]>
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> |
b20a3503 | 52 | |
0d1836c3 MN |
53 | #include <asm/tlbflush.h> |
54 | ||
7b2a2d4a MG |
55 | #define CREATE_TRACE_POINTS |
56 | #include <trace/events/migrate.h> | |
57 | ||
b20a3503 CL |
58 | #include "internal.h" |
59 | ||
9e5bcd61 | 60 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
bda807d4 MK |
61 | { |
62 | struct address_space *mapping; | |
63 | ||
64 | /* | |
65 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
66 | * or just got freed under us. | |
67 | * | |
68 | * In case we 'win' a race for a movable page being freed under us and | |
69 | * raise its refcount preventing __free_pages() from doing its job | |
70 | * the put_page() at the end of this block will take care of | |
71 | * release this page, thus avoiding a nasty leakage. | |
72 | */ | |
73 | if (unlikely(!get_page_unless_zero(page))) | |
74 | goto out; | |
75 | ||
76 | /* | |
77 | * Check PageMovable before holding a PG_lock because page's owner | |
78 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
8bb4e7a2 | 79 | * so unconditionally grabbing the lock ruins page's owner side. |
bda807d4 MK |
80 | */ |
81 | if (unlikely(!__PageMovable(page))) | |
82 | goto out_putpage; | |
83 | /* | |
84 | * As movable pages are not isolated from LRU lists, concurrent | |
85 | * compaction threads can race against page migration functions | |
86 | * as well as race against the releasing a page. | |
87 | * | |
88 | * In order to avoid having an already isolated movable page | |
89 | * being (wrongly) re-isolated while it is under migration, | |
90 | * or to avoid attempting to isolate pages being released, | |
91 | * lets be sure we have the page lock | |
92 | * before proceeding with the movable page isolation steps. | |
93 | */ | |
94 | if (unlikely(!trylock_page(page))) | |
95 | goto out_putpage; | |
96 | ||
97 | if (!PageMovable(page) || PageIsolated(page)) | |
98 | goto out_no_isolated; | |
99 | ||
100 | mapping = page_mapping(page); | |
101 | VM_BUG_ON_PAGE(!mapping, page); | |
102 | ||
103 | if (!mapping->a_ops->isolate_page(page, mode)) | |
104 | goto out_no_isolated; | |
105 | ||
106 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
107 | WARN_ON_ONCE(PageIsolated(page)); | |
108 | __SetPageIsolated(page); | |
109 | unlock_page(page); | |
110 | ||
9e5bcd61 | 111 | return 0; |
bda807d4 MK |
112 | |
113 | out_no_isolated: | |
114 | unlock_page(page); | |
115 | out_putpage: | |
116 | put_page(page); | |
117 | out: | |
9e5bcd61 | 118 | return -EBUSY; |
bda807d4 MK |
119 | } |
120 | ||
606a6f71 | 121 | static void putback_movable_page(struct page *page) |
bda807d4 MK |
122 | { |
123 | struct address_space *mapping; | |
124 | ||
bda807d4 MK |
125 | mapping = page_mapping(page); |
126 | mapping->a_ops->putback_page(page); | |
127 | __ClearPageIsolated(page); | |
128 | } | |
129 | ||
5733c7d1 RA |
130 | /* |
131 | * Put previously isolated pages back onto the appropriate lists | |
132 | * from where they were once taken off for compaction/migration. | |
133 | * | |
59c82b70 JK |
134 | * This function shall be used whenever the isolated pageset has been |
135 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
136 | * and isolate_huge_page(). | |
5733c7d1 RA |
137 | */ |
138 | void putback_movable_pages(struct list_head *l) | |
139 | { | |
140 | struct page *page; | |
141 | struct page *page2; | |
142 | ||
b20a3503 | 143 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
144 | if (unlikely(PageHuge(page))) { |
145 | putback_active_hugepage(page); | |
146 | continue; | |
147 | } | |
e24f0b8f | 148 | list_del(&page->lru); |
bda807d4 MK |
149 | /* |
150 | * We isolated non-lru movable page so here we can use | |
151 | * __PageMovable because LRU page's mapping cannot have | |
152 | * PAGE_MAPPING_MOVABLE. | |
153 | */ | |
b1123ea6 | 154 | if (unlikely(__PageMovable(page))) { |
bda807d4 MK |
155 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
156 | lock_page(page); | |
157 | if (PageMovable(page)) | |
158 | putback_movable_page(page); | |
159 | else | |
160 | __ClearPageIsolated(page); | |
161 | unlock_page(page); | |
162 | put_page(page); | |
163 | } else { | |
e8db67eb | 164 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 165 | page_is_file_lru(page), -thp_nr_pages(page)); |
fc280fe8 | 166 | putback_lru_page(page); |
bda807d4 | 167 | } |
b20a3503 | 168 | } |
b20a3503 CL |
169 | } |
170 | ||
0697212a CL |
171 | /* |
172 | * Restore a potential migration pte to a working pte entry | |
173 | */ | |
e4b82222 | 174 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
e9995ef9 | 175 | unsigned long addr, void *old) |
0697212a | 176 | { |
3fe87967 KS |
177 | struct page_vma_mapped_walk pvmw = { |
178 | .page = old, | |
179 | .vma = vma, | |
180 | .address = addr, | |
181 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
182 | }; | |
183 | struct page *new; | |
184 | pte_t pte; | |
0697212a | 185 | swp_entry_t entry; |
0697212a | 186 | |
3fe87967 KS |
187 | VM_BUG_ON_PAGE(PageTail(page), page); |
188 | while (page_vma_mapped_walk(&pvmw)) { | |
4b0ece6f NH |
189 | if (PageKsm(page)) |
190 | new = page; | |
191 | else | |
192 | new = page - pvmw.page->index + | |
193 | linear_page_index(vma, pvmw.address); | |
0697212a | 194 | |
616b8371 ZY |
195 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
196 | /* PMD-mapped THP migration entry */ | |
197 | if (!pvmw.pte) { | |
198 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
199 | remove_migration_pmd(&pvmw, new); | |
200 | continue; | |
201 | } | |
202 | #endif | |
203 | ||
3fe87967 KS |
204 | get_page(new); |
205 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
206 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
207 | pte = pte_mksoft_dirty(pte); | |
0697212a | 208 | |
3fe87967 KS |
209 | /* |
210 | * Recheck VMA as permissions can change since migration started | |
211 | */ | |
212 | entry = pte_to_swp_entry(*pvmw.pte); | |
213 | if (is_write_migration_entry(entry)) | |
214 | pte = maybe_mkwrite(pte, vma); | |
f45ec5ff PX |
215 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
216 | pte = pte_mkuffd_wp(pte); | |
d3cb8bf6 | 217 | |
6128763f RC |
218 | if (unlikely(is_device_private_page(new))) { |
219 | entry = make_device_private_entry(new, pte_write(pte)); | |
220 | pte = swp_entry_to_pte(entry); | |
3d321bf8 RC |
221 | if (pte_swp_soft_dirty(*pvmw.pte)) |
222 | pte = pte_swp_mksoft_dirty(pte); | |
6128763f RC |
223 | if (pte_swp_uffd_wp(*pvmw.pte)) |
224 | pte = pte_swp_mkuffd_wp(pte); | |
d2b2c6dd | 225 | } |
a5430dda | 226 | |
3ef8fd7f | 227 | #ifdef CONFIG_HUGETLB_PAGE |
3fe87967 KS |
228 | if (PageHuge(new)) { |
229 | pte = pte_mkhuge(pte); | |
230 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
383321ab | 231 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
3fe87967 KS |
232 | if (PageAnon(new)) |
233 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
234 | else | |
235 | page_dup_rmap(new, true); | |
383321ab AK |
236 | } else |
237 | #endif | |
238 | { | |
239 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
04e62a29 | 240 | |
383321ab AK |
241 | if (PageAnon(new)) |
242 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
243 | else | |
244 | page_add_file_rmap(new, false); | |
245 | } | |
3fe87967 KS |
246 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
247 | mlock_vma_page(new); | |
248 | ||
e125fe40 KS |
249 | if (PageTransHuge(page) && PageMlocked(page)) |
250 | clear_page_mlock(page); | |
251 | ||
3fe87967 KS |
252 | /* No need to invalidate - it was non-present before */ |
253 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
254 | } | |
51afb12b | 255 | |
e4b82222 | 256 | return true; |
0697212a CL |
257 | } |
258 | ||
04e62a29 CL |
259 | /* |
260 | * Get rid of all migration entries and replace them by | |
261 | * references to the indicated page. | |
262 | */ | |
e388466d | 263 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
04e62a29 | 264 | { |
051ac83a JK |
265 | struct rmap_walk_control rwc = { |
266 | .rmap_one = remove_migration_pte, | |
267 | .arg = old, | |
268 | }; | |
269 | ||
e388466d KS |
270 | if (locked) |
271 | rmap_walk_locked(new, &rwc); | |
272 | else | |
273 | rmap_walk(new, &rwc); | |
04e62a29 CL |
274 | } |
275 | ||
0697212a CL |
276 | /* |
277 | * Something used the pte of a page under migration. We need to | |
278 | * get to the page and wait until migration is finished. | |
279 | * When we return from this function the fault will be retried. | |
0697212a | 280 | */ |
e66f17ff | 281 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
30dad309 | 282 | spinlock_t *ptl) |
0697212a | 283 | { |
30dad309 | 284 | pte_t pte; |
0697212a CL |
285 | swp_entry_t entry; |
286 | struct page *page; | |
287 | ||
30dad309 | 288 | spin_lock(ptl); |
0697212a CL |
289 | pte = *ptep; |
290 | if (!is_swap_pte(pte)) | |
291 | goto out; | |
292 | ||
293 | entry = pte_to_swp_entry(pte); | |
294 | if (!is_migration_entry(entry)) | |
295 | goto out; | |
296 | ||
297 | page = migration_entry_to_page(entry); | |
298 | ||
e286781d | 299 | /* |
89eb946a | 300 | * Once page cache replacement of page migration started, page_count |
9a1ea439 HD |
301 | * is zero; but we must not call put_and_wait_on_page_locked() without |
302 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
e286781d NP |
303 | */ |
304 | if (!get_page_unless_zero(page)) | |
305 | goto out; | |
0697212a | 306 | pte_unmap_unlock(ptep, ptl); |
48054625 | 307 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
0697212a CL |
308 | return; |
309 | out: | |
310 | pte_unmap_unlock(ptep, ptl); | |
311 | } | |
312 | ||
30dad309 NH |
313 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
314 | unsigned long address) | |
315 | { | |
316 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
317 | pte_t *ptep = pte_offset_map(pmd, address); | |
318 | __migration_entry_wait(mm, ptep, ptl); | |
319 | } | |
320 | ||
cb900f41 KS |
321 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
322 | struct mm_struct *mm, pte_t *pte) | |
30dad309 | 323 | { |
cb900f41 | 324 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
30dad309 NH |
325 | __migration_entry_wait(mm, pte, ptl); |
326 | } | |
327 | ||
616b8371 ZY |
328 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
329 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
330 | { | |
331 | spinlock_t *ptl; | |
332 | struct page *page; | |
333 | ||
334 | ptl = pmd_lock(mm, pmd); | |
335 | if (!is_pmd_migration_entry(*pmd)) | |
336 | goto unlock; | |
337 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
338 | if (!get_page_unless_zero(page)) | |
339 | goto unlock; | |
340 | spin_unlock(ptl); | |
48054625 | 341 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
616b8371 ZY |
342 | return; |
343 | unlock: | |
344 | spin_unlock(ptl); | |
345 | } | |
346 | #endif | |
347 | ||
f900482d | 348 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
0b3901b3 JK |
349 | { |
350 | int expected_count = 1; | |
351 | ||
352 | /* | |
f1f4f3ab | 353 | * Device private pages have an extra refcount as they are |
0b3901b3 JK |
354 | * ZONE_DEVICE pages. |
355 | */ | |
356 | expected_count += is_device_private_page(page); | |
f900482d | 357 | if (mapping) |
6c357848 | 358 | expected_count += thp_nr_pages(page) + page_has_private(page); |
0b3901b3 JK |
359 | |
360 | return expected_count; | |
361 | } | |
362 | ||
b20a3503 | 363 | /* |
c3fcf8a5 | 364 | * Replace the page in the mapping. |
5b5c7120 CL |
365 | * |
366 | * The number of remaining references must be: | |
367 | * 1 for anonymous pages without a mapping | |
368 | * 2 for pages with a mapping | |
266cf658 | 369 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 370 | */ |
36bc08cc | 371 | int migrate_page_move_mapping(struct address_space *mapping, |
37109694 | 372 | struct page *newpage, struct page *page, int extra_count) |
b20a3503 | 373 | { |
89eb946a | 374 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
42cb14b1 HD |
375 | struct zone *oldzone, *newzone; |
376 | int dirty; | |
f900482d | 377 | int expected_count = expected_page_refs(mapping, page) + extra_count; |
5c447d27 | 378 | int nr = thp_nr_pages(page); |
8763cb45 | 379 | |
6c5240ae | 380 | if (!mapping) { |
0e8c7d0f | 381 | /* Anonymous page without mapping */ |
8e321fef | 382 | if (page_count(page) != expected_count) |
6c5240ae | 383 | return -EAGAIN; |
cf4b769a HD |
384 | |
385 | /* No turning back from here */ | |
cf4b769a HD |
386 | newpage->index = page->index; |
387 | newpage->mapping = page->mapping; | |
388 | if (PageSwapBacked(page)) | |
fa9949da | 389 | __SetPageSwapBacked(newpage); |
cf4b769a | 390 | |
78bd5209 | 391 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
392 | } |
393 | ||
42cb14b1 HD |
394 | oldzone = page_zone(page); |
395 | newzone = page_zone(newpage); | |
396 | ||
89eb946a | 397 | xas_lock_irq(&xas); |
89eb946a MW |
398 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
399 | xas_unlock_irq(&xas); | |
e23ca00b | 400 | return -EAGAIN; |
b20a3503 CL |
401 | } |
402 | ||
fe896d18 | 403 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 404 | xas_unlock_irq(&xas); |
e286781d NP |
405 | return -EAGAIN; |
406 | } | |
407 | ||
b20a3503 | 408 | /* |
cf4b769a HD |
409 | * Now we know that no one else is looking at the page: |
410 | * no turning back from here. | |
b20a3503 | 411 | */ |
cf4b769a HD |
412 | newpage->index = page->index; |
413 | newpage->mapping = page->mapping; | |
5c447d27 | 414 | page_ref_add(newpage, nr); /* add cache reference */ |
6326fec1 NP |
415 | if (PageSwapBacked(page)) { |
416 | __SetPageSwapBacked(newpage); | |
417 | if (PageSwapCache(page)) { | |
418 | SetPageSwapCache(newpage); | |
419 | set_page_private(newpage, page_private(page)); | |
420 | } | |
421 | } else { | |
422 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
b20a3503 CL |
423 | } |
424 | ||
42cb14b1 HD |
425 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
426 | dirty = PageDirty(page); | |
427 | if (dirty) { | |
428 | ClearPageDirty(page); | |
429 | SetPageDirty(newpage); | |
430 | } | |
431 | ||
89eb946a | 432 | xas_store(&xas, newpage); |
e71769ae NH |
433 | if (PageTransHuge(page)) { |
434 | int i; | |
e71769ae | 435 | |
5c447d27 | 436 | for (i = 1; i < nr; i++) { |
89eb946a | 437 | xas_next(&xas); |
4101196b | 438 | xas_store(&xas, newpage); |
e71769ae | 439 | } |
e71769ae | 440 | } |
7cf9c2c7 NP |
441 | |
442 | /* | |
937a94c9 JG |
443 | * Drop cache reference from old page by unfreezing |
444 | * to one less reference. | |
7cf9c2c7 NP |
445 | * We know this isn't the last reference. |
446 | */ | |
5c447d27 | 447 | page_ref_unfreeze(page, expected_count - nr); |
7cf9c2c7 | 448 | |
