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>
73 #include <asm/mmu_context.h>
74 #include <asm/pgalloc.h>
75 #include <linux/uaccess.h>
77 #include <asm/tlbflush.h>
78 #include <asm/pgtable.h>
82 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
83 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
86 #ifndef CONFIG_NEED_MULTIPLE_NODES
87 /* use the per-pgdat data instead for discontigmem - mbligh */
88 unsigned long max_mapnr;
89 EXPORT_SYMBOL(max_mapnr);
92 EXPORT_SYMBOL(mem_map);
96 * A number of key systems in x86 including ioremap() rely on the assumption
97 * that high_memory defines the upper bound on direct map memory, then end
98 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
99 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
103 EXPORT_SYMBOL(high_memory);
106 * Randomize the address space (stacks, mmaps, brk, etc.).
108 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
109 * as ancient (libc5 based) binaries can segfault. )
111 int randomize_va_space __read_mostly =
112 #ifdef CONFIG_COMPAT_BRK
118 static int __init disable_randmaps(char *s)
120 randomize_va_space = 0;
123 __setup("norandmaps", disable_randmaps);
125 unsigned long zero_pfn __read_mostly;
126 EXPORT_SYMBOL(zero_pfn);
128 unsigned long highest_memmap_pfn __read_mostly;
131 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
133 static int __init init_zero_pfn(void)
135 zero_pfn = page_to_pfn(ZERO_PAGE(0));
138 core_initcall(init_zero_pfn);
141 #if defined(SPLIT_RSS_COUNTING)
143 void sync_mm_rss(struct mm_struct *mm)
147 for (i = 0; i < NR_MM_COUNTERS; i++) {
148 if (current->rss_stat.count[i]) {
149 add_mm_counter(mm, i, current->rss_stat.count[i]);
150 current->rss_stat.count[i] = 0;
153 current->rss_stat.events = 0;
156 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
158 struct task_struct *task = current;
160 if (likely(task->mm == mm))
161 task->rss_stat.count[member] += val;
163 add_mm_counter(mm, member, val);
165 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
166 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
168 /* sync counter once per 64 page faults */
169 #define TASK_RSS_EVENTS_THRESH (64)
170 static void check_sync_rss_stat(struct task_struct *task)
172 if (unlikely(task != current))
174 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
175 sync_mm_rss(task->mm);
177 #else /* SPLIT_RSS_COUNTING */
179 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
180 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
182 static void check_sync_rss_stat(struct task_struct *task)
186 #endif /* SPLIT_RSS_COUNTING */
188 #ifdef HAVE_GENERIC_MMU_GATHER
190 static bool tlb_next_batch(struct mmu_gather *tlb)
192 struct mmu_gather_batch *batch;
196 tlb->active = batch->next;
200 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
203 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
210 batch->max = MAX_GATHER_BATCH;
212 tlb->active->next = batch;
219 * Called to initialize an (on-stack) mmu_gather structure for page-table
220 * tear-down from @mm. The @fullmm argument is used when @mm is without
221 * users and we're going to destroy the full address space (exit/execve).
223 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
227 /* Is it from 0 to ~0? */
228 tlb->fullmm = !(start | (end+1));
229 tlb->need_flush_all = 0;
230 tlb->local.next = NULL;
232 tlb->local.max = ARRAY_SIZE(tlb->__pages);
233 tlb->active = &tlb->local;
234 tlb->batch_count = 0;
236 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
241 __tlb_reset_range(tlb);
244 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
250 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
251 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
252 tlb_table_flush(tlb);
254 __tlb_reset_range(tlb);
257 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
259 struct mmu_gather_batch *batch;
261 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
262 free_pages_and_swap_cache(batch->pages, batch->nr);
265 tlb->active = &tlb->local;
268 void tlb_flush_mmu(struct mmu_gather *tlb)
270 tlb_flush_mmu_tlbonly(tlb);
271 tlb_flush_mmu_free(tlb);
275 * Called at the end of the shootdown operation to free up any resources
276 * that were required.
278 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
280 struct mmu_gather_batch *batch, *next;
284 /* keep the page table cache within bounds */
287 for (batch = tlb->local.next; batch; batch = next) {
289 free_pages((unsigned long)batch, 0);
291 tlb->local.next = NULL;
295 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
296 * handling the additional races in SMP caused by other CPUs caching valid
297 * mappings in their TLBs. Returns the number of free page slots left.
298 * When out of page slots we must call tlb_flush_mmu().
299 *returns true if the caller should flush.
301 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
303 struct mmu_gather_batch *batch;
305 VM_BUG_ON(!tlb->end);
306 VM_WARN_ON(tlb->page_size != page_size);
310 * Add the page and check if we are full. If so
313 batch->pages[batch->nr++] = page;
314 if (batch->nr == batch->max) {
315 if (!tlb_next_batch(tlb))
319 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
324 #endif /* HAVE_GENERIC_MMU_GATHER */
326 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
329 * See the comment near struct mmu_table_batch.
332 static void tlb_remove_table_smp_sync(void *arg)
334 /* Simply deliver the interrupt */
337 static void tlb_remove_table_one(void *table)
340 * This isn't an RCU grace period and hence the page-tables cannot be
341 * assumed to be actually RCU-freed.
343 * It is however sufficient for software page-table walkers that rely on
344 * IRQ disabling. See the comment near struct mmu_table_batch.
346 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
347 __tlb_remove_table(table);
350 static void tlb_remove_table_rcu(struct rcu_head *head)
352 struct mmu_table_batch *batch;
355 batch = container_of(head, struct mmu_table_batch, rcu);
357 for (i = 0; i < batch->nr; i++)
358 __tlb_remove_table(batch->tables[i]);
360 free_page((unsigned long)batch);
363 void tlb_table_flush(struct mmu_gather *tlb)
365 struct mmu_table_batch **batch = &tlb->batch;
368 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
373 void tlb_remove_table(struct mmu_gather *tlb, void *table)
375 struct mmu_table_batch **batch = &tlb->batch;
378 * When there's less then two users of this mm there cannot be a
379 * concurrent page-table walk.
381 if (atomic_read(&tlb->mm->mm_users) < 2) {
382 __tlb_remove_table(table);
386 if (*batch == NULL) {
387 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
388 if (*batch == NULL) {
389 tlb_remove_table_one(table);
394 (*batch)->tables[(*batch)->nr++] = table;
395 if ((*batch)->nr == MAX_TABLE_BATCH)
396 tlb_table_flush(tlb);
399 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
402 * Note: this doesn't free the actual pages themselves. That
403 * has been handled earlier when unmapping all the memory regions.
