Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / mm / gup.c
blob1b46e6e74881d3ce634511d98e4f177625b5501c
1 #include <linux/kernel.h>
2 #include <linux/errno.h>
3 #include <linux/err.h>
4 #include <linux/spinlock.h>
6 #include <linux/mm.h>
7 #include <linux/memremap.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/swapops.h>
13 #include <linux/sched/signal.h>
14 #include <linux/rwsem.h>
15 #include <linux/hugetlb.h>
17 #include <asm/mmu_context.h>
18 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
21 #include "internal.h"
23 static struct page *no_page_table(struct vm_area_struct *vma,
24 unsigned int flags)
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
34 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
35 return ERR_PTR(-EFAULT);
36 return NULL;
39 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
40 pte_t *pte, unsigned int flags)
42 /* No page to get reference */
43 if (flags & FOLL_GET)
44 return -EFAULT;
46 if (flags & FOLL_TOUCH) {
47 pte_t entry = *pte;
49 if (flags & FOLL_WRITE)
50 entry = pte_mkdirty(entry);
51 entry = pte_mkyoung(entry);
53 if (!pte_same(*pte, entry)) {
54 set_pte_at(vma->vm_mm, address, pte, entry);
55 update_mmu_cache(vma, address, pte);
59 /* Proper page table entry exists, but no corresponding struct page */
60 return -EEXIST;
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
69 return pte_write(pte) ||
70 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
73 static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
76 struct mm_struct *mm = vma->vm_mm;
77 struct dev_pagemap *pgmap = NULL;
78 struct page *page;
79 spinlock_t *ptl;
80 pte_t *ptep, pte;
82 retry:
83 if (unlikely(pmd_bad(*pmd)))
84 return no_page_table(vma, flags);
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
87 pte = *ptep;
88 if (!pte_present(pte)) {
89 swp_entry_t entry;
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
95 if (likely(!(flags & FOLL_MIGRATION)))
96 goto no_page;
97 if (pte_none(pte))
98 goto no_page;
99 entry = pte_to_swp_entry(pte);
100 if (!is_migration_entry(entry))
101 goto no_page;
102 pte_unmap_unlock(ptep, ptl);
103 migration_entry_wait(mm, pmd, address);
104 goto retry;
106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
107 goto no_page;
108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
109 pte_unmap_unlock(ptep, ptl);
110 return NULL;
113 page = vm_normal_page(vma, address, pte);
114 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
119 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
120 if (pgmap)
121 page = pte_page(pte);
122 else
123 goto no_page;
124 } else if (unlikely(!page)) {
125 if (flags & FOLL_DUMP) {
126 /* Avoid special (like zero) pages in core dumps */
127 page = ERR_PTR(-EFAULT);
128 goto out;
131 if (is_zero_pfn(pte_pfn(pte))) {
132 page = pte_page(pte);
133 } else {
134 int ret;
136 ret = follow_pfn_pte(vma, address, ptep, flags);
137 page = ERR_PTR(ret);
138 goto out;
142 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
143 int ret;
144 get_page(page);
145 pte_unmap_unlock(ptep, ptl);
146 lock_page(page);
147 ret = split_huge_page(page);
148 unlock_page(page);
149 put_page(page);
150 if (ret)
151 return ERR_PTR(ret);
152 goto retry;
155 if (flags & FOLL_GET) {
156 get_page(page);
158 /* drop the pgmap reference now that we hold the page */
159 if (pgmap) {
160 put_dev_pagemap(pgmap);
161 pgmap = NULL;
164 if (flags & FOLL_TOUCH) {
165 if ((flags & FOLL_WRITE) &&
166 !pte_dirty(pte) && !PageDirty(page))
167 set_page_dirty(page);
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
173 mark_page_accessed(page);
175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
178 goto out;
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
189 if (page->mapping && trylock_page(page)) {
190 lru_add_drain(); /* push cached pages to LRU */
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
197 mlock_vma_page(page);
198 unlock_page(page);
201 out:
202 pte_unmap_unlock(ptep, ptl);
203 return page;
204 no_page:
205 pte_unmap_unlock(ptep, ptl);
206 if (!pte_none(pte))
207 return NULL;
208 return no_page_table(vma, flags);
211 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
212 unsigned long address, pud_t *pudp,
213 unsigned int flags, unsigned int *page_mask)
215 pmd_t *pmd;
216 spinlock_t *ptl;
217 struct page *page;
218 struct mm_struct *mm = vma->vm_mm;
220 pmd = pmd_offset(pudp, address);
221 if (pmd_none(*pmd))
222 return no_page_table(vma, flags);
223 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
224 page = follow_huge_pmd(mm, address, pmd, flags);
225 if (page)
226 return page;
227 return no_page_table(vma, flags);
229 if (is_hugepd(__hugepd(pmd_val(*pmd)))) {
230 page = follow_huge_pd(vma, address,
231 __hugepd(pmd_val(*pmd)), flags,
232 PMD_SHIFT);
233 if (page)
234 return page;
235 return no_page_table(vma, flags);
237 retry:
238 if (!pmd_present(*pmd)) {
239 if (likely(!(flags & FOLL_MIGRATION)))
240 return no_page_table(vma, flags);
241 VM_BUG_ON(thp_migration_supported() &&
242 !is_pmd_migration_entry(*pmd));
243 if (is_pmd_migration_entry(*pmd))
244 pmd_migration_entry_wait(mm, pmd);
245 goto retry;
247 if (pmd_devmap(*pmd)) {
248 ptl = pmd_lock(mm, pmd);
249 page = follow_devmap_pmd(vma, address, pmd, flags);
250 spin_unlock(ptl);
251 if (page)
252 return page;
254 if (likely(!pmd_trans_huge(*pmd)))
255 return follow_page_pte(vma, address, pmd, flags);
257 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
