1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/err.h>
5 #include <linux/spinlock.h>
6
7 #include <linux/mm.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
20
21 #include <asm/mmu_context.h>
22 #include <asm/pgtable.h>
23 #include <asm/tlbflush.h>
24
25 #include "internal.h"
26
27 struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
30 };
31
32 /**
33 * put_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
34 * @pages: array of pages to be maybe marked dirty, and definitely released.
35 * @npages: number of pages in the @pages array.
36 * @make_dirty: whether to mark the pages dirty
37 *
38 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
39 * variants called on that page.
40 *
41 * For each page in the @pages array, make that page (or its head page, if a
42 * compound page) dirty, if @make_dirty is true, and if the page was previously
43 * listed as clean. In any case, releases all pages using put_user_page(),
44 * possibly via put_user_pages(), for the non-dirty case.
45 *
46 * Please see the put_user_page() documentation for details.
47 *
48 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
49 * required, then the caller should a) verify that this is really correct,
50 * because _lock() is usually required, and b) hand code it:
51 * set_page_dirty_lock(), put_user_page().
52 *
53 */
54 void put_user_pages_dirty_lock(struct page **pages, unsigned long npages,
55 bool make_dirty)
56 {
57 unsigned long index;
58
59 /*
60 * TODO: this can be optimized for huge pages: if a series of pages is
61 * physically contiguous and part of the same compound page, then a
62 * single operation to the head page should suffice.
63 */
64
65 if (!make_dirty) {
66 put_user_pages(pages, npages);
67 return;
68 }
69
70 for (index = 0; index < npages; index++) {
71 struct page *page = compound_head(pages[index]);
72 /*
73 * Checking PageDirty at this point may race with
74 * clear_page_dirty_for_io(), but that's OK. Two key
75 * cases:
76 *
77 * 1) This code sees the page as already dirty, so it
78 * skips the call to set_page_dirty(). That could happen
79 * because clear_page_dirty_for_io() called
80 * page_mkclean(), followed by set_page_dirty().
81 * However, now the page is going to get written back,
82 * which meets the original intention of setting it
83 * dirty, so all is well: clear_page_dirty_for_io() goes
84 * on to call TestClearPageDirty(), and write the page
85 * back.
86 *
87 * 2) This code sees the page as clean, so it calls
88 * set_page_dirty(). The page stays dirty, despite being
89 * written back, so it gets written back again in the
90 * next writeback cycle. This is harmless.
91 */
92 if (!PageDirty(page))
93 set_page_dirty_lock(page);
94 put_user_page(page);
95 }
96 }
97 EXPORT_SYMBOL(put_user_pages_dirty_lock);
98
99 /**
100 * put_user_pages() - release an array of gup-pinned pages.
101 * @pages: array of pages to be marked dirty and released.
102 * @npages: number of pages in the @pages array.
103 *
104 * For each page in the @pages array, release the page using put_user_page().
105 *
106 * Please see the put_user_page() documentation for details.
107 */
108 void put_user_pages(struct page **pages, unsigned long npages)
109 {
110 unsigned long index;
111
112 /*
113 * TODO: this can be optimized for huge pages: if a series of pages is
114 * physically contiguous and part of the same compound page, then a
115 * single operation to the head page should suffice.
116 */
117 for (index = 0; index < npages; index++)
118 put_user_page(pages[index]);
119 }
120 EXPORT_SYMBOL(put_user_pages);
121
122 #ifdef CONFIG_MMU
123 static struct page *no_page_table(struct vm_area_struct *vma,
124 unsigned int flags)
125 {
126 /*
127 * When core dumping an enormous anonymous area that nobody
128 * has touched so far, we don't want to allocate unnecessary pages or
129 * page tables. Return error instead of NULL to skip handle_mm_fault,
130 * then get_dump_page() will return NULL to leave a hole in the dump.
131 * But we can only make this optimization where a hole would surely
132 * be zero-filled if handle_mm_fault() actually did handle it.
133 */
134 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
135 return ERR_PTR(-EFAULT);
136 return NULL;
137 }
138
139 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
140 pte_t *pte, unsigned int flags)
141 {
142 /* No page to get reference */
143 if (flags & FOLL_GET)
144 return -EFAULT;
145
146 if (flags & FOLL_TOUCH) {
147 pte_t entry = *pte;
148
149 if (flags & FOLL_WRITE)
150 entry = pte_mkdirty(entry);
151 entry = pte_mkyoung(entry);
152
153 if (!pte_same(*pte, entry)) {
154 set_pte_at(vma->vm_mm, address, pte, entry);
155 update_mmu_cache(vma, address, pte);
156 }
157 }
158
159 /* Proper page table entry exists, but no corresponding struct page */
160 return -EEXIST;
161 }
162
163 /*
164 * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
165 * but only after we've gone through a COW cycle and they are dirty.
166 */
167 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
168 {
169 return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
170 }
171
172 /*
173 * A (separate) COW fault might break the page the other way and
174 * get_user_pages() would return the page from what is now the wrong
175 * VM. So we need to force a COW break at GUP time even for reads.
176 */
177 static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
178 {
179 return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
180 }
181
182 static struct page *follow_page_pte(struct vm_area_struct *vma,
183 unsigned long address, pmd_t *pmd, unsigned int flags,
184 struct dev_pagemap **pgmap)
185 {
186 struct mm_struct *mm = vma->vm_mm;
187 struct page *page;
188 spinlock_t *ptl;
189 pte_t *ptep, pte;
190
191 retry:
192 if (unlikely(pmd_bad(*pmd)))
193 return no_page_table(vma, flags);
194
195 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
196 pte = *ptep;
197 if (!pte_present(pte)) {
198 swp_entry_t entry;
199 /*
200 * KSM's break_ksm() relies upon recognizing a ksm page
201 * even while it is being migrated, so for that case we
202 * need migration_entry_wait().
203 */
204 if (likely(!(flags & FOLL_MIGRATION)))
205 goto no_page;
206 if (pte_none(pte))
207 goto no_page;
208 entry = pte_to_swp_entry(pte);
209 if (!is_migration_entry(entry))
210 goto no_page;
211 pte_unmap_unlock(ptep, ptl);
212 migration_entry_wait(mm, pmd, address);
213 goto retry;
214 }
215 if ((flags & FOLL_NUMA) && pte_protnone(pte))
216 goto no_page;
217 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
218 pte_unmap_unlock(ptep, ptl);
219 return NULL;
220 }
221
222 page = vm_normal_page(vma, address, pte);
223 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
224 /*
225 * Only return device mapping pages in the FOLL_GET case since
226 * they are only valid while holding the pgmap reference.
227 */
228 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
229 if (*pgmap)
230 page = pte_page(pte);
231 else
232 goto no_page;
233 } else if (unlikely(!page)) {
234 if (flags & FOLL_DUMP) {
235 /* Avoid special (like zero) pages in core dumps */
236 page = ERR_PTR(-EFAULT);
237 goto out;
238 }
239
240 if (is_zero_pfn(pte_pfn(pte))) {
241 page = pte_page(pte);
242 } else {
243 int ret;
244
245 ret = follow_pfn_pte(vma, address, ptep, flags);
246 page = ERR_PTR(ret);
247 goto out;
248 }
249 }
250
251 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
252 int ret;
253 get_page(page);
254 pte_unmap_unlock(ptep, ptl);
255 lock_page(page);
256 ret = split_huge_page(page);
257 unlock_page(page);
258 put_page(page);
259 if (ret)
260 return ERR_PTR(ret);
261 goto retry;
262 }
263
264 if (flags & FOLL_GET) {
265 if (unlikely(!try_get_page(page))) {
266 page = ERR_PTR(-ENOMEM);
267 goto out;
268 }
269 }
270 if (flags & FOLL_TOUCH) {
271 if ((flags & FOLL_WRITE) &&
272 !pte_dirty(pte) && !PageDirty(page))
273 set_page_dirty(page);
274 /*
275 * pte_mkyoung() would be more correct here, but atomic care
276 * is needed to avoid losing the dirty bit: it is easier to use
277 * mark_page_accessed().
278 */
279 mark_page_accessed(page);
280 }
281 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
282 /* Do not mlock pte-mapped THP */
283 if (PageTransCompound(page))
284 goto out;
285
286 /*
287 * The preliminary mapping check is mainly to avoid the
288 * pointless overhead of lock_page on the ZERO_PAGE
289 * which might bounce very badly if there is contention.
290 *
291 * If the page is already locked, we don't need to
292 * handle it now - vmscan will handle it later if and
293 * when it attempts to reclaim the page.
294 */
295 if (page->mapping && trylock_page(page)) {
296 lru_add_drain(); /* push cached pages to LRU */
297 /*
298 * Because we lock page here, and migration is
299 * blocked by the pte's page reference, and we
300 * know the page is still mapped, we don't even
301 * need to check for file-cache page truncation.
302 */
303 mlock_vma_page(page);
304 unlock_page(page);
305 }
306 }
307 out:
308 pte_unmap_unlock(ptep, ptl);
309 return page;
310 no_page:
311 pte_unmap_unlock(ptep, ptl);
312 if (!pte_none(pte))
313 return NULL;
314 return no_page_table(vma, flags);
315 }
316
317 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
318 unsigned long address, pud_t *pudp,
319 unsigned int flags,
320 struct follow_page_context *ctx)
321 {
322 pmd_t *pmd, pmdval;
323 spinlock_t *ptl;
324 struct page *page;
325 struct mm_struct *mm = vma->vm_mm;
326
327 pmd = pmd_offset(pudp, address);
328 /*
329 * The READ_ONCE() will stabilize the pmdval in a register or
330 * on the stack so that it will stop changing under the code.
