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> |
3565fce3 | 8 | #include <linux/memremap.h> |
4bbd4c77 KS |
9 | #include <linux/pagemap.h> |
10 | #include <linux/rmap.h> | |
11 | #include <linux/swap.h> | |
12 | #include <linux/swapops.h> | |
1507f512 | 13 | #include <linux/secretmem.h> |
4bbd4c77 | 14 | |
174cd4b1 | 15 | #include <linux/sched/signal.h> |
2667f50e | 16 | #include <linux/rwsem.h> |
f30c59e9 | 17 | #include <linux/hugetlb.h> |
9a4e9f3b AK |
18 | #include <linux/migrate.h> |
19 | #include <linux/mm_inline.h> | |
20 | #include <linux/sched/mm.h> | |
a6e79df9 | 21 | #include <linux/shmem_fs.h> |
1027e443 | 22 | |
33a709b2 | 23 | #include <asm/mmu_context.h> |
1027e443 | 24 | #include <asm/tlbflush.h> |
2667f50e | 25 | |
4bbd4c77 KS |
26 | #include "internal.h" |
27 | ||
df06b37f KB |
28 | struct follow_page_context { |
29 | struct dev_pagemap *pgmap; | |
30 | unsigned int page_mask; | |
31 | }; | |
32 | ||
b6a2619c DH |
33 | static inline void sanity_check_pinned_pages(struct page **pages, |
34 | unsigned long npages) | |
35 | { | |
36 | if (!IS_ENABLED(CONFIG_DEBUG_VM)) | |
37 | return; | |
38 | ||
39 | /* | |
40 | * We only pin anonymous pages if they are exclusive. Once pinned, we | |
41 | * can no longer turn them possibly shared and PageAnonExclusive() will | |
42 | * stick around until the page is freed. | |
43 | * | |
44 | * We'd like to verify that our pinned anonymous pages are still mapped | |
45 | * exclusively. The issue with anon THP is that we don't know how | |
46 | * they are/were mapped when pinning them. However, for anon | |
47 | * THP we can assume that either the given page (PTE-mapped THP) or | |
48 | * the head page (PMD-mapped THP) should be PageAnonExclusive(). If | |
49 | * neither is the case, there is certainly something wrong. | |
50 | */ | |
51 | for (; npages; npages--, pages++) { | |
52 | struct page *page = *pages; | |
53 | struct folio *folio = page_folio(page); | |
54 | ||
c8070b78 DH |
55 | if (is_zero_page(page) || |
56 | !folio_test_anon(folio)) | |
b6a2619c DH |
57 | continue; |
58 | if (!folio_test_large(folio) || folio_test_hugetlb(folio)) | |
59 | VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page); | |
60 | else | |
61 | /* Either a PTE-mapped or a PMD-mapped THP. */ | |
62 | VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) && | |
63 | !PageAnonExclusive(page), page); | |
64 | } | |
65 | } | |
66 | ||
cd1adf1b | 67 | /* |
ece1ed7b | 68 | * Return the folio with ref appropriately incremented, |
cd1adf1b | 69 | * or NULL if that failed. |
a707cdd5 | 70 | */ |
ece1ed7b | 71 | static inline struct folio *try_get_folio(struct page *page, int refs) |
a707cdd5 | 72 | { |
ece1ed7b | 73 | struct folio *folio; |
a707cdd5 | 74 | |
59409373 | 75 | retry: |
ece1ed7b MWO |
76 | folio = page_folio(page); |
77 | if (WARN_ON_ONCE(folio_ref_count(folio) < 0)) | |
a707cdd5 | 78 | return NULL; |
ece1ed7b | 79 | if (unlikely(!folio_ref_try_add_rcu(folio, refs))) |
a707cdd5 | 80 | return NULL; |
c24d3732 JH |
81 | |
82 | /* | |
ece1ed7b MWO |
83 | * At this point we have a stable reference to the folio; but it |
84 | * could be that between calling page_folio() and the refcount | |
85 | * increment, the folio was split, in which case we'd end up | |
86 | * holding a reference on a folio that has nothing to do with the page | |
c24d3732 | 87 | * we were given anymore. |
ece1ed7b MWO |
88 | * So now that the folio is stable, recheck that the page still |
89 | * belongs to this folio. | |
c24d3732 | 90 | */ |
ece1ed7b | 91 | if (unlikely(page_folio(page) != folio)) { |
53e45c4f | 92 | if (!put_devmap_managed_folio_refs(folio, refs)) |
f4f451a1 | 93 | folio_put_refs(folio, refs); |
59409373 | 94 | goto retry; |
c24d3732 JH |
95 | } |
96 | ||
ece1ed7b | 97 | return folio; |
a707cdd5 JH |
98 | } |
99 | ||
3967db22 | 100 | /** |
ece1ed7b | 101 | * try_grab_folio() - Attempt to get or pin a folio. |
3967db22 | 102 | * @page: pointer to page to be grabbed |
ece1ed7b | 103 | * @refs: the value to (effectively) add to the folio's refcount |
3967db22 JH |
104 | * @flags: gup flags: these are the FOLL_* flag values. |
105 | * | |
3faa52c0 | 106 | * "grab" names in this file mean, "look at flags to decide whether to use |
ece1ed7b | 107 | * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount. |
3faa52c0 JH |
108 | * |
109 | * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the | |
110 | * same time. (That's true throughout the get_user_pages*() and | |
111 | * pin_user_pages*() APIs.) Cases: | |
112 | * | |
ece1ed7b | 113 | * FOLL_GET: folio's refcount will be incremented by @refs. |
3967db22 | 114 | * |
ece1ed7b | 115 | * FOLL_PIN on large folios: folio's refcount will be incremented by |
94688e8e | 116 | * @refs, and its pincount will be incremented by @refs. |
3967db22 | 117 | * |
ece1ed7b | 118 | * FOLL_PIN on single-page folios: folio's refcount will be incremented by |
5232c63f | 119 | * @refs * GUP_PIN_COUNTING_BIAS. |
3faa52c0 | 120 | * |
ece1ed7b MWO |
121 | * Return: The folio containing @page (with refcount appropriately |
122 | * incremented) for success, or NULL upon failure. If neither FOLL_GET | |
123 | * nor FOLL_PIN was set, that's considered failure, and furthermore, | |
124 | * a likely bug in the caller, so a warning is also emitted. | |
3faa52c0 | 125 | */ |
ece1ed7b | 126 | struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags) |
3faa52c0 | 127 | { |
503670ee VMO |
128 | struct folio *folio; |
129 | ||
130 | if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0)) | |
131 | return NULL; | |
132 | ||
4003f107 LG |
133 | if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page))) |
134 | return NULL; | |
135 | ||
3faa52c0 | 136 | if (flags & FOLL_GET) |
ece1ed7b | 137 | return try_get_folio(page, refs); |
ece1ed7b | 138 | |
503670ee | 139 | /* FOLL_PIN is set */ |
c8070b78 | 140 | |
6e17c6de LT |
141 | /* |
142 | * Don't take a pin on the zero page - it's not going anywhere | |
143 | * and it is used in a *lot* of places. | |
144 | */ | |
145 | if (is_zero_page(page)) | |
146 | return page_folio(page); | |
df3a0a21 | 147 | |
6e17c6de | 148 | folio = try_get_folio(page, refs); |
503670ee VMO |
149 | if (!folio) |
150 | return NULL; | |
c24d3732 | 151 | |
503670ee VMO |
152 | /* |
153 | * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a | |
154 | * right zone, so fail and let the caller fall back to the slow | |
155 | * path. | |
156 | */ | |
157 | if (unlikely((flags & FOLL_LONGTERM) && | |
158 | !folio_is_longterm_pinnable(folio))) { | |
53e45c4f | 159 | if (!put_devmap_managed_folio_refs(folio, refs)) |
503670ee VMO |
160 | folio_put_refs(folio, refs); |
161 | return NULL; | |
3faa52c0 | 162 | } |
088b8aa5 | 163 | |
503670ee VMO |
164 | /* |
165 | * When pinning a large folio, use an exact count to track it. | |
166 | * | |
167 | * However, be sure to *also* increment the normal folio | |
168 | * refcount field at least once, so that the folio really | |
169 | * is pinned. That's why the refcount from the earlier | |
170 | * try_get_folio() is left intact. | |
171 | */ | |
172 | if (folio_test_large(folio)) | |
173 | atomic_add(refs, &folio->_pincount); | |
174 | else | |
175 | folio_ref_add(folio, | |
176 | refs * (GUP_PIN_COUNTING_BIAS - 1)); | |
177 | /* | |
178 | * Adjust the pincount before re-checking the PTE for changes. | |
179 | * This is essentially a smp_mb() and is paired with a memory | |
e3b4b137 | 180 | * barrier in folio_try_share_anon_rmap_*(). |
503670ee VMO |
181 | */ |
182 | smp_mb__after_atomic(); | |
47e29d32 | 183 | |
503670ee | 184 | node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs); |
3faa52c0 | 185 | |
503670ee | 186 | return folio; |
3faa52c0 JH |
187 | } |
188 | ||
d8ddc099 | 189 | static void gup_put_folio(struct folio *folio, int refs, unsigned int flags) |
4509b42c JG |
190 | { |
191 | if (flags & FOLL_PIN) { | |
c8070b78 DH |
192 | if (is_zero_folio(folio)) |
193 | return; | |
d8ddc099 MWO |
194 | node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs); |
195 | if (folio_test_large(folio)) | |
94688e8e | 196 | atomic_sub(refs, &folio->_pincount); |
4509b42c JG |
197 | else |
198 | refs *= GUP_PIN_COUNTING_BIAS; | |
199 | } | |
200 | ||
53e45c4f | 201 | if (!put_devmap_managed_folio_refs(folio, refs)) |
f4f451a1 | 202 | folio_put_refs(folio, refs); |
4509b42c JG |
203 | } |
204 | ||
3faa52c0 JH |
205 | /** |
206 | * try_grab_page() - elevate a page's refcount by a flag-dependent amount | |
5fec0719 MWO |
207 | * @page: pointer to page to be grabbed |
208 | * @flags: gup flags: these are the FOLL_* flag values. | |
3faa52c0 JH |
209 | * |
210 | * This might not do anything at all, depending on the flags argument. | |
211 | * | |
212 | * "grab" names in this file mean, "look at flags to decide whether to use | |
213 | * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount. | |
214 | * | |
3faa52c0 | 215 | * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same |
ece1ed7b | 216 | * time. Cases: please see the try_grab_folio() documentation, with |
3967db22 | 217 | * "refs=1". |
3faa52c0 | 218 | * |
0f089235 LG |
219 | * Return: 0 for success, or if no action was required (if neither FOLL_PIN |
220 | * nor FOLL_GET was set, nothing is done). A negative error code for failure: | |
221 | * | |
222 | * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not | |
223 | * be grabbed. | |
3faa52c0 | 224 | */ |
0f089235 | 225 | int __must_check try_grab_page(struct page *page, unsigned int flags) |
3faa52c0 | 226 | { |
5fec0719 MWO |
227 | struct folio *folio = page_folio(page); |
228 | ||
5fec0719 | 229 | if (WARN_ON_ONCE(folio_ref_count(folio) <= 0)) |
0f089235 | 230 | return -ENOMEM; |
3faa52c0 | 231 | |
4003f107 LG |
232 | if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page))) |
233 | return -EREMOTEIO; | |
3faa52c0 | 234 | |
c36c04c2 | 235 | if (flags & FOLL_GET) |
5fec0719 | 236 | folio_ref_inc(folio); |
c36c04c2 | 237 | else if (flags & FOLL_PIN) { |
c8070b78 DH |
238 | /* |
239 | * Don't take a pin on the zero page - it's not going anywhere | |
240 | * and it is used in a *lot* of places. | |
241 | */ | |
242 | if (is_zero_page(page)) | |
243 | return 0; | |
244 | ||
c36c04c2 | 245 | /* |
5fec0719 | 246 | * Similar to try_grab_folio(): be sure to *also* |
78d9d6ce MWO |
247 | * increment the normal page refcount field at least once, |
248 | * so that the page really is pinned. | |
c36c04c2 | 249 | */ |
5fec0719 MWO |
250 | if (folio_test_large(folio)) { |
251 | folio_ref_add(folio, 1); | |
94688e8e | 252 | atomic_add(1, &folio->_pincount); |
8ea2979c | 253 | } else { |
5fec0719 | 254 | folio_ref_add(folio, GUP_PIN_COUNTING_BIAS); |
8ea2979c | 255 | } |
c36c04c2 | 256 | |
5fec0719 | 257 | node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1); |
c36c04c2 JH |
258 | } |
259 | ||
0f089235 | 260 | return 0; |
3faa52c0 JH |
261 | } |
262 | ||
3faa52c0 JH |
263 | /** |
264 | * unpin_user_page() - release a dma-pinned page | |
265 | * @page: pointer to page to be released | |
266 | * | |
267 | * Pages that were pinned via pin_user_pages*() must be released via either | |
268 | * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so | |
269 | * that such pages can be separately tracked and uniquely handled. In | |
270 | * particular, interactions with RDMA and filesystems need special handling. | |
271 | */ | |
272 | void unpin_user_page(struct page *page) | |
273 | { | |
b6a2619c | 274 | sanity_check_pinned_pages(&page, 1); |
d8ddc099 | 275 | gup_put_folio(page_folio(page), 1, FOLL_PIN); |
3faa52c0 JH |
276 | } |
277 | EXPORT_SYMBOL(unpin_user_page); | |
278 | ||
1101fb8f DH |
279 | /** |
280 | * folio_add_pin - Try to get an additional pin on a pinned folio | |
281 | * @folio: The folio to be pinned | |
282 | * | |
283 | * Get an additional pin on a folio we already have a pin on. Makes no change | |
284 | * if the folio is a zero_page. | |
285 | */ | |
286 | void folio_add_pin(struct folio *folio) | |
287 | { | |
288 | if (is_zero_folio(folio)) | |
289 | return; | |
290 | ||
291 | /* | |
292 | * Similar to try_grab_folio(): be sure to *also* increment the normal | |
293 | * page refcount field at least once, so that the page really is | |
294 | * pinned. | |
295 | */ | |
296 | if (folio_test_large(folio)) { | |
297 | WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1); | |
298 | folio_ref_inc(folio); | |
299 | atomic_inc(&folio->_pincount); | |
300 | } else { | |
301 | WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS); | |
302 | folio_ref_add(folio, GUP_PIN_COUNTING_BIAS); | |
303 | } | |
304 | } | |
305 | ||
659508f9 | 306 | static inline struct folio *gup_folio_range_next(struct page *start, |
8f39f5fc | 307 | unsigned long npages, unsigned long i, unsigned int *ntails) |
458a4f78 | 308 | { |
659508f9 MWO |
309 | struct page *next = nth_page(start, i); |
310 | struct folio *folio = page_folio(next); | |
458a4f78 JM |
311 | unsigned int nr = 1; |
312 | ||
659508f9 | 313 | if (folio_test_large(folio)) |
4c654229 | 314 | nr = min_t(unsigned int, npages - i, |
659508f9 | 315 | folio_nr_pages(folio) - folio_page_idx(folio, next)); |
458a4f78 | 316 | |
458a4f78 | 317 | *ntails = nr; |
659508f9 | 318 | return folio; |
458a4f78 JM |
319 | } |
320 | ||
12521c76 | 321 | static inline struct folio *gup_folio_next(struct page **list, |
28297dbc | 322 | unsigned long npages, unsigned long i, unsigned int *ntails) |
8745d7f6 | 323 | { |
12521c76 | 324 | struct folio *folio = page_folio(list[i]); |
8745d7f6 JM |
325 | unsigned int nr; |
326 | ||
8745d7f6 | 327 | for (nr = i + 1; nr < npages; nr++) { |
12521c76 | 328 | if (page_folio(list[nr]) != folio) |
8745d7f6 JM |
329 | break; |
330 | } | |
331 | ||
8745d7f6 | 332 | *ntails = nr - i; |
12521c76 | 333 | return folio; |
8745d7f6 JM |
334 | } |
335 | ||
fc1d8e7c | 336 | /** |
f1f6a7dd | 337 | * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages |
2d15eb31 | 338 | * @pages: array of pages to be maybe marked dirty, and definitely released. |
fc1d8e7c | 339 | * @npages: number of pages in the @pages array. |
2d15eb31 | 340 | * @make_dirty: whether to mark the pages dirty |
fc1d8e7c JH |
341 | * |
342 | * "gup-pinned page" refers to a page that has had one of the get_user_pages() | |
343 | * variants called on that page. | |
344 | * | |
345 | * For each page in the @pages array, make that page (or its head page, if a | |
2d15eb31 | 346 | * compound page) dirty, if @make_dirty is true, and if the page was previously |
f1f6a7dd JH |
347 | * listed as clean. In any case, releases all pages using unpin_user_page(), |
348 | * possibly via unpin_user_pages(), for the non-dirty case. | |
fc1d8e7c | 349 | * |
f1f6a7dd | 350 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 351 | * |
2d15eb31 | 352 | * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is |
353 | * required, then the caller should a) verify that this is really correct, | |
354 | * because _lock() is usually required, and b) hand code it: | |
f1f6a7dd | 355 | * set_page_dirty_lock(), unpin_user_page(). |
fc1d8e7c JH |
356 | * |
357 | */ | |
f1f6a7dd JH |
358 | void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages, |
359 | bool make_dirty) | |
fc1d8e7c | 360 | { |
12521c76 MWO |
361 | unsigned long i; |
362 | struct folio *folio; | |
363 | unsigned int nr; | |
2d15eb31 | 364 | |
365 | if (!make_dirty) { | |
f1f6a7dd | 366 | unpin_user_pages(pages, npages); |
2d15eb31 | 367 | return; |
368 | } | |
369 | ||
b6a2619c | 370 | sanity_check_pinned_pages(pages, npages); |
12521c76 MWO |
371 | for (i = 0; i < npages; i += nr) { |
372 | folio = gup_folio_next(pages, npages, i, &nr); | |
2d15eb31 | 373 | /* |
374 | * Checking PageDirty at this point may race with | |
375 | * clear_page_dirty_for_io(), but that's OK. Two key | |
376 | * cases: | |
377 | * | |
378 | * 1) This code sees the page as already dirty, so it | |
379 | * skips the call to set_page_dirty(). That could happen | |
380 | * because clear_page_dirty_for_io() called | |
381 | * page_mkclean(), followed by set_page_dirty(). | |
382 | * However, now the page is going to get written back, | |
383 | * which meets the original intention of setting it | |
384 | * dirty, so all is well: clear_page_dirty_for_io() goes | |
385 | * on to call TestClearPageDirty(), and write the page | |
386 | * back. | |
387 | * | |
388 | * 2) This code sees the page as clean, so it calls | |
389 | * set_page_dirty(). The page stays dirty, despite being | |
390 | * written back, so it gets written back again in the | |
391 | * next writeback cycle. This is harmless. | |
392 | */ | |
12521c76 MWO |
393 | if (!folio_test_dirty(folio)) { |
394 | folio_lock(folio); | |
395 | folio_mark_dirty(folio); | |
396 | folio_unlock(folio); | |
397 | } | |
398 | gup_put_folio(folio, nr, FOLL_PIN); | |
2d15eb31 | 399 | } |
fc1d8e7c | 400 | } |
f1f6a7dd | 401 | EXPORT_SYMBOL(unpin_user_pages_dirty_lock); |
fc1d8e7c | 402 | |
458a4f78 JM |
403 | /** |
404 | * unpin_user_page_range_dirty_lock() - release and optionally dirty | |
405 | * gup-pinned page range | |
406 | * | |
407 | * @page: the starting page of a range maybe marked dirty, and definitely released. | |
408 | * @npages: number of consecutive pages to release. | |
409 | * @make_dirty: whether to mark the pages dirty | |
410 | * | |
411 | * "gup-pinned page range" refers to a range of pages that has had one of the | |
412 | * pin_user_pages() variants called on that page. | |
413 | * | |
414 | * For the page ranges defined by [page .. page+npages], make that range (or | |
415 | * its head pages, if a compound page) dirty, if @make_dirty is true, and if the | |
416 | * page range was previously listed as clean. | |
417 | * | |
418 | * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is | |
419 | * required, then the caller should a) verify that this is really correct, | |
420 | * because _lock() is usually required, and b) hand code it: | |
421 | * set_page_dirty_lock(), unpin_user_page(). | |
422 | * | |
423 | */ | |
424 | void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages, | |
425 | bool make_dirty) | |
426 | { | |
659508f9 MWO |
427 | unsigned long i; |
428 | struct folio *folio; | |
429 | unsigned int nr; | |
430 | ||
431 | for (i = 0; i < npages; i += nr) { | |
432 | folio = gup_folio_range_next(page, npages, i, &nr); | |
433 | if (make_dirty && !folio_test_dirty(folio)) { | |
434 | folio_lock(folio); | |
435 | folio_mark_dirty(folio); | |
436 | folio_unlock(folio); | |
437 | } | |
438 | gup_put_folio(folio, nr, FOLL_PIN); | |
458a4f78 JM |
439 | } |
440 | } | |
441 | EXPORT_SYMBOL(unpin_user_page_range_dirty_lock); | |
442 | ||
23babe19 | 443 | static void gup_fast_unpin_user_pages(struct page **pages, unsigned long npages) |
b6a2619c DH |
444 | { |
445 | unsigned long i; | |
446 | struct folio *folio; | |
447 | unsigned int nr; | |
448 | ||
449 | /* | |