258 return no_page_table(vma, flags);
260 retry_locked:
261 ptl = pmd_lock(mm, pmd);
262 if (unlikely(!pmd_present(*pmd))) {
263 spin_unlock(ptl);
264 if (likely(!(flags & FOLL_MIGRATION)))
265 return no_page_table(vma, flags);
266 pmd_migration_entry_wait(mm, pmd);
267 goto retry_locked;
269 if (unlikely(!pmd_trans_huge(*pmd))) {
270 spin_unlock(ptl);
271 return follow_page_pte(vma, address, pmd, flags);
273 if (flags & FOLL_SPLIT) {
274 int ret;
275 page = pmd_page(*pmd);
276 if (is_huge_zero_page(page)) {
277 spin_unlock(ptl);
278 ret = 0;
279 split_huge_pmd(vma, pmd, address);
280 if (pmd_trans_unstable(pmd))
281 ret = -EBUSY;
282 } else {
283 get_page(page);
284 spin_unlock(ptl);
285 lock_page(page);
286 ret = split_huge_page(page);
287 unlock_page(page);
288 put_page(page);
289 if (pmd_none(*pmd))
290 return no_page_table(vma, flags);
293 return ret ? ERR_PTR(ret) :
294 follow_page_pte(vma, address, pmd, flags);
296 page = follow_trans_huge_pmd(vma, address, pmd, flags);
297 spin_unlock(ptl);
298 *page_mask = HPAGE_PMD_NR - 1;
299 return page;
303 static struct page *follow_pud_mask(struct vm_area_struct *vma,
304 unsigned long address, p4d_t *p4dp,
305 unsigned int flags, unsigned int *page_mask)
307 pud_t *pud;
308 spinlock_t *ptl;
309 struct page *page;
310 struct mm_struct *mm = vma->vm_mm;
312 pud = pud_offset(p4dp, address);
313 if (pud_none(*pud))
314 return no_page_table(vma, flags);
315 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
316 page = follow_huge_pud(mm, address, pud, flags);
317 if (page)
318 return page;
319 return no_page_table(vma, flags);
321 if (is_hugepd(__hugepd(pud_val(*pud)))) {
322 page = follow_huge_pd(vma, address,
323 __hugepd(pud_val(*pud)), flags,
324 PUD_SHIFT);
325 if (page)
326 return page;
327 return no_page_table(vma, flags);
329 if (pud_devmap(*pud)) {
330 ptl = pud_lock(mm, pud);
331 page = follow_devmap_pud(vma, address, pud, flags);
332 spin_unlock(ptl);
333 if (page)
334 return page;
336 if (unlikely(pud_bad(*pud)))
337 return no_page_table(vma, flags);
339 return follow_pmd_mask(vma, address, pud, flags, page_mask);
343 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
344 unsigned long address, pgd_t *pgdp,
345 unsigned int flags, unsigned int *page_mask)
347 p4d_t *p4d;
348 struct page *page;
350 p4d = p4d_offset(pgdp, address);
351 if (p4d_none(*p4d))
352 return no_page_table(vma, flags);
353 BUILD_BUG_ON(p4d_huge(*p4d));
354 if (unlikely(p4d_bad(*p4d)))
355 return no_page_table(vma, flags);
357 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
358 page = follow_huge_pd(vma, address,
359 __hugepd(p4d_val(*p4d)), flags,
360 P4D_SHIFT);
361 if (page)
362 return page;
363 return no_page_table(vma, flags);
365 return follow_pud_mask(vma, address, p4d, flags, page_mask);
369 * follow_page_mask - look up a page descriptor from a user-virtual address
370 * @vma: vm_area_struct mapping @address
371 * @address: virtual address to look up
372 * @flags: flags modifying lookup behaviour
373 * @page_mask: on output, *page_mask is set according to the size of the page
375 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
377 * Returns the mapped (struct page *), %NULL if no mapping exists, or
378 * an error pointer if there is a mapping to something not represented
379 * by a page descriptor (see also vm_normal_page()).
381 struct page *follow_page_mask(struct vm_area_struct *vma,
382 unsigned long address, unsigned int flags,
383 unsigned int *page_mask)
385 pgd_t *pgd;
386 struct page *page;
387 struct mm_struct *mm = vma->vm_mm;
389 *page_mask = 0;
391 /* make this handle hugepd */
392 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
393 if (!IS_ERR(page)) {
394 BUG_ON(flags & FOLL_GET);
395 return page;
398 pgd = pgd_offset(mm, address);
400 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
401 return no_page_table(vma, flags);
403 if (pgd_huge(*pgd)) {
404 page = follow_huge_pgd(mm, address, pgd, flags);
405 if (page)
406 return page;
407 return no_page_table(vma, flags);
409 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
410 page = follow_huge_pd(vma, address,
411 __hugepd(pgd_val(*pgd)), flags,
412 PGDIR_SHIFT);
413 if (page)
414 return page;
415 return no_page_table(vma, flags);
418 return follow_p4d_mask(vma, address, pgd, flags, page_mask);
421 static int get_gate_page(struct mm_struct *mm, unsigned long address,
422 unsigned int gup_flags, struct vm_area_struct **vma,
423 struct page **page)
425 pgd_t *pgd;
426 p4d_t *p4d;
427 pud_t *pud;
428 pmd_t *pmd;
429 pte_t *pte;
430 int ret = -EFAULT;
432 /* user gate pages are read-only */
433 if (gup_flags & FOLL_WRITE)
434 return -EFAULT;
435 if (address > TASK_SIZE)
436 pgd = pgd_offset_k(address);
437 else
438 pgd = pgd_offset_gate(mm, address);
439 BUG_ON(pgd_none(*pgd));
440 p4d = p4d_offset(pgd, address);
441 BUG_ON(p4d_none(*p4d));
442 pud = pud_offset(p4d, address);
443 BUG_ON(pud_none(*pud));
444 pmd = pmd_offset(pud, address);
445 if (!pmd_present(*pmd))
446 return -EFAULT;
447 VM_BUG_ON(pmd_trans_huge(*pmd));
448 pte = pte_offset_map(pmd, address);
449 if (pte_none(*pte))
450 goto unmap;
451 *vma = get_gate_vma(mm);
452 if (!page)
453 goto out;
454 *page = vm_normal_page(*vma, address, *pte);
455 if (!*page) {
456 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
457 goto unmap;
458 *page = pte_page(*pte);
461 * This should never happen (a device public page in the gate
462 * area).