331 */
332 pmdval = READ_ONCE(*pmd);
333 if (pmd_none(pmdval))
334 return no_page_table(vma, flags);
335 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
336 page = follow_huge_pmd(mm, address, pmd, flags);
337 if (page)
338 return page;
339 return no_page_table(vma, flags);
340 }
341 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
342 page = follow_huge_pd(vma, address,
343 __hugepd(pmd_val(pmdval)), flags,
344 PMD_SHIFT);
345 if (page)
346 return page;
347 return no_page_table(vma, flags);
348 }
349 retry:
350 if (!pmd_present(pmdval)) {
351 if (likely(!(flags & FOLL_MIGRATION)))
352 return no_page_table(vma, flags);
353 VM_BUG_ON(thp_migration_supported() &&
354 !is_pmd_migration_entry(pmdval));
355 if (is_pmd_migration_entry(pmdval))
356 pmd_migration_entry_wait(mm, pmd);
357 pmdval = READ_ONCE(*pmd);
358 /*
359 * MADV_DONTNEED may convert the pmd to null because
360 * mmap_sem is held in read mode
361 */
362 if (pmd_none(pmdval))
363 return no_page_table(vma, flags);
364 goto retry;
365 }
366 if (pmd_devmap(pmdval)) {
367 ptl = pmd_lock(mm, pmd);
368 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
369 spin_unlock(ptl);
370 if (page)
371 return page;
372 }
373 if (likely(!pmd_trans_huge(pmdval)))
374 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
375
376 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
377 return no_page_table(vma, flags);
378
379 retry_locked:
380 ptl = pmd_lock(mm, pmd);
381 if (unlikely(pmd_none(*pmd))) {
382 spin_unlock(ptl);
383 return no_page_table(vma, flags);
384 }
385 if (unlikely(!pmd_present(*pmd))) {
386 spin_unlock(ptl);
387 if (likely(!(flags & FOLL_MIGRATION)))
388 return no_page_table(vma, flags);
389 pmd_migration_entry_wait(mm, pmd);
390 goto retry_locked;
391 }
392 if (unlikely(!pmd_trans_huge(*pmd))) {
393 spin_unlock(ptl);
394 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
395 }
396 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
397 int ret;
398 page = pmd_page(*pmd);
399 if (is_huge_zero_page(page)) {
400 spin_unlock(ptl);
401 ret = 0;
402 split_huge_pmd(vma, pmd, address);
403 if (pmd_trans_unstable(pmd))
404 ret = -EBUSY;
405 } else if (flags & FOLL_SPLIT) {
406 if (unlikely(!try_get_page(page))) {
407 spin_unlock(ptl);
408 return ERR_PTR(-ENOMEM);
409 }
410 spin_unlock(ptl);
411 lock_page(page);
412 ret = split_huge_page(page);
413 unlock_page(page);
414 put_page(page);
415 if (pmd_none(*pmd))
416 return no_page_table(vma, flags);
417 } else { /* flags & FOLL_SPLIT_PMD */
418 spin_unlock(ptl);
419 split_huge_pmd(vma, pmd, address);
420 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
421 }
422
423 return ret ? ERR_PTR(ret) :
424 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
425 }
426 page = follow_trans_huge_pmd(vma, address, pmd, flags);
427 spin_unlock(ptl);
428 ctx->page_mask = HPAGE_PMD_NR - 1;
429 return page;
430 }
431
432 static struct page *follow_pud_mask(struct vm_area_struct *vma,
433 unsigned long address, p4d_t *p4dp,
434 unsigned int flags,
435 struct follow_page_context *ctx)
436 {
437 pud_t *pud;
438 spinlock_t *ptl;
439 struct page *page;
440 struct mm_struct *mm = vma->vm_mm;
441
442 pud = pud_offset(p4dp, address);
443 if (pud_none(*pud))
444 return no_page_table(vma, flags);
445 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
446 page = follow_huge_pud(mm, address, pud, flags);
447 if (page)
448 return page;
449 return no_page_table(vma, flags);
450 }
451 if (is_hugepd(__hugepd(pud_val(*pud)))) {
452 page = follow_huge_pd(vma, address,
453 __hugepd(pud_val(*pud)), flags,
454 PUD_SHIFT);
455 if (page)
456 return page;
457 return no_page_table(vma, flags);
458 }
459 if (pud_devmap(*pud)) {
460 ptl = pud_lock(mm, pud);
461 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
462 spin_unlock(ptl);
463 if (page)
464 return page;
465 }
466 if (unlikely(pud_bad(*pud)))
467 return no_page_table(vma, flags);
468
469 return follow_pmd_mask(vma, address, pud, flags, ctx);
470 }
471
472 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
473 unsigned long address, pgd_t *pgdp,
474 unsigned int flags,
475 struct follow_page_context *ctx)
476 {
477 p4d_t *p4d;
478 struct page *page;
479
480 p4d = p4d_offset(pgdp, address);
481 if (p4d_none(*p4d))
482 return no_page_table(vma, flags);
483 BUILD_BUG_ON(p4d_huge(*p4d));
484 if (unlikely(p4d_bad(*p4d)))
485 return no_page_table(vma, flags);
486
487 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
488 page = follow_huge_pd(vma, address,
489 __hugepd(p4d_val(*p4d)), flags,
490 P4D_SHIFT);
491 if (page)
492 return page;
493 return no_page_table(vma, flags);
494 }
495 return follow_pud_mask(vma, address, p4d, flags, ctx);
496 }
497
498 /**
499 * follow_page_mask - look up a page descriptor from a user-virtual address
500 * @vma: vm_area_struct mapping @address
501 * @address: virtual address to look up
502 * @flags: flags modifying lookup behaviour
503 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
504 * pointer to output page_mask
505 *
506 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
507 *
508 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
509 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
510 *
511 * On output, the @ctx->page_mask is set according to the size of the page.
512 *
513 * Return: the mapped (struct page *), %NULL if no mapping exists, or
514 * an error pointer if there is a mapping to something not represented
515 * by a page descriptor (see also vm_normal_page()).
516 */
517 static struct page *follow_page_mask(struct vm_area_struct *vma,
518 unsigned long address, unsigned int flags,
519 struct follow_page_context *ctx)
520 {
521 pgd_t *pgd;
522 struct page *page;
523 struct mm_struct *mm = vma->vm_mm;
524
525 ctx->page_mask = 0;
526
527 /* make this handle hugepd */
528 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
529 if (!IS_ERR(page)) {
530 BUG_ON(flags & FOLL_GET);
531 return page;
532 }
533
534 pgd = pgd_offset(mm, address);
535
536 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
537 return no_page_table(vma, flags);
538
539 if (pgd_huge(*pgd)) {
540 page = follow_huge_pgd(mm, address, pgd, flags);
541 if (page)
542 return page;
543 return no_page_table(vma, flags);
544 }
545 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
546 page = follow_huge_pd(vma, address,
547 __hugepd(pgd_val(*pgd)), flags,
548 PGDIR_SHIFT);
549 if (page)
550 return page;
551 return no_page_table(vma, flags);
552 }
553
554 return follow_p4d_mask(vma, address, pgd, flags, ctx);
555 }
556
557 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
558 unsigned int foll_flags)
559 {
560 struct follow_page_context ctx = { NULL };
561 struct page *page;
562
563 page = follow_page_mask(vma, address, foll_flags, &ctx);
564 if (ctx.pgmap)
565 put_dev_pagemap(ctx.pgmap);
566 return page;
567 }
568
569 static int get_gate_page(struct mm_struct *mm, unsigned long address,
570 unsigned int gup_flags, struct vm_area_struct **vma,
571 struct page **page)
572 {
573 pgd_t *pgd;
574 p4d_t *p4d;
575 pud_t *pud;
576 pmd_t *pmd;
577 pte_t *pte;
578 int ret = -EFAULT;
579
580 /* user gate pages are read-only */
581 if (gup_flags & FOLL_WRITE)
582 return -EFAULT;
583 if (address > TASK_SIZE)
584 pgd = pgd_offset_k(address);
585 else
586 pgd = pgd_offset_gate(mm, address);
587 if (pgd_none(*pgd))
588 return -EFAULT;
589 p4d = p4d_offset(pgd, address);
590 if (p4d_none(*p4d))
591 return -EFAULT;
592 pud = pud_offset(p4d, address);
593 if (pud_none(*pud))
594 return -EFAULT;
595 pmd = pmd_offset(pud, address);
596 if (!pmd_present(*pmd))
597 return -EFAULT;
598 VM_BUG_ON(pmd_trans_huge(*pmd));
599 pte = pte_offset_map(pmd, address);
600 if (pte_none(*pte))
601 goto unmap;
602 *vma = get_gate_vma(mm);
603 if (!page)
604 goto out;
605 *page = vm_normal_page(*vma, address, *pte);
606 if (!*page) {
607 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
608 goto unmap;
609 *page = pte_page(*pte);
610 }
611 if (unlikely(!try_get_page(*page))) {
612 ret = -ENOMEM;
613 goto unmap;
614 }
615 out:
616 ret = 0;
617 unmap:
618 pte_unmap(pte);
619 return ret;
620 }
621
622 /*
623 * mmap_sem must be held on entry. If @nonblocking != NULL and
624 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
625 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
626 */
627 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
628 unsigned long address, unsigned int *flags, int *nonblocking)
629 {
630 unsigned int fault_flags = 0;
631 vm_fault_t ret;
632
633 /* mlock all present pages, but do not fault in new pages */
634 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
635 return -ENOENT;
636 if (*flags & FOLL_WRITE)
637 fault_flags |= FAULT_FLAG_WRITE;
638 if (*flags & FOLL_REMOTE)
639 fault_flags |= FAULT_FLAG_REMOTE;
640 if (nonblocking)
641 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
642 if (*flags & FOLL_NOWAIT)
643 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
644 if (*flags & FOLL_TRIED) {
645 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
646 fault_flags |= FAULT_FLAG_TRIED;
647 }
648
649 ret = handle_mm_fault(vma, address, fault_flags);
650 if (ret & VM_FAULT_ERROR) {
651 int err = vm_fault_to_errno(ret, *flags);
652
653 if (err)
654 return err;
655 BUG();
656 }
657
658 if (tsk) {
659 if (ret & VM_FAULT_MAJOR)
660 tsk->maj_flt++;
661 else
662 tsk->min_flt++;
663 }
664
665 if (ret & VM_FAULT_RETRY) {
666 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
667 *nonblocking = 0;
668 return -EBUSY;
669 }
670
671 /*
672 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
673 * necessary, even if maybe_mkwrite decided not to set pte_write. We
674 * can thus safely do subsequent page lookups as if they were reads.