89eb946a | 449 | xas_unlock(&xas); |
42cb14b1 HD |
450 | /* Leave irq disabled to prevent preemption while updating stats */ |
451 | ||
0e8c7d0f CL |
452 | /* |
453 | * If moved to a different zone then also account | |
454 | * the page for that zone. Other VM counters will be | |
455 | * taken care of when we establish references to the | |
456 | * new page and drop references to the old page. | |
457 | * | |
458 | * Note that anonymous pages are accounted for | |
4b9d0fab | 459 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
0e8c7d0f CL |
460 | * are mapped to swap space. |
461 | */ | |
42cb14b1 | 462 | if (newzone != oldzone) { |
0d1c2072 JW |
463 | struct lruvec *old_lruvec, *new_lruvec; |
464 | struct mem_cgroup *memcg; | |
465 | ||
466 | memcg = page_memcg(page); | |
467 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); | |
468 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
469 | ||
5c447d27 SB |
470 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
471 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); | |
42cb14b1 | 472 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
5c447d27 SB |
473 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
474 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); | |
42cb14b1 | 475 | } |
b6038942 SB |
476 | #ifdef CONFIG_SWAP |
477 | if (PageSwapCache(page)) { | |
478 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); | |
479 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); | |
480 | } | |
481 | #endif | |
f56753ac | 482 | if (dirty && mapping_can_writeback(mapping)) { |
5c447d27 SB |
483 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
484 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); | |
485 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); | |
486 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); | |
42cb14b1 | 487 | } |
4b02108a | 488 | } |
42cb14b1 | 489 | local_irq_enable(); |
b20a3503 | 490 | |
78bd5209 | 491 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 492 | } |
1118dce7 | 493 | EXPORT_SYMBOL(migrate_page_move_mapping); |
b20a3503 | 494 | |
290408d4 NH |
495 | /* |
496 | * The expected number of remaining references is the same as that | |
497 | * of migrate_page_move_mapping(). | |
498 | */ | |
499 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
500 | struct page *newpage, struct page *page) | |
501 | { | |
89eb946a | 502 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
290408d4 | 503 | int expected_count; |
290408d4 | 504 | |
89eb946a | 505 | xas_lock_irq(&xas); |
290408d4 | 506 | expected_count = 2 + page_has_private(page); |
89eb946a MW |
507 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
508 | xas_unlock_irq(&xas); | |
290408d4 NH |
509 | return -EAGAIN; |
510 | } | |
511 | ||
fe896d18 | 512 | if (!page_ref_freeze(page, expected_count)) { |
89eb946a | 513 | xas_unlock_irq(&xas); |
290408d4 NH |
514 | return -EAGAIN; |
515 | } | |
516 | ||
cf4b769a HD |
517 | newpage->index = page->index; |
518 | newpage->mapping = page->mapping; | |
6a93ca8f | 519 | |
290408d4 NH |
520 | get_page(newpage); |
521 | ||
89eb946a | 522 | xas_store(&xas, newpage); |
290408d4 | 523 | |
fe896d18 | 524 | page_ref_unfreeze(page, expected_count - 1); |
290408d4 | 525 | |
89eb946a | 526 | xas_unlock_irq(&xas); |
6a93ca8f | 527 | |
78bd5209 | 528 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
529 | } |
530 | ||
30b0a105 DH |
531 | /* |
532 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
533 | * arithmetic will work across the entire page. We need something more | |
534 | * specialized. | |
535 | */ | |
536 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
537 | int nr_pages) | |
538 | { | |
539 | int i; | |
540 | struct page *dst_base = dst; | |
541 | struct page *src_base = src; | |
542 | ||
543 | for (i = 0; i < nr_pages; ) { | |
544 | cond_resched(); | |
545 | copy_highpage(dst, src); | |
546 | ||
547 | i++; | |
548 | dst = mem_map_next(dst, dst_base, i); | |
549 | src = mem_map_next(src, src_base, i); | |
550 | } | |
551 | } | |
552 | ||
553 | static void copy_huge_page(struct page *dst, struct page *src) | |
554 | { | |
555 | int i; | |
556 | int nr_pages; | |
557 | ||
558 | if (PageHuge(src)) { | |
559 | /* hugetlbfs page */ | |
560 | struct hstate *h = page_hstate(src); | |
561 | nr_pages = pages_per_huge_page(h); | |
562 | ||
563 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
564 | __copy_gigantic_page(dst, src, nr_pages); | |
565 | return; | |
566 | } | |
567 | } else { | |
568 | /* thp page */ | |
569 | BUG_ON(!PageTransHuge(src)); | |
6c357848 | 570 | nr_pages = thp_nr_pages(src); |
30b0a105 DH |
571 | } |
572 | ||
573 | for (i = 0; i < nr_pages; i++) { | |
574 | cond_resched(); | |
575 | copy_highpage(dst + i, src + i); | |
576 | } | |
577 | } | |
578 | ||
b20a3503 CL |
579 | /* |
580 | * Copy the page to its new location | |
581 | */ | |
2916ecc0 | 582 | void migrate_page_states(struct page *newpage, struct page *page) |
b20a3503 | 583 | { |
7851a45c RR |
584 | int cpupid; |
585 | ||
b20a3503 CL |
586 | if (PageError(page)) |
587 | SetPageError(newpage); | |
588 | if (PageReferenced(page)) | |
589 | SetPageReferenced(newpage); | |
590 | if (PageUptodate(page)) | |
591 | SetPageUptodate(newpage); | |
894bc310 | 592 | if (TestClearPageActive(page)) { |
309381fe | 593 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
b20a3503 | 594 | SetPageActive(newpage); |
418b27ef LS |
595 | } else if (TestClearPageUnevictable(page)) |
596 | SetPageUnevictable(newpage); | |
1899ad18 JW |
597 | if (PageWorkingset(page)) |
598 | SetPageWorkingset(newpage); | |
b20a3503 CL |
599 | if (PageChecked(page)) |
600 | SetPageChecked(newpage); | |
601 | if (PageMappedToDisk(page)) | |
602 | SetPageMappedToDisk(newpage); | |
603 | ||
42cb14b1 HD |
604 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
605 | if (PageDirty(page)) | |
606 | SetPageDirty(newpage); | |
b20a3503 | 607 | |
33c3fc71 VD |
608 | if (page_is_young(page)) |
609 | set_page_young(newpage); | |
610 | if (page_is_idle(page)) | |
611 | set_page_idle(newpage); | |
612 | ||
7851a45c RR |
613 | /* |
614 | * Copy NUMA information to the new page, to prevent over-eager | |
615 | * future migrations of this same page. | |
616 | */ | |
617 | cpupid = page_cpupid_xchg_last(page, -1); | |
618 | page_cpupid_xchg_last(newpage, cpupid); | |
619 | ||
e9995ef9 | 620 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
621 | /* |
622 | * Please do not reorder this without considering how mm/ksm.c's | |
623 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
624 | */ | |
b3b3a99c NH |
625 | if (PageSwapCache(page)) |
626 | ClearPageSwapCache(page); | |
b20a3503 CL |
627 | ClearPagePrivate(page); |
628 | set_page_private(page, 0); | |
b20a3503 CL |
629 | |
630 | /* | |
631 | * If any waiters have accumulated on the new page then | |
632 | * wake them up. | |
633 | */ | |
634 | if (PageWriteback(newpage)) | |
635 | end_page_writeback(newpage); | |
d435edca | 636 | |
6aeff241 YS |
637 | /* |
638 | * PG_readahead shares the same bit with PG_reclaim. The above | |
639 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
640 | * bit after that. | |
641 | */ | |
642 | if (PageReadahead(page)) | |
643 | SetPageReadahead(newpage); | |
644 | ||
d435edca | 645 | copy_page_owner(page, newpage); |
74485cf2 | 646 | |
a333e3e7 HD |
647 | if (!PageHuge(page)) |
648 | mem_cgroup_migrate(page, newpage); | |
b20a3503 | 649 | } |
2916ecc0 JG |
650 | EXPORT_SYMBOL(migrate_page_states); |
651 | ||
652 | void migrate_page_copy(struct page *newpage, struct page *page) | |
653 | { | |
654 | if (PageHuge(page) || PageTransHuge(page)) | |
655 | copy_huge_page(newpage, page); | |
656 | else | |
657 | copy_highpage(newpage, page); | |
658 | ||
659 | migrate_page_states(newpage, page); | |
660 | } | |
1118dce7 | 661 | EXPORT_SYMBOL(migrate_page_copy); |
b20a3503 | 662 | |
1d8b85cc CL |
663 | /************************************************************ |
664 | * Migration functions | |
665 | ***********************************************************/ | |
666 | ||
b20a3503 | 667 | /* |
bda807d4 | 668 | * Common logic to directly migrate a single LRU page suitable for |
266cf658 | 669 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
670 | * |
671 | * Pages are locked upon entry and exit. | |
672 | */ | |
2d1db3b1 | 673 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
674 | struct page *newpage, struct page *page, |
675 | enum migrate_mode mode) | |
b20a3503 CL |
676 | { |
677 | int rc; | |
678 | ||
679 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
680 | ||
37109694 | 681 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
b20a3503 | 682 | |
78bd5209 | 683 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
684 | return rc; |
685 | ||
2916ecc0 JG |
686 | if (mode != MIGRATE_SYNC_NO_COPY) |
687 | migrate_page_copy(newpage, page); | |
688 | else | |
689 | migrate_page_states(newpage, page); | |
78bd5209 | 690 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
691 | } |
692 | EXPORT_SYMBOL(migrate_page); | |
693 | ||
9361401e | 694 | #ifdef CONFIG_BLOCK |
84ade7c1 JK |
695 | /* Returns true if all buffers are successfully locked */ |
696 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
697 | enum migrate_mode mode) | |
698 | { | |
699 | struct buffer_head *bh = head; | |
700 | ||
701 | /* Simple case, sync compaction */ | |
702 | if (mode != MIGRATE_ASYNC) { | |
703 | do { | |
84ade7c1 JK |
704 | lock_buffer(bh); |
705 | bh = bh->b_this_page; | |
706 | ||
707 | } while (bh != head); | |
708 | ||
709 | return true; | |
710 | } | |
711 | ||
712 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
713 | do { | |
84ade7c1 JK |
714 | if (!trylock_buffer(bh)) { |
715 | /* | |
716 | * We failed to lock the buffer and cannot stall in | |
717 | * async migration. Release the taken locks | |
718 | */ | |
719 | struct buffer_head *failed_bh = bh; | |
84ade7c1 JK |
720 | bh = head; |
721 | while (bh != failed_bh) { | |
722 | unlock_buffer(bh); | |
84ade7c1 JK |
723 | bh = bh->b_this_page; |
724 | } | |
725 | return false; | |
726 | } | |
727 | ||
728 | bh = bh->b_this_page; | |
729 | } while (bh != head); | |
730 | return true; | |
731 | } | |
732 | ||
89cb0888 JK |
733 | static int __buffer_migrate_page(struct address_space *mapping, |
734 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
735 | bool check_refs) | |
1d8b85cc | 736 | { |
1d8b85cc CL |
737 | struct buffer_head *bh, *head; |
738 | int rc; | |
cc4f11e6 | 739 | int expected_count; |
1d8b85cc | 740 | |
1d8b85cc | 741 | if (!page_has_buffers(page)) |
a6bc32b8 | 742 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc | 743 | |
cc4f11e6 | 744 | /* Check whether page does not have extra refs before we do more work */ |
f900482d | 745 | expected_count = expected_page_refs(mapping, page); |
cc4f11e6 JK |
746 | if (page_count(page) != expected_count) |
747 | return -EAGAIN; | |
1d8b85cc | 748 | |
cc4f11e6 JK |
749 | head = page_buffers(page); |
750 | if (!buffer_migrate_lock_buffers(head, mode)) | |
751 | return -EAGAIN; | |
1d8b85cc | 752 | |
89cb0888 JK |
753 | if (check_refs) { |
754 | bool busy; | |
755 | bool invalidated = false; | |
756 | ||
757 | recheck_buffers: | |
758 | busy = false; | |
759 | spin_lock(&mapping->private_lock); | |
760 | bh = head; | |
761 | do { | |
762 | if (atomic_read(&bh->b_count)) { | |
763 | busy = true; | |
764 | break; | |
765 | } | |
766 | bh = bh->b_this_page; | |
767 | } while (bh != head); | |
89cb0888 JK |
768 | if (busy) { |
769 | if (invalidated) { | |
770 | rc = -EAGAIN; | |
771 | goto unlock_buffers; | |
772 | } | |
ebdf4de5 | 773 | spin_unlock(&mapping->private_lock); |
89cb0888 JK |
774 | invalidate_bh_lrus(); |
775 | invalidated = true; | |
776 | goto recheck_buffers; | |
777 | } | |
778 | } | |
779 | ||
37109694 | 780 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
78bd5209 | 781 | if (rc != MIGRATEPAGE_SUCCESS) |
cc4f11e6 | 782 | goto unlock_buffers; |
1d8b85cc | 783 | |
cd0f3715 | 784 | attach_page_private(newpage, detach_page_private(page)); |
1d8b85cc CL |
785 | |
786 | bh = head; | |
787 | do { | |
788 | set_bh_page(bh, newpage, bh_offset(bh)); | |
789 | bh = bh->b_this_page; | |
790 | ||
791 | } while (bh != head); | |
792 | ||
2916ecc0 JG |
793 | if (mode != MIGRATE_SYNC_NO_COPY) |
794 | migrate_page_copy(newpage, page); | |
795 | else | |
796 | migrate_page_states(newpage, page); | |
1d8b85cc | 797 | |
cc4f11e6 JK |
798 | rc = MIGRATEPAGE_SUCCESS; |
799 | unlock_buffers: | |
ebdf4de5 JK |
800 | if (check_refs) |
801 | spin_unlock(&mapping->private_lock); | |
1d8b85cc CL |
802 | bh = head; |
803 | do { | |
804 | unlock_buffer(bh); | |
1d8b85cc CL |
805 | bh = bh->b_this_page; |
806 | ||
807 | } while (bh != head); | |
808 | ||
cc4f11e6 | 809 | return rc; |
1d8b85cc | 810 | } |
89cb0888 JK |
811 | |
812 | /* | |
813 | * Migration function for pages with buffers. This function can only be used | |
814 | * if the underlying filesystem guarantees that no other references to "page" | |
815 | * exist. For example attached buffer heads are accessed only under page lock. | |
816 | */ | |
817 | int buffer_migrate_page(struct address_space *mapping, | |
818 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
819 | { | |
820 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
821 | } | |
1d8b85cc | 822 | EXPORT_SYMBOL(buffer_migrate_page); |
89cb0888 JK |
823 | |
824 | /* | |
825 | * Same as above except that this variant is more careful and checks that there | |
826 | * are also no buffer head references. This function is the right one for | |
827 | * mappings where buffer heads are directly looked up and referenced (such as | |
828 | * block device mappings). | |
829 | */ | |
830 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
831 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
832 | { | |
833 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
834 | } | |
9361401e | 835 | #endif |
1d8b85cc | 836 | |
04e62a29 CL |
837 | /* |
838 | * Writeback a page to clean the dirty state | |
839 | */ | |
840 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 841 | { |
04e62a29 CL |
842 | struct writeback_control wbc = { |
843 | .sync_mode = WB_SYNC_NONE, | |
844 | .nr_to_write = 1, | |
845 | .range_start = 0, | |
846 | .range_end = LLONG_MAX, | |
04e62a29 CL |
847 | .for_reclaim = 1 |
848 | }; | |
849 | int rc; | |
850 | ||
851 | if (!mapping->a_ops->writepage) | |
852 | /* No write method for the address space */ | |
853 | return -EINVAL; | |
854 | ||
855 | if (!clear_page_dirty_for_io(page)) | |
856 | /* Someone else already triggered a write */ | |
857 | return -EAGAIN; | |
858 | ||
8351a6e4 | 859 | /* |
04e62a29 CL |
