1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
86 #include "pgalloc-track.h"
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
95 unsigned long max_mapnr;
96 EXPORT_SYMBOL(max_mapnr);
99 EXPORT_SYMBOL(mem_map);
103 * A number of key systems in x86 including ioremap() rely on the assumption
104 * that high_memory defines the upper bound on direct map memory, then end
105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
110 EXPORT_SYMBOL(high_memory);
113 * Randomize the address space (stacks, mmaps, brk, etc.).
115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116 * as ancient (libc5 based) binaries can segfault. )
118 int randomize_va_space __read_mostly =
119 #ifdef CONFIG_COMPAT_BRK
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
129 * Those arches which don't have hw access flag feature need to
130 * implement their own helper. By default, "true" means pagefault
131 * will be hit on old pte.
137 #ifndef arch_wants_old_prefaulted_pte
138 static inline bool arch_wants_old_prefaulted_pte(void)
141 * Transitioning a PTE from 'old' to 'young' can be expensive on
142 * some architectures, even if it's performed in hardware. By
143 * default, "false" means prefaulted entries will be 'young'.
149 static int __init disable_randmaps(char *s)
151 randomize_va_space = 0;
154 __setup("norandmaps", disable_randmaps);
156 unsigned long zero_pfn __read_mostly;
157 EXPORT_SYMBOL(zero_pfn);
159 unsigned long highest_memmap_pfn __read_mostly;
162 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
164 static int __init init_zero_pfn(void)
166 zero_pfn = page_to_pfn(ZERO_PAGE(0));
169 early_initcall(init_zero_pfn);
171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
173 trace_rss_stat(mm, member, count);
176 #if defined(SPLIT_RSS_COUNTING)
178 void sync_mm_rss(struct mm_struct *mm)
182 for (i = 0; i < NR_MM_COUNTERS; i++) {
183 if (current->rss_stat.count[i]) {
184 add_mm_counter(mm, i, current->rss_stat.count[i]);
185 current->rss_stat.count[i] = 0;
188 current->rss_stat.events = 0;
191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
193 struct task_struct *task = current;
195 if (likely(task->mm == mm))
196 task->rss_stat.count[member] += val;
198 add_mm_counter(mm, member, val);
200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
203 /* sync counter once per 64 page faults */
204 #define TASK_RSS_EVENTS_THRESH (64)
205 static void check_sync_rss_stat(struct task_struct *task)
207 if (unlikely(task != current))
209 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
210 sync_mm_rss(task->mm);
212 #else /* SPLIT_RSS_COUNTING */
214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
217 static void check_sync_rss_stat(struct task_struct *task)
221 #endif /* SPLIT_RSS_COUNTING */
224 * Note: this doesn't free the actual pages themselves. That
225 * has been handled earlier when unmapping all the memory regions.
227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
230 pgtable_t token = pmd_pgtable(*pmd);
232 pte_free_tlb(tlb, token, addr);
233 mm_dec_nr_ptes(tlb->mm);
236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
237 unsigned long addr, unsigned long end,
238 unsigned long floor, unsigned long ceiling)
245 pmd = pmd_offset(pud, addr);
247 next = pmd_addr_end(addr, end);
248 if (pmd_none_or_clear_bad(pmd))
250 free_pte_range(tlb, pmd, addr);
251 } while (pmd++, addr = next, addr != end);
261 if (end - 1 > ceiling - 1)
264 pmd = pmd_offset(pud, start);
266 pmd_free_tlb(tlb, pmd, start);
267 mm_dec_nr_pmds(tlb->mm);
270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
271 unsigned long addr, unsigned long end,
272 unsigned long floor, unsigned long ceiling)
279 pud = pud_offset(p4d, addr);
281 next = pud_addr_end(addr, end);
282 if (pud_none_or_clear_bad(pud))
284 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
285 } while (pud++, addr = next, addr != end);
295 if (end - 1 > ceiling - 1)
298 pud = pud_offset(p4d, start);
300 pud_free_tlb(tlb, pud, start);
301 mm_dec_nr_puds(tlb->mm);
304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
305 unsigned long addr, unsigned long end,
306 unsigned long floor, unsigned long ceiling)
313 p4d = p4d_offset(pgd, addr);
315 next = p4d_addr_end(addr, end);
316 if (p4d_none_or_clear_bad(p4d))
318 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
319 } while (p4d++, addr = next, addr != end);
325 ceiling &= PGDIR_MASK;
329 if (end - 1 > ceiling - 1)
332 p4d = p4d_offset(pgd, start);
334 p4d_free_tlb(tlb, p4d, start);
338 * This function frees user-level page tables of a process.
340 void free_pgd_range(struct mmu_gather *tlb,
341 unsigned long addr, unsigned long end,
342 unsigned long floor, unsigned long ceiling)
348 * The next few lines have given us lots of grief...
350 * Why are we testing PMD* at this top level? Because often
351 * there will be no work to do at all, and we'd prefer not to
352 * go all the way down to the bottom just to discover that.
354 * Why all these "- 1"s? Because 0 represents both the bottom
355 * of the address space and the top of it (using -1 for the
356 * top wouldn't help much: the masks would do the wrong thing).
357 * The rule is that addr 0 and floor 0 refer to the bottom of
358 * the address space, but end 0 and ceiling 0 refer to the top
359 * Comparisons need to use "end - 1" and "ceiling - 1" (though
360 * that end 0 case should be mythical).
362 * Wherever addr is brought up or ceiling brought down, we must
363 * be careful to reject "the opposite 0" before it confuses the
364 * subsequent tests. But what about where end is brought down
365 * by PMD_SIZE below? no, end can't go down to 0 there.
367 * Whereas we round start (addr) and ceiling down, by different
368 * masks at different levels, in order to test whether a table
369 * now has no other vmas using it, so can be freed, we don't
370 * bother to round floor or end up - the tests don't need that.
384 if (end - 1 > ceiling - 1)
389 * We add page table cache pages with PAGE_SIZE,
390 * (see pte_free_tlb()), flush the tlb if we need
392 tlb_change_page_size(tlb, PAGE_SIZE);
393 pgd = pgd_offset(tlb->mm, addr);
395 next = pgd_addr_end(addr, end);
396 if (pgd_none_or_clear_bad(pgd))
398 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
399 } while (pgd++, addr = next, addr != end);
402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
403 unsigned long floor, unsigned long ceiling)
406 struct vm_area_struct *next = vma->vm_next;
407 unsigned long addr = vma->vm_start;
410 * Hide vma from rmap and truncate_pagecache before freeing
413 unlink_anon_vmas(vma);
414 unlink_file_vma(vma);
416 if (is_vm_hugetlb_page(vma)) {
417 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
418 floor, next ? next->vm_start : ceiling);
421 * Optimization: gather nearby vmas into one call down
423 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
424 && !is_vm_hugetlb_page(next)) {
427 unlink_anon_vmas(vma);
428 unlink_file_vma(vma);
430 free_pgd_range(tlb, addr, vma->vm_end,
431 floor, next ? next->vm_start : ceiling);
437 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
440 pgtable_t new = pte_alloc_one(mm);
445 * Ensure all pte setup (eg. pte page lock and page clearing) are
446 * visible before the pte is made visible to other CPUs by being
447 * put into page tables.
449 * The other side of the story is the pointer chasing in the page
450 * table walking code (when walking the page table without locking;
451 * ie. most of the time). Fortunately, these data accesses consist
452 * of a chain of data-dependent loads, meaning most CPUs (alpha
453 * being the notable exception) will already guarantee loads are
454 * seen in-order. See the alpha page table accessors for the
455 * smp_rmb() barriers in page table walking code.
457 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
459 ptl = pmd_lock(mm, pmd);
460 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
462 pmd_populate(mm, pmd, new);
471 int __pte_alloc_kernel(pmd_t *pmd)
473 pte_t *new = pte_alloc_one_kernel(&init_mm);
477 smp_wmb(); /* See comment in __pte_alloc */
479 spin_lock(&init_mm.page_table_lock);
480 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
481 pmd_populate_kernel(&init_mm, pmd, new);
484 spin_unlock(&init_mm.page_table_lock);
486 pte_free_kernel(&init_mm, new);
490 static inline void init_rss_vec(int *rss)
492 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
495 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
499 if (current->mm == mm)
501 for (i = 0; i < NR_MM_COUNTERS; i++)
503 add_mm_counter(mm, i, rss[i]);
507 * This function is called to print an error when a bad pte
508 * is found. For example, we might have a PFN-mapped pte in
509 * a region that doesn't allow it.
511 * The calling function must still handle the error.
513 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
514 pte_t pte, struct page *page)
516 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
517 p4d_t *p4d = p4d_offset(pgd, addr);
518 pud_t *pud = pud_offset(p4d, addr);
519 pmd_t *pmd = pmd_offset(pud, addr);
520 struct address_space *mapping;
522 static unsigned long resume;
523 static unsigned long nr_shown;
524 static unsigned long nr_unshown;
527 * Allow a burst of 60 reports, then keep quiet for that minute;
528 * or allow a steady drip of one report per second.
530 if (nr_shown == 60) {
531 if (time_before(jiffies, resume)) {
536 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
543 resume = jiffies + 60 * HZ;
545 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
546 index = linear_page_index(vma, addr);
548 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
550 (long long)pte_val(pte), (long long)pmd_val(*pmd));
552 dump_page(page, "bad pte");
553 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
554 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
555 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
557 vma->vm_ops ? vma->vm_ops->fault : NULL,
558 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
559 mapping ? mapping->a_ops->readpage : NULL);
561 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
565 * vm_normal_page -- This function gets the "struct page" associated with a pte.
567 * "Special" mappings do not wish to be associated with a "struct page" (either
568 * it doesn't exist, or it exists but they don't want to touch it). In this
569 * case, NULL is returned here. "Normal" mappings do have a struct page.
571 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
572 * pte bit, in which case this function is trivial. Secondly, an architecture
573 * may not have a spare pte bit, which requires a more complicated scheme,
576 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
577 * special mapping (even if there are underlying and valid "struct pages").
578 * COWed pages of a VM_PFNMAP are always normal.
580 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
581 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
582 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
583 * mapping will always honor the rule
585 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
587 * And for normal mappings this is false.
589 * This restricts such mappings to be a linear translation from virtual address
590 * to pfn. To get around this restriction, we allow arbitrary mappings so long
591 * as the vma is not a COW mapping; in that case, we know that all ptes are
592 * special (because none can have been COWed).
595 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
597 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
598 * page" backing, however the difference is that _all_ pages with a struct
599 * page (that is, those where pfn_valid is true) are refcounted and considered
600 * normal pages by the VM. The disadvantage is that pages are refcounted
601 * (which can be slower and simply not an option for some PFNMAP users). The
602 * advantage is that we don't have to follow the strict linearity rule of
603 * PFNMAP mappings in order to support COWable mappings.
606 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
609 unsigned long pfn = pte_pfn(pte);
611 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
612 if (likely(!pte_special(pte)))
614 if (vma->vm_ops && vma->vm_ops->find_special_page)
615 return vma->vm_ops->find_special_page(vma, addr);
616 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
618 if (is_zero_pfn(pfn))
623 print_bad_pte(vma, addr, pte, NULL);
627 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
629 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
630 if (vma->vm_flags & VM_MIXEDMAP) {
636 off = (addr - vma->vm_start) >> PAGE_SHIFT;
637 if (pfn == vma->vm_pgoff + off)
639 if (!is_cow_mapping(vma->vm_flags))
644 if (is_zero_pfn(pfn))
648 if (unlikely(pfn > highest_memmap_pfn)) {
649 print_bad_pte(vma, addr, pte, NULL);
654 * NOTE! We still have PageReserved() pages in the page tables.
655 * eg. VDSO mappings can cause them to exist.
658 return pfn_to_page(pfn);
661 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
662 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
665 unsigned long pfn = pmd_pfn(pmd);
668 * There is no pmd_special() but there may be special pmds, e.g.
669 * in a direct-access (dax) mapping, so let's just replicate the
670 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
672 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
673 if (vma->vm_flags & VM_MIXEDMAP) {
679 off = (addr - vma->vm_start) >> PAGE_SHIFT;
680 if (pfn == vma->vm_pgoff + off)
682 if (!is_cow_mapping(vma->vm_flags))
689 if (is_huge_zero_pmd(pmd))
691 if (unlikely(pfn > highest_memmap_pfn))
695 * NOTE! We still have PageReserved() pages in the page tables.
696 * eg. VDSO mappings can cause them to exist.
699 return pfn_to_page(pfn);
704 * copy one vm_area from one task to the other. Assumes the page tables
705 * already present in the new task to be cleared in the whole range
706 * covered by this vma.
710 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
711 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
712 unsigned long addr, int *rss)
714 unsigned long vm_flags = vma->vm_flags;
715 pte_t pte = *src_pte;
717 swp_entry_t entry = pte_to_swp_entry(pte);
719 if (likely(!non_swap_entry(entry))) {
720 if (swap_duplicate(entry) < 0)
723 /* make sure dst_mm is on swapoff's mmlist. */
724 if (unlikely(list_empty(&dst_mm->mmlist))) {
725 spin_lock(&mmlist_lock);
726 if (list_empty(&dst_mm->mmlist))
727 list_add(&dst_mm->mmlist,
729 spin_unlock(&mmlist_lock);
732 } else if (is_migration_entry(entry)) {
733 page = migration_entry_to_page(entry);
735 rss[mm_counter(page)]++;
737 if (is_write_migration_entry(entry) &&
738 is_cow_mapping(vm_flags)) {
740 * COW mappings require pages in both
741 * parent and child to be set to read.
743 make_migration_entry_read(&entry);
744 pte = swp_entry_to_pte(entry);
745 if (pte_swp_soft_dirty(*src_pte))
746 pte = pte_swp_mksoft_dirty(pte);
747 if (pte_swp_uffd_wp(*src_pte))
748 pte = pte_swp_mkuffd_wp(pte);
749 set_pte_at(src_mm, addr, src_pte, pte);
751 } else if (is_device_private_entry(entry)) {
752 page = device_private_entry_to_page(entry);
755 * Update rss count even for unaddressable pages, as
756 * they should treated just like normal pages in this
759 * We will likely want to have some new rss counters
760 * for unaddressable pages, at some point. But for now
761 * keep things as they are.
764 rss[mm_counter(page)]++;
765 page_dup_rmap(page, false);
768 * We do not preserve soft-dirty information, because so
769 * far, checkpoint/restore is the only feature that
770 * requires that. And checkpoint/restore does not work
771 * when a device driver is involved (you cannot easily
772 * save and restore device driver state).