675 * But only do so when looping for pte_write is futile: in some cases
676 * userspace may also be wanting to write to the gotten user page,
677 * which a read fault here might prevent (a readonly page might get
678 * reCOWed by userspace write).
679 */
680 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
681 *flags |= FOLL_COW;
682 return 0;
683 }
684
685 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
686 {
687 vm_flags_t vm_flags = vma->vm_flags;
688 int write = (gup_flags & FOLL_WRITE);
689 int foreign = (gup_flags & FOLL_REMOTE);
690
691 if (vm_flags & (VM_IO | VM_PFNMAP))
692 return -EFAULT;
693
694 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
695 return -EFAULT;
696
697 if (write) {
698 if (!(vm_flags & VM_WRITE)) {
699 if (!(gup_flags & FOLL_FORCE))
700 return -EFAULT;
701 /*
702 * We used to let the write,force case do COW in a
703 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
704 * set a breakpoint in a read-only mapping of an
705 * executable, without corrupting the file (yet only
706 * when that file had been opened for writing!).
707 * Anon pages in shared mappings are surprising: now
708 * just reject it.
709 */
710 if (!is_cow_mapping(vm_flags))
711 return -EFAULT;
712 }
713 } else if (!(vm_flags & VM_READ)) {
714 if (!(gup_flags & FOLL_FORCE))
715 return -EFAULT;
716 /*
717 * Is there actually any vma we can reach here which does not
718 * have VM_MAYREAD set?
719 */
720 if (!(vm_flags & VM_MAYREAD))
721 return -EFAULT;
722 }
723 /*
724 * gups are always data accesses, not instruction
725 * fetches, so execute=false here
726 */
727 if (!arch_vma_access_permitted(vma, write, false, foreign))
728 return -EFAULT;
729 return 0;
730 }
731
732 /**
733 * __get_user_pages() - pin user pages in memory
734 * @tsk: task_struct of target task
735 * @mm: mm_struct of target mm
736 * @start: starting user address
737 * @nr_pages: number of pages from start to pin
738 * @gup_flags: flags modifying pin behaviour
739 * @pages: array that receives pointers to the pages pinned.
740 * Should be at least nr_pages long. Or NULL, if caller
741 * only intends to ensure the pages are faulted in.
742 * @vmas: array of pointers to vmas corresponding to each page.
743 * Or NULL if the caller does not require them.
744 * @nonblocking: whether waiting for disk IO or mmap_sem contention
745 *
746 * Returns number of pages pinned. This may be fewer than the number
747 * requested. If nr_pages is 0 or negative, returns 0. If no pages
748 * were pinned, returns -errno. Each page returned must be released
749 * with a put_page() call when it is finished with. vmas will only
750 * remain valid while mmap_sem is held.
751 *
752 * Must be called with mmap_sem held. It may be released. See below.
753 *
754 * __get_user_pages walks a process's page tables and takes a reference to
755 * each struct page that each user address corresponds to at a given
756 * instant. That is, it takes the page that would be accessed if a user
757 * thread accesses the given user virtual address at that instant.
758 *
759 * This does not guarantee that the page exists in the user mappings when
760 * __get_user_pages returns, and there may even be a completely different
761 * page there in some cases (eg. if mmapped pagecache has been invalidated
762 * and subsequently re faulted). However it does guarantee that the page
763 * won't be freed completely. And mostly callers simply care that the page
764 * contains data that was valid *at some point in time*. Typically, an IO
765 * or similar operation cannot guarantee anything stronger anyway because
766 * locks can't be held over the syscall boundary.
767 *
768 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
769 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
770 * appropriate) must be called after the page is finished with, and
771 * before put_page is called.
772 *
773 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
774 * or mmap_sem contention, and if waiting is needed to pin all pages,
775 * *@nonblocking will be set to 0. Further, if @gup_flags does not
776 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
777 * this case.
778 *
779 * A caller using such a combination of @nonblocking and @gup_flags
780 * must therefore hold the mmap_sem for reading only, and recognize
781 * when it's been released. Otherwise, it must be held for either
782 * reading or writing and will not be released.
783 *
784 * In most cases, get_user_pages or get_user_pages_fast should be used
785 * instead of __get_user_pages. __get_user_pages should be used only if
786 * you need some special @gup_flags.
787 */
788 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
789 unsigned long start, unsigned long nr_pages,
790 unsigned int gup_flags, struct page **pages,
791 struct vm_area_struct **vmas, int *nonblocking)
792 {
793 long ret = 0, i = 0;
794 struct vm_area_struct *vma = NULL;
795 struct follow_page_context ctx = { NULL };
796
797 if (!nr_pages)
798 return 0;
799
800 start = untagged_addr(start);
801
802 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
803
804 /*
805 * If FOLL_FORCE is set then do not force a full fault as the hinting
806 * fault information is unrelated to the reference behaviour of a task
807 * using the address space
808 */
809 if (!(gup_flags & FOLL_FORCE))
810 gup_flags |= FOLL_NUMA;
811
812 do {
813 struct page *page;
814 unsigned int foll_flags = gup_flags;
815 unsigned int page_increm;
816
817 /* first iteration or cross vma bound */
818 if (!vma || start >= vma->vm_end) {
819 vma = find_extend_vma(mm, start);
820 if (!vma && in_gate_area(mm, start)) {
821 ret = get_gate_page(mm, start & PAGE_MASK,
822 gup_flags, &vma,
823 pages ? &pages[i] : NULL);
824 if (ret)
825 goto out;
826 ctx.page_mask = 0;
827 goto next_page;
828 }
829
830 if (!vma || check_vma_flags(vma, gup_flags)) {
831 ret = -EFAULT;
832 goto out;
833 }
834 if (is_vm_hugetlb_page(vma)) {
835 if (should_force_cow_break(vma, foll_flags))
836 foll_flags |= FOLL_WRITE;
837 i = follow_hugetlb_page(mm, vma, pages, vmas,
838 &start, &nr_pages, i,
839 foll_flags, nonblocking);
840 continue;
841 }
842 }
843
844 if (should_force_cow_break(vma, foll_flags))
845 foll_flags |= FOLL_WRITE;
846
847 retry:
848 /*
849 * If we have a pending SIGKILL, don't keep faulting pages and
850 * potentially allocating memory.
851 */
852 if (fatal_signal_pending(current)) {
853 ret = -ERESTARTSYS;
854 goto out;
855 }
856 cond_resched();
857
858 page = follow_page_mask(vma, start, foll_flags, &ctx);
859 if (!page) {
860 ret = faultin_page(tsk, vma, start, &foll_flags,
861 nonblocking);
862 switch (ret) {
863 case 0:
864 goto retry;
865 case -EBUSY:
866 ret = 0;
867 /* FALLTHRU */
868 case -EFAULT:
869 case -ENOMEM:
870 case -EHWPOISON:
871 goto out;
872 case -ENOENT:
873 goto next_page;
874 }
875 BUG();
876 } else if (PTR_ERR(page) == -EEXIST) {
877 /*
878 * Proper page table entry exists, but no corresponding
879 * struct page.