464 if (is_device_public_page(*page))
465 goto unmap;
467 get_page(*page);
468 out:
469 ret = 0;
470 unmap:
471 pte_unmap(pte);
472 return ret;
476 * mmap_sem must be held on entry. If @nonblocking != NULL and
477 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
478 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
480 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
481 unsigned long address, unsigned int *flags, int *nonblocking)
483 unsigned int fault_flags = 0;
484 int ret;
486 /* mlock all present pages, but do not fault in new pages */
487 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
488 return -ENOENT;
489 if (*flags & FOLL_WRITE)
490 fault_flags |= FAULT_FLAG_WRITE;
491 if (*flags & FOLL_REMOTE)
492 fault_flags |= FAULT_FLAG_REMOTE;
493 if (nonblocking)
494 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
495 if (*flags & FOLL_NOWAIT)
496 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
497 if (*flags & FOLL_TRIED) {
498 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
499 fault_flags |= FAULT_FLAG_TRIED;
502 ret = handle_mm_fault(vma, address, fault_flags);
503 if (ret & VM_FAULT_ERROR) {
504 int err = vm_fault_to_errno(ret, *flags);
506 if (err)
507 return err;
508 BUG();
511 if (tsk) {
512 if (ret & VM_FAULT_MAJOR)
513 tsk->maj_flt++;
514 else
515 tsk->min_flt++;
518 if (ret & VM_FAULT_RETRY) {
519 if (nonblocking)
520 *nonblocking = 0;
521 return -EBUSY;
525 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
526 * necessary, even if maybe_mkwrite decided not to set pte_write. We
527 * can thus safely do subsequent page lookups as if they were reads.
528 * But only do so when looping for pte_write is futile: in some cases
529 * userspace may also be wanting to write to the gotten user page,
530 * which a read fault here might prevent (a readonly page might get
531 * reCOWed by userspace write).
533 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
534 *flags |= FOLL_COW;
535 return 0;
538 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
540 vm_flags_t vm_flags = vma->vm_flags;
541 int write = (gup_flags & FOLL_WRITE);
542 int foreign = (gup_flags & FOLL_REMOTE);
544 if (vm_flags & (VM_IO | VM_PFNMAP))
545 return -EFAULT;
547 if (write) {
548 if (!(vm_flags & VM_WRITE)) {
549 if (!(gup_flags & FOLL_FORCE))
550 return -EFAULT;
552 * We used to let the write,force case do COW in a
553 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
554 * set a breakpoint in a read-only mapping of an
555 * executable, without corrupting the file (yet only
556 * when that file had been opened for writing!).
557 * Anon pages in shared mappings are surprising: now
558 * just reject it.
560 if (!is_cow_mapping(vm_flags))
561 return -EFAULT;
563 } else if (!(vm_flags & VM_READ)) {
564 if (!(gup_flags & FOLL_FORCE))
565 return -EFAULT;
567 * Is there actually any vma we can reach here which does not
568 * have VM_MAYREAD set?
570 if (!(vm_flags & VM_MAYREAD))
571 return -EFAULT;
574 * gups are always data accesses, not instruction
575 * fetches, so execute=false here
577 if (!arch_vma_access_permitted(vma, write, false, foreign))
578 return -EFAULT;
579 return 0;
583 * __get_user_pages() - pin user pages in memory
584 * @tsk: task_struct of target task
585 * @mm: mm_struct of target mm
586 * @start: starting user address
587 * @nr_pages: number of pages from start to pin
588 * @gup_flags: flags modifying pin behaviour
589 * @pages: array that receives pointers to the pages pinned.
590 * Should be at least nr_pages long. Or NULL, if caller
591 * only intends to ensure the pages are faulted in.
592 * @vmas: array of pointers to vmas corresponding to each page.
593 * Or NULL if the caller does not require them.
594 * @nonblocking: whether waiting for disk IO or mmap_sem contention
596 * Returns number of pages pinned. This may be fewer than the number
597 * requested. If nr_pages is 0 or negative, returns 0. If no pages
598 * were pinned, returns -errno. Each page returned must be released
599 * with a put_page() call when it is finished with. vmas will only
600 * remain valid while mmap_sem is held.
602 * Must be called with mmap_sem held. It may be released. See below.
604 * __get_user_pages walks a process's page tables and takes a reference to
605 * each struct page that each user address corresponds to at a given
606 * instant. That is, it takes the page that would be accessed if a user
607 * thread accesses the given user virtual address at that instant.
609 * This does not guarantee that the page exists in the user mappings when
610 * __get_user_pages returns, and there may even be a completely different
611 * page there in some cases (eg. if mmapped pagecache has been invalidated
612 * and subsequently re faulted). However it does guarantee that the page
613 * won't be freed completely. And mostly callers simply care that the page
614 * contains data that was valid *at some point in time*. Typically, an IO
615 * or similar operation cannot guarantee anything stronger anyway because
616 * locks can't be held over the syscall boundary.
618 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
619 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
620 * appropriate) must be called after the page is finished with, and
621 * before put_page is called.
623 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
624 * or mmap_sem contention, and if waiting is needed to pin all pages,
625 * *@nonblocking will be set to 0. Further, if @gup_flags does not
626 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
627 * this case.
629 * A caller using such a combination of @nonblocking and @gup_flags
630 * must therefore hold the mmap_sem for reading only, and recognize
631 * when it's been released. Otherwise, it must be held for either
632 * reading or writing and will not be released.
634 * In most cases, get_user_pages or get_user_pages_fast should be used
635 * instead of __get_user_pages. __get_user_pages should be used only if
636 * you need some special @gup_flags.
638 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
639 unsigned long start, unsigned long nr_pages,
640 unsigned int gup_flags, struct page **pages,
641 struct vm_area_struct **vmas, int *nonblocking)
643 long i = 0;
644 unsigned int page_mask;
645 struct vm_area_struct *vma = NULL;
647 if (!nr_pages)
648 return 0;
650 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
653 * If FOLL_FORCE is set then do not force a full fault as the hinting
654 * fault information is unrelated to the reference behaviour of a task
655 * using the address space
657 if (!(gup_flags & FOLL_FORCE))
658 gup_flags |= FOLL_NUMA;
660 do {
661 struct page *page;
662 unsigned int foll_flags = gup_flags;
663 unsigned int page_increm;
665 /* first iteration or cross vma bound */
666 if (!vma || start >= vma->vm_end) {
667 vma = find_extend_vma(mm, start);
668 if (!vma && in_gate_area(mm, start)) {
669 int ret;
670 ret = get_gate_page(mm, start & PAGE_MASK,
671 gup_flags, &vma,
672 pages ? &pages[i] : NULL);
673 if (ret)
674 return i ? : ret;
675 page_mask = 0;
676 goto next_page;
679 if (!vma || check_vma_flags(vma, gup_flags))
680 return i ? : -EFAULT;
681 if (is_vm_hugetlb_page(vma)) {
682 i = follow_hugetlb_page(mm, vma, pages, vmas,
683 &start, &nr_pages, i,
684 gup_flags, nonblocking);
685 continue;
688 retry:
690 * If we have a pending SIGKILL, don't keep faulting pages and
691 * potentially allocating memory.
