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> |
b20a3503 | 53 | |
0d1836c3 MN |
54 | #include <asm/tlbflush.h> |
55 | ||
7b2a2d4a MG |
56 | #define CREATE_TRACE_POINTS |
57 | #include <trace/events/migrate.h> | |
58 | ||
b20a3503 CL |
59 | #include "internal.h" |
60 | ||
9e5bcd61 | 61 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
62 | { |
63 | struct address_space *mapping; | |
64 | ||
65 | /* | |
66 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
67 | * or just got freed under us. | |
68 | * | |
69 | * In case we 'win' a race for a movable page being freed under us and | |
70 | * raise its refcount preventing __free_pages() from doing its job | |
71 | * the put_page() at the end of this block will take care of | |
72 | * release this page, thus avoiding a nasty leakage. | |
73 | */ | |
74 | if (unlikely(!get_page_unless_zero(page))) | |
75 | goto out; | |
76 | ||
77 | /* | |
78 | * Check PageMovable before holding a PG_lock because page's owner | |
79 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 80 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
81 | */ |
82 | if (unlikely(!__PageMovable(page))) | |
83 | goto out_putpage; | |
84 | /* | |
85 | * As movable pages are not isolated from LRU lists, concurrent | |
86 | * compaction threads can race against page migration functions | |
87 | * as well as race against the releasing a page. | |
88 | * | |
89 | * In order to avoid having an already isolated movable page | |
90 | * being (wrongly) re-isolated while it is under migration, | |
91 | * or to avoid attempting to isolate pages being released, | |
92 | * lets be sure we have the page lock | |
93 | * before proceeding with the movable page isolation steps. | |
94 | */ | |
95 | if (unlikely(!trylock_page(page))) | |
96 | goto out_putpage; | |
97 | ||
98 | if (!PageMovable(page) || PageIsolated(page)) | |
99 | goto out_no_isolated; | |
100 | ||
101 | mapping = page_mapping(page); | |
102 | VM_BUG_ON_PAGE(!mapping, page); | |
103 | ||
104 | if (!mapping->a_ops->isolate_page(page, mode)) | |
105 | goto out_no_isolated; | |
106 | ||
107 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
108 | WARN_ON_ONCE(PageIsolated(page)); | |
109 | __SetPageIsolated(page); | |
110 | unlock_page(page); | |
111 | ||
9e5bcd61 | 112 | return 0; |
bda807d4 MK |
113 | |
114 | out_no_isolated: | |
115 | unlock_page(page); | |
116 | out_putpage: | |
117 | put_page(page); | |
118 | out: | |
9e5bcd61 | 119 | return -EBUSY; |
bda807d4 MK |
120 | } |
121 | ||
606a6f71 | 122 | static void putback_movable_page(struct page *page) |
bda807d4 MK |
123 | { |
124 | struct address_space *mapping; | |
125 | ||
bda807d4 MK |
126 | mapping = page_mapping(page); |
127 | mapping->a_ops->putback_page(page); | |
128 | __ClearPageIsolated(page); | |
129 | } | |
130 | ||
5733c7d1 RA |
131 | /* |
132 | * Put previously isolated pages back onto the appropriate lists | |
133 | * from where they were once taken off for compaction/migration. | |
134 | * | |
59c82b70 JK |
135 | * This function shall be used whenever the isolated pageset has been |
136 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
137 | * and isolate_huge_page(). | |
5733c7d1 RA |
138 | */ |
139 | void putback_movable_pages(struct list_head *l) | |
140 | { | |
141 | struct page *page; | |
142 | struct page *page2; | |
143 | ||
b20a3503 | 144 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
145 | if (unlikely(PageHuge(page))) { |
146 | putback_active_hugepage(page); | |
147 | continue; | |
148 | } | |
e24f0b8f | 149 | list_del(&page->lru); |
bda807d4 MK |
150 | /* |
151 | * We isolated non-lru movable page so here we can use | |
152 | * __PageMovable because LRU page's mapping cannot have | |
153 | * PAGE_MAPPING_MOVABLE. | |
154 | */ | |
b1123ea6 | 155 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
156 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
157 | lock_page(page); | |
158 | if (PageMovable(page)) | |
159 | putback_movable_page(page); | |
160 | else | |
161 | __ClearPageIsolated(page); | |
162 | unlock_page(page); | |
163 | put_page(page); | |
164 | } else { | |
e8db67eb | 165 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 166 | page_is_file_lru(page), -thp_nr_pages(page)); |
fc280fe8 | 167 | putback_lru_page(page); |
bda807d4 | 168 | } |
b20a3503 | 169 | } |
b20a3503 CL |
170 | } |
171 | ||
0697212a CL |
172 | /* |
173 | * Restore a potential migration pte to a working pte entry | |
174 | */ | |
e4b82222 | 175 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 176 | unsigned long addr, void *old) |
0697212a | 177 | { |
3fe87967 KS |
178 | struct page_vma_mapped_walk pvmw = { |
179 | .page = old, | |
180 | .vma = vma, | |
181 | .address = addr, | |
182 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
183 | }; | |
184 | struct page *new; | |
185 | pte_t pte; | |
0697212a | 186 | swp_entry_t entry; |
0697212a | 187 | |
3fe87967 KS |
188 | VM_BUG_ON_PAGE(PageTail(page), page); |
189 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
190 | if (PageKsm(page)) |
191 | new = page; | |
192 | else | |
193 | new = page - pvmw.page->index + | |
194 | linear_page_index(vma, pvmw.address); | |
0697212a | 195 | |
616b8371 ZY |
196 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
197 | /* PMD-mapped THP migration entry */ | |
198 | if (!pvmw.pte) { | |
199 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
200 | remove_migration_pmd(&pvmw, new); | |
201 | continue; | |
202 | } | |
203 | #endif | |
204 | ||
3fe87967 KS |
205 | get_page(new); |
206 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
207 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
208 | pte = pte_mksoft_dirty(pte); | |
0697212a | 209 | |
3fe87967 KS |
210 | /* |
211 | * Recheck VMA as permissions can change since migration started | |
212 | */ | |
213 | entry = pte_to_swp_entry(*pvmw.pte); | |
4dd845b5 | 214 | if (is_writable_migration_entry(entry)) |
3fe87967 | 215 | pte = maybe_mkwrite(pte, vma); |
f45ec5ff PX |
216 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
217 | pte = pte_mkuffd_wp(pte); | |
d3cb8bf6 | 218 | |
6128763f | 219 | if (unlikely(is_device_private_page(new))) { |
4dd845b5 AP |
220 | if (pte_write(pte)) |
221 | entry = make_writable_device_private_entry( | |
222 | page_to_pfn(new)); | |
223 | else | |
224 | entry = make_readable_device_private_entry( | |
225 | page_to_pfn(new)); | |
6128763f | 226 | pte = swp_entry_to_pte(entry); |
3d321bf8 RC |
227 | if (pte_swp_soft_dirty(*pvmw.pte)) |
228 | pte = pte_swp_mksoft_dirty(pte); | |
6128763f RC |
229 | if (pte_swp_uffd_wp(*pvmw.pte)) |
230 | pte = pte_swp_mkuffd_wp(pte); | |
d2b2c6dd | 231 | } |
a5430dda | 232 | |
3ef8fd7f | 233 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 | 234 | if (PageHuge(new)) { |
79c1c594 CL |
235 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
236 | ||
3fe87967 | 237 | pte = pte_mkhuge(pte); |
79c1c594 | 238 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
383321ab | 239 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
240 | if (PageAnon(new)) |
241 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
242 | else | |
243 | page_dup_rmap(new, true); | |
383321ab AK |
244 | } else |
245 | #endif | |
246 | { | |
247 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 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); | |
253 | } | |
3fe87967 KS |
254 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
255 | mlock_vma_page(new); | |
256 | ||
e125fe40 KS |
257 | if (PageTransHuge(page) && PageMlocked(page)) |
258 | clear_page_mlock(page); | |
259 | ||
3fe87967 KS |
260 | /* No need to invalidate - it was non-present before */ |
261 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
262 | } | |
51afb12b | 263 | |
e4b82222 | 264 | return true; |
0697212a CL |
265 | } |
266 | ||
04e62a29 CL |
267 | /* |
268 | * Get rid of all migration entries and replace them by | |
269 | * references to the indicated page. | |
270 | */ | |
e388466d | 271 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 272 | { |
051ac83a JK |
273 | struct rmap_walk_control rwc = { |
274 | .rmap_one = remove_migration_pte, | |
275 | .arg = old, | |
276 | }; | |
277 | ||
e388466d KS |
278 | if (locked) |
279 | rmap_walk_locked(new, &rwc); | |
280 | else | |
281 | rmap_walk(new, &rwc); | |
04e62a29 CL |
282 | } |
283 | ||
0697212a CL |
284 | /* |
285 | * Something used the pte of a page under migration. We need to | |
286 | * get to the page and wait until migration is finished. | |
287 | * When we return from this function the fault will be retried. | |
0697212a | 288 | */ |
e66f17ff | 289 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 290 | spinlock_t *ptl) |
0697212a | 291 | { |
30dad309 | 292 | pte_t pte; |
0697212a CL |
293 | swp_entry_t entry; |
294 | struct page *page; | |
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 | ||
af5cdaf8 | 305 | page = pfn_swap_entry_to_page(entry); |
ffc90cbb | 306 | page = compound_head(page); |
0697212a | 307 | |
e286781d | 308 | /* |
89eb946a | 309 | * Once page cache replacement of page migration started, page_count |
9a1ea439 HD |
310 | * is zero; but we must not call put_and_wait_on_page_locked() without |
311 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
e286781d NP |
312 | */ |
313 | if (!get_page_unless_zero(page)) | |
314 | goto out; | |
0697212a | 315 | pte_unmap_unlock(ptep, ptl); |
48054625 | 316 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
0697212a CL |
317 | return; |
318 | out: | |
319 | pte_unmap_unlock(ptep, ptl); | |
320 | } | |
321 | ||
30dad309 NH |
322 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
323 | unsigned long address) | |
324 | { | |
325 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
326 | pte_t *ptep = pte_offset_map(pmd, address); | |
327 | __migration_entry_wait(mm, ptep, ptl); | |
328 | } | |
329 | ||
cb900f41 KS |
330 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
331 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 332 | { |
cb900f41 | 333 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
334 | __migration_entry_wait(mm, pte, ptl); |
335 | } | |
336 | ||
616b8371 ZY |
337 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
338 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
339 | { | |
340 | spinlock_t *ptl; | |
341 | struct page *page; | |
342 | ||
343 | ptl = pmd_lock(mm, pmd); | |
344 | if (!is_pmd_migration_entry(*pmd)) | |
345 | goto unlock; | |
af5cdaf8 | 346 | page = pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd)); |
616b8371 ZY |
347 | if (!get_page_unless_zero(page)) |
348 | goto unlock; | |
349 | spin_unlock(ptl); | |
48054625 | 350 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
616b8371 ZY |
351 | return; |
352 | unlock: | |
353 | spin_unlock(ptl); | |
354 | } | |
355 | #endif | |
356 | ||
f900482d | 357 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
0b3901b3 JK |
358 | { |
359 | int expected_count = 1; | |
360 | ||
361 | /* | |
f1f4f3ab | 362 | * Device private pages have an extra refcount as they are |
0b3901b3 JK |
363 | * ZONE_DEVICE pages. |
364 | */ | |
365 | expected_count += is_device_private_page(page); | |
f900482d | 366 | if (mapping) |
6c357848 | 367 | expected_count += thp_nr_pages(page) + page_has_private(page); |
0b3901b3 JK |
368 | |
369 | return expected_count; | |
370 | } | |
371 | ||
b20a3503 | 372 | /* |
c3fcf8a5 | 373 | * Replace the page in the mapping. |
5b5c7120 CL |
374 | * |
375 | * The number of remaining references must be: | |
376 | * 1 for anonymous pages without a mapping | |
377 | * 2 for pages with a mapping | |
266cf658 | 378 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 379 | */ |
36bc08cc | 380 | int migrate_page_move_mapping(struct address_space *mapping, |
37109694 | 381 | struct page *newpage, struct page *page, int extra_count) |
b20a3503 | 382 | { |
89eb946a | 383 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
42cb14b1 HD |
384 | struct zone *oldzone, *newzone; |
385 | int dirty; | |
f900482d | 386 | int expected_count = expected_page_refs(mapping, page) + extra_count; |
5c447d27 | 387 | int nr = thp_nr_pages(page); |
8763cb45 | 388 | |
6c5240ae | 389 | if (!mapping) { |
0e8c7d0f | 390 | /* Anonymous page without mapping */ |
8e321fef | 391 | if (page_count(page) != expected_count) |
6c5240ae | 392 | return -EAGAIN; |
cf4b769a HD |
393 | |
394 | /* No turning back from here */ | |
cf4b769a HD |
395 | newpage->index = page->index; |
396 | newpage->mapping = page->mapping; | |
397 | if (PageSwapBacked(page)) | |
fa9949da | 398 | __SetPageSwapBacked(newpage); |
cf4b769a | 399 | |
78bd5209 | 400 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
401 | } |
402 | ||
42cb14b1 HD |
403 | oldzone = page_zone(page); |
404 | newzone = page_zone(newpage); | |
405 | ||
89eb946a | 406 | xas_lock_irq(&xas); |
89eb946a MW |
407 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
408 | xas_unlock_irq(&xas); | |
e23ca00b | 409 | return -EAGAIN; |
b20a3503 CL |
410 | } |
411 | ||
fe896d18 | 412 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 413 | xas_unlock_irq(&xas); |
e286781d NP |
414 | return -EAGAIN; |
415 | } | |
416 | ||
b20a3503 | 417 | /* |
cf4b769a HD |
418 | * Now we know that no one else is looking at the page: |
419 | * no turning back from here. | |
b20a3503 | 420 | */ |
cf4b769a HD |
421 | newpage->index = page->index; |
422 | newpage->mapping = page->mapping; | |
5c447d27 | 423 | page_ref_add(newpage, nr); /* add cache reference */ |
6326fec1 NP |
424 | if (PageSwapBacked(page)) { |
425 | __SetPageSwapBacked(newpage); | |
426 | if (PageSwapCache(page)) { | |
427 | SetPageSwapCache(newpage); | |
428 | set_page_private(newpage, page_private(page)); | |
429 | } | |
430 | } else { | |
431 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
432 | } |
433 | ||
42cb14b1 HD |
434 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
435 | dirty = PageDirty(page); | |
436 | if (dirty) { | |
437 | ClearPageDirty(page); | |
438 | SetPageDirty(newpage); | |
439 | } | |
440 | ||
89eb946a | 441 | xas_store(&xas, newpage); |
e71769ae NH |
442 | if (PageTransHuge(page)) { |
443 | int i; | |
e71769ae | 444 | |
5c447d27 | 445 | for (i = 1; i < nr; i++) { |
89eb946a | 446 | xas_next(&xas); |
4101196b | 447 | xas_store(&xas, newpage); |
e71769ae | 448 | } |
e71769ae | 449 | } |
7cf9c2c7 NP |
450 | |
451 | /* | |
937a94c9 JG |
452 | * Drop cache reference from old page by unfreezing |
453 | * to one less reference. | |
7cf9c2c7 NP |
454 | * We know this isn't the last reference. |
455 | */ | |
5c447d27 | 456 | page_ref_unfreeze(page, expected_count - nr); |
7cf9c2c7 | 457 | |
89eb946a | 458 | xas_unlock(&xas); |
42cb14b1 HD |
459 | /* Leave irq disabled to prevent preemption while updating stats */ |
460 | ||
0e8c7d0f CL |
461 | /* |
462 | * If moved to a different zone then also account | |
463 | * the page for that zone. Other VM counters will be | |
464 | * taken care of when we establish references to the | |
465 | * new page and drop references to the old page. | |
466 | * | |
467 | * Note that anonymous pages are accounted for | |
4b9d0fab | 468 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
469 | * are mapped to swap space. |
470 | */ | |
42cb14b1 | 471 | if (newzone != oldzone) { |
0d1c2072 JW |
472 | struct lruvec *old_lruvec, *new_lruvec; |
473 | struct mem_cgroup *memcg; | |
474 | ||
475 | memcg = page_memcg(page); | |
476 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); | |
477 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
478 | ||
5c447d27 SB |
479 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
