Commit | Line | Data |
---|---|---|
76cbbead CH |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Device Memory Migration functionality. | |
4 | * | |
5 | * Originally written by Jérôme Glisse. | |
6 | */ | |
7 | #include <linux/export.h> | |
8 | #include <linux/memremap.h> | |
9 | #include <linux/migrate.h> | |
fd35ca3d | 10 | #include <linux/mm.h> |
76cbbead CH |
11 | #include <linux/mm_inline.h> |
12 | #include <linux/mmu_notifier.h> | |
13 | #include <linux/oom.h> | |
14 | #include <linux/pagewalk.h> | |
15 | #include <linux/rmap.h> | |
16 | #include <linux/swapops.h> | |
17 | #include <asm/tlbflush.h> | |
18 | #include "internal.h" | |
19 | ||
20 | static int migrate_vma_collect_skip(unsigned long start, | |
21 | unsigned long end, | |
22 | struct mm_walk *walk) | |
23 | { | |
24 | struct migrate_vma *migrate = walk->private; | |
25 | unsigned long addr; | |
26 | ||
27 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
28 | migrate->dst[migrate->npages] = 0; | |
29 | migrate->src[migrate->npages++] = 0; | |
30 | } | |
31 | ||
32 | return 0; | |
33 | } | |
34 | ||
35 | static int migrate_vma_collect_hole(unsigned long start, | |
36 | unsigned long end, | |
37 | __always_unused int depth, | |
38 | struct mm_walk *walk) | |
39 | { | |
40 | struct migrate_vma *migrate = walk->private; | |
41 | unsigned long addr; | |
42 | ||
43 | /* Only allow populating anonymous memory. */ | |
44 | if (!vma_is_anonymous(walk->vma)) | |
45 | return migrate_vma_collect_skip(start, end, walk); | |
46 | ||
47 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
48 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; | |
49 | migrate->dst[migrate->npages] = 0; | |
50 | migrate->npages++; | |
51 | migrate->cpages++; | |
52 | } | |
53 | ||
54 | return 0; | |
55 | } | |
56 | ||
57 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
58 | unsigned long start, | |
59 | unsigned long end, | |
60 | struct mm_walk *walk) | |
61 | { | |
62 | struct migrate_vma *migrate = walk->private; | |
63 | struct vm_area_struct *vma = walk->vma; | |
64 | struct mm_struct *mm = vma->vm_mm; | |
65 | unsigned long addr = start, unmapped = 0; | |
66 | spinlock_t *ptl; | |
67 | pte_t *ptep; | |
68 | ||
69 | again: | |
70 | if (pmd_none(*pmdp)) | |
71 | return migrate_vma_collect_hole(start, end, -1, walk); | |
72 | ||
73 | if (pmd_trans_huge(*pmdp)) { | |
74 | struct page *page; | |
75 | ||
76 | ptl = pmd_lock(mm, pmdp); | |
77 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
78 | spin_unlock(ptl); | |
79 | goto again; | |
80 | } | |
81 | ||
82 | page = pmd_page(*pmdp); | |
83 | if (is_huge_zero_page(page)) { | |
84 | spin_unlock(ptl); | |
85 | split_huge_pmd(vma, pmdp, addr); | |
76cbbead CH |
86 | } else { |
87 | int ret; | |
88 | ||
89 | get_page(page); | |
90 | spin_unlock(ptl); | |
91 | if (unlikely(!trylock_page(page))) | |
92 | return migrate_vma_collect_skip(start, end, | |
93 | walk); | |
94 | ret = split_huge_page(page); | |
95 | unlock_page(page); | |
96 | put_page(page); | |
97 | if (ret) | |
98 | return migrate_vma_collect_skip(start, end, | |
99 | walk); | |
76cbbead CH |
100 | } |
101 | } | |
102 | ||
76cbbead | 103 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); |
4b56069c HD |
104 | if (!ptep) |
105 | goto again; | |
76cbbead CH |
106 | arch_enter_lazy_mmu_mode(); |
107 | ||
108 | for (; addr < end; addr += PAGE_SIZE, ptep++) { | |
109 | unsigned long mpfn = 0, pfn; | |
110 | struct page *page; | |
111 | swp_entry_t entry; | |
112 | pte_t pte; | |
113 | ||
c33c7948 | 114 | pte = ptep_get(ptep); |
76cbbead CH |
115 | |
116 | if (pte_none(pte)) { | |
117 | if (vma_is_anonymous(vma)) { | |
118 | mpfn = MIGRATE_PFN_MIGRATE; | |
119 | migrate->cpages++; | |
120 | } | |
121 | goto next; | |
122 | } | |
123 | ||
124 | if (!pte_present(pte)) { | |
125 | /* | |
126 | * Only care about unaddressable device page special | |
127 | * page table entry. Other special swap entries are not | |
128 | * migratable, and we ignore regular swapped page. | |
129 | */ | |
130 | entry = pte_to_swp_entry(pte); | |
131 | if (!is_device_private_entry(entry)) | |
132 | goto next; | |
133 | ||
134 | page = pfn_swap_entry_to_page(entry); | |
135 | if (!(migrate->flags & | |
136 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
137 | page->pgmap->owner != migrate->pgmap_owner) | |
138 | goto next; | |
139 | ||
140 | mpfn = migrate_pfn(page_to_pfn(page)) | | |
141 | MIGRATE_PFN_MIGRATE; | |
142 | if (is_writable_device_private_entry(entry)) | |