405 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
408 pgtable_t token = pmd_pgtable(*pmd);
410 pte_free_tlb(tlb, token, addr);
411 atomic_long_dec(&tlb->mm->nr_ptes);
414 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
415 unsigned long addr, unsigned long end,
416 unsigned long floor, unsigned long ceiling)
423 pmd = pmd_offset(pud, addr);
425 next = pmd_addr_end(addr, end);
426 if (pmd_none_or_clear_bad(pmd))
428 free_pte_range(tlb, pmd, addr);
429 } while (pmd++, addr = next, addr != end);
439 if (end - 1 > ceiling - 1)
442 pmd = pmd_offset(pud, start);
444 pmd_free_tlb(tlb, pmd, start);
445 mm_dec_nr_pmds(tlb->mm);
448 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
449 unsigned long addr, unsigned long end,
450 unsigned long floor, unsigned long ceiling)
457 pud = pud_offset(p4d, addr);
459 next = pud_addr_end(addr, end);
460 if (pud_none_or_clear_bad(pud))
462 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
463 } while (pud++, addr = next, addr != end);
473 if (end - 1 > ceiling - 1)
476 pud = pud_offset(p4d, start);
478 pud_free_tlb(tlb, pud, start);
481 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
482 unsigned long addr, unsigned long end,
483 unsigned long floor, unsigned long ceiling)
490 p4d = p4d_offset(pgd, addr);
492 next = p4d_addr_end(addr, end);
493 if (p4d_none_or_clear_bad(p4d))
495 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
496 } while (p4d++, addr = next, addr != end);
502 ceiling &= PGDIR_MASK;
506 if (end - 1 > ceiling - 1)
509 p4d = p4d_offset(pgd, start);
511 p4d_free_tlb(tlb, p4d, start);
515 * This function frees user-level page tables of a process.
517 void free_pgd_range(struct mmu_gather *tlb,
518 unsigned long addr, unsigned long end,
519 unsigned long floor, unsigned long ceiling)
525 * The next few lines have given us lots of grief...
527 * Why are we testing PMD* at this top level? Because often
528 * there will be no work to do at all, and we'd prefer not to
529 * go all the way down to the bottom just to discover that.
531 * Why all these "- 1"s? Because 0 represents both the bottom
532 * of the address space and the top of it (using -1 for the
533 * top wouldn't help much: the masks would do the wrong thing).
534 * The rule is that addr 0 and floor 0 refer to the bottom of
535 * the address space, but end 0 and ceiling 0 refer to the top
536 * Comparisons need to use "end - 1" and "ceiling - 1" (though
537 * that end 0 case should be mythical).
539 * Wherever addr is brought up or ceiling brought down, we must
540 * be careful to reject "the opposite 0" before it confuses the
541 * subsequent tests. But what about where end is brought down
542 * by PMD_SIZE below? no, end can't go down to 0 there.
544 * Whereas we round start (addr) and ceiling down, by different
545 * masks at different levels, in order to test whether a table
546 * now has no other vmas using it, so can be freed, we don't
547 * bother to round floor or end up - the tests don't need that.
561 if (end - 1 > ceiling - 1)
566 * We add page table cache pages with PAGE_SIZE,
567 * (see pte_free_tlb()), flush the tlb if we need
569 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
570 pgd = pgd_offset(tlb->mm, addr);
572 next = pgd_addr_end(addr, end);
573 if (pgd_none_or_clear_bad(pgd))
575 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
576 } while (pgd++, addr = next, addr != end);
579 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
580 unsigned long floor, unsigned long ceiling)
583 struct vm_area_struct *next = vma->vm_next;
584 unsigned long addr = vma->vm_start;
587 * Hide vma from rmap and truncate_pagecache before freeing
590 unlink_anon_vmas(vma);
591 unlink_file_vma(vma);
593 if (is_vm_hugetlb_page(vma)) {
594 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
595 floor, next ? next->vm_start : ceiling);
598 * Optimization: gather nearby vmas into one call down
600 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
601 && !is_vm_hugetlb_page(next)) {
604 unlink_anon_vmas(vma);
605 unlink_file_vma(vma);
607 free_pgd_range(tlb, addr, vma->vm_end,
608 floor, next ? next->vm_start : ceiling);
614 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
617 pgtable_t new = pte_alloc_one(mm, address);
622 * Ensure all pte setup (eg. pte page lock and page clearing) are
623 * visible before the pte is made visible to other CPUs by being
624 * put into page tables.
626 * The other side of the story is the pointer chasing in the page
627 * table walking code (when walking the page table without locking;
628 * ie. most of the time). Fortunately, these data accesses consist
629 * of a chain of data-dependent loads, meaning most CPUs (alpha
630 * being the notable exception) will already guarantee loads are
631 * seen in-order. See the alpha page table accessors for the
632 * smp_read_barrier_depends() barriers in page table walking code.
634 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
636 ptl = pmd_lock(mm, pmd);
637 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
638 atomic_long_inc(&mm->nr_ptes);
639 pmd_populate(mm, pmd, new);
648 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
650 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
654 smp_wmb(); /* See comment in __pte_alloc */
656 spin_lock(&init_mm.page_table_lock);
657 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
658 pmd_populate_kernel(&init_mm, pmd, new);
661 spin_unlock(&init_mm.page_table_lock);
663 pte_free_kernel(&init_mm, new);
667 static inline void init_rss_vec(int *rss)
669 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
672 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
676 if (current->mm == mm)
678 for (i = 0; i < NR_MM_COUNTERS; i++)
680 add_mm_counter(mm, i, rss[i]);
684 * This function is called to print an error when a bad pte
685 * is found. For example, we might have a PFN-mapped pte in
686 * a region that doesn't allow it.
688 * The calling function must still handle the error.
690 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
691 pte_t pte, struct page *page)
693 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
694 p4d_t *p4d = p4d_offset(pgd, addr);
695 pud_t *pud = pud_offset(p4d, addr);
696 pmd_t *pmd = pmd_offset(pud, addr);
697 struct address_space *mapping;
699 static unsigned long resume;
700 static unsigned long nr_shown;
701 static unsigned long nr_unshown;
704 * Allow a burst of 60 reports, then keep quiet for that minute;
705 * or allow a steady drip of one report per second.