774 if (is_write_device_private_entry(entry) &&
775 is_cow_mapping(vm_flags)) {
776 make_device_private_entry_read(&entry);
777 pte = swp_entry_to_pte(entry);
778 if (pte_swp_uffd_wp(*src_pte))
779 pte = pte_swp_mkuffd_wp(pte);
780 set_pte_at(src_mm, addr, src_pte, pte);
783 set_pte_at(dst_mm, addr, dst_pte, pte);
788 * Copy a present and normal page if necessary.
790 * NOTE! The usual case is that this doesn't need to do
791 * anything, and can just return a positive value. That
792 * will let the caller know that it can just increase
793 * the page refcount and re-use the pte the traditional
796 * But _if_ we need to copy it because it needs to be
797 * pinned in the parent (and the child should get its own
798 * copy rather than just a reference to the same page),
799 * we'll do that here and return zero to let the caller
802 * And if we need a pre-allocated page but don't yet have
803 * one, return a negative error to let the preallocation
804 * code know so that it can do so outside the page table
808 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
809 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
810 struct page **prealloc, pte_t pte, struct page *page)
812 struct page *new_page;
815 * What we want to do is to check whether this page may
816 * have been pinned by the parent process. If so,
817 * instead of wrprotect the pte on both sides, we copy
818 * the page immediately so that we'll always guarantee
819 * the pinned page won't be randomly replaced in the
822 * The page pinning checks are just "has this mm ever
823 * seen pinning", along with the (inexact) check of
824 * the page count. That might give false positives for
825 * for pinning, but it will work correctly.
827 if (likely(!page_needs_cow_for_dma(src_vma, page)))
830 new_page = *prealloc;
835 * We have a prealloc page, all good! Take it
836 * over and copy the page & arm it.
839 copy_user_highpage(new_page, page, addr, src_vma);
840 __SetPageUptodate(new_page);
841 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
842 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
843 rss[mm_counter(new_page)]++;
845 /* All done, just insert the new page copy in the child */
846 pte = mk_pte(new_page, dst_vma->vm_page_prot);
847 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
848 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
853 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
854 * is required to copy this pte.
857 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
858 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
859 struct page **prealloc)
861 struct mm_struct *src_mm = src_vma->vm_mm;
862 unsigned long vm_flags = src_vma->vm_flags;
863 pte_t pte = *src_pte;
866 page = vm_normal_page(src_vma, addr, pte);
870 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
871 addr, rss, prealloc, pte, page);
876 page_dup_rmap(page, false);
877 rss[mm_counter(page)]++;
881 * If it's a COW mapping, write protect it both
882 * in the parent and the child
884 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
885 ptep_set_wrprotect(src_mm, addr, src_pte);
886 pte = pte_wrprotect(pte);
890 * If it's a shared mapping, mark it clean in
893 if (vm_flags & VM_SHARED)
894 pte = pte_mkclean(pte);
895 pte = pte_mkold(pte);
898 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
899 * does not have the VM_UFFD_WP, which means that the uffd
900 * fork event is not enabled.
902 if (!(vm_flags & VM_UFFD_WP))
903 pte = pte_clear_uffd_wp(pte);
905 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
909 static inline struct page *
910 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
913 struct page *new_page;
915 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
919 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
923 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
929 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
930 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
933 struct mm_struct *dst_mm = dst_vma->vm_mm;
934 struct mm_struct *src_mm = src_vma->vm_mm;
935 pte_t *orig_src_pte, *orig_dst_pte;
936 pte_t *src_pte, *dst_pte;
937 spinlock_t *src_ptl, *dst_ptl;
938 int progress, ret = 0;
939 int rss[NR_MM_COUNTERS];
940 swp_entry_t entry = (swp_entry_t){0};
941 struct page *prealloc = NULL;
947 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
952 src_pte = pte_offset_map(src_pmd, addr);
953 src_ptl = pte_lockptr(src_mm, src_pmd);
954 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
955 orig_src_pte = src_pte;
956 orig_dst_pte = dst_pte;
957 arch_enter_lazy_mmu_mode();
961 * We are holding two locks at this point - either of them
962 * could generate latencies in another task on another CPU.
964 if (progress >= 32) {
966 if (need_resched() ||
967 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
970 if (pte_none(*src_pte)) {
974 if (unlikely(!pte_present(*src_pte))) {
975 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
983 /* copy_present_pte() will clear `*prealloc' if consumed */
984 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
985 addr, rss, &prealloc);
987 * If we need a pre-allocated page for this pte, drop the
988 * locks, allocate, and try again.
990 if (unlikely(ret == -EAGAIN))
992 if (unlikely(prealloc)) {
994 * pre-alloc page cannot be reused by next time so as
995 * to strictly follow mempolicy (e.g., alloc_page_vma()
996 * will allocate page according to address). This
997 * could only happen if one pinned pte changed.
1003 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1005 arch_leave_lazy_mmu_mode();
1006 spin_unlock(src_ptl);
1007 pte_unmap(orig_src_pte);
1008 add_mm_rss_vec(dst_mm, rss);
1009 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1013 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1019 WARN_ON_ONCE(ret != -EAGAIN);
1020 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1023 /* We've captured and resolved the error. Reset, try again. */
1029 if (unlikely(prealloc))
1035 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1036 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1039 struct mm_struct *dst_mm = dst_vma->vm_mm;
1040 struct mm_struct *src_mm = src_vma->vm_mm;
1041 pmd_t *src_pmd, *dst_pmd;
1044 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1047 src_pmd = pmd_offset(src_pud, addr);
1049 next = pmd_addr_end(addr, end);
1050 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1051 || pmd_devmap(*src_pmd)) {
1053 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1054 err = copy_huge_pmd(dst_mm, src_mm,
1055 dst_pmd, src_pmd, addr, src_vma);
1062 if (pmd_none_or_clear_bad(src_pmd))
1064 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1067 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1072 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1073 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1076 struct mm_struct *dst_mm = dst_vma->vm_mm;
1077 struct mm_struct *src_mm = src_vma->vm_mm;
1078 pud_t *src_pud, *dst_pud;
1081 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1084 src_pud = pud_offset(src_p4d, addr);
1086 next = pud_addr_end(addr, end);
1087 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1090 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1091 err = copy_huge_pud(dst_mm, src_mm,
1092 dst_pud, src_pud, addr, src_vma);
1099 if (pud_none_or_clear_bad(src_pud))
1101 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1104 } while (dst_pud++, src_pud++, addr = next, addr != end);
1109 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1110 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1113 struct mm_struct *dst_mm = dst_vma->vm_mm;
1114 p4d_t *src_p4d, *dst_p4d;
1117 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1120 src_p4d = p4d_offset(src_pgd, addr);
1122 next = p4d_addr_end(addr, end);
1123 if (p4d_none_or_clear_bad(src_p4d))
1125 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1128 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1133 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1135 pgd_t *src_pgd, *dst_pgd;
1137 unsigned long addr = src_vma->vm_start;
1138 unsigned long end = src_vma->vm_end;
1139 struct mm_struct *dst_mm = dst_vma->vm_mm;
1140 struct mm_struct *src_mm = src_vma->vm_mm;
1141 struct mmu_notifier_range range;
1146 * Don't copy ptes where a page fault will fill them correctly.
1147 * Fork becomes much lighter when there are big shared or private
1148 * readonly mappings. The tradeoff is that copy_page_range is more
1149 * efficient than faulting.
1151 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1155 if (is_vm_hugetlb_page(src_vma))
1156 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1158 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1160 * We do not free on error cases below as remove_vma
1161 * gets called on error from higher level routine
1163 ret = track_pfn_copy(src_vma);
1169 * We need to invalidate the secondary MMU mappings only when
1170 * there could be a permission downgrade on the ptes of the
1171 * parent mm. And a permission downgrade will only happen if
1172 * is_cow_mapping() returns true.
1174 is_cow = is_cow_mapping(src_vma->vm_flags);
1177 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1178 0, src_vma, src_mm, addr, end);
1179 mmu_notifier_invalidate_range_start(&range);
1181 * Disabling preemption is not needed for the write side, as
1182 * the read side doesn't spin, but goes to the mmap_lock.
1184 * Use the raw variant of the seqcount_t write API to avoid
1185 * lockdep complaining about preemptibility.
1187 mmap_assert_write_locked(src_mm);
1188 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1192 dst_pgd = pgd_offset(dst_mm, addr);
1193 src_pgd = pgd_offset(src_mm, addr);
1195 next = pgd_addr_end(addr, end);
1196 if (pgd_none_or_clear_bad(src_pgd))
1198 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1203 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1206 raw_write_seqcount_end(&src_mm->write_protect_seq);
1207 mmu_notifier_invalidate_range_end(&range);
1212 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1213 struct vm_area_struct *vma, pmd_t *pmd,
1214 unsigned long addr, unsigned long end,
1215 struct zap_details *details)
1217 struct mm_struct *mm = tlb->mm;
1218 int force_flush = 0;
1219 int rss[NR_MM_COUNTERS];
1225 tlb_change_page_size(tlb, PAGE_SIZE);
1228 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1230 flush_tlb_batched_pending(mm);
1231 arch_enter_lazy_mmu_mode();
1234 if (pte_none(ptent))
1240 if (pte_present(ptent)) {
1243 page = vm_normal_page(vma, addr, ptent);
1244 if (unlikely(details) && page) {
1246 * unmap_shared_mapping_pages() wants to
1247 * invalidate cache without truncating:
1248 * unmap shared but keep private pages.
1250 if (details->check_mapping &&
1251 details->check_mapping != page_rmapping(page))
1254 ptent = ptep_get_and_clear_full(mm, addr, pte,
1256 tlb_remove_tlb_entry(tlb, pte, addr);
1257 if (unlikely(!page))
1260 if (!PageAnon(page)) {
1261 if (pte_dirty(ptent)) {
1263 set_page_dirty(page);
1265 if (pte_young(ptent) &&
1266 likely(!(vma->vm_flags & VM_SEQ_READ)))
1267 mark_page_accessed(page);
1269 rss[mm_counter(page)]--;
1270 page_remove_rmap(page, false);
1271 if (unlikely(page_mapcount(page) < 0))
1272 print_bad_pte(vma, addr, ptent, page);
1273 if (unlikely(__tlb_remove_page(tlb, page))) {
1281 entry = pte_to_swp_entry(ptent);
1282 if (is_device_private_entry(entry)) {
1283 struct page *page = device_private_entry_to_page(entry);
1285 if (unlikely(details && details->check_mapping)) {
1287 * unmap_shared_mapping_pages() wants to
1288 * invalidate cache without truncating:
1289 * unmap shared but keep private pages.
1291 if (details->check_mapping !=
1292 page_rmapping(page))
1296 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1297 rss[mm_counter(page)]--;
1298 page_remove_rmap(page, false);
1303 /* If details->check_mapping, we leave swap entries. */
1304 if (unlikely(details))
1307 if (!non_swap_entry(entry))
1309 else if (is_migration_entry(entry)) {
1312 page = migration_entry_to_page(entry);
1313 rss[mm_counter(page)]--;
1315 if (unlikely(!free_swap_and_cache(entry)))
1316 print_bad_pte(vma, addr, ptent, NULL);
1317 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1318 } while (pte++, addr += PAGE_SIZE, addr != end);
1320 add_mm_rss_vec(mm, rss);
1321 arch_leave_lazy_mmu_mode();
1323 /* Do the actual TLB flush before dropping ptl */
1325 tlb_flush_mmu_tlbonly(tlb);
1326 pte_unmap_unlock(start_pte, ptl);
1329 * If we forced a TLB flush (either due to running out of
1330 * batch buffers or because we needed to flush dirty TLB
1331 * entries before releasing the ptl), free the batched
1332 * memory too. Restart if we didn't do everything.
1347 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1348 struct vm_area_struct *vma, pud_t *pud,
1349 unsigned long addr, unsigned long end,
1350 struct zap_details *details)
1355 pmd = pmd_offset(pud, addr);
1357 next = pmd_addr_end(addr, end);
1358 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1359 if (next - addr != HPAGE_PMD_SIZE)
1360 __split_huge_pmd(vma, pmd, addr, false, NULL);
1361 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1366 * Here there can be other concurrent MADV_DONTNEED or
1367 * trans huge page faults running, and if the pmd is
1368 * none or trans huge it can change under us. This is
1369 * because MADV_DONTNEED holds the mmap_lock in read
1372 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1374 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1377 } while (pmd++, addr = next, addr != end);
1382 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1383 struct vm_area_struct *vma, p4d_t *p4d,
1384 unsigned long addr, unsigned long end,
1385 struct zap_details *details)
1390 pud = pud_offset(p4d, addr);
1392 next = pud_addr_end(addr, end);
1393 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1394 if (next - addr != HPAGE_PUD_SIZE) {
1395 mmap_assert_locked(tlb->mm);
1396 split_huge_pud(vma, pud, addr);
1397 } else if (zap_huge_pud(tlb, vma, pud, addr))
1401 if (pud_none_or_clear_bad(pud))
1403 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1406 } while (pud++, addr = next, addr != end);
1411 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1412 struct vm_area_struct *vma, pgd_t *pgd,
1413 unsigned long addr, unsigned long end,
1414 struct zap_details *details)
1419 p4d = p4d_offset(pgd, addr);
1421 next = p4d_addr_end(addr, end);
1422 if (p4d_none_or_clear_bad(p4d))
1424 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1425 } while (p4d++, addr = next, addr != end);
1430 void unmap_page_range(struct mmu_gather *tlb,
1431 struct vm_area_struct *vma,
1432 unsigned long addr, unsigned long end,
1433 struct zap_details *details)
1438 BUG_ON(addr >= end);
1439 tlb_start_vma(tlb, vma);
1440 pgd = pgd_offset(vma->vm_mm, addr);
1442 next = pgd_addr_end(addr, end);
1443 if (pgd_none_or_clear_bad(pgd))
1445 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1446 } while (pgd++, addr = next, addr != end);
1447 tlb_end_vma(tlb, vma);
1451 static void unmap_single_vma(struct mmu_gather *tlb,
1452 struct vm_area_struct *vma, unsigned long start_addr,
1453 unsigned long end_addr,
1454 struct zap_details *details)
1456 unsigned long start = max(vma->vm_start, start_addr);
1459 if (start >= vma->vm_end)
1461 end = min(vma->vm_end, end_addr);
1462 if (end <= vma->vm_start)
1466 uprobe_munmap(vma, start, end);
1468 if (unlikely(vma->vm_flags & VM_PFNMAP))
1469 untrack_pfn(vma, 0, 0);
1472 if (unlikely(is_vm_hugetlb_page(vma))) {
1474 * It is undesirable to test vma->vm_file as it
1475 * should be non-null for valid hugetlb area.