860 | * A dirty page may imply that the underlying filesystem has |
861 | * the page on some queue. So the page must be clean for | |
862 | * migration. Writeout may mean we loose the lock and the | |
863 | * page state is no longer what we checked for earlier. | |
864 | * At this point we know that the migration attempt cannot | |
865 | * be successful. | |
8351a6e4 | 866 | */ |
e388466d | 867 | remove_migration_ptes(page, page, false); |
8351a6e4 | 868 | |
04e62a29 | 869 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 870 | |
04e62a29 CL |
871 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
872 | /* unlocked. Relock */ | |
873 | lock_page(page); | |
874 | ||
bda8550d | 875 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
876 | } |
877 | ||
878 | /* | |
879 | * Default handling if a filesystem does not provide a migration function. | |
880 | */ | |
881 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 882 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 883 | { |
b969c4ab | 884 | if (PageDirty(page)) { |
a6bc32b8 | 885 | /* Only writeback pages in full synchronous migration */ |
2916ecc0 JG |
886 | switch (mode) { |
887 | case MIGRATE_SYNC: | |
888 | case MIGRATE_SYNC_NO_COPY: | |
889 | break; | |
890 | default: | |
b969c4ab | 891 | return -EBUSY; |
2916ecc0 | 892 | } |
04e62a29 | 893 | return writeout(mapping, page); |
b969c4ab | 894 | } |
8351a6e4 CL |
895 | |
896 | /* | |
897 | * Buffers may be managed in a filesystem specific way. | |
898 | * We must have no buffers or drop them. | |
899 | */ | |
266cf658 | 900 | if (page_has_private(page) && |
8351a6e4 | 901 | !try_to_release_page(page, GFP_KERNEL)) |
806031bb | 902 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
8351a6e4 | 903 | |
a6bc32b8 | 904 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
905 | } |
906 | ||
e24f0b8f CL |
907 | /* |
908 | * Move a page to a newly allocated page | |
909 | * The page is locked and all ptes have been successfully removed. | |
910 | * | |
911 | * The new page will have replaced the old page if this function | |
912 | * is successful. | |
894bc310 LS |
913 | * |
914 | * Return value: | |
915 | * < 0 - error code | |
78bd5209 | 916 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 917 | */ |
3fe2011f | 918 | static int move_to_new_page(struct page *newpage, struct page *page, |
5c3f9a67 | 919 | enum migrate_mode mode) |
e24f0b8f CL |
920 | { |
921 | struct address_space *mapping; | |
bda807d4 MK |
922 | int rc = -EAGAIN; |
923 | bool is_lru = !__PageMovable(page); | |
e24f0b8f | 924 | |
7db7671f HD |
925 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
926 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
e24f0b8f | 927 | |
e24f0b8f | 928 | mapping = page_mapping(page); |
bda807d4 MK |
929 | |
930 | if (likely(is_lru)) { | |
931 | if (!mapping) | |
932 | rc = migrate_page(mapping, newpage, page, mode); | |
933 | else if (mapping->a_ops->migratepage) | |
934 | /* | |
935 | * Most pages have a mapping and most filesystems | |
936 | * provide a migratepage callback. Anonymous pages | |
937 | * are part of swap space which also has its own | |
938 | * migratepage callback. This is the most common path | |
939 | * for page migration. | |
940 | */ | |
941 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
942 | page, mode); | |
943 | else | |
944 | rc = fallback_migrate_page(mapping, newpage, | |
945 | page, mode); | |
946 | } else { | |
e24f0b8f | 947 | /* |
bda807d4 MK |
948 | * In case of non-lru page, it could be released after |
949 | * isolation step. In that case, we shouldn't try migration. | |
e24f0b8f | 950 | */ |
bda807d4 MK |
951 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
952 | if (!PageMovable(page)) { | |
953 | rc = MIGRATEPAGE_SUCCESS; | |
954 | __ClearPageIsolated(page); | |
955 | goto out; | |
956 | } | |
957 | ||
958 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
959 | page, mode); | |
960 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
961 | !PageIsolated(page)); | |
962 | } | |
e24f0b8f | 963 | |
5c3f9a67 HD |
964 | /* |
965 | * When successful, old pagecache page->mapping must be cleared before | |
966 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
967 | */ | |
968 | if (rc == MIGRATEPAGE_SUCCESS) { | |
bda807d4 MK |
969 | if (__PageMovable(page)) { |
970 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
971 | ||
972 | /* | |
973 | * We clear PG_movable under page_lock so any compactor | |
974 | * cannot try to migrate this page. | |
975 | */ | |
976 | __ClearPageIsolated(page); | |
977 | } | |
978 | ||
979 | /* | |
c23a0c99 | 980 | * Anonymous and movable page->mapping will be cleared by |
bda807d4 MK |
981 | * free_pages_prepare so don't reset it here for keeping |
982 | * the type to work PageAnon, for example. | |
983 | */ | |
984 | if (!PageMappingFlags(page)) | |
5c3f9a67 | 985 | page->mapping = NULL; |
d2b2c6dd | 986 | |
25b2995a | 987 | if (likely(!is_zone_device_page(newpage))) |
d2b2c6dd LP |
988 | flush_dcache_page(newpage); |
989 | ||
3fe2011f | 990 | } |
bda807d4 | 991 | out: |
e24f0b8f CL |
992 | return rc; |
993 | } | |
994 | ||
0dabec93 | 995 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 996 | int force, enum migrate_mode mode) |
e24f0b8f | 997 | { |
0dabec93 | 998 | int rc = -EAGAIN; |
2ebba6b7 | 999 | int page_was_mapped = 0; |
3f6c8272 | 1000 | struct anon_vma *anon_vma = NULL; |
bda807d4 | 1001 | bool is_lru = !__PageMovable(page); |
95a402c3 | 1002 | |
529ae9aa | 1003 | if (!trylock_page(page)) { |
a6bc32b8 | 1004 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 1005 | goto out; |
3e7d3449 MG |
1006 | |
1007 | /* | |
1008 | * It's not safe for direct compaction to call lock_page. | |
1009 | * For example, during page readahead pages are added locked | |
1010 | * to the LRU. Later, when the IO completes the pages are | |
1011 | * marked uptodate and unlocked. However, the queueing | |
1012 | * could be merging multiple pages for one bio (e.g. | |
d4388340 | 1013 | * mpage_readahead). If an allocation happens for the |
3e7d3449 MG |
1014 | * second or third page, the process can end up locking |
1015 | * the same page twice and deadlocking. Rather than | |
1016 | * trying to be clever about what pages can be locked, | |
1017 | * avoid the use of lock_page for direct compaction | |
1018 | * altogether. | |
1019 | */ | |
1020 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 1021 | goto out; |
3e7d3449 | 1022 | |
e24f0b8f CL |
1023 | lock_page(page); |
1024 | } | |
1025 | ||
1026 | if (PageWriteback(page)) { | |
11bc82d6 | 1027 | /* |
fed5b64a | 1028 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
1029 | * necessary to wait for PageWriteback. In the async case, |
1030 | * the retry loop is too short and in the sync-light case, | |
1031 | * the overhead of stalling is too much | |
11bc82d6 | 1032 | */ |
2916ecc0 JG |
1033 | switch (mode) { |
1034 | case MIGRATE_SYNC: | |
1035 | case MIGRATE_SYNC_NO_COPY: | |
1036 | break; | |
1037 | default: | |
11bc82d6 | 1038 | rc = -EBUSY; |
0a31bc97 | 1039 | goto out_unlock; |
11bc82d6 AA |
1040 | } |
1041 | if (!force) | |
0a31bc97 | 1042 | goto out_unlock; |
e24f0b8f CL |
1043 | wait_on_page_writeback(page); |
1044 | } | |
03f15c86 | 1045 | |
e24f0b8f | 1046 | /* |
dc386d4d KH |
1047 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1048 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 1049 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 1050 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
1051 | * File Caches may use write_page() or lock_page() in migration, then, |
1052 | * just care Anon page here. | |
03f15c86 HD |
1053 | * |
1054 | * Only page_get_anon_vma() understands the subtleties of | |
1055 | * getting a hold on an anon_vma from outside one of its mms. | |
1056 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1057 | * because that implies that the anon page is no longer mapped | |
1058 | * (and cannot be remapped so long as we hold the page lock). | |
dc386d4d | 1059 | */ |
03f15c86 | 1060 | if (PageAnon(page) && !PageKsm(page)) |
746b18d4 | 1061 | anon_vma = page_get_anon_vma(page); |
62e1c553 | 1062 | |
7db7671f HD |
1063 | /* |
1064 | * Block others from accessing the new page when we get around to | |
1065 | * establishing additional references. We are usually the only one | |
1066 | * holding a reference to newpage at this point. We used to have a BUG | |
1067 | * here if trylock_page(newpage) fails, but would like to allow for | |
1068 | * cases where there might be a race with the previous use of newpage. | |
1069 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1070 | */ | |
1071 | if (unlikely(!trylock_page(newpage))) | |
1072 | goto out_unlock; | |
1073 | ||
bda807d4 MK |
1074 | if (unlikely(!is_lru)) { |
1075 | rc = move_to_new_page(newpage, page, mode); | |
1076 | goto out_unlock_both; | |
1077 | } | |
1078 | ||
dc386d4d | 1079 | /* |
62e1c553 SL |
1080 | * Corner case handling: |
1081 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1082 | * and treated as swapcache but it has no rmap yet. | |
1083 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1084 | * trigger a BUG. So handle it here. | |
d12b8951 | 1085 | * 2. An orphaned page (see truncate_cleanup_page) might have |
62e1c553 SL |
1086 | * fs-private metadata. The page can be picked up due to memory |
1087 | * offlining. Everywhere else except page reclaim, the page is | |
1088 | * invisible to the vm, so the page can not be migrated. So try to | |
1089 | * free the metadata, so the page can be freed. | |
e24f0b8f | 1090 | */ |
62e1c553 | 1091 | if (!page->mapping) { |
309381fe | 1092 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1ce82b69 | 1093 | if (page_has_private(page)) { |
62e1c553 | 1094 | try_to_free_buffers(page); |
7db7671f | 1095 | goto out_unlock_both; |
62e1c553 | 1096 | } |
7db7671f HD |
1097 | } else if (page_mapped(page)) { |
1098 | /* Establish migration ptes */ | |
03f15c86 HD |
1099 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1100 | page); | |
013339df | 1101 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK); |
2ebba6b7 HD |
1102 | page_was_mapped = 1; |
1103 | } | |
dc386d4d | 1104 | |
e6a1530d | 1105 | if (!page_mapped(page)) |
5c3f9a67 | 1106 | rc = move_to_new_page(newpage, page, mode); |
e24f0b8f | 1107 | |
5c3f9a67 HD |
1108 | if (page_was_mapped) |
1109 | remove_migration_ptes(page, | |
e388466d | 1110 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
3f6c8272 | 1111 | |
7db7671f HD |
1112 | out_unlock_both: |
1113 | unlock_page(newpage); | |
1114 | out_unlock: | |
3f6c8272 | 1115 | /* Drop an anon_vma reference if we took one */ |
76545066 | 1116 | if (anon_vma) |
9e60109f | 1117 | put_anon_vma(anon_vma); |
e24f0b8f | 1118 | unlock_page(page); |
0dabec93 | 1119 | out: |
c6c919eb MK |
1120 | /* |
1121 | * If migration is successful, decrease refcount of the newpage | |
1122 | * which will not free the page because new page owner increased | |
1123 | * refcounter. As well, if it is LRU page, add the page to LRU | |
e0a352fa DH |
1124 | * list in here. Use the old state of the isolated source page to |
1125 | * determine if we migrated a LRU page. newpage was already unlocked | |
1126 | * and possibly modified by its owner - don't rely on the page | |
1127 | * state. | |
c6c919eb MK |
1128 | */ |
1129 | if (rc == MIGRATEPAGE_SUCCESS) { | |
e0a352fa | 1130 | if (unlikely(!is_lru)) |
c6c919eb MK |
1131 | put_page(newpage); |
1132 | else | |
1133 | putback_lru_page(newpage); | |
1134 | } | |
1135 | ||
0dabec93 MK |
1136 | return rc; |
1137 | } | |
95a402c3 | 1138 | |
0dabec93 MK |
1139 | /* |
1140 | * Obtain the lock on page, remove all ptes and migrate the page | |
1141 | * to the newly allocated page in newpage. | |
1142 | */ | |
6ec4476a | 1143 | static int unmap_and_move(new_page_t get_new_page, |
ef2a5153 GU |
1144 | free_page_t put_new_page, |
1145 | unsigned long private, struct page *page, | |
add05cec | 1146 | int force, enum migrate_mode mode, |
dd4ae78a YS |
1147 | enum migrate_reason reason, |
1148 | struct list_head *ret) | |
0dabec93 | 1149 | { |
2def7424 | 1150 | int rc = MIGRATEPAGE_SUCCESS; |
74d4a579 | 1151 | struct page *newpage = NULL; |
0dabec93 | 1152 | |
94723aaf | 1153 | if (!thp_migration_supported() && PageTransHuge(page)) |
d532e2e5 | 1154 | return -ENOSYS; |
94723aaf | 1155 | |
0dabec93 MK |
1156 | if (page_count(page) == 1) { |
1157 | /* page was freed from under us. So we are done. */ | |
c6c919eb MK |
1158 | ClearPageActive(page); |
1159 | ClearPageUnevictable(page); | |
bda807d4 MK |
1160 | if (unlikely(__PageMovable(page))) { |
1161 | lock_page(page); | |
1162 | if (!PageMovable(page)) | |
1163 | __ClearPageIsolated(page); | |
1164 | unlock_page(page); | |
1165 | } | |
0dabec93 MK |
1166 | goto out; |
1167 | } | |
1168 | ||
74d4a579 YS |
1169 | newpage = get_new_page(page, private); |
1170 | if (!newpage) | |
1171 | return -ENOMEM; | |
1172 | ||
9c620e2b | 1173 | rc = __unmap_and_move(page, newpage, force, mode); |
c6c919eb | 1174 | if (rc == MIGRATEPAGE_SUCCESS) |
7cd12b4a | 1175 | set_page_owner_migrate_reason(newpage, reason); |
bf6bddf1 | 1176 | |
0dabec93 | 1177 | out: |
e24f0b8f | 1178 | if (rc != -EAGAIN) { |
0dabec93 MK |
1179 | /* |
1180 | * A page that has been migrated has all references | |
1181 | * removed and will be freed. A page that has not been | |
c23a0c99 | 1182 | * migrated will have kept its references and be restored. |
0dabec93 MK |
1183 | */ |
1184 | list_del(&page->lru); | |
dd4ae78a | 1185 | } |
6afcf8ef | 1186 | |
dd4ae78a YS |
1187 | /* |
1188 | * If migration is successful, releases reference grabbed during | |
1189 | * isolation. Otherwise, restore the page to right list unless | |
1190 | * we want to retry. | |
1191 | */ | |
1192 | if (rc == MIGRATEPAGE_SUCCESS) { | |
6afcf8ef ML |
1193 | /* |
1194 | * Compaction can migrate also non-LRU pages which are | |
1195 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1196 | * as __PageMovable | |
1197 | */ | |
1198 | if (likely(!__PageMovable(page))) | |
e8db67eb | 1199 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
6c357848 | 1200 | page_is_file_lru(page), -thp_nr_pages(page)); |
c6c919eb | 1201 | |
79f5f8fa | 1202 | if (reason != MR_MEMORY_FAILURE) |
d7e69488 | 1203 | /* |
79f5f8fa | 1204 | * We release the page in page_handle_poison. |
d7e69488 | 1205 | */ |
79f5f8fa | 1206 | put_page(page); |
c6c919eb | 1207 | } else { |
dd4ae78a YS |
1208 | if (rc != -EAGAIN) |
1209 | list_add_tail(&page->lru, ret); | |
bda807d4 | 1210 | |
c6c919eb MK |
1211 | if (put_new_page) |
1212 | put_new_page(newpage, private); | |
1213 | else | |
1214 | put_page(newpage); | |