450 | * Don't perform any sanity checks because we might have raced with | |
451 | * fork() and some anonymous pages might now actually be shared -- | |
452 | * which is why we're unpinning after all. | |
453 | */ | |
454 | for (i = 0; i < npages; i += nr) { | |
455 | folio = gup_folio_next(pages, npages, i, &nr); | |
456 | gup_put_folio(folio, nr, FOLL_PIN); | |
457 | } | |
458 | } | |
459 | ||
fc1d8e7c | 460 | /** |
f1f6a7dd | 461 | * unpin_user_pages() - release an array of gup-pinned pages. |
fc1d8e7c JH |
462 | * @pages: array of pages to be marked dirty and released. |
463 | * @npages: number of pages in the @pages array. | |
464 | * | |
f1f6a7dd | 465 | * For each page in the @pages array, release the page using unpin_user_page(). |
fc1d8e7c | 466 | * |
f1f6a7dd | 467 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 468 | */ |
f1f6a7dd | 469 | void unpin_user_pages(struct page **pages, unsigned long npages) |
fc1d8e7c | 470 | { |
12521c76 MWO |
471 | unsigned long i; |
472 | struct folio *folio; | |
473 | unsigned int nr; | |
fc1d8e7c | 474 | |
146608bb JH |
475 | /* |
476 | * If this WARN_ON() fires, then the system *might* be leaking pages (by | |
477 | * leaving them pinned), but probably not. More likely, gup/pup returned | |
478 | * a hard -ERRNO error to the caller, who erroneously passed it here. | |
479 | */ | |
480 | if (WARN_ON(IS_ERR_VALUE(npages))) | |
481 | return; | |
31b912de | 482 | |
b6a2619c | 483 | sanity_check_pinned_pages(pages, npages); |
12521c76 MWO |
484 | for (i = 0; i < npages; i += nr) { |
485 | folio = gup_folio_next(pages, npages, i, &nr); | |
486 | gup_put_folio(folio, nr, FOLL_PIN); | |
e7602748 | 487 | } |
fc1d8e7c | 488 | } |
f1f6a7dd | 489 | EXPORT_SYMBOL(unpin_user_pages); |
fc1d8e7c | 490 | |
a458b76a AA |
491 | /* |
492 | * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's | |
493 | * lifecycle. Avoid setting the bit unless necessary, or it might cause write | |
494 | * cache bouncing on large SMP machines for concurrent pinned gups. | |
495 | */ | |
496 | static inline void mm_set_has_pinned_flag(unsigned long *mm_flags) | |
497 | { | |
498 | if (!test_bit(MMF_HAS_PINNED, mm_flags)) | |
499 | set_bit(MMF_HAS_PINNED, mm_flags); | |
500 | } | |
501 | ||
050a9adc | 502 | #ifdef CONFIG_MMU |
a12083d7 | 503 | |
25176ad0 | 504 | #if defined(CONFIG_ARCH_HAS_HUGEPD) || defined(CONFIG_HAVE_GUP_FAST) |
a12083d7 PX |
505 | static int record_subpages(struct page *page, unsigned long sz, |
506 | unsigned long addr, unsigned long end, | |
507 | struct page **pages) | |
508 | { | |
509 | struct page *start_page; | |
510 | int nr; | |
511 | ||
512 | start_page = nth_page(page, (addr & (sz - 1)) >> PAGE_SHIFT); | |
513 | for (nr = 0; addr != end; nr++, addr += PAGE_SIZE) | |
514 | pages[nr] = nth_page(start_page, nr); | |
515 | ||
516 | return nr; | |
517 | } | |
25176ad0 | 518 | #endif /* CONFIG_ARCH_HAS_HUGEPD || CONFIG_HAVE_GUP_FAST */ |
a12083d7 PX |
519 | |
520 | #ifdef CONFIG_ARCH_HAS_HUGEPD | |
521 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, | |
522 | unsigned long sz) | |
523 | { | |
524 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
525 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
526 | } | |
527 | ||
23babe19 DH |
528 | static int gup_fast_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, |
529 | unsigned long end, unsigned int flags, struct page **pages, | |
530 | int *nr) | |
a12083d7 PX |
531 | { |
532 | unsigned long pte_end; | |
533 | struct page *page; | |
534 | struct folio *folio; | |
535 | pte_t pte; | |
536 | int refs; | |
537 | ||
538 | pte_end = (addr + sz) & ~(sz-1); | |
539 | if (pte_end < end) | |
540 | end = pte_end; | |
541 | ||
542 | pte = huge_ptep_get(ptep); | |
543 | ||
544 | if (!pte_access_permitted(pte, flags & FOLL_WRITE)) | |
545 | return 0; | |
546 | ||
547 | /* hugepages are never "special" */ | |
548 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
549 | ||
550 | page = pte_page(pte); | |
551 | refs = record_subpages(page, sz, addr, end, pages + *nr); | |
552 | ||
553 | folio = try_grab_folio(page, refs, flags); | |
554 | if (!folio) | |
555 | return 0; | |
556 | ||
557 | if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) { | |
558 | gup_put_folio(folio, refs, flags); | |
559 | return 0; | |
560 | } | |
561 | ||
562 | if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) { | |
563 | gup_put_folio(folio, refs, flags); | |
564 | return 0; | |
565 | } | |
566 | ||
567 | *nr += refs; | |
568 | folio_set_referenced(folio); | |
569 | return 1; | |
570 | } | |
571 | ||
572 | /* | |
573 | * NOTE: currently GUP for a hugepd is only possible on hugetlbfs file | |
574 | * systems on Power, which does not have issue with folio writeback against | |
575 | * GUP updates. When hugepd will be extended to support non-hugetlbfs or | |
576 | * even anonymous memory, we need to do extra check as what we do with most | |
577 | * of the other folios. See writable_file_mapping_allowed() and | |
578 | * gup_fast_folio_allowed() for more information. | |
579 | */ | |
23babe19 | 580 | static int gup_fast_hugepd(hugepd_t hugepd, unsigned long addr, |
a12083d7 PX |
581 | unsigned int pdshift, unsigned long end, unsigned int flags, |
582 | struct page **pages, int *nr) | |
583 | { | |
584 | pte_t *ptep; | |
585 | unsigned long sz = 1UL << hugepd_shift(hugepd); | |
586 | unsigned long next; | |
587 | ||
588 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
589 | do { | |
590 | next = hugepte_addr_end(addr, end, sz); | |
23babe19 | 591 | if (!gup_fast_hugepte(ptep, sz, addr, end, flags, pages, nr)) |
a12083d7 PX |
592 | return 0; |
593 | } while (ptep++, addr = next, addr != end); | |
594 | ||
595 | return 1; | |
596 | } | |
597 | ||
598 | static struct page *follow_hugepd(struct vm_area_struct *vma, hugepd_t hugepd, | |
599 | unsigned long addr, unsigned int pdshift, | |
600 | unsigned int flags, | |
601 | struct follow_page_context *ctx) | |
602 | { | |
603 | struct page *page; | |
604 | struct hstate *h; | |
605 | spinlock_t *ptl; | |
606 | int nr = 0, ret; | |
607 | pte_t *ptep; | |
608 | ||
609 | /* Only hugetlb supports hugepd */ | |
610 | if (WARN_ON_ONCE(!is_vm_hugetlb_page(vma))) | |
611 | return ERR_PTR(-EFAULT); | |
612 | ||
613 | h = hstate_vma(vma); | |
614 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
615 | ptl = huge_pte_lock(h, vma->vm_mm, ptep); | |
23babe19 DH |
616 | ret = gup_fast_hugepd(hugepd, addr, pdshift, addr + PAGE_SIZE, |
617 | flags, &page, &nr); | |
a12083d7 PX |
618 | spin_unlock(ptl); |
619 | ||
620 | if (ret) { | |
621 | WARN_ON_ONCE(nr != 1); | |
622 | ctx->page_mask = (1U << huge_page_order(h)) - 1; | |
623 | return page; | |
624 | } | |
625 | ||
626 | return NULL; | |
627 | } | |
628 | #else /* CONFIG_ARCH_HAS_HUGEPD */ | |
23babe19 | 629 | static inline int gup_fast_hugepd(hugepd_t hugepd, unsigned long addr, |
a12083d7 PX |
630 | unsigned int pdshift, unsigned long end, unsigned int flags, |
631 | struct page **pages, int *nr) | |
632 | { | |
633 | return 0; | |
634 | } | |
635 | ||
636 | static struct page *follow_hugepd(struct vm_area_struct *vma, hugepd_t hugepd, | |
637 | unsigned long addr, unsigned int pdshift, | |
638 | unsigned int flags, | |
639 | struct follow_page_context *ctx) | |
640 | { | |
641 | return NULL; | |
642 | } | |
643 | #endif /* CONFIG_ARCH_HAS_HUGEPD */ | |
644 | ||
645 | ||
69e68b4f | 646 | static struct page *no_page_table(struct vm_area_struct *vma, |
878b0c45 | 647 | unsigned int flags, unsigned long address) |
4bbd4c77 | 648 | { |
878b0c45 PX |
649 | if (!(flags & FOLL_DUMP)) |
650 | return NULL; | |
651 | ||
69e68b4f | 652 | /* |
878b0c45 | 653 | * When core dumping, we don't want to allocate unnecessary pages or |
69e68b4f KS |
654 | * page tables. Return error instead of NULL to skip handle_mm_fault, |
655 | * then get_dump_page() will return NULL to leave a hole in the dump. | |
656 | * But we can only make this optimization where a hole would surely | |
657 | * be zero-filled if handle_mm_fault() actually did handle it. | |
658 | */ | |
878b0c45 PX |
659 | if (is_vm_hugetlb_page(vma)) { |
660 | struct hstate *h = hstate_vma(vma); | |
661 | ||
662 | if (!hugetlbfs_pagecache_present(h, vma, address)) | |
663 | return ERR_PTR(-EFAULT); | |
664 | } else if ((vma_is_anonymous(vma) || !vma->vm_ops->fault)) { | |
69e68b4f | 665 | return ERR_PTR(-EFAULT); |
878b0c45 PX |
666 | } |
667 | ||
69e68b4f KS |
668 | return NULL; |
669 | } | |
4bbd4c77 | 670 | |
1b167618 PX |
671 | #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES |
672 | static struct page *follow_huge_pud(struct vm_area_struct *vma, | |
673 | unsigned long addr, pud_t *pudp, | |
674 | int flags, struct follow_page_context *ctx) | |
675 | { | |
676 | struct mm_struct *mm = vma->vm_mm; | |
677 | struct page *page; | |
678 | pud_t pud = *pudp; | |
679 | unsigned long pfn = pud_pfn(pud); | |
680 | int ret; | |
681 | ||
682 | assert_spin_locked(pud_lockptr(mm, pudp)); | |
683 | ||
684 | if ((flags & FOLL_WRITE) && !pud_write(pud)) | |
685 | return NULL; | |
686 | ||
687 | if (!pud_present(pud)) | |
688 | return NULL; | |
689 | ||
690 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
691 | ||
692 | if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) && | |
693 | pud_devmap(pud)) { | |
694 | /* | |
695 | * device mapped pages can only be returned if the caller | |
696 | * will manage the page reference count. | |
697 | * | |
698 | * At least one of FOLL_GET | FOLL_PIN must be set, so | |
699 | * assert that here: | |
700 | */ | |
701 | if (!(flags & (FOLL_GET | FOLL_PIN))) | |
702 | return ERR_PTR(-EEXIST); | |
703 | ||
704 | if (flags & FOLL_TOUCH) | |
705 | touch_pud(vma, addr, pudp, flags & FOLL_WRITE); | |
706 | ||
707 | ctx->pgmap = get_dev_pagemap(pfn, ctx->pgmap); | |
708 | if (!ctx->pgmap) | |
709 | return ERR_PTR(-EFAULT); | |
710 | } | |
711 | ||
712 | page = pfn_to_page(pfn); | |
713 | ||
714 | if (!pud_devmap(pud) && !pud_write(pud) && | |
715 | gup_must_unshare(vma, flags, page)) | |
716 | return ERR_PTR(-EMLINK); | |
717 | ||
718 | ret = try_grab_page(page, flags); | |
719 | if (ret) | |
720 | page = ERR_PTR(ret); | |
721 | else | |
722 | ctx->page_mask = HPAGE_PUD_NR - 1; | |
723 | ||
724 | return page; | |
725 | } | |
4418c522 PX |
726 | |
727 | /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ | |
728 | static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, | |
729 | struct vm_area_struct *vma, | |
730 | unsigned int flags) | |
731 | { | |
732 | /* If the pmd is writable, we can write to the page. */ | |
733 | if (pmd_write(pmd)) | |
734 | return true; | |
735 | ||
736 | /* Maybe FOLL_FORCE is set to override it? */ | |
737 | if (!(flags & FOLL_FORCE)) | |
738 | return false; | |
739 | ||
740 | /* But FOLL_FORCE has no effect on shared mappings */ | |
741 | if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) | |
742 | return false; | |
743 | ||
744 | /* ... or read-only private ones */ | |
745 | if (!(vma->vm_flags & VM_MAYWRITE)) | |
746 | return false; | |
747 | ||
748 | /* ... or already writable ones that just need to take a write fault */ | |
749 | if (vma->vm_flags & VM_WRITE) | |
750 | return false; | |
751 | ||
752 | /* | |
753 | * See can_change_pte_writable(): we broke COW and could map the page | |
754 | * writable if we have an exclusive anonymous page ... | |
755 | */ | |
756 | if (!page || !PageAnon(page) || !PageAnonExclusive(page)) | |
757 | return false; | |
758 | ||
759 | /* ... and a write-fault isn't required for other reasons. */ | |
760 | if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) | |
761 | return false; | |
762 | return !userfaultfd_huge_pmd_wp(vma, pmd); | |
763 | } | |
764 | ||
765 | static struct page *follow_huge_pmd(struct vm_area_struct *vma, | |
766 | unsigned long addr, pmd_t *pmd, | |
767 | unsigned int flags, | |
768 | struct follow_page_context *ctx) | |
769 | { | |
770 | struct mm_struct *mm = vma->vm_mm; | |
771 | pmd_t pmdval = *pmd; | |
772 | struct page *page; | |
773 | int ret; | |
774 | ||
775 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
776 | ||
777 | page = pmd_page(pmdval); | |
778 | if ((flags & FOLL_WRITE) && | |
779 | !can_follow_write_pmd(pmdval, page, vma, flags)) | |
780 | return NULL; | |
781 | ||
782 | /* Avoid dumping huge zero page */ | |
783 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(pmdval)) | |
784 | return ERR_PTR(-EFAULT); | |
785 | ||
786 | if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags)) | |
787 | return NULL; | |
788 | ||
789 | if (!pmd_write(pmdval) && gup_must_unshare(vma, flags, page)) | |
790 | return ERR_PTR(-EMLINK); | |
791 | ||
792 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && | |
793 | !PageAnonExclusive(page), page); | |
794 | ||
795 | ret = try_grab_page(page, flags); | |
796 | if (ret) | |
797 | return ERR_PTR(ret); | |
798 | ||
799 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
800 | if (pmd_trans_huge(pmdval) && (flags & FOLL_TOUCH)) | |
801 | touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); | |
802 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
803 | ||
804 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; | |
805 | ctx->page_mask = HPAGE_PMD_NR - 1; | |
806 | ||
807 | return page; | |
808 | } | |
809 | ||
1b167618 PX |
810 | #else /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */ |
811 | static struct page *follow_huge_pud(struct vm_area_struct *vma, | |
812 | unsigned long addr, pud_t *pudp, | |
813 | int flags, struct follow_page_context *ctx) | |
814 | { | |
815 | return NULL; | |
816 | } | |
4418c522 PX |
817 | |
818 | static struct page *follow_huge_pmd(struct vm_area_struct *vma, | |
819 | unsigned long addr, pmd_t *pmd, | |
820 | unsigned int flags, | |
821 | struct follow_page_context *ctx) | |
822 | { | |
823 | return NULL; | |
824 | } | |
1b167618 PX |
825 | #endif /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */ |
826 | ||
1027e443 KS |
827 | static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address, |
828 | pte_t *pte, unsigned int flags) | |
829 | { | |
1027e443 | 830 | if (flags & FOLL_TOUCH) { |
c33c7948 RR |
831 | pte_t orig_entry = ptep_get(pte); |
832 | pte_t entry = orig_entry; | |
1027e443 KS |
833 | |
834 | if (flags & FOLL_WRITE) | |
835 | entry = pte_mkdirty(entry); | |
836 | entry = pte_mkyoung(entry); | |
837 | ||
c33c7948 | 838 | if (!pte_same(orig_entry, entry)) { |
1027e443 KS |
839 | set_pte_at(vma->vm_mm, address, pte, entry); |
840 | update_mmu_cache(vma, address, pte); | |
841 | } | |
842 | } | |
843 | ||
844 | /* Proper page table entry exists, but no corresponding struct page */ | |
845 | return -EEXIST; | |
846 | } | |
847 | ||
5535be30 DH |
848 | /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */ |
849 | static inline bool can_follow_write_pte(pte_t pte, struct page *page, | |
850 | struct vm_area_struct *vma, | |
851 | unsigned int flags) | |
19be0eaf | 852 | { |
5535be30 DH |
853 | /* If the pte is writable, we can write to the page. */ |
854 | if (pte_write(pte)) | |
855 | return true; | |
856 | ||
857 | /* Maybe FOLL_FORCE is set to override it? */ | |
858 | if (!(flags & FOLL_FORCE)) | |
859 | return false; | |
860 | ||
861 | /* But FOLL_FORCE has no effect on shared mappings */ | |
862 | if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) | |
863 | return false; | |
864 | ||
865 | /* ... or read-only private ones */ | |
866 | if (!(vma->vm_flags & VM_MAYWRITE)) | |
867 | return false; | |
868 | ||
869 | /* ... or already writable ones that just need to take a write fault */ | |
870 | if (vma->vm_flags & VM_WRITE) | |
871 | return false; | |
872 | ||
873 | /* | |
874 | * See can_change_pte_writable(): we broke COW and could map the page | |
875 | * writable if we have an exclusive anonymous page ... | |
876 | */ | |
877 | if (!page || !PageAnon(page) || !PageAnonExclusive(page)) | |
878 | return false; | |
879 | ||
880 | /* ... and a write-fault isn't required for other reasons. */ | |
881 | if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte)) | |
882 | return false; | |
883 | return !userfaultfd_pte_wp(vma, pte); | |
19be0eaf LT |
884 | } |
885 | ||
69e68b4f | 886 | static struct page *follow_page_pte(struct vm_area_struct *vma, |
df06b37f KB |
887 | unsigned long address, pmd_t *pmd, unsigned int flags, |
888 | struct dev_pagemap **pgmap) | |
69e68b4f KS |
889 | { |
890 | struct mm_struct *mm = vma->vm_mm; | |
891 | struct page *page; | |
892 | spinlock_t *ptl; | |
893 | pte_t *ptep, pte; | |
f28d4363 | 894 | int ret; |
4bbd4c77 | 895 | |
eddb1c22 JH |
896 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
897 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
898 | (FOLL_PIN | FOLL_GET))) | |
899 | return ERR_PTR(-EINVAL); | |
4bbd4c77 KS |
900 | |
901 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
04dee9e8 | 902 | if (!ptep) |
878b0c45 | 903 | return no_page_table(vma, flags, address); |
c33c7948 | 904 | pte = ptep_get(ptep); |
f7355e99 DH |
905 | if (!pte_present(pte)) |
906 | goto no_page; | |
d74943a2 | 907 | if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags)) |
4bbd4c77 | 908 | goto no_page; |
4bbd4c77 KS |
909 | |
910 | page = vm_normal_page(vma, address, pte); | |
5535be30 DH |
911 | |
912 | /* | |
913 | * We only care about anon pages in can_follow_write_pte() and don't | |
914 | * have to worry about pte_devmap() because they are never anon. | |
915 | */ | |
916 | if ((flags & FOLL_WRITE) && | |
917 | !can_follow_write_pte(pte, page, vma, flags)) { | |
918 | page = NULL; | |
919 | goto out; | |
920 | } | |
921 | ||
3faa52c0 | 922 | if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) { |
3565fce3 | 923 | /* |
3faa52c0 JH |
924 | * Only return device mapping pages in the FOLL_GET or FOLL_PIN |
925 | * case since they are only valid while holding the pgmap | |
926 | * reference. | |
3565fce3 | 927 | */ |
df06b37f KB |
928 | *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap); |
929 | if (*pgmap) | |
3565fce3 DW |
930 | page = pte_page(pte); |
931 | else | |
932 | goto no_page; | |
933 | } else if (unlikely(!page)) { | |
1027e443 KS |
934 | if (flags & FOLL_DUMP) { |
935 | /* Avoid special (like zero) pages in core dumps */ | |
936 | page = ERR_PTR(-EFAULT); | |
937 | goto out; | |
938 | } | |
939 | ||
940 | if (is_zero_pfn(pte_pfn(pte))) { | |
941 | page = pte_page(pte); | |
942 | } else { | |
1027e443 KS |
943 | ret = follow_pfn_pte(vma, address, ptep, flags); |
944 | page = ERR_PTR(ret); | |
945 | goto out; | |
946 | } | |
4bbd4c77 KS |
947 | } |
948 | ||
84209e87 | 949 | if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) { |
a7f22660 DH |
950 | page = ERR_PTR(-EMLINK); |
951 | goto out; | |
952 | } | |
b6a2619c DH |
953 | |
954 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && | |
955 | !PageAnonExclusive(page), page); | |
956 | ||