1476 * However, vm_file will be NULL in the error
1477 * cleanup path of mmap_region. When
1478 * hugetlbfs ->mmap method fails,
1479 * mmap_region() nullifies vma->vm_file
1480 * before calling this function to clean up.
1481 * Since no pte has actually been setup, it is
1482 * safe to do nothing in this case.
1485 i_mmap_lock_write(vma->vm_file->f_mapping);
1486 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1487 i_mmap_unlock_write(vma->vm_file->f_mapping);
1490 unmap_page_range(tlb, vma, start, end, details);
1495 * unmap_vmas - unmap a range of memory covered by a list of vma's
1496 * @tlb: address of the caller's struct mmu_gather
1497 * @vma: the starting vma
1498 * @start_addr: virtual address at which to start unmapping
1499 * @end_addr: virtual address at which to end unmapping
1501 * Unmap all pages in the vma list.
1503 * Only addresses between `start' and `end' will be unmapped.
1505 * The VMA list must be sorted in ascending virtual address order.
1507 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1508 * range after unmap_vmas() returns. So the only responsibility here is to
1509 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1510 * drops the lock and schedules.
1512 void unmap_vmas(struct mmu_gather *tlb,
1513 struct vm_area_struct *vma, unsigned long start_addr,
1514 unsigned long end_addr)
1516 struct mmu_notifier_range range;
1518 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1519 start_addr, end_addr);
1520 mmu_notifier_invalidate_range_start(&range);
1521 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1522 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1523 mmu_notifier_invalidate_range_end(&range);
1527 * zap_page_range - remove user pages in a given range
1528 * @vma: vm_area_struct holding the applicable pages
1529 * @start: starting address of pages to zap
1530 * @size: number of bytes to zap
1532 * Caller must protect the VMA list
1534 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1537 struct mmu_notifier_range range;
1538 struct mmu_gather tlb;
1541 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1542 start, start + size);
1543 tlb_gather_mmu(&tlb, vma->vm_mm);
1544 update_hiwater_rss(vma->vm_mm);
1545 mmu_notifier_invalidate_range_start(&range);
1546 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1547 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1548 mmu_notifier_invalidate_range_end(&range);
1549 tlb_finish_mmu(&tlb);
1553 * zap_page_range_single - remove user pages in a given range
1554 * @vma: vm_area_struct holding the applicable pages
1555 * @address: starting address of pages to zap
1556 * @size: number of bytes to zap
1557 * @details: details of shared cache invalidation
1559 * The range must fit into one VMA.
1561 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1562 unsigned long size, struct zap_details *details)
1564 struct mmu_notifier_range range;
1565 struct mmu_gather tlb;
1568 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1569 address, address + size);
1570 tlb_gather_mmu(&tlb, vma->vm_mm);
1571 update_hiwater_rss(vma->vm_mm);
1572 mmu_notifier_invalidate_range_start(&range);
1573 unmap_single_vma(&tlb, vma, address, range.end, details);
1574 mmu_notifier_invalidate_range_end(&range);
1575 tlb_finish_mmu(&tlb);
1579 * zap_vma_ptes - remove ptes mapping the vma
1580 * @vma: vm_area_struct holding ptes to be zapped
1581 * @address: starting address of pages to zap
1582 * @size: number of bytes to zap
1584 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1586 * The entire address range must be fully contained within the vma.
1589 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1592 if (address < vma->vm_start || address + size > vma->vm_end ||
1593 !(vma->vm_flags & VM_PFNMAP))
1596 zap_page_range_single(vma, address, size, NULL);
1598 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1600 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1607 pgd = pgd_offset(mm, addr);
1608 p4d = p4d_alloc(mm, pgd, addr);
1611 pud = pud_alloc(mm, p4d, addr);
1614 pmd = pmd_alloc(mm, pud, addr);
1618 VM_BUG_ON(pmd_trans_huge(*pmd));
1622 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1625 pmd_t *pmd = walk_to_pmd(mm, addr);
1629 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1632 static int validate_page_before_insert(struct page *page)
1634 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1636 flush_dcache_page(page);
1640 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1641 unsigned long addr, struct page *page, pgprot_t prot)
1643 if (!pte_none(*pte))
1645 /* Ok, finally just insert the thing.. */
1647 inc_mm_counter_fast(mm, mm_counter_file(page));
1648 page_add_file_rmap(page, false);
1649 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1654 * This is the old fallback for page remapping.
1656 * For historical reasons, it only allows reserved pages. Only
1657 * old drivers should use this, and they needed to mark their
1658 * pages reserved for the old functions anyway.
1660 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1661 struct page *page, pgprot_t prot)
1663 struct mm_struct *mm = vma->vm_mm;
1668 retval = validate_page_before_insert(page);
1672 pte = get_locked_pte(mm, addr, &ptl);
1675 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1676 pte_unmap_unlock(pte, ptl);
1682 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1683 unsigned long addr, struct page *page, pgprot_t prot)
1687 if (!page_count(page))
1689 err = validate_page_before_insert(page);
1692 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1695 /* insert_pages() amortizes the cost of spinlock operations
1696 * when inserting pages in a loop. Arch *must* define pte_index.
1698 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1699 struct page **pages, unsigned long *num, pgprot_t prot)
1702 pte_t *start_pte, *pte;
1703 spinlock_t *pte_lock;
1704 struct mm_struct *const mm = vma->vm_mm;
1705 unsigned long curr_page_idx = 0;
1706 unsigned long remaining_pages_total = *num;
1707 unsigned long pages_to_write_in_pmd;
1711 pmd = walk_to_pmd(mm, addr);
1715 pages_to_write_in_pmd = min_t(unsigned long,
1716 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1718 /* Allocate the PTE if necessary; takes PMD lock once only. */
1720 if (pte_alloc(mm, pmd))
1723 while (pages_to_write_in_pmd) {
1725 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1727 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1728 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1729 int err = insert_page_in_batch_locked(mm, pte,
1730 addr, pages[curr_page_idx], prot);
1731 if (unlikely(err)) {
1732 pte_unmap_unlock(start_pte, pte_lock);
1734 remaining_pages_total -= pte_idx;
1740 pte_unmap_unlock(start_pte, pte_lock);
1741 pages_to_write_in_pmd -= batch_size;
1742 remaining_pages_total -= batch_size;
1744 if (remaining_pages_total)
1748 *num = remaining_pages_total;
1751 #endif /* ifdef pte_index */
1754 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1755 * @vma: user vma to map to
1756 * @addr: target start user address of these pages
1757 * @pages: source kernel pages
1758 * @num: in: number of pages to map. out: number of pages that were *not*
1759 * mapped. (0 means all pages were successfully mapped).
1761 * Preferred over vm_insert_page() when inserting multiple pages.
1763 * In case of error, we may have mapped a subset of the provided
1764 * pages. It is the caller's responsibility to account for this case.
1766 * The same restrictions apply as in vm_insert_page().
1768 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1769 struct page **pages, unsigned long *num)
1772 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1774 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1776 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1777 BUG_ON(mmap_read_trylock(vma->vm_mm));
1778 BUG_ON(vma->vm_flags & VM_PFNMAP);
1779 vma->vm_flags |= VM_MIXEDMAP;
1781 /* Defer page refcount checking till we're about to map that page. */
1782 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1784 unsigned long idx = 0, pgcount = *num;
1787 for (; idx < pgcount; ++idx) {
1788 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1792 *num = pgcount - idx;
1794 #endif /* ifdef pte_index */
1796 EXPORT_SYMBOL(vm_insert_pages);
1799 * vm_insert_page - insert single page into user vma
1800 * @vma: user vma to map to
1801 * @addr: target user address of this page
1802 * @page: source kernel page
1804 * This allows drivers to insert individual pages they've allocated
1807 * The page has to be a nice clean _individual_ kernel allocation.
1808 * If you allocate a compound page, you need to have marked it as
1809 * such (__GFP_COMP), or manually just split the page up yourself
1810 * (see split_page()).
1812 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1813 * took an arbitrary page protection parameter. This doesn't allow
1814 * that. Your vma protection will have to be set up correctly, which
1815 * means that if you want a shared writable mapping, you'd better
1816 * ask for a shared writable mapping!
1818 * The page does not need to be reserved.
1820 * Usually this function is called from f_op->mmap() handler
1821 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1822 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1823 * function from other places, for example from page-fault handler.
1825 * Return: %0 on success, negative error code otherwise.
1827 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1830 if (addr < vma->vm_start || addr >= vma->vm_end)
1832 if (!page_count(page))
1834 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1835 BUG_ON(mmap_read_trylock(vma->vm_mm));
1836 BUG_ON(vma->vm_flags & VM_PFNMAP);
1837 vma->vm_flags |= VM_MIXEDMAP;
1839 return insert_page(vma, addr, page, vma->vm_page_prot);
1841 EXPORT_SYMBOL(vm_insert_page);
1844 * __vm_map_pages - maps range of kernel pages into user vma
1845 * @vma: user vma to map to
1846 * @pages: pointer to array of source kernel pages
1847 * @num: number of pages in page array
1848 * @offset: user's requested vm_pgoff
1850 * This allows drivers to map range of kernel pages into a user vma.
1852 * Return: 0 on success and error code otherwise.
1854 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1855 unsigned long num, unsigned long offset)
1857 unsigned long count = vma_pages(vma);
1858 unsigned long uaddr = vma->vm_start;
1861 /* Fail if the user requested offset is beyond the end of the object */
1865 /* Fail if the user requested size exceeds available object size */
1866 if (count > num - offset)
1869 for (i = 0; i < count; i++) {
1870 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1880 * vm_map_pages - maps range of kernel pages starts with non zero offset
1881 * @vma: user vma to map to
1882 * @pages: pointer to array of source kernel pages
1883 * @num: number of pages in page array
1885 * Maps an object consisting of @num pages, catering for the user's
1886 * requested vm_pgoff
1888 * If we fail to insert any page into the vma, the function will return
1889 * immediately leaving any previously inserted pages present. Callers
1890 * from the mmap handler may immediately return the error as their caller
1891 * will destroy the vma, removing any successfully inserted pages. Other
1892 * callers should make their own arrangements for calling unmap_region().
1894 * Context: Process context. Called by mmap handlers.
1895 * Return: 0 on success and error code otherwise.
1897 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1900 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1902 EXPORT_SYMBOL(vm_map_pages);
1905 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1906 * @vma: user vma to map to
1907 * @pages: pointer to array of source kernel pages
1908 * @num: number of pages in page array
1910 * Similar to vm_map_pages(), except that it explicitly sets the offset
1911 * to 0. This function is intended for the drivers that did not consider
1914 * Context: Process context. Called by mmap handlers.
1915 * Return: 0 on success and error code otherwise.
1917 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1920 return __vm_map_pages(vma, pages, num, 0);
1922 EXPORT_SYMBOL(vm_map_pages_zero);
1924 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1925 pfn_t pfn, pgprot_t prot, bool mkwrite)
1927 struct mm_struct *mm = vma->vm_mm;
1931 pte = get_locked_pte(mm, addr, &ptl);
1933 return VM_FAULT_OOM;
1934 if (!pte_none(*pte)) {
1937 * For read faults on private mappings the PFN passed
1938 * in may not match the PFN we have mapped if the
1939 * mapped PFN is a writeable COW page. In the mkwrite
1940 * case we are creating a writable PTE for a shared
1941 * mapping and we expect the PFNs to match. If they
1942 * don't match, we are likely racing with block
1943 * allocation and mapping invalidation so just skip the
1946 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1947 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1950 entry = pte_mkyoung(*pte);
1951 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1952 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1953 update_mmu_cache(vma, addr, pte);
1958 /* Ok, finally just insert the thing.. */
1959 if (pfn_t_devmap(pfn))
1960 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1962 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1965 entry = pte_mkyoung(entry);
1966 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1969 set_pte_at(mm, addr, pte, entry);
1970 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1973 pte_unmap_unlock(pte, ptl);
1974 return VM_FAULT_NOPAGE;
1978 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1979 * @vma: user vma to map to
1980 * @addr: target user address of this page
1981 * @pfn: source kernel pfn
1982 * @pgprot: pgprot flags for the inserted page
1984 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1985 * to override pgprot on a per-page basis.
1987 * This only makes sense for IO mappings, and it makes no sense for
1988 * COW mappings. In general, using multiple vmas is preferable;
1989 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1992 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1993 * a value of @pgprot different from that of @vma->vm_page_prot.
1995 * Context: Process context. May allocate using %GFP_KERNEL.
1996 * Return: vm_fault_t value.
1998 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1999 unsigned long pfn, pgprot_t pgprot)
2002 * Technically, architectures with pte_special can avoid all these
2003 * restrictions (same for remap_pfn_range). However we would like
2004 * consistency in testing and feature parity among all, so we should
2005 * try to keep these invariants in place for everybody.
2007 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2008 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2009 (VM_PFNMAP|VM_MIXEDMAP));
2010 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2011 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2013 if (addr < vma->vm_start || addr >= vma->vm_end)
2014 return VM_FAULT_SIGBUS;
2016 if (!pfn_modify_allowed(pfn, pgprot))
2017 return VM_FAULT_SIGBUS;
2019 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2021 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2024 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2027 * vmf_insert_pfn - insert single pfn into user vma
2028 * @vma: user vma to map to
2029 * @addr: target user address of this page
2030 * @pfn: source kernel pfn
2032 * Similar to vm_insert_page, this allows drivers to insert individual pages
2033 * they've allocated into a user vma. Same comments apply.
2035 * This function should only be called from a vm_ops->fault handler, and
2036 * in that case the handler should return the result of this function.
2038 * vma cannot be a COW mapping.
2040 * As this is called only for pages that do not currently exist, we
2041 * do not need to flush old virtual caches or the TLB.
2043 * Context: Process context. May allocate using %GFP_KERNEL.
2044 * Return: vm_fault_t value.
2046 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2049 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2051 EXPORT_SYMBOL(vmf_insert_pfn);
2053 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2055 /* these checks mirror the abort conditions in vm_normal_page */
2056 if (vma->vm_flags & VM_MIXEDMAP)
2058 if (pfn_t_devmap(pfn))
2060 if (pfn_t_special(pfn))
2062 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2067 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2068 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2073 BUG_ON(!vm_mixed_ok(vma, pfn));
2075 if (addr < vma->vm_start || addr >= vma->vm_end)
2076 return VM_FAULT_SIGBUS;
2078 track_pfn_insert(vma, &pgprot, pfn);
2080 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2081 return VM_FAULT_SIGBUS;
2084 * If we don't have pte special, then we have to use the pfn_valid()
2085 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2086 * refcount the page if pfn_valid is true (hence insert_page rather
2087 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2088 * without pte special, it would there be refcounted as a normal page.