480 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); | |
42cb14b1 | 481 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
5c447d27 SB |
482 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
483 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); | |
42cb14b1 | 484 | } |
b6038942 SB |
485 | #ifdef CONFIG_SWAP |
486 | if (PageSwapCache(page)) { | |
487 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); | |
488 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); | |
489 | } | |
490 | #endif | |
f56753ac | 491 | if (dirty && mapping_can_writeback(mapping)) { |
5c447d27 SB |
492 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
493 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); | |
494 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); | |
495 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); | |
42cb14b1 | 496 | } |
4b02108a | 497 | } |
42cb14b1 | 498 | local_irq_enable(); |
b20a3503 | 499 | |
78bd5209 | 500 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 501 | } |
1118dce7 | 502 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 503 | |
290408d4 NH |
504 | /* |
505 | * The expected number of remaining references is the same as that | |
506 | * of migrate_page_move_mapping(). | |
507 | */ | |
508 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
509 | struct page *newpage, struct page *page) | |
510 | { | |
89eb946a | 511 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 512 | int expected_count; |
290408d4 | 513 | |
89eb946a | 514 | xas_lock_irq(&xas); |
290408d4 | 515 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
516 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
517 | xas_unlock_irq(&xas); | |
290408d4 NH |
518 | return -EAGAIN; |
519 | } | |
520 | ||
fe896d18 | 521 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 522 | xas_unlock_irq(&xas); |
290408d4 NH |
523 | return -EAGAIN; |
524 | } | |
525 | ||
cf4b769a HD |
526 | newpage->index = page->index; |
527 | newpage->mapping = page->mapping; | |
6a93ca8f | 528 | |
290408d4 NH |
529 | get_page(newpage); |
530 | ||
89eb946a | 531 | xas_store(&xas, newpage); |
290408d4 | 532 | |
fe896d18 | 533 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 534 | |
89eb946a | 535 | xas_unlock_irq(&xas); |
6a93ca8f | 536 | |
78bd5209 | 537 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
538 | } |
539 | ||
b20a3503 CL |
540 | /* |
541 | * Copy the page to its new location | |
542 | */ | |
2916ecc0 | 543 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 544 | { |
7851a45c RR |
545 | int cpupid; |
546 | ||
b20a3503 CL |
547 | if (PageError(page)) |
548 | SetPageError(newpage); | |
549 | if (PageReferenced(page)) | |
550 | SetPageReferenced(newpage); | |
551 | if (PageUptodate(page)) | |
552 | SetPageUptodate(newpage); | |
894bc310 | 553 | if (TestClearPageActive(page)) { |
309381fe | 554 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 555 | SetPageActive(newpage); |
418b27ef LS |
556 | } else if (TestClearPageUnevictable(page)) |
557 | SetPageUnevictable(newpage); | |
1899ad18 JW |
558 | if (PageWorkingset(page)) |
559 | SetPageWorkingset(newpage); | |
b20a3503 CL |
560 | if (PageChecked(page)) |
561 | SetPageChecked(newpage); | |
562 | if (PageMappedToDisk(page)) | |
563 | SetPageMappedToDisk(newpage); | |
564 | ||
42cb14b1 HD |
565 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
566 | if (PageDirty(page)) | |
567 | SetPageDirty(newpage); | |
b20a3503 | 568 | |
33c3fc71 VD |
569 | if (page_is_young(page)) |
570 | set_page_young(newpage); | |
571 | if (page_is_idle(page)) | |
572 | set_page_idle(newpage); | |
573 | ||
7851a45c RR |
574 | /* |
575 | * Copy NUMA information to the new page, to prevent over-eager | |
576 | * future migrations of this same page. | |
577 | */ | |
578 | cpupid = page_cpupid_xchg_last(page, -1); | |
579 | page_cpupid_xchg_last(newpage, cpupid); | |
580 | ||
e9995ef9 | 581 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
582 | /* |
583 | * Please do not reorder this without considering how mm/ksm.c's | |
584 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
585 | */ | |
b3b3a99c NH |
586 | if (PageSwapCache(page)) |
587 | ClearPageSwapCache(page); | |
b20a3503 | 588 | ClearPagePrivate(page); |
ad2fa371 MS |
589 | |
590 | /* page->private contains hugetlb specific flags */ | |
591 | if (!PageHuge(page)) | |
592 | set_page_private(page, 0); | |
b20a3503 CL |
593 | |
594 | /* | |
595 | * If any waiters have accumulated on the new page then | |
596 | * wake them up. | |
597 | */ | |
598 | if (PageWriteback(newpage)) | |
599 | end_page_writeback(newpage); | |
d435edca | 600 | |
6aeff241 YS |
601 | /* |
602 | * PG_readahead shares the same bit with PG_reclaim. The above | |
603 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
604 | * bit after that. | |
605 | */ | |
606 | if (PageReadahead(page)) | |
607 | SetPageReadahead(newpage); | |
608 | ||
d435edca | 609 | copy_page_owner(page, newpage); |
74485cf2 | 610 | |
a333e3e7 HD |
611 | if (!PageHuge(page)) |
612 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 613 | } |
2916ecc0 JG |
614 | EXPORT_SYMBOL(migrate_page_states); |
615 | ||
616 | void migrate_page_copy(struct page *newpage, struct page *page) | |
617 | { | |
618 | if (PageHuge(page) || PageTransHuge(page)) | |
619 | copy_huge_page(newpage, page); | |
620 | else | |
621 | copy_highpage(newpage, page); | |
622 | ||
623 | migrate_page_states(newpage, page); | |
624 | } | |
1118dce7 | 625 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 626 | |
1d8b85cc CL |
627 | /************************************************************ |
628 | * Migration functions | |
629 | ***********************************************************/ | |
630 | ||
b20a3503 | 631 | /* |
bda807d4 | 632 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 633 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
634 | * |
635 | * Pages are locked upon entry and exit. | |
636 | */ | |
2d1db3b1 | 637 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
638 | struct page *newpage, struct page *page, |
639 | enum migrate_mode mode) | |
b20a3503 CL |
640 | { |
641 | int rc; | |
642 | ||
643 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
644 | ||
37109694 | 645 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
b20a3503 | 646 | |
78bd5209 | 647 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
648 | return rc; |
649 | ||
2916ecc0 JG |
650 | if (mode != MIGRATE_SYNC_NO_COPY) |
651 | migrate_page_copy(newpage, page); | |
652 | else | |
653 | migrate_page_states(newpage, page); | |
78bd5209 | 654 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
655 | } |
656 | EXPORT_SYMBOL(migrate_page); | |
657 | ||
9361401e | 658 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
659 | /* Returns true if all buffers are successfully locked */ |
660 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
661 | enum migrate_mode mode) | |
662 | { | |
663 | struct buffer_head *bh = head; | |
664 | ||
665 | /* Simple case, sync compaction */ | |
666 | if (mode != MIGRATE_ASYNC) { | |
667 | do { | |
84ade7c1 JK |
668 | lock_buffer(bh); |
669 | bh = bh->b_this_page; | |
670 | ||
671 | } while (bh != head); | |
672 | ||
673 | return true; | |
674 | } | |
675 | ||
676 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
677 | do { | |
84ade7c1 JK |
678 | if (!trylock_buffer(bh)) { |
679 | /* | |
680 | * We failed to lock the buffer and cannot stall in | |
681 | * async migration. Release the taken locks | |
682 | */ | |
683 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
684 | bh = head; |
685 | while (bh != failed_bh) { | |
686 | unlock_buffer(bh); | |
84ade7c1 JK |
687 | bh = bh->b_this_page; |
688 | } | |
689 | return false; | |
690 | } | |
691 | ||
692 | bh = bh->b_this_page; | |
693 | } while (bh != head); | |
694 | return true; | |
695 | } | |
696 | ||
89cb0888 JK |
697 | static int __buffer_migrate_page(struct address_space *mapping, |
698 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
699 | bool check_refs) | |
1d8b85cc | 700 | { |
1d8b85cc CL |
701 | struct buffer_head *bh, *head; |
702 | int rc; | |
cc4f11e6 | 703 | int expected_count; |
1d8b85cc | 704 | |
1d8b85cc | 705 | if (!page_has_buffers(page)) |
a6bc32b8 | 706 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 707 | |
cc4f11e6 | 708 | /* Check whether page does not have extra refs before we do more work */ |
f900482d | 709 | expected_count = expected_page_refs(mapping, page); |
cc4f11e6 JK |
710 | if (page_count(page) != expected_count) |
711 | return -EAGAIN; | |
1d8b85cc | 712 | |
cc4f11e6 JK |
713 | head = page_buffers(page); |
714 | if (!buffer_migrate_lock_buffers(head, mode)) | |
715 | return -EAGAIN; | |
1d8b85cc | 716 | |
89cb0888 JK |
717 | if (check_refs) { |
718 | bool busy; | |
719 | bool invalidated = false; | |
720 | ||
721 | recheck_buffers: | |
722 | busy = false; | |
723 | spin_lock(&mapping->private_lock); | |
724 | bh = head; | |
725 | do { | |
726 | if (atomic_read(&bh->b_count)) { | |
727 | busy = true; | |
728 | break; | |
729 | } | |
730 | bh = bh->b_this_page; | |
731 | } while (bh != head); | |
89cb0888 JK |
732 | if (busy) { |
733 | if (invalidated) { | |
734 | rc = -EAGAIN; | |
735 | goto unlock_buffers; | |
736 | } | |
ebdf4de5 | 737 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
738 | invalidate_bh_lrus(); |
739 | invalidated = true; | |
740 | goto recheck_buffers; | |
741 | } | |
742 | } | |
743 | ||
37109694 | 744 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
78bd5209 | 745 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 746 | goto unlock_buffers; |
1d8b85cc | 747 | |
cd0f3715 | 748 | attach_page_private(newpage, detach_page_private(page)); |
1d8b85cc CL |
749 | |
750 | bh = head; | |
751 | do { | |
752 | set_bh_page(bh, newpage, bh_offset(bh)); | |
753 | bh = bh->b_this_page; | |
754 | ||
755 | } while (bh != head); | |
756 | ||
2916ecc0 JG |
757 | if (mode != MIGRATE_SYNC_NO_COPY) |
758 | migrate_page_copy(newpage, page); | |
759 | else | |
760 | migrate_page_states(newpage, page); | |
1d8b85cc | 761 | |
cc4f11e6 JK |
762 | rc = MIGRATEPAGE_SUCCESS; |
763 | unlock_buffers: | |
ebdf4de5 JK |
764 | if (check_refs) |
765 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
766 | bh = head; |
767 | do { | |
768 | unlock_buffer(bh); | |
1d8b85cc CL |
769 | bh = bh->b_this_page; |
770 | ||
771 | } while (bh != head); | |
772 | ||
cc4f11e6 | 773 | return rc; |
1d8b85cc | 774 | } |
89cb0888 JK |
775 | |
776 | /* | |
777 | * Migration function for pages with buffers. This function can only be used | |
778 | * if the underlying filesystem guarantees that no other references to "page" | |
779 | * exist. For example attached buffer heads are accessed only under page lock. | |
780 | */ | |
781 | int buffer_migrate_page(struct address_space *mapping, | |
782 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
783 | { | |
784 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
785 | } | |
1d8b85cc | 786 | EXPORT_SYMBOL(buffer_migrate_page); |
89cb0888 JK |
787 | |
788 | /* | |
789 | * Same as above except that this variant is more careful and checks that there | |
790 | * are also no buffer head references. This function is the right one for | |
791 | * mappings where buffer heads are directly looked up and referenced (such as | |
792 | * block device mappings). | |
793 | */ | |
794 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
795 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
796 | { | |
797 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
798 | } | |
9361401e | 799 | #endif |
1d8b85cc | 800 | |
04e62a29 CL |
801 | /* |
802 | * Writeback a page to clean the dirty state | |
803 | */ | |
804 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 805 | { |
04e62a29 CL |
806 | struct writeback_control wbc = { |
807 | .sync_mode = WB_SYNC_NONE, | |
808 | .nr_to_write = 1, | |
809 | .range_start = 0, | |
810 | .range_end = LLONG_MAX, | |
04e62a29 CL |
811 | .for_reclaim = 1 |
812 | }; | |
813 | int rc; | |
814 | ||
815 | if (!mapping->a_ops->writepage) | |
816 | /* No write method for the address space */ | |
817 | return -EINVAL; | |
818 | ||
819 | if (!clear_page_dirty_for_io(page)) | |
820 | /* Someone else already triggered a write */ | |
821 | return -EAGAIN; | |
822 | ||
8351a6e4 | 823 | /* |
04e62a29 CL |
824 | * A dirty page may imply that the underlying filesystem has |
825 | * the page on some queue. So the page must be clean for | |
826 | * migration. Writeout may mean we loose the lock and the | |
827 | * page state is no longer what we checked for earlier. | |
828 | * At this point we know that the migration attempt cannot | |
829 | * be successful. | |
8351a6e4 | 830 | */ |
e388466d | 831 | remove_migration_ptes(page, page, false); |
8351a6e4 | 832 | |
04e62a29 | 833 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 834 | |
04e62a29 CL |
835 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
836 | /* unlocked. Relock */ | |
837 | lock_page(page); | |
838 | ||
bda8550d | 839 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
840 | } |
841 | ||
842 | /* | |
843 | * Default handling if a filesystem does not provide a migration function. | |
844 | */ | |
845 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 846 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 847 | { |
b969c4ab | 848 | if (PageDirty(page)) { |
a6bc32b8 | 849 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
850 | switch (mode) { |
851 | case MIGRATE_SYNC: | |
852 | case MIGRATE_SYNC_NO_COPY: | |
853 | break; | |
854 | default: | |
b969c4ab | 855 | return -EBUSY; |
2916ecc0 | 856 | } |
04e62a29 | 857 | return writeout(mapping, page); |
b969c4ab | 858 | } |
8351a6e4 CL |
859 | |
860 | /* | |
861 | * Buffers may be managed in a filesystem specific way. | |
862 | * We must have no buffers or drop them. | |
863 | */ | |
266cf658 | 864 | if (page_has_private(page) && |
8351a6e4 | 865 | !try_to_release_page(page, GFP_KERNEL)) |
806031bb | 866 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 867 | |
a6bc32b8 | 868 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
869 | } |
870 | ||
e24f0b8f CL |
871 | /* |
872 | * Move a page to a newly allocated page | |
873 | * The page is locked and all ptes have been successfully removed. | |
874 | * | |
875 | * The new page will have replaced the old page if this function | |
876 | * is successful. | |
894bc310 LS |
877 | * |
878 | * Return value: | |
879 | * < 0 - error code | |
78bd5209 | 880 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 881 | */ |
3fe2011f | 882 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 883 | enum migrate_mode mode) |
e24f0b8f CL |
884 | { |
885 | struct address_space *mapping; | |
bda807d4 MK |
886 | int rc = -EAGAIN; |
887 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 888 | |
7db7671f HD |
889 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
890 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 891 | |
e24f0b8f | 892 | mapping = page_mapping(page); |
bda807d4 MK |
893 | |
894 | if (likely(is_lru)) { | |
895 | if (!mapping) | |
896 | rc = migrate_page(mapping, newpage, page, mode); | |
897 | else if (mapping->a_ops->migratepage) | |
898 | /* | |
899 | * Most pages have a mapping and most filesystems | |
900 | * provide a migratepage callback. Anonymous pages | |
901 | * are part of swap space which also has its own | |
902 | * migratepage callback. This is the most common path | |
903 | * for page migration. | |
904 | */ | |
905 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
906 | page, mode); | |
907 | else | |