143 | mpfn |= MIGRATE_PFN_WRITE; | |
144 | } else { | |
76cbbead | 145 | pfn = pte_pfn(pte); |
dd19e6d8 AS |
146 | if (is_zero_pfn(pfn) && |
147 | (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) { | |
76cbbead CH |
148 | mpfn = MIGRATE_PFN_MIGRATE; |
149 | migrate->cpages++; | |
150 | goto next; | |
151 | } | |
152 | page = vm_normal_page(migrate->vma, addr, pte); | |
dd19e6d8 AS |
153 | if (page && !is_zone_device_page(page) && |
154 | !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) | |
155 | goto next; | |
156 | else if (page && is_device_coherent_page(page) && | |
157 | (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) || | |
158 | page->pgmap->owner != migrate->pgmap_owner)) | |
159 | goto next; | |
76cbbead CH |
160 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
161 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
162 | } | |
163 | ||
164 | /* FIXME support THP */ | |
165 | if (!page || !page->mapping || PageTransCompound(page)) { | |
166 | mpfn = 0; | |
167 | goto next; | |
168 | } | |
169 | ||
170 | /* | |
171 | * By getting a reference on the page we pin it and that blocks | |
172 | * any kind of migration. Side effect is that it "freezes" the | |
173 | * pte. | |
174 | * | |
175 | * We drop this reference after isolating the page from the lru | |
176 | * for non device page (device page are not on the lru and thus | |
177 | * can't be dropped from it). | |
178 | */ | |
179 | get_page(page); | |
180 | ||
181 | /* | |
0742e490 AP |
182 | * We rely on trylock_page() to avoid deadlock between |
183 | * concurrent migrations where each is waiting on the others | |
184 | * page lock. If we can't immediately lock the page we fail this | |
185 | * migration as it is only best effort anyway. | |
186 | * | |
187 | * If we can lock the page it's safe to set up a migration entry | |
188 | * now. In the common case where the page is mapped once in a | |
189 | * single process setting up the migration entry now is an | |
190 | * optimisation to avoid walking the rmap later with | |
191 | * try_to_migrate(). | |
76cbbead CH |
192 | */ |
193 | if (trylock_page(page)) { | |
6c287605 | 194 | bool anon_exclusive; |
76cbbead CH |
195 | pte_t swp_pte; |
196 | ||
c33c7948 | 197 | flush_cache_page(vma, addr, pte_pfn(pte)); |
6c287605 DH |
198 | anon_exclusive = PageAnon(page) && PageAnonExclusive(page); |
199 | if (anon_exclusive) { | |
fd35ca3d | 200 | pte = ptep_clear_flush(vma, addr, ptep); |
6c287605 DH |
201 | |
202 | if (page_try_share_anon_rmap(page)) { | |
203 | set_pte_at(mm, addr, ptep, pte); | |
204 | unlock_page(page); | |
205 | put_page(page); | |
206 | mpfn = 0; | |
207 | goto next; | |
208 | } | |
209 | } else { | |
fd35ca3d | 210 | pte = ptep_get_and_clear(mm, addr, ptep); |
6c287605 DH |
211 | } |
212 | ||
76cbbead | 213 | migrate->cpages++; |
76cbbead | 214 | |
fd35ca3d AP |
215 | /* Set the dirty flag on the folio now the pte is gone. */ |
216 | if (pte_dirty(pte)) | |
217 | folio_mark_dirty(page_folio(page)); | |
218 | ||
76cbbead CH |
219 | /* Setup special migration page table entry */ |
220 | if (mpfn & MIGRATE_PFN_WRITE) | |
221 | entry = make_writable_migration_entry( | |
222 | page_to_pfn(page)); | |
6c287605 DH |
223 | else if (anon_exclusive) |
224 | entry = make_readable_exclusive_migration_entry( | |
225 | page_to_pfn(page)); | |
76cbbead CH |
226 | else |
227 | entry = make_readable_migration_entry( | |
228 | page_to_pfn(page)); | |
2e346877 PX |
229 | if (pte_present(pte)) { |
230 | if (pte_young(pte)) | |
231 | entry = make_migration_entry_young(entry); | |
232 | if (pte_dirty(pte)) | |
233 | entry = make_migration_entry_dirty(entry); | |
234 | } | |
76cbbead CH |
235 | swp_pte = swp_entry_to_pte(entry); |
236 | if (pte_present(pte)) { | |
237 | if (pte_soft_dirty(pte)) | |
238 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
239 | if (pte_uffd_wp(pte)) | |
240 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
241 | } else { | |
242 | if (pte_swp_soft_dirty(pte)) | |
243 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
244 | if (pte_swp_uffd_wp(pte)) | |
245 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
246 | } | |
247 | set_pte_at(mm, addr, ptep, swp_pte); | |
248 | ||
249 | /* | |
250 | * This is like regular unmap: we remove the rmap and | |
251 | * drop page refcount. Page won't be freed, as we took | |
252 | * a reference just above. | |
253 | */ | |
254 | page_remove_rmap(page, vma, false); | |
255 | put_page(page); | |
256 | ||