2090 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2091 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2095 * At this point we are committed to insert_page()
2096 * regardless of whether the caller specified flags that
2097 * result in pfn_t_has_page() == false.
2099 page = pfn_to_page(pfn_t_to_pfn(pfn));
2100 err = insert_page(vma, addr, page, pgprot);
2102 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2106 return VM_FAULT_OOM;
2107 if (err < 0 && err != -EBUSY)
2108 return VM_FAULT_SIGBUS;
2110 return VM_FAULT_NOPAGE;
2114 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2115 * @vma: user vma to map to
2116 * @addr: target user address of this page
2117 * @pfn: source kernel pfn
2118 * @pgprot: pgprot flags for the inserted page
2120 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2121 * to override pgprot on a per-page basis.
2123 * Typically this function should be used by drivers to set caching- and
2124 * encryption bits different than those of @vma->vm_page_prot, because
2125 * the caching- or encryption mode may not be known at mmap() time.
2126 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2127 * to set caching and encryption bits for those vmas (except for COW pages).
2128 * This is ensured by core vm only modifying these page table entries using
2129 * functions that don't touch caching- or encryption bits, using pte_modify()
2130 * if needed. (See for example mprotect()).
2131 * Also when new page-table entries are created, this is only done using the
2132 * fault() callback, and never using the value of vma->vm_page_prot,
2133 * except for page-table entries that point to anonymous pages as the result
2136 * Context: Process context. May allocate using %GFP_KERNEL.
2137 * Return: vm_fault_t value.
2139 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2140 pfn_t pfn, pgprot_t pgprot)
2142 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2144 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2146 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2149 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2151 EXPORT_SYMBOL(vmf_insert_mixed);
2154 * If the insertion of PTE failed because someone else already added a
2155 * different entry in the mean time, we treat that as success as we assume
2156 * the same entry was actually inserted.
2158 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2159 unsigned long addr, pfn_t pfn)
2161 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2163 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2166 * maps a range of physical memory into the requested pages. the old
2167 * mappings are removed. any references to nonexistent pages results
2168 * in null mappings (currently treated as "copy-on-access")
2170 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2171 unsigned long addr, unsigned long end,
2172 unsigned long pfn, pgprot_t prot)
2174 pte_t *pte, *mapped_pte;
2178 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2181 arch_enter_lazy_mmu_mode();
2183 BUG_ON(!pte_none(*pte));
2184 if (!pfn_modify_allowed(pfn, prot)) {
2188 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2190 } while (pte++, addr += PAGE_SIZE, addr != end);
2191 arch_leave_lazy_mmu_mode();
2192 pte_unmap_unlock(mapped_pte, ptl);
2196 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2197 unsigned long addr, unsigned long end,
2198 unsigned long pfn, pgprot_t prot)
2204 pfn -= addr >> PAGE_SHIFT;
2205 pmd = pmd_alloc(mm, pud, addr);
2208 VM_BUG_ON(pmd_trans_huge(*pmd));
2210 next = pmd_addr_end(addr, end);
2211 err = remap_pte_range(mm, pmd, addr, next,
2212 pfn + (addr >> PAGE_SHIFT), prot);
2215 } while (pmd++, addr = next, addr != end);
2219 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2220 unsigned long addr, unsigned long end,
2221 unsigned long pfn, pgprot_t prot)
2227 pfn -= addr >> PAGE_SHIFT;
2228 pud = pud_alloc(mm, p4d, addr);
2232 next = pud_addr_end(addr, end);
2233 err = remap_pmd_range(mm, pud, addr, next,
2234 pfn + (addr >> PAGE_SHIFT), prot);
2237 } while (pud++, addr = next, addr != end);
2241 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2242 unsigned long addr, unsigned long end,
2243 unsigned long pfn, pgprot_t prot)
2249 pfn -= addr >> PAGE_SHIFT;
2250 p4d = p4d_alloc(mm, pgd, addr);
2254 next = p4d_addr_end(addr, end);
2255 err = remap_pud_range(mm, p4d, addr, next,
2256 pfn + (addr >> PAGE_SHIFT), prot);
2259 } while (p4d++, addr = next, addr != end);
2264 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2265 * must have pre-validated the caching bits of the pgprot_t.
2267 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2268 unsigned long pfn, unsigned long size, pgprot_t prot)
2272 unsigned long end = addr + PAGE_ALIGN(size);
2273 struct mm_struct *mm = vma->vm_mm;
2276 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2280 * Physically remapped pages are special. Tell the
2281 * rest of the world about it:
2282 * VM_IO tells people not to look at these pages
2283 * (accesses can have side effects).
2284 * VM_PFNMAP tells the core MM that the base pages are just
2285 * raw PFN mappings, and do not have a "struct page" associated
2288 * Disable vma merging and expanding with mremap().
2290 * Omit vma from core dump, even when VM_IO turned off.
2292 * There's a horrible special case to handle copy-on-write
2293 * behaviour that some programs depend on. We mark the "original"
2294 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2295 * See vm_normal_page() for details.
2297 if (is_cow_mapping(vma->vm_flags)) {
2298 if (addr != vma->vm_start || end != vma->vm_end)
2300 vma->vm_pgoff = pfn;
2303 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2305 BUG_ON(addr >= end);
2306 pfn -= addr >> PAGE_SHIFT;
2307 pgd = pgd_offset(mm, addr);
2308 flush_cache_range(vma, addr, end);
2310 next = pgd_addr_end(addr, end);
2311 err = remap_p4d_range(mm, pgd, addr, next,
2312 pfn + (addr >> PAGE_SHIFT), prot);
2315 } while (pgd++, addr = next, addr != end);
2321 * remap_pfn_range - remap kernel memory to userspace
2322 * @vma: user vma to map to
2323 * @addr: target page aligned user address to start at
2324 * @pfn: page frame number of kernel physical memory address
2325 * @size: size of mapping area
2326 * @prot: page protection flags for this mapping
2328 * Note: this is only safe if the mm semaphore is held when called.
2330 * Return: %0 on success, negative error code otherwise.
2332 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2333 unsigned long pfn, unsigned long size, pgprot_t prot)
2337 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2341 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2343 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2346 EXPORT_SYMBOL(remap_pfn_range);
2349 * vm_iomap_memory - remap memory to userspace
2350 * @vma: user vma to map to
2351 * @start: start of the physical memory to be mapped
2352 * @len: size of area
2354 * This is a simplified io_remap_pfn_range() for common driver use. The
2355 * driver just needs to give us the physical memory range to be mapped,
2356 * we'll figure out the rest from the vma information.
2358 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2359 * whatever write-combining details or similar.
2361 * Return: %0 on success, negative error code otherwise.
2363 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2365 unsigned long vm_len, pfn, pages;
2367 /* Check that the physical memory area passed in looks valid */
2368 if (start + len < start)
2371 * You *really* shouldn't map things that aren't page-aligned,
2372 * but we've historically allowed it because IO memory might
2373 * just have smaller alignment.
2375 len += start & ~PAGE_MASK;
2376 pfn = start >> PAGE_SHIFT;
2377 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2378 if (pfn + pages < pfn)
2381 /* We start the mapping 'vm_pgoff' pages into the area */
2382 if (vma->vm_pgoff > pages)
2384 pfn += vma->vm_pgoff;
2385 pages -= vma->vm_pgoff;
2387 /* Can we fit all of the mapping? */
2388 vm_len = vma->vm_end - vma->vm_start;
2389 if (vm_len >> PAGE_SHIFT > pages)
2392 /* Ok, let it rip */
2393 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2395 EXPORT_SYMBOL(vm_iomap_memory);
2397 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2398 unsigned long addr, unsigned long end,
2399 pte_fn_t fn, void *data, bool create,
2400 pgtbl_mod_mask *mask)
2402 pte_t *pte, *mapped_pte;
2407 mapped_pte = pte = (mm == &init_mm) ?
2408 pte_alloc_kernel_track(pmd, addr, mask) :
2409 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2413 mapped_pte = pte = (mm == &init_mm) ?
2414 pte_offset_kernel(pmd, addr) :
2415 pte_offset_map_lock(mm, pmd, addr, &ptl);
2418 BUG_ON(pmd_huge(*pmd));
2420 arch_enter_lazy_mmu_mode();
2424 if (create || !pte_none(*pte)) {
2425 err = fn(pte++, addr, data);
2429 } while (addr += PAGE_SIZE, addr != end);
2431 *mask |= PGTBL_PTE_MODIFIED;
2433 arch_leave_lazy_mmu_mode();
2436 pte_unmap_unlock(mapped_pte, ptl);
2440 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2441 unsigned long addr, unsigned long end,
2442 pte_fn_t fn, void *data, bool create,
2443 pgtbl_mod_mask *mask)
2449 BUG_ON(pud_huge(*pud));
2452 pmd = pmd_alloc_track(mm, pud, addr, mask);
2456 pmd = pmd_offset(pud, addr);
2459 next = pmd_addr_end(addr, end);
2460 if (create || !pmd_none_or_clear_bad(pmd)) {
2461 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2466 } while (pmd++, addr = next, addr != end);
2470 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2471 unsigned long addr, unsigned long end,
2472 pte_fn_t fn, void *data, bool create,
2473 pgtbl_mod_mask *mask)
2480 pud = pud_alloc_track(mm, p4d, addr, mask);
2484 pud = pud_offset(p4d, addr);
2487 next = pud_addr_end(addr, end);
2488 if (create || !pud_none_or_clear_bad(pud)) {
2489 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2494 } while (pud++, addr = next, addr != end);
2498 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2499 unsigned long addr, unsigned long end,
2500 pte_fn_t fn, void *data, bool create,
2501 pgtbl_mod_mask *mask)
2508 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2512 p4d = p4d_offset(pgd, addr);
2515 next = p4d_addr_end(addr, end);
2516 if (create || !p4d_none_or_clear_bad(p4d)) {
2517 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2522 } while (p4d++, addr = next, addr != end);
2526 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2527 unsigned long size, pte_fn_t fn,
2528 void *data, bool create)
2531 unsigned long start = addr, next;
2532 unsigned long end = addr + size;
2533 pgtbl_mod_mask mask = 0;
2536 if (WARN_ON(addr >= end))
2539 pgd = pgd_offset(mm, addr);
2541 next = pgd_addr_end(addr, end);
2542 if (!create && pgd_none_or_clear_bad(pgd))
2544 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2547 } while (pgd++, addr = next, addr != end);
2549 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2550 arch_sync_kernel_mappings(start, start + size);
2556 * Scan a region of virtual memory, filling in page tables as necessary
2557 * and calling a provided function on each leaf page table.
2559 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2560 unsigned long size, pte_fn_t fn, void *data)
2562 return __apply_to_page_range(mm, addr, size, fn, data, true);
2564 EXPORT_SYMBOL_GPL(apply_to_page_range);
2567 * Scan a region of virtual memory, calling a provided function on
2568 * each leaf page table where it exists.
2570 * Unlike apply_to_page_range, this does _not_ fill in page tables
2571 * where they are absent.
2573 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2574 unsigned long size, pte_fn_t fn, void *data)
2576 return __apply_to_page_range(mm, addr, size, fn, data, false);
2578 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2581 * handle_pte_fault chooses page fault handler according to an entry which was
2582 * read non-atomically. Before making any commitment, on those architectures
2583 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2584 * parts, do_swap_page must check under lock before unmapping the pte and
2585 * proceeding (but do_wp_page is only called after already making such a check;
2586 * and do_anonymous_page can safely check later on).
2588 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2589 pte_t *page_table, pte_t orig_pte)
2592 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2593 if (sizeof(pte_t) > sizeof(unsigned long)) {
2594 spinlock_t *ptl = pte_lockptr(mm, pmd);
2596 same = pte_same(*page_table, orig_pte);
2600 pte_unmap(page_table);
2604 static inline bool cow_user_page(struct page *dst, struct page *src,
2605 struct vm_fault *vmf)
2610 bool locked = false;
2611 struct vm_area_struct *vma = vmf->vma;
2612 struct mm_struct *mm = vma->vm_mm;
2613 unsigned long addr = vmf->address;
2616 copy_user_highpage(dst, src, addr, vma);
2621 * If the source page was a PFN mapping, we don't have
2622 * a "struct page" for it. We do a best-effort copy by
2623 * just copying from the original user address. If that
2624 * fails, we just zero-fill it. Live with it.
2626 kaddr = kmap_atomic(dst);
2627 uaddr = (void __user *)(addr & PAGE_MASK);
2630 * On architectures with software "accessed" bits, we would
2631 * take a double page fault, so mark it accessed here.
2633 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2636 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2638 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2640 * Other thread has already handled the fault
2641 * and update local tlb only
2643 update_mmu_tlb(vma, addr, vmf->pte);
2648 entry = pte_mkyoung(vmf->orig_pte);
2649 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2650 update_mmu_cache(vma, addr, vmf->pte);
2654 * This really shouldn't fail, because the page is there
2655 * in the page tables. But it might just be unreadable,
2656 * in which case we just give up and fill the result with
2659 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2663 /* Re-validate under PTL if the page is still mapped */
2664 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2666 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2667 /* The PTE changed under us, update local tlb */
2668 update_mmu_tlb(vma, addr, vmf->pte);
2674 * The same page can be mapped back since last copy attempt.
2675 * Try to copy again under PTL.
2677 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2679 * Give a warn in case there can be some obscure
2692 pte_unmap_unlock(vmf->pte, vmf->ptl);
2693 kunmap_atomic(kaddr);
2694 flush_dcache_page(dst);
2699 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2701 struct file *vm_file = vma->vm_file;
2704 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2707 * Special mappings (e.g. VDSO) do not have any file so fake
2708 * a default GFP_KERNEL for them.
2714 * Notify the address space that the page is about to become writable so that
2715 * it can prohibit this or wait for the page to get into an appropriate state.
2717 * We do this without the lock held, so that it can sleep if it needs to.
2719 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2722 struct page *page = vmf->page;
2723 unsigned int old_flags = vmf->flags;
2725 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2727 if (vmf->vma->vm_file &&
2728 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2729 return VM_FAULT_SIGBUS;
2731 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2732 /* Restore original flags so that caller is not surprised */
2733 vmf->flags = old_flags;
2734 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2736 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2738 if (!page->mapping) {
2740 return 0; /* retry */
2742 ret |= VM_FAULT_LOCKED;
2744 VM_BUG_ON_PAGE(!PageLocked(page), page);
2749 * Handle dirtying of a page in shared file mapping on a write fault.