908 | rc = fallback_migrate_page(mapping, newpage, | |
909 | page, mode); | |
910 | } else { | |
e24f0b8f | 911 | /* |
bda807d4 MK |
912 | * In case of non-lru page, it could be released after |
913 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 914 | */ |
bda807d4 MK |
915 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
916 | if (!PageMovable(page)) { | |
917 | rc = MIGRATEPAGE_SUCCESS; | |
918 | __ClearPageIsolated(page); | |
919 | goto out; | |
920 | } | |
921 | ||
922 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
923 | page, mode); | |
924 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
925 | !PageIsolated(page)); | |
926 | } | |
e24f0b8f | 927 | |
5c3f9a67 HD |
928 | /* |
929 | * When successful, old pagecache page->mapping must be cleared before | |
930 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
931 | */ | |
932 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
933 | if (__PageMovable(page)) { |
934 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
935 | ||
936 | /* | |
937 | * We clear PG_movable under page_lock so any compactor | |
938 | * cannot try to migrate this page. | |
939 | */ | |
940 | __ClearPageIsolated(page); | |
941 | } | |
942 | ||
943 | /* | |
c23a0c99 | 944 | * Anonymous and movable page->mapping will be cleared by |
bda807d4 MK |
945 | * free_pages_prepare so don't reset it here for keeping |
946 | * the type to work PageAnon, for example. | |
947 | */ | |
948 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 949 | page->mapping = NULL; |
d2b2c6dd | 950 | |
25b2995a | 951 | if (likely(!is_zone_device_page(newpage))) |
d2b2c6dd LP |
952 | flush_dcache_page(newpage); |
953 | ||
3fe2011f | 954 | } |
bda807d4 | 955 | out: |
e24f0b8f CL |
956 | return rc; |
957 | } | |
958 | ||
0dabec93 | 959 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 960 | int force, enum migrate_mode mode) |
e24f0b8f | 961 | { |
0dabec93 | 962 | int rc = -EAGAIN; |
213ecb31 | 963 | bool page_was_mapped = false; |
3f6c8272 | 964 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 965 | bool is_lru = !__PageMovable(page); |
95a402c3 | 966 | |
529ae9aa | 967 | if (!trylock_page(page)) { |
a6bc32b8 | 968 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 969 | goto out; |
3e7d3449 MG |
970 | |
971 | /* | |
972 | * It's not safe for direct compaction to call lock_page. | |
973 | * For example, during page readahead pages are added locked | |
974 | * to the LRU. Later, when the IO completes the pages are | |
975 | * marked uptodate and unlocked. However, the queueing | |
976 | * could be merging multiple pages for one bio (e.g. | |
d4388340 | 977 | * mpage_readahead). If an allocation happens for the |
3e7d3449 MG |
978 | * second or third page, the process can end up locking |
979 | * the same page twice and deadlocking. Rather than | |
980 | * trying to be clever about what pages can be locked, | |
981 | * avoid the use of lock_page for direct compaction | |
982 | * altogether. | |
983 | */ | |
984 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 985 | goto out; |
3e7d3449 | 986 | |
e24f0b8f CL |
987 | lock_page(page); |
988 | } | |
989 | ||
990 | if (PageWriteback(page)) { | |
11bc82d6 | 991 | /* |
fed5b64a | 992 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
993 | * necessary to wait for PageWriteback. In the async case, |
994 | * the retry loop is too short and in the sync-light case, | |
995 | * the overhead of stalling is too much | |
11bc82d6 | 996 | */ |
2916ecc0 JG |
997 | switch (mode) { |
998 | case MIGRATE_SYNC: | |
999 | case MIGRATE_SYNC_NO_COPY: | |
1000 | break; | |
1001 | default: | |
11bc82d6 | 1002 | rc = -EBUSY; |
0a31bc97 | 1003 | goto out_unlock; |
11bc82d6 AA |
1004 | } |
1005 | if (!force) | |
0a31bc97 | 1006 | goto out_unlock; |
e24f0b8f CL |
1007 | wait_on_page_writeback(page); |
1008 | } | |
03f15c86 | 1009 | |
e24f0b8f | 1010 | /* |
68a9843f | 1011 | * By try_to_migrate(), page->mapcount goes down to 0 here. In this case, |
dc386d4d | 1012 | * we cannot notice that anon_vma is freed while we migrates a page. |
1ce82b69 | 1013 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1014 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1015 | * File Caches may use write_page() or lock_page() in migration, then, |
1016 | * just care Anon page here. | |
03f15c86 HD |
1017 | * |
1018 | * Only page_get_anon_vma() understands the subtleties of | |
1019 | * getting a hold on an anon_vma from outside one of its mms. | |
1020 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1021 | * because that implies that the anon page is no longer mapped | |
1022 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1023 | */ |
03f15c86 | 1024 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1025 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1026 | |
7db7671f HD |
1027 | /* |
1028 | * Block others from accessing the new page when we get around to | |
1029 | * establishing additional references. We are usually the only one | |
1030 | * holding a reference to newpage at this point. We used to have a BUG | |
1031 | * here if trylock_page(newpage) fails, but would like to allow for | |
1032 | * cases where there might be a race with the previous use of newpage. | |
1033 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1034 | */ | |
1035 | if (unlikely(!trylock_page(newpage))) | |
1036 | goto out_unlock; | |
1037 | ||
bda807d4 MK |
1038 | if (unlikely(!is_lru)) { |
1039 | rc = move_to_new_page(newpage, page, mode); | |
1040 | goto out_unlock_both; | |
1041 | } | |
1042 | ||
dc386d4d | 1043 | /* |
62e1c553 SL |
1044 | * Corner case handling: |
1045 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1046 | * and treated as swapcache but it has no rmap yet. | |
1047 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1048 | * trigger a BUG. So handle it here. | |
d12b8951 | 1049 | * 2. An orphaned page (see truncate_cleanup_page) might have |
62e1c553 SL |
1050 | * fs-private metadata. The page can be picked up due to memory |
1051 | * offlining. Everywhere else except page reclaim, the page is | |
1052 | * invisible to the vm, so the page can not be migrated. So try to | |
1053 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1054 | */ |
62e1c553 | 1055 | if (!page->mapping) { |
309381fe | 1056 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1057 | if (page_has_private(page)) { |
62e1c553 | 1058 | try_to_free_buffers(page); |
7db7671f | 1059 | goto out_unlock_both; |
62e1c553 | 1060 | } |
7db7671f HD |
1061 | } else if (page_mapped(page)) { |
1062 | /* Establish migration ptes */ | |
03f15c86 HD |
1063 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1064 | page); | |
a98a2f0c | 1065 | try_to_migrate(page, 0); |
213ecb31 | 1066 | page_was_mapped = true; |
2ebba6b7 | 1067 | } |
dc386d4d | 1068 | |
e6a1530d | 1069 | if (!page_mapped(page)) |
5c3f9a67 | 1070 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1071 | |
5c3f9a67 HD |
1072 | if (page_was_mapped) |
1073 | remove_migration_ptes(page, | |
e388466d | 1074 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1075 | |
7db7671f HD |
1076 | out_unlock_both: |
1077 | unlock_page(newpage); | |
1078 | out_unlock: | |
3f6c8272 | 1079 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1080 | if (anon_vma) |
9e60109f | 1081 | put_anon_vma(anon_vma); |
e24f0b8f | 1082 | unlock_page(page); |
0dabec93 | 1083 | out: |
c6c919eb MK |
1084 | /* |
1085 | * If migration is successful, decrease refcount of the newpage | |
1086 | * which will not free the page because new page owner increased | |
1087 | * refcounter. As well, if it is LRU page, add the page to LRU | |
e0a352fa DH |
1088 | * list in here. Use the old state of the isolated source page to |
1089 | * determine if we migrated a LRU page. newpage was already unlocked | |
1090 | * and possibly modified by its owner - don't rely on the page | |
1091 | * state. | |
c6c919eb MK |
1092 | */ |
1093 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e0a352fa | 1094 | if (unlikely(!is_lru)) |
c6c919eb MK |
1095 | put_page(newpage); |
1096 | else | |
1097 | putback_lru_page(newpage); | |
1098 | } | |
1099 | ||
0dabec93 MK |
1100 | return rc; |
1101 | } | |
95a402c3 | 1102 | |
79c28a41 DH |
1103 | |
1104 | /* | |
1105 | * node_demotion[] example: | |
1106 | * | |
1107 | * Consider a system with two sockets. Each socket has | |
1108 | * three classes of memory attached: fast, medium and slow. | |
1109 | * Each memory class is placed in its own NUMA node. The | |
1110 | * CPUs are placed in the node with the "fast" memory. The | |
1111 | * 6 NUMA nodes (0-5) might be split among the sockets like | |
1112 | * this: | |
1113 | * | |
1114 | * Socket A: 0, 1, 2 | |
1115 | * Socket B: 3, 4, 5 | |
1116 | * | |
1117 | * When Node 0 fills up, its memory should be migrated to | |
1118 | * Node 1. When Node 1 fills up, it should be migrated to | |
1119 | * Node 2. The migration path start on the nodes with the | |
1120 | * processors (since allocations default to this node) and | |
1121 | * fast memory, progress through medium and end with the | |
1122 | * slow memory: | |
1123 | * | |
1124 | * 0 -> 1 -> 2 -> stop | |
1125 | * 3 -> 4 -> 5 -> stop | |
1126 | * | |
1127 | * This is represented in the node_demotion[] like this: | |
1128 | * | |
1129 | * { 1, // Node 0 migrates to 1 | |
1130 | * 2, // Node 1 migrates to 2 | |
1131 | * -1, // Node 2 does not migrate | |
1132 | * 4, // Node 3 migrates to 4 | |
1133 | * 5, // Node 4 migrates to 5 | |
1134 | * -1} // Node 5 does not migrate | |
1135 | */ | |
1136 | ||
1137 | /* | |
1138 | * Writes to this array occur without locking. Cycles are | |
1139 | * not allowed: Node X demotes to Y which demotes to X... | |
1140 | * | |
1141 | * If multiple reads are performed, a single rcu_read_lock() | |
1142 | * must be held over all reads to ensure that no cycles are | |
1143 | * observed. | |
1144 | */ | |
1145 | static int node_demotion[MAX_NUMNODES] __read_mostly = | |
1146 | {[0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE}; | |
1147 | ||
1148 | /** | |
1149 | * next_demotion_node() - Get the next node in the demotion path | |
1150 | * @node: The starting node to lookup the next node | |
1151 | * | |
c9bd7d18 | 1152 | * Return: node id for next memory node in the demotion path hierarchy |
79c28a41 DH |
1153 | * from @node; NUMA_NO_NODE if @node is terminal. This does not keep |
1154 | * @node online or guarantee that it *continues* to be the next demotion | |
1155 | * target. | |
1156 | */ | |
1157 | int next_demotion_node(int node) | |
1158 | { | |
1159 | int target; | |
1160 | ||
1161 | /* | |
1162 | * node_demotion[] is updated without excluding this | |
1163 | * function from running. RCU doesn't provide any | |
1164 | * compiler barriers, so the READ_ONCE() is required | |
1165 | * to avoid compiler reordering or read merging. | |
1166 | * | |
1167 | * Make sure to use RCU over entire code blocks if | |
1168 | * node_demotion[] reads need to be consistent. | |
1169 | */ | |
1170 | rcu_read_lock(); | |
1171 | target = READ_ONCE(node_demotion[node]); | |
1172 | rcu_read_unlock(); | |
1173 | ||
1174 | return target; | |
1175 | } | |
1176 | ||
0dabec93 MK |
1177 | /* |
1178 | * Obtain the lock on page, remove all ptes and migrate the page | |
1179 | * to the newly allocated page in newpage. | |
1180 | */ | |
6ec4476a | 1181 | static int unmap_and_move(new_page_t get_new_page, |
ef2a5153 GU |
1182 | free_page_t put_new_page, |
1183 | unsigned long private, struct page *page, | |
add05cec | 1184 | int force, enum migrate_mode mode, |
dd4ae78a YS |
1185 | enum migrate_reason reason, |
1186 | struct list_head *ret) | |
0dabec93 | 1187 | { |
2def7424 | 1188 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1189 | struct page *newpage = NULL; |
0dabec93 | 1190 | |
94723aaf | 1191 | if (!thp_migration_supported() && PageTransHuge(page)) |
d532e2e5 | 1192 | return -ENOSYS; |
94723aaf | 1193 | |
0dabec93 MK |
1194 | if (page_count(page) == 1) { |
1195 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1196 | ClearPageActive(page); |
1197 | ClearPageUnevictable(page); | |
bda807d4 MK |
1198 | if (unlikely(__PageMovable(page))) { |
1199 | lock_page(page); | |
1200 | if (!PageMovable(page)) | |
1201 | __ClearPageIsolated(page); | |
1202 | unlock_page(page); | |
1203 | } | |
0dabec93 MK |
1204 | goto out; |
1205 | } | |
1206 | ||
74d4a579 YS |
1207 | newpage = get_new_page(page, private); |
1208 | if (!newpage) | |
1209 | return -ENOMEM; | |
1210 | ||
9c620e2b | 1211 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1212 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1213 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1214 | |
0dabec93 | 1215 | out: |
e24f0b8f | 1216 | if (rc != -EAGAIN) { |
0dabec93 MK |
1217 | /* |
1218 | * A page that has been migrated has all references | |
1219 | * removed and will be freed. A page that has not been | |
c23a0c99 | 1220 | * migrated will have kept its references and be restored. |
0dabec93 MK |
1221 | */ |
1222 | list_del(&page->lru); | |
dd4ae78a | 1223 | } |
6afcf8ef | 1224 | |
dd4ae78a YS |
1225 | /* |
1226 | * If migration is successful, releases reference grabbed during | |
1227 | * isolation. Otherwise, restore the page to right list unless | |
1228 | * we want to retry. | |
1229 | */ | |
1230 | if (rc == MIGRATEPAGE_SUCCESS) { | |
6afcf8ef ML |
1231 | /* |
1232 | * Compaction can migrate also non-LRU pages which are | |
1233 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1234 | * as __PageMovable | |
1235 | */ | |
1236 | if (likely(!__PageMovable(page))) | |
e8db67eb | 1237 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 1238 | page_is_file_lru(page), -thp_nr_pages(page)); |
c6c919eb | 1239 | |
79f5f8fa | 1240 | if (reason != MR_MEMORY_FAILURE) |
d7e69488 | 1241 | /* |
79f5f8fa | 1242 | * We release the page in page_handle_poison. |
d7e69488 | 1243 | */ |
79f5f8fa | 1244 | put_page(page); |
c6c919eb | 1245 | } else { |
dd4ae78a YS |
1246 | if (rc != -EAGAIN) |
1247 | list_add_tail(&page->lru, ret); | |
bda807d4 | 1248 | |
c6c919eb MK |
1249 | if (put_new_page) |
1250 | put_new_page(newpage, private); | |
1251 | else | |
1252 | put_page(newpage); | |
e24f0b8f | 1253 | } |
68711a74 | 1254 | |
e24f0b8f CL |
1255 | return rc; |
1256 | } | |
1257 | ||
290408d4 NH |
1258 | /* |
1259 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1260 | * | |
1261 | * This function doesn't wait the completion of hugepage I/O | |
1262 | * because there is no race between I/O and migration for hugepage. | |
1263 | * Note that currently hugepage I/O occurs only in direct I/O | |
1264 | * where no lock is held and PG_writeback is irrelevant, | |
1265 | * and writeback status of all subpages are counted in the reference | |
1266 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1267 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1268 | * This means that when we try to migrate hugepage whose subpages are | |
1269 | * doing direct I/O, some references remain after try_to_unmap() and | |
1270 | * hugepage migration fails without data corruption. | |
1271 | * | |
1272 | * There is also no race when direct I/O is issued on the page under migration, | |
1273 | * because then pte is replaced with migration swap entry and direct I/O code | |
1274 | * will wait in the page fault for migration to complete. | |
1275 | */ | |
1276 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1277 | free_page_t put_new_page, unsigned long private, |
1278 | struct page *hpage, int force, | |
dd4ae78a YS |
1279 | enum migrate_mode mode, int reason, |
1280 | struct list_head *ret) | |
290408d4 | 1281 | { |