257 | if (pte_present(pte)) | |
258 | unmapped++; | |
259 | } else { | |
260 | put_page(page); | |
261 | mpfn = 0; | |
262 | } | |
263 | ||
264 | next: | |
265 | migrate->dst[migrate->npages] = 0; | |
266 | migrate->src[migrate->npages++] = mpfn; | |
267 | } | |
76cbbead CH |
268 | |
269 | /* Only flush the TLB if we actually modified any entries */ | |
270 | if (unmapped) | |
271 | flush_tlb_range(walk->vma, start, end); | |
272 | ||
60bae737 AP |
273 | arch_leave_lazy_mmu_mode(); |
274 | pte_unmap_unlock(ptep - 1, ptl); | |
275 | ||
76cbbead CH |
276 | return 0; |
277 | } | |
278 | ||
279 | static const struct mm_walk_ops migrate_vma_walk_ops = { | |
280 | .pmd_entry = migrate_vma_collect_pmd, | |
281 | .pte_hole = migrate_vma_collect_hole, | |
282 | }; | |
283 | ||
284 | /* | |
285 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
286 | * @migrate: migrate struct containing all migration information | |
287 | * | |
288 | * This will walk the CPU page table. For each virtual address backed by a | |
289 | * valid page, it updates the src array and takes a reference on the page, in | |
290 | * order to pin the page until we lock it and unmap it. | |
291 | */ | |
292 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
293 | { | |
294 | struct mmu_notifier_range range; | |
295 | ||
296 | /* | |
297 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
298 | * that the registered device driver can skip invalidating device | |
299 | * private page mappings that won't be migrated. | |
300 | */ | |
301 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0, | |
7d4a8be0 | 302 | migrate->vma->vm_mm, migrate->start, migrate->end, |
76cbbead CH |
303 | migrate->pgmap_owner); |
304 | mmu_notifier_invalidate_range_start(&range); | |
305 | ||
306 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, | |
307 | &migrate_vma_walk_ops, migrate); | |
308 | ||
309 | mmu_notifier_invalidate_range_end(&range); | |
310 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
311 | } | |
312 | ||
313 | /* | |
314 | * migrate_vma_check_page() - check if page is pinned or not | |
315 | * @page: struct page to check | |
316 | * | |
317 | * Pinned pages cannot be migrated. This is the same test as in | |
318 | * folio_migrate_mapping(), except that here we allow migration of a | |
319 | * ZONE_DEVICE page. | |
320 | */ | |
16ce101d | 321 | static bool migrate_vma_check_page(struct page *page, struct page *fault_page) |
76cbbead CH |
322 | { |
323 | /* | |
324 | * One extra ref because caller holds an extra reference, either from | |
325 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
326 | * a device page. | |
327 | */ | |
16ce101d | 328 | int extra = 1 + (page == fault_page); |
76cbbead CH |
329 | |
330 | /* | |
331 | * FIXME support THP (transparent huge page), it is bit more complex to | |
332 | * check them than regular pages, because they can be mapped with a pmd | |
333 | * or with a pte (split pte mapping). | |
334 | */ | |
335 | if (PageCompound(page)) | |
336 | return false; | |
337 | ||
338 | /* Page from ZONE_DEVICE have one extra reference */ | |
339 | if (is_zone_device_page(page)) | |
340 | extra++; | |
341 | ||
342 | /* For file back page */ | |
343 | if (page_mapping(page)) | |
344 | extra += 1 + page_has_private(page); | |
345 | ||
346 | if ((page_count(page) - extra) > page_mapcount(page)) | |
347 | return false; | |
348 | ||
349 | return true; | |
350 | } | |
351 | ||
352 | /* | |
44af0b45 AP |
353 | * Unmaps pages for migration. Returns number of source pfns marked as |
354 | * migrating. | |
76cbbead | 355 | */ |
241f6885 AP |
356 | static unsigned long migrate_device_unmap(unsigned long *src_pfns, |
357 | unsigned long npages, | |
358 | struct page *fault_page) | |
76cbbead | 359 | { |
76cbbead CH |
360 | unsigned long i, restore = 0; |
361 | bool allow_drain = true; | |
241f6885 | 362 | unsigned long unmapped = 0; |
76cbbead CH |
363 | |
364 | lru_add_drain(); | |
365 | ||
366 | for (i = 0; i < npages; i++) { | |
241f6885 | 367 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4b8554c5 | 368 | struct folio *folio; |
76cbbead | 369 | |
44af0b45 AP |
370 | if (!page) { |
371 | if (src_pfns[i] & MIGRATE_PFN_MIGRATE) | |
372 | unmapped++; | |
76cbbead | 373 | continue; |
44af0b45 | 374 | } |
76cbbead CH |
375 | |
376 | /* ZONE_DEVICE pages are not on LRU */ | |
377 | if (!is_zone_device_page(page)) { | |
378 | if (!PageLRU(page) && allow_drain) { | |
1fec6890 | 379 | /* Drain CPU's lru cache */ |
76cbbead CH |