707 if (nr_shown == 60) {
708 if (time_before(jiffies, resume)) {
713 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
720 resume = jiffies + 60 * HZ;
722 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
723 index = linear_page_index(vma, addr);
725 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
727 (long long)pte_val(pte), (long long)pmd_val(*pmd));
729 dump_page(page, "bad pte");
730 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
731 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
733 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
735 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
737 vma->vm_ops ? vma->vm_ops->fault : NULL,
738 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
739 mapping ? mapping->a_ops->readpage : NULL);
741 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
745 * vm_normal_page -- This function gets the "struct page" associated with a pte.
747 * "Special" mappings do not wish to be associated with a "struct page" (either
748 * it doesn't exist, or it exists but they don't want to touch it). In this
749 * case, NULL is returned here. "Normal" mappings do have a struct page.
751 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
752 * pte bit, in which case this function is trivial. Secondly, an architecture
753 * may not have a spare pte bit, which requires a more complicated scheme,
756 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
757 * special mapping (even if there are underlying and valid "struct pages").
758 * COWed pages of a VM_PFNMAP are always normal.
760 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
761 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
762 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
763 * mapping will always honor the rule
765 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
767 * And for normal mappings this is false.
769 * This restricts such mappings to be a linear translation from virtual address
770 * to pfn. To get around this restriction, we allow arbitrary mappings so long
771 * as the vma is not a COW mapping; in that case, we know that all ptes are
772 * special (because none can have been COWed).
775 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
777 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
778 * page" backing, however the difference is that _all_ pages with a struct
779 * page (that is, those where pfn_valid is true) are refcounted and considered
780 * normal pages by the VM. The disadvantage is that pages are refcounted
781 * (which can be slower and simply not an option for some PFNMAP users). The
782 * advantage is that we don't have to follow the strict linearity rule of
783 * PFNMAP mappings in order to support COWable mappings.
786 #ifdef __HAVE_ARCH_PTE_SPECIAL
787 # define HAVE_PTE_SPECIAL 1
789 # define HAVE_PTE_SPECIAL 0
791 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
794 unsigned long pfn = pte_pfn(pte);
796 if (HAVE_PTE_SPECIAL) {
797 if (likely(!pte_special(pte)))
799 if (vma->vm_ops && vma->vm_ops->find_special_page)
800 return vma->vm_ops->find_special_page(vma, addr);
801 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
803 if (!is_zero_pfn(pfn))
804 print_bad_pte(vma, addr, pte, NULL);
808 /* !HAVE_PTE_SPECIAL case follows: */
810 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
811 if (vma->vm_flags & VM_MIXEDMAP) {
817 off = (addr - vma->vm_start) >> PAGE_SHIFT;
818 if (pfn == vma->vm_pgoff + off)
820 if (!is_cow_mapping(vma->vm_flags))
825 if (is_zero_pfn(pfn))
828 if (unlikely(pfn > highest_memmap_pfn)) {
829 print_bad_pte(vma, addr, pte, NULL);
834 * NOTE! We still have PageReserved() pages in the page tables.
835 * eg. VDSO mappings can cause them to exist.
838 return pfn_to_page(pfn);
841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
842 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
845 unsigned long pfn = pmd_pfn(pmd);
848 * There is no pmd_special() but there may be special pmds, e.g.
849 * in a direct-access (dax) mapping, so let's just replicate the
850 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
852 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
853 if (vma->vm_flags & VM_MIXEDMAP) {
859 off = (addr - vma->vm_start) >> PAGE_SHIFT;
860 if (pfn == vma->vm_pgoff + off)
862 if (!is_cow_mapping(vma->vm_flags))
867 if (is_zero_pfn(pfn))
869 if (unlikely(pfn > highest_memmap_pfn))
873 * NOTE! We still have PageReserved() pages in the page tables.
874 * eg. VDSO mappings can cause them to exist.
877 return pfn_to_page(pfn);
882 * copy one vm_area from one task to the other. Assumes the page tables
883 * already present in the new task to be cleared in the whole range
884 * covered by this vma.
887 static inline unsigned long
888 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
889 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
890 unsigned long addr, int *rss)
892 unsigned long vm_flags = vma->vm_flags;
893 pte_t pte = *src_pte;
896 /* pte contains position in swap or file, so copy. */
897 if (unlikely(!pte_present(pte))) {
898 swp_entry_t entry = pte_to_swp_entry(pte);
900 if (likely(!non_swap_entry(entry))) {
901 if (swap_duplicate(entry) < 0)
904 /* make sure dst_mm is on swapoff's mmlist. */
905 if (unlikely(list_empty(&dst_mm->mmlist))) {
906 spin_lock(&mmlist_lock);
907 if (list_empty(&dst_mm->mmlist))
908 list_add(&dst_mm->mmlist,
910 spin_unlock(&mmlist_lock);
913 } else if (is_migration_entry(entry)) {
914 page = migration_entry_to_page(entry);
916 rss[mm_counter(page)]++;
918 if (is_write_migration_entry(entry) &&
919 is_cow_mapping(vm_flags)) {
921 * COW mappings require pages in both
922 * parent and child to be set to read.
924 make_migration_entry_read(&entry);
925 pte = swp_entry_to_pte(entry);
926 if (pte_swp_soft_dirty(*src_pte))
927 pte = pte_swp_mksoft_dirty(pte);
928 set_pte_at(src_mm, addr, src_pte, pte);
935 * If it's a COW mapping, write protect it both
936 * in the parent and the child
938 if (is_cow_mapping(vm_flags)) {
939 ptep_set_wrprotect(src_mm, addr, src_pte);
940 pte = pte_wrprotect(pte);
944 * If it's a shared mapping, mark it clean in
947 if (vm_flags & VM_SHARED)
948 pte = pte_mkclean(pte);
949 pte = pte_mkold(pte);
951 page = vm_normal_page(vma, addr, pte);
954 page_dup_rmap(page, false);
955 rss[mm_counter(page)]++;
959 set_pte_at(dst_mm, addr, dst_pte, pte);
963 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
964 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
965 unsigned long addr, unsigned long end)
967 pte_t *orig_src_pte, *orig_dst_pte;
968 pte_t *src_pte, *dst_pte;
969 spinlock_t *src_ptl, *dst_ptl;
971 int rss[NR_MM_COUNTERS];
972 swp_entry_t entry = (swp_entry_t){0};
977 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
980 src_pte = pte_offset_map(src_pmd, addr);
981 src_ptl = pte_lockptr(src_mm, src_pmd);
982 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
983 orig_src_pte = src_pte;
984 orig_dst_pte = dst_pte;
985 arch_enter_lazy_mmu_mode();
989 * We are holding two locks at this point - either of them
990 * could generate latencies in another task on another CPU.