3faa52c0 | 957 | /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */ |
0f089235 LG |
958 | ret = try_grab_page(page, flags); |
959 | if (unlikely(ret)) { | |
960 | page = ERR_PTR(ret); | |
3faa52c0 | 961 | goto out; |
8fde12ca | 962 | } |
4003f107 | 963 | |
f28d4363 CI |
964 | /* |
965 | * We need to make the page accessible if and only if we are going | |
966 | * to access its content (the FOLL_PIN case). Please see | |
967 | * Documentation/core-api/pin_user_pages.rst for details. | |
968 | */ | |
969 | if (flags & FOLL_PIN) { | |
970 | ret = arch_make_page_accessible(page); | |
971 | if (ret) { | |
972 | unpin_user_page(page); | |
973 | page = ERR_PTR(ret); | |
974 | goto out; | |
975 | } | |
976 | } | |
4bbd4c77 KS |
977 | if (flags & FOLL_TOUCH) { |
978 | if ((flags & FOLL_WRITE) && | |
979 | !pte_dirty(pte) && !PageDirty(page)) | |
980 | set_page_dirty(page); | |
981 | /* | |
982 | * pte_mkyoung() would be more correct here, but atomic care | |
983 | * is needed to avoid losing the dirty bit: it is easier to use | |
984 | * mark_page_accessed(). | |
985 | */ | |
986 | mark_page_accessed(page); | |
987 | } | |
1027e443 | 988 | out: |
4bbd4c77 | 989 | pte_unmap_unlock(ptep, ptl); |
4bbd4c77 | 990 | return page; |
4bbd4c77 KS |
991 | no_page: |
992 | pte_unmap_unlock(ptep, ptl); | |
993 | if (!pte_none(pte)) | |
69e68b4f | 994 | return NULL; |
878b0c45 | 995 | return no_page_table(vma, flags, address); |
69e68b4f KS |
996 | } |
997 | ||
080dbb61 AK |
998 | static struct page *follow_pmd_mask(struct vm_area_struct *vma, |
999 | unsigned long address, pud_t *pudp, | |
df06b37f KB |
1000 | unsigned int flags, |
1001 | struct follow_page_context *ctx) | |
69e68b4f | 1002 | { |
68827280 | 1003 | pmd_t *pmd, pmdval; |
69e68b4f KS |
1004 | spinlock_t *ptl; |
1005 | struct page *page; | |
1006 | struct mm_struct *mm = vma->vm_mm; | |
1007 | ||
080dbb61 | 1008 | pmd = pmd_offset(pudp, address); |
26e1a0c3 | 1009 | pmdval = pmdp_get_lockless(pmd); |
68827280 | 1010 | if (pmd_none(pmdval)) |
878b0c45 | 1011 | return no_page_table(vma, flags, address); |
f7355e99 | 1012 | if (!pmd_present(pmdval)) |
878b0c45 | 1013 | return no_page_table(vma, flags, address); |
a12083d7 PX |
1014 | if (unlikely(is_hugepd(__hugepd(pmd_val(pmdval))))) |
1015 | return follow_hugepd(vma, __hugepd(pmd_val(pmdval)), | |
1016 | address, PMD_SHIFT, flags, ctx); | |
68827280 | 1017 | if (pmd_devmap(pmdval)) { |
3565fce3 | 1018 | ptl = pmd_lock(mm, pmd); |
df06b37f | 1019 | page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap); |
3565fce3 DW |
1020 | spin_unlock(ptl); |
1021 | if (page) | |
1022 | return page; | |
878b0c45 | 1023 | return no_page_table(vma, flags, address); |
3565fce3 | 1024 | } |
4418c522 | 1025 | if (likely(!pmd_leaf(pmdval))) |
df06b37f | 1026 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 1027 | |
d74943a2 | 1028 | if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags)) |
878b0c45 | 1029 | return no_page_table(vma, flags, address); |
db08f203 | 1030 | |
6742d293 | 1031 | ptl = pmd_lock(mm, pmd); |
4418c522 PX |
1032 | pmdval = *pmd; |
1033 | if (unlikely(!pmd_present(pmdval))) { | |
84c3fc4e | 1034 | spin_unlock(ptl); |
878b0c45 | 1035 | return no_page_table(vma, flags, address); |
84c3fc4e | 1036 | } |
4418c522 | 1037 | if (unlikely(!pmd_leaf(pmdval))) { |
6742d293 | 1038 | spin_unlock(ptl); |
df06b37f | 1039 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 1040 | } |
4418c522 | 1041 | if (pmd_trans_huge(pmdval) && (flags & FOLL_SPLIT_PMD)) { |
2378118b HD |
1042 | spin_unlock(ptl); |
1043 | split_huge_pmd(vma, pmd, address); | |
1044 | /* If pmd was left empty, stuff a page table in there quickly */ | |
1045 | return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) : | |
df06b37f | 1046 | follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
69e68b4f | 1047 | } |
4418c522 | 1048 | page = follow_huge_pmd(vma, address, pmd, flags, ctx); |
6742d293 | 1049 | spin_unlock(ptl); |
6742d293 | 1050 | return page; |
4bbd4c77 KS |
1051 | } |
1052 | ||
080dbb61 AK |
1053 | static struct page *follow_pud_mask(struct vm_area_struct *vma, |
1054 | unsigned long address, p4d_t *p4dp, | |
df06b37f KB |
1055 | unsigned int flags, |
1056 | struct follow_page_context *ctx) | |
080dbb61 | 1057 | { |
caf8cab7 | 1058 | pud_t *pudp, pud; |
080dbb61 AK |
1059 | spinlock_t *ptl; |
1060 | struct page *page; | |
1061 | struct mm_struct *mm = vma->vm_mm; | |
1062 | ||
caf8cab7 PX |
1063 | pudp = pud_offset(p4dp, address); |
1064 | pud = READ_ONCE(*pudp); | |
1b167618 | 1065 | if (!pud_present(pud)) |
878b0c45 | 1066 | return no_page_table(vma, flags, address); |
a12083d7 PX |
1067 | if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) |
1068 | return follow_hugepd(vma, __hugepd(pud_val(pud)), | |
1069 | address, PUD_SHIFT, flags, ctx); | |
1b167618 | 1070 | if (pud_leaf(pud)) { |
caf8cab7 | 1071 | ptl = pud_lock(mm, pudp); |
1b167618 | 1072 | page = follow_huge_pud(vma, address, pudp, flags, ctx); |
080dbb61 AK |
1073 | spin_unlock(ptl); |
1074 | if (page) | |
1075 | return page; | |
878b0c45 | 1076 | return no_page_table(vma, flags, address); |
080dbb61 | 1077 | } |
caf8cab7 | 1078 | if (unlikely(pud_bad(pud))) |
878b0c45 | 1079 | return no_page_table(vma, flags, address); |
080dbb61 | 1080 | |
caf8cab7 | 1081 | return follow_pmd_mask(vma, address, pudp, flags, ctx); |
080dbb61 AK |
1082 | } |
1083 | ||
080dbb61 AK |
1084 | static struct page *follow_p4d_mask(struct vm_area_struct *vma, |
1085 | unsigned long address, pgd_t *pgdp, | |
df06b37f KB |
1086 | unsigned int flags, |
1087 | struct follow_page_context *ctx) | |
080dbb61 | 1088 | { |
e6fd5564 | 1089 | p4d_t *p4dp, p4d; |
080dbb61 | 1090 | |
e6fd5564 PX |
1091 | p4dp = p4d_offset(pgdp, address); |
1092 | p4d = READ_ONCE(*p4dp); | |
1965e933 | 1093 | BUILD_BUG_ON(p4d_leaf(p4d)); |
a12083d7 PX |
1094 | |
1095 | if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) | |
1096 | return follow_hugepd(vma, __hugepd(p4d_val(p4d)), | |
1097 | address, P4D_SHIFT, flags, ctx); | |
1098 | ||
1099 | if (!p4d_present(p4d) || p4d_bad(p4d)) | |
878b0c45 | 1100 | return no_page_table(vma, flags, address); |
080dbb61 | 1101 | |
e6fd5564 | 1102 | return follow_pud_mask(vma, address, p4dp, flags, ctx); |
080dbb61 AK |
1103 | } |
1104 | ||
1105 | /** | |
1106 | * follow_page_mask - look up a page descriptor from a user-virtual address | |
1107 | * @vma: vm_area_struct mapping @address | |
1108 | * @address: virtual address to look up | |
1109 | * @flags: flags modifying lookup behaviour | |
78179556 MR |
1110 | * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a |
1111 | * pointer to output page_mask | |
080dbb61 AK |
1112 | * |
1113 | * @flags can have FOLL_ flags set, defined in <linux/mm.h> | |
1114 | * | |
78179556 MR |
1115 | * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches |
1116 | * the device's dev_pagemap metadata to avoid repeating expensive lookups. | |
1117 | * | |
a7f22660 DH |
1118 | * When getting an anonymous page and the caller has to trigger unsharing |
1119 | * of a shared anonymous page first, -EMLINK is returned. The caller should | |
1120 | * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only | |
1121 | * relevant with FOLL_PIN and !FOLL_WRITE. | |
1122 | * | |
78179556 MR |
1123 | * On output, the @ctx->page_mask is set according to the size of the page. |
1124 | * | |
1125 | * Return: the mapped (struct page *), %NULL if no mapping exists, or | |
080dbb61 AK |
1126 | * an error pointer if there is a mapping to something not represented |
1127 | * by a page descriptor (see also vm_normal_page()). | |
1128 | */ | |
a7030aea | 1129 | static struct page *follow_page_mask(struct vm_area_struct *vma, |
080dbb61 | 1130 | unsigned long address, unsigned int flags, |
df06b37f | 1131 | struct follow_page_context *ctx) |
080dbb61 AK |
1132 | { |
1133 | pgd_t *pgd; | |
080dbb61 | 1134 | struct mm_struct *mm = vma->vm_mm; |
9cb28da5 | 1135 | struct page *page; |
080dbb61 | 1136 | |
9cb28da5 | 1137 | vma_pgtable_walk_begin(vma); |
080dbb61 | 1138 | |
9cb28da5 | 1139 | ctx->page_mask = 0; |
080dbb61 AK |
1140 | pgd = pgd_offset(mm, address); |
1141 | ||
a12083d7 PX |
1142 | if (unlikely(is_hugepd(__hugepd(pgd_val(*pgd))))) |
1143 | page = follow_hugepd(vma, __hugepd(pgd_val(*pgd)), | |
1144 | address, PGDIR_SHIFT, flags, ctx); | |
1145 | else if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) | |
1146 | page = no_page_table(vma, flags, address); | |
1147 | else | |
1148 | page = follow_p4d_mask(vma, address, pgd, flags, ctx); | |
080dbb61 | 1149 | |
9cb28da5 PX |
1150 | vma_pgtable_walk_end(vma); |
1151 | ||
a12083d7 | 1152 | return page; |
df06b37f KB |
1153 | } |
1154 | ||
1155 | struct page *follow_page(struct vm_area_struct *vma, unsigned long address, | |
1156 | unsigned int foll_flags) | |
1157 | { | |
1158 | struct follow_page_context ctx = { NULL }; | |
1159 | struct page *page; | |
1160 | ||
1507f512 MR |
1161 | if (vma_is_secretmem(vma)) |
1162 | return NULL; | |
1163 | ||
d64e2dbc | 1164 | if (WARN_ON_ONCE(foll_flags & FOLL_PIN)) |
8909691b DH |
1165 | return NULL; |
1166 | ||
d74943a2 DH |
1167 | /* |
1168 | * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect | |
1169 | * to fail on PROT_NONE-mapped pages. | |
1170 | */ | |
df06b37f KB |
1171 | page = follow_page_mask(vma, address, foll_flags, &ctx); |
1172 | if (ctx.pgmap) | |
1173 | put_dev_pagemap(ctx.pgmap); | |
1174 | return page; | |
080dbb61 AK |
1175 | } |
1176 | ||
f2b495ca KS |
1177 | static int get_gate_page(struct mm_struct *mm, unsigned long address, |
1178 | unsigned int gup_flags, struct vm_area_struct **vma, | |
1179 | struct page **page) | |
1180 | { | |
1181 | pgd_t *pgd; | |
c2febafc | 1182 | p4d_t *p4d; |
f2b495ca KS |
1183 | pud_t *pud; |
1184 | pmd_t *pmd; | |
1185 | pte_t *pte; | |
c33c7948 | 1186 | pte_t entry; |
f2b495ca KS |
1187 | int ret = -EFAULT; |
1188 | ||
1189 | /* user gate pages are read-only */ | |
1190 | if (gup_flags & FOLL_WRITE) | |
1191 | return -EFAULT; | |
1192 | if (address > TASK_SIZE) | |
1193 | pgd = pgd_offset_k(address); | |
1194 | else | |
1195 | pgd = pgd_offset_gate(mm, address); | |
b5d1c39f AL |
1196 | if (pgd_none(*pgd)) |
1197 | return -EFAULT; | |
c2febafc | 1198 | p4d = p4d_offset(pgd, address); |
b5d1c39f AL |
1199 | if (p4d_none(*p4d)) |
1200 | return -EFAULT; | |
c2febafc | 1201 | pud = pud_offset(p4d, address); |
b5d1c39f AL |
1202 | if (pud_none(*pud)) |
1203 | return -EFAULT; | |
f2b495ca | 1204 | pmd = pmd_offset(pud, address); |
84c3fc4e | 1205 | if (!pmd_present(*pmd)) |
f2b495ca | 1206 | return -EFAULT; |
f2b495ca | 1207 | pte = pte_offset_map(pmd, address); |
04dee9e8 HD |
1208 | if (!pte) |
1209 | return -EFAULT; | |
c33c7948 RR |
1210 | entry = ptep_get(pte); |
1211 | if (pte_none(entry)) | |
f2b495ca KS |
1212 | goto unmap; |
1213 | *vma = get_gate_vma(mm); | |
1214 | if (!page) | |
1215 | goto out; | |
c33c7948 | 1216 | *page = vm_normal_page(*vma, address, entry); |
f2b495ca | 1217 | if (!*page) { |
c33c7948 | 1218 | if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry))) |
f2b495ca | 1219 | goto unmap; |
c33c7948 | 1220 | *page = pte_page(entry); |
f2b495ca | 1221 | } |
0f089235 LG |
1222 | ret = try_grab_page(*page, gup_flags); |
1223 | if (unlikely(ret)) | |
8fde12ca | 1224 | goto unmap; |
f2b495ca KS |
1225 | out: |
1226 | ret = 0; | |
1227 | unmap: | |
1228 | pte_unmap(pte); | |
1229 | return ret; | |
1230 | } | |
1231 | ||
9a95f3cf | 1232 | /* |
9a863a6a JG |
1233 | * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not |
1234 | * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set | |
1235 | * to 0 and -EBUSY returned. | |
9a95f3cf | 1236 | */ |
64019a2e | 1237 | static int faultin_page(struct vm_area_struct *vma, |
a7f22660 DH |
1238 | unsigned long address, unsigned int *flags, bool unshare, |
1239 | int *locked) | |
16744483 | 1240 | { |
16744483 | 1241 | unsigned int fault_flags = 0; |
2b740303 | 1242 | vm_fault_t ret; |
16744483 | 1243 | |
55b8fe70 AG |
1244 | if (*flags & FOLL_NOFAULT) |
1245 | return -EFAULT; | |
16744483 KS |
1246 | if (*flags & FOLL_WRITE) |
1247 | fault_flags |= FAULT_FLAG_WRITE; | |
1b2ee126 DH |
1248 | if (*flags & FOLL_REMOTE) |
1249 | fault_flags |= FAULT_FLAG_REMOTE; | |
f04740f5 | 1250 | if (*flags & FOLL_UNLOCKABLE) { |
71335f37 | 1251 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; |
93c5c61d PX |
1252 | /* |
1253 | * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set | |
1254 | * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE. | |
1255 | * That's because some callers may not be prepared to | |
1256 | * handle early exits caused by non-fatal signals. | |
1257 | */ | |
1258 | if (*flags & FOLL_INTERRUPTIBLE) | |
1259 | fault_flags |= FAULT_FLAG_INTERRUPTIBLE; | |
1260 | } | |
16744483 KS |
1261 | if (*flags & FOLL_NOWAIT) |
1262 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT; | |
234b239b | 1263 | if (*flags & FOLL_TRIED) { |
4426e945 PX |
1264 | /* |
1265 | * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED | |
1266 | * can co-exist | |
1267 | */ | |
234b239b ALC |
1268 | fault_flags |= FAULT_FLAG_TRIED; |
1269 | } | |
a7f22660 DH |
1270 | if (unshare) { |
1271 | fault_flags |= FAULT_FLAG_UNSHARE; | |
1272 | /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */ | |
1273 | VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE); | |
1274 | } | |
16744483 | 1275 | |
bce617ed | 1276 | ret = handle_mm_fault(vma, address, fault_flags, NULL); |
d9272525 PX |
1277 | |
1278 | if (ret & VM_FAULT_COMPLETED) { | |
1279 | /* | |
1280 | * With FAULT_FLAG_RETRY_NOWAIT we'll never release the | |
1281 | * mmap lock in the page fault handler. Sanity check this. | |
1282 | */ | |
1283 | WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT); | |
9a863a6a JG |
1284 | *locked = 0; |
1285 | ||
d9272525 PX |
1286 | /* |
1287 | * We should do the same as VM_FAULT_RETRY, but let's not | |
1288 | * return -EBUSY since that's not reflecting the reality of | |
1289 | * what has happened - we've just fully completed a page | |
1290 | * fault, with the mmap lock released. Use -EAGAIN to show | |
1291 | * that we want to take the mmap lock _again_. | |
1292 | */ | |
1293 | return -EAGAIN; | |
1294 | } | |
1295 | ||
16744483 | 1296 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
1297 | int err = vm_fault_to_errno(ret, *flags); |
1298 | ||
1299 | if (err) | |
1300 | return err; | |
16744483 KS |
1301 | BUG(); |
1302 | } | |
1303 | ||
16744483 | 1304 | if (ret & VM_FAULT_RETRY) { |
9a863a6a | 1305 | if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 1306 | *locked = 0; |
16744483 KS |
1307 | return -EBUSY; |
1308 | } | |
1309 | ||
16744483 KS |
1310 | return 0; |
1311 | } | |
1312 | ||
8ac26843 LS |
1313 | /* |
1314 | * Writing to file-backed mappings which require folio dirty tracking using GUP | |
1315 | * is a fundamentally broken operation, as kernel write access to GUP mappings | |
1316 | * do not adhere to the semantics expected by a file system. | |
1317 | * | |
1318 | * Consider the following scenario:- | |
1319 | * | |
1320 | * 1. A folio is written to via GUP which write-faults the memory, notifying | |
1321 | * the file system and dirtying the folio. | |
1322 | * 2. Later, writeback is triggered, resulting in the folio being cleaned and | |
1323 | * the PTE being marked read-only. | |
1324 | * 3. The GUP caller writes to the folio, as it is mapped read/write via the | |
1325 | * direct mapping. | |
1326 | * 4. The GUP caller, now done with the page, unpins it and sets it dirty | |
1327 | * (though it does not have to). | |
1328 | * | |
1329 | * This results in both data being written to a folio without writenotify, and | |
1330 | * the folio being dirtied unexpectedly (if the caller decides to do so). | |
1331 | */ | |
1332 | static bool writable_file_mapping_allowed(struct vm_area_struct *vma, | |
1333 | unsigned long gup_flags) | |
1334 | { | |
1335 | /* | |
1336 | * If we aren't pinning then no problematic write can occur. A long term | |
1337 | * pin is the most egregious case so this is the case we disallow. | |
1338 | */ | |
1339 | if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) != | |
1340 | (FOLL_PIN | FOLL_LONGTERM)) | |
1341 | return true; | |
1342 | ||
1343 | /* | |
1344 | * If the VMA does not require dirty tracking then no problematic write | |
1345 | * can occur either. | |
1346 | */ | |
1347 | return !vma_needs_dirty_tracking(vma); | |
1348 | } | |
1349 | ||
fa5bb209 KS |
1350 | static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags) |
1351 | { | |
1352 | vm_flags_t vm_flags = vma->vm_flags; | |
1b2ee126 DH |
1353 | int write = (gup_flags & FOLL_WRITE); |
1354 | int foreign = (gup_flags & FOLL_REMOTE); | |
8ac26843 | 1355 | bool vma_anon = vma_is_anonymous(vma); |
fa5bb209 KS |
1356 | |
1357 | if (vm_flags & (VM_IO | VM_PFNMAP)) | |
1358 | return -EFAULT; | |
1359 | ||
8ac26843 | 1360 | if ((gup_flags & FOLL_ANON) && !vma_anon) |
7f7ccc2c WT |
1361 | return -EFAULT; |
1362 | ||
52650c8b JG |
1363 | if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma)) |
1364 | return -EOPNOTSUPP; | |
1365 | ||
1507f512 MR |
1366 | if (vma_is_secretmem(vma)) |
1367 | return -EFAULT; | |
1368 | ||
1b2ee126 | 1369 | if (write) { |
8ac26843 LS |
1370 | if (!vma_anon && |
1371 | !writable_file_mapping_allowed(vma, gup_flags)) | |
1372 | return -EFAULT; | |
1373 | ||
6beb9958 | 1374 | if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) { |
fa5bb209 KS |
1375 | if (!(gup_flags & FOLL_FORCE)) |
1376 | return -EFAULT; | |
f347454d DH |
1377 | /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */ |
1378 | if (is_vm_hugetlb_page(vma)) | |
1379 | return -EFAULT; | |
fa5bb209 KS |
1380 | /* |
1381 | * We used to let the write,force case do COW in a | |
1382 | * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could | |
1383 | * set a breakpoint in a read-only mapping of an | |
1384 | * executable, without corrupting the file (yet only | |
1385 | * when that file had been opened for writing!). | |
1386 | * Anon pages in shared mappings are surprising: now | |
1387 | * just reject it. | |
1388 | */ | |
46435364 | 1389 | if (!is_cow_mapping(vm_flags)) |
fa5bb209 | 1390 | return -EFAULT; |
fa5bb209 KS |
1391 | } |
1392 | } else if (!(vm_flags & VM_READ)) { | |
1393 | if (!(gup_flags & FOLL_FORCE)) | |
1394 | return -EFAULT; | |
1395 | /* | |
1396 | * Is there actually any vma we can reach here which does not | |
1397 | * have VM_MAYREAD set? | |
1398 | */ | |
1399 | if (!(vm_flags & VM_MAYREAD)) | |
1400 | return -EFAULT; | |
1401 | } | |
d61172b4 DH |
1402 | /* |
1403 | * gups are always data accesses, not instruction | |
1404 | * fetches, so execute=false here | |
1405 | */ | |
1406 | if (!arch_vma_access_permitted(vma, write, false, foreign)) | |
33a709b2 | 1407 | return -EFAULT; |
fa5bb209 KS |
1408 | return 0; |
1409 | } | |
1410 | ||