380 | lru_add_drain_all(); |
381 | allow_drain = false; | |
382 | } | |
383 | ||
f7f9c00d | 384 | if (!isolate_lru_page(page)) { |
241f6885 | 385 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
386 | restore++; |
387 | continue; | |
388 | } | |
389 | ||
390 | /* Drop the reference we took in collect */ | |
391 | put_page(page); | |
392 | } | |
393 | ||
4b8554c5 MWO |
394 | folio = page_folio(page); |
395 | if (folio_mapped(folio)) | |
396 | try_to_migrate(folio, 0); | |
76cbbead | 397 | |
16ce101d | 398 | if (page_mapped(page) || |
241f6885 | 399 | !migrate_vma_check_page(page, fault_page)) { |
76cbbead CH |
400 | if (!is_zone_device_page(page)) { |
401 | get_page(page); | |
402 | putback_lru_page(page); | |
403 | } | |
404 | ||
241f6885 | 405 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
406 | restore++; |
407 | continue; | |
408 | } | |
241f6885 AP |
409 | |
410 | unmapped++; | |
76cbbead CH |
411 | } |
412 | ||
413 | for (i = 0; i < npages && restore; i++) { | |
241f6885 | 414 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4eecb8b9 | 415 | struct folio *folio; |
76cbbead | 416 | |
241f6885 | 417 | if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
76cbbead CH |
418 | continue; |
419 | ||
4eecb8b9 MWO |
420 | folio = page_folio(page); |
421 | remove_migration_ptes(folio, folio, false); | |
76cbbead | 422 | |
241f6885 | 423 | src_pfns[i] = 0; |
4eecb8b9 MWO |
424 | folio_unlock(folio); |
425 | folio_put(folio); | |
76cbbead CH |
426 | restore--; |
427 | } | |
241f6885 AP |
428 | |
429 | return unmapped; | |
430 | } | |
431 | ||
432 | /* | |
433 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
434 | * @migrate: migrate struct containing all migration information | |
435 | * | |
436 | * Isolate pages from the LRU and replace mappings (CPU page table pte) with a | |
437 | * special migration pte entry and check if it has been pinned. Pinned pages are | |
438 | * restored because we cannot migrate them. | |
439 | * | |
440 | * This is the last step before we call the device driver callback to allocate | |
441 | * destination memory and copy contents of original page over to new page. | |
442 | */ | |
443 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
444 | { | |
445 | migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages, | |
446 | migrate->fault_page); | |
76cbbead CH |
447 | } |
448 | ||
449 | /** | |
450 | * migrate_vma_setup() - prepare to migrate a range of memory | |
451 | * @args: contains the vma, start, and pfns arrays for the migration | |
452 | * | |
453 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
454 | * without an error. | |
455 | * | |
456 | * Prepare to migrate a range of memory virtual address range by collecting all | |
457 | * the pages backing each virtual address in the range, saving them inside the | |
458 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
459 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
460 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
461 | * corresponding src array entry. Then restores any pages that are pinned, by | |
462 | * remapping and unlocking those pages. | |
463 | * | |
464 | * The caller should then allocate destination memory and copy source memory to | |
465 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
466 | * flag set). Once these are allocated and copied, the caller must update each | |
467 | * corresponding entry in the dst array with the pfn value of the destination | |
468 | * page and with MIGRATE_PFN_VALID. Destination pages must be locked via | |
469 | * lock_page(). | |
470 | * | |
471 | * Note that the caller does not have to migrate all the pages that are marked | |
472 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
473 | * device memory to system memory. If the caller cannot migrate a device page | |
474 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
475 | * consequences for the userspace process, so it must be avoided if at all | |
476 | * possible. | |
477 | * | |
478 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
479 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
480 | * allowing the caller to allocate device memory for those unbacked virtual | |
481 | * addresses. For this the caller simply has to allocate device memory and | |
482 | * properly set the destination entry like for regular migration. Note that | |
483 | * this can still fail, and thus inside the device driver you must check if the | |