880 */
881 goto next_page;
882 } else if (IS_ERR(page)) {
883 ret = PTR_ERR(page);
884 goto out;
885 }
886 if (pages) {
887 pages[i] = page;
888 flush_anon_page(vma, page, start);
889 flush_dcache_page(page);
890 ctx.page_mask = 0;
891 }
892 next_page:
893 if (vmas) {
894 vmas[i] = vma;
895 ctx.page_mask = 0;
896 }
897 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
898 if (page_increm > nr_pages)
899 page_increm = nr_pages;
900 i += page_increm;
901 start += page_increm * PAGE_SIZE;
902 nr_pages -= page_increm;
903 } while (nr_pages);
904 out:
905 if (ctx.pgmap)
906 put_dev_pagemap(ctx.pgmap);
907 return i ? i : ret;
908 }
909
910 static bool vma_permits_fault(struct vm_area_struct *vma,
911 unsigned int fault_flags)
912 {
913 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
914 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
915 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
916
917 if (!(vm_flags & vma->vm_flags))
918 return false;
919
920 /*
921 * The architecture might have a hardware protection
922 * mechanism other than read/write that can deny access.
923 *
924 * gup always represents data access, not instruction
925 * fetches, so execute=false here:
926 */
927 if (!arch_vma_access_permitted(vma, write, false, foreign))
928 return false;
929
930 return true;
931 }
932
933 /*
934 * fixup_user_fault() - manually resolve a user page fault
935 * @tsk: the task_struct to use for page fault accounting, or
936 * NULL if faults are not to be recorded.
937 * @mm: mm_struct of target mm
938 * @address: user address
939 * @fault_flags:flags to pass down to handle_mm_fault()
940 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
941 * does not allow retry
942 *
943 * This is meant to be called in the specific scenario where for locking reasons
944 * we try to access user memory in atomic context (within a pagefault_disable()
945 * section), this returns -EFAULT, and we want to resolve the user fault before
946 * trying again.
947 *
948 * Typically this is meant to be used by the futex code.
949 *
950 * The main difference with get_user_pages() is that this function will
951 * unconditionally call handle_mm_fault() which will in turn perform all the
952 * necessary SW fixup of the dirty and young bits in the PTE, while
953 * get_user_pages() only guarantees to update these in the struct page.
954 *
955 * This is important for some architectures where those bits also gate the
956 * access permission to the page because they are maintained in software. On
957 * such architectures, gup() will not be enough to make a subsequent access
958 * succeed.
959 *
960 * This function will not return with an unlocked mmap_sem. So it has not the
961 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
962 */
963 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
964 unsigned long address, unsigned int fault_flags,
965 bool *unlocked)
966 {
967 struct vm_area_struct *vma;
968 vm_fault_t ret, major = 0;
969
970 address = untagged_addr(address);
971
972 if (unlocked)
973 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
974
975 retry:
976 vma = find_extend_vma(mm, address);
977 if (!vma || address < vma->vm_start)
978 return -EFAULT;
979
980 if (!vma_permits_fault(vma, fault_flags))
981 return -EFAULT;
982
983 ret = handle_mm_fault(vma, address, fault_flags);
984 major |= ret & VM_FAULT_MAJOR;
985 if (ret & VM_FAULT_ERROR) {
986 int err = vm_fault_to_errno(ret, 0);
987
988 if (err)
989 return err;
990 BUG();
991 }
992
993 if (ret & VM_FAULT_RETRY) {
994 down_read(&mm->mmap_sem);
995 if (!(fault_flags & FAULT_FLAG_TRIED)) {
996 *unlocked = true;
997 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
998 fault_flags |= FAULT_FLAG_TRIED;
999 goto retry;
1000 }
1001 }
1002
1003 if (tsk) {
1004 if (major)
1005 tsk->maj_flt++;
1006 else
1007 tsk->min_flt++;
1008 }
1009 return 0;
1010 }
1011 EXPORT_SYMBOL_GPL(fixup_user_fault);
1012
1013 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
1014 struct mm_struct *mm,
1015 unsigned long start,
1016 unsigned long nr_pages,
1017 struct page **pages,
1018 struct vm_area_struct **vmas,
1019 int *locked,
1020 unsigned int flags)
1021 {
1022 long ret, pages_done;
1023 bool lock_dropped;
1024
1025 if (locked) {
1026 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1027 BUG_ON(vmas);
1028 /* check caller initialized locked */
1029 BUG_ON(*locked != 1);
1030 }
1031
1032 if (pages)
1033 flags |= FOLL_GET;
1034
1035 pages_done = 0;
1036 lock_dropped = false;
1037 for (;;) {
1038 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1039 vmas, locked);
1040 if (!locked)
1041 /* VM_FAULT_RETRY couldn't trigger, bypass */
1042 return ret;
1043
1044 /* VM_FAULT_RETRY cannot return errors */
1045 if (!*locked) {
1046 BUG_ON(ret < 0);
1047 BUG_ON(ret >= nr_pages);
1048 }
1049
1050 if (ret > 0) {
1051 nr_pages -= ret;
1052 pages_done += ret;
1053 if (!nr_pages)
1054 break;
1055 }
1056 if (*locked) {
1057 /*
1058 * VM_FAULT_RETRY didn't trigger or it was a
1059 * FOLL_NOWAIT.
1060 */
1061 if (!pages_done)
1062 pages_done = ret;
1063 break;
1064 }
1065 /*
1066 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1067 * For the prefault case (!pages) we only update counts.
1068 */
1069 if (likely(pages))
1070 pages += ret;
1071 start += ret << PAGE_SHIFT;
1072
1073 /*
1074 * Repeat on the address that fired VM_FAULT_RETRY
1075 * without FAULT_FLAG_ALLOW_RETRY but with
1076 * FAULT_FLAG_TRIED.
1077 */
1078 *locked = 1;
1079 lock_dropped = true;
1080 down_read(&mm->mmap_sem);
1081 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1082 pages, NULL, NULL);
1083 if (ret != 1) {
1084 BUG_ON(ret > 1);
1085 if (!pages_done)
1086 pages_done = ret;
1087 break;
1088 }
1089 nr_pages--;
1090 pages_done++;
1091 if (!nr_pages)
1092 break;
1093 if (likely(pages))
1094 pages++;
1095 start += PAGE_SIZE;
1096 }
1097 if (lock_dropped && *locked) {
1098 /*
1099 * We must let the caller know we temporarily dropped the lock
1100 * and so the critical section protected by it was lost.
1101 */
1102 up_read(&mm->mmap_sem);
1103 *locked = 0;
1104 }
1105 return pages_done;
1106 }
1107
1108 /*
1109 * get_user_pages_remote() - pin user pages in memory
1110 * @tsk: the task_struct to use for page fault accounting, or
1111 * NULL if faults are not to be recorded.
1112 * @mm: mm_struct of target mm
1113 * @start: starting user address
1114 * @nr_pages: number of pages from start to pin
1115 * @gup_flags: flags modifying lookup behaviour
1116 * @pages: array that receives pointers to the pages pinned.
1117 * Should be at least nr_pages long. Or NULL, if caller
1118 * only intends to ensure the pages are faulted in.
1119 * @vmas: array of pointers to vmas corresponding to each page.
1120 * Or NULL if the caller does not require them.
1121 * @locked: pointer to lock flag indicating whether lock is held and
1122 * subsequently whether VM_FAULT_RETRY functionality can be
1123 * utilised. Lock must initially be held.
1124 *
1125 * Returns number of pages pinned. This may be fewer than the number
1126 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1127 * were pinned, returns -errno. Each page returned must be released
1128 * with a put_page() call when it is finished with. vmas will only
1129 * remain valid while mmap_sem is held.
1130 *
1131 * Must be called with mmap_sem held for read or write.
1132 *
1133 * get_user_pages walks a process's page tables and takes a reference to
1134 * each struct page that each user address corresponds to at a given
1135 * instant. That is, it takes the page that would be accessed if a user
1136 * thread accesses the given user virtual address at that instant.
1137 *
1138 * This does not guarantee that the page exists in the user mappings when
1139 * get_user_pages returns, and there may even be a completely different
1140 * page there in some cases (eg. if mmapped pagecache has been invalidated
1141 * and subsequently re faulted). However it does guarantee that the page
1142 * won't be freed completely. And mostly callers simply care that the page
1143 * contains data that was valid *at some point in time*. Typically, an IO
1144 * or similar operation cannot guarantee anything stronger anyway because
1145 * locks can't be held over the syscall boundary.
1146 *
1147 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1148 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1149 * be called after the page is finished with, and before put_page is called.
1150 *
1151 * get_user_pages is typically used for fewer-copy IO operations, to get a
1152 * handle on the memory by some means other than accesses via the user virtual
1153 * addresses. The pages may be submitted for DMA to devices or accessed via
1154 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1155 * use the correct cache flushing APIs.
1156 *
1157 * See also get_user_pages_fast, for performance critical applications.
1158 *
1159 * get_user_pages should be phased out in favor of
1160 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1161 * should use get_user_pages because it cannot pass
1162 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1163 */
1164 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1165 unsigned long start, unsigned long nr_pages,
1166 unsigned int gup_flags, struct page **pages,
1167 struct vm_area_struct **vmas, int *locked)
1168 {
1169 /*
1170 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1171 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1172 * vmas. As there are no users of this flag in this call we simply
1173 * disallow this option for now.
1174 */
1175 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1176 return -EINVAL;
1177
1178 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1179 locked,
1180 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1181 }
1182 EXPORT_SYMBOL(get_user_pages_remote);
1183
1184 /**
1185 * populate_vma_page_range() - populate a range of pages in the vma.
1186 * @vma: target vma
1187 * @start: start address
1188 * @end: end address
1189 * @nonblocking:
1190 *
1191 * This takes care of mlocking the pages too if VM_LOCKED is set.