2751 * The function expects the page to be locked and unlocks it.
2753 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2755 struct vm_area_struct *vma = vmf->vma;
2756 struct address_space *mapping;
2757 struct page *page = vmf->page;
2759 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2761 dirtied = set_page_dirty(page);
2762 VM_BUG_ON_PAGE(PageAnon(page), page);
2764 * Take a local copy of the address_space - page.mapping may be zeroed
2765 * by truncate after unlock_page(). The address_space itself remains
2766 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2767 * release semantics to prevent the compiler from undoing this copying.
2769 mapping = page_rmapping(page);
2773 file_update_time(vma->vm_file);
2776 * Throttle page dirtying rate down to writeback speed.
2778 * mapping may be NULL here because some device drivers do not
2779 * set page.mapping but still dirty their pages
2781 * Drop the mmap_lock before waiting on IO, if we can. The file
2782 * is pinning the mapping, as per above.
2784 if ((dirtied || page_mkwrite) && mapping) {
2787 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2788 balance_dirty_pages_ratelimited(mapping);
2791 return VM_FAULT_RETRY;
2799 * Handle write page faults for pages that can be reused in the current vma
2801 * This can happen either due to the mapping being with the VM_SHARED flag,
2802 * or due to us being the last reference standing to the page. In either
2803 * case, all we need to do here is to mark the page as writable and update
2804 * any related book-keeping.
2806 static inline void wp_page_reuse(struct vm_fault *vmf)
2807 __releases(vmf->ptl)
2809 struct vm_area_struct *vma = vmf->vma;
2810 struct page *page = vmf->page;
2813 * Clear the pages cpupid information as the existing
2814 * information potentially belongs to a now completely
2815 * unrelated process.
2818 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2820 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2821 entry = pte_mkyoung(vmf->orig_pte);
2822 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2823 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2824 update_mmu_cache(vma, vmf->address, vmf->pte);
2825 pte_unmap_unlock(vmf->pte, vmf->ptl);
2826 count_vm_event(PGREUSE);
2830 * Handle the case of a page which we actually need to copy to a new page.
2832 * Called with mmap_lock locked and the old page referenced, but
2833 * without the ptl held.
2835 * High level logic flow:
2837 * - Allocate a page, copy the content of the old page to the new one.
2838 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2839 * - Take the PTL. If the pte changed, bail out and release the allocated page
2840 * - If the pte is still the way we remember it, update the page table and all
2841 * relevant references. This includes dropping the reference the page-table
2842 * held to the old page, as well as updating the rmap.
2843 * - In any case, unlock the PTL and drop the reference we took to the old page.
2845 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2847 struct vm_area_struct *vma = vmf->vma;
2848 struct mm_struct *mm = vma->vm_mm;
2849 struct page *old_page = vmf->page;
2850 struct page *new_page = NULL;
2852 int page_copied = 0;
2853 struct mmu_notifier_range range;
2855 if (unlikely(anon_vma_prepare(vma)))
2858 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2859 new_page = alloc_zeroed_user_highpage_movable(vma,
2864 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2869 if (!cow_user_page(new_page, old_page, vmf)) {
2871 * COW failed, if the fault was solved by other,
2872 * it's fine. If not, userspace would re-fault on
2873 * the same address and we will handle the fault
2874 * from the second attempt.
2883 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2885 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2887 __SetPageUptodate(new_page);
2889 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2890 vmf->address & PAGE_MASK,
2891 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2892 mmu_notifier_invalidate_range_start(&range);
2895 * Re-check the pte - we dropped the lock
2897 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2898 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2900 if (!PageAnon(old_page)) {
2901 dec_mm_counter_fast(mm,
2902 mm_counter_file(old_page));
2903 inc_mm_counter_fast(mm, MM_ANONPAGES);
2906 inc_mm_counter_fast(mm, MM_ANONPAGES);
2908 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2909 entry = mk_pte(new_page, vma->vm_page_prot);
2910 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2913 * Clear the pte entry and flush it first, before updating the
2914 * pte with the new entry, to keep TLBs on different CPUs in
2915 * sync. This code used to set the new PTE then flush TLBs, but
2916 * that left a window where the new PTE could be loaded into
2917 * some TLBs while the old PTE remains in others.
2919 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2920 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2921 lru_cache_add_inactive_or_unevictable(new_page, vma);
2923 * We call the notify macro here because, when using secondary
2924 * mmu page tables (such as kvm shadow page tables), we want the
2925 * new page to be mapped directly into the secondary page table.
2927 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2928 update_mmu_cache(vma, vmf->address, vmf->pte);
2931 * Only after switching the pte to the new page may
2932 * we remove the mapcount here. Otherwise another
2933 * process may come and find the rmap count decremented
2934 * before the pte is switched to the new page, and
2935 * "reuse" the old page writing into it while our pte
2936 * here still points into it and can be read by other
2939 * The critical issue is to order this
2940 * page_remove_rmap with the ptp_clear_flush above.
2941 * Those stores are ordered by (if nothing else,)
2942 * the barrier present in the atomic_add_negative
2943 * in page_remove_rmap.
2945 * Then the TLB flush in ptep_clear_flush ensures that
2946 * no process can access the old page before the
2947 * decremented mapcount is visible. And the old page
2948 * cannot be reused until after the decremented
2949 * mapcount is visible. So transitively, TLBs to
2950 * old page will be flushed before it can be reused.
2952 page_remove_rmap(old_page, false);
2955 /* Free the old page.. */
2956 new_page = old_page;
2959 update_mmu_tlb(vma, vmf->address, vmf->pte);
2965 pte_unmap_unlock(vmf->pte, vmf->ptl);
2967 * No need to double call mmu_notifier->invalidate_range() callback as
2968 * the above ptep_clear_flush_notify() did already call it.
2970 mmu_notifier_invalidate_range_only_end(&range);
2973 * Don't let another task, with possibly unlocked vma,
2974 * keep the mlocked page.
2976 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2977 lock_page(old_page); /* LRU manipulation */
2978 if (PageMlocked(old_page))
2979 munlock_vma_page(old_page);
2980 unlock_page(old_page);
2984 return page_copied ? VM_FAULT_WRITE : 0;
2990 return VM_FAULT_OOM;
2994 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2995 * writeable once the page is prepared
2997 * @vmf: structure describing the fault
2999 * This function handles all that is needed to finish a write page fault in a
3000 * shared mapping due to PTE being read-only once the mapped page is prepared.
3001 * It handles locking of PTE and modifying it.
3003 * The function expects the page to be locked or other protection against
3004 * concurrent faults / writeback (such as DAX radix tree locks).
3006 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3007 * we acquired PTE lock.
3009 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3011 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3012 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3015 * We might have raced with another page fault while we released the
3016 * pte_offset_map_lock.
3018 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3019 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3020 pte_unmap_unlock(vmf->pte, vmf->ptl);
3021 return VM_FAULT_NOPAGE;
3028 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3031 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3033 struct vm_area_struct *vma = vmf->vma;
3035 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3038 pte_unmap_unlock(vmf->pte, vmf->ptl);
3039 vmf->flags |= FAULT_FLAG_MKWRITE;
3040 ret = vma->vm_ops->pfn_mkwrite(vmf);
3041 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3043 return finish_mkwrite_fault(vmf);
3046 return VM_FAULT_WRITE;
3049 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3050 __releases(vmf->ptl)
3052 struct vm_area_struct *vma = vmf->vma;
3053 vm_fault_t ret = VM_FAULT_WRITE;
3055 get_page(vmf->page);
3057 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3060 pte_unmap_unlock(vmf->pte, vmf->ptl);
3061 tmp = do_page_mkwrite(vmf);
3062 if (unlikely(!tmp || (tmp &
3063 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3064 put_page(vmf->page);
3067 tmp = finish_mkwrite_fault(vmf);
3068 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3069 unlock_page(vmf->page);
3070 put_page(vmf->page);
3075 lock_page(vmf->page);
3077 ret |= fault_dirty_shared_page(vmf);
3078 put_page(vmf->page);
3084 * This routine handles present pages, when users try to write
3085 * to a shared page. It is done by copying the page to a new address
3086 * and decrementing the shared-page counter for the old page.
3088 * Note that this routine assumes that the protection checks have been
3089 * done by the caller (the low-level page fault routine in most cases).
3090 * Thus we can safely just mark it writable once we've done any necessary
3093 * We also mark the page dirty at this point even though the page will
3094 * change only once the write actually happens. This avoids a few races,
3095 * and potentially makes it more efficient.
3097 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3098 * but allow concurrent faults), with pte both mapped and locked.
3099 * We return with mmap_lock still held, but pte unmapped and unlocked.
3101 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3102 __releases(vmf->ptl)
3104 struct vm_area_struct *vma = vmf->vma;
3106 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3107 pte_unmap_unlock(vmf->pte, vmf->ptl);
3108 return handle_userfault(vmf, VM_UFFD_WP);
3112 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3113 * is flushed in this case before copying.
3115 if (unlikely(userfaultfd_wp(vmf->vma) &&
3116 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3117 flush_tlb_page(vmf->vma, vmf->address);
3119 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3122 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3125 * We should not cow pages in a shared writeable mapping.
3126 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3128 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3129 (VM_WRITE|VM_SHARED))
3130 return wp_pfn_shared(vmf);
3132 pte_unmap_unlock(vmf->pte, vmf->ptl);
3133 return wp_page_copy(vmf);
3137 * Take out anonymous pages first, anonymous shared vmas are
3138 * not dirty accountable.
3140 if (PageAnon(vmf->page)) {
3141 struct page *page = vmf->page;
3143 /* PageKsm() doesn't necessarily raise the page refcount */
3144 if (PageKsm(page) || page_count(page) != 1)
3146 if (!trylock_page(page))
3148 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3153 * Ok, we've got the only map reference, and the only
3154 * page count reference, and the page is locked,
3155 * it's dark out, and we're wearing sunglasses. Hit it.
3159 return VM_FAULT_WRITE;
3160 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3161 (VM_WRITE|VM_SHARED))) {
3162 return wp_page_shared(vmf);
3166 * Ok, we need to copy. Oh, well..
3168 get_page(vmf->page);
3170 pte_unmap_unlock(vmf->pte, vmf->ptl);
3171 return wp_page_copy(vmf);
3174 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3175 unsigned long start_addr, unsigned long end_addr,
3176 struct zap_details *details)
3178 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3181 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3182 struct zap_details *details)
3184 struct vm_area_struct *vma;
3185 pgoff_t vba, vea, zba, zea;
3187 vma_interval_tree_foreach(vma, root,
3188 details->first_index, details->last_index) {
3190 vba = vma->vm_pgoff;
3191 vea = vba + vma_pages(vma) - 1;
3192 zba = details->first_index;
3195 zea = details->last_index;
3199 unmap_mapping_range_vma(vma,
3200 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3201 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3207 * unmap_mapping_pages() - Unmap pages from processes.
3208 * @mapping: The address space containing pages to be unmapped.
3209 * @start: Index of first page to be unmapped.
3210 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3211 * @even_cows: Whether to unmap even private COWed pages.
3213 * Unmap the pages in this address space from any userspace process which
3214 * has them mmaped. Generally, you want to remove COWed pages as well when
3215 * a file is being truncated, but not when invalidating pages from the page
3218 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3219 pgoff_t nr, bool even_cows)
3221 struct zap_details details = { };
3223 details.check_mapping = even_cows ? NULL : mapping;
3224 details.first_index = start;
3225 details.last_index = start + nr - 1;
3226 if (details.last_index < details.first_index)
3227 details.last_index = ULONG_MAX;
3229 i_mmap_lock_write(mapping);
3230 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3231 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3232 i_mmap_unlock_write(mapping);
3236 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3237 * address_space corresponding to the specified byte range in the underlying
3240 * @mapping: the address space containing mmaps to be unmapped.
3241 * @holebegin: byte in first page to unmap, relative to the start of
3242 * the underlying file. This will be rounded down to a PAGE_SIZE
3243 * boundary. Note that this is different from truncate_pagecache(), which
3244 * must keep the partial page. In contrast, we must get rid of
3246 * @holelen: size of prospective hole in bytes. This will be rounded
3247 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3249 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3250 * but 0 when invalidating pagecache, don't throw away private data.
3252 void unmap_mapping_range(struct address_space *mapping,
3253 loff_t const holebegin, loff_t const holelen, int even_cows)
3255 pgoff_t hba = holebegin >> PAGE_SHIFT;
3256 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3258 /* Check for overflow. */
3259 if (sizeof(holelen) > sizeof(hlen)) {
3261 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3262 if (holeend & ~(long long)ULONG_MAX)
3263 hlen = ULONG_MAX - hba + 1;
3266 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3268 EXPORT_SYMBOL(unmap_mapping_range);
3271 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3272 * but allow concurrent faults), and pte mapped but not yet locked.
3273 * We return with pte unmapped and unlocked.
3275 * We return with the mmap_lock locked or unlocked in the same cases
3276 * as does filemap_fault().
3278 vm_fault_t do_swap_page(struct vm_fault *vmf)
3280 struct vm_area_struct *vma = vmf->vma;
3281 struct page *page = NULL, *swapcache;
3287 void *shadow = NULL;
3289 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3292 entry = pte_to_swp_entry(vmf->orig_pte);
3293 if (unlikely(non_swap_entry(entry))) {
3294 if (is_migration_entry(entry)) {
3295 migration_entry_wait(vma->vm_mm, vmf->pmd,
3297 } else if (is_device_private_entry(entry)) {
3298 vmf->page = device_private_entry_to_page(entry);
3299 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3300 } else if (is_hwpoison_entry(entry)) {
3301 ret = VM_FAULT_HWPOISON;
3303 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3304 ret = VM_FAULT_SIGBUS;
3310 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3311 page = lookup_swap_cache(entry, vma, vmf->address);
3315 struct swap_info_struct *si = swp_swap_info(entry);
3317 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3318 __swap_count(entry) == 1) {
3319 /* skip swapcache */
3320 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3323 __SetPageLocked(page);
3324 __SetPageSwapBacked(page);
3326 if (mem_cgroup_swapin_charge_page(page,
3327 vma->vm_mm, GFP_KERNEL, entry)) {
3331 mem_cgroup_swapin_uncharge_swap(entry);
3333 shadow = get_shadow_from_swap_cache(entry);
3335 workingset_refault(page, shadow);
3337 lru_cache_add(page);
3339 /* To provide entry to swap_readpage() */
3340 set_page_private(page, entry.val);
3341 swap_readpage(page, true);
3342 set_page_private(page, 0);
3345 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3352 * Back out if somebody else faulted in this pte
3353 * while we released the pte lock.
