 1365 | thp_retry = 0; |
b20a3503 | 1366 | |
e24f0b8f | 1367 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1368 | retry: |
1a5bae25 AK |
1369 | /* |
1370 | * THP statistics is based on the source huge page. | |
1371 | * Capture required information that might get lost | |
1372 | * during migration. | |
1373 | */ | |
6c5c7b9f | 1374 | is_thp = PageTransHuge(page) && !PageHuge(page); |
5d39a7eb | 1375 | nr_subpages = compound_nr(page); |
e24f0b8f | 1376 | cond_resched(); |
2d1db3b1 | 1377 | |
31caf665 NH |
1378 | if (PageHuge(page)) |
1379 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1380 | put_new_page, private, page, |
dd4ae78a YS |
1381 | pass > 2, mode, reason, |
1382 | &ret_pages); | |
31caf665 | 1383 | else |
68711a74 | 1384 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec | 1385 | private, page, pass > 2, mode, |
dd4ae78a YS |
1386 | reason, &ret_pages); |
1387 | /* | |
1388 | * The rules are: | |
1389 | * Success: non hugetlb page will be freed, hugetlb | |
1390 | * page will be put back | |
1391 | * -EAGAIN: stay on the from list | |
1392 | * -ENOMEM: stay on the from list | |
1393 | * Other errno: put on ret_pages list then splice to | |
1394 | * from list | |
1395 | */ | |
e24f0b8f | 1396 | switch(rc) { |
d532e2e5 YS |
1397 | /* |
1398 | * THP migration might be unsupported or the | |
1399 | * allocation could've failed so we should | |
1400 | * retry on the same page with the THP split | |
1401 | * to base pages. | |
1402 | * | |
1403 | * Head page is retried immediately and tail | |
1404 | * pages are added to the tail of the list so | |
1405 | * we encounter them after the rest of the list | |
1406 | * is processed. | |
1407 | */ | |
1408 | case -ENOSYS: | |
1409 | /* THP migration is unsupported */ | |
1410 | if (is_thp) { | |
b5bade97 BW |
1411 | nr_thp_failed++; |
1412 | if (!try_split_thp(page, &page2, &thp_split_pages)) { | |
d532e2e5 YS |
1413 | nr_thp_split++; |
1414 | goto retry; | |
1415 | } | |
1416 | ||
b5bade97 | 1417 | nr_failed_pages += nr_subpages; |
d532e2e5 YS |
1418 | break; |
1419 | } | |
1420 | ||
1421 | /* Hugetlb migration is unsupported */ | |
b5bade97 BW |
1422 | if (!no_subpage_counting) |
1423 | nr_failed++; | |
5d39a7eb | 1424 | nr_failed_pages += nr_subpages; |
d532e2e5 | 1425 | break; |
95a402c3 | 1426 | case -ENOMEM: |
94723aaf | 1427 | /* |
d532e2e5 YS |
1428 | * When memory is low, don't bother to try to migrate |
1429 | * other pages, just exit. | |
b0b515bf | 1430 | * THP NUMA faulting doesn't split THP to retry. |
94723aaf | 1431 | */ |
b0b515bf | 1432 | if (is_thp && !nosplit) { |
b5bade97 BW |
1433 | nr_thp_failed++; |
1434 | if (!try_split_thp(page, &page2, &thp_split_pages)) { | |
1a5bae25 | 1435 | nr_thp_split++; |
94723aaf MH |
1436 | goto retry; |
1437 | } | |
6c5c7b9f | 1438 | |
b5bade97 | 1439 | nr_failed_pages += nr_subpages; |
1a5bae25 AK |
1440 | goto out; |
1441 | } | |
b5bade97 BW |
1442 | |
1443 | if (!no_subpage_counting) | |
1444 | nr_failed++; | |
5d39a7eb | 1445 | nr_failed_pages += nr_subpages; |
95a402c3 | 1446 | goto out; |
e24f0b8f | 1447 | case -EAGAIN: |
1a5bae25 AK |
1448 | if (is_thp) { |
1449 | thp_retry++; | |
1450 | break; | |
1451 | } | |
2d1db3b1 | 1452 | retry++; |
e24f0b8f | 1453 | break; |
78bd5209 | 1454 | case MIGRATEPAGE_SUCCESS: |
5d39a7eb | 1455 | nr_succeeded += nr_subpages; |
1a5bae25 AK |
1456 | if (is_thp) { |
1457 | nr_thp_succeeded++; | |
1a5bae25 AK |
1458 | break; |
1459 | } | |
e24f0b8f CL |
1460 | break; |
1461 | default: | |
354a3363 | 1462 | /* |
d532e2e5 | 1463 | * Permanent failure (-EBUSY, etc.): |
354a3363 NH |
1464 | * unlike -EAGAIN case, the failed page is |
1465 | * removed from migration page list and not | |
1466 | * retried in the next outer loop. | |
1467 | */ | |
1a5bae25 AK |
1468 | if (is_thp) { |
1469 | nr_thp_failed++; | |
b5bade97 | 1470 | nr_failed_pages += nr_subpages; |
1a5bae25 AK |
1471 | break; |
1472 | } | |
b5bade97 BW |
1473 | |
1474 | if (!no_subpage_counting) | |
1475 | nr_failed++; | |
5d39a7eb | 1476 | nr_failed_pages += nr_subpages; |
e24f0b8f | 1477 | break; |
2d1db3b1 | 1478 | } |
b20a3503 CL |
1479 | } |
1480 | } | |
b5bade97 | 1481 | nr_failed += retry; |
1a5bae25 | 1482 | nr_thp_failed += thp_retry; |
b5bade97 BW |
1483 | /* |
1484 | * Try to migrate subpages of fail-to-migrate THPs, no nr_failed | |
1485 | * counting in this round, since all subpages of a THP is counted | |
1486 | * as 1 failure in the first round. | |
1487 | */ | |
1488 | if (!list_empty(&thp_split_pages)) { | |
1489 | /* | |
1490 | * Move non-migrated pages (after 10 retries) to ret_pages | |
1491 | * to avoid migrating them again. | |
1492 | */ | |
1493 | list_splice_init(from, &ret_pages); | |
1494 | list_splice_init(&thp_split_pages, from); | |
1495 | no_subpage_counting = true; | |
1496 | retry = 1; | |
1497 | goto thp_subpage_migration; | |
1498 | } | |
1499 | ||
1500 | rc = nr_failed + nr_thp_failed; | |
95a402c3 | 1501 | out: |
dd4ae78a YS |
1502 | /* |
1503 | * Put the permanent failure page back to migration list, they | |
1504 | * will be put back to the right list by the caller. | |
1505 | */ | |
1506 | list_splice(&ret_pages, from); | |
1507 | ||
1a5bae25 | 1508 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
b5bade97 | 1509 | count_vm_events(PGMIGRATE_FAIL, nr_failed_pages); |
1a5bae25 AK |
1510 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); |
1511 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); | |
1512 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); | |
b5bade97 | 1513 | trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded, |
1a5bae25 | 1514 | nr_thp_failed, nr_thp_split, mode, reason); |
7b2a2d4a | 1515 | |
5ac95884 YS |
1516 | if (ret_succeeded) |
1517 | *ret_succeeded = nr_succeeded; | |
1518 | ||
78bd5209 | 1519 | return rc; |
b20a3503 | 1520 | } |
95a402c3 | 1521 | |
19fc7bed | 1522 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
b4b38223 | 1523 | { |
19fc7bed JK |
1524 | struct migration_target_control *mtc; |
1525 | gfp_t gfp_mask; | |
b4b38223 JK |
1526 | unsigned int order = 0; |
1527 | struct page *new_page = NULL; | |
19fc7bed JK |
1528 | int nid; |
1529 | int zidx; | |
1530 | ||
1531 | mtc = (struct migration_target_control *)private; | |
1532 | gfp_mask = mtc->gfp_mask; | |
1533 | nid = mtc->nid; | |
1534 | if (nid == NUMA_NO_NODE) | |
1535 | nid = page_to_nid(page); | |
b4b38223 | 1536 | |
d92bbc27 JK |
1537 | if (PageHuge(page)) { |
1538 | struct hstate *h = page_hstate(compound_head(page)); | |
1539 | ||
19fc7bed JK |
1540 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
1541 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); | |
d92bbc27 | 1542 | } |
b4b38223 JK |
1543 | |
1544 | if (PageTransHuge(page)) { | |
9933a0c8 JK |
1545 | /* |
1546 | * clear __GFP_RECLAIM to make the migration callback | |
1547 | * consistent with regular THP allocations. | |
1548 | */ | |
1549 | gfp_mask &= ~__GFP_RECLAIM; | |
b4b38223 JK |
1550 | gfp_mask |= GFP_TRANSHUGE; |
1551 | order = HPAGE_PMD_ORDER; | |
1552 | } | |
19fc7bed JK |
1553 | zidx = zone_idx(page_zone(page)); |
1554 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) | |
b4b38223 JK |
1555 | gfp_mask |= __GFP_HIGHMEM; |
1556 | ||
84172f4b | 1557 | new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask); |
b4b38223 JK |
1558 | |
1559 | if (new_page && PageTransHuge(new_page)) | |
1560 | prep_transhuge_page(new_page); | |
1561 | ||
1562 | return new_page; | |
1563 | } | |
1564 | ||
742755a1 | 1565 | #ifdef CONFIG_NUMA |
742755a1 | 1566 | |
a49bd4d7 | 1567 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1568 | { |
a49bd4d7 MH |
1569 | while (nr-- > 0) { |
1570 | if (put_user(value, status + start)) | |
1571 | return -EFAULT; | |
1572 | start++; | |
1573 | } | |
1574 | ||
1575 | return 0; | |
1576 | } | |
1577 | ||
1578 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1579 | struct list_head *pagelist, int node) | |
1580 | { | |
1581 | int err; | |
a0976311 JK |
1582 | struct migration_target_control mtc = { |
1583 | .nid = node, | |
1584 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
1585 | }; | |
a49bd4d7 | 1586 | |
a0976311 | 1587 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
5ac95884 | 1588 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); |
a49bd4d7 MH |
1589 | if (err) |
1590 | putback_movable_pages(pagelist); | |
1591 | return err; | |
742755a1 CL |
1592 | } |
1593 | ||
1594 | /* | |
a49bd4d7 MH |
1595 | * Resolves the given address to a struct page, isolates it from the LRU and |
1596 | * puts it to the given pagelist. | |
e0153fc2 YS |
1597 | * Returns: |
1598 | * errno - if the page cannot be found/isolated | |
1599 | * 0 - when it doesn't have to be migrated because it is already on the | |
1600 | * target node | |
1601 | * 1 - when it has been queued | |
742755a1 | 1602 | */ |
a49bd4d7 MH |
1603 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1604 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1605 | { |
a49bd4d7 MH |
1606 | struct vm_area_struct *vma; |
1607 | struct page *page; | |
742755a1 | 1608 | int err; |
742755a1 | 1609 | |
d8ed45c5 | 1610 | mmap_read_lock(mm); |
a49bd4d7 MH |
1611 | err = -EFAULT; |
1612 | vma = find_vma(mm, addr); | |
1613 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1614 | goto out; | |
742755a1 | 1615 | |
a49bd4d7 | 1616 | /* FOLL_DUMP to ignore special (like zero) pages */ |
87d2762e | 1617 | page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); |
89f5b7da | 1618 | |
a49bd4d7 MH |
