Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
4bbd4c77 KS |
2 | #include <linux/kernel.h> |
3 | #include <linux/errno.h> | |
4 | #include <linux/err.h> | |
5 | #include <linux/spinlock.h> | |
6 | ||
4bbd4c77 | 7 | #include <linux/mm.h> |
89c1905d | 8 | #include <linux/memfd.h> |
3565fce3 | 9 | #include <linux/memremap.h> |
4bbd4c77 KS |
10 | #include <linux/pagemap.h> |
11 | #include <linux/rmap.h> | |
12 | #include <linux/swap.h> | |
13 | #include <linux/swapops.h> | |
1507f512 | 14 | #include <linux/secretmem.h> |
4bbd4c77 | 15 | |
174cd4b1 | 16 | #include <linux/sched/signal.h> |
2667f50e | 17 | #include <linux/rwsem.h> |
f30c59e9 | 18 | #include <linux/hugetlb.h> |
9a4e9f3b AK |
19 | #include <linux/migrate.h> |
20 | #include <linux/mm_inline.h> | |
89c1905d | 21 | #include <linux/pagevec.h> |
9a4e9f3b | 22 | #include <linux/sched/mm.h> |
a6e79df9 | 23 | #include <linux/shmem_fs.h> |
1027e443 | 24 | |
33a709b2 | 25 | #include <asm/mmu_context.h> |
1027e443 | 26 | #include <asm/tlbflush.h> |
2667f50e | 27 | |
4bbd4c77 KS |
28 | #include "internal.h" |
29 | ||
df06b37f KB |
30 | struct follow_page_context { |
31 | struct dev_pagemap *pgmap; | |
32 | unsigned int page_mask; | |
33 | }; | |
34 | ||
b6a2619c DH |
35 | static inline void sanity_check_pinned_pages(struct page **pages, |
36 | unsigned long npages) | |
37 | { | |
38 | if (!IS_ENABLED(CONFIG_DEBUG_VM)) | |
39 | return; | |
40 | ||
41 | /* | |
42 | * We only pin anonymous pages if they are exclusive. Once pinned, we | |
43 | * can no longer turn them possibly shared and PageAnonExclusive() will | |
44 | * stick around until the page is freed. | |
45 | * | |
46 | * We'd like to verify that our pinned anonymous pages are still mapped | |
47 | * exclusively. The issue with anon THP is that we don't know how | |
48 | * they are/were mapped when pinning them. However, for anon | |
49 | * THP we can assume that either the given page (PTE-mapped THP) or | |
50 | * the head page (PMD-mapped THP) should be PageAnonExclusive(). If | |
51 | * neither is the case, there is certainly something wrong. | |
52 | */ | |
53 | for (; npages; npages--, pages++) { | |
54 | struct page *page = *pages; | |
55 | struct folio *folio = page_folio(page); | |
56 | ||
c8070b78 DH |
57 | if (is_zero_page(page) || |
58 | !folio_test_anon(folio)) | |
b6a2619c DH |
59 | continue; |
60 | if (!folio_test_large(folio) || folio_test_hugetlb(folio)) | |
61 | VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page); | |
62 | else | |
63 | /* Either a PTE-mapped or a PMD-mapped THP. */ | |
64 | VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) && | |
65 | !PageAnonExclusive(page), page); | |
66 | } | |
67 | } | |
68 | ||
cd1adf1b | 69 | /* |
ece1ed7b | 70 | * Return the folio with ref appropriately incremented, |
cd1adf1b | 71 | * or NULL if that failed. |
a707cdd5 | 72 | */ |
ece1ed7b | 73 | static inline struct folio *try_get_folio(struct page *page, int refs) |
a707cdd5 | 74 | { |
ece1ed7b | 75 | struct folio *folio; |
a707cdd5 | 76 | |
59409373 | 77 | retry: |
ece1ed7b MWO |
78 | folio = page_folio(page); |
79 | if (WARN_ON_ONCE(folio_ref_count(folio) < 0)) | |
a707cdd5 | 80 | return NULL; |
fa2690af | 81 | if (unlikely(!folio_ref_try_add(folio, refs))) |
a707cdd5 | 82 | return NULL; |
c24d3732 JH |
83 | |
84 | /* | |
ece1ed7b MWO |
85 | * At this point we have a stable reference to the folio; but it |
86 | * could be that between calling page_folio() and the refcount | |
87 | * increment, the folio was split, in which case we'd end up | |
88 | * holding a reference on a folio that has nothing to do with the page | |
c24d3732 | 89 | * we were given anymore. |
ece1ed7b MWO |
90 | * So now that the folio is stable, recheck that the page still |
91 | * belongs to this folio. | |
c24d3732 | 92 | */ |
ece1ed7b | 93 | if (unlikely(page_folio(page) != folio)) { |
53e45c4f | 94 | if (!put_devmap_managed_folio_refs(folio, refs)) |
f4f451a1 | 95 | folio_put_refs(folio, refs); |
59409373 | 96 | goto retry; |
c24d3732 JH |
97 | } |
98 | ||
ece1ed7b | 99 | return folio; |
a707cdd5 JH |
100 | } |
101 | ||
d8ddc099 | 102 | static void gup_put_folio(struct folio *folio, int refs, unsigned int flags) |
4509b42c JG |
103 | { |
104 | if (flags & FOLL_PIN) { | |
c8070b78 DH |
105 | if (is_zero_folio(folio)) |
106 | return; | |
d8ddc099 MWO |
107 | node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs); |
108 | if (folio_test_large(folio)) | |
94688e8e | 109 | atomic_sub(refs, &folio->_pincount); |
4509b42c JG |
110 | else |
111 | refs *= GUP_PIN_COUNTING_BIAS; | |
112 | } | |
113 | ||
53e45c4f | 114 | if (!put_devmap_managed_folio_refs(folio, refs)) |
f4f451a1 | 115 | folio_put_refs(folio, refs); |
4509b42c JG |
116 | } |
117 | ||
3faa52c0 | 118 | /** |
f442fa61 YS |
119 | * try_grab_folio() - add a folio's refcount by a flag-dependent amount |
120 | * @folio: pointer to folio to be grabbed | |
121 | * @refs: the value to (effectively) add to the folio's refcount | |
122 | * @flags: gup flags: these are the FOLL_* flag values | |
3faa52c0 JH |
123 | * |
124 | * This might not do anything at all, depending on the flags argument. | |
125 | * | |
126 | * "grab" names in this file mean, "look at flags to decide whether to use | |
f442fa61 | 127 | * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount. |
3faa52c0 | 128 | * |
3faa52c0 | 129 | * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same |
f442fa61 | 130 | * time. |
3faa52c0 | 131 | * |
0f089235 LG |
132 | * Return: 0 for success, or if no action was required (if neither FOLL_PIN |
133 | * nor FOLL_GET was set, nothing is done). A negative error code for failure: | |
134 | * | |
f442fa61 | 135 | * -ENOMEM FOLL_GET or FOLL_PIN was set, but the folio could not |
0f089235 | 136 | * be grabbed. |
f442fa61 YS |
137 | * |
138 | * It is called when we have a stable reference for the folio, typically in | |
139 | * GUP slow path. | |
3faa52c0 | 140 | */ |
f442fa61 YS |
141 | int __must_check try_grab_folio(struct folio *folio, int refs, |
142 | unsigned int flags) | |
3faa52c0 | 143 | { |
5fec0719 | 144 | if (WARN_ON_ONCE(folio_ref_count(folio) <= 0)) |
0f089235 | 145 | return -ENOMEM; |
3faa52c0 | 146 | |
f442fa61 | 147 | if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(&folio->page))) |
4003f107 | 148 | return -EREMOTEIO; |
3faa52c0 | 149 | |
c36c04c2 | 150 | if (flags & FOLL_GET) |
f442fa61 | 151 | folio_ref_add(folio, refs); |
c36c04c2 | 152 | else if (flags & FOLL_PIN) { |
c8070b78 DH |
153 | /* |
154 | * Don't take a pin on the zero page - it's not going anywhere | |
155 | * and it is used in a *lot* of places. | |
156 | */ | |
f442fa61 | 157 | if (is_zero_folio(folio)) |
c8070b78 DH |
158 | return 0; |
159 | ||
c36c04c2 | 160 | /* |
f442fa61 | 161 | * Increment the normal page refcount field at least once, |
78d9d6ce | 162 | * so that the page really is pinned. |
c36c04c2 | 163 | */ |
5fec0719 | 164 | if (folio_test_large(folio)) { |
f442fa61 YS |
165 | folio_ref_add(folio, refs); |
166 | atomic_add(refs, &folio->_pincount); | |
8ea2979c | 167 | } else { |
f442fa61 | 168 | folio_ref_add(folio, refs * GUP_PIN_COUNTING_BIAS); |
8ea2979c | 169 | } |
c36c04c2 | 170 | |
f442fa61 | 171 | node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs); |
c36c04c2 JH |
172 | } |
173 | ||
0f089235 | 174 | return 0; |
3faa52c0 JH |
175 | } |
176 | ||
3faa52c0 JH |
177 | /** |
178 | * unpin_user_page() - release a dma-pinned page | |
179 | * @page: pointer to page to be released | |
180 | * | |
181 | * Pages that were pinned via pin_user_pages*() must be released via either | |
182 | * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so | |
183 | * that such pages can be separately tracked and uniquely handled. In | |
184 | * particular, interactions with RDMA and filesystems need special handling. | |
185 | */ | |
186 | void unpin_user_page(struct page *page) | |
187 | { | |
b6a2619c | 188 | sanity_check_pinned_pages(&page, 1); |
d8ddc099 | 189 | gup_put_folio(page_folio(page), 1, FOLL_PIN); |
3faa52c0 JH |
190 | } |
191 | EXPORT_SYMBOL(unpin_user_page); | |
192 | ||
6cc04054 VK |
193 | /** |
194 | * unpin_folio() - release a dma-pinned folio | |
195 | * @folio: pointer to folio to be released | |
196 | * | |
197 | * Folios that were pinned via memfd_pin_folios() or other similar routines | |
198 | * must be released either using unpin_folio() or unpin_folios(). | |
199 | */ | |
200 | void unpin_folio(struct folio *folio) | |
201 | { | |
202 | gup_put_folio(folio, 1, FOLL_PIN); | |
203 | } | |
204 | EXPORT_SYMBOL_GPL(unpin_folio); | |
205 | ||
1101fb8f DH |
206 | /** |
207 | * folio_add_pin - Try to get an additional pin on a pinned folio | |
208 | * @folio: The folio to be pinned | |
209 | * | |
210 | * Get an additional pin on a folio we already have a pin on. Makes no change | |
211 | * if the folio is a zero_page. | |
212 | */ | |
213 | void folio_add_pin(struct folio *folio) | |
214 | { | |
215 | if (is_zero_folio(folio)) | |
216 | return; | |
217 | ||
218 | /* | |
219 | * Similar to try_grab_folio(): be sure to *also* increment the normal | |
220 | * page refcount field at least once, so that the page really is | |
221 | * pinned. | |
222 | */ | |
223 | if (folio_test_large(folio)) { | |
224 | WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1); | |
225 | folio_ref_inc(folio); | |
226 | atomic_inc(&folio->_pincount); | |
227 | } else { | |
228 | WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS); | |
229 | folio_ref_add(folio, GUP_PIN_COUNTING_BIAS); | |
230 | } | |
231 | } | |
232 | ||
659508f9 | 233 | static inline struct folio *gup_folio_range_next(struct page *start, |
8f39f5fc | 234 | unsigned long npages, unsigned long i, unsigned int *ntails) |
458a4f78 | 235 | { |
659508f9 MWO |
236 | struct page *next = nth_page(start, i); |
237 | struct folio *folio = page_folio(next); | |
458a4f78 JM |
238 | unsigned int nr = 1; |
239 | ||
659508f9 | 240 | if (folio_test_large(folio)) |
4c654229 | 241 | nr = min_t(unsigned int, npages - i, |
659508f9 | 242 | folio_nr_pages(folio) - folio_page_idx(folio, next)); |
458a4f78 | 243 | |
458a4f78 | 244 | *ntails = nr; |
659508f9 | 245 | return folio; |
458a4f78 JM |
246 | } |
247 | ||
12521c76 | 248 | static inline struct folio *gup_folio_next(struct page **list, |
28297dbc | 249 | unsigned long npages, unsigned long i, unsigned int *ntails) |
8745d7f6 | 250 | { |
12521c76 | 251 | struct folio *folio = page_folio(list[i]); |
8745d7f6 JM |
252 | unsigned int nr; |
253 | ||
8745d7f6 | 254 | for (nr = i + 1; nr < npages; nr++) { |
12521c76 | 255 | if (page_folio(list[nr]) != folio) |
8745d7f6 JM |
256 | break; |
257 | } | |
258 | ||
8745d7f6 | 259 | *ntails = nr - i; |
12521c76 | 260 | return folio; |
8745d7f6 JM |
261 | } |
262 | ||
fc1d8e7c | 263 | /** |
f1f6a7dd | 264 | * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages |
2d15eb31 | 265 | * @pages: array of pages to be maybe marked dirty, and definitely released. |
fc1d8e7c | 266 | * @npages: number of pages in the @pages array. |
2d15eb31 | 267 | * @make_dirty: whether to mark the pages dirty |
fc1d8e7c JH |
268 | * |
269 | * "gup-pinned page" refers to a page that has had one of the get_user_pages() | |
270 | * variants called on that page. | |
271 | * | |
272 | * For each page in the @pages array, make that page (or its head page, if a | |
2d15eb31 | 273 | * compound page) dirty, if @make_dirty is true, and if the page was previously |
f1f6a7dd JH |
274 | * listed as clean. In any case, releases all pages using unpin_user_page(), |
275 | * possibly via unpin_user_pages(), for the non-dirty case. | |
fc1d8e7c | 276 | * |
f1f6a7dd | 277 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 278 | * |
2d15eb31 | 279 | * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is |
280 | * required, then the caller should a) verify that this is really correct, | |
281 | * because _lock() is usually required, and b) hand code it: | |
f1f6a7dd | 282 | * set_page_dirty_lock(), unpin_user_page(). |
fc1d8e7c JH |
283 | * |
284 | */ | |
f1f6a7dd JH |
285 | void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages, |
286 | bool make_dirty) | |
fc1d8e7c | 287 | { |
12521c76 MWO |
288 | unsigned long i; |
289 | struct folio *folio; | |
290 | unsigned int nr; | |
2d15eb31 | 291 | |
292 | if (!make_dirty) { | |
f1f6a7dd | 293 | unpin_user_pages(pages, npages); |
2d15eb31 | 294 | return; |
295 | } | |
296 | ||
b6a2619c | 297 | sanity_check_pinned_pages(pages, npages); |
12521c76 MWO |
298 | for (i = 0; i < npages; i += nr) { |
299 | folio = gup_folio_next(pages, npages, i, &nr); | |
2d15eb31 | 300 | /* |
301 | * Checking PageDirty at this point may race with | |
302 | * clear_page_dirty_for_io(), but that's OK. Two key | |
303 | * cases: | |
304 | * | |
305 | * 1) This code sees the page as already dirty, so it | |
306 | * skips the call to set_page_dirty(). That could happen | |
307 | * because clear_page_dirty_for_io() called | |
a929e0d1 | 308 | * folio_mkclean(), followed by set_page_dirty(). |
2d15eb31 | 309 | * However, now the page is going to get written back, |
310 | * which meets the original intention of setting it | |
311 | * dirty, so all is well: clear_page_dirty_for_io() goes | |
312 | * on to call TestClearPageDirty(), and write the page | |
313 | * back. | |
314 | * | |
315 | * 2) This code sees the page as clean, so it calls | |
316 | * set_page_dirty(). The page stays dirty, despite being | |
317 | * written back, so it gets written back again in the | |
318 | * next writeback cycle. This is harmless. | |
319 | */ | |
12521c76 MWO |
320 | if (!folio_test_dirty(folio)) { |
321 | folio_lock(folio); | |
322 | folio_mark_dirty(folio); | |
323 | folio_unlock(folio); | |
324 | } | |
325 | gup_put_folio(folio, nr, FOLL_PIN); | |
2d15eb31 | 326 | } |
fc1d8e7c | 327 | } |
f1f6a7dd | 328 | EXPORT_SYMBOL(unpin_user_pages_dirty_lock); |
fc1d8e7c | 329 | |
458a4f78 JM |
330 | /** |
331 | * unpin_user_page_range_dirty_lock() - release and optionally dirty | |
332 | * gup-pinned page range | |
333 | * | |
334 | * @page: the starting page of a range maybe marked dirty, and definitely released. | |
335 | * @npages: number of consecutive pages to release. | |
336 | * @make_dirty: whether to mark the pages dirty | |
337 | * | |
338 | * "gup-pinned page range" refers to a range of pages that has had one of the | |
339 | * pin_user_pages() variants called on that page. | |
340 | * | |
341 | * For the page ranges defined by [page .. page+npages], make that range (or | |
342 | * its head pages, if a compound page) dirty, if @make_dirty is true, and if the | |
343 | * page range was previously listed as clean. | |
344 | * | |
345 | * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is | |
346 | * required, then the caller should a) verify that this is really correct, | |
347 | * because _lock() is usually required, and b) hand code it: | |
348 | * set_page_dirty_lock(), unpin_user_page(). | |
349 | * | |
350 | */ | |
351 | void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages, | |
352 | bool make_dirty) | |
353 | { | |
659508f9 MWO |
354 | unsigned long i; |
355 | struct folio *folio; | |
356 | unsigned int nr; | |
357 | ||
358 | for (i = 0; i < npages; i += nr) { | |
359 | folio = gup_folio_range_next(page, npages, i, &nr); | |
360 | if (make_dirty && !folio_test_dirty(folio)) { | |
361 | folio_lock(folio); | |
362 | folio_mark_dirty(folio); | |
363 | folio_unlock(folio); | |
364 | } | |
365 | gup_put_folio(folio, nr, FOLL_PIN); | |
458a4f78 JM |
366 | } |
367 | } | |
368 | EXPORT_SYMBOL(unpin_user_page_range_dirty_lock); | |
369 | ||
23babe19 | 370 | static void gup_fast_unpin_user_pages(struct page **pages, unsigned long npages) |
b6a2619c DH |
371 | { |
372 | unsigned long i; | |
373 | struct folio *folio; | |
374 | unsigned int nr; | |
375 | ||
376 | /* | |
377 | * Don't perform any sanity checks because we might have raced with | |
378 | * fork() and some anonymous pages might now actually be shared -- | |
379 | * which is why we're unpinning after all. | |
380 | */ | |
381 | for (i = 0; i < npages; i += nr) { | |
382 | folio = gup_folio_next(pages, npages, i, &nr); | |
383 | gup_put_folio(folio, nr, FOLL_PIN); | |
384 | } | |
385 | } | |
386 | ||
fc1d8e7c | 387 | /** |
f1f6a7dd | 388 | * unpin_user_pages() - release an array of gup-pinned pages. |
fc1d8e7c JH |
389 | * @pages: array of pages to be marked dirty and released. |
390 | * @npages: number of pages in the @pages array. | |
391 | * | |
f1f6a7dd | 392 | * For each page in the @pages array, release the page using unpin_user_page(). |
fc1d8e7c | 393 | * |
f1f6a7dd | 394 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 395 | */ |
f1f6a7dd | 396 | void unpin_user_pages(struct page **pages, unsigned long npages) |
fc1d8e7c | 397 | { |
12521c76 MWO |
398 | unsigned long i; |
399 | struct folio *folio; | |
400 | unsigned int nr; | |
fc1d8e7c | 401 | |
146608bb JH |
402 | /* |
403 | * If this WARN_ON() fires, then the system *might* be leaking pages (by | |
404 | * leaving them pinned), but probably not. More likely, gup/pup returned | |
405 | * a hard -ERRNO error to the caller, who erroneously passed it here. | |
406 | */ | |
407 | if (WARN_ON(IS_ERR_VALUE(npages))) | |
408 | return; | |
31b912de | 409 | |
b6a2619c | 410 | sanity_check_pinned_pages(pages, npages); |
12521c76 MWO |
411 | for (i = 0; i < npages; i += nr) { |
412 | folio = gup_folio_next(pages, npages, i, &nr); | |
413 | gup_put_folio(folio, nr, FOLL_PIN); | |
e7602748 | 414 | } |
fc1d8e7c | 415 | } |
f1f6a7dd | 416 | EXPORT_SYMBOL(unpin_user_pages); |
fc1d8e7c | 417 | |
d3bfbfb1 KK |
418 | /** |
419 | * unpin_user_folio() - release pages of a folio | |
420 | * @folio: pointer to folio to be released | |
421 | * @npages: number of pages of same folio | |
422 | * | |
423 | * Release npages of the folio | |
424 | */ | |
425 | void unpin_user_folio(struct folio *folio, unsigned long npages) | |
426 | { | |
427 | gup_put_folio(folio, npages, FOLL_PIN); | |
428 | } | |
429 | EXPORT_SYMBOL(unpin_user_folio); | |
430 | ||
6cc04054 VK |
431 | /** |
432 | * unpin_folios() - release an array of gup-pinned folios. | |
433 | * @folios: array of folios to be marked dirty and released. | |
434 | * @nfolios: number of folios in the @folios array. | |
435 | * | |
436 | * For each folio in the @folios array, release the folio using gup_put_folio. | |
437 | * | |
438 | * Please see the unpin_folio() documentation for details. | |
439 | */ | |
440 | void unpin_folios(struct folio **folios, unsigned long nfolios) | |
441 | { | |
442 | unsigned long i = 0, j; | |
443 | ||
444 | /* | |
445 | * If this WARN_ON() fires, then the system *might* be leaking folios | |
446 | * (by leaving them pinned), but probably not. More likely, gup/pup | |
447 | * returned a hard -ERRNO error to the caller, who erroneously passed | |
448 | * it here. | |
449 | */ | |
450 | if (WARN_ON(IS_ERR_VALUE(nfolios))) | |
451 | return; | |
452 | ||
453 | while (i < nfolios) { | |
454 | for (j = i + 1; j < nfolios; j++) | |
455 | if (folios[i] != folios[j]) | |
456 | break; | |
457 | ||
458 | if (folios[i]) | |
459 | gup_put_folio(folios[i], j - i, FOLL_PIN); | |
460 | i = j; | |
461 | } | |