3355 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3356 vmf->address, &vmf->ptl);
3357 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3359 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3363 /* Had to read the page from swap area: Major fault */
3364 ret = VM_FAULT_MAJOR;
3365 count_vm_event(PGMAJFAULT);
3366 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3367 } else if (PageHWPoison(page)) {
3369 * hwpoisoned dirty swapcache pages are kept for killing
3370 * owner processes (which may be unknown at hwpoison time)
3372 ret = VM_FAULT_HWPOISON;
3373 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3377 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3379 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3381 ret |= VM_FAULT_RETRY;
3386 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3387 * release the swapcache from under us. The page pin, and pte_same
3388 * test below, are not enough to exclude that. Even if it is still
3389 * swapcache, we need to check that the page's swap has not changed.
3391 if (unlikely((!PageSwapCache(page) ||
3392 page_private(page) != entry.val)) && swapcache)
3395 page = ksm_might_need_to_copy(page, vma, vmf->address);
3396 if (unlikely(!page)) {
3402 cgroup_throttle_swaprate(page, GFP_KERNEL);
3405 * Back out if somebody else already faulted in this pte.
3407 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3409 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3412 if (unlikely(!PageUptodate(page))) {
3413 ret = VM_FAULT_SIGBUS;
3418 * The page isn't present yet, go ahead with the fault.
3420 * Be careful about the sequence of operations here.
3421 * To get its accounting right, reuse_swap_page() must be called
3422 * while the page is counted on swap but not yet in mapcount i.e.
3423 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3424 * must be called after the swap_free(), or it will never succeed.
3427 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3428 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3429 pte = mk_pte(page, vma->vm_page_prot);
3430 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3431 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3432 vmf->flags &= ~FAULT_FLAG_WRITE;
3433 ret |= VM_FAULT_WRITE;
3434 exclusive = RMAP_EXCLUSIVE;
3436 flush_icache_page(vma, page);
3437 if (pte_swp_soft_dirty(vmf->orig_pte))
3438 pte = pte_mksoft_dirty(pte);
3439 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3440 pte = pte_mkuffd_wp(pte);
3441 pte = pte_wrprotect(pte);
3443 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3444 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3445 vmf->orig_pte = pte;
3447 /* ksm created a completely new copy */
3448 if (unlikely(page != swapcache && swapcache)) {
3449 page_add_new_anon_rmap(page, vma, vmf->address, false);
3450 lru_cache_add_inactive_or_unevictable(page, vma);
3452 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3456 if (mem_cgroup_swap_full(page) ||
3457 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3458 try_to_free_swap(page);
3460 if (page != swapcache && swapcache) {
3462 * Hold the lock to avoid the swap entry to be reused
3463 * until we take the PT lock for the pte_same() check
3464 * (to avoid false positives from pte_same). For
3465 * further safety release the lock after the swap_free
3466 * so that the swap count won't change under a
3467 * parallel locked swapcache.
3469 unlock_page(swapcache);
3470 put_page(swapcache);
3473 if (vmf->flags & FAULT_FLAG_WRITE) {
3474 ret |= do_wp_page(vmf);
3475 if (ret & VM_FAULT_ERROR)
3476 ret &= VM_FAULT_ERROR;
3480 /* No need to invalidate - it was non-present before */
3481 update_mmu_cache(vma, vmf->address, vmf->pte);
3483 pte_unmap_unlock(vmf->pte, vmf->ptl);
3487 pte_unmap_unlock(vmf->pte, vmf->ptl);
3492 if (page != swapcache && swapcache) {
3493 unlock_page(swapcache);
3494 put_page(swapcache);
3500 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3501 * but allow concurrent faults), and pte mapped but not yet locked.
3502 * We return with mmap_lock still held, but pte unmapped and unlocked.
3504 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3506 struct vm_area_struct *vma = vmf->vma;
3511 /* File mapping without ->vm_ops ? */
3512 if (vma->vm_flags & VM_SHARED)
3513 return VM_FAULT_SIGBUS;
3516 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3517 * pte_offset_map() on pmds where a huge pmd might be created
3518 * from a different thread.
3520 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3521 * parallel threads are excluded by other means.
3523 * Here we only have mmap_read_lock(mm).
3525 if (pte_alloc(vma->vm_mm, vmf->pmd))
3526 return VM_FAULT_OOM;
3528 /* See comment in handle_pte_fault() */
3529 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3532 /* Use the zero-page for reads */
3533 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3534 !mm_forbids_zeropage(vma->vm_mm)) {
3535 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3536 vma->vm_page_prot));
3537 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3538 vmf->address, &vmf->ptl);
3539 if (!pte_none(*vmf->pte)) {
3540 update_mmu_tlb(vma, vmf->address, vmf->pte);
3543 ret = check_stable_address_space(vma->vm_mm);
3546 /* Deliver the page fault to userland, check inside PT lock */
3547 if (userfaultfd_missing(vma)) {
3548 pte_unmap_unlock(vmf->pte, vmf->ptl);
3549 return handle_userfault(vmf, VM_UFFD_MISSING);
3554 /* Allocate our own private page. */
3555 if (unlikely(anon_vma_prepare(vma)))
3557 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3561 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3563 cgroup_throttle_swaprate(page, GFP_KERNEL);
3566 * The memory barrier inside __SetPageUptodate makes sure that
3567 * preceding stores to the page contents become visible before
3568 * the set_pte_at() write.
3570 __SetPageUptodate(page);
3572 entry = mk_pte(page, vma->vm_page_prot);
3573 if (vma->vm_flags & VM_WRITE)
3574 entry = pte_mkwrite(pte_mkdirty(entry));
3576 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3578 if (!pte_none(*vmf->pte)) {
3579 update_mmu_cache(vma, vmf->address, vmf->pte);
3583 ret = check_stable_address_space(vma->vm_mm);
3587 /* Deliver the page fault to userland, check inside PT lock */
3588 if (userfaultfd_missing(vma)) {
3589 pte_unmap_unlock(vmf->pte, vmf->ptl);
3591 return handle_userfault(vmf, VM_UFFD_MISSING);
3594 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3595 page_add_new_anon_rmap(page, vma, vmf->address, false);
3596 lru_cache_add_inactive_or_unevictable(page, vma);
3598 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3600 /* No need to invalidate - it was non-present before */
3601 update_mmu_cache(vma, vmf->address, vmf->pte);
3603 pte_unmap_unlock(vmf->pte, vmf->ptl);
3611 return VM_FAULT_OOM;
3615 * The mmap_lock must have been held on entry, and may have been
3616 * released depending on flags and vma->vm_ops->fault() return value.
3617 * See filemap_fault() and __lock_page_retry().
3619 static vm_fault_t __do_fault(struct vm_fault *vmf)
3621 struct vm_area_struct *vma = vmf->vma;
3625 * Preallocate pte before we take page_lock because this might lead to
3626 * deadlocks for memcg reclaim which waits for pages under writeback:
3628 * SetPageWriteback(A)
3634 * wait_on_page_writeback(A)
3635 * SetPageWriteback(B)
3637 * # flush A, B to clear the writeback
3639 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3640 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3641 if (!vmf->prealloc_pte)
3642 return VM_FAULT_OOM;
3643 smp_wmb(); /* See comment in __pte_alloc() */
3646 ret = vma->vm_ops->fault(vmf);
3647 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3648 VM_FAULT_DONE_COW)))
3651 if (unlikely(PageHWPoison(vmf->page))) {
3652 if (ret & VM_FAULT_LOCKED)
3653 unlock_page(vmf->page);
3654 put_page(vmf->page);
3656 return VM_FAULT_HWPOISON;
3659 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3660 lock_page(vmf->page);
3662 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3667 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3668 static void deposit_prealloc_pte(struct vm_fault *vmf)
3670 struct vm_area_struct *vma = vmf->vma;
3672 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3674 * We are going to consume the prealloc table,
3675 * count that as nr_ptes.
3677 mm_inc_nr_ptes(vma->vm_mm);
3678 vmf->prealloc_pte = NULL;
3681 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3683 struct vm_area_struct *vma = vmf->vma;
3684 bool write = vmf->flags & FAULT_FLAG_WRITE;
3685 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3688 vm_fault_t ret = VM_FAULT_FALLBACK;
3690 if (!transhuge_vma_suitable(vma, haddr))
3693 page = compound_head(page);
3694 if (compound_order(page) != HPAGE_PMD_ORDER)
3698 * Archs like ppc64 need additonal space to store information
3699 * related to pte entry. Use the preallocated table for that.
3701 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3702 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3703 if (!vmf->prealloc_pte)
3704 return VM_FAULT_OOM;
3705 smp_wmb(); /* See comment in __pte_alloc() */
3708 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3709 if (unlikely(!pmd_none(*vmf->pmd)))
3712 for (i = 0; i < HPAGE_PMD_NR; i++)
3713 flush_icache_page(vma, page + i);
3715 entry = mk_huge_pmd(page, vma->vm_page_prot);
3717 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3719 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3720 page_add_file_rmap(page, true);
3722 * deposit and withdraw with pmd lock held
3724 if (arch_needs_pgtable_deposit())
3725 deposit_prealloc_pte(vmf);
3727 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3729 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3731 /* fault is handled */
3733 count_vm_event(THP_FILE_MAPPED);
3735 spin_unlock(vmf->ptl);
3739 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3741 return VM_FAULT_FALLBACK;
3745 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3747 struct vm_area_struct *vma = vmf->vma;
3748 bool write = vmf->flags & FAULT_FLAG_WRITE;
3749 bool prefault = vmf->address != addr;
3752 flush_icache_page(vma, page);
3753 entry = mk_pte(page, vma->vm_page_prot);
3755 if (prefault && arch_wants_old_prefaulted_pte())
3756 entry = pte_mkold(entry);
3759 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3760 /* copy-on-write page */
3761 if (write && !(vma->vm_flags & VM_SHARED)) {
3762 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3763 page_add_new_anon_rmap(page, vma, addr, false);
3764 lru_cache_add_inactive_or_unevictable(page, vma);
3766 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3767 page_add_file_rmap(page, false);
3769 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3773 * finish_fault - finish page fault once we have prepared the page to fault
3775 * @vmf: structure describing the fault
3777 * This function handles all that is needed to finish a page fault once the
3778 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3779 * given page, adds reverse page mapping, handles memcg charges and LRU
3782 * The function expects the page to be locked and on success it consumes a
3783 * reference of a page being mapped (for the PTE which maps it).
3785 * Return: %0 on success, %VM_FAULT_ code in case of error.
3787 vm_fault_t finish_fault(struct vm_fault *vmf)
3789 struct vm_area_struct *vma = vmf->vma;
3793 /* Did we COW the page? */
3794 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3795 page = vmf->cow_page;
3800 * check even for read faults because we might have lost our CoWed
3803 if (!(vma->vm_flags & VM_SHARED)) {
3804 ret = check_stable_address_space(vma->vm_mm);
3809 if (pmd_none(*vmf->pmd)) {
3810 if (PageTransCompound(page)) {
3811 ret = do_set_pmd(vmf, page);
3812 if (ret != VM_FAULT_FALLBACK)
3816 if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
3817 return VM_FAULT_OOM;
3820 /* See comment in handle_pte_fault() */
3821 if (pmd_devmap_trans_unstable(vmf->pmd))
3824 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3825 vmf->address, &vmf->ptl);
3827 /* Re-check under ptl */
3828 if (likely(pte_none(*vmf->pte)))
3829 do_set_pte(vmf, page, vmf->address);
3831 ret = VM_FAULT_NOPAGE;
3833 update_mmu_tlb(vma, vmf->address, vmf->pte);
3834 pte_unmap_unlock(vmf->pte, vmf->ptl);
3838 static unsigned long fault_around_bytes __read_mostly =
3839 rounddown_pow_of_two(65536);
3841 #ifdef CONFIG_DEBUG_FS
3842 static int fault_around_bytes_get(void *data, u64 *val)
3844 *val = fault_around_bytes;
3849 * fault_around_bytes must be rounded down to the nearest page order as it's
3850 * what do_fault_around() expects to see.
3852 static int fault_around_bytes_set(void *data, u64 val)
3854 if (val / PAGE_SIZE > PTRS_PER_PTE)
3856 if (val > PAGE_SIZE)
3857 fault_around_bytes = rounddown_pow_of_two(val);
3859 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3862 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3863 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3865 static int __init fault_around_debugfs(void)
3867 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3868 &fault_around_bytes_fops);
3871 late_initcall(fault_around_debugfs);
3875 * do_fault_around() tries to map few pages around the fault address. The hope
3876 * is that the pages will be needed soon and this will lower the number of
3879 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3880 * not ready to be mapped: not up-to-date, locked, etc.
3882 * This function is called with the page table lock taken. In the split ptlock
3883 * case the page table lock only protects only those entries which belong to
3884 * the page table corresponding to the fault address.
3886 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3889 * fault_around_bytes defines how many bytes we'll try to map.
3890 * do_fault_around() expects it to be set to a power of two less than or equal
3893 * The virtual address of the area that we map is naturally aligned to
3894 * fault_around_bytes rounded down to the machine page size
3895 * (and therefore to page order). This way it's easier to guarantee
3896 * that we don't cross page table boundaries.
3898 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3900 unsigned long address = vmf->address, nr_pages, mask;
3901 pgoff_t start_pgoff = vmf->pgoff;
3905 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3906 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3908 address = max(address & mask, vmf->vma->vm_start);
3909 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3913 * end_pgoff is either the end of the page table, the end of
3914 * the vma or nr_pages from start_pgoff, depending what is nearest.