2def7424 | 1282 | int rc = -EAGAIN; |
2ebba6b7 | 1283 | int page_was_mapped = 0; |
32665f2b | 1284 | struct page *new_hpage; |
290408d4 | 1285 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1286 | struct address_space *mapping = NULL; |
290408d4 | 1287 | |
83467efb | 1288 | /* |
7ed2c31d | 1289 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1290 | * This check is necessary because some callers of hugepage migration |
1291 | * like soft offline and memory hotremove don't walk through page | |
1292 | * tables or check whether the hugepage is pmd-based or not before | |
1293 | * kicking migration. | |
1294 | */ | |
100873d7 | 1295 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
dd4ae78a | 1296 | list_move_tail(&hpage->lru, ret); |
83467efb | 1297 | return -ENOSYS; |
32665f2b | 1298 | } |
83467efb | 1299 | |
71a64f61 MS |
1300 | if (page_count(hpage) == 1) { |
1301 | /* page was freed from under us. So we are done. */ | |
1302 | putback_active_hugepage(hpage); | |
1303 | return MIGRATEPAGE_SUCCESS; | |
1304 | } | |
1305 | ||
666feb21 | 1306 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1307 | if (!new_hpage) |
1308 | return -ENOMEM; | |
1309 | ||
290408d4 | 1310 | if (!trylock_page(hpage)) { |
2916ecc0 | 1311 | if (!force) |
290408d4 | 1312 | goto out; |
2916ecc0 JG |
1313 | switch (mode) { |
1314 | case MIGRATE_SYNC: | |
1315 | case MIGRATE_SYNC_NO_COPY: | |
1316 | break; | |
1317 | default: | |
1318 | goto out; | |
1319 | } | |
290408d4 NH |
1320 | lock_page(hpage); |
1321 | } | |
1322 | ||
cb6acd01 MK |
1323 | /* |
1324 | * Check for pages which are in the process of being freed. Without | |
1325 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1326 | * be called and we could leak usage counts for subpools. | |
1327 | */ | |
6acfb5ba | 1328 | if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) { |
cb6acd01 MK |
1329 | rc = -EBUSY; |
1330 | goto out_unlock; | |
1331 | } | |
1332 | ||
746b18d4 PZ |
1333 | if (PageAnon(hpage)) |
1334 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1335 | |
7db7671f HD |
1336 | if (unlikely(!trylock_page(new_hpage))) |
1337 | goto put_anon; | |
1338 | ||
2ebba6b7 | 1339 | if (page_mapped(hpage)) { |
336bf30e | 1340 | bool mapping_locked = false; |
a98a2f0c | 1341 | enum ttu_flags ttu = 0; |
336bf30e MK |
1342 | |
1343 | if (!PageAnon(hpage)) { | |
1344 | /* | |
1345 | * In shared mappings, try_to_unmap could potentially | |
1346 | * call huge_pmd_unshare. Because of this, take | |
1347 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1348 | * to let lower levels know we have taken the lock. | |
1349 | */ | |
1350 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1351 | if (unlikely(!mapping)) | |
1352 | goto unlock_put_anon; | |
1353 | ||
1354 | mapping_locked = true; | |
1355 | ttu |= TTU_RMAP_LOCKED; | |
1356 | } | |
c0d0381a | 1357 | |
a98a2f0c | 1358 | try_to_migrate(hpage, ttu); |
2ebba6b7 | 1359 | page_was_mapped = 1; |
336bf30e MK |
1360 | |
1361 | if (mapping_locked) | |
1362 | i_mmap_unlock_write(mapping); | |
2ebba6b7 | 1363 | } |
290408d4 NH |
1364 | |
1365 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1366 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1367 | |
336bf30e | 1368 | if (page_was_mapped) |
5c3f9a67 | 1369 | remove_migration_ptes(hpage, |
336bf30e | 1370 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1371 | |
c0d0381a | 1372 | unlock_put_anon: |
7db7671f HD |
1373 | unlock_page(new_hpage); |
1374 | ||
1375 | put_anon: | |
fd4a4663 | 1376 | if (anon_vma) |
9e60109f | 1377 | put_anon_vma(anon_vma); |
8e6ac7fa | 1378 | |
2def7424 | 1379 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1380 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1381 | put_new_page = NULL; |
1382 | } | |
8e6ac7fa | 1383 | |
cb6acd01 | 1384 | out_unlock: |
290408d4 | 1385 | unlock_page(hpage); |
09761333 | 1386 | out: |
dd4ae78a | 1387 | if (rc == MIGRATEPAGE_SUCCESS) |
b8ec1cee | 1388 | putback_active_hugepage(hpage); |
a04840c6 | 1389 | else if (rc != -EAGAIN) |
dd4ae78a | 1390 | list_move_tail(&hpage->lru, ret); |
68711a74 DR |
1391 | |
1392 | /* | |
1393 | * If migration was not successful and there's a freeing callback, use | |
1394 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1395 | * isolation. | |
1396 | */ | |
2def7424 | 1397 | if (put_new_page) |
68711a74 DR |
1398 | put_new_page(new_hpage, private); |
1399 | else | |
3aaa76e1 | 1400 | putback_active_hugepage(new_hpage); |
68711a74 | 1401 | |
290408d4 NH |
1402 | return rc; |
1403 | } | |
1404 | ||
d532e2e5 YS |
1405 | static inline int try_split_thp(struct page *page, struct page **page2, |
1406 | struct list_head *from) | |
1407 | { | |
1408 | int rc = 0; | |
1409 | ||
1410 | lock_page(page); | |
1411 | rc = split_huge_page_to_list(page, from); | |
1412 | unlock_page(page); | |
1413 | if (!rc) | |
1414 | list_safe_reset_next(page, *page2, lru); | |
1415 | ||
1416 | return rc; | |
1417 | } | |
1418 | ||
b20a3503 | 1419 | /* |
c73e5c9c SB |
1420 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1421 | * supplied as the target for the page migration | |
b20a3503 | 1422 | * |
c73e5c9c SB |
1423 | * @from: The list of pages to be migrated. |
1424 | * @get_new_page: The function used to allocate free pages to be used | |
1425 | * as the target of the page migration. | |
68711a74 DR |
1426 | * @put_new_page: The function used to free target pages if migration |
1427 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1428 | * @private: Private data to be passed on to get_new_page() |
1429 | * @mode: The migration mode that specifies the constraints for | |
1430 | * page migration, if any. | |
1431 | * @reason: The reason for page migration. | |
5ac95884 YS |
1432 | * @ret_succeeded: Set to the number of pages migrated successfully if |
1433 | * the caller passes a non-NULL pointer. | |
b20a3503 | 1434 | * |
c73e5c9c SB |
1435 | * The function returns after 10 attempts or if no pages are movable any more |
1436 | * because the list has become empty or no retryable pages exist any more. | |
dd4ae78a YS |
1437 | * It is caller's responsibility to call putback_movable_pages() to return pages |
1438 | * to the LRU or free list only if ret != 0. | |
b20a3503 | 1439 | * |
c73e5c9c | 1440 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1441 | */ |
9c620e2b | 1442 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 | 1443 | free_page_t put_new_page, unsigned long private, |
5ac95884 | 1444 | enum migrate_mode mode, int reason, unsigned int *ret_succeeded) |
b20a3503 | 1445 | { |
e24f0b8f | 1446 | int retry = 1; |
1a5bae25 | 1447 | int thp_retry = 1; |
b20a3503 | 1448 | int nr_failed = 0; |
5647bc29 | 1449 | int nr_succeeded = 0; |
1a5bae25 AK |
1450 | int nr_thp_succeeded = 0; |
1451 | int nr_thp_failed = 0; | |
1452 | int nr_thp_split = 0; | |
b20a3503 | 1453 | int pass = 0; |
1a5bae25 | 1454 | bool is_thp = false; |
b20a3503 CL |
1455 | struct page *page; |
1456 | struct page *page2; | |
1457 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1a5bae25 | 1458 | int rc, nr_subpages; |
dd4ae78a | 1459 | LIST_HEAD(ret_pages); |
b0b515bf | 1460 | bool nosplit = (reason == MR_NUMA_MISPLACED); |
b20a3503 | 1461 | |
7bc1aec5 LM |
1462 | trace_mm_migrate_pages_start(mode, reason); |
1463 | ||
b20a3503 CL |
1464 | if (!swapwrite) |
1465 | current->flags |= PF_SWAPWRITE; | |
1466 | ||
1a5bae25 | 1467 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { |
e24f0b8f | 1468 | retry = 0; |
1a5bae25 | 1469 | thp_retry = 0; |
b20a3503 | 1470 | |
e24f0b8f | 1471 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1472 | retry: |
1a5bae25 AK |
1473 | /* |
1474 | * THP statistics is based on the source huge page. | |
1475 | * Capture required information that might get lost | |
1476 | * during migration. | |
1477 | */ | |
6c5c7b9f | 1478 | is_thp = PageTransHuge(page) && !PageHuge(page); |
6c357848 | 1479 | nr_subpages = thp_nr_pages(page); |
e24f0b8f | 1480 | cond_resched(); |
2d1db3b1 | 1481 | |
31caf665 NH |
1482 | if (PageHuge(page)) |
1483 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1484 | put_new_page, private, page, |
dd4ae78a YS |
1485 | pass > 2, mode, reason, |
1486 | &ret_pages); | |
31caf665 | 1487 | else |
68711a74 | 1488 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec | 1489 | private, page, pass > 2, mode, |
dd4ae78a YS |
1490 | reason, &ret_pages); |
1491 | /* | |
1492 | * The rules are: | |
1493 | * Success: non hugetlb page will be freed, hugetlb | |
1494 | * page will be put back | |
1495 | * -EAGAIN: stay on the from list | |
1496 | * -ENOMEM: stay on the from list | |
1497 | * Other errno: put on ret_pages list then splice to | |
1498 | * from list | |
1499 | */ | |
e24f0b8f | 1500 | switch(rc) { |
d532e2e5 YS |
1501 | /* |
1502 | * THP migration might be unsupported or the | |
1503 | * allocation could've failed so we should | |
1504 | * retry on the same page with the THP split | |
1505 | * to base pages. | |
1506 | * | |
1507 | * Head page is retried immediately and tail | |
1508 | * pages are added to the tail of the list so | |
1509 | * we encounter them after the rest of the list | |
1510 | * is processed. | |
1511 | */ | |
1512 | case -ENOSYS: | |
1513 | /* THP migration is unsupported */ | |
1514 | if (is_thp) { | |
1515 | if (!try_split_thp(page, &page2, from)) { | |
1516 | nr_thp_split++; | |
1517 | goto retry; | |
1518 | } | |
1519 | ||
1520 | nr_thp_failed++; | |
1521 | nr_failed += nr_subpages; | |
1522 | break; | |
1523 | } | |
1524 | ||
1525 | /* Hugetlb migration is unsupported */ | |
1526 | nr_failed++; | |
1527 | break; | |
95a402c3 | 1528 | case -ENOMEM: |
94723aaf | 1529 | /* |
d532e2e5 YS |
1530 | * When memory is low, don't bother to try to migrate |
1531 | * other pages, just exit. | |
b0b515bf | 1532 | * THP NUMA faulting doesn't split THP to retry. |
94723aaf | 1533 | */ |
b0b515bf | 1534 | if (is_thp && !nosplit) { |
d532e2e5 | 1535 | if (!try_split_thp(page, &page2, from)) { |
1a5bae25 | 1536 | nr_thp_split++; |
94723aaf MH |
1537 | goto retry; |
1538 | } | |
6c5c7b9f | 1539 | |
1a5bae25 AK |
1540 | nr_thp_failed++; |
1541 | nr_failed += nr_subpages; | |
1542 | goto out; | |
1543 | } | |
dfef2ef4 | 1544 | nr_failed++; |
95a402c3 | 1545 | goto out; |
e24f0b8f | 1546 | case -EAGAIN: |
1a5bae25 AK |
1547 | if (is_thp) { |
1548 | thp_retry++; | |
1549 | break; | |
1550 | } | |
2d1db3b1 | 1551 | retry++; |
e24f0b8f | 1552 | break; |
78bd5209 | 1553 | case MIGRATEPAGE_SUCCESS: |
1a5bae25 AK |
1554 | if (is_thp) { |
1555 | nr_thp_succeeded++; | |
1556 | nr_succeeded += nr_subpages; | |
1557 | break; | |
1558 | } | |
5647bc29 | 1559 | nr_succeeded++; |
e24f0b8f CL |
1560 | break; |
1561 | default: | |
354a3363 | 1562 | /* |
d532e2e5 | 1563 | * Permanent failure (-EBUSY, etc.): |
354a3363 NH |
1564 | * unlike -EAGAIN case, the failed page is |
1565 | * removed from migration page list and not | |
1566 | * retried in the next outer loop. | |
1567 | */ | |
1a5bae25 AK |
1568 | if (is_thp) { |
1569 | nr_thp_failed++; | |
1570 | nr_failed += nr_subpages; | |
1571 | break; | |
1572 | } | |
2d1db3b1 | 1573 | nr_failed++; |
e24f0b8f | 1574 | break; |
2d1db3b1 | 1575 | } |
b20a3503 CL |
1576 | } |
1577 | } | |
1a5bae25 AK |
1578 | nr_failed += retry + thp_retry; |
1579 | nr_thp_failed += thp_retry; | |
f2f81fb2 | 1580 | rc = nr_failed; |
95a402c3 | 1581 | out: |
dd4ae78a YS |
1582 | /* |
1583 | * Put the permanent failure page back to migration list, they | |
1584 | * will be put back to the right list by the caller. | |
1585 | */ | |
1586 | list_splice(&ret_pages, from); | |
1587 | ||
1a5bae25 AK |
1588 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
1589 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1590 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); | |
1591 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); | |
1592 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); | |
1593 | trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded, | |
1594 | nr_thp_failed, nr_thp_split, mode, reason); | |
7b2a2d4a | 1595 | |
b20a3503 CL |
1596 | if (!swapwrite) |
1597 | current->flags &= ~PF_SWAPWRITE; | |
1598 | ||
5ac95884 YS |
1599 | if (ret_succeeded) |
1600 | *ret_succeeded = nr_succeeded; | |
1601 | ||
78bd5209 | 1602 | return rc; |
b20a3503 | 1603 | } |
95a402c3 | 1604 | |
19fc7bed | 1605 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
b4b38223 | 1606 | { |
19fc7bed JK |
1607 | struct migration_target_control *mtc; |
1608 | gfp_t gfp_mask; | |
b4b38223 JK |
1609 | unsigned int order = 0; |
1610 | struct page *new_page = NULL; | |
19fc7bed JK |
1611 | int nid; |
1612 | int zidx; | |
1613 | ||
1614 | mtc = (struct migration_target_control *)private; | |
1615 | gfp_mask = mtc->gfp_mask; | |
1616 | nid = mtc->nid; | |
1617 | if (nid == NUMA_NO_NODE) | |
1618 | nid = page_to_nid(page); | |
b4b38223 | 1619 | |
d92bbc27 JK |
1620 | if (PageHuge(page)) { |
1621 | struct hstate *h = page_hstate(compound_head(page)); | |
1622 | ||
19fc7bed JK |
1623 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
1624 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); | |
d92bbc27 | 1625 | } |
b4b38223 JK |
1626 | |
1627 | if (PageTransHuge(page)) { | |
9933a0c8 JK |
1628 | /* |
1629 | * clear __GFP_RECLAIM to make the migration callback | |
1630 | * consistent with regular THP allocations. | |
1631 | */ | |
1632 | gfp_mask &= ~__GFP_RECLAIM; | |
b4b38223 JK |
1633 | gfp_mask |= GFP_TRANSHUGE; |
1634 | order = HPAGE_PMD_ORDER; | |
1635 | } | |
19fc7bed JK |
1636 | zidx = zone_idx(page_zone(page)); |
1637 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) | |
b4b38223 JK |
1638 | gfp_mask |= __GFP_HIGHMEM; |
1639 | ||
84172f4b | 1640 | new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask); |
b4b38223 JK |
1641 | |
1642 | if (new_page && PageTransHuge(new_page)) | |
1643 | prep_transhuge_page(new_page); | |
1644 | ||
1645 | return new_page; | |
1646 | } | |
1647 | ||
742755a1 | 1648 | #ifdef CONFIG_NUMA |
742755a1 | 1649 | |
a49bd4d7 | 1650 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1651 | { |
a49bd4d7 MH |
1652 | while (nr-- > 0) { |
1653 | if (put_user(value, status + start)) | |
1654 | return -EFAULT; | |
1655 | start++; | |
1656 | } | |
1657 | ||
1658 | return 0; | |
1659 | } | |
1660 | ||
1661 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1662 | struct list_head *pagelist, int node) | |
1663 | { | |
1664 | int err; | |
a0976311 JK |
1665 | struct migration_target_control mtc = { |
1666 | .nid = node, | |
1667 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
1668 | }; | |
a49bd4d7 | 1669 | |
a0976311 | 1670 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
5ac95884 | 1671 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); |
a49bd4d7 MH |
1672 | if (err) |
1673 | putback_movable_pages(pagelist); | |
1674 | return err; | |
742755a1 CL |
1675 | } |
1676 | ||
1677 | /* | |
a49bd4d7 MH |
1678 | * Resolves the given address to a struct page, isolates it from the LRU and |
1679 | * puts it to the given pagelist. | |
e0153fc2 YS |
1680 | * Returns: |
1681 | * errno - if the page cannot be found/isolated | |
1682 | * 0 - when it doesn't have to be migrated because it is already on the | |
1683 | * target node | |
1684 | * 1 - when it has been queued | |
742755a1 | 1685 | */ |
a49bd4d7 MH |
1686 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1687 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1688 | { |
a49bd4d7 MH |
1689 | struct vm_area_struct *vma; |
1690 | struct page *page; | |
1691 | unsigned int follflags; | |
742755a1 | 1692 | int err; |
742755a1 | 1693 | |