462 | } | |
463 | EXPORT_SYMBOL_GPL(unpin_folios); | |
464 | ||
a458b76a AA |
465 | /* |
466 | * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's | |
467 | * lifecycle. Avoid setting the bit unless necessary, or it might cause write | |
468 | * cache bouncing on large SMP machines for concurrent pinned gups. | |
469 | */ | |
470 | static inline void mm_set_has_pinned_flag(unsigned long *mm_flags) | |
471 | { | |
472 | if (!test_bit(MMF_HAS_PINNED, mm_flags)) | |
473 | set_bit(MMF_HAS_PINNED, mm_flags); | |
474 | } | |
475 | ||
050a9adc | 476 | #ifdef CONFIG_MMU |
a12083d7 | 477 | |
8268614b | 478 | #ifdef CONFIG_HAVE_GUP_FAST |
a12083d7 PX |
479 | static int record_subpages(struct page *page, unsigned long sz, |
480 | unsigned long addr, unsigned long end, | |
481 | struct page **pages) | |
482 | { | |
483 | struct page *start_page; | |
484 | int nr; | |
485 | ||
486 | start_page = nth_page(page, (addr & (sz - 1)) >> PAGE_SHIFT); | |
487 | for (nr = 0; addr != end; nr++, addr += PAGE_SIZE) | |
488 | pages[nr] = nth_page(start_page, nr); | |
489 | ||
490 | return nr; | |
491 | } | |
f442fa61 YS |
492 | |
493 | /** | |
494 | * try_grab_folio_fast() - Attempt to get or pin a folio in fast path. | |
495 | * @page: pointer to page to be grabbed | |
496 | * @refs: the value to (effectively) add to the folio's refcount | |
497 | * @flags: gup flags: these are the FOLL_* flag values. | |
498 | * | |
499 | * "grab" names in this file mean, "look at flags to decide whether to use | |
500 | * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount. | |
501 | * | |
502 | * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the | |
503 | * same time. (That's true throughout the get_user_pages*() and | |
504 | * pin_user_pages*() APIs.) Cases: | |
505 | * | |
506 | * FOLL_GET: folio's refcount will be incremented by @refs. | |
507 | * | |
508 | * FOLL_PIN on large folios: folio's refcount will be incremented by | |
509 | * @refs, and its pincount will be incremented by @refs. | |
510 | * | |
511 | * FOLL_PIN on single-page folios: folio's refcount will be incremented by | |
512 | * @refs * GUP_PIN_COUNTING_BIAS. | |
513 | * | |
514 | * Return: The folio containing @page (with refcount appropriately | |
515 | * incremented) for success, or NULL upon failure. If neither FOLL_GET | |
516 | * nor FOLL_PIN was set, that's considered failure, and furthermore, | |
517 | * a likely bug in the caller, so a warning is also emitted. | |
518 | * | |
519 | * It uses add ref unless zero to elevate the folio refcount and must be called | |
520 | * in fast path only. | |
521 | */ | |
522 | static struct folio *try_grab_folio_fast(struct page *page, int refs, | |
523 | unsigned int flags) | |
524 | { | |
525 | struct folio *folio; | |
526 | ||
527 | /* Raise warn if it is not called in fast GUP */ | |
528 | VM_WARN_ON_ONCE(!irqs_disabled()); | |
529 | ||
530 | if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0)) | |
531 | return NULL; | |
532 | ||
533 | if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page))) | |
534 | return NULL; | |
535 | ||
536 | if (flags & FOLL_GET) | |
537 | return try_get_folio(page, refs); | |
538 | ||
539 | /* FOLL_PIN is set */ | |
540 | ||
541 | /* | |
542 | * Don't take a pin on the zero page - it's not going anywhere | |
543 | * and it is used in a *lot* of places. | |
544 | */ | |
545 | if (is_zero_page(page)) | |
546 | return page_folio(page); | |
547 | ||
548 | folio = try_get_folio(page, refs); | |
549 | if (!folio) | |
550 | return NULL; | |
551 | ||
552 | /* | |
553 | * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a | |
554 | * right zone, so fail and let the caller fall back to the slow | |
555 | * path. | |
556 | */ | |
557 | if (unlikely((flags & FOLL_LONGTERM) && | |
558 | !folio_is_longterm_pinnable(folio))) { | |
559 | if (!put_devmap_managed_folio_refs(folio, refs)) | |
560 | folio_put_refs(folio, refs); | |
561 | return NULL; | |
562 | } | |
563 | ||
564 | /* | |
565 | * When pinning a large folio, use an exact count to track it. | |
566 | * | |
567 | * However, be sure to *also* increment the normal folio | |
568 | * refcount field at least once, so that the folio really | |
569 | * is pinned. That's why the refcount from the earlier | |
570 | * try_get_folio() is left intact. | |
571 | */ | |
572 | if (folio_test_large(folio)) | |
573 | atomic_add(refs, &folio->_pincount); | |
574 | else | |
575 | folio_ref_add(folio, | |
576 | refs * (GUP_PIN_COUNTING_BIAS - 1)); | |
577 | /* | |
578 | * Adjust the pincount before re-checking the PTE for changes. | |
579 | * This is essentially a smp_mb() and is paired with a memory | |
580 | * barrier in folio_try_share_anon_rmap_*(). | |
581 | */ | |
582 | smp_mb__after_atomic(); | |
583 | ||
584 | node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs); | |
585 | ||
586 | return folio; | |
587 | } | |
8268614b | 588 | #endif /* CONFIG_HAVE_GUP_FAST */ |
a12083d7 | 589 | |
69e68b4f | 590 | static struct page *no_page_table(struct vm_area_struct *vma, |
878b0c45 | 591 | unsigned int flags, unsigned long address) |
4bbd4c77 | 592 | { |
878b0c45 PX |
593 | if (!(flags & FOLL_DUMP)) |
594 | return NULL; | |
595 | ||
69e68b4f | 596 | /* |
878b0c45 | 597 | * When core dumping, we don't want to allocate unnecessary pages or |
69e68b4f KS |
598 | * page tables. Return error instead of NULL to skip handle_mm_fault, |
599 | * then get_dump_page() will return NULL to leave a hole in the dump. | |
600 | * But we can only make this optimization where a hole would surely | |
601 | * be zero-filled if handle_mm_fault() actually did handle it. | |
602 | */ | |
878b0c45 PX |
603 | if (is_vm_hugetlb_page(vma)) { |
604 | struct hstate *h = hstate_vma(vma); | |
605 | ||
606 | if (!hugetlbfs_pagecache_present(h, vma, address)) | |
607 | return ERR_PTR(-EFAULT); | |
608 | } else if ((vma_is_anonymous(vma) || !vma->vm_ops->fault)) { | |
69e68b4f | 609 | return ERR_PTR(-EFAULT); |
878b0c45 PX |
610 | } |
611 | ||
69e68b4f KS |
612 | return NULL; |
613 | } | |
4bbd4c77 | 614 | |
1b167618 PX |
615 | #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES |
616 | static struct page *follow_huge_pud(struct vm_area_struct *vma, | |
617 | unsigned long addr, pud_t *pudp, | |
618 | int flags, struct follow_page_context *ctx) | |
619 | { | |
620 | struct mm_struct *mm = vma->vm_mm; | |
621 | struct page *page; | |
622 | pud_t pud = *pudp; | |
623 | unsigned long pfn = pud_pfn(pud); | |
624 | int ret; | |
625 | ||
626 | assert_spin_locked(pud_lockptr(mm, pudp)); | |
627 | ||
628 | if ((flags & FOLL_WRITE) && !pud_write(pud)) | |
629 | return NULL; | |
630 | ||
631 | if (!pud_present(pud)) | |
632 | return NULL; | |
633 | ||
634 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
635 | ||
636 | if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) && | |
637 | pud_devmap(pud)) { | |
638 | /* | |
639 | * device mapped pages can only be returned if the caller | |
640 | * will manage the page reference count. | |
641 | * | |
642 | * At least one of FOLL_GET | FOLL_PIN must be set, so | |
643 | * assert that here: | |
644 | */ | |
645 | if (!(flags & (FOLL_GET | FOLL_PIN))) | |
646 | return ERR_PTR(-EEXIST); | |
647 | ||
648 | if (flags & FOLL_TOUCH) | |
649 | touch_pud(vma, addr, pudp, flags & FOLL_WRITE); | |
650 | ||
651 | ctx->pgmap = get_dev_pagemap(pfn, ctx->pgmap); | |
652 | if (!ctx->pgmap) | |
653 | return ERR_PTR(-EFAULT); | |
654 | } | |
655 | ||
656 | page = pfn_to_page(pfn); | |
657 | ||
658 | if (!pud_devmap(pud) && !pud_write(pud) && | |
659 | gup_must_unshare(vma, flags, page)) | |
660 | return ERR_PTR(-EMLINK); | |
661 | ||
f442fa61 | 662 | ret = try_grab_folio(page_folio(page), 1, flags); |
1b167618 PX |
663 | if (ret) |
664 | page = ERR_PTR(ret); | |
665 | else | |
666 | ctx->page_mask = HPAGE_PUD_NR - 1; | |
667 | ||
668 | return page; | |
669 | } | |
4418c522 PX |
670 | |
671 | /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ | |
672 | static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, | |
673 | struct vm_area_struct *vma, | |
674 | unsigned int flags) | |
675 | { | |
676 | /* If the pmd is writable, we can write to the page. */ | |
677 | if (pmd_write(pmd)) | |
678 | return true; | |
679 | ||
680 | /* Maybe FOLL_FORCE is set to override it? */ | |
681 | if (!(flags & FOLL_FORCE)) | |
682 | return false; | |
683 | ||
684 | /* But FOLL_FORCE has no effect on shared mappings */ | |
685 | if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) | |
686 | return false; | |
687 | ||
688 | /* ... or read-only private ones */ | |
689 | if (!(vma->vm_flags & VM_MAYWRITE)) | |
690 | return false; | |
691 | ||
692 | /* ... or already writable ones that just need to take a write fault */ | |
693 | if (vma->vm_flags & VM_WRITE) | |
694 | return false; | |
695 | ||
696 | /* | |
697 | * See can_change_pte_writable(): we broke COW and could map the page | |
698 | * writable if we have an exclusive anonymous page ... | |
699 | */ | |
700 | if (!page || !PageAnon(page) || !PageAnonExclusive(page)) | |
701 | return false; | |
702 | ||
703 | /* ... and a write-fault isn't required for other reasons. */ | |
f38ee285 | 704 | if (pmd_needs_soft_dirty_wp(vma, pmd)) |
4418c522 PX |
705 | return false; |
706 | return !userfaultfd_huge_pmd_wp(vma, pmd); | |
707 | } | |
708 | ||
709 | static struct page *follow_huge_pmd(struct vm_area_struct *vma, | |
710 | unsigned long addr, pmd_t *pmd, | |
711 | unsigned int flags, | |
712 | struct follow_page_context *ctx) | |
713 | { | |
714 | struct mm_struct *mm = vma->vm_mm; | |
715 | pmd_t pmdval = *pmd; | |
716 | struct page *page; | |
717 | int ret; | |
718 | ||
719 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
720 | ||
721 | page = pmd_page(pmdval); | |
722 | if ((flags & FOLL_WRITE) && | |
723 | !can_follow_write_pmd(pmdval, page, vma, flags)) | |
724 | return NULL; | |
725 | ||
726 | /* Avoid dumping huge zero page */ | |
727 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(pmdval)) | |
728 | return ERR_PTR(-EFAULT); | |
729 | ||
730 | if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags)) | |
731 | return NULL; | |
732 | ||
733 | if (!pmd_write(pmdval) && gup_must_unshare(vma, flags, page)) | |
734 | return ERR_PTR(-EMLINK); | |
735 | ||
736 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && | |
737 | !PageAnonExclusive(page), page); | |
738 | ||
f442fa61 | 739 | ret = try_grab_folio(page_folio(page), 1, flags); |
4418c522 PX |
740 | if (ret) |
741 | return ERR_PTR(ret); | |
742 | ||
743 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
744 | if (pmd_trans_huge(pmdval) && (flags & FOLL_TOUCH)) | |
745 | touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); | |
746 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
747 | ||
748 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; | |
749 | ctx->page_mask = HPAGE_PMD_NR - 1; | |
750 | ||
751 | return page; | |
752 | } | |
753 | ||
1b167618 PX |
754 | #else /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */ |
755 | static struct page *follow_huge_pud(struct vm_area_struct *vma, | |
756 | unsigned long addr, pud_t *pudp, | |
757 | int flags, struct follow_page_context *ctx) | |
758 | { | |
759 | return NULL; | |
760 | } | |
4418c522 PX |
761 | |
762 | static struct page *follow_huge_pmd(struct vm_area_struct *vma, | |
763 | unsigned long addr, pmd_t *pmd, | |
764 | unsigned int flags, | |
765 | struct follow_page_context *ctx) | |
766 | { | |
767 | return NULL; | |
768 | } | |
1b167618 PX |
769 | #endif /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */ |
770 | ||
1027e443 KS |
771 | static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address, |
772 | pte_t *pte, unsigned int flags) | |
773 | { | |
1027e443 | 774 | if (flags & FOLL_TOUCH) { |
c33c7948 RR |
775 | pte_t orig_entry = ptep_get(pte); |
776 | pte_t entry = orig_entry; | |
1027e443 KS |
777 | |
778 | if (flags & FOLL_WRITE) | |
779 | entry = pte_mkdirty(entry); | |
780 | entry = pte_mkyoung(entry); | |
781 | ||
c33c7948 | 782 | if (!pte_same(orig_entry, entry)) { |
1027e443 KS |
783 | set_pte_at(vma->vm_mm, address, pte, entry); |
784 | update_mmu_cache(vma, address, pte); | |
785 | } | |
786 | } | |
787 | ||
788 | /* Proper page table entry exists, but no corresponding struct page */ | |
789 | return -EEXIST; | |
790 | } | |
791 | ||
5535be30 DH |
792 | /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */ |
793 | static inline bool can_follow_write_pte(pte_t pte, struct page *page, | |
794 | struct vm_area_struct *vma, | |
795 | unsigned int flags) | |
19be0eaf | 796 | { |
5535be30 DH |
797 | /* If the pte is writable, we can write to the page. */ |
798 | if (pte_write(pte)) | |
799 | return true; | |
800 | ||
801 | /* Maybe FOLL_FORCE is set to override it? */ | |
802 | if (!(flags & FOLL_FORCE)) | |
803 | return false; | |
804 | ||
805 | /* But FOLL_FORCE has no effect on shared mappings */ | |
806 | if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) | |
807 | return false; | |
808 | ||
809 | /* ... or read-only private ones */ | |
810 | if (!(vma->vm_flags & VM_MAYWRITE)) | |
811 | return false; | |
812 | ||
813 | /* ... or already writable ones that just need to take a write fault */ | |
814 | if (vma->vm_flags & VM_WRITE) | |
815 | return false; | |
816 | ||
817 | /* | |
818 | * See can_change_pte_writable(): we broke COW and could map the page | |
819 | * writable if we have an exclusive anonymous page ... | |
820 | */ | |
821 | if (!page || !PageAnon(page) || !PageAnonExclusive(page)) | |
822 | return false; | |
823 | ||
824 | /* ... and a write-fault isn't required for other reasons. */ | |
f38ee285 | 825 | if (pte_needs_soft_dirty_wp(vma, pte)) |
5535be30 DH |
826 | return false; |
827 | return !userfaultfd_pte_wp(vma, pte); | |
19be0eaf LT |
828 | } |
829 | ||
69e68b4f | 830 | static struct page *follow_page_pte(struct vm_area_struct *vma, |
df06b37f KB |
831 | unsigned long address, pmd_t *pmd, unsigned int flags, |
832 | struct dev_pagemap **pgmap) | |
69e68b4f KS |
833 | { |
834 | struct mm_struct *mm = vma->vm_mm; | |
835 | struct page *page; | |
836 | spinlock_t *ptl; | |
837 | pte_t *ptep, pte; | |
f28d4363 | 838 | int ret; |
4bbd4c77 | 839 | |
eddb1c22 JH |
840 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
841 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
842 | (FOLL_PIN | FOLL_GET))) | |
843 | return ERR_PTR(-EINVAL); | |
4bbd4c77 KS |
844 | |
845 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
04dee9e8 | 846 | if (!ptep) |
878b0c45 | 847 | return no_page_table(vma, flags, address); |
c33c7948 | 848 | pte = ptep_get(ptep); |
f7355e99 DH |
849 | if (!pte_present(pte)) |
850 | goto no_page; | |
d74943a2 | 851 | if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags)) |
4bbd4c77 | 852 | goto no_page; |
4bbd4c77 KS |
853 | |
854 | page = vm_normal_page(vma, address, pte); | |
5535be30 DH |
855 | |
856 | /* | |
857 | * We only care about anon pages in can_follow_write_pte() and don't | |
858 | * have to worry about pte_devmap() because they are never anon. | |
859 | */ | |
860 | if ((flags & FOLL_WRITE) && | |
861 | !can_follow_write_pte(pte, page, vma, flags)) { | |
862 | page = NULL; | |
863 | goto out; | |
864 | } | |
865 | ||
3faa52c0 | 866 | if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) { |
3565fce3 | 867 | /* |
3faa52c0 JH |
868 | * Only return device mapping pages in the FOLL_GET or FOLL_PIN |
869 | * case since they are only valid while holding the pgmap | |
870 | * reference. | |
3565fce3 | 871 | */ |
df06b37f KB |
872 | *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap); |
873 | if (*pgmap) | |
3565fce3 DW |
874 | page = pte_page(pte); |
875 | else | |
876 | goto no_page; | |
877 | } else if (unlikely(!page)) { | |
1027e443 KS |
878 | if (flags & FOLL_DUMP) { |
879 | /* Avoid special (like zero) pages in core dumps */ | |
880 | page = ERR_PTR(-EFAULT); | |
881 | goto out; | |
882 | } | |
883 | ||
884 | if (is_zero_pfn(pte_pfn(pte))) { | |
885 | page = pte_page(pte); | |
886 | } else { | |
1027e443 KS |
887 | ret = follow_pfn_pte(vma, address, ptep, flags); |
888 | page = ERR_PTR(ret); | |
889 | goto out; | |
890 | } | |
4bbd4c77 KS |
891 | } |
892 | ||
84209e87 | 893 | if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) { |
a7f22660 DH |
894 | page = ERR_PTR(-EMLINK); |
895 | goto out; | |
896 | } | |
b6a2619c DH |
897 | |
898 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && | |
899 | !PageAnonExclusive(page), page); | |
900 | ||
f442fa61 YS |
901 | /* try_grab_folio() does nothing unless FOLL_GET or FOLL_PIN is set. */ |
902 | ret = try_grab_folio(page_folio(page), 1, flags); | |
0f089235 LG |
903 | if (unlikely(ret)) { |
904 | page = ERR_PTR(ret); | |
3faa52c0 | 905 | goto out; |
8fde12ca | 906 | } |
4003f107 | 907 | |
f28d4363 CI |
908 | /* |
909 | * We need to make the page accessible if and only if we are going | |
910 | * to access its content (the FOLL_PIN case). Please see | |
911 | * Documentation/core-api/pin_user_pages.rst for details. | |
912 | */ | |
913 | if (flags & FOLL_PIN) { | |
914 | ret = arch_make_page_accessible(page); | |
915 | if (ret) { | |
916 | unpin_user_page(page); | |
917 | page = ERR_PTR(ret); | |
918 | goto out; | |
919 | } | |
920 | } | |
4bbd4c77 KS |
921 | if (flags & FOLL_TOUCH) { |
922 | if ((flags & FOLL_WRITE) && | |
923 | !pte_dirty(pte) && !PageDirty(page)) | |
924 | set_page_dirty(page); | |
925 | /* | |
926 | * pte_mkyoung() would be more correct here, but atomic care | |
927 | * is needed to avoid losing the dirty bit: it is easier to use | |
928 | * mark_page_accessed(). | |
929 | */ | |
930 | mark_page_accessed(page); | |
931 | } | |
1027e443 | 932 | out: |
4bbd4c77 | 933 | pte_unmap_unlock(ptep, ptl); |
4bbd4c77 | 934 | return page; |
4bbd4c77 KS |
935 | no_page: |
936 | pte_unmap_unlock(ptep, ptl); | |
937 | if (!pte_none(pte)) | |
69e68b4f | 938 | return NULL; |
878b0c45 | 939 | return no_page_table(vma, flags, address); |
69e68b4f KS |
940 | } |
941 | ||
080dbb61 AK |
942 | static struct page *follow_pmd_mask(struct vm_area_struct *vma, |
943 | unsigned long address, pud_t *pudp, | |
df06b37f KB |
944 | unsigned int flags, |
945 | struct follow_page_context *ctx) | |
69e68b4f | 946 | { |
68827280 | 947 | pmd_t *pmd, pmdval; |
69e68b4f KS |
948 | spinlock_t *ptl; |
949 | struct page *page; | |
950 | struct mm_struct *mm = vma->vm_mm; | |
951 | ||
080dbb61 | 952 | pmd = pmd_offset(pudp, address); |
26e1a0c3 | 953 | pmdval = pmdp_get_lockless(pmd); |
68827280 | 954 | if (pmd_none(pmdval)) |
878b0c45 | 955 | return no_page_table(vma, flags, address); |
f7355e99 | 956 | if (!pmd_present(pmdval)) |
878b0c45 | 957 | return no_page_table(vma, flags, address); |
68827280 | 958 | if (pmd_devmap(pmdval)) { |
3565fce3 | 959 | ptl = pmd_lock(mm, pmd); |
df06b37f | 960 | page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap); |
3565fce3 DW |
961 | spin_unlock(ptl); |
962 | if (page) | |
963 | return page; | |
878b0c45 | 964 | return no_page_table(vma, flags, address); |
3565fce3 | 965 | } |
4418c522 | 966 | if (likely(!pmd_leaf(pmdval))) |
df06b37f | 967 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 968 | |
d74943a2 | 969 | if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags)) |
