4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/export.h>
55 #include <linux/delayacct.h>
56 #include <linux/init.h>
57 #include <linux/pfn_t.h>
58 #include <linux/writeback.h>
59 #include <linux/memcontrol.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/kallsyms.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
93 EXPORT_SYMBOL(mem_map);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
119 static int __init disable_randmaps(char *s)
121 randomize_va_space = 0;
124 __setup("norandmaps", disable_randmaps);
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
129 unsigned long highest_memmap_pfn __read_mostly;
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 static int __init init_zero_pfn(void)
136 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 core_initcall(init_zero_pfn);
142 #if defined(SPLIT_RSS_COUNTING)
144 void sync_mm_rss(struct mm_struct *mm)
148 for (i = 0; i < NR_MM_COUNTERS; i++) {
149 if (current->rss_stat.count[i]) {
150 add_mm_counter(mm, i, current->rss_stat.count[i]);
151 current->rss_stat.count[i] = 0;
154 current->rss_stat.events = 0;
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 struct task_struct *task = current;
161 if (likely(task->mm == mm))
162 task->rss_stat.count[member] += val;
164 add_mm_counter(mm, member, val);
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH (64)
171 static void check_sync_rss_stat(struct task_struct *task)
173 if (unlikely(task != current))
175 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176 sync_mm_rss(task->mm);
178 #else /* SPLIT_RSS_COUNTING */
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183 static void check_sync_rss_stat(struct task_struct *task)
187 #endif /* SPLIT_RSS_COUNTING */
189 #ifdef HAVE_GENERIC_MMU_GATHER
191 static bool tlb_next_batch(struct mmu_gather *tlb)
193 struct mmu_gather_batch *batch;
197 tlb->active = batch->next;
201 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
204 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
211 batch->max = MAX_GATHER_BATCH;
213 tlb->active->next = batch;
219 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220 unsigned long start, unsigned long end)
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238 __tlb_reset_range(tlb);
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
251 __tlb_reset_range(tlb);
254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
256 struct mmu_gather_batch *batch;
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
262 tlb->active = &tlb->local;
265 void tlb_flush_mmu(struct mmu_gather *tlb)
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276 unsigned long start, unsigned long end, bool force)
278 struct mmu_gather_batch *batch, *next;
281 __tlb_adjust_range(tlb, start, end - start);
285 /* keep the page table cache within bounds */
288 for (batch = tlb->local.next; batch; batch = next) {
290 free_pages((unsigned long)batch, 0);
292 tlb->local.next = NULL;
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
302 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 struct mmu_gather_batch *batch;
306 VM_BUG_ON(!tlb->end);
307 VM_WARN_ON(tlb->page_size != page_size);
311 * Add the page and check if we are full. If so
314 batch->pages[batch->nr++] = page;
315 if (batch->nr == batch->max) {
316 if (!tlb_next_batch(tlb))
320 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
325 #endif /* HAVE_GENERIC_MMU_GATHER */
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
330 * See the comment near struct mmu_table_batch.
333 static void tlb_remove_table_smp_sync(void *arg)
335 /* Simply deliver the interrupt */
338 static void tlb_remove_table_one(void *table)
341 * This isn't an RCU grace period and hence the page-tables cannot be
342 * assumed to be actually RCU-freed.
344 * It is however sufficient for software page-table walkers that rely on
345 * IRQ disabling. See the comment near struct mmu_table_batch.
347 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
348 __tlb_remove_table(table);
351 static void tlb_remove_table_rcu(struct rcu_head *head)
353 struct mmu_table_batch *batch;
356 batch = container_of(head, struct mmu_table_batch, rcu);
358 for (i = 0; i < batch->nr; i++)
359 __tlb_remove_table(batch->tables[i]);
361 free_page((unsigned long)batch);
364 void tlb_table_flush(struct mmu_gather *tlb)
366 struct mmu_table_batch **batch = &tlb->batch;
369 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
374 void tlb_remove_table(struct mmu_gather *tlb, void *table)
376 struct mmu_table_batch **batch = &tlb->batch;
379 * When there's less then two users of this mm there cannot be a
380 * concurrent page-table walk.
382 if (atomic_read(&tlb->mm->mm_users) < 2) {
383 __tlb_remove_table(table);
387 if (*batch == NULL) {
388 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
389 if (*batch == NULL) {
390 tlb_remove_table_one(table);
395 (*batch)->tables[(*batch)->nr++] = table;
396 if ((*batch)->nr == MAX_TABLE_BATCH)
397 tlb_table_flush(tlb);
400 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
403 * Called to initialize an (on-stack) mmu_gather structure for page-table
404 * tear-down from @mm. The @fullmm argument is used when @mm is without
405 * users and we're going to destroy the full address space (exit/execve).
407 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
408 unsigned long start, unsigned long end)
410 arch_tlb_gather_mmu(tlb, mm, start, end);
411 inc_tlb_flush_pending(tlb->mm);
414 void tlb_finish_mmu(struct mmu_gather *tlb,
415 unsigned long start, unsigned long end)
418 * If there are parallel threads are doing PTE changes on same range
419 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
420 * flush by batching, a thread has stable TLB entry can fail to flush
421 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
422 * forcefully if we detect parallel PTE batching threads.
424 bool force = mm_tlb_flush_nested(tlb->mm);
426 arch_tlb_finish_mmu(tlb, start, end, force);
427 dec_tlb_flush_pending(tlb->mm);
431 * Note: this doesn't free the actual pages themselves. That
432 * has been handled earlier when unmapping all the memory regions.
434 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
437 pgtable_t token = pmd_pgtable(*pmd);
439 pte_free_tlb(tlb, token, addr);
440 atomic_long_dec(&tlb->mm->nr_ptes);
443 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
444 unsigned long addr, unsigned long end,
445 unsigned long floor, unsigned long ceiling)
452 pmd = pmd_offset(pud, addr);
454 next = pmd_addr_end(addr, end);
455 if (pmd_none_or_clear_bad(pmd))
457 free_pte_range(tlb, pmd, addr);
458 } while (pmd++, addr = next, addr != end);
468 if (end - 1 > ceiling - 1)
471 pmd = pmd_offset(pud, start);
473 pmd_free_tlb(tlb, pmd, start);
474 mm_dec_nr_pmds(tlb->mm);
477 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
478 unsigned long addr, unsigned long end,
479 unsigned long floor, unsigned long ceiling)
486 pud = pud_offset(p4d, addr);
488 next = pud_addr_end(addr, end);
489 if (pud_none_or_clear_bad(pud))
491 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
492 } while (pud++, addr = next, addr != end);
502 if (end - 1 > ceiling - 1)
505 pud = pud_offset(p4d, start);
507 pud_free_tlb(tlb, pud, start);
510 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
511 unsigned long addr, unsigned long end,
512 unsigned long floor, unsigned long ceiling)
519 p4d = p4d_offset(pgd, addr);
521 next = p4d_addr_end(addr, end);
522 if (p4d_none_or_clear_bad(p4d))
524 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
525 } while (p4d++, addr = next, addr != end);
531 ceiling &= PGDIR_MASK;
535 if (end - 1 > ceiling - 1)
538 p4d = p4d_offset(pgd, start);
540 p4d_free_tlb(tlb, p4d, start);
544 * This function frees user-level page tables of a process.
546 void free_pgd_range(struct mmu_gather *tlb,
547 unsigned long addr, unsigned long end,
548 unsigned long floor, unsigned long ceiling)
554 * The next few lines have given us lots of grief...
556 * Why are we testing PMD* at this top level? Because often
557 * there will be no work to do at all, and we'd prefer not to
558 * go all the way down to the bottom just to discover that.
560 * Why all these "- 1"s? Because 0 represents both the bottom
561 * of the address space and the top of it (using -1 for the
562 * top wouldn't help much: the masks would do the wrong thing).
563 * The rule is that addr 0 and floor 0 refer to the bottom of
564 * the address space, but end 0 and ceiling 0 refer to the top
565 * Comparisons need to use "end - 1" and "ceiling - 1" (though
566 * that end 0 case should be mythical).
568 * Wherever addr is brought up or ceiling brought down, we must
569 * be careful to reject "the opposite 0" before it confuses the
570 * subsequent tests. But what about where end is brought down
571 * by PMD_SIZE below? no, end can't go down to 0 there.
573 * Whereas we round start (addr) and ceiling down, by different
574 * masks at different levels, in order to test whether a table
575 * now has no other vmas using it, so can be freed, we don't
576 * bother to round floor or end up - the tests don't need that.
590 if (end - 1 > ceiling - 1)
595 * We add page table cache pages with PAGE_SIZE,
596 * (see pte_free_tlb()), flush the tlb if we need
598 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
599 pgd = pgd_offset(tlb->mm, addr);
601 next = pgd_addr_end(addr, end);
602 if (pgd_none_or_clear_bad(pgd))
604 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
605 } while (pgd++, addr = next, addr != end);
608 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
609 unsigned long floor, unsigned long ceiling)
612 struct vm_area_struct *next = vma->vm_next;
613 unsigned long addr = vma->vm_start;
616 * Hide vma from rmap and truncate_pagecache before freeing
619 unlink_anon_vmas(vma);
620 unlink_file_vma(vma);
622 if (is_vm_hugetlb_page(vma)) {
623 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
624 floor, next ? next->vm_start : ceiling);
627 * Optimization: gather nearby vmas into one call down
629 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
630 && !is_vm_hugetlb_page(next)) {
633 unlink_anon_vmas(vma);
634 unlink_file_vma(vma);
636 free_pgd_range(tlb, addr, vma->vm_end,
637 floor, next ? next->vm_start : ceiling);
643 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
646 pgtable_t new = pte_alloc_one(mm, address);
651 * Ensure all pte setup (eg. pte page lock and page clearing) are
652 * visible before the pte is made visible to other CPUs by being
653 * put into page tables.
655 * The other side of the story is the pointer chasing in the page
656 * table walking code (when walking the page table without locking;
657 * ie. most of the time). Fortunately, these data accesses consist
658 * of a chain of data-dependent loads, meaning most CPUs (alpha
659 * being the notable exception) will already guarantee loads are
660 * seen in-order. See the alpha page table accessors for the
661 * smp_read_barrier_depends() barriers in page table walking code.
663 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
665 ptl = pmd_lock(mm, pmd);
666 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
667 atomic_long_inc(&mm->nr_ptes);
668 pmd_populate(mm, pmd, new);
677 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
679 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
683 smp_wmb(); /* See comment in __pte_alloc */
685 spin_lock(&init_mm.page_table_lock);
686 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
687 pmd_populate_kernel(&init_mm, pmd, new);
690 spin_unlock(&init_mm.page_table_lock);
692 pte_free_kernel(&init_mm, new);
696 static inline void init_rss_vec(int *rss)
698 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
701 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
705 if (current->mm == mm)
707 for (i = 0; i < NR_MM_COUNTERS; i++)
709 add_mm_counter(mm, i, rss[i]);
713 * This function is called to print an error when a bad pte
714 * is found. For example, we might have a PFN-mapped pte in
715 * a region that doesn't allow it.
717 * The calling function must still handle the error.
719 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
720 pte_t pte, struct page *page)
722 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
723 p4d_t *p4d = p4d_offset(pgd, addr);
724 pud_t *pud = pud_offset(p4d, addr);
725 pmd_t *pmd = pmd_offset(pud, addr);
726 struct address_space *mapping;
728 static unsigned long resume;
729 static unsigned long nr_shown;
730 static unsigned long nr_unshown;
733 * Allow a burst of 60 reports, then keep quiet for that minute;
734 * or allow a steady drip of one report per second.
736 if (nr_shown == 60) {
737 if (time_before(jiffies, resume)) {
742 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
749 resume = jiffies + 60 * HZ;
751 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
752 index = linear_page_index(vma, addr);
754 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
756 (long long)pte_val(pte), (long long)pmd_val(*pmd));
758 dump_page(page, "bad pte");
759 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
760 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
762 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
764 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
766 vma->vm_ops ? vma->vm_ops->fault : NULL,
767 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
768 mapping ? mapping->a_ops->readpage : NULL);
770 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
774 * vm_normal_page -- This function gets the "struct page" associated with a pte.
776 * "Special" mappings do not wish to be associated with a "struct page" (either
777 * it doesn't exist, or it exists but they don't want to touch it). In this
778 * case, NULL is returned here. "Normal" mappings do have a struct page.