d8ed45c5 | 1694 | mmap_read_lock(mm); |
a49bd4d7 MH |
1695 | err = -EFAULT; |
1696 | vma = find_vma(mm, addr); | |
1697 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1698 | goto out; | |
742755a1 | 1699 | |
a49bd4d7 MH |
1700 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1701 | follflags = FOLL_GET | FOLL_DUMP; | |
a49bd4d7 | 1702 | page = follow_page(vma, addr, follflags); |
89f5b7da | 1703 | |
a49bd4d7 MH |
1704 | err = PTR_ERR(page); |
1705 | if (IS_ERR(page)) | |
1706 | goto out; | |
89f5b7da | 1707 | |
a49bd4d7 MH |
1708 | err = -ENOENT; |
1709 | if (!page) | |
1710 | goto out; | |
742755a1 | 1711 | |
a49bd4d7 MH |
1712 | err = 0; |
1713 | if (page_to_nid(page) == node) | |
1714 | goto out_putpage; | |
742755a1 | 1715 | |
a49bd4d7 MH |
1716 | err = -EACCES; |
1717 | if (page_mapcount(page) > 1 && !migrate_all) | |
1718 | goto out_putpage; | |
742755a1 | 1719 | |
a49bd4d7 MH |
1720 | if (PageHuge(page)) { |
1721 | if (PageHead(page)) { | |
1722 | isolate_huge_page(page, pagelist); | |
e0153fc2 | 1723 | err = 1; |
e632a938 | 1724 | } |
a49bd4d7 MH |
1725 | } else { |
1726 | struct page *head; | |
e632a938 | 1727 | |
e8db67eb NH |
1728 | head = compound_head(page); |
1729 | err = isolate_lru_page(head); | |
cf608ac1 | 1730 | if (err) |
a49bd4d7 | 1731 | goto out_putpage; |
742755a1 | 1732 | |
e0153fc2 | 1733 | err = 1; |
a49bd4d7 MH |
1734 | list_add_tail(&head->lru, pagelist); |
1735 | mod_node_page_state(page_pgdat(head), | |
9de4f22a | 1736 | NR_ISOLATED_ANON + page_is_file_lru(head), |
6c357848 | 1737 | thp_nr_pages(head)); |
a49bd4d7 MH |
1738 | } |
1739 | out_putpage: | |
1740 | /* | |
1741 | * Either remove the duplicate refcount from | |
1742 | * isolate_lru_page() or drop the page ref if it was | |
1743 | * not isolated. | |
1744 | */ | |
1745 | put_page(page); | |
1746 | out: | |
d8ed45c5 | 1747 | mmap_read_unlock(mm); |
742755a1 CL |
1748 | return err; |
1749 | } | |
1750 | ||
7ca8783a WY |
1751 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
1752 | struct list_head *pagelist, int __user *status, | |
1753 | int start, int i, unsigned long nr_pages) | |
1754 | { | |
1755 | int err; | |
1756 | ||
5d7ae891 WY |
1757 | if (list_empty(pagelist)) |
1758 | return 0; | |
1759 | ||
7ca8783a WY |
1760 | err = do_move_pages_to_node(mm, pagelist, node); |
1761 | if (err) { | |
1762 | /* | |
1763 | * Positive err means the number of failed | |
1764 | * pages to migrate. Since we are going to | |
1765 | * abort and return the number of non-migrated | |
ab9dd4f8 | 1766 | * pages, so need to include the rest of the |
7ca8783a WY |
1767 | * nr_pages that have not been attempted as |
1768 | * well. | |
1769 | */ | |
1770 | if (err > 0) | |
1771 | err += nr_pages - i - 1; | |
1772 | return err; | |
1773 | } | |
1774 | return store_status(status, start, node, i - start); | |
1775 | } | |
1776 | ||
5e9a0f02 BG |
1777 | /* |
1778 | * Migrate an array of page address onto an array of nodes and fill | |
1779 | * the corresponding array of status. | |
1780 | */ | |
3268c63e | 1781 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1782 | unsigned long nr_pages, |
1783 | const void __user * __user *pages, | |
1784 | const int __user *nodes, | |
1785 | int __user *status, int flags) | |
1786 | { | |
a49bd4d7 MH |
1787 | int current_node = NUMA_NO_NODE; |
1788 | LIST_HEAD(pagelist); | |
1789 | int start, i; | |
1790 | int err = 0, err1; | |
35282a2d | 1791 | |
361a2a22 | 1792 | lru_cache_disable(); |
35282a2d | 1793 | |
a49bd4d7 MH |
1794 | for (i = start = 0; i < nr_pages; i++) { |
1795 | const void __user *p; | |
1796 | unsigned long addr; | |
1797 | int node; | |
3140a227 | 1798 | |
a49bd4d7 MH |
1799 | err = -EFAULT; |
1800 | if (get_user(p, pages + i)) | |
1801 | goto out_flush; | |
1802 | if (get_user(node, nodes + i)) | |
1803 | goto out_flush; | |
057d3389 | 1804 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1805 | |
1806 | err = -ENODEV; | |
1807 | if (node < 0 || node >= MAX_NUMNODES) | |
1808 | goto out_flush; | |
1809 | if (!node_state(node, N_MEMORY)) | |
1810 | goto out_flush; | |
5e9a0f02 | 1811 | |
a49bd4d7 MH |
1812 | err = -EACCES; |
1813 | if (!node_isset(node, task_nodes)) | |
1814 | goto out_flush; | |
1815 | ||
1816 | if (current_node == NUMA_NO_NODE) { | |
1817 | current_node = node; | |
1818 | start = i; | |
1819 | } else if (node != current_node) { | |
7ca8783a WY |
1820 | err = move_pages_and_store_status(mm, current_node, |
1821 | &pagelist, status, start, i, nr_pages); | |
a49bd4d7 MH |
1822 | if (err) |
1823 | goto out; | |
1824 | start = i; | |
1825 | current_node = node; | |
3140a227 BG |
1826 | } |
1827 | ||
a49bd4d7 MH |
1828 | /* |
1829 | * Errors in the page lookup or isolation are not fatal and we simply | |
1830 | * report them via status | |
1831 | */ | |
1832 | err = add_page_for_migration(mm, addr, current_node, | |
1833 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 | 1834 | |
d08221a0 | 1835 | if (err > 0) { |
e0153fc2 YS |
1836 | /* The page is successfully queued for migration */ |
1837 | continue; | |
1838 | } | |
3140a227 | 1839 | |
d08221a0 WY |
1840 | /* |
1841 | * If the page is already on the target node (!err), store the | |
1842 | * node, otherwise, store the err. | |
1843 | */ | |
1844 | err = store_status(status, i, err ? : current_node, 1); | |
a49bd4d7 MH |
1845 | if (err) |
1846 | goto out_flush; | |
5e9a0f02 | 1847 | |
7ca8783a WY |
1848 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
1849 | status, start, i, nr_pages); | |
4afdacec WY |
1850 | if (err) |
1851 | goto out; | |
a49bd4d7 | 1852 | current_node = NUMA_NO_NODE; |
3140a227 | 1853 | } |
a49bd4d7 MH |
1854 | out_flush: |
1855 | /* Make sure we do not overwrite the existing error */ | |
7ca8783a WY |
1856 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
1857 | status, start, i, nr_pages); | |
dfe9aa23 | 1858 | if (err >= 0) |
a49bd4d7 | 1859 | err = err1; |
5e9a0f02 | 1860 | out: |
361a2a22 | 1861 | lru_cache_enable(); |
5e9a0f02 BG |
1862 | return err; |
1863 | } | |
1864 | ||
742755a1 | 1865 | /* |
2f007e74 | 1866 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1867 | */ |
80bba129 BG |
1868 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1869 | const void __user **pages, int *status) | |
742755a1 | 1870 | { |
2f007e74 | 1871 | unsigned long i; |
2f007e74 | 1872 | |
d8ed45c5 | 1873 | mmap_read_lock(mm); |
742755a1 | 1874 | |
2f007e74 | 1875 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1876 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1877 | struct vm_area_struct *vma; |
1878 | struct page *page; | |
c095adbc | 1879 | int err = -EFAULT; |
2f007e74 | 1880 | |
059b8b48 LH |
1881 | vma = vma_lookup(mm, addr); |
1882 | if (!vma) | |
742755a1 CL |
1883 | goto set_status; |
1884 | ||
d899844e KS |
1885 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1886 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1887 | |
1888 | err = PTR_ERR(page); | |
1889 | if (IS_ERR(page)) | |
1890 | goto set_status; | |
1891 | ||
d899844e | 1892 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1893 | set_status: |
80bba129 BG |
1894 | *status = err; |
1895 | ||
1896 | pages++; | |
1897 | status++; | |
1898 | } | |
1899 | ||
d8ed45c5 | 1900 | mmap_read_unlock(mm); |
80bba129 BG |
1901 | } |
1902 | ||
5b1b561b AB |
1903 | static int get_compat_pages_array(const void __user *chunk_pages[], |
1904 | const void __user * __user *pages, | |
1905 | unsigned long chunk_nr) | |
1906 | { | |
1907 | compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; | |
1908 | compat_uptr_t p; | |
1909 | int i; | |
1910 | ||
1911 | for (i = 0; i < chunk_nr; i++) { | |
1912 | if (get_user(p, pages32 + i)) | |
1913 | return -EFAULT; | |
1914 | chunk_pages[i] = compat_ptr(p); | |
1915 | } | |
1916 | ||
1917 | return 0; | |
1918 | } | |
1919 | ||
80bba129 BG |
1920 | /* |
1921 | * Determine the nodes of a user array of pages and store it in | |
1922 | * a user array of status. | |
1923 | */ | |
1924 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1925 | const void __user * __user *pages, | |
1926 | int __user *status) | |
1927 | { | |
1928 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1929 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1930 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1931 | |
87b8d1ad PA |
1932 | while (nr_pages) { |
1933 | unsigned long chunk_nr; | |
80bba129 | 1934 | |
87b8d1ad PA |
1935 | chunk_nr = nr_pages; |
1936 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1937 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1938 | ||
5b1b561b AB |
1939 | if (in_compat_syscall()) { |
1940 | if (get_compat_pages_array(chunk_pages, pages, | |
1941 | chunk_nr)) | |
1942 | break; | |
1943 | } else { | |
1944 | if (copy_from_user(chunk_pages, pages, | |
1945 | chunk_nr * sizeof(*chunk_pages))) | |
1946 | break; | |
1947 | } | |
80bba129 BG |
1948 | |
1949 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1950 | ||
87b8d1ad PA |
1951 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1952 | break; | |
742755a1 | 1953 | |
87b8d1ad PA |
1954 | pages += chunk_nr; |
1955 | status += chunk_nr; | |
1956 | nr_pages -= chunk_nr; | |
1957 | } | |
1958 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1959 | } |
1960 | ||
4dc200ce | 1961 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
742755a1 | 1962 | { |
742755a1 | 1963 | struct task_struct *task; |
742755a1 | 1964 | struct mm_struct *mm; |
742755a1 | 1965 | |
4dc200ce ML |
1966 | /* |
1967 | * There is no need to check if current process has the right to modify | |
1968 | * the specified process when they are same. | |
1969 | */ | |
1970 | if (!pid) { | |
1971 | mmget(current->mm); | |
1972 | *mem_nodes = cpuset_mems_allowed(current); | |
1973 | return current->mm; | |
1974 | } | |
742755a1 CL |
1975 | |
1976 | /* Find the mm_struct */ | |
a879bf58 | 1977 | rcu_read_lock(); |
4dc200ce | 1978 | task = find_task_by_vpid(pid); |
742755a1 | 1979 | if (!task) { |
a879bf58 | 1980 | rcu_read_unlock(); |
4dc200ce | 1981 | return ERR_PTR(-ESRCH); |
742755a1 | 1982 | } |
3268c63e | 1983 | get_task_struct(task); |
742755a1 CL |
1984 | |
1985 | /* | |
1986 | * Check if this process has the right to modify the specified | |
197e7e52 | 1987 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1988 | */ |
197e7e52 | 1989 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1990 | rcu_read_unlock(); |
4dc200ce | 1991 | mm = ERR_PTR(-EPERM); |
5e9a0f02 | 1992 | goto out; |
742755a1 | 1993 | } |
c69e8d9c | 1994 | rcu_read_unlock(); |
742755a1 | 1995 | |
4dc200ce ML |
1996 | mm = ERR_PTR(security_task_movememory(task)); |
1997 | if (IS_ERR(mm)) | |
5e9a0f02 | 1998 | goto out; |
4dc200ce | 1999 | *mem_nodes = cpuset_mems_allowed(task); |
3268c63e | 2000 | mm = get_task_mm(task); |
4dc200ce | 2001 | out: |
3268c63e | 2002 | put_task_struct(task); |
6e8b09ea | 2003 | if (!mm) |
4dc200ce ML |
2004 | mm = ERR_PTR(-EINVAL); |
2005 | return mm; | |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * Move a list of pages in the address space of the currently executing | |
2010 | * process. | |
2011 | */ | |
2012 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
2013 | const void __user * __user *pages, | |
2014 | const int __user *nodes, | |
2015 | int __user *status, int flags) | |
2016 | { | |
2017 | struct mm_struct *mm; | |
2018 | int err; | |
2019 | nodemask_t task_nodes; | |
2020 | ||
2021 | /* Check flags */ | |
2022 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
6e8b09ea SL |
2023 | return -EINVAL; |
2024 | ||
4dc200ce ML |
2025 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
2026 | return -EPERM; | |
2027 | ||
2028 | mm = find_mm_struct(pid, &task_nodes); | |
2029 | if (IS_ERR(mm)) | |
2030 | return PTR_ERR(mm); | |
2031 | ||
6e8b09ea SL |
2032 | if (nodes) |
2033 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
2034 | nodes, status, flags); | |
2035 | else | |
2036 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 2037 | |
742755a1 CL |
2038 | mmput(mm); |
2039 | return err; | |
2040 | } | |
742755a1 | 2041 | |
7addf443 DB |
2042 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
2043 | const void __user * __user *, pages, | |
2044 | const int __user *, nodes, | |
2045 | int __user *, status, int, flags) | |
2046 | { | |
2047 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
2048 | } | |
2049 | ||
7039e1db PZ |
2050 | #ifdef CONFIG_NUMA_BALANCING |
2051 | /* | |
2052 | * Returns true if this is a safe migration target node for misplaced NUMA | |
2053 | * pages. Currently it only checks the watermarks which crude | |
2054 | */ | |
2055 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 2056 | unsigned long nr_migrate_pages) |
7039e1db PZ |
2057 | { |
2058 | int z; | |
599d0c95 | 2059 | |
7039e1db PZ |
2060 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
2061 | struct zone *zone = pgdat->node_zones + z; | |
2062 | ||
2063 | if (!populated_zone(zone)) | |
2064 | continue; | |
2065 | ||
7039e1db PZ |
2066 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
2067 | if (!zone_watermark_ok(zone, 0, | |
2068 | high_wmark_pages(zone) + | |
2069 | nr_migrate_pages, | |
bfe9d006 | 2070 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
2071 | continue; |
2072 | return true; | |
2073 | } | |
2074 | return false; | |
2075 | } | |
2076 | ||
2077 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 2078 | unsigned long data) |
7039e1db PZ |
2079 | { |
2080 | int nid = (int) data; | |
2081 | struct page *newpage; | |
2082 | ||
96db800f | 2083 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
2084 | (GFP_HIGHUSER_MOVABLE | |
2085 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
2086 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 2087 | ~__GFP_RECLAIM, 0); |
bac0382c | 2088 | |
7039e1db PZ |
2089 | return newpage; |
2090 | } | |
2091 | ||
c5b5a3dd YS |
2092 | static struct page *alloc_misplaced_dst_page_thp(struct page *page, |
2093 | unsigned long data) | |
2094 | { | |
2095 | int nid = (int) data; | |
2096 | struct page *newpage; | |
2097 | ||
2098 | newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), | |
2099 | HPAGE_PMD_ORDER); | |
2100 | if (!newpage) | |
2101 | goto out; | |
2102 | ||
2103 | prep_transhuge_page(newpage); | |
2104 | ||
2105 | out: | |
2106 | return newpage; | |
2107 | } | |
2108 | ||
1c30e017 | 2109 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 2110 | { |
340ef390 | 2111 | int page_lru; |
2b9b624f | 2112 | int nr_pages = thp_nr_pages(page); |
a8f60772 | 2113 | |
309381fe | 2114 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 2115 | |
662aeea7 YS |
2116 | /* Do not migrate THP mapped by multiple processes */ |
2117 | if (PageTransHuge(page) && total_mapcount(page) > 1) | |
2118 | return 0; | |
2119 | ||
7039e1db | 2120 | /* Avoid migrating to a node that is nearly full */ |
2b9b624f | 2121 | if (!migrate_balanced_pgdat(pgdat, nr_pages)) |
340ef390 | 2122 | return 0; |
7039e1db | 2123 | |
340ef390 HD |
2124 | if (isolate_lru_page(page)) |
2125 | return 0; | |
7039e1db | 2126 | |
9de4f22a | 2127 | page_lru = page_is_file_lru(page); |
599d0c95 | 2128 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
2b9b624f | 2129 | nr_pages); |
340ef390 | 2130 | |
149c33e1 | 2131 | /* |
340ef390 HD |
2132 | * Isolating the page has taken another reference, so the |
2133 | * caller's reference can be safely dropped without the page | |
2134 | * disappearing underneath us during migration. | |
149c33e1 MG |
2135 | */ |
2136 | put_page(page); | |
340ef390 | 2137 | return 1; |
b32967ff MG |
2138 | } |
2139 | ||
2140 | /* | |