878b0c45 | 970 | return no_page_table(vma, flags, address); |
db08f203 | 971 | |
6742d293 | 972 | ptl = pmd_lock(mm, pmd); |
4418c522 PX |
973 | pmdval = *pmd; |
974 | if (unlikely(!pmd_present(pmdval))) { | |
84c3fc4e | 975 | spin_unlock(ptl); |
878b0c45 | 976 | return no_page_table(vma, flags, address); |
84c3fc4e | 977 | } |
4418c522 | 978 | if (unlikely(!pmd_leaf(pmdval))) { |
6742d293 | 979 | spin_unlock(ptl); |
df06b37f | 980 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 981 | } |
4418c522 | 982 | if (pmd_trans_huge(pmdval) && (flags & FOLL_SPLIT_PMD)) { |
2378118b HD |
983 | spin_unlock(ptl); |
984 | split_huge_pmd(vma, pmd, address); | |
985 | /* If pmd was left empty, stuff a page table in there quickly */ | |
986 | return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) : | |
df06b37f | 987 | follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
69e68b4f | 988 | } |
4418c522 | 989 | page = follow_huge_pmd(vma, address, pmd, flags, ctx); |
6742d293 | 990 | spin_unlock(ptl); |
6742d293 | 991 | return page; |
4bbd4c77 KS |
992 | } |
993 | ||
080dbb61 AK |
994 | static struct page *follow_pud_mask(struct vm_area_struct *vma, |
995 | unsigned long address, p4d_t *p4dp, | |
df06b37f KB |
996 | unsigned int flags, |
997 | struct follow_page_context *ctx) | |
080dbb61 | 998 | { |
caf8cab7 | 999 | pud_t *pudp, pud; |
080dbb61 AK |
1000 | spinlock_t *ptl; |
1001 | struct page *page; | |
1002 | struct mm_struct *mm = vma->vm_mm; | |
1003 | ||
caf8cab7 PX |
1004 | pudp = pud_offset(p4dp, address); |
1005 | pud = READ_ONCE(*pudp); | |
1b167618 | 1006 | if (!pud_present(pud)) |
878b0c45 | 1007 | return no_page_table(vma, flags, address); |
1b167618 | 1008 | if (pud_leaf(pud)) { |
caf8cab7 | 1009 | ptl = pud_lock(mm, pudp); |
1b167618 | 1010 | page = follow_huge_pud(vma, address, pudp, flags, ctx); |
080dbb61 AK |
1011 | spin_unlock(ptl); |
1012 | if (page) | |
1013 | return page; | |
878b0c45 | 1014 | return no_page_table(vma, flags, address); |
080dbb61 | 1015 | } |
caf8cab7 | 1016 | if (unlikely(pud_bad(pud))) |
878b0c45 | 1017 | return no_page_table(vma, flags, address); |
080dbb61 | 1018 | |
caf8cab7 | 1019 | return follow_pmd_mask(vma, address, pudp, flags, ctx); |
080dbb61 AK |
1020 | } |
1021 | ||
080dbb61 AK |
1022 | static struct page *follow_p4d_mask(struct vm_area_struct *vma, |
1023 | unsigned long address, pgd_t *pgdp, | |
df06b37f KB |
1024 | unsigned int flags, |
1025 | struct follow_page_context *ctx) | |
080dbb61 | 1026 | { |
e6fd5564 | 1027 | p4d_t *p4dp, p4d; |
080dbb61 | 1028 | |
e6fd5564 PX |
1029 | p4dp = p4d_offset(pgdp, address); |
1030 | p4d = READ_ONCE(*p4dp); | |
1965e933 | 1031 | BUILD_BUG_ON(p4d_leaf(p4d)); |
a12083d7 | 1032 | |
a12083d7 | 1033 | if (!p4d_present(p4d) || p4d_bad(p4d)) |
878b0c45 | 1034 | return no_page_table(vma, flags, address); |
080dbb61 | 1035 | |
e6fd5564 | 1036 | return follow_pud_mask(vma, address, p4dp, flags, ctx); |
080dbb61 AK |
1037 | } |
1038 | ||
1039 | /** | |
1040 | * follow_page_mask - look up a page descriptor from a user-virtual address | |
1041 | * @vma: vm_area_struct mapping @address | |
1042 | * @address: virtual address to look up | |
1043 | * @flags: flags modifying lookup behaviour | |
78179556 MR |
1044 | * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a |
1045 | * pointer to output page_mask | |
080dbb61 AK |
1046 | * |
1047 | * @flags can have FOLL_ flags set, defined in <linux/mm.h> | |
1048 | * | |
78179556 MR |
1049 | * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches |
1050 | * the device's dev_pagemap metadata to avoid repeating expensive lookups. | |
1051 | * | |
a7f22660 DH |
1052 | * When getting an anonymous page and the caller has to trigger unsharing |
1053 | * of a shared anonymous page first, -EMLINK is returned. The caller should | |
1054 | * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only | |
1055 | * relevant with FOLL_PIN and !FOLL_WRITE. | |
1056 | * | |
78179556 MR |
1057 | * On output, the @ctx->page_mask is set according to the size of the page. |
1058 | * | |
1059 | * Return: the mapped (struct page *), %NULL if no mapping exists, or | |
080dbb61 AK |
1060 | * an error pointer if there is a mapping to something not represented |
1061 | * by a page descriptor (see also vm_normal_page()). | |
1062 | */ | |
a7030aea | 1063 | static struct page *follow_page_mask(struct vm_area_struct *vma, |
080dbb61 | 1064 | unsigned long address, unsigned int flags, |
df06b37f | 1065 | struct follow_page_context *ctx) |
080dbb61 AK |
1066 | { |
1067 | pgd_t *pgd; | |
080dbb61 | 1068 | struct mm_struct *mm = vma->vm_mm; |
9cb28da5 | 1069 | struct page *page; |
080dbb61 | 1070 | |
9cb28da5 | 1071 | vma_pgtable_walk_begin(vma); |
080dbb61 | 1072 | |
9cb28da5 | 1073 | ctx->page_mask = 0; |
080dbb61 AK |
1074 | pgd = pgd_offset(mm, address); |
1075 | ||
8268614b | 1076 | if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) |
a12083d7 PX |
1077 | page = no_page_table(vma, flags, address); |
1078 | else | |
1079 | page = follow_p4d_mask(vma, address, pgd, flags, ctx); | |
080dbb61 | 1080 | |
9cb28da5 PX |
1081 | vma_pgtable_walk_end(vma); |
1082 | ||
a12083d7 | 1083 | return page; |
df06b37f KB |
1084 | } |
1085 | ||
1086 | struct page *follow_page(struct vm_area_struct *vma, unsigned long address, | |
1087 | unsigned int foll_flags) | |
1088 | { | |
1089 | struct follow_page_context ctx = { NULL }; | |
1090 | struct page *page; | |
1091 | ||
1507f512 MR |
1092 | if (vma_is_secretmem(vma)) |
1093 | return NULL; | |
1094 | ||
d64e2dbc | 1095 | if (WARN_ON_ONCE(foll_flags & FOLL_PIN)) |
8909691b DH |
1096 | return NULL; |
1097 | ||
d74943a2 DH |
1098 | /* |
1099 | * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect | |
1100 | * to fail on PROT_NONE-mapped pages. | |
1101 | */ | |
df06b37f KB |
1102 | page = follow_page_mask(vma, address, foll_flags, &ctx); |
1103 | if (ctx.pgmap) | |
1104 | put_dev_pagemap(ctx.pgmap); | |
1105 | return page; | |
080dbb61 AK |
1106 | } |
1107 | ||
f2b495ca KS |
1108 | static int get_gate_page(struct mm_struct *mm, unsigned long address, |
1109 | unsigned int gup_flags, struct vm_area_struct **vma, | |
1110 | struct page **page) | |
1111 | { | |
1112 | pgd_t *pgd; | |
c2febafc | 1113 | p4d_t *p4d; |
f2b495ca KS |
1114 | pud_t *pud; |
1115 | pmd_t *pmd; | |
1116 | pte_t *pte; | |
c33c7948 | 1117 | pte_t entry; |
f2b495ca KS |
1118 | int ret = -EFAULT; |
1119 | ||
1120 | /* user gate pages are read-only */ | |
1121 | if (gup_flags & FOLL_WRITE) | |
1122 | return -EFAULT; | |
1123 | if (address > TASK_SIZE) | |
1124 | pgd = pgd_offset_k(address); | |
1125 | else | |
1126 | pgd = pgd_offset_gate(mm, address); | |
b5d1c39f AL |
1127 | if (pgd_none(*pgd)) |
1128 | return -EFAULT; | |
c2febafc | 1129 | p4d = p4d_offset(pgd, address); |
b5d1c39f AL |
1130 | if (p4d_none(*p4d)) |
1131 | return -EFAULT; | |
c2febafc | 1132 | pud = pud_offset(p4d, address); |
b5d1c39f AL |
1133 | if (pud_none(*pud)) |
1134 | return -EFAULT; | |
f2b495ca | 1135 | pmd = pmd_offset(pud, address); |
84c3fc4e | 1136 | if (!pmd_present(*pmd)) |
f2b495ca | 1137 | return -EFAULT; |
f2b495ca | 1138 | pte = pte_offset_map(pmd, address); |
04dee9e8 HD |
1139 | if (!pte) |
1140 | return -EFAULT; | |
c33c7948 RR |
1141 | entry = ptep_get(pte); |
1142 | if (pte_none(entry)) | |
f2b495ca KS |
1143 | goto unmap; |
1144 | *vma = get_gate_vma(mm); | |
1145 | if (!page) | |
1146 | goto out; | |
c33c7948 | 1147 | *page = vm_normal_page(*vma, address, entry); |
f2b495ca | 1148 | if (!*page) { |
c33c7948 | 1149 | if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry))) |
f2b495ca | 1150 | goto unmap; |
c33c7948 | 1151 | *page = pte_page(entry); |
f2b495ca | 1152 | } |
f442fa61 | 1153 | ret = try_grab_folio(page_folio(*page), 1, gup_flags); |
0f089235 | 1154 | if (unlikely(ret)) |
8fde12ca | 1155 | goto unmap; |
f2b495ca KS |
1156 | out: |
1157 | ret = 0; | |
1158 | unmap: | |
1159 | pte_unmap(pte); | |
1160 | return ret; | |
1161 | } | |
1162 | ||
9a95f3cf | 1163 | /* |
9a863a6a JG |
1164 | * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not |
1165 | * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set | |
1166 | * to 0 and -EBUSY returned. | |
9a95f3cf | 1167 | */ |
64019a2e | 1168 | static int faultin_page(struct vm_area_struct *vma, |
a7f22660 DH |
1169 | unsigned long address, unsigned int *flags, bool unshare, |
1170 | int *locked) | |
16744483 | 1171 | { |
16744483 | 1172 | unsigned int fault_flags = 0; |
2b740303 | 1173 | vm_fault_t ret; |
16744483 | 1174 | |
55b8fe70 AG |
1175 | if (*flags & FOLL_NOFAULT) |
1176 | return -EFAULT; | |
16744483 KS |
1177 | if (*flags & FOLL_WRITE) |
1178 | fault_flags |= FAULT_FLAG_WRITE; | |
1b2ee126 DH |
1179 | if (*flags & FOLL_REMOTE) |
1180 | fault_flags |= FAULT_FLAG_REMOTE; | |
f04740f5 | 1181 | if (*flags & FOLL_UNLOCKABLE) { |
71335f37 | 1182 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; |
93c5c61d PX |
1183 | /* |
1184 | * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set | |
1185 | * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE. | |
1186 | * That's because some callers may not be prepared to | |
1187 | * handle early exits caused by non-fatal signals. | |
1188 | */ | |
1189 | if (*flags & FOLL_INTERRUPTIBLE) | |
1190 | fault_flags |= FAULT_FLAG_INTERRUPTIBLE; | |
1191 | } | |
16744483 KS |
1192 | if (*flags & FOLL_NOWAIT) |
1193 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT; | |
234b239b | 1194 | if (*flags & FOLL_TRIED) { |
4426e945 PX |
1195 | /* |
1196 | * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED | |
1197 | * can co-exist | |
1198 | */ | |
234b239b ALC |
1199 | fault_flags |= FAULT_FLAG_TRIED; |
1200 | } | |
a7f22660 DH |
1201 | if (unshare) { |
1202 | fault_flags |= FAULT_FLAG_UNSHARE; | |
1203 | /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */ | |
1204 | VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE); | |
1205 | } | |
16744483 | 1206 | |
bce617ed | 1207 | ret = handle_mm_fault(vma, address, fault_flags, NULL); |
d9272525 PX |
1208 | |
1209 | if (ret & VM_FAULT_COMPLETED) { | |
1210 | /* | |
1211 | * With FAULT_FLAG_RETRY_NOWAIT we'll never release the | |
1212 | * mmap lock in the page fault handler. Sanity check this. | |
1213 | */ | |
1214 | WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT); | |
9a863a6a JG |
1215 | *locked = 0; |
1216 | ||
d9272525 PX |
1217 | /* |
1218 | * We should do the same as VM_FAULT_RETRY, but let's not | |
1219 | * return -EBUSY since that's not reflecting the reality of | |
1220 | * what has happened - we've just fully completed a page | |
1221 | * fault, with the mmap lock released. Use -EAGAIN to show | |
1222 | * that we want to take the mmap lock _again_. | |
1223 | */ | |
1224 | return -EAGAIN; | |
1225 | } | |
1226 | ||
16744483 | 1227 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
1228 | int err = vm_fault_to_errno(ret, *flags); |
1229 | ||
1230 | if (err) | |
1231 | return err; | |
16744483 KS |
1232 | BUG(); |
1233 | } | |
1234 | ||
16744483 | 1235 | if (ret & VM_FAULT_RETRY) { |
9a863a6a | 1236 | if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 1237 | *locked = 0; |
16744483 KS |
1238 | return -EBUSY; |
1239 | } | |
1240 | ||
16744483 KS |
1241 | return 0; |
1242 | } | |
1243 | ||
8ac26843 LS |
1244 | /* |
1245 | * Writing to file-backed mappings which require folio dirty tracking using GUP | |
1246 | * is a fundamentally broken operation, as kernel write access to GUP mappings | |
1247 | * do not adhere to the semantics expected by a file system. | |
1248 | * | |
1249 | * Consider the following scenario:- | |
1250 | * | |
1251 | * 1. A folio is written to via GUP which write-faults the memory, notifying | |
1252 | * the file system and dirtying the folio. | |
1253 | * 2. Later, writeback is triggered, resulting in the folio being cleaned and | |
1254 | * the PTE being marked read-only. | |
1255 | * 3. The GUP caller writes to the folio, as it is mapped read/write via the | |
1256 | * direct mapping. | |
1257 | * 4. The GUP caller, now done with the page, unpins it and sets it dirty | |
1258 | * (though it does not have to). | |
1259 | * | |
1260 | * This results in both data being written to a folio without writenotify, and | |
1261 | * the folio being dirtied unexpectedly (if the caller decides to do so). | |
1262 | */ | |
1263 | static bool writable_file_mapping_allowed(struct vm_area_struct *vma, | |
1264 | unsigned long gup_flags) | |
1265 | { | |
1266 | /* | |
1267 | * If we aren't pinning then no problematic write can occur. A long term | |
1268 | * pin is the most egregious case so this is the case we disallow. | |
1269 | */ | |
1270 | if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) != | |
1271 | (FOLL_PIN | FOLL_LONGTERM)) | |
1272 | return true; | |
1273 | ||
1274 | /* | |
1275 | * If the VMA does not require dirty tracking then no problematic write | |
1276 | * can occur either. | |
1277 | */ | |
1278 | return !vma_needs_dirty_tracking(vma); | |
1279 | } | |
1280 | ||
fa5bb209 KS |
1281 | static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags) |
1282 | { | |
1283 | vm_flags_t vm_flags = vma->vm_flags; | |
1b2ee126 DH |
1284 | int write = (gup_flags & FOLL_WRITE); |
1285 | int foreign = (gup_flags & FOLL_REMOTE); | |
8ac26843 | 1286 | bool vma_anon = vma_is_anonymous(vma); |
fa5bb209 KS |
1287 | |
1288 | if (vm_flags & (VM_IO | VM_PFNMAP)) | |
1289 | return -EFAULT; | |
1290 | ||
8ac26843 | 1291 | if ((gup_flags & FOLL_ANON) && !vma_anon) |
7f7ccc2c WT |
1292 | return -EFAULT; |
1293 | ||
52650c8b JG |
1294 | if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma)) |
1295 | return -EOPNOTSUPP; | |
1296 | ||
1507f512 MR |
1297 | if (vma_is_secretmem(vma)) |
1298 | return -EFAULT; | |
1299 | ||
1b2ee126 | 1300 | if (write) { |
8ac26843 LS |
1301 | if (!vma_anon && |
1302 | !writable_file_mapping_allowed(vma, gup_flags)) | |
1303 | return -EFAULT; | |
1304 | ||
6beb9958 | 1305 | if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) { |
fa5bb209 KS |
1306 | if (!(gup_flags & FOLL_FORCE)) |
1307 | return -EFAULT; | |
f347454d DH |
1308 | /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */ |
1309 | if (is_vm_hugetlb_page(vma)) | |
1310 | return -EFAULT; | |
fa5bb209 KS |
1311 | /* |
1312 | * We used to let the write,force case do COW in a | |
1313 | * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could | |
1314 | * set a breakpoint in a read-only mapping of an | |
1315 | * executable, without corrupting the file (yet only | |
1316 | * when that file had been opened for writing!). | |
1317 | * Anon pages in shared mappings are surprising: now | |
1318 | * just reject it. | |
1319 | */ | |
46435364 | 1320 | if (!is_cow_mapping(vm_flags)) |
fa5bb209 | 1321 | return -EFAULT; |
fa5bb209 KS |
1322 | } |
1323 | } else if (!(vm_flags & VM_READ)) { | |
1324 | if (!(gup_flags & FOLL_FORCE)) | |
1325 | return -EFAULT; | |
1326 | /* | |
1327 | * Is there actually any vma we can reach here which does not | |
1328 | * have VM_MAYREAD set? | |
1329 | */ | |
1330 | if (!(vm_flags & VM_MAYREAD)) | |
1331 | return -EFAULT; | |
1332 | } | |
d61172b4 DH |
1333 | /* |
1334 | * gups are always data accesses, not instruction | |
1335 | * fetches, so execute=false here | |
1336 | */ | |
1337 | if (!arch_vma_access_permitted(vma, write, false, foreign)) | |
33a709b2 | 1338 | return -EFAULT; |
fa5bb209 KS |
1339 | return 0; |
1340 | } | |
1341 | ||
6cd06ab1 LT |
1342 | /* |
1343 | * This is "vma_lookup()", but with a warning if we would have | |
1344 | * historically expanded the stack in the GUP code. | |
1345 | */ | |
1346 | static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm, | |
1347 | unsigned long addr) | |
1348 | { | |
1349 | #ifdef CONFIG_STACK_GROWSUP | |
1350 | return vma_lookup(mm, addr); | |
1351 | #else | |
1352 | static volatile unsigned long next_warn; | |
1353 | struct vm_area_struct *vma; | |
1354 | unsigned long now, next; | |
1355 | ||
1356 | vma = find_vma(mm, addr); | |
1357 | if (!vma || (addr >= vma->vm_start)) | |
1358 | return vma; | |
1359 | ||
1360 | /* Only warn for half-way relevant accesses */ | |
1361 | if (!(vma->vm_flags & VM_GROWSDOWN)) | |
1362 | return NULL; | |
1363 | if (vma->vm_start - addr > 65536) | |
1364 | return NULL; | |
1365 | ||
1366 | /* Let's not warn more than once an hour.. */ | |
1367 | now = jiffies; next = next_warn; | |
1368 | if (next && time_before(now, next)) | |
1369 | return NULL; | |
1370 | next_warn = now + 60*60*HZ; | |
1371 | ||
1372 | /* Let people know things may have changed. */ | |
1373 | pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n", | |
1374 | current->comm, task_pid_nr(current), | |
1375 | vma->vm_start, vma->vm_end, addr); | |
1376 | dump_stack(); | |
1377 | return NULL; | |
1378 | #endif | |
1379 | } | |
1380 | ||
4bbd4c77 KS |
1381 | /** |
1382 | * __get_user_pages() - pin user pages in memory | |
4bbd4c77 KS |
1383 | * @mm: mm_struct of target mm |
1384 | * @start: starting user address | |
1385 | * @nr_pages: number of pages from start to pin | |
1386 | * @gup_flags: flags modifying pin behaviour | |
1387 | * @pages: array that receives pointers to the pages pinned. | |
1388 | * Should be at least nr_pages long. Or NULL, if caller | |
1389 | * only intends to ensure the pages are faulted in. | |
c1e8d7c6 | 1390 | * @locked: whether we're still with the mmap_lock held |
4bbd4c77 | 1391 | * |
d2dfbe47 LX |
1392 | * Returns either number of pages pinned (which may be less than the |
1393 | * number requested), or an error. Details about the return value: | |
1394 | * | |
1395 | * -- If nr_pages is 0, returns 0. | |
1396 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
1397 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
1398 | * pages pinned. Again, this may be less than nr_pages. | |
2d3a36a4 | 1399 | * -- 0 return value is possible when the fault would need to be retried. |
d2dfbe47 LX |
1400 | * |
1401 | * The caller is responsible for releasing returned @pages, via put_page(). | |
1402 | * | |
c1e8d7c6 | 1403 | * Must be called with mmap_lock held. It may be released. See below. |
4bbd4c77 KS |
1404 | * |
1405 | * __get_user_pages walks a process's page tables and takes a reference to | |
1406 | * each struct page that each user address corresponds to at a given | |
1407 | * instant. That is, it takes the page that would be accessed if a user | |
1408 | * thread accesses the given user virtual address at that instant. | |
1409 | * | |
1410 | * This does not guarantee that the page exists in the user mappings when | |
1411 | * __get_user_pages returns, and there may even be a completely different | |
1412 | * page there in some cases (eg. if mmapped pagecache has been invalidated | |
c5acf1f6 | 1413 | * and subsequently re-faulted). However it does guarantee that the page |
4bbd4c77 KS |
1414 | * won't be freed completely. And mostly callers simply care that the page |
1415 | * contains data that was valid *at some point in time*. Typically, an IO | |
1416 | * or similar operation cannot guarantee anything stronger anyway because | |
1417 | * locks can't be held over the syscall boundary. | |
1418 | * | |