484 | * migration was successful for those entries after calling migrate_vma_pages(), | |
485 | * just like for regular migration. | |
486 | * | |
487 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
488 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
489 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
490 | * then migrate_vma_pages() to migrate struct page information from the source | |
491 | * struct page to the destination struct page. If it fails to migrate the | |
492 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
493 | * src array. | |
494 | * | |
495 | * At this point all successfully migrated pages have an entry in the src | |
496 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
497 | * array entry with MIGRATE_PFN_VALID flag set. | |
498 | * | |
499 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
500 | * successfully migrated, and which were not. Successfully migrated pages will | |
501 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
502 | * | |
503 | * It is safe to update device page table after migrate_vma_pages() because | |
504 | * both destination and source page are still locked, and the mmap_lock is held | |
505 | * in read mode (hence no one can unmap the range being migrated). | |
506 | * | |
507 | * Once the caller is done cleaning up things and updating its page table (if it | |
508 | * chose to do so, this is not an obligation) it finally calls | |
509 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
510 | * for successfully migrated pages or otherwise restore the CPU page table to | |
511 | * point to the original source pages. | |
512 | */ | |
513 | int migrate_vma_setup(struct migrate_vma *args) | |
514 | { | |
515 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
516 | ||
517 | args->start &= PAGE_MASK; | |
518 | args->end &= PAGE_MASK; | |
519 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
520 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
521 | return -EINVAL; | |
522 | if (nr_pages <= 0) | |
523 | return -EINVAL; | |
524 | if (args->start < args->vma->vm_start || | |
525 | args->start >= args->vma->vm_end) | |
526 | return -EINVAL; | |
527 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
528 | return -EINVAL; | |
529 | if (!args->src || !args->dst) | |
530 | return -EINVAL; | |
16ce101d AP |
531 | if (args->fault_page && !is_device_private_page(args->fault_page)) |
532 | return -EINVAL; | |
76cbbead CH |
533 | |
534 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
535 | args->cpages = 0; | |
536 | args->npages = 0; | |
537 | ||
538 | migrate_vma_collect(args); | |
539 | ||
540 | if (args->cpages) | |
541 | migrate_vma_unmap(args); | |
542 | ||
543 | /* | |
544 | * At this point pages are locked and unmapped, and thus they have | |
545 | * stable content and can safely be copied to destination memory that | |
546 | * is allocated by the drivers. | |
547 | */ | |
548 | return 0; | |
549 | ||
550 | } | |
551 | EXPORT_SYMBOL(migrate_vma_setup); | |
552 | ||
553 | /* | |
554 | * This code closely matches the code in: | |
555 | * __handle_mm_fault() | |
556 | * handle_pte_fault() | |
557 | * do_anonymous_page() | |
558 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
f25cbb7a | 559 | * private or coherent page. |
76cbbead CH |
560 | */ |
561 | static void migrate_vma_insert_page(struct migrate_vma *migrate, | |
562 | unsigned long addr, | |
563 | struct page *page, | |
564 | unsigned long *src) | |
565 | { | |
566 | struct vm_area_struct *vma = migrate->vma; | |
567 | struct mm_struct *mm = vma->vm_mm; | |
568 | bool flush = false; | |
569 | spinlock_t *ptl; | |
570 | pte_t entry; | |
571 | pgd_t *pgdp; | |
572 | p4d_t *p4dp; | |
573 | pud_t *pudp; | |
574 | pmd_t *pmdp; | |
575 | pte_t *ptep; | |
c33c7948 | 576 | pte_t orig_pte; |
76cbbead CH |
577 | |
578 | /* Only allow populating anonymous memory */ | |
579 | if (!vma_is_anonymous(vma)) | |
580 | goto abort; | |
581 | ||
582 | pgdp = pgd_offset(mm, addr); | |
583 | p4dp = p4d_alloc(mm, pgdp, addr); | |
584 | if (!p4dp) | |
585 | goto abort; | |
586 | pudp = pud_alloc(mm, p4dp, addr); | |
587 | if (!pudp) | |
588 | goto abort; | |
589 | pmdp = pmd_alloc(mm, pudp, addr); | |
590 | if (!pmdp) | |
591 | goto abort; | |
76cbbead CH |
592 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) |
593 | goto abort; | |
76cbbead CH |
594 | if (pte_alloc(mm, pmdp)) |
595 | goto abort; | |
76cbbead CH |
596 | if (unlikely(anon_vma_prepare(vma))) |
597 | goto abort; | |
598 | if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL)) | |