1619 | err = PTR_ERR(page); |
1620 | if (IS_ERR(page)) | |
1621 | goto out; | |
89f5b7da | 1622 | |
a49bd4d7 MH |
1623 | err = -ENOENT; |
1624 | if (!page) | |
1625 | goto out; | |
742755a1 | 1626 | |
a49bd4d7 MH |
1627 | err = 0; |
1628 | if (page_to_nid(page) == node) | |
1629 | goto out_putpage; | |
742755a1 | 1630 | |
a49bd4d7 MH |
1631 | err = -EACCES; |
1632 | if (page_mapcount(page) > 1 && !migrate_all) | |
1633 | goto out_putpage; | |
742755a1 | 1634 | |
a49bd4d7 MH |
1635 | if (PageHuge(page)) { |
1636 | if (PageHead(page)) { | |
1637 | isolate_huge_page(page, pagelist); | |
e0153fc2 | 1638 | err = 1; |
e632a938 | 1639 | } |
a49bd4d7 MH |
1640 | } else { |
1641 | struct page *head; | |
e632a938 | 1642 | |
e8db67eb NH |
1643 | head = compound_head(page); |
1644 | err = isolate_lru_page(head); | |
cf608ac1 | 1645 | if (err) |
a49bd4d7 | 1646 | goto out_putpage; |
742755a1 | 1647 | |
e0153fc2 | 1648 | err = 1; |
a49bd4d7 MH |
1649 | list_add_tail(&head->lru, pagelist); |
1650 | mod_node_page_state(page_pgdat(head), | |
9de4f22a | 1651 | NR_ISOLATED_ANON + page_is_file_lru(head), |
6c357848 | 1652 | thp_nr_pages(head)); |
a49bd4d7 MH |
1653 | } |
1654 | out_putpage: | |
1655 | /* | |
1656 | * Either remove the duplicate refcount from | |
1657 | * isolate_lru_page() or drop the page ref if it was | |
1658 | * not isolated. | |
1659 | */ | |
1660 | put_page(page); | |
1661 | out: | |
d8ed45c5 | 1662 | mmap_read_unlock(mm); |
742755a1 CL |
1663 | return err; |
1664 | } | |
1665 | ||
7ca8783a WY |
1666 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
1667 | struct list_head *pagelist, int __user *status, | |
1668 | int start, int i, unsigned long nr_pages) | |
1669 | { | |
1670 | int err; | |
1671 | ||
5d7ae891 WY |
1672 | if (list_empty(pagelist)) |
1673 | return 0; | |
1674 | ||
7ca8783a WY |
1675 | err = do_move_pages_to_node(mm, pagelist, node); |
1676 | if (err) { | |
1677 | /* | |
1678 | * Positive err means the number of failed | |
1679 | * pages to migrate. Since we are going to | |
1680 | * abort and return the number of non-migrated | |
ab9dd4f8 | 1681 | * pages, so need to include the rest of the |
7ca8783a WY |
1682 | * nr_pages that have not been attempted as |
1683 | * well. | |
1684 | */ | |
1685 | if (err > 0) | |
1686 | err += nr_pages - i - 1; | |
1687 | return err; | |
1688 | } | |
1689 | return store_status(status, start, node, i - start); | |
1690 | } | |
1691 | ||
5e9a0f02 BG |
1692 | /* |
1693 | * Migrate an array of page address onto an array of nodes and fill | |
1694 | * the corresponding array of status. | |
1695 | */ | |
3268c63e | 1696 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1697 | unsigned long nr_pages, |
1698 | const void __user * __user *pages, | |
1699 | const int __user *nodes, | |
1700 | int __user *status, int flags) | |
1701 | { | |
a49bd4d7 MH |
1702 | int current_node = NUMA_NO_NODE; |
1703 | LIST_HEAD(pagelist); | |
1704 | int start, i; | |
1705 | int err = 0, err1; | |
35282a2d | 1706 | |
361a2a22 | 1707 | lru_cache_disable(); |
35282a2d | 1708 | |
a49bd4d7 MH |
1709 | for (i = start = 0; i < nr_pages; i++) { |
1710 | const void __user *p; | |
1711 | unsigned long addr; | |
1712 | int node; | |
3140a227 | 1713 | |
a49bd4d7 MH |
1714 | err = -EFAULT; |
1715 | if (get_user(p, pages + i)) | |
1716 | goto out_flush; | |
1717 | if (get_user(node, nodes + i)) | |
1718 | goto out_flush; | |
057d3389 | 1719 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1720 | |
1721 | err = -ENODEV; | |
1722 | if (node < 0 || node >= MAX_NUMNODES) | |
1723 | goto out_flush; | |
1724 | if (!node_state(node, N_MEMORY)) | |
1725 | goto out_flush; | |
5e9a0f02 | 1726 | |
a49bd4d7 MH |
1727 | err = -EACCES; |
1728 | if (!node_isset(node, task_nodes)) | |
1729 | goto out_flush; | |
1730 | ||
1731 | if (current_node == NUMA_NO_NODE) { | |
1732 | current_node = node; | |
1733 | start = i; | |
1734 | } else if (node != current_node) { | |
7ca8783a WY |
1735 | err = move_pages_and_store_status(mm, current_node, |
1736 | &pagelist, status, start, i, nr_pages); | |
a49bd4d7 MH |
1737 | if (err) |
1738 | goto out; | |
1739 | start = i; | |
1740 | current_node = node; | |
3140a227 BG |
1741 | } |
1742 | ||
a49bd4d7 MH |
1743 | /* |
1744 | * Errors in the page lookup or isolation are not fatal and we simply | |
1745 | * report them via status | |
1746 | */ | |
1747 | err = add_page_for_migration(mm, addr, current_node, | |
1748 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 | 1749 | |
d08221a0 | 1750 | if (err > 0) { |
e0153fc2 YS |
1751 | /* The page is successfully queued for migration */ |
1752 | continue; | |
1753 | } | |
3140a227 | 1754 | |
65462462 JH |
1755 | /* |
1756 | * The move_pages() man page does not have an -EEXIST choice, so | |
1757 | * use -EFAULT instead. | |
1758 | */ | |
1759 | if (err == -EEXIST) | |
1760 | err = -EFAULT; | |
1761 | ||
d08221a0 WY |
1762 | /* |
1763 | * If the page is already on the target node (!err), store the | |
1764 | * node, otherwise, store the err. | |
1765 | */ | |
1766 | err = store_status(status, i, err ? : current_node, 1); | |
a49bd4d7 MH |
1767 | if (err) |
1768 | goto out_flush; | |
5e9a0f02 | 1769 | |
7ca8783a WY |
1770 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
1771 | status, start, i, nr_pages); | |
4afdacec WY |
1772 | if (err) |
1773 | goto out; | |
a49bd4d7 | 1774 | current_node = NUMA_NO_NODE; |
3140a227 | 1775 | } |
a49bd4d7 MH |
1776 | out_flush: |
1777 | /* Make sure we do not overwrite the existing error */ | |
7ca8783a WY |
1778 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
1779 | status, start, i, nr_pages); | |
dfe9aa23 | 1780 | if (err >= 0) |
a49bd4d7 | 1781 | err = err1; |
5e9a0f02 | 1782 | out: |
361a2a22 | 1783 | lru_cache_enable(); |
5e9a0f02 BG |
1784 | return err; |
1785 | } | |
1786 | ||
742755a1 | 1787 | /* |
2f007e74 | 1788 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1789 | */ |
80bba129 BG |
1790 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1791 | const void __user **pages, int *status) | |
742755a1 | 1792 | { |
2f007e74 | 1793 | unsigned long i; |
2f007e74 | 1794 | |
d8ed45c5 | 1795 | mmap_read_lock(mm); |
742755a1 | 1796 | |
2f007e74 | 1797 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1798 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1799 | struct vm_area_struct *vma; |
1800 | struct page *page; | |
c095adbc | 1801 | int err = -EFAULT; |
2f007e74 | 1802 | |
059b8b48 LH |
1803 | vma = vma_lookup(mm, addr); |
1804 | if (!vma) | |
742755a1 CL |
1805 | goto set_status; |
1806 | ||
d899844e KS |
1807 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1808 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1809 | |
1810 | err = PTR_ERR(page); | |
1811 | if (IS_ERR(page)) | |
1812 | goto set_status; | |
1813 | ||
d899844e | 1814 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1815 | set_status: |
80bba129 BG |
1816 | *status = err; |
1817 | ||
1818 | pages++; | |
1819 | status++; | |
1820 | } | |
1821 | ||
d8ed45c5 | 1822 | mmap_read_unlock(mm); |
80bba129 BG |
1823 | } |
1824 | ||
5b1b561b AB |
1825 | static int get_compat_pages_array(const void __user *chunk_pages[], |
1826 | const void __user * __user *pages, | |
1827 | unsigned long chunk_nr) | |
1828 | { | |
1829 | compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; | |
1830 | compat_uptr_t p; | |
1831 | int i; | |
1832 | ||
1833 | for (i = 0; i < chunk_nr; i++) { | |
1834 | if (get_user(p, pages32 + i)) | |
1835 | return -EFAULT; | |
1836 | chunk_pages[i] = compat_ptr(p); | |
1837 | } | |
1838 | ||
1839 | return 0; | |
1840 | } | |
1841 | ||
80bba129 BG |
1842 | /* |
1843 | * Determine the nodes of a user array of pages and store it in | |
1844 | * a user array of status. | |
1845 | */ | |
1846 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1847 | const void __user * __user *pages, | |
1848 | int __user *status) | |
1849 | { | |
1850 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1851 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1852 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1853 | |
87b8d1ad PA |
1854 | while (nr_pages) { |
1855 | unsigned long chunk_nr; | |
80bba129 | 1856 | |
87b8d1ad PA |
1857 | chunk_nr = nr_pages; |
1858 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1859 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1860 | ||
5b1b561b AB |
1861 | if (in_compat_syscall()) { |
1862 | if (get_compat_pages_array(chunk_pages, pages, | |
1863 | chunk_nr)) | |
1864 | break; | |
1865 | } else { | |
1866 | if (copy_from_user(chunk_pages, pages, | |
1867 | chunk_nr * sizeof(*chunk_pages))) | |
1868 | break; | |
1869 | } | |
80bba129 BG |
1870 | |
1871 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1872 | ||
87b8d1ad PA |
1873 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1874 | break; | |
742755a1 | 1875 | |
87b8d1ad PA |
1876 | pages += chunk_nr; |
1877 | status += chunk_nr; | |
1878 | nr_pages -= chunk_nr; | |
1879 | } | |
1880 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1881 | } |
1882 | ||
4dc200ce | 1883 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
742755a1 | 1884 | { |
742755a1 | 1885 | struct task_struct *task; |
742755a1 | 1886 | struct mm_struct *mm; |
742755a1 | 1887 | |
4dc200ce ML |
1888 | /* |
1889 | * There is no need to check if current process has the right to modify | |
1890 | * the specified process when they are same. | |
1891 | */ | |
1892 | if (!pid) { | |
1893 | mmget(current->mm); | |
1894 | *mem_nodes = cpuset_mems_allowed(current); | |
