1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
6 #include <linux/hugetlb.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/swapops.h>
15 static struct page *no_page_table(struct vm_area_struct *vma,
19 * When core dumping an enormous anonymous area that nobody
20 * has touched so far, we don't want to allocate unnecessary pages or
21 * page tables. Return error instead of NULL to skip handle_mm_fault,
22 * then get_dump_page() will return NULL to leave a hole in the dump.
23 * But we can only make this optimization where a hole would surely
24 * be zero-filled if handle_mm_fault() actually did handle it.
26 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
27 return ERR_PTR(-EFAULT);
31 static struct page *follow_page_pte(struct vm_area_struct *vma,
32 unsigned long address, pmd_t *pmd, unsigned int flags)
34 struct mm_struct *mm = vma->vm_mm;
40 if (unlikely(pmd_bad(*pmd)))
41 return no_page_table(vma, flags);
43 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
45 if (!pte_present(pte)) {
48 * KSM's break_ksm() relies upon recognizing a ksm page
49 * even while it is being migrated, so for that case we
50 * need migration_entry_wait().
52 if (likely(!(flags & FOLL_MIGRATION)))
54 if (pte_none(pte) || pte_file(pte))
56 entry = pte_to_swp_entry(pte);
57 if (!is_migration_entry(entry))
59 pte_unmap_unlock(ptep, ptl);
60 migration_entry_wait(mm, pmd, address);
63 if ((flags & FOLL_NUMA) && pte_numa(pte))
65 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
66 pte_unmap_unlock(ptep, ptl);
70 page = vm_normal_page(vma, address, pte);
71 if (unlikely(!page)) {
72 if ((flags & FOLL_DUMP) ||
73 !is_zero_pfn(pte_pfn(pte)))
80 if (flags & FOLL_TOUCH) {
81 if ((flags & FOLL_WRITE) &&
82 !pte_dirty(pte) && !PageDirty(page))
85 * pte_mkyoung() would be more correct here, but atomic care
86 * is needed to avoid losing the dirty bit: it is easier to use
87 * mark_page_accessed().
89 mark_page_accessed(page);
91 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
93 * The preliminary mapping check is mainly to avoid the
94 * pointless overhead of lock_page on the ZERO_PAGE
95 * which might bounce very badly if there is contention.
97 * If the page is already locked, we don't need to
98 * handle it now - vmscan will handle it later if and
99 * when it attempts to reclaim the page.
101 if (page->mapping && trylock_page(page)) {
102 lru_add_drain(); /* push cached pages to LRU */
104 * Because we lock page here, and migration is
105 * blocked by the pte's page reference, and we
106 * know the page is still mapped, we don't even
107 * need to check for file-cache page truncation.
109 mlock_vma_page(page);
113 pte_unmap_unlock(ptep, ptl);
116 pte_unmap_unlock(ptep, ptl);
117 return ERR_PTR(-EFAULT);
120 pte_unmap_unlock(ptep, ptl);
123 return no_page_table(vma, flags);
127 * follow_page_mask - look up a page descriptor from a user-virtual address
128 * @vma: vm_area_struct mapping @address
129 * @address: virtual address to look up
130 * @flags: flags modifying lookup behaviour
131 * @page_mask: on output, *page_mask is set according to the size of the page
133 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
135 * Returns the mapped (struct page *), %NULL if no mapping exists, or
136 * an error pointer if there is a mapping to something not represented
137 * by a page descriptor (see also vm_normal_page()).
139 struct page *follow_page_mask(struct vm_area_struct *vma,
140 unsigned long address, unsigned int flags,
141 unsigned int *page_mask)
148 struct mm_struct *mm = vma->vm_mm;
152 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
154 BUG_ON(flags & FOLL_GET);
158 pgd = pgd_offset(mm, address);
159 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
160 return no_page_table(vma, flags);
162 pud = pud_offset(pgd, address);
164 return no_page_table(vma, flags);
165 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
166 if (flags & FOLL_GET)
168 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
171 if (unlikely(pud_bad(*pud)))
172 return no_page_table(vma, flags);
174 pmd = pmd_offset(pud, address);
176 return no_page_table(vma, flags);
177 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
178 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
179 if (flags & FOLL_GET) {
181 * Refcount on tail pages are not well-defined and
182 * shouldn't be taken. The caller should handle a NULL
183 * return when trying to follow tail pages.
