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> | |
13 | ||
174cd4b1 | 14 | #include <linux/sched/signal.h> |
2667f50e | 15 | #include <linux/rwsem.h> |
f30c59e9 | 16 | #include <linux/hugetlb.h> |
9a4e9f3b AK |
17 | #include <linux/migrate.h> |
18 | #include <linux/mm_inline.h> | |
19 | #include <linux/sched/mm.h> | |
1027e443 | 20 | |
33a709b2 | 21 | #include <asm/mmu_context.h> |
2667f50e | 22 | #include <asm/pgtable.h> |
1027e443 | 23 | #include <asm/tlbflush.h> |
2667f50e | 24 | |
4bbd4c77 KS |
25 | #include "internal.h" |
26 | ||
df06b37f KB |
27 | struct follow_page_context { |
28 | struct dev_pagemap *pgmap; | |
29 | unsigned int page_mask; | |
30 | }; | |
31 | ||
47e29d32 JH |
32 | static void hpage_pincount_add(struct page *page, int refs) |
33 | { | |
34 | VM_BUG_ON_PAGE(!hpage_pincount_available(page), page); | |
35 | VM_BUG_ON_PAGE(page != compound_head(page), page); | |
36 | ||
37 | atomic_add(refs, compound_pincount_ptr(page)); | |
38 | } | |
39 | ||
40 | static void hpage_pincount_sub(struct page *page, int refs) | |
41 | { | |
42 | VM_BUG_ON_PAGE(!hpage_pincount_available(page), page); | |
43 | VM_BUG_ON_PAGE(page != compound_head(page), page); | |
44 | ||
45 | atomic_sub(refs, compound_pincount_ptr(page)); | |
46 | } | |
47 | ||
a707cdd5 JH |
48 | /* |
49 | * Return the compound head page with ref appropriately incremented, | |
50 | * or NULL if that failed. | |
51 | */ | |
52 | static inline struct page *try_get_compound_head(struct page *page, int refs) | |
53 | { | |
54 | struct page *head = compound_head(page); | |
55 | ||
56 | if (WARN_ON_ONCE(page_ref_count(head) < 0)) | |
57 | return NULL; | |
58 | if (unlikely(!page_cache_add_speculative(head, refs))) | |
59 | return NULL; | |
60 | return head; | |
61 | } | |
62 | ||
3faa52c0 JH |
63 | /* |
64 | * try_grab_compound_head() - attempt to elevate a page's refcount, by a | |
65 | * flags-dependent amount. | |
66 | * | |
67 | * "grab" names in this file mean, "look at flags to decide whether to use | |
68 | * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount. | |
69 | * | |
70 | * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the | |
71 | * same time. (That's true throughout the get_user_pages*() and | |
72 | * pin_user_pages*() APIs.) Cases: | |
73 | * | |
74 | * FOLL_GET: page's refcount will be incremented by 1. | |
75 | * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS. | |
76 | * | |
77 | * Return: head page (with refcount appropriately incremented) for success, or | |
78 | * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's | |
79 | * considered failure, and furthermore, a likely bug in the caller, so a warning | |
80 | * is also emitted. | |
81 | */ | |
82 | static __maybe_unused struct page *try_grab_compound_head(struct page *page, | |
83 | int refs, | |
84 | unsigned int flags) | |
85 | { | |
86 | if (flags & FOLL_GET) | |
87 | return try_get_compound_head(page, refs); | |
88 | else if (flags & FOLL_PIN) { | |
1970dc6f JH |
89 | int orig_refs = refs; |
90 | ||
47e29d32 JH |
91 | /* |
92 | * When pinning a compound page of order > 1 (which is what | |
93 | * hpage_pincount_available() checks for), use an exact count to | |
94 | * track it, via hpage_pincount_add/_sub(). | |
95 | * | |
96 | * However, be sure to *also* increment the normal page refcount | |
97 | * field at least once, so that the page really is pinned. | |
98 | */ | |
99 | if (!hpage_pincount_available(page)) | |
100 | refs *= GUP_PIN_COUNTING_BIAS; | |
101 | ||
102 | page = try_get_compound_head(page, refs); | |
103 | if (!page) | |
104 | return NULL; | |
105 | ||
106 | if (hpage_pincount_available(page)) | |
107 | hpage_pincount_add(page, refs); | |
108 | ||
1970dc6f JH |
109 | mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, |
110 | orig_refs); | |
111 | ||
47e29d32 | 112 | return page; |
3faa52c0 JH |
113 | } |
114 | ||
115 | WARN_ON_ONCE(1); | |
116 | return NULL; | |
117 | } | |
118 | ||
119 | /** | |
120 | * try_grab_page() - elevate a page's refcount by a flag-dependent amount | |
121 | * | |
122 | * This might not do anything at all, depending on the flags argument. | |
123 | * | |
124 | * "grab" names in this file mean, "look at flags to decide whether to use | |
125 | * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount. | |
126 | * | |
127 | * @page: pointer to page to be grabbed | |
128 | * @flags: gup flags: these are the FOLL_* flag values. | |
129 | * | |
130 | * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same | |
131 | * time. Cases: | |
132 | * | |
133 | * FOLL_GET: page's refcount will be incremented by 1. | |
134 | * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS. | |
135 | * | |
136 | * Return: true for success, or if no action was required (if neither FOLL_PIN | |
137 | * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or | |
138 | * FOLL_PIN was set, but the page could not be grabbed. | |
139 | */ | |
140 | bool __must_check try_grab_page(struct page *page, unsigned int flags) | |
141 | { | |
142 | WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN)); | |
143 | ||
144 | if (flags & FOLL_GET) | |
145 | return try_get_page(page); | |
146 | else if (flags & FOLL_PIN) { | |
47e29d32 JH |
147 | int refs = 1; |
148 | ||
3faa52c0 JH |
149 | page = compound_head(page); |
150 | ||
151 | if (WARN_ON_ONCE(page_ref_count(page) <= 0)) | |
152 | return false; | |
153 | ||
47e29d32 JH |
154 | if (hpage_pincount_available(page)) |
155 | hpage_pincount_add(page, 1); | |
156 | else | |
157 | refs = GUP_PIN_COUNTING_BIAS; | |
158 | ||
159 | /* | |
160 | * Similar to try_grab_compound_head(): even if using the | |
161 | * hpage_pincount_add/_sub() routines, be sure to | |
162 | * *also* increment the normal page refcount field at least | |
163 | * once, so that the page really is pinned. | |
164 | */ | |
165 | page_ref_add(page, refs); | |
1970dc6f JH |
166 | |
167 | mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1); | |
3faa52c0 JH |
168 | } |
169 | ||
170 | return true; | |
171 | } | |
172 | ||
173 | #ifdef CONFIG_DEV_PAGEMAP_OPS | |
174 | static bool __unpin_devmap_managed_user_page(struct page *page) | |
175 | { | |
47e29d32 | 176 | int count, refs = 1; |
3faa52c0 JH |
177 | |
178 | if (!page_is_devmap_managed(page)) | |
179 | return false; | |
180 | ||
47e29d32 JH |
181 | if (hpage_pincount_available(page)) |
182 | hpage_pincount_sub(page, 1); | |
183 | else | |
184 | refs = GUP_PIN_COUNTING_BIAS; | |
185 | ||
186 | count = page_ref_sub_return(page, refs); | |
3faa52c0 | 187 | |
1970dc6f | 188 | mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1); |
3faa52c0 JH |
189 | /* |
190 | * devmap page refcounts are 1-based, rather than 0-based: if | |
191 | * refcount is 1, then the page is free and the refcount is | |
192 | * stable because nobody holds a reference on the page. | |
193 | */ | |
194 | if (count == 1) | |
195 | free_devmap_managed_page(page); | |
196 | else if (!count) | |
197 | __put_page(page); | |
198 | ||
199 | return true; | |
200 | } | |
201 | #else | |
202 | static bool __unpin_devmap_managed_user_page(struct page *page) | |
203 | { | |
204 | return false; | |
205 | } | |
206 | #endif /* CONFIG_DEV_PAGEMAP_OPS */ | |
207 | ||
208 | /** | |
209 | * unpin_user_page() - release a dma-pinned page | |
210 | * @page: pointer to page to be released | |
211 | * | |
212 | * Pages that were pinned via pin_user_pages*() must be released via either | |
213 | * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so | |
214 | * that such pages can be separately tracked and uniquely handled. In | |
215 | * particular, interactions with RDMA and filesystems need special handling. | |
216 | */ | |
217 | void unpin_user_page(struct page *page) | |
218 | { | |
47e29d32 JH |
219 | int refs = 1; |
220 | ||
3faa52c0 JH |
221 | page = compound_head(page); |
222 | ||
223 | /* | |
224 | * For devmap managed pages we need to catch refcount transition from | |
225 | * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the | |
226 | * page is free and we need to inform the device driver through | |
227 | * callback. See include/linux/memremap.h and HMM for details. | |
228 | */ | |
229 | if (__unpin_devmap_managed_user_page(page)) | |
230 | return; | |
231 | ||
47e29d32 JH |
232 | if (hpage_pincount_available(page)) |
233 | hpage_pincount_sub(page, 1); | |
234 | else | |
235 | refs = GUP_PIN_COUNTING_BIAS; | |
236 | ||
237 | if (page_ref_sub_and_test(page, refs)) | |
3faa52c0 | 238 | __put_page(page); |
1970dc6f JH |
239 | |
240 | mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1); | |
3faa52c0 JH |
241 | } |
242 | EXPORT_SYMBOL(unpin_user_page); | |
243 | ||
fc1d8e7c | 244 | /** |
f1f6a7dd | 245 | * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages |
2d15eb31 | 246 | * @pages: array of pages to be maybe marked dirty, and definitely released. |
fc1d8e7c | 247 | * @npages: number of pages in the @pages array. |
2d15eb31 | 248 | * @make_dirty: whether to mark the pages dirty |
fc1d8e7c JH |
249 | * |
250 | * "gup-pinned page" refers to a page that has had one of the get_user_pages() | |
251 | * variants called on that page. | |
252 | * | |
253 | * For each page in the @pages array, make that page (or its head page, if a | |
2d15eb31 | 254 | * compound page) dirty, if @make_dirty is true, and if the page was previously |
f1f6a7dd JH |
255 | * listed as clean. In any case, releases all pages using unpin_user_page(), |
256 | * possibly via unpin_user_pages(), for the non-dirty case. | |
fc1d8e7c | 257 | * |
f1f6a7dd | 258 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 259 | * |
2d15eb31 | 260 | * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is |
261 | * required, then the caller should a) verify that this is really correct, | |
262 | * because _lock() is usually required, and b) hand code it: | |
f1f6a7dd | 263 | * set_page_dirty_lock(), unpin_user_page(). |
fc1d8e7c JH |
264 | * |
265 | */ | |
f1f6a7dd JH |
266 | void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages, |
267 | bool make_dirty) | |
fc1d8e7c | 268 | { |
2d15eb31 | 269 | unsigned long index; |
fc1d8e7c | 270 | |
2d15eb31 | 271 | /* |
272 | * TODO: this can be optimized for huge pages: if a series of pages is | |
273 | * physically contiguous and part of the same compound page, then a | |
274 | * single operation to the head page should suffice. | |
275 | */ | |
276 | ||
277 | if (!make_dirty) { | |
f1f6a7dd | 278 | unpin_user_pages(pages, npages); |
2d15eb31 | 279 | return; |
280 | } | |
281 | ||
282 | for (index = 0; index < npages; index++) { | |
283 | struct page *page = compound_head(pages[index]); | |
284 | /* | |
285 | * Checking PageDirty at this point may race with | |
286 | * clear_page_dirty_for_io(), but that's OK. Two key | |
287 | * cases: | |
288 | * | |
289 | * 1) This code sees the page as already dirty, so it | |
290 | * skips the call to set_page_dirty(). That could happen | |
291 | * because clear_page_dirty_for_io() called | |
292 | * page_mkclean(), followed by set_page_dirty(). | |
293 | * However, now the page is going to get written back, | |
294 | * which meets the original intention of setting it | |
295 | * dirty, so all is well: clear_page_dirty_for_io() goes | |
296 | * on to call TestClearPageDirty(), and write the page | |
297 | * back. | |
298 | * | |
299 | * 2) This code sees the page as clean, so it calls | |
300 | * set_page_dirty(). The page stays dirty, despite being | |
301 | * written back, so it gets written back again in the | |
302 | * next writeback cycle. This is harmless. | |
303 | */ | |
304 | if (!PageDirty(page)) | |
305 | set_page_dirty_lock(page); | |
f1f6a7dd | 306 | unpin_user_page(page); |
2d15eb31 | 307 | } |
fc1d8e7c | 308 | } |
f1f6a7dd | 309 | EXPORT_SYMBOL(unpin_user_pages_dirty_lock); |
fc1d8e7c JH |
310 | |
311 | /** | |
f1f6a7dd | 312 | * unpin_user_pages() - release an array of gup-pinned pages. |
fc1d8e7c JH |
313 | * @pages: array of pages to be marked dirty and released. |
314 | * @npages: number of pages in the @pages array. | |
315 | * | |
f1f6a7dd | 316 | * For each page in the @pages array, release the page using unpin_user_page(). |
fc1d8e7c | 317 | * |
f1f6a7dd | 318 | * Please see the unpin_user_page() documentation for details. |
fc1d8e7c | 319 | */ |
f1f6a7dd | 320 | void unpin_user_pages(struct page **pages, unsigned long npages) |
fc1d8e7c JH |
321 | { |
322 | unsigned long index; | |
323 | ||
324 | /* | |
325 | * TODO: this can be optimized for huge pages: if a series of pages is | |
326 | * physically contiguous and part of the same compound page, then a | |
327 | * single operation to the head page should suffice. | |
328 | */ | |
329 | for (index = 0; index < npages; index++) | |
f1f6a7dd | 330 | unpin_user_page(pages[index]); |
fc1d8e7c | 331 | } |
f1f6a7dd | 332 | EXPORT_SYMBOL(unpin_user_pages); |
fc1d8e7c | 333 | |
050a9adc | 334 | #ifdef CONFIG_MMU |
69e68b4f KS |
335 | static struct page *no_page_table(struct vm_area_struct *vma, |
336 | unsigned int flags) | |
4bbd4c77 | 337 | { |
69e68b4f KS |
338 | /* |
339 | * When core dumping an enormous anonymous area that nobody | |
340 | * has touched so far, we don't want to allocate unnecessary pages or | |
341 | * page tables. Return error instead of NULL to skip handle_mm_fault, | |
342 | * then get_dump_page() will return NULL to leave a hole in the dump. | |
343 | * But we can only make this optimization where a hole would surely | |
344 | * be zero-filled if handle_mm_fault() actually did handle it. | |
345 | */ | |
346 | if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault)) | |
347 | return ERR_PTR(-EFAULT); | |
348 | return NULL; | |
349 | } | |
4bbd4c77 | 350 | |
1027e443 KS |
351 | static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address, |
352 | pte_t *pte, unsigned int flags) | |
353 | { | |
354 | /* No page to get reference */ | |
355 | if (flags & FOLL_GET) | |
356 | return -EFAULT; | |
357 | ||
358 | if (flags & FOLL_TOUCH) { | |
359 | pte_t entry = *pte; | |
360 | ||
361 | if (flags & FOLL_WRITE) | |
362 | entry = pte_mkdirty(entry); | |
363 | entry = pte_mkyoung(entry); | |
364 | ||
365 | if (!pte_same(*pte, entry)) { | |
366 | set_pte_at(vma->vm_mm, address, pte, entry); | |
367 | update_mmu_cache(vma, address, pte); | |
368 | } | |
369 | } | |
370 | ||
371 | /* Proper page table entry exists, but no corresponding struct page */ | |