1895 | return current->mm; | |
1896 | } | |
742755a1 CL |
1897 | |
1898 | /* Find the mm_struct */ | |
a879bf58 | 1899 | rcu_read_lock(); |
4dc200ce | 1900 | task = find_task_by_vpid(pid); |
742755a1 | 1901 | if (!task) { |
a879bf58 | 1902 | rcu_read_unlock(); |
4dc200ce | 1903 | return ERR_PTR(-ESRCH); |
742755a1 | 1904 | } |
3268c63e | 1905 | get_task_struct(task); |
742755a1 CL |
1906 | |
1907 | /* | |
1908 | * Check if this process has the right to modify the specified | |
197e7e52 | 1909 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1910 | */ |
197e7e52 | 1911 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1912 | rcu_read_unlock(); |
4dc200ce | 1913 | mm = ERR_PTR(-EPERM); |
5e9a0f02 | 1914 | goto out; |
742755a1 | 1915 | } |
c69e8d9c | 1916 | rcu_read_unlock(); |
742755a1 | 1917 | |
4dc200ce ML |
1918 | mm = ERR_PTR(security_task_movememory(task)); |
1919 | if (IS_ERR(mm)) | |
5e9a0f02 | 1920 | goto out; |
4dc200ce | 1921 | *mem_nodes = cpuset_mems_allowed(task); |
3268c63e | 1922 | mm = get_task_mm(task); |
4dc200ce | 1923 | out: |
3268c63e | 1924 | put_task_struct(task); |
6e8b09ea | 1925 | if (!mm) |
4dc200ce ML |
1926 | mm = ERR_PTR(-EINVAL); |
1927 | return mm; | |
1928 | } | |
1929 | ||
1930 | /* | |
1931 | * Move a list of pages in the address space of the currently executing | |
1932 | * process. | |
1933 | */ | |
1934 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
1935 | const void __user * __user *pages, | |
1936 | const int __user *nodes, | |
1937 | int __user *status, int flags) | |
1938 | { | |
1939 | struct mm_struct *mm; | |
1940 | int err; | |
1941 | nodemask_t task_nodes; | |
1942 | ||
1943 | /* Check flags */ | |
1944 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
6e8b09ea SL |
1945 | return -EINVAL; |
1946 | ||
4dc200ce ML |
1947 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
1948 | return -EPERM; | |
1949 | ||
1950 | mm = find_mm_struct(pid, &task_nodes); | |
1951 | if (IS_ERR(mm)) | |
1952 | return PTR_ERR(mm); | |
1953 | ||
6e8b09ea SL |
1954 | if (nodes) |
1955 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1956 | nodes, status, flags); | |
1957 | else | |
1958 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1959 | |
742755a1 CL |
1960 | mmput(mm); |
1961 | return err; | |
1962 | } | |
742755a1 | 1963 | |
7addf443 DB |
1964 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1965 | const void __user * __user *, pages, | |
1966 | const int __user *, nodes, | |
1967 | int __user *, status, int, flags) | |
1968 | { | |
1969 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1970 | } | |
1971 | ||
7039e1db PZ |
1972 | #ifdef CONFIG_NUMA_BALANCING |
1973 | /* | |
1974 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1975 | * pages. Currently it only checks the watermarks which crude | |
1976 | */ | |
1977 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1978 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1979 | { |
1980 | int z; | |
599d0c95 | 1981 | |
7039e1db PZ |
1982 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1983 | struct zone *zone = pgdat->node_zones + z; | |
1984 | ||
1985 | if (!populated_zone(zone)) | |
1986 | continue; | |
1987 | ||
7039e1db PZ |
1988 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1989 | if (!zone_watermark_ok(zone, 0, | |
1990 | high_wmark_pages(zone) + | |
1991 | nr_migrate_pages, | |
bfe9d006 | 1992 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
1993 | continue; |
1994 | return true; | |
1995 | } | |
1996 | return false; | |
1997 | } | |
1998 | ||
1999 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 2000 | unsigned long data) |
7039e1db PZ |
2001 | { |
2002 | int nid = (int) data; | |
2003 | struct page *newpage; | |
2004 | ||
96db800f | 2005 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
2006 | (GFP_HIGHUSER_MOVABLE | |
2007 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
2008 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 2009 | ~__GFP_RECLAIM, 0); |
bac0382c | 2010 | |
7039e1db PZ |
2011 | return newpage; |
2012 | } | |
2013 | ||
c5b5a3dd YS |
2014 | static struct page *alloc_misplaced_dst_page_thp(struct page *page, |
2015 | unsigned long data) | |
2016 | { | |
2017 | int nid = (int) data; | |
2018 | struct page *newpage; | |
2019 | ||
2020 | newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), | |
2021 | HPAGE_PMD_ORDER); | |
2022 | if (!newpage) | |
2023 | goto out; | |
2024 | ||
2025 | prep_transhuge_page(newpage); | |
2026 | ||
2027 | out: | |
2028 | return newpage; | |
2029 | } | |
2030 | ||
1c30e017 | 2031 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 2032 | { |
340ef390 | 2033 | int page_lru; |
2b9b624f | 2034 | int nr_pages = thp_nr_pages(page); |
c574bbe9 | 2035 | int order = compound_order(page); |
a8f60772 | 2036 | |
c574bbe9 | 2037 | VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); |
3abef4e6 | 2038 | |
662aeea7 YS |
2039 | /* Do not migrate THP mapped by multiple processes */ |
2040 | if (PageTransHuge(page) && total_mapcount(page) > 1) | |
2041 | return 0; | |
2042 | ||
7039e1db | 2043 | /* Avoid migrating to a node that is nearly full */ |
c574bbe9 HY |
2044 | if (!migrate_balanced_pgdat(pgdat, nr_pages)) { |
2045 | int z; | |
2046 | ||
2047 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) | |
2048 | return 0; | |
2049 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
2050 | if (populated_zone(pgdat->node_zones + z)) | |
2051 | break; | |
2052 | } | |
2053 | wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); | |
340ef390 | 2054 | return 0; |
c574bbe9 | 2055 | } |
7039e1db | 2056 | |
340ef390 HD |
2057 | if (isolate_lru_page(page)) |
2058 | return 0; | |
7039e1db | 2059 | |
9de4f22a | 2060 | page_lru = page_is_file_lru(page); |
599d0c95 | 2061 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
2b9b624f | 2062 | nr_pages); |
340ef390 | 2063 | |
149c33e1 | 2064 | /* |
340ef390 HD |
2065 | * Isolating the page has taken another reference, so the |
2066 | * caller's reference can be safely dropped without the page | |
2067 | * disappearing underneath us during migration. | |
149c33e1 MG |
2068 | */ |
2069 | put_page(page); | |
340ef390 | 2070 | return 1; |
b32967ff MG |
2071 | } |
2072 | ||
2073 | /* | |
2074 | * Attempt to migrate a misplaced page to the specified destination | |
2075 | * node. Caller is expected to have an elevated reference count on | |
2076 | * the page that will be dropped by this function before returning. | |
2077 | */ | |
1bc115d8 MG |
2078 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
2079 | int node) | |
b32967ff MG |
2080 | { |
2081 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 2082 | int isolated; |
b32967ff | 2083 | int nr_remaining; |
e39bb6be | 2084 | unsigned int nr_succeeded; |
b32967ff | 2085 | LIST_HEAD(migratepages); |
c5b5a3dd YS |
2086 | new_page_t *new; |
2087 | bool compound; | |
b5916c02 | 2088 | int nr_pages = thp_nr_pages(page); |
c5b5a3dd YS |
2089 | |
2090 | /* | |
2091 | * PTE mapped THP or HugeTLB page can't reach here so the page could | |
2092 | * be either base page or THP. And it must be head page if it is | |
2093 | * THP. | |
2094 | */ | |
2095 | compound = PageTransHuge(page); | |
2096 | ||
2097 | if (compound) | |
2098 | new = alloc_misplaced_dst_page_thp; | |
2099 | else | |
2100 | new = alloc_misplaced_dst_page; | |
b32967ff MG |
2101 | |
2102 | /* | |
1bc115d8 MG |
2103 | * Don't migrate file pages that are mapped in multiple processes |
2104 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 2105 | */ |
7ee820ee ML |
2106 | if (page_mapcount(page) != 1 && page_is_file_lru(page) && |
2107 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 2108 | goto out; |
b32967ff | 2109 | |
09a913a7 MG |
2110 | /* |
2111 | * Also do not migrate dirty pages as not all filesystems can move | |
2112 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2113 | */ | |
9de4f22a | 2114 | if (page_is_file_lru(page) && PageDirty(page)) |
09a913a7 MG |
2115 | goto out; |
2116 | ||
b32967ff MG |
2117 | isolated = numamigrate_isolate_page(pgdat, page); |
2118 | if (!isolated) | |
2119 | goto out; | |
2120 | ||
2121 | list_add(&page->lru, &migratepages); | |
c5b5a3dd | 2122 | nr_remaining = migrate_pages(&migratepages, *new, NULL, node, |
e39bb6be HY |
2123 | MIGRATE_ASYNC, MR_NUMA_MISPLACED, |
2124 | &nr_succeeded); | |
b32967ff | 2125 | if (nr_remaining) { |
59c82b70 JK |
2126 | if (!list_empty(&migratepages)) { |
2127 | list_del(&page->lru); | |
c5fc5c3a YS |
2128 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
2129 | page_is_file_lru(page), -nr_pages); | |
59c82b70 JK |
2130 | putback_lru_page(page); |
2131 | } | |
b32967ff | 2132 | isolated = 0; |
e39bb6be HY |
2133 | } |
2134 | if (nr_succeeded) { | |
2135 | count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); | |
2136 | if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) | |
2137 | mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, | |
2138 | nr_succeeded); | |
2139 | } | |
7039e1db | 2140 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2141 | return isolated; |
340ef390 HD |
2142 | |
2143 | out: | |
2144 | put_page(page); | |
2145 | return 0; | |
7039e1db | 2146 | } |
220018d3 | 2147 | #endif /* CONFIG_NUMA_BALANCING */ |
7039e1db | 2148 | #endif /* CONFIG_NUMA */ |
8763cb45 | 2149 | |
9b2ed9cb | 2150 | #ifdef CONFIG_DEVICE_PRIVATE |
843e1be1 | 2151 | static int migrate_vma_collect_skip(unsigned long start, |
8763cb45 JG |
2152 | unsigned long end, |
2153 | struct mm_walk *walk) | |