1419 | * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If | |
1420 | * the page is written to, set_page_dirty (or set_page_dirty_lock, as | |
1421 | * appropriate) must be called after the page is finished with, and | |
1422 | * before put_page is called. | |
1423 | * | |
9a863a6a JG |
1424 | * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may |
1425 | * be released. If this happens *@locked will be set to 0 on return. | |
9a95f3cf | 1426 | * |
9a863a6a JG |
1427 | * A caller using such a combination of @gup_flags must therefore hold the |
1428 | * mmap_lock for reading only, and recognize when it's been released. Otherwise, | |
1429 | * it must be held for either reading or writing and will not be released. | |
4bbd4c77 KS |
1430 | * |
1431 | * In most cases, get_user_pages or get_user_pages_fast should be used | |
1432 | * instead of __get_user_pages. __get_user_pages should be used only if | |
1433 | * you need some special @gup_flags. | |
1434 | */ | |
64019a2e | 1435 | static long __get_user_pages(struct mm_struct *mm, |
4bbd4c77 KS |
1436 | unsigned long start, unsigned long nr_pages, |
1437 | unsigned int gup_flags, struct page **pages, | |
b2cac248 | 1438 | int *locked) |
4bbd4c77 | 1439 | { |
df06b37f | 1440 | long ret = 0, i = 0; |
fa5bb209 | 1441 | struct vm_area_struct *vma = NULL; |
df06b37f | 1442 | struct follow_page_context ctx = { NULL }; |
4bbd4c77 KS |
1443 | |
1444 | if (!nr_pages) | |
1445 | return 0; | |
1446 | ||
428e106a | 1447 | start = untagged_addr_remote(mm, start); |
f9652594 | 1448 | |
eddb1c22 | 1449 | VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN))); |
4bbd4c77 | 1450 | |
4bbd4c77 | 1451 | do { |
fa5bb209 KS |
1452 | struct page *page; |
1453 | unsigned int foll_flags = gup_flags; | |
1454 | unsigned int page_increm; | |
1455 | ||
1456 | /* first iteration or cross vma bound */ | |
1457 | if (!vma || start >= vma->vm_end) { | |
631426ba DH |
1458 | /* |
1459 | * MADV_POPULATE_(READ|WRITE) wants to handle VMA | |
1460 | * lookups+error reporting differently. | |
1461 | */ | |
1462 | if (gup_flags & FOLL_MADV_POPULATE) { | |
1463 | vma = vma_lookup(mm, start); | |
1464 | if (!vma) { | |
1465 | ret = -ENOMEM; | |
1466 | goto out; | |
1467 | } | |
1468 | if (check_vma_flags(vma, gup_flags)) { | |
1469 | ret = -EINVAL; | |
1470 | goto out; | |
1471 | } | |
1472 | goto retry; | |
1473 | } | |
6cd06ab1 | 1474 | vma = gup_vma_lookup(mm, start); |
fa5bb209 | 1475 | if (!vma && in_gate_area(mm, start)) { |
fa5bb209 KS |
1476 | ret = get_gate_page(mm, start & PAGE_MASK, |
1477 | gup_flags, &vma, | |
ffe1e786 | 1478 | pages ? &page : NULL); |
fa5bb209 | 1479 | if (ret) |
08be37b7 | 1480 | goto out; |
df06b37f | 1481 | ctx.page_mask = 0; |
fa5bb209 KS |
1482 | goto next_page; |
1483 | } | |
4bbd4c77 | 1484 | |
52650c8b | 1485 | if (!vma) { |
df06b37f KB |
1486 | ret = -EFAULT; |
1487 | goto out; | |
1488 | } | |
52650c8b JG |
1489 | ret = check_vma_flags(vma, gup_flags); |
1490 | if (ret) | |
1491 | goto out; | |
fa5bb209 KS |
1492 | } |
1493 | retry: | |
1494 | /* | |
1495 | * If we have a pending SIGKILL, don't keep faulting pages and | |
1496 | * potentially allocating memory. | |
1497 | */ | |
fa45f116 | 1498 | if (fatal_signal_pending(current)) { |
d180870d | 1499 | ret = -EINTR; |
df06b37f KB |
1500 | goto out; |
1501 | } | |
fa5bb209 | 1502 | cond_resched(); |
df06b37f KB |
1503 | |
1504 | page = follow_page_mask(vma, start, foll_flags, &ctx); | |
a7f22660 DH |
1505 | if (!page || PTR_ERR(page) == -EMLINK) { |
1506 | ret = faultin_page(vma, start, &foll_flags, | |
1507 | PTR_ERR(page) == -EMLINK, locked); | |
fa5bb209 KS |
1508 | switch (ret) { |
1509 | case 0: | |
1510 | goto retry; | |
df06b37f | 1511 | case -EBUSY: |
d9272525 | 1512 | case -EAGAIN: |
df06b37f | 1513 | ret = 0; |
e4a9bc58 | 1514 | fallthrough; |
fa5bb209 KS |
1515 | case -EFAULT: |
1516 | case -ENOMEM: | |
1517 | case -EHWPOISON: | |
df06b37f | 1518 | goto out; |
4bbd4c77 | 1519 | } |
fa5bb209 | 1520 | BUG(); |
1027e443 KS |
1521 | } else if (PTR_ERR(page) == -EEXIST) { |
1522 | /* | |
1523 | * Proper page table entry exists, but no corresponding | |
65462462 JH |
1524 | * struct page. If the caller expects **pages to be |
1525 | * filled in, bail out now, because that can't be done | |
1526 | * for this page. | |
1027e443 | 1527 | */ |
65462462 JH |
1528 | if (pages) { |
1529 | ret = PTR_ERR(page); | |
1530 | goto out; | |
1531 | } | |
1027e443 | 1532 | } else if (IS_ERR(page)) { |
df06b37f KB |
1533 | ret = PTR_ERR(page); |
1534 | goto out; | |
1027e443 | 1535 | } |
ffe1e786 | 1536 | next_page: |
df06b37f | 1537 | page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); |
fa5bb209 KS |
1538 | if (page_increm > nr_pages) |
1539 | page_increm = nr_pages; | |
57edfcfd PX |
1540 | |
1541 | if (pages) { | |
1542 | struct page *subpage; | |
1543 | unsigned int j; | |
1544 | ||
1545 | /* | |
1546 | * This must be a large folio (and doesn't need to | |
1547 | * be the whole folio; it can be part of it), do | |
1548 | * the refcount work for all the subpages too. | |
1549 | * | |
1550 | * NOTE: here the page may not be the head page | |
1551 | * e.g. when start addr is not thp-size aligned. | |
1552 | * try_grab_folio() should have taken care of tail | |
1553 | * pages. | |
1554 | */ | |
1555 | if (page_increm > 1) { | |
f442fa61 | 1556 | struct folio *folio = page_folio(page); |
57edfcfd PX |
1557 | |
1558 | /* | |
1559 | * Since we already hold refcount on the | |
1560 | * large folio, this should never fail. | |
1561 | */ | |
f442fa61 YS |
1562 | if (try_grab_folio(folio, page_increm - 1, |
1563 | foll_flags)) { | |
57edfcfd PX |
1564 | /* |
1565 | * Release the 1st page ref if the | |
1566 | * folio is problematic, fail hard. | |
1567 | */ | |
f442fa61 | 1568 | gup_put_folio(folio, 1, |
57edfcfd PX |
1569 | foll_flags); |
1570 | ret = -EFAULT; | |
1571 | goto out; | |
1572 | } | |
1573 | } | |
1574 | ||
1575 | for (j = 0; j < page_increm; j++) { | |
1576 | subpage = nth_page(page, j); | |
1577 | pages[i + j] = subpage; | |
1578 | flush_anon_page(vma, subpage, start + j * PAGE_SIZE); | |
1579 | flush_dcache_page(subpage); | |
1580 | } | |
1581 | } | |
1582 | ||
fa5bb209 KS |
1583 | i += page_increm; |
1584 | start += page_increm * PAGE_SIZE; | |
1585 | nr_pages -= page_increm; | |
4bbd4c77 | 1586 | } while (nr_pages); |
df06b37f KB |
1587 | out: |
1588 | if (ctx.pgmap) | |
1589 | put_dev_pagemap(ctx.pgmap); | |
1590 | return i ? i : ret; | |
4bbd4c77 | 1591 | } |
4bbd4c77 | 1592 | |
771ab430 TK |
1593 | static bool vma_permits_fault(struct vm_area_struct *vma, |
1594 | unsigned int fault_flags) | |
d4925e00 | 1595 | { |
1b2ee126 DH |
1596 | bool write = !!(fault_flags & FAULT_FLAG_WRITE); |
1597 | bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE); | |
33a709b2 | 1598 | vm_flags_t vm_flags = write ? VM_WRITE : VM_READ; |
d4925e00 DH |
1599 | |
1600 | if (!(vm_flags & vma->vm_flags)) | |
1601 | return false; | |
1602 | ||
33a709b2 DH |
1603 | /* |
1604 | * The architecture might have a hardware protection | |
1b2ee126 | 1605 | * mechanism other than read/write that can deny access. |
d61172b4 DH |
1606 | * |
1607 | * gup always represents data access, not instruction | |
1608 | * fetches, so execute=false here: | |
33a709b2 | 1609 | */ |
d61172b4 | 1610 | if (!arch_vma_access_permitted(vma, write, false, foreign)) |
33a709b2 DH |
1611 | return false; |
1612 | ||
d4925e00 DH |
1613 | return true; |
1614 | } | |
1615 | ||
adc8cb40 | 1616 | /** |
4bbd4c77 | 1617 | * fixup_user_fault() - manually resolve a user page fault |
4bbd4c77 KS |
1618 | * @mm: mm_struct of target mm |
1619 | * @address: user address | |
1620 | * @fault_flags:flags to pass down to handle_mm_fault() | |
c1e8d7c6 | 1621 | * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller |
548b6a1e MC |
1622 | * does not allow retry. If NULL, the caller must guarantee |
1623 | * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY. | |
4bbd4c77 KS |
1624 | * |
1625 | * This is meant to be called in the specific scenario where for locking reasons | |
1626 | * we try to access user memory in atomic context (within a pagefault_disable() | |
1627 | * section), this returns -EFAULT, and we want to resolve the user fault before | |
1628 | * trying again. | |
1629 | * | |
1630 | * Typically this is meant to be used by the futex code. | |
1631 | * | |
1632 | * The main difference with get_user_pages() is that this function will | |
1633 | * unconditionally call handle_mm_fault() which will in turn perform all the | |
1634 | * necessary SW fixup of the dirty and young bits in the PTE, while | |
4a9e1cda | 1635 | * get_user_pages() only guarantees to update these in the struct page. |
4bbd4c77 KS |
1636 | * |
1637 | * This is important for some architectures where those bits also gate the | |
1638 | * access permission to the page because they are maintained in software. On | |
1639 | * such architectures, gup() will not be enough to make a subsequent access | |
1640 | * succeed. | |
1641 | * | |
c1e8d7c6 ML |
1642 | * This function will not return with an unlocked mmap_lock. So it has not the |
1643 | * same semantics wrt the @mm->mmap_lock as does filemap_fault(). | |
4bbd4c77 | 1644 | */ |
64019a2e | 1645 | int fixup_user_fault(struct mm_struct *mm, |
4a9e1cda DD |
1646 | unsigned long address, unsigned int fault_flags, |
1647 | bool *unlocked) | |
4bbd4c77 KS |
1648 | { |
1649 | struct vm_area_struct *vma; | |
8fed2f3c | 1650 | vm_fault_t ret; |
4a9e1cda | 1651 | |
428e106a | 1652 | address = untagged_addr_remote(mm, address); |
f9652594 | 1653 | |
4a9e1cda | 1654 | if (unlocked) |
71335f37 | 1655 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; |
4bbd4c77 | 1656 | |
4a9e1cda | 1657 | retry: |
6cd06ab1 | 1658 | vma = gup_vma_lookup(mm, address); |
8d7071af | 1659 | if (!vma) |
4bbd4c77 KS |
1660 | return -EFAULT; |
1661 | ||
d4925e00 | 1662 | if (!vma_permits_fault(vma, fault_flags)) |
4bbd4c77 KS |
1663 | return -EFAULT; |
1664 | ||
475f4dfc PX |
1665 | if ((fault_flags & FAULT_FLAG_KILLABLE) && |
1666 | fatal_signal_pending(current)) | |
1667 | return -EINTR; | |
1668 | ||
bce617ed | 1669 | ret = handle_mm_fault(vma, address, fault_flags, NULL); |
d9272525 PX |
1670 | |
1671 | if (ret & VM_FAULT_COMPLETED) { | |
1672 | /* | |
1673 | * NOTE: it's a pity that we need to retake the lock here | |
1674 | * to pair with the unlock() in the callers. Ideally we | |
1675 | * could tell the callers so they do not need to unlock. | |
1676 | */ | |
1677 | mmap_read_lock(mm); | |
1678 | *unlocked = true; | |
1679 | return 0; | |
1680 | } | |
1681 | ||
4bbd4c77 | 1682 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
1683 | int err = vm_fault_to_errno(ret, 0); |
1684 | ||
1685 | if (err) | |
1686 | return err; | |
4bbd4c77 KS |
1687 | BUG(); |
1688 | } | |
4a9e1cda DD |
1689 | |
1690 | if (ret & VM_FAULT_RETRY) { | |
d8ed45c5 | 1691 | mmap_read_lock(mm); |
475f4dfc PX |
1692 | *unlocked = true; |
1693 | fault_flags |= FAULT_FLAG_TRIED; | |
1694 | goto retry; | |
4a9e1cda DD |
1695 | } |
1696 | ||
4bbd4c77 KS |
1697 | return 0; |
1698 | } | |
add6a0cd | 1699 | EXPORT_SYMBOL_GPL(fixup_user_fault); |
4bbd4c77 | 1700 | |
93c5c61d PX |
1701 | /* |
1702 | * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is | |
1703 | * specified, it'll also respond to generic signals. The caller of GUP | |
1704 | * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption. | |
1705 | */ | |
1706 | static bool gup_signal_pending(unsigned int flags) | |
1707 | { | |
1708 | if (fatal_signal_pending(current)) | |
1709 | return true; | |
1710 | ||
1711 | if (!(flags & FOLL_INTERRUPTIBLE)) | |
1712 | return false; | |
1713 | ||
1714 | return signal_pending(current); | |
1715 | } | |
1716 | ||
2d3a36a4 | 1717 | /* |
b2a72dff JG |
1718 | * Locking: (*locked == 1) means that the mmap_lock has already been acquired by |
1719 | * the caller. This function may drop the mmap_lock. If it does so, then it will | |
1720 | * set (*locked = 0). | |
1721 | * | |
1722 | * (*locked == 0) means that the caller expects this function to acquire and | |
1723 | * drop the mmap_lock. Therefore, the value of *locked will still be zero when | |
1724 | * the function returns, even though it may have changed temporarily during | |
1725 | * function execution. | |
1726 | * | |
1727 | * Please note that this function, unlike __get_user_pages(), will not return 0 | |
1728 | * for nr_pages > 0, unless FOLL_NOWAIT is used. | |
2d3a36a4 | 1729 | */ |
64019a2e | 1730 | static __always_inline long __get_user_pages_locked(struct mm_struct *mm, |
f0818f47 AA |
1731 | unsigned long start, |
1732 | unsigned long nr_pages, | |
f0818f47 | 1733 | struct page **pages, |
e716712f | 1734 | int *locked, |
0fd71a56 | 1735 | unsigned int flags) |
f0818f47 | 1736 | { |
f0818f47 | 1737 | long ret, pages_done; |
b2a72dff | 1738 | bool must_unlock = false; |
f0818f47 | 1739 | |
9c4b2142 LS |
1740 | if (!nr_pages) |
1741 | return 0; | |
1742 | ||
b2a72dff JG |
1743 | /* |
1744 | * The internal caller expects GUP to manage the lock internally and the | |
1745 | * lock must be released when this returns. | |
1746 | */ | |
9a863a6a | 1747 | if (!*locked) { |
b2a72dff JG |
1748 | if (mmap_read_lock_killable(mm)) |
1749 | return -EAGAIN; | |
1750 | must_unlock = true; | |
1751 | *locked = 1; | |
f0818f47 | 1752 | } |
961ba472 JG |
1753 | else |
1754 | mmap_assert_locked(mm); | |
f0818f47 | 1755 | |
a458b76a AA |
1756 | if (flags & FOLL_PIN) |
1757 | mm_set_has_pinned_flag(&mm->flags); | |
008cfe44 | 1758 | |
eddb1c22 JH |
1759 | /* |
1760 | * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior | |
1761 | * is to set FOLL_GET if the caller wants pages[] filled in (but has | |
1762 | * carelessly failed to specify FOLL_GET), so keep doing that, but only | |
1763 | * for FOLL_GET, not for the newer FOLL_PIN. | |
1764 | * | |
1765 | * FOLL_PIN always expects pages to be non-null, but no need to assert | |
1766 | * that here, as any failures will be obvious enough. | |
1767 | */ | |
1768 | if (pages && !(flags & FOLL_PIN)) | |
f0818f47 | 1769 | flags |= FOLL_GET; |
f0818f47 AA |
1770 | |
1771 | pages_done = 0; | |
f0818f47 | 1772 | for (;;) { |
64019a2e | 1773 | ret = __get_user_pages(mm, start, nr_pages, flags, pages, |
b2cac248 | 1774 | locked); |
f04740f5 | 1775 | if (!(flags & FOLL_UNLOCKABLE)) { |
f0818f47 | 1776 | /* VM_FAULT_RETRY couldn't trigger, bypass */ |
f04740f5 JG |
1777 | pages_done = ret; |
1778 | break; | |
1779 | } | |
f0818f47 | 1780 | |
d9272525 | 1781 | /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */ |
f0818f47 AA |
1782 | if (!*locked) { |
1783 | BUG_ON(ret < 0); | |
1784 | BUG_ON(ret >= nr_pages); | |
1785 | } | |
1786 | ||
f0818f47 AA |
1787 | if (ret > 0) { |
1788 | nr_pages -= ret; | |
1789 | pages_done += ret; | |
1790 | if (!nr_pages) | |
1791 | break; | |
1792 | } | |
1793 | if (*locked) { | |
96312e61 AA |
1794 | /* |
1795 | * VM_FAULT_RETRY didn't trigger or it was a | |
1796 | * FOLL_NOWAIT. | |
1797 | */ | |
f0818f47 AA |
1798 | if (!pages_done) |
1799 | pages_done = ret; | |
1800 | break; | |
1801 | } | |
df17277b MR |
1802 | /* |
1803 | * VM_FAULT_RETRY triggered, so seek to the faulting offset. | |
1804 | * For the prefault case (!pages) we only update counts. | |
1805 | */ | |
1806 | if (likely(pages)) | |
1807 | pages += ret; | |
f0818f47 | 1808 | start += ret << PAGE_SHIFT; |
b2a72dff JG |
1809 | |
1810 | /* The lock was temporarily dropped, so we must unlock later */ | |
1811 | must_unlock = true; | |
f0818f47 | 1812 | |
4426e945 | 1813 | retry: |
f0818f47 AA |
1814 | /* |
1815 | * Repeat on the address that fired VM_FAULT_RETRY | |
4426e945 PX |
1816 | * with both FAULT_FLAG_ALLOW_RETRY and |
1817 | * FAULT_FLAG_TRIED. Note that GUP can be interrupted | |
93c5c61d PX |
1818 | * by fatal signals of even common signals, depending on |
1819 | * the caller's request. So we need to check it before we | |
4426e945 | 1820 | * start trying again otherwise it can loop forever. |
f0818f47 | 1821 | */ |
93c5c61d | 1822 | if (gup_signal_pending(flags)) { |
ae46d2aa HD |
1823 | if (!pages_done) |
1824 | pages_done = -EINTR; | |
4426e945 | 1825 | break; |
ae46d2aa | 1826 | } |
4426e945 | 1827 | |
d8ed45c5 | 1828 | ret = mmap_read_lock_killable(mm); |
71335f37 PX |
1829 | if (ret) { |
1830 | BUG_ON(ret > 0); | |
1831 | if (!pages_done) | |
1832 | pages_done = ret; | |
1833 | break; | |
1834 | } | |
4426e945 | 1835 | |
c7b6a566 | 1836 | *locked = 1; |
64019a2e | 1837 | ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED, |
b2cac248 | 1838 | pages, locked); |
4426e945 PX |
1839 | if (!*locked) { |
1840 | /* Continue to retry until we succeeded */ | |
1841 | BUG_ON(ret != 0); | |
1842 | goto retry; | |
1843 | } | |
f0818f47 AA |
1844 | if (ret != 1) { |
1845 | BUG_ON(ret > 1); | |
1846 | if (!pages_done) | |
1847 | pages_done = ret; | |
1848 | break; | |
1849 | } | |
1850 | nr_pages--; | |
1851 | pages_done++; | |
1852 | if (!nr_pages) | |
1853 | break; | |
df17277b MR |
1854 | if (likely(pages)) |
1855 | pages++; | |
f0818f47 AA |
1856 | start += PAGE_SIZE; |
1857 | } | |
b2a72dff | 1858 | if (must_unlock && *locked) { |
f0818f47 | 1859 | /* |
b2a72dff JG |
1860 | * We either temporarily dropped the lock, or the caller |
1861 | * requested that we both acquire and drop the lock. Either way, | |
1862 | * we must now unlock, and notify the caller of that state. | |
f0818f47 | 1863 | */ |
d8ed45c5 | 1864 | mmap_read_unlock(mm); |
f0818f47 AA |
1865 | *locked = 0; |
1866 | } | |
9c4b2142 LS |
1867 | |
1868 | /* | |
1869 | * Failing to pin anything implies something has gone wrong (except when | |
1870 | * FOLL_NOWAIT is specified). | |
1871 | */ | |
1872 | if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT))) | |
1873 | return -EFAULT; | |
1874 | ||
f0818f47 AA |
1875 | return pages_done; |
1876 | } | |
1877 | ||
d3649f68 CH |
1878 | /** |
1879 | * populate_vma_page_range() - populate a range of pages in the vma. | |
1880 | * @vma: target vma | |
1881 | * @start: start address | |
1882 | * @end: end address | |
c1e8d7c6 | 1883 | * @locked: whether the mmap_lock is still held |
d3649f68 CH |
1884 | * |
1885 | * This takes care of mlocking the pages too if VM_LOCKED is set. | |
1886 | * | |
0a36f7f8 TY |
1887 | * Return either number of pages pinned in the vma, or a negative error |
1888 | * code on error. | |
d3649f68 | 1889 | * |
c1e8d7c6 | 1890 | * vma->vm_mm->mmap_lock must be held. |
d3649f68 | 1891 | * |
4f6da934 | 1892 | * If @locked is NULL, it may be held for read or write and will |
d3649f68 CH |
1893 | * be unperturbed. |
1894 | * | |
4f6da934 PX |
1895 | * If @locked is non-NULL, it must held for read only and may be |
1896 | * released. If it's released, *@locked will be set to 0. | |
d3649f68 CH |
1897 | */ |
1898 | long populate_vma_page_range(struct vm_area_struct *vma, | |
4f6da934 | 1899 | unsigned long start, unsigned long end, int *locked) |
d3649f68 CH |
1900 | { |
1901 | struct mm_struct *mm = vma->vm_mm; | |
1902 | unsigned long nr_pages = (end - start) / PAGE_SIZE; | |
9a863a6a | 1903 | int local_locked = 1; |
d3649f68 | 1904 | int gup_flags; |
ece369c7 | 1905 | long ret; |
d3649f68 | 1906 | |
be51eb18 ML |