6cd06ab1 LT |
1411 | /* |
1412 | * This is "vma_lookup()", but with a warning if we would have | |
1413 | * historically expanded the stack in the GUP code. | |
1414 | */ | |
1415 | static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm, | |
1416 | unsigned long addr) | |
1417 | { | |
1418 | #ifdef CONFIG_STACK_GROWSUP | |
1419 | return vma_lookup(mm, addr); | |
1420 | #else | |
1421 | static volatile unsigned long next_warn; | |
1422 | struct vm_area_struct *vma; | |
1423 | unsigned long now, next; | |
1424 | ||
1425 | vma = find_vma(mm, addr); | |
1426 | if (!vma || (addr >= vma->vm_start)) | |
1427 | return vma; | |
1428 | ||
1429 | /* Only warn for half-way relevant accesses */ | |
1430 | if (!(vma->vm_flags & VM_GROWSDOWN)) | |
1431 | return NULL; | |
1432 | if (vma->vm_start - addr > 65536) | |
1433 | return NULL; | |
1434 | ||
1435 | /* Let's not warn more than once an hour.. */ | |
1436 | now = jiffies; next = next_warn; | |
1437 | if (next && time_before(now, next)) | |
1438 | return NULL; | |
1439 | next_warn = now + 60*60*HZ; | |
1440 | ||
1441 | /* Let people know things may have changed. */ | |
1442 | pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n", | |
1443 | current->comm, task_pid_nr(current), | |
1444 | vma->vm_start, vma->vm_end, addr); | |
1445 | dump_stack(); | |
1446 | return NULL; | |
1447 | #endif | |
1448 | } | |
1449 | ||
4bbd4c77 KS |
1450 | /** |
1451 | * __get_user_pages() - pin user pages in memory | |
4bbd4c77 KS |
1452 | * @mm: mm_struct of target mm |
1453 | * @start: starting user address | |
1454 | * @nr_pages: number of pages from start to pin | |
1455 | * @gup_flags: flags modifying pin behaviour | |
1456 | * @pages: array that receives pointers to the pages pinned. | |
1457 | * Should be at least nr_pages long. Or NULL, if caller | |
1458 | * only intends to ensure the pages are faulted in. | |
c1e8d7c6 | 1459 | * @locked: whether we're still with the mmap_lock held |
4bbd4c77 | 1460 | * |
d2dfbe47 LX |
1461 | * Returns either number of pages pinned (which may be less than the |
1462 | * number requested), or an error. Details about the return value: | |
1463 | * | |
1464 | * -- If nr_pages is 0, returns 0. | |
1465 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
1466 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
1467 | * pages pinned. Again, this may be less than nr_pages. | |
2d3a36a4 | 1468 | * -- 0 return value is possible when the fault would need to be retried. |
d2dfbe47 LX |
1469 | * |
1470 | * The caller is responsible for releasing returned @pages, via put_page(). | |
1471 | * | |
c1e8d7c6 | 1472 | * Must be called with mmap_lock held. It may be released. See below. |
4bbd4c77 KS |
1473 | * |
1474 | * __get_user_pages walks a process's page tables and takes a reference to | |
1475 | * each struct page that each user address corresponds to at a given | |
1476 | * instant. That is, it takes the page that would be accessed if a user | |
1477 | * thread accesses the given user virtual address at that instant. | |
1478 | * | |
1479 | * This does not guarantee that the page exists in the user mappings when | |
1480 | * __get_user_pages returns, and there may even be a completely different | |
1481 | * page there in some cases (eg. if mmapped pagecache has been invalidated | |
c5acf1f6 | 1482 | * and subsequently re-faulted). However it does guarantee that the page |
4bbd4c77 KS |
1483 | * won't be freed completely. And mostly callers simply care that the page |
1484 | * contains data that was valid *at some point in time*. Typically, an IO | |
1485 | * or similar operation cannot guarantee anything stronger anyway because | |
1486 | * locks can't be held over the syscall boundary. | |
1487 | * | |
1488 | * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If | |
1489 | * the page is written to, set_page_dirty (or set_page_dirty_lock, as | |
1490 | * appropriate) must be called after the page is finished with, and | |
1491 | * before put_page is called. | |
1492 | * | |
9a863a6a JG |
1493 | * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may |
1494 | * be released. If this happens *@locked will be set to 0 on return. | |
9a95f3cf | 1495 | * |
9a863a6a JG |
1496 | * A caller using such a combination of @gup_flags must therefore hold the |
1497 | * mmap_lock for reading only, and recognize when it's been released. Otherwise, | |
1498 | * it must be held for either reading or writing and will not be released. | |
4bbd4c77 KS |
1499 | * |
1500 | * In most cases, get_user_pages or get_user_pages_fast should be used | |
1501 | * instead of __get_user_pages. __get_user_pages should be used only if | |
1502 | * you need some special @gup_flags. | |
1503 | */ | |
64019a2e | 1504 | static long __get_user_pages(struct mm_struct *mm, |
4bbd4c77 KS |
1505 | unsigned long start, unsigned long nr_pages, |
1506 | unsigned int gup_flags, struct page **pages, | |
b2cac248 | 1507 | int *locked) |
4bbd4c77 | 1508 | { |
df06b37f | 1509 | long ret = 0, i = 0; |
fa5bb209 | 1510 | struct vm_area_struct *vma = NULL; |
df06b37f | 1511 | struct follow_page_context ctx = { NULL }; |
4bbd4c77 KS |
1512 | |
1513 | if (!nr_pages) | |
1514 | return 0; | |
1515 | ||
428e106a | 1516 | start = untagged_addr_remote(mm, start); |
f9652594 | 1517 | |
eddb1c22 | 1518 | VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN))); |
4bbd4c77 | 1519 | |
4bbd4c77 | 1520 | do { |
fa5bb209 KS |
1521 | struct page *page; |
1522 | unsigned int foll_flags = gup_flags; | |
1523 | unsigned int page_increm; | |
1524 | ||
1525 | /* first iteration or cross vma bound */ | |
1526 | if (!vma || start >= vma->vm_end) { | |
631426ba DH |
1527 | /* |
1528 | * MADV_POPULATE_(READ|WRITE) wants to handle VMA | |
1529 | * lookups+error reporting differently. | |
1530 | */ | |
1531 | if (gup_flags & FOLL_MADV_POPULATE) { | |
1532 | vma = vma_lookup(mm, start); | |
1533 | if (!vma) { | |
1534 | ret = -ENOMEM; | |
1535 | goto out; | |
1536 | } | |
1537 | if (check_vma_flags(vma, gup_flags)) { | |
1538 | ret = -EINVAL; | |
1539 | goto out; | |
1540 | } | |
1541 | goto retry; | |
1542 | } | |
6cd06ab1 | 1543 | vma = gup_vma_lookup(mm, start); |
fa5bb209 | 1544 | if (!vma && in_gate_area(mm, start)) { |
fa5bb209 KS |
1545 | ret = get_gate_page(mm, start & PAGE_MASK, |
1546 | gup_flags, &vma, | |
ffe1e786 | 1547 | pages ? &page : NULL); |
fa5bb209 | 1548 | if (ret) |
08be37b7 | 1549 | goto out; |
df06b37f | 1550 | ctx.page_mask = 0; |
fa5bb209 KS |
1551 | goto next_page; |
1552 | } | |
4bbd4c77 | 1553 | |
52650c8b | 1554 | if (!vma) { |
df06b37f KB |
1555 | ret = -EFAULT; |
1556 | goto out; | |
1557 | } | |
52650c8b JG |
1558 | ret = check_vma_flags(vma, gup_flags); |
1559 | if (ret) | |
1560 | goto out; | |
fa5bb209 KS |
1561 | } |
1562 | retry: | |
1563 | /* | |
1564 | * If we have a pending SIGKILL, don't keep faulting pages and | |
1565 | * potentially allocating memory. | |
1566 | */ | |
fa45f116 | 1567 | if (fatal_signal_pending(current)) { |
d180870d | 1568 | ret = -EINTR; |
df06b37f KB |
1569 | goto out; |
1570 | } | |
fa5bb209 | 1571 | cond_resched(); |
df06b37f KB |
1572 | |
1573 | page = follow_page_mask(vma, start, foll_flags, &ctx); | |
a7f22660 DH |
1574 | if (!page || PTR_ERR(page) == -EMLINK) { |
1575 | ret = faultin_page(vma, start, &foll_flags, | |
1576 | PTR_ERR(page) == -EMLINK, locked); | |
fa5bb209 KS |
1577 | switch (ret) { |
1578 | case 0: | |
1579 | goto retry; | |
df06b37f | 1580 | case -EBUSY: |
d9272525 | 1581 | case -EAGAIN: |
df06b37f | 1582 | ret = 0; |
e4a9bc58 | 1583 | fallthrough; |
fa5bb209 KS |
1584 | case -EFAULT: |
1585 | case -ENOMEM: | |
1586 | case -EHWPOISON: | |
df06b37f | 1587 | goto out; |
4bbd4c77 | 1588 | } |
fa5bb209 | 1589 | BUG(); |
1027e443 KS |
1590 | } else if (PTR_ERR(page) == -EEXIST) { |
1591 | /* | |
1592 | * Proper page table entry exists, but no corresponding | |
65462462 JH |
1593 | * struct page. If the caller expects **pages to be |
1594 | * filled in, bail out now, because that can't be done | |
1595 | * for this page. | |
1027e443 | 1596 | */ |
65462462 JH |
1597 | if (pages) { |
1598 | ret = PTR_ERR(page); | |
1599 | goto out; | |
1600 | } | |
1027e443 | 1601 | } else if (IS_ERR(page)) { |
df06b37f KB |
1602 | ret = PTR_ERR(page); |
1603 | goto out; | |
1027e443 | 1604 | } |
ffe1e786 | 1605 | next_page: |
df06b37f | 1606 | page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); |
fa5bb209 KS |
1607 | if (page_increm > nr_pages) |
1608 | page_increm = nr_pages; | |
57edfcfd PX |
1609 | |
1610 | if (pages) { | |
1611 | struct page *subpage; | |
1612 | unsigned int j; | |
1613 | ||
1614 | /* | |
1615 | * This must be a large folio (and doesn't need to | |
1616 | * be the whole folio; it can be part of it), do | |
1617 | * the refcount work for all the subpages too. | |
1618 | * | |
1619 | * NOTE: here the page may not be the head page | |
1620 | * e.g. when start addr is not thp-size aligned. | |
1621 | * try_grab_folio() should have taken care of tail | |
1622 | * pages. | |
1623 | */ | |
1624 | if (page_increm > 1) { | |
1625 | struct folio *folio; | |
1626 | ||
1627 | /* | |
1628 | * Since we already hold refcount on the | |
1629 | * large folio, this should never fail. | |
1630 | */ | |
1631 | folio = try_grab_folio(page, page_increm - 1, | |
1632 | foll_flags); | |
1633 | if (WARN_ON_ONCE(!folio)) { | |
1634 | /* | |
1635 | * Release the 1st page ref if the | |
1636 | * folio is problematic, fail hard. | |
1637 | */ | |
1638 | gup_put_folio(page_folio(page), 1, | |
1639 | foll_flags); | |
1640 | ret = -EFAULT; | |
1641 | goto out; | |
1642 | } | |
1643 | } | |
1644 | ||
1645 | for (j = 0; j < page_increm; j++) { | |
1646 | subpage = nth_page(page, j); | |
1647 | pages[i + j] = subpage; | |
1648 | flush_anon_page(vma, subpage, start + j * PAGE_SIZE); | |
1649 | flush_dcache_page(subpage); | |
1650 | } | |
1651 | } | |
1652 | ||
fa5bb209 KS |
1653 | i += page_increm; |
1654 | start += page_increm * PAGE_SIZE; | |
1655 | nr_pages -= page_increm; | |
4bbd4c77 | 1656 | } while (nr_pages); |
df06b37f KB |
1657 | out: |
1658 | if (ctx.pgmap) | |
1659 | put_dev_pagemap(ctx.pgmap); | |
1660 | return i ? i : ret; | |
4bbd4c77 | 1661 | } |
4bbd4c77 | 1662 | |
771ab430 TK |
1663 | static bool vma_permits_fault(struct vm_area_struct *vma, |
1664 | unsigned int fault_flags) | |
d4925e00 | 1665 | { |
1b2ee126 DH |
1666 | bool write = !!(fault_flags & FAULT_FLAG_WRITE); |
1667 | bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE); | |
33a709b2 | 1668 | vm_flags_t vm_flags = write ? VM_WRITE : VM_READ; |
d4925e00 DH |
1669 | |
1670 | if (!(vm_flags & vma->vm_flags)) | |
1671 | return false; | |
1672 | ||
33a709b2 DH |
1673 | /* |
1674 | * The architecture might have a hardware protection | |
1b2ee126 | 1675 | * mechanism other than read/write that can deny access. |
d61172b4 DH |
1676 | * |
1677 | * gup always represents data access, not instruction | |
1678 | * fetches, so execute=false here: | |
33a709b2 | 1679 | */ |
d61172b4 | 1680 | if (!arch_vma_access_permitted(vma, write, false, foreign)) |
33a709b2 DH |
1681 | return false; |
1682 | ||
d4925e00 DH |
1683 | return true; |
1684 | } | |
1685 | ||
adc8cb40 | 1686 | /** |
4bbd4c77 | 1687 | * fixup_user_fault() - manually resolve a user page fault |
4bbd4c77 KS |
1688 | * @mm: mm_struct of target mm |
1689 | * @address: user address | |
1690 | * @fault_flags:flags to pass down to handle_mm_fault() | |
c1e8d7c6 | 1691 | * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller |
548b6a1e MC |
1692 | * does not allow retry. If NULL, the caller must guarantee |
1693 | * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY. | |
4bbd4c77 KS |
1694 | * |
1695 | * This is meant to be called in the specific scenario where for locking reasons | |
1696 | * we try to access user memory in atomic context (within a pagefault_disable() | |
1697 | * section), this returns -EFAULT, and we want to resolve the user fault before | |
1698 | * trying again. | |
1699 | * | |
1700 | * Typically this is meant to be used by the futex code. | |
1701 | * | |
1702 | * The main difference with get_user_pages() is that this function will | |
1703 | * unconditionally call handle_mm_fault() which will in turn perform all the | |
1704 | * necessary SW fixup of the dirty and young bits in the PTE, while | |
4a9e1cda | 1705 | * get_user_pages() only guarantees to update these in the struct page. |
4bbd4c77 KS |
1706 | * |
1707 | * This is important for some architectures where those bits also gate the | |
1708 | * access permission to the page because they are maintained in software. On | |
1709 | * such architectures, gup() will not be enough to make a subsequent access | |
1710 | * succeed. | |
1711 | * | |
c1e8d7c6 ML |
1712 | * This function will not return with an unlocked mmap_lock. So it has not the |
1713 | * same semantics wrt the @mm->mmap_lock as does filemap_fault(). | |
4bbd4c77 | 1714 | */ |
64019a2e | 1715 | int fixup_user_fault(struct mm_struct *mm, |
4a9e1cda DD |
1716 | unsigned long address, unsigned int fault_flags, |
1717 | bool *unlocked) | |
4bbd4c77 KS |
1718 | { |
1719 | struct vm_area_struct *vma; | |
8fed2f3c | 1720 | vm_fault_t ret; |
4a9e1cda | 1721 | |
428e106a | 1722 | address = untagged_addr_remote(mm, address); |
f9652594 | 1723 | |
4a9e1cda | 1724 | if (unlocked) |
71335f37 | 1725 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; |
4bbd4c77 | 1726 | |
4a9e1cda | 1727 | retry: |
6cd06ab1 | 1728 | vma = gup_vma_lookup(mm, address); |
8d7071af | 1729 | if (!vma) |
4bbd4c77 KS |
1730 | return -EFAULT; |
1731 | ||
d4925e00 | 1732 | if (!vma_permits_fault(vma, fault_flags)) |
4bbd4c77 KS |
1733 | return -EFAULT; |
1734 | ||
475f4dfc PX |
1735 | if ((fault_flags & FAULT_FLAG_KILLABLE) && |
1736 | fatal_signal_pending(current)) | |
1737 | return -EINTR; | |
1738 | ||
bce617ed | 1739 | ret = handle_mm_fault(vma, address, fault_flags, NULL); |
d9272525 PX |
1740 | |
1741 | if (ret & VM_FAULT_COMPLETED) { | |
1742 | /* | |
1743 | * NOTE: it's a pity that we need to retake the lock here | |
1744 | * to pair with the unlock() in the callers. Ideally we | |
1745 | * could tell the callers so they do not need to unlock. | |
1746 | */ | |
1747 | mmap_read_lock(mm); | |
1748 | *unlocked = true; | |
1749 | return 0; | |
1750 | } | |
1751 | ||
4bbd4c77 | 1752 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
1753 | int err = vm_fault_to_errno(ret, 0); |
1754 | ||
1755 | if (err) | |
1756 | return err; | |
4bbd4c77 KS |
1757 | BUG(); |
1758 | } | |
4a9e1cda DD |
1759 | |
1760 | if (ret & VM_FAULT_RETRY) { | |
d8ed45c5 | 1761 | mmap_read_lock(mm); |
475f4dfc PX |
1762 | *unlocked = true; |
1763 | fault_flags |= FAULT_FLAG_TRIED; | |
1764 | goto retry; | |
4a9e1cda DD |
1765 | } |
1766 | ||
4bbd4c77 KS |
1767 | return 0; |
1768 | } | |
add6a0cd | 1769 | EXPORT_SYMBOL_GPL(fixup_user_fault); |
4bbd4c77 | 1770 | |
93c5c61d PX |
1771 | /* |
1772 | * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is | |
1773 | * specified, it'll also respond to generic signals. The caller of GUP | |
1774 | * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption. | |
1775 | */ | |
1776 | static bool gup_signal_pending(unsigned int flags) | |
1777 | { | |
1778 | if (fatal_signal_pending(current)) | |
1779 | return true; | |
1780 | ||
1781 | if (!(flags & FOLL_INTERRUPTIBLE)) | |
1782 | return false; | |
1783 | ||
1784 | return signal_pending(current); | |
1785 | } | |
1786 | ||
2d3a36a4 | 1787 | /* |
b2a72dff JG |
1788 | * Locking: (*locked == 1) means that the mmap_lock has already been acquired by |
1789 | * the caller. This function may drop the mmap_lock. If it does so, then it will | |
1790 | * set (*locked = 0). | |
1791 | * | |
1792 | * (*locked == 0) means that the caller expects this function to acquire and | |
1793 | * drop the mmap_lock. Therefore, the value of *locked will still be zero when | |
1794 | * the function returns, even though it may have changed temporarily during | |
1795 | * function execution. | |
1796 | * | |
1797 | * Please note that this function, unlike __get_user_pages(), will not return 0 | |
1798 | * for nr_pages > 0, unless FOLL_NOWAIT is used. | |
2d3a36a4 | 1799 | */ |
64019a2e | 1800 | static __always_inline long __get_user_pages_locked(struct mm_struct *mm, |
f0818f47 AA |
1801 | unsigned long start, |
1802 | unsigned long nr_pages, | |
f0818f47 | 1803 | struct page **pages, |
e716712f | 1804 | int *locked, |
0fd71a56 | 1805 | unsigned int flags) |
f0818f47 | 1806 | { |
f0818f47 | 1807 | long ret, pages_done; |
b2a72dff | 1808 | bool must_unlock = false; |
f0818f47 | 1809 | |
9c4b2142 LS |
1810 | if (!nr_pages) |
1811 | return 0; | |
1812 | ||
b2a72dff JG |
1813 | /* |
1814 | * The internal caller expects GUP to manage the lock internally and the | |
1815 | * lock must be released when this returns. | |
1816 | */ | |
9a863a6a | 1817 | if (!*locked) { |
b2a72dff JG |
1818 | if (mmap_read_lock_killable(mm)) |
1819 | return -EAGAIN; | |
1820 | must_unlock = true; | |
1821 | *locked = 1; | |
f0818f47 | 1822 | } |
961ba472 JG |
1823 | else |
1824 | mmap_assert_locked(mm); | |
f0818f47 | 1825 | |
a458b76a AA |
1826 | if (flags & FOLL_PIN) |
1827 | mm_set_has_pinned_flag(&mm->flags); | |
008cfe44 | 1828 | |
eddb1c22 JH |
1829 | /* |
1830 | * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior | |
1831 | * is to set FOLL_GET if the caller wants pages[] filled in (but has | |
1832 | * carelessly failed to specify FOLL_GET), so keep doing that, but only | |
1833 | * for FOLL_GET, not for the newer FOLL_PIN. | |
1834 | * | |
1835 | * FOLL_PIN always expects pages to be non-null, but no need to assert | |
1836 | * that here, as any failures will be obvious enough. | |
1837 | */ | |
1838 | if (pages && !(flags & FOLL_PIN)) | |
f0818f47 | 1839 | flags |= FOLL_GET; |
f0818f47 AA |
1840 | |
1841 | pages_done = 0; | |
f0818f47 | 1842 | for (;;) { |
64019a2e | 1843 | ret = __get_user_pages(mm, start, nr_pages, flags, pages, |
b2cac248 | 1844 | locked); |
f04740f5 | 1845 | if (!(flags & FOLL_UNLOCKABLE)) { |
f0818f47 | 1846 | /* VM_FAULT_RETRY couldn't trigger, bypass */ |
f04740f5 JG |
1847 | pages_done = ret; |
1848 | break; | |
1849 | } | |
f0818f47 | 1850 | |
d9272525 | 1851 | /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */ |
f0818f47 AA |
1852 | if (!*locked) { |
1853 | BUG_ON(ret < 0); | |
1854 | BUG_ON(ret >= nr_pages); | |
1855 | } | |
1856 | ||
f0818f47 AA |
1857 | if (ret > 0) { |
1858 | nr_pages -= ret; | |
1859 | pages_done += ret; | |
1860 | if (!nr_pages) | |
1861 | break; | |
1862 | } | |
1863 | if (*locked) { | |
96312e61 AA |
1864 | /* |
1865 | * VM_FAULT_RETRY didn't trigger or it was a | |
1866 | * FOLL_NOWAIT. | |
1867 | */ | |
f0818f47 AA |
1868 | if (!pages_done) |
1869 | pages_done = ret; | |
1870 | break; | |
1871 | } | |
df17277b MR |
1872 | /* |