372 | return -EEXIST; | |
373 | } | |
374 | ||
19be0eaf LT |
375 | /* |
376 | * FOLL_FORCE can write to even unwritable pte's, but only | |
377 | * after we've gone through a COW cycle and they are dirty. | |
378 | */ | |
379 | static inline bool can_follow_write_pte(pte_t pte, unsigned int flags) | |
380 | { | |
f6f37321 | 381 | return pte_write(pte) || |
19be0eaf LT |
382 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte)); |
383 | } | |
384 | ||
69e68b4f | 385 | static struct page *follow_page_pte(struct vm_area_struct *vma, |
df06b37f KB |
386 | unsigned long address, pmd_t *pmd, unsigned int flags, |
387 | struct dev_pagemap **pgmap) | |
69e68b4f KS |
388 | { |
389 | struct mm_struct *mm = vma->vm_mm; | |
390 | struct page *page; | |
391 | spinlock_t *ptl; | |
392 | pte_t *ptep, pte; | |
f28d4363 | 393 | int ret; |
4bbd4c77 | 394 | |
eddb1c22 JH |
395 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
396 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
397 | (FOLL_PIN | FOLL_GET))) | |
398 | return ERR_PTR(-EINVAL); | |
69e68b4f | 399 | retry: |
4bbd4c77 | 400 | if (unlikely(pmd_bad(*pmd))) |
69e68b4f | 401 | return no_page_table(vma, flags); |
4bbd4c77 KS |
402 | |
403 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
4bbd4c77 KS |
404 | pte = *ptep; |
405 | if (!pte_present(pte)) { | |
406 | swp_entry_t entry; | |
407 | /* | |
408 | * KSM's break_ksm() relies upon recognizing a ksm page | |
409 | * even while it is being migrated, so for that case we | |
410 | * need migration_entry_wait(). | |
411 | */ | |
412 | if (likely(!(flags & FOLL_MIGRATION))) | |
413 | goto no_page; | |
0661a336 | 414 | if (pte_none(pte)) |
4bbd4c77 KS |
415 | goto no_page; |
416 | entry = pte_to_swp_entry(pte); | |
417 | if (!is_migration_entry(entry)) | |
418 | goto no_page; | |
419 | pte_unmap_unlock(ptep, ptl); | |
420 | migration_entry_wait(mm, pmd, address); | |
69e68b4f | 421 | goto retry; |
4bbd4c77 | 422 | } |
8a0516ed | 423 | if ((flags & FOLL_NUMA) && pte_protnone(pte)) |
4bbd4c77 | 424 | goto no_page; |
19be0eaf | 425 | if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) { |
69e68b4f KS |
426 | pte_unmap_unlock(ptep, ptl); |
427 | return NULL; | |
428 | } | |
4bbd4c77 KS |
429 | |
430 | page = vm_normal_page(vma, address, pte); | |
3faa52c0 | 431 | if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) { |
3565fce3 | 432 | /* |
3faa52c0 JH |
433 | * Only return device mapping pages in the FOLL_GET or FOLL_PIN |
434 | * case since they are only valid while holding the pgmap | |
435 | * reference. | |
3565fce3 | 436 | */ |
df06b37f KB |
437 | *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap); |
438 | if (*pgmap) | |
3565fce3 DW |
439 | page = pte_page(pte); |
440 | else | |
441 | goto no_page; | |
442 | } else if (unlikely(!page)) { | |
1027e443 KS |
443 | if (flags & FOLL_DUMP) { |
444 | /* Avoid special (like zero) pages in core dumps */ | |
445 | page = ERR_PTR(-EFAULT); | |
446 | goto out; | |
447 | } | |
448 | ||
449 | if (is_zero_pfn(pte_pfn(pte))) { | |
450 | page = pte_page(pte); | |
451 | } else { | |
1027e443 KS |
452 | ret = follow_pfn_pte(vma, address, ptep, flags); |
453 | page = ERR_PTR(ret); | |
454 | goto out; | |
455 | } | |
4bbd4c77 KS |
456 | } |
457 | ||
6742d293 | 458 | if (flags & FOLL_SPLIT && PageTransCompound(page)) { |
6742d293 KS |
459 | get_page(page); |
460 | pte_unmap_unlock(ptep, ptl); | |
461 | lock_page(page); | |
462 | ret = split_huge_page(page); | |
463 | unlock_page(page); | |
464 | put_page(page); | |
465 | if (ret) | |
466 | return ERR_PTR(ret); | |
467 | goto retry; | |
468 | } | |
469 | ||
3faa52c0 JH |
470 | /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */ |
471 | if (unlikely(!try_grab_page(page, flags))) { | |
472 | page = ERR_PTR(-ENOMEM); | |
473 | goto out; | |
8fde12ca | 474 | } |
f28d4363 CI |
475 | /* |
476 | * We need to make the page accessible if and only if we are going | |
477 | * to access its content (the FOLL_PIN case). Please see | |
478 | * Documentation/core-api/pin_user_pages.rst for details. | |
479 | */ | |
480 | if (flags & FOLL_PIN) { | |
481 | ret = arch_make_page_accessible(page); | |
482 | if (ret) { | |
483 | unpin_user_page(page); | |
484 | page = ERR_PTR(ret); | |
485 | goto out; | |
486 | } | |
487 | } | |
4bbd4c77 KS |
488 | if (flags & FOLL_TOUCH) { |
489 | if ((flags & FOLL_WRITE) && | |
490 | !pte_dirty(pte) && !PageDirty(page)) | |
491 | set_page_dirty(page); | |
492 | /* | |
493 | * pte_mkyoung() would be more correct here, but atomic care | |
494 | * is needed to avoid losing the dirty bit: it is easier to use | |
495 | * mark_page_accessed(). | |
496 | */ | |
497 | mark_page_accessed(page); | |
498 | } | |
de60f5f1 | 499 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
500 | /* Do not mlock pte-mapped THP */ |
501 | if (PageTransCompound(page)) | |
502 | goto out; | |
503 | ||
4bbd4c77 KS |
504 | /* |
505 | * The preliminary mapping check is mainly to avoid the | |
506 | * pointless overhead of lock_page on the ZERO_PAGE | |
507 | * which might bounce very badly if there is contention. | |
508 | * | |
509 | * If the page is already locked, we don't need to | |
510 | * handle it now - vmscan will handle it later if and | |
511 | * when it attempts to reclaim the page. | |
512 | */ | |
513 | if (page->mapping && trylock_page(page)) { | |
514 | lru_add_drain(); /* push cached pages to LRU */ | |
515 | /* | |
516 | * Because we lock page here, and migration is | |
517 | * blocked by the pte's page reference, and we | |
518 | * know the page is still mapped, we don't even | |
519 | * need to check for file-cache page truncation. | |
520 | */ | |
521 | mlock_vma_page(page); | |
522 | unlock_page(page); | |
523 | } | |
524 | } | |
1027e443 | 525 | out: |
4bbd4c77 | 526 | pte_unmap_unlock(ptep, ptl); |
4bbd4c77 | 527 | return page; |
4bbd4c77 KS |
528 | no_page: |
529 | pte_unmap_unlock(ptep, ptl); | |
530 | if (!pte_none(pte)) | |
69e68b4f KS |
531 | return NULL; |
532 | return no_page_table(vma, flags); | |
533 | } | |
534 | ||
080dbb61 AK |
535 | static struct page *follow_pmd_mask(struct vm_area_struct *vma, |
536 | unsigned long address, pud_t *pudp, | |
df06b37f KB |
537 | unsigned int flags, |
538 | struct follow_page_context *ctx) | |
69e68b4f | 539 | { |
68827280 | 540 | pmd_t *pmd, pmdval; |
69e68b4f KS |
541 | spinlock_t *ptl; |
542 | struct page *page; | |
543 | struct mm_struct *mm = vma->vm_mm; | |
544 | ||
080dbb61 | 545 | pmd = pmd_offset(pudp, address); |
68827280 HY |
546 | /* |
547 | * The READ_ONCE() will stabilize the pmdval in a register or | |
548 | * on the stack so that it will stop changing under the code. | |
549 | */ | |
550 | pmdval = READ_ONCE(*pmd); | |
551 | if (pmd_none(pmdval)) | |
69e68b4f | 552 | return no_page_table(vma, flags); |
be9d3045 | 553 | if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) { |
e66f17ff NH |
554 | page = follow_huge_pmd(mm, address, pmd, flags); |
555 | if (page) | |
556 | return page; | |
557 | return no_page_table(vma, flags); | |
69e68b4f | 558 | } |
68827280 | 559 | if (is_hugepd(__hugepd(pmd_val(pmdval)))) { |
4dc71451 | 560 | page = follow_huge_pd(vma, address, |
68827280 | 561 | __hugepd(pmd_val(pmdval)), flags, |
4dc71451 AK |
562 | PMD_SHIFT); |
563 | if (page) | |
564 | return page; | |
565 | return no_page_table(vma, flags); | |
566 | } | |
84c3fc4e | 567 | retry: |
68827280 | 568 | if (!pmd_present(pmdval)) { |
84c3fc4e ZY |
569 | if (likely(!(flags & FOLL_MIGRATION))) |
570 | return no_page_table(vma, flags); | |
571 | VM_BUG_ON(thp_migration_supported() && | |
68827280 HY |
572 | !is_pmd_migration_entry(pmdval)); |
573 | if (is_pmd_migration_entry(pmdval)) | |
84c3fc4e | 574 | pmd_migration_entry_wait(mm, pmd); |
68827280 HY |
575 | pmdval = READ_ONCE(*pmd); |
576 | /* | |
577 | * MADV_DONTNEED may convert the pmd to null because | |
578 | * mmap_sem is held in read mode | |
579 | */ | |
580 | if (pmd_none(pmdval)) | |
581 | return no_page_table(vma, flags); | |
84c3fc4e ZY |
582 | goto retry; |
583 | } | |
68827280 | 584 | if (pmd_devmap(pmdval)) { |
3565fce3 | 585 | ptl = pmd_lock(mm, pmd); |
df06b37f | 586 | page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap); |
3565fce3 DW |
587 | spin_unlock(ptl); |
588 | if (page) | |
589 | return page; | |
590 | } | |
68827280 | 591 | if (likely(!pmd_trans_huge(pmdval))) |
df06b37f | 592 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 593 | |
68827280 | 594 | if ((flags & FOLL_NUMA) && pmd_protnone(pmdval)) |
db08f203 AK |
595 | return no_page_table(vma, flags); |
596 | ||
84c3fc4e | 597 | retry_locked: |
6742d293 | 598 | ptl = pmd_lock(mm, pmd); |
68827280 HY |
599 | if (unlikely(pmd_none(*pmd))) { |
600 | spin_unlock(ptl); | |
601 | return no_page_table(vma, flags); | |
602 | } | |
84c3fc4e ZY |
603 | if (unlikely(!pmd_present(*pmd))) { |
604 | spin_unlock(ptl); | |
605 | if (likely(!(flags & FOLL_MIGRATION))) | |
606 | return no_page_table(vma, flags); | |
607 | pmd_migration_entry_wait(mm, pmd); | |
608 | goto retry_locked; | |
609 | } | |
6742d293 KS |
610 | if (unlikely(!pmd_trans_huge(*pmd))) { |
611 | spin_unlock(ptl); | |
df06b37f | 612 | return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
6742d293 | 613 | } |
bfe7b00d | 614 | if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) { |
6742d293 KS |
615 | int ret; |
616 | page = pmd_page(*pmd); | |
617 | if (is_huge_zero_page(page)) { | |
618 | spin_unlock(ptl); | |
619 | ret = 0; | |
78ddc534 | 620 | split_huge_pmd(vma, pmd, address); |
337d9abf NH |
621 | if (pmd_trans_unstable(pmd)) |
622 | ret = -EBUSY; | |
bfe7b00d | 623 | } else if (flags & FOLL_SPLIT) { |
8fde12ca LT |
624 | if (unlikely(!try_get_page(page))) { |
625 | spin_unlock(ptl); | |
626 | return ERR_PTR(-ENOMEM); | |
627 | } | |
69e68b4f | 628 | spin_unlock(ptl); |
6742d293 KS |
629 | lock_page(page); |
630 | ret = split_huge_page(page); | |
631 | unlock_page(page); | |
632 | put_page(page); | |
baa355fd KS |
633 | if (pmd_none(*pmd)) |
634 | return no_page_table(vma, flags); | |
bfe7b00d SL |
635 | } else { /* flags & FOLL_SPLIT_PMD */ |
636 | spin_unlock(ptl); | |
637 | split_huge_pmd(vma, pmd, address); | |
638 | ret = pte_alloc(mm, pmd) ? -ENOMEM : 0; | |
6742d293 KS |
639 | } |
640 | ||
641 | return ret ? ERR_PTR(ret) : | |
df06b37f | 642 | follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); |
69e68b4f | 643 | } |
6742d293 KS |
644 | page = follow_trans_huge_pmd(vma, address, pmd, flags); |
645 | spin_unlock(ptl); | |
df06b37f | 646 | ctx->page_mask = HPAGE_PMD_NR - 1; |
6742d293 | 647 | return page; |
4bbd4c77 KS |
648 | } |
649 | ||
080dbb61 AK |
650 | static struct page *follow_pud_mask(struct vm_area_struct *vma, |
651 | unsigned long address, p4d_t *p4dp, | |
df06b37f KB |
652 | unsigned int flags, |
653 | struct follow_page_context *ctx) | |
080dbb61 AK |
654 | { |
655 | pud_t *pud; | |
656 | spinlock_t *ptl; | |
657 | struct page *page; | |
658 | struct mm_struct *mm = vma->vm_mm; | |
659 | ||
660 | pud = pud_offset(p4dp, address); | |
661 | if (pud_none(*pud)) | |
662 | return no_page_table(vma, flags); | |
be9d3045 | 663 | if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) { |
080dbb61 AK |
664 | page = follow_huge_pud(mm, address, pud, flags); |
665 | if (page) | |
666 | return page; | |
667 | return no_page_table(vma, flags); | |
668 | } | |
4dc71451 AK |
669 | if (is_hugepd(__hugepd(pud_val(*pud)))) { |
670 | page = follow_huge_pd(vma, address, | |
671 | __hugepd(pud_val(*pud)), flags, | |
672 | PUD_SHIFT); | |
673 | if (page) | |
674 | return page; | |
675 | return no_page_table(vma, flags); | |
676 | } | |
080dbb61 AK |
677 | if (pud_devmap(*pud)) { |
678 | ptl = pud_lock(mm, pud); | |
df06b37f | 679 | page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap); |
080dbb61 AK |
680 | spin_unlock(ptl); |
681 | if (page) | |
682 | return page; | |
683 | } | |
684 | if (unlikely(pud_bad(*pud))) | |
685 | return no_page_table(vma, flags); | |
686 | ||
df06b37f | 687 | return follow_pmd_mask(vma, address, pud, flags, ctx); |
080dbb61 AK |
688 | } |
689 | ||
080dbb61 AK |
690 | static struct page *follow_p4d_mask(struct vm_area_struct *vma, |
691 | unsigned long address, pgd_t *pgdp, | |
df06b37f KB |
692 | unsigned int flags, |
693 | struct follow_page_context *ctx) | |
080dbb61 AK |
694 | { |
695 | p4d_t *p4d; | |
4dc71451 | 696 | struct page *page; |
080dbb61 AK |
697 | |
698 | p4d = p4d_offset(pgdp, address); | |
699 | if (p4d_none(*p4d)) | |
700 | return no_page_table(vma, flags); | |
701 | BUILD_BUG_ON(p4d_huge(*p4d)); | |
702 | if (unlikely(p4d_bad(*p4d))) | |
703 | return no_page_table(vma, flags); | |
704 | ||
4dc71451 AK |
705 | if (is_hugepd(__hugepd(p4d_val(*p4d)))) { |
706 | page = follow_huge_pd(vma, address, | |
707 | __hugepd(p4d_val(*p4d)), flags, | |
708 | P4D_SHIFT); | |
709 | if (page) | |
710 | return page; | |
711 | return no_page_table(vma, flags); | |
712 | } | |
df06b37f | 713 | return follow_pud_mask(vma, address, p4d, flags, ctx); |
080dbb61 AK |
714 | } |
715 | ||
716 | /** | |
717 | * follow_page_mask - look up a page descriptor from a user-virtual address | |
718 | * @vma: vm_area_struct mapping @address | |
719 | * @address: virtual address to look up | |
720 | * @flags: flags modifying lookup behaviour | |
78179556 MR |
721 | * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a |
722 | * pointer to output page_mask | |
080dbb61 AK |
723 | * |
724 | * @flags can have FOLL_ flags set, defined in <linux/mm.h> | |
725 | * | |
78179556 MR |
726 | * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches |
727 | * the device's dev_pagemap metadata to avoid repeating expensive lookups. | |
728 | * | |
729 | * On output, the @ctx->page_mask is set according to the size of the page. | |
730 | * | |
731 | * Return: the mapped (struct page *), %NULL if no mapping exists, or | |
080dbb61 AK |
732 | * an error pointer if there is a mapping to something not represented |
733 | * by a page descriptor (see also vm_normal_page()). | |
734 | */ | |
a7030aea | 735 | static struct page *follow_page_mask(struct vm_area_struct *vma, |
080dbb61 | 736 | unsigned long address, unsigned int flags, |
df06b37f | 737 | struct follow_page_context *ctx) |
080dbb61 AK |
738 | { |
739 | pgd_t *pgd; | |
740 | struct page *page; | |
741 | struct mm_struct *mm = vma->vm_mm; | |
742 | ||
df06b37f | 743 | ctx->page_mask = 0; |