599 | goto abort; | |
600 | ||
601 | /* | |
602 | * The memory barrier inside __SetPageUptodate makes sure that | |
603 | * preceding stores to the page contents become visible before | |
604 | * the set_pte_at() write. | |
605 | */ | |
606 | __SetPageUptodate(page); | |
607 | ||
608 | if (is_device_private_page(page)) { | |
609 | swp_entry_t swp_entry; | |
610 | ||
611 | if (vma->vm_flags & VM_WRITE) | |
612 | swp_entry = make_writable_device_private_entry( | |
613 | page_to_pfn(page)); | |
614 | else | |
615 | swp_entry = make_readable_device_private_entry( | |
616 | page_to_pfn(page)); | |
617 | entry = swp_entry_to_pte(swp_entry); | |
618 | } else { | |
f25cbb7a AS |
619 | if (is_zone_device_page(page) && |
620 | !is_device_coherent_page(page)) { | |
76cbbead CH |
621 | pr_warn_once("Unsupported ZONE_DEVICE page type.\n"); |
622 | goto abort; | |
623 | } | |
624 | entry = mk_pte(page, vma->vm_page_prot); | |
625 | if (vma->vm_flags & VM_WRITE) | |
626 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
627 | } | |
628 | ||
629 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
4b56069c HD |
630 | if (!ptep) |
631 | goto abort; | |
c33c7948 RR |
632 | orig_pte = ptep_get(ptep); |
633 | ||
76cbbead CH |
634 | if (check_stable_address_space(mm)) |
635 | goto unlock_abort; | |
636 | ||
c33c7948 RR |
637 | if (pte_present(orig_pte)) { |
638 | unsigned long pfn = pte_pfn(orig_pte); | |
76cbbead CH |
639 | |
640 | if (!is_zero_pfn(pfn)) | |
641 | goto unlock_abort; | |
642 | flush = true; | |
c33c7948 | 643 | } else if (!pte_none(orig_pte)) |
76cbbead CH |
644 | goto unlock_abort; |
645 | ||
646 | /* | |
647 | * Check for userfaultfd but do not deliver the fault. Instead, | |
648 | * just back off. | |
649 | */ | |
650 | if (userfaultfd_missing(vma)) | |
651 | goto unlock_abort; | |
652 | ||
653 | inc_mm_counter(mm, MM_ANONPAGES); | |
40f2bbf7 | 654 | page_add_new_anon_rmap(page, vma, addr); |
76cbbead CH |
655 | if (!is_zone_device_page(page)) |
656 | lru_cache_add_inactive_or_unevictable(page, vma); | |
657 | get_page(page); | |
658 | ||
659 | if (flush) { | |
c33c7948 | 660 | flush_cache_page(vma, addr, pte_pfn(orig_pte)); |
76cbbead CH |
661 | ptep_clear_flush_notify(vma, addr, ptep); |
662 | set_pte_at_notify(mm, addr, ptep, entry); | |
663 | update_mmu_cache(vma, addr, ptep); | |
664 | } else { | |
665 | /* No need to invalidate - it was non-present before */ | |
666 | set_pte_at(mm, addr, ptep, entry); | |
667 | update_mmu_cache(vma, addr, ptep); | |
668 | } | |
669 | ||
670 | pte_unmap_unlock(ptep, ptl); | |
671 | *src = MIGRATE_PFN_MIGRATE; | |
672 | return; | |
673 | ||
674 | unlock_abort: | |
675 | pte_unmap_unlock(ptep, ptl); | |
676 | abort: | |
677 | *src &= ~MIGRATE_PFN_MIGRATE; | |
678 | } | |
679 | ||
e778406b | 680 | static void __migrate_device_pages(unsigned long *src_pfns, |
241f6885 AP |
681 | unsigned long *dst_pfns, unsigned long npages, |
682 | struct migrate_vma *migrate) | |
76cbbead | 683 | { |
76cbbead | 684 | struct mmu_notifier_range range; |
241f6885 | 685 | unsigned long i; |
76cbbead CH |
686 | bool notified = false; |
687 | ||
241f6885 AP |
688 | for (i = 0; i < npages; i++) { |
689 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); | |
690 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
691 | struct address_space *mapping; |
692 | int r; | |
693 | ||
694 | if (!newpage) { | |
241f6885 | 695 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
696 | continue; |
697 | } | |
698 | ||
699 | if (!page) { | |
241f6885 AP |
700 | unsigned long addr; |
701 | ||
e778406b AP |
702 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
703 | continue; | |
704 | ||
b05a79d4 AP |
705 | /* |
706 | * The only time there is no vma is when called from | |
707 | * migrate_device_coherent_page(). However this isn't | |
708 | * called if the page could not be unmapped. | |
709 | */ | |
241f6885 AP |
710 | VM_BUG_ON(!migrate); |
711 | addr = migrate->start + i*PAGE_SIZE; | |
76cbbead CH |
712 | if (!notified) { |
713 | notified = true; | |
714 | ||
715 | mmu_notifier_range_init_owner(&range, | |
7d4a8be0 | 716 | MMU_NOTIFY_MIGRATE, 0, |
76cbbead CH |
717 | migrate->vma->vm_mm, addr, migrate->end, |
718 | migrate->pgmap_owner); | |
719 | mmu_notifier_invalidate_range_start(&range); | |
720 | } | |
721 | migrate_vma_insert_page(migrate, addr, newpage, | |
241f6885 | 722 | &src_pfns[i]); |
76cbbead CH |
723 | continue; |
724 | } | |
725 | ||