2154 | { | |
2155 | struct migrate_vma *migrate = walk->private; | |
2156 | unsigned long addr; | |
2157 | ||
872ea707 | 2158 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
8315ada7 | 2159 | migrate->dst[migrate->npages] = 0; |
843e1be1 | 2160 | migrate->src[migrate->npages++] = 0; |
8315ada7 JG |
2161 | } |
2162 | ||
2163 | return 0; | |
2164 | } | |
2165 | ||
843e1be1 | 2166 | static int migrate_vma_collect_hole(unsigned long start, |
8315ada7 | 2167 | unsigned long end, |
843e1be1 | 2168 | __always_unused int depth, |
8315ada7 JG |
2169 | struct mm_walk *walk) |
2170 | { | |
2171 | struct migrate_vma *migrate = walk->private; | |
2172 | unsigned long addr; | |
2173 | ||
843e1be1 ML |
2174 | /* Only allow populating anonymous memory. */ |
2175 | if (!vma_is_anonymous(walk->vma)) | |
2176 | return migrate_vma_collect_skip(start, end, walk); | |
2177 | ||
872ea707 | 2178 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
843e1be1 | 2179 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8763cb45 | 2180 | migrate->dst[migrate->npages] = 0; |
843e1be1 ML |
2181 | migrate->npages++; |
2182 | migrate->cpages++; | |
8763cb45 JG |
2183 | } |
2184 | ||
2185 | return 0; | |
2186 | } | |
2187 | ||
2188 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2189 | unsigned long start, | |
2190 | unsigned long end, | |
2191 | struct mm_walk *walk) | |
2192 | { | |
2193 | struct migrate_vma *migrate = walk->private; | |
2194 | struct vm_area_struct *vma = walk->vma; | |
2195 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2196 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2197 | spinlock_t *ptl; |
2198 | pte_t *ptep; | |
2199 | ||
2200 | again: | |
2201 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2202 | return migrate_vma_collect_hole(start, end, -1, walk); |
8763cb45 JG |
2203 | |
2204 | if (pmd_trans_huge(*pmdp)) { | |
2205 | struct page *page; | |
2206 | ||
2207 | ptl = pmd_lock(mm, pmdp); | |
2208 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2209 | spin_unlock(ptl); | |
2210 | goto again; | |
2211 | } | |
2212 | ||
2213 | page = pmd_page(*pmdp); | |
2214 | if (is_huge_zero_page(page)) { | |
2215 | spin_unlock(ptl); | |
2216 | split_huge_pmd(vma, pmdp, addr); | |
2217 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2218 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2219 | walk); |
2220 | } else { | |
2221 | int ret; | |
2222 | ||
2223 | get_page(page); | |
2224 | spin_unlock(ptl); | |
2225 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2226 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2227 | walk); |
2228 | ret = split_huge_page(page); | |
2229 | unlock_page(page); | |
2230 | put_page(page); | |
8315ada7 JG |
2231 | if (ret) |
2232 | return migrate_vma_collect_skip(start, end, | |
2233 | walk); | |
2234 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2235 | return migrate_vma_collect_hole(start, end, -1, |
8763cb45 JG |
2236 | walk); |
2237 | } | |
2238 | } | |
2239 | ||
2240 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2241 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2242 | |
2243 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2244 | arch_enter_lazy_mmu_mode(); |
2245 | ||
8763cb45 | 2246 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
800bb1c8 | 2247 | unsigned long mpfn = 0, pfn; |
8763cb45 | 2248 | struct page *page; |
8c3328f1 | 2249 | swp_entry_t entry; |
8763cb45 JG |
2250 | pte_t pte; |
2251 | ||
2252 | pte = *ptep; | |
8763cb45 | 2253 | |
a5430dda | 2254 | if (pte_none(pte)) { |
0744f280 RC |
2255 | if (vma_is_anonymous(vma)) { |
2256 | mpfn = MIGRATE_PFN_MIGRATE; | |
2257 | migrate->cpages++; | |
2258 | } | |
8763cb45 JG |
2259 | goto next; |
2260 | } | |
2261 | ||
a5430dda | 2262 | if (!pte_present(pte)) { |
a5430dda JG |
2263 | /* |
2264 | * Only care about unaddressable device page special | |
2265 | * page table entry. Other special swap entries are not | |
2266 | * migratable, and we ignore regular swapped page. | |
2267 | */ | |
2268 | entry = pte_to_swp_entry(pte); | |
2269 | if (!is_device_private_entry(entry)) | |
2270 | goto next; | |
2271 | ||
af5cdaf8 | 2272 | page = pfn_swap_entry_to_page(entry); |
5143192c RC |
2273 | if (!(migrate->flags & |
2274 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
2275 | page->pgmap->owner != migrate->pgmap_owner) | |
800bb1c8 CH |
2276 | goto next; |
2277 | ||
06d462be CH |
2278 | mpfn = migrate_pfn(page_to_pfn(page)) | |
2279 | MIGRATE_PFN_MIGRATE; | |
4dd845b5 | 2280 | if (is_writable_device_private_entry(entry)) |
a5430dda JG |
2281 | mpfn |= MIGRATE_PFN_WRITE; |
2282 | } else { | |
5143192c | 2283 | if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) |
800bb1c8 | 2284 | goto next; |
276f756d | 2285 | pfn = pte_pfn(pte); |
8315ada7 JG |
2286 | if (is_zero_pfn(pfn)) { |
2287 | mpfn = MIGRATE_PFN_MIGRATE; | |
2288 | migrate->cpages++; | |
8315ada7 JG |
2289 | goto next; |
2290 | } | |
25b2995a | 2291 | page = vm_normal_page(migrate->vma, addr, pte); |
a5430dda JG |
2292 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2293 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2294 | } | |
2295 | ||
8763cb45 | 2296 | /* FIXME support THP */ |
8763cb45 | 2297 | if (!page || !page->mapping || PageTransCompound(page)) { |
276f756d | 2298 | mpfn = 0; |
8763cb45 JG |
2299 | goto next; |
2300 | } | |
2301 | ||
2302 | /* | |
2303 | * By getting a reference on the page we pin it and that blocks | |
2304 | * any kind of migration. Side effect is that it "freezes" the | |
2305 | * pte. | |
2306 | * | |
2307 | * We drop this reference after isolating the page from the lru | |
2308 | * for non device page (device page are not on the lru and thus | |
2309 | * can't be dropped from it). | |
2310 | */ | |
2311 | get_page(page); | |
8763cb45 | 2312 | |
8c3328f1 JG |
2313 | /* |
2314 | * Optimize for the common case where page is only mapped once | |
2315 | * in one process. If we can lock the page, then we can safely | |
2316 | * set up a special migration page table entry now. | |
2317 | */ | |
2318 | if (trylock_page(page)) { | |
2319 | pte_t swp_pte; | |
2320 | ||
ab09243a | 2321 | migrate->cpages++; |
8c3328f1 JG |
2322 | ptep_get_and_clear(mm, addr, ptep); |
2323 | ||
2324 | /* Setup special migration page table entry */ | |
4dd845b5 AP |
2325 | if (mpfn & MIGRATE_PFN_WRITE) |
2326 | entry = make_writable_migration_entry( | |
2327 | page_to_pfn(page)); | |
2328 | else | |
2329 | entry = make_readable_migration_entry( | |
2330 | page_to_pfn(page)); | |
8c3328f1 | 2331 | swp_pte = swp_entry_to_pte(entry); |
ad7df764 AP |
2332 | if (pte_present(pte)) { |
2333 | if (pte_soft_dirty(pte)) | |
2334 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2335 | if (pte_uffd_wp(pte)) | |
2336 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2337 | } else { | |
2338 | if (pte_swp_soft_dirty(pte)) | |
2339 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2340 | if (pte_swp_uffd_wp(pte)) | |
2341 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2342 | } | |
8c3328f1 JG |
2343 | set_pte_at(mm, addr, ptep, swp_pte); |
2344 | ||
2345 | /* | |
2346 | * This is like regular unmap: we remove the rmap and | |
2347 | * drop page refcount. Page won't be freed, as we took | |
2348 | * a reference just above. | |
2349 | */ | |
2350 | page_remove_rmap(page, false); | |
2351 | put_page(page); | |
a5430dda JG |
2352 | |
2353 | if (pte_present(pte)) | |
2354 | unmapped++; | |
ab09243a AP |
2355 | } else { |
2356 | put_page(page); | |
2357 | mpfn = 0; | |
8c3328f1 JG |
2358 | } |
2359 | ||
8763cb45 | 2360 | next: |
a5430dda | 2361 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2362 | migrate->src[migrate->npages++] = mpfn; |
2363 | } | |
8c3328f1 | 2364 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2365 | pte_unmap_unlock(ptep - 1, ptl); |
2366 | ||
8c3328f1 JG |
2367 | /* Only flush the TLB if we actually modified any entries */ |
2368 | if (unmapped) | |
2369 | flush_tlb_range(walk->vma, start, end); | |
2370 | ||
8763cb45 JG |
2371 | return 0; |
2372 | } | |
2373 | ||
7b86ac33 CH |
2374 | static const struct mm_walk_ops migrate_vma_walk_ops = { |
2375 | .pmd_entry = migrate_vma_collect_pmd, | |
2376 | .pte_hole = migrate_vma_collect_hole, | |
2377 | }; | |
2378 | ||
8763cb45 JG |
2379 | /* |
2380 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2381 | * @migrate: migrate struct containing all migration information | |
2382 | * | |
2383 | * This will walk the CPU page table. For each virtual address backed by a | |
2384 | * valid page, it updates the src array and takes a reference on the page, in | |
2385 | * order to pin the page until we lock it and unmap it. | |
2386 | */ | |
2387 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2388 | { | |
ac46d4f3 | 2389 | struct mmu_notifier_range range; |
8763cb45 | 2390 | |
998427b3 RC |
2391 | /* |
2392 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
2393 | * that the registered device driver can skip invalidating device | |
2394 | * private page mappings that won't be migrated. | |
2395 | */ | |
6b49bf6d AP |
2396 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0, |
2397 | migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end, | |
c1a06df6 | 2398 | migrate->pgmap_owner); |
ac46d4f3 | 2399 | mmu_notifier_invalidate_range_start(&range); |
8763cb45 | 2400 | |
7b86ac33 CH |
2401 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, |
2402 | &migrate_vma_walk_ops, migrate); | |
2403 | ||
2404 | mmu_notifier_invalidate_range_end(&range); | |