780 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
781 * pte bit, in which case this function is trivial. Secondly, an architecture
782 * may not have a spare pte bit, which requires a more complicated scheme,
785 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
786 * special mapping (even if there are underlying and valid "struct pages").
787 * COWed pages of a VM_PFNMAP are always normal.
789 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
790 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
791 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
792 * mapping will always honor the rule
794 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
796 * And for normal mappings this is false.
798 * This restricts such mappings to be a linear translation from virtual address
799 * to pfn. To get around this restriction, we allow arbitrary mappings so long
800 * as the vma is not a COW mapping; in that case, we know that all ptes are
801 * special (because none can have been COWed).
804 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
806 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
807 * page" backing, however the difference is that _all_ pages with a struct
808 * page (that is, those where pfn_valid is true) are refcounted and considered
809 * normal pages by the VM. The disadvantage is that pages are refcounted
810 * (which can be slower and simply not an option for some PFNMAP users). The
811 * advantage is that we don't have to follow the strict linearity rule of
812 * PFNMAP mappings in order to support COWable mappings.
815 #ifdef __HAVE_ARCH_PTE_SPECIAL
816 # define HAVE_PTE_SPECIAL 1
818 # define HAVE_PTE_SPECIAL 0
820 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
823 unsigned long pfn = pte_pfn(pte);
825 if (HAVE_PTE_SPECIAL) {
826 if (likely(!pte_special(pte)))
828 if (vma->vm_ops && vma->vm_ops->find_special_page)
829 return vma->vm_ops->find_special_page(vma, addr);
830 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
832 if (!is_zero_pfn(pfn))
833 print_bad_pte(vma, addr, pte, NULL);
837 /* !HAVE_PTE_SPECIAL case follows: */
839 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
840 if (vma->vm_flags & VM_MIXEDMAP) {
846 off = (addr - vma->vm_start) >> PAGE_SHIFT;
847 if (pfn == vma->vm_pgoff + off)
849 if (!is_cow_mapping(vma->vm_flags))
854 if (is_zero_pfn(pfn))
857 if (unlikely(pfn > highest_memmap_pfn)) {
858 print_bad_pte(vma, addr, pte, NULL);
863 * NOTE! We still have PageReserved() pages in the page tables.
864 * eg. VDSO mappings can cause them to exist.
867 return pfn_to_page(pfn);
870 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
871 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
874 unsigned long pfn = pmd_pfn(pmd);
877 * There is no pmd_special() but there may be special pmds, e.g.
878 * in a direct-access (dax) mapping, so let's just replicate the
879 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
881 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
882 if (vma->vm_flags & VM_MIXEDMAP) {
888 off = (addr - vma->vm_start) >> PAGE_SHIFT;
889 if (pfn == vma->vm_pgoff + off)
891 if (!is_cow_mapping(vma->vm_flags))
896 if (is_zero_pfn(pfn))
898 if (unlikely(pfn > highest_memmap_pfn))
902 * NOTE! We still have PageReserved() pages in the page tables.
903 * eg. VDSO mappings can cause them to exist.
906 return pfn_to_page(pfn);
911 * copy one vm_area from one task to the other. Assumes the page tables
912 * already present in the new task to be cleared in the whole range
913 * covered by this vma.
916 static inline unsigned long
917 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
918 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
919 unsigned long addr, int *rss)
921 unsigned long vm_flags = vma->vm_flags;
922 pte_t pte = *src_pte;
925 /* pte contains position in swap or file, so copy. */
926 if (unlikely(!pte_present(pte))) {
927 swp_entry_t entry = pte_to_swp_entry(pte);
929 if (likely(!non_swap_entry(entry))) {
930 if (swap_duplicate(entry) < 0)
933 /* make sure dst_mm is on swapoff's mmlist. */
934 if (unlikely(list_empty(&dst_mm->mmlist))) {
935 spin_lock(&mmlist_lock);
936 if (list_empty(&dst_mm->mmlist))
937 list_add(&dst_mm->mmlist,
939 spin_unlock(&mmlist_lock);
942 } else if (is_migration_entry(entry)) {
943 page = migration_entry_to_page(entry);
945 rss[mm_counter(page)]++;
947 if (is_write_migration_entry(entry) &&
948 is_cow_mapping(vm_flags)) {
950 * COW mappings require pages in both
951 * parent and child to be set to read.
953 make_migration_entry_read(&entry);
954 pte = swp_entry_to_pte(entry);
955 if (pte_swp_soft_dirty(*src_pte))
956 pte = pte_swp_mksoft_dirty(pte);
957 set_pte_at(src_mm, addr, src_pte, pte);
964 * If it's a COW mapping, write protect it both
965 * in the parent and the child
967 if (is_cow_mapping(vm_flags)) {
968 ptep_set_wrprotect(src_mm, addr, src_pte);
969 pte = pte_wrprotect(pte);
973 * If it's a shared mapping, mark it clean in
976 if (vm_flags & VM_SHARED)
977 pte = pte_mkclean(pte);
978 pte = pte_mkold(pte);
980 page = vm_normal_page(vma, addr, pte);
983 page_dup_rmap(page, false);
984 rss[mm_counter(page)]++;
988 set_pte_at(dst_mm, addr, dst_pte, pte);
992 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
993 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
994 unsigned long addr, unsigned long end)
996 pte_t *orig_src_pte, *orig_dst_pte;
997 pte_t *src_pte, *dst_pte;
998 spinlock_t *src_ptl, *dst_ptl;
1000 int rss[NR_MM_COUNTERS];
1001 swp_entry_t entry = (swp_entry_t){0};
1006 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1009 src_pte = pte_offset_map(src_pmd, addr);
1010 src_ptl = pte_lockptr(src_mm, src_pmd);
1011 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1012 orig_src_pte = src_pte;
1013 orig_dst_pte = dst_pte;
1014 arch_enter_lazy_mmu_mode();
1018 * We are holding two locks at this point - either of them
1019 * could generate latencies in another task on another CPU.
1021 if (progress >= 32) {
1023 if (need_resched() ||
1024 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1027 if (pte_none(*src_pte)) {
1031 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1036 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1038 arch_leave_lazy_mmu_mode();
1039 spin_unlock(src_ptl);
1040 pte_unmap(orig_src_pte);
1041 add_mm_rss_vec(dst_mm, rss);
1042 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1046 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1055 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1056 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1057 unsigned long addr, unsigned long end)
1059 pmd_t *src_pmd, *dst_pmd;
1062 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1065 src_pmd = pmd_offset(src_pud, addr);
1067 next = pmd_addr_end(addr, end);
1068 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1070 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1071 err = copy_huge_pmd(dst_mm, src_mm,
1072 dst_pmd, src_pmd, addr, vma);
1079 if (pmd_none_or_clear_bad(src_pmd))
1081 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1084 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1088 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1089 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1090 unsigned long addr, unsigned long end)
1092 pud_t *src_pud, *dst_pud;
1095 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1098 src_pud = pud_offset(src_p4d, addr);
1100 next = pud_addr_end(addr, end);
1101 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1104 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1105 err = copy_huge_pud(dst_mm, src_mm,
1106 dst_pud, src_pud, addr, vma);
1113 if (pud_none_or_clear_bad(src_pud))
1115 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1118 } while (dst_pud++, src_pud++, addr = next, addr != end);
1122 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1123 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1124 unsigned long addr, unsigned long end)
1126 p4d_t *src_p4d, *dst_p4d;
1129 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1132 src_p4d = p4d_offset(src_pgd, addr);
1134 next = p4d_addr_end(addr, end);
1135 if (p4d_none_or_clear_bad(src_p4d))
1137 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1140 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1144 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1145 struct vm_area_struct *vma)
1147 pgd_t *src_pgd, *dst_pgd;
1149 unsigned long addr = vma->vm_start;
1150 unsigned long end = vma->vm_end;
1151 unsigned long mmun_start; /* For mmu_notifiers */
1152 unsigned long mmun_end; /* For mmu_notifiers */
1157 * Don't copy ptes where a page fault will fill them correctly.
1158 * Fork becomes much lighter when there are big shared or private
1159 * readonly mappings. The tradeoff is that copy_page_range is more
1160 * efficient than faulting.
1162 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1166 if (is_vm_hugetlb_page(vma))
1167 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1169 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1171 * We do not free on error cases below as remove_vma
1172 * gets called on error from higher level routine
1174 ret = track_pfn_copy(vma);
1180 * We need to invalidate the secondary MMU mappings only when
1181 * there could be a permission downgrade on the ptes of the
1182 * parent mm. And a permission downgrade will only happen if
1183 * is_cow_mapping() returns true.
1185 is_cow = is_cow_mapping(vma->vm_flags);
1189 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1193 dst_pgd = pgd_offset(dst_mm, addr);
1194 src_pgd = pgd_offset(src_mm, addr);
1196 next = pgd_addr_end(addr, end);
1197 if (pgd_none_or_clear_bad(src_pgd))
1199 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1200 vma, addr, next))) {
1204 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1207 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1211 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1212 struct vm_area_struct *vma, pmd_t *pmd,
1213 unsigned long addr, unsigned long end,
1214 struct zap_details *details)
1216 struct mm_struct *mm = tlb->mm;
1217 int force_flush = 0;
1218 int rss[NR_MM_COUNTERS];
1224 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1227 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1229 flush_tlb_batched_pending(mm);
1230 arch_enter_lazy_mmu_mode();
1233 if (pte_none(ptent))
1236 if (pte_present(ptent)) {
1239 page = vm_normal_page(vma, addr, ptent);
1240 if (unlikely(details) && page) {
1242 * unmap_shared_mapping_pages() wants to
1243 * invalidate cache without truncating:
1244 * unmap shared but keep private pages.
1246 if (details->check_mapping &&
1247 details->check_mapping != page_rmapping(page))
1250 ptent = ptep_get_and_clear_full(mm, addr, pte,
1252 tlb_remove_tlb_entry(tlb, pte, addr);
1253 if (unlikely(!page))
1256 if (!PageAnon(page)) {
1257 if (pte_dirty(ptent)) {
1259 set_page_dirty(page);
1261 if (pte_young(ptent) &&
1262 likely(!(vma->vm_flags & VM_SEQ_READ)))
1263 mark_page_accessed(page);
1265 rss[mm_counter(page)]--;
1266 page_remove_rmap(page, false);
1267 if (unlikely(page_mapcount(page) < 0))
1268 print_bad_pte(vma, addr, ptent, page);
1269 if (unlikely(__tlb_remove_page(tlb, page))) {
1276 /* If details->check_mapping, we leave swap entries. */
1277 if (unlikely(details))
1280 entry = pte_to_swp_entry(ptent);
1281 if (!non_swap_entry(entry))
1283 else if (is_migration_entry(entry)) {
1286 page = migration_entry_to_page(entry);
1287 rss[mm_counter(page)]--;
1289 if (unlikely(!free_swap_and_cache(entry)))
1290 print_bad_pte(vma, addr, ptent, NULL);
1291 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1292 } while (pte++, addr += PAGE_SIZE, addr != end);
1294 add_mm_rss_vec(mm, rss);
1295 arch_leave_lazy_mmu_mode();
1297 /* Do the actual TLB flush before dropping ptl */
1299 tlb_flush_mmu_tlbonly(tlb);
1300 pte_unmap_unlock(start_pte, ptl);
1303 * If we forced a TLB flush (either due to running out of
1304 * batch buffers or because we needed to flush dirty TLB
1305 * entries before releasing the ptl), free the batched
1306 * memory too. Restart if we didn't do everything.