080dbb61 AK |
744 | |
745 | /* make this handle hugepd */ | |
746 | page = follow_huge_addr(mm, address, flags & FOLL_WRITE); | |
747 | if (!IS_ERR(page)) { | |
3faa52c0 | 748 | WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN)); |
080dbb61 AK |
749 | return page; |
750 | } | |
751 | ||
752 | pgd = pgd_offset(mm, address); | |
753 | ||
754 | if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) | |
755 | return no_page_table(vma, flags); | |
756 | ||
faaa5b62 AK |
757 | if (pgd_huge(*pgd)) { |
758 | page = follow_huge_pgd(mm, address, pgd, flags); | |
759 | if (page) | |
760 | return page; | |
761 | return no_page_table(vma, flags); | |
762 | } | |
4dc71451 AK |
763 | if (is_hugepd(__hugepd(pgd_val(*pgd)))) { |
764 | page = follow_huge_pd(vma, address, | |
765 | __hugepd(pgd_val(*pgd)), flags, | |
766 | PGDIR_SHIFT); | |
767 | if (page) | |
768 | return page; | |
769 | return no_page_table(vma, flags); | |
770 | } | |
faaa5b62 | 771 | |
df06b37f KB |
772 | return follow_p4d_mask(vma, address, pgd, flags, ctx); |
773 | } | |
774 | ||
775 | struct page *follow_page(struct vm_area_struct *vma, unsigned long address, | |
776 | unsigned int foll_flags) | |
777 | { | |
778 | struct follow_page_context ctx = { NULL }; | |
779 | struct page *page; | |
780 | ||
781 | page = follow_page_mask(vma, address, foll_flags, &ctx); | |
782 | if (ctx.pgmap) | |
783 | put_dev_pagemap(ctx.pgmap); | |
784 | return page; | |
080dbb61 AK |
785 | } |
786 | ||
f2b495ca KS |
787 | static int get_gate_page(struct mm_struct *mm, unsigned long address, |
788 | unsigned int gup_flags, struct vm_area_struct **vma, | |
789 | struct page **page) | |
790 | { | |
791 | pgd_t *pgd; | |
c2febafc | 792 | p4d_t *p4d; |
f2b495ca KS |
793 | pud_t *pud; |
794 | pmd_t *pmd; | |
795 | pte_t *pte; | |
796 | int ret = -EFAULT; | |
797 | ||
798 | /* user gate pages are read-only */ | |
799 | if (gup_flags & FOLL_WRITE) | |
800 | return -EFAULT; | |
801 | if (address > TASK_SIZE) | |
802 | pgd = pgd_offset_k(address); | |
803 | else | |
804 | pgd = pgd_offset_gate(mm, address); | |
b5d1c39f AL |
805 | if (pgd_none(*pgd)) |
806 | return -EFAULT; | |
c2febafc | 807 | p4d = p4d_offset(pgd, address); |
b5d1c39f AL |
808 | if (p4d_none(*p4d)) |
809 | return -EFAULT; | |
c2febafc | 810 | pud = pud_offset(p4d, address); |
b5d1c39f AL |
811 | if (pud_none(*pud)) |
812 | return -EFAULT; | |
f2b495ca | 813 | pmd = pmd_offset(pud, address); |
84c3fc4e | 814 | if (!pmd_present(*pmd)) |
f2b495ca KS |
815 | return -EFAULT; |
816 | VM_BUG_ON(pmd_trans_huge(*pmd)); | |
817 | pte = pte_offset_map(pmd, address); | |
818 | if (pte_none(*pte)) | |
819 | goto unmap; | |
820 | *vma = get_gate_vma(mm); | |
821 | if (!page) | |
822 | goto out; | |
823 | *page = vm_normal_page(*vma, address, *pte); | |
824 | if (!*page) { | |
825 | if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte))) | |
826 | goto unmap; | |
827 | *page = pte_page(*pte); | |
828 | } | |
8fde12ca LT |
829 | if (unlikely(!try_get_page(*page))) { |
830 | ret = -ENOMEM; | |
831 | goto unmap; | |
832 | } | |
f2b495ca KS |
833 | out: |
834 | ret = 0; | |
835 | unmap: | |
836 | pte_unmap(pte); | |
837 | return ret; | |
838 | } | |
839 | ||
9a95f3cf PC |
840 | /* |
841 | * mmap_sem must be held on entry. If @nonblocking != NULL and | |
842 | * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released. | |
843 | * If it is, *@nonblocking will be set to 0 and -EBUSY returned. | |
844 | */ | |
16744483 KS |
845 | static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma, |
846 | unsigned long address, unsigned int *flags, int *nonblocking) | |
847 | { | |
16744483 | 848 | unsigned int fault_flags = 0; |
2b740303 | 849 | vm_fault_t ret; |
16744483 | 850 | |
de60f5f1 EM |
851 | /* mlock all present pages, but do not fault in new pages */ |
852 | if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK) | |
853 | return -ENOENT; | |
16744483 KS |
854 | if (*flags & FOLL_WRITE) |
855 | fault_flags |= FAULT_FLAG_WRITE; | |
1b2ee126 DH |
856 | if (*flags & FOLL_REMOTE) |
857 | fault_flags |= FAULT_FLAG_REMOTE; | |
16744483 KS |
858 | if (nonblocking) |
859 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
860 | if (*flags & FOLL_NOWAIT) | |
861 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT; | |
234b239b ALC |
862 | if (*flags & FOLL_TRIED) { |
863 | VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY); | |
864 | fault_flags |= FAULT_FLAG_TRIED; | |
865 | } | |
16744483 | 866 | |
dcddffd4 | 867 | ret = handle_mm_fault(vma, address, fault_flags); |
16744483 | 868 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
869 | int err = vm_fault_to_errno(ret, *flags); |
870 | ||
871 | if (err) | |
872 | return err; | |
16744483 KS |
873 | BUG(); |
874 | } | |
875 | ||
876 | if (tsk) { | |
877 | if (ret & VM_FAULT_MAJOR) | |
878 | tsk->maj_flt++; | |
879 | else | |
880 | tsk->min_flt++; | |
881 | } | |
882 | ||
883 | if (ret & VM_FAULT_RETRY) { | |
96312e61 | 884 | if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
16744483 KS |
885 | *nonblocking = 0; |
886 | return -EBUSY; | |
887 | } | |
888 | ||
889 | /* | |
890 | * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when | |
891 | * necessary, even if maybe_mkwrite decided not to set pte_write. We | |
892 | * can thus safely do subsequent page lookups as if they were reads. | |
893 | * But only do so when looping for pte_write is futile: in some cases | |
894 | * userspace may also be wanting to write to the gotten user page, | |
895 | * which a read fault here might prevent (a readonly page might get | |
896 | * reCOWed by userspace write). | |
897 | */ | |
898 | if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE)) | |
2923117b | 899 | *flags |= FOLL_COW; |
16744483 KS |
900 | return 0; |
901 | } | |
902 | ||
fa5bb209 KS |
903 | static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags) |
904 | { | |
905 | vm_flags_t vm_flags = vma->vm_flags; | |
1b2ee126 DH |
906 | int write = (gup_flags & FOLL_WRITE); |
907 | int foreign = (gup_flags & FOLL_REMOTE); | |
fa5bb209 KS |
908 | |
909 | if (vm_flags & (VM_IO | VM_PFNMAP)) | |
910 | return -EFAULT; | |
911 | ||
7f7ccc2c WT |
912 | if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma)) |
913 | return -EFAULT; | |
914 | ||
1b2ee126 | 915 | if (write) { |
fa5bb209 KS |
916 | if (!(vm_flags & VM_WRITE)) { |
917 | if (!(gup_flags & FOLL_FORCE)) | |
918 | return -EFAULT; | |
919 | /* | |
920 | * We used to let the write,force case do COW in a | |
921 | * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could | |
922 | * set a breakpoint in a read-only mapping of an | |
923 | * executable, without corrupting the file (yet only | |
924 | * when that file had been opened for writing!). | |
925 | * Anon pages in shared mappings are surprising: now | |
926 | * just reject it. | |
927 | */ | |
46435364 | 928 | if (!is_cow_mapping(vm_flags)) |
fa5bb209 | 929 | return -EFAULT; |
fa5bb209 KS |
930 | } |
931 | } else if (!(vm_flags & VM_READ)) { | |
932 | if (!(gup_flags & FOLL_FORCE)) | |
933 | return -EFAULT; | |
934 | /* | |
935 | * Is there actually any vma we can reach here which does not | |
936 | * have VM_MAYREAD set? | |
937 | */ | |
938 | if (!(vm_flags & VM_MAYREAD)) | |
939 | return -EFAULT; | |
940 | } | |
d61172b4 DH |
941 | /* |
942 | * gups are always data accesses, not instruction | |
943 | * fetches, so execute=false here | |
944 | */ | |
945 | if (!arch_vma_access_permitted(vma, write, false, foreign)) | |
33a709b2 | 946 | return -EFAULT; |
fa5bb209 KS |
947 | return 0; |
948 | } | |
949 | ||
4bbd4c77 KS |
950 | /** |
951 | * __get_user_pages() - pin user pages in memory | |
952 | * @tsk: task_struct of target task | |
953 | * @mm: mm_struct of target mm | |
954 | * @start: starting user address | |
955 | * @nr_pages: number of pages from start to pin | |
956 | * @gup_flags: flags modifying pin behaviour | |
957 | * @pages: array that receives pointers to the pages pinned. | |
958 | * Should be at least nr_pages long. Or NULL, if caller | |
959 | * only intends to ensure the pages are faulted in. | |
960 | * @vmas: array of pointers to vmas corresponding to each page. | |
961 | * Or NULL if the caller does not require them. | |
962 | * @nonblocking: whether waiting for disk IO or mmap_sem contention | |
963 | * | |
d2dfbe47 LX |
964 | * Returns either number of pages pinned (which may be less than the |
965 | * number requested), or an error. Details about the return value: | |
966 | * | |
967 | * -- If nr_pages is 0, returns 0. | |
968 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
969 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
970 | * pages pinned. Again, this may be less than nr_pages. | |
971 | * | |
972 | * The caller is responsible for releasing returned @pages, via put_page(). | |
973 | * | |
974 | * @vmas are valid only as long as mmap_sem is held. | |
4bbd4c77 | 975 | * |
9a95f3cf | 976 | * Must be called with mmap_sem held. It may be released. See below. |
4bbd4c77 KS |
977 | * |
978 | * __get_user_pages walks a process's page tables and takes a reference to | |
979 | * each struct page that each user address corresponds to at a given | |
980 | * instant. That is, it takes the page that would be accessed if a user | |
981 | * thread accesses the given user virtual address at that instant. | |
982 | * | |
983 | * This does not guarantee that the page exists in the user mappings when | |
984 | * __get_user_pages returns, and there may even be a completely different | |
985 | * page there in some cases (eg. if mmapped pagecache has been invalidated | |
986 | * and subsequently re faulted). However it does guarantee that the page | |
987 | * won't be freed completely. And mostly callers simply care that the page | |
988 | * contains data that was valid *at some point in time*. Typically, an IO | |
989 | * or similar operation cannot guarantee anything stronger anyway because | |
990 | * locks can't be held over the syscall boundary. | |
991 | * | |
992 | * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If | |
993 | * the page is written to, set_page_dirty (or set_page_dirty_lock, as | |
994 | * appropriate) must be called after the page is finished with, and | |
995 | * before put_page is called. | |
996 | * | |
997 | * If @nonblocking != NULL, __get_user_pages will not wait for disk IO | |
998 | * or mmap_sem contention, and if waiting is needed to pin all pages, | |
9a95f3cf PC |
999 | * *@nonblocking will be set to 0. Further, if @gup_flags does not |
1000 | * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in | |
1001 | * this case. | |
1002 | * | |
1003 | * A caller using such a combination of @nonblocking and @gup_flags | |
1004 | * must therefore hold the mmap_sem for reading only, and recognize | |
1005 | * when it's been released. Otherwise, it must be held for either | |
1006 | * reading or writing and will not be released. | |
4bbd4c77 KS |
1007 | * |
1008 | * In most cases, get_user_pages or get_user_pages_fast should be used | |
1009 | * instead of __get_user_pages. __get_user_pages should be used only if | |
1010 | * you need some special @gup_flags. | |
1011 | */ | |
0d731759 | 1012 | static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm, |
4bbd4c77 KS |
1013 | unsigned long start, unsigned long nr_pages, |
1014 | unsigned int gup_flags, struct page **pages, | |
1015 | struct vm_area_struct **vmas, int *nonblocking) | |
1016 | { | |
df06b37f | 1017 | long ret = 0, i = 0; |
fa5bb209 | 1018 | struct vm_area_struct *vma = NULL; |
df06b37f | 1019 | struct follow_page_context ctx = { NULL }; |
4bbd4c77 KS |
1020 | |
1021 | if (!nr_pages) | |
1022 | return 0; | |
1023 | ||
f9652594 AK |
1024 | start = untagged_addr(start); |
1025 | ||
eddb1c22 | 1026 | VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN))); |
4bbd4c77 KS |
1027 | |
1028 | /* | |
1029 | * If FOLL_FORCE is set then do not force a full fault as the hinting | |
1030 | * fault information is unrelated to the reference behaviour of a task | |
1031 | * using the address space | |
1032 | */ | |
1033 | if (!(gup_flags & FOLL_FORCE)) | |
1034 | gup_flags |= FOLL_NUMA; | |
1035 | ||
4bbd4c77 | 1036 | do { |
fa5bb209 KS |
1037 | struct page *page; |
1038 | unsigned int foll_flags = gup_flags; | |
1039 | unsigned int page_increm; | |
1040 | ||
1041 | /* first iteration or cross vma bound */ | |
1042 | if (!vma || start >= vma->vm_end) { | |
1043 | vma = find_extend_vma(mm, start); | |
1044 | if (!vma && in_gate_area(mm, start)) { | |
fa5bb209 KS |
1045 | ret = get_gate_page(mm, start & PAGE_MASK, |
1046 | gup_flags, &vma, | |
1047 | pages ? &pages[i] : NULL); | |
1048 | if (ret) | |
08be37b7 | 1049 | goto out; |
df06b37f | 1050 | ctx.page_mask = 0; |
fa5bb209 KS |
1051 | goto next_page; |
1052 | } | |
4bbd4c77 | 1053 | |
df06b37f KB |
1054 | if (!vma || check_vma_flags(vma, gup_flags)) { |
1055 | ret = -EFAULT; | |
1056 | goto out; | |
1057 | } | |
fa5bb209 KS |
1058 | if (is_vm_hugetlb_page(vma)) { |
1059 | i = follow_hugetlb_page(mm, vma, pages, vmas, | |
1060 | &start, &nr_pages, i, | |
87ffc118 | 1061 | gup_flags, nonblocking); |
fa5bb209 | 1062 | continue; |
4bbd4c77 | 1063 | } |
fa5bb209 KS |
1064 | } |
1065 | retry: | |
1066 | /* | |
1067 | * If we have a pending SIGKILL, don't keep faulting pages and | |
1068 | * potentially allocating memory. | |
1069 | */ | |
fa45f116 | 1070 | if (fatal_signal_pending(current)) { |
df06b37f KB |
1071 | ret = -ERESTARTSYS; |
1072 | goto out; | |
1073 | } | |
fa5bb209 | 1074 | cond_resched(); |
df06b37f KB |
1075 | |
1076 | page = follow_page_mask(vma, start, foll_flags, &ctx); | |
fa5bb209 | 1077 | if (!page) { |
fa5bb209 KS |
1078 | ret = faultin_page(tsk, vma, start, &foll_flags, |
1079 | nonblocking); | |
1080 | switch (ret) { | |
1081 | case 0: | |
1082 | goto retry; | |
df06b37f KB |
1083 | case -EBUSY: |
1084 | ret = 0; | |
1085 | /* FALLTHRU */ | |
fa5bb209 KS |
1086 | case -EFAULT: |
1087 | case -ENOMEM: | |
1088 | case -EHWPOISON: | |
df06b37f | 1089 | goto out; |
fa5bb209 KS |
1090 | case -ENOENT: |
1091 | goto next_page; | |
4bbd4c77 | 1092 | } |
fa5bb209 | 1093 | BUG(); |
1027e443 KS |
1094 | } else if (PTR_ERR(page) == -EEXIST) { |
1095 | /* | |
1096 | * Proper page table entry exists, but no corresponding | |
1097 | * struct page. | |
1098 | */ | |
1099 | goto next_page; | |
1100 | } else if (IS_ERR(page)) { | |