726 | mapping = page_mapping(page); | |
727 | ||
f25cbb7a AS |
728 | if (is_device_private_page(newpage) || |
729 | is_device_coherent_page(newpage)) { | |
76cbbead | 730 | /* |
f25cbb7a AS |
731 | * For now only support anonymous memory migrating to |
732 | * device private or coherent memory. | |
76cbbead CH |
733 | */ |
734 | if (mapping) { | |
241f6885 | 735 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
736 | continue; |
737 | } | |
738 | } else if (is_zone_device_page(newpage)) { | |
739 | /* | |
740 | * Other types of ZONE_DEVICE page are not supported. | |
741 | */ | |
241f6885 | 742 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
743 | continue; |
744 | } | |
745 | ||
241f6885 | 746 | if (migrate && migrate->fault_page == page) |
16ce101d AP |
747 | r = migrate_folio_extra(mapping, page_folio(newpage), |
748 | page_folio(page), | |
749 | MIGRATE_SYNC_NO_COPY, 1); | |
750 | else | |
751 | r = migrate_folio(mapping, page_folio(newpage), | |
752 | page_folio(page), MIGRATE_SYNC_NO_COPY); | |
76cbbead | 753 | if (r != MIGRATEPAGE_SUCCESS) |
241f6885 | 754 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
755 | } |
756 | ||
757 | /* | |
758 | * No need to double call mmu_notifier->invalidate_range() callback as | |
759 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
760 | * did already call it. | |
761 | */ | |
762 | if (notified) | |
763 | mmu_notifier_invalidate_range_only_end(&range); | |
764 | } | |
76cbbead | 765 | |
e778406b AP |
766 | /** |
767 | * migrate_device_pages() - migrate meta-data from src page to dst page | |
768 | * @src_pfns: src_pfns returned from migrate_device_range() | |
769 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
770 | * @npages: number of pages in the range | |
771 | * | |
772 | * Equivalent to migrate_vma_pages(). This is called to migrate struct page | |
773 | * meta-data from source struct page to destination. | |
774 | */ | |
775 | void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns, | |
776 | unsigned long npages) | |
777 | { | |
778 | __migrate_device_pages(src_pfns, dst_pfns, npages, NULL); | |
779 | } | |
780 | EXPORT_SYMBOL(migrate_device_pages); | |
781 | ||
76cbbead | 782 | /** |
241f6885 | 783 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
76cbbead CH |
784 | * @migrate: migrate struct containing all migration information |
785 | * | |
241f6885 AP |
786 | * This migrates struct page meta-data from source struct page to destination |
787 | * struct page. This effectively finishes the migration from source page to the | |
788 | * destination page. | |
76cbbead | 789 | */ |
241f6885 AP |
790 | void migrate_vma_pages(struct migrate_vma *migrate) |
791 | { | |
e778406b | 792 | __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate); |
241f6885 AP |
793 | } |
794 | EXPORT_SYMBOL(migrate_vma_pages); | |
795 | ||
e778406b AP |
796 | /* |
797 | * migrate_device_finalize() - complete page migration | |
798 | * @src_pfns: src_pfns returned from migrate_device_range() | |
799 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
800 | * @npages: number of pages in the range | |
801 | * | |
802 | * Completes migration of the page by removing special migration entries. | |
803 | * Drivers must ensure copying of page data is complete and visible to the CPU | |
804 | * before calling this. | |
805 | */ | |
806 | void migrate_device_finalize(unsigned long *src_pfns, | |
807 | unsigned long *dst_pfns, unsigned long npages) | |
76cbbead | 808 | { |
76cbbead CH |
809 | unsigned long i; |
810 | ||
811 | for (i = 0; i < npages; i++) { | |
4eecb8b9 | 812 | struct folio *dst, *src; |
241f6885 AP |
813 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); |
814 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
815 | |
816 | if (!page) { | |
817 | if (newpage) { | |
818 | unlock_page(newpage); | |
819 | put_page(newpage); | |
820 | } | |
821 | continue; | |
822 | } | |
823 | ||
241f6885 | 824 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
76cbbead CH |
825 | if (newpage) { |
826 | unlock_page(newpage); | |
827 | put_page(newpage); | |
828 | } | |
829 | newpage = page; | |
830 | } | |
831 | ||
4eecb8b9 MWO |
832 | src = page_folio(page); |
833 | dst = page_folio(newpage); | |
834 | remove_migration_ptes(src, dst, false); | |
835 | folio_unlock(src); | |
76cbbead CH |
836 | |
837 | if (is_zone_device_page(page)) | |
838 | put_page(page); | |
839 | else | |
840 | putback_lru_page(page); | |
841 | ||
842 | if (newpage != page) { | |