693 if (unlikely(fatal_signal_pending(current)))
694 return i ? i : -ERESTARTSYS;
695 cond_resched();
696 page = follow_page_mask(vma, start, foll_flags, &page_mask);
697 if (!page) {
698 int ret;
699 ret = faultin_page(tsk, vma, start, &foll_flags,
700 nonblocking);
701 switch (ret) {
702 case 0:
703 goto retry;
704 case -EFAULT:
705 case -ENOMEM:
706 case -EHWPOISON:
707 return i ? i : ret;
708 case -EBUSY:
709 return i;
710 case -ENOENT:
711 goto next_page;
713 BUG();
714 } else if (PTR_ERR(page) == -EEXIST) {
716 * Proper page table entry exists, but no corresponding
717 * struct page.
719 goto next_page;
720 } else if (IS_ERR(page)) {
721 return i ? i : PTR_ERR(page);
723 if (pages) {
724 pages[i] = page;
725 flush_anon_page(vma, page, start);
726 flush_dcache_page(page);
727 page_mask = 0;
729 next_page:
730 if (vmas) {
731 vmas[i] = vma;
732 page_mask = 0;
734 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
735 if (page_increm > nr_pages)
736 page_increm = nr_pages;
737 i += page_increm;
738 start += page_increm * PAGE_SIZE;
739 nr_pages -= page_increm;
740 } while (nr_pages);
741 return i;
744 static bool vma_permits_fault(struct vm_area_struct *vma,
745 unsigned int fault_flags)
747 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
748 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
749 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
751 if (!(vm_flags & vma->vm_flags))
752 return false;
755 * The architecture might have a hardware protection
756 * mechanism other than read/write that can deny access.
758 * gup always represents data access, not instruction
759 * fetches, so execute=false here:
761 if (!arch_vma_access_permitted(vma, write, false, foreign))
762 return false;
764 return true;
768 * fixup_user_fault() - manually resolve a user page fault
769 * @tsk: the task_struct to use for page fault accounting, or
770 * NULL if faults are not to be recorded.
771 * @mm: mm_struct of target mm
772 * @address: user address
773 * @fault_flags:flags to pass down to handle_mm_fault()
774 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
775 * does not allow retry
777 * This is meant to be called in the specific scenario where for locking reasons
778 * we try to access user memory in atomic context (within a pagefault_disable()
779 * section), this returns -EFAULT, and we want to resolve the user fault before
780 * trying again.
782 * Typically this is meant to be used by the futex code.
784 * The main difference with get_user_pages() is that this function will
785 * unconditionally call handle_mm_fault() which will in turn perform all the
786 * necessary SW fixup of the dirty and young bits in the PTE, while
787 * get_user_pages() only guarantees to update these in the struct page.
789 * This is important for some architectures where those bits also gate the
790 * access permission to the page because they are maintained in software. On
791 * such architectures, gup() will not be enough to make a subsequent access
792 * succeed.
794 * This function will not return with an unlocked mmap_sem. So it has not the
795 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
797 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
798 unsigned long address, unsigned int fault_flags,
799 bool *unlocked)
801 struct vm_area_struct *vma;
802 int ret, major = 0;
804 if (unlocked)
805 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
807 retry:
808 vma = find_extend_vma(mm, address);
809 if (!vma || address < vma->vm_start)
810 return -EFAULT;
812 if (!vma_permits_fault(vma, fault_flags))
813 return -EFAULT;
815 ret = handle_mm_fault(vma, address, fault_flags);
816 major |= ret & VM_FAULT_MAJOR;
817 if (ret & VM_FAULT_ERROR) {
818 int err = vm_fault_to_errno(ret, 0);
820 if (err)
821 return err;
822 BUG();
825 if (ret & VM_FAULT_RETRY) {
826 down_read(&mm->mmap_sem);
827 if (!(fault_flags & FAULT_FLAG_TRIED)) {
828 *unlocked = true;
829 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
830 fault_flags |= FAULT_FLAG_TRIED;
831 goto retry;
835 if (tsk) {
836 if (major)
837 tsk->maj_flt++;
838 else
839 tsk->min_flt++;
841 return 0;
843 EXPORT_SYMBOL_GPL(fixup_user_fault);
845 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
846 struct mm_struct *mm,
847 unsigned long start,
848 unsigned long nr_pages,
849 struct page **pages,
850 struct vm_area_struct **vmas,
851 int *locked,
852 unsigned int flags)
854 long ret, pages_done;
855 bool lock_dropped;
857 if (locked) {
858 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
859 BUG_ON(vmas);
860 /* check caller initialized locked */
861 BUG_ON(*locked != 1);
864 if (pages)
865 flags |= FOLL_GET;
867 pages_done = 0;
868 lock_dropped = false;
869 for (;;) {
870 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
871 vmas, locked);
872 if (!locked)
873 /* VM_FAULT_RETRY couldn't trigger, bypass */
874 return ret;
876 /* VM_FAULT_RETRY cannot return errors */
877 if (!*locked) {
878 BUG_ON(ret < 0);
879 BUG_ON(ret >= nr_pages);
882 if (!pages)
883 /* If it's a prefault don't insist harder */
884 return ret;
886 if (ret > 0) {
887 nr_pages -= ret;
888 pages_done += ret;
889 if (!nr_pages)
890 break;
892 if (*locked) {
893 /* VM_FAULT_RETRY didn't trigger */
894 if (!pages_done)
895 pages_done = ret;
896 break;
898 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
899 pages += ret;
900 start += ret << PAGE_SHIFT;
903 * Repeat on the address that fired VM_FAULT_RETRY
904 * without FAULT_FLAG_ALLOW_RETRY but with
905 * FAULT_FLAG_TRIED.