8763cb45 JG |
2405 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2406 | } | |
2407 | ||
2408 | /* | |
2409 | * migrate_vma_check_page() - check if page is pinned or not | |
2410 | * @page: struct page to check | |
2411 | * | |
2412 | * Pinned pages cannot be migrated. This is the same test as in | |
3417013e | 2413 | * folio_migrate_mapping(), except that here we allow migration of a |
8763cb45 JG |
2414 | * ZONE_DEVICE page. |
2415 | */ | |
2416 | static bool migrate_vma_check_page(struct page *page) | |
2417 | { | |
2418 | /* | |
2419 | * One extra ref because caller holds an extra reference, either from | |
2420 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2421 | * a device page. | |
2422 | */ | |
2423 | int extra = 1; | |
2424 | ||
2425 | /* | |
2426 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2427 | * check them than regular pages, because they can be mapped with a pmd | |
2428 | * or with a pte (split pte mapping). | |
2429 | */ | |
2430 | if (PageCompound(page)) | |
2431 | return false; | |
2432 | ||
a5430dda | 2433 | /* Page from ZONE_DEVICE have one extra reference */ |
ffa65753 AP |
2434 | if (is_zone_device_page(page)) |
2435 | extra++; | |
a5430dda | 2436 | |
df6ad698 JG |
2437 | /* For file back page */ |
2438 | if (page_mapping(page)) | |
2439 | extra += 1 + page_has_private(page); | |
2440 | ||
8763cb45 JG |
2441 | if ((page_count(page) - extra) > page_mapcount(page)) |
2442 | return false; | |
2443 | ||
2444 | return true; | |
2445 | } | |
2446 | ||
2447 | /* | |
ab09243a | 2448 | * migrate_vma_unmap() - replace page mapping with special migration pte entry |
8763cb45 JG |
2449 | * @migrate: migrate struct containing all migration information |
2450 | * | |
ab09243a AP |
2451 | * Isolate pages from the LRU and replace mappings (CPU page table pte) with a |
2452 | * special migration pte entry and check if it has been pinned. Pinned pages are | |
2453 | * restored because we cannot migrate them. | |
2454 | * | |
2455 | * This is the last step before we call the device driver callback to allocate | |
2456 | * destination memory and copy contents of original page over to new page. | |
8763cb45 | 2457 | */ |
ab09243a | 2458 | static void migrate_vma_unmap(struct migrate_vma *migrate) |
8763cb45 JG |
2459 | { |
2460 | const unsigned long npages = migrate->npages; | |
f1e8db04 | 2461 | unsigned long i, restore = 0; |
8763cb45 | 2462 | bool allow_drain = true; |
8763cb45 JG |
2463 | |
2464 | lru_add_drain(); | |
2465 | ||
ab09243a | 2466 | for (i = 0; i < npages; i++) { |
8763cb45 JG |
2467 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2468 | ||
2469 | if (!page) | |
2470 | continue; | |
2471 | ||
a5430dda JG |
2472 | /* ZONE_DEVICE pages are not on LRU */ |
2473 | if (!is_zone_device_page(page)) { | |
2474 | if (!PageLRU(page) && allow_drain) { | |
2475 | /* Drain CPU's pagevec */ | |
2476 | lru_add_drain_all(); | |
2477 | allow_drain = false; | |
2478 | } | |
8763cb45 | 2479 | |
a5430dda | 2480 | if (isolate_lru_page(page)) { |
ab09243a AP |
2481 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2482 | migrate->cpages--; | |
2483 | restore++; | |
a5430dda | 2484 | continue; |
8c3328f1 | 2485 | } |
a5430dda JG |
2486 | |
2487 | /* Drop the reference we took in collect */ | |
2488 | put_page(page); | |
8763cb45 JG |
2489 | } |
2490 | ||
ab09243a AP |
2491 | if (page_mapped(page)) |
2492 | try_to_migrate(page, 0); | |
8c3328f1 | 2493 | |
ab09243a AP |
2494 | if (page_mapped(page) || !migrate_vma_check_page(page)) { |
2495 | if (!is_zone_device_page(page)) { | |
2496 | get_page(page); | |
2497 | putback_lru_page(page); | |
8c3328f1 | 2498 | } |
8c3328f1 | 2499 | |
ab09243a AP |
2500 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2501 | migrate->cpages--; | |
2502 | restore++; | |
8c3328f1 | 2503 | continue; |
8763cb45 JG |
2504 | } |
2505 | } | |
2506 | ||
f1e8db04 | 2507 | for (i = 0; i < npages && restore; i++) { |
8763cb45 JG |
2508 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2509 | ||
2510 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2511 | continue; | |
2512 | ||
2513 | remove_migration_ptes(page, page, false); | |
2514 | ||
2515 | migrate->src[i] = 0; | |
2516 | unlock_page(page); | |
ab09243a | 2517 | put_page(page); |
8763cb45 | 2518 | restore--; |
8763cb45 JG |
2519 | } |
2520 | } | |
2521 | ||
a7d1f22b CH |
2522 | /** |
2523 | * migrate_vma_setup() - prepare to migrate a range of memory | |
eaf444de | 2524 | * @args: contains the vma, start, and pfns arrays for the migration |
a7d1f22b CH |
2525 | * |
2526 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2527 | * without an error. | |
2528 | * | |
2529 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2530 | * the pages backing each virtual address in the range, saving them inside the | |
2531 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2532 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2533 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2534 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2535 | * remapping and unlocking those pages. | |
2536 | * | |
2537 | * The caller should then allocate destination memory and copy source memory to | |
2538 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2539 | * flag set). Once these are allocated and copied, the caller must update each | |
2540 | * corresponding entry in the dst array with the pfn value of the destination | |
ab09243a AP |
2541 | * page and with MIGRATE_PFN_VALID. Destination pages must be locked via |
2542 | * lock_page(). | |
a7d1f22b CH |
2543 | * |
2544 | * Note that the caller does not have to migrate all the pages that are marked | |
2545 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2546 | * device memory to system memory. If the caller cannot migrate a device page | |
2547 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2548 | * consequences for the userspace process, so it must be avoided if at all | |
2549 | * possible. | |
2550 | * | |
2551 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2552 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
f0953a1b IM |
2553 | * allowing the caller to allocate device memory for those unbacked virtual |
2554 | * addresses. For this the caller simply has to allocate device memory and | |
a7d1f22b | 2555 | * properly set the destination entry like for regular migration. Note that |
f0953a1b IM |
2556 | * this can still fail, and thus inside the device driver you must check if the |
2557 | * migration was successful for those entries after calling migrate_vma_pages(), | |
a7d1f22b CH |
2558 | * just like for regular migration. |
2559 | * | |
2560 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2561 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2562 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2563 | * then migrate_vma_pages() to migrate struct page information from the source | |
2564 | * struct page to the destination struct page. If it fails to migrate the | |
2565 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2566 | * src array. | |
2567 | * | |
2568 | * At this point all successfully migrated pages have an entry in the src | |
2569 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2570 | * array entry with MIGRATE_PFN_VALID flag set. | |
2571 | * | |
2572 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2573 | * successfully migrated, and which were not. Successfully migrated pages will | |
2574 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2575 | * | |
2576 | * It is safe to update device page table after migrate_vma_pages() because | |
c1e8d7c6 | 2577 | * both destination and source page are still locked, and the mmap_lock is held |
a7d1f22b CH |
2578 | * in read mode (hence no one can unmap the range being migrated). |
2579 | * | |
2580 | * Once the caller is done cleaning up things and updating its page table (if it | |
2581 | * chose to do so, this is not an obligation) it finally calls | |
2582 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2583 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2584 | * point to the original source pages. | |
2585 | */ | |
2586 | int migrate_vma_setup(struct migrate_vma *args) | |
2587 | { | |
2588 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2589 | ||
2590 | args->start &= PAGE_MASK; | |
2591 | args->end &= PAGE_MASK; | |
2592 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2593 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2594 | return -EINVAL; | |
2595 | if (nr_pages <= 0) | |
2596 | return -EINVAL; | |
2597 | if (args->start < args->vma->vm_start || | |
2598 | args->start >= args->vma->vm_end) | |
2599 | return -EINVAL; | |
2600 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2601 | return -EINVAL; | |
2602 | if (!args->src || !args->dst) | |
2603 | return -EINVAL; | |
2604 | ||
2605 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2606 | args->cpages = 0; | |
2607 | args->npages = 0; | |
2608 | ||
2609 | migrate_vma_collect(args); | |
2610 | ||
a7d1f22b CH |
2611 | if (args->cpages) |
2612 | migrate_vma_unmap(args); | |
2613 | ||
2614 | /* | |
2615 | * At this point pages are locked and unmapped, and thus they have | |
2616 | * stable content and can safely be copied to destination memory that | |
2617 | * is allocated by the drivers. | |
2618 | */ | |
2619 | return 0; | |
2620 | ||
2621 | } | |
2622 | EXPORT_SYMBOL(migrate_vma_setup); | |
2623 | ||
34290e2c RC |
2624 | /* |
2625 | * This code closely matches the code in: | |