1873 | * VM_FAULT_RETRY triggered, so seek to the faulting offset. | |
1874 | * For the prefault case (!pages) we only update counts. | |
1875 | */ | |
1876 | if (likely(pages)) | |
1877 | pages += ret; | |
f0818f47 | 1878 | start += ret << PAGE_SHIFT; |
b2a72dff JG |
1879 | |
1880 | /* The lock was temporarily dropped, so we must unlock later */ | |
1881 | must_unlock = true; | |
f0818f47 | 1882 | |
4426e945 | 1883 | retry: |
f0818f47 AA |
1884 | /* |
1885 | * Repeat on the address that fired VM_FAULT_RETRY | |
4426e945 PX |
1886 | * with both FAULT_FLAG_ALLOW_RETRY and |
1887 | * FAULT_FLAG_TRIED. Note that GUP can be interrupted | |
93c5c61d PX |
1888 | * by fatal signals of even common signals, depending on |
1889 | * the caller's request. So we need to check it before we | |
4426e945 | 1890 | * start trying again otherwise it can loop forever. |
f0818f47 | 1891 | */ |
93c5c61d | 1892 | if (gup_signal_pending(flags)) { |
ae46d2aa HD |
1893 | if (!pages_done) |
1894 | pages_done = -EINTR; | |
4426e945 | 1895 | break; |
ae46d2aa | 1896 | } |
4426e945 | 1897 | |
d8ed45c5 | 1898 | ret = mmap_read_lock_killable(mm); |
71335f37 PX |
1899 | if (ret) { |
1900 | BUG_ON(ret > 0); | |
1901 | if (!pages_done) | |
1902 | pages_done = ret; | |
1903 | break; | |
1904 | } | |
4426e945 | 1905 | |
c7b6a566 | 1906 | *locked = 1; |
64019a2e | 1907 | ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED, |
b2cac248 | 1908 | pages, locked); |
4426e945 PX |
1909 | if (!*locked) { |
1910 | /* Continue to retry until we succeeded */ | |
1911 | BUG_ON(ret != 0); | |
1912 | goto retry; | |
1913 | } | |
f0818f47 AA |
1914 | if (ret != 1) { |
1915 | BUG_ON(ret > 1); | |
1916 | if (!pages_done) | |
1917 | pages_done = ret; | |
1918 | break; | |
1919 | } | |
1920 | nr_pages--; | |
1921 | pages_done++; | |
1922 | if (!nr_pages) | |
1923 | break; | |
df17277b MR |
1924 | if (likely(pages)) |
1925 | pages++; | |
f0818f47 AA |
1926 | start += PAGE_SIZE; |
1927 | } | |
b2a72dff | 1928 | if (must_unlock && *locked) { |
f0818f47 | 1929 | /* |
b2a72dff JG |
1930 | * We either temporarily dropped the lock, or the caller |
1931 | * requested that we both acquire and drop the lock. Either way, | |
1932 | * we must now unlock, and notify the caller of that state. | |
f0818f47 | 1933 | */ |
d8ed45c5 | 1934 | mmap_read_unlock(mm); |
f0818f47 AA |
1935 | *locked = 0; |
1936 | } | |
9c4b2142 LS |
1937 | |
1938 | /* | |
1939 | * Failing to pin anything implies something has gone wrong (except when | |
1940 | * FOLL_NOWAIT is specified). | |
1941 | */ | |
1942 | if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT))) | |
1943 | return -EFAULT; | |
1944 | ||
f0818f47 AA |
1945 | return pages_done; |
1946 | } | |
1947 | ||
d3649f68 CH |
1948 | /** |
1949 | * populate_vma_page_range() - populate a range of pages in the vma. | |
1950 | * @vma: target vma | |
1951 | * @start: start address | |
1952 | * @end: end address | |
c1e8d7c6 | 1953 | * @locked: whether the mmap_lock is still held |
d3649f68 CH |
1954 | * |
1955 | * This takes care of mlocking the pages too if VM_LOCKED is set. | |
1956 | * | |
0a36f7f8 TY |
1957 | * Return either number of pages pinned in the vma, or a negative error |
1958 | * code on error. | |
d3649f68 | 1959 | * |
c1e8d7c6 | 1960 | * vma->vm_mm->mmap_lock must be held. |
d3649f68 | 1961 | * |
4f6da934 | 1962 | * If @locked is NULL, it may be held for read or write and will |
d3649f68 CH |
1963 | * be unperturbed. |
1964 | * | |
4f6da934 PX |
1965 | * If @locked is non-NULL, it must held for read only and may be |
1966 | * released. If it's released, *@locked will be set to 0. | |
d3649f68 CH |
1967 | */ |
1968 | long populate_vma_page_range(struct vm_area_struct *vma, | |
4f6da934 | 1969 | unsigned long start, unsigned long end, int *locked) |
d3649f68 CH |
1970 | { |
1971 | struct mm_struct *mm = vma->vm_mm; | |
1972 | unsigned long nr_pages = (end - start) / PAGE_SIZE; | |
9a863a6a | 1973 | int local_locked = 1; |
d3649f68 | 1974 | int gup_flags; |
ece369c7 | 1975 | long ret; |
d3649f68 | 1976 | |
be51eb18 ML |
1977 | VM_BUG_ON(!PAGE_ALIGNED(start)); |
1978 | VM_BUG_ON(!PAGE_ALIGNED(end)); | |
d3649f68 CH |
1979 | VM_BUG_ON_VMA(start < vma->vm_start, vma); |
1980 | VM_BUG_ON_VMA(end > vma->vm_end, vma); | |
42fc5414 | 1981 | mmap_assert_locked(mm); |
d3649f68 | 1982 | |
b67bf49c HD |
1983 | /* |
1984 | * Rightly or wrongly, the VM_LOCKONFAULT case has never used | |
1985 | * faultin_page() to break COW, so it has no work to do here. | |
1986 | */ | |
d3649f68 | 1987 | if (vma->vm_flags & VM_LOCKONFAULT) |
b67bf49c HD |
1988 | return nr_pages; |
1989 | ||
1096bc93 LT |
1990 | /* ... similarly, we've never faulted in PROT_NONE pages */ |
1991 | if (!vma_is_accessible(vma)) | |
1992 | return -EFAULT; | |
1993 | ||
b67bf49c | 1994 | gup_flags = FOLL_TOUCH; |
d3649f68 CH |
1995 | /* |
1996 | * We want to touch writable mappings with a write fault in order | |
1997 | * to break COW, except for shared mappings because these don't COW | |
1998 | * and we would not want to dirty them for nothing. | |
1096bc93 LT |
1999 | * |
2000 | * Otherwise, do a read fault, and use FOLL_FORCE in case it's not | |
2001 | * readable (ie write-only or executable). | |
d3649f68 CH |
2002 | */ |
2003 | if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE) | |
2004 | gup_flags |= FOLL_WRITE; | |
1096bc93 | 2005 | else |
d3649f68 CH |
2006 | gup_flags |= FOLL_FORCE; |
2007 | ||
f04740f5 JG |
2008 | if (locked) |
2009 | gup_flags |= FOLL_UNLOCKABLE; | |
2010 | ||
d3649f68 CH |
2011 | /* |
2012 | * We made sure addr is within a VMA, so the following will | |
2013 | * not result in a stack expansion that recurses back here. | |
2014 | */ | |
ece369c7 | 2015 | ret = __get_user_pages(mm, start, nr_pages, gup_flags, |
b2cac248 | 2016 | NULL, locked ? locked : &local_locked); |
ece369c7 HD |
2017 | lru_add_drain(); |
2018 | return ret; | |
d3649f68 CH |
2019 | } |
2020 | ||
4ca9b385 | 2021 | /* |
631426ba DH |
2022 | * faultin_page_range() - populate (prefault) page tables inside the |
2023 | * given range readable/writable | |
4ca9b385 DH |
2024 | * |
2025 | * This takes care of mlocking the pages, too, if VM_LOCKED is set. | |
2026 | * | |
631426ba | 2027 | * @mm: the mm to populate page tables in |
4ca9b385 DH |
2028 | * @start: start address |
2029 | * @end: end address | |
2030 | * @write: whether to prefault readable or writable | |
2031 | * @locked: whether the mmap_lock is still held | |
2032 | * | |
631426ba DH |
2033 | * Returns either number of processed pages in the MM, or a negative error |
2034 | * code on error (see __get_user_pages()). Note that this function reports | |
2035 | * errors related to VMAs, such as incompatible mappings, as expected by | |
2036 | * MADV_POPULATE_(READ|WRITE). | |
4ca9b385 | 2037 | * |
631426ba DH |
2038 | * The range must be page-aligned. |
2039 | * | |
2040 | * mm->mmap_lock must be held. If it's released, *@locked will be set to 0. | |
4ca9b385 | 2041 | */ |
631426ba DH |
2042 | long faultin_page_range(struct mm_struct *mm, unsigned long start, |
2043 | unsigned long end, bool write, int *locked) | |
4ca9b385 | 2044 | { |
4ca9b385 DH |
2045 | unsigned long nr_pages = (end - start) / PAGE_SIZE; |
2046 | int gup_flags; | |
ece369c7 | 2047 | long ret; |
4ca9b385 DH |
2048 | |
2049 | VM_BUG_ON(!PAGE_ALIGNED(start)); | |
2050 | VM_BUG_ON(!PAGE_ALIGNED(end)); | |
4ca9b385 DH |
2051 | mmap_assert_locked(mm); |
2052 | ||
2053 | /* | |
2054 | * FOLL_TOUCH: Mark page accessed and thereby young; will also mark | |
2055 | * the page dirty with FOLL_WRITE -- which doesn't make a | |
2056 | * difference with !FOLL_FORCE, because the page is writable | |
2057 | * in the page table. | |
2058 | * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit | |
2059 | * a poisoned page. | |
4ca9b385 DH |
2060 | * !FOLL_FORCE: Require proper access permissions. |
2061 | */ | |
631426ba DH |
2062 | gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE | |
2063 | FOLL_MADV_POPULATE; | |
4ca9b385 DH |
2064 | if (write) |
2065 | gup_flags |= FOLL_WRITE; | |
2066 | ||
631426ba DH |
2067 | ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked, |
2068 | gup_flags); | |
ece369c7 HD |
2069 | lru_add_drain(); |
2070 | return ret; | |
4ca9b385 DH |
2071 | } |
2072 | ||
d3649f68 CH |
2073 | /* |
2074 | * __mm_populate - populate and/or mlock pages within a range of address space. | |
2075 | * | |
2076 | * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap | |
2077 | * flags. VMAs must be already marked with the desired vm_flags, and | |
c1e8d7c6 | 2078 | * mmap_lock must not be held. |
d3649f68 CH |
2079 | */ |
2080 | int __mm_populate(unsigned long start, unsigned long len, int ignore_errors) | |
2081 | { | |
2082 | struct mm_struct *mm = current->mm; | |
2083 | unsigned long end, nstart, nend; | |
2084 | struct vm_area_struct *vma = NULL; | |
2085 | int locked = 0; | |
2086 | long ret = 0; | |
2087 | ||
2088 | end = start + len; | |
2089 | ||
2090 | for (nstart = start; nstart < end; nstart = nend) { | |
2091 | /* | |
2092 | * We want to fault in pages for [nstart; end) address range. | |
2093 | * Find first corresponding VMA. | |
2094 | */ | |
2095 | if (!locked) { | |
2096 | locked = 1; | |
d8ed45c5 | 2097 | mmap_read_lock(mm); |
c4d1a92d | 2098 | vma = find_vma_intersection(mm, nstart, end); |
d3649f68 | 2099 | } else if (nstart >= vma->vm_end) |
c4d1a92d LH |
2100 | vma = find_vma_intersection(mm, vma->vm_end, end); |
2101 | ||
2102 | if (!vma) | |
d3649f68 CH |
2103 | break; |
2104 | /* | |
2105 | * Set [nstart; nend) to intersection of desired address | |
2106 | * range with the first VMA. Also, skip undesirable VMA types. | |
2107 | */ | |
2108 | nend = min(end, vma->vm_end); | |
2109 | if (vma->vm_flags & (VM_IO | VM_PFNMAP)) | |
2110 | continue; | |
2111 | if (nstart < vma->vm_start) | |
2112 | nstart = vma->vm_start; | |
2113 | /* | |
2114 | * Now fault in a range of pages. populate_vma_page_range() | |
2115 | * double checks the vma flags, so that it won't mlock pages | |
2116 | * if the vma was already munlocked. | |
2117 | */ | |
2118 | ret = populate_vma_page_range(vma, nstart, nend, &locked); | |
2119 | if (ret < 0) { | |
2120 | if (ignore_errors) { | |
2121 | ret = 0; | |
2122 | continue; /* continue at next VMA */ | |
2123 | } | |
2124 | break; | |
2125 | } | |
2126 | nend = nstart + ret * PAGE_SIZE; | |
2127 | ret = 0; | |
2128 | } | |
2129 | if (locked) | |
d8ed45c5 | 2130 | mmap_read_unlock(mm); |
d3649f68 CH |
2131 | return ret; /* 0 or negative error code */ |
2132 | } | |
050a9adc | 2133 | #else /* CONFIG_MMU */ |
64019a2e | 2134 | static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start, |
050a9adc | 2135 | unsigned long nr_pages, struct page **pages, |
b2cac248 | 2136 | int *locked, unsigned int foll_flags) |
050a9adc CH |
2137 | { |
2138 | struct vm_area_struct *vma; | |
b2a72dff | 2139 | bool must_unlock = false; |
050a9adc | 2140 | unsigned long vm_flags; |
24dc20c7 | 2141 | long i; |
050a9adc | 2142 | |
b2a72dff JG |
2143 | if (!nr_pages) |
2144 | return 0; | |
2145 | ||
2146 | /* | |
2147 | * The internal caller expects GUP to manage the lock internally and the | |
2148 | * lock must be released when this returns. | |
2149 | */ | |
9a863a6a | 2150 | if (!*locked) { |
b2a72dff JG |
2151 | if (mmap_read_lock_killable(mm)) |
2152 | return -EAGAIN; | |
2153 | must_unlock = true; | |
2154 | *locked = 1; | |
2155 | } | |
2156 | ||
050a9adc CH |
2157 | /* calculate required read or write permissions. |
2158 | * If FOLL_FORCE is set, we only require the "MAY" flags. | |
2159 | */ | |
2160 | vm_flags = (foll_flags & FOLL_WRITE) ? | |
2161 | (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); | |
2162 | vm_flags &= (foll_flags & FOLL_FORCE) ? | |
2163 | (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); | |
2164 | ||
2165 | for (i = 0; i < nr_pages; i++) { | |
2166 | vma = find_vma(mm, start); | |
2167 | if (!vma) | |
b2a72dff | 2168 | break; |
050a9adc CH |
2169 | |
2170 | /* protect what we can, including chardevs */ | |
2171 | if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) || | |
2172 | !(vm_flags & vma->vm_flags)) | |
b2a72dff | 2173 | break; |
050a9adc CH |
2174 | |
2175 | if (pages) { | |
396a400b | 2176 | pages[i] = virt_to_page((void *)start); |
050a9adc CH |
2177 | if (pages[i]) |
2178 | get_page(pages[i]); | |
2179 | } | |
b2cac248 | 2180 | |
050a9adc CH |
2181 | start = (start + PAGE_SIZE) & PAGE_MASK; |
2182 | } | |
2183 | ||
b2a72dff JG |
2184 | if (must_unlock && *locked) { |
2185 | mmap_read_unlock(mm); | |
2186 | *locked = 0; | |
2187 | } | |
050a9adc | 2188 | |
050a9adc CH |
2189 | return i ? : -EFAULT; |
2190 | } | |
2191 | #endif /* !CONFIG_MMU */ | |
d3649f68 | 2192 | |
bb523b40 AG |
2193 | /** |
2194 | * fault_in_writeable - fault in userspace address range for writing | |
2195 | * @uaddr: start of address range | |
2196 | * @size: size of address range | |
2197 | * | |
2198 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2199 | * copy_from_user()). | |
2200 | */ | |
2201 | size_t fault_in_writeable(char __user *uaddr, size_t size) | |
2202 | { | |
2203 | char __user *start = uaddr, *end; | |
2204 | ||
2205 | if (unlikely(size == 0)) | |
2206 | return 0; | |
677b2a8c CL |
2207 | if (!user_write_access_begin(uaddr, size)) |
2208 | return size; | |
bb523b40 | 2209 | if (!PAGE_ALIGNED(uaddr)) { |
677b2a8c | 2210 | unsafe_put_user(0, uaddr, out); |
bb523b40 AG |
2211 | uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr); |
2212 | } | |
2213 | end = (char __user *)PAGE_ALIGN((unsigned long)start + size); | |
2214 | if (unlikely(end < start)) | |
2215 | end = NULL; | |
2216 | while (uaddr != end) { | |
677b2a8c | 2217 | unsafe_put_user(0, uaddr, out); |
bb523b40 AG |
2218 | uaddr += PAGE_SIZE; |
2219 | } | |
2220 | ||
2221 | out: | |
677b2a8c | 2222 | user_write_access_end(); |
bb523b40 AG |
2223 | if (size > uaddr - start) |
2224 | return size - (uaddr - start); | |
2225 | return 0; | |
2226 | } | |
2227 | EXPORT_SYMBOL(fault_in_writeable); | |
2228 | ||
da32b581 CM |
2229 | /** |
2230 | * fault_in_subpage_writeable - fault in an address range for writing | |
2231 | * @uaddr: start of address range | |
2232 | * @size: size of address range | |
2233 | * | |
2234 | * Fault in a user address range for writing while checking for permissions at | |
2235 | * sub-page granularity (e.g. arm64 MTE). This function should be used when | |
2236 | * the caller cannot guarantee forward progress of a copy_to_user() loop. | |
2237 | * | |
2238 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2239 | * copy_from_user()). | |
2240 | */ | |
2241 | size_t fault_in_subpage_writeable(char __user *uaddr, size_t size) | |
2242 | { | |
2243 | size_t faulted_in; | |
2244 | ||
2245 | /* | |
2246 | * Attempt faulting in at page granularity first for page table | |
2247 | * permission checking. The arch-specific probe_subpage_writeable() | |
2248 | * functions may not check for this. | |
2249 | */ | |
2250 | faulted_in = size - fault_in_writeable(uaddr, size); | |
2251 | if (faulted_in) | |
2252 | faulted_in -= probe_subpage_writeable(uaddr, faulted_in); | |
2253 | ||
2254 | return size - faulted_in; | |
2255 | } | |
2256 | EXPORT_SYMBOL(fault_in_subpage_writeable); | |
2257 | ||
cdd591fc AG |
2258 | /* |
2259 | * fault_in_safe_writeable - fault in an address range for writing | |
2260 | * @uaddr: start of address range | |
2261 | * @size: length of address range | |
2262 | * | |
fe673d3f LT |
2263 | * Faults in an address range for writing. This is primarily useful when we |
2264 | * already know that some or all of the pages in the address range aren't in | |
2265 | * memory. | |
cdd591fc | 2266 | * |
fe673d3f | 2267 | * Unlike fault_in_writeable(), this function is non-destructive. |
cdd591fc AG |
2268 | * |
2269 | * Note that we don't pin or otherwise hold the pages referenced that we fault | |
2270 | * in. There's no guarantee that they'll stay in memory for any duration of | |
2271 | * time. | |
2272 | * | |
2273 | * Returns the number of bytes not faulted in, like copy_to_user() and | |
2274 | * copy_from_user(). | |
2275 | */ | |
2276 | size_t fault_in_safe_writeable(const char __user *uaddr, size_t size) | |
2277 | { | |
fe673d3f | 2278 | unsigned long start = (unsigned long)uaddr, end; |
cdd591fc | 2279 | struct mm_struct *mm = current->mm; |
fe673d3f | 2280 | bool unlocked = false; |
cdd591fc | 2281 | |
fe673d3f LT |
2282 | if (unlikely(size == 0)) |
2283 | return 0; | |
cdd591fc | 2284 | end = PAGE_ALIGN(start + size); |
fe673d3f | 2285 | if (end < start) |
cdd591fc | 2286 | end = 0; |
cdd591fc | 2287 | |
fe673d3f LT |
2288 | mmap_read_lock(mm); |
2289 | do { | |
2290 | if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked)) | |
cdd591fc | 2291 | break; |
fe673d3f LT |
2292 | start = (start + PAGE_SIZE) & PAGE_MASK; |
2293 | } while (start != end); | |
2294 | mmap_read_unlock(mm); | |
2295 | ||
2296 | if (size > (unsigned long)uaddr - start) | |
2297 | return size - ((unsigned long)uaddr - start); | |
2298 | return 0; | |
cdd591fc AG |
2299 | } |
2300 | EXPORT_SYMBOL(fault_in_safe_writeable); | |
2301 | ||
bb523b40 AG |
2302 | /** |
2303 | * fault_in_readable - fault in userspace address range for reading | |
2304 | * @uaddr: start of user address range | |
2305 | * @size: size of user address range | |
2306 | * | |
2307 | * Returns the number of bytes not faulted in (like copy_to_user() and | |
2308 | * copy_from_user()). | |
2309 | */ | |
2310 | size_t fault_in_readable(const char __user *uaddr, size_t size) | |
2311 | { | |
2312 | const char __user *start = uaddr, *end; | |
2313 | volatile char c; | |
2314 | ||
2315 | if (unlikely(size == 0)) | |
2316 | return 0; | |
677b2a8c CL |
2317 | if (!user_read_access_begin(uaddr, size)) |
2318 | return size; | |
bb523b40 | 2319 | if (!PAGE_ALIGNED(uaddr)) { |
677b2a8c | 2320 | unsafe_get_user(c, uaddr, out); |
bb523b40 AG |
2321 | uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr); |
2322 | } | |
2323 | end = (const char __user *)PAGE_ALIGN((unsigned long)start + size); | |
2324 | if (unlikely(end < start)) | |
2325 | end = NULL; | |
2326 | while (uaddr != end) { | |
677b2a8c | 2327 | unsafe_get_user(c, uaddr, out); |
bb523b40 AG |
2328 | uaddr += PAGE_SIZE; |
2329 | } | |
2330 | ||
2331 | out: | |
677b2a8c | 2332 | user_read_access_end(); |
bb523b40 AG |
2333 | (void)c; |
2334 | if (size > uaddr - start) | |
2335 | return size - (uaddr - start); | |
2336 | return 0; | |
2337 | } | |
2338 | EXPORT_SYMBOL(fault_in_readable); | |
2339 | ||
8f942eea JH |
2340 | /** |
2341 | * get_dump_page() - pin user page in memory while writing it to core dump | |
2342 | * @addr: user address | |
2343 | * | |
2344 | * Returns struct page pointer of user page pinned for dump, | |
2345 | * to be freed afterwards by put_page(). | |
2346 | * | |