e24f0b8f | 1215 | } |
68711a74 | 1216 | |
e24f0b8f CL |
1217 | return rc; |
1218 | } | |
1219 | ||
290408d4 NH |
1220 | /* |
1221 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1222 | * | |
1223 | * This function doesn't wait the completion of hugepage I/O | |
1224 | * because there is no race between I/O and migration for hugepage. | |
1225 | * Note that currently hugepage I/O occurs only in direct I/O | |
1226 | * where no lock is held and PG_writeback is irrelevant, | |
1227 | * and writeback status of all subpages are counted in the reference | |
1228 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1229 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1230 | * This means that when we try to migrate hugepage whose subpages are | |
1231 | * doing direct I/O, some references remain after try_to_unmap() and | |
1232 | * hugepage migration fails without data corruption. | |
1233 | * | |
1234 | * There is also no race when direct I/O is issued on the page under migration, | |
1235 | * because then pte is replaced with migration swap entry and direct I/O code | |
1236 | * will wait in the page fault for migration to complete. | |
1237 | */ | |
1238 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
68711a74 DR |
1239 | free_page_t put_new_page, unsigned long private, |
1240 | struct page *hpage, int force, | |
dd4ae78a YS |
1241 | enum migrate_mode mode, int reason, |
1242 | struct list_head *ret) | |
290408d4 | 1243 | { |
2def7424 | 1244 | int rc = -EAGAIN; |
2ebba6b7 | 1245 | int page_was_mapped = 0; |
32665f2b | 1246 | struct page *new_hpage; |
290408d4 | 1247 | struct anon_vma *anon_vma = NULL; |
c0d0381a | 1248 | struct address_space *mapping = NULL; |
290408d4 | 1249 | |
83467efb | 1250 | /* |
7ed2c31d | 1251 | * Migratability of hugepages depends on architectures and their size. |
83467efb NH |
1252 | * This check is necessary because some callers of hugepage migration |
1253 | * like soft offline and memory hotremove don't walk through page | |
1254 | * tables or check whether the hugepage is pmd-based or not before | |
1255 | * kicking migration. | |
1256 | */ | |
100873d7 | 1257 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
dd4ae78a | 1258 | list_move_tail(&hpage->lru, ret); |
83467efb | 1259 | return -ENOSYS; |
32665f2b | 1260 | } |
83467efb | 1261 | |
71a64f61 MS |
1262 | if (page_count(hpage) == 1) { |
1263 | /* page was freed from under us. So we are done. */ | |
1264 | putback_active_hugepage(hpage); | |
1265 | return MIGRATEPAGE_SUCCESS; | |
1266 | } | |
1267 | ||
666feb21 | 1268 | new_hpage = get_new_page(hpage, private); |
290408d4 NH |
1269 | if (!new_hpage) |
1270 | return -ENOMEM; | |
1271 | ||
290408d4 | 1272 | if (!trylock_page(hpage)) { |
2916ecc0 | 1273 | if (!force) |
290408d4 | 1274 | goto out; |
2916ecc0 JG |
1275 | switch (mode) { |
1276 | case MIGRATE_SYNC: | |
1277 | case MIGRATE_SYNC_NO_COPY: | |
1278 | break; | |
1279 | default: | |
1280 | goto out; | |
1281 | } | |
290408d4 NH |
1282 | lock_page(hpage); |
1283 | } | |
1284 | ||
cb6acd01 MK |
1285 | /* |
1286 | * Check for pages which are in the process of being freed. Without | |
1287 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1288 | * be called and we could leak usage counts for subpools. | |
1289 | */ | |
1290 | if (page_private(hpage) && !page_mapping(hpage)) { | |
1291 | rc = -EBUSY; | |
1292 | goto out_unlock; | |
1293 | } | |
1294 | ||
746b18d4 PZ |
1295 | if (PageAnon(hpage)) |
1296 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 | 1297 | |
7db7671f HD |
1298 | if (unlikely(!trylock_page(new_hpage))) |
1299 | goto put_anon; | |
1300 | ||
2ebba6b7 | 1301 | if (page_mapped(hpage)) { |
336bf30e | 1302 | bool mapping_locked = false; |
013339df | 1303 | enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK; |
336bf30e MK |
1304 | |
1305 | if (!PageAnon(hpage)) { | |
1306 | /* | |
1307 | * In shared mappings, try_to_unmap could potentially | |
1308 | * call huge_pmd_unshare. Because of this, take | |
1309 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1310 | * to let lower levels know we have taken the lock. | |
1311 | */ | |
1312 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1313 | if (unlikely(!mapping)) | |
1314 | goto unlock_put_anon; | |
1315 | ||
1316 | mapping_locked = true; | |
1317 | ttu |= TTU_RMAP_LOCKED; | |
1318 | } | |
c0d0381a | 1319 | |
336bf30e | 1320 | try_to_unmap(hpage, ttu); |
2ebba6b7 | 1321 | page_was_mapped = 1; |
336bf30e MK |
1322 | |
1323 | if (mapping_locked) | |
1324 | i_mmap_unlock_write(mapping); | |
2ebba6b7 | 1325 | } |
290408d4 NH |
1326 | |
1327 | if (!page_mapped(hpage)) | |
5c3f9a67 | 1328 | rc = move_to_new_page(new_hpage, hpage, mode); |
290408d4 | 1329 | |
336bf30e | 1330 | if (page_was_mapped) |
5c3f9a67 | 1331 | remove_migration_ptes(hpage, |
336bf30e | 1332 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
290408d4 | 1333 | |
c0d0381a | 1334 | unlock_put_anon: |
7db7671f HD |
1335 | unlock_page(new_hpage); |
1336 | ||
1337 | put_anon: | |
fd4a4663 | 1338 | if (anon_vma) |
9e60109f | 1339 | put_anon_vma(anon_vma); |
8e6ac7fa | 1340 | |
2def7424 | 1341 | if (rc == MIGRATEPAGE_SUCCESS) { |
ab5ac90a | 1342 | move_hugetlb_state(hpage, new_hpage, reason); |
2def7424 HD |
1343 | put_new_page = NULL; |
1344 | } | |
8e6ac7fa | 1345 | |
cb6acd01 | 1346 | out_unlock: |
290408d4 | 1347 | unlock_page(hpage); |
09761333 | 1348 | out: |
dd4ae78a | 1349 | if (rc == MIGRATEPAGE_SUCCESS) |
b8ec1cee | 1350 | putback_active_hugepage(hpage); |
dd4ae78a YS |
1351 | else if (rc != -EAGAIN && rc != MIGRATEPAGE_SUCCESS) |
1352 | list_move_tail(&hpage->lru, ret); | |
68711a74 DR |
1353 | |
1354 | /* | |
1355 | * If migration was not successful and there's a freeing callback, use | |
1356 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1357 | * isolation. | |
1358 | */ | |
2def7424 | 1359 | if (put_new_page) |
68711a74 DR |
1360 | put_new_page(new_hpage, private); |
1361 | else | |
3aaa76e1 | 1362 | putback_active_hugepage(new_hpage); |
68711a74 | 1363 | |
290408d4 NH |
1364 | return rc; |
1365 | } | |
1366 | ||
d532e2e5 YS |
1367 | static inline int try_split_thp(struct page *page, struct page **page2, |
1368 | struct list_head *from) | |
1369 | { | |
1370 | int rc = 0; | |
1371 | ||
1372 | lock_page(page); | |
1373 | rc = split_huge_page_to_list(page, from); | |
1374 | unlock_page(page); | |
1375 | if (!rc) | |
1376 | list_safe_reset_next(page, *page2, lru); | |
1377 | ||
1378 | return rc; | |
1379 | } | |
1380 | ||
b20a3503 | 1381 | /* |
c73e5c9c SB |
1382 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1383 | * supplied as the target for the page migration | |
b20a3503 | 1384 | * |
c73e5c9c SB |
1385 | * @from: The list of pages to be migrated. |
1386 | * @get_new_page: The function used to allocate free pages to be used | |
1387 | * as the target of the page migration. | |
68711a74 DR |
1388 | * @put_new_page: The function used to free target pages if migration |
1389 | * fails, or NULL if no special handling is necessary. | |
c73e5c9c SB |
1390 | * @private: Private data to be passed on to get_new_page() |
1391 | * @mode: The migration mode that specifies the constraints for | |
1392 | * page migration, if any. | |
1393 | * @reason: The reason for page migration. | |
b20a3503 | 1394 | * |
c73e5c9c SB |
1395 | * The function returns after 10 attempts or if no pages are movable any more |
1396 | * because the list has become empty or no retryable pages exist any more. | |
dd4ae78a YS |
1397 | * It is caller's responsibility to call putback_movable_pages() to return pages |
1398 | * to the LRU or free list only if ret != 0. | |
b20a3503 | 1399 | * |
c73e5c9c | 1400 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1401 | */ |
9c620e2b | 1402 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
68711a74 DR |
1403 | free_page_t put_new_page, unsigned long private, |
1404 | enum migrate_mode mode, int reason) | |
b20a3503 | 1405 | { |
e24f0b8f | 1406 | int retry = 1; |
1a5bae25 | 1407 | int thp_retry = 1; |
b20a3503 | 1408 | int nr_failed = 0; |
5647bc29 | 1409 | int nr_succeeded = 0; |
1a5bae25 AK |
1410 | int nr_thp_succeeded = 0; |
1411 | int nr_thp_failed = 0; | |
1412 | int nr_thp_split = 0; | |
b20a3503 | 1413 | int pass = 0; |
1a5bae25 | 1414 | bool is_thp = false; |
b20a3503 CL |
1415 | struct page *page; |
1416 | struct page *page2; | |
1417 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1a5bae25 | 1418 | int rc, nr_subpages; |
dd4ae78a | 1419 | LIST_HEAD(ret_pages); |
b20a3503 CL |
1420 | |
1421 | if (!swapwrite) | |
1422 | current->flags |= PF_SWAPWRITE; | |
1423 | ||
1a5bae25 | 1424 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { |
e24f0b8f | 1425 | retry = 0; |
1a5bae25 | 1426 | thp_retry = 0; |
b20a3503 | 1427 | |
e24f0b8f | 1428 | list_for_each_entry_safe(page, page2, from, lru) { |
94723aaf | 1429 | retry: |
1a5bae25 AK |
1430 | /* |
1431 | * THP statistics is based on the source huge page. | |
1432 | * Capture required information that might get lost | |
1433 | * during migration. | |
1434 | */ | |
6c5c7b9f | 1435 | is_thp = PageTransHuge(page) && !PageHuge(page); |
6c357848 | 1436 | nr_subpages = thp_nr_pages(page); |
e24f0b8f | 1437 | cond_resched(); |
2d1db3b1 | 1438 | |
31caf665 NH |
1439 | if (PageHuge(page)) |
1440 | rc = unmap_and_move_huge_page(get_new_page, | |
68711a74 | 1441 | put_new_page, private, page, |
dd4ae78a YS |
1442 | pass > 2, mode, reason, |
1443 | &ret_pages); | |
31caf665 | 1444 | else |
68711a74 | 1445 | rc = unmap_and_move(get_new_page, put_new_page, |
add05cec | 1446 | private, page, pass > 2, mode, |
dd4ae78a YS |
1447 | reason, &ret_pages); |
1448 | /* | |
1449 | * The rules are: | |
1450 | * Success: non hugetlb page will be freed, hugetlb | |
1451 | * page will be put back | |
1452 | * -EAGAIN: stay on the from list | |
1453 | * -ENOMEM: stay on the from list | |
1454 | * Other errno: put on ret_pages list then splice to | |
1455 | * from list | |
1456 | */ | |
e24f0b8f | 1457 | switch(rc) { |
d532e2e5 YS |
1458 | /* |
1459 | * THP migration might be unsupported or the | |
1460 | * allocation could've failed so we should | |
1461 | * retry on the same page with the THP split | |
1462 | * to base pages. | |
1463 | * | |
1464 | * Head page is retried immediately and tail | |
1465 | * pages are added to the tail of the list so | |
1466 | * we encounter them after the rest of the list | |
1467 | * is processed. | |
1468 | */ | |
1469 | case -ENOSYS: | |
1470 | /* THP migration is unsupported */ | |
1471 | if (is_thp) { | |
1472 | if (!try_split_thp(page, &page2, from)) { | |
1473 | nr_thp_split++; | |
1474 | goto retry; | |
1475 | } | |
1476 | ||
1477 | nr_thp_failed++; | |
1478 | nr_failed += nr_subpages; | |
1479 | break; | |
1480 | } | |
1481 | ||
1482 | /* Hugetlb migration is unsupported */ | |
1483 | nr_failed++; | |
1484 | break; | |
95a402c3 | 1485 | case -ENOMEM: |
94723aaf | 1486 | /* |
d532e2e5 YS |
1487 | * When memory is low, don't bother to try to migrate |
1488 | * other pages, just exit. | |
94723aaf | 1489 | */ |
6c5c7b9f | 1490 | if (is_thp) { |
d532e2e5 | 1491 | if (!try_split_thp(page, &page2, from)) { |
1a5bae25 | 1492 | nr_thp_split++; |
94723aaf MH |
1493 | goto retry; |
1494 | } | |
6c5c7b9f | 1495 | |
1a5bae25 AK |
1496 | nr_thp_failed++; |
1497 | nr_failed += nr_subpages; | |
1498 | goto out; | |
1499 | } | |
dfef2ef4 | 1500 | nr_failed++; |
95a402c3 | 1501 | goto out; |
e24f0b8f | 1502 | case -EAGAIN: |
1a5bae25 AK |
1503 | if (is_thp) { |
1504 | thp_retry++; | |
1505 | break; | |
1506 | } | |
2d1db3b1 | 1507 | retry++; |
e24f0b8f | 1508 | break; |
78bd5209 | 1509 | case MIGRATEPAGE_SUCCESS: |
1a5bae25 AK |
1510 | if (is_thp) { |
1511 | nr_thp_succeeded++; | |
1512 | nr_succeeded += nr_subpages; | |
1513 | break; | |
1514 | } | |
5647bc29 | 1515 | nr_succeeded++; |
e24f0b8f CL |
1516 | break; |
1517 | default: | |
354a3363 | 1518 | /* |
d532e2e5 | 1519 | * Permanent failure (-EBUSY, etc.): |
354a3363 NH |
1520 | * unlike -EAGAIN case, the failed page is |
1521 | * removed from migration page list and not | |
1522 | * retried in the next outer loop. | |
1523 | */ | |
1a5bae25 AK |
1524 | if (is_thp) { |
1525 | nr_thp_failed++; | |
1526 | nr_failed += nr_subpages; | |
1527 | break; | |
1528 | } | |
2d1db3b1 | 1529 | nr_failed++; |
e24f0b8f | 1530 | break; |
2d1db3b1 | 1531 | } |
b20a3503 CL |
1532 | } |
1533 | } | |
1a5bae25 AK |
1534 | nr_failed += retry + thp_retry; |
1535 | nr_thp_failed += thp_retry; | |
f2f81fb2 | 1536 | rc = nr_failed; |
95a402c3 | 1537 | out: |
dd4ae78a YS |
1538 | /* |
1539 | * Put the permanent failure page back to migration list, they | |
1540 | * will be put back to the right list by the caller. | |
1541 | */ | |
1542 | list_splice(&ret_pages, from); | |
1543 | ||
1a5bae25 AK |
1544 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
1545 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1546 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); | |
1547 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); | |
1548 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); | |
1549 | trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded, | |
1550 | nr_thp_failed, nr_thp_split, mode, reason); | |
7b2a2d4a | 1551 | |
b20a3503 CL |
1552 | if (!swapwrite) |
1553 | current->flags &= ~PF_SWAPWRITE; | |
1554 | ||
78bd5209 | 1555 | return rc; |
b20a3503 | 1556 | } |
95a402c3 | 1557 | |
19fc7bed | 1558 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
b4b38223 | 1559 | { |
19fc7bed JK |
1560 | struct migration_target_control *mtc; |
1561 | gfp_t gfp_mask; | |
b4b38223 JK |
1562 | unsigned int order = 0; |
1563 | struct page *new_page = NULL; | |
19fc7bed JK |
1564 | int nid; |
1565 | int zidx; | |
1566 | ||
1567 | mtc = (struct migration_target_control *)private; | |
1568 | gfp_mask = mtc->gfp_mask; | |
1569 | nid = mtc->nid; | |
1570 | if (nid == NUMA_NO_NODE) | |
1571 | nid = page_to_nid(page); | |
b4b38223 | 1572 | |
d92bbc27 JK |
1573 | if (PageHuge(page)) { |
1574 | struct hstate *h = page_hstate(compound_head(page)); | |
1575 | ||
19fc7bed JK |
1576 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
1577 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); | |
d92bbc27 | 1578 | } |
b4b38223 JK |
1579 | |
1580 | if (PageTransHuge(page)) { | |
9933a0c8 JK |
1581 | /* |
1582 | * clear __GFP_RECLAIM to make the migration callback | |
1583 | * consistent with regular THP allocations. | |
1584 | */ | |
1585 | gfp_mask &= ~__GFP_RECLAIM; | |
b4b38223 JK |
1586 | gfp_mask |= GFP_TRANSHUGE; |
1587 | order = HPAGE_PMD_ORDER; | |
1588 | } | |
19fc7bed JK |
1589 | zidx = zone_idx(page_zone(page)); |
1590 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) | |