992 if (progress >= 32) {
994 if (need_resched() ||
995 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
998 if (pte_none(*src_pte)) {
1002 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1007 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1009 arch_leave_lazy_mmu_mode();
1010 spin_unlock(src_ptl);
1011 pte_unmap(orig_src_pte);
1012 add_mm_rss_vec(dst_mm, rss);
1013 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1017 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1026 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1027 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1028 unsigned long addr, unsigned long end)
1030 pmd_t *src_pmd, *dst_pmd;
1033 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1036 src_pmd = pmd_offset(src_pud, addr);
1038 next = pmd_addr_end(addr, end);
1039 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1041 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1042 err = copy_huge_pmd(dst_mm, src_mm,
1043 dst_pmd, src_pmd, addr, vma);
1050 if (pmd_none_or_clear_bad(src_pmd))
1052 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1055 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1059 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1060 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1061 unsigned long addr, unsigned long end)
1063 pud_t *src_pud, *dst_pud;
1066 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1069 src_pud = pud_offset(src_p4d, addr);
1071 next = pud_addr_end(addr, end);
1072 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1075 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1076 err = copy_huge_pud(dst_mm, src_mm,
1077 dst_pud, src_pud, addr, vma);
1084 if (pud_none_or_clear_bad(src_pud))
1086 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1089 } while (dst_pud++, src_pud++, addr = next, addr != end);
1093 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1094 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1095 unsigned long addr, unsigned long end)
1097 p4d_t *src_p4d, *dst_p4d;
1100 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1103 src_p4d = p4d_offset(src_pgd, addr);
1105 next = p4d_addr_end(addr, end);
1106 if (p4d_none_or_clear_bad(src_p4d))
1108 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1111 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1115 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1116 struct vm_area_struct *vma)
1118 pgd_t *src_pgd, *dst_pgd;
1120 unsigned long addr = vma->vm_start;
1121 unsigned long end = vma->vm_end;
1122 unsigned long mmun_start; /* For mmu_notifiers */
1123 unsigned long mmun_end; /* For mmu_notifiers */
1128 * Don't copy ptes where a page fault will fill them correctly.
1129 * Fork becomes much lighter when there are big shared or private
1130 * readonly mappings. The tradeoff is that copy_page_range is more
1131 * efficient than faulting.
1133 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1137 if (is_vm_hugetlb_page(vma))
1138 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1140 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1142 * We do not free on error cases below as remove_vma
1143 * gets called on error from higher level routine
1145 ret = track_pfn_copy(vma);
1151 * We need to invalidate the secondary MMU mappings only when
1152 * there could be a permission downgrade on the ptes of the
1153 * parent mm. And a permission downgrade will only happen if
1154 * is_cow_mapping() returns true.
1156 is_cow = is_cow_mapping(vma->vm_flags);
1160 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1164 dst_pgd = pgd_offset(dst_mm, addr);
1165 src_pgd = pgd_offset(src_mm, addr);
1167 next = pgd_addr_end(addr, end);
1168 if (pgd_none_or_clear_bad(src_pgd))
1170 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1171 vma, addr, next))) {
1175 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1178 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1182 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1183 struct vm_area_struct *vma, pmd_t *pmd,
1184 unsigned long addr, unsigned long end,
1185 struct zap_details *details)
1187 struct mm_struct *mm = tlb->mm;
1188 int force_flush = 0;
1189 int rss[NR_MM_COUNTERS];
1195 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1198 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1200 flush_tlb_batched_pending(mm);
1201 arch_enter_lazy_mmu_mode();
1204 if (pte_none(ptent))
1207 if (pte_present(ptent)) {
1210 page = vm_normal_page(vma, addr, ptent);
1211 if (unlikely(details) && page) {
1213 * unmap_shared_mapping_pages() wants to
1214 * invalidate cache without truncating:
1215 * unmap shared but keep private pages.
1217 if (details->check_mapping &&
1218 details->check_mapping != page_rmapping(page))
1221 ptent = ptep_get_and_clear_full(mm, addr, pte,
1223 tlb_remove_tlb_entry(tlb, pte, addr);
1224 if (unlikely(!page))
1227 if (!PageAnon(page)) {
1228 if (pte_dirty(ptent)) {
1230 set_page_dirty(page);
1232 if (pte_young(ptent) &&
1233 likely(!(vma->vm_flags & VM_SEQ_READ)))
1234 mark_page_accessed(page);
1236 rss[mm_counter(page)]--;
1237 page_remove_rmap(page, false);
1238 if (unlikely(page_mapcount(page) < 0))
1239 print_bad_pte(vma, addr, ptent, page);
1240 if (unlikely(__tlb_remove_page(tlb, page))) {
1247 /* If details->check_mapping, we leave swap entries. */
1248 if (unlikely(details))
1251 entry = pte_to_swp_entry(ptent);
1252 if (!non_swap_entry(entry))
1254 else if (is_migration_entry(entry)) {
1257 page = migration_entry_to_page(entry);
1258 rss[mm_counter(page)]--;
1260 if (unlikely(!free_swap_and_cache(entry)))
1261 print_bad_pte(vma, addr, ptent, NULL);
1262 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1263 } while (pte++, addr += PAGE_SIZE, addr != end);
1265 add_mm_rss_vec(mm, rss);
1266 arch_leave_lazy_mmu_mode();
1268 /* Do the actual TLB flush before dropping ptl */
1270 tlb_flush_mmu_tlbonly(tlb);
1271 pte_unmap_unlock(start_pte, ptl);
1274 * If we forced a TLB flush (either due to running out of
1275 * batch buffers or because we needed to flush dirty TLB
1276 * entries before releasing the ptl), free the batched
1277 * memory too. Restart if we didn't do everything.