2626 | * __handle_mm_fault() | |
2627 | * handle_pte_fault() | |
2628 | * do_anonymous_page() | |
2629 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
2630 | * private page. | |
2631 | */ | |
8315ada7 JG |
2632 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2633 | unsigned long addr, | |
2634 | struct page *page, | |
d85c6db4 | 2635 | unsigned long *src) |
8315ada7 JG |
2636 | { |
2637 | struct vm_area_struct *vma = migrate->vma; | |
2638 | struct mm_struct *mm = vma->vm_mm; | |
8315ada7 JG |
2639 | bool flush = false; |
2640 | spinlock_t *ptl; | |
2641 | pte_t entry; | |
2642 | pgd_t *pgdp; | |
2643 | p4d_t *p4dp; | |
2644 | pud_t *pudp; | |
2645 | pmd_t *pmdp; | |
2646 | pte_t *ptep; | |
2647 | ||
2648 | /* Only allow populating anonymous memory */ | |
2649 | if (!vma_is_anonymous(vma)) | |
2650 | goto abort; | |
2651 | ||
2652 | pgdp = pgd_offset(mm, addr); | |
2653 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2654 | if (!p4dp) | |
2655 | goto abort; | |
2656 | pudp = pud_alloc(mm, p4dp, addr); | |
2657 | if (!pudp) | |
2658 | goto abort; | |
2659 | pmdp = pmd_alloc(mm, pudp, addr); | |
2660 | if (!pmdp) | |
2661 | goto abort; | |
2662 | ||
2663 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2664 | goto abort; | |
2665 | ||
2666 | /* | |
2667 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2668 | * pte_offset_map() on pmds where a huge pmd might be created | |
2669 | * from a different thread. | |
2670 | * | |
3e4e28c5 | 2671 | * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when |
8315ada7 JG |
2672 | * parallel threads are excluded by other means. |
2673 | * | |
3e4e28c5 | 2674 | * Here we only have mmap_read_lock(mm). |
8315ada7 | 2675 | */ |
4cf58924 | 2676 | if (pte_alloc(mm, pmdp)) |
8315ada7 JG |
2677 | goto abort; |
2678 | ||
2679 | /* See the comment in pte_alloc_one_map() */ | |
2680 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2681 | goto abort; | |
2682 | ||
2683 | if (unlikely(anon_vma_prepare(vma))) | |
2684 | goto abort; | |
8f425e4e | 2685 | if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL)) |
8315ada7 JG |
2686 | goto abort; |
2687 | ||
2688 | /* | |
2689 | * The memory barrier inside __SetPageUptodate makes sure that | |
2690 | * preceding stores to the page contents become visible before | |
2691 | * the set_pte_at() write. | |
2692 | */ | |
2693 | __SetPageUptodate(page); | |
2694 | ||
df6ad698 JG |
2695 | if (is_zone_device_page(page)) { |
2696 | if (is_device_private_page(page)) { | |
2697 | swp_entry_t swp_entry; | |
2698 | ||
4dd845b5 AP |
2699 | if (vma->vm_flags & VM_WRITE) |
2700 | swp_entry = make_writable_device_private_entry( | |
2701 | page_to_pfn(page)); | |
2702 | else | |
2703 | swp_entry = make_readable_device_private_entry( | |
2704 | page_to_pfn(page)); | |
df6ad698 | 2705 | entry = swp_entry_to_pte(swp_entry); |
34f5e9b9 ML |
2706 | } else { |
2707 | /* | |
2708 | * For now we only support migrating to un-addressable | |
2709 | * device memory. | |
2710 | */ | |
2711 | pr_warn_once("Unsupported ZONE_DEVICE page type.\n"); | |
2712 | goto abort; | |
df6ad698 | 2713 | } |
8315ada7 JG |
2714 | } else { |
2715 | entry = mk_pte(page, vma->vm_page_prot); | |
2716 | if (vma->vm_flags & VM_WRITE) | |
2717 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2718 | } | |
2719 | ||
2720 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2721 | ||
34290e2c RC |
2722 | if (check_stable_address_space(mm)) |
2723 | goto unlock_abort; | |
2724 | ||
8315ada7 JG |
2725 | if (pte_present(*ptep)) { |
2726 | unsigned long pfn = pte_pfn(*ptep); | |
2727 | ||
c23a0c99 RC |
2728 | if (!is_zero_pfn(pfn)) |
2729 | goto unlock_abort; | |
8315ada7 | 2730 | flush = true; |
c23a0c99 RC |
2731 | } else if (!pte_none(*ptep)) |
2732 | goto unlock_abort; | |
8315ada7 JG |
2733 | |
2734 | /* | |
c23a0c99 | 2735 | * Check for userfaultfd but do not deliver the fault. Instead, |
8315ada7 JG |
2736 | * just back off. |
2737 | */ | |
c23a0c99 RC |
2738 | if (userfaultfd_missing(vma)) |
2739 | goto unlock_abort; | |
8315ada7 JG |
2740 | |
2741 | inc_mm_counter(mm, MM_ANONPAGES); | |
be5d0a74 | 2742 | page_add_new_anon_rmap(page, vma, addr, false); |
8315ada7 | 2743 | if (!is_zone_device_page(page)) |
b518154e | 2744 | lru_cache_add_inactive_or_unevictable(page, vma); |
8315ada7 JG |
2745 | get_page(page); |
2746 | ||
2747 | if (flush) { | |
2748 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2749 | ptep_clear_flush_notify(vma, addr, ptep); | |
2750 | set_pte_at_notify(mm, addr, ptep, entry); | |
2751 | update_mmu_cache(vma, addr, ptep); | |
2752 | } else { | |
2753 | /* No need to invalidate - it was non-present before */ | |
2754 | set_pte_at(mm, addr, ptep, entry); | |
2755 | update_mmu_cache(vma, addr, ptep); | |
2756 | } | |
2757 | ||
2758 | pte_unmap_unlock(ptep, ptl); | |
2759 | *src = MIGRATE_PFN_MIGRATE; | |
2760 | return; | |
2761 | ||
c23a0c99 RC |
2762 | unlock_abort: |
2763 | pte_unmap_unlock(ptep, ptl); | |
8315ada7 JG |
2764 | abort: |
2765 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2766 | } | |
2767 | ||
a7d1f22b | 2768 | /** |
8763cb45 JG |
2769 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
2770 | * @migrate: migrate struct containing all migration information | |
2771 | * | |
2772 | * This migrates struct page meta-data from source struct page to destination | |
2773 | * struct page. This effectively finishes the migration from source page to the | |
2774 | * destination page. | |
2775 | */ | |
a7d1f22b | 2776 | void migrate_vma_pages(struct migrate_vma *migrate) |
8763cb45 JG |
2777 | { |
2778 | const unsigned long npages = migrate->npages; | |
2779 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
2780 | struct mmu_notifier_range range; |
2781 | unsigned long addr, i; | |
8315ada7 | 2782 | bool notified = false; |
8763cb45 JG |
2783 | |
2784 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2785 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2786 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2787 | struct address_space *mapping; | |
2788 | int r; | |
2789 | ||
8315ada7 JG |
2790 | if (!newpage) { |
2791 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2792 | continue; |
8315ada7 JG |
2793 | } |
2794 | ||
2795 | if (!page) { | |
c23a0c99 | 2796 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
8315ada7 | 2797 | continue; |
8315ada7 | 2798 | if (!notified) { |
8315ada7 | 2799 | notified = true; |
ac46d4f3 | 2800 | |
6b49bf6d AP |
2801 | mmu_notifier_range_init_owner(&range, |
2802 | MMU_NOTIFY_MIGRATE, 0, migrate->vma, | |
2803 | migrate->vma->vm_mm, addr, migrate->end, | |
5e5dda81 | 2804 | migrate->pgmap_owner); |
ac46d4f3 | 2805 | mmu_notifier_invalidate_range_start(&range); |
8315ada7 JG |
2806 | } |
2807 | migrate_vma_insert_page(migrate, addr, newpage, | |
d85c6db4 | 2808 | &migrate->src[i]); |
8763cb45 | 2809 | continue; |
8315ada7 | 2810 | } |
8763cb45 JG |
2811 | |
2812 | mapping = page_mapping(page); | |
2813 | ||
a5430dda JG |
2814 | if (is_zone_device_page(newpage)) { |
2815 | if (is_device_private_page(newpage)) { | |
2816 | /* | |
2817 | * For now only support private anonymous when | |
2818 | * migrating to un-addressable device memory. | |
2819 | */ | |
2820 | if (mapping) { | |
2821 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2822 | continue; | |
2823 | } | |
25b2995a | 2824 | } else { |
a5430dda JG |
2825 | /* |
2826 | * Other types of ZONE_DEVICE page are not | |
2827 | * supported. | |
2828 | */ | |
2829 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2830 | continue; | |
2831 | } | |
2832 | } | |
2833 | ||
8763cb45 JG |
2834 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2835 | if (r != MIGRATEPAGE_SUCCESS) | |
2836 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2837 | } | |
8315ada7 | 2838 | |
4645b9fe JG |
2839 | /* |
2840 | * No need to double call mmu_notifier->invalidate_range() callback as | |
2841 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
2842 | * did already call it. | |
2843 | */ | |
8315ada7 | 2844 | if (notified) |
ac46d4f3 | 2845 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 | 2846 | } |
a7d1f22b | 2847 | EXPORT_SYMBOL(migrate_vma_pages); |
8763cb45 | 2848 | |
a7d1f22b | 2849 | /** |
8763cb45 JG |
2850 | * migrate_vma_finalize() - restore CPU page table entry |
2851 | * @migrate: migrate struct containing all migration information | |
2852 | * | |
2853 | * This replaces the special migration pte entry with either a mapping to the | |
2854 | * new page if migration was successful for that page, or to the original page | |
2855 | * otherwise. | |
2856 | * | |
2857 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2858 | * refcount, for device pages. | |
2859 | */ | |
a7d1f22b | 2860 | void migrate_vma_finalize(struct migrate_vma *migrate) |
8763cb45 JG |
2861 | { |
2862 | const unsigned long npages = migrate->npages; | |
2863 | unsigned long i; | |
2864 | ||
2865 | for (i = 0; i < npages; i++) { | |
2866 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2867 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2868 | ||
8315ada7 JG |
2869 | if (!page) { |
2870 | if (newpage) { | |
2871 | unlock_page(newpage); | |
2872 | put_page(newpage); | |
2873 | } | |
8763cb45 | 2874 | continue; |
8315ada7 JG |
2875 | } |
2876 | ||
8763cb45 JG |
2877 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2878 | if (newpage) { | |
2879 | unlock_page(newpage); | |
2880 | put_page(newpage); | |
2881 | } | |
2882 | newpage = page; | |
2883 | } | |
2884 | ||