b4b38223 JK |
1591 | gfp_mask |= __GFP_HIGHMEM; |
1592 | ||
84172f4b | 1593 | new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask); |
b4b38223 JK |
1594 | |
1595 | if (new_page && PageTransHuge(new_page)) | |
1596 | prep_transhuge_page(new_page); | |
1597 | ||
1598 | return new_page; | |
1599 | } | |
1600 | ||
742755a1 | 1601 | #ifdef CONFIG_NUMA |
742755a1 | 1602 | |
a49bd4d7 | 1603 | static int store_status(int __user *status, int start, int value, int nr) |
742755a1 | 1604 | { |
a49bd4d7 MH |
1605 | while (nr-- > 0) { |
1606 | if (put_user(value, status + start)) | |
1607 | return -EFAULT; | |
1608 | start++; | |
1609 | } | |
1610 | ||
1611 | return 0; | |
1612 | } | |
1613 | ||
1614 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1615 | struct list_head *pagelist, int node) | |
1616 | { | |
1617 | int err; | |
a0976311 JK |
1618 | struct migration_target_control mtc = { |
1619 | .nid = node, | |
1620 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
1621 | }; | |
a49bd4d7 | 1622 | |
a0976311 JK |
1623 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
1624 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL); | |
a49bd4d7 MH |
1625 | if (err) |
1626 | putback_movable_pages(pagelist); | |
1627 | return err; | |
742755a1 CL |
1628 | } |
1629 | ||
1630 | /* | |
a49bd4d7 MH |
1631 | * Resolves the given address to a struct page, isolates it from the LRU and |
1632 | * puts it to the given pagelist. | |
e0153fc2 YS |
1633 | * Returns: |
1634 | * errno - if the page cannot be found/isolated | |
1635 | * 0 - when it doesn't have to be migrated because it is already on the | |
1636 | * target node | |
1637 | * 1 - when it has been queued | |
742755a1 | 1638 | */ |
a49bd4d7 MH |
1639 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1640 | int node, struct list_head *pagelist, bool migrate_all) | |
742755a1 | 1641 | { |
a49bd4d7 MH |
1642 | struct vm_area_struct *vma; |
1643 | struct page *page; | |
1644 | unsigned int follflags; | |
742755a1 | 1645 | int err; |
742755a1 | 1646 | |
d8ed45c5 | 1647 | mmap_read_lock(mm); |
a49bd4d7 MH |
1648 | err = -EFAULT; |
1649 | vma = find_vma(mm, addr); | |
1650 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1651 | goto out; | |
742755a1 | 1652 | |
a49bd4d7 MH |
1653 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1654 | follflags = FOLL_GET | FOLL_DUMP; | |
a49bd4d7 | 1655 | page = follow_page(vma, addr, follflags); |
89f5b7da | 1656 | |
a49bd4d7 MH |
1657 | err = PTR_ERR(page); |
1658 | if (IS_ERR(page)) | |
1659 | goto out; | |
89f5b7da | 1660 | |
a49bd4d7 MH |
1661 | err = -ENOENT; |
1662 | if (!page) | |
1663 | goto out; | |
742755a1 | 1664 | |
a49bd4d7 MH |
1665 | err = 0; |
1666 | if (page_to_nid(page) == node) | |
1667 | goto out_putpage; | |
742755a1 | 1668 | |
a49bd4d7 MH |
1669 | err = -EACCES; |
1670 | if (page_mapcount(page) > 1 && !migrate_all) | |
1671 | goto out_putpage; | |
742755a1 | 1672 | |
a49bd4d7 MH |
1673 | if (PageHuge(page)) { |
1674 | if (PageHead(page)) { | |
1675 | isolate_huge_page(page, pagelist); | |
e0153fc2 | 1676 | err = 1; |
e632a938 | 1677 | } |
a49bd4d7 MH |
1678 | } else { |
1679 | struct page *head; | |
e632a938 | 1680 | |
e8db67eb NH |
1681 | head = compound_head(page); |
1682 | err = isolate_lru_page(head); | |
cf608ac1 | 1683 | if (err) |
a49bd4d7 | 1684 | goto out_putpage; |
742755a1 | 1685 | |
e0153fc2 | 1686 | err = 1; |
a49bd4d7 MH |
1687 | list_add_tail(&head->lru, pagelist); |
1688 | mod_node_page_state(page_pgdat(head), | |
9de4f22a | 1689 | NR_ISOLATED_ANON + page_is_file_lru(head), |
6c357848 | 1690 | thp_nr_pages(head)); |
a49bd4d7 MH |
1691 | } |
1692 | out_putpage: | |
1693 | /* | |
1694 | * Either remove the duplicate refcount from | |
1695 | * isolate_lru_page() or drop the page ref if it was | |
1696 | * not isolated. | |
1697 | */ | |
1698 | put_page(page); | |
1699 | out: | |
d8ed45c5 | 1700 | mmap_read_unlock(mm); |
742755a1 CL |
1701 | return err; |
1702 | } | |
1703 | ||
7ca8783a WY |
1704 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
1705 | struct list_head *pagelist, int __user *status, | |
1706 | int start, int i, unsigned long nr_pages) | |
1707 | { | |
1708 | int err; | |
1709 | ||
5d7ae891 WY |
1710 | if (list_empty(pagelist)) |
1711 | return 0; | |
1712 | ||
7ca8783a WY |
1713 | err = do_move_pages_to_node(mm, pagelist, node); |
1714 | if (err) { | |
1715 | /* | |
1716 | * Positive err means the number of failed | |
1717 | * pages to migrate. Since we are going to | |
1718 | * abort and return the number of non-migrated | |
ab9dd4f8 | 1719 | * pages, so need to include the rest of the |
7ca8783a WY |
1720 | * nr_pages that have not been attempted as |
1721 | * well. | |
1722 | */ | |
1723 | if (err > 0) | |
1724 | err += nr_pages - i - 1; | |
1725 | return err; | |
1726 | } | |
1727 | return store_status(status, start, node, i - start); | |
1728 | } | |
1729 | ||
5e9a0f02 BG |
1730 | /* |
1731 | * Migrate an array of page address onto an array of nodes and fill | |
1732 | * the corresponding array of status. | |
1733 | */ | |
3268c63e | 1734 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1735 | unsigned long nr_pages, |
1736 | const void __user * __user *pages, | |
1737 | const int __user *nodes, | |
1738 | int __user *status, int flags) | |
1739 | { | |
a49bd4d7 MH |
1740 | int current_node = NUMA_NO_NODE; |
1741 | LIST_HEAD(pagelist); | |
1742 | int start, i; | |
1743 | int err = 0, err1; | |
35282a2d | 1744 | |
361a2a22 | 1745 | lru_cache_disable(); |
35282a2d | 1746 | |
a49bd4d7 MH |
1747 | for (i = start = 0; i < nr_pages; i++) { |
1748 | const void __user *p; | |
1749 | unsigned long addr; | |
1750 | int node; | |
3140a227 | 1751 | |
a49bd4d7 MH |
1752 | err = -EFAULT; |
1753 | if (get_user(p, pages + i)) | |
1754 | goto out_flush; | |
1755 | if (get_user(node, nodes + i)) | |
1756 | goto out_flush; | |
057d3389 | 1757 | addr = (unsigned long)untagged_addr(p); |
a49bd4d7 MH |
1758 | |
1759 | err = -ENODEV; | |
1760 | if (node < 0 || node >= MAX_NUMNODES) | |
1761 | goto out_flush; | |
1762 | if (!node_state(node, N_MEMORY)) | |
1763 | goto out_flush; | |
5e9a0f02 | 1764 | |
a49bd4d7 MH |
1765 | err = -EACCES; |
1766 | if (!node_isset(node, task_nodes)) | |
1767 | goto out_flush; | |
1768 | ||
1769 | if (current_node == NUMA_NO_NODE) { | |
1770 | current_node = node; | |
1771 | start = i; | |
1772 | } else if (node != current_node) { | |
7ca8783a WY |
1773 | err = move_pages_and_store_status(mm, current_node, |
1774 | &pagelist, status, start, i, nr_pages); | |
a49bd4d7 MH |
1775 | if (err) |
1776 | goto out; | |
1777 | start = i; | |
1778 | current_node = node; | |
3140a227 BG |
1779 | } |
1780 | ||
a49bd4d7 MH |
1781 | /* |
1782 | * Errors in the page lookup or isolation are not fatal and we simply | |
1783 | * report them via status | |
1784 | */ | |
1785 | err = add_page_for_migration(mm, addr, current_node, | |
1786 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
e0153fc2 | 1787 | |
d08221a0 | 1788 | if (err > 0) { |
e0153fc2 YS |
1789 | /* The page is successfully queued for migration */ |
1790 | continue; | |
1791 | } | |
3140a227 | 1792 | |
d08221a0 WY |
1793 | /* |
1794 | * If the page is already on the target node (!err), store the | |
1795 | * node, otherwise, store the err. | |
1796 | */ | |
1797 | err = store_status(status, i, err ? : current_node, 1); | |
a49bd4d7 MH |
1798 | if (err) |
1799 | goto out_flush; | |
5e9a0f02 | 1800 | |
7ca8783a WY |
1801 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
1802 | status, start, i, nr_pages); | |
4afdacec WY |
1803 | if (err) |
1804 | goto out; | |
a49bd4d7 | 1805 | current_node = NUMA_NO_NODE; |
3140a227 | 1806 | } |
a49bd4d7 MH |
1807 | out_flush: |
1808 | /* Make sure we do not overwrite the existing error */ | |
7ca8783a WY |
1809 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
1810 | status, start, i, nr_pages); | |
dfe9aa23 | 1811 | if (err >= 0) |
a49bd4d7 | 1812 | err = err1; |
5e9a0f02 | 1813 | out: |
361a2a22 | 1814 | lru_cache_enable(); |
5e9a0f02 BG |
1815 | return err; |
1816 | } | |
1817 | ||
742755a1 | 1818 | /* |
2f007e74 | 1819 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1820 | */ |
80bba129 BG |
1821 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1822 | const void __user **pages, int *status) | |
742755a1 | 1823 | { |
2f007e74 | 1824 | unsigned long i; |
2f007e74 | 1825 | |
d8ed45c5 | 1826 | mmap_read_lock(mm); |
742755a1 | 1827 | |
2f007e74 | 1828 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1829 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1830 | struct vm_area_struct *vma; |
1831 | struct page *page; | |
c095adbc | 1832 | int err = -EFAULT; |
2f007e74 BG |
1833 | |
1834 | vma = find_vma(mm, addr); | |
70384dc6 | 1835 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1836 | goto set_status; |
1837 | ||
d899844e KS |
1838 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1839 | page = follow_page(vma, addr, FOLL_DUMP); | |
89f5b7da LT |
1840 | |
1841 | err = PTR_ERR(page); | |
1842 | if (IS_ERR(page)) | |
1843 | goto set_status; | |
1844 | ||
d899844e | 1845 | err = page ? page_to_nid(page) : -ENOENT; |
742755a1 | 1846 | set_status: |
80bba129 BG |
1847 | *status = err; |
1848 | ||
1849 | pages++; | |
1850 | status++; | |
1851 | } | |
1852 | ||
d8ed45c5 | 1853 | mmap_read_unlock(mm); |
80bba129 BG |
1854 | } |
1855 | ||
1856 | /* | |
1857 | * Determine the nodes of a user array of pages and store it in | |
1858 | * a user array of status. | |
1859 | */ | |
1860 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1861 | const void __user * __user *pages, | |
1862 | int __user *status) | |
1863 | { | |
1864 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1865 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1866 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1867 | |
87b8d1ad PA |
1868 | while (nr_pages) { |
1869 | unsigned long chunk_nr; | |
80bba129 | 1870 | |
87b8d1ad PA |
1871 | chunk_nr = nr_pages; |
1872 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1873 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1874 | ||
1875 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1876 | break; | |
80bba129 BG |
1877 | |
1878 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1879 | ||
87b8d1ad PA |
1880 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1881 | break; | |
742755a1 | 1882 | |
87b8d1ad PA |
1883 | pages += chunk_nr; |
1884 | status += chunk_nr; | |
1885 | nr_pages -= chunk_nr; | |
1886 | } | |
1887 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1888 | } |
1889 | ||
4dc200ce | 1890 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
742755a1 | 1891 | { |
742755a1 | 1892 | struct task_struct *task; |
742755a1 | 1893 | struct mm_struct *mm; |
742755a1 | 1894 | |
4dc200ce ML |
1895 | /* |
1896 | * There is no need to check if current process has the right to modify | |
1897 | * the specified process when they are same. | |
1898 | */ | |
1899 | if (!pid) { | |
1900 | mmget(current->mm); | |
1901 | *mem_nodes = cpuset_mems_allowed(current); | |
1902 | return current->mm; | |
1903 | } | |
742755a1 CL |
1904 | |
1905 | /* Find the mm_struct */ | |
a879bf58 | 1906 | rcu_read_lock(); |
4dc200ce | 1907 | task = find_task_by_vpid(pid); |
742755a1 | 1908 | if (!task) { |
a879bf58 | 1909 | rcu_read_unlock(); |
4dc200ce | 1910 | return ERR_PTR(-ESRCH); |
742755a1 | 1911 | } |
3268c63e | 1912 | get_task_struct(task); |
742755a1 CL |
1913 | |
1914 | /* | |
1915 | * Check if this process has the right to modify the specified | |
197e7e52 | 1916 | * process. Use the regular "ptrace_may_access()" checks. |
742755a1 | 1917 | */ |
197e7e52 | 1918 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
c69e8d9c | 1919 | rcu_read_unlock(); |
4dc200ce | 1920 | mm = ERR_PTR(-EPERM); |
5e9a0f02 | 1921 | goto out; |
742755a1 | 1922 | } |
c69e8d9c | 1923 | rcu_read_unlock(); |
742755a1 | 1924 | |
4dc200ce ML |
1925 | mm = ERR_PTR(security_task_movememory(task)); |
1926 | if (IS_ERR(mm)) | |
5e9a0f02 | 1927 | goto out; |
4dc200ce | 1928 | *mem_nodes = cpuset_mems_allowed(task); |
3268c63e | 1929 | mm = get_task_mm(task); |
4dc200ce | 1930 | out: |
3268c63e | 1931 | put_task_struct(task); |
6e8b09ea | 1932 | if (!mm) |
4dc200ce ML |
1933 | mm = ERR_PTR(-EINVAL); |
1934 | return mm; | |
1935 | } | |
1936 | ||
1937 | /* | |
1938 | * Move a list of pages in the address space of the currently executing | |
1939 | * process. | |
1940 | */ | |
1941 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
1942 | const void __user * __user *pages, | |
1943 | const int __user *nodes, | |
1944 | int __user *status, int flags) | |
1945 | { | |
1946 | struct mm_struct *mm; | |
1947 | int err; | |
1948 | nodemask_t task_nodes; | |
1949 | ||
1950 | /* Check flags */ | |
1951 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
6e8b09ea SL |
1952 | return -EINVAL; |
1953 | ||
4dc200ce ML |
1954 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
1955 | return -EPERM; | |
1956 | ||
1957 | mm = find_mm_struct(pid, &task_nodes); | |
1958 | if (IS_ERR(mm)) | |
1959 | return PTR_ERR(mm); | |
1960 | ||
6e8b09ea SL |
1961 | if (nodes) |
1962 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1963 | nodes, status, flags); | |
1964 | else | |
1965 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1966 | |
742755a1 CL |
1967 | mmput(mm); |
1968 | return err; | |
1969 | } | |
742755a1 | 1970 | |
7addf443 DB |
1971 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1972 | const void __user * __user *, pages, | |
1973 | const int __user *, nodes, | |
1974 | int __user *, status, int, flags) | |
1975 | { | |
1976 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1977 | } | |
1978 | ||
1979 | #ifdef CONFIG_COMPAT | |
1980 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, | |
1981 | compat_uptr_t __user *, pages32, | |
1982 | const int __user *, nodes, | |
1983 | int __user *, status, | |
1984 | int, flags) | |
1985 | { | |
1986 | const void __user * __user *pages; | |
1987 | int i; | |
1988 | ||
1989 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); | |
1990 | for (i = 0; i < nr_pages; i++) { | |
1991 | compat_uptr_t p; | |
1992 | ||
1993 | if (get_user(p, pages32 + i) || | |
1994 | put_user(compat_ptr(p), pages + i)) | |
1995 | return -EFAULT; | |
1996 | } | |
1997 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1998 | } | |
1999 | #endif /* CONFIG_COMPAT */ | |
2000 | ||
7039e1db PZ |
2001 | #ifdef CONFIG_NUMA_BALANCING |
2002 | /* | |
2003 | * Returns true if this is a safe migration target node for misplaced NUMA | |
2004 | * pages. Currently it only checks the watermarks which crude | |