df06b37f KB |
1101 | ret = PTR_ERR(page); |
1102 | goto out; | |
1027e443 | 1103 | } |
fa5bb209 KS |
1104 | if (pages) { |
1105 | pages[i] = page; | |
1106 | flush_anon_page(vma, page, start); | |
1107 | flush_dcache_page(page); | |
df06b37f | 1108 | ctx.page_mask = 0; |
4bbd4c77 | 1109 | } |
4bbd4c77 | 1110 | next_page: |
fa5bb209 KS |
1111 | if (vmas) { |
1112 | vmas[i] = vma; | |
df06b37f | 1113 | ctx.page_mask = 0; |
fa5bb209 | 1114 | } |
df06b37f | 1115 | page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); |
fa5bb209 KS |
1116 | if (page_increm > nr_pages) |
1117 | page_increm = nr_pages; | |
1118 | i += page_increm; | |
1119 | start += page_increm * PAGE_SIZE; | |
1120 | nr_pages -= page_increm; | |
4bbd4c77 | 1121 | } while (nr_pages); |
df06b37f KB |
1122 | out: |
1123 | if (ctx.pgmap) | |
1124 | put_dev_pagemap(ctx.pgmap); | |
1125 | return i ? i : ret; | |
4bbd4c77 | 1126 | } |
4bbd4c77 | 1127 | |
771ab430 TK |
1128 | static bool vma_permits_fault(struct vm_area_struct *vma, |
1129 | unsigned int fault_flags) | |
d4925e00 | 1130 | { |
1b2ee126 DH |
1131 | bool write = !!(fault_flags & FAULT_FLAG_WRITE); |
1132 | bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE); | |
33a709b2 | 1133 | vm_flags_t vm_flags = write ? VM_WRITE : VM_READ; |
d4925e00 DH |
1134 | |
1135 | if (!(vm_flags & vma->vm_flags)) | |
1136 | return false; | |
1137 | ||
33a709b2 DH |
1138 | /* |
1139 | * The architecture might have a hardware protection | |
1b2ee126 | 1140 | * mechanism other than read/write that can deny access. |
d61172b4 DH |
1141 | * |
1142 | * gup always represents data access, not instruction | |
1143 | * fetches, so execute=false here: | |
33a709b2 | 1144 | */ |
d61172b4 | 1145 | if (!arch_vma_access_permitted(vma, write, false, foreign)) |
33a709b2 DH |
1146 | return false; |
1147 | ||
d4925e00 DH |
1148 | return true; |
1149 | } | |
1150 | ||
4bbd4c77 KS |
1151 | /* |
1152 | * fixup_user_fault() - manually resolve a user page fault | |
1153 | * @tsk: the task_struct to use for page fault accounting, or | |
1154 | * NULL if faults are not to be recorded. | |
1155 | * @mm: mm_struct of target mm | |
1156 | * @address: user address | |
1157 | * @fault_flags:flags to pass down to handle_mm_fault() | |
4a9e1cda DD |
1158 | * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller |
1159 | * does not allow retry | |
4bbd4c77 KS |
1160 | * |
1161 | * This is meant to be called in the specific scenario where for locking reasons | |
1162 | * we try to access user memory in atomic context (within a pagefault_disable() | |
1163 | * section), this returns -EFAULT, and we want to resolve the user fault before | |
1164 | * trying again. | |
1165 | * | |
1166 | * Typically this is meant to be used by the futex code. | |
1167 | * | |
1168 | * The main difference with get_user_pages() is that this function will | |
1169 | * unconditionally call handle_mm_fault() which will in turn perform all the | |
1170 | * necessary SW fixup of the dirty and young bits in the PTE, while | |
4a9e1cda | 1171 | * get_user_pages() only guarantees to update these in the struct page. |
4bbd4c77 KS |
1172 | * |
1173 | * This is important for some architectures where those bits also gate the | |
1174 | * access permission to the page because they are maintained in software. On | |
1175 | * such architectures, gup() will not be enough to make a subsequent access | |
1176 | * succeed. | |
1177 | * | |
4a9e1cda DD |
1178 | * This function will not return with an unlocked mmap_sem. So it has not the |
1179 | * same semantics wrt the @mm->mmap_sem as does filemap_fault(). | |
4bbd4c77 KS |
1180 | */ |
1181 | int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm, | |
4a9e1cda DD |
1182 | unsigned long address, unsigned int fault_flags, |
1183 | bool *unlocked) | |
4bbd4c77 KS |
1184 | { |
1185 | struct vm_area_struct *vma; | |
2b740303 | 1186 | vm_fault_t ret, major = 0; |
4a9e1cda | 1187 | |
f9652594 AK |
1188 | address = untagged_addr(address); |
1189 | ||
4a9e1cda DD |
1190 | if (unlocked) |
1191 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4bbd4c77 | 1192 | |
4a9e1cda | 1193 | retry: |
4bbd4c77 KS |
1194 | vma = find_extend_vma(mm, address); |
1195 | if (!vma || address < vma->vm_start) | |
1196 | return -EFAULT; | |
1197 | ||
d4925e00 | 1198 | if (!vma_permits_fault(vma, fault_flags)) |
4bbd4c77 KS |
1199 | return -EFAULT; |
1200 | ||
dcddffd4 | 1201 | ret = handle_mm_fault(vma, address, fault_flags); |
4a9e1cda | 1202 | major |= ret & VM_FAULT_MAJOR; |
4bbd4c77 | 1203 | if (ret & VM_FAULT_ERROR) { |
9a291a7c JM |
1204 | int err = vm_fault_to_errno(ret, 0); |
1205 | ||
1206 | if (err) | |
1207 | return err; | |
4bbd4c77 KS |
1208 | BUG(); |
1209 | } | |
4a9e1cda DD |
1210 | |
1211 | if (ret & VM_FAULT_RETRY) { | |
1212 | down_read(&mm->mmap_sem); | |
1213 | if (!(fault_flags & FAULT_FLAG_TRIED)) { | |
1214 | *unlocked = true; | |
1215 | fault_flags &= ~FAULT_FLAG_ALLOW_RETRY; | |
1216 | fault_flags |= FAULT_FLAG_TRIED; | |
1217 | goto retry; | |
1218 | } | |
1219 | } | |
1220 | ||
4bbd4c77 | 1221 | if (tsk) { |
4a9e1cda | 1222 | if (major) |
4bbd4c77 KS |
1223 | tsk->maj_flt++; |
1224 | else | |
1225 | tsk->min_flt++; | |
1226 | } | |
1227 | return 0; | |
1228 | } | |
add6a0cd | 1229 | EXPORT_SYMBOL_GPL(fixup_user_fault); |
4bbd4c77 | 1230 | |
f0818f47 AA |
1231 | static __always_inline long __get_user_pages_locked(struct task_struct *tsk, |
1232 | struct mm_struct *mm, | |
1233 | unsigned long start, | |
1234 | unsigned long nr_pages, | |
f0818f47 AA |
1235 | struct page **pages, |
1236 | struct vm_area_struct **vmas, | |
e716712f | 1237 | int *locked, |
0fd71a56 | 1238 | unsigned int flags) |
f0818f47 | 1239 | { |
f0818f47 AA |
1240 | long ret, pages_done; |
1241 | bool lock_dropped; | |
1242 | ||
1243 | if (locked) { | |
1244 | /* if VM_FAULT_RETRY can be returned, vmas become invalid */ | |
1245 | BUG_ON(vmas); | |
1246 | /* check caller initialized locked */ | |
1247 | BUG_ON(*locked != 1); | |
1248 | } | |
1249 | ||
eddb1c22 JH |
1250 | /* |
1251 | * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior | |
1252 | * is to set FOLL_GET if the caller wants pages[] filled in (but has | |
1253 | * carelessly failed to specify FOLL_GET), so keep doing that, but only | |
1254 | * for FOLL_GET, not for the newer FOLL_PIN. | |
1255 | * | |
1256 | * FOLL_PIN always expects pages to be non-null, but no need to assert | |
1257 | * that here, as any failures will be obvious enough. | |
1258 | */ | |
1259 | if (pages && !(flags & FOLL_PIN)) | |
f0818f47 | 1260 | flags |= FOLL_GET; |
f0818f47 AA |
1261 | |
1262 | pages_done = 0; | |
1263 | lock_dropped = false; | |
1264 | for (;;) { | |
1265 | ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages, | |
1266 | vmas, locked); | |
1267 | if (!locked) | |
1268 | /* VM_FAULT_RETRY couldn't trigger, bypass */ | |
1269 | return ret; | |
1270 | ||
1271 | /* VM_FAULT_RETRY cannot return errors */ | |
1272 | if (!*locked) { | |
1273 | BUG_ON(ret < 0); | |
1274 | BUG_ON(ret >= nr_pages); | |
1275 | } | |
1276 | ||
f0818f47 AA |
1277 | if (ret > 0) { |
1278 | nr_pages -= ret; | |
1279 | pages_done += ret; | |
1280 | if (!nr_pages) | |
1281 | break; | |
1282 | } | |
1283 | if (*locked) { | |
96312e61 AA |
1284 | /* |
1285 | * VM_FAULT_RETRY didn't trigger or it was a | |
1286 | * FOLL_NOWAIT. | |
1287 | */ | |
f0818f47 AA |
1288 | if (!pages_done) |
1289 | pages_done = ret; | |
1290 | break; | |
1291 | } | |
df17277b MR |
1292 | /* |
1293 | * VM_FAULT_RETRY triggered, so seek to the faulting offset. | |
1294 | * For the prefault case (!pages) we only update counts. | |
1295 | */ | |
1296 | if (likely(pages)) | |
1297 | pages += ret; | |
f0818f47 AA |
1298 | start += ret << PAGE_SHIFT; |
1299 | ||
1300 | /* | |
1301 | * Repeat on the address that fired VM_FAULT_RETRY | |
1302 | * without FAULT_FLAG_ALLOW_RETRY but with | |
1303 | * FAULT_FLAG_TRIED. | |
1304 | */ | |
1305 | *locked = 1; | |
1306 | lock_dropped = true; | |
1307 | down_read(&mm->mmap_sem); | |
1308 | ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED, | |
1309 | pages, NULL, NULL); | |
1310 | if (ret != 1) { | |
1311 | BUG_ON(ret > 1); | |
1312 | if (!pages_done) | |
1313 | pages_done = ret; | |
1314 | break; | |
1315 | } | |
1316 | nr_pages--; | |
1317 | pages_done++; | |
1318 | if (!nr_pages) | |
1319 | break; | |
df17277b MR |
1320 | if (likely(pages)) |
1321 | pages++; | |
f0818f47 AA |
1322 | start += PAGE_SIZE; |
1323 | } | |
e716712f | 1324 | if (lock_dropped && *locked) { |
f0818f47 AA |
1325 | /* |
1326 | * We must let the caller know we temporarily dropped the lock | |
1327 | * and so the critical section protected by it was lost. | |
1328 | */ | |
1329 | up_read(&mm->mmap_sem); | |
1330 | *locked = 0; | |
1331 | } | |
1332 | return pages_done; | |
1333 | } | |
1334 | ||
d3649f68 CH |
1335 | /** |
1336 | * populate_vma_page_range() - populate a range of pages in the vma. | |
1337 | * @vma: target vma | |
1338 | * @start: start address | |
1339 | * @end: end address | |
1340 | * @nonblocking: | |
1341 | * | |
1342 | * This takes care of mlocking the pages too if VM_LOCKED is set. | |
1343 | * | |
1344 | * return 0 on success, negative error code on error. | |
1345 | * | |
1346 | * vma->vm_mm->mmap_sem must be held. | |
1347 | * | |
1348 | * If @nonblocking is NULL, it may be held for read or write and will | |
1349 | * be unperturbed. | |
1350 | * | |
1351 | * If @nonblocking is non-NULL, it must held for read only and may be | |
1352 | * released. If it's released, *@nonblocking will be set to 0. | |
1353 | */ | |
1354 | long populate_vma_page_range(struct vm_area_struct *vma, | |
1355 | unsigned long start, unsigned long end, int *nonblocking) | |
1356 | { | |
1357 | struct mm_struct *mm = vma->vm_mm; | |
1358 | unsigned long nr_pages = (end - start) / PAGE_SIZE; | |
1359 | int gup_flags; | |
1360 | ||
1361 | VM_BUG_ON(start & ~PAGE_MASK); | |
1362 | VM_BUG_ON(end & ~PAGE_MASK); | |
1363 | VM_BUG_ON_VMA(start < vma->vm_start, vma); | |
1364 | VM_BUG_ON_VMA(end > vma->vm_end, vma); | |
1365 | VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm); | |
1366 | ||
1367 | gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK; | |
1368 | if (vma->vm_flags & VM_LOCKONFAULT) | |
1369 | gup_flags &= ~FOLL_POPULATE; | |
1370 | /* | |
1371 | * We want to touch writable mappings with a write fault in order | |
1372 | * to break COW, except for shared mappings because these don't COW | |
1373 | * and we would not want to dirty them for nothing. | |
1374 | */ | |
1375 | if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE) | |
1376 | gup_flags |= FOLL_WRITE; | |
1377 | ||
1378 | /* | |
1379 | * We want mlock to succeed for regions that have any permissions | |
1380 | * other than PROT_NONE. | |
1381 | */ | |
1382 | if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC)) | |
1383 | gup_flags |= FOLL_FORCE; | |
1384 | ||
1385 | /* | |
1386 | * We made sure addr is within a VMA, so the following will | |
1387 | * not result in a stack expansion that recurses back here. | |
1388 | */ | |
1389 | return __get_user_pages(current, mm, start, nr_pages, gup_flags, | |
1390 | NULL, NULL, nonblocking); | |
1391 | } | |
1392 | ||
1393 | /* | |
1394 | * __mm_populate - populate and/or mlock pages within a range of address space. | |
1395 | * | |
1396 | * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap | |
1397 | * flags. VMAs must be already marked with the desired vm_flags, and | |
1398 | * mmap_sem must not be held. | |
1399 | */ | |
1400 | int __mm_populate(unsigned long start, unsigned long len, int ignore_errors) | |
1401 | { | |
1402 | struct mm_struct *mm = current->mm; | |
1403 | unsigned long end, nstart, nend; | |
1404 | struct vm_area_struct *vma = NULL; | |
1405 | int locked = 0; | |
1406 | long ret = 0; | |
1407 | ||
1408 | end = start + len; | |
1409 | ||
1410 | for (nstart = start; nstart < end; nstart = nend) { | |
1411 | /* | |
1412 | * We want to fault in pages for [nstart; end) address range. | |
1413 | * Find first corresponding VMA. | |
1414 | */ | |
1415 | if (!locked) { | |
1416 | locked = 1; | |
1417 | down_read(&mm->mmap_sem); | |
1418 | vma = find_vma(mm, nstart); | |
1419 | } else if (nstart >= vma->vm_end) | |
1420 | vma = vma->vm_next; | |
1421 | if (!vma || vma->vm_start >= end) | |
1422 | break; | |
1423 | /* | |
1424 | * Set [nstart; nend) to intersection of desired address | |
1425 | * range with the first VMA. Also, skip undesirable VMA types. | |
1426 | */ | |
1427 | nend = min(end, vma->vm_end); | |
1428 | if (vma->vm_flags & (VM_IO | VM_PFNMAP)) | |
1429 | continue; | |
1430 | if (nstart < vma->vm_start) | |
1431 | nstart = vma->vm_start; | |
1432 | /* | |
1433 | * Now fault in a range of pages. populate_vma_page_range() | |
1434 | * double checks the vma flags, so that it won't mlock pages | |
1435 | * if the vma was already munlocked. | |
1436 | */ | |
1437 | ret = populate_vma_page_range(vma, nstart, nend, &locked); | |
1438 | if (ret < 0) { | |
1439 | if (ignore_errors) { | |
1440 | ret = 0; | |
1441 | continue; /* continue at next VMA */ | |
1442 | } | |
1443 | break; | |
1444 | } | |
1445 | nend = nstart + ret * PAGE_SIZE; | |
1446 | ret = 0; | |
1447 | } | |
1448 | if (locked) | |
1449 | up_read(&mm->mmap_sem); | |
1450 | return ret; /* 0 or negative error code */ | |
1451 | } | |
1452 | ||
1453 | /** | |
1454 | * get_dump_page() - pin user page in memory while writing it to core dump | |
1455 | * @addr: user address | |
1456 | * | |
1457 | * Returns struct page pointer of user page pinned for dump, | |
1458 | * to be freed afterwards by put_page(). | |
1459 | * | |
1460 | * Returns NULL on any kind of failure - a hole must then be inserted into | |
1461 | * the corefile, to preserve alignment with its headers; and also returns | |
1462 | * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found - | |
1463 | * allowing a hole to be left in the corefile to save diskspace. | |
1464 | * | |
1465 | * Called without mmap_sem, but after all other threads have been killed. | |
1466 | */ | |
1467 | #ifdef CONFIG_ELF_CORE | |
1468 | struct page *get_dump_page(unsigned long addr) | |
1469 | { | |
1470 | struct vm_area_struct *vma; | |
1471 | struct page *page; | |
1472 | ||
1473 | if (__get_user_pages(current, current->mm, addr, 1, | |
1474 | FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma, | |