843 | unlock_page(newpage); | |
844 | if (is_zone_device_page(newpage)) | |
845 | put_page(newpage); | |
846 | else | |
847 | putback_lru_page(newpage); | |
848 | } | |
849 | } | |
850 | } | |
e778406b | 851 | EXPORT_SYMBOL(migrate_device_finalize); |
241f6885 AP |
852 | |
853 | /** | |
854 | * migrate_vma_finalize() - restore CPU page table entry | |
855 | * @migrate: migrate struct containing all migration information | |
856 | * | |
857 | * This replaces the special migration pte entry with either a mapping to the | |
858 | * new page if migration was successful for that page, or to the original page | |
859 | * otherwise. | |
860 | * | |
861 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
862 | * refcount, for device pages. | |
863 | */ | |
864 | void migrate_vma_finalize(struct migrate_vma *migrate) | |
865 | { | |
866 | migrate_device_finalize(migrate->src, migrate->dst, migrate->npages); | |
867 | } | |
76cbbead | 868 | EXPORT_SYMBOL(migrate_vma_finalize); |
b05a79d4 | 869 | |
e778406b AP |
870 | /** |
871 | * migrate_device_range() - migrate device private pfns to normal memory. | |
872 | * @src_pfns: array large enough to hold migrating source device private pfns. | |
873 | * @start: starting pfn in the range to migrate. | |
874 | * @npages: number of pages to migrate. | |
875 | * | |
876 | * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that | |
877 | * instead of looking up pages based on virtual address mappings a range of | |
878 | * device pfns that should be migrated to system memory is used instead. | |
879 | * | |
880 | * This is useful when a driver needs to free device memory but doesn't know the | |
881 | * virtual mappings of every page that may be in device memory. For example this | |
882 | * is often the case when a driver is being unloaded or unbound from a device. | |
883 | * | |
884 | * Like migrate_vma_setup() this function will take a reference and lock any | |
885 | * migrating pages that aren't free before unmapping them. Drivers may then | |
886 | * allocate destination pages and start copying data from the device to CPU | |
887 | * memory before calling migrate_device_pages(). | |
888 | */ | |
889 | int migrate_device_range(unsigned long *src_pfns, unsigned long start, | |
890 | unsigned long npages) | |
891 | { | |
892 | unsigned long i, pfn; | |
893 | ||
894 | for (pfn = start, i = 0; i < npages; pfn++, i++) { | |
895 | struct page *page = pfn_to_page(pfn); | |
896 | ||
897 | if (!get_page_unless_zero(page)) { | |
898 | src_pfns[i] = 0; | |
899 | continue; | |
900 | } | |
901 | ||
902 | if (!trylock_page(page)) { | |
903 | src_pfns[i] = 0; | |
904 | put_page(page); | |
905 | continue; | |
906 | } | |
907 | ||
908 | src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; | |
909 | } | |
910 | ||
911 | migrate_device_unmap(src_pfns, npages, NULL); | |
912 | ||
913 | return 0; | |
914 | } | |
915 | EXPORT_SYMBOL(migrate_device_range); | |
916 | ||
b05a79d4 AP |
917 | /* |
918 | * Migrate a device coherent page back to normal memory. The caller should have | |
919 | * a reference on page which will be copied to the new page if migration is | |
920 | * successful or dropped on failure. | |
921 | */ | |
922 | int migrate_device_coherent_page(struct page *page) | |
923 | { | |
924 | unsigned long src_pfn, dst_pfn = 0; | |
b05a79d4 AP |
925 | struct page *dpage; |
926 | ||
927 | WARN_ON_ONCE(PageCompound(page)); | |
928 | ||
929 | lock_page(page); | |
930 | src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE; | |
b05a79d4 AP |
931 | |
932 | /* | |
933 | * We don't have a VMA and don't need to walk the page tables to find | |
934 | * the source page. So call migrate_vma_unmap() directly to unmap the | |
935 | * page as migrate_vma_setup() will fail if args.vma == NULL. | |
936 | */ | |
241f6885 | 937 | migrate_device_unmap(&src_pfn, 1, NULL); |
b05a79d4 AP |
938 | if (!(src_pfn & MIGRATE_PFN_MIGRATE)) |
939 | return -EBUSY; | |
940 | ||
941 | dpage = alloc_page(GFP_USER | __GFP_NOWARN); | |
942 | if (dpage) { | |
943 | lock_page(dpage); | |
944 | dst_pfn = migrate_pfn(page_to_pfn(dpage)); | |
945 | } | |
946 | ||
e778406b | 947 | migrate_device_pages(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
948 | if (src_pfn & MIGRATE_PFN_MIGRATE) |
949 | copy_highpage(dpage, page); | |
241f6885 | 950 | migrate_device_finalize(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
951 | |
952 | if (src_pfn & MIGRATE_PFN_MIGRATE) | |
953 | return 0; | |
954 | return -EBUSY; | |
955 | } |