1192 *
1193 * return 0 on success, negative error code on error.
1194 *
1195 * vma->vm_mm->mmap_sem must be held.
1196 *
1197 * If @nonblocking is NULL, it may be held for read or write and will
1198 * be unperturbed.
1199 *
1200 * If @nonblocking is non-NULL, it must held for read only and may be
1201 * released. If it's released, *@nonblocking will be set to 0.
1202 */
1203 long populate_vma_page_range(struct vm_area_struct *vma,
1204 unsigned long start, unsigned long end, int *nonblocking)
1205 {
1206 struct mm_struct *mm = vma->vm_mm;
1207 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1208 int gup_flags;
1209
1210 VM_BUG_ON(start & ~PAGE_MASK);
1211 VM_BUG_ON(end & ~PAGE_MASK);
1212 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1213 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1214 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1215
1216 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1217 if (vma->vm_flags & VM_LOCKONFAULT)
1218 gup_flags &= ~FOLL_POPULATE;
1219 /*
1220 * We want to touch writable mappings with a write fault in order
1221 * to break COW, except for shared mappings because these don't COW
1222 * and we would not want to dirty them for nothing.
1223 */
1224 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1225 gup_flags |= FOLL_WRITE;
1226
1227 /*
1228 * We want mlock to succeed for regions that have any permissions
1229 * other than PROT_NONE.
1230 */
1231 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1232 gup_flags |= FOLL_FORCE;
1233
1234 /*
1235 * We made sure addr is within a VMA, so the following will
1236 * not result in a stack expansion that recurses back here.
1237 */
1238 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1239 NULL, NULL, nonblocking);
1240 }
1241
1242 /*
1243 * __mm_populate - populate and/or mlock pages within a range of address space.
1244 *
1245 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1246 * flags. VMAs must be already marked with the desired vm_flags, and
1247 * mmap_sem must not be held.
1248 */
1249 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1250 {
1251 struct mm_struct *mm = current->mm;
1252 unsigned long end, nstart, nend;
1253 struct vm_area_struct *vma = NULL;
1254 int locked = 0;
1255 long ret = 0;
1256
1257 end = start + len;
1258
1259 for (nstart = start; nstart < end; nstart = nend) {
1260 /*
1261 * We want to fault in pages for [nstart; end) address range.
1262 * Find first corresponding VMA.
1263 */
1264 if (!locked) {
1265 locked = 1;
1266 down_read(&mm->mmap_sem);
1267 vma = find_vma(mm, nstart);
1268 } else if (nstart >= vma->vm_end)
1269 vma = vma->vm_next;
1270 if (!vma || vma->vm_start >= end)
1271 break;
1272 /*
1273 * Set [nstart; nend) to intersection of desired address
1274 * range with the first VMA. Also, skip undesirable VMA types.
1275 */
1276 nend = min(end, vma->vm_end);
1277 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1278 continue;
1279 if (nstart < vma->vm_start)
1280 nstart = vma->vm_start;
1281 /*
1282 * Now fault in a range of pages. populate_vma_page_range()
1283 * double checks the vma flags, so that it won't mlock pages
1284 * if the vma was already munlocked.
1285 */
1286 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1287 if (ret < 0) {
1288 if (ignore_errors) {
1289 ret = 0;
1290 continue; /* continue at next VMA */
1291 }
1292 break;
1293 }
1294 nend = nstart + ret * PAGE_SIZE;
1295 ret = 0;
1296 }
1297 if (locked)
1298 up_read(&mm->mmap_sem);
1299 return ret; /* 0 or negative error code */
1300 }
1301
1302 /**
1303 * get_dump_page() - pin user page in memory while writing it to core dump
1304 * @addr: user address
1305 *
1306 * Returns struct page pointer of user page pinned for dump,
1307 * to be freed afterwards by put_page().
1308 *
1309 * Returns NULL on any kind of failure - a hole must then be inserted into
1310 * the corefile, to preserve alignment with its headers; and also returns
1311 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1312 * allowing a hole to be left in the corefile to save diskspace.
1313 *
1314 * Called without mmap_sem, but after all other threads have been killed.
1315 */
1316 #ifdef CONFIG_ELF_CORE
1317 struct page *get_dump_page(unsigned long addr)
1318 {
1319 struct vm_area_struct *vma;
1320 struct page *page;
1321
1322 if (__get_user_pages(current, current->mm, addr, 1,
1323 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1324 NULL) < 1)
1325 return NULL;
1326 flush_cache_page(vma, addr, page_to_pfn(page));
1327 return page;
1328 }
1329 #endif /* CONFIG_ELF_CORE */
1330 #else /* CONFIG_MMU */
1331 static long __get_user_pages_locked(struct task_struct *tsk,
1332 struct mm_struct *mm, unsigned long start,
1333 unsigned long nr_pages, struct page **pages,
1334 struct vm_area_struct **vmas, int *locked,
1335 unsigned int foll_flags)
1336 {
1337 struct vm_area_struct *vma;
1338 unsigned long vm_flags;
1339 int i;
1340
1341 /* calculate required read or write permissions.
1342 * If FOLL_FORCE is set, we only require the "MAY" flags.
1343 */
1344 vm_flags = (foll_flags & FOLL_WRITE) ?
1345 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1346 vm_flags &= (foll_flags & FOLL_FORCE) ?
1347 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1348
1349 for (i = 0; i < nr_pages; i++) {
1350 vma = find_vma(mm, start);
1351 if (!vma)
1352 goto finish_or_fault;
1353
1354 /* protect what we can, including chardevs */
1355 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1356 !(vm_flags & vma->vm_flags))
1357 goto finish_or_fault;
1358
1359 if (pages) {
1360 pages[i] = virt_to_page(start);
1361 if (pages[i])
1362 get_page(pages[i]);
1363 }
1364 if (vmas)
1365 vmas[i] = vma;
1366 start = (start + PAGE_SIZE) & PAGE_MASK;
1367 }
1368
1369 return i;
1370
1371 finish_or_fault:
1372 return i ? : -EFAULT;
1373 }
1374 #endif /* !CONFIG_MMU */
1375
1376 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1377 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1378 {
1379 long i;
1380 struct vm_area_struct *vma_prev = NULL;
1381
1382 for (i = 0; i < nr_pages; i++) {
1383 struct vm_area_struct *vma = vmas[i];
1384
1385 if (vma == vma_prev)
1386 continue;
1387
1388 vma_prev = vma;
1389
1390 if (vma_is_fsdax(vma))
1391 return true;
1392 }
1393 return false;
1394 }
1395
1396 #ifdef CONFIG_CMA
1397 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1398 {
1399 /*
1400 * We want to make sure we allocate the new page from the same node
1401 * as the source page.
1402 */
1403 int nid = page_to_nid(page);
1404 /*
1405 * Trying to allocate a page for migration. Ignore allocation
1406 * failure warnings. We don't force __GFP_THISNODE here because
1407 * this node here is the node where we have CMA reservation and
1408 * in some case these nodes will have really less non movable
1409 * allocation memory.
1410 */
1411 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1412
1413 if (PageHighMem(page))
1414 gfp_mask |= __GFP_HIGHMEM;
1415
1416 #ifdef CONFIG_HUGETLB_PAGE
1417 if (PageHuge(page)) {
1418 struct hstate *h = page_hstate(page);
1419 /*
1420 * We don't want to dequeue from the pool because pool pages will
1421 * mostly be from the CMA region.
1422 */
1423 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1424 }
1425 #endif
1426 if (PageTransHuge(page)) {
1427 struct page *thp;
1428 /*
1429 * ignore allocation failure warnings
1430 */
1431 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1432
1433 /*
1434 * Remove the movable mask so that we don't allocate from
1435 * CMA area again.
1436 */
1437 thp_gfpmask &= ~__GFP_MOVABLE;
1438 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1439 if (!thp)
1440 return NULL;
1441 prep_transhuge_page(thp);
1442 return thp;
1443 }
1444
1445 return __alloc_pages_node(nid, gfp_mask, 0);
1446 }
1447
1448 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1449 struct mm_struct *mm,
1450 unsigned long start,
1451 unsigned long nr_pages,
1452 struct page **pages,
1453 struct vm_area_struct **vmas,
1454 unsigned int gup_flags)
1455 {
1456 unsigned long i;
1457 unsigned long step;
1458 bool drain_allow = true;
1459 bool migrate_allow = true;
1460 LIST_HEAD(cma_page_list);
1461
1462 check_again:
1463 for (i = 0; i < nr_pages;) {
1464
1465 struct page *head = compound_head(pages[i]);
1466
1467 /*
1468 * gup may start from a tail page. Advance step by the left
1469 * part.
1470 */
1471 step = compound_nr(head) - (pages[i] - head);
1472 /*
1473 * If we get a page from the CMA zone, since we are going to
1474 * be pinning these entries, we might as well move them out
1475 * of the CMA zone if possible.