1907 | VM_BUG_ON(!PAGE_ALIGNED(start)); |
1908 | VM_BUG_ON(!PAGE_ALIGNED(end)); | |
d3649f68 CH |
1909 | VM_BUG_ON_VMA(start < vma->vm_start, vma); |
1910 | VM_BUG_ON_VMA(end > vma->vm_end, vma); | |
42fc5414 | 1911 | mmap_assert_locked(mm); |
d3649f68 | 1912 | |
b67bf49c HD |
1913 | /* |
1914 | * Rightly or wrongly, the VM_LOCKONFAULT case has never used | |
1915 | * faultin_page() to break COW, so it has no work to do here. | |
1916 | */ | |
d3649f68 | 1917 | if (vma->vm_flags & VM_LOCKONFAULT) |
b67bf49c HD |
1918 | return nr_pages; |
1919 | ||
1096bc93 LT |
1920 | /* ... similarly, we've never faulted in PROT_NONE pages */ |
1921 | if (!vma_is_accessible(vma)) | |
1922 | return -EFAULT; | |
1923 | ||
b67bf49c | 1924 | gup_flags = FOLL_TOUCH; |
d3649f68 CH |
1925 | /* |
1926 | * We want to touch writable mappings with a write fault in order | |
1927 | * to break COW, except for shared mappings because these don't COW | |
1928 | * and we would not want to dirty them for nothing. | |
1096bc93 LT |
1929 | * |
1930 | * Otherwise, do a read fault, and use FOLL_FORCE in case it's not | |
1931 | * readable (ie write-only or executable). | |
d3649f68 CH |
1932 | */ |
1933 | if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE) | |
1934 | gup_flags |= FOLL_WRITE; | |
1096bc93 | 1935 | else |
d3649f68 CH |
1936 | gup_flags |= FOLL_FORCE; |
1937 | ||
f04740f5 JG |
1938 | if (locked) |
1939 | gup_flags |= FOLL_UNLOCKABLE; | |
1940 | ||
d3649f68 CH |
1941 | /* |
1942 | * We made sure addr is within a VMA, so the following will | |
1943 | * not result in a stack expansion that recurses back here. | |
1944 | */ | |
ece369c7 | 1945 | ret = __get_user_pages(mm, start, nr_pages, gup_flags, |
b2cac248 | 1946 | NULL, locked ? locked : &local_locked); |
ece369c7 HD |
1947 | lru_add_drain(); |
1948 | return ret; | |
d3649f68 CH |
1949 | } |
1950 | ||
4ca9b385 | 1951 | /* |
631426ba DH |
1952 | * faultin_page_range() - populate (prefault) page tables inside the |
1953 | * given range readable/writable | |
4ca9b385 DH |
1954 | * |
1955 | * This takes care of mlocking the pages, too, if VM_LOCKED is set. | |
1956 | * | |
631426ba | 1957 | * @mm: the mm to populate page tables in |
4ca9b385 DH |
1958 | * @start: start address |
1959 | * @end: end address | |
1960 | * @write: whether to prefault readable or writable | |
1961 | * @locked: whether the mmap_lock is still held | |
1962 | * | |
631426ba DH |
1963 | * Returns either number of processed pages in the MM, or a negative error |
1964 | * code on error (see __get_user_pages()). Note that this function reports | |
1965 | * errors related to VMAs, such as incompatible mappings, as expected by | |
1966 | * MADV_POPULATE_(READ|WRITE). | |
4ca9b385 | 1967 | * |
631426ba DH |
1968 | * The range must be page-aligned. |
1969 | * | |
1970 | * mm->mmap_lock must be held. If it's released, *@locked will be set to 0. | |
4ca9b385 | 1971 | */ |
631426ba DH |
1972 | long faultin_page_range(struct mm_struct *mm, unsigned long start, |
1973 | unsigned long end, bool write, int *locked) | |
4ca9b385 | 1974 | { |
4ca9b385 DH |
1975 | unsigned long nr_pages = (end - start) / PAGE_SIZE; |
1976 | int gup_flags; | |
ece369c7 | 1977 | long ret; |
4ca9b385 DH |
1978 | |
1979 | VM_BUG_ON(!PAGE_ALIGNED(start)); | |
1980 | VM_BUG_ON(!PAGE_ALIGNED(end)); | |
4ca9b385 DH |
1981 | mmap_assert_locked(mm); |
1982 | ||
1983 | /* | |
1984 | * FOLL_TOUCH: Mark page accessed and thereby young; will also mark | |
1985 | * the page dirty with FOLL_WRITE -- which doesn't make a | |
1986 | * difference with !FOLL_FORCE, because the page is writable | |
1987 | * in the page table. | |
1988 | * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit | |
1989 | * a poisoned page. | |
4ca9b385 DH |
1990 | * !FOLL_FORCE: Require proper access permissions. |
1991 | */ | |
631426ba DH |
1992 | gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE | |
1993 | FOLL_MADV_POPULATE; | |
4ca9b385 DH |
1994 | if (write) |
1995 | gup_flags |= FOLL_WRITE; | |
1996 | ||
631426ba DH |
1997 | ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked, |
1998 | gup_flags); | |
ece369c7 HD |
1999 | lru_add_drain(); |
2000 | return ret; | |
4ca9b385 DH |
2001 | } |
2002 | ||
d3649f68 CH |
2003 | /* |
2004 | * __mm_populate - populate and/or mlock pages within a range of address space. | |
2005 | * | |
2006 | * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap | |
2007 | * flags. VMAs must be already marked with the desired vm_flags, and | |
c1e8d7c6 | 2008 | * mmap_lock must not be held. |
d3649f68 CH |
2009 | */ |
2010 | int __mm_populate(unsigned long start, unsigned long len, int ignore_errors) | |
2011 | { | |
2012 | struct mm_struct *mm = current->mm; | |
2013 | unsigned long end, nstart, nend; | |
2014 | struct vm_area_struct *vma = NULL; | |
2015 | int locked = 0; | |
2016 | long ret = 0; | |
2017 | ||
2018 | end = start + len; | |
2019 | ||
2020 | for (nstart = start; nstart < end; nstart = nend) { | |
2021 | /* | |
2022 | * We want to fault in pages for [nstart; end) address range. | |
2023 | * Find first corresponding VMA. | |
2024 | */ | |
2025 | if (!locked) { | |
2026 | locked = 1; | |
d8ed45c5 | 2027 | mmap_read_lock(mm); |
c4d1a92d | 2028 | vma = find_vma_intersection(mm, nstart, end); |
d3649f68 | 2029 | } else if (nstart >= vma->vm_end) |
c4d1a92d LH |
2030 | vma = find_vma_intersection(mm, vma->vm_end, end); |
2031 | ||
2032 | if (!vma) | |
d3649f68 CH |
2033 | break; |
2034 | /* | |
2035 | * Set [nstart; nend) to intersection of desired address | |
2036 | * range with the first VMA. Also, skip undesirable VMA types. | |
2037 | */ | |
2038 | nend = min(end, vma->vm_end); | |
2039 | if (vma->vm_flags & (VM_IO | VM_PFNMAP)) | |
2040 | continue; | |
2041 | if (nstart < vma->vm_start) | |
2042 | nstart = vma->vm_start; | |
2043 | /* | |
2044 | * Now fault in a range of pages. populate_vma_page_range() | |
2045 | * double checks the vma flags, so that it won't mlock pages | |
2046 | * if the vma was already munlocked. | |
2047 | */ | |
2048 | ret = populate_vma_page_range(vma, nstart, nend, &locked); | |
2049 | if (ret < 0) { | |
2050 | if (ignore_errors) { | |
2051 | ret = 0; | |
2052 | continue; /* continue at next VMA */ | |
2053 | } | |
2054 | break; | |
2055 | } | |
2056 | nend = nstart + ret * PAGE_SIZE; | |
2057 | ret = 0; | |
2058 | } | |
2059 | if (locked) | |
d8ed45c5 | 2060 | mmap_read_unlock(mm); |
d3649f68 CH |
2061 | return ret; /* 0 or negative error code */ |
2062 | } | |
050a9adc | 2063 | #else /* CONFIG_MMU */ |
64019a2e | 2064 | static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start, |
050a9adc | 2065 | unsigned long nr_pages, struct page **pages, |
b2cac248 | 2066 | int *locked, unsigned int foll_flags) |
050a9adc CH |
2067 | { |
2068 | struct vm_area_struct *vma; | |
b2a72dff | 2069 | bool must_unlock = false; |
050a9adc | 2070 | unsigned long vm_flags; |
24dc20c7 | 2071 | long i; |
050a9adc | 2072 | |
b2a72dff JG |
2073 | if (!nr_pages) |
2074 | return 0; | |
2075 | ||
2076 | /* | |
2077 | * The internal caller expects GUP to manage the lock internally and the | |
2078 | * lock must be released when this returns. | |
2079 | */ | |
9a863a6a | 2080 | if (!*locked) { |
b2a72dff JG |
2081 | if (mmap_read_lock_killable(mm)) |
2082 | return -EAGAIN; | |
2083 | must_unlock = true; | |
2084 | *locked = 1; | |
2085 | } | |
2086 | ||
050a9adc CH |
2087 | /* calculate required read or write permissions. |
2088 | * If FOLL_FORCE is set, we only require the "MAY" flags. | |
2089 | */ | |
2090 | vm_flags = (foll_flags & FOLL_WRITE) ? | |
2091 | (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); | |
2092 | vm_flags &= (foll_flags & FOLL_FORCE) ? | |
2093 | (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); | |
2094 | ||
2095 | for (i = 0; i < nr_pages; i++) { | |
2096 | vma = find_vma(mm, start); | |
2097 | if (!vma) | |
b2a72dff | 2098 | break; |
050a9adc CH |
2099 | |
2100 | /* protect what we can, including chardevs */ | |
2101 | if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) || | |
2102 | !(vm_flags & vma->vm_flags)) | |
b2a72dff | 2103 | break; |
050a9adc CH |
2104 | |
2105 | if (pages) { | |
396a400b | 2106 | pages[i] = virt_to_page((void *)start); |
050a9adc CH |
2107 | if (pages[i]) |
2108 | get_page(pages[i]); | |
2109 | } | |
b2cac248 | 2110 | |
050a9adc CH |
2111 | start = (start + PAGE_SIZE) & PAGE_MASK; |
2112 | } | |
2113 | ||
b2a72dff JG |
2114 | if (must_unlock && *locked) { |
2115 | mmap_read_unlock(mm); | |
2116 | *locked = 0; | |
2117 | } | |
050a9adc | 2118 | |
050a9adc CH |
2119 | return i ? : -EFAULT; |
2120 | } | |
2121 | #endif /* !CONFIG_MMU */ | |
d3649f68 | 2122 | |
bb523b40 AG |
2123 | /** |
2124 | * fault_in_writeable - fault in userspace address range for writing | |
2125 | * @uaddr: start of address range | |
2126 | * @size: size of address range | |
2127 | * | |
2128 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2129 | * copy_from_user()). | |
2130 | */ | |
2131 | size_t fault_in_writeable(char __user *uaddr, size_t size) | |
2132 | { | |
2133 | char __user *start = uaddr, *end; | |
2134 | ||
2135 | if (unlikely(size == 0)) | |
2136 | return 0; | |
677b2a8c CL |
2137 | if (!user_write_access_begin(uaddr, size)) |
2138 | return size; | |
bb523b40 | 2139 | if (!PAGE_ALIGNED(uaddr)) { |
677b2a8c | 2140 | unsafe_put_user(0, uaddr, out); |
bb523b40 AG |
2141 | uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr); |
2142 | } | |
2143 | end = (char __user *)PAGE_ALIGN((unsigned long)start + size); | |
2144 | if (unlikely(end < start)) | |
2145 | end = NULL; | |
2146 | while (uaddr != end) { | |
677b2a8c | 2147 | unsafe_put_user(0, uaddr, out); |
bb523b40 AG |
2148 | uaddr += PAGE_SIZE; |
2149 | } | |
2150 | ||
2151 | out: | |
677b2a8c | 2152 | user_write_access_end(); |
bb523b40 AG |
2153 | if (size > uaddr - start) |
2154 | return size - (uaddr - start); | |
2155 | return 0; | |
2156 | } | |
2157 | EXPORT_SYMBOL(fault_in_writeable); | |
2158 | ||
da32b581 CM |
2159 | /** |
2160 | * fault_in_subpage_writeable - fault in an address range for writing | |
2161 | * @uaddr: start of address range | |
2162 | * @size: size of address range | |
2163 | * | |
2164 | * Fault in a user address range for writing while checking for permissions at | |
2165 | * sub-page granularity (e.g. arm64 MTE). This function should be used when | |
2166 | * the caller cannot guarantee forward progress of a copy_to_user() loop. | |
2167 | * | |
2168 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2169 | * copy_from_user()). | |
2170 | */ | |
2171 | size_t fault_in_subpage_writeable(char __user *uaddr, size_t size) | |
2172 | { | |
2173 | size_t faulted_in; | |
2174 | ||
2175 | /* | |
2176 | * Attempt faulting in at page granularity first for page table | |
2177 | * permission checking. The arch-specific probe_subpage_writeable() | |
2178 | * functions may not check for this. | |
2179 | */ | |
2180 | faulted_in = size - fault_in_writeable(uaddr, size); | |
2181 | if (faulted_in) | |
2182 | faulted_in -= probe_subpage_writeable(uaddr, faulted_in); | |
2183 | ||
2184 | return size - faulted_in; | |
2185 | } | |
2186 | EXPORT_SYMBOL(fault_in_subpage_writeable); | |
2187 | ||
cdd591fc AG |
2188 | /* |
2189 | * fault_in_safe_writeable - fault in an address range for writing | |
2190 | * @uaddr: start of address range | |
2191 | * @size: length of address range | |
2192 | * | |
fe673d3f LT |
2193 | * Faults in an address range for writing. This is primarily useful when we |
2194 | * already know that some or all of the pages in the address range aren't in | |
2195 | * memory. | |
cdd591fc | 2196 | * |
fe673d3f | 2197 | * Unlike fault_in_writeable(), this function is non-destructive. |
cdd591fc AG |
2198 | * |
2199 | * Note that we don't pin or otherwise hold the pages referenced that we fault | |
2200 | * in. There's no guarantee that they'll stay in memory for any duration of | |
2201 | * time. | |
2202 | * | |
2203 | * Returns the number of bytes not faulted in, like copy_to_user() and | |
2204 | * copy_from_user(). | |
2205 | */ | |
2206 | size_t fault_in_safe_writeable(const char __user *uaddr, size_t size) | |
2207 | { | |
fe673d3f | 2208 | unsigned long start = (unsigned long)uaddr, end; |
cdd591fc | 2209 | struct mm_struct *mm = current->mm; |
fe673d3f | 2210 | bool unlocked = false; |
cdd591fc | 2211 | |
fe673d3f LT |
2212 | if (unlikely(size == 0)) |
2213 | return 0; | |
cdd591fc | 2214 | end = PAGE_ALIGN(start + size); |
fe673d3f | 2215 | if (end < start) |
cdd591fc | 2216 | end = 0; |
cdd591fc | 2217 | |
fe673d3f LT |
2218 | mmap_read_lock(mm); |
2219 | do { | |
2220 | if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked)) | |
cdd591fc | 2221 | break; |
fe673d3f LT |
2222 | start = (start + PAGE_SIZE) & PAGE_MASK; |
2223 | } while (start != end); | |
2224 | mmap_read_unlock(mm); | |
2225 | ||
2226 | if (size > (unsigned long)uaddr - start) | |
2227 | return size - ((unsigned long)uaddr - start); | |
2228 | return 0; | |
cdd591fc AG |
2229 | } |
2230 | EXPORT_SYMBOL(fault_in_safe_writeable); | |
2231 | ||
bb523b40 AG |
2232 | /** |
2233 | * fault_in_readable - fault in userspace address range for reading | |
2234 | * @uaddr: start of user address range | |
2235 | * @size: size of user address range | |
2236 | * | |
2237 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2238 | * copy_from_user()). | |
2239 | */ | |
2240 | size_t fault_in_readable(const char __user *uaddr, size_t size) | |
2241 | { | |
2242 | const char __user *start = uaddr, *end; | |
2243 | volatile char c; | |
2244 | ||
2245 | if (unlikely(size == 0)) | |
2246 | return 0; | |
677b2a8c CL |
2247 | if (!user_read_access_begin(uaddr, size)) |
2248 | return size; | |
bb523b40 | 2249 | if (!PAGE_ALIGNED(uaddr)) { |
677b2a8c | 2250 | unsafe_get_user(c, uaddr, out); |
bb523b40 AG |
2251 | uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr); |
2252 | } | |
2253 | end = (const char __user *)PAGE_ALIGN((unsigned long)start + size); | |
2254 | if (unlikely(end < start)) | |
2255 | end = NULL; | |
2256 | while (uaddr != end) { | |
677b2a8c | 2257 | unsafe_get_user(c, uaddr, out); |
bb523b40 AG |
2258 | uaddr += PAGE_SIZE; |
2259 | } | |
2260 | ||
2261 | out: | |
677b2a8c | 2262 | user_read_access_end(); |
bb523b40 AG |
2263 | (void)c; |
2264 | if (size > uaddr - start) | |
2265 | return size - (uaddr - start); | |
2266 | return 0; | |
2267 | } | |
2268 | EXPORT_SYMBOL(fault_in_readable); | |
2269 | ||
8f942eea JH |
2270 | /** |
2271 | * get_dump_page() - pin user page in memory while writing it to core dump | |
2272 | * @addr: user address | |
2273 | * | |
2274 | * Returns struct page pointer of user page pinned for dump, | |
2275 | * to be freed afterwards by put_page(). | |
2276 | * | |
2277 | * Returns NULL on any kind of failure - a hole must then be inserted into | |
2278 | * the corefile, to preserve alignment with its headers; and also returns | |
2279 | * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found - | |
f0953a1b | 2280 | * allowing a hole to be left in the corefile to save disk space. |
8f942eea | 2281 | * |
7f3bfab5 | 2282 | * Called without mmap_lock (takes and releases the mmap_lock by itself). |
8f942eea JH |
2283 | */ |
2284 | #ifdef CONFIG_ELF_CORE | |
2285 | struct page *get_dump_page(unsigned long addr) | |
2286 | { | |
8f942eea | 2287 | struct page *page; |
b2a72dff | 2288 | int locked = 0; |
7f3bfab5 | 2289 | int ret; |
8f942eea | 2290 | |
b2cac248 | 2291 | ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked, |
7f3bfab5 | 2292 | FOLL_FORCE | FOLL_DUMP | FOLL_GET); |
7f3bfab5 | 2293 | return (ret == 1) ? page : NULL; |
8f942eea JH |
2294 | } |
2295 | #endif /* CONFIG_ELF_CORE */ | |
2296 | ||
d1e153fe | 2297 | #ifdef CONFIG_MIGRATION |
f68749ec | 2298 | /* |
53ba78de | 2299 | * Returns the number of collected folios. Return value is always >= 0. |
f68749ec | 2300 | */ |
53ba78de VK |
2301 | static unsigned long collect_longterm_unpinnable_folios( |
2302 | struct list_head *movable_folio_list, | |
2303 | unsigned long nr_folios, | |
2304 | struct folio **folios) | |
9a4e9f3b | 2305 | { |
67e139b0 | 2306 | unsigned long i, collected = 0; |
1b7f7e58 | 2307 | struct folio *prev_folio = NULL; |
67e139b0 | 2308 | bool drain_allow = true; |
9a4e9f3b | 2309 | |
53ba78de VK |
2310 | for (i = 0; i < nr_folios; i++) { |
2311 | struct folio *folio = folios[i]; | |
f9f38f78 | 2312 | |
1b7f7e58 | 2313 | if (folio == prev_folio) |
83c02c23 | 2314 | continue; |
1b7f7e58 | 2315 | prev_folio = folio; |
f9f38f78 | 2316 | |
67e139b0 AP |
2317 | if (folio_is_longterm_pinnable(folio)) |
2318 | continue; | |
b05a79d4 | 2319 | |
67e139b0 | 2320 | collected++; |
b05a79d4 | 2321 | |
67e139b0 | 2322 | if (folio_is_device_coherent(folio)) |
f9f38f78 CH |
2323 | continue; |
2324 | ||
1b7f7e58 | 2325 | if (folio_test_hugetlb(folio)) { |
53ba78de | 2326 | isolate_hugetlb(folio, movable_folio_list); |
f9f38f78 CH |
2327 | continue; |
2328 | } | |
9a4e9f3b | 2329 | |
1b7f7e58 | 2330 | if (!folio_test_lru(folio) && drain_allow) { |
f9f38f78 CH |
2331 | lru_add_drain_all(); |
2332 | drain_allow = false; | |
2333 | } | |
2334 | ||
be2d5756 | 2335 | if (!folio_isolate_lru(folio)) |
f9f38f78 | 2336 | continue; |
67e139b0 | 2337 | |
53ba78de | 2338 | list_add_tail(&folio->lru, movable_folio_list); |
1b7f7e58 MWO |
2339 | node_stat_mod_folio(folio, |
2340 | NR_ISOLATED_ANON + folio_is_file_lru(folio), | |
2341 | folio_nr_pages(folio)); | |
9a4e9f3b AK |
2342 | } |
2343 | ||
67e139b0 AP |
2344 | return collected; |
2345 | } | |
2346 | ||
2347 | /* | |
53ba78de VK |
2348 | * Unpins all folios and migrates device coherent folios and movable_folio_list. |
2349 | * Returns -EAGAIN if all folios were successfully migrated or -errno for | |
2350 | * failure (or partial success). | |
67e139b0 | 2351 | */ |
53ba78de VK |
2352 | static int migrate_longterm_unpinnable_folios( |
2353 | struct list_head *movable_folio_list, | |
2354 | unsigned long nr_folios, | |
2355 | struct folio **folios) | |
67e139b0 AP |
2356 | { |
2357 | int ret; | |
2358 | unsigned long i; | |
6e7f34eb | 2359 | |
53ba78de VK |
2360 | for (i = 0; i < nr_folios; i++) { |
2361 | struct folio *folio = folios[i]; | |
67e139b0 AP |
2362 | |
2363 | if (folio_is_device_coherent(folio)) { | |
2364 | /* | |
53ba78de VK |
2365 | * Migration will fail if the folio is pinned, so |
2366 | * convert the pin on the source folio to a normal | |
2367 | * reference. | |
67e139b0 | 2368 | */ |
53ba78de | 2369 | folios[i] = NULL; |
67e139b0 AP |
2370 | folio_get(folio); |
2371 | gup_put_folio(folio, 1, FOLL_PIN); | |
2372 | ||
2373 | if (migrate_device_coherent_page(&folio->page)) { | |
2374 | ret = -EBUSY; | |
2375 | goto err; | |
2376 | } | |
2377 | ||
b05a79d4 | 2378 | continue; |
67e139b0 | 2379 | } |
b05a79d4 | 2380 | |
67e139b0 | 2381 | /* |
53ba78de | 2382 | * We can't migrate folios with unexpected references, so drop |
67e139b0 | 2383 | * the reference obtained by __get_user_pages_locked(). |
53ba78de | 2384 | * Migrating folios have been added to movable_folio_list after |
67e139b0 | 2385 | * calling folio_isolate_lru() which takes a reference so the |
53ba78de | 2386 | * folio won't be freed if it's migrating. |