2885 | remove_migration_ptes(page, newpage, false); | |
2886 | unlock_page(page); | |
8763cb45 | 2887 | |
a5430dda JG |
2888 | if (is_zone_device_page(page)) |
2889 | put_page(page); | |
2890 | else | |
2891 | putback_lru_page(page); | |
8763cb45 JG |
2892 | |
2893 | if (newpage != page) { | |
2894 | unlock_page(newpage); | |
a5430dda JG |
2895 | if (is_zone_device_page(newpage)) |
2896 | put_page(newpage); | |
2897 | else | |
2898 | putback_lru_page(newpage); | |
8763cb45 JG |
2899 | } |
2900 | } | |
2901 | } | |
a7d1f22b | 2902 | EXPORT_SYMBOL(migrate_vma_finalize); |
9b2ed9cb | 2903 | #endif /* CONFIG_DEVICE_PRIVATE */ |
79c28a41 | 2904 | |
dcee9bf5 HY |
2905 | /* |
2906 | * node_demotion[] example: | |
2907 | * | |
2908 | * Consider a system with two sockets. Each socket has | |
2909 | * three classes of memory attached: fast, medium and slow. | |
2910 | * Each memory class is placed in its own NUMA node. The | |
2911 | * CPUs are placed in the node with the "fast" memory. The | |
2912 | * 6 NUMA nodes (0-5) might be split among the sockets like | |
2913 | * this: | |
2914 | * | |
2915 | * Socket A: 0, 1, 2 | |
2916 | * Socket B: 3, 4, 5 | |
2917 | * | |
2918 | * When Node 0 fills up, its memory should be migrated to | |
2919 | * Node 1. When Node 1 fills up, it should be migrated to | |
2920 | * Node 2. The migration path start on the nodes with the | |
2921 | * processors (since allocations default to this node) and | |
2922 | * fast memory, progress through medium and end with the | |
2923 | * slow memory: | |
2924 | * | |
2925 | * 0 -> 1 -> 2 -> stop | |
2926 | * 3 -> 4 -> 5 -> stop | |
2927 | * | |
2928 | * This is represented in the node_demotion[] like this: | |
2929 | * | |
2930 | * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1 | |
2931 | * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2 | |
2932 | * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate | |
2933 | * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4 | |
2934 | * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5 | |
2935 | * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate | |
2936 | * | |
2937 | * Moreover some systems may have multiple slow memory nodes. | |
2938 | * Suppose a system has one socket with 3 memory nodes, node 0 | |
2939 | * is fast memory type, and node 1/2 both are slow memory | |
2940 | * type, and the distance between fast memory node and slow | |
2941 | * memory node is same. So the migration path should be: | |
2942 | * | |
2943 | * 0 -> 1/2 -> stop | |
2944 | * | |
2945 | * This is represented in the node_demotion[] like this: | |
2946 | * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2 | |
2947 | * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate | |
2948 | * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate | |
2949 | */ | |
2950 | ||
2951 | /* | |
2952 | * Writes to this array occur without locking. Cycles are | |
2953 | * not allowed: Node X demotes to Y which demotes to X... | |
2954 | * | |
2955 | * If multiple reads are performed, a single rcu_read_lock() | |
2956 | * must be held over all reads to ensure that no cycles are | |
2957 | * observed. | |
2958 | */ | |
2959 | #define DEFAULT_DEMOTION_TARGET_NODES 15 | |
2960 | ||
2961 | #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES | |
2962 | #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1) | |
2963 | #else | |
2964 | #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES | |
2965 | #endif | |
2966 | ||
2967 | struct demotion_nodes { | |
2968 | unsigned short nr; | |
2969 | short nodes[DEMOTION_TARGET_NODES]; | |
2970 | }; | |
2971 | ||
2972 | static struct demotion_nodes *node_demotion __read_mostly; | |
2973 | ||
2974 | /** | |
2975 | * next_demotion_node() - Get the next node in the demotion path | |
2976 | * @node: The starting node to lookup the next node | |
2977 | * | |
2978 | * Return: node id for next memory node in the demotion path hierarchy | |
2979 | * from @node; NUMA_NO_NODE if @node is terminal. This does not keep | |
2980 | * @node online or guarantee that it *continues* to be the next demotion | |
2981 | * target. | |
2982 | */ | |
2983 | int next_demotion_node(int node) | |
2984 | { | |
2985 | struct demotion_nodes *nd; | |
2986 | unsigned short target_nr, index; | |
2987 | int target; | |
2988 | ||
2989 | if (!node_demotion) | |
2990 | return NUMA_NO_NODE; | |
2991 | ||
2992 | nd = &node_demotion[node]; | |
2993 | ||
2994 | /* | |
2995 | * node_demotion[] is updated without excluding this | |
2996 | * function from running. RCU doesn't provide any | |
2997 | * compiler barriers, so the READ_ONCE() is required | |
2998 | * to avoid compiler reordering or read merging. | |
2999 | * | |
3000 | * Make sure to use RCU over entire code blocks if | |
3001 | * node_demotion[] reads need to be consistent. | |
3002 | */ | |
3003 | rcu_read_lock(); | |
3004 | target_nr = READ_ONCE(nd->nr); | |
3005 | ||
3006 | switch (target_nr) { | |
3007 | case 0: | |
3008 | target = NUMA_NO_NODE; | |
3009 | goto out; | |
3010 | case 1: | |
3011 | index = 0; | |
3012 | break; | |
3013 | default: | |
3014 | /* | |
3015 | * If there are multiple target nodes, just select one | |
3016 | * target node randomly. | |
3017 | * | |
3018 | * In addition, we can also use round-robin to select | |
3019 | * target node, but we should introduce another variable | |
3020 | * for node_demotion[] to record last selected target node, | |
3021 | * that may cause cache ping-pong due to the changing of | |
3022 | * last target node. Or introducing per-cpu data to avoid | |
3023 | * caching issue, which seems more complicated. So selecting | |
3024 | * target node randomly seems better until now. | |
3025 | */ | |
3026 | index = get_random_int() % target_nr; | |
3027 | break; | |
3028 | } | |
3029 | ||
3030 | target = READ_ONCE(nd->nodes[index]); | |
3031 | ||
3032 | out: | |
3033 | rcu_read_unlock(); | |
3034 | return target; | |
3035 | } | |
3036 | ||
76af6a05 | 3037 | #if defined(CONFIG_HOTPLUG_CPU) |
79c28a41 DH |
3038 | /* Disable reclaim-based migration. */ |
3039 | static void __disable_all_migrate_targets(void) | |
3040 | { | |
ac16ec83 | 3041 | int node, i; |
79c28a41 | 3042 | |
ac16ec83 BW |
3043 | if (!node_demotion) |
3044 | return; | |
79c28a41 | 3045 | |
ac16ec83 BW |
3046 | for_each_online_node(node) { |
3047 | node_demotion[node].nr = 0; | |
3048 | for (i = 0; i < DEMOTION_TARGET_NODES; i++) | |
3049 | node_demotion[node].nodes[i] = NUMA_NO_NODE; | |
3050 | } | |
79c28a41 DH |
3051 | } |
3052 | ||
3053 | static void disable_all_migrate_targets(void) | |
3054 | { | |
3055 | __disable_all_migrate_targets(); | |
3056 | ||
3057 | /* | |
3058 | * Ensure that the "disable" is visible across the system. | |
3059 | * Readers will see either a combination of before+disable | |
3060 | * state or disable+after. They will never see before and | |
3061 | * after state together. | |
3062 | * | |
3063 | * The before+after state together might have cycles and | |
3064 | * could cause readers to do things like loop until this | |
3065 | * function finishes. This ensures they can only see a | |
3066 | * single "bad" read and would, for instance, only loop | |
3067 | * once. | |
3068 | */ | |
3069 | synchronize_rcu(); | |
3070 | } | |
3071 | ||
3072 | /* | |
3073 | * Find an automatic demotion target for 'node'. | |
3074 | * Failing here is OK. It might just indicate | |
3075 | * being at the end of a chain. | |
3076 | */ | |
ac16ec83 BW |
3077 | static int establish_migrate_target(int node, nodemask_t *used, |
3078 | int best_distance) | |
79c28a41 | 3079 | { |
ac16ec83 BW |
3080 | int migration_target, index, val; |
3081 | struct demotion_nodes *nd; | |
79c28a41 | 3082 | |
ac16ec83 | 3083 | if (!node_demotion) |
79c28a41 DH |
3084 | return NUMA_NO_NODE; |
3085 | ||
ac16ec83 BW |
3086 | nd = &node_demotion[node]; |
3087 | ||
79c28a41 DH |
3088 | migration_target = find_next_best_node(node, used); |
3089 | if (migration_target == NUMA_NO_NODE) | |
3090 | return NUMA_NO_NODE; | |
3091 | ||
ac16ec83 BW |
3092 | /* |
3093 | * If the node has been set a migration target node before, | |
3094 | * which means it's the best distance between them. Still | |
3095 | * check if this node can be demoted to other target nodes | |
3096 | * if they have a same best distance. | |
3097 | */ | |
3098 | if (best_distance != -1) { | |
3099 | val = node_distance(node, migration_target); | |
3100 | if (val > best_distance) | |
fc89213a | 3101 | goto out_clear; |
ac16ec83 BW |
3102 | } |
3103 | ||
3104 | index = nd->nr; | |
3105 | if (WARN_ONCE(index >= DEMOTION_TARGET_NODES, | |
3106 | "Exceeds maximum demotion target nodes\n")) | |
fc89213a | 3107 | goto out_clear; |
ac16ec83 BW |
3108 | |
3109 | nd->nodes[index] = migration_target; | |
3110 | nd->nr++; | |
79c28a41 DH |
3111 | |
3112 | return migration_target; | |
fc89213a HY |
3113 | out_clear: |
3114 | node_clear(migration_target, *used); | |
3115 | return NUMA_NO_NODE; | |
79c28a41 DH |
3116 | } |
3117 | ||
3118 | /* | |
3119 | * When memory fills up on a node, memory contents can be | |
3120 | * automatically migrated to another node instead of | |
3121 | * discarded at reclaim. | |
3122 | * | |
3123 | * Establish a "migration path" which will start at nodes | |
3124 | * with CPUs and will follow the priorities used to build the | |
3125 | * page allocator zonelists. | |
3126 | * | |
3127 | * The difference here is that cycles must be avoided. If | |
3128 | * node0 migrates to node1, then neither node1, nor anything | |