2141 | * Attempt to migrate a misplaced page to the specified destination | |
2142 | * node. Caller is expected to have an elevated reference count on | |
2143 | * the page that will be dropped by this function before returning. | |
2144 | */ | |
1bc115d8 MG |
2145 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
2146 | int node) | |
b32967ff MG |
2147 | { |
2148 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 2149 | int isolated; |
b32967ff MG |
2150 | int nr_remaining; |
2151 | LIST_HEAD(migratepages); | |
c5b5a3dd YS |
2152 | new_page_t *new; |
2153 | bool compound; | |
b5916c02 | 2154 | int nr_pages = thp_nr_pages(page); |
c5b5a3dd YS |
2155 | |
2156 | /* | |
2157 | * PTE mapped THP or HugeTLB page can't reach here so the page could | |
2158 | * be either base page or THP. And it must be head page if it is | |
2159 | * THP. | |
2160 | */ | |
2161 | compound = PageTransHuge(page); | |
2162 | ||
2163 | if (compound) | |
2164 | new = alloc_misplaced_dst_page_thp; | |
2165 | else | |
2166 | new = alloc_misplaced_dst_page; | |
b32967ff MG |
2167 | |
2168 | /* | |
1bc115d8 MG |
2169 | * Don't migrate file pages that are mapped in multiple processes |
2170 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 2171 | */ |
7ee820ee ML |
2172 | if (page_mapcount(page) != 1 && page_is_file_lru(page) && |
2173 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 2174 | goto out; |
b32967ff | 2175 | |
09a913a7 MG |
2176 | /* |
2177 | * Also do not migrate dirty pages as not all filesystems can move | |
2178 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2179 | */ | |
9de4f22a | 2180 | if (page_is_file_lru(page) && PageDirty(page)) |
09a913a7 MG |
2181 | goto out; |
2182 | ||
b32967ff MG |
2183 | isolated = numamigrate_isolate_page(pgdat, page); |
2184 | if (!isolated) | |
2185 | goto out; | |
2186 | ||
2187 | list_add(&page->lru, &migratepages); | |
c5b5a3dd | 2188 | nr_remaining = migrate_pages(&migratepages, *new, NULL, node, |
5ac95884 | 2189 | MIGRATE_ASYNC, MR_NUMA_MISPLACED, NULL); |
b32967ff | 2190 | if (nr_remaining) { |
59c82b70 JK |
2191 | if (!list_empty(&migratepages)) { |
2192 | list_del(&page->lru); | |
c5fc5c3a YS |
2193 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
2194 | page_is_file_lru(page), -nr_pages); | |
59c82b70 JK |
2195 | putback_lru_page(page); |
2196 | } | |
b32967ff MG |
2197 | isolated = 0; |
2198 | } else | |
c5fc5c3a | 2199 | count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages); |
7039e1db | 2200 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2201 | return isolated; |
340ef390 HD |
2202 | |
2203 | out: | |
2204 | put_page(page); | |
2205 | return 0; | |
7039e1db | 2206 | } |
220018d3 | 2207 | #endif /* CONFIG_NUMA_BALANCING */ |
7039e1db | 2208 | #endif /* CONFIG_NUMA */ |
8763cb45 | 2209 | |
9b2ed9cb | 2210 | #ifdef CONFIG_DEVICE_PRIVATE |
843e1be1 | 2211 | static int migrate_vma_collect_skip(unsigned long start, |
8763cb45 JG |
2212 | unsigned long end, |
2213 | struct mm_walk *walk) | |
2214 | { | |
2215 | struct migrate_vma *migrate = walk->private; | |
2216 | unsigned long addr; | |
2217 | ||
872ea707 | 2218 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
8315ada7 | 2219 | migrate->dst[migrate->npages] = 0; |
843e1be1 | 2220 | migrate->src[migrate->npages++] = 0; |
8315ada7 JG |
2221 | } |
2222 | ||
2223 | return 0; | |
2224 | } | |
2225 | ||
843e1be1 | 2226 | static int migrate_vma_collect_hole(unsigned long start, |
8315ada7 | 2227 | unsigned long end, |
843e1be1 | 2228 | __always_unused int depth, |
8315ada7 JG |
2229 | struct mm_walk *walk) |
2230 | { | |
2231 | struct migrate_vma *migrate = walk->private; | |
2232 | unsigned long addr; | |
2233 | ||
843e1be1 ML |
2234 | /* Only allow populating anonymous memory. */ |
2235 | if (!vma_is_anonymous(walk->vma)) | |
2236 | return migrate_vma_collect_skip(start, end, walk); | |
2237 | ||
872ea707 | 2238 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
843e1be1 | 2239 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8763cb45 | 2240 | migrate->dst[migrate->npages] = 0; |
843e1be1 ML |
2241 | migrate->npages++; |
2242 | migrate->cpages++; | |
8763cb45 JG |
2243 | } |
2244 | ||
2245 | return 0; | |
2246 | } | |
2247 | ||
2248 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2249 | unsigned long start, | |
2250 | unsigned long end, | |
2251 | struct mm_walk *walk) | |
2252 | { | |
2253 | struct migrate_vma *migrate = walk->private; | |
2254 | struct vm_area_struct *vma = walk->vma; | |
2255 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2256 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2257 | spinlock_t *ptl; |
2258 | pte_t *ptep; | |
2259 | ||
2260 | again: | |
2261 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2262 | return migrate_vma_collect_hole(start, end, -1, walk); |
8763cb45 JG |
2263 | |
2264 | if (pmd_trans_huge(*pmdp)) { | |
2265 | struct page *page; | |
2266 | ||
2267 | ptl = pmd_lock(mm, pmdp); | |
2268 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2269 | spin_unlock(ptl); | |
2270 | goto again; | |
2271 | } | |
2272 | ||
2273 | page = pmd_page(*pmdp); | |
2274 | if (is_huge_zero_page(page)) { | |
2275 | spin_unlock(ptl); | |
2276 | split_huge_pmd(vma, pmdp, addr); | |
2277 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2278 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2279 | walk); |
2280 | } else { | |
2281 | int ret; | |
2282 | ||
2283 | get_page(page); | |
2284 | spin_unlock(ptl); | |
2285 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2286 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2287 | walk); |
2288 | ret = split_huge_page(page); | |
2289 | unlock_page(page); | |
2290 | put_page(page); | |
8315ada7 JG |
2291 | if (ret) |
2292 | return migrate_vma_collect_skip(start, end, | |
2293 | walk); | |
2294 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2295 | return migrate_vma_collect_hole(start, end, -1, |
8763cb45 JG |
2296 | walk); |
2297 | } | |
2298 | } | |
2299 | ||
2300 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2301 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2302 | |
2303 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2304 | arch_enter_lazy_mmu_mode(); |
2305 | ||
8763cb45 | 2306 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
800bb1c8 | 2307 | unsigned long mpfn = 0, pfn; |
8763cb45 | 2308 | struct page *page; |
8c3328f1 | 2309 | swp_entry_t entry; |
8763cb45 JG |
2310 | pte_t pte; |
2311 | ||
2312 | pte = *ptep; | |
8763cb45 | 2313 | |
a5430dda | 2314 | if (pte_none(pte)) { |
0744f280 RC |
2315 | if (vma_is_anonymous(vma)) { |
2316 | mpfn = MIGRATE_PFN_MIGRATE; | |
2317 | migrate->cpages++; | |
2318 | } | |
8763cb45 JG |
2319 | goto next; |
2320 | } | |
2321 | ||
a5430dda | 2322 | if (!pte_present(pte)) { |
a5430dda JG |
2323 | /* |
2324 | * Only care about unaddressable device page special | |
2325 | * page table entry. Other special swap entries are not | |
2326 | * migratable, and we ignore regular swapped page. | |
2327 | */ | |
2328 | entry = pte_to_swp_entry(pte); | |
2329 | if (!is_device_private_entry(entry)) | |
2330 | goto next; | |
2331 | ||
af5cdaf8 | 2332 | page = pfn_swap_entry_to_page(entry); |
5143192c RC |
2333 | if (!(migrate->flags & |
2334 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
2335 | page->pgmap->owner != migrate->pgmap_owner) | |
800bb1c8 CH |
2336 | goto next; |
2337 | ||
06d462be CH |
2338 | mpfn = migrate_pfn(page_to_pfn(page)) | |
2339 | MIGRATE_PFN_MIGRATE; | |
4dd845b5 | 2340 | if (is_writable_device_private_entry(entry)) |
a5430dda JG |
2341 | mpfn |= MIGRATE_PFN_WRITE; |
2342 | } else { | |
5143192c | 2343 | if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) |
800bb1c8 | 2344 | goto next; |
276f756d | 2345 | pfn = pte_pfn(pte); |
8315ada7 JG |
2346 | if (is_zero_pfn(pfn)) { |
2347 | mpfn = MIGRATE_PFN_MIGRATE; | |
2348 | migrate->cpages++; | |
8315ada7 JG |
2349 | goto next; |
2350 | } | |
25b2995a | 2351 | page = vm_normal_page(migrate->vma, addr, pte); |
a5430dda JG |
2352 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2353 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2354 | } | |
2355 | ||
8763cb45 | 2356 | /* FIXME support THP */ |
8763cb45 | 2357 | if (!page || !page->mapping || PageTransCompound(page)) { |
276f756d | 2358 | mpfn = 0; |
8763cb45 JG |
2359 | goto next; |
2360 | } | |
2361 | ||
2362 | /* | |
2363 | * By getting a reference on the page we pin it and that blocks | |
2364 | * any kind of migration. Side effect is that it "freezes" the | |
2365 | * pte. | |
2366 | * | |
2367 | * We drop this reference after isolating the page from the lru | |
2368 | * for non device page (device page are not on the lru and thus | |
2369 | * can't be dropped from it). | |
2370 | */ | |
2371 | get_page(page); | |
2372 | migrate->cpages++; | |
8763cb45 | 2373 | |
8c3328f1 JG |
2374 | /* |
2375 | * Optimize for the common case where page is only mapped once | |
2376 | * in one process. If we can lock the page, then we can safely | |
2377 | * set up a special migration page table entry now. | |
2378 | */ | |
2379 | if (trylock_page(page)) { | |
2380 | pte_t swp_pte; | |
2381 | ||
2382 | mpfn |= MIGRATE_PFN_LOCKED; | |
2383 | ptep_get_and_clear(mm, addr, ptep); | |
2384 | ||
2385 | /* Setup special migration page table entry */ | |
4dd845b5 AP |
2386 | if (mpfn & MIGRATE_PFN_WRITE) |
2387 | entry = make_writable_migration_entry( | |
2388 | page_to_pfn(page)); | |
2389 | else | |
2390 | entry = make_readable_migration_entry( | |
2391 | page_to_pfn(page)); | |
8c3328f1 | 2392 | swp_pte = swp_entry_to_pte(entry); |
ad7df764 AP |
2393 | if (pte_present(pte)) { |
2394 | if (pte_soft_dirty(pte)) | |
2395 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2396 | if (pte_uffd_wp(pte)) | |
2397 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2398 | } else { | |
2399 | if (pte_swp_soft_dirty(pte)) | |
2400 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2401 | if (pte_swp_uffd_wp(pte)) | |
2402 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2403 | } | |
8c3328f1 JG |
2404 | set_pte_at(mm, addr, ptep, swp_pte); |
2405 | ||
2406 | /* | |
2407 | * This is like regular unmap: we remove the rmap and | |
2408 | * drop page refcount. Page won't be freed, as we took | |
2409 | * a reference just above. | |
2410 | */ | |
2411 | page_remove_rmap(page, false); | |
2412 | put_page(page); | |
a5430dda JG |
2413 | |
2414 | if (pte_present(pte)) | |
2415 | unmapped++; | |
8c3328f1 JG |
2416 | } |
2417 | ||
8763cb45 | 2418 | next: |
a5430dda | 2419 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2420 | migrate->src[migrate->npages++] = mpfn; |
2421 | } | |
8c3328f1 | 2422 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2423 | pte_unmap_unlock(ptep - 1, ptl); |
2424 | ||
8c3328f1 JG |
2425 | /* Only flush the TLB if we actually modified any entries */ |
2426 | if (unmapped) | |
2427 | flush_tlb_range(walk->vma, start, end); | |
2428 | ||
8763cb45 JG |
2429 | return 0; |
2430 | } | |
2431 | ||
7b86ac33 CH |
2432 | static const struct mm_walk_ops migrate_vma_walk_ops = { |
2433 | .pmd_entry = migrate_vma_collect_pmd, | |
2434 | .pte_hole = migrate_vma_collect_hole, | |
2435 | }; | |
2436 | ||
8763cb45 JG |
2437 | /* |
2438 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2439 | * @migrate: migrate struct containing all migration information | |
2440 | * | |
2441 | * This will walk the CPU page table. For each virtual address backed by a | |
2442 | * valid page, it updates the src array and takes a reference on the page, in | |
2443 | * order to pin the page until we lock it and unmap it. | |
2444 | */ | |
2445 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2446 | { | |
ac46d4f3 | 2447 | struct mmu_notifier_range range; |
8763cb45 | 2448 | |
998427b3 RC |
2449 | /* |
2450 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
2451 | * that the registered device driver can skip invalidating device | |
2452 | * private page mappings that won't be migrated. | |
2453 | */ | |
6b49bf6d AP |
2454 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0, |
2455 | migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end, | |
c1a06df6 | 2456 | migrate->pgmap_owner); |
ac46d4f3 | 2457 | mmu_notifier_invalidate_range_start(&range); |
8763cb45 | 2458 | |
7b86ac33 CH |
2459 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, |
2460 | &migrate_vma_walk_ops, migrate); | |
2461 | ||
2462 | mmu_notifier_invalidate_range_end(&range); | |
8763cb45 JG |
2463 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2464 | } | |
2465 | ||
2466 | /* | |
2467 | * migrate_vma_check_page() - check if page is pinned or not | |
2468 | * @page: struct page to check | |
2469 | * | |
2470 | * Pinned pages cannot be migrated. This is the same test as in | |
2471 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2472 | * ZONE_DEVICE page. | |
2473 | */ | |
2474 | static bool migrate_vma_check_page(struct page *page) | |
2475 | { | |
2476 | /* | |
2477 | * One extra ref because caller holds an extra reference, either from | |
2478 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2479 | * a device page. | |
2480 | */ | |
2481 | int extra = 1; | |
2482 | ||
2483 | /* | |
2484 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2485 | * check them than regular pages, because they can be mapped with a pmd | |
2486 | * or with a pte (split pte mapping). | |
2487 | */ | |
2488 | if (PageCompound(page)) | |
2489 | return false; | |
2490 | ||
a5430dda JG |
2491 | /* Page from ZONE_DEVICE have one extra reference */ |
2492 | if (is_zone_device_page(page)) { | |
2493 | /* | |
2494 | * Private page can never be pin as they have no valid pte and | |
2495 | * GUP will fail for those. Yet if there is a pending migration | |
2496 | * a thread might try to wait on the pte migration entry and | |
2497 | * will bump the page reference count. Sadly there is no way to | |
2498 | * differentiate a regular pin from migration wait. Hence to | |
2499 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
8958b249 | 2500 | * infinite loop (one stopping migration because the other is |
a5430dda JG |
2501 | * waiting on pte migration entry). We always return true here. |
2502 | * | |
2503 | * FIXME proper solution is to rework migration_entry_wait() so | |
2504 | * it does not need to take a reference on page. | |
2505 | */ | |
25b2995a | 2506 | return is_device_private_page(page); |
a5430dda JG |
2507 | } |
2508 | ||
df6ad698 JG |
2509 | /* For file back page */ |
2510 | if (page_mapping(page)) | |
2511 | extra += 1 + page_has_private(page); | |
2512 | ||
8763cb45 JG |
2513 | if ((page_count(page) - extra) > page_mapcount(page)) |
2514 | return false; | |
2515 | ||
2516 | return true; | |
2517 | } | |
2518 | ||
2519 | /* | |
2520 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2521 | * @migrate: migrate struct containing all migration information | |
2522 | * | |
2523 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2524 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2525 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2526 | * migrated by concurrent kernel threads. | |
2527 | */ | |
2528 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2529 | { | |
2530 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2531 | const unsigned long start = migrate->start; |
2532 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2533 | bool allow_drain = true; |
8763cb45 JG |
2534 | |
2535 | lru_add_drain(); | |
2536 | ||
2537 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2538 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2539 | bool remap = true; |