1281 tlb_flush_mmu_free(tlb);
1289 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1290 struct vm_area_struct *vma, pud_t *pud,
1291 unsigned long addr, unsigned long end,
1292 struct zap_details *details)
1297 pmd = pmd_offset(pud, addr);
1299 next = pmd_addr_end(addr, end);
1300 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1301 if (next - addr != HPAGE_PMD_SIZE) {
1302 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1303 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1304 __split_huge_pmd(vma, pmd, addr, false, NULL);
1305 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1310 * Here there can be other concurrent MADV_DONTNEED or
1311 * trans huge page faults running, and if the pmd is
1312 * none or trans huge it can change under us. This is
1313 * because MADV_DONTNEED holds the mmap_sem in read
1316 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1318 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1321 } while (pmd++, addr = next, addr != end);
1326 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1327 struct vm_area_struct *vma, p4d_t *p4d,
1328 unsigned long addr, unsigned long end,
1329 struct zap_details *details)
1334 pud = pud_offset(p4d, addr);
1336 next = pud_addr_end(addr, end);
1337 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1338 if (next - addr != HPAGE_PUD_SIZE) {
1339 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1340 split_huge_pud(vma, pud, addr);
1341 } else if (zap_huge_pud(tlb, vma, pud, addr))
1345 if (pud_none_or_clear_bad(pud))
1347 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1350 } while (pud++, addr = next, addr != end);
1355 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1356 struct vm_area_struct *vma, pgd_t *pgd,
1357 unsigned long addr, unsigned long end,
1358 struct zap_details *details)
1363 p4d = p4d_offset(pgd, addr);
1365 next = p4d_addr_end(addr, end);
1366 if (p4d_none_or_clear_bad(p4d))
1368 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1369 } while (p4d++, addr = next, addr != end);
1374 void unmap_page_range(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma,
1376 unsigned long addr, unsigned long end,
1377 struct zap_details *details)
1382 BUG_ON(addr >= end);
1383 tlb_start_vma(tlb, vma);
1384 pgd = pgd_offset(vma->vm_mm, addr);
1386 next = pgd_addr_end(addr, end);
1387 if (pgd_none_or_clear_bad(pgd))
1389 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1390 } while (pgd++, addr = next, addr != end);
1391 tlb_end_vma(tlb, vma);
1395 static void unmap_single_vma(struct mmu_gather *tlb,
1396 struct vm_area_struct *vma, unsigned long start_addr,
1397 unsigned long end_addr,
1398 struct zap_details *details)
1400 unsigned long start = max(vma->vm_start, start_addr);
1403 if (start >= vma->vm_end)
1405 end = min(vma->vm_end, end_addr);
1406 if (end <= vma->vm_start)
1410 uprobe_munmap(vma, start, end);
1412 if (unlikely(vma->vm_flags & VM_PFNMAP))
1413 untrack_pfn(vma, 0, 0);
1416 if (unlikely(is_vm_hugetlb_page(vma))) {
1418 * It is undesirable to test vma->vm_file as it
1419 * should be non-null for valid hugetlb area.
1420 * However, vm_file will be NULL in the error
1421 * cleanup path of mmap_region. When
1422 * hugetlbfs ->mmap method fails,
1423 * mmap_region() nullifies vma->vm_file
1424 * before calling this function to clean up.
1425 * Since no pte has actually been setup, it is
1426 * safe to do nothing in this case.
1429 i_mmap_lock_write(vma->vm_file->f_mapping);
1430 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1431 i_mmap_unlock_write(vma->vm_file->f_mapping);
1434 unmap_page_range(tlb, vma, start, end, details);
1439 * unmap_vmas - unmap a range of memory covered by a list of vma's
1440 * @tlb: address of the caller's struct mmu_gather
1441 * @vma: the starting vma
1442 * @start_addr: virtual address at which to start unmapping
1443 * @end_addr: virtual address at which to end unmapping
1445 * Unmap all pages in the vma list.
1447 * Only addresses between `start' and `end' will be unmapped.
1449 * The VMA list must be sorted in ascending virtual address order.
1451 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1452 * range after unmap_vmas() returns. So the only responsibility here is to
1453 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1454 * drops the lock and schedules.
1456 void unmap_vmas(struct mmu_gather *tlb,
1457 struct vm_area_struct *vma, unsigned long start_addr,
1458 unsigned long end_addr)
1460 struct mm_struct *mm = vma->vm_mm;
1462 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1463 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1464 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1465 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1469 * zap_page_range - remove user pages in a given range
1470 * @vma: vm_area_struct holding the applicable pages
1471 * @start: starting address of pages to zap
1472 * @size: number of bytes to zap
1474 * Caller must protect the VMA list
1476 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1479 struct mm_struct *mm = vma->vm_mm;
1480 struct mmu_gather tlb;
1481 unsigned long end = start + size;
1484 tlb_gather_mmu(&tlb, mm, start, end);
1485 update_hiwater_rss(mm);
1486 mmu_notifier_invalidate_range_start(mm, start, end);
1487 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1488 unmap_single_vma(&tlb, vma, start, end, NULL);
1489 mmu_notifier_invalidate_range_end(mm, start, end);
1490 tlb_finish_mmu(&tlb, start, end);
1494 * zap_page_range_single - remove user pages in a given range
1495 * @vma: vm_area_struct holding the applicable pages
1496 * @address: starting address of pages to zap
1497 * @size: number of bytes to zap
1498 * @details: details of shared cache invalidation
1500 * The range must fit into one VMA.
1502 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1503 unsigned long size, struct zap_details *details)
1505 struct mm_struct *mm = vma->vm_mm;
1506 struct mmu_gather tlb;
1507 unsigned long end = address + size;
1510 tlb_gather_mmu(&tlb, mm, address, end);
1511 update_hiwater_rss(mm);
1512 mmu_notifier_invalidate_range_start(mm, address, end);
1513 unmap_single_vma(&tlb, vma, address, end, details);
1514 mmu_notifier_invalidate_range_end(mm, address, end);
1515 tlb_finish_mmu(&tlb, address, end);
1519 * zap_vma_ptes - remove ptes mapping the vma
1520 * @vma: vm_area_struct holding ptes to be zapped
1521 * @address: starting address of pages to zap
1522 * @size: number of bytes to zap
1524 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1526 * The entire address range must be fully contained within the vma.
1528 * Returns 0 if successful.
1530 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1533 if (address < vma->vm_start || address + size > vma->vm_end ||
1534 !(vma->vm_flags & VM_PFNMAP))
1536 zap_page_range_single(vma, address, size, NULL);
1539 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1541 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1549 pgd = pgd_offset(mm, addr);
1550 p4d = p4d_alloc(mm, pgd, addr);
1553 pud = pud_alloc(mm, p4d, addr);
1556 pmd = pmd_alloc(mm, pud, addr);
1560 VM_BUG_ON(pmd_trans_huge(*pmd));
1561 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1565 * This is the old fallback for page remapping.