192 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
193 return no_page_table(vma, flags);
194 if (pmd_trans_huge(*pmd)) {
195 if (flags & FOLL_SPLIT) {
196 split_huge_page_pmd(vma, address, pmd);
197 return follow_page_pte(vma, address, pmd, flags);
199 ptl = pmd_lock(mm, pmd);
200 if (likely(pmd_trans_huge(*pmd))) {
201 if (unlikely(pmd_trans_splitting(*pmd))) {
203 wait_split_huge_page(vma->anon_vma, pmd);
205 page = follow_trans_huge_pmd(vma, address,
208 *page_mask = HPAGE_PMD_NR - 1;
214 return follow_page_pte(vma, address, pmd, flags);
217 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
219 return stack_guard_page_start(vma, addr) ||
220 stack_guard_page_end(vma, addr+PAGE_SIZE);
223 static int get_gate_page(struct mm_struct *mm, unsigned long address,
224 unsigned int gup_flags, struct vm_area_struct **vma,
233 /* user gate pages are read-only */
234 if (gup_flags & FOLL_WRITE)
236 if (address > TASK_SIZE)
237 pgd = pgd_offset_k(address);
239 pgd = pgd_offset_gate(mm, address);
240 BUG_ON(pgd_none(*pgd));
241 pud = pud_offset(pgd, address);
242 BUG_ON(pud_none(*pud));
243 pmd = pmd_offset(pud, address);
246 VM_BUG_ON(pmd_trans_huge(*pmd));
247 pte = pte_offset_map(pmd, address);
250 *vma = get_gate_vma(mm);
253 *page = vm_normal_page(*vma, address, *pte);
255 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
257 *page = pte_page(*pte);
268 * __get_user_pages() - pin user pages in memory
269 * @tsk: task_struct of target task
270 * @mm: mm_struct of target mm
271 * @start: starting user address
272 * @nr_pages: number of pages from start to pin
273 * @gup_flags: flags modifying pin behaviour
274 * @pages: array that receives pointers to the pages pinned.
275 * Should be at least nr_pages long. Or NULL, if caller
276 * only intends to ensure the pages are faulted in.
277 * @vmas: array of pointers to vmas corresponding to each page.
278 * Or NULL if the caller does not require them.
279 * @nonblocking: whether waiting for disk IO or mmap_sem contention
281 * Returns number of pages pinned. This may be fewer than the number
282 * requested. If nr_pages is 0 or negative, returns 0. If no pages
283 * were pinned, returns -errno. Each page returned must be released
284 * with a put_page() call when it is finished with. vmas will only
285 * remain valid while mmap_sem is held.
287 * Must be called with mmap_sem held for read or write.
289 * __get_user_pages walks a process's page tables and takes a reference to
290 * each struct page that each user address corresponds to at a given
291 * instant. That is, it takes the page that would be accessed if a user
292 * thread accesses the given user virtual address at that instant.
294 * This does not guarantee that the page exists in the user mappings when
295 * __get_user_pages returns, and there may even be a completely different
296 * page there in some cases (eg. if mmapped pagecache has been invalidated
297 * and subsequently re faulted). However it does guarantee that the page
298 * won't be freed completely. And mostly callers simply care that the page
299 * contains data that was valid *at some point in time*. Typically, an IO
300 * or similar operation cannot guarantee anything stronger anyway because
301 * locks can't be held over the syscall boundary.
303 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
304 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
305 * appropriate) must be called after the page is finished with, and
306 * before put_page is called.
308 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
309 * or mmap_sem contention, and if waiting is needed to pin all pages,
310 * *@nonblocking will be set to 0.
312 * In most cases, get_user_pages or get_user_pages_fast should be used
313 * instead of __get_user_pages. __get_user_pages should be used only if
314 * you need some special @gup_flags.