1476 */
1477 if (is_migrate_cma_page(head)) {
1478 if (PageHuge(head))
1479 isolate_huge_page(head, &cma_page_list);
1480 else {
1481 if (!PageLRU(head) && drain_allow) {
1482 lru_add_drain_all();
1483 drain_allow = false;
1484 }
1485
1486 if (!isolate_lru_page(head)) {
1487 list_add_tail(&head->lru, &cma_page_list);
1488 mod_node_page_state(page_pgdat(head),
1489 NR_ISOLATED_ANON +
1490 page_is_file_cache(head),
1491 hpage_nr_pages(head));
1492 }
1493 }
1494 }
1495
1496 i += step;
1497 }
1498
1499 if (!list_empty(&cma_page_list)) {
1500 /*
1501 * drop the above get_user_pages reference.
1502 */
1503 for (i = 0; i < nr_pages; i++)
1504 put_page(pages[i]);
1505
1506 if (migrate_pages(&cma_page_list, new_non_cma_page,
1507 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1508 /*
1509 * some of the pages failed migration. Do get_user_pages
1510 * without migration.
1511 */
1512 migrate_allow = false;
1513
1514 if (!list_empty(&cma_page_list))
1515 putback_movable_pages(&cma_page_list);
1516 }
1517 /*
1518 * We did migrate all the pages, Try to get the page references
1519 * again migrating any new CMA pages which we failed to isolate
1520 * earlier.
1521 */
1522 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1523 pages, vmas, NULL,
1524 gup_flags);
1525
1526 if ((nr_pages > 0) && migrate_allow) {
1527 drain_allow = true;
1528 goto check_again;
1529 }
1530 }
1531
1532 return nr_pages;
1533 }
1534 #else
1535 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1536 struct mm_struct *mm,
1537 unsigned long start,
1538 unsigned long nr_pages,
1539 struct page **pages,
1540 struct vm_area_struct **vmas,
1541 unsigned int gup_flags)
1542 {
1543 return nr_pages;
1544 }
1545 #endif /* CONFIG_CMA */
1546
1547 /*
1548 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1549 * allows us to process the FOLL_LONGTERM flag.
1550 */
1551 static long __gup_longterm_locked(struct task_struct *tsk,
1552 struct mm_struct *mm,
1553 unsigned long start,
1554 unsigned long nr_pages,
1555 struct page **pages,
1556 struct vm_area_struct **vmas,
1557 unsigned int gup_flags)
1558 {
1559 struct vm_area_struct **vmas_tmp = vmas;
1560 unsigned long flags = 0;
1561 long rc, i;
1562
1563 if (gup_flags & FOLL_LONGTERM) {
1564 if (!pages)
1565 return -EINVAL;
1566
1567 if (!vmas_tmp) {
1568 vmas_tmp = kcalloc(nr_pages,
1569 sizeof(struct vm_area_struct *),
1570 GFP_KERNEL);
1571 if (!vmas_tmp)
1572 return -ENOMEM;
1573 }
1574 flags = memalloc_nocma_save();
1575 }
1576
1577 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1578 vmas_tmp, NULL, gup_flags);
1579
1580 if (gup_flags & FOLL_LONGTERM) {
1581 memalloc_nocma_restore(flags);
1582 if (rc < 0)
1583 goto out;
1584
1585 if (check_dax_vmas(vmas_tmp, rc)) {
1586 for (i = 0; i < rc; i++)
1587 put_page(pages[i]);
1588 rc = -EOPNOTSUPP;
1589 goto out;
1590 }
1591
1592 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1593 vmas_tmp, gup_flags);
1594 }
1595
1596 out:
1597 if (vmas_tmp != vmas)
1598 kfree(vmas_tmp);
1599 return rc;
1600 }
1601 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1602 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1603 struct mm_struct *mm,
1604 unsigned long start,
1605 unsigned long nr_pages,
1606 struct page **pages,
1607 struct vm_area_struct **vmas,
1608 unsigned int flags)
1609 {
1610 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1611 NULL, flags);
1612 }
1613 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1614
1615 /*
1616 * This is the same as get_user_pages_remote(), just with a
1617 * less-flexible calling convention where we assume that the task
1618 * and mm being operated on are the current task's and don't allow
1619 * passing of a locked parameter. We also obviously don't pass
1620 * FOLL_REMOTE in here.
1621 */
1622 long get_user_pages(unsigned long start, unsigned long nr_pages,
1623 unsigned int gup_flags, struct page **pages,
1624 struct vm_area_struct **vmas)
1625 {
1626 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1627 pages, vmas, gup_flags | FOLL_TOUCH);
1628 }
1629 EXPORT_SYMBOL(get_user_pages);
1630
1631 /*
1632 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1633 * paths better by using either get_user_pages_locked() or
1634 * get_user_pages_unlocked().
1635 *
1636 * get_user_pages_locked() is suitable to replace the form:
1637 *
1638 * down_read(&mm->mmap_sem);
1639 * do_something()
1640 * get_user_pages(tsk, mm, ..., pages, NULL);
1641 * up_read(&mm->mmap_sem);
1642 *
1643 * to:
1644 *
1645 * int locked = 1;
1646 * down_read(&mm->mmap_sem);
1647 * do_something()
1648 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1649 * if (locked)
1650 * up_read(&mm->mmap_sem);
1651 */
1652 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1653 unsigned int gup_flags, struct page **pages,
1654 int *locked)
1655 {
1656 /*
1657 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1658 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1659 * vmas. As there are no users of this flag in this call we simply
1660 * disallow this option for now.
1661 */
1662 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1663 return -EINVAL;
1664
1665 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1666 pages, NULL, locked,
1667 gup_flags | FOLL_TOUCH);
1668 }
1669 EXPORT_SYMBOL(get_user_pages_locked);
1670
1671 /*
1672 * get_user_pages_unlocked() is suitable to replace the form:
1673 *
1674 * down_read(&mm->mmap_sem);
1675 * get_user_pages(tsk, mm, ..., pages, NULL);
1676 * up_read(&mm->mmap_sem);
1677 *
1678 * with:
1679 *
1680 * get_user_pages_unlocked(tsk, mm, ..., pages);
1681 *
1682 * It is functionally equivalent to get_user_pages_fast so
1683 * get_user_pages_fast should be used instead if specific gup_flags
1684 * (e.g. FOLL_FORCE) are not required.
1685 */
1686 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1687 struct page **pages, unsigned int gup_flags)
1688 {
1689 struct mm_struct *mm = current->mm;
1690 int locked = 1;
1691 long ret;
1692
1693 /*
1694 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1695 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1696 * vmas. As there are no users of this flag in this call we simply
1697 * disallow this option for now.
1698 */
1699 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1700 return -EINVAL;
1701
1702 down_read(&mm->mmap_sem);
1703 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1704 &locked, gup_flags | FOLL_TOUCH);
1705 if (locked)
1706 up_read(&mm->mmap_sem);
1707 return ret;
1708 }
1709 EXPORT_SYMBOL(get_user_pages_unlocked);
1710
1711 /*
1712 * Fast GUP
1713 *
1714 * get_user_pages_fast attempts to pin user pages by walking the page
1715 * tables directly and avoids taking locks. Thus the walker needs to be
1716 * protected from page table pages being freed from under it, and should
1717 * block any THP splits.
1718 *
1719 * One way to achieve this is to have the walker disable interrupts, and
1720 * rely on IPIs from the TLB flushing code blocking before the page table
1721 * pages are freed. This is unsuitable for architectures that do not need
1722 * to broadcast an IPI when invalidating TLBs.
1723 *
1724 * Another way to achieve this is to batch up page table containing pages
1725 * belonging to more than one mm_user, then rcu_sched a callback to free those
1726 * pages. Disabling interrupts will allow the fast_gup walker to both block
1727 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1728 * (which is a relatively rare event). The code below adopts this strategy.
1729 *
1730 * Before activating this code, please be aware that the following assumptions
1731 * are currently made:
1732 *
1733 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1734 * free pages containing page tables or TLB flushing requires IPI broadcast.
1735 *
1736 * *) ptes can be read atomically by the architecture.
1737 *
1738 * *) access_ok is sufficient to validate userspace address ranges.
1739 *
1740 * The last two assumptions can be relaxed by the addition of helper functions.
1741 *
1742 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1743 */
1744 #ifdef CONFIG_HAVE_FAST_GUP
1745 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1746 /*
1747 * WARNING: only to be used in the get_user_pages_fast() implementation.
1748 *
1749 * With get_user_pages_fast(), we walk down the pagetables without taking any
1750 * locks. For this we would like to load the pointers atomically, but sometimes
1751 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1752 * we do have is the guarantee that a PTE will only either go from not present
1753 * to present, or present to not present or both -- it will not switch to a
1754 * completely different present page without a TLB flush in between; something
1755 * that we are blocking by holding interrupts off.
1756 *
1757 * Setting ptes from not present to present goes:
1758 *
1759 * ptep->pte_high = h;
1760 * smp_wmb();
1761 * ptep->pte_low = l;
1762 *
1763 * And present to not present goes:
1764 *
1765 * ptep->pte_low = 0;
1766 * smp_wmb();
1767 * ptep->pte_high = 0;
1768 *
1769 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1770 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1771 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1772 * picked up a changed pte high. We might have gotten rubbish values from
1773 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1774 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1775 * operates on present ptes we're safe.
1776 */
1777 static inline pte_t gup_get_pte(pte_t *ptep)
1778 {
1779 pte_t pte;
1780
1781 do {
1782 pte.pte_low = ptep->pte_low;
1783 smp_rmb();
1784 pte.pte_high = ptep->pte_high;
1785 smp_rmb();
1786 } while (unlikely(pte.pte_low != ptep->pte_low));
1787
1788 return pte;
1789 }
1790 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1791 /*
1792 * We require that the PTE can be read atomically.