1567 * For historical reasons, it only allows reserved pages. Only
1568 * old drivers should use this, and they needed to mark their
1569 * pages reserved for the old functions anyway.
1571 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1572 struct page *page, pgprot_t prot)
1574 struct mm_struct *mm = vma->vm_mm;
1583 flush_dcache_page(page);
1584 pte = get_locked_pte(mm, addr, &ptl);
1588 if (!pte_none(*pte))
1591 /* Ok, finally just insert the thing.. */
1593 inc_mm_counter_fast(mm, mm_counter_file(page));
1594 page_add_file_rmap(page, false);
1595 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1598 pte_unmap_unlock(pte, ptl);
1601 pte_unmap_unlock(pte, ptl);
1607 * vm_insert_page - insert single page into user vma
1608 * @vma: user vma to map to
1609 * @addr: target user address of this page
1610 * @page: source kernel page
1612 * This allows drivers to insert individual pages they've allocated
1615 * The page has to be a nice clean _individual_ kernel allocation.
1616 * If you allocate a compound page, you need to have marked it as
1617 * such (__GFP_COMP), or manually just split the page up yourself
1618 * (see split_page()).
1620 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1621 * took an arbitrary page protection parameter. This doesn't allow
1622 * that. Your vma protection will have to be set up correctly, which
1623 * means that if you want a shared writable mapping, you'd better
1624 * ask for a shared writable mapping!
1626 * The page does not need to be reserved.
1628 * Usually this function is called from f_op->mmap() handler
1629 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1630 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1631 * function from other places, for example from page-fault handler.
1633 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1636 if (addr < vma->vm_start || addr >= vma->vm_end)
1638 if (!page_count(page))
1640 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1641 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1642 BUG_ON(vma->vm_flags & VM_PFNMAP);
1643 vma->vm_flags |= VM_MIXEDMAP;
1645 return insert_page(vma, addr, page, vma->vm_page_prot);
1647 EXPORT_SYMBOL(vm_insert_page);
1649 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1650 pfn_t pfn, pgprot_t prot)
1652 struct mm_struct *mm = vma->vm_mm;
1658 pte = get_locked_pte(mm, addr, &ptl);
1662 if (!pte_none(*pte))
1665 /* Ok, finally just insert the thing.. */
1666 if (pfn_t_devmap(pfn))
1667 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1669 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1670 set_pte_at(mm, addr, pte, entry);
1671 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1675 pte_unmap_unlock(pte, ptl);
1681 * vm_insert_pfn - insert single pfn into user vma
1682 * @vma: user vma to map to
1683 * @addr: target user address of this page
1684 * @pfn: source kernel pfn
1686 * Similar to vm_insert_page, this allows drivers to insert individual pages
1687 * they've allocated into a user vma. Same comments apply.
1689 * This function should only be called from a vm_ops->fault handler, and
1690 * in that case the handler should return NULL.
1692 * vma cannot be a COW mapping.
1694 * As this is called only for pages that do not currently exist, we
1695 * do not need to flush old virtual caches or the TLB.
1697 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1700 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1702 EXPORT_SYMBOL(vm_insert_pfn);
1705 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1706 * @vma: user vma to map to
1707 * @addr: target user address of this page
1708 * @pfn: source kernel pfn
1709 * @pgprot: pgprot flags for the inserted page
1711 * This is exactly like vm_insert_pfn, except that it allows drivers to
1712 * to override pgprot on a per-page basis.
1714 * This only makes sense for IO mappings, and it makes no sense for
1715 * cow mappings. In general, using multiple vmas is preferable;
1716 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1719 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1720 unsigned long pfn, pgprot_t pgprot)
1724 * Technically, architectures with pte_special can avoid all these
1725 * restrictions (same for remap_pfn_range). However we would like
1726 * consistency in testing and feature parity among all, so we should
1727 * try to keep these invariants in place for everybody.
1729 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1730 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1731 (VM_PFNMAP|VM_MIXEDMAP));
1732 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1733 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1735 if (addr < vma->vm_start || addr >= vma->vm_end)
1738 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1740 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1744 EXPORT_SYMBOL(vm_insert_pfn_prot);
1746 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1749 pgprot_t pgprot = vma->vm_page_prot;
1751 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1753 if (addr < vma->vm_start || addr >= vma->vm_end)
1756 track_pfn_insert(vma, &pgprot, pfn);
1759 * If we don't have pte special, then we have to use the pfn_valid()
1760 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1761 * refcount the page if pfn_valid is true (hence insert_page rather
1762 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1763 * without pte special, it would there be refcounted as a normal page.
1765 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1769 * At this point we are committed to insert_page()
1770 * regardless of whether the caller specified flags that
1771 * result in pfn_t_has_page() == false.
1773 page = pfn_to_page(pfn_t_to_pfn(pfn));
1774 return insert_page(vma, addr, page, pgprot);
1776 return insert_pfn(vma, addr, pfn, pgprot);
1778 EXPORT_SYMBOL(vm_insert_mixed);
1781 * maps a range of physical memory into the requested pages. the old
1782 * mappings are removed. any references to nonexistent pages results
1783 * in null mappings (currently treated as "copy-on-access")
1785 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1786 unsigned long addr, unsigned long end,
1787 unsigned long pfn, pgprot_t prot)
1792 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1795 arch_enter_lazy_mmu_mode();
1797 BUG_ON(!pte_none(*pte));
1798 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1800 } while (pte++, addr += PAGE_SIZE, addr != end);
1801 arch_leave_lazy_mmu_mode();
1802 pte_unmap_unlock(pte - 1, ptl);
1806 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1807 unsigned long addr, unsigned long end,
1808 unsigned long pfn, pgprot_t prot)
1813 pfn -= addr >> PAGE_SHIFT;
1814 pmd = pmd_alloc(mm, pud, addr);
1817 VM_BUG_ON(pmd_trans_huge(*pmd));
1819 next = pmd_addr_end(addr, end);
1820 if (remap_pte_range(mm, pmd, addr, next,
1821 pfn + (addr >> PAGE_SHIFT), prot))
1823 } while (pmd++, addr = next, addr != end);
1827 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1828 unsigned long addr, unsigned long end,
1829 unsigned long pfn, pgprot_t prot)
1834 pfn -= addr >> PAGE_SHIFT;
1835 pud = pud_alloc(mm, p4d, addr);
1839 next = pud_addr_end(addr, end);
1840 if (remap_pmd_range(mm, pud, addr, next,
1841 pfn + (addr >> PAGE_SHIFT), prot))
1843 } while (pud++, addr = next, addr != end);
1847 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1848 unsigned long addr, unsigned long end,
1849 unsigned long pfn, pgprot_t prot)
1854 pfn -= addr >> PAGE_SHIFT;
1855 p4d = p4d_alloc(mm, pgd, addr);
1859 next = p4d_addr_end(addr, end);
1860 if (remap_pud_range(mm, p4d, addr, next,
1861 pfn + (addr >> PAGE_SHIFT), prot))
1863 } while (p4d++, addr = next, addr != end);
1868 * remap_pfn_range - remap kernel memory to userspace
1869 * @vma: user vma to map to
1870 * @addr: target user address to start at
1871 * @pfn: physical address of kernel memory
1872 * @size: size of map area
1873 * @prot: page protection flags for this mapping
1875 * Note: this is only safe if the mm semaphore is held when called.