2347 | * Returns NULL on any kind of failure - a hole must then be inserted into | |
2348 | * the corefile, to preserve alignment with its headers; and also returns | |
2349 | * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found - | |
f0953a1b | 2350 | * allowing a hole to be left in the corefile to save disk space. |
8f942eea | 2351 | * |
7f3bfab5 | 2352 | * Called without mmap_lock (takes and releases the mmap_lock by itself). |
8f942eea JH |
2353 | */ |
2354 | #ifdef CONFIG_ELF_CORE | |
2355 | struct page *get_dump_page(unsigned long addr) | |
2356 | { | |
8f942eea | 2357 | struct page *page; |
b2a72dff | 2358 | int locked = 0; |
7f3bfab5 | 2359 | int ret; |
8f942eea | 2360 | |
b2cac248 | 2361 | ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked, |
7f3bfab5 | 2362 | FOLL_FORCE | FOLL_DUMP | FOLL_GET); |
7f3bfab5 | 2363 | return (ret == 1) ? page : NULL; |
8f942eea JH |
2364 | } |
2365 | #endif /* CONFIG_ELF_CORE */ | |
2366 | ||
d1e153fe | 2367 | #ifdef CONFIG_MIGRATION |
f68749ec | 2368 | /* |
67e139b0 | 2369 | * Returns the number of collected pages. Return value is always >= 0. |
f68749ec | 2370 | */ |
67e139b0 AP |
2371 | static unsigned long collect_longterm_unpinnable_pages( |
2372 | struct list_head *movable_page_list, | |
2373 | unsigned long nr_pages, | |
2374 | struct page **pages) | |
9a4e9f3b | 2375 | { |
67e139b0 | 2376 | unsigned long i, collected = 0; |
1b7f7e58 | 2377 | struct folio *prev_folio = NULL; |
67e139b0 | 2378 | bool drain_allow = true; |
9a4e9f3b | 2379 | |
83c02c23 | 2380 | for (i = 0; i < nr_pages; i++) { |
1b7f7e58 | 2381 | struct folio *folio = page_folio(pages[i]); |
f9f38f78 | 2382 | |
1b7f7e58 | 2383 | if (folio == prev_folio) |
83c02c23 | 2384 | continue; |
1b7f7e58 | 2385 | prev_folio = folio; |
f9f38f78 | 2386 | |
67e139b0 AP |
2387 | if (folio_is_longterm_pinnable(folio)) |
2388 | continue; | |
b05a79d4 | 2389 | |
67e139b0 | 2390 | collected++; |
b05a79d4 | 2391 | |
67e139b0 | 2392 | if (folio_is_device_coherent(folio)) |
f9f38f78 CH |
2393 | continue; |
2394 | ||
1b7f7e58 | 2395 | if (folio_test_hugetlb(folio)) { |
6aa3a920 | 2396 | isolate_hugetlb(folio, movable_page_list); |
f9f38f78 CH |
2397 | continue; |
2398 | } | |
9a4e9f3b | 2399 | |
1b7f7e58 | 2400 | if (!folio_test_lru(folio) && drain_allow) { |
f9f38f78 CH |
2401 | lru_add_drain_all(); |
2402 | drain_allow = false; | |
2403 | } | |
2404 | ||
be2d5756 | 2405 | if (!folio_isolate_lru(folio)) |
f9f38f78 | 2406 | continue; |
67e139b0 AP |
2407 | |
2408 | list_add_tail(&folio->lru, movable_page_list); | |
1b7f7e58 MWO |
2409 | node_stat_mod_folio(folio, |
2410 | NR_ISOLATED_ANON + folio_is_file_lru(folio), | |
2411 | folio_nr_pages(folio)); | |
9a4e9f3b AK |
2412 | } |
2413 | ||
67e139b0 AP |
2414 | return collected; |
2415 | } | |
2416 | ||
2417 | /* | |
2418 | * Unpins all pages and migrates device coherent pages and movable_page_list. | |
2419 | * Returns -EAGAIN if all pages were successfully migrated or -errno for failure | |
2420 | * (or partial success). | |
2421 | */ | |
2422 | static int migrate_longterm_unpinnable_pages( | |
2423 | struct list_head *movable_page_list, | |
2424 | unsigned long nr_pages, | |
2425 | struct page **pages) | |
2426 | { | |
2427 | int ret; | |
2428 | unsigned long i; | |
6e7f34eb | 2429 | |
b05a79d4 | 2430 | for (i = 0; i < nr_pages; i++) { |
67e139b0 AP |
2431 | struct folio *folio = page_folio(pages[i]); |
2432 | ||
2433 | if (folio_is_device_coherent(folio)) { | |
2434 | /* | |
2435 | * Migration will fail if the page is pinned, so convert | |
2436 | * the pin on the source page to a normal reference. | |
2437 | */ | |
2438 | pages[i] = NULL; | |
2439 | folio_get(folio); | |
2440 | gup_put_folio(folio, 1, FOLL_PIN); | |
2441 | ||
2442 | if (migrate_device_coherent_page(&folio->page)) { | |
2443 | ret = -EBUSY; | |
2444 | goto err; | |
2445 | } | |
2446 | ||
b05a79d4 | 2447 | continue; |
67e139b0 | 2448 | } |
b05a79d4 | 2449 | |
67e139b0 AP |
2450 | /* |
2451 | * We can't migrate pages with unexpected references, so drop | |
2452 | * the reference obtained by __get_user_pages_locked(). | |
2453 | * Migrating pages have been added to movable_page_list after | |
2454 | * calling folio_isolate_lru() which takes a reference so the | |
2455 | * page won't be freed if it's migrating. | |
2456 | */ | |
f6d299ec | 2457 | unpin_user_page(pages[i]); |
67e139b0 | 2458 | pages[i] = NULL; |
f68749ec | 2459 | } |
f9f38f78 | 2460 | |
67e139b0 | 2461 | if (!list_empty(movable_page_list)) { |
f9f38f78 CH |
2462 | struct migration_target_control mtc = { |
2463 | .nid = NUMA_NO_NODE, | |
2464 | .gfp_mask = GFP_USER | __GFP_NOWARN, | |
e42dfe4e | 2465 | .reason = MR_LONGTERM_PIN, |
f9f38f78 CH |
2466 | }; |
2467 | ||
67e139b0 AP |
2468 | if (migrate_pages(movable_page_list, alloc_migration_target, |
2469 | NULL, (unsigned long)&mtc, MIGRATE_SYNC, | |
2470 | MR_LONGTERM_PIN, NULL)) { | |
f9f38f78 | 2471 | ret = -ENOMEM; |
67e139b0 AP |
2472 | goto err; |
2473 | } | |
9a4e9f3b AK |
2474 | } |
2475 | ||
67e139b0 AP |
2476 | putback_movable_pages(movable_page_list); |
2477 | ||
2478 | return -EAGAIN; | |
2479 | ||
2480 | err: | |
2481 | for (i = 0; i < nr_pages; i++) | |
2482 | if (pages[i]) | |
2483 | unpin_user_page(pages[i]); | |
2484 | putback_movable_pages(movable_page_list); | |
24a95998 | 2485 | |
67e139b0 AP |
2486 | return ret; |
2487 | } | |
2488 | ||
2489 | /* | |
2490 | * Check whether all pages are *allowed* to be pinned. Rather confusingly, all | |
2491 | * pages in the range are required to be pinned via FOLL_PIN, before calling | |
2492 | * this routine. | |
2493 | * | |
2494 | * If any pages in the range are not allowed to be pinned, then this routine | |
2495 | * will migrate those pages away, unpin all the pages in the range and return | |
2496 | * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then | |
2497 | * call this routine again. | |
2498 | * | |
2499 | * If an error other than -EAGAIN occurs, this indicates a migration failure. | |
2500 | * The caller should give up, and propagate the error back up the call stack. | |
2501 | * | |
2502 | * If everything is OK and all pages in the range are allowed to be pinned, then | |
2503 | * this routine leaves all pages pinned and returns zero for success. | |
2504 | */ | |
2505 | static long check_and_migrate_movable_pages(unsigned long nr_pages, | |
2506 | struct page **pages) | |
2507 | { | |
2508 | unsigned long collected; | |
2509 | LIST_HEAD(movable_page_list); | |
2510 | ||
2511 | collected = collect_longterm_unpinnable_pages(&movable_page_list, | |
2512 | nr_pages, pages); | |
2513 | if (!collected) | |
2514 | return 0; | |
2515 | ||
2516 | return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages, | |
2517 | pages); | |
9a4e9f3b AK |
2518 | } |
2519 | #else | |
f68749ec | 2520 | static long check_and_migrate_movable_pages(unsigned long nr_pages, |
f6d299ec | 2521 | struct page **pages) |
9a4e9f3b | 2522 | { |
24a95998 | 2523 | return 0; |
9a4e9f3b | 2524 | } |
d1e153fe | 2525 | #endif /* CONFIG_MIGRATION */ |
9a4e9f3b | 2526 | |
2bb6d283 | 2527 | /* |
932f4a63 IW |
2528 | * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which |
2529 | * allows us to process the FOLL_LONGTERM flag. | |
2bb6d283 | 2530 | */ |
64019a2e | 2531 | static long __gup_longterm_locked(struct mm_struct *mm, |
932f4a63 IW |
2532 | unsigned long start, |
2533 | unsigned long nr_pages, | |
2534 | struct page **pages, | |
53b2d09b | 2535 | int *locked, |
932f4a63 | 2536 | unsigned int gup_flags) |
2bb6d283 | 2537 | { |
f68749ec | 2538 | unsigned int flags; |
24a95998 | 2539 | long rc, nr_pinned_pages; |
2bb6d283 | 2540 | |
f68749ec | 2541 | if (!(gup_flags & FOLL_LONGTERM)) |
b2cac248 | 2542 | return __get_user_pages_locked(mm, start, nr_pages, pages, |
53b2d09b | 2543 | locked, gup_flags); |
67e139b0 | 2544 | |
f68749ec PT |
2545 | flags = memalloc_pin_save(); |
2546 | do { | |
24a95998 | 2547 | nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages, |
b2cac248 | 2548 | pages, locked, |
24a95998 AP |
2549 | gup_flags); |
2550 | if (nr_pinned_pages <= 0) { | |
2551 | rc = nr_pinned_pages; | |
f68749ec | 2552 | break; |
24a95998 | 2553 | } |
d64e2dbc JG |
2554 | |
2555 | /* FOLL_LONGTERM implies FOLL_PIN */ | |
f6d299ec | 2556 | rc = check_and_migrate_movable_pages(nr_pinned_pages, pages); |
24a95998 | 2557 | } while (rc == -EAGAIN); |
f68749ec | 2558 | memalloc_pin_restore(flags); |
24a95998 | 2559 | return rc ? rc : nr_pinned_pages; |
2bb6d283 | 2560 | } |
932f4a63 | 2561 | |
d64e2dbc JG |
2562 | /* |
2563 | * Check that the given flags are valid for the exported gup/pup interface, and | |
2564 | * update them with the required flags that the caller must have set. | |
2565 | */ | |
b2cac248 LS |
2566 | static bool is_valid_gup_args(struct page **pages, int *locked, |
2567 | unsigned int *gup_flags_p, unsigned int to_set) | |
447f3e45 | 2568 | { |
d64e2dbc JG |
2569 | unsigned int gup_flags = *gup_flags_p; |
2570 | ||
447f3e45 | 2571 | /* |
d64e2dbc JG |
2572 | * These flags not allowed to be specified externally to the gup |
2573 | * interfaces: | |
0f20bba1 | 2574 | * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only |
d64e2dbc | 2575 | * - FOLL_REMOTE is internal only and used on follow_page() |
f04740f5 | 2576 | * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL |
447f3e45 | 2577 | */ |
0f20bba1 | 2578 | if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS)) |
d64e2dbc JG |
2579 | return false; |
2580 | ||
2581 | gup_flags |= to_set; | |
f04740f5 JG |
2582 | if (locked) { |
2583 | /* At the external interface locked must be set */ | |
2584 | if (WARN_ON_ONCE(*locked != 1)) | |
2585 | return false; | |
2586 | ||
2587 | gup_flags |= FOLL_UNLOCKABLE; | |
2588 | } | |
d64e2dbc JG |
2589 | |
2590 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
2591 | if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) == | |
2592 | (FOLL_PIN | FOLL_GET))) | |
2593 | return false; | |
2594 | ||
2595 | /* LONGTERM can only be specified when pinning */ | |
2596 | if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM))) | |
2597 | return false; | |
2598 | ||
2599 | /* Pages input must be given if using GET/PIN */ | |
2600 | if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages)) | |
447f3e45 | 2601 | return false; |
d64e2dbc | 2602 | |
d64e2dbc JG |
2603 | /* We want to allow the pgmap to be hot-unplugged at all times */ |
2604 | if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) && | |
2605 | (gup_flags & FOLL_PCI_P2PDMA))) | |
2606 | return false; | |
2607 | ||
d64e2dbc | 2608 | *gup_flags_p = gup_flags; |
447f3e45 BS |
2609 | return true; |
2610 | } | |
2611 | ||
22bf29b6 | 2612 | #ifdef CONFIG_MMU |
adc8cb40 | 2613 | /** |
c4237f8b | 2614 | * get_user_pages_remote() - pin user pages in memory |
c4237f8b JH |
2615 | * @mm: mm_struct of target mm |
2616 | * @start: starting user address | |
2617 | * @nr_pages: number of pages from start to pin | |
2618 | * @gup_flags: flags modifying lookup behaviour | |
2619 | * @pages: array that receives pointers to the pages pinned. | |
2620 | * Should be at least nr_pages long. Or NULL, if caller | |
2621 | * only intends to ensure the pages are faulted in. | |
c4237f8b JH |
2622 | * @locked: pointer to lock flag indicating whether lock is held and |
2623 | * subsequently whether VM_FAULT_RETRY functionality can be | |
2624 | * utilised. Lock must initially be held. | |
2625 | * | |
2626 | * Returns either number of pages pinned (which may be less than the | |
2627 | * number requested), or an error. Details about the return value: | |
2628 | * | |
2629 | * -- If nr_pages is 0, returns 0. | |
2630 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
2631 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
2632 | * pages pinned. Again, this may be less than nr_pages. | |
2633 | * | |
2634 | * The caller is responsible for releasing returned @pages, via put_page(). | |
2635 | * | |
c1e8d7c6 | 2636 | * Must be called with mmap_lock held for read or write. |
c4237f8b | 2637 | * |
adc8cb40 SJ |
2638 | * get_user_pages_remote walks a process's page tables and takes a reference |
2639 | * to each struct page that each user address corresponds to at a given | |
c4237f8b JH |
2640 | * instant. That is, it takes the page that would be accessed if a user |
2641 | * thread accesses the given user virtual address at that instant. | |
2642 | * | |
2643 | * This does not guarantee that the page exists in the user mappings when | |
adc8cb40 | 2644 | * get_user_pages_remote returns, and there may even be a completely different |
c4237f8b | 2645 | * page there in some cases (eg. if mmapped pagecache has been invalidated |
5da1a868 | 2646 | * and subsequently re-faulted). However it does guarantee that the page |
c4237f8b JH |
2647 | * won't be freed completely. And mostly callers simply care that the page |
2648 | * contains data that was valid *at some point in time*. Typically, an IO | |
2649 | * or similar operation cannot guarantee anything stronger anyway because | |
2650 | * locks can't be held over the syscall boundary. | |
2651 | * | |
2652 | * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page | |
2653 | * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must | |
2654 | * be called after the page is finished with, and before put_page is called. | |
2655 | * | |
adc8cb40 SJ |
2656 | * get_user_pages_remote is typically used for fewer-copy IO operations, |
2657 | * to get a handle on the memory by some means other than accesses | |
2658 | * via the user virtual addresses. The pages may be submitted for | |
2659 | * DMA to devices or accessed via their kernel linear mapping (via the | |
2660 | * kmap APIs). Care should be taken to use the correct cache flushing APIs. | |
c4237f8b JH |
2661 | * |
2662 | * See also get_user_pages_fast, for performance critical applications. | |
2663 | * | |
adc8cb40 | 2664 | * get_user_pages_remote should be phased out in favor of |
c4237f8b | 2665 | * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing |
adc8cb40 | 2666 | * should use get_user_pages_remote because it cannot pass |
c4237f8b JH |
2667 | * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault. |
2668 | */ | |
64019a2e | 2669 | long get_user_pages_remote(struct mm_struct *mm, |
c4237f8b JH |
2670 | unsigned long start, unsigned long nr_pages, |
2671 | unsigned int gup_flags, struct page **pages, | |
ca5e8632 | 2672 | int *locked) |
c4237f8b | 2673 | { |
9a863a6a JG |
2674 | int local_locked = 1; |
2675 | ||
b2cac248 | 2676 | if (!is_valid_gup_args(pages, locked, &gup_flags, |
d64e2dbc | 2677 | FOLL_TOUCH | FOLL_REMOTE)) |
eddb1c22 JH |
2678 | return -EINVAL; |
2679 | ||
b2cac248 | 2680 | return __get_user_pages_locked(mm, start, nr_pages, pages, |
9a863a6a | 2681 | locked ? locked : &local_locked, |
d64e2dbc | 2682 | gup_flags); |
c4237f8b JH |
2683 | } |
2684 | EXPORT_SYMBOL(get_user_pages_remote); | |
2685 | ||
eddb1c22 | 2686 | #else /* CONFIG_MMU */ |
64019a2e | 2687 | long get_user_pages_remote(struct mm_struct *mm, |
eddb1c22 JH |
2688 | unsigned long start, unsigned long nr_pages, |
2689 | unsigned int gup_flags, struct page **pages, | |
ca5e8632 | 2690 | int *locked) |
eddb1c22 JH |
2691 | { |
2692 | return 0; | |
2693 | } | |
2694 | #endif /* !CONFIG_MMU */ | |
2695 | ||
adc8cb40 SJ |
2696 | /** |
2697 | * get_user_pages() - pin user pages in memory | |
2698 | * @start: starting user address | |
2699 | * @nr_pages: number of pages from start to pin | |
2700 | * @gup_flags: flags modifying lookup behaviour | |
2701 | * @pages: array that receives pointers to the pages pinned. | |
2702 | * Should be at least nr_pages long. Or NULL, if caller | |
2703 | * only intends to ensure the pages are faulted in. | |
adc8cb40 | 2704 | * |
64019a2e PX |
2705 | * This is the same as get_user_pages_remote(), just with a less-flexible |
2706 | * calling convention where we assume that the mm being operated on belongs to | |
2707 | * the current task, and doesn't allow passing of a locked parameter. We also | |
2708 | * obviously don't pass FOLL_REMOTE in here. | |
932f4a63 IW |
2709 | */ |
2710 | long get_user_pages(unsigned long start, unsigned long nr_pages, | |
54d02069 | 2711 | unsigned int gup_flags, struct page **pages) |
932f4a63 | 2712 | { |
9a863a6a JG |
2713 | int locked = 1; |
2714 | ||
b2cac248 | 2715 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH)) |
eddb1c22 JH |
2716 | return -EINVAL; |
2717 | ||
afa3c33e | 2718 | return __get_user_pages_locked(current->mm, start, nr_pages, pages, |
b2cac248 | 2719 | &locked, gup_flags); |
932f4a63 IW |
2720 | } |
2721 | EXPORT_SYMBOL(get_user_pages); | |
2bb6d283 | 2722 | |
acc3c8d1 | 2723 | /* |
d3649f68 | 2724 | * get_user_pages_unlocked() is suitable to replace the form: |
acc3c8d1 | 2725 | * |
3e4e28c5 | 2726 | * mmap_read_lock(mm); |
64019a2e | 2727 | * get_user_pages(mm, ..., pages, NULL); |
3e4e28c5 | 2728 | * mmap_read_unlock(mm); |
d3649f68 CH |
2729 | * |
2730 | * with: | |
2731 | * | |
64019a2e | 2732 | * get_user_pages_unlocked(mm, ..., pages); |
d3649f68 CH |
2733 | * |
2734 | * It is functionally equivalent to get_user_pages_fast so | |
2735 | * get_user_pages_fast should be used instead if specific gup_flags | |
2736 | * (e.g. FOLL_FORCE) are not required. | |
acc3c8d1 | 2737 | */ |
d3649f68 CH |
2738 | long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages, |
2739 | struct page **pages, unsigned int gup_flags) | |
acc3c8d1 | 2740 | { |
b2a72dff | 2741 | int locked = 0; |
acc3c8d1 | 2742 | |
b2cac248 | 2743 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
f04740f5 | 2744 | FOLL_TOUCH | FOLL_UNLOCKABLE)) |
d64e2dbc JG |
2745 | return -EINVAL; |
2746 | ||
afa3c33e | 2747 | return __get_user_pages_locked(current->mm, start, nr_pages, pages, |
b2cac248 | 2748 | &locked, gup_flags); |
4bbd4c77 | 2749 | } |
d3649f68 | 2750 | EXPORT_SYMBOL(get_user_pages_unlocked); |
2667f50e SC |
2751 | |
2752 | /* | |
23babe19 | 2753 | * GUP-fast |
2667f50e SC |
2754 | * |
2755 | * get_user_pages_fast attempts to pin user pages by walking the page | |
2756 | * tables directly and avoids taking locks. Thus the walker needs to be | |
2757 | * protected from page table pages being freed from under it, and should | |
2758 | * block any THP splits. | |
2759 | * | |
2760 | * One way to achieve this is to have the walker disable interrupts, and | |
2761 | * rely on IPIs from the TLB flushing code blocking before the page table | |
2762 | * pages are freed. This is unsuitable for architectures that do not need | |
2763 | * to broadcast an IPI when invalidating TLBs. | |
2764 | * | |
2765 | * Another way to achieve this is to batch up page table containing pages | |
2766 | * belonging to more than one mm_user, then rcu_sched a callback to free those | |
23babe19 | 2767 | * pages. Disabling interrupts will allow the gup_fast() walker to both block |