907 *locked = 1;
908 lock_dropped = true;
909 down_read(&mm->mmap_sem);
910 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
911 pages, NULL, NULL);
912 if (ret != 1) {
913 BUG_ON(ret > 1);
914 if (!pages_done)
915 pages_done = ret;
916 break;
918 nr_pages--;
919 pages_done++;
920 if (!nr_pages)
921 break;
922 pages++;
923 start += PAGE_SIZE;
925 if (lock_dropped && *locked) {
927 * We must let the caller know we temporarily dropped the lock
928 * and so the critical section protected by it was lost.
930 up_read(&mm->mmap_sem);
931 *locked = 0;
933 return pages_done;
937 * We can leverage the VM_FAULT_RETRY functionality in the page fault
938 * paths better by using either get_user_pages_locked() or
939 * get_user_pages_unlocked().
941 * get_user_pages_locked() is suitable to replace the form:
943 * down_read(&mm->mmap_sem);
944 * do_something()
945 * get_user_pages(tsk, mm, ..., pages, NULL);
946 * up_read(&mm->mmap_sem);
948 * to:
950 * int locked = 1;
951 * down_read(&mm->mmap_sem);
952 * do_something()
953 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
954 * if (locked)
955 * up_read(&mm->mmap_sem);
957 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
958 unsigned int gup_flags, struct page **pages,
959 int *locked)
961 return __get_user_pages_locked(current, current->mm, start, nr_pages,
962 pages, NULL, locked,
963 gup_flags | FOLL_TOUCH);
965 EXPORT_SYMBOL(get_user_pages_locked);
968 * get_user_pages_unlocked() is suitable to replace the form:
970 * down_read(&mm->mmap_sem);
971 * get_user_pages(tsk, mm, ..., pages, NULL);
972 * up_read(&mm->mmap_sem);
974 * with:
976 * get_user_pages_unlocked(tsk, mm, ..., pages);
978 * It is functionally equivalent to get_user_pages_fast so
979 * get_user_pages_fast should be used instead if specific gup_flags
980 * (e.g. FOLL_FORCE) are not required.
982 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
983 struct page **pages, unsigned int gup_flags)
985 struct mm_struct *mm = current->mm;
986 int locked = 1;
987 long ret;
989 down_read(&mm->mmap_sem);
990 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
991 &locked, gup_flags | FOLL_TOUCH);
992 if (locked)
993 up_read(&mm->mmap_sem);
994 return ret;
996 EXPORT_SYMBOL(get_user_pages_unlocked);
999 * get_user_pages_remote() - pin user pages in memory
1000 * @tsk: the task_struct to use for page fault accounting, or
1001 * NULL if faults are not to be recorded.
1002 * @mm: mm_struct of target mm
1003 * @start: starting user address
1004 * @nr_pages: number of pages from start to pin
1005 * @gup_flags: flags modifying lookup behaviour
1006 * @pages: array that receives pointers to the pages pinned.
1007 * Should be at least nr_pages long. Or NULL, if caller
1008 * only intends to ensure the pages are faulted in.
1009 * @vmas: array of pointers to vmas corresponding to each page.
1010 * Or NULL if the caller does not require them.
1011 * @locked: pointer to lock flag indicating whether lock is held and
1012 * subsequently whether VM_FAULT_RETRY functionality can be
1013 * utilised. Lock must initially be held.
1015 * Returns number of pages pinned. This may be fewer than the number
1016 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1017 * were pinned, returns -errno. Each page returned must be released
1018 * with a put_page() call when it is finished with. vmas will only
1019 * remain valid while mmap_sem is held.
1021 * Must be called with mmap_sem held for read or write.
1023 * get_user_pages walks a process's page tables and takes a reference to
1024 * each struct page that each user address corresponds to at a given
1025 * instant. That is, it takes the page that would be accessed if a user
1026 * thread accesses the given user virtual address at that instant.
1028 * This does not guarantee that the page exists in the user mappings when
1029 * get_user_pages returns, and there may even be a completely different
1030 * page there in some cases (eg. if mmapped pagecache has been invalidated
1031 * and subsequently re faulted). However it does guarantee that the page
1032 * won't be freed completely. And mostly callers simply care that the page
1033 * contains data that was valid *at some point in time*. Typically, an IO
1034 * or similar operation cannot guarantee anything stronger anyway because
1035 * locks can't be held over the syscall boundary.
1037 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1038 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1039 * be called after the page is finished with, and before put_page is called.
1041 * get_user_pages is typically used for fewer-copy IO operations, to get a
1042 * handle on the memory by some means other than accesses via the user virtual
1043 * addresses. The pages may be submitted for DMA to devices or accessed via
1044 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1045 * use the correct cache flushing APIs.
1047 * See also get_user_pages_fast, for performance critical applications.
1049 * get_user_pages should be phased out in favor of
1050 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1051 * should use get_user_pages because it cannot pass
1052 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1054 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1055 unsigned long start, unsigned long nr_pages,
1056 unsigned int gup_flags, struct page **pages,
1057 struct vm_area_struct **vmas, int *locked)
1059 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1060 locked,
1061 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1063 EXPORT_SYMBOL(get_user_pages_remote);
1066 * This is the same as get_user_pages_remote(), just with a
1067 * less-flexible calling convention where we assume that the task
1068 * and mm being operated on are the current task's and don't allow
1069 * passing of a locked parameter. We also obviously don't pass
1070 * FOLL_REMOTE in here.
1072 long get_user_pages(unsigned long start, unsigned long nr_pages,
1073 unsigned int gup_flags, struct page **pages,
1074 struct vm_area_struct **vmas)
1076 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1077 pages, vmas, NULL,
1078 gup_flags | FOLL_TOUCH);
1080 EXPORT_SYMBOL(get_user_pages);
1082 #ifdef CONFIG_FS_DAX
1084 * This is the same as get_user_pages() in that it assumes we are
1085 * operating on the current task's mm, but it goes further to validate
1086 * that the vmas associated with the address range are suitable for
1087 * longterm elevated page reference counts. For example, filesystem-dax
1088 * mappings are subject to the lifetime enforced by the filesystem and
1089 * we need guarantees that longterm users like RDMA and V4L2 only
1090 * establish mappings that have a kernel enforced revocation mechanism.
1092 * "longterm" == userspace controlled elevated page count lifetime.
1093 * Contrast this to iov_iter_get_pages() usages which are transient.