2005 | */ | |
2006 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 2007 | unsigned long nr_migrate_pages) |
7039e1db PZ |
2008 | { |
2009 | int z; | |
599d0c95 | 2010 | |
7039e1db PZ |
2011 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
2012 | struct zone *zone = pgdat->node_zones + z; | |
2013 | ||
2014 | if (!populated_zone(zone)) | |
2015 | continue; | |
2016 | ||
7039e1db PZ |
2017 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
2018 | if (!zone_watermark_ok(zone, 0, | |
2019 | high_wmark_pages(zone) + | |
2020 | nr_migrate_pages, | |
bfe9d006 | 2021 | ZONE_MOVABLE, 0)) |
7039e1db PZ |
2022 | continue; |
2023 | return true; | |
2024 | } | |
2025 | return false; | |
2026 | } | |
2027 | ||
2028 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
666feb21 | 2029 | unsigned long data) |
7039e1db PZ |
2030 | { |
2031 | int nid = (int) data; | |
2032 | struct page *newpage; | |
2033 | ||
96db800f | 2034 | newpage = __alloc_pages_node(nid, |
e97ca8e5 JW |
2035 | (GFP_HIGHUSER_MOVABLE | |
2036 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
2037 | __GFP_NORETRY | __GFP_NOWARN) & | |
8479eba7 | 2038 | ~__GFP_RECLAIM, 0); |
bac0382c | 2039 | |
7039e1db PZ |
2040 | return newpage; |
2041 | } | |
2042 | ||
1c30e017 | 2043 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
b32967ff | 2044 | { |
340ef390 | 2045 | int page_lru; |
a8f60772 | 2046 | |
309381fe | 2047 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
3abef4e6 | 2048 | |
7039e1db | 2049 | /* Avoid migrating to a node that is nearly full */ |
d8c6546b | 2050 | if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) |
340ef390 | 2051 | return 0; |
7039e1db | 2052 | |
340ef390 HD |
2053 | if (isolate_lru_page(page)) |
2054 | return 0; | |
7039e1db | 2055 | |
340ef390 HD |
2056 | /* |
2057 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
2058 | * check on page_count(), so we must do it here, now that the page | |
2059 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
2060 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
2061 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
2062 | */ | |
2063 | if (PageTransHuge(page) && page_count(page) != 3) { | |
2064 | putback_lru_page(page); | |
2065 | return 0; | |
7039e1db PZ |
2066 | } |
2067 | ||
9de4f22a | 2068 | page_lru = page_is_file_lru(page); |
599d0c95 | 2069 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
6c357848 | 2070 | thp_nr_pages(page)); |
340ef390 | 2071 | |
149c33e1 | 2072 | /* |
340ef390 HD |
2073 | * Isolating the page has taken another reference, so the |
2074 | * caller's reference can be safely dropped without the page | |
2075 | * disappearing underneath us during migration. | |
149c33e1 MG |
2076 | */ |
2077 | put_page(page); | |
340ef390 | 2078 | return 1; |
b32967ff MG |
2079 | } |
2080 | ||
de466bd6 MG |
2081 | bool pmd_trans_migrating(pmd_t pmd) |
2082 | { | |
2083 | struct page *page = pmd_page(pmd); | |
2084 | return PageLocked(page); | |
2085 | } | |
2086 | ||
c77c5cba YS |
2087 | static inline bool is_shared_exec_page(struct vm_area_struct *vma, |
2088 | struct page *page) | |
2089 | { | |
2090 | if (page_mapcount(page) != 1 && | |
2091 | (page_is_file_lru(page) || vma_is_shmem(vma)) && | |
2092 | (vma->vm_flags & VM_EXEC)) | |
2093 | return true; | |
2094 | ||
2095 | return false; | |
2096 | } | |
2097 | ||
b32967ff MG |
2098 | /* |
2099 | * Attempt to migrate a misplaced page to the specified destination | |
2100 | * node. Caller is expected to have an elevated reference count on | |
2101 | * the page that will be dropped by this function before returning. | |
2102 | */ | |
1bc115d8 MG |
2103 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
2104 | int node) | |
b32967ff MG |
2105 | { |
2106 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 2107 | int isolated; |
b32967ff MG |
2108 | int nr_remaining; |
2109 | LIST_HEAD(migratepages); | |
2110 | ||
2111 | /* | |
1bc115d8 MG |
2112 | * Don't migrate file pages that are mapped in multiple processes |
2113 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 2114 | */ |
c77c5cba | 2115 | if (is_shared_exec_page(vma, page)) |
b32967ff | 2116 | goto out; |
b32967ff | 2117 | |
09a913a7 MG |
2118 | /* |
2119 | * Also do not migrate dirty pages as not all filesystems can move | |
2120 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2121 | */ | |
9de4f22a | 2122 | if (page_is_file_lru(page) && PageDirty(page)) |
09a913a7 MG |
2123 | goto out; |
2124 | ||
b32967ff MG |
2125 | isolated = numamigrate_isolate_page(pgdat, page); |
2126 | if (!isolated) | |
2127 | goto out; | |
2128 | ||
2129 | list_add(&page->lru, &migratepages); | |
9c620e2b | 2130 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
68711a74 DR |
2131 | NULL, node, MIGRATE_ASYNC, |
2132 | MR_NUMA_MISPLACED); | |
b32967ff | 2133 | if (nr_remaining) { |
59c82b70 JK |
2134 | if (!list_empty(&migratepages)) { |
2135 | list_del(&page->lru); | |
599d0c95 | 2136 | dec_node_page_state(page, NR_ISOLATED_ANON + |
9de4f22a | 2137 | page_is_file_lru(page)); |
59c82b70 JK |
2138 | putback_lru_page(page); |
2139 | } | |
b32967ff MG |
2140 | isolated = 0; |
2141 | } else | |
2142 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 2143 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 2144 | return isolated; |
340ef390 HD |
2145 | |
2146 | out: | |
2147 | put_page(page); | |
2148 | return 0; | |
7039e1db | 2149 | } |
220018d3 | 2150 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 2151 | |
220018d3 | 2152 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
2153 | /* |
2154 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
2155 | * before returning. | |
2156 | */ | |
b32967ff MG |
2157 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
2158 | struct vm_area_struct *vma, | |
2159 | pmd_t *pmd, pmd_t entry, | |
2160 | unsigned long address, | |
2161 | struct page *page, int node) | |
2162 | { | |
c4088ebd | 2163 | spinlock_t *ptl; |
b32967ff MG |
2164 | pg_data_t *pgdat = NODE_DATA(node); |
2165 | int isolated = 0; | |
2166 | struct page *new_page = NULL; | |
9de4f22a | 2167 | int page_lru = page_is_file_lru(page); |
7066f0f9 | 2168 | unsigned long start = address & HPAGE_PMD_MASK; |
b32967ff | 2169 | |
c77c5cba YS |
2170 | if (is_shared_exec_page(vma, page)) |
2171 | goto out; | |
2172 | ||
b32967ff | 2173 | new_page = alloc_pages_node(node, |
25160354 | 2174 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
e97ca8e5 | 2175 | HPAGE_PMD_ORDER); |
340ef390 HD |
2176 | if (!new_page) |
2177 | goto out_fail; | |
9a982250 | 2178 | prep_transhuge_page(new_page); |
340ef390 | 2179 | |
b32967ff | 2180 | isolated = numamigrate_isolate_page(pgdat, page); |
340ef390 | 2181 | if (!isolated) { |
b32967ff | 2182 | put_page(new_page); |
340ef390 | 2183 | goto out_fail; |
b32967ff | 2184 | } |
b0943d61 | 2185 | |
b32967ff | 2186 | /* Prepare a page as a migration target */ |
48c935ad | 2187 | __SetPageLocked(new_page); |
d44d363f SL |
2188 | if (PageSwapBacked(page)) |
2189 | __SetPageSwapBacked(new_page); | |
b32967ff MG |
2190 | |
2191 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2192 | new_page->mapping = page->mapping; | |
2193 | new_page->index = page->index; | |
7eef5f97 AA |
2194 | /* flush the cache before copying using the kernel virtual address */ |
2195 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); | |
b32967ff MG |
2196 | migrate_page_copy(new_page, page); |
2197 | WARN_ON(PageLRU(new_page)); | |
2198 | ||
2199 | /* Recheck the target PMD */ | |
c4088ebd | 2200 | ptl = pmd_lock(mm, pmd); |
f4e177d1 | 2201 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
c4088ebd | 2202 | spin_unlock(ptl); |
b32967ff MG |
2203 | |
2204 | /* Reverse changes made by migrate_page_copy() */ | |
2205 | if (TestClearPageActive(new_page)) | |
2206 | SetPageActive(page); | |
2207 | if (TestClearPageUnevictable(new_page)) | |
2208 | SetPageUnevictable(page); | |
b32967ff MG |
2209 | |
2210 | unlock_page(new_page); | |
2211 | put_page(new_page); /* Free it */ | |
2212 | ||
a54a407f MG |
2213 | /* Retake the callers reference and putback on LRU */ |
2214 | get_page(page); | |
b32967ff | 2215 | putback_lru_page(page); |
599d0c95 | 2216 | mod_node_page_state(page_pgdat(page), |
a54a407f | 2217 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
eb4489f6 MG |
2218 | |
2219 | goto out_unlock; | |
b32967ff MG |
2220 | } |
2221 | ||
10102459 | 2222 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
f55e1014 | 2223 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
b32967ff | 2224 | |
2b4847e7 | 2225 | /* |
d7c33934 AA |
2226 | * Overwrite the old entry under pagetable lock and establish |
2227 | * the new PTE. Any parallel GUP will either observe the old | |
2228 | * page blocking on the page lock, block on the page table | |
2229 | * lock or observe the new page. The SetPageUptodate on the | |
2230 | * new page and page_add_new_anon_rmap guarantee the copy is | |
2231 | * visible before the pagetable update. | |
2b4847e7 | 2232 | */ |
7066f0f9 | 2233 | page_add_anon_rmap(new_page, vma, start, true); |
d7c33934 AA |
2234 | /* |
2235 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB | |
2236 | * has already been flushed globally. So no TLB can be currently | |
2237 | * caching this non present pmd mapping. There's no need to clear the | |
2238 | * pmd before doing set_pmd_at(), nor to flush the TLB after | |
2239 | * set_pmd_at(). Clearing the pmd here would introduce a race | |
2240 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the | |
c1e8d7c6 | 2241 | * mmap_lock for reading. If the pmd is set to NULL at any given time, |
d7c33934 AA |
2242 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this |
2243 | * pmd. | |
2244 | */ | |
7066f0f9 | 2245 | set_pmd_at(mm, start, pmd, entry); |
ce4a9cc5 | 2246 | update_mmu_cache_pmd(vma, address, &entry); |
2b4847e7 | 2247 | |
f4e177d1 | 2248 | page_ref_unfreeze(page, 2); |
51afb12b | 2249 | mlock_migrate_page(new_page, page); |
d281ee61 | 2250 | page_remove_rmap(page, true); |
7cd12b4a | 2251 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2b4847e7 | 2252 | |
c4088ebd | 2253 | spin_unlock(ptl); |
b32967ff | 2254 | |
11de9927 MG |
2255 | /* Take an "isolate" reference and put new page on the LRU. */ |
2256 | get_page(new_page); | |
2257 | putback_lru_page(new_page); | |
2258 | ||
b32967ff MG |
2259 | unlock_page(new_page); |
2260 | unlock_page(page); | |
2261 | put_page(page); /* Drop the rmap reference */ | |
2262 | put_page(page); /* Drop the LRU isolation reference */ | |
2263 | ||
2264 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2265 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2266 | ||
599d0c95 | 2267 | mod_node_page_state(page_pgdat(page), |
b32967ff MG |
2268 | NR_ISOLATED_ANON + page_lru, |
2269 | -HPAGE_PMD_NR); | |
2270 | return isolated; | |
2271 | ||
340ef390 HD |
2272 | out_fail: |
2273 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2b4847e7 MG |
2274 | ptl = pmd_lock(mm, pmd); |
2275 | if (pmd_same(*pmd, entry)) { | |
4d942466 | 2276 | entry = pmd_modify(entry, vma->vm_page_prot); |
7066f0f9 | 2277 | set_pmd_at(mm, start, pmd, entry); |
2b4847e7 MG |
2278 | update_mmu_cache_pmd(vma, address, &entry); |
2279 | } | |
2280 | spin_unlock(ptl); | |
a54a407f | 2281 | |
eb4489f6 | 2282 | out_unlock: |
340ef390 | 2283 | unlock_page(page); |
c77c5cba | 2284 | out: |
b32967ff | 2285 | put_page(page); |
b32967ff MG |
2286 | return 0; |
2287 | } | |
7039e1db PZ |
2288 | #endif /* CONFIG_NUMA_BALANCING */ |
2289 | ||
2290 | #endif /* CONFIG_NUMA */ | |
8763cb45 | 2291 | |
9b2ed9cb | 2292 | #ifdef CONFIG_DEVICE_PRIVATE |
8763cb45 JG |
2293 | static int migrate_vma_collect_hole(unsigned long start, |
2294 | unsigned long end, | |
b7a16c7a | 2295 | __always_unused int depth, |
8763cb45 JG |
2296 | struct mm_walk *walk) |
2297 | { | |
2298 | struct migrate_vma *migrate = walk->private; | |
2299 | unsigned long addr; | |
2300 | ||
0744f280 RC |
2301 | /* Only allow populating anonymous memory. */ |
2302 | if (!vma_is_anonymous(walk->vma)) { | |
2303 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
2304 | migrate->src[migrate->npages] = 0; | |
2305 | migrate->dst[migrate->npages] = 0; | |
2306 | migrate->npages++; | |
2307 | } | |
2308 | return 0; | |
2309 | } | |
2310 | ||
872ea707 | 2311 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
e20d103b | 2312 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
8315ada7 | 2313 | migrate->dst[migrate->npages] = 0; |
e20d103b | 2314 | migrate->npages++; |
8315ada7 JG |
2315 | migrate->cpages++; |
2316 | } | |
2317 | ||
2318 | return 0; | |
2319 | } | |
2320 | ||
2321 | static int migrate_vma_collect_skip(unsigned long start, | |
2322 | unsigned long end, | |
2323 | struct mm_walk *walk) | |
2324 | { | |
2325 | struct migrate_vma *migrate = walk->private; | |
2326 | unsigned long addr; | |
2327 | ||
872ea707 | 2328 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
8763cb45 JG |
2329 | migrate->dst[migrate->npages] = 0; |
2330 | migrate->src[migrate->npages++] = 0; | |
2331 | } | |
2332 | ||
2333 | return 0; | |
2334 | } | |
2335 | ||
2336 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2337 | unsigned long start, | |
2338 | unsigned long end, | |
2339 | struct mm_walk *walk) | |
2340 | { | |
2341 | struct migrate_vma *migrate = walk->private; | |
2342 | struct vm_area_struct *vma = walk->vma; | |
2343 | struct mm_struct *mm = vma->vm_mm; | |
8c3328f1 | 2344 | unsigned long addr = start, unmapped = 0; |
8763cb45 JG |
2345 | spinlock_t *ptl; |
2346 | pte_t *ptep; | |
2347 | ||
2348 | again: | |
2349 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2350 | return migrate_vma_collect_hole(start, end, -1, walk); |
8763cb45 JG |
2351 | |
2352 | if (pmd_trans_huge(*pmdp)) { | |
2353 | struct page *page; | |
2354 | ||
2355 | ptl = pmd_lock(mm, pmdp); | |
2356 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2357 | spin_unlock(ptl); | |
2358 | goto again; | |
2359 | } | |
2360 | ||
2361 | page = pmd_page(*pmdp); | |
2362 | if (is_huge_zero_page(page)) { | |
2363 | spin_unlock(ptl); | |
2364 | split_huge_pmd(vma, pmdp, addr); | |
2365 | if (pmd_trans_unstable(pmdp)) | |
8315ada7 | 2366 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2367 | walk); |
2368 | } else { | |
2369 | int ret; | |
2370 | ||
2371 | get_page(page); | |
2372 | spin_unlock(ptl); | |
2373 | if (unlikely(!trylock_page(page))) | |
8315ada7 | 2374 | return migrate_vma_collect_skip(start, end, |
8763cb45 JG |
2375 | walk); |
2376 | ret = split_huge_page(page); | |
2377 | unlock_page(page); | |
2378 | put_page(page); | |
8315ada7 JG |
2379 | if (ret) |
2380 | return migrate_vma_collect_skip(start, end, | |