2667f50e SC |
2768 | * the rcu_sched callback, and an IPI that we broadcast for splitting THPs |
2769 | * (which is a relatively rare event). The code below adopts this strategy. | |
2770 | * | |
2771 | * Before activating this code, please be aware that the following assumptions | |
2772 | * are currently made: | |
2773 | * | |
ff2e6d72 | 2774 | * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to |
e585513b | 2775 | * free pages containing page tables or TLB flushing requires IPI broadcast. |
2667f50e | 2776 | * |
2667f50e SC |
2777 | * *) ptes can be read atomically by the architecture. |
2778 | * | |
2779 | * *) access_ok is sufficient to validate userspace address ranges. | |
2780 | * | |
2781 | * The last two assumptions can be relaxed by the addition of helper functions. | |
2782 | * | |
2783 | * This code is based heavily on the PowerPC implementation by Nick Piggin. | |
2784 | */ | |
25176ad0 | 2785 | #ifdef CONFIG_HAVE_GUP_FAST |
3faa52c0 | 2786 | |
a6e79df9 | 2787 | /* |
f002882c DH |
2788 | * Used in the GUP-fast path to determine whether GUP is permitted to work on |
2789 | * a specific folio. | |
a6e79df9 LS |
2790 | * |
2791 | * This call assumes the caller has pinned the folio, that the lowest page table | |
2792 | * level still points to this folio, and that interrupts have been disabled. | |
2793 | * | |
f002882c DH |
2794 | * GUP-fast must reject all secretmem folios. |
2795 | * | |
a6e79df9 LS |
2796 | * Writing to pinned file-backed dirty tracked folios is inherently problematic |
2797 | * (see comment describing the writable_file_mapping_allowed() function). We | |
2798 | * therefore try to avoid the most egregious case of a long-term mapping doing | |
2799 | * so. | |
2800 | * | |
2801 | * This function cannot be as thorough as that one as the VMA is not available | |
2802 | * in the fast path, so instead we whitelist known good cases and if in doubt, | |
2803 | * fall back to the slow path. | |
2804 | */ | |
f002882c | 2805 | static bool gup_fast_folio_allowed(struct folio *folio, unsigned int flags) |
a6e79df9 | 2806 | { |
f002882c | 2807 | bool reject_file_backed = false; |
a6e79df9 | 2808 | struct address_space *mapping; |
f002882c | 2809 | bool check_secretmem = false; |
a6e79df9 LS |
2810 | unsigned long mapping_flags; |
2811 | ||
2812 | /* | |
2813 | * If we aren't pinning then no problematic write can occur. A long term | |
2814 | * pin is the most egregious case so this is the one we disallow. | |
2815 | */ | |
f002882c | 2816 | if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) == |
a6e79df9 | 2817 | (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) |
f002882c DH |
2818 | reject_file_backed = true; |
2819 | ||
2820 | /* We hold a folio reference, so we can safely access folio fields. */ | |
a6e79df9 | 2821 | |
f002882c DH |
2822 | /* secretmem folios are always order-0 folios. */ |
2823 | if (IS_ENABLED(CONFIG_SECRETMEM) && !folio_test_large(folio)) | |
2824 | check_secretmem = true; | |
2825 | ||
2826 | if (!reject_file_backed && !check_secretmem) | |
2827 | return true; | |
a6e79df9 LS |
2828 | |
2829 | if (WARN_ON_ONCE(folio_test_slab(folio))) | |
2830 | return false; | |
2831 | ||
f002882c | 2832 | /* hugetlb neither requires dirty-tracking nor can be secretmem. */ |
a6e79df9 LS |
2833 | if (folio_test_hugetlb(folio)) |
2834 | return true; | |
2835 | ||
2836 | /* | |
2837 | * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods | |
2838 | * cannot proceed, which means no actions performed under RCU can | |
2839 | * proceed either. | |
2840 | * | |
2841 | * inodes and thus their mappings are freed under RCU, which means the | |
2842 | * mapping cannot be freed beneath us and thus we can safely dereference | |
2843 | * it. | |
2844 | */ | |
2845 | lockdep_assert_irqs_disabled(); | |
2846 | ||
2847 | /* | |
2848 | * However, there may be operations which _alter_ the mapping, so ensure | |
2849 | * we read it once and only once. | |
2850 | */ | |
2851 | mapping = READ_ONCE(folio->mapping); | |
2852 | ||
2853 | /* | |
2854 | * The mapping may have been truncated, in any case we cannot determine | |
2855 | * if this mapping is safe - fall back to slow path to determine how to | |
2856 | * proceed. | |
2857 | */ | |
2858 | if (!mapping) | |
2859 | return false; | |
2860 | ||
2861 | /* Anonymous folios pose no problem. */ | |
2862 | mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS; | |
2863 | if (mapping_flags) | |
2864 | return mapping_flags & PAGE_MAPPING_ANON; | |
2865 | ||
2866 | /* | |
2867 | * At this point, we know the mapping is non-null and points to an | |
f002882c | 2868 | * address_space object. |
a6e79df9 | 2869 | */ |
f002882c DH |
2870 | if (check_secretmem && secretmem_mapping(mapping)) |
2871 | return false; | |
2872 | /* The only remaining allowed file system is shmem. */ | |
2873 | return !reject_file_backed || shmem_mapping(mapping); | |
a6e79df9 LS |
2874 | } |
2875 | ||
23babe19 DH |
2876 | static void __maybe_unused gup_fast_undo_dev_pagemap(int *nr, int nr_start, |
2877 | unsigned int flags, struct page **pages) | |
b59f65fa KS |
2878 | { |
2879 | while ((*nr) - nr_start) { | |
2880 | struct page *page = pages[--(*nr)]; | |
2881 | ||
2882 | ClearPageReferenced(page); | |
3faa52c0 JH |
2883 | if (flags & FOLL_PIN) |
2884 | unpin_user_page(page); | |
2885 | else | |
2886 | put_page(page); | |
b59f65fa KS |
2887 | } |
2888 | } | |
2889 | ||
3010a5ea | 2890 | #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL |
70cbc3cc | 2891 | /* |
23babe19 | 2892 | * GUP-fast relies on pte change detection to avoid concurrent pgtable |
70cbc3cc YS |
2893 | * operations. |
2894 | * | |
23babe19 | 2895 | * To pin the page, GUP-fast needs to do below in order: |
70cbc3cc YS |
2896 | * (1) pin the page (by prefetching pte), then (2) check pte not changed. |
2897 | * | |
2898 | * For the rest of pgtable operations where pgtable updates can be racy | |
23babe19 | 2899 | * with GUP-fast, we need to do (1) clear pte, then (2) check whether page |
70cbc3cc YS |
2900 | * is pinned. |
2901 | * | |
2902 | * Above will work for all pte-level operations, including THP split. | |
2903 | * | |
23babe19 | 2904 | * For THP collapse, it's a bit more complicated because GUP-fast may be |
70cbc3cc YS |
2905 | * walking a pgtable page that is being freed (pte is still valid but pmd |
2906 | * can be cleared already). To avoid race in such condition, we need to | |
2907 | * also check pmd here to make sure pmd doesn't change (corresponds to | |
2908 | * pmdp_collapse_flush() in the THP collapse code path). | |
2909 | */ | |
23babe19 DH |
2910 | static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, |
2911 | unsigned long end, unsigned int flags, struct page **pages, | |
2912 | int *nr) | |
2667f50e | 2913 | { |
b59f65fa KS |
2914 | struct dev_pagemap *pgmap = NULL; |
2915 | int nr_start = *nr, ret = 0; | |
2667f50e | 2916 | pte_t *ptep, *ptem; |
2667f50e SC |
2917 | |
2918 | ptem = ptep = pte_offset_map(&pmd, addr); | |
04dee9e8 HD |
2919 | if (!ptep) |
2920 | return 0; | |
2667f50e | 2921 | do { |
2a4a06da | 2922 | pte_t pte = ptep_get_lockless(ptep); |
b0496fe4 MWO |
2923 | struct page *page; |
2924 | struct folio *folio; | |
2667f50e | 2925 | |
d74943a2 DH |
2926 | /* |
2927 | * Always fallback to ordinary GUP on PROT_NONE-mapped pages: | |
2928 | * pte_access_permitted() better should reject these pages | |
2929 | * either way: otherwise, GUP-fast might succeed in | |
2930 | * cases where ordinary GUP would fail due to VMA access | |
2931 | * permissions. | |
2932 | */ | |
2933 | if (pte_protnone(pte)) | |
e7884f8e KS |
2934 | goto pte_unmap; |
2935 | ||
b798bec4 | 2936 | if (!pte_access_permitted(pte, flags & FOLL_WRITE)) |
e7884f8e KS |
2937 | goto pte_unmap; |
2938 | ||
b59f65fa | 2939 | if (pte_devmap(pte)) { |
7af75561 IW |
2940 | if (unlikely(flags & FOLL_LONGTERM)) |
2941 | goto pte_unmap; | |
2942 | ||
b59f65fa KS |
2943 | pgmap = get_dev_pagemap(pte_pfn(pte), pgmap); |
2944 | if (unlikely(!pgmap)) { | |
23babe19 | 2945 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
b59f65fa KS |
2946 | goto pte_unmap; |
2947 | } | |
2948 | } else if (pte_special(pte)) | |
2667f50e SC |
2949 | goto pte_unmap; |
2950 | ||
2951 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
2952 | page = pte_page(pte); | |
2953 | ||
b0496fe4 MWO |
2954 | folio = try_grab_folio(page, 1, flags); |
2955 | if (!folio) | |
2667f50e SC |
2956 | goto pte_unmap; |
2957 | ||
70cbc3cc | 2958 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) || |
c33c7948 | 2959 | unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) { |
b0496fe4 | 2960 | gup_put_folio(folio, 1, flags); |
2667f50e SC |
2961 | goto pte_unmap; |
2962 | } | |
2963 | ||
f002882c | 2964 | if (!gup_fast_folio_allowed(folio, flags)) { |
b0496fe4 | 2965 | gup_put_folio(folio, 1, flags); |
2667f50e SC |
2966 | goto pte_unmap; |
2967 | } | |
2968 | ||
84209e87 | 2969 | if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) { |
a7f22660 DH |
2970 | gup_put_folio(folio, 1, flags); |
2971 | goto pte_unmap; | |
2972 | } | |
2973 | ||
f28d4363 CI |
2974 | /* |
2975 | * We need to make the page accessible if and only if we are | |
2976 | * going to access its content (the FOLL_PIN case). Please | |
2977 | * see Documentation/core-api/pin_user_pages.rst for | |
2978 | * details. | |
2979 | */ | |
2980 | if (flags & FOLL_PIN) { | |
2981 | ret = arch_make_page_accessible(page); | |
2982 | if (ret) { | |
b0496fe4 | 2983 | gup_put_folio(folio, 1, flags); |
f28d4363 CI |
2984 | goto pte_unmap; |
2985 | } | |
2986 | } | |
b0496fe4 | 2987 | folio_set_referenced(folio); |
2667f50e SC |
2988 | pages[*nr] = page; |
2989 | (*nr)++; | |
2667f50e SC |
2990 | } while (ptep++, addr += PAGE_SIZE, addr != end); |
2991 | ||
2992 | ret = 1; | |
2993 | ||
2994 | pte_unmap: | |
832d7aa0 CH |
2995 | if (pgmap) |
2996 | put_dev_pagemap(pgmap); | |
2667f50e SC |
2997 | pte_unmap(ptem); |
2998 | return ret; | |
2999 | } | |
3000 | #else | |
3001 | ||
3002 | /* | |
3003 | * If we can't determine whether or not a pte is special, then fail immediately | |
3004 | * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not | |
3005 | * to be special. | |
3006 | * | |
3007 | * For a futex to be placed on a THP tail page, get_futex_key requires a | |
dadbb612 | 3008 | * get_user_pages_fast_only implementation that can pin pages. Thus it's still |
23babe19 | 3009 | * useful to have gup_fast_pmd_leaf even if we can't operate on ptes. |
2667f50e | 3010 | */ |
23babe19 DH |
3011 | static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, |
3012 | unsigned long end, unsigned int flags, struct page **pages, | |
3013 | int *nr) | |
2667f50e SC |
3014 | { |
3015 | return 0; | |
3016 | } | |
3010a5ea | 3017 | #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */ |
2667f50e | 3018 | |
17596731 | 3019 | #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
23babe19 DH |
3020 | static int gup_fast_devmap_leaf(unsigned long pfn, unsigned long addr, |
3021 | unsigned long end, unsigned int flags, struct page **pages, int *nr) | |
b59f65fa KS |
3022 | { |
3023 | int nr_start = *nr; | |
3024 | struct dev_pagemap *pgmap = NULL; | |
3025 | ||
3026 | do { | |
3027 | struct page *page = pfn_to_page(pfn); | |
3028 | ||
3029 | pgmap = get_dev_pagemap(pfn, pgmap); | |
3030 | if (unlikely(!pgmap)) { | |
23babe19 | 3031 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
6401c4eb | 3032 | break; |
b59f65fa | 3033 | } |
4003f107 LG |
3034 | |
3035 | if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) { | |
23babe19 | 3036 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
4003f107 LG |
3037 | break; |
3038 | } | |
3039 | ||
b59f65fa KS |
3040 | SetPageReferenced(page); |
3041 | pages[*nr] = page; | |
0f089235 | 3042 | if (unlikely(try_grab_page(page, flags))) { |
23babe19 | 3043 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
6401c4eb | 3044 | break; |
3faa52c0 | 3045 | } |
b59f65fa KS |
3046 | (*nr)++; |
3047 | pfn++; | |
3048 | } while (addr += PAGE_SIZE, addr != end); | |
832d7aa0 | 3049 | |
6401c4eb | 3050 | put_dev_pagemap(pgmap); |
20b7fee7 | 3051 | return addr == end; |
b59f65fa KS |
3052 | } |
3053 | ||
23babe19 DH |
3054 | static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3055 | unsigned long end, unsigned int flags, struct page **pages, | |
3056 | int *nr) | |
b59f65fa KS |
3057 | { |
3058 | unsigned long fault_pfn; | |
a9b6de77 DW |
3059 | int nr_start = *nr; |
3060 | ||
3061 | fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
23babe19 | 3062 | if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 3063 | return 0; |
b59f65fa | 3064 | |
a9b6de77 | 3065 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { |
23babe19 | 3066 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
3067 | return 0; |
3068 | } | |
3069 | return 1; | |
b59f65fa KS |
3070 | } |
3071 | ||
23babe19 DH |
3072 | static int gup_fast_devmap_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr, |
3073 | unsigned long end, unsigned int flags, struct page **pages, | |
3074 | int *nr) | |
b59f65fa KS |
3075 | { |
3076 | unsigned long fault_pfn; | |
a9b6de77 DW |
3077 | int nr_start = *nr; |
3078 | ||
3079 | fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
23babe19 | 3080 | if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 3081 | return 0; |
b59f65fa | 3082 | |
a9b6de77 | 3083 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { |
23babe19 | 3084 | gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
3085 | return 0; |
3086 | } | |
3087 | return 1; | |
b59f65fa KS |
3088 | } |
3089 | #else | |
23babe19 DH |
3090 | static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3091 | unsigned long end, unsigned int flags, struct page **pages, | |
3092 | int *nr) | |
b59f65fa KS |
3093 | { |
3094 | BUILD_BUG(); | |
3095 | return 0; | |
3096 | } | |
3097 | ||
23babe19 DH |
3098 | static int gup_fast_devmap_pud_leaf(pud_t pud, pud_t *pudp, unsigned long addr, |
3099 | unsigned long end, unsigned int flags, struct page **pages, | |
3100 | int *nr) | |
b59f65fa KS |
3101 | { |
3102 | BUILD_BUG(); | |
3103 | return 0; | |
3104 | } | |
3105 | #endif | |
3106 | ||
23babe19 DH |
3107 | static int gup_fast_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
3108 | unsigned long end, unsigned int flags, struct page **pages, | |
3109 | int *nr) | |
2667f50e | 3110 | { |
667ed1f7 MWO |
3111 | struct page *page; |
3112 | struct folio *folio; | |
2667f50e SC |
3113 | int refs; |
3114 | ||
b798bec4 | 3115 | if (!pmd_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
3116 | return 0; |
3117 | ||
7af75561 IW |
3118 | if (pmd_devmap(orig)) { |
3119 | if (unlikely(flags & FOLL_LONGTERM)) | |
3120 | return 0; | |
23babe19 DH |
3121 | return gup_fast_devmap_pmd_leaf(orig, pmdp, addr, end, flags, |
3122 | pages, nr); | |
7af75561 | 3123 | } |
b59f65fa | 3124 | |
f3c94c62 PX |
3125 | page = pmd_page(orig); |
3126 | refs = record_subpages(page, PMD_SIZE, addr, end, pages + *nr); | |
2667f50e | 3127 | |
667ed1f7 MWO |
3128 | folio = try_grab_folio(page, refs, flags); |
3129 | if (!folio) | |
2667f50e | 3130 | return 0; |
2667f50e SC |
3131 | |
3132 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { | |
667ed1f7 | 3133 | gup_put_folio(folio, refs, flags); |
2667f50e SC |
3134 | return 0; |
3135 | } | |
3136 | ||
f002882c | 3137 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3138 | gup_put_folio(folio, refs, flags); |
3139 | return 0; | |
3140 | } | |
84209e87 | 3141 | if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
a7f22660 DH |
3142 | gup_put_folio(folio, refs, flags); |
3143 | return 0; | |
3144 | } | |
3145 | ||
a43e9820 | 3146 | *nr += refs; |
667ed1f7 | 3147 | folio_set_referenced(folio); |
2667f50e SC |
3148 | return 1; |
3149 | } | |
3150 | ||
23babe19 DH |
3151 | static int gup_fast_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr, |
3152 | unsigned long end, unsigned int flags, struct page **pages, | |
3153 | int *nr) | |
2667f50e | 3154 | { |
83afb52e MWO |
3155 | struct page *page; |
3156 | struct folio *folio; | |
2667f50e SC |
3157 | int refs; |
3158 | ||
b798bec4 | 3159 | if (!pud_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
3160 | return 0; |
3161 | ||
7af75561 IW |
3162 | if (pud_devmap(orig)) { |
3163 | if (unlikely(flags & FOLL_LONGTERM)) | |
3164 | return 0; | |
23babe19 DH |
3165 | return gup_fast_devmap_pud_leaf(orig, pudp, addr, end, flags, |
3166 | pages, nr); | |
7af75561 | 3167 | } |
b59f65fa | 3168 | |
f3c94c62 PX |
3169 | page = pud_page(orig); |
3170 | refs = record_subpages(page, PUD_SIZE, addr, end, pages + *nr); | |
2667f50e | 3171 | |
83afb52e MWO |
3172 | folio = try_grab_folio(page, refs, flags); |
3173 | if (!folio) | |
2667f50e | 3174 | return 0; |
2667f50e SC |
3175 | |
3176 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { | |
83afb52e | 3177 | gup_put_folio(folio, refs, flags); |
2667f50e SC |
3178 | return 0; |
3179 | } | |
3180 | ||
f002882c | 3181 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3182 | gup_put_folio(folio, refs, flags); |
3183 | return 0; | |
3184 | } | |
3185 | ||
84209e87 | 3186 | if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
a7f22660 DH |
3187 | gup_put_folio(folio, refs, flags); |
3188 | return 0; | |
3189 | } | |
3190 | ||
a43e9820 | 3191 | *nr += refs; |
83afb52e | 3192 | folio_set_referenced(folio); |
2667f50e SC |
3193 | return 1; |
3194 | } | |
3195 | ||
23babe19 DH |
3196 | static int gup_fast_pgd_leaf(pgd_t orig, pgd_t *pgdp, unsigned long addr, |
3197 | unsigned long end, unsigned int flags, struct page **pages, | |
3198 | int *nr) | |
f30c59e9 AK |
3199 | { |
3200 | int refs; | |
2d7919a2 MWO |
3201 | struct page *page; |
3202 | struct folio *folio; | |
f30c59e9 | 3203 | |
b798bec4 | 3204 | if (!pgd_access_permitted(orig, flags & FOLL_WRITE)) |
f30c59e9 AK |
3205 | return 0; |
3206 | ||
b59f65fa | 3207 | BUILD_BUG_ON(pgd_devmap(orig)); |
a43e9820 | 3208 | |
f3c94c62 PX |
3209 | page = pgd_page(orig); |
3210 | refs = record_subpages(page, PGDIR_SIZE, addr, end, pages + *nr); | |
f30c59e9 | 3211 | |
2d7919a2 MWO |
3212 | folio = try_grab_folio(page, refs, flags); |
3213 | if (!folio) | |
f30c59e9 | 3214 | return 0; |
f30c59e9 AK |
3215 | |
3216 | if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) { | |
2d7919a2 | 3217 | gup_put_folio(folio, refs, flags); |
f30c59e9 AK |
3218 | return 0; |
3219 | } | |
3220 | ||
31115034 LS |
3221 | if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) { |
3222 | gup_put_folio(folio, refs, flags); | |
3223 | return 0; | |
3224 | } | |
3225 | ||
f002882c | 3226 | if (!gup_fast_folio_allowed(folio, flags)) { |
a6e79df9 LS |
3227 | gup_put_folio(folio, refs, flags); |
3228 | return 0; | |
3229 | } | |
3230 | ||
a43e9820 | 3231 | *nr += refs; |
2d7919a2 | 3232 | folio_set_referenced(folio); |
f30c59e9 AK |
3233 | return 1; |
3234 | } | |