316 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
317 unsigned long start, unsigned long nr_pages,
318 unsigned int gup_flags, struct page **pages,
319 struct vm_area_struct **vmas, int *nonblocking)
322 unsigned long vm_flags;
323 unsigned int page_mask;
328 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
331 * If FOLL_FORCE is set then do not force a full fault as the hinting
332 * fault information is unrelated to the reference behaviour of a task
333 * using the address space
335 if (!(gup_flags & FOLL_FORCE))
336 gup_flags |= FOLL_NUMA;
341 struct vm_area_struct *vma;
343 vma = find_extend_vma(mm, start);
344 if (!vma && in_gate_area(mm, start)) {
346 ret = get_gate_page(mm, start & PAGE_MASK, gup_flags,
347 &vma, pages ? &pages[i] : NULL);
356 vm_flags = vma->vm_flags;
357 if (vm_flags & (VM_IO | VM_PFNMAP))
360 if (gup_flags & FOLL_WRITE) {
361 if (!(vm_flags & VM_WRITE)) {
362 if (!(gup_flags & FOLL_FORCE))
365 * We used to let the write,force case do COW
366 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
367 * ptrace could set a breakpoint in a read-only
368 * mapping of an executable, without corrupting
369 * the file (yet only when that file had been
370 * opened for writing!). Anon pages in shared
371 * mappings are surprising: now just reject it.
373 if (!is_cow_mapping(vm_flags)) {
374 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
379 if (!(vm_flags & VM_READ)) {
380 if (!(gup_flags & FOLL_FORCE))
383 * Is there actually any vma we can reach here
384 * which does not have VM_MAYREAD set?
386 if (!(vm_flags & VM_MAYREAD))
391 if (is_vm_hugetlb_page(vma)) {
392 i = follow_hugetlb_page(mm, vma, pages, vmas,
393 &start, &nr_pages, i, gup_flags);
399 unsigned int foll_flags = gup_flags;
400 unsigned int page_increm;
403 * If we have a pending SIGKILL, don't keep faulting
404 * pages and potentially allocating memory.
406 if (unlikely(fatal_signal_pending(current)))
407 return i ? i : -ERESTARTSYS;
410 while (!(page = follow_page_mask(vma, start,
411 foll_flags, &page_mask))) {
413 unsigned int fault_flags = 0;
415 /* For mlock, just skip the stack guard page. */
416 if (foll_flags & FOLL_MLOCK) {
417 if (stack_guard_page(vma, start))
420 if (foll_flags & FOLL_WRITE)
421 fault_flags |= FAULT_FLAG_WRITE;
423 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
424 if (foll_flags & FOLL_NOWAIT)
425 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
427 ret = handle_mm_fault(mm, vma, start,
430 if (ret & VM_FAULT_ERROR) {
431 if (ret & VM_FAULT_OOM)
432 return i ? i : -ENOMEM;
433 if (ret & (VM_FAULT_HWPOISON |
434 VM_FAULT_HWPOISON_LARGE)) {
437 else if (gup_flags & FOLL_HWPOISON)
442 if (ret & VM_FAULT_SIGBUS)
448 if (ret & VM_FAULT_MAJOR)
454 if (ret & VM_FAULT_RETRY) {
461 * The VM_FAULT_WRITE bit tells us that
462 * do_wp_page has broken COW when necessary,
463 * even if maybe_mkwrite decided not to set
464 * pte_write. We can thus safely do subsequent
465 * page lookups as if they were reads. But only
466 * do so when looping for pte_write is futile:
467 * in some cases userspace may also be wanting
468 * to write to the gotten user page, which a
469 * read fault here might prevent (a readonly
470 * page might get reCOWed by userspace write).
472 if ((ret & VM_FAULT_WRITE) &&
473 !(vma->vm_flags & VM_WRITE))
474 foll_flags &= ~FOLL_WRITE;
479 return i ? i : PTR_ERR(page);
483 flush_anon_page(vma, page, start);
484 flush_dcache_page(page);
492 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
493 if (page_increm > nr_pages)
494 page_increm = nr_pages;
496 start += page_increm * PAGE_SIZE;
497 nr_pages -= page_increm;
498 } while (nr_pages && start < vma->vm_end);
502 return i ? : -EFAULT;
504 EXPORT_SYMBOL(__get_user_pages);
507 * fixup_user_fault() - manually resolve a user page fault
508 * @tsk: the task_struct to use for page fault accounting, or
509 * NULL if faults are not to be recorded.
510 * @mm: mm_struct of target mm
511 * @address: user address
512 * @fault_flags:flags to pass down to handle_mm_fault()
514 * This is meant to be called in the specific scenario where for locking reasons
515 * we try to access user memory in atomic context (within a pagefault_disable()
516 * section), this returns -EFAULT, and we want to resolve the user fault before
519 * Typically this is meant to be used by the futex code.