1877 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1878 unsigned long pfn, unsigned long size, pgprot_t prot)
1882 unsigned long end = addr + PAGE_ALIGN(size);
1883 struct mm_struct *mm = vma->vm_mm;
1884 unsigned long remap_pfn = pfn;
1888 * Physically remapped pages are special. Tell the
1889 * rest of the world about it:
1890 * VM_IO tells people not to look at these pages
1891 * (accesses can have side effects).
1892 * VM_PFNMAP tells the core MM that the base pages are just
1893 * raw PFN mappings, and do not have a "struct page" associated
1896 * Disable vma merging and expanding with mremap().
1898 * Omit vma from core dump, even when VM_IO turned off.
1900 * There's a horrible special case to handle copy-on-write
1901 * behaviour that some programs depend on. We mark the "original"
1902 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1903 * See vm_normal_page() for details.
1905 if (is_cow_mapping(vma->vm_flags)) {
1906 if (addr != vma->vm_start || end != vma->vm_end)
1908 vma->vm_pgoff = pfn;
1911 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1915 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1917 BUG_ON(addr >= end);
1918 pfn -= addr >> PAGE_SHIFT;
1919 pgd = pgd_offset(mm, addr);
1920 flush_cache_range(vma, addr, end);
1922 next = pgd_addr_end(addr, end);
1923 err = remap_p4d_range(mm, pgd, addr, next,
1924 pfn + (addr >> PAGE_SHIFT), prot);
1927 } while (pgd++, addr = next, addr != end);
1930 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1934 EXPORT_SYMBOL(remap_pfn_range);
1937 * vm_iomap_memory - remap memory to userspace
1938 * @vma: user vma to map to
1939 * @start: start of area
1940 * @len: size of area
1942 * This is a simplified io_remap_pfn_range() for common driver use. The
1943 * driver just needs to give us the physical memory range to be mapped,
1944 * we'll figure out the rest from the vma information.
1946 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1947 * whatever write-combining details or similar.
1949 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1951 unsigned long vm_len, pfn, pages;
1953 /* Check that the physical memory area passed in looks valid */
1954 if (start + len < start)
1957 * You *really* shouldn't map things that aren't page-aligned,
1958 * but we've historically allowed it because IO memory might
1959 * just have smaller alignment.
1961 len += start & ~PAGE_MASK;
1962 pfn = start >> PAGE_SHIFT;
1963 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1964 if (pfn + pages < pfn)
1967 /* We start the mapping 'vm_pgoff' pages into the area */
1968 if (vma->vm_pgoff > pages)
1970 pfn += vma->vm_pgoff;
1971 pages -= vma->vm_pgoff;
1973 /* Can we fit all of the mapping? */
1974 vm_len = vma->vm_end - vma->vm_start;
1975 if (vm_len >> PAGE_SHIFT > pages)
1978 /* Ok, let it rip */
1979 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1981 EXPORT_SYMBOL(vm_iomap_memory);
1983 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1984 unsigned long addr, unsigned long end,
1985 pte_fn_t fn, void *data)
1990 spinlock_t *uninitialized_var(ptl);
1992 pte = (mm == &init_mm) ?
1993 pte_alloc_kernel(pmd, addr) :
1994 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1998 BUG_ON(pmd_huge(*pmd));
2000 arch_enter_lazy_mmu_mode();
2002 token = pmd_pgtable(*pmd);
2005 err = fn(pte++, token, addr, data);
2008 } while (addr += PAGE_SIZE, addr != end);
2010 arch_leave_lazy_mmu_mode();
2013 pte_unmap_unlock(pte-1, ptl);
2017 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2018 unsigned long addr, unsigned long end,
2019 pte_fn_t fn, void *data)
2025 BUG_ON(pud_huge(*pud));
2027 pmd = pmd_alloc(mm, pud, addr);
2031 next = pmd_addr_end(addr, end);
2032 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2035 } while (pmd++, addr = next, addr != end);
2039 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2040 unsigned long addr, unsigned long end,
2041 pte_fn_t fn, void *data)
2047 pud = pud_alloc(mm, p4d, addr);
2051 next = pud_addr_end(addr, end);
2052 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2055 } while (pud++, addr = next, addr != end);
2059 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2060 unsigned long addr, unsigned long end,
2061 pte_fn_t fn, void *data)
2067 p4d = p4d_alloc(mm, pgd, addr);
2071 next = p4d_addr_end(addr, end);
2072 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2075 } while (p4d++, addr = next, addr != end);
2080 * Scan a region of virtual memory, filling in page tables as necessary
2081 * and calling a provided function on each leaf page table.
2083 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2084 unsigned long size, pte_fn_t fn, void *data)
2088 unsigned long end = addr + size;
2091 if (WARN_ON(addr >= end))
2094 pgd = pgd_offset(mm, addr);
2096 next = pgd_addr_end(addr, end);
2097 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2100 } while (pgd++, addr = next, addr != end);
2104 EXPORT_SYMBOL_GPL(apply_to_page_range);
2107 * handle_pte_fault chooses page fault handler according to an entry which was
2108 * read non-atomically. Before making any commitment, on those architectures
2109 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2110 * parts, do_swap_page must check under lock before unmapping the pte and
2111 * proceeding (but do_wp_page is only called after already making such a check;
2112 * and do_anonymous_page can safely check later on).