1310 tlb_flush_mmu_free(tlb);
1318 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1319 struct vm_area_struct *vma, pud_t *pud,
1320 unsigned long addr, unsigned long end,
1321 struct zap_details *details)
1326 pmd = pmd_offset(pud, addr);
1328 next = pmd_addr_end(addr, end);
1329 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1330 if (next - addr != HPAGE_PMD_SIZE) {
1331 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1332 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1333 __split_huge_pmd(vma, pmd, addr, false, NULL);
1334 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1339 * Here there can be other concurrent MADV_DONTNEED or
1340 * trans huge page faults running, and if the pmd is
1341 * none or trans huge it can change under us. This is
1342 * because MADV_DONTNEED holds the mmap_sem in read
1345 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1347 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1350 } while (pmd++, addr = next, addr != end);
1355 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1356 struct vm_area_struct *vma, p4d_t *p4d,
1357 unsigned long addr, unsigned long end,
1358 struct zap_details *details)
1363 pud = pud_offset(p4d, addr);
1365 next = pud_addr_end(addr, end);
1366 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1367 if (next - addr != HPAGE_PUD_SIZE) {
1368 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1369 split_huge_pud(vma, pud, addr);
1370 } else if (zap_huge_pud(tlb, vma, pud, addr))
1374 if (pud_none_or_clear_bad(pud))
1376 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1379 } while (pud++, addr = next, addr != end);
1384 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1385 struct vm_area_struct *vma, pgd_t *pgd,
1386 unsigned long addr, unsigned long end,
1387 struct zap_details *details)
1392 p4d = p4d_offset(pgd, addr);
1394 next = p4d_addr_end(addr, end);
1395 if (p4d_none_or_clear_bad(p4d))
1397 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1398 } while (p4d++, addr = next, addr != end);
1403 void unmap_page_range(struct mmu_gather *tlb,
1404 struct vm_area_struct *vma,
1405 unsigned long addr, unsigned long end,
1406 struct zap_details *details)
1411 BUG_ON(addr >= end);
1412 tlb_start_vma(tlb, vma);
1413 pgd = pgd_offset(vma->vm_mm, addr);
1415 next = pgd_addr_end(addr, end);
1416 if (pgd_none_or_clear_bad(pgd))
1418 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1419 } while (pgd++, addr = next, addr != end);
1420 tlb_end_vma(tlb, vma);
1424 static void unmap_single_vma(struct mmu_gather *tlb,
1425 struct vm_area_struct *vma, unsigned long start_addr,
1426 unsigned long end_addr,
1427 struct zap_details *details)
1429 unsigned long start = max(vma->vm_start, start_addr);
1432 if (start >= vma->vm_end)
1434 end = min(vma->vm_end, end_addr);
1435 if (end <= vma->vm_start)
1439 uprobe_munmap(vma, start, end);
1441 if (unlikely(vma->vm_flags & VM_PFNMAP))
1442 untrack_pfn(vma, 0, 0);
1445 if (unlikely(is_vm_hugetlb_page(vma))) {
1447 * It is undesirable to test vma->vm_file as it
1448 * should be non-null for valid hugetlb area.
1449 * However, vm_file will be NULL in the error
1450 * cleanup path of mmap_region. When
1451 * hugetlbfs ->mmap method fails,
1452 * mmap_region() nullifies vma->vm_file
1453 * before calling this function to clean up.
1454 * Since no pte has actually been setup, it is
1455 * safe to do nothing in this case.
1458 i_mmap_lock_write(vma->vm_file->f_mapping);
1459 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1460 i_mmap_unlock_write(vma->vm_file->f_mapping);
1463 unmap_page_range(tlb, vma, start, end, details);
1468 * unmap_vmas - unmap a range of memory covered by a list of vma's
1469 * @tlb: address of the caller's struct mmu_gather
1470 * @vma: the starting vma
1471 * @start_addr: virtual address at which to start unmapping
1472 * @end_addr: virtual address at which to end unmapping
1474 * Unmap all pages in the vma list.
1476 * Only addresses between `start' and `end' will be unmapped.
1478 * The VMA list must be sorted in ascending virtual address order.
1480 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1481 * range after unmap_vmas() returns. So the only responsibility here is to
1482 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1483 * drops the lock and schedules.
1485 void unmap_vmas(struct mmu_gather *tlb,
1486 struct vm_area_struct *vma, unsigned long start_addr,
1487 unsigned long end_addr)
1489 struct mm_struct *mm = vma->vm_mm;
1491 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1492 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1493 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1494 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1498 * zap_page_range - remove user pages in a given range
1499 * @vma: vm_area_struct holding the applicable pages
1500 * @start: starting address of pages to zap
1501 * @size: number of bytes to zap
1503 * Caller must protect the VMA list
1505 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1508 struct mm_struct *mm = vma->vm_mm;
1509 struct mmu_gather tlb;
1510 unsigned long end = start + size;
1513 tlb_gather_mmu(&tlb, mm, start, end);
1514 update_hiwater_rss(mm);
1515 mmu_notifier_invalidate_range_start(mm, start, end);
1516 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1517 unmap_single_vma(&tlb, vma, start, end, NULL);
1520 * zap_page_range does not specify whether mmap_sem should be
1521 * held for read or write. That allows parallel zap_page_range
1522 * operations to unmap a PTE and defer a flush meaning that
1523 * this call observes pte_none and fails to flush the TLB.
1524 * Rather than adding a complex API, ensure that no stale
1525 * TLB entries exist when this call returns.
1527 flush_tlb_range(vma, start, end);
1530 mmu_notifier_invalidate_range_end(mm, start, end);
1531 tlb_finish_mmu(&tlb, start, end);
1535 * zap_page_range_single - remove user pages in a given range
1536 * @vma: vm_area_struct holding the applicable pages
1537 * @address: starting address of pages to zap
1538 * @size: number of bytes to zap
1539 * @details: details of shared cache invalidation
1541 * The range must fit into one VMA.
1543 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1544 unsigned long size, struct zap_details *details)
1546 struct mm_struct *mm = vma->vm_mm;
1547 struct mmu_gather tlb;
1548 unsigned long end = address + size;
1551 tlb_gather_mmu(&tlb, mm, address, end);
1552 update_hiwater_rss(mm);
1553 mmu_notifier_invalidate_range_start(mm, address, end);
1554 unmap_single_vma(&tlb, vma, address, end, details);
1555 mmu_notifier_invalidate_range_end(mm, address, end);
1556 tlb_finish_mmu(&tlb, address, end);
1560 * zap_vma_ptes - remove ptes mapping the vma
1561 * @vma: vm_area_struct holding ptes to be zapped
1562 * @address: starting address of pages to zap
1563 * @size: number of bytes to zap
1565 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1567 * The entire address range must be fully contained within the vma.
1569 * Returns 0 if successful.
1571 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1574 if (address < vma->vm_start || address + size > vma->vm_end ||
1575 !(vma->vm_flags & VM_PFNMAP))
1577 zap_page_range_single(vma, address, size, NULL);
1580 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1582 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1590 pgd = pgd_offset(mm, addr);
1591 p4d = p4d_alloc(mm, pgd, addr);
1594 pud = pud_alloc(mm, p4d, addr);
1597 pmd = pmd_alloc(mm, pud, addr);
1601 VM_BUG_ON(pmd_trans_huge(*pmd));
1602 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1606 * This is the old fallback for page remapping.
1608 * For historical reasons, it only allows reserved pages. Only
1609 * old drivers should use this, and they needed to mark their
1610 * pages reserved for the old functions anyway.
1612 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1613 struct page *page, pgprot_t prot)
1615 struct mm_struct *mm = vma->vm_mm;
1624 flush_dcache_page(page);
1625 pte = get_locked_pte(mm, addr, &ptl);
1629 if (!pte_none(*pte))
1632 /* Ok, finally just insert the thing.. */
1634 inc_mm_counter_fast(mm, mm_counter_file(page));
1635 page_add_file_rmap(page, false);
1636 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1639 pte_unmap_unlock(pte, ptl);
1642 pte_unmap_unlock(pte, ptl);
1648 * vm_insert_page - insert single page into user vma
1649 * @vma: user vma to map to
1650 * @addr: target user address of this page
1651 * @page: source kernel page
1653 * This allows drivers to insert individual pages they've allocated
1656 * The page has to be a nice clean _individual_ kernel allocation.
1657 * If you allocate a compound page, you need to have marked it as
1658 * such (__GFP_COMP), or manually just split the page up yourself
1659 * (see split_page()).
1661 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1662 * took an arbitrary page protection parameter. This doesn't allow
1663 * that. Your vma protection will have to be set up correctly, which
1664 * means that if you want a shared writable mapping, you'd better
1665 * ask for a shared writable mapping!
1667 * The page does not need to be reserved.
1669 * Usually this function is called from f_op->mmap() handler
1670 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1671 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1672 * function from other places, for example from page-fault handler.
1674 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1677 if (addr < vma->vm_start || addr >= vma->vm_end)
1679 if (!page_count(page))
1681 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1682 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1683 BUG_ON(vma->vm_flags & VM_PFNMAP);
1684 vma->vm_flags |= VM_MIXEDMAP;
1686 return insert_page(vma, addr, page, vma->vm_page_prot);
1688 EXPORT_SYMBOL(vm_insert_page);
1690 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1691 pfn_t pfn, pgprot_t prot, bool mkwrite)
1693 struct mm_struct *mm = vma->vm_mm;
1699 pte = get_locked_pte(mm, addr, &ptl);
1703 if (!pte_none(*pte)) {
1706 * For read faults on private mappings the PFN passed
1707 * in may not match the PFN we have mapped if the
1708 * mapped PFN is a writeable COW page. In the mkwrite
1709 * case we are creating a writable PTE for a shared
1710 * mapping and we expect the PFNs to match.
1712 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1720 /* Ok, finally just insert the thing.. */
1721 if (pfn_t_devmap(pfn))
1722 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1724 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1728 entry = pte_mkyoung(entry);
1729 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1732 set_pte_at(mm, addr, pte, entry);
1733 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1737 pte_unmap_unlock(pte, ptl);
1743 * vm_insert_pfn - insert single pfn into user vma
1744 * @vma: user vma to map to
1745 * @addr: target user address of this page
1746 * @pfn: source kernel pfn
1748 * Similar to vm_insert_page, this allows drivers to insert individual pages
1749 * they've allocated into a user vma. Same comments apply.
1751 * This function should only be called from a vm_ops->fault handler, and
1752 * in that case the handler should return NULL.
1754 * vma cannot be a COW mapping.
1756 * As this is called only for pages that do not currently exist, we
1757 * do not need to flush old virtual caches or the TLB.
1759 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1762 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1764 EXPORT_SYMBOL(vm_insert_pfn);
1767 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1768 * @vma: user vma to map to
1769 * @addr: target user address of this page
1770 * @pfn: source kernel pfn
1771 * @pgprot: pgprot flags for the inserted page
1773 * This is exactly like vm_insert_pfn, except that it allows drivers to
1774 * to override pgprot on a per-page basis.
1776 * This only makes sense for IO mappings, and it makes no sense for
1777 * cow mappings. In general, using multiple vmas is preferable;
1778 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1781 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1782 unsigned long pfn, pgprot_t pgprot)
1786 * Technically, architectures with pte_special can avoid all these
1787 * restrictions (same for remap_pfn_range). However we would like
1788 * consistency in testing and feature parity among all, so we should
1789 * try to keep these invariants in place for everybody.
1791 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1792 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1793 (VM_PFNMAP|VM_MIXEDMAP));
1794 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1795 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1797 if (addr < vma->vm_start || addr >= vma->vm_end)
1800 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1802 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1807 EXPORT_SYMBOL(vm_insert_pfn_prot);
1809 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1810 pfn_t pfn, bool mkwrite)
1812 pgprot_t pgprot = vma->vm_page_prot;
1814 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1816 if (addr < vma->vm_start || addr >= vma->vm_end)
1819 track_pfn_insert(vma, &pgprot, pfn);
1822 * If we don't have pte special, then we have to use the pfn_valid()
1823 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1824 * refcount the page if pfn_valid is true (hence insert_page rather
1825 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1826 * without pte special, it would there be refcounted as a normal page.
1828 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1832 * At this point we are committed to insert_page()
1833 * regardless of whether the caller specified flags that
1834 * result in pfn_t_has_page() == false.