1475 | NULL) < 1) | |
1476 | return NULL; | |
1477 | flush_cache_page(vma, addr, page_to_pfn(page)); | |
1478 | return page; | |
1479 | } | |
1480 | #endif /* CONFIG_ELF_CORE */ | |
050a9adc CH |
1481 | #else /* CONFIG_MMU */ |
1482 | static long __get_user_pages_locked(struct task_struct *tsk, | |
1483 | struct mm_struct *mm, unsigned long start, | |
1484 | unsigned long nr_pages, struct page **pages, | |
1485 | struct vm_area_struct **vmas, int *locked, | |
1486 | unsigned int foll_flags) | |
1487 | { | |
1488 | struct vm_area_struct *vma; | |
1489 | unsigned long vm_flags; | |
1490 | int i; | |
1491 | ||
1492 | /* calculate required read or write permissions. | |
1493 | * If FOLL_FORCE is set, we only require the "MAY" flags. | |
1494 | */ | |
1495 | vm_flags = (foll_flags & FOLL_WRITE) ? | |
1496 | (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); | |
1497 | vm_flags &= (foll_flags & FOLL_FORCE) ? | |
1498 | (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); | |
1499 | ||
1500 | for (i = 0; i < nr_pages; i++) { | |
1501 | vma = find_vma(mm, start); | |
1502 | if (!vma) | |
1503 | goto finish_or_fault; | |
1504 | ||
1505 | /* protect what we can, including chardevs */ | |
1506 | if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) || | |
1507 | !(vm_flags & vma->vm_flags)) | |
1508 | goto finish_or_fault; | |
1509 | ||
1510 | if (pages) { | |
1511 | pages[i] = virt_to_page(start); | |
1512 | if (pages[i]) | |
1513 | get_page(pages[i]); | |
1514 | } | |
1515 | if (vmas) | |
1516 | vmas[i] = vma; | |
1517 | start = (start + PAGE_SIZE) & PAGE_MASK; | |
1518 | } | |
1519 | ||
1520 | return i; | |
1521 | ||
1522 | finish_or_fault: | |
1523 | return i ? : -EFAULT; | |
1524 | } | |
1525 | #endif /* !CONFIG_MMU */ | |
d3649f68 | 1526 | |
9a4e9f3b | 1527 | #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA) |
9a4e9f3b AK |
1528 | static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages) |
1529 | { | |
1530 | long i; | |
1531 | struct vm_area_struct *vma_prev = NULL; | |
1532 | ||
1533 | for (i = 0; i < nr_pages; i++) { | |
1534 | struct vm_area_struct *vma = vmas[i]; | |
1535 | ||
1536 | if (vma == vma_prev) | |
1537 | continue; | |
1538 | ||
1539 | vma_prev = vma; | |
1540 | ||
1541 | if (vma_is_fsdax(vma)) | |
1542 | return true; | |
1543 | } | |
1544 | return false; | |
1545 | } | |
9a4e9f3b AK |
1546 | |
1547 | #ifdef CONFIG_CMA | |
1548 | static struct page *new_non_cma_page(struct page *page, unsigned long private) | |
1549 | { | |
1550 | /* | |
1551 | * We want to make sure we allocate the new page from the same node | |
1552 | * as the source page. | |
1553 | */ | |
1554 | int nid = page_to_nid(page); | |
1555 | /* | |
1556 | * Trying to allocate a page for migration. Ignore allocation | |
1557 | * failure warnings. We don't force __GFP_THISNODE here because | |
1558 | * this node here is the node where we have CMA reservation and | |
1559 | * in some case these nodes will have really less non movable | |
1560 | * allocation memory. | |
1561 | */ | |
1562 | gfp_t gfp_mask = GFP_USER | __GFP_NOWARN; | |
1563 | ||
1564 | if (PageHighMem(page)) | |
1565 | gfp_mask |= __GFP_HIGHMEM; | |
1566 | ||
1567 | #ifdef CONFIG_HUGETLB_PAGE | |
1568 | if (PageHuge(page)) { | |
1569 | struct hstate *h = page_hstate(page); | |
1570 | /* | |
1571 | * We don't want to dequeue from the pool because pool pages will | |
1572 | * mostly be from the CMA region. | |
1573 | */ | |
1574 | return alloc_migrate_huge_page(h, gfp_mask, nid, NULL); | |
1575 | } | |
1576 | #endif | |
1577 | if (PageTransHuge(page)) { | |
1578 | struct page *thp; | |
1579 | /* | |
1580 | * ignore allocation failure warnings | |
1581 | */ | |
1582 | gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN; | |
1583 | ||
1584 | /* | |
1585 | * Remove the movable mask so that we don't allocate from | |
1586 | * CMA area again. | |
1587 | */ | |
1588 | thp_gfpmask &= ~__GFP_MOVABLE; | |
1589 | thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER); | |
1590 | if (!thp) | |
1591 | return NULL; | |
1592 | prep_transhuge_page(thp); | |
1593 | return thp; | |
1594 | } | |
1595 | ||
1596 | return __alloc_pages_node(nid, gfp_mask, 0); | |
1597 | } | |
1598 | ||
932f4a63 IW |
1599 | static long check_and_migrate_cma_pages(struct task_struct *tsk, |
1600 | struct mm_struct *mm, | |
1601 | unsigned long start, | |
1602 | unsigned long nr_pages, | |
9a4e9f3b | 1603 | struct page **pages, |
932f4a63 IW |
1604 | struct vm_area_struct **vmas, |
1605 | unsigned int gup_flags) | |
9a4e9f3b | 1606 | { |
aa712399 PL |
1607 | unsigned long i; |
1608 | unsigned long step; | |
9a4e9f3b AK |
1609 | bool drain_allow = true; |
1610 | bool migrate_allow = true; | |
1611 | LIST_HEAD(cma_page_list); | |
b96cc655 | 1612 | long ret = nr_pages; |
9a4e9f3b AK |
1613 | |
1614 | check_again: | |
aa712399 PL |
1615 | for (i = 0; i < nr_pages;) { |
1616 | ||
1617 | struct page *head = compound_head(pages[i]); | |
1618 | ||
1619 | /* | |
1620 | * gup may start from a tail page. Advance step by the left | |
1621 | * part. | |
1622 | */ | |
d8c6546b | 1623 | step = compound_nr(head) - (pages[i] - head); |
9a4e9f3b AK |
1624 | /* |
1625 | * If we get a page from the CMA zone, since we are going to | |
1626 | * be pinning these entries, we might as well move them out | |
1627 | * of the CMA zone if possible. | |
1628 | */ | |
aa712399 PL |
1629 | if (is_migrate_cma_page(head)) { |
1630 | if (PageHuge(head)) | |
9a4e9f3b | 1631 | isolate_huge_page(head, &cma_page_list); |
aa712399 | 1632 | else { |
9a4e9f3b AK |
1633 | if (!PageLRU(head) && drain_allow) { |
1634 | lru_add_drain_all(); | |
1635 | drain_allow = false; | |
1636 | } | |
1637 | ||
1638 | if (!isolate_lru_page(head)) { | |
1639 | list_add_tail(&head->lru, &cma_page_list); | |
1640 | mod_node_page_state(page_pgdat(head), | |
1641 | NR_ISOLATED_ANON + | |
1642 | page_is_file_cache(head), | |
1643 | hpage_nr_pages(head)); | |
1644 | } | |
1645 | } | |
1646 | } | |
aa712399 PL |
1647 | |
1648 | i += step; | |
9a4e9f3b AK |
1649 | } |
1650 | ||
1651 | if (!list_empty(&cma_page_list)) { | |
1652 | /* | |
1653 | * drop the above get_user_pages reference. | |
1654 | */ | |
1655 | for (i = 0; i < nr_pages; i++) | |
1656 | put_page(pages[i]); | |
1657 | ||
1658 | if (migrate_pages(&cma_page_list, new_non_cma_page, | |
1659 | NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) { | |
1660 | /* | |
1661 | * some of the pages failed migration. Do get_user_pages | |
1662 | * without migration. | |
1663 | */ | |
1664 | migrate_allow = false; | |
1665 | ||
1666 | if (!list_empty(&cma_page_list)) | |
1667 | putback_movable_pages(&cma_page_list); | |
1668 | } | |
1669 | /* | |
932f4a63 IW |
1670 | * We did migrate all the pages, Try to get the page references |
1671 | * again migrating any new CMA pages which we failed to isolate | |
1672 | * earlier. | |
9a4e9f3b | 1673 | */ |
b96cc655 | 1674 | ret = __get_user_pages_locked(tsk, mm, start, nr_pages, |
932f4a63 IW |
1675 | pages, vmas, NULL, |
1676 | gup_flags); | |
1677 | ||
b96cc655 | 1678 | if ((ret > 0) && migrate_allow) { |
1679 | nr_pages = ret; | |
9a4e9f3b AK |
1680 | drain_allow = true; |
1681 | goto check_again; | |
1682 | } | |
1683 | } | |
1684 | ||
b96cc655 | 1685 | return ret; |
9a4e9f3b AK |
1686 | } |
1687 | #else | |
932f4a63 IW |
1688 | static long check_and_migrate_cma_pages(struct task_struct *tsk, |
1689 | struct mm_struct *mm, | |
1690 | unsigned long start, | |
1691 | unsigned long nr_pages, | |
1692 | struct page **pages, | |
1693 | struct vm_area_struct **vmas, | |
1694 | unsigned int gup_flags) | |
9a4e9f3b AK |
1695 | { |
1696 | return nr_pages; | |
1697 | } | |
050a9adc | 1698 | #endif /* CONFIG_CMA */ |
9a4e9f3b | 1699 | |
2bb6d283 | 1700 | /* |
932f4a63 IW |
1701 | * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which |
1702 | * allows us to process the FOLL_LONGTERM flag. | |
2bb6d283 | 1703 | */ |
932f4a63 IW |
1704 | static long __gup_longterm_locked(struct task_struct *tsk, |
1705 | struct mm_struct *mm, | |
1706 | unsigned long start, | |
1707 | unsigned long nr_pages, | |
1708 | struct page **pages, | |
1709 | struct vm_area_struct **vmas, | |
1710 | unsigned int gup_flags) | |
2bb6d283 | 1711 | { |
932f4a63 IW |
1712 | struct vm_area_struct **vmas_tmp = vmas; |
1713 | unsigned long flags = 0; | |
2bb6d283 DW |
1714 | long rc, i; |
1715 | ||
932f4a63 IW |
1716 | if (gup_flags & FOLL_LONGTERM) { |
1717 | if (!pages) | |
1718 | return -EINVAL; | |
1719 | ||
1720 | if (!vmas_tmp) { | |
1721 | vmas_tmp = kcalloc(nr_pages, | |
1722 | sizeof(struct vm_area_struct *), | |
1723 | GFP_KERNEL); | |
1724 | if (!vmas_tmp) | |
1725 | return -ENOMEM; | |
1726 | } | |
1727 | flags = memalloc_nocma_save(); | |
2bb6d283 DW |
1728 | } |
1729 | ||
932f4a63 IW |
1730 | rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, |
1731 | vmas_tmp, NULL, gup_flags); | |
2bb6d283 | 1732 | |
932f4a63 IW |
1733 | if (gup_flags & FOLL_LONGTERM) { |
1734 | memalloc_nocma_restore(flags); | |
1735 | if (rc < 0) | |
1736 | goto out; | |
1737 | ||
1738 | if (check_dax_vmas(vmas_tmp, rc)) { | |
1739 | for (i = 0; i < rc; i++) | |
1740 | put_page(pages[i]); | |
1741 | rc = -EOPNOTSUPP; | |
1742 | goto out; | |
1743 | } | |
1744 | ||
1745 | rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages, | |
1746 | vmas_tmp, gup_flags); | |
9a4e9f3b | 1747 | } |
2bb6d283 | 1748 | |
2bb6d283 | 1749 | out: |
932f4a63 IW |
1750 | if (vmas_tmp != vmas) |
1751 | kfree(vmas_tmp); | |
2bb6d283 DW |
1752 | return rc; |
1753 | } | |
932f4a63 IW |
1754 | #else /* !CONFIG_FS_DAX && !CONFIG_CMA */ |
1755 | static __always_inline long __gup_longterm_locked(struct task_struct *tsk, | |
1756 | struct mm_struct *mm, | |
1757 | unsigned long start, | |
1758 | unsigned long nr_pages, | |
1759 | struct page **pages, | |
1760 | struct vm_area_struct **vmas, | |
1761 | unsigned int flags) | |
1762 | { | |
1763 | return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas, | |
1764 | NULL, flags); | |
1765 | } | |
1766 | #endif /* CONFIG_FS_DAX || CONFIG_CMA */ | |
1767 | ||
22bf29b6 JH |
1768 | #ifdef CONFIG_MMU |
1769 | static long __get_user_pages_remote(struct task_struct *tsk, | |
1770 | struct mm_struct *mm, | |
1771 | unsigned long start, unsigned long nr_pages, | |
1772 | unsigned int gup_flags, struct page **pages, | |
1773 | struct vm_area_struct **vmas, int *locked) | |
1774 | { | |
1775 | /* | |
1776 | * Parts of FOLL_LONGTERM behavior are incompatible with | |
1777 | * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on | |
1778 | * vmas. However, this only comes up if locked is set, and there are | |
1779 | * callers that do request FOLL_LONGTERM, but do not set locked. So, | |
1780 | * allow what we can. | |
1781 | */ | |
1782 | if (gup_flags & FOLL_LONGTERM) { | |
1783 | if (WARN_ON_ONCE(locked)) | |
1784 | return -EINVAL; | |
1785 | /* | |
1786 | * This will check the vmas (even if our vmas arg is NULL) | |
1787 | * and return -ENOTSUPP if DAX isn't allowed in this case: | |
1788 | */ | |
1789 | return __gup_longterm_locked(tsk, mm, start, nr_pages, pages, | |
1790 | vmas, gup_flags | FOLL_TOUCH | | |
1791 | FOLL_REMOTE); | |
1792 | } | |
1793 | ||
1794 | return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas, | |
1795 | locked, | |
1796 | gup_flags | FOLL_TOUCH | FOLL_REMOTE); | |
1797 | } | |
1798 | ||
c4237f8b JH |
1799 | /* |
1800 | * get_user_pages_remote() - pin user pages in memory | |
1801 | * @tsk: the task_struct to use for page fault accounting, or | |
1802 | * NULL if faults are not to be recorded. | |
1803 | * @mm: mm_struct of target mm | |
1804 | * @start: starting user address | |
1805 | * @nr_pages: number of pages from start to pin | |
1806 | * @gup_flags: flags modifying lookup behaviour | |
1807 | * @pages: array that receives pointers to the pages pinned. | |
1808 | * Should be at least nr_pages long. Or NULL, if caller | |
1809 | * only intends to ensure the pages are faulted in. | |
1810 | * @vmas: array of pointers to vmas corresponding to each page. | |
1811 | * Or NULL if the caller does not require them. | |
1812 | * @locked: pointer to lock flag indicating whether lock is held and | |
1813 | * subsequently whether VM_FAULT_RETRY functionality can be | |
1814 | * utilised. Lock must initially be held. | |
1815 | * | |
1816 | * Returns either number of pages pinned (which may be less than the | |
1817 | * number requested), or an error. Details about the return value: | |
1818 | * | |
1819 | * -- If nr_pages is 0, returns 0. | |
1820 | * -- If nr_pages is >0, but no pages were pinned, returns -errno. | |
1821 | * -- If nr_pages is >0, and some pages were pinned, returns the number of | |
1822 | * pages pinned. Again, this may be less than nr_pages. | |
1823 | * | |
1824 | * The caller is responsible for releasing returned @pages, via put_page(). | |
1825 | * | |
1826 | * @vmas are valid only as long as mmap_sem is held. | |
1827 | * | |
1828 | * Must be called with mmap_sem held for read or write. | |
1829 | * | |
1830 | * get_user_pages walks a process's page tables and takes a reference to | |
1831 | * each struct page that each user address corresponds to at a given | |
1832 | * instant. That is, it takes the page that would be accessed if a user | |
1833 | * thread accesses the given user virtual address at that instant. | |
1834 | * | |
1835 | * This does not guarantee that the page exists in the user mappings when | |
1836 | * get_user_pages returns, and there may even be a completely different | |
1837 | * page there in some cases (eg. if mmapped pagecache has been invalidated | |
1838 | * and subsequently re faulted). However it does guarantee that the page | |
1839 | * won't be freed completely. And mostly callers simply care that the page | |
1840 | * contains data that was valid *at some point in time*. Typically, an IO | |
1841 | * or similar operation cannot guarantee anything stronger anyway because | |
1842 | * locks can't be held over the syscall boundary. | |
1843 | * | |
1844 | * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page | |
1845 | * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must | |
1846 | * be called after the page is finished with, and before put_page is called. | |
1847 | * | |
1848 | * get_user_pages is typically used for fewer-copy IO operations, to get a | |