3916 end_pgoff = start_pgoff -
3917 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3919 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3920 start_pgoff + nr_pages - 1);
3922 if (pmd_none(*vmf->pmd)) {
3923 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3924 if (!vmf->prealloc_pte)
3925 return VM_FAULT_OOM;
3926 smp_wmb(); /* See comment in __pte_alloc() */
3929 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3932 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3934 struct vm_area_struct *vma = vmf->vma;
3938 * Let's call ->map_pages() first and use ->fault() as fallback
3939 * if page by the offset is not ready to be mapped (cold cache or
3942 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3943 ret = do_fault_around(vmf);
3948 ret = __do_fault(vmf);
3949 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3952 ret |= finish_fault(vmf);
3953 unlock_page(vmf->page);
3954 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3955 put_page(vmf->page);
3959 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3961 struct vm_area_struct *vma = vmf->vma;
3964 if (unlikely(anon_vma_prepare(vma)))
3965 return VM_FAULT_OOM;
3967 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3969 return VM_FAULT_OOM;
3971 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
3972 put_page(vmf->cow_page);
3973 return VM_FAULT_OOM;
3975 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
3977 ret = __do_fault(vmf);
3978 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3980 if (ret & VM_FAULT_DONE_COW)
3983 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3984 __SetPageUptodate(vmf->cow_page);
3986 ret |= finish_fault(vmf);
3987 unlock_page(vmf->page);
3988 put_page(vmf->page);
3989 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3993 put_page(vmf->cow_page);
3997 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3999 struct vm_area_struct *vma = vmf->vma;
4000 vm_fault_t ret, tmp;
4002 ret = __do_fault(vmf);
4003 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4007 * Check if the backing address space wants to know that the page is
4008 * about to become writable
4010 if (vma->vm_ops->page_mkwrite) {
4011 unlock_page(vmf->page);
4012 tmp = do_page_mkwrite(vmf);
4013 if (unlikely(!tmp ||
4014 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4015 put_page(vmf->page);
4020 ret |= finish_fault(vmf);
4021 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4023 unlock_page(vmf->page);
4024 put_page(vmf->page);
4028 ret |= fault_dirty_shared_page(vmf);
4033 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4034 * but allow concurrent faults).
4035 * The mmap_lock may have been released depending on flags and our
4036 * return value. See filemap_fault() and __lock_page_or_retry().
4037 * If mmap_lock is released, vma may become invalid (for example
4038 * by other thread calling munmap()).
4040 static vm_fault_t do_fault(struct vm_fault *vmf)
4042 struct vm_area_struct *vma = vmf->vma;
4043 struct mm_struct *vm_mm = vma->vm_mm;
4047 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4049 if (!vma->vm_ops->fault) {
4051 * If we find a migration pmd entry or a none pmd entry, which
4052 * should never happen, return SIGBUS
4054 if (unlikely(!pmd_present(*vmf->pmd)))
4055 ret = VM_FAULT_SIGBUS;
4057 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4062 * Make sure this is not a temporary clearing of pte
4063 * by holding ptl and checking again. A R/M/W update
4064 * of pte involves: take ptl, clearing the pte so that
4065 * we don't have concurrent modification by hardware
4066 * followed by an update.
4068 if (unlikely(pte_none(*vmf->pte)))
4069 ret = VM_FAULT_SIGBUS;
4071 ret = VM_FAULT_NOPAGE;
4073 pte_unmap_unlock(vmf->pte, vmf->ptl);
4075 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4076 ret = do_read_fault(vmf);
4077 else if (!(vma->vm_flags & VM_SHARED))
4078 ret = do_cow_fault(vmf);
4080 ret = do_shared_fault(vmf);
4082 /* preallocated pagetable is unused: free it */
4083 if (vmf->prealloc_pte) {
4084 pte_free(vm_mm, vmf->prealloc_pte);
4085 vmf->prealloc_pte = NULL;
4090 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4091 unsigned long addr, int page_nid,
4096 count_vm_numa_event(NUMA_HINT_FAULTS);
4097 if (page_nid == numa_node_id()) {
4098 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4099 *flags |= TNF_FAULT_LOCAL;
4102 return mpol_misplaced(page, vma, addr);
4105 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4107 struct vm_area_struct *vma = vmf->vma;
4108 struct page *page = NULL;
4109 int page_nid = NUMA_NO_NODE;
4113 bool was_writable = pte_savedwrite(vmf->orig_pte);
4117 * The "pte" at this point cannot be used safely without
4118 * validation through pte_unmap_same(). It's of NUMA type but
4119 * the pfn may be screwed if the read is non atomic.
4121 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4122 spin_lock(vmf->ptl);
4123 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4124 pte_unmap_unlock(vmf->pte, vmf->ptl);
4128 /* Get the normal PTE */
4129 old_pte = ptep_get(vmf->pte);
4130 pte = pte_modify(old_pte, vma->vm_page_prot);
4132 page = vm_normal_page(vma, vmf->address, pte);
4136 /* TODO: handle PTE-mapped THP */
4137 if (PageCompound(page))
4141 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4142 * much anyway since they can be in shared cache state. This misses
4143 * the case where a mapping is writable but the process never writes
4144 * to it but pte_write gets cleared during protection updates and
4145 * pte_dirty has unpredictable behaviour between PTE scan updates,
4146 * background writeback, dirty balancing and application behaviour.
4149 flags |= TNF_NO_GROUP;
4152 * Flag if the page is shared between multiple address spaces. This
4153 * is later used when determining whether to group tasks together
4155 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4156 flags |= TNF_SHARED;
4158 last_cpupid = page_cpupid_last(page);
4159 page_nid = page_to_nid(page);
4160 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4162 if (target_nid == NUMA_NO_NODE) {
4166 pte_unmap_unlock(vmf->pte, vmf->ptl);
4168 /* Migrate to the requested node */
4169 if (migrate_misplaced_page(page, vma, target_nid)) {
4170 page_nid = target_nid;
4171 flags |= TNF_MIGRATED;
4173 flags |= TNF_MIGRATE_FAIL;
4174 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4175 spin_lock(vmf->ptl);
4176 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4177 pte_unmap_unlock(vmf->pte, vmf->ptl);
4184 if (page_nid != NUMA_NO_NODE)
4185 task_numa_fault(last_cpupid, page_nid, 1, flags);
4189 * Make it present again, depending on how arch implements
4190 * non-accessible ptes, some can allow access by kernel mode.
4192 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4193 pte = pte_modify(old_pte, vma->vm_page_prot);
4194 pte = pte_mkyoung(pte);
4196 pte = pte_mkwrite(pte);
4197 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4198 update_mmu_cache(vma, vmf->address, vmf->pte);
4199 pte_unmap_unlock(vmf->pte, vmf->ptl);
4203 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4205 if (vma_is_anonymous(vmf->vma))
4206 return do_huge_pmd_anonymous_page(vmf);
4207 if (vmf->vma->vm_ops->huge_fault)
4208 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4209 return VM_FAULT_FALLBACK;
4212 /* `inline' is required to avoid gcc 4.1.2 build error */
4213 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4215 if (vma_is_anonymous(vmf->vma)) {
4216 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4217 return handle_userfault(vmf, VM_UFFD_WP);
4218 return do_huge_pmd_wp_page(vmf, orig_pmd);
4220 if (vmf->vma->vm_ops->huge_fault) {
4221 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4223 if (!(ret & VM_FAULT_FALLBACK))
4227 /* COW or write-notify handled on pte level: split pmd. */
4228 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4230 return VM_FAULT_FALLBACK;
4233 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4235 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4236 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4237 /* No support for anonymous transparent PUD pages yet */
4238 if (vma_is_anonymous(vmf->vma))
4240 if (vmf->vma->vm_ops->huge_fault) {
4241 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4243 if (!(ret & VM_FAULT_FALLBACK))
4247 /* COW or write-notify not handled on PUD level: split pud.*/
4248 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4249 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4250 return VM_FAULT_FALLBACK;
4253 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4255 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4256 /* No support for anonymous transparent PUD pages yet */
4257 if (vma_is_anonymous(vmf->vma))
4258 return VM_FAULT_FALLBACK;
4259 if (vmf->vma->vm_ops->huge_fault)
4260 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4261 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4262 return VM_FAULT_FALLBACK;
4266 * These routines also need to handle stuff like marking pages dirty
4267 * and/or accessed for architectures that don't do it in hardware (most
4268 * RISC architectures). The early dirtying is also good on the i386.
4270 * There is also a hook called "update_mmu_cache()" that architectures
4271 * with external mmu caches can use to update those (ie the Sparc or
4272 * PowerPC hashed page tables that act as extended TLBs).
4274 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4275 * concurrent faults).
4277 * The mmap_lock may have been released depending on flags and our return value.
4278 * See filemap_fault() and __lock_page_or_retry().
4280 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4284 if (unlikely(pmd_none(*vmf->pmd))) {
4286 * Leave __pte_alloc() until later: because vm_ops->fault may
4287 * want to allocate huge page, and if we expose page table
4288 * for an instant, it will be difficult to retract from
4289 * concurrent faults and from rmap lookups.
4294 * If a huge pmd materialized under us just retry later. Use
4295 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4296 * of pmd_trans_huge() to ensure the pmd didn't become
4297 * pmd_trans_huge under us and then back to pmd_none, as a
4298 * result of MADV_DONTNEED running immediately after a huge pmd
4299 * fault in a different thread of this mm, in turn leading to a
4300 * misleading pmd_trans_huge() retval. All we have to ensure is
4301 * that it is a regular pmd that we can walk with
4302 * pte_offset_map() and we can do that through an atomic read
4303 * in C, which is what pmd_trans_unstable() provides.
4305 if (pmd_devmap_trans_unstable(vmf->pmd))
4308 * A regular pmd is established and it can't morph into a huge
4309 * pmd from under us anymore at this point because we hold the
4310 * mmap_lock read mode and khugepaged takes it in write mode.
4311 * So now it's safe to run pte_offset_map().
4313 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4314 vmf->orig_pte = *vmf->pte;
4317 * some architectures can have larger ptes than wordsize,
4318 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4319 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4320 * accesses. The code below just needs a consistent view
4321 * for the ifs and we later double check anyway with the
4322 * ptl lock held. So here a barrier will do.
4325 if (pte_none(vmf->orig_pte)) {
4326 pte_unmap(vmf->pte);
4332 if (vma_is_anonymous(vmf->vma))
4333 return do_anonymous_page(vmf);
4335 return do_fault(vmf);
4338 if (!pte_present(vmf->orig_pte))
4339 return do_swap_page(vmf);
4341 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4342 return do_numa_page(vmf);
4344 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4345 spin_lock(vmf->ptl);
4346 entry = vmf->orig_pte;
4347 if (unlikely(!pte_same(*vmf->pte, entry))) {
4348 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4351 if (vmf->flags & FAULT_FLAG_WRITE) {
4352 if (!pte_write(entry))
4353 return do_wp_page(vmf);
4354 entry = pte_mkdirty(entry);
4356 entry = pte_mkyoung(entry);
4357 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4358 vmf->flags & FAULT_FLAG_WRITE)) {
4359 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4361 /* Skip spurious TLB flush for retried page fault */
4362 if (vmf->flags & FAULT_FLAG_TRIED)
4365 * This is needed only for protection faults but the arch code
4366 * is not yet telling us if this is a protection fault or not.
4367 * This still avoids useless tlb flushes for .text page faults
4370 if (vmf->flags & FAULT_FLAG_WRITE)
4371 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4374 pte_unmap_unlock(vmf->pte, vmf->ptl);
4379 * By the time we get here, we already hold the mm semaphore
4381 * The mmap_lock may have been released depending on flags and our
4382 * return value. See filemap_fault() and __lock_page_or_retry().
4384 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4385 unsigned long address, unsigned int flags)
4387 struct vm_fault vmf = {
4389 .address = address & PAGE_MASK,
4391 .pgoff = linear_page_index(vma, address),
4392 .gfp_mask = __get_fault_gfp_mask(vma),
4394 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4395 struct mm_struct *mm = vma->vm_mm;
4400 pgd = pgd_offset(mm, address);
4401 p4d = p4d_alloc(mm, pgd, address);
4403 return VM_FAULT_OOM;
4405 vmf.pud = pud_alloc(mm, p4d, address);
4407 return VM_FAULT_OOM;
4409 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4410 ret = create_huge_pud(&vmf);
4411 if (!(ret & VM_FAULT_FALLBACK))
4414 pud_t orig_pud = *vmf.pud;
4417 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4419 /* NUMA case for anonymous PUDs would go here */
4421 if (dirty && !pud_write(orig_pud)) {
4422 ret = wp_huge_pud(&vmf, orig_pud);
4423 if (!(ret & VM_FAULT_FALLBACK))
4426 huge_pud_set_accessed(&vmf, orig_pud);
4432 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4434 return VM_FAULT_OOM;
4436 /* Huge pud page fault raced with pmd_alloc? */
4437 if (pud_trans_unstable(vmf.pud))
4440 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4441 ret = create_huge_pmd(&vmf);
4442 if (!(ret & VM_FAULT_FALLBACK))
4445 pmd_t orig_pmd = *vmf.pmd;
4448 if (unlikely(is_swap_pmd(orig_pmd))) {
4449 VM_BUG_ON(thp_migration_supported() &&
4450 !is_pmd_migration_entry(orig_pmd));
4451 if (is_pmd_migration_entry(orig_pmd))
4452 pmd_migration_entry_wait(mm, vmf.pmd);
4455 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4456 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4457 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4459 if (dirty && !pmd_write(orig_pmd)) {
4460 ret = wp_huge_pmd(&vmf, orig_pmd);
4461 if (!(ret & VM_FAULT_FALLBACK))
4464 huge_pmd_set_accessed(&vmf, orig_pmd);
4470 return handle_pte_fault(&vmf);
4474 * mm_account_fault - Do page fault accountings
4476 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4477 * of perf event counters, but we'll still do the per-task accounting to
4478 * the task who triggered this page fault.
4479 * @address: the faulted address.
4480 * @flags: the fault flags.
4481 * @ret: the fault retcode.
4483 * This will take care of most of the page fault accountings. Meanwhile, it
4484 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4485 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4486 * still be in per-arch page fault handlers at the entry of page fault.
4488 static inline void mm_account_fault(struct pt_regs *regs,
4489 unsigned long address, unsigned int flags,
4495 * We don't do accounting for some specific faults:
4497 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4498 * includes arch_vma_access_permitted() failing before reaching here.
4499 * So this is not a "this many hardware page faults" counter. We
4500 * should use the hw profiling for that.
4502 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4503 * once they're completed.
4505 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4509 * We define the fault as a major fault when the final successful fault
4510 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4511 * handle it immediately previously).
4513 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4521 * If the fault is done for GUP, regs will be NULL. We only do the
4522 * accounting for the per thread fault counters who triggered the
4523 * fault, and we skip the perf event updates.
4529 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4531 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4535 * By the time we get here, we already hold the mm semaphore
4537 * The mmap_lock may have been released depending on flags and our
4538 * return value. See filemap_fault() and __lock_page_or_retry().