1836 page = pfn_to_page(pfn_t_to_pfn(pfn));
1837 return insert_page(vma, addr, page, pgprot);
1839 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1842 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1845 return __vm_insert_mixed(vma, addr, pfn, false);
1848 EXPORT_SYMBOL(vm_insert_mixed);
1850 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1853 return __vm_insert_mixed(vma, addr, pfn, true);
1855 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1858 * maps a range of physical memory into the requested pages. the old
1859 * mappings are removed. any references to nonexistent pages results
1860 * in null mappings (currently treated as "copy-on-access")
1862 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1863 unsigned long addr, unsigned long end,
1864 unsigned long pfn, pgprot_t prot)
1869 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1872 arch_enter_lazy_mmu_mode();
1874 BUG_ON(!pte_none(*pte));
1875 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1877 } while (pte++, addr += PAGE_SIZE, addr != end);
1878 arch_leave_lazy_mmu_mode();
1879 pte_unmap_unlock(pte - 1, ptl);
1883 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1884 unsigned long addr, unsigned long end,
1885 unsigned long pfn, pgprot_t prot)
1890 pfn -= addr >> PAGE_SHIFT;
1891 pmd = pmd_alloc(mm, pud, addr);
1894 VM_BUG_ON(pmd_trans_huge(*pmd));
1896 next = pmd_addr_end(addr, end);
1897 if (remap_pte_range(mm, pmd, addr, next,
1898 pfn + (addr >> PAGE_SHIFT), prot))
1900 } while (pmd++, addr = next, addr != end);
1904 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1905 unsigned long addr, unsigned long end,
1906 unsigned long pfn, pgprot_t prot)
1911 pfn -= addr >> PAGE_SHIFT;
1912 pud = pud_alloc(mm, p4d, addr);
1916 next = pud_addr_end(addr, end);
1917 if (remap_pmd_range(mm, pud, addr, next,
1918 pfn + (addr >> PAGE_SHIFT), prot))
1920 } while (pud++, addr = next, addr != end);
1924 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1925 unsigned long addr, unsigned long end,
1926 unsigned long pfn, pgprot_t prot)
1931 pfn -= addr >> PAGE_SHIFT;
1932 p4d = p4d_alloc(mm, pgd, addr);
1936 next = p4d_addr_end(addr, end);
1937 if (remap_pud_range(mm, p4d, addr, next,
1938 pfn + (addr >> PAGE_SHIFT), prot))
1940 } while (p4d++, addr = next, addr != end);
1945 * remap_pfn_range - remap kernel memory to userspace
1946 * @vma: user vma to map to
1947 * @addr: target user address to start at
1948 * @pfn: physical address of kernel memory
1949 * @size: size of map area
1950 * @prot: page protection flags for this mapping
1952 * Note: this is only safe if the mm semaphore is held when called.
1954 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1955 unsigned long pfn, unsigned long size, pgprot_t prot)
1959 unsigned long end = addr + PAGE_ALIGN(size);
1960 struct mm_struct *mm = vma->vm_mm;
1961 unsigned long remap_pfn = pfn;
1965 * Physically remapped pages are special. Tell the
1966 * rest of the world about it:
1967 * VM_IO tells people not to look at these pages
1968 * (accesses can have side effects).
1969 * VM_PFNMAP tells the core MM that the base pages are just
1970 * raw PFN mappings, and do not have a "struct page" associated
1973 * Disable vma merging and expanding with mremap().
1975 * Omit vma from core dump, even when VM_IO turned off.
1977 * There's a horrible special case to handle copy-on-write
1978 * behaviour that some programs depend on. We mark the "original"
1979 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1980 * See vm_normal_page() for details.
1982 if (is_cow_mapping(vma->vm_flags)) {
1983 if (addr != vma->vm_start || end != vma->vm_end)
1985 vma->vm_pgoff = pfn;
1988 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1992 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1994 BUG_ON(addr >= end);
1995 pfn -= addr >> PAGE_SHIFT;
1996 pgd = pgd_offset(mm, addr);
1997 flush_cache_range(vma, addr, end);
1999 next = pgd_addr_end(addr, end);
2000 err = remap_p4d_range(mm, pgd, addr, next,
2001 pfn + (addr >> PAGE_SHIFT), prot);
2004 } while (pgd++, addr = next, addr != end);
2007 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2011 EXPORT_SYMBOL(remap_pfn_range);
2014 * vm_iomap_memory - remap memory to userspace
2015 * @vma: user vma to map to
2016 * @start: start of area
2017 * @len: size of area
2019 * This is a simplified io_remap_pfn_range() for common driver use. The
2020 * driver just needs to give us the physical memory range to be mapped,
2021 * we'll figure out the rest from the vma information.
2023 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2024 * whatever write-combining details or similar.
2026 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2028 unsigned long vm_len, pfn, pages;
2030 /* Check that the physical memory area passed in looks valid */
2031 if (start + len < start)
2034 * You *really* shouldn't map things that aren't page-aligned,
2035 * but we've historically allowed it because IO memory might
2036 * just have smaller alignment.
2038 len += start & ~PAGE_MASK;
2039 pfn = start >> PAGE_SHIFT;
2040 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2041 if (pfn + pages < pfn)
2044 /* We start the mapping 'vm_pgoff' pages into the area */
2045 if (vma->vm_pgoff > pages)
2047 pfn += vma->vm_pgoff;
2048 pages -= vma->vm_pgoff;
2050 /* Can we fit all of the mapping? */
2051 vm_len = vma->vm_end - vma->vm_start;
2052 if (vm_len >> PAGE_SHIFT > pages)
2055 /* Ok, let it rip */
2056 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2058 EXPORT_SYMBOL(vm_iomap_memory);
2060 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2061 unsigned long addr, unsigned long end,
2062 pte_fn_t fn, void *data)
2067 spinlock_t *uninitialized_var(ptl);
2069 pte = (mm == &init_mm) ?
2070 pte_alloc_kernel(pmd, addr) :
2071 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2075 BUG_ON(pmd_huge(*pmd));
2077 arch_enter_lazy_mmu_mode();
2079 token = pmd_pgtable(*pmd);
2082 err = fn(pte++, token, addr, data);
2085 } while (addr += PAGE_SIZE, addr != end);
2087 arch_leave_lazy_mmu_mode();
2090 pte_unmap_unlock(pte-1, ptl);
2094 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2095 unsigned long addr, unsigned long end,
2096 pte_fn_t fn, void *data)
2102 BUG_ON(pud_huge(*pud));
2104 pmd = pmd_alloc(mm, pud, addr);
2108 next = pmd_addr_end(addr, end);
2109 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2112 } while (pmd++, addr = next, addr != end);
2116 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2117 unsigned long addr, unsigned long end,
2118 pte_fn_t fn, void *data)
2124 pud = pud_alloc(mm, p4d, addr);
2128 next = pud_addr_end(addr, end);
2129 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2132 } while (pud++, addr = next, addr != end);
2136 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2137 unsigned long addr, unsigned long end,
2138 pte_fn_t fn, void *data)
2144 p4d = p4d_alloc(mm, pgd, addr);
2148 next = p4d_addr_end(addr, end);
2149 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2152 } while (p4d++, addr = next, addr != end);
2157 * Scan a region of virtual memory, filling in page tables as necessary
2158 * and calling a provided function on each leaf page table.
2160 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2161 unsigned long size, pte_fn_t fn, void *data)
2165 unsigned long end = addr + size;
2168 if (WARN_ON(addr >= end))
2171 pgd = pgd_offset(mm, addr);
2173 next = pgd_addr_end(addr, end);
2174 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2177 } while (pgd++, addr = next, addr != end);
2181 EXPORT_SYMBOL_GPL(apply_to_page_range);
2184 * handle_pte_fault chooses page fault handler according to an entry which was
2185 * read non-atomically. Before making any commitment, on those architectures
2186 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2187 * parts, do_swap_page must check under lock before unmapping the pte and
2188 * proceeding (but do_wp_page is only called after already making such a check;
2189 * and do_anonymous_page can safely check later on).
2191 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2192 pte_t *page_table, pte_t orig_pte)
2195 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2196 if (sizeof(pte_t) > sizeof(unsigned long)) {
2197 spinlock_t *ptl = pte_lockptr(mm, pmd);
2199 same = pte_same(*page_table, orig_pte);
2203 pte_unmap(page_table);
2207 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2209 debug_dma_assert_idle(src);
2212 * If the source page was a PFN mapping, we don't have
2213 * a "struct page" for it. We do a best-effort copy by
2214 * just copying from the original user address. If that
2215 * fails, we just zero-fill it. Live with it.
2217 if (unlikely(!src)) {
2218 void *kaddr = kmap_atomic(dst);
2219 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2222 * This really shouldn't fail, because the page is there
2223 * in the page tables. But it might just be unreadable,
2224 * in which case we just give up and fill the result with
2227 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2229 kunmap_atomic(kaddr);
2230 flush_dcache_page(dst);
2232 copy_user_highpage(dst, src, va, vma);
2235 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2237 struct file *vm_file = vma->vm_file;
2240 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2243 * Special mappings (e.g. VDSO) do not have any file so fake
2244 * a default GFP_KERNEL for them.
2250 * Notify the address space that the page is about to become writable so that
2251 * it can prohibit this or wait for the page to get into an appropriate state.
2253 * We do this without the lock held, so that it can sleep if it needs to.
2255 static int do_page_mkwrite(struct vm_fault *vmf)
2258 struct page *page = vmf->page;
2259 unsigned int old_flags = vmf->flags;
2261 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2263 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2264 /* Restore original flags so that caller is not surprised */
2265 vmf->flags = old_flags;
2266 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2268 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2270 if (!page->mapping) {
2272 return 0; /* retry */
2274 ret |= VM_FAULT_LOCKED;
2276 VM_BUG_ON_PAGE(!PageLocked(page), page);
2281 * Handle dirtying of a page in shared file mapping on a write fault.
2283 * The function expects the page to be locked and unlocks it.
2285 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2288 struct address_space *mapping;
2290 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2292 dirtied = set_page_dirty(page);
2293 VM_BUG_ON_PAGE(PageAnon(page), page);
2295 * Take a local copy of the address_space - page.mapping may be zeroed
2296 * by truncate after unlock_page(). The address_space itself remains
2297 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2298 * release semantics to prevent the compiler from undoing this copying.
2300 mapping = page_rmapping(page);
2303 if ((dirtied || page_mkwrite) && mapping) {
2305 * Some device drivers do not set page.mapping
2306 * but still dirty their pages
2308 balance_dirty_pages_ratelimited(mapping);
2312 file_update_time(vma->vm_file);
2316 * Handle write page faults for pages that can be reused in the current vma
2318 * This can happen either due to the mapping being with the VM_SHARED flag,
2319 * or due to us being the last reference standing to the page. In either
2320 * case, all we need to do here is to mark the page as writable and update
2321 * any related book-keeping.
2323 static inline void wp_page_reuse(struct vm_fault *vmf)
2324 __releases(vmf->ptl)
2326 struct vm_area_struct *vma = vmf->vma;
2327 struct page *page = vmf->page;
2330 * Clear the pages cpupid information as the existing
2331 * information potentially belongs to a now completely
2332 * unrelated process.
2335 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2337 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2338 entry = pte_mkyoung(vmf->orig_pte);
2339 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2340 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2341 update_mmu_cache(vma, vmf->address, vmf->pte);
2342 pte_unmap_unlock(vmf->pte, vmf->ptl);
2346 * Handle the case of a page which we actually need to copy to a new page.
2348 * Called with mmap_sem locked and the old page referenced, but
2349 * without the ptl held.
2351 * High level logic flow:
2353 * - Allocate a page, copy the content of the old page to the new one.
2354 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2355 * - Take the PTL. If the pte changed, bail out and release the allocated page
2356 * - If the pte is still the way we remember it, update the page table and all
2357 * relevant references. This includes dropping the reference the page-table
2358 * held to the old page, as well as updating the rmap.
2359 * - In any case, unlock the PTL and drop the reference we took to the old page.
2361 static int wp_page_copy(struct vm_fault *vmf)
2363 struct vm_area_struct *vma = vmf->vma;
2364 struct mm_struct *mm = vma->vm_mm;
2365 struct page *old_page = vmf->page;
2366 struct page *new_page = NULL;
2368 int page_copied = 0;
2369 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2370 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2371 struct mem_cgroup *memcg;
2373 if (unlikely(anon_vma_prepare(vma)))
2376 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2377 new_page = alloc_zeroed_user_highpage_movable(vma,
2382 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2386 cow_user_page(new_page, old_page, vmf->address, vma);
2389 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2392 __SetPageUptodate(new_page);
2394 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2397 * Re-check the pte - we dropped the lock
2399 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2400 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2402 if (!PageAnon(old_page)) {
2403 dec_mm_counter_fast(mm,
2404 mm_counter_file(old_page));
2405 inc_mm_counter_fast(mm, MM_ANONPAGES);
2408 inc_mm_counter_fast(mm, MM_ANONPAGES);
2410 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2411 entry = mk_pte(new_page, vma->vm_page_prot);
2412 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2414 * Clear the pte entry and flush it first, before updating the
2415 * pte with the new entry. This will avoid a race condition
2416 * seen in the presence of one thread doing SMC and another
2419 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2420 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2421 mem_cgroup_commit_charge(new_page, memcg, false, false);
2422 lru_cache_add_active_or_unevictable(new_page, vma);
2424 * We call the notify macro here because, when using secondary
2425 * mmu page tables (such as kvm shadow page tables), we want the
2426 * new page to be mapped directly into the secondary page table.