1849 | * handle on the memory by some means other than accesses via the user virtual | |
1850 | * addresses. The pages may be submitted for DMA to devices or accessed via | |
1851 | * their kernel linear mapping (via the kmap APIs). Care should be taken to | |
1852 | * use the correct cache flushing APIs. | |
1853 | * | |
1854 | * See also get_user_pages_fast, for performance critical applications. | |
1855 | * | |
1856 | * get_user_pages should be phased out in favor of | |
1857 | * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing | |
1858 | * should use get_user_pages because it cannot pass | |
1859 | * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault. | |
1860 | */ | |
1861 | long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm, | |
1862 | unsigned long start, unsigned long nr_pages, | |
1863 | unsigned int gup_flags, struct page **pages, | |
1864 | struct vm_area_struct **vmas, int *locked) | |
1865 | { | |
eddb1c22 JH |
1866 | /* |
1867 | * FOLL_PIN must only be set internally by the pin_user_pages*() APIs, | |
1868 | * never directly by the caller, so enforce that with an assertion: | |
1869 | */ | |
1870 | if (WARN_ON_ONCE(gup_flags & FOLL_PIN)) | |
1871 | return -EINVAL; | |
1872 | ||
22bf29b6 JH |
1873 | return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags, |
1874 | pages, vmas, locked); | |
c4237f8b JH |
1875 | } |
1876 | EXPORT_SYMBOL(get_user_pages_remote); | |
1877 | ||
eddb1c22 JH |
1878 | #else /* CONFIG_MMU */ |
1879 | long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm, | |
1880 | unsigned long start, unsigned long nr_pages, | |
1881 | unsigned int gup_flags, struct page **pages, | |
1882 | struct vm_area_struct **vmas, int *locked) | |
1883 | { | |
1884 | return 0; | |
1885 | } | |
3faa52c0 JH |
1886 | |
1887 | static long __get_user_pages_remote(struct task_struct *tsk, | |
1888 | struct mm_struct *mm, | |
1889 | unsigned long start, unsigned long nr_pages, | |
1890 | unsigned int gup_flags, struct page **pages, | |
1891 | struct vm_area_struct **vmas, int *locked) | |
1892 | { | |
1893 | return 0; | |
1894 | } | |
eddb1c22 JH |
1895 | #endif /* !CONFIG_MMU */ |
1896 | ||
932f4a63 IW |
1897 | /* |
1898 | * This is the same as get_user_pages_remote(), just with a | |
1899 | * less-flexible calling convention where we assume that the task | |
1900 | * and mm being operated on are the current task's and don't allow | |
1901 | * passing of a locked parameter. We also obviously don't pass | |
1902 | * FOLL_REMOTE in here. | |
1903 | */ | |
1904 | long get_user_pages(unsigned long start, unsigned long nr_pages, | |
1905 | unsigned int gup_flags, struct page **pages, | |
1906 | struct vm_area_struct **vmas) | |
1907 | { | |
eddb1c22 JH |
1908 | /* |
1909 | * FOLL_PIN must only be set internally by the pin_user_pages*() APIs, | |
1910 | * never directly by the caller, so enforce that with an assertion: | |
1911 | */ | |
1912 | if (WARN_ON_ONCE(gup_flags & FOLL_PIN)) | |
1913 | return -EINVAL; | |
1914 | ||
932f4a63 IW |
1915 | return __gup_longterm_locked(current, current->mm, start, nr_pages, |
1916 | pages, vmas, gup_flags | FOLL_TOUCH); | |
1917 | } | |
1918 | EXPORT_SYMBOL(get_user_pages); | |
2bb6d283 | 1919 | |
d3649f68 CH |
1920 | /* |
1921 | * We can leverage the VM_FAULT_RETRY functionality in the page fault | |
1922 | * paths better by using either get_user_pages_locked() or | |
1923 | * get_user_pages_unlocked(). | |
acc3c8d1 | 1924 | * |
d3649f68 | 1925 | * get_user_pages_locked() is suitable to replace the form: |
acc3c8d1 | 1926 | * |
d3649f68 CH |
1927 | * down_read(&mm->mmap_sem); |
1928 | * do_something() | |
1929 | * get_user_pages(tsk, mm, ..., pages, NULL); | |
1930 | * up_read(&mm->mmap_sem); | |
acc3c8d1 | 1931 | * |
d3649f68 | 1932 | * to: |
acc3c8d1 | 1933 | * |
d3649f68 CH |
1934 | * int locked = 1; |
1935 | * down_read(&mm->mmap_sem); | |
1936 | * do_something() | |
1937 | * get_user_pages_locked(tsk, mm, ..., pages, &locked); | |
1938 | * if (locked) | |
1939 | * up_read(&mm->mmap_sem); | |
acc3c8d1 | 1940 | */ |
d3649f68 CH |
1941 | long get_user_pages_locked(unsigned long start, unsigned long nr_pages, |
1942 | unsigned int gup_flags, struct page **pages, | |
1943 | int *locked) | |
acc3c8d1 | 1944 | { |
acc3c8d1 | 1945 | /* |
d3649f68 CH |
1946 | * FIXME: Current FOLL_LONGTERM behavior is incompatible with |
1947 | * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on | |
1948 | * vmas. As there are no users of this flag in this call we simply | |
1949 | * disallow this option for now. | |
acc3c8d1 | 1950 | */ |
d3649f68 CH |
1951 | if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM)) |
1952 | return -EINVAL; | |
acc3c8d1 | 1953 | |
d3649f68 CH |
1954 | return __get_user_pages_locked(current, current->mm, start, nr_pages, |
1955 | pages, NULL, locked, | |
1956 | gup_flags | FOLL_TOUCH); | |
acc3c8d1 | 1957 | } |
d3649f68 | 1958 | EXPORT_SYMBOL(get_user_pages_locked); |
acc3c8d1 KS |
1959 | |
1960 | /* | |
d3649f68 | 1961 | * get_user_pages_unlocked() is suitable to replace the form: |
acc3c8d1 | 1962 | * |
d3649f68 CH |
1963 | * down_read(&mm->mmap_sem); |
1964 | * get_user_pages(tsk, mm, ..., pages, NULL); | |
1965 | * up_read(&mm->mmap_sem); | |
1966 | * | |
1967 | * with: | |
1968 | * | |
1969 | * get_user_pages_unlocked(tsk, mm, ..., pages); | |
1970 | * | |
1971 | * It is functionally equivalent to get_user_pages_fast so | |
1972 | * get_user_pages_fast should be used instead if specific gup_flags | |
1973 | * (e.g. FOLL_FORCE) are not required. | |
acc3c8d1 | 1974 | */ |
d3649f68 CH |
1975 | long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages, |
1976 | struct page **pages, unsigned int gup_flags) | |
acc3c8d1 KS |
1977 | { |
1978 | struct mm_struct *mm = current->mm; | |
d3649f68 CH |
1979 | int locked = 1; |
1980 | long ret; | |
acc3c8d1 | 1981 | |
d3649f68 CH |
1982 | /* |
1983 | * FIXME: Current FOLL_LONGTERM behavior is incompatible with | |
1984 | * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on | |
1985 | * vmas. As there are no users of this flag in this call we simply | |
1986 | * disallow this option for now. | |
1987 | */ | |
1988 | if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM)) | |
1989 | return -EINVAL; | |
acc3c8d1 | 1990 | |
d3649f68 CH |
1991 | down_read(&mm->mmap_sem); |
1992 | ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL, | |
1993 | &locked, gup_flags | FOLL_TOUCH); | |
acc3c8d1 KS |
1994 | if (locked) |
1995 | up_read(&mm->mmap_sem); | |
d3649f68 | 1996 | return ret; |
4bbd4c77 | 1997 | } |
d3649f68 | 1998 | EXPORT_SYMBOL(get_user_pages_unlocked); |
2667f50e SC |
1999 | |
2000 | /* | |
67a929e0 | 2001 | * Fast GUP |
2667f50e SC |
2002 | * |
2003 | * get_user_pages_fast attempts to pin user pages by walking the page | |
2004 | * tables directly and avoids taking locks. Thus the walker needs to be | |
2005 | * protected from page table pages being freed from under it, and should | |
2006 | * block any THP splits. | |
2007 | * | |
2008 | * One way to achieve this is to have the walker disable interrupts, and | |
2009 | * rely on IPIs from the TLB flushing code blocking before the page table | |
2010 | * pages are freed. This is unsuitable for architectures that do not need | |
2011 | * to broadcast an IPI when invalidating TLBs. | |
2012 | * | |
2013 | * Another way to achieve this is to batch up page table containing pages | |
2014 | * belonging to more than one mm_user, then rcu_sched a callback to free those | |
2015 | * pages. Disabling interrupts will allow the fast_gup walker to both block | |
2016 | * the rcu_sched callback, and an IPI that we broadcast for splitting THPs | |
2017 | * (which is a relatively rare event). The code below adopts this strategy. | |
2018 | * | |
2019 | * Before activating this code, please be aware that the following assumptions | |
2020 | * are currently made: | |
2021 | * | |
ff2e6d72 | 2022 | * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to |
e585513b | 2023 | * free pages containing page tables or TLB flushing requires IPI broadcast. |
2667f50e | 2024 | * |
2667f50e SC |
2025 | * *) ptes can be read atomically by the architecture. |
2026 | * | |
2027 | * *) access_ok is sufficient to validate userspace address ranges. | |
2028 | * | |
2029 | * The last two assumptions can be relaxed by the addition of helper functions. | |
2030 | * | |
2031 | * This code is based heavily on the PowerPC implementation by Nick Piggin. | |
2032 | */ | |
67a929e0 | 2033 | #ifdef CONFIG_HAVE_FAST_GUP |
3faa52c0 JH |
2034 | |
2035 | static void put_compound_head(struct page *page, int refs, unsigned int flags) | |
2036 | { | |
47e29d32 | 2037 | if (flags & FOLL_PIN) { |
1970dc6f JH |
2038 | mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, |
2039 | refs); | |
2040 | ||
47e29d32 JH |
2041 | if (hpage_pincount_available(page)) |
2042 | hpage_pincount_sub(page, refs); | |
2043 | else | |
2044 | refs *= GUP_PIN_COUNTING_BIAS; | |
2045 | } | |
3faa52c0 JH |
2046 | |
2047 | VM_BUG_ON_PAGE(page_ref_count(page) < refs, page); | |
2048 | /* | |
2049 | * Calling put_page() for each ref is unnecessarily slow. Only the last | |
2050 | * ref needs a put_page(). | |
2051 | */ | |
2052 | if (refs > 1) | |
2053 | page_ref_sub(page, refs - 1); | |
2054 | put_page(page); | |
2055 | } | |
2056 | ||
39656e83 | 2057 | #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH |
3faa52c0 | 2058 | |
39656e83 CH |
2059 | /* |
2060 | * WARNING: only to be used in the get_user_pages_fast() implementation. | |
2061 | * | |
2062 | * With get_user_pages_fast(), we walk down the pagetables without taking any | |
2063 | * locks. For this we would like to load the pointers atomically, but sometimes | |
2064 | * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What | |
2065 | * we do have is the guarantee that a PTE will only either go from not present | |
2066 | * to present, or present to not present or both -- it will not switch to a | |
2067 | * completely different present page without a TLB flush in between; something | |
2068 | * that we are blocking by holding interrupts off. | |
2069 | * | |
2070 | * Setting ptes from not present to present goes: | |
2071 | * | |
2072 | * ptep->pte_high = h; | |
2073 | * smp_wmb(); | |
2074 | * ptep->pte_low = l; | |
2075 | * | |
2076 | * And present to not present goes: | |
2077 | * | |
2078 | * ptep->pte_low = 0; | |
2079 | * smp_wmb(); | |
2080 | * ptep->pte_high = 0; | |
2081 | * | |
2082 | * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'. | |
2083 | * We load pte_high *after* loading pte_low, which ensures we don't see an older | |
2084 | * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't | |
2085 | * picked up a changed pte high. We might have gotten rubbish values from | |
2086 | * pte_low and pte_high, but we are guaranteed that pte_low will not have the | |
2087 | * present bit set *unless* it is 'l'. Because get_user_pages_fast() only | |
2088 | * operates on present ptes we're safe. | |
2089 | */ | |
2090 | static inline pte_t gup_get_pte(pte_t *ptep) | |
2091 | { | |
2092 | pte_t pte; | |
2667f50e | 2093 | |
39656e83 CH |
2094 | do { |
2095 | pte.pte_low = ptep->pte_low; | |
2096 | smp_rmb(); | |
2097 | pte.pte_high = ptep->pte_high; | |
2098 | smp_rmb(); | |
2099 | } while (unlikely(pte.pte_low != ptep->pte_low)); | |
2100 | ||
2101 | return pte; | |
2102 | } | |
2103 | #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */ | |
0005d20b | 2104 | /* |
39656e83 | 2105 | * We require that the PTE can be read atomically. |
0005d20b KS |
2106 | */ |
2107 | static inline pte_t gup_get_pte(pte_t *ptep) | |
2108 | { | |
2109 | return READ_ONCE(*ptep); | |
2110 | } | |
39656e83 | 2111 | #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */ |
0005d20b | 2112 | |
790c7369 | 2113 | static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start, |
3b78d834 | 2114 | unsigned int flags, |
790c7369 | 2115 | struct page **pages) |
b59f65fa KS |
2116 | { |
2117 | while ((*nr) - nr_start) { | |
2118 | struct page *page = pages[--(*nr)]; | |
2119 | ||
2120 | ClearPageReferenced(page); | |
3faa52c0 JH |
2121 | if (flags & FOLL_PIN) |
2122 | unpin_user_page(page); | |
2123 | else | |
2124 | put_page(page); | |
b59f65fa KS |
2125 | } |
2126 | } | |
2127 | ||
3010a5ea | 2128 | #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL |
2667f50e | 2129 | static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end, |
b798bec4 | 2130 | unsigned int flags, struct page **pages, int *nr) |
2667f50e | 2131 | { |
b59f65fa KS |
2132 | struct dev_pagemap *pgmap = NULL; |
2133 | int nr_start = *nr, ret = 0; | |
2667f50e | 2134 | pte_t *ptep, *ptem; |
2667f50e SC |
2135 | |
2136 | ptem = ptep = pte_offset_map(&pmd, addr); | |
2137 | do { | |
0005d20b | 2138 | pte_t pte = gup_get_pte(ptep); |
7aef4172 | 2139 | struct page *head, *page; |
2667f50e SC |
2140 | |
2141 | /* | |
2142 | * Similar to the PMD case below, NUMA hinting must take slow | |
8a0516ed | 2143 | * path using the pte_protnone check. |
2667f50e | 2144 | */ |
e7884f8e KS |
2145 | if (pte_protnone(pte)) |
2146 | goto pte_unmap; | |
2147 | ||
b798bec4 | 2148 | if (!pte_access_permitted(pte, flags & FOLL_WRITE)) |
e7884f8e KS |
2149 | goto pte_unmap; |
2150 | ||
b59f65fa | 2151 | if (pte_devmap(pte)) { |
7af75561 IW |
2152 | if (unlikely(flags & FOLL_LONGTERM)) |
2153 | goto pte_unmap; | |
2154 | ||
b59f65fa KS |
2155 | pgmap = get_dev_pagemap(pte_pfn(pte), pgmap); |
2156 | if (unlikely(!pgmap)) { | |
3b78d834 | 2157 | undo_dev_pagemap(nr, nr_start, flags, pages); |
b59f65fa KS |
2158 | goto pte_unmap; |
2159 | } | |
2160 | } else if (pte_special(pte)) | |
2667f50e SC |
2161 | goto pte_unmap; |
2162 | ||
2163 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
2164 | page = pte_page(pte); | |
2165 | ||
3faa52c0 | 2166 | head = try_grab_compound_head(page, 1, flags); |
8fde12ca | 2167 | if (!head) |
2667f50e SC |
2168 | goto pte_unmap; |
2169 | ||
2170 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
3faa52c0 | 2171 | put_compound_head(head, 1, flags); |
2667f50e SC |
2172 | goto pte_unmap; |
2173 | } | |
2174 | ||
7aef4172 | 2175 | VM_BUG_ON_PAGE(compound_head(page) != head, page); |
e9348053 | 2176 | |
f28d4363 CI |
2177 | /* |
2178 | * We need to make the page accessible if and only if we are | |
2179 | * going to access its content (the FOLL_PIN case). Please | |