67e139b0 | 2387 | */ |
53ba78de VK |
2388 | unpin_folio(folios[i]); |
2389 | folios[i] = NULL; | |
f68749ec | 2390 | } |
f9f38f78 | 2391 | |
53ba78de | 2392 | if (!list_empty(movable_folio_list)) { |
f9f38f78 CH |
2393 | struct migration_target_control mtc = { |
2394 | .nid = NUMA_NO_NODE, | |
2395 | .gfp_mask = GFP_USER | __GFP_NOWARN, | |
e42dfe4e | 2396 | .reason = MR_LONGTERM_PIN, |
f9f38f78 CH |
2397 | }; |
2398 | ||
53ba78de | 2399 | if (migrate_pages(movable_folio_list, alloc_migration_target, |
67e139b0 AP |
2400 | NULL, (unsigned long)&mtc, MIGRATE_SYNC, |
2401 | MR_LONGTERM_PIN, NULL)) { | |
f9f38f78 | 2402 | ret = -ENOMEM; |
67e139b0 AP |
2403 | goto err; |
2404 | } | |
9a4e9f3b AK |
2405 | } |
2406 | ||
53ba78de | 2407 | putback_movable_pages(movable_folio_list); |
67e139b0 AP |
2408 | |
2409 | return -EAGAIN; | |
2410 | ||
2411 | err: | |
53ba78de VK |
2412 | unpin_folios(folios, nr_folios); |
2413 | putback_movable_pages(movable_folio_list); | |
24a95998 | 2414 | |
67e139b0 AP |
2415 | return ret; |
2416 | } | |
2417 | ||
2418 | /* | |
53ba78de VK |
2419 | * Check whether all folios are *allowed* to be pinned indefinitely (longterm). |
2420 | * Rather confusingly, all folios in the range are required to be pinned via | |
2421 | * FOLL_PIN, before calling this routine. | |
67e139b0 | 2422 | * |
53ba78de VK |
2423 | * If any folios in the range are not allowed to be pinned, then this routine |
2424 | * will migrate those folios away, unpin all the folios in the range and return | |
67e139b0 AP |
2425 | * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then |
2426 | * call this routine again. | |
2427 | * | |
2428 | * If an error other than -EAGAIN occurs, this indicates a migration failure. | |
2429 | * The caller should give up, and propagate the error back up the call stack. | |
2430 | * | |
53ba78de VK |
2431 | * If everything is OK and all folios in the range are allowed to be pinned, |
2432 | * then this routine leaves all folios pinned and returns zero for success. | |
67e139b0 | 2433 | */ |
53ba78de VK |
2434 | static long check_and_migrate_movable_folios(unsigned long nr_folios, |
2435 | struct folio **folios) | |
67e139b0 AP |
2436 | { |
2437 | unsigned long collected; | |
53ba78de | 2438 | LIST_HEAD(movable_folio_list); |
67e139b0 | 2439 | |
53ba78de VK |
2440 | collected = collect_longterm_unpinnable_folios(&movable_folio_list, |
2441 | nr_folios, folios); | |
67e139b0 AP |
2442 | if (!collected) |
2443 | return 0; | |
2444 | ||
53ba78de VK |
2445 | return migrate_longterm_unpinnable_folios(&movable_folio_list, |
2446 | nr_folios, folios); | |
2447 | } | |
2448 | ||
2449 | /* | |
2450 | * This routine just converts all the pages in the @pages array to folios and | |
2451 | * calls check_and_migrate_movable_folios() to do the heavy lifting. | |
2452 | * | |
2453 | * Please see the check_and_migrate_movable_folios() documentation for details. | |
2454 | */ | |
2455 | static long check_and_migrate_movable_pages(unsigned long nr_pages, | |
2456 | struct page **pages) | |
2457 | { | |
2458 | struct folio **folios; | |
2459 | long i, ret; | |
2460 | ||
2461 | folios = kmalloc_array(nr_pages, sizeof(*folios), GFP_KERNEL); | |
2462 | if (!folios) | |
2463 | return -ENOMEM; | |
2464 | ||
2465 | for (i = 0; i < nr_pages; i++) | |
2466 | folios[i] = page_folio(pages[i]); | |
2467 | ||
2468 | ret = check_and_migrate_movable_folios(nr_pages, folios); | |
2469 | ||
2470 | kfree(folios); | |
2471 | return ret; | |
9a4e9f3b AK |
2472 | } |
2473 | #else | |
f68749ec | 2474 | static long check_and_migrate_movable_pages(unsigned long nr_pages, |
f6d299ec | 2475 | struct page **pages) |
9a4e9f3b | 2476 | { |
24a95998 | 2477 | return 0; |
9a4e9f3b | 2478 | } |
53ba78de VK |
2479 | |
2480 | static long check_and_migrate_movable_folios(unsigned long nr_folios, | |
2481 | struct folio **folios) | |
2482 | { | |
2483 | return 0; | |
2484 | } | |
d1e153fe | 2485 | #endif /* CONFIG_MIGRATION */ |
9a4e9f3b | 2486 | |
2bb6d283 | 2487 | /* |
932f4a63 IW |
2488 | * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which |
2489 | * allows us to process the FOLL_LONGTERM flag. | |
2bb6d283 | 2490 | */ |
64019a2e | 2491 | static long __gup_longterm_locked(struct mm_struct *mm, |
932f4a63 IW |
2492 | unsigned long start, |
2493 | unsigned long nr_pages, | |
2494 | struct page **pages, | |
53b2d09b | 2495 | int *locked, |
932f4a63 | 2496 | unsigned int gup_flags) |
2bb6d283 | 2497 | { |
f68749ec | 2498 | unsigned int flags; |
24a95998 | 2499 | long rc, nr_pinned_pages; |
2bb6d283 | 2500 | |
f68749ec | 2501 | if (!(gup_flags & FOLL_LONGTERM)) |
b2cac248 | 2502 | return __get_user_pages_locked(mm, start, nr_pages, pages, |
53b2d09b | 2503 | locked, gup_flags); |
67e139b0 | 2504 | |
f68749ec PT |
2505 | flags = memalloc_pin_save(); |
2506 | do { | |
24a95998 | 2507 | nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages, |
b2cac248 | 2508 | pages, locked, |
24a95998 AP |
2509 | gup_flags); |
2510 | if (nr_pinned_pages <= 0) { | |
2511 | rc = nr_pinned_pages; | |
f68749ec | 2512 | break; |
24a95998 | 2513 | } |
d64e2dbc JG |
2514 | |
2515 | /* FOLL_LONGTERM implies FOLL_PIN */ | |
f6d299ec | 2516 | rc = check_and_migrate_movable_pages(nr_pinned_pages, pages); |
24a95998 | 2517 | } while (rc == -EAGAIN); |
f68749ec | 2518 | memalloc_pin_restore(flags); |
24a95998 | 2519 | return rc ? rc : nr_pinned_pages; |
2bb6d283 | 2520 | } |
932f4a63 | 2521 | |
d64e2dbc JG |
2522 | /* |
2523 | * Check that the given flags are valid for the exported gup/pup interface, and | |
2524 | * update them with the required flags that the caller must have set. | |
2525 | */ | |
b2cac248 LS |
2526 | static bool is_valid_gup_args(struct page **pages, int *locked, |
2527 | unsigned int *gup_flags_p, unsigned int to_set) | |
447f3e45 | 2528 | { |
d64e2dbc JG |
2529 | unsigned int gup_flags = *gup_flags_p; |
2530 | ||
447f3e45 | 2531 | /* |
d64e2dbc JG |
2532 | * These flags not allowed to be specified externally to the gup |
2533 | * interfaces: | |
0f20bba1 | 2534 | * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only |
d64e2dbc | 2535 | * - FOLL_REMOTE is internal only and used on follow_page() |
f04740f5 | 2536 | * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL |
447f3e45 | 2537 | */ |
0f20bba1 | 2538 | if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS)) |
d64e2dbc JG |
2539 | return false; |
2540 | ||
2541 | gup_flags |= to_set; | |
f04740f5 JG |
2542 | if (locked) { |
2543 | /* At the external interface locked must be set */ | |
2544 | if (WARN_ON_ONCE(*locked != 1)) | |
2545 | return false; | |
2546 | ||
2547 | gup_flags |= FOLL_UNLOCKABLE; | |
2548 | } | |
d64e2dbc JG |
2549 | |
2550 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
2551 | if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) == | |
2552 | (FOLL_PIN | FOLL_GET))) | |
2553 | return false; | |
2554 | ||
2555 | /* LONGTERM can only be specified when pinning */ | |
2556 | if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM))) | |
2557 | return false; | |
2558 | ||
2559 | /* Pages input must be given if using GET/PIN */ | |
2560 | if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages)) | |
447f3e45 | 2561 | return false; |
d64e2dbc | 2562 | |
d64e2dbc JG |
2563 | /* We want to allow the pgmap to be hot-unplugged at all times */ |
2564 | if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) && | |
2565 | (gup_flags & FOLL_PCI_P2PDMA))) | |
2566 | return false; | |
2567 | ||
d64e2dbc | 2568 | *gup_flags_p = gup_flags; |
447f3e45 BS |
2569 | return true; |
2570 | } | |
2571 | ||
22bf29b6 | 2572 | #ifdef CONFIG_MMU |
adc8cb40 | 2573 | /** |
c4237f8b | 2574 | * get_user_pages_remote() - pin user pages in memory |
c4237f8b JH |
2575 | * @mm: mm_struct of target mm |
2576 | * @start: starting user address | |
2577 | * @nr_pages: number of pages from start to pin | |
2578 | * @gup_flags: flags modifying lookup behaviour | |
2579 | * @pages: array that receives pointers to the pages pinned. | |
2580 | * Should be at least nr_pages long. Or NULL, if caller | |
2581 | * only intends to ensure the pages are faulted in. | |
c4237f8b JH |
2582 | * @locked: pointer to lock flag indicating whether lock is held and |
2583 | * subsequently whether VM_FAULT_RETRY functionality can be | |
2584 | * utilised. Lock must initially be held. | |
2585 | * | |
2586 | * Returns either number of pages pinned (which may be less than the | |
2587 | * number requested), or an error. Details about the return value: | |
2588 | * | |
2589 | * -- If nr_pages is 0, returns 0. | |
2590 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
2591 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
2592 | * pages pinned. Again, this may be less than nr_pages. | |
2593 | * | |
2594 | * The caller is responsible for releasing returned @pages, via put_page(). | |
2595 | * | |
c1e8d7c6 | 2596 | * Must be called with mmap_lock held for read or write. |
c4237f8b | 2597 | * |
adc8cb40 SJ |
2598 | * get_user_pages_remote walks a process's page tables and takes a reference |
2599 | * to each struct page that each user address corresponds to at a given | |
c4237f8b JH |
2600 | * instant. That is, it takes the page that would be accessed if a user |
2601 | * thread accesses the given user virtual address at that instant. | |
2602 | * | |
2603 | * This does not guarantee that the page exists in the user mappings when | |
adc8cb40 | 2604 | * get_user_pages_remote returns, and there may even be a completely different |
c4237f8b | 2605 | * page there in some cases (eg. if mmapped pagecache has been invalidated |
5da1a868 | 2606 | * and subsequently re-faulted). However it does guarantee that the page |
c4237f8b JH |
2607 | * won't be freed completely. And mostly callers simply care that the page |
2608 | * contains data that was valid *at some point in time*. Typically, an IO | |
2609 | * or similar operation cannot guarantee anything stronger anyway because | |
2610 | * locks can't be held over the syscall boundary. | |
2611 | * | |
2612 | * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page | |
2613 | * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must | |
2614 | * be called after the page is finished with, and before put_page is called. | |
2615 | * | |
adc8cb40 SJ |
2616 | * get_user_pages_remote is typically used for fewer-copy IO operations, |
2617 | * to get a handle on the memory by some means other than accesses | |
2618 | * via the user virtual addresses. The pages may be submitted for | |
2619 | * DMA to devices or accessed via their kernel linear mapping (via the | |
2620 | * kmap APIs). Care should be taken to use the correct cache flushing APIs. | |
c4237f8b JH |
2621 | * |
2622 | * See also get_user_pages_fast, for performance critical applications. | |
2623 | * | |
adc8cb40 | 2624 | * get_user_pages_remote should be phased out in favor of |
c4237f8b | 2625 | * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing |
adc8cb40 | 2626 | * should use get_user_pages_remote because it cannot pass |
c4237f8b JH |
2627 | * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault. |
2628 | */ | |
64019a2e | 2629 | long get_user_pages_remote(struct mm_struct *mm, |
c4237f8b JH |
2630 | unsigned long start, unsigned long nr_pages, |
2631 | unsigned int gup_flags, struct page **pages, | |
ca5e8632 | 2632 | int *locked) |
c4237f8b | 2633 | { |
9a863a6a JG |
2634 | int local_locked = 1; |
2635 | ||
b2cac248 | 2636 | if (!is_valid_gup_args(pages, locked, &gup_flags, |
d64e2dbc | 2637 | FOLL_TOUCH | FOLL_REMOTE)) |
eddb1c22 JH |
2638 | return -EINVAL; |
2639 | ||
b2cac248 | 2640 | return __get_user_pages_locked(mm, start, nr_pages, pages, |
9a863a6a | 2641 | locked ? locked : &local_locked, |
d64e2dbc | 2642 | gup_flags); |
c4237f8b JH |
2643 | } |
2644 | EXPORT_SYMBOL(get_user_pages_remote); | |
2645 | ||
eddb1c22 | 2646 | #else /* CONFIG_MMU */ |
64019a2e | 2647 | long get_user_pages_remote(struct mm_struct *mm, |
eddb1c22 JH |
2648 | unsigned long start, unsigned long nr_pages, |
2649 | unsigned int gup_flags, struct page **pages, | |
ca5e8632 | 2650 | int *locked) |
eddb1c22 JH |
2651 | { |
2652 | return 0; | |
2653 | } | |
2654 | #endif /* !CONFIG_MMU */ | |
2655 | ||
adc8cb40 SJ |
2656 | /** |
2657 | * get_user_pages() - pin user pages in memory | |
2658 | * @start: starting user address | |
2659 | * @nr_pages: number of pages from start to pin | |
2660 | * @gup_flags: flags modifying lookup behaviour | |
2661 | * @pages: array that receives pointers to the pages pinned. | |
2662 | * Should be at least nr_pages long. Or NULL, if caller | |
2663 | * only intends to ensure the pages are faulted in. | |
adc8cb40 | 2664 | * |
64019a2e PX |
2665 | * This is the same as get_user_pages_remote(), just with a less-flexible |
2666 | * calling convention where we assume that the mm being operated on belongs to | |
2667 | * the current task, and doesn't allow passing of a locked parameter. We also | |
2668 | * obviously don't pass FOLL_REMOTE in here. | |
932f4a63 IW |
2669 | */ |
2670 | long get_user_pages(unsigned long start, unsigned long nr_pages, | |
54d02069 | 2671 | unsigned int gup_flags, struct page **pages) |
932f4a63 | 2672 | { |
9a863a6a JG |
2673 | int locked = 1; |
2674 | ||
b2cac248 | 2675 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH)) |
eddb1c22 JH |
2676 | return -EINVAL; |
2677 | ||
afa3c33e | 2678 | return __get_user_pages_locked(current->mm, start, nr_pages, pages, |
b2cac248 | 2679 | &locked, gup_flags); |
932f4a63 IW |
2680 | } |
2681 | EXPORT_SYMBOL(get_user_pages); | |
2bb6d283 | 2682 | |
acc3c8d1 | 2683 | /* |
d3649f68 | 2684 | * get_user_pages_unlocked() is suitable to replace the form: |
acc3c8d1 | 2685 | * |
3e4e28c5 | 2686 | * mmap_read_lock(mm); |
64019a2e | 2687 | * get_user_pages(mm, ..., pages, NULL); |
3e4e28c5 | 2688 | * mmap_read_unlock(mm); |
d3649f68 CH |
2689 | * |
2690 | * with: | |
2691 | * | |
64019a2e | 2692 | * get_user_pages_unlocked(mm, ..., pages); |
d3649f68 CH |
2693 | * |
2694 | * It is functionally equivalent to get_user_pages_fast so | |
2695 | * get_user_pages_fast should be used instead if specific gup_flags | |
2696 | * (e.g. FOLL_FORCE) are not required. | |
acc3c8d1 | 2697 | */ |
d3649f68 CH |
2698 | long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages, |
2699 | struct page **pages, unsigned int gup_flags) | |
acc3c8d1 | 2700 | { |
b2a72dff | 2701 | int locked = 0; |
acc3c8d1 | 2702 | |
b2cac248 | 2703 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
f04740f5 | 2704 | FOLL_TOUCH | FOLL_UNLOCKABLE)) |
d64e2dbc JG |
2705 | return -EINVAL; |
2706 | ||
afa3c33e | 2707 | return __get_user_pages_locked(current->mm, start, nr_pages, pages, |
b2cac248 | 2708 | &locked, gup_flags); |
4bbd4c77 | 2709 | } |
d3649f68 | 2710 | EXPORT_SYMBOL(get_user_pages_unlocked); |
2667f50e SC |
2711 | |
2712 | /* | |
23babe19 | 2713 | * GUP-fast |
2667f50e SC |
2714 | * |
2715 | * get_user_pages_fast attempts to pin user pages by walking the page | |
2716 | * tables directly and avoids taking locks. Thus the walker needs to be | |
2717 | * protected from page table pages being freed from under it, and should | |
2718 | * block any THP splits. | |
2719 | * | |
2720 | * One way to achieve this is to have the walker disable interrupts, and | |
2721 | * rely on IPIs from the TLB flushing code blocking before the page table | |
2722 | * pages are freed. This is unsuitable for architectures that do not need | |
2723 | * to broadcast an IPI when invalidating TLBs. | |
2724 | * | |
2725 | * Another way to achieve this is to batch up page table containing pages | |
2726 | * belonging to more than one mm_user, then rcu_sched a callback to free those | |
23babe19 | 2727 | * pages. Disabling interrupts will allow the gup_fast() walker to both block |
2667f50e SC |
2728 | * the rcu_sched callback, and an IPI that we broadcast for splitting THPs |
2729 | * (which is a relatively rare event). The code below adopts this strategy. | |
2730 | * | |
2731 | * Before activating this code, please be aware that the following assumptions | |
2732 | * are currently made: | |
2733 | * | |
ff2e6d72 | 2734 | * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to |
e585513b | 2735 | * free pages containing page tables or TLB flushing requires IPI broadcast. |
2667f50e | 2736 | * |
2667f50e SC |
2737 | * *) ptes can be read atomically by the architecture. |
2738 | * | |
2739 | * *) access_ok is sufficient to validate userspace address ranges. | |
2740 | * | |
2741 | * The last two assumptions can be relaxed by the addition of helper functions. | |
2742 | * | |
2743 | * This code is based heavily on the PowerPC implementation by Nick Piggin. | |
2744 | */ | |
25176ad0 | 2745 | #ifdef CONFIG_HAVE_GUP_FAST |
a6e79df9 | 2746 | /* |
f002882c DH |
2747 | * Used in the GUP-fast path to determine whether GUP is permitted to work on |
2748 | * a specific folio. | |
a6e79df9 LS |
2749 | * |
2750 | * This call assumes the caller has pinned the folio, that the lowest page table | |
2751 | * level still points to this folio, and that interrupts have been disabled. | |
2752 | * | |
f002882c DH |
2753 | * GUP-fast must reject all secretmem folios. |
2754 | * | |
a6e79df9 LS |
2755 | * Writing to pinned file-backed dirty tracked folios is inherently problematic |
2756 | * (see comment describing the writable_file_mapping_allowed() function). We | |
2757 | * therefore try to avoid the most egregious case of a long-term mapping doing | |
2758 | * so. | |
2759 | * | |
2760 | * This function cannot be as thorough as that one as the VMA is not available | |
2761 | * in the fast path, so instead we whitelist known good cases and if in doubt, | |
2762 | * fall back to the slow path. | |
2763 | */ | |
f002882c | 2764 | static bool gup_fast_folio_allowed(struct folio *folio, unsigned int flags) |
a6e79df9 | 2765 | { |
f002882c | 2766 | bool reject_file_backed = false; |
a6e79df9 | 2767 | struct address_space *mapping; |
f002882c | 2768 | bool check_secretmem = false; |
a6e79df9 LS |
2769 | unsigned long mapping_flags; |
2770 | ||
2771 | /* | |
2772 | * If we aren't pinning then no problematic write can occur. A long term | |
2773 | * pin is the most egregious case so this is the one we disallow. | |
2774 | */ | |
f002882c | 2775 | if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) == |
a6e79df9 | 2776 | (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) |
f002882c DH |
2777 | reject_file_backed = true; |
2778 | ||
2779 | /* We hold a folio reference, so we can safely access folio fields. */ | |
a6e79df9 | 2780 | |
f002882c DH |
2781 | /* secretmem folios are always order-0 folios. */ |
2782 | if (IS_ENABLED(CONFIG_SECRETMEM) && !folio_test_large(folio)) | |
2783 | check_secretmem = true; | |
2784 | ||
2785 | if (!reject_file_backed && !check_secretmem) | |
2786 | return true; | |
a6e79df9 LS |
2787 | |
2788 | if (WARN_ON_ONCE(folio_test_slab(folio))) | |
2789 | return false; | |
2790 | ||
f002882c | 2791 | /* hugetlb neither requires dirty-tracking nor can be secretmem. */ |
a6e79df9 LS |
2792 | if (folio_test_hugetlb(folio)) |
2793 | return true; | |
2794 | ||
2795 | /* | |
2796 | * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods | |
2797 | * cannot proceed, which means no actions performed under RCU can | |
2798 | * proceed either. | |
2799 | * | |
2800 | * inodes and thus their mappings are freed under RCU, which means the | |
2801 | * mapping cannot be freed beneath us and thus we can safely dereference | |
2802 | * it. | |
2803 | */ | |
2804 | lockdep_assert_irqs_disabled(); | |
2805 | ||
2806 | /* | |
2807 | * However, there may be operations which _alter_ the mapping, so ensure | |
2808 | * we read it once and only once. | |
2809 | */ | |
2810 | mapping = READ_ONCE(folio->mapping); | |
2811 | ||
2812 | /* | |
2813 | * The mapping may have been truncated, in any case we cannot determine | |
2814 | * if this mapping is safe - fall back to slow path to determine how to | |
2815 | * proceed. | |
2816 | */ | |
2817 | if (!mapping) | |
2818 | return false; | |
2819 | ||
2820 | /* Anonymous folios pose no problem. */ | |
2821 | mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS; | |
2822 | if (mapping_flags) | |
2823 | return mapping_flags & PAGE_MAPPING_ANON; | |
2824 | ||
2825 | /* | |
2826 | * At this point, we know the mapping is non-null and points to an | |
f002882c | 2827 | * address_space object. |
a6e79df9 | 2828 | */ |
f002882c DH |
2829 | if (check_secretmem && secretmem_mapping(mapping)) |
2830 | return false; | |
2831 | /* The only remaining allowed file system is shmem. */ | |
2832 | return !reject_file_backed || shmem_mapping(mapping); | |
a6e79df9 LS |
2833 | } |
2834 | ||
23babe19 DH |
2835 | static void __maybe_unused gup_fast_undo_dev_pagemap(int *nr, int nr_start, |
2836 | unsigned int flags, struct page **pages) | |
b59f65fa KS |
2837 | { |
2838 | while ((*nr) - nr_start) { | |
9cbe4954 | 2839 | struct folio *folio = page_folio(pages[--(*nr)]); |
b59f65fa | 2840 | |
9cbe4954 MWO |
2841 | folio_clear_referenced(folio); |
2842 | gup_put_folio(folio, 1, flags); | |
b59f65fa KS |
2843 | } |
2844 | } | |
2845 | ||
3010a5ea | 2846 | #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL |
70cbc3cc | 2847 | /* |
23babe19 | 2848 | * GUP-fast relies on pte change detection to avoid concurrent pgtable |
70cbc3cc YS |
2849 | * operations. |
2850 | * | |
23babe19 | 2851 | * To pin the page, GUP-fast needs to do below in order: |
70cbc3cc YS |
2852 | * (1) pin the page (by prefetching pte), then (2) check pte not changed. |
2853 | * | |
2854 | * For the rest of pgtable operations where pgtable updates can be racy | |
23babe19 | 2855 | * with GUP-fast, we need to do (1) clear pte, then (2) check whether page |
70cbc3cc YS |
2856 | * is pinned. |
2857 | * | |
2858 | * Above will work for all pte-level operations, including THP split. | |
2859 | * | |
23babe19 | 2860 | * For THP collapse, it's a bit more complicated because GUP-fast may be |
70cbc3cc YS |
2861 | * walking a pgtable page that is being freed (pte is still valid but pmd |
2862 | * can be cleared already). To avoid race in such condition, we need to | |
2863 | * also check pmd here to make sure pmd doesn't change (corresponds to | |
2864 | * pmdp_collapse_flush() in the THP collapse code path). | |
2865 | */ | |
23babe19 DH |
2866 | static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, |
2867 | unsigned long end, unsigned int flags, struct page **pages, | |
2868 | int *nr) | |
2667f50e | 2869 | { |
b59f65fa KS |
2870 | struct dev_pagemap *pgmap = NULL; |
2871 | int nr_start = *nr, ret = 0; | |
2667f50e | 2872 | pte_t *ptep, *ptem; |
2667f50e SC |
2873 | |
2874 | ptem = ptep = pte_offset_map(&pmd, addr); | |
04dee9e8 HD |
2875 | if (!ptep) |
2876 | return 0; | |
2667f50e | 2877 | do { |
2a4a06da | 2878 | pte_t pte = ptep_get_lockless(ptep); |
b0496fe4 MWO |
2879 | struct page *page; |
2880 | struct folio *folio; | |
2667f50e | 2881 | |
d74943a2 DH |
2882 | /* |
2883 | * Always fallback to ordinary GUP on PROT_NONE-mapped pages: | |
2884 | * pte_access_permitted() better should reject these pages | |
2885 | * either way: otherwise, GUP-fast might succeed in | |
2886 | * cases where ordinary GUP would fail due to VMA access | |
2887 | * permissions. | |
2888 | */ | |
2889 | if (pte_protnone(pte)) | |
e7884f8e KS |
2890 | goto pte_unmap; |
2891 | ||
b798bec4 | 2892 | if (!pte_access_permitted(pte, flags & FOLL_WRITE)) |
e7884f8e KS |
2893 | goto pte_unmap; |
2894 | ||
b59f65fa | 2895 | if (pte_devmap(pte)) { |
7af75561 IW |
2896 | if (unlikely(flags & FOLL_LONGTERM)) |
2897 | goto pte_unmap; | |
2898 | ||
b59f65fa KS |
2899 | pgmap = get_dev_pagemap(pte_pfn(pte), pgmap); |
2900 | if (unlikely(!pgmap)) { | |
23babe19 | 2901 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
b59f65fa KS |
2902 | goto pte_unmap; |
2903 | } | |
2904 | } else if (pte_special(pte)) | |
2667f50e SC |
2905 | goto pte_unmap; |
2906 | ||
2907 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
2908 | page = pte_page(pte); | |
2909 | ||
f442fa61 | 2910 | folio = try_grab_folio_fast(page, 1, flags); |
b0496fe4 | 2911 | if (!folio) |
2667f50e SC |
2912 | goto pte_unmap; |
2913 | ||
70cbc3cc | 2914 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) || |
c33c7948 | 2915 | unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) { |
b0496fe4 | 2916 | gup_put_folio(folio, 1, flags); |
2667f50e SC |
2917 | goto pte_unmap; |
2918 | } | |
2919 | ||
f002882c | 2920 | if (!gup_fast_folio_allowed(folio, flags)) { |
b0496fe4 | 2921 | gup_put_folio(folio, 1, flags); |
2667f50e SC |
2922 | goto pte_unmap; |
2923 | } | |
2924 | ||
84209e87 | 2925 | if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) { |
a7f22660 DH |
2926 | gup_put_folio(folio, 1, flags); |
2927 | goto pte_unmap; | |
2928 | } | |
2929 | ||
f28d4363 CI |
2930 | /* |
2931 | * We need to make the page accessible if and only if we are | |
2932 | * going to access its content (the FOLL_PIN case). Please | |
2933 | * see Documentation/core-api/pin_user_pages.rst for | |
2934 | * details. | |
2935 | */ | |
2936 | if (flags & FOLL_PIN) { | |
2937 | ret = arch_make_page_accessible(page); | |
2938 | if (ret) { | |
b0496fe4 | 2939 | gup_put_folio(folio, 1, flags); |
f28d4363 CI |
2940 | goto pte_unmap; |
2941 | } | |
2942 | } | |
b0496fe4 | 2943 | folio_set_referenced(folio); |
2667f50e SC |
2944 | pages[*nr] = page; |
2945 | (*nr)++; | |
2667f50e SC |
2946 | } while (ptep++, addr += PAGE_SIZE, addr != end); |
2947 | ||
2948 | ret = 1; | |
2949 | ||
2950 | pte_unmap: | |
832d7aa0 CH |
2951 | if (pgmap) |
2952 | put_dev_pagemap(pgmap); | |
2667f50e SC |
2953 | pte_unmap(ptem); |
2954 | return ret; | |
2955 | } | |
2956 | #else | |
2957 | ||
2958 | /* | |
2959 | * If we can't determine whether or not a pte is special, then fail immediately | |
2960 | * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not | |
2961 | * to be special. | |
2962 | * | |
2963 | * For a futex to be placed on a THP tail page, get_futex_key requires a | |
dadbb612 | 2964 | * get_user_pages_fast_only implementation that can pin pages. Thus it's still |
23babe19 | 2965 | * useful to have gup_fast_pmd_leaf even if we can't operate on ptes. |
2667f50e | 2966 | */ |
23babe19 DH |
2967 | static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, |
2968 | unsigned long end, unsigned int flags, struct page **pages, | |
2969 | int *nr) | |
2667f50e SC |
2970 | { |
2971 | return 0; | |
2972 | } | |
3010a5ea | 2973 | #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */ |
2667f50e | 2974 | |
17596731 | 2975 | #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
23babe19 DH |
2976 | static int gup_fast_devmap_leaf(unsigned long pfn, unsigned long addr, |
2977 | unsigned long end, unsigned int flags, struct page **pages, int *nr) | |
b59f65fa KS |
2978 | { |
2979 | int nr_start = *nr; | |
2980 | struct dev_pagemap *pgmap = NULL; | |
2981 | ||
2982 | do { | |
9cbe4954 | 2983 | struct folio *folio; |
b59f65fa KS |
2984 | struct page *page = pfn_to_page(pfn); |
2985 | ||
2986 | pgmap = get_dev_pagemap(pfn, pgmap); | |
2987 | if (unlikely(!pgmap)) { | |
23babe19 | 2988 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
6401c4eb | 2989 | break; |
b59f65fa | 2990 | } |
4003f107 LG |
2991 | |
2992 | if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) { | |
23babe19 | 2993 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
4003f107 LG |
2994 | break; |
2995 | } | |
2996 | ||
f442fa61 | 2997 | folio = try_grab_folio_fast(page, 1, flags); |
9cbe4954 | 2998 | if (!folio) { |
23babe19 | 2999 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
6401c4eb | 3000 | break; |
3faa52c0 | 3001 | } |
9cbe4954 MWO |
3002 | folio_set_referenced(folio); |
3003 | pages[*nr] = page; | |
b59f65fa KS |
3004 | (*nr)++; |
3005 | pfn++; | |
3006 | } while (addr += PAGE_SIZE, addr != end); | |
832d7aa0 | 3007 | |
6401c4eb | 3008 | put_dev_pagemap(pgmap); |
20b7fee7 | 3009 | return addr == end; |
b59f65fa KS |
3010 | } |
3011 | ||
23babe19 DH |
3012 | static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3013 | unsigned long end, unsigned int flags, struct page **pages, | |
3014 | int *nr) | |
b59f65fa KS |
3015 | { |
3016 | unsigned long fault_pfn; | |
a9b6de77 DW |
3017 | int nr_start = *nr; |
3018 | ||
3019 | fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
23babe19 | 3020 | if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 3021 | return 0; |
b59f65fa | 3022 | |
a9b6de77 | 3023 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { |
23babe19 | 3024 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
3025 | return 0; |
3026 | } | |
3027 | return 1; | |
b59f65fa KS |
3028 | } |
3029 | ||
23babe19 DH |
3030 | static int gup_fast_devmap_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr, |
3031 | unsigned long end, unsigned int flags, struct page **pages, | |
3032 | int *nr) | |
b59f65fa KS |
3033 | { |
3034 | unsigned long fault_pfn; | |
a9b6de77 DW |
3035 | int nr_start = *nr; |
3036 | ||
3037 | fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
23babe19 | 3038 | if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 3039 | return 0; |
b59f65fa | 3040 | |
a9b6de77 | 3041 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { |
23babe19 | 3042 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
3043 | return 0; |
3044 | } | |
3045 | return 1; | |
b59f65fa KS |
3046 | } |
3047 | #else | |
23babe19 DH |
3048 | static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3049 | unsigned long end, unsigned int flags, struct page **pages, | |
3050 | int *nr) | |
b59f65fa KS |
3051 | { |
3052 | BUILD_BUG(); | |
3053 | return 0; | |
3054 | } | |
3055 | ||
23babe19 DH |
3056 | static int gup_fast_devmap_pud_leaf(pud_t pud, pud_t *pudp, unsigned long addr, |
3057 | unsigned long end, unsigned int flags, struct page **pages, | |
3058 | int *nr) | |
b59f65fa KS |
3059 | { |
3060 | BUILD_BUG(); | |
3061 | return 0; | |
3062 | } | |
3063 | #endif | |
3064 | ||
23babe19 DH |
3065 | static int gup_fast_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3066 | unsigned long end, unsigned int flags, struct page **pages, | |
3067 | int *nr) | |
2667f50e | 3068 | { |
667ed1f7 MWO |
3069 | struct page *page; |
3070 | struct folio *folio; | |
2667f50e SC |
3071 | int refs; |
3072 | ||
b798bec4 | 3073 | if (!pmd_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
3074 | return 0; |
3075 | ||
7af75561 IW |
3076 | if (pmd_devmap(orig)) { |
3077 | if (unlikely(flags & FOLL_LONGTERM)) | |
3078 | return 0; | |
23babe19 DH |
3079 | return gup_fast_devmap_pmd_leaf(orig, pmdp, addr, end, flags, |
3080 | pages, nr); | |
7af75561 | 3081 | } |
b59f65fa | 3082 | |
f3c94c62 PX |
3083 | page = pmd_page(orig); |
3084 | refs = record_subpages(page, PMD_SIZE, addr, end, pages + *nr); | |
2667f50e | 3085 | |
f442fa61 | 3086 | folio = try_grab_folio_fast(page, refs, flags); |
667ed1f7 | 3087 | if (!folio) |
2667f50e | 3088 | return 0; |
2667f50e SC |
3089 | |
3090 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { | |
667ed1f7 | 3091 | gup_put_folio(folio, refs, flags); |
2667f50e SC |
3092 | return 0; |
3093 | } | |
3094 | ||
f002882c | 3095 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3096 | gup_put_folio(folio, refs, flags); |
3097 | return 0; | |
3098 | } | |
84209e87 | 3099 | if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
a7f22660 DH |
3100 | gup_put_folio(folio, refs, flags); |
3101 | return 0; | |
3102 | } | |
3103 | ||
a43e9820 | 3104 | *nr += refs; |
667ed1f7 | 3105 | folio_set_referenced(folio); |
2667f50e SC |
3106 | return 1; |
3107 | } | |
3108 | ||
23babe19 DH |
3109 | static int gup_fast_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr, |
3110 | unsigned long end, unsigned int flags, struct page **pages, | |
3111 | int *nr) | |
2667f50e | 3112 | { |
83afb52e MWO |
3113 | struct page *page; |
3114 | struct folio *folio; | |
2667f50e SC |
3115 | int refs; |
3116 | ||
b798bec4 | 3117 | if (!pud_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
3118 | return 0; |
3119 | ||
7af75561 IW |
3120 | if (pud_devmap(orig)) { |
3121 | if (unlikely(flags & FOLL_LONGTERM)) | |
3122 | return 0; | |
23babe19 DH |
3123 | return gup_fast_devmap_pud_leaf(orig, pudp, addr, end, flags, |
3124 | pages, nr); | |
7af75561 | 3125 | } |
b59f65fa | 3126 | |
f3c94c62 PX |
3127 | page = pud_page(orig); |
3128 | refs = record_subpages(page, PUD_SIZE, addr, end, pages + *nr); | |
2667f50e | 3129 | |
f442fa61 | 3130 | folio = try_grab_folio_fast(page, refs, flags); |
83afb52e | 3131 | if (!folio) |
2667f50e | 3132 | return 0; |
2667f50e SC |
3133 | |
3134 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { | |
83afb52e | 3135 | gup_put_folio(folio, refs, flags); |
2667f50e SC |
3136 | return 0; |
3137 | } | |
3138 | ||
f002882c | 3139 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3140 | gup_put_folio(folio, refs, flags); |
3141 | return 0; | |
3142 | } | |
3143 | ||
84209e87 | 3144 | if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
a7f22660 DH |
3145 | gup_put_folio(folio, refs, flags); |
3146 | return 0; | |
3147 | } | |
3148 | ||
a43e9820 | 3149 | *nr += refs; |
83afb52e | 3150 | folio_set_referenced(folio); |
2667f50e SC |
3151 | return 1; |
3152 | } | |
3153 | ||
23babe19 DH |
3154 | static int gup_fast_pgd_leaf(pgd_t orig, pgd_t *pgdp, unsigned long addr, |
3155 | unsigned long end, unsigned int flags, struct page **pages, | |
3156 | int *nr) | |
f30c59e9 AK |
3157 | { |
3158 | int refs; | |
2d7919a2 MWO |
3159 | struct page *page; |
3160 | struct folio *folio; | |
f30c59e9 | 3161 | |
b798bec4 | 3162 | if (!pgd_access_permitted(orig, flags & FOLL_WRITE)) |
f30c59e9 AK |
3163 | return 0; |
3164 | ||
b59f65fa | 3165 | BUILD_BUG_ON(pgd_devmap(orig)); |
a43e9820 | 3166 | |
f3c94c62 PX |
3167 | page = pgd_page(orig); |
3168 | refs = record_subpages(page, PGDIR_SIZE, addr, end, pages + *nr); | |
f30c59e9 | 3169 | |
f442fa61 | 3170 | folio = try_grab_folio_fast(page, refs, flags); |
2d7919a2 | 3171 | if (!folio) |
f30c59e9 | 3172 | return 0; |
f30c59e9 AK |
3173 | |
3174 | if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) { | |
2d7919a2 | 3175 | gup_put_folio(folio, refs, flags); |
f30c59e9 AK |
3176 | return 0; |
3177 | } | |
3178 | ||
31115034 LS |
3179 | if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
3180 | gup_put_folio(folio, refs, flags); | |
3181 | return 0; | |
3182 | } | |
3183 | ||
f002882c | 3184 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3185 | gup_put_folio(folio, refs, flags); |
3186 | return 0; | |
3187 | } | |
3188 | ||
a43e9820 | 3189 | *nr += refs; |
2d7919a2 | 3190 | folio_set_referenced(folio); |
f30c59e9 AK |
3191 | return 1; |
3192 | } | |
3193 | ||
23babe19 DH |
3194 | static int gup_fast_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, |
3195 | unsigned long end, unsigned int flags, struct page **pages, | |
3196 | int *nr) | |
2667f50e SC |
3197 | { |
3198 | unsigned long next; | |
3199 | pmd_t *pmdp; | |
3200 | ||
d3f7b1bb | 3201 | pmdp = pmd_offset_lockless(pudp, pud, addr); |
2667f50e | 3202 | do { |
1180e732 | 3203 | pmd_t pmd = pmdp_get_lockless(pmdp); |
2667f50e SC |
3204 | |
3205 | next = pmd_addr_end(addr, end); | |
84c3fc4e | 3206 | if (!pmd_present(pmd)) |
2667f50e SC |
3207 | return 0; |
3208 | ||
7db86dc3 | 3209 | if (unlikely(pmd_leaf(pmd))) { |
23babe19 | 3210 | /* See gup_fast_pte_range() */ |
d74943a2 | 3211 | if (pmd_protnone(pmd)) |
2667f50e SC |
3212 | return 0; |
3213 | ||
23babe19 | 3214 | if (!gup_fast_pmd_leaf(pmd, pmdp, addr, next, flags, |
2667f50e SC |
3215 | pages, nr)) |
3216 | return 0; | |
3217 | ||
23babe19 DH |
3218 | } else if (!gup_fast_pte_range(pmd, pmdp, addr, next, flags, |
3219 | pages, nr)) | |
2923117b | 3220 | return 0; |
2667f50e SC |
3221 | } while (pmdp++, addr = next, addr != end); |
3222 | ||
3223 | return 1; | |
3224 | } | |
3225 | ||
23babe19 DH |
3226 | static int gup_fast_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, |
3227 | unsigned long end, unsigned int flags, struct page **pages, | |
3228 | int *nr) | |
2667f50e SC |
3229 | { |
3230 | unsigned long next; | |
3231 | pud_t *pudp; | |
3232 | ||
d3f7b1bb | 3233 | pudp = pud_offset_lockless(p4dp, p4d, addr); |
2667f50e | 3234 | do { |
e37c6982 | 3235 | pud_t pud = READ_ONCE(*pudp); |
2667f50e SC |
3236 | |
3237 | next = pud_addr_end(addr, end); | |
15494520 | 3238 | if (unlikely(!pud_present(pud))) |
2667f50e | 3239 | return 0; |
7db86dc3 | 3240 | if (unlikely(pud_leaf(pud))) { |
23babe19 DH |
3241 | if (!gup_fast_pud_leaf(pud, pudp, addr, next, flags, |
3242 | pages, nr)) | |
f30c59e9 | 3243 | return 0; |
23babe19 DH |
3244 | } else if (!gup_fast_pmd_range(pudp, pud, addr, next, flags, |
3245 | pages, nr)) | |
2667f50e SC |
3246 | return 0; |
3247 | } while (pudp++, addr = next, addr != end); | |
3248 | ||
3249 | return 1; | |
3250 | } | |
3251 | ||
23babe19 DH |
3252 | static int gup_fast_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, |
3253 | unsigned long end, unsigned int flags, struct page **pages, | |
3254 | int *nr) | |
c2febafc KS |
3255 | { |
3256 | unsigned long next; | |
3257 | p4d_t *p4dp; | |
3258 | ||
d3f7b1bb | 3259 | p4dp = p4d_offset_lockless(pgdp, pgd, addr); |
c2febafc KS |
3260 | do { |
3261 | p4d_t p4d = READ_ONCE(*p4dp); | |
3262 | ||
3263 | next = p4d_addr_end(addr, end); | |
089f9214 | 3264 | if (!p4d_present(p4d)) |
c2febafc | 3265 | return 0; |
1965e933 | 3266 | BUILD_BUG_ON(p4d_leaf(p4d)); |
8268614b CL |
3267 | if (!gup_fast_pud_range(p4dp, p4d, addr, next, flags, |
3268 | pages, nr)) | |
c2febafc KS |
3269 | return 0; |
3270 | } while (p4dp++, addr = next, addr != end); | |
3271 | ||
3272 | return 1; | |
3273 | } | |
3274 | ||
23babe19 | 3275 | static void gup_fast_pgd_range(unsigned long addr, unsigned long end, |
b798bec4 | 3276 | unsigned int flags, struct page **pages, int *nr) |
5b65c467 KS |
3277 | { |
3278 | unsigned long next; | |
3279 | pgd_t *pgdp; | |
3280 | ||
3281 | pgdp = pgd_offset(current->mm, addr); | |
3282 | do { | |
3283 | pgd_t pgd = READ_ONCE(*pgdp); | |
3284 | ||
3285 | next = pgd_addr_end(addr, end); | |
3286 | if (pgd_none(pgd)) | |
3287 | return; | |
7db86dc3 | 3288 | if (unlikely(pgd_leaf(pgd))) { |
23babe19 DH |
3289 | if (!gup_fast_pgd_leaf(pgd, pgdp, addr, next, flags, |
3290 | pages, nr)) | |
5b65c467 | 3291 | return; |
23babe19 DH |