2428 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2429 update_mmu_cache(vma, vmf->address, vmf->pte);
2432 * Only after switching the pte to the new page may
2433 * we remove the mapcount here. Otherwise another
2434 * process may come and find the rmap count decremented
2435 * before the pte is switched to the new page, and
2436 * "reuse" the old page writing into it while our pte
2437 * here still points into it and can be read by other
2440 * The critical issue is to order this
2441 * page_remove_rmap with the ptp_clear_flush above.
2442 * Those stores are ordered by (if nothing else,)
2443 * the barrier present in the atomic_add_negative
2444 * in page_remove_rmap.
2446 * Then the TLB flush in ptep_clear_flush ensures that
2447 * no process can access the old page before the
2448 * decremented mapcount is visible. And the old page
2449 * cannot be reused until after the decremented
2450 * mapcount is visible. So transitively, TLBs to
2451 * old page will be flushed before it can be reused.
2453 page_remove_rmap(old_page, false);
2456 /* Free the old page.. */
2457 new_page = old_page;
2460 mem_cgroup_cancel_charge(new_page, memcg, false);
2466 pte_unmap_unlock(vmf->pte, vmf->ptl);
2467 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2470 * Don't let another task, with possibly unlocked vma,
2471 * keep the mlocked page.
2473 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2474 lock_page(old_page); /* LRU manipulation */
2475 if (PageMlocked(old_page))
2476 munlock_vma_page(old_page);
2477 unlock_page(old_page);
2481 return page_copied ? VM_FAULT_WRITE : 0;
2487 return VM_FAULT_OOM;
2491 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2492 * writeable once the page is prepared
2494 * @vmf: structure describing the fault
2496 * This function handles all that is needed to finish a write page fault in a
2497 * shared mapping due to PTE being read-only once the mapped page is prepared.
2498 * It handles locking of PTE and modifying it. The function returns
2499 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2502 * The function expects the page to be locked or other protection against
2503 * concurrent faults / writeback (such as DAX radix tree locks).
2505 int finish_mkwrite_fault(struct vm_fault *vmf)
2507 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2508 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2511 * We might have raced with another page fault while we released the
2512 * pte_offset_map_lock.
2514 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2515 pte_unmap_unlock(vmf->pte, vmf->ptl);
2516 return VM_FAULT_NOPAGE;
2523 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2526 static int wp_pfn_shared(struct vm_fault *vmf)
2528 struct vm_area_struct *vma = vmf->vma;
2530 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2533 pte_unmap_unlock(vmf->pte, vmf->ptl);
2534 vmf->flags |= FAULT_FLAG_MKWRITE;
2535 ret = vma->vm_ops->pfn_mkwrite(vmf);
2536 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2538 return finish_mkwrite_fault(vmf);
2541 return VM_FAULT_WRITE;
2544 static int wp_page_shared(struct vm_fault *vmf)
2545 __releases(vmf->ptl)
2547 struct vm_area_struct *vma = vmf->vma;
2549 get_page(vmf->page);
2551 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2554 pte_unmap_unlock(vmf->pte, vmf->ptl);
2555 tmp = do_page_mkwrite(vmf);
2556 if (unlikely(!tmp || (tmp &
2557 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2558 put_page(vmf->page);
2561 tmp = finish_mkwrite_fault(vmf);
2562 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2563 unlock_page(vmf->page);
2564 put_page(vmf->page);
2569 lock_page(vmf->page);
2571 fault_dirty_shared_page(vma, vmf->page);
2572 put_page(vmf->page);
2574 return VM_FAULT_WRITE;
2578 * This routine handles present pages, when users try to write
2579 * to a shared page. It is done by copying the page to a new address
2580 * and decrementing the shared-page counter for the old page.
2582 * Note that this routine assumes that the protection checks have been
2583 * done by the caller (the low-level page fault routine in most cases).
2584 * Thus we can safely just mark it writable once we've done any necessary
2587 * We also mark the page dirty at this point even though the page will
2588 * change only once the write actually happens. This avoids a few races,
2589 * and potentially makes it more efficient.
2591 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2592 * but allow concurrent faults), with pte both mapped and locked.
2593 * We return with mmap_sem still held, but pte unmapped and unlocked.
2595 static int do_wp_page(struct vm_fault *vmf)
2596 __releases(vmf->ptl)
2598 struct vm_area_struct *vma = vmf->vma;
2600 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2603 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2606 * We should not cow pages in a shared writeable mapping.
2607 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2609 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2610 (VM_WRITE|VM_SHARED))
2611 return wp_pfn_shared(vmf);
2613 pte_unmap_unlock(vmf->pte, vmf->ptl);
2614 return wp_page_copy(vmf);
2618 * Take out anonymous pages first, anonymous shared vmas are
2619 * not dirty accountable.
2621 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2622 int total_map_swapcount;
2623 if (!trylock_page(vmf->page)) {
2624 get_page(vmf->page);
2625 pte_unmap_unlock(vmf->pte, vmf->ptl);
2626 lock_page(vmf->page);
2627 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2628 vmf->address, &vmf->ptl);
2629 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2630 unlock_page(vmf->page);
2631 pte_unmap_unlock(vmf->pte, vmf->ptl);
2632 put_page(vmf->page);
2635 put_page(vmf->page);
2637 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2638 if (total_map_swapcount == 1) {
2640 * The page is all ours. Move it to
2641 * our anon_vma so the rmap code will
2642 * not search our parent or siblings.
2643 * Protected against the rmap code by
2646 page_move_anon_rmap(vmf->page, vma);
2648 unlock_page(vmf->page);
2650 return VM_FAULT_WRITE;
2652 unlock_page(vmf->page);
2653 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2654 (VM_WRITE|VM_SHARED))) {
2655 return wp_page_shared(vmf);
2659 * Ok, we need to copy. Oh, well..
2661 get_page(vmf->page);
2663 pte_unmap_unlock(vmf->pte, vmf->ptl);
2664 return wp_page_copy(vmf);
2667 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2668 unsigned long start_addr, unsigned long end_addr,
2669 struct zap_details *details)
2671 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2674 static inline void unmap_mapping_range_tree(struct rb_root *root,
2675 struct zap_details *details)
2677 struct vm_area_struct *vma;
2678 pgoff_t vba, vea, zba, zea;
2680 vma_interval_tree_foreach(vma, root,
2681 details->first_index, details->last_index) {
2683 vba = vma->vm_pgoff;
2684 vea = vba + vma_pages(vma) - 1;
2685 zba = details->first_index;
2688 zea = details->last_index;
2692 unmap_mapping_range_vma(vma,
2693 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2694 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2700 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2701 * address_space corresponding to the specified page range in the underlying
2704 * @mapping: the address space containing mmaps to be unmapped.
2705 * @holebegin: byte in first page to unmap, relative to the start of
2706 * the underlying file. This will be rounded down to a PAGE_SIZE
2707 * boundary. Note that this is different from truncate_pagecache(), which
2708 * must keep the partial page. In contrast, we must get rid of
2710 * @holelen: size of prospective hole in bytes. This will be rounded
2711 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2713 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2714 * but 0 when invalidating pagecache, don't throw away private data.
2716 void unmap_mapping_range(struct address_space *mapping,
2717 loff_t const holebegin, loff_t const holelen, int even_cows)
2719 struct zap_details details = { };
2720 pgoff_t hba = holebegin >> PAGE_SHIFT;
2721 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2723 /* Check for overflow. */
2724 if (sizeof(holelen) > sizeof(hlen)) {
2726 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2727 if (holeend & ~(long long)ULONG_MAX)
2728 hlen = ULONG_MAX - hba + 1;
2731 details.check_mapping = even_cows ? NULL : mapping;
2732 details.first_index = hba;
2733 details.last_index = hba + hlen - 1;
2734 if (details.last_index < details.first_index)
2735 details.last_index = ULONG_MAX;
2737 i_mmap_lock_write(mapping);
2738 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2739 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2740 i_mmap_unlock_write(mapping);
2742 EXPORT_SYMBOL(unmap_mapping_range);
2745 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2746 * but allow concurrent faults), and pte mapped but not yet locked.
2747 * We return with pte unmapped and unlocked.
2749 * We return with the mmap_sem locked or unlocked in the same cases
2750 * as does filemap_fault().
2752 int do_swap_page(struct vm_fault *vmf)
2754 struct vm_area_struct *vma = vmf->vma;
2755 struct page *page = NULL, *swapcache;
2756 struct mem_cgroup *memcg;
2757 struct vma_swap_readahead swap_ra;
2763 bool vma_readahead = swap_use_vma_readahead();
2766 page = swap_readahead_detect(vmf, &swap_ra);
2767 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2773 entry = pte_to_swp_entry(vmf->orig_pte);
2774 if (unlikely(non_swap_entry(entry))) {
2775 if (is_migration_entry(entry)) {
2776 migration_entry_wait(vma->vm_mm, vmf->pmd,
2778 } else if (is_hwpoison_entry(entry)) {
2779 ret = VM_FAULT_HWPOISON;
2781 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2782 ret = VM_FAULT_SIGBUS;
2786 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2788 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2792 page = do_swap_page_readahead(entry,
2793 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2795 page = swapin_readahead(entry,
2796 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2799 * Back out if somebody else faulted in this pte
2800 * while we released the pte lock.
2802 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2803 vmf->address, &vmf->ptl);
2804 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2806 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2810 /* Had to read the page from swap area: Major fault */
2811 ret = VM_FAULT_MAJOR;
2812 count_vm_event(PGMAJFAULT);
2813 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2814 } else if (PageHWPoison(page)) {
2816 * hwpoisoned dirty swapcache pages are kept for killing
2817 * owner processes (which may be unknown at hwpoison time)
2819 ret = VM_FAULT_HWPOISON;
2820 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2826 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2828 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2830 ret |= VM_FAULT_RETRY;
2835 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2836 * release the swapcache from under us. The page pin, and pte_same
2837 * test below, are not enough to exclude that. Even if it is still
2838 * swapcache, we need to check that the page's swap has not changed.
2840 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2843 page = ksm_might_need_to_copy(page, vma, vmf->address);
2844 if (unlikely(!page)) {
2850 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2857 * Back out if somebody else already faulted in this pte.
2859 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2861 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2864 if (unlikely(!PageUptodate(page))) {
2865 ret = VM_FAULT_SIGBUS;
2870 * The page isn't present yet, go ahead with the fault.
2872 * Be careful about the sequence of operations here.
2873 * To get its accounting right, reuse_swap_page() must be called
2874 * while the page is counted on swap but not yet in mapcount i.e.
2875 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2876 * must be called after the swap_free(), or it will never succeed.
2879 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2880 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2881 pte = mk_pte(page, vma->vm_page_prot);
2882 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2883 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2884 vmf->flags &= ~FAULT_FLAG_WRITE;
2885 ret |= VM_FAULT_WRITE;
2886 exclusive = RMAP_EXCLUSIVE;
2888 flush_icache_page(vma, page);
2889 if (pte_swp_soft_dirty(vmf->orig_pte))
2890 pte = pte_mksoft_dirty(pte);
2891 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2892 vmf->orig_pte = pte;
2893 if (page == swapcache) {
2894 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2895 mem_cgroup_commit_charge(page, memcg, true, false);
2896 activate_page(page);
2897 } else { /* ksm created a completely new copy */
2898 page_add_new_anon_rmap(page, vma, vmf->address, false);
2899 mem_cgroup_commit_charge(page, memcg, false, false);
2900 lru_cache_add_active_or_unevictable(page, vma);
2904 if (mem_cgroup_swap_full(page) ||
2905 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2906 try_to_free_swap(page);
2908 if (page != swapcache) {
2910 * Hold the lock to avoid the swap entry to be reused
2911 * until we take the PT lock for the pte_same() check
2912 * (to avoid false positives from pte_same). For
2913 * further safety release the lock after the swap_free
2914 * so that the swap count won't change under a
2915 * parallel locked swapcache.