8763cb45 JG |
2540 | |
2541 | if (!page) | |
2542 | continue; | |
2543 | ||
8c3328f1 JG |
2544 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2545 | /* | |
2546 | * Because we are migrating several pages there can be | |
2547 | * a deadlock between 2 concurrent migration where each | |
2548 | * are waiting on each other page lock. | |
2549 | * | |
2550 | * Make migrate_vma() a best effort thing and backoff | |
2551 | * for any page we can not lock right away. | |
2552 | */ | |
2553 | if (!trylock_page(page)) { | |
2554 | migrate->src[i] = 0; | |
2555 | migrate->cpages--; | |
2556 | put_page(page); | |
2557 | continue; | |
2558 | } | |
2559 | remap = false; | |
2560 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2561 | } |
8763cb45 | 2562 | |
a5430dda JG |
2563 | /* ZONE_DEVICE pages are not on LRU */ |
2564 | if (!is_zone_device_page(page)) { | |
2565 | if (!PageLRU(page) && allow_drain) { | |
2566 | /* Drain CPU's pagevec */ | |
2567 | lru_add_drain_all(); | |
2568 | allow_drain = false; | |
2569 | } | |
8763cb45 | 2570 | |
a5430dda JG |
2571 | if (isolate_lru_page(page)) { |
2572 | if (remap) { | |
2573 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2574 | migrate->cpages--; | |
2575 | restore++; | |
2576 | } else { | |
2577 | migrate->src[i] = 0; | |
2578 | unlock_page(page); | |
2579 | migrate->cpages--; | |
2580 | put_page(page); | |
2581 | } | |
2582 | continue; | |
8c3328f1 | 2583 | } |
a5430dda JG |
2584 | |
2585 | /* Drop the reference we took in collect */ | |
2586 | put_page(page); | |
8763cb45 JG |
2587 | } |
2588 | ||
2589 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2590 | if (remap) { |
2591 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2592 | migrate->cpages--; | |
2593 | restore++; | |
8763cb45 | 2594 | |
a5430dda JG |
2595 | if (!is_zone_device_page(page)) { |
2596 | get_page(page); | |
2597 | putback_lru_page(page); | |
2598 | } | |
8c3328f1 JG |
2599 | } else { |
2600 | migrate->src[i] = 0; | |
2601 | unlock_page(page); | |
2602 | migrate->cpages--; | |
2603 | ||
a5430dda JG |
2604 | if (!is_zone_device_page(page)) |
2605 | putback_lru_page(page); | |
2606 | else | |
2607 | put_page(page); | |
8c3328f1 | 2608 | } |
8763cb45 JG |
2609 | } |
2610 | } | |
8c3328f1 JG |
2611 | |
2612 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2613 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2614 | ||
2615 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2616 | continue; | |
2617 | ||
2618 | remove_migration_pte(page, migrate->vma, addr, page); | |
2619 | ||
2620 | migrate->src[i] = 0; | |
2621 | unlock_page(page); | |
2622 | put_page(page); | |
2623 | restore--; | |
2624 | } | |
8763cb45 JG |
2625 | } |
2626 | ||
2627 | /* | |
2628 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2629 | * @migrate: migrate struct containing all migration information | |
2630 | * | |
2631 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2632 | * and check again if it has been pinned. Pinned pages are restored because we | |
2633 | * cannot migrate them. | |
2634 | * | |
2635 | * This is the last step before we call the device driver callback to allocate | |
2636 | * destination memory and copy contents of original page over to new page. | |
2637 | */ | |
2638 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2639 | { | |
8763cb45 JG |
2640 | const unsigned long npages = migrate->npages; |
2641 | const unsigned long start = migrate->start; | |
2642 | unsigned long addr, i, restore = 0; | |
2643 | ||
2644 | for (i = 0; i < npages; i++) { | |
2645 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2646 | ||
2647 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2648 | continue; | |
2649 | ||
8c3328f1 | 2650 | if (page_mapped(page)) { |
a98a2f0c | 2651 | try_to_migrate(page, 0); |
8c3328f1 JG |
2652 | if (page_mapped(page)) |
2653 | goto restore; | |
8763cb45 | 2654 | } |
8c3328f1 JG |
2655 | |
2656 | if (migrate_vma_check_page(page)) | |
2657 | continue; | |
2658 | ||
2659 | restore: | |
2660 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2661 | migrate->cpages--; | |
2662 | restore++; | |
8763cb45 JG |
2663 | } |
2664 | ||
2665 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2666 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2667 | ||
2668 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2669 | continue; | |
2670 | ||
2671 | remove_migration_ptes(page, page, false); | |
2672 | ||
2673 | migrate->src[i] = 0; | |
2674 | unlock_page(page); | |
2675 | restore--; | |
2676 | ||
a5430dda JG |
2677 | if (is_zone_device_page(page)) |
2678 | put_page(page); | |
2679 | else | |
2680 | putback_lru_page(page); | |
8763cb45 JG |
2681 | } |
2682 | } | |
2683 | ||
a7d1f22b CH |
2684 | /** |
2685 | * migrate_vma_setup() - prepare to migrate a range of memory | |
eaf444de | 2686 | * @args: contains the vma, start, and pfns arrays for the migration |
a7d1f22b CH |
2687 | * |
2688 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2689 | * without an error. | |
2690 | * | |
2691 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2692 | * the pages backing each virtual address in the range, saving them inside the | |
2693 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2694 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2695 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2696 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2697 | * remapping and unlocking those pages. | |
2698 | * | |
2699 | * The caller should then allocate destination memory and copy source memory to | |
2700 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2701 | * flag set). Once these are allocated and copied, the caller must update each | |
2702 | * corresponding entry in the dst array with the pfn value of the destination | |
2703 | * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set | |
2704 | * (destination pages must have their struct pages locked, via lock_page()). | |
2705 | * | |
2706 | * Note that the caller does not have to migrate all the pages that are marked | |
2707 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2708 | * device memory to system memory. If the caller cannot migrate a device page | |
2709 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2710 | * consequences for the userspace process, so it must be avoided if at all | |
2711 | * possible. | |
2712 | * | |
2713 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2714 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
f0953a1b IM |
2715 | * allowing the caller to allocate device memory for those unbacked virtual |
2716 | * addresses. For this the caller simply has to allocate device memory and | |
a7d1f22b | 2717 | * properly set the destination entry like for regular migration. Note that |
f0953a1b IM |
2718 | * this can still fail, and thus inside the device driver you must check if the |
2719 | * migration was successful for those entries after calling migrate_vma_pages(), | |
a7d1f22b CH |
2720 | * just like for regular migration. |
2721 | * | |
2722 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2723 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2724 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2725 | * then migrate_vma_pages() to migrate struct page information from the source | |
2726 | * struct page to the destination struct page. If it fails to migrate the | |
2727 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2728 | * src array. | |
2729 | * | |
2730 | * At this point all successfully migrated pages have an entry in the src | |
2731 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2732 | * array entry with MIGRATE_PFN_VALID flag set. | |
2733 | * | |
2734 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2735 | * successfully migrated, and which were not. Successfully migrated pages will | |
2736 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2737 | * | |
2738 | * It is safe to update device page table after migrate_vma_pages() because | |
c1e8d7c6 | 2739 | * both destination and source page are still locked, and the mmap_lock is held |
a7d1f22b CH |
2740 | * in read mode (hence no one can unmap the range being migrated). |
2741 | * | |
2742 | * Once the caller is done cleaning up things and updating its page table (if it | |
2743 | * chose to do so, this is not an obligation) it finally calls | |
2744 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2745 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2746 | * point to the original source pages. | |
2747 | */ | |
2748 | int migrate_vma_setup(struct migrate_vma *args) | |
2749 | { | |
2750 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2751 | ||
2752 | args->start &= PAGE_MASK; | |
2753 | args->end &= PAGE_MASK; | |
2754 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2755 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2756 | return -EINVAL; | |
2757 | if (nr_pages <= 0) | |
2758 | return -EINVAL; | |
2759 | if (args->start < args->vma->vm_start || | |
2760 | args->start >= args->vma->vm_end) | |
2761 | return -EINVAL; | |
2762 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2763 | return -EINVAL; | |
2764 | if (!args->src || !args->dst) | |
2765 | return -EINVAL; | |
2766 | ||
2767 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2768 | args->cpages = 0; | |
2769 | args->npages = 0; | |
2770 | ||
2771 | migrate_vma_collect(args); | |
2772 | ||
2773 | if (args->cpages) | |
2774 | migrate_vma_prepare(args); | |
2775 | if (args->cpages) | |
2776 | migrate_vma_unmap(args); | |
2777 | ||
2778 | /* | |
2779 | * At this point pages are locked and unmapped, and thus they have | |
2780 | * stable content and can safely be copied to destination memory that | |
2781 | * is allocated by the drivers. | |
2782 | */ | |
2783 | return 0; | |
2784 | ||
2785 | } | |
2786 | EXPORT_SYMBOL(migrate_vma_setup); | |
2787 | ||
34290e2c RC |
2788 | /* |
2789 | * This code closely matches the code in: | |
2790 | * __handle_mm_fault() | |
2791 | * handle_pte_fault() | |
2792 | * do_anonymous_page() | |
2793 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
2794 | * private page. | |
2795 | */ | |
8315ada7 JG |
2796 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2797 | unsigned long addr, | |
2798 | struct page *page, | |
d85c6db4 | 2799 | unsigned long *src) |
8315ada7 JG |
2800 | { |
2801 | struct vm_area_struct *vma = migrate->vma; | |
2802 | struct mm_struct *mm = vma->vm_mm; | |
8315ada7 JG |
2803 | bool flush = false; |
2804 | spinlock_t *ptl; | |
2805 | pte_t entry; | |
2806 | pgd_t *pgdp; | |
2807 | p4d_t *p4dp; | |
2808 | pud_t *pudp; | |
2809 | pmd_t *pmdp; | |
2810 | pte_t *ptep; | |
2811 | ||
2812 | /* Only allow populating anonymous memory */ | |
2813 | if (!vma_is_anonymous(vma)) | |
2814 | goto abort; | |
2815 | ||
2816 | pgdp = pgd_offset(mm, addr); | |
2817 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2818 | if (!p4dp) | |
2819 | goto abort; | |
2820 | pudp = pud_alloc(mm, p4dp, addr); | |
2821 | if (!pudp) | |
2822 | goto abort; | |
2823 | pmdp = pmd_alloc(mm, pudp, addr); | |
2824 | if (!pmdp) | |
2825 | goto abort; | |
2826 | ||
2827 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2828 | goto abort; | |
2829 | ||
2830 | /* | |
2831 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2832 | * pte_offset_map() on pmds where a huge pmd might be created | |
2833 | * from a different thread. | |
2834 | * | |
3e4e28c5 | 2835 | * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when |
8315ada7 JG |
2836 | * parallel threads are excluded by other means. |
2837 | * | |
3e4e28c5 | 2838 | * Here we only have mmap_read_lock(mm). |
8315ada7 | 2839 | */ |
4cf58924 | 2840 | if (pte_alloc(mm, pmdp)) |
8315ada7 JG |
2841 | goto abort; |
2842 | ||
2843 | /* See the comment in pte_alloc_one_map() */ | |
2844 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2845 | goto abort; | |
2846 | ||
2847 | if (unlikely(anon_vma_prepare(vma))) | |
2848 | goto abort; | |
d9eb1ea2 | 2849 | if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) |
8315ada7 JG |
2850 | goto abort; |
2851 | ||
2852 | /* | |
2853 | * The memory barrier inside __SetPageUptodate makes sure that | |
2854 | * preceding stores to the page contents become visible before | |
2855 | * the set_pte_at() write. | |
2856 | */ | |
2857 | __SetPageUptodate(page); | |
2858 | ||
df6ad698 JG |
2859 | if (is_zone_device_page(page)) { |
2860 | if (is_device_private_page(page)) { | |
2861 | swp_entry_t swp_entry; | |
2862 | ||
4dd845b5 AP |
2863 | if (vma->vm_flags & VM_WRITE) |
2864 | swp_entry = make_writable_device_private_entry( | |
2865 | page_to_pfn(page)); | |
2866 | else | |
2867 | swp_entry = make_readable_device_private_entry( | |
2868 | page_to_pfn(page)); | |
df6ad698 | 2869 | entry = swp_entry_to_pte(swp_entry); |
34f5e9b9 ML |
2870 | } else { |
2871 | /* | |
2872 | * For now we only support migrating to un-addressable | |
2873 | * device memory. | |
2874 | */ | |
2875 | pr_warn_once("Unsupported ZONE_DEVICE page type.\n"); | |
2876 | goto abort; | |
df6ad698 | 2877 | } |
8315ada7 JG |
2878 | } else { |
2879 | entry = mk_pte(page, vma->vm_page_prot); | |
2880 | if (vma->vm_flags & VM_WRITE) | |
2881 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2882 | } | |
2883 | ||
2884 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2885 | ||
34290e2c RC |
2886 | if (check_stable_address_space(mm)) |
2887 | goto unlock_abort; | |
2888 | ||
8315ada7 JG |
2889 | if (pte_present(*ptep)) { |
2890 | unsigned long pfn = pte_pfn(*ptep); | |
2891 | ||
c23a0c99 RC |
2892 | if (!is_zero_pfn(pfn)) |
2893 | goto unlock_abort; | |
8315ada7 | 2894 | flush = true; |
c23a0c99 RC |
2895 | } else if (!pte_none(*ptep)) |
2896 | goto unlock_abort; | |
8315ada7 JG |
2897 | |
2898 | /* | |
c23a0c99 | 2899 | * Check for userfaultfd but do not deliver the fault. Instead, |
8315ada7 JG |
2900 | * just back off. |
2901 | */ | |
c23a0c99 RC |
2902 | if (userfaultfd_missing(vma)) |
2903 | goto unlock_abort; | |
8315ada7 JG |
2904 | |
2905 | inc_mm_counter(mm, MM_ANONPAGES); | |
be5d0a74 | 2906 | page_add_new_anon_rmap(page, vma, addr, false); |
8315ada7 | 2907 | if (!is_zone_device_page(page)) |
b518154e | 2908 | lru_cache_add_inactive_or_unevictable(page, vma); |
8315ada7 JG |
2909 | get_page(page); |
2910 | ||
2911 | if (flush) { | |
2912 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2913 | ptep_clear_flush_notify(vma, addr, ptep); | |
2914 | set_pte_at_notify(mm, addr, ptep, entry); | |
2915 | update_mmu_cache(vma, addr, ptep); | |
2916 | } else { | |
2917 | /* No need to invalidate - it was non-present before */ | |
2918 | set_pte_at(mm, addr, ptep, entry); | |
2919 | update_mmu_cache(vma, addr, ptep); | |
2920 | } | |
2921 | ||
2922 | pte_unmap_unlock(ptep, ptl); | |
2923 | *src = MIGRATE_PFN_MIGRATE; | |
2924 | return; | |
2925 | ||
c23a0c99 RC |
2926 | unlock_abort: |
2927 | pte_unmap_unlock(ptep, ptl); | |
8315ada7 JG |
2928 | abort: |
2929 | *src &= ~MIGRATE_PFN_MIGRATE; | |
2930 | } | |
2931 | ||
a7d1f22b | 2932 | /** |
8763cb45 JG |
2933 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
2934 | * @migrate: migrate struct containing all migration information | |
2935 | * | |
2936 | * This migrates struct page meta-data from source struct page to destination | |
2937 | * struct page. This effectively finishes the migration from source page to the | |
2938 | * destination page. | |
2939 | */ | |
a7d1f22b | 2940 | void migrate_vma_pages(struct migrate_vma *migrate) |
8763cb45 JG |
2941 | { |
2942 | const unsigned long npages = migrate->npages; | |
2943 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