1095 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1096 unsigned int gup_flags, struct page **pages,
1097 struct vm_area_struct **vmas_arg)
1099 struct vm_area_struct **vmas = vmas_arg;
1100 struct vm_area_struct *vma_prev = NULL;
1101 long rc, i;
1103 if (!pages)
1104 return -EINVAL;
1106 if (!vmas) {
1107 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1108 GFP_KERNEL);
1109 if (!vmas)
1110 return -ENOMEM;
1113 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1115 for (i = 0; i < rc; i++) {
1116 struct vm_area_struct *vma = vmas[i];
1118 if (vma == vma_prev)
1119 continue;
1121 vma_prev = vma;
1123 if (vma_is_fsdax(vma))
1124 break;
1128 * Either get_user_pages() failed, or the vma validation
1129 * succeeded, in either case we don't need to put_page() before
1130 * returning.
1132 if (i >= rc)
1133 goto out;
1135 for (i = 0; i < rc; i++)
1136 put_page(pages[i]);
1137 rc = -EOPNOTSUPP;
1138 out:
1139 if (vmas != vmas_arg)
1140 kfree(vmas);
1141 return rc;
1143 EXPORT_SYMBOL(get_user_pages_longterm);
1144 #endif /* CONFIG_FS_DAX */
1147 * populate_vma_page_range() - populate a range of pages in the vma.
1148 * @vma: target vma
1149 * @start: start address
1150 * @end: end address
1151 * @nonblocking:
1153 * This takes care of mlocking the pages too if VM_LOCKED is set.
1155 * return 0 on success, negative error code on error.
1157 * vma->vm_mm->mmap_sem must be held.
1159 * If @nonblocking is NULL, it may be held for read or write and will
1160 * be unperturbed.
1162 * If @nonblocking is non-NULL, it must held for read only and may be
1163 * released. If it's released, *@nonblocking will be set to 0.
1165 long populate_vma_page_range(struct vm_area_struct *vma,
1166 unsigned long start, unsigned long end, int *nonblocking)
1168 struct mm_struct *mm = vma->vm_mm;
1169 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1170 int gup_flags;
1172 VM_BUG_ON(start & ~PAGE_MASK);
1173 VM_BUG_ON(end & ~PAGE_MASK);
1174 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1175 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1176 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1178 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1179 if (vma->vm_flags & VM_LOCKONFAULT)
1180 gup_flags &= ~FOLL_POPULATE;
1182 * We want to touch writable mappings with a write fault in order
1183 * to break COW, except for shared mappings because these don't COW
1184 * and we would not want to dirty them for nothing.
1186 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1187 gup_flags |= FOLL_WRITE;
1190 * We want mlock to succeed for regions that have any permissions
1191 * other than PROT_NONE.
1193 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1194 gup_flags |= FOLL_FORCE;
1197 * We made sure addr is within a VMA, so the following will
1198 * not result in a stack expansion that recurses back here.
1200 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1201 NULL, NULL, nonblocking);
1205 * __mm_populate - populate and/or mlock pages within a range of address space.
1207 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1208 * flags. VMAs must be already marked with the desired vm_flags, and
1209 * mmap_sem must not be held.
1211 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1213 struct mm_struct *mm = current->mm;
1214 unsigned long end, nstart, nend;
1215 struct vm_area_struct *vma = NULL;
1216 int locked = 0;
1217 long ret = 0;
1219 VM_BUG_ON(start & ~PAGE_MASK);
1220 VM_BUG_ON(len != PAGE_ALIGN(len));
1221 end = start + len;
1223 for (nstart = start; nstart < end; nstart = nend) {
1225 * We want to fault in pages for [nstart; end) address range.
1226 * Find first corresponding VMA.
1228 if (!locked) {
1229 locked = 1;
1230 down_read(&mm->mmap_sem);
1231 vma = find_vma(mm, nstart);
1232 } else if (nstart >= vma->vm_end)
1233 vma = vma->vm_next;
1234 if (!vma || vma->vm_start >= end)
1235 break;
1237 * Set [nstart; nend) to intersection of desired address
1238 * range with the first VMA. Also, skip undesirable VMA types.
1240 nend = min(end, vma->vm_end);
1241 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1242 continue;
1243 if (nstart < vma->vm_start)
1244 nstart = vma->vm_start;
1246 * Now fault in a range of pages. populate_vma_page_range()
1247 * double checks the vma flags, so that it won't mlock pages
1248 * if the vma was already munlocked.
1250 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1251 if (ret < 0) {
1252 if (ignore_errors) {
1253 ret = 0;
1254 continue; /* continue at next VMA */
1256 break;
1258 nend = nstart + ret * PAGE_SIZE;
1259 ret = 0;
1261 if (locked)
1262 up_read(&mm->mmap_sem);
1263 return ret; /* 0 or negative error code */
1267 * get_dump_page() - pin user page in memory while writing it to core dump
1268 * @addr: user address
1270 * Returns struct page pointer of user page pinned for dump,
1271 * to be freed afterwards by put_page().
1273 * Returns NULL on any kind of failure - a hole must then be inserted into
1274 * the corefile, to preserve alignment with its headers; and also returns
1275 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1276 * allowing a hole to be left in the corefile to save diskspace.
1278 * Called without mmap_sem, but after all other threads have been killed.
1280 #ifdef CONFIG_ELF_CORE
1281 struct page *get_dump_page(unsigned long addr)
1283 struct vm_area_struct *vma;
1284 struct page *page;
1286 if (__get_user_pages(current, current->mm, addr, 1,
1287 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1288 NULL) < 1)
1289 return NULL;
1290 flush_cache_page(vma, addr, page_to_pfn(page));
1291 return page;
1293 #endif /* CONFIG_ELF_CORE */
1296 * Generic Fast GUP
1298 * get_user_pages_fast attempts to pin user pages by walking the page
1299 * tables directly and avoids taking locks. Thus the walker needs to be
1300 * protected from page table pages being freed from under it, and should
1301 * block any THP splits.
1303 * One way to achieve this is to have the walker disable interrupts, and
1304 * rely on IPIs from the TLB flushing code blocking before the page table
1305 * pages are freed. This is unsuitable for architectures that do not need
1306 * to broadcast an IPI when invalidating TLBs.