2180 | * see Documentation/core-api/pin_user_pages.rst for | |
2181 | * details. | |
2182 | */ | |
2183 | if (flags & FOLL_PIN) { | |
2184 | ret = arch_make_page_accessible(page); | |
2185 | if (ret) { | |
2186 | unpin_user_page(page); | |
2187 | goto pte_unmap; | |
2188 | } | |
2189 | } | |
e9348053 | 2190 | SetPageReferenced(page); |
2667f50e SC |
2191 | pages[*nr] = page; |
2192 | (*nr)++; | |
2193 | ||
2194 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
2195 | ||
2196 | ret = 1; | |
2197 | ||
2198 | pte_unmap: | |
832d7aa0 CH |
2199 | if (pgmap) |
2200 | put_dev_pagemap(pgmap); | |
2667f50e SC |
2201 | pte_unmap(ptem); |
2202 | return ret; | |
2203 | } | |
2204 | #else | |
2205 | ||
2206 | /* | |
2207 | * If we can't determine whether or not a pte is special, then fail immediately | |
2208 | * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not | |
2209 | * to be special. | |
2210 | * | |
2211 | * For a futex to be placed on a THP tail page, get_futex_key requires a | |
2212 | * __get_user_pages_fast implementation that can pin pages. Thus it's still | |
2213 | * useful to have gup_huge_pmd even if we can't operate on ptes. | |
2214 | */ | |
2215 | static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end, | |
b798bec4 | 2216 | unsigned int flags, struct page **pages, int *nr) |
2667f50e SC |
2217 | { |
2218 | return 0; | |
2219 | } | |
3010a5ea | 2220 | #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */ |
2667f50e | 2221 | |
17596731 | 2222 | #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
b59f65fa | 2223 | static int __gup_device_huge(unsigned long pfn, unsigned long addr, |
86dfbed4 JH |
2224 | unsigned long end, unsigned int flags, |
2225 | struct page **pages, int *nr) | |
b59f65fa KS |
2226 | { |
2227 | int nr_start = *nr; | |
2228 | struct dev_pagemap *pgmap = NULL; | |
2229 | ||
2230 | do { | |
2231 | struct page *page = pfn_to_page(pfn); | |
2232 | ||
2233 | pgmap = get_dev_pagemap(pfn, pgmap); | |
2234 | if (unlikely(!pgmap)) { | |
3b78d834 | 2235 | undo_dev_pagemap(nr, nr_start, flags, pages); |
b59f65fa KS |
2236 | return 0; |
2237 | } | |
2238 | SetPageReferenced(page); | |
2239 | pages[*nr] = page; | |
3faa52c0 JH |
2240 | if (unlikely(!try_grab_page(page, flags))) { |
2241 | undo_dev_pagemap(nr, nr_start, flags, pages); | |
2242 | return 0; | |
2243 | } | |
b59f65fa KS |
2244 | (*nr)++; |
2245 | pfn++; | |
2246 | } while (addr += PAGE_SIZE, addr != end); | |
832d7aa0 CH |
2247 | |
2248 | if (pgmap) | |
2249 | put_dev_pagemap(pgmap); | |
b59f65fa KS |
2250 | return 1; |
2251 | } | |
2252 | ||
a9b6de77 | 2253 | static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
86dfbed4 JH |
2254 | unsigned long end, unsigned int flags, |
2255 | struct page **pages, int *nr) | |
b59f65fa KS |
2256 | { |
2257 | unsigned long fault_pfn; | |
a9b6de77 DW |
2258 | int nr_start = *nr; |
2259 | ||
2260 | fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
86dfbed4 | 2261 | if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 2262 | return 0; |
b59f65fa | 2263 | |
a9b6de77 | 2264 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { |
3b78d834 | 2265 | undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
2266 | return 0; |
2267 | } | |
2268 | return 1; | |
b59f65fa KS |
2269 | } |
2270 | ||
a9b6de77 | 2271 | static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr, |
86dfbed4 JH |
2272 | unsigned long end, unsigned int flags, |
2273 | struct page **pages, int *nr) | |
b59f65fa KS |
2274 | { |
2275 | unsigned long fault_pfn; | |
a9b6de77 DW |
2276 | int nr_start = *nr; |
2277 | ||
2278 | fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
86dfbed4 | 2279 | if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr)) |
a9b6de77 | 2280 | return 0; |
b59f65fa | 2281 | |
a9b6de77 | 2282 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { |
3b78d834 | 2283 | undo_dev_pagemap(nr, nr_start, flags, pages); |
a9b6de77 DW |
2284 | return 0; |
2285 | } | |
2286 | return 1; | |
b59f65fa KS |
2287 | } |
2288 | #else | |
a9b6de77 | 2289 | static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
86dfbed4 JH |
2290 | unsigned long end, unsigned int flags, |
2291 | struct page **pages, int *nr) | |
b59f65fa KS |
2292 | { |
2293 | BUILD_BUG(); | |
2294 | return 0; | |
2295 | } | |
2296 | ||
a9b6de77 | 2297 | static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr, |
86dfbed4 JH |
2298 | unsigned long end, unsigned int flags, |
2299 | struct page **pages, int *nr) | |
b59f65fa KS |
2300 | { |
2301 | BUILD_BUG(); | |
2302 | return 0; | |
2303 | } | |
2304 | #endif | |
2305 | ||
a43e9820 JH |
2306 | static int record_subpages(struct page *page, unsigned long addr, |
2307 | unsigned long end, struct page **pages) | |
2308 | { | |
2309 | int nr; | |
2310 | ||
2311 | for (nr = 0; addr != end; addr += PAGE_SIZE) | |
2312 | pages[nr++] = page++; | |
2313 | ||
2314 | return nr; | |
2315 | } | |
2316 | ||
cbd34da7 CH |
2317 | #ifdef CONFIG_ARCH_HAS_HUGEPD |
2318 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, | |
2319 | unsigned long sz) | |
2320 | { | |
2321 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
2322 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
2323 | } | |
2324 | ||
2325 | static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, | |
0cd22afd JH |
2326 | unsigned long end, unsigned int flags, |
2327 | struct page **pages, int *nr) | |
cbd34da7 CH |
2328 | { |
2329 | unsigned long pte_end; | |
2330 | struct page *head, *page; | |
2331 | pte_t pte; | |
2332 | int refs; | |
2333 | ||
2334 | pte_end = (addr + sz) & ~(sz-1); | |
2335 | if (pte_end < end) | |
2336 | end = pte_end; | |
2337 | ||
2338 | pte = READ_ONCE(*ptep); | |
2339 | ||
0cd22afd | 2340 | if (!pte_access_permitted(pte, flags & FOLL_WRITE)) |
cbd34da7 CH |
2341 | return 0; |
2342 | ||
2343 | /* hugepages are never "special" */ | |
2344 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
2345 | ||
cbd34da7 | 2346 | head = pte_page(pte); |
cbd34da7 | 2347 | page = head + ((addr & (sz-1)) >> PAGE_SHIFT); |
a43e9820 | 2348 | refs = record_subpages(page, addr, end, pages + *nr); |
cbd34da7 | 2349 | |
3faa52c0 | 2350 | head = try_grab_compound_head(head, refs, flags); |
a43e9820 | 2351 | if (!head) |
cbd34da7 | 2352 | return 0; |
cbd34da7 CH |
2353 | |
2354 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
3b78d834 | 2355 | put_compound_head(head, refs, flags); |
cbd34da7 CH |
2356 | return 0; |
2357 | } | |
2358 | ||
a43e9820 | 2359 | *nr += refs; |
520b4a44 | 2360 | SetPageReferenced(head); |
cbd34da7 CH |
2361 | return 1; |
2362 | } | |
2363 | ||
2364 | static int gup_huge_pd(hugepd_t hugepd, unsigned long addr, | |
0cd22afd | 2365 | unsigned int pdshift, unsigned long end, unsigned int flags, |
cbd34da7 CH |
2366 | struct page **pages, int *nr) |
2367 | { | |
2368 | pte_t *ptep; | |
2369 | unsigned long sz = 1UL << hugepd_shift(hugepd); | |
2370 | unsigned long next; | |
2371 | ||
2372 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
2373 | do { | |
2374 | next = hugepte_addr_end(addr, end, sz); | |
0cd22afd | 2375 | if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr)) |
cbd34da7 CH |
2376 | return 0; |
2377 | } while (ptep++, addr = next, addr != end); | |
2378 | ||
2379 | return 1; | |
2380 | } | |
2381 | #else | |
2382 | static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr, | |
0cd22afd | 2383 | unsigned int pdshift, unsigned long end, unsigned int flags, |
cbd34da7 CH |
2384 | struct page **pages, int *nr) |
2385 | { | |
2386 | return 0; | |
2387 | } | |
2388 | #endif /* CONFIG_ARCH_HAS_HUGEPD */ | |
2389 | ||
2667f50e | 2390 | static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, |
0cd22afd JH |
2391 | unsigned long end, unsigned int flags, |
2392 | struct page **pages, int *nr) | |
2667f50e | 2393 | { |
ddc58f27 | 2394 | struct page *head, *page; |
2667f50e SC |
2395 | int refs; |
2396 | ||
b798bec4 | 2397 | if (!pmd_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
2398 | return 0; |
2399 | ||
7af75561 IW |
2400 | if (pmd_devmap(orig)) { |
2401 | if (unlikely(flags & FOLL_LONGTERM)) | |
2402 | return 0; | |
86dfbed4 JH |
2403 | return __gup_device_huge_pmd(orig, pmdp, addr, end, flags, |
2404 | pages, nr); | |
7af75561 | 2405 | } |
b59f65fa | 2406 | |
d63206ee | 2407 | page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); |
a43e9820 | 2408 | refs = record_subpages(page, addr, end, pages + *nr); |
2667f50e | 2409 | |
3faa52c0 | 2410 | head = try_grab_compound_head(pmd_page(orig), refs, flags); |
a43e9820 | 2411 | if (!head) |
2667f50e | 2412 | return 0; |
2667f50e SC |
2413 | |
2414 | if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { | |
3b78d834 | 2415 | put_compound_head(head, refs, flags); |
2667f50e SC |
2416 | return 0; |
2417 | } | |
2418 | ||
a43e9820 | 2419 | *nr += refs; |
e9348053 | 2420 | SetPageReferenced(head); |
2667f50e SC |
2421 | return 1; |
2422 | } | |
2423 | ||
2424 | static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr, | |
86dfbed4 JH |
2425 | unsigned long end, unsigned int flags, |
2426 | struct page **pages, int *nr) | |
2667f50e | 2427 | { |
ddc58f27 | 2428 | struct page *head, *page; |
2667f50e SC |
2429 | int refs; |
2430 | ||
b798bec4 | 2431 | if (!pud_access_permitted(orig, flags & FOLL_WRITE)) |
2667f50e SC |
2432 | return 0; |
2433 | ||
7af75561 IW |
2434 | if (pud_devmap(orig)) { |
2435 | if (unlikely(flags & FOLL_LONGTERM)) | |
2436 | return 0; | |
86dfbed4 JH |
2437 | return __gup_device_huge_pud(orig, pudp, addr, end, flags, |
2438 | pages, nr); | |
7af75561 | 2439 | } |
b59f65fa | 2440 | |
d63206ee | 2441 | page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); |
a43e9820 | 2442 | refs = record_subpages(page, addr, end, pages + *nr); |
2667f50e | 2443 | |
3faa52c0 | 2444 | head = try_grab_compound_head(pud_page(orig), refs, flags); |
a43e9820 | 2445 | if (!head) |
2667f50e | 2446 | return 0; |
2667f50e SC |
2447 | |
2448 | if (unlikely(pud_val(orig) != pud_val(*pudp))) { | |
3b78d834 | 2449 | put_compound_head(head, refs, flags); |
2667f50e SC |
2450 | return 0; |
2451 | } | |
2452 | ||
a43e9820 | 2453 | *nr += refs; |
e9348053 | 2454 | SetPageReferenced(head); |
2667f50e SC |
2455 | return 1; |
2456 | } | |
2457 | ||
f30c59e9 | 2458 | static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr, |
b798bec4 | 2459 | unsigned long end, unsigned int flags, |
f30c59e9 AK |
2460 | struct page **pages, int *nr) |
2461 | { | |
2462 | int refs; | |
ddc58f27 | 2463 | struct page *head, *page; |
f30c59e9 | 2464 | |
b798bec4 | 2465 | if (!pgd_access_permitted(orig, flags & FOLL_WRITE)) |
f30c59e9 AK |
2466 | return 0; |
2467 | ||
b59f65fa | 2468 | BUILD_BUG_ON(pgd_devmap(orig)); |
a43e9820 | 2469 | |
d63206ee | 2470 | page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT); |
a43e9820 | 2471 | refs = record_subpages(page, addr, end, pages + *nr); |
f30c59e9 | 2472 | |
3faa52c0 | 2473 | head = try_grab_compound_head(pgd_page(orig), refs, flags); |
a43e9820 | 2474 | if (!head) |
f30c59e9 | 2475 | return 0; |
f30c59e9 AK |
2476 | |
2477 | if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) { | |
3b78d834 | 2478 | put_compound_head(head, refs, flags); |
f30c59e9 AK |
2479 | return 0; |
2480 | } | |
2481 | ||
a43e9820 | 2482 | *nr += refs; |
e9348053 | 2483 | SetPageReferenced(head); |
f30c59e9 AK |
2484 | return 1; |
2485 | } | |
2486 | ||
2667f50e | 2487 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, |
b798bec4 | 2488 | unsigned int flags, struct page **pages, int *nr) |
2667f50e SC |
2489 | { |
2490 | unsigned long next; | |
2491 | pmd_t *pmdp; | |
2492 | ||
2493 | pmdp = pmd_offset(&pud, addr); | |
2494 | do { | |
38c5ce93 | 2495 | pmd_t pmd = READ_ONCE(*pmdp); |
2667f50e SC |
2496 | |
2497 | next = pmd_addr_end(addr, end); | |
84c3fc4e | 2498 | if (!pmd_present(pmd)) |
2667f50e SC |
2499 | return 0; |
2500 | ||
414fd080 YZ |
2501 | if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) || |
2502 | pmd_devmap(pmd))) { | |
2667f50e SC |
2503 | /* |
2504 | * NUMA hinting faults need to be handled in the GUP | |
2505 | * slowpath for accounting purposes and so that they | |
2506 | * can be serialised against THP migration. | |
2507 | */ | |
8a0516ed | 2508 | if (pmd_protnone(pmd)) |
2667f50e SC |
2509 | return 0; |
2510 | ||
b798bec4 | 2511 | if (!gup_huge_pmd(pmd, pmdp, addr, next, flags, |
2667f50e SC |
2512 | pages, nr)) |
2513 | return 0; | |
2514 | ||
f30c59e9 AK |
2515 | } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) { |
2516 | /* | |
2517 | * architecture have different format for hugetlbfs | |
2518 | * pmd format and THP pmd format | |
2519 | */ | |
2520 | if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr, | |
b798bec4 | 2521 | PMD_SHIFT, next, flags, pages, nr)) |
f30c59e9 | 2522 | return 0; |
b798bec4 | 2523 | } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr)) |
2923117b | 2524 | return 0; |
2667f50e SC |
2525 | } while (pmdp++, addr = next, addr != end); |
2526 | ||
2527 | return 1; | |
2528 | } | |
2529 | ||
c2febafc | 2530 | static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end, |
b798bec4 | 2531 | unsigned int flags, struct page **pages, int *nr) |
2667f50e SC |
2532 | { |
2533 | unsigned long next; | |
2534 | pud_t *pudp; | |
2535 | ||
c2febafc | 2536 | pudp = pud_offset(&p4d, addr); |
2667f50e | 2537 | do { |
e37c6982 | 2538 | pud_t pud = READ_ONCE(*pudp); |
2667f50e SC |
2539 | |
2540 | next = pud_addr_end(addr, end); | |
15494520 | 2541 | if (unlikely(!pud_present(pud))) |
2667f50e | 2542 | return 0; |
f30c59e9 | 2543 | if (unlikely(pud_huge(pud))) { |
b798bec4 | 2544 | if (!gup_huge_pud(pud, pudp, addr, next, flags, |
f30c59e9 AK |
2545 | pages, nr)) |
2546 | return 0; | |
2547 | } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) { | |
2548 | if (!gup_huge_pd(__hugepd(pud_val(pud)), addr, | |
b798bec4 | 2549 | PUD_SHIFT, next, flags, pages, nr)) |
2667f50e | 2550 | return 0; |
b798bec4 | 2551 | } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr)) |
2667f50e SC |
2552 | return 0; |
2553 | } while (pudp++, addr = next, addr != end); | |
2554 | ||
2555 | return 1; | |
2556 | } | |
2557 | ||
c2febafc | 2558 | static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end, |
b798bec4 | 2559 | unsigned int flags, struct page **pages, int *nr) |
c2febafc KS |
2560 | { |
2561 | unsigned long next; | |
2562 | p4d_t *p4dp; | |
2563 | ||