2944 | struct mmu_notifier_range range; |
2945 | unsigned long addr, i; | |
8315ada7 | 2946 | bool notified = false; |
8763cb45 JG |
2947 | |
2948 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2949 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2950 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2951 | struct address_space *mapping; | |
2952 | int r; | |
2953 | ||
8315ada7 JG |
2954 | if (!newpage) { |
2955 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 2956 | continue; |
8315ada7 JG |
2957 | } |
2958 | ||
2959 | if (!page) { | |
c23a0c99 | 2960 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
8315ada7 | 2961 | continue; |
8315ada7 | 2962 | if (!notified) { |
8315ada7 | 2963 | notified = true; |
ac46d4f3 | 2964 | |
6b49bf6d AP |
2965 | mmu_notifier_range_init_owner(&range, |
2966 | MMU_NOTIFY_MIGRATE, 0, migrate->vma, | |
2967 | migrate->vma->vm_mm, addr, migrate->end, | |
5e5dda81 | 2968 | migrate->pgmap_owner); |
ac46d4f3 | 2969 | mmu_notifier_invalidate_range_start(&range); |
8315ada7 JG |
2970 | } |
2971 | migrate_vma_insert_page(migrate, addr, newpage, | |
d85c6db4 | 2972 | &migrate->src[i]); |
8763cb45 | 2973 | continue; |
8315ada7 | 2974 | } |
8763cb45 JG |
2975 | |
2976 | mapping = page_mapping(page); | |
2977 | ||
a5430dda JG |
2978 | if (is_zone_device_page(newpage)) { |
2979 | if (is_device_private_page(newpage)) { | |
2980 | /* | |
2981 | * For now only support private anonymous when | |
2982 | * migrating to un-addressable device memory. | |
2983 | */ | |
2984 | if (mapping) { | |
2985 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2986 | continue; | |
2987 | } | |
25b2995a | 2988 | } else { |
a5430dda JG |
2989 | /* |
2990 | * Other types of ZONE_DEVICE page are not | |
2991 | * supported. | |
2992 | */ | |
2993 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2994 | continue; | |
2995 | } | |
2996 | } | |
2997 | ||
8763cb45 JG |
2998 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2999 | if (r != MIGRATEPAGE_SUCCESS) | |
3000 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3001 | } | |
8315ada7 | 3002 | |
4645b9fe JG |
3003 | /* |
3004 | * No need to double call mmu_notifier->invalidate_range() callback as | |
3005 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
3006 | * did already call it. | |
3007 | */ | |
8315ada7 | 3008 | if (notified) |
ac46d4f3 | 3009 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 | 3010 | } |
a7d1f22b | 3011 | EXPORT_SYMBOL(migrate_vma_pages); |
8763cb45 | 3012 | |
a7d1f22b | 3013 | /** |
8763cb45 JG |
3014 | * migrate_vma_finalize() - restore CPU page table entry |
3015 | * @migrate: migrate struct containing all migration information | |
3016 | * | |
3017 | * This replaces the special migration pte entry with either a mapping to the | |
3018 | * new page if migration was successful for that page, or to the original page | |
3019 | * otherwise. | |
3020 | * | |
3021 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
3022 | * refcount, for device pages. | |
3023 | */ | |
a7d1f22b | 3024 | void migrate_vma_finalize(struct migrate_vma *migrate) |
8763cb45 JG |
3025 | { |
3026 | const unsigned long npages = migrate->npages; | |
3027 | unsigned long i; | |
3028 | ||
3029 | for (i = 0; i < npages; i++) { | |
3030 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
3031 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
3032 | ||
8315ada7 JG |
3033 | if (!page) { |
3034 | if (newpage) { | |
3035 | unlock_page(newpage); | |
3036 | put_page(newpage); | |
3037 | } | |
8763cb45 | 3038 | continue; |
8315ada7 JG |
3039 | } |
3040 | ||
8763cb45 JG |
3041 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
3042 | if (newpage) { | |
3043 | unlock_page(newpage); | |
3044 | put_page(newpage); | |
3045 | } | |
3046 | newpage = page; | |
3047 | } | |
3048 | ||
3049 | remove_migration_ptes(page, newpage, false); | |
3050 | unlock_page(page); | |
8763cb45 | 3051 | |
a5430dda JG |
3052 | if (is_zone_device_page(page)) |
3053 | put_page(page); | |
3054 | else | |
3055 | putback_lru_page(page); | |
8763cb45 JG |
3056 | |
3057 | if (newpage != page) { | |
3058 | unlock_page(newpage); | |
a5430dda JG |
3059 | if (is_zone_device_page(newpage)) |
3060 | put_page(newpage); | |
3061 | else | |
3062 | putback_lru_page(newpage); | |
8763cb45 JG |
3063 | } |
3064 | } | |
3065 | } | |
a7d1f22b | 3066 | EXPORT_SYMBOL(migrate_vma_finalize); |
9b2ed9cb | 3067 | #endif /* CONFIG_DEVICE_PRIVATE */ |
79c28a41 | 3068 | |
76af6a05 | 3069 | #if defined(CONFIG_HOTPLUG_CPU) |
79c28a41 DH |
3070 | /* Disable reclaim-based migration. */ |
3071 | static void __disable_all_migrate_targets(void) | |
3072 | { | |
3073 | int node; | |
3074 | ||
3075 | for_each_online_node(node) | |
3076 | node_demotion[node] = NUMA_NO_NODE; | |
3077 | } | |
3078 | ||
3079 | static void disable_all_migrate_targets(void) | |
3080 | { | |
3081 | __disable_all_migrate_targets(); | |
3082 | ||
3083 | /* | |
3084 | * Ensure that the "disable" is visible across the system. | |
3085 | * Readers will see either a combination of before+disable | |
3086 | * state or disable+after. They will never see before and | |
3087 | * after state together. | |
3088 | * | |
3089 | * The before+after state together might have cycles and | |
3090 | * could cause readers to do things like loop until this | |
3091 | * function finishes. This ensures they can only see a | |
3092 | * single "bad" read and would, for instance, only loop | |
3093 | * once. | |
3094 | */ | |
3095 | synchronize_rcu(); | |
3096 | } | |
3097 | ||
3098 | /* | |
3099 | * Find an automatic demotion target for 'node'. | |
3100 | * Failing here is OK. It might just indicate | |
3101 | * being at the end of a chain. | |
3102 | */ | |
3103 | static int establish_migrate_target(int node, nodemask_t *used) | |
3104 | { | |
3105 | int migration_target; | |
3106 | ||
3107 | /* | |
3108 | * Can not set a migration target on a | |
3109 | * node with it already set. | |
3110 | * | |
3111 | * No need for READ_ONCE() here since this | |
3112 | * in the write path for node_demotion[]. | |
3113 | * This should be the only thread writing. | |
3114 | */ | |
3115 | if (node_demotion[node] != NUMA_NO_NODE) | |
3116 | return NUMA_NO_NODE; | |
3117 | ||
3118 | migration_target = find_next_best_node(node, used); | |
3119 | if (migration_target == NUMA_NO_NODE) | |
3120 | return NUMA_NO_NODE; | |
3121 | ||
3122 | node_demotion[node] = migration_target; | |
3123 | ||
3124 | return migration_target; | |
3125 | } | |
3126 | ||
3127 | /* | |
3128 | * When memory fills up on a node, memory contents can be | |
3129 | * automatically migrated to another node instead of | |
3130 | * discarded at reclaim. | |
3131 | * | |
3132 | * Establish a "migration path" which will start at nodes | |
3133 | * with CPUs and will follow the priorities used to build the | |
3134 | * page allocator zonelists. | |
3135 | * | |
3136 | * The difference here is that cycles must be avoided. If | |
3137 | * node0 migrates to node1, then neither node1, nor anything | |
3138 | * node1 migrates to can migrate to node0. | |
3139 | * | |
3140 | * This function can run simultaneously with readers of | |
3141 | * node_demotion[]. However, it can not run simultaneously | |
3142 | * with itself. Exclusion is provided by memory hotplug events | |
3143 | * being single-threaded. | |
3144 | */ | |
3145 | static void __set_migration_target_nodes(void) | |
3146 | { | |
3147 | nodemask_t next_pass = NODE_MASK_NONE; | |
3148 | nodemask_t this_pass = NODE_MASK_NONE; | |
3149 | nodemask_t used_targets = NODE_MASK_NONE; | |
3150 | int node; | |
3151 | ||
3152 | /* | |
3153 | * Avoid any oddities like cycles that could occur | |
3154 | * from changes in the topology. This will leave | |
3155 | * a momentary gap when migration is disabled. | |
3156 | */ | |
3157 | disable_all_migrate_targets(); | |
3158 | ||
3159 | /* | |
3160 | * Allocations go close to CPUs, first. Assume that | |
3161 | * the migration path starts at the nodes with CPUs. | |
3162 | */ | |
3163 | next_pass = node_states[N_CPU]; | |
3164 | again: | |
3165 | this_pass = next_pass; | |
3166 | next_pass = NODE_MASK_NONE; | |
3167 | /* | |
3168 | * To avoid cycles in the migration "graph", ensure | |
3169 | * that migration sources are not future targets by | |
3170 | * setting them in 'used_targets'. Do this only | |
3171 | * once per pass so that multiple source nodes can | |
3172 | * share a target node. | |
3173 | * | |
3174 | * 'used_targets' will become unavailable in future | |
3175 | * passes. This limits some opportunities for | |
3176 | * multiple source nodes to share a destination. | |
3177 | */ | |
3178 | nodes_or(used_targets, used_targets, this_pass); | |
3179 | for_each_node_mask(node, this_pass) { | |
3180 | int target_node = establish_migrate_target(node, &used_targets); | |
3181 | ||
3182 | if (target_node == NUMA_NO_NODE) | |
3183 | continue; | |
3184 | ||
3185 | /* | |
3186 | * Visit targets from this pass in the next pass. | |
3187 | * Eventually, every node will have been part of | |
3188 | * a pass, and will become set in 'used_targets'. | |
3189 | */ | |
3190 | node_set(target_node, next_pass); | |
3191 | } | |
3192 | /* | |
3193 | * 'next_pass' contains nodes which became migration | |
3194 | * targets in this pass. Make additional passes until | |
3195 | * no more migrations targets are available. | |
3196 | */ | |
3197 | if (!nodes_empty(next_pass)) | |
3198 | goto again; | |
3199 | } | |
3200 | ||
3201 | /* | |
3202 | * For callers that do not hold get_online_mems() already. | |
3203 | */ | |
79c28a41 DH |
3204 | static void set_migration_target_nodes(void) |
3205 | { | |
3206 | get_online_mems(); | |
3207 | __set_migration_target_nodes(); | |
3208 | put_online_mems(); | |
3209 | } | |
884a6e5d | 3210 | |
884a6e5d DH |
3211 | /* |
3212 | * This leaves migrate-on-reclaim transiently disabled between | |
3213 | * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs | |
3214 | * whether reclaim-based migration is enabled or not, which | |
3215 | * ensures that the user can turn reclaim-based migration at | |
3216 | * any time without needing to recalculate migration targets. | |
3217 | * | |
3218 | * These callbacks already hold get_online_mems(). That is why | |
3219 | * __set_migration_target_nodes() can be used as opposed to | |
3220 | * set_migration_target_nodes(). | |
3221 | */ | |
3222 | static int __meminit migrate_on_reclaim_callback(struct notifier_block *self, | |
295be91f | 3223 | unsigned long action, void *_arg) |
884a6e5d | 3224 | { |
295be91f DH |
3225 | struct memory_notify *arg = _arg; |
3226 | ||
3227 | /* | |
3228 | * Only update the node migration order when a node is | |
3229 | * changing status, like online->offline. This avoids | |
3230 | * the overhead of synchronize_rcu() in most cases. | |
3231 | */ | |
3232 | if (arg->status_change_nid < 0) | |
3233 | return notifier_from_errno(0); | |
3234 | ||
884a6e5d DH |
3235 | switch (action) { |
3236 | case MEM_GOING_OFFLINE: | |
3237 | /* | |
3238 | * Make sure there are not transient states where | |
3239 | * an offline node is a migration target. This | |
3240 | * will leave migration disabled until the offline | |
3241 | * completes and the MEM_OFFLINE case below runs. | |
3242 | */ | |
3243 | disable_all_migrate_targets(); | |
3244 | break; | |
3245 | case MEM_OFFLINE: | |
3246 | case MEM_ONLINE: | |
3247 | /* | |
3248 | * Recalculate the target nodes once the node | |
3249 | * reaches its final state (online or offline). | |
3250 | */ | |
3251 | __set_migration_target_nodes(); | |
3252 | break; | |
3253 | case MEM_CANCEL_OFFLINE: | |
3254 | /* | |
3255 | * MEM_GOING_OFFLINE disabled all the migration | |
3256 | * targets. Reenable them. | |
3257 | */ | |
3258 | __set_migration_target_nodes(); | |
3259 | break; | |
3260 | case MEM_GOING_ONLINE: | |
3261 | case MEM_CANCEL_ONLINE: | |
3262 | break; | |
3263 | } | |
3264 | ||
3265 | return notifier_from_errno(0); | |
3266 | } | |
3267 | ||
76af6a05 DH |
3268 | /* |
3269 | * React to hotplug events that might affect the migration targets | |
3270 | * like events that online or offline NUMA nodes. | |
3271 | * | |
3272 | * The ordering is also currently dependent on which nodes have | |
3273 | * CPUs. That means we need CPU on/offline notification too. | |
3274 | */ | |
3275 | static int migration_online_cpu(unsigned int cpu) | |
3276 | { | |
3277 | set_migration_target_nodes(); | |
3278 | return 0; | |
3279 | } | |
3280 | ||
3281 | static int migration_offline_cpu(unsigned int cpu) | |
3282 | { | |
3283 | set_migration_target_nodes(); | |
3284 | return 0; | |
3285 | } | |
3286 | ||
884a6e5d DH |
3287 | static int __init migrate_on_reclaim_init(void) |
3288 | { | |
3289 | int ret; | |
3290 | ||
a6a0251c HY |
3291 | ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline", |
3292 | NULL, migration_offline_cpu); | |
884a6e5d DH |
3293 | /* |
3294 | * In the unlikely case that this fails, the automatic | |
3295 | * migration targets may become suboptimal for nodes | |
3296 | * where N_CPU changes. With such a small impact in a | |
3297 | * rare case, do not bother trying to do anything special. | |
3298 | */ | |
3299 | WARN_ON(ret < 0); | |
a6a0251c HY |
3300 | ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online", |
3301 | migration_online_cpu, NULL); | |
3302 | WARN_ON(ret < 0); | |
884a6e5d DH |
3303 | |
3304 | hotplug_memory_notifier(migrate_on_reclaim_callback, 100); | |
3305 | return 0; | |
3306 | } | |
3307 | late_initcall(migrate_on_reclaim_init); | |
76af6a05 | 3308 | #endif /* CONFIG_HOTPLUG_CPU */ |
20f9ba4f YS |
3309 | |
3310 | bool numa_demotion_enabled = false; | |
3311 | ||
3312 | #ifdef CONFIG_SYSFS | |
3313 | static ssize_t numa_demotion_enabled_show(struct kobject *kobj, | |
3314 | struct kobj_attribute *attr, char *buf) | |
3315 | { | |
3316 | return sysfs_emit(buf, "%s\n", | |
3317 | numa_demotion_enabled ? "true" : "false"); | |
3318 | } | |
3319 | ||
3320 | static ssize_t numa_demotion_enabled_store(struct kobject *kobj, | |
3321 | struct kobj_attribute *attr, | |
3322 | const char *buf, size_t count) | |
3323 | { | |
3324 | if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1)) | |
3325 | numa_demotion_enabled = true; | |
3326 | else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1)) | |
3327 | numa_demotion_enabled = false; | |
3328 | else | |
3329 | return -EINVAL; | |
3330 | ||
3331 | return count; | |
3332 | } | |
3333 | ||
3334 | static struct kobj_attribute numa_demotion_enabled_attr = | |
3335 | __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show, | |
3336 | numa_demotion_enabled_store); | |
3337 | ||
3338 | static struct attribute *numa_attrs[] = { | |
3339 | &numa_demotion_enabled_attr.attr, | |
3340 | NULL, | |
3341 | }; | |
3342 | ||
3343 | static const struct attribute_group numa_attr_group = { | |
3344 | .attrs = numa_attrs, | |
3345 | }; | |
3346 | ||
3347 | static int __init numa_init_sysfs(void) | |
3348 | { | |
3349 | int err; | |
3350 | struct kobject *numa_kobj; | |
3351 | ||
3352 | numa_kobj = kobject_create_and_add("numa", mm_kobj); | |
3353 | if (!numa_kobj) { | |
3354 | pr_err("failed to create numa kobject\n"); | |
3355 | return -ENOMEM; | |
3356 | } | |
3357 | err = sysfs_create_group(numa_kobj, &numa_attr_group); | |
3358 | if (err) { | |
3359 | pr_err("failed to register numa group\n"); | |
3360 | goto delete_obj; | |
3361 | } | |
3362 | return 0; | |
3363 | ||
3364 | delete_obj: | |
3365 | kobject_put(numa_kobj); | |
3366 | return err; | |
3367 | } | |
3368 | subsys_initcall(numa_init_sysfs); | |
3369 | #endif |