3235 | ||
23babe19 DH |
3236 | static int gup_fast_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, |
3237 | unsigned long end, unsigned int flags, struct page **pages, | |
3238 | int *nr) | |
2667f50e SC |
3239 | { |
3240 | unsigned long next; | |
3241 | pmd_t *pmdp; | |
3242 | ||
d3f7b1bb | 3243 | pmdp = pmd_offset_lockless(pudp, pud, addr); |
2667f50e | 3244 | do { |
1180e732 | 3245 | pmd_t pmd = pmdp_get_lockless(pmdp); |
2667f50e SC |
3246 | |
3247 | next = pmd_addr_end(addr, end); | |
84c3fc4e | 3248 | if (!pmd_present(pmd)) |
2667f50e SC |
3249 | return 0; |
3250 | ||
7db86dc3 | 3251 | if (unlikely(pmd_leaf(pmd))) { |
23babe19 | 3252 | /* See gup_fast_pte_range() */ |
d74943a2 | 3253 | if (pmd_protnone(pmd)) |
2667f50e SC |
3254 | return 0; |
3255 | ||
23babe19 | 3256 | if (!gup_fast_pmd_leaf(pmd, pmdp, addr, next, flags, |
2667f50e SC |
3257 | pages, nr)) |
3258 | return 0; | |
3259 | ||
f30c59e9 AK |
3260 | } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) { |
3261 | /* | |
3262 | * architecture have different format for hugetlbfs | |
3263 | * pmd format and THP pmd format | |
3264 | */ | |
23babe19 DH |
3265 | if (!gup_fast_hugepd(__hugepd(pmd_val(pmd)), addr, |
3266 | PMD_SHIFT, next, flags, pages, nr)) | |
f30c59e9 | 3267 | return 0; |
23babe19 DH |
3268 | } else if (!gup_fast_pte_range(pmd, pmdp, addr, next, flags, |
3269 | pages, nr)) | |
2923117b | 3270 | return 0; |
2667f50e SC |
3271 | } while (pmdp++, addr = next, addr != end); |
3272 | ||
3273 | return 1; | |
3274 | } | |
3275 | ||
23babe19 DH |
3276 | static int gup_fast_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, |
3277 | unsigned long end, unsigned int flags, struct page **pages, | |
3278 | int *nr) | |
2667f50e SC |
3279 | { |
3280 | unsigned long next; | |
3281 | pud_t *pudp; | |
3282 | ||
d3f7b1bb | 3283 | pudp = pud_offset_lockless(p4dp, p4d, addr); |
2667f50e | 3284 | do { |
e37c6982 | 3285 | pud_t pud = READ_ONCE(*pudp); |
2667f50e SC |
3286 | |
3287 | next = pud_addr_end(addr, end); | |
15494520 | 3288 | if (unlikely(!pud_present(pud))) |
2667f50e | 3289 | return 0; |
7db86dc3 | 3290 | if (unlikely(pud_leaf(pud))) { |
23babe19 DH |
3291 | if (!gup_fast_pud_leaf(pud, pudp, addr, next, flags, |
3292 | pages, nr)) | |
f30c59e9 AK |
3293 | return 0; |
3294 | } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) { | |
23babe19 DH |
3295 | if (!gup_fast_hugepd(__hugepd(pud_val(pud)), addr, |
3296 | PUD_SHIFT, next, flags, pages, nr)) | |
2667f50e | 3297 | return 0; |
23babe19 DH |
3298 | } else if (!gup_fast_pmd_range(pudp, pud, addr, next, flags, |
3299 | pages, nr)) | |
2667f50e SC |
3300 | return 0; |
3301 | } while (pudp++, addr = next, addr != end); | |
3302 | ||
3303 | return 1; | |
3304 | } | |
3305 | ||
23babe19 DH |
3306 | static int gup_fast_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, |
3307 | unsigned long end, unsigned int flags, struct page **pages, | |
3308 | int *nr) | |
c2febafc KS |
3309 | { |
3310 | unsigned long next; | |
3311 | p4d_t *p4dp; | |
3312 | ||
d3f7b1bb | 3313 | p4dp = p4d_offset_lockless(pgdp, pgd, addr); |
c2febafc KS |
3314 | do { |
3315 | p4d_t p4d = READ_ONCE(*p4dp); | |
3316 | ||
3317 | next = p4d_addr_end(addr, end); | |
089f9214 | 3318 | if (!p4d_present(p4d)) |
c2febafc | 3319 | return 0; |
1965e933 | 3320 | BUILD_BUG_ON(p4d_leaf(p4d)); |
c2febafc | 3321 | if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) { |
23babe19 DH |
3322 | if (!gup_fast_hugepd(__hugepd(p4d_val(p4d)), addr, |
3323 | P4D_SHIFT, next, flags, pages, nr)) | |
c2febafc | 3324 | return 0; |
23babe19 DH |
3325 | } else if (!gup_fast_pud_range(p4dp, p4d, addr, next, flags, |
3326 | pages, nr)) | |
c2febafc KS |
3327 | return 0; |
3328 | } while (p4dp++, addr = next, addr != end); | |
3329 | ||
3330 | return 1; | |
3331 | } | |
3332 | ||
23babe19 | 3333 | static void gup_fast_pgd_range(unsigned long addr, unsigned long end, |
b798bec4 | 3334 | unsigned int flags, struct page **pages, int *nr) |
5b65c467 KS |
3335 | { |
3336 | unsigned long next; | |
3337 | pgd_t *pgdp; | |
3338 | ||
3339 | pgdp = pgd_offset(current->mm, addr); | |
3340 | do { | |
3341 | pgd_t pgd = READ_ONCE(*pgdp); | |
3342 | ||
3343 | next = pgd_addr_end(addr, end); | |
3344 | if (pgd_none(pgd)) | |
3345 | return; | |
7db86dc3 | 3346 | if (unlikely(pgd_leaf(pgd))) { |
23babe19 DH |
3347 | if (!gup_fast_pgd_leaf(pgd, pgdp, addr, next, flags, |
3348 | pages, nr)) | |
5b65c467 KS |
3349 | return; |
3350 | } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) { | |
23babe19 DH |
3351 | if (!gup_fast_hugepd(__hugepd(pgd_val(pgd)), addr, |
3352 | PGDIR_SHIFT, next, flags, pages, nr)) | |
5b65c467 | 3353 | return; |
23babe19 DH |
3354 | } else if (!gup_fast_p4d_range(pgdp, pgd, addr, next, flags, |
3355 | pages, nr)) | |
5b65c467 KS |
3356 | return; |
3357 | } while (pgdp++, addr = next, addr != end); | |
3358 | } | |
050a9adc | 3359 | #else |
23babe19 | 3360 | static inline void gup_fast_pgd_range(unsigned long addr, unsigned long end, |
050a9adc CH |
3361 | unsigned int flags, struct page **pages, int *nr) |
3362 | { | |
3363 | } | |
25176ad0 | 3364 | #endif /* CONFIG_HAVE_GUP_FAST */ |
5b65c467 KS |
3365 | |
3366 | #ifndef gup_fast_permitted | |
3367 | /* | |
dadbb612 | 3368 | * Check if it's allowed to use get_user_pages_fast_only() for the range, or |
5b65c467 KS |
3369 | * we need to fall back to the slow version: |
3370 | */ | |
26f4c328 | 3371 | static bool gup_fast_permitted(unsigned long start, unsigned long end) |
5b65c467 | 3372 | { |
26f4c328 | 3373 | return true; |
5b65c467 KS |
3374 | } |
3375 | #endif | |
3376 | ||
23babe19 DH |
3377 | static unsigned long gup_fast(unsigned long start, unsigned long end, |
3378 | unsigned int gup_flags, struct page **pages) | |
c28b1fc7 JG |
3379 | { |
3380 | unsigned long flags; | |
3381 | int nr_pinned = 0; | |
57efa1fe | 3382 | unsigned seq; |
c28b1fc7 | 3383 | |
25176ad0 | 3384 | if (!IS_ENABLED(CONFIG_HAVE_GUP_FAST) || |
c28b1fc7 JG |
3385 | !gup_fast_permitted(start, end)) |
3386 | return 0; | |
3387 | ||
57efa1fe JG |
3388 | if (gup_flags & FOLL_PIN) { |
3389 | seq = raw_read_seqcount(¤t->mm->write_protect_seq); | |
3390 | if (seq & 1) | |
3391 | return 0; | |
3392 | } | |
3393 | ||
c28b1fc7 JG |
3394 | /* |
3395 | * Disable interrupts. The nested form is used, in order to allow full, | |
3396 | * general purpose use of this routine. | |
3397 | * | |
3398 | * With interrupts disabled, we block page table pages from being freed | |
3399 | * from under us. See struct mmu_table_batch comments in | |
3400 | * include/asm-generic/tlb.h for more details. | |
3401 | * | |
3402 | * We do not adopt an rcu_read_lock() here as we also want to block IPIs | |
3403 | * that come from THPs splitting. | |
3404 | */ | |
3405 | local_irq_save(flags); | |
23babe19 | 3406 | gup_fast_pgd_range(start, end, gup_flags, pages, &nr_pinned); |
c28b1fc7 | 3407 | local_irq_restore(flags); |
57efa1fe JG |
3408 | |
3409 | /* | |
3410 | * When pinning pages for DMA there could be a concurrent write protect | |
23babe19 | 3411 | * from fork() via copy_page_range(), in this case always fail GUP-fast. |
57efa1fe JG |
3412 | */ |
3413 | if (gup_flags & FOLL_PIN) { | |
3414 | if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) { | |
23babe19 | 3415 | gup_fast_unpin_user_pages(pages, nr_pinned); |
57efa1fe | 3416 | return 0; |
b6a2619c DH |
3417 | } else { |
3418 | sanity_check_pinned_pages(pages, nr_pinned); | |
57efa1fe JG |
3419 | } |
3420 | } | |
c28b1fc7 JG |
3421 | return nr_pinned; |
3422 | } | |
3423 | ||
23babe19 DH |
3424 | static int gup_fast_fallback(unsigned long start, unsigned long nr_pages, |
3425 | unsigned int gup_flags, struct page **pages) | |
2667f50e | 3426 | { |
c28b1fc7 JG |
3427 | unsigned long len, end; |
3428 | unsigned long nr_pinned; | |
b2a72dff | 3429 | int locked = 0; |
c28b1fc7 | 3430 | int ret; |
2667f50e | 3431 | |
f4000fdf | 3432 | if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM | |
376a34ef | 3433 | FOLL_FORCE | FOLL_PIN | FOLL_GET | |
4003f107 | 3434 | FOLL_FAST_ONLY | FOLL_NOFAULT | |
d74943a2 | 3435 | FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT))) |
817be129 CH |
3436 | return -EINVAL; |
3437 | ||
a458b76a AA |
3438 | if (gup_flags & FOLL_PIN) |
3439 | mm_set_has_pinned_flag(¤t->mm->flags); | |
008cfe44 | 3440 | |
f81cd178 | 3441 | if (!(gup_flags & FOLL_FAST_ONLY)) |
da1c55f1 | 3442 | might_lock_read(¤t->mm->mmap_lock); |
f81cd178 | 3443 | |
f455c854 | 3444 | start = untagged_addr(start) & PAGE_MASK; |
c28b1fc7 JG |
3445 | len = nr_pages << PAGE_SHIFT; |
3446 | if (check_add_overflow(start, len, &end)) | |
9883c7f8 | 3447 | return -EOVERFLOW; |
6014bc27 LT |
3448 | if (end > TASK_SIZE_MAX) |
3449 | return -EFAULT; | |
96d4f267 | 3450 | if (unlikely(!access_ok((void __user *)start, len))) |
c61611f7 | 3451 | return -EFAULT; |
73e10a61 | 3452 | |
23babe19 | 3453 | nr_pinned = gup_fast(start, end, gup_flags, pages); |
c28b1fc7 JG |
3454 | if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY) |
3455 | return nr_pinned; | |
2667f50e | 3456 | |
c28b1fc7 JG |
3457 | /* Slow path: try to get the remaining pages with get_user_pages */ |
3458 | start += nr_pinned << PAGE_SHIFT; | |
3459 | pages += nr_pinned; | |
b2a72dff | 3460 | ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned, |
b2cac248 | 3461 | pages, &locked, |
f04740f5 | 3462 | gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE); |
c28b1fc7 JG |
3463 | if (ret < 0) { |
3464 | /* | |
3465 | * The caller has to unpin the pages we already pinned so | |
3466 | * returning -errno is not an option | |
3467 | */ | |
3468 | if (nr_pinned) | |
3469 | return nr_pinned; | |
3470 | return ret; | |
2667f50e | 3471 | } |
c28b1fc7 | 3472 | return ret + nr_pinned; |
2667f50e | 3473 | } |
c28b1fc7 | 3474 | |
dadbb612 SJ |
3475 | /** |
3476 | * get_user_pages_fast_only() - pin user pages in memory | |
3477 | * @start: starting user address | |
3478 | * @nr_pages: number of pages from start to pin | |
3479 | * @gup_flags: flags modifying pin behaviour | |
3480 | * @pages: array that receives pointers to the pages pinned. | |
3481 | * Should be at least nr_pages long. | |
3482 | * | |
9e1f0580 JH |
3483 | * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to |
3484 | * the regular GUP. | |
9e1f0580 JH |
3485 | * |
3486 | * If the architecture does not support this function, simply return with no | |
3487 | * pages pinned. | |
3488 | * | |
3489 | * Careful, careful! COW breaking can go either way, so a non-write | |
3490 | * access can get ambiguous page results. If you call this function without | |
3491 | * 'write' set, you'd better be sure that you're ok with that ambiguity. | |
3492 | */ | |
dadbb612 SJ |
3493 | int get_user_pages_fast_only(unsigned long start, int nr_pages, |
3494 | unsigned int gup_flags, struct page **pages) | |
9e1f0580 | 3495 | { |
9e1f0580 JH |
3496 | /* |
3497 | * Internally (within mm/gup.c), gup fast variants must set FOLL_GET, | |
3498 | * because gup fast is always a "pin with a +1 page refcount" request. | |
376a34ef JH |
3499 | * |
3500 | * FOLL_FAST_ONLY is required in order to match the API description of | |
3501 | * this routine: no fall back to regular ("slow") GUP. | |
9e1f0580 | 3502 | */ |
b2cac248 | 3503 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
d64e2dbc JG |
3504 | FOLL_GET | FOLL_FAST_ONLY)) |
3505 | return -EINVAL; | |
9e1f0580 | 3506 | |
23babe19 | 3507 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
9e1f0580 | 3508 | } |
dadbb612 | 3509 | EXPORT_SYMBOL_GPL(get_user_pages_fast_only); |
9e1f0580 | 3510 | |
eddb1c22 JH |
3511 | /** |
3512 | * get_user_pages_fast() - pin user pages in memory | |
3faa52c0 JH |
3513 | * @start: starting user address |
3514 | * @nr_pages: number of pages from start to pin | |
3515 | * @gup_flags: flags modifying pin behaviour | |
3516 | * @pages: array that receives pointers to the pages pinned. | |
3517 | * Should be at least nr_pages long. | |
eddb1c22 | 3518 | * |
c1e8d7c6 | 3519 | * Attempt to pin user pages in memory without taking mm->mmap_lock. |
eddb1c22 JH |
3520 | * If not successful, it will fall back to taking the lock and |
3521 | * calling get_user_pages(). | |
3522 | * | |
3523 | * Returns number of pages pinned. This may be fewer than the number requested. | |
3524 | * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns | |
3525 | * -errno. | |
3526 | */ | |
3527 | int get_user_pages_fast(unsigned long start, int nr_pages, | |
3528 | unsigned int gup_flags, struct page **pages) | |
3529 | { | |
94202f12 JH |
3530 | /* |
3531 | * The caller may or may not have explicitly set FOLL_GET; either way is | |
3532 | * OK. However, internally (within mm/gup.c), gup fast variants must set | |
3533 | * FOLL_GET, because gup fast is always a "pin with a +1 page refcount" | |
3534 | * request. | |
3535 | */ | |
b2cac248 | 3536 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET)) |
d64e2dbc | 3537 | return -EINVAL; |
23babe19 | 3538 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
eddb1c22 | 3539 | } |
050a9adc | 3540 | EXPORT_SYMBOL_GPL(get_user_pages_fast); |
eddb1c22 JH |
3541 | |
3542 | /** | |
3543 | * pin_user_pages_fast() - pin user pages in memory without taking locks | |
3544 | * | |
3faa52c0 JH |
3545 | * @start: starting user address |
3546 | * @nr_pages: number of pages from start to pin | |
3547 | * @gup_flags: flags modifying pin behaviour | |
3548 | * @pages: array that receives pointers to the pages pinned. | |
3549 | * Should be at least nr_pages long. | |
3550 | * | |
3551 | * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See | |
3552 | * get_user_pages_fast() for documentation on the function arguments, because | |
3553 | * the arguments here are identical. | |
3554 | * | |
3555 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3556 | * see Documentation/core-api/pin_user_pages.rst for further details. |
c8070b78 DH |
3557 | * |
3558 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3559 | * pins in it and unpin_user_page() will not remove pins from it. | |
eddb1c22 JH |
3560 | */ |
3561 | int pin_user_pages_fast(unsigned long start, int nr_pages, | |
3562 | unsigned int gup_flags, struct page **pages) | |
3563 | { | |
b2cac248 | 3564 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN)) |
3faa52c0 | 3565 | return -EINVAL; |
23babe19 | 3566 | return gup_fast_fallback(start, nr_pages, gup_flags, pages); |
eddb1c22 JH |
3567 | } |
3568 | EXPORT_SYMBOL_GPL(pin_user_pages_fast); | |
3569 | ||
3570 | /** | |
64019a2e | 3571 | * pin_user_pages_remote() - pin pages of a remote process |
eddb1c22 | 3572 | * |
3faa52c0 JH |
3573 | * @mm: mm_struct of target mm |
3574 | * @start: starting user address | |
3575 | * @nr_pages: number of pages from start to pin | |
3576 | * @gup_flags: flags modifying lookup behaviour | |
3577 | * @pages: array that receives pointers to the pages pinned. | |
0768c8de | 3578 | * Should be at least nr_pages long. |
3faa52c0 JH |
3579 | * @locked: pointer to lock flag indicating whether lock is held and |
3580 | * subsequently whether VM_FAULT_RETRY functionality can be | |
3581 | * utilised. Lock must initially be held. | |
3582 | * | |
3583 | * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See | |
3584 | * get_user_pages_remote() for documentation on the function arguments, because | |
3585 | * the arguments here are identical. | |
3586 | * | |
3587 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3588 | * see Documentation/core-api/pin_user_pages.rst for details. |
c8070b78 DH |
3589 | * |
3590 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3591 | * pins in it and unpin_user_page*() will not remove pins from it. | |
eddb1c22 | 3592 | */ |
64019a2e | 3593 | long pin_user_pages_remote(struct mm_struct *mm, |
eddb1c22 JH |
3594 | unsigned long start, unsigned long nr_pages, |
3595 | unsigned int gup_flags, struct page **pages, | |
0b295316 | 3596 | int *locked) |
eddb1c22 | 3597 | { |
9a863a6a JG |
3598 | int local_locked = 1; |
3599 | ||
b2cac248 | 3600 | if (!is_valid_gup_args(pages, locked, &gup_flags, |
d64e2dbc JG |
3601 | FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE)) |
3602 | return 0; | |
b2cac248 | 3603 | return __gup_longterm_locked(mm, start, nr_pages, pages, |
9a863a6a | 3604 | locked ? locked : &local_locked, |
d64e2dbc | 3605 | gup_flags); |
eddb1c22 JH |
3606 | } |
3607 | EXPORT_SYMBOL(pin_user_pages_remote); | |
3608 | ||
3609 | /** | |
3610 | * pin_user_pages() - pin user pages in memory for use by other devices | |
3611 | * | |
3faa52c0 JH |
3612 | * @start: starting user address |
3613 | * @nr_pages: number of pages from start to pin | |
3614 | * @gup_flags: flags modifying lookup behaviour | |
3615 | * @pages: array that receives pointers to the pages pinned. | |
0768c8de | 3616 | * Should be at least nr_pages long. |
3faa52c0 JH |
3617 | * |
3618 | * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and | |
3619 | * FOLL_PIN is set. | |
3620 | * | |
3621 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
72ef5e52 | 3622 | * see Documentation/core-api/pin_user_pages.rst for details. |
c8070b78 DH |
3623 | * |
3624 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3625 | * pins in it and unpin_user_page*() will not remove pins from it. | |
eddb1c22 JH |
3626 | */ |
3627 | long pin_user_pages(unsigned long start, unsigned long nr_pages, | |
4c630f30 | 3628 | unsigned int gup_flags, struct page **pages) |
eddb1c22 | 3629 | { |
9a863a6a JG |
3630 | int locked = 1; |
3631 | ||
b2cac248 | 3632 | if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN)) |
d64e2dbc | 3633 | return 0; |
64019a2e | 3634 | return __gup_longterm_locked(current->mm, start, nr_pages, |
b2cac248 | 3635 | pages, &locked, gup_flags); |
eddb1c22 JH |
3636 | } |
3637 | EXPORT_SYMBOL(pin_user_pages); | |
91429023 JH |
3638 | |
3639 | /* | |
3640 | * pin_user_pages_unlocked() is the FOLL_PIN variant of | |
3641 | * get_user_pages_unlocked(). Behavior is the same, except that this one sets | |
3642 | * FOLL_PIN and rejects FOLL_GET. | |
c8070b78 DH |
3643 | * |
3644 | * Note that if a zero_page is amongst the returned pages, it will not have | |
3645 | * pins in it and unpin_user_page*() will not remove pins from it. | |
91429023 JH |
3646 | */ |
3647 | long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages, | |
3648 | struct page **pages, unsigned int gup_flags) | |
3649 | { | |
b2a72dff | 3650 | int locked = 0; |
91429023 | 3651 | |
b2cac248 | 3652 | if (!is_valid_gup_args(pages, NULL, &gup_flags, |
f04740f5 | 3653 | FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE)) |
d64e2dbc | 3654 | return 0; |
0768c8de | 3655 | |
b2cac248 | 3656 | return __gup_longterm_locked(current->mm, start, nr_pages, pages, |
b2a72dff | 3657 | &locked, gup_flags); |
91429023 JH |
3658 | } |
3659 | EXPORT_SYMBOL(pin_user_pages_unlocked); |