2917 unlock_page(swapcache);
2918 put_page(swapcache);
2921 if (vmf->flags & FAULT_FLAG_WRITE) {
2922 ret |= do_wp_page(vmf);
2923 if (ret & VM_FAULT_ERROR)
2924 ret &= VM_FAULT_ERROR;
2928 /* No need to invalidate - it was non-present before */
2929 update_mmu_cache(vma, vmf->address, vmf->pte);
2931 pte_unmap_unlock(vmf->pte, vmf->ptl);
2935 mem_cgroup_cancel_charge(page, memcg, false);
2936 pte_unmap_unlock(vmf->pte, vmf->ptl);
2941 if (page != swapcache) {
2942 unlock_page(swapcache);
2943 put_page(swapcache);
2949 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2950 * but allow concurrent faults), and pte mapped but not yet locked.
2951 * We return with mmap_sem still held, but pte unmapped and unlocked.
2953 static int do_anonymous_page(struct vm_fault *vmf)
2955 struct vm_area_struct *vma = vmf->vma;
2956 struct mem_cgroup *memcg;
2961 /* File mapping without ->vm_ops ? */
2962 if (vma->vm_flags & VM_SHARED)
2963 return VM_FAULT_SIGBUS;
2966 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2967 * pte_offset_map() on pmds where a huge pmd might be created
2968 * from a different thread.
2970 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2971 * parallel threads are excluded by other means.
2973 * Here we only have down_read(mmap_sem).
2975 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2976 return VM_FAULT_OOM;
2978 /* See the comment in pte_alloc_one_map() */
2979 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2982 /* Use the zero-page for reads */
2983 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2984 !mm_forbids_zeropage(vma->vm_mm)) {
2985 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2986 vma->vm_page_prot));
2987 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2988 vmf->address, &vmf->ptl);
2989 if (!pte_none(*vmf->pte))
2991 ret = check_stable_address_space(vma->vm_mm);
2994 /* Deliver the page fault to userland, check inside PT lock */
2995 if (userfaultfd_missing(vma)) {
2996 pte_unmap_unlock(vmf->pte, vmf->ptl);
2997 return handle_userfault(vmf, VM_UFFD_MISSING);
3002 /* Allocate our own private page. */
3003 if (unlikely(anon_vma_prepare(vma)))
3005 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3009 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3013 * The memory barrier inside __SetPageUptodate makes sure that
3014 * preceeding stores to the page contents become visible before
3015 * the set_pte_at() write.
3017 __SetPageUptodate(page);
3019 entry = mk_pte(page, vma->vm_page_prot);
3020 if (vma->vm_flags & VM_WRITE)
3021 entry = pte_mkwrite(pte_mkdirty(entry));
3023 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3025 if (!pte_none(*vmf->pte))
3028 ret = check_stable_address_space(vma->vm_mm);
3032 /* Deliver the page fault to userland, check inside PT lock */
3033 if (userfaultfd_missing(vma)) {
3034 pte_unmap_unlock(vmf->pte, vmf->ptl);
3035 mem_cgroup_cancel_charge(page, memcg, false);
3037 return handle_userfault(vmf, VM_UFFD_MISSING);
3040 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3041 page_add_new_anon_rmap(page, vma, vmf->address, false);
3042 mem_cgroup_commit_charge(page, memcg, false, false);
3043 lru_cache_add_active_or_unevictable(page, vma);
3045 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3047 /* No need to invalidate - it was non-present before */
3048 update_mmu_cache(vma, vmf->address, vmf->pte);
3050 pte_unmap_unlock(vmf->pte, vmf->ptl);
3053 mem_cgroup_cancel_charge(page, memcg, false);
3059 return VM_FAULT_OOM;
3063 * The mmap_sem must have been held on entry, and may have been
3064 * released depending on flags and vma->vm_ops->fault() return value.
3065 * See filemap_fault() and __lock_page_retry().
3067 static int __do_fault(struct vm_fault *vmf)
3069 struct vm_area_struct *vma = vmf->vma;
3072 ret = vma->vm_ops->fault(vmf);
3073 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3074 VM_FAULT_DONE_COW)))
3077 if (unlikely(PageHWPoison(vmf->page))) {
3078 if (ret & VM_FAULT_LOCKED)
3079 unlock_page(vmf->page);
3080 put_page(vmf->page);
3082 return VM_FAULT_HWPOISON;
3085 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3086 lock_page(vmf->page);
3088 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3094 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3095 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3096 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3097 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3099 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3101 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3104 static int pte_alloc_one_map(struct vm_fault *vmf)
3106 struct vm_area_struct *vma = vmf->vma;
3108 if (!pmd_none(*vmf->pmd))
3110 if (vmf->prealloc_pte) {
3111 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3112 if (unlikely(!pmd_none(*vmf->pmd))) {
3113 spin_unlock(vmf->ptl);
3117 atomic_long_inc(&vma->vm_mm->nr_ptes);
3118 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3119 spin_unlock(vmf->ptl);
3120 vmf->prealloc_pte = NULL;
3121 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3122 return VM_FAULT_OOM;
3126 * If a huge pmd materialized under us just retry later. Use
3127 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3128 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3129 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3130 * running immediately after a huge pmd fault in a different thread of
3131 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3132 * All we have to ensure is that it is a regular pmd that we can walk
3133 * with pte_offset_map() and we can do that through an atomic read in
3134 * C, which is what pmd_trans_unstable() provides.
3136 if (pmd_devmap_trans_unstable(vmf->pmd))
3137 return VM_FAULT_NOPAGE;
3140 * At this point we know that our vmf->pmd points to a page of ptes
3141 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3142 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3143 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3144 * be valid and we will re-check to make sure the vmf->pte isn't
3145 * pte_none() under vmf->ptl protection when we return to
3148 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3153 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3155 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3156 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3157 unsigned long haddr)
3159 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3160 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3162 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3167 static void deposit_prealloc_pte(struct vm_fault *vmf)
3169 struct vm_area_struct *vma = vmf->vma;
3171 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3173 * We are going to consume the prealloc table,
3174 * count that as nr_ptes.
3176 atomic_long_inc(&vma->vm_mm->nr_ptes);
3177 vmf->prealloc_pte = NULL;
3180 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3182 struct vm_area_struct *vma = vmf->vma;
3183 bool write = vmf->flags & FAULT_FLAG_WRITE;
3184 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3188 if (!transhuge_vma_suitable(vma, haddr))
3189 return VM_FAULT_FALLBACK;
3191 ret = VM_FAULT_FALLBACK;
3192 page = compound_head(page);
3195 * Archs like ppc64 need additonal space to store information
3196 * related to pte entry. Use the preallocated table for that.
3198 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3199 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3200 if (!vmf->prealloc_pte)
3201 return VM_FAULT_OOM;
3202 smp_wmb(); /* See comment in __pte_alloc() */
3205 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3206 if (unlikely(!pmd_none(*vmf->pmd)))
3209 for (i = 0; i < HPAGE_PMD_NR; i++)
3210 flush_icache_page(vma, page + i);
3212 entry = mk_huge_pmd(page, vma->vm_page_prot);
3214 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3216 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3217 page_add_file_rmap(page, true);
3219 * deposit and withdraw with pmd lock held
3221 if (arch_needs_pgtable_deposit())
3222 deposit_prealloc_pte(vmf);
3224 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3226 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3228 /* fault is handled */
3230 count_vm_event(THP_FILE_MAPPED);
3232 spin_unlock(vmf->ptl);
3236 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3244 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3245 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3247 * @vmf: fault environment
3248 * @memcg: memcg to charge page (only for private mappings)
3249 * @page: page to map
3251 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3254 * Target users are page handler itself and implementations of
3255 * vm_ops->map_pages.
3257 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3260 struct vm_area_struct *vma = vmf->vma;
3261 bool write = vmf->flags & FAULT_FLAG_WRITE;
3265 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3266 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3268 VM_BUG_ON_PAGE(memcg, page);
3270 ret = do_set_pmd(vmf, page);
3271 if (ret != VM_FAULT_FALLBACK)
3276 ret = pte_alloc_one_map(vmf);
3281 /* Re-check under ptl */
3282 if (unlikely(!pte_none(*vmf->pte)))
3283 return VM_FAULT_NOPAGE;
3285 flush_icache_page(vma, page);
3286 entry = mk_pte(page, vma->vm_page_prot);
3288 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3289 /* copy-on-write page */
3290 if (write && !(vma->vm_flags & VM_SHARED)) {
3291 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3292 page_add_new_anon_rmap(page, vma, vmf->address, false);
3293 mem_cgroup_commit_charge(page, memcg, false, false);
3294 lru_cache_add_active_or_unevictable(page, vma);
3296 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3297 page_add_file_rmap(page, false);
3299 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3301 /* no need to invalidate: a not-present page won't be cached */
3302 update_mmu_cache(vma, vmf->address, vmf->pte);
3309 * finish_fault - finish page fault once we have prepared the page to fault
3311 * @vmf: structure describing the fault
3313 * This function handles all that is needed to finish a page fault once the
3314 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3315 * given page, adds reverse page mapping, handles memcg charges and LRU
3316 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3319 * The function expects the page to be locked and on success it consumes a
3320 * reference of a page being mapped (for the PTE which maps it).
3322 int finish_fault(struct vm_fault *vmf)
3327 /* Did we COW the page? */
3328 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3329 !(vmf->vma->vm_flags & VM_SHARED))
3330 page = vmf->cow_page;
3335 * check even for read faults because we might have lost our CoWed
3338 if (!(vmf->vma->vm_flags & VM_SHARED))
3339 ret = check_stable_address_space(vmf->vma->vm_mm);
3341 ret = alloc_set_pte(vmf, vmf->memcg, page);
3343 pte_unmap_unlock(vmf->pte, vmf->ptl);
3347 static unsigned long fault_around_bytes __read_mostly =
3348 rounddown_pow_of_two(65536);
3350 #ifdef CONFIG_DEBUG_FS
3351 static int fault_around_bytes_get(void *data, u64 *val)
3353 *val = fault_around_bytes;
3358 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3359 * rounded down to nearest page order. It's what do_fault_around() expects to
3362 static int fault_around_bytes_set(void *data, u64 val)
3364 if (val / PAGE_SIZE > PTRS_PER_PTE)
3366 if (val > PAGE_SIZE)
3367 fault_around_bytes = rounddown_pow_of_two(val);
3369 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3372 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3373 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3375 static int __init fault_around_debugfs(void)
3379 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3380 &fault_around_bytes_fops);
3382 pr_warn("Failed to create fault_around_bytes in debugfs");
3385 late_initcall(fault_around_debugfs);
3389 * do_fault_around() tries to map few pages around the fault address. The hope
3390 * is that the pages will be needed soon and this will lower the number of
3393 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3394 * not ready to be mapped: not up-to-date, locked, etc.
3396 * This function is called with the page table lock taken. In the split ptlock
3397 * case the page table lock only protects only those entries which belong to
3398 * the page table corresponding to the fault address.
3400 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3403 * fault_around_pages() defines how many pages we'll try to map.
3404 * do_fault_around() expects it to return a power of two less than or equal to
3407 * The virtual address of the area that we map is naturally aligned to the
3408 * fault_around_pages() value (and therefore to page order). This way it's
3409 * easier to guarantee that we don't cross page table boundaries.
3411 static int do_fault_around(struct vm_fault *vmf)
3413 unsigned long address = vmf->address, nr_pages, mask;
3414 pgoff_t start_pgoff = vmf->pgoff;
3418 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3419 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3421 vmf->address = max(address & mask, vmf->vma->vm_start);
3422 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3426 * end_pgoff is either end of page table or end of vma
3427 * or fault_around_pages() from start_pgoff, depending what is nearest.