2564 | p4dp = p4d_offset(&pgd, addr); | |
2565 | do { | |
2566 | p4d_t p4d = READ_ONCE(*p4dp); | |
2567 | ||
2568 | next = p4d_addr_end(addr, end); | |
2569 | if (p4d_none(p4d)) | |
2570 | return 0; | |
2571 | BUILD_BUG_ON(p4d_huge(p4d)); | |
2572 | if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) { | |
2573 | if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr, | |
b798bec4 | 2574 | P4D_SHIFT, next, flags, pages, nr)) |
c2febafc | 2575 | return 0; |
b798bec4 | 2576 | } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr)) |
c2febafc KS |
2577 | return 0; |
2578 | } while (p4dp++, addr = next, addr != end); | |
2579 | ||
2580 | return 1; | |
2581 | } | |
2582 | ||
5b65c467 | 2583 | static void gup_pgd_range(unsigned long addr, unsigned long end, |
b798bec4 | 2584 | unsigned int flags, struct page **pages, int *nr) |
5b65c467 KS |
2585 | { |
2586 | unsigned long next; | |
2587 | pgd_t *pgdp; | |
2588 | ||
2589 | pgdp = pgd_offset(current->mm, addr); | |
2590 | do { | |
2591 | pgd_t pgd = READ_ONCE(*pgdp); | |
2592 | ||
2593 | next = pgd_addr_end(addr, end); | |
2594 | if (pgd_none(pgd)) | |
2595 | return; | |
2596 | if (unlikely(pgd_huge(pgd))) { | |
b798bec4 | 2597 | if (!gup_huge_pgd(pgd, pgdp, addr, next, flags, |
5b65c467 KS |
2598 | pages, nr)) |
2599 | return; | |
2600 | } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) { | |
2601 | if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr, | |
b798bec4 | 2602 | PGDIR_SHIFT, next, flags, pages, nr)) |
5b65c467 | 2603 | return; |
b798bec4 | 2604 | } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr)) |
5b65c467 KS |
2605 | return; |
2606 | } while (pgdp++, addr = next, addr != end); | |
2607 | } | |
050a9adc CH |
2608 | #else |
2609 | static inline void gup_pgd_range(unsigned long addr, unsigned long end, | |
2610 | unsigned int flags, struct page **pages, int *nr) | |
2611 | { | |
2612 | } | |
2613 | #endif /* CONFIG_HAVE_FAST_GUP */ | |
5b65c467 KS |
2614 | |
2615 | #ifndef gup_fast_permitted | |
2616 | /* | |
2617 | * Check if it's allowed to use __get_user_pages_fast() for the range, or | |
2618 | * we need to fall back to the slow version: | |
2619 | */ | |
26f4c328 | 2620 | static bool gup_fast_permitted(unsigned long start, unsigned long end) |
5b65c467 | 2621 | { |
26f4c328 | 2622 | return true; |
5b65c467 KS |
2623 | } |
2624 | #endif | |
2625 | ||
2667f50e SC |
2626 | /* |
2627 | * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to | |
d0811078 MT |
2628 | * the regular GUP. |
2629 | * Note a difference with get_user_pages_fast: this always returns the | |
2630 | * number of pages pinned, 0 if no pages were pinned. | |
050a9adc CH |
2631 | * |
2632 | * If the architecture does not support this function, simply return with no | |
2633 | * pages pinned. | |
2667f50e SC |
2634 | */ |
2635 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
2636 | struct page **pages) | |
2637 | { | |
d4faa402 | 2638 | unsigned long len, end; |
5b65c467 | 2639 | unsigned long flags; |
4628b063 | 2640 | int nr_pinned = 0; |
94202f12 JH |
2641 | /* |
2642 | * Internally (within mm/gup.c), gup fast variants must set FOLL_GET, | |
2643 | * because gup fast is always a "pin with a +1 page refcount" request. | |
2644 | */ | |
2645 | unsigned int gup_flags = FOLL_GET; | |
2646 | ||
2647 | if (write) | |
2648 | gup_flags |= FOLL_WRITE; | |
2667f50e | 2649 | |
f455c854 | 2650 | start = untagged_addr(start) & PAGE_MASK; |
2667f50e SC |
2651 | len = (unsigned long) nr_pages << PAGE_SHIFT; |
2652 | end = start + len; | |
2653 | ||
26f4c328 CH |
2654 | if (end <= start) |
2655 | return 0; | |
96d4f267 | 2656 | if (unlikely(!access_ok((void __user *)start, len))) |
2667f50e SC |
2657 | return 0; |
2658 | ||
2659 | /* | |
2660 | * Disable interrupts. We use the nested form as we can already have | |
2661 | * interrupts disabled by get_futex_key. | |
2662 | * | |
2663 | * With interrupts disabled, we block page table pages from being | |
2ebe8228 FW |
2664 | * freed from under us. See struct mmu_table_batch comments in |
2665 | * include/asm-generic/tlb.h for more details. | |
2667f50e SC |
2666 | * |
2667 | * We do not adopt an rcu_read_lock(.) here as we also want to | |
2668 | * block IPIs that come from THPs splitting. | |
2669 | */ | |
2670 | ||
050a9adc CH |
2671 | if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && |
2672 | gup_fast_permitted(start, end)) { | |
5b65c467 | 2673 | local_irq_save(flags); |
4628b063 | 2674 | gup_pgd_range(start, end, gup_flags, pages, &nr_pinned); |
5b65c467 KS |
2675 | local_irq_restore(flags); |
2676 | } | |
2667f50e | 2677 | |
4628b063 | 2678 | return nr_pinned; |
2667f50e | 2679 | } |
050a9adc | 2680 | EXPORT_SYMBOL_GPL(__get_user_pages_fast); |
2667f50e | 2681 | |
7af75561 IW |
2682 | static int __gup_longterm_unlocked(unsigned long start, int nr_pages, |
2683 | unsigned int gup_flags, struct page **pages) | |
2684 | { | |
2685 | int ret; | |
2686 | ||
2687 | /* | |
2688 | * FIXME: FOLL_LONGTERM does not work with | |
2689 | * get_user_pages_unlocked() (see comments in that function) | |
2690 | */ | |
2691 | if (gup_flags & FOLL_LONGTERM) { | |
2692 | down_read(¤t->mm->mmap_sem); | |
2693 | ret = __gup_longterm_locked(current, current->mm, | |
2694 | start, nr_pages, | |
2695 | pages, NULL, gup_flags); | |
2696 | up_read(¤t->mm->mmap_sem); | |
2697 | } else { | |
2698 | ret = get_user_pages_unlocked(start, nr_pages, | |
2699 | pages, gup_flags); | |
2700 | } | |
2701 | ||
2702 | return ret; | |
2703 | } | |
2704 | ||
eddb1c22 JH |
2705 | static int internal_get_user_pages_fast(unsigned long start, int nr_pages, |
2706 | unsigned int gup_flags, | |
2707 | struct page **pages) | |
2667f50e | 2708 | { |
5b65c467 | 2709 | unsigned long addr, len, end; |
4628b063 | 2710 | int nr_pinned = 0, ret = 0; |
2667f50e | 2711 | |
f4000fdf | 2712 | if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM | |
94202f12 | 2713 | FOLL_FORCE | FOLL_PIN | FOLL_GET))) |
817be129 CH |
2714 | return -EINVAL; |
2715 | ||
f455c854 | 2716 | start = untagged_addr(start) & PAGE_MASK; |
5b65c467 KS |
2717 | addr = start; |
2718 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
2719 | end = start + len; | |
2720 | ||
26f4c328 | 2721 | if (end <= start) |
c61611f7 | 2722 | return 0; |
96d4f267 | 2723 | if (unlikely(!access_ok((void __user *)start, len))) |
c61611f7 | 2724 | return -EFAULT; |
73e10a61 | 2725 | |
050a9adc CH |
2726 | if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && |
2727 | gup_fast_permitted(start, end)) { | |
5b65c467 | 2728 | local_irq_disable(); |
4628b063 | 2729 | gup_pgd_range(addr, end, gup_flags, pages, &nr_pinned); |
5b65c467 | 2730 | local_irq_enable(); |
4628b063 | 2731 | ret = nr_pinned; |
73e10a61 | 2732 | } |
2667f50e | 2733 | |
4628b063 | 2734 | if (nr_pinned < nr_pages) { |
2667f50e | 2735 | /* Try to get the remaining pages with get_user_pages */ |
4628b063 PL |
2736 | start += nr_pinned << PAGE_SHIFT; |
2737 | pages += nr_pinned; | |
2667f50e | 2738 | |
4628b063 | 2739 | ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, |
7af75561 | 2740 | gup_flags, pages); |
2667f50e SC |
2741 | |
2742 | /* Have to be a bit careful with return values */ | |
4628b063 | 2743 | if (nr_pinned > 0) { |
2667f50e | 2744 | if (ret < 0) |
4628b063 | 2745 | ret = nr_pinned; |
2667f50e | 2746 | else |
4628b063 | 2747 | ret += nr_pinned; |
2667f50e SC |
2748 | } |
2749 | } | |
2750 | ||
2751 | return ret; | |
2752 | } | |
eddb1c22 JH |
2753 | |
2754 | /** | |
2755 | * get_user_pages_fast() - pin user pages in memory | |
3faa52c0 JH |
2756 | * @start: starting user address |
2757 | * @nr_pages: number of pages from start to pin | |
2758 | * @gup_flags: flags modifying pin behaviour | |
2759 | * @pages: array that receives pointers to the pages pinned. | |
2760 | * Should be at least nr_pages long. | |
eddb1c22 JH |
2761 | * |
2762 | * Attempt to pin user pages in memory without taking mm->mmap_sem. | |
2763 | * If not successful, it will fall back to taking the lock and | |
2764 | * calling get_user_pages(). | |
2765 | * | |
2766 | * Returns number of pages pinned. This may be fewer than the number requested. | |
2767 | * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns | |
2768 | * -errno. | |
2769 | */ | |
2770 | int get_user_pages_fast(unsigned long start, int nr_pages, | |
2771 | unsigned int gup_flags, struct page **pages) | |
2772 | { | |
2773 | /* | |
2774 | * FOLL_PIN must only be set internally by the pin_user_pages*() APIs, | |
2775 | * never directly by the caller, so enforce that: | |
2776 | */ | |
2777 | if (WARN_ON_ONCE(gup_flags & FOLL_PIN)) | |
2778 | return -EINVAL; | |
2779 | ||
94202f12 JH |
2780 | /* |
2781 | * The caller may or may not have explicitly set FOLL_GET; either way is | |
2782 | * OK. However, internally (within mm/gup.c), gup fast variants must set | |
2783 | * FOLL_GET, because gup fast is always a "pin with a +1 page refcount" | |
2784 | * request. | |
2785 | */ | |
2786 | gup_flags |= FOLL_GET; | |
eddb1c22 JH |
2787 | return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages); |
2788 | } | |
050a9adc | 2789 | EXPORT_SYMBOL_GPL(get_user_pages_fast); |
eddb1c22 JH |
2790 | |
2791 | /** | |
2792 | * pin_user_pages_fast() - pin user pages in memory without taking locks | |
2793 | * | |
3faa52c0 JH |
2794 | * @start: starting user address |
2795 | * @nr_pages: number of pages from start to pin | |
2796 | * @gup_flags: flags modifying pin behaviour | |
2797 | * @pages: array that receives pointers to the pages pinned. | |
2798 | * Should be at least nr_pages long. | |
2799 | * | |
2800 | * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See | |
2801 | * get_user_pages_fast() for documentation on the function arguments, because | |
2802 | * the arguments here are identical. | |
2803 | * | |
2804 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
2805 | * see Documentation/vm/pin_user_pages.rst for further details. | |
eddb1c22 JH |
2806 | * |
2807 | * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It | |
2808 | * is NOT intended for Case 2 (RDMA: long-term pins). | |
2809 | */ | |
2810 | int pin_user_pages_fast(unsigned long start, int nr_pages, | |
2811 | unsigned int gup_flags, struct page **pages) | |
2812 | { | |
3faa52c0 JH |
2813 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
2814 | if (WARN_ON_ONCE(gup_flags & FOLL_GET)) | |
2815 | return -EINVAL; | |
2816 | ||
2817 | gup_flags |= FOLL_PIN; | |
2818 | return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages); | |
eddb1c22 JH |
2819 | } |
2820 | EXPORT_SYMBOL_GPL(pin_user_pages_fast); | |
2821 | ||
2822 | /** | |
2823 | * pin_user_pages_remote() - pin pages of a remote process (task != current) | |
2824 | * | |
3faa52c0 JH |
2825 | * @tsk: the task_struct to use for page fault accounting, or |
2826 | * NULL if faults are not to be recorded. | |
2827 | * @mm: mm_struct of target mm | |
2828 | * @start: starting user address | |
2829 | * @nr_pages: number of pages from start to pin | |
2830 | * @gup_flags: flags modifying lookup behaviour | |
2831 | * @pages: array that receives pointers to the pages pinned. | |
2832 | * Should be at least nr_pages long. Or NULL, if caller | |
2833 | * only intends to ensure the pages are faulted in. | |
2834 | * @vmas: array of pointers to vmas corresponding to each page. | |
2835 | * Or NULL if the caller does not require them. | |
2836 | * @locked: pointer to lock flag indicating whether lock is held and | |
2837 | * subsequently whether VM_FAULT_RETRY functionality can be | |
2838 | * utilised. Lock must initially be held. | |
2839 | * | |
2840 | * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See | |
2841 | * get_user_pages_remote() for documentation on the function arguments, because | |
2842 | * the arguments here are identical. | |
2843 | * | |
2844 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
2845 | * see Documentation/vm/pin_user_pages.rst for details. | |
eddb1c22 JH |
2846 | * |
2847 | * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It | |
2848 | * is NOT intended for Case 2 (RDMA: long-term pins). | |
2849 | */ | |
2850 | long pin_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm, | |
2851 | unsigned long start, unsigned long nr_pages, | |
2852 | unsigned int gup_flags, struct page **pages, | |
2853 | struct vm_area_struct **vmas, int *locked) | |
2854 | { | |
3faa52c0 JH |
2855 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
2856 | if (WARN_ON_ONCE(gup_flags & FOLL_GET)) | |
2857 | return -EINVAL; | |
2858 | ||
2859 | gup_flags |= FOLL_PIN; | |
2860 | return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags, | |
2861 | pages, vmas, locked); | |
eddb1c22 JH |
2862 | } |
2863 | EXPORT_SYMBOL(pin_user_pages_remote); | |
2864 | ||
2865 | /** | |
2866 | * pin_user_pages() - pin user pages in memory for use by other devices | |
2867 | * | |
3faa52c0 JH |
2868 | * @start: starting user address |
2869 | * @nr_pages: number of pages from start to pin | |
2870 | * @gup_flags: flags modifying lookup behaviour | |
2871 | * @pages: array that receives pointers to the pages pinned. | |
2872 | * Should be at least nr_pages long. Or NULL, if caller | |
2873 | * only intends to ensure the pages are faulted in. | |
2874 | * @vmas: array of pointers to vmas corresponding to each page. | |
2875 | * Or NULL if the caller does not require them. | |
2876 | * | |
2877 | * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and | |
2878 | * FOLL_PIN is set. | |
2879 | * | |
2880 | * FOLL_PIN means that the pages must be released via unpin_user_page(). Please | |
2881 | * see Documentation/vm/pin_user_pages.rst for details. | |
eddb1c22 JH |
2882 | * |
2883 | * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It | |
2884 | * is NOT intended for Case 2 (RDMA: long-term pins). | |
2885 | */ | |
2886 | long pin_user_pages(unsigned long start, unsigned long nr_pages, | |
2887 | unsigned int gup_flags, struct page **pages, | |
2888 | struct vm_area_struct **vmas) | |
2889 | { | |
3faa52c0 JH |
2890 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
2891 | if (WARN_ON_ONCE(gup_flags & FOLL_GET)) | |
2892 | return -EINVAL; | |
2893 | ||
2894 | gup_flags |= FOLL_PIN; | |
2895 | return __gup_longterm_locked(current, current->mm, start, nr_pages, | |
2896 | pages, vmas, gup_flags); | |
eddb1c22 JH |
2897 | } |
2898 | EXPORT_SYMBOL(pin_user_pages); |