2381 | walk); | |
2382 | if (pmd_none(*pmdp)) | |
b7a16c7a | 2383 | return migrate_vma_collect_hole(start, end, -1, |
8763cb45 JG |
2384 | walk); |
2385 | } | |
2386 | } | |
2387 | ||
2388 | if (unlikely(pmd_bad(*pmdp))) | |
8315ada7 | 2389 | return migrate_vma_collect_skip(start, end, walk); |
8763cb45 JG |
2390 | |
2391 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
8c3328f1 JG |
2392 | arch_enter_lazy_mmu_mode(); |
2393 | ||
8763cb45 | 2394 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
800bb1c8 | 2395 | unsigned long mpfn = 0, pfn; |
8763cb45 | 2396 | struct page *page; |
8c3328f1 | 2397 | swp_entry_t entry; |
8763cb45 JG |
2398 | pte_t pte; |
2399 | ||
2400 | pte = *ptep; | |
8763cb45 | 2401 | |
a5430dda | 2402 | if (pte_none(pte)) { |
0744f280 RC |
2403 | if (vma_is_anonymous(vma)) { |
2404 | mpfn = MIGRATE_PFN_MIGRATE; | |
2405 | migrate->cpages++; | |
2406 | } | |
8763cb45 JG |
2407 | goto next; |
2408 | } | |
2409 | ||
a5430dda | 2410 | if (!pte_present(pte)) { |
a5430dda JG |
2411 | /* |
2412 | * Only care about unaddressable device page special | |
2413 | * page table entry. Other special swap entries are not | |
2414 | * migratable, and we ignore regular swapped page. | |
2415 | */ | |
2416 | entry = pte_to_swp_entry(pte); | |
2417 | if (!is_device_private_entry(entry)) | |
2418 | goto next; | |
2419 | ||
2420 | page = device_private_entry_to_page(entry); | |
5143192c RC |
2421 | if (!(migrate->flags & |
2422 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
2423 | page->pgmap->owner != migrate->pgmap_owner) | |
800bb1c8 CH |
2424 | goto next; |
2425 | ||
06d462be CH |
2426 | mpfn = migrate_pfn(page_to_pfn(page)) | |
2427 | MIGRATE_PFN_MIGRATE; | |
a5430dda JG |
2428 | if (is_write_device_private_entry(entry)) |
2429 | mpfn |= MIGRATE_PFN_WRITE; | |
2430 | } else { | |
5143192c | 2431 | if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) |
800bb1c8 | 2432 | goto next; |
276f756d | 2433 | pfn = pte_pfn(pte); |
8315ada7 JG |
2434 | if (is_zero_pfn(pfn)) { |
2435 | mpfn = MIGRATE_PFN_MIGRATE; | |
2436 | migrate->cpages++; | |
8315ada7 JG |
2437 | goto next; |
2438 | } | |
25b2995a | 2439 | page = vm_normal_page(migrate->vma, addr, pte); |
a5430dda JG |
2440 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2441 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2442 | } | |
2443 | ||
8763cb45 | 2444 | /* FIXME support THP */ |
8763cb45 | 2445 | if (!page || !page->mapping || PageTransCompound(page)) { |
276f756d | 2446 | mpfn = 0; |
8763cb45 JG |
2447 | goto next; |
2448 | } | |
2449 | ||
2450 | /* | |
2451 | * By getting a reference on the page we pin it and that blocks | |
2452 | * any kind of migration. Side effect is that it "freezes" the | |
2453 | * pte. | |
2454 | * | |
2455 | * We drop this reference after isolating the page from the lru | |
2456 | * for non device page (device page are not on the lru and thus | |
2457 | * can't be dropped from it). | |
2458 | */ | |
2459 | get_page(page); | |
2460 | migrate->cpages++; | |
8763cb45 | 2461 | |
8c3328f1 JG |
2462 | /* |
2463 | * Optimize for the common case where page is only mapped once | |
2464 | * in one process. If we can lock the page, then we can safely | |
2465 | * set up a special migration page table entry now. | |
2466 | */ | |
2467 | if (trylock_page(page)) { | |
2468 | pte_t swp_pte; | |
2469 | ||
2470 | mpfn |= MIGRATE_PFN_LOCKED; | |
2471 | ptep_get_and_clear(mm, addr, ptep); | |
2472 | ||
2473 | /* Setup special migration page table entry */ | |
07707125 RC |
2474 | entry = make_migration_entry(page, mpfn & |
2475 | MIGRATE_PFN_WRITE); | |
8c3328f1 | 2476 | swp_pte = swp_entry_to_pte(entry); |
ad7df764 AP |
2477 | if (pte_present(pte)) { |
2478 | if (pte_soft_dirty(pte)) | |
2479 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2480 | if (pte_uffd_wp(pte)) | |
2481 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2482 | } else { | |
2483 | if (pte_swp_soft_dirty(pte)) | |
2484 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2485 | if (pte_swp_uffd_wp(pte)) | |
2486 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2487 | } | |
8c3328f1 JG |
2488 | set_pte_at(mm, addr, ptep, swp_pte); |
2489 | ||
2490 | /* | |
2491 | * This is like regular unmap: we remove the rmap and | |
2492 | * drop page refcount. Page won't be freed, as we took | |
2493 | * a reference just above. | |
2494 | */ | |
2495 | page_remove_rmap(page, false); | |
2496 | put_page(page); | |
a5430dda JG |
2497 | |
2498 | if (pte_present(pte)) | |
2499 | unmapped++; | |
8c3328f1 JG |
2500 | } |
2501 | ||
8763cb45 | 2502 | next: |
a5430dda | 2503 | migrate->dst[migrate->npages] = 0; |
8763cb45 JG |
2504 | migrate->src[migrate->npages++] = mpfn; |
2505 | } | |
8c3328f1 | 2506 | arch_leave_lazy_mmu_mode(); |
8763cb45 JG |
2507 | pte_unmap_unlock(ptep - 1, ptl); |
2508 | ||
8c3328f1 JG |
2509 | /* Only flush the TLB if we actually modified any entries */ |
2510 | if (unmapped) | |
2511 | flush_tlb_range(walk->vma, start, end); | |
2512 | ||
8763cb45 JG |
2513 | return 0; |
2514 | } | |
2515 | ||
7b86ac33 CH |
2516 | static const struct mm_walk_ops migrate_vma_walk_ops = { |
2517 | .pmd_entry = migrate_vma_collect_pmd, | |
2518 | .pte_hole = migrate_vma_collect_hole, | |
2519 | }; | |
2520 | ||
8763cb45 JG |
2521 | /* |
2522 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2523 | * @migrate: migrate struct containing all migration information | |
2524 | * | |
2525 | * This will walk the CPU page table. For each virtual address backed by a | |
2526 | * valid page, it updates the src array and takes a reference on the page, in | |
2527 | * order to pin the page until we lock it and unmap it. | |
2528 | */ | |
2529 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2530 | { | |
ac46d4f3 | 2531 | struct mmu_notifier_range range; |
8763cb45 | 2532 | |
998427b3 RC |
2533 | /* |
2534 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
2535 | * that the registered device driver can skip invalidating device | |
2536 | * private page mappings that won't be migrated. | |
2537 | */ | |
c1a06df6 RC |
2538 | mmu_notifier_range_init_migrate(&range, 0, migrate->vma, |
2539 | migrate->vma->vm_mm, migrate->start, migrate->end, | |
2540 | migrate->pgmap_owner); | |
ac46d4f3 | 2541 | mmu_notifier_invalidate_range_start(&range); |
8763cb45 | 2542 | |
7b86ac33 CH |
2543 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, |
2544 | &migrate_vma_walk_ops, migrate); | |
2545 | ||
2546 | mmu_notifier_invalidate_range_end(&range); | |
8763cb45 JG |
2547 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2548 | } | |
2549 | ||
2550 | /* | |
2551 | * migrate_vma_check_page() - check if page is pinned or not | |
2552 | * @page: struct page to check | |
2553 | * | |
2554 | * Pinned pages cannot be migrated. This is the same test as in | |
2555 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2556 | * ZONE_DEVICE page. | |
2557 | */ | |
2558 | static bool migrate_vma_check_page(struct page *page) | |
2559 | { | |
2560 | /* | |
2561 | * One extra ref because caller holds an extra reference, either from | |
2562 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2563 | * a device page. | |
2564 | */ | |
2565 | int extra = 1; | |
2566 | ||
2567 | /* | |
2568 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2569 | * check them than regular pages, because they can be mapped with a pmd | |
2570 | * or with a pte (split pte mapping). | |
2571 | */ | |
2572 | if (PageCompound(page)) | |
2573 | return false; | |
2574 | ||
a5430dda JG |
2575 | /* Page from ZONE_DEVICE have one extra reference */ |
2576 | if (is_zone_device_page(page)) { | |
2577 | /* | |
2578 | * Private page can never be pin as they have no valid pte and | |
2579 | * GUP will fail for those. Yet if there is a pending migration | |
2580 | * a thread might try to wait on the pte migration entry and | |
2581 | * will bump the page reference count. Sadly there is no way to | |
2582 | * differentiate a regular pin from migration wait. Hence to | |
2583 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
8958b249 | 2584 | * infinite loop (one stopping migration because the other is |
a5430dda JG |
2585 | * waiting on pte migration entry). We always return true here. |
2586 | * | |
2587 | * FIXME proper solution is to rework migration_entry_wait() so | |
2588 | * it does not need to take a reference on page. | |
2589 | */ | |
25b2995a | 2590 | return is_device_private_page(page); |
a5430dda JG |
2591 | } |
2592 | ||
df6ad698 JG |
2593 | /* For file back page */ |
2594 | if (page_mapping(page)) | |
2595 | extra += 1 + page_has_private(page); | |
2596 | ||
8763cb45 JG |
2597 | if ((page_count(page) - extra) > page_mapcount(page)) |
2598 | return false; | |
2599 | ||
2600 | return true; | |
2601 | } | |
2602 | ||
2603 | /* | |
2604 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2605 | * @migrate: migrate struct containing all migration information | |
2606 | * | |
2607 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2608 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2609 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2610 | * migrated by concurrent kernel threads. | |
2611 | */ | |
2612 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2613 | { | |
2614 | const unsigned long npages = migrate->npages; | |
8c3328f1 JG |
2615 | const unsigned long start = migrate->start; |
2616 | unsigned long addr, i, restore = 0; | |
8763cb45 | 2617 | bool allow_drain = true; |
8763cb45 JG |
2618 | |
2619 | lru_add_drain(); | |
2620 | ||
2621 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2622 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
8c3328f1 | 2623 | bool remap = true; |
8763cb45 JG |
2624 | |
2625 | if (!page) | |
2626 | continue; | |
2627 | ||
8c3328f1 JG |
2628 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2629 | /* | |
2630 | * Because we are migrating several pages there can be | |
2631 | * a deadlock between 2 concurrent migration where each | |
2632 | * are waiting on each other page lock. | |
2633 | * | |
2634 | * Make migrate_vma() a best effort thing and backoff | |
2635 | * for any page we can not lock right away. | |
2636 | */ | |
2637 | if (!trylock_page(page)) { | |
2638 | migrate->src[i] = 0; | |
2639 | migrate->cpages--; | |
2640 | put_page(page); | |
2641 | continue; | |
2642 | } | |
2643 | remap = false; | |
2644 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
8763cb45 | 2645 | } |
8763cb45 | 2646 | |
a5430dda JG |
2647 | /* ZONE_DEVICE pages are not on LRU */ |
2648 | if (!is_zone_device_page(page)) { | |
2649 | if (!PageLRU(page) && allow_drain) { | |
2650 | /* Drain CPU's pagevec */ | |
2651 | lru_add_drain_all(); | |
2652 | allow_drain = false; | |
2653 | } | |
8763cb45 | 2654 | |
a5430dda JG |
2655 | if (isolate_lru_page(page)) { |
2656 | if (remap) { | |
2657 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2658 | migrate->cpages--; | |
2659 | restore++; | |
2660 | } else { | |
2661 | migrate->src[i] = 0; | |
2662 | unlock_page(page); | |
2663 | migrate->cpages--; | |
2664 | put_page(page); | |
2665 | } | |
2666 | continue; | |
8c3328f1 | 2667 | } |
a5430dda JG |
2668 | |
2669 | /* Drop the reference we took in collect */ | |
2670 | put_page(page); | |
8763cb45 JG |
2671 | } |
2672 | ||
2673 | if (!migrate_vma_check_page(page)) { | |
8c3328f1 JG |
2674 | if (remap) { |
2675 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2676 | migrate->cpages--; | |
2677 | restore++; | |
8763cb45 | 2678 | |
a5430dda JG |
2679 | if (!is_zone_device_page(page)) { |
2680 | get_page(page); | |
2681 | putback_lru_page(page); | |
2682 | } | |
8c3328f1 JG |
2683 | } else { |
2684 | migrate->src[i] = 0; | |
2685 | unlock_page(page); | |
2686 | migrate->cpages--; | |
2687 | ||
a5430dda JG |
2688 | if (!is_zone_device_page(page)) |
2689 | putback_lru_page(page); | |
2690 | else | |
2691 | put_page(page); | |
8c3328f1 | 2692 | } |
8763cb45 JG |
2693 | } |
2694 | } | |
8c3328f1 JG |
2695 | |
2696 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2697 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2698 | ||
2699 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2700 | continue; | |
2701 | ||
2702 | remove_migration_pte(page, migrate->vma, addr, page); | |
2703 | ||
2704 | migrate->src[i] = 0; | |
2705 | unlock_page(page); | |
2706 | put_page(page); | |
2707 | restore--; | |
2708 | } | |
8763cb45 JG |
2709 | } |
2710 | ||
2711 | /* | |
2712 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2713 | * @migrate: migrate struct containing all migration information | |
2714 | * | |
2715 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2716 | * and check again if it has been pinned. Pinned pages are restored because we | |
2717 | * cannot migrate them. | |
2718 | * | |
2719 | * This is the last step before we call the device driver callback to allocate | |
2720 | * destination memory and copy contents of original page over to new page. | |
2721 | */ | |
2722 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2723 | { | |
013339df | 2724 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK; |
8763cb45 JG |
2725 | const unsigned long npages = migrate->npages; |
2726 | const unsigned long start = migrate->start; | |
2727 | unsigned long addr, i, restore = 0; | |
2728 | ||
2729 | for (i = 0; i < npages; i++) { | |
2730 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2731 | ||
2732 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2733 | continue; | |
2734 | ||
8c3328f1 JG |
2735 | if (page_mapped(page)) { |
2736 | try_to_unmap(page, flags); | |
2737 | if (page_mapped(page)) | |
2738 | goto restore; | |
8763cb45 | 2739 | } |
8c3328f1 JG |
2740 | |
2741 | if (migrate_vma_check_page(page)) | |
2742 | continue; | |
2743 | ||
2744 | restore: | |
2745 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2746 | migrate->cpages--; | |
2747 | restore++; | |
8763cb45 JG |
2748 | } |
2749 | ||
2750 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2751 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2752 | ||
2753 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2754 | continue; | |
2755 | ||
2756 | remove_migration_ptes(page, page, false); | |
2757 | ||
2758 | migrate->src[i] = 0; | |
2759 | unlock_page(page); | |
2760 | restore--; | |
2761 | ||
a5430dda JG |
2762 | if (is_zone_device_page(page)) |
2763 | put_page(page); | |
2764 | else | |
2765 | putback_lru_page(page); | |
8763cb45 JG |
2766 | } |
2767 | } | |
2768 | ||
a7d1f22b CH |
2769 | /** |
2770 | * migrate_vma_setup() - prepare to migrate a range of memory | |
eaf444de | 2771 | * @args: contains the vma, start, and pfns arrays for the migration |
a7d1f22b CH |