3429 end_pgoff = start_pgoff -
3430 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3432 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3433 start_pgoff + nr_pages - 1);
3435 if (pmd_none(*vmf->pmd)) {
3436 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3438 if (!vmf->prealloc_pte)
3440 smp_wmb(); /* See comment in __pte_alloc() */
3443 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3445 /* Huge page is mapped? Page fault is solved */
3446 if (pmd_trans_huge(*vmf->pmd)) {
3447 ret = VM_FAULT_NOPAGE;
3451 /* ->map_pages() haven't done anything useful. Cold page cache? */
3455 /* check if the page fault is solved */
3456 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3457 if (!pte_none(*vmf->pte))
3458 ret = VM_FAULT_NOPAGE;
3459 pte_unmap_unlock(vmf->pte, vmf->ptl);
3461 vmf->address = address;
3466 static int do_read_fault(struct vm_fault *vmf)
3468 struct vm_area_struct *vma = vmf->vma;
3472 * Let's call ->map_pages() first and use ->fault() as fallback
3473 * if page by the offset is not ready to be mapped (cold cache or
3476 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3477 ret = do_fault_around(vmf);
3482 ret = __do_fault(vmf);
3483 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3486 ret |= finish_fault(vmf);
3487 unlock_page(vmf->page);
3488 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3489 put_page(vmf->page);
3493 static int do_cow_fault(struct vm_fault *vmf)
3495 struct vm_area_struct *vma = vmf->vma;
3498 if (unlikely(anon_vma_prepare(vma)))
3499 return VM_FAULT_OOM;
3501 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3503 return VM_FAULT_OOM;
3505 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3506 &vmf->memcg, false)) {
3507 put_page(vmf->cow_page);
3508 return VM_FAULT_OOM;
3511 ret = __do_fault(vmf);
3512 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3514 if (ret & VM_FAULT_DONE_COW)
3517 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3518 __SetPageUptodate(vmf->cow_page);
3520 ret |= finish_fault(vmf);
3521 unlock_page(vmf->page);
3522 put_page(vmf->page);
3523 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3527 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3528 put_page(vmf->cow_page);
3532 static int do_shared_fault(struct vm_fault *vmf)
3534 struct vm_area_struct *vma = vmf->vma;
3537 ret = __do_fault(vmf);
3538 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3542 * Check if the backing address space wants to know that the page is
3543 * about to become writable
3545 if (vma->vm_ops->page_mkwrite) {
3546 unlock_page(vmf->page);
3547 tmp = do_page_mkwrite(vmf);
3548 if (unlikely(!tmp ||
3549 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3550 put_page(vmf->page);
3555 ret |= finish_fault(vmf);
3556 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3558 unlock_page(vmf->page);
3559 put_page(vmf->page);
3563 fault_dirty_shared_page(vma, vmf->page);
3568 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3569 * but allow concurrent faults).
3570 * The mmap_sem may have been released depending on flags and our
3571 * return value. See filemap_fault() and __lock_page_or_retry().
3573 static int do_fault(struct vm_fault *vmf)
3575 struct vm_area_struct *vma = vmf->vma;
3578 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3579 if (!vma->vm_ops->fault)
3580 ret = VM_FAULT_SIGBUS;
3581 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3582 ret = do_read_fault(vmf);
3583 else if (!(vma->vm_flags & VM_SHARED))
3584 ret = do_cow_fault(vmf);
3586 ret = do_shared_fault(vmf);
3588 /* preallocated pagetable is unused: free it */
3589 if (vmf->prealloc_pte) {
3590 pte_free(vma->vm_mm, vmf->prealloc_pte);
3591 vmf->prealloc_pte = NULL;
3596 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3597 unsigned long addr, int page_nid,
3602 count_vm_numa_event(NUMA_HINT_FAULTS);
3603 if (page_nid == numa_node_id()) {
3604 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3605 *flags |= TNF_FAULT_LOCAL;
3608 return mpol_misplaced(page, vma, addr);
3611 static int do_numa_page(struct vm_fault *vmf)
3613 struct vm_area_struct *vma = vmf->vma;
3614 struct page *page = NULL;
3618 bool migrated = false;
3620 bool was_writable = pte_savedwrite(vmf->orig_pte);
3624 * The "pte" at this point cannot be used safely without
3625 * validation through pte_unmap_same(). It's of NUMA type but
3626 * the pfn may be screwed if the read is non atomic.
3628 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3629 spin_lock(vmf->ptl);
3630 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3631 pte_unmap_unlock(vmf->pte, vmf->ptl);
3636 * Make it present again, Depending on how arch implementes non
3637 * accessible ptes, some can allow access by kernel mode.
3639 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3640 pte = pte_modify(pte, vma->vm_page_prot);
3641 pte = pte_mkyoung(pte);
3643 pte = pte_mkwrite(pte);
3644 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3645 update_mmu_cache(vma, vmf->address, vmf->pte);
3647 page = vm_normal_page(vma, vmf->address, pte);
3649 pte_unmap_unlock(vmf->pte, vmf->ptl);
3653 /* TODO: handle PTE-mapped THP */
3654 if (PageCompound(page)) {
3655 pte_unmap_unlock(vmf->pte, vmf->ptl);
3660 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3661 * much anyway since they can be in shared cache state. This misses
3662 * the case where a mapping is writable but the process never writes
3663 * to it but pte_write gets cleared during protection updates and
3664 * pte_dirty has unpredictable behaviour between PTE scan updates,
3665 * background writeback, dirty balancing and application behaviour.
3667 if (!pte_write(pte))
3668 flags |= TNF_NO_GROUP;
3671 * Flag if the page is shared between multiple address spaces. This
3672 * is later used when determining whether to group tasks together
3674 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3675 flags |= TNF_SHARED;
3677 last_cpupid = page_cpupid_last(page);
3678 page_nid = page_to_nid(page);
3679 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3681 pte_unmap_unlock(vmf->pte, vmf->ptl);
3682 if (target_nid == -1) {
3687 /* Migrate to the requested node */
3688 migrated = migrate_misplaced_page(page, vma, target_nid);
3690 page_nid = target_nid;
3691 flags |= TNF_MIGRATED;
3693 flags |= TNF_MIGRATE_FAIL;
3697 task_numa_fault(last_cpupid, page_nid, 1, flags);
3701 static inline int create_huge_pmd(struct vm_fault *vmf)
3703 if (vma_is_anonymous(vmf->vma))
3704 return do_huge_pmd_anonymous_page(vmf);
3705 if (vmf->vma->vm_ops->huge_fault)
3706 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3707 return VM_FAULT_FALLBACK;
3710 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3712 if (vma_is_anonymous(vmf->vma))
3713 return do_huge_pmd_wp_page(vmf, orig_pmd);
3714 if (vmf->vma->vm_ops->huge_fault)
3715 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3717 /* COW handled on pte level: split pmd */
3718 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3719 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3721 return VM_FAULT_FALLBACK;
3724 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3726 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3729 static int create_huge_pud(struct vm_fault *vmf)
3731 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3732 /* No support for anonymous transparent PUD pages yet */
3733 if (vma_is_anonymous(vmf->vma))
3734 return VM_FAULT_FALLBACK;
3735 if (vmf->vma->vm_ops->huge_fault)
3736 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3737 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3738 return VM_FAULT_FALLBACK;
3741 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3743 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3744 /* No support for anonymous transparent PUD pages yet */
3745 if (vma_is_anonymous(vmf->vma))
3746 return VM_FAULT_FALLBACK;
3747 if (vmf->vma->vm_ops->huge_fault)
3748 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3749 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3750 return VM_FAULT_FALLBACK;
3754 * These routines also need to handle stuff like marking pages dirty
3755 * and/or accessed for architectures that don't do it in hardware (most
3756 * RISC architectures). The early dirtying is also good on the i386.
3758 * There is also a hook called "update_mmu_cache()" that architectures
3759 * with external mmu caches can use to update those (ie the Sparc or
3760 * PowerPC hashed page tables that act as extended TLBs).
3762 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3763 * concurrent faults).
3765 * The mmap_sem may have been released depending on flags and our return value.
3766 * See filemap_fault() and __lock_page_or_retry().
3768 static int handle_pte_fault(struct vm_fault *vmf)
3772 if (unlikely(pmd_none(*vmf->pmd))) {
3774 * Leave __pte_alloc() until later: because vm_ops->fault may
3775 * want to allocate huge page, and if we expose page table
3776 * for an instant, it will be difficult to retract from
3777 * concurrent faults and from rmap lookups.
3781 /* See comment in pte_alloc_one_map() */
3782 if (pmd_devmap_trans_unstable(vmf->pmd))
3785 * A regular pmd is established and it can't morph into a huge
3786 * pmd from under us anymore at this point because we hold the
3787 * mmap_sem read mode and khugepaged takes it in write mode.
3788 * So now it's safe to run pte_offset_map().
3790 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3791 vmf->orig_pte = *vmf->pte;
3794 * some architectures can have larger ptes than wordsize,
3795 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3796 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3797 * atomic accesses. The code below just needs a consistent
3798 * view for the ifs and we later double check anyway with the
3799 * ptl lock held. So here a barrier will do.
3802 if (pte_none(vmf->orig_pte)) {
3803 pte_unmap(vmf->pte);
3809 if (vma_is_anonymous(vmf->vma))
3810 return do_anonymous_page(vmf);
3812 return do_fault(vmf);
3815 if (!pte_present(vmf->orig_pte))
3816 return do_swap_page(vmf);
3818 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3819 return do_numa_page(vmf);
3821 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3822 spin_lock(vmf->ptl);
3823 entry = vmf->orig_pte;
3824 if (unlikely(!pte_same(*vmf->pte, entry)))
3826 if (vmf->flags & FAULT_FLAG_WRITE) {
3827 if (!pte_write(entry))
3828 return do_wp_page(vmf);
3829 entry = pte_mkdirty(entry);
3831 entry = pte_mkyoung(entry);
3832 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3833 vmf->flags & FAULT_FLAG_WRITE)) {
3834 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3837 * This is needed only for protection faults but the arch code
3838 * is not yet telling us if this is a protection fault or not.
3839 * This still avoids useless tlb flushes for .text page faults
3842 if (vmf->flags & FAULT_FLAG_WRITE)
3843 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3846 pte_unmap_unlock(vmf->pte, vmf->ptl);
3851 * By the time we get here, we already hold the mm semaphore
3853 * The mmap_sem may have been released depending on flags and our
3854 * return value. See filemap_fault() and __lock_page_or_retry().
3856 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3859 struct vm_fault vmf = {
3861 .address = address & PAGE_MASK,
3863 .pgoff = linear_page_index(vma, address),
3864 .gfp_mask = __get_fault_gfp_mask(vma),
3866 struct mm_struct *mm = vma->vm_mm;
3871 pgd = pgd_offset(mm, address);
3872 p4d = p4d_alloc(mm, pgd, address);
3874 return VM_FAULT_OOM;
3876 vmf.pud = pud_alloc(mm, p4d, address);
3878 return VM_FAULT_OOM;
3879 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3880 ret = create_huge_pud(&vmf);
3881 if (!(ret & VM_FAULT_FALLBACK))
3884 pud_t orig_pud = *vmf.pud;
3887 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3888 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3890 /* NUMA case for anonymous PUDs would go here */
3892 if (dirty && !pud_write(orig_pud)) {
3893 ret = wp_huge_pud(&vmf, orig_pud);
3894 if (!(ret & VM_FAULT_FALLBACK))
3897 huge_pud_set_accessed(&vmf, orig_pud);
3903 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3905 return VM_FAULT_OOM;
3906 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3907 ret = create_huge_pmd(&vmf);
3908 if (!(ret & VM_FAULT_FALLBACK))
3911 pmd_t orig_pmd = *vmf.pmd;
3914 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3915 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3916 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3918 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3919 !pmd_write(orig_pmd)) {
3920 ret = wp_huge_pmd(&vmf, orig_pmd);
3921 if (!(ret & VM_FAULT_FALLBACK))
3924 huge_pmd_set_accessed(&vmf, orig_pmd);
3930 return handle_pte_fault(&vmf);
3934 * By the time we get here, we already hold the mm semaphore
3936 * The mmap_sem may have been released depending on flags and our
3937 * return value. See filemap_fault() and __lock_page_or_retry().
3939 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3944 __set_current_state(TASK_RUNNING);
3946 count_vm_event(PGFAULT);
3947 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3949 /* do counter updates before entering really critical section. */
3950 check_sync_rss_stat(current);
3953 * Enable the memcg OOM handling for faults triggered in user
3954 * space. Kernel faults are handled more gracefully.