ac16ec83 BW |
3129 | * node1 migrates to can migrate to node0. Also one node can |
3130 | * be migrated to multiple nodes if the target nodes all have | |
3131 | * a same best-distance against the source node. | |
79c28a41 DH |
3132 | * |
3133 | * This function can run simultaneously with readers of | |
3134 | * node_demotion[]. However, it can not run simultaneously | |
3135 | * with itself. Exclusion is provided by memory hotplug events | |
3136 | * being single-threaded. | |
3137 | */ | |
3138 | static void __set_migration_target_nodes(void) | |
3139 | { | |
3140 | nodemask_t next_pass = NODE_MASK_NONE; | |
3141 | nodemask_t this_pass = NODE_MASK_NONE; | |
3142 | nodemask_t used_targets = NODE_MASK_NONE; | |
ac16ec83 | 3143 | int node, best_distance; |
79c28a41 DH |
3144 | |
3145 | /* | |
3146 | * Avoid any oddities like cycles that could occur | |
3147 | * from changes in the topology. This will leave | |
3148 | * a momentary gap when migration is disabled. | |
3149 | */ | |
3150 | disable_all_migrate_targets(); | |
3151 | ||
3152 | /* | |
3153 | * Allocations go close to CPUs, first. Assume that | |
3154 | * the migration path starts at the nodes with CPUs. | |
3155 | */ | |
3156 | next_pass = node_states[N_CPU]; | |
3157 | again: | |
3158 | this_pass = next_pass; | |
3159 | next_pass = NODE_MASK_NONE; | |
3160 | /* | |
3161 | * To avoid cycles in the migration "graph", ensure | |
3162 | * that migration sources are not future targets by | |
3163 | * setting them in 'used_targets'. Do this only | |
3164 | * once per pass so that multiple source nodes can | |
3165 | * share a target node. | |
3166 | * | |
3167 | * 'used_targets' will become unavailable in future | |
3168 | * passes. This limits some opportunities for | |
3169 | * multiple source nodes to share a destination. | |
3170 | */ | |
3171 | nodes_or(used_targets, used_targets, this_pass); | |
79c28a41 | 3172 | |
ac16ec83 BW |
3173 | for_each_node_mask(node, this_pass) { |
3174 | best_distance = -1; | |
79c28a41 DH |
3175 | |
3176 | /* | |
ac16ec83 BW |
3177 | * Try to set up the migration path for the node, and the target |
3178 | * migration nodes can be multiple, so doing a loop to find all | |
3179 | * the target nodes if they all have a best node distance. | |
79c28a41 | 3180 | */ |
ac16ec83 BW |
3181 | do { |
3182 | int target_node = | |
3183 | establish_migrate_target(node, &used_targets, | |
3184 | best_distance); | |
3185 | ||
3186 | if (target_node == NUMA_NO_NODE) | |
3187 | break; | |
3188 | ||
3189 | if (best_distance == -1) | |
3190 | best_distance = node_distance(node, target_node); | |
3191 | ||
3192 | /* | |
3193 | * Visit targets from this pass in the next pass. | |
3194 | * Eventually, every node will have been part of | |
3195 | * a pass, and will become set in 'used_targets'. | |
3196 | */ | |
3197 | node_set(target_node, next_pass); | |
3198 | } while (1); | |
79c28a41 DH |
3199 | } |
3200 | /* | |
3201 | * 'next_pass' contains nodes which became migration | |
3202 | * targets in this pass. Make additional passes until | |
3203 | * no more migrations targets are available. | |
3204 | */ | |
3205 | if (!nodes_empty(next_pass)) | |
3206 | goto again; | |
3207 | } | |
3208 | ||
3209 | /* | |
3210 | * For callers that do not hold get_online_mems() already. | |
3211 | */ | |
79c28a41 DH |
3212 | static void set_migration_target_nodes(void) |
3213 | { | |
3214 | get_online_mems(); | |
3215 | __set_migration_target_nodes(); | |
3216 | put_online_mems(); | |
3217 | } | |
884a6e5d | 3218 | |
884a6e5d DH |
3219 | /* |
3220 | * This leaves migrate-on-reclaim transiently disabled between | |
3221 | * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs | |
3222 | * whether reclaim-based migration is enabled or not, which | |
3223 | * ensures that the user can turn reclaim-based migration at | |
3224 | * any time without needing to recalculate migration targets. | |
3225 | * | |
3226 | * These callbacks already hold get_online_mems(). That is why | |
3227 | * __set_migration_target_nodes() can be used as opposed to | |
3228 | * set_migration_target_nodes(). | |
3229 | */ | |
3230 | static int __meminit migrate_on_reclaim_callback(struct notifier_block *self, | |
295be91f | 3231 | unsigned long action, void *_arg) |
884a6e5d | 3232 | { |
295be91f DH |
3233 | struct memory_notify *arg = _arg; |
3234 | ||
3235 | /* | |
3236 | * Only update the node migration order when a node is | |
3237 | * changing status, like online->offline. This avoids | |
3238 | * the overhead of synchronize_rcu() in most cases. | |
3239 | */ | |
3240 | if (arg->status_change_nid < 0) | |
3241 | return notifier_from_errno(0); | |
3242 | ||
884a6e5d DH |
3243 | switch (action) { |
3244 | case MEM_GOING_OFFLINE: | |
3245 | /* | |
3246 | * Make sure there are not transient states where | |
3247 | * an offline node is a migration target. This | |
3248 | * will leave migration disabled until the offline | |
3249 | * completes and the MEM_OFFLINE case below runs. | |
3250 | */ | |
3251 | disable_all_migrate_targets(); | |
3252 | break; | |
3253 | case MEM_OFFLINE: | |
3254 | case MEM_ONLINE: | |
3255 | /* | |
3256 | * Recalculate the target nodes once the node | |
3257 | * reaches its final state (online or offline). | |
3258 | */ | |
3259 | __set_migration_target_nodes(); | |
3260 | break; | |
3261 | case MEM_CANCEL_OFFLINE: | |
3262 | /* | |
3263 | * MEM_GOING_OFFLINE disabled all the migration | |
3264 | * targets. Reenable them. | |
3265 | */ | |
3266 | __set_migration_target_nodes(); | |
3267 | break; | |
3268 | case MEM_GOING_ONLINE: | |
3269 | case MEM_CANCEL_ONLINE: | |
3270 | break; | |
3271 | } | |
3272 | ||
3273 | return notifier_from_errno(0); | |
3274 | } | |
3275 | ||
76af6a05 DH |
3276 | /* |
3277 | * React to hotplug events that might affect the migration targets | |
3278 | * like events that online or offline NUMA nodes. | |
3279 | * | |
3280 | * The ordering is also currently dependent on which nodes have | |
3281 | * CPUs. That means we need CPU on/offline notification too. | |
3282 | */ | |
3283 | static int migration_online_cpu(unsigned int cpu) | |
3284 | { | |
3285 | set_migration_target_nodes(); | |
3286 | return 0; | |
3287 | } | |
3288 | ||
3289 | static int migration_offline_cpu(unsigned int cpu) | |
3290 | { | |
3291 | set_migration_target_nodes(); | |
3292 | return 0; | |
3293 | } | |
3294 | ||
884a6e5d DH |
3295 | static int __init migrate_on_reclaim_init(void) |
3296 | { | |
3297 | int ret; | |
3298 | ||
ac16ec83 BW |
3299 | node_demotion = kmalloc_array(nr_node_ids, |
3300 | sizeof(struct demotion_nodes), | |
3301 | GFP_KERNEL); | |
3302 | WARN_ON(!node_demotion); | |
3303 | ||
a6a0251c HY |
3304 | ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline", |
3305 | NULL, migration_offline_cpu); | |
884a6e5d DH |
3306 | /* |
3307 | * In the unlikely case that this fails, the automatic | |
3308 | * migration targets may become suboptimal for nodes | |
3309 | * where N_CPU changes. With such a small impact in a | |
3310 | * rare case, do not bother trying to do anything special. | |
3311 | */ | |
3312 | WARN_ON(ret < 0); | |
a6a0251c HY |
3313 | ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online", |
3314 | migration_online_cpu, NULL); | |
3315 | WARN_ON(ret < 0); | |
884a6e5d DH |
3316 | |
3317 | hotplug_memory_notifier(migrate_on_reclaim_callback, 100); | |
3318 | return 0; | |
3319 | } | |
3320 | late_initcall(migrate_on_reclaim_init); | |
76af6a05 | 3321 | #endif /* CONFIG_HOTPLUG_CPU */ |
20f9ba4f YS |
3322 | |
3323 | bool numa_demotion_enabled = false; | |
3324 | ||
3325 | #ifdef CONFIG_SYSFS | |
3326 | static ssize_t numa_demotion_enabled_show(struct kobject *kobj, | |
3327 | struct kobj_attribute *attr, char *buf) | |
3328 | { | |
3329 | return sysfs_emit(buf, "%s\n", | |
3330 | numa_demotion_enabled ? "true" : "false"); | |
3331 | } | |
3332 | ||
3333 | static ssize_t numa_demotion_enabled_store(struct kobject *kobj, | |
3334 | struct kobj_attribute *attr, | |
3335 | const char *buf, size_t count) | |
3336 | { | |
3337 | if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1)) | |
3338 | numa_demotion_enabled = true; | |
3339 | else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1)) | |
3340 | numa_demotion_enabled = false; | |
3341 | else | |
3342 | return -EINVAL; | |
3343 | ||
3344 | return count; | |
3345 | } | |
3346 | ||
3347 | static struct kobj_attribute numa_demotion_enabled_attr = | |
3348 | __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show, | |
3349 | numa_demotion_enabled_store); | |
3350 | ||
3351 | static struct attribute *numa_attrs[] = { | |
3352 | &numa_demotion_enabled_attr.attr, | |
3353 | NULL, | |
3354 | }; | |
3355 | ||
3356 | static const struct attribute_group numa_attr_group = { | |
3357 | .attrs = numa_attrs, | |
3358 | }; | |
3359 | ||
3360 | static int __init numa_init_sysfs(void) | |
3361 | { | |
3362 | int err; | |
3363 | struct kobject *numa_kobj; | |
3364 | ||
3365 | numa_kobj = kobject_create_and_add("numa", mm_kobj); | |
3366 | if (!numa_kobj) { | |
3367 | pr_err("failed to create numa kobject\n"); | |
3368 | return -ENOMEM; | |
3369 | } | |
3370 | err = sysfs_create_group(numa_kobj, &numa_attr_group); | |
3371 | if (err) { | |
3372 | pr_err("failed to register numa group\n"); | |
3373 | goto delete_obj; | |
3374 | } | |
3375 | return 0; | |
3376 | ||
3377 | delete_obj: | |
3378 | kobject_put(numa_kobj); | |
3379 | return err; | |
3380 | } | |
3381 | subsys_initcall(numa_init_sysfs); | |
3382 | #endif |