1793 */
1794 static inline pte_t gup_get_pte(pte_t *ptep)
1795 {
1796 return READ_ONCE(*ptep);
1797 }
1798 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1799
1800 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1801 struct page **pages)
1802 {
1803 while ((*nr) - nr_start) {
1804 struct page *page = pages[--(*nr)];
1805
1806 ClearPageReferenced(page);
1807 put_page(page);
1808 }
1809 }
1810
1811 /*
1812 * Return the compund head page with ref appropriately incremented,
1813 * or NULL if that failed.
1814 */
1815 static inline struct page *try_get_compound_head(struct page *page, int refs)
1816 {
1817 struct page *head = compound_head(page);
1818 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1819 return NULL;
1820 if (unlikely(!page_cache_add_speculative(head, refs)))
1821 return NULL;
1822 return head;
1823 }
1824
1825 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1826 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1827 unsigned int flags, struct page **pages, int *nr)
1828 {
1829 struct dev_pagemap *pgmap = NULL;
1830 int nr_start = *nr, ret = 0;
1831 pte_t *ptep, *ptem;
1832
1833 ptem = ptep = pte_offset_map(&pmd, addr);
1834 do {
1835 pte_t pte = gup_get_pte(ptep);
1836 struct page *head, *page;
1837
1838 /*
1839 * Similar to the PMD case below, NUMA hinting must take slow
1840 * path using the pte_protnone check.
1841 */
1842 if (pte_protnone(pte))
1843 goto pte_unmap;
1844
1845 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1846 goto pte_unmap;
1847
1848 if (pte_devmap(pte)) {
1849 if (unlikely(flags & FOLL_LONGTERM))
1850 goto pte_unmap;
1851
1852 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1853 if (unlikely(!pgmap)) {
1854 undo_dev_pagemap(nr, nr_start, pages);
1855 goto pte_unmap;
1856 }
1857 } else if (pte_special(pte))
1858 goto pte_unmap;
1859
1860 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1861 page = pte_page(pte);
1862
1863 head = try_get_compound_head(page, 1);
1864 if (!head)
1865 goto pte_unmap;
1866
1867 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1868 put_page(head);
1869 goto pte_unmap;
1870 }
1871
1872 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1873
1874 SetPageReferenced(page);
1875 pages[*nr] = page;
1876 (*nr)++;
1877
1878 } while (ptep++, addr += PAGE_SIZE, addr != end);
1879
1880 ret = 1;
1881
1882 pte_unmap:
1883 if (pgmap)
1884 put_dev_pagemap(pgmap);
1885 pte_unmap(ptem);
1886 return ret;
1887 }
1888 #else
1889
1890 /*
1891 * If we can't determine whether or not a pte is special, then fail immediately
1892 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1893 * to be special.
1894 *
1895 * For a futex to be placed on a THP tail page, get_futex_key requires a
1896 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1897 * useful to have gup_huge_pmd even if we can't operate on ptes.
1898 */
1899 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1900 unsigned int flags, struct page **pages, int *nr)
1901 {
1902 return 0;
1903 }
1904 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1905
1906 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1907 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1908 unsigned long end, struct page **pages, int *nr)
1909 {
1910 int nr_start = *nr;
1911 struct dev_pagemap *pgmap = NULL;
1912
1913 do {
1914 struct page *page = pfn_to_page(pfn);
1915
1916 pgmap = get_dev_pagemap(pfn, pgmap);
1917 if (unlikely(!pgmap)) {
1918 undo_dev_pagemap(nr, nr_start, pages);
1919 return 0;
1920 }
1921 SetPageReferenced(page);
1922 pages[*nr] = page;
1923 get_page(page);
1924 (*nr)++;
1925 pfn++;
1926 } while (addr += PAGE_SIZE, addr != end);
1927
1928 if (pgmap)
1929 put_dev_pagemap(pgmap);
1930 return 1;
1931 }
1932
1933 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1934 unsigned long end, struct page **pages, int *nr)
1935 {
1936 unsigned long fault_pfn;
1937 int nr_start = *nr;
1938
1939 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1940 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1941 return 0;
1942
1943 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1944 undo_dev_pagemap(nr, nr_start, pages);
1945 return 0;
1946 }
1947 return 1;
1948 }
1949
1950 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1951 unsigned long end, struct page **pages, int *nr)
1952 {
1953 unsigned long fault_pfn;
1954 int nr_start = *nr;
1955
1956 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1957 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1958 return 0;
1959
1960 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1961 undo_dev_pagemap(nr, nr_start, pages);
1962 return 0;
1963 }
1964 return 1;
1965 }
1966 #else
1967 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1968 unsigned long end, struct page **pages, int *nr)
1969 {
1970 BUILD_BUG();
1971 return 0;
1972 }
1973
1974 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1975 unsigned long end, struct page **pages, int *nr)
1976 {
1977 BUILD_BUG();
1978 return 0;
1979 }
1980 #endif
1981
1982 #ifdef CONFIG_ARCH_HAS_HUGEPD
1983 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1984 unsigned long sz)
1985 {
1986 unsigned long __boundary = (addr + sz) & ~(sz-1);
1987 return (__boundary - 1 < end - 1) ? __boundary : end;
1988 }
1989
1990 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1991 unsigned long end, unsigned int flags,
1992 struct page **pages, int *nr)
1993 {
1994 unsigned long pte_end;
1995 struct page *head, *page;
1996 pte_t pte;
1997 int refs;
1998
1999 pte_end = (addr + sz) & ~(sz-1);
2000 if (pte_end < end)
2001 end = pte_end;
2002
2003 pte = READ_ONCE(*ptep);
2004
2005 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2006 return 0;
2007
2008 /* hugepages are never "special" */
2009 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2010
2011 refs = 0;
2012 head = pte_page(pte);
2013
2014 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2015 do {
2016 VM_BUG_ON(compound_head(page) != head);
2017 pages[*nr] = page;
2018 (*nr)++;
2019 page++;
2020 refs++;
2021 } while (addr += PAGE_SIZE, addr != end);
2022
2023 head = try_get_compound_head(head, refs);
2024 if (!head) {
2025 *nr -= refs;
2026 return 0;
2027 }
2028
2029 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2030 /* Could be optimized better */
2031 *nr -= refs;
2032 while (refs--)
2033 put_page(head);
2034 return 0;
2035 }
2036
2037 SetPageReferenced(head);
2038 return 1;
2039 }
2040
2041 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2042 unsigned int pdshift, unsigned long end, unsigned int flags,
2043 struct page **pages, int *nr)
2044 {
2045 pte_t *ptep;
2046 unsigned long sz = 1UL << hugepd_shift(hugepd);
2047 unsigned long next;
2048
2049 ptep = hugepte_offset(hugepd, addr, pdshift);
2050 do {
2051 next = hugepte_addr_end(addr, end, sz);
2052 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2053 return 0;
2054 } while (ptep++, addr = next, addr != end);
2055
2056 return 1;
2057 }
2058 #else
2059 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2060 unsigned int pdshift, unsigned long end, unsigned int flags,
2061 struct page **pages, int *nr)
2062 {
2063 return 0;
2064 }
2065 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2066
2067 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2068 unsigned long end, unsigned int flags,
2069 struct page **pages, int *nr)
2070 {
2071 struct page *head, *page;
2072 int refs;
2073
2074 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2075 return 0;
2076
2077 if (pmd_devmap(orig)) {
2078 if (unlikely(flags & FOLL_LONGTERM))
2079 return 0;
2080 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2081 }
2082
2083 refs = 0;
2084 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2085 do {
2086 pages[*nr] = page;
2087 (*nr)++;
2088 page++;
2089 refs++;
2090 } while (addr += PAGE_SIZE, addr != end);
2091
2092 head = try_get_compound_head(pmd_page(orig), refs);
2093 if (!head) {
2094 *nr -= refs;
2095 return 0;
2096 }
2097
2098 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2099 *nr -= refs;
2100 while (refs--)
2101 put_page(head);
2102 return 0;
2103 }
2104
2105 SetPageReferenced(head);
2106 return 1;
2107 }
2108
2109 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2110 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2111 {
2112 struct page *head, *page;
2113 int refs;
2114
2115 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2116 return 0;
2117
2118 if (pud_devmap(orig)) {
2119 if (unlikely(flags & FOLL_LONGTERM))
2120 return 0;
2121 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2122 }
2123
2124 refs = 0;
2125 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2126 do {
2127 pages[*nr] = page;
2128 (*nr)++;
2129 page++;
2130 refs++;
2131 } while (addr += PAGE_SIZE, addr != end);
2132
2133 head = try_get_compound_head(pud_page(orig), refs);
2134 if (!head) {
2135 *nr -= refs;
2136 return 0;
2137 }
2138
2139 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2140 *nr -= refs;
2141 while (refs--)
2142 put_page(head);
2143 return 0;
2144 }
2145
2146 SetPageReferenced(head);
2147 return 1;
2148 }
2149
2150 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2151 unsigned long end, unsigned int flags,
2152 struct page **pages, int *nr)
2153 {
2154 int refs;
2155 struct page *head, *page;
2156
2157 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2158 return 0;
2159
2160 BUILD_BUG_ON(pgd_devmap(orig));
2161 refs = 0;
2162 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2163 do {
2164 pages[*nr] = page;
2165 (*nr)++;
2166 page++;
2167 refs++;
2168 } while (addr += PAGE_SIZE, addr != end);
2169
2170 head = try_get_compound_head(pgd_page(orig), refs);
2171 if (!head) {
2172 *nr -= refs;
2173 return 0;
2174 }
2175
2176 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2177 *nr -= refs;
2178 while (refs--)
2179 put_page(head);
2180 return 0;
2181 }
2182
2183 SetPageReferenced(head);
2184 return 1;
2185 }
2186
2187 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2188 unsigned int flags, struct page **pages, int *nr)
2189 {
2190 unsigned long next;
2191 pmd_t *pmdp;
2192
2193 pmdp = pmd_offset(&pud, addr);
2194 do {
2195 pmd_t pmd = READ_ONCE(*pmdp);
2196
2197 next = pmd_addr_end(addr, end);
2198 if (!pmd_present(pmd))
2199 return 0;
2200
2201 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2202 pmd_devmap(pmd))) {
2203 /*
2204 * NUMA hinting faults need to be handled in the GUP
2205 * slowpath for accounting purposes and so that they
2206 * can be serialised against THP migration.