3956 if (flags & FAULT_FLAG_USER)
3957 mem_cgroup_oom_enable();
3959 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3960 flags & FAULT_FLAG_INSTRUCTION,
3961 flags & FAULT_FLAG_REMOTE))
3962 return VM_FAULT_SIGSEGV;
3964 if (unlikely(is_vm_hugetlb_page(vma)))
3965 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3967 ret = __handle_mm_fault(vma, address, flags);
3969 if (flags & FAULT_FLAG_USER) {
3970 mem_cgroup_oom_disable();
3972 * The task may have entered a memcg OOM situation but
3973 * if the allocation error was handled gracefully (no
3974 * VM_FAULT_OOM), there is no need to kill anything.
3975 * Just clean up the OOM state peacefully.
3977 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3978 mem_cgroup_oom_synchronize(false);
3983 EXPORT_SYMBOL_GPL(handle_mm_fault);
3985 #ifndef __PAGETABLE_P4D_FOLDED
3987 * Allocate p4d page table.
3988 * We've already handled the fast-path in-line.
3990 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3992 p4d_t *new = p4d_alloc_one(mm, address);
3996 smp_wmb(); /* See comment in __pte_alloc */
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4006 #endif /* __PAGETABLE_P4D_FOLDED */
4008 #ifndef __PAGETABLE_PUD_FOLDED
4010 * Allocate page upper directory.
4011 * We've already handled the fast-path in-line.
4013 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4015 pud_t *new = pud_alloc_one(mm, address);
4019 smp_wmb(); /* See comment in __pte_alloc */
4021 spin_lock(&mm->page_table_lock);
4022 #ifndef __ARCH_HAS_5LEVEL_HACK
4023 if (p4d_present(*p4d)) /* Another has populated it */
4026 p4d_populate(mm, p4d, new);
4028 if (pgd_present(*p4d)) /* Another has populated it */
4031 pgd_populate(mm, p4d, new);
4032 #endif /* __ARCH_HAS_5LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4036 #endif /* __PAGETABLE_PUD_FOLDED */
4038 #ifndef __PAGETABLE_PMD_FOLDED
4040 * Allocate page middle directory.
4041 * We've already handled the fast-path in-line.
4043 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4046 pmd_t *new = pmd_alloc_one(mm, address);
4050 smp_wmb(); /* See comment in __pte_alloc */
4052 ptl = pud_lock(mm, pud);
4053 #ifndef __ARCH_HAS_4LEVEL_HACK
4054 if (!pud_present(*pud)) {
4056 pud_populate(mm, pud, new);
4057 } else /* Another has populated it */
4060 if (!pgd_present(*pud)) {
4062 pgd_populate(mm, pud, new);
4063 } else /* Another has populated it */
4065 #endif /* __ARCH_HAS_4LEVEL_HACK */
4069 #endif /* __PAGETABLE_PMD_FOLDED */
4071 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4072 unsigned long *start, unsigned long *end,
4073 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4081 pgd = pgd_offset(mm, address);
4082 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4085 p4d = p4d_offset(pgd, address);
4086 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4089 pud = pud_offset(p4d, address);
4090 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4093 pmd = pmd_offset(pud, address);
4094 VM_BUG_ON(pmd_trans_huge(*pmd));
4096 if (pmd_huge(*pmd)) {
4101 *start = address & PMD_MASK;
4102 *end = *start + PMD_SIZE;
4103 mmu_notifier_invalidate_range_start(mm, *start, *end);
4105 *ptlp = pmd_lock(mm, pmd);
4106 if (pmd_huge(*pmd)) {
4112 mmu_notifier_invalidate_range_end(mm, *start, *end);
4115 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4119 *start = address & PAGE_MASK;
4120 *end = *start + PAGE_SIZE;
4121 mmu_notifier_invalidate_range_start(mm, *start, *end);
4123 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4124 if (!pte_present(*ptep))
4129 pte_unmap_unlock(ptep, *ptlp);
4131 mmu_notifier_invalidate_range_end(mm, *start, *end);
4136 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4137 pte_t **ptepp, spinlock_t **ptlp)
4141 /* (void) is needed to make gcc happy */
4142 (void) __cond_lock(*ptlp,
4143 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4144 ptepp, NULL, ptlp)));
4148 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4149 unsigned long *start, unsigned long *end,
4150 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4154 /* (void) is needed to make gcc happy */
4155 (void) __cond_lock(*ptlp,
4156 !(res = __follow_pte_pmd(mm, address, start, end,
4157 ptepp, pmdpp, ptlp)));
4160 EXPORT_SYMBOL(follow_pte_pmd);
4163 * follow_pfn - look up PFN at a user virtual address
4164 * @vma: memory mapping
4165 * @address: user virtual address
4166 * @pfn: location to store found PFN
4168 * Only IO mappings and raw PFN mappings are allowed.
4170 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4172 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4179 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4182 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4185 *pfn = pte_pfn(*ptep);
4186 pte_unmap_unlock(ptep, ptl);
4189 EXPORT_SYMBOL(follow_pfn);
4191 #ifdef CONFIG_HAVE_IOREMAP_PROT
4192 int follow_phys(struct vm_area_struct *vma,
4193 unsigned long address, unsigned int flags,
4194 unsigned long *prot, resource_size_t *phys)
4200 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4203 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4207 if ((flags & FOLL_WRITE) && !pte_write(pte))
4210 *prot = pgprot_val(pte_pgprot(pte));
4211 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4215 pte_unmap_unlock(ptep, ptl);
4220 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4221 void *buf, int len, int write)
4223 resource_size_t phys_addr;
4224 unsigned long prot = 0;
4225 void __iomem *maddr;
4226 int offset = addr & (PAGE_SIZE-1);
4228 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4231 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4233 memcpy_toio(maddr + offset, buf, len);
4235 memcpy_fromio(buf, maddr + offset, len);
4240 EXPORT_SYMBOL_GPL(generic_access_phys);
4244 * Access another process' address space as given in mm. If non-NULL, use the
4245 * given task for page fault accounting.
4247 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4248 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4250 struct vm_area_struct *vma;
4251 void *old_buf = buf;
4252 int write = gup_flags & FOLL_WRITE;
4254 down_read(&mm->mmap_sem);
4255 /* ignore errors, just check how much was successfully transferred */
4257 int bytes, ret, offset;
4259 struct page *page = NULL;
4261 ret = get_user_pages_remote(tsk, mm, addr, 1,
4262 gup_flags, &page, &vma, NULL);
4264 #ifndef CONFIG_HAVE_IOREMAP_PROT
4268 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4269 * we can access using slightly different code.
4271 vma = find_vma(mm, addr);
4272 if (!vma || vma->vm_start > addr)
4274 if (vma->vm_ops && vma->vm_ops->access)
4275 ret = vma->vm_ops->access(vma, addr, buf,
4283 offset = addr & (PAGE_SIZE-1);
4284 if (bytes > PAGE_SIZE-offset)
4285 bytes = PAGE_SIZE-offset;
4289 copy_to_user_page(vma, page, addr,
4290 maddr + offset, buf, bytes);
4291 set_page_dirty_lock(page);
4293 copy_from_user_page(vma, page, addr,
4294 buf, maddr + offset, bytes);
4303 up_read(&mm->mmap_sem);
4305 return buf - old_buf;
4309 * access_remote_vm - access another process' address space
4310 * @mm: the mm_struct of the target address space
4311 * @addr: start address to access
4312 * @buf: source or destination buffer
4313 * @len: number of bytes to transfer
4314 * @gup_flags: flags modifying lookup behaviour
4316 * The caller must hold a reference on @mm.
4318 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4319 void *buf, int len, unsigned int gup_flags)
4321 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4325 * Access another process' address space.
4326 * Source/target buffer must be kernel space,
4327 * Do not walk the page table directly, use get_user_pages
4329 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4330 void *buf, int len, unsigned int gup_flags)
4332 struct mm_struct *mm;
4335 mm = get_task_mm(tsk);
4339 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4345 EXPORT_SYMBOL_GPL(access_process_vm);
4348 * Print the name of a VMA.
4350 void print_vma_addr(char *prefix, unsigned long ip)
4352 struct mm_struct *mm = current->mm;
4353 struct vm_area_struct *vma;
4356 * Do not print if we are in atomic
4357 * contexts (in exception stacks, etc.):
4359 if (preempt_count())
4362 down_read(&mm->mmap_sem);
4363 vma = find_vma(mm, ip);
4364 if (vma && vma->vm_file) {
4365 struct file *f = vma->vm_file;
4366 char *buf = (char *)__get_free_page(GFP_KERNEL);
4370 p = file_path(f, buf, PAGE_SIZE);
4373 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4375 vma->vm_end - vma->vm_start);
4376 free_page((unsigned long)buf);
4379 up_read(&mm->mmap_sem);
4382 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4383 void __might_fault(const char *file, int line)
4386 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4387 * holding the mmap_sem, this is safe because kernel memory doesn't
4388 * get paged out, therefore we'll never actually fault, and the
4389 * below annotations will generate false positives.
4391 if (uaccess_kernel())
4393 if (pagefault_disabled())
4395 __might_sleep(file, line, 0);
4396 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4398 might_lock_read(¤t->mm->mmap_sem);
4401 EXPORT_SYMBOL(__might_fault);
4404 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4405 static void clear_gigantic_page(struct page *page,
4407 unsigned int pages_per_huge_page)
4410 struct page *p = page;
4413 for (i = 0; i < pages_per_huge_page;
4414 i++, p = mem_map_next(p, page, i)) {
4416 clear_user_highpage(p, addr + i * PAGE_SIZE);
4419 void clear_huge_page(struct page *page,
4420 unsigned long addr_hint, unsigned int pages_per_huge_page)
4423 unsigned long addr = addr_hint &
4424 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4426 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4427 clear_gigantic_page(page, addr, pages_per_huge_page);
4431 /* Clear sub-page to access last to keep its cache lines hot */
4433 n = (addr_hint - addr) / PAGE_SIZE;
4434 if (2 * n <= pages_per_huge_page) {
4435 /* If sub-page to access in first half of huge page */
4438 /* Clear sub-pages at the end of huge page */
4439 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4441 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4444 /* If sub-page to access in second half of huge page */
4445 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4446 l = pages_per_huge_page - n;
4447 /* Clear sub-pages at the begin of huge page */
4448 for (i = 0; i < base; i++) {
4450 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4454 * Clear remaining sub-pages in left-right-left-right pattern
4455 * towards the sub-page to access
4457 for (i = 0; i < l; i++) {
4458 int left_idx = base + i;
4459 int right_idx = base + 2 * l - 1 - i;
4462 clear_user_highpage(page + left_idx,
4463 addr + left_idx * PAGE_SIZE);
4465 clear_user_highpage(page + right_idx,
4466 addr + right_idx * PAGE_SIZE);
4470 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4472 struct vm_area_struct *vma,
4473 unsigned int pages_per_huge_page)
4476 struct page *dst_base = dst;
4477 struct page *src_base = src;
4479 for (i = 0; i < pages_per_huge_page; ) {
4481 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4484 dst = mem_map_next(dst, dst_base, i);
4485 src = mem_map_next(src, src_base, i);
4489 void copy_user_huge_page(struct page *dst, struct page *src,
4490 unsigned long addr, struct vm_area_struct *vma,
4491 unsigned int pages_per_huge_page)
4495 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4496 copy_user_gigantic_page(dst, src, addr, vma,
4497 pages_per_huge_page);
4502 for (i = 0; i < pages_per_huge_page; i++) {
4504 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4508 long copy_huge_page_from_user(struct page *dst_page,
4509 const void __user *usr_src,
4510 unsigned int pages_per_huge_page,
4511 bool allow_pagefault)
4513 void *src = (void *)usr_src;
4515 unsigned long i, rc = 0;
4516 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4518 for (i = 0; i < pages_per_huge_page; i++) {
4519 if (allow_pagefault)
4520 page_kaddr = kmap(dst_page + i);
4522 page_kaddr = kmap_atomic(dst_page + i);
4523 rc = copy_from_user(page_kaddr,
4524 (const void __user *)(src + i * PAGE_SIZE),
4526 if (allow_pagefault)
4527 kunmap(dst_page + i);
4529 kunmap_atomic(page_kaddr);
4531 ret_val -= (PAGE_SIZE - rc);
4539 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4541 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4543 static struct kmem_cache *page_ptl_cachep;
4545 void __init ptlock_cache_init(void)
4547 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4551 bool ptlock_alloc(struct page *page)
4555 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4562 void ptlock_free(struct page *page)
4564 kmem_cache_free(page_ptl_cachep, page->ptl);