521 * The main difference with get_user_pages() is that this function will
522 * unconditionally call handle_mm_fault() which will in turn perform all the
523 * necessary SW fixup of the dirty and young bits in the PTE, while
524 * handle_mm_fault() only guarantees to update these in the struct page.
526 * This is important for some architectures where those bits also gate the
527 * access permission to the page because they are maintained in software. On
528 * such architectures, gup() will not be enough to make a subsequent access
531 * This should be called with the mm_sem held for read.
533 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
534 unsigned long address, unsigned int fault_flags)
536 struct vm_area_struct *vma;
540 vma = find_extend_vma(mm, address);
541 if (!vma || address < vma->vm_start)
544 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
545 if (!(vm_flags & vma->vm_flags))
548 ret = handle_mm_fault(mm, vma, address, fault_flags);
549 if (ret & VM_FAULT_ERROR) {
550 if (ret & VM_FAULT_OOM)
552 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
554 if (ret & VM_FAULT_SIGBUS)
559 if (ret & VM_FAULT_MAJOR)
568 * get_user_pages() - pin user pages in memory
569 * @tsk: the task_struct to use for page fault accounting, or
570 * NULL if faults are not to be recorded.
571 * @mm: mm_struct of target mm
572 * @start: starting user address
573 * @nr_pages: number of pages from start to pin
574 * @write: whether pages will be written to by the caller
575 * @force: whether to force access even when user mapping is currently
576 * protected (but never forces write access to shared mapping).
577 * @pages: array that receives pointers to the pages pinned.
578 * Should be at least nr_pages long. Or NULL, if caller
579 * only intends to ensure the pages are faulted in.
580 * @vmas: array of pointers to vmas corresponding to each page.
581 * Or NULL if the caller does not require them.
583 * Returns number of pages pinned. This may be fewer than the number
584 * requested. If nr_pages is 0 or negative, returns 0. If no pages
585 * were pinned, returns -errno. Each page returned must be released
586 * with a put_page() call when it is finished with. vmas will only
587 * remain valid while mmap_sem is held.
589 * Must be called with mmap_sem held for read or write.
591 * get_user_pages walks a process's page tables and takes a reference to
592 * each struct page that each user address corresponds to at a given
593 * instant. That is, it takes the page that would be accessed if a user
594 * thread accesses the given user virtual address at that instant.
596 * This does not guarantee that the page exists in the user mappings when
597 * get_user_pages returns, and there may even be a completely different
598 * page there in some cases (eg. if mmapped pagecache has been invalidated
599 * and subsequently re faulted). However it does guarantee that the page
600 * won't be freed completely. And mostly callers simply care that the page
601 * contains data that was valid *at some point in time*. Typically, an IO
602 * or similar operation cannot guarantee anything stronger anyway because
603 * locks can't be held over the syscall boundary.
605 * If write=0, the page must not be written to. If the page is written to,
606 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
607 * after the page is finished with, and before put_page is called.
609 * get_user_pages is typically used for fewer-copy IO operations, to get a
610 * handle on the memory by some means other than accesses via the user virtual
611 * addresses. The pages may be submitted for DMA to devices or accessed via
612 * their kernel linear mapping (via the kmap APIs). Care should be taken to
613 * use the correct cache flushing APIs.
615 * See also get_user_pages_fast, for performance critical applications.
617 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
618 unsigned long start, unsigned long nr_pages, int write,
619 int force, struct page **pages, struct vm_area_struct **vmas)
621 int flags = FOLL_TOUCH;
630 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
633 EXPORT_SYMBOL(get_user_pages);
636 * get_dump_page() - pin user page in memory while writing it to core dump
637 * @addr: user address
639 * Returns struct page pointer of user page pinned for dump,
640 * to be freed afterwards by page_cache_release() or put_page().
642 * Returns NULL on any kind of failure - a hole must then be inserted into
643 * the corefile, to preserve alignment with its headers; and also returns
644 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
645 * allowing a hole to be left in the corefile to save diskspace.
647 * Called without mmap_sem, but after all other threads have been killed.
649 #ifdef CONFIG_ELF_CORE
650 struct page *get_dump_page(unsigned long addr)
652 struct vm_area_struct *vma;
655 if (__get_user_pages(current, current->mm, addr, 1,
656 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
659 flush_cache_page(vma, addr, page_to_pfn(page));
662 #endif /* CONFIG_ELF_CORE */