1308 * Another way to achieve this is to batch up page table containing pages
1309 * belonging to more than one mm_user, then rcu_sched a callback to free those
1310 * pages. Disabling interrupts will allow the fast_gup walker to both block
1311 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1312 * (which is a relatively rare event). The code below adopts this strategy.
1314 * Before activating this code, please be aware that the following assumptions
1315 * are currently made:
1317 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1318 * free pages containing page tables or TLB flushing requires IPI broadcast.
1320 * *) ptes can be read atomically by the architecture.
1322 * *) access_ok is sufficient to validate userspace address ranges.
1324 * The last two assumptions can be relaxed by the addition of helper functions.
1326 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1328 #ifdef CONFIG_HAVE_GENERIC_GUP
1330 #ifndef gup_get_pte
1332 * We assume that the PTE can be read atomically. If this is not the case for
1333 * your architecture, please provide the helper.
1335 static inline pte_t gup_get_pte(pte_t *ptep)
1337 return READ_ONCE(*ptep);
1339 #endif
1341 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1343 while ((*nr) - nr_start) {
1344 struct page *page = pages[--(*nr)];
1346 ClearPageReferenced(page);
1347 put_page(page);
1351 #ifdef __HAVE_ARCH_PTE_SPECIAL
1352 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1353 int write, struct page **pages, int *nr)
1355 struct dev_pagemap *pgmap = NULL;
1356 int nr_start = *nr, ret = 0;
1357 pte_t *ptep, *ptem;
1359 ptem = ptep = pte_offset_map(&pmd, addr);
1360 do {
1361 pte_t pte = gup_get_pte(ptep);
1362 struct page *head, *page;
1365 * Similar to the PMD case below, NUMA hinting must take slow
1366 * path using the pte_protnone check.
1368 if (pte_protnone(pte))
1369 goto pte_unmap;
1371 if (!pte_access_permitted(pte, write))
1372 goto pte_unmap;
1374 if (pte_devmap(pte)) {
1375 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1376 if (unlikely(!pgmap)) {
1377 undo_dev_pagemap(nr, nr_start, pages);
1378 goto pte_unmap;
1380 } else if (pte_special(pte))
1381 goto pte_unmap;
1383 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1384 page = pte_page(pte);
1385 head = compound_head(page);
1387 if (!page_cache_get_speculative(head))
1388 goto pte_unmap;
1390 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1391 put_page(head);
1392 goto pte_unmap;
1395 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1397 SetPageReferenced(page);
1398 pages[*nr] = page;
1399 (*nr)++;
1401 } while (ptep++, addr += PAGE_SIZE, addr != end);
1403 ret = 1;
1405 pte_unmap:
1406 if (pgmap)
1407 put_dev_pagemap(pgmap);
1408 pte_unmap(ptem);
1409 return ret;
1411 #else
1414 * If we can't determine whether or not a pte is special, then fail immediately
1415 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1416 * to be special.
1418 * For a futex to be placed on a THP tail page, get_futex_key requires a
1419 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1420 * useful to have gup_huge_pmd even if we can't operate on ptes.
1422 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1423 int write, struct page **pages, int *nr)
1425 return 0;
1427 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1429 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1430 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1431 unsigned long end, struct page **pages, int *nr)
1433 int nr_start = *nr;
1434 struct dev_pagemap *pgmap = NULL;
1436 do {
1437 struct page *page = pfn_to_page(pfn);
1439 pgmap = get_dev_pagemap(pfn, pgmap);
1440 if (unlikely(!pgmap)) {
1441 undo_dev_pagemap(nr, nr_start, pages);
1442 return 0;
1444 SetPageReferenced(page);
1445 pages[*nr] = page;
1446 get_page(page);
1447 (*nr)++;
1448 pfn++;
1449 } while (addr += PAGE_SIZE, addr != end);
1451 if (pgmap)
1452 put_dev_pagemap(pgmap);
1453 return 1;
1456 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1457 unsigned long end, struct page **pages, int *nr)
1459 unsigned long fault_pfn;
1461 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1462 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1465 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1466 unsigned long end, struct page **pages, int *nr)
1468 unsigned long fault_pfn;
1470 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1471 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1473 #else
1474 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1475 unsigned long end, struct page **pages, int *nr)
1477 BUILD_BUG();
1478 return 0;
1481 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1482 unsigned long end, struct page **pages, int *nr)
1484 BUILD_BUG();
1485 return 0;
1487 #endif
1489 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1490 unsigned long end, int write, struct page **pages, int *nr)
1492 struct page *head, *page;
1493 int refs;
1495 if (!pmd_access_permitted(orig, write))
1496 return 0;
1498 if (pmd_devmap(orig))
1499 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1501 refs = 0;
1502 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1503 do {
1504 pages[*nr] = page;
1505 (*nr)++;
1506 page++;
1507 refs++;
1508 } while (addr += PAGE_SIZE, addr != end);
1510 head = compound_head(pmd_page(orig));
1511 if (!page_cache_add_speculative(head, refs)) {
1512 *nr -= refs;
1513 return 0;
1516 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1517 *nr -= refs;
1518 while (refs--)
1519 put_page(head);
1520 return 0;
1523 SetPageReferenced(head);
1524 return 1;
1527 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1528 unsigned long end, int write, struct page **pages, int *nr)
1530 struct page *head, *page;
1531 int refs;
1533 if (!pud_access_permitted(orig, write))
1534 return 0;
1536 if (pud_devmap(orig))
1537 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1539 refs = 0;
1540 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1541 do {
1542 pages[*nr] = page;
1543 (*nr)++;
1544 page++;
1545 refs++;
1546 } while (addr += PAGE_SIZE, addr != end);
1548 head = compound_head(pud_page(orig));
1549 if (!page_cache_add_speculative(head, refs)) {
1550 *nr -= refs;
1551 return 0;
1554 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1555 *nr -= refs;
1556 while (refs--)
1557 put_page(head);
1558 return 0;
1561 SetPageReferenced(head);
1562 return 1;
1565 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1566 unsigned long end, int write,
1567 struct page **pages, int *nr)
1569 int refs;
1570 struct page *head, *page;
1572 if (!pgd_access_permitted(orig, write))
1573 return 0;
1575 BUILD_BUG_ON(pgd_devmap(orig));
1576 refs = 0;
1577 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1578 do {
1579 pages[*nr] = page;
1580 (*nr)++;
1581 page++;
1582 refs++;
1583 } while (addr += PAGE_SIZE, addr != end);
1585 head = compound_head(pgd_page(orig));
1586 if (!page_cache_add_speculative(head, refs)) {
1587 *nr -= refs;
1588 return 0;
1591 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1592 *nr -= refs;
1593 while (refs--)
1594 put_page(head);
1595 return 0;
1598 SetPageReferenced(head);
1599 return 1;
1602 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1603 int write, struct page **pages, int *nr)
1605 unsigned long next;
1606 pmd_t *pmdp;
1608 pmdp = pmd_offset(&pud, addr);
1609 do {
1610 pmd_t pmd = READ_ONCE(*pmdp);
1612 next = pmd_addr_end(addr, end);
1613 if (!pmd_present(pmd))
1614 return 0;
1616 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1618 * NUMA hinting faults need to be handled in the GUP
1619 * slowpath for accounting purposes and so that they
1620 * can be serialised against THP migration.