2772 | * |
2773 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2774 | * without an error. | |
2775 | * | |
2776 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2777 | * the pages backing each virtual address in the range, saving them inside the | |
2778 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2779 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2780 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2781 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2782 | * remapping and unlocking those pages. | |
2783 | * | |
2784 | * The caller should then allocate destination memory and copy source memory to | |
2785 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2786 | * flag set). Once these are allocated and copied, the caller must update each | |
2787 | * corresponding entry in the dst array with the pfn value of the destination | |
2788 | * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set | |
2789 | * (destination pages must have their struct pages locked, via lock_page()). | |
2790 | * | |
2791 | * Note that the caller does not have to migrate all the pages that are marked | |
2792 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2793 | * device memory to system memory. If the caller cannot migrate a device page | |
2794 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2795 | * consequences for the userspace process, so it must be avoided if at all | |
2796 | * possible. | |
2797 | * | |
2798 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2799 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
2800 | * allowing the caller to allocate device memory for those unback virtual | |
2801 | * address. For this the caller simply has to allocate device memory and | |
2802 | * properly set the destination entry like for regular migration. Note that | |
2803 | * this can still fails and thus inside the device driver must check if the | |
2804 | * migration was successful for those entries after calling migrate_vma_pages() | |
2805 | * just like for regular migration. | |
2806 | * | |
2807 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2808 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2809 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2810 | * then migrate_vma_pages() to migrate struct page information from the source | |
2811 | * struct page to the destination struct page. If it fails to migrate the | |
2812 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2813 | * src array. | |
2814 | * | |
2815 | * At this point all successfully migrated pages have an entry in the src | |
2816 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2817 | * array entry with MIGRATE_PFN_VALID flag set. | |
2818 | * | |
2819 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2820 | * successfully migrated, and which were not. Successfully migrated pages will | |
2821 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2822 | * | |
2823 | * It is safe to update device page table after migrate_vma_pages() because | |
c1e8d7c6 | 2824 | * both destination and source page are still locked, and the mmap_lock is held |
a7d1f22b CH |
2825 | * in read mode (hence no one can unmap the range being migrated). |
2826 | * | |
2827 | * Once the caller is done cleaning up things and updating its page table (if it | |
2828 | * chose to do so, this is not an obligation) it finally calls | |
2829 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2830 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2831 | * point to the original source pages. | |
2832 | */ | |
2833 | int migrate_vma_setup(struct migrate_vma *args) | |
2834 | { | |
2835 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2836 | ||
2837 | args->start &= PAGE_MASK; | |
2838 | args->end &= PAGE_MASK; | |
2839 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2840 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2841 | return -EINVAL; | |
2842 | if (nr_pages <= 0) | |
2843 | return -EINVAL; | |
2844 | if (args->start < args->vma->vm_start || | |
2845 | args->start >= args->vma->vm_end) | |
2846 | return -EINVAL; | |
2847 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2848 | return -EINVAL; | |
2849 | if (!args->src || !args->dst) | |
2850 | return -EINVAL; | |
2851 | ||
2852 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2853 | args->cpages = 0; | |
2854 | args->npages = 0; | |
2855 | ||
2856 | migrate_vma_collect(args); | |
2857 | ||
2858 | if (args->cpages) | |
2859 | migrate_vma_prepare(args); | |
2860 | if (args->cpages) | |
2861 | migrate_vma_unmap(args); | |
2862 | ||
2863 | /* | |
2864 | * At this point pages are locked and unmapped, and thus they have | |
2865 | * stable content and can safely be copied to destination memory that | |
2866 | * is allocated by the drivers. | |
2867 | */ | |
2868 | return 0; | |
2869 | ||
2870 | } | |
2871 | EXPORT_SYMBOL(migrate_vma_setup); | |
2872 | ||
34290e2c RC |
2873 | /* |
2874 | * This code closely matches the code in: | |
2875 | * __handle_mm_fault() | |
2876 | * handle_pte_fault() | |
2877 | * do_anonymous_page() | |
2878 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
2879 | * private page. | |
2880 | */ | |
8315ada7 JG |
2881 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2882 | unsigned long addr, | |
2883 | struct page *page, | |
d85c6db4 | 2884 | unsigned long *src) |
8315ada7 JG |
2885 | { |
2886 | struct vm_area_struct *vma = migrate->vma; | |
2887 | struct mm_struct *mm = vma->vm_mm; | |
8315ada7 JG |
2888 | bool flush = false; |
2889 | spinlock_t *ptl; | |
2890 | pte_t entry; | |
2891 | pgd_t *pgdp; | |
2892 | p4d_t *p4dp; | |
2893 | pud_t *pudp; | |
2894 | pmd_t *pmdp; | |
2895 | pte_t *ptep; | |
2896 | ||
2897 | /* Only allow populating anonymous memory */ | |
2898 | if (!vma_is_anonymous(vma)) | |
2899 | goto abort; | |
2900 | ||
2901 | pgdp = pgd_offset(mm, addr); | |
2902 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2903 | if (!p4dp) | |
2904 | goto abort; | |
2905 | pudp = pud_alloc(mm, p4dp, addr); | |
2906 | if (!pudp) | |
2907 | goto abort; | |
2908 | pmdp = pmd_alloc(mm, pudp, addr); | |
2909 | if (!pmdp) | |
2910 | goto abort; | |
2911 | ||
2912 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2913 | goto abort; | |
2914 | ||
2915 | /* | |
2916 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2917 | * pte_offset_map() on pmds where a huge pmd might be created | |
2918 | * from a different thread. | |
2919 | * | |
3e4e28c5 | 2920 | * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when |
8315ada7 JG |
2921 | * parallel threads are excluded by other means. |
2922 | * | |
3e4e28c5 | 2923 | * Here we only have mmap_read_lock(mm). |
8315ada7 | 2924 | */ |
4cf58924 | 2925 | if (pte_alloc(mm, pmdp)) |
8315ada7 JG |
2926 | goto abort; |
2927 | ||
2928 | /* See the comment in pte_alloc_one_map() */ | |
2929 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2930 | goto abort; | |
2931 | ||
2932 | if (unlikely(anon_vma_prepare(vma))) | |
2933 | goto abort; | |
d9eb1ea2 | 2934 | if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) |
8315ada7 JG |
2935 | goto abort; |
2936 | ||
2937 | /* | |
2938 | * The memory barrier inside __SetPageUptodate makes sure that | |
2939 | * preceding stores to the page contents become visible before | |
2940 | * the set_pte_at() write. | |
2941 | */ | |
2942 | __SetPageUptodate(page); | |
2943 | ||
df6ad698 JG |
2944 | if (is_zone_device_page(page)) { |
2945 | if (is_device_private_page(page)) { | |
2946 | swp_entry_t swp_entry; | |
2947 | ||
2948 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2949 | entry = swp_entry_to_pte(swp_entry); | |
df6ad698 | 2950 | } |
8315ada7 JG |
2951 | } else { |
2952 | entry = mk_pte(page, vma->vm_page_prot); | |
2953 | if (vma->vm_flags & VM_WRITE) | |
2954 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2955 | } | |
2956 | ||
2957 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2958 | ||
34290e2c RC |
2959 | if (check_stable_address_space(mm)) |
2960 | goto unlock_abort; | |
2961 | ||
8315ada7 JG |
2962 | if (pte_present(*ptep)) { |
2963 | unsigned long pfn = pte_pfn(*ptep); | |
2964 | ||
c23a0c99 RC |
2965 | if (!is_zero_pfn(pfn)) |
2966 | goto unlock_abort; | |
8315ada7 | 2967 | flush = true; |
c23a0c99 RC |
2968 | } else if (!pte_none(*ptep)) |
2969 | goto unlock_abort; | |
8315ada7 JG |
2970 | |
2971 | /* | |
c23a0c99 | 2972 | * Check for userfaultfd but do not deliver the fault. Instead, |
8315ada7 JG |
2973 | * just back off. |
2974 | */ | |
c23a0c99 RC |
2975 | if (userfaultfd_missing(vma)) |
2976 | goto unlock_abort; | |
8315ada7 JG |
2977 | |
2978 | inc_mm_counter(mm, MM_ANONPAGES); | |
be5d0a74 | 2979 | page_add_new_anon_rmap(page, vma, addr, false); |
8315ada7 | 2980 | if (!is_zone_device_page(page)) |
b518154e | 2981 | lru_cache_add_inactive_or_unevictable(page, vma); |
8315ada7 JG |
2982 | get_page(page); |
2983 | ||
2984 | if (flush) { | |
2985 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2986 | ptep_clear_flush_notify(vma, addr, ptep); | |
2987 | set_pte_at_notify(mm, addr, ptep, entry); | |
2988 | update_mmu_cache(vma, addr, ptep); | |
2989 | } else { | |
2990 | /* No need to invalidate - it was non-present before */ | |
2991 | set_pte_at(mm, addr, ptep, entry); | |
2992 | update_mmu_cache(vma, addr, ptep); | |
2993 | } | |
2994 | ||
2995 | pte_unmap_unlock(ptep, ptl); | |
2996 | *src = MIGRATE_PFN_MIGRATE; | |
2997 | return; | |
2998 | ||
c23a0c99 RC |
2999 | unlock_abort: |
3000 | pte_unmap_unlock(ptep, ptl); | |
8315ada7 JG |
3001 | abort: |
3002 | *src &= ~MIGRATE_PFN_MIGRATE; | |
3003 | } | |
3004 | ||
a7d1f22b | 3005 | /** |
8763cb45 JG |
3006 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
3007 | * @migrate: migrate struct containing all migration information | |
3008 | * | |
3009 | * This migrates struct page meta-data from source struct page to destination | |
3010 | * struct page. This effectively finishes the migration from source page to the | |
3011 | * destination page. | |
3012 | */ | |
a7d1f22b | 3013 | void migrate_vma_pages(struct migrate_vma *migrate) |
8763cb45 JG |
3014 | { |
3015 | const unsigned long npages = migrate->npages; | |
3016 | const unsigned long start = migrate->start; | |
ac46d4f3 JG |
3017 | struct mmu_notifier_range range; |
3018 | unsigned long addr, i; | |
8315ada7 | 3019 | bool notified = false; |
8763cb45 JG |
3020 | |
3021 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
3022 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
3023 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
3024 | struct address_space *mapping; | |
3025 | int r; | |
3026 | ||
8315ada7 JG |
3027 | if (!newpage) { |
3028 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
8763cb45 | 3029 | continue; |
8315ada7 JG |
3030 | } |
3031 | ||
3032 | if (!page) { | |
c23a0c99 | 3033 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
8315ada7 | 3034 | continue; |
8315ada7 | 3035 | if (!notified) { |
8315ada7 | 3036 | notified = true; |
ac46d4f3 | 3037 | |
5e5dda81 RC |
3038 | mmu_notifier_range_init_migrate(&range, 0, |
3039 | migrate->vma, migrate->vma->vm_mm, | |
3040 | addr, migrate->end, | |
3041 | migrate->pgmap_owner); | |
ac46d4f3 | 3042 | mmu_notifier_invalidate_range_start(&range); |
8315ada7 JG |
3043 | } |
3044 | migrate_vma_insert_page(migrate, addr, newpage, | |
d85c6db4 | 3045 | &migrate->src[i]); |
8763cb45 | 3046 | continue; |
8315ada7 | 3047 | } |
8763cb45 JG |
3048 | |
3049 | mapping = page_mapping(page); | |
3050 | ||
a5430dda JG |
3051 | if (is_zone_device_page(newpage)) { |
3052 | if (is_device_private_page(newpage)) { | |
3053 | /* | |
3054 | * For now only support private anonymous when | |
3055 | * migrating to un-addressable device memory. | |
3056 | */ | |
3057 | if (mapping) { | |
3058 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3059 | continue; | |
3060 | } | |
25b2995a | 3061 | } else { |
a5430dda JG |
3062 | /* |
3063 | * Other types of ZONE_DEVICE page are not | |
3064 | * supported. | |
3065 | */ | |
3066 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3067 | continue; | |
3068 | } | |
3069 | } | |
3070 | ||
8763cb45 JG |
3071 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
3072 | if (r != MIGRATEPAGE_SUCCESS) | |
3073 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3074 | } | |
8315ada7 | 3075 | |
4645b9fe JG |
3076 | /* |
3077 | * No need to double call mmu_notifier->invalidate_range() callback as | |
3078 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
3079 | * did already call it. | |
3080 | */ | |
8315ada7 | 3081 | if (notified) |
ac46d4f3 | 3082 | mmu_notifier_invalidate_range_only_end(&range); |
8763cb45 | 3083 | } |
a7d1f22b | 3084 | EXPORT_SYMBOL(migrate_vma_pages); |
8763cb45 | 3085 | |
a7d1f22b | 3086 | /** |
8763cb45 JG |
3087 | * migrate_vma_finalize() - restore CPU page table entry |
3088 | * @migrate: migrate struct containing all migration information | |
3089 | * | |
3090 | * This replaces the special migration pte entry with either a mapping to the | |
3091 | * new page if migration was successful for that page, or to the original page | |
3092 | * otherwise. | |
3093 | * | |
3094 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
3095 | * refcount, for device pages. | |
3096 | */ | |
a7d1f22b | 3097 | void migrate_vma_finalize(struct migrate_vma *migrate) |
8763cb45 JG |
3098 | { |
3099 | const unsigned long npages = migrate->npages; | |
3100 | unsigned long i; | |
3101 | ||
3102 | for (i = 0; i < npages; i++) { | |
3103 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
3104 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
3105 | ||
8315ada7 JG |
3106 | if (!page) { |
3107 | if (newpage) { | |
3108 | unlock_page(newpage); | |
3109 | put_page(newpage); | |
3110 | } | |
8763cb45 | 3111 | continue; |
8315ada7 JG |
3112 | } |
3113 | ||
8763cb45 JG |
3114 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
3115 | if (newpage) { | |
3116 | unlock_page(newpage); | |
3117 | put_page(newpage); | |
3118 | } | |
3119 | newpage = page; | |
3120 | } | |
3121 | ||
3122 | remove_migration_ptes(page, newpage, false); | |
3123 | unlock_page(page); | |
8763cb45 | 3124 | |
a5430dda JG |
3125 | if (is_zone_device_page(page)) |
3126 | put_page(page); | |
3127 | else | |
3128 | putback_lru_page(page); | |
8763cb45 JG |
3129 | |
3130 | if (newpage != page) { | |
3131 | unlock_page(newpage); | |
a5430dda JG |
3132 | if (is_zone_device_page(newpage)) |
3133 | put_page(newpage); | |
3134 | else | |
3135 | putback_lru_page(newpage); | |
8763cb45 JG |
3136 | } |
3137 | } | |
3138 | } | |
a7d1f22b | 3139 | EXPORT_SYMBOL(migrate_vma_finalize); |
9b2ed9cb | 3140 | #endif /* CONFIG_DEVICE_PRIVATE */ |