2207 */
2208 if (pmd_protnone(pmd))
2209 return 0;
2210
2211 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2212 pages, nr))
2213 return 0;
2214
2215 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2216 /*
2217 * architecture have different format for hugetlbfs
2218 * pmd format and THP pmd format
2219 */
2220 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2221 PMD_SHIFT, next, flags, pages, nr))
2222 return 0;
2223 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2224 return 0;
2225 } while (pmdp++, addr = next, addr != end);
2226
2227 return 1;
2228 }
2229
2230 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2231 unsigned int flags, struct page **pages, int *nr)
2232 {
2233 unsigned long next;
2234 pud_t *pudp;
2235
2236 pudp = pud_offset(&p4d, addr);
2237 do {
2238 pud_t pud = READ_ONCE(*pudp);
2239
2240 next = pud_addr_end(addr, end);
2241 if (pud_none(pud))
2242 return 0;
2243 if (unlikely(pud_huge(pud))) {
2244 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2245 pages, nr))
2246 return 0;
2247 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2248 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2249 PUD_SHIFT, next, flags, pages, nr))
2250 return 0;
2251 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2252 return 0;
2253 } while (pudp++, addr = next, addr != end);
2254
2255 return 1;
2256 }
2257
2258 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2259 unsigned int flags, struct page **pages, int *nr)
2260 {
2261 unsigned long next;
2262 p4d_t *p4dp;
2263
2264 p4dp = p4d_offset(&pgd, addr);
2265 do {
2266 p4d_t p4d = READ_ONCE(*p4dp);
2267
2268 next = p4d_addr_end(addr, end);
2269 if (p4d_none(p4d))
2270 return 0;
2271 BUILD_BUG_ON(p4d_huge(p4d));
2272 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2273 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2274 P4D_SHIFT, next, flags, pages, nr))
2275 return 0;
2276 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2277 return 0;
2278 } while (p4dp++, addr = next, addr != end);
2279
2280 return 1;
2281 }
2282
2283 static void gup_pgd_range(unsigned long addr, unsigned long end,
2284 unsigned int flags, struct page **pages, int *nr)
2285 {
2286 unsigned long next;
2287 pgd_t *pgdp;
2288
2289 pgdp = pgd_offset(current->mm, addr);
2290 do {
2291 pgd_t pgd = READ_ONCE(*pgdp);
2292
2293 next = pgd_addr_end(addr, end);
2294 if (pgd_none(pgd))
2295 return;
2296 if (unlikely(pgd_huge(pgd))) {
2297 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2298 pages, nr))
2299 return;
2300 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2301 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2302 PGDIR_SHIFT, next, flags, pages, nr))
2303 return;
2304 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2305 return;
2306 } while (pgdp++, addr = next, addr != end);
2307 }
2308 #else
2309 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2310 unsigned int flags, struct page **pages, int *nr)
2311 {
2312 }
2313 #endif /* CONFIG_HAVE_FAST_GUP */
2314
2315 #ifndef gup_fast_permitted
2316 /*
2317 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2318 * we need to fall back to the slow version:
2319 */
2320 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2321 {
2322 return true;
2323 }
2324 #endif
2325
2326 /*
2327 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2328 * the regular GUP.
2329 * Note a difference with get_user_pages_fast: this always returns the
2330 * number of pages pinned, 0 if no pages were pinned.
2331 *
2332 * If the architecture does not support this function, simply return with no
2333 * pages pinned.
2334 *
2335 * Careful, careful! COW breaking can go either way, so a non-write
2336 * access can get ambiguous page results. If you call this function without
2337 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2338 */
2339 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2340 struct page **pages)
2341 {
2342 unsigned long len, end;
2343 unsigned long flags;
2344 int nr = 0;
2345
2346 start = untagged_addr(start) & PAGE_MASK;
2347 len = (unsigned long) nr_pages << PAGE_SHIFT;
2348 end = start + len;
2349
2350 if (end <= start)
2351 return 0;
2352 if (unlikely(!access_ok((void __user *)start, len)))
2353 return 0;
2354
2355 /*
2356 * Disable interrupts. We use the nested form as we can already have
2357 * interrupts disabled by get_futex_key.
2358 *
2359 * With interrupts disabled, we block page table pages from being
2360 * freed from under us. See struct mmu_table_batch comments in
2361 * include/asm-generic/tlb.h for more details.
2362 *
2363 * We do not adopt an rcu_read_lock(.) here as we also want to
2364 * block IPIs that come from THPs splitting.
2365 *
2366 * NOTE! We allow read-only gup_fast() here, but you'd better be
2367 * careful about possible COW pages. You'll get _a_ COW page, but
2368 * not necessarily the one you intended to get depending on what
2369 * COW event happens after this. COW may break the page copy in a
2370 * random direction.
2371 */
2372
2373 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2374 gup_fast_permitted(start, end)) {
2375 local_irq_save(flags);
2376 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2377 local_irq_restore(flags);
2378 }
2379
2380 return nr;
2381 }
2382 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2383
2384 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2385 unsigned int gup_flags, struct page **pages)
2386 {
2387 int ret;
2388
2389 /*
2390 * FIXME: FOLL_LONGTERM does not work with
2391 * get_user_pages_unlocked() (see comments in that function)
2392 */
2393 if (gup_flags & FOLL_LONGTERM) {
2394 down_read(¤t->mm->mmap_sem);
2395 ret = __gup_longterm_locked(current, current->mm,
2396 start, nr_pages,
2397 pages, NULL, gup_flags);
2398 up_read(¤t->mm->mmap_sem);
2399 } else {
2400 ret = get_user_pages_unlocked(start, nr_pages,
2401 pages, gup_flags);
2402 }
2403
2404 return ret;
2405 }
2406
2407 /**
2408 * get_user_pages_fast() - pin user pages in memory
2409 * @start: starting user address
2410 * @nr_pages: number of pages from start to pin
2411 * @gup_flags: flags modifying pin behaviour
2412 * @pages: array that receives pointers to the pages pinned.
2413 * Should be at least nr_pages long.
2414 *
2415 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2416 * If not successful, it will fall back to taking the lock and
2417 * calling get_user_pages().
2418 *
2419 * Returns number of pages pinned. This may be fewer than the number
2420 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2421 * were pinned, returns -errno.
2422 */
2423 int get_user_pages_fast(unsigned long start, int nr_pages,
2424 unsigned int gup_flags, struct page **pages)
2425 {
2426 unsigned long addr, len, end;
2427 int nr = 0, ret = 0;
2428
2429 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2430 FOLL_FORCE)))
2431 return -EINVAL;
2432
2433 start = untagged_addr(start) & PAGE_MASK;
2434 addr = start;
2435 len = (unsigned long) nr_pages << PAGE_SHIFT;
2436 end = start + len;
2437
2438 if (end <= start)
2439 return 0;
2440 if (unlikely(!access_ok((void __user *)start, len)))
2441 return -EFAULT;
2442
2443 /*
2444 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
2445 * because get_user_pages() may need to cause an early COW in
2446 * order to avoid confusing the normal COW routines. So only
2447 * targets that are already writable are safe to do by just
2448 * looking at the page tables.
2449 */
2450 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2451 gup_fast_permitted(start, end)) {
2452 local_irq_disable();
2453 gup_pgd_range(addr, end, gup_flags | FOLL_WRITE, pages, &nr);
2454 local_irq_enable();
2455 ret = nr;
2456 }
2457
2458 if (nr < nr_pages) {
2459 /* Try to get the remaining pages with get_user_pages */
2460 start += nr << PAGE_SHIFT;
2461 pages += nr;
2462
2463 ret = __gup_longterm_unlocked(start, nr_pages - nr,
2464 gup_flags, pages);
2465
2466 /* Have to be a bit careful with return values */
2467 if (nr > 0) {
2468 if (ret < 0)
2469 ret = nr;
2470 else
2471 ret += nr;
2472 }
2473 }
2474
2475 return ret;
2476 }
2477 EXPORT_SYMBOL_GPL(get_user_pages_fast);