1622 if (pmd_protnone(pmd))
1623 return 0;
1625 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1626 pages, nr))
1627 return 0;
1629 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1631 * architecture have different format for hugetlbfs
1632 * pmd format and THP pmd format
1634 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1635 PMD_SHIFT, next, write, pages, nr))
1636 return 0;
1637 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1638 return 0;
1639 } while (pmdp++, addr = next, addr != end);
1641 return 1;
1644 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1645 int write, struct page **pages, int *nr)
1647 unsigned long next;
1648 pud_t *pudp;
1650 pudp = pud_offset(&p4d, addr);
1651 do {
1652 pud_t pud = READ_ONCE(*pudp);
1654 next = pud_addr_end(addr, end);
1655 if (pud_none(pud))
1656 return 0;
1657 if (unlikely(pud_huge(pud))) {
1658 if (!gup_huge_pud(pud, pudp, addr, next, write,
1659 pages, nr))
1660 return 0;
1661 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1662 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1663 PUD_SHIFT, next, write, pages, nr))
1664 return 0;
1665 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1666 return 0;
1667 } while (pudp++, addr = next, addr != end);
1669 return 1;
1672 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1673 int write, struct page **pages, int *nr)
1675 unsigned long next;
1676 p4d_t *p4dp;
1678 p4dp = p4d_offset(&pgd, addr);
1679 do {
1680 p4d_t p4d = READ_ONCE(*p4dp);
1682 next = p4d_addr_end(addr, end);
1683 if (p4d_none(p4d))
1684 return 0;
1685 BUILD_BUG_ON(p4d_huge(p4d));
1686 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1687 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1688 P4D_SHIFT, next, write, pages, nr))
1689 return 0;
1690 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1691 return 0;
1692 } while (p4dp++, addr = next, addr != end);
1694 return 1;
1697 static void gup_pgd_range(unsigned long addr, unsigned long end,
1698 int write, struct page **pages, int *nr)
1700 unsigned long next;
1701 pgd_t *pgdp;
1703 pgdp = pgd_offset(current->mm, addr);
1704 do {
1705 pgd_t pgd = READ_ONCE(*pgdp);
1707 next = pgd_addr_end(addr, end);
1708 if (pgd_none(pgd))
1709 return;
1710 if (unlikely(pgd_huge(pgd))) {
1711 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1712 pages, nr))
1713 return;
1714 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1715 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1716 PGDIR_SHIFT, next, write, pages, nr))
1717 return;
1718 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1719 return;
1720 } while (pgdp++, addr = next, addr != end);
1723 #ifndef gup_fast_permitted
1725 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1726 * we need to fall back to the slow version:
1728 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1730 unsigned long len, end;
1732 len = (unsigned long) nr_pages << PAGE_SHIFT;
1733 end = start + len;
1734 return end >= start;
1736 #endif
1739 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1740 * the regular GUP. It will only return non-negative values.
1742 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1743 struct page **pages)
1745 unsigned long addr, len, end;
1746 unsigned long flags;
1747 int nr = 0;
1749 start &= PAGE_MASK;
1750 addr = start;
1751 len = (unsigned long) nr_pages << PAGE_SHIFT;
1752 end = start + len;
1754 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1755 (void __user *)start, len)))
1756 return 0;
1759 * Disable interrupts. We use the nested form as we can already have
1760 * interrupts disabled by get_futex_key.
1762 * With interrupts disabled, we block page table pages from being
1763 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1764 * for more details.
1766 * We do not adopt an rcu_read_lock(.) here as we also want to
1767 * block IPIs that come from THPs splitting.
1770 if (gup_fast_permitted(start, nr_pages, write)) {
1771 local_irq_save(flags);
1772 gup_pgd_range(addr, end, write, pages, &nr);
1773 local_irq_restore(flags);
1776 return nr;
1780 * get_user_pages_fast() - pin user pages in memory
1781 * @start: starting user address
1782 * @nr_pages: number of pages from start to pin
1783 * @write: whether pages will be written to
1784 * @pages: array that receives pointers to the pages pinned.
1785 * Should be at least nr_pages long.
1787 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1788 * If not successful, it will fall back to taking the lock and
1789 * calling get_user_pages().
1791 * Returns number of pages pinned. This may be fewer than the number
1792 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1793 * were pinned, returns -errno.
1795 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1796 struct page **pages)
1798 unsigned long addr, len, end;
1799 int nr = 0, ret = 0;
1801 start &= PAGE_MASK;
1802 addr = start;
1803 len = (unsigned long) nr_pages << PAGE_SHIFT;
1804 end = start + len;
1806 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1807 (void __user *)start, len)))
1808 return 0;
1810 if (gup_fast_permitted(start, nr_pages, write)) {
1811 local_irq_disable();
1812 gup_pgd_range(addr, end, write, pages, &nr);
1813 local_irq_enable();
1814 ret = nr;
1817 if (nr < nr_pages) {
1818 /* Try to get the remaining pages with get_user_pages */
1819 start += nr << PAGE_SHIFT;
1820 pages += nr;
1822 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1823 write ? FOLL_WRITE : 0);
1825 /* Have to be a bit careful with return values */
1826 if (nr > 0) {
1827 if (ret < 0)
1828 ret = nr;
1829 else
1830 ret += nr;
1834 return ret;
1837 #endif /* CONFIG_HAVE_GENERIC_GUP */