3292 | } else if (!gup_fast_p4d_range(pgdp, pgd, addr, next, flags, |
3293 | pages, nr)) | |
5b65c467 KS |
3294 | return; |
3295 | } while (pgdp++, addr = next, addr != end); | |
3296 | } | |
050a9adc | 3297 | #else |
23babe19 | 3298 | static inline void gup_fast_pgd_range(unsigned long addr, unsigned long end, |
050a9adc CH |
3299 | unsigned int flags, struct page **pages, int *nr) |
3300 | { | |
3301 | } | |
25176ad0 | 3302 | #endif /* CONFIG_HAVE_GUP_FAST */ |
5b65c467 KS |
3303 | |
3304 | #ifndef gup_fast_permitted | |
3305 | /* | |
dadbb612 | 3306 | * Check if it's allowed to use get_user_pages_fast_only() for the range, or |
5b65c467 KS |
3307 | * we need to fall back to the slow version: |
3308 | */ | |
26f4c328 | 3309 | static bool gup_fast_permitted(unsigned long start, unsigned long end) |
5b65c467 | 3310 | { |
26f4c328 | 3311 | return true; |
5b65c467 KS |
3312 | } |
3313 | #endif | |
3314 | ||
23babe19 DH |
3315 | static unsigned long gup_fast(unsigned long start, unsigned long end, |
3316 | unsigned int gup_flags, struct page **pages) | |
c28b1fc7 JG |
3317 | { |
3318 | unsigned long flags; | |
3319 | int nr_pinned = 0; | |
57efa1fe | 3320 | unsigned seq; |
c28b1fc7 | 3321 | |
25176ad0 | 3322 | if (!IS_ENABLED(CONFIG_HAVE_GUP_FAST) || |
c28b1fc7 JG |
3323 | !gup_fast_permitted(start, end)) |
3324 | return 0; | |
3325 | ||
57efa1fe JG |
3326 | if (gup_flags & FOLL_PIN) { |
3327 | seq = raw_read_seqcount(¤t->mm->write_protect_seq); | |
3328 | if (seq & 1) | |
3329 | return 0; | |
3330 | } | |
3331 | ||
c28b1fc7 JG |
3332 | /* |
3333 | * Disable interrupts. The nested form is used, in order to allow full, | |
3334 | * general purpose use of this routine. | |
3335 | * | |
3336 | * With interrupts disabled, we block page table pages from being freed | |
3337 | * from under us. See struct mmu_table_batch comments in | |
3338 | * include/asm-generic/tlb.h for more details. | |
3339 | * | |
3340 | * We do not adopt an rcu_read_lock() here as we also want to block IPIs | |
3341 | * that come from THPs splitting. | |
3342 | */ | |
3343 | local_irq_save(flags); | |
23babe19 | 3344 | gup_fast_pgd_range(start, end, gup_flags, pages, &nr_pinned); |
c28b1fc7 | 3345 | local_irq_restore(flags); |
57efa1fe JG |
3346 | |
3347 | /* | |
3348 | * When pinning pages for DMA there could be a concurrent write protect | |
23babe19 | 3349 | * from fork() via copy_page_range(), in this case always fail GUP-fast. |
57efa1fe JG |
3350 | */ |
3351 | if (gup_flags & FOLL_PIN) { | |
3352 | if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) { | |
23babe19 | 3353 | gup_fast_unpin_user_pages(pages, nr_pinned); |
57efa1fe | 3354 | return 0; |
b6a2619c DH |
3355 | } else { |
3356 | sanity_check_pinned_pages(pages, nr_pinned); | |
57efa1fe JG |
3357 | } |
3358 | } | |
c28b1fc7 JG |
3359 | return nr_pinned; |
3360 | } | |
3361 | ||
23babe19 DH |
3362 | static int gup_fast_fallback(unsigned long start, unsigned long nr_pages, |
3363 | unsigned int gup_flags, struct page **pages) | |
2667f50e | 3364 | { |
c28b1fc7 JG |
3365 | unsigned long len, end; |
3366 | unsigned long nr_pinned; | |
b2a72dff | 3367 | int locked = 0; |
c28b1fc7 | 3368 | int ret; |
2667f50e | 3369 | |
f4000fdf | 3370 | if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM | |
376a34ef | 3371 | FOLL_FORCE | FOLL_PIN | FOLL_GET | |
4003f107 | 3372 | FOLL_FAST_ONLY | FOLL_NOFAULT | |
d74943a2 | 3373 | FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT))) |
817be129 CH |
3374 | return -EINVAL; |
3375 | ||
a458b76a AA |
3376 | if (gup_flags & FOLL_PIN) |
3377 | mm_set_has_pinned_flag(¤t->mm->flags); | |
008cfe44 | 3378 | |
f81cd178 | 3379 | if (!(gup_flags & FOLL_FAST_ONLY)) |
da1c55f1 | 3380 | might_lock_read(¤t->mm->mmap_lock); |
f81cd178 | 3381 | |
f455c854 | 3382 | start = untagged_addr(start) & PAGE_MASK; |
c28b1fc7 JG |
3383 | len = nr_pages << PAGE_SHIFT; |
3384 | if (check_add_overflow(start, len, &end)) | |
9883c7f8 | 3385 | return -EOVERFLOW; |
6014bc27 LT |
3386 | if (end > TASK_SIZE_MAX) |
3387 | return -EFAULT; | |
96d4f267 | 3388 | if (unlikely(!access_ok((void __user *)start, len))) |
c61611f7 | 3389 | return -EFAULT; |
73e10a61 | 3390 | |
23babe19 | 3391 | nr_pinned = gup_fast(start, end, gup_flags, pages); |
c28b1fc7 JG |
3392 | if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY) |
3393 | return nr_pinned; | |
2667f50e | 3394 | |
c28b1fc7 JG |
3395 | /* Slow path: try to get the remaining pages with get_user_pages */ |
3396 | start += nr_pinned << PAGE_SHIFT; | |
3397 | pages += nr_pinned; | |
b2a72dff | 3398 | ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned, |
b2cac248 | 3399 | pages, &locked, |
f04740f5 | 3400 | gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE); |
c28b1fc7 JG |
3401 | if (ret < 0) { |
3402 | /* | |
3403 | * The caller has to unpin the pages we already pinned so | |
3404 | * returning -errno is not an option | |
3405 | */ | |
3406 | if (nr_pinned) | |
3407 | return nr_pinned; | |
3408 | return ret; | |
2667f50e | 3409 | } |
c28b1fc7 | 3410 | return ret + nr_pinned; |
2667f50e | 3411 | } |
c28b1fc7 | 3412 | |
dadbb612 SJ |
3413 | /** |
3414 | * get_user_pages_fast_only() - pin user pages in memory | |
3415 | * @start: starting user address | |
3416 | * @nr_pages: number of pages from start to pin | |
3417 | * @gup_flags: flags modifying pin behaviour | |
3418 | * @pages: array that receives pointers to the pages pinned. | |
3419 | * Should be at least nr_pages long. | |
3420 | * | |
9e1f0580 JH |
3421 | * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to |
3422 | * the regular GUP. | |
9e1f0580 JH |
3423 | * |
3424 | * If the architecture does not support this function, simply return with no | |
3425 | * pages pinned. | |
3426 | * | |
3427 | * Careful, careful! COW breaking can go either way, so a non-write | |
3428 | * access can get ambiguous page results. If you call this function without | |
3429 | * 'write' set, you'd better be sure that you're ok with that ambiguity. | |
3430 | */ | |
dadbb612 SJ |
3431 | int get_user_pages_fast_only(unsigned long start, int nr_pages, |
3432 | unsigned int gup_flags, struct page **pages) | |
9e1f0580 | 3433 | { |
9e1f0580 JH |
3434 | /* |
3435 | * Internally (within mm/gup.c), gup fast variants must set FOLL_GET, | |
3436 | * because gup fast is always a "pin with a +1 page refcount" request. | |
376a34ef JH |
3437 | * |
3438 | * FOLL_FAST_ONLY is required in order to match the API description of | |
3439 | * this routine: no fall back to regular ("slow") GUP. | |
9e1f0580 | 3440 | */ |
b2cac248 | 3441 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
d64e2dbc JG |
3442 | FOLL_GET | FOLL_FAST_ONLY)) |
3443 | return -EINVAL; | |
9e1f0580 | 3444 | |
23babe19 | 3445 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
9e1f0580 | 3446 | } |
dadbb612 | 3447 | EXPORT_SYMBOL_GPL(get_user_pages_fast_only); |
9e1f0580 | 3448 | |
eddb1c22 JH |
3449 | /** |
3450 | * get_user_pages_fast() - pin user pages in memory | |
3faa52c0 JH |
3451 | * @start: starting user address |
3452 | * @nr_pages: number of pages from start to pin | |
3453 | * @gup_flags: flags modifying pin behaviour | |
3454 | * @pages: array that receives pointers to the pages pinned. | |
3455 | * Should be at least nr_pages long. | |
eddb1c22 | 3456 | * |
c1e8d7c6 | 3457 | * Attempt to pin user pages in memory without taking mm->mmap_lock. |
eddb1c22 JH |
3458 | * If not successful, it will fall back to taking the lock and |
3459 | * calling get_user_pages(). | |
3460 | * | |
3461 | * Returns number of pages pinned. This may be fewer than the number requested. | |
3462 | * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns | |
3463 | * -errno. | |
3464 | */ | |
3465 | int get_user_pages_fast(unsigned long start, int nr_pages, | |
3466 | unsigned int gup_flags, struct page **pages) | |
3467 | { | |
94202f12 JH |
3468 | /* |
3469 | * The caller may or may not have explicitly set FOLL_GET; either way is | |
3470 | * OK. However, internally (within mm/gup.c), gup fast variants must set | |
3471 | * FOLL_GET, because gup fast is always a "pin with a +1 page refcount" | |
3472 | * request. | |
3473 | */ | |
b2cac248 | 3474 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET)) |
d64e2dbc | 3475 | return -EINVAL; |
23babe19 | 3476 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
eddb1c22 | 3477 | } |
050a9adc | 3478 | EXPORT_SYMBOL_GPL(get_user_pages_fast); |
eddb1c22 JH |
3479 | |
3480 | /** | |
3481 | * pin_user_pages_fast() - pin user pages in memory without taking locks | |
3482 | * | |
3faa52c0 JH |
3483 | * @start: starting user address |
3484 | * @nr_pages: number of pages from start to pin | |
3485 | * @gup_flags: flags modifying pin behaviour | |
3486 | * @pages: array that receives pointers to the pages pinned. | |
3487 | * Should be at least nr_pages long. | |
3488 | * | |
3489 | * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See | |
3490 | * get_user_pages_fast() for documentation on the function arguments, because | |
3491 | * the arguments here are identical. | |
3492 | * | |
3493 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3494 | * see Documentation/core-api/pin_user_pages.rst for further details. |
c8070b78 DH |
3495 | * |
3496 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3497 | * pins in it and unpin_user_page() will not remove pins from it. | |
eddb1c22 JH |
3498 | */ |
3499 | int pin_user_pages_fast(unsigned long start, int nr_pages, | |
3500 | unsigned int gup_flags, struct page **pages) | |
3501 | { | |
b2cac248 | 3502 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN)) |
3faa52c0 | 3503 | return -EINVAL; |
23babe19 | 3504 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
eddb1c22 JH |
3505 | } |
3506 | EXPORT_SYMBOL_GPL(pin_user_pages_fast); | |
3507 | ||
3508 | /** | |
64019a2e | 3509 | * pin_user_pages_remote() - pin pages of a remote process |
eddb1c22 | 3510 | * |
3faa52c0 JH |
3511 | * @mm: mm_struct of target mm |
3512 | * @start: starting user address | |
3513 | * @nr_pages: number of pages from start to pin | |
3514 | * @gup_flags: flags modifying lookup behaviour | |
3515 | * @pages: array that receives pointers to the pages pinned. | |
0768c8de | 3516 | * Should be at least nr_pages long. |
3faa52c0 JH |
3517 | * @locked: pointer to lock flag indicating whether lock is held and |
3518 | * subsequently whether VM_FAULT_RETRY functionality can be | |
3519 | * utilised. Lock must initially be held. | |
3520 | * | |
3521 | * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See | |
3522 | * get_user_pages_remote() for documentation on the function arguments, because | |
3523 | * the arguments here are identical. | |
3524 | * | |
3525 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3526 | * see Documentation/core-api/pin_user_pages.rst for details. |
c8070b78 DH |
3527 | * |
3528 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3529 | * pins in it and unpin_user_page*() will not remove pins from it. | |
eddb1c22 | 3530 | */ |
64019a2e | 3531 | long pin_user_pages_remote(struct mm_struct *mm, |
eddb1c22 JH |
3532 | unsigned long start, unsigned long nr_pages, |
3533 | unsigned int gup_flags, struct page **pages, | |
0b295316 | 3534 | int *locked) |
eddb1c22 | 3535 | { |
9a863a6a JG |
3536 | int local_locked = 1; |
3537 | ||
b2cac248 | 3538 | if (!is_valid_gup_args(pages, locked, &gup_flags, |
d64e2dbc JG |
3539 | FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE)) |
3540 | return 0; | |
b2cac248 | 3541 | return __gup_longterm_locked(mm, start, nr_pages, pages, |
9a863a6a | 3542 | locked ? locked : &local_locked, |
d64e2dbc | 3543 | gup_flags); |
eddb1c22 JH |
3544 | } |
3545 | EXPORT_SYMBOL(pin_user_pages_remote); | |
3546 | ||
3547 | /** | |
3548 | * pin_user_pages() - pin user pages in memory for use by other devices | |
3549 | * | |
3faa52c0 JH |
3550 | * @start: starting user address |
3551 | * @nr_pages: number of pages from start to pin | |
3552 | * @gup_flags: flags modifying lookup behaviour | |
3553 | * @pages: array that receives pointers to the pages pinned. | |
0768c8de | 3554 | * Should be at least nr_pages long. |
3faa52c0 JH |
3555 | * |
3556 | * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and | |
3557 | * FOLL_PIN is set. | |
3558 | * | |
3559 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3560 | * see Documentation/core-api/pin_user_pages.rst for details. |
c8070b78 DH |
3561 | * |
3562 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3563 | * pins in it and unpin_user_page*() will not remove pins from it. | |
eddb1c22 JH |
3564 | */ |
3565 | long pin_user_pages(unsigned long start, unsigned long nr_pages, | |
4c630f30 | 3566 | unsigned int gup_flags, struct page **pages) |
eddb1c22 | 3567 | { |
9a863a6a JG |
3568 | int locked = 1; |
3569 | ||
b2cac248 | 3570 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN)) |
d64e2dbc | 3571 | return 0; |
64019a2e | 3572 | return __gup_longterm_locked(current->mm, start, nr_pages, |
b2cac248 | 3573 | pages, &locked, gup_flags); |
eddb1c22 JH |
3574 | } |
3575 | EXPORT_SYMBOL(pin_user_pages); | |
91429023 JH |
3576 | |
3577 | /* | |
3578 | * pin_user_pages_unlocked() is the FOLL_PIN variant of | |
3579 | * get_user_pages_unlocked(). Behavior is the same, except that this one sets | |
3580 | * FOLL_PIN and rejects FOLL_GET. | |
c8070b78 DH |
3581 | * |
3582 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3583 | * pins in it and unpin_user_page*() will not remove pins from it. | |
91429023 JH |
3584 | */ |
3585 | long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages, | |
3586 | struct page **pages, unsigned int gup_flags) | |
3587 | { | |
b2a72dff | 3588 | int locked = 0; |
91429023 | 3589 | |
b2cac248 | 3590 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
f04740f5 | 3591 | FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE)) |
d64e2dbc | 3592 | return 0; |
0768c8de | 3593 | |
b2cac248 | 3594 | return __gup_longterm_locked(current->mm, start, nr_pages, pages, |
b2a72dff | 3595 | &locked, gup_flags); |
91429023 JH |
3596 | } |
3597 | EXPORT_SYMBOL(pin_user_pages_unlocked); | |
89c1905d VK |
3598 | |
3599 | /** | |
3600 | * memfd_pin_folios() - pin folios associated with a memfd | |
3601 | * @memfd: the memfd whose folios are to be pinned | |
3602 | * @start: the first memfd offset | |
3603 | * @end: the last memfd offset (inclusive) | |
3604 | * @folios: array that receives pointers to the folios pinned | |
3605 | * @max_folios: maximum number of entries in @folios | |
3606 | * @offset: the offset into the first folio | |
3607 | * | |
3608 | * Attempt to pin folios associated with a memfd in the contiguous range | |
3609 | * [start, end]. Given that a memfd is either backed by shmem or hugetlb, | |
3610 | * the folios can either be found in the page cache or need to be allocated | |
3611 | * if necessary. Once the folios are located, they are all pinned via | |
3612 | * FOLL_PIN and @offset is populatedwith the offset into the first folio. | |
3613 | * And, eventually, these pinned folios must be released either using | |
3614 | * unpin_folios() or unpin_folio(). | |
3615 | * | |
3616 | * It must be noted that the folios may be pinned for an indefinite amount | |
3617 | * of time. And, in most cases, the duration of time they may stay pinned | |
3618 | * would be controlled by the userspace. This behavior is effectively the | |
3619 | * same as using FOLL_LONGTERM with other GUP APIs. | |
3620 | * | |
3621 | * Returns number of folios pinned, which could be less than @max_folios | |
3622 | * as it depends on the folio sizes that cover the range [start, end]. | |
3623 | * If no folios were pinned, it returns -errno. | |
3624 | */ | |
3625 | long memfd_pin_folios(struct file *memfd, loff_t start, loff_t end, | |
3626 | struct folio **folios, unsigned int max_folios, | |
3627 | pgoff_t *offset) | |
3628 | { | |
3629 | unsigned int flags, nr_folios, nr_found; | |
3630 | unsigned int i, pgshift = PAGE_SHIFT; | |
3631 | pgoff_t start_idx, end_idx, next_idx; | |
3632 | struct folio *folio = NULL; | |
3633 | struct folio_batch fbatch; | |
3634 | struct hstate *h; | |
3635 | long ret = -EINVAL; | |
3636 | ||
3637 | if (start < 0 || start > end || !max_folios) | |
3638 | return -EINVAL; | |
3639 | ||
3640 | if (!memfd) | |
3641 | return -EINVAL; | |
3642 | ||
3643 | if (!shmem_file(memfd) && !is_file_hugepages(memfd)) | |
3644 | return -EINVAL; | |
3645 | ||
3646 | if (end >= i_size_read(file_inode(memfd))) | |
3647 | return -EINVAL; | |
3648 | ||
3649 | if (is_file_hugepages(memfd)) { | |
3650 | h = hstate_file(memfd); | |
3651 | pgshift = huge_page_shift(h); | |
3652 | } | |
3653 | ||
3654 | flags = memalloc_pin_save(); | |
3655 | do { | |
3656 | nr_folios = 0; | |
3657 | start_idx = start >> pgshift; | |
3658 | end_idx = end >> pgshift; | |
3659 | if (is_file_hugepages(memfd)) { | |
3660 | start_idx <<= huge_page_order(h); | |
3661 | end_idx <<= huge_page_order(h); | |
3662 | } | |
3663 | ||
3664 | folio_batch_init(&fbatch); | |
3665 | while (start_idx <= end_idx && nr_folios < max_folios) { | |
3666 | /* | |
3667 | * In most cases, we should be able to find the folios | |
3668 | * in the page cache. If we cannot find them for some | |
3669 | * reason, we try to allocate them and add them to the | |
3670 | * page cache. | |
3671 | */ | |
3672 | nr_found = filemap_get_folios_contig(memfd->f_mapping, | |
3673 | &start_idx, | |
3674 | end_idx, | |
3675 | &fbatch); | |
3676 | if (folio) { | |
3677 | folio_put(folio); | |
3678 | folio = NULL; | |
3679 | } | |
3680 | ||
3681 | next_idx = 0; | |
3682 | for (i = 0; i < nr_found; i++) { | |
3683 | /* | |
3684 | * As there can be multiple entries for a | |
3685 | * given folio in the batch returned by | |
3686 | * filemap_get_folios_contig(), the below | |
3687 | * check is to ensure that we pin and return a | |
3688 | * unique set of folios between start and end. | |
3689 | */ | |
3690 | if (next_idx && | |
3691 | next_idx != folio_index(fbatch.folios[i])) | |
3692 | continue; | |
3693 | ||
3694 | folio = page_folio(&fbatch.folios[i]->page); | |
3695 | ||
3696 | if (try_grab_folio(folio, 1, FOLL_PIN)) { | |
3697 | folio_batch_release(&fbatch); | |
3698 | ret = -EINVAL; | |
3699 | goto err; | |
3700 | } | |
3701 | ||
3702 | if (nr_folios == 0) | |
3703 | *offset = offset_in_folio(folio, start); | |
3704 | ||
3705 | folios[nr_folios] = folio; | |
3706 | next_idx = folio_next_index(folio); | |
3707 | if (++nr_folios == max_folios) | |
3708 | break; | |
3709 | } | |
3710 | ||
3711 | folio = NULL; | |
3712 | folio_batch_release(&fbatch); | |
3713 | if (!nr_found) { | |
3714 | folio = memfd_alloc_folio(memfd, start_idx); | |
3715 | if (IS_ERR(folio)) { | |
3716 | ret = PTR_ERR(folio); | |
3717 | if (ret != -EEXIST) | |
3718 | goto err; | |
3719 | } | |
3720 | } | |
3721 | } | |
3722 | ||
3723 | ret = check_and_migrate_movable_folios(nr_folios, folios); | |
3724 | } while (ret == -EAGAIN); | |
3725 | ||
3726 | memalloc_pin_restore(flags); | |
3727 | return ret ? ret : nr_folios; | |
3728 | err: | |
3729 | memalloc_pin_restore(flags); | |
3730 | unpin_folios(folios, nr_folios); | |
3731 | ||
3732 | return ret; | |
3733 | } | |
3734 | EXPORT_SYMBOL_GPL(memfd_pin_folios); |