4540 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4541 unsigned int flags, struct pt_regs *regs)
4545 __set_current_state(TASK_RUNNING);
4547 count_vm_event(PGFAULT);
4548 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4550 /* do counter updates before entering really critical section. */
4551 check_sync_rss_stat(current);
4553 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4554 flags & FAULT_FLAG_INSTRUCTION,
4555 flags & FAULT_FLAG_REMOTE))
4556 return VM_FAULT_SIGSEGV;
4559 * Enable the memcg OOM handling for faults triggered in user
4560 * space. Kernel faults are handled more gracefully.
4562 if (flags & FAULT_FLAG_USER)
4563 mem_cgroup_enter_user_fault();
4565 if (unlikely(is_vm_hugetlb_page(vma)))
4566 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4568 ret = __handle_mm_fault(vma, address, flags);
4570 if (flags & FAULT_FLAG_USER) {
4571 mem_cgroup_exit_user_fault();
4573 * The task may have entered a memcg OOM situation but
4574 * if the allocation error was handled gracefully (no
4575 * VM_FAULT_OOM), there is no need to kill anything.
4576 * Just clean up the OOM state peacefully.
4578 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4579 mem_cgroup_oom_synchronize(false);
4582 mm_account_fault(regs, address, flags, ret);
4586 EXPORT_SYMBOL_GPL(handle_mm_fault);
4588 #ifndef __PAGETABLE_P4D_FOLDED
4590 * Allocate p4d page table.
4591 * We've already handled the fast-path in-line.
4593 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4595 p4d_t *new = p4d_alloc_one(mm, address);
4599 smp_wmb(); /* See comment in __pte_alloc */
4601 spin_lock(&mm->page_table_lock);
4602 if (pgd_present(*pgd)) /* Another has populated it */
4605 pgd_populate(mm, pgd, new);
4606 spin_unlock(&mm->page_table_lock);
4609 #endif /* __PAGETABLE_P4D_FOLDED */
4611 #ifndef __PAGETABLE_PUD_FOLDED
4613 * Allocate page upper directory.
4614 * We've already handled the fast-path in-line.
4616 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4618 pud_t *new = pud_alloc_one(mm, address);
4622 smp_wmb(); /* See comment in __pte_alloc */
4624 spin_lock(&mm->page_table_lock);
4625 if (!p4d_present(*p4d)) {
4627 p4d_populate(mm, p4d, new);
4628 } else /* Another has populated it */
4630 spin_unlock(&mm->page_table_lock);
4633 #endif /* __PAGETABLE_PUD_FOLDED */
4635 #ifndef __PAGETABLE_PMD_FOLDED
4637 * Allocate page middle directory.
4638 * We've already handled the fast-path in-line.
4640 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4643 pmd_t *new = pmd_alloc_one(mm, address);
4647 smp_wmb(); /* See comment in __pte_alloc */
4649 ptl = pud_lock(mm, pud);
4650 if (!pud_present(*pud)) {
4652 pud_populate(mm, pud, new);
4653 } else /* Another has populated it */
4658 #endif /* __PAGETABLE_PMD_FOLDED */
4660 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4661 struct mmu_notifier_range *range, pte_t **ptepp,
4662 pmd_t **pmdpp, spinlock_t **ptlp)
4670 pgd = pgd_offset(mm, address);
4671 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4674 p4d = p4d_offset(pgd, address);
4675 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4678 pud = pud_offset(p4d, address);
4679 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4682 pmd = pmd_offset(pud, address);
4683 VM_BUG_ON(pmd_trans_huge(*pmd));
4685 if (pmd_huge(*pmd)) {
4690 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4691 NULL, mm, address & PMD_MASK,
4692 (address & PMD_MASK) + PMD_SIZE);
4693 mmu_notifier_invalidate_range_start(range);
4695 *ptlp = pmd_lock(mm, pmd);
4696 if (pmd_huge(*pmd)) {
4702 mmu_notifier_invalidate_range_end(range);
4705 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4709 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4710 address & PAGE_MASK,
4711 (address & PAGE_MASK) + PAGE_SIZE);
4712 mmu_notifier_invalidate_range_start(range);
4714 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4715 if (!pte_present(*ptep))
4720 pte_unmap_unlock(ptep, *ptlp);
4722 mmu_notifier_invalidate_range_end(range);
4728 * follow_pte - look up PTE at a user virtual address
4729 * @mm: the mm_struct of the target address space
4730 * @address: user virtual address
4731 * @ptepp: location to store found PTE
4732 * @ptlp: location to store the lock for the PTE
4734 * On a successful return, the pointer to the PTE is stored in @ptepp;
4735 * the corresponding lock is taken and its location is stored in @ptlp.
4736 * The contents of the PTE are only stable until @ptlp is released;
4737 * any further use, if any, must be protected against invalidation
4738 * with MMU notifiers.
4740 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4741 * should be taken for read.
4743 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4744 * it is not a good general-purpose API.
4746 * Return: zero on success, -ve otherwise.
4748 int follow_pte(struct mm_struct *mm, unsigned long address,
4749 pte_t **ptepp, spinlock_t **ptlp)
4751 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4753 EXPORT_SYMBOL_GPL(follow_pte);
4756 * follow_pfn - look up PFN at a user virtual address
4757 * @vma: memory mapping
4758 * @address: user virtual address
4759 * @pfn: location to store found PFN
4761 * Only IO mappings and raw PFN mappings are allowed.
4763 * This function does not allow the caller to read the permissions
4764 * of the PTE. Do not use it.
4766 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4768 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4775 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4778 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4781 *pfn = pte_pfn(*ptep);
4782 pte_unmap_unlock(ptep, ptl);
4785 EXPORT_SYMBOL(follow_pfn);
4787 #ifdef CONFIG_HAVE_IOREMAP_PROT
4788 int follow_phys(struct vm_area_struct *vma,
4789 unsigned long address, unsigned int flags,
4790 unsigned long *prot, resource_size_t *phys)
4796 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4799 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4803 if ((flags & FOLL_WRITE) && !pte_write(pte))
4806 *prot = pgprot_val(pte_pgprot(pte));
4807 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4811 pte_unmap_unlock(ptep, ptl);
4817 * generic_access_phys - generic implementation for iomem mmap access
4818 * @vma: the vma to access
4819 * @addr: userspace addres, not relative offset within @vma
4820 * @buf: buffer to read/write
4821 * @len: length of transfer
4822 * @write: set to FOLL_WRITE when writing, otherwise reading
4824 * This is a generic implementation for &vm_operations_struct.access for an
4825 * iomem mapping. This callback is used by access_process_vm() when the @vma is
4828 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4829 void *buf, int len, int write)
4831 resource_size_t phys_addr;
4832 unsigned long prot = 0;
4833 void __iomem *maddr;
4836 int offset = offset_in_page(addr);
4839 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4843 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4846 pte_unmap_unlock(ptep, ptl);
4848 prot = pgprot_val(pte_pgprot(pte));
4849 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4851 if ((write & FOLL_WRITE) && !pte_write(pte))
4854 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4858 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4861 if (!pte_same(pte, *ptep)) {
4862 pte_unmap_unlock(ptep, ptl);
4869 memcpy_toio(maddr + offset, buf, len);
4871 memcpy_fromio(buf, maddr + offset, len);
4873 pte_unmap_unlock(ptep, ptl);
4879 EXPORT_SYMBOL_GPL(generic_access_phys);
4883 * Access another process' address space as given in mm.
4885 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
4886 int len, unsigned int gup_flags)
4888 struct vm_area_struct *vma;
4889 void *old_buf = buf;
4890 int write = gup_flags & FOLL_WRITE;
4892 if (mmap_read_lock_killable(mm))
4895 /* ignore errors, just check how much was successfully transferred */
4897 int bytes, ret, offset;
4899 struct page *page = NULL;
4901 ret = get_user_pages_remote(mm, addr, 1,
4902 gup_flags, &page, &vma, NULL);
4904 #ifndef CONFIG_HAVE_IOREMAP_PROT
4908 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4909 * we can access using slightly different code.
4911 vma = find_vma(mm, addr);
4912 if (!vma || vma->vm_start > addr)
4914 if (vma->vm_ops && vma->vm_ops->access)
4915 ret = vma->vm_ops->access(vma, addr, buf,
4923 offset = addr & (PAGE_SIZE-1);
4924 if (bytes > PAGE_SIZE-offset)
4925 bytes = PAGE_SIZE-offset;
4929 copy_to_user_page(vma, page, addr,
4930 maddr + offset, buf, bytes);
4931 set_page_dirty_lock(page);
4933 copy_from_user_page(vma, page, addr,
4934 buf, maddr + offset, bytes);
4943 mmap_read_unlock(mm);
4945 return buf - old_buf;
4949 * access_remote_vm - access another process' address space
4950 * @mm: the mm_struct of the target address space
4951 * @addr: start address to access
4952 * @buf: source or destination buffer
4953 * @len: number of bytes to transfer
4954 * @gup_flags: flags modifying lookup behaviour
4956 * The caller must hold a reference on @mm.
4958 * Return: number of bytes copied from source to destination.
4960 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4961 void *buf, int len, unsigned int gup_flags)
4963 return __access_remote_vm(mm, addr, buf, len, gup_flags);
4967 * Access another process' address space.
4968 * Source/target buffer must be kernel space,
4969 * Do not walk the page table directly, use get_user_pages
4971 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4972 void *buf, int len, unsigned int gup_flags)
4974 struct mm_struct *mm;
4977 mm = get_task_mm(tsk);
4981 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
4987 EXPORT_SYMBOL_GPL(access_process_vm);
4990 * Print the name of a VMA.
4992 void print_vma_addr(char *prefix, unsigned long ip)
4994 struct mm_struct *mm = current->mm;
4995 struct vm_area_struct *vma;
4998 * we might be running from an atomic context so we cannot sleep
5000 if (!mmap_read_trylock(mm))
5003 vma = find_vma(mm, ip);
5004 if (vma && vma->vm_file) {
5005 struct file *f = vma->vm_file;
5006 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5010 p = file_path(f, buf, PAGE_SIZE);
5013 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5015 vma->vm_end - vma->vm_start);
5016 free_page((unsigned long)buf);
5019 mmap_read_unlock(mm);
5022 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5023 void __might_fault(const char *file, int line)
5026 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5027 * holding the mmap_lock, this is safe because kernel memory doesn't
5028 * get paged out, therefore we'll never actually fault, and the
5029 * below annotations will generate false positives.
5031 if (uaccess_kernel())
5033 if (pagefault_disabled())
5035 __might_sleep(file, line, 0);
5036 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5038 might_lock_read(¤t->mm->mmap_lock);
5041 EXPORT_SYMBOL(__might_fault);
5044 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5046 * Process all subpages of the specified huge page with the specified
5047 * operation. The target subpage will be processed last to keep its
5050 static inline void process_huge_page(
5051 unsigned long addr_hint, unsigned int pages_per_huge_page,
5052 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5056 unsigned long addr = addr_hint &
5057 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5059 /* Process target subpage last to keep its cache lines hot */
5061 n = (addr_hint - addr) / PAGE_SIZE;
5062 if (2 * n <= pages_per_huge_page) {
5063 /* If target subpage in first half of huge page */
5066 /* Process subpages at the end of huge page */
5067 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5069 process_subpage(addr + i * PAGE_SIZE, i, arg);
5072 /* If target subpage in second half of huge page */
5073 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5074 l = pages_per_huge_page - n;
5075 /* Process subpages at the begin of huge page */
5076 for (i = 0; i < base; i++) {
5078 process_subpage(addr + i * PAGE_SIZE, i, arg);
5082 * Process remaining subpages in left-right-left-right pattern
5083 * towards the target subpage
5085 for (i = 0; i < l; i++) {
5086 int left_idx = base + i;
5087 int right_idx = base + 2 * l - 1 - i;
5090 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5092 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5096 static void clear_gigantic_page(struct page *page,
5098 unsigned int pages_per_huge_page)
5101 struct page *p = page;
5104 for (i = 0; i < pages_per_huge_page;
5105 i++, p = mem_map_next(p, page, i)) {
5107 clear_user_highpage(p, addr + i * PAGE_SIZE);
5111 static void clear_subpage(unsigned long addr, int idx, void *arg)
5113 struct page *page = arg;
5115 clear_user_highpage(page + idx, addr);
5118 void clear_huge_page(struct page *page,
5119 unsigned long addr_hint, unsigned int pages_per_huge_page)
5121 unsigned long addr = addr_hint &
5122 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5124 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5125 clear_gigantic_page(page, addr, pages_per_huge_page);
5129 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5132 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5134 struct vm_area_struct *vma,
5135 unsigned int pages_per_huge_page)
5138 struct page *dst_base = dst;
5139 struct page *src_base = src;
5141 for (i = 0; i < pages_per_huge_page; ) {
5143 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5146 dst = mem_map_next(dst, dst_base, i);
5147 src = mem_map_next(src, src_base, i);
5151 struct copy_subpage_arg {
5154 struct vm_area_struct *vma;
5157 static void copy_subpage(unsigned long addr, int idx, void *arg)
5159 struct copy_subpage_arg *copy_arg = arg;
5161 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5162 addr, copy_arg->vma);
5165 void copy_user_huge_page(struct page *dst, struct page *src,
5166 unsigned long addr_hint, struct vm_area_struct *vma,
5167 unsigned int pages_per_huge_page)
5169 unsigned long addr = addr_hint &
5170 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5171 struct copy_subpage_arg arg = {
5177 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5178 copy_user_gigantic_page(dst, src, addr, vma,
5179 pages_per_huge_page);
5183 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5186 long copy_huge_page_from_user(struct page *dst_page,
5187 const void __user *usr_src,
5188 unsigned int pages_per_huge_page,
5189 bool allow_pagefault)
5191 void *src = (void *)usr_src;
5193 unsigned long i, rc = 0;
5194 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5195 struct page *subpage = dst_page;
5197 for (i = 0; i < pages_per_huge_page;
5198 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5199 if (allow_pagefault)
5200 page_kaddr = kmap(subpage);
5202 page_kaddr = kmap_atomic(subpage);
5203 rc = copy_from_user(page_kaddr,
5204 (const void __user *)(src + i * PAGE_SIZE),
5206 if (allow_pagefault)
5209 kunmap_atomic(page_kaddr);
5211 ret_val -= (PAGE_SIZE - rc);
5219 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5221 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5223 static struct kmem_cache *page_ptl_cachep;
5225 void __init ptlock_cache_init(void)
5227 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5231 bool ptlock_alloc(struct page *page)
5235 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5242 void ptlock_free(struct page *page)
5244 kmem_cache_free(page_ptl_cachep, page->ptl);