2114 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2115 pte_t *page_table, pte_t orig_pte)
2118 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2119 if (sizeof(pte_t) > sizeof(unsigned long)) {
2120 spinlock_t *ptl = pte_lockptr(mm, pmd);
2122 same = pte_same(*page_table, orig_pte);
2126 pte_unmap(page_table);
2130 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2132 debug_dma_assert_idle(src);
2135 * If the source page was a PFN mapping, we don't have
2136 * a "struct page" for it. We do a best-effort copy by
2137 * just copying from the original user address. If that
2138 * fails, we just zero-fill it. Live with it.
2140 if (unlikely(!src)) {
2141 void *kaddr = kmap_atomic(dst);
2142 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2145 * This really shouldn't fail, because the page is there
2146 * in the page tables. But it might just be unreadable,
2147 * in which case we just give up and fill the result with
2150 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2152 kunmap_atomic(kaddr);
2153 flush_dcache_page(dst);
2155 copy_user_highpage(dst, src, va, vma);
2158 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2160 struct file *vm_file = vma->vm_file;
2163 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2166 * Special mappings (e.g. VDSO) do not have any file so fake
2167 * a default GFP_KERNEL for them.
2173 * Notify the address space that the page is about to become writable so that
2174 * it can prohibit this or wait for the page to get into an appropriate state.
2176 * We do this without the lock held, so that it can sleep if it needs to.
2178 static int do_page_mkwrite(struct vm_fault *vmf)
2181 struct page *page = vmf->page;
2182 unsigned int old_flags = vmf->flags;
2184 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2186 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2187 /* Restore original flags so that caller is not surprised */
2188 vmf->flags = old_flags;
2189 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2191 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2193 if (!page->mapping) {
2195 return 0; /* retry */
2197 ret |= VM_FAULT_LOCKED;
2199 VM_BUG_ON_PAGE(!PageLocked(page), page);
2204 * Handle dirtying of a page in shared file mapping on a write fault.
2206 * The function expects the page to be locked and unlocks it.
2208 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2211 struct address_space *mapping;
2213 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2215 dirtied = set_page_dirty(page);
2216 VM_BUG_ON_PAGE(PageAnon(page), page);
2218 * Take a local copy of the address_space - page.mapping may be zeroed
2219 * by truncate after unlock_page(). The address_space itself remains
2220 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2221 * release semantics to prevent the compiler from undoing this copying.
2223 mapping = page_rmapping(page);
2226 if ((dirtied || page_mkwrite) && mapping) {
2228 * Some device drivers do not set page.mapping
2229 * but still dirty their pages
2231 balance_dirty_pages_ratelimited(mapping);
2235 file_update_time(vma->vm_file);
2239 * Handle write page faults for pages that can be reused in the current vma
2241 * This can happen either due to the mapping being with the VM_SHARED flag,
2242 * or due to us being the last reference standing to the page. In either
2243 * case, all we need to do here is to mark the page as writable and update
2244 * any related book-keeping.
2246 static inline void wp_page_reuse(struct vm_fault *vmf)
2247 __releases(vmf->ptl)
2249 struct vm_area_struct *vma = vmf->vma;
2250 struct page *page = vmf->page;
2253 * Clear the pages cpupid information as the existing
2254 * information potentially belongs to a now completely
2255 * unrelated process.
2258 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2260 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2261 entry = pte_mkyoung(vmf->orig_pte);
2262 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2263 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2264 update_mmu_cache(vma, vmf->address, vmf->pte);
2265 pte_unmap_unlock(vmf->pte, vmf->ptl);
2269 * Handle the case of a page which we actually need to copy to a new page.
2271 * Called with mmap_sem locked and the old page referenced, but
2272 * without the ptl held.
2274 * High level logic flow:
2276 * - Allocate a page, copy the content of the old page to the new one.
2277 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2278 * - Take the PTL. If the pte changed, bail out and release the allocated page
2279 * - If the pte is still the way we remember it, update the page table and all
2280 * relevant references. This includes dropping the reference the page-table
2281 * held to the old page, as well as updating the rmap.
2282 * - In any case, unlock the PTL and drop the reference we took to the old page.
2284 static int wp_page_copy(struct vm_fault *vmf)
2286 struct vm_area_struct *vma = vmf->vma;
2287 struct mm_struct *mm = vma->vm_mm;
2288 struct page *old_page = vmf->page;
2289 struct page *new_page = NULL;
2291 int page_copied = 0;
2292 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2293 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2294 struct mem_cgroup *memcg;
2296 if (unlikely(anon_vma_prepare(vma)))
2299 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2300 new_page = alloc_zeroed_user_highpage_movable(vma,
2305 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2309 cow_user_page(new_page, old_page, vmf->address, vma);
2312 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2315 __SetPageUptodate(new_page);
2317 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2320 * Re-check the pte - we dropped the lock
2322 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2323 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2325 if (!PageAnon(old_page)) {
2326 dec_mm_counter_fast(mm,
2327 mm_counter_file(old_page));
2328 inc_mm_counter_fast(mm, MM_ANONPAGES);
2331 inc_mm_counter_fast(mm, MM_ANONPAGES);
2333 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2334 entry = mk_pte(new_page, vma->vm_page_prot);
2335 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2337 * Clear the pte entry and flush it first, before updating the
2338 * pte with the new entry. This will avoid a race condition
2339 * seen in the presence of one thread doing SMC and another
2342 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2343 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2344 mem_cgroup_commit_charge(new_page, memcg, false, false);
2345 lru_cache_add_active_or_unevictable(new_page, vma);
2347 * We call the notify macro here because, when using secondary
2348 * mmu page tables (such as kvm shadow page tables), we want the
2349 * new page to be mapped directly into the secondary page table.
2351 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2352 update_mmu_cache(vma, vmf->address, vmf->pte);
2355 * Only after switching the pte to the new page may
2356 * we remove the mapcount here. Otherwise another
2357 * process may come and find the rmap count decremented
2358 * before the pte is switched to the new page, and
2359 * "reuse" the old page writing into it while our pte
2360 * here still points into it and can be read by other
2363 * The critical issue is to order this
2364 * page_remove_rmap with the ptp_clear_flush above.
2365 * Those stores are ordered by (if nothing else,)
2366 * the barrier present in the atomic_add_negative
2367 * in page_remove_rmap.
2369 * Then the TLB flush in ptep_clear_flush ensures that
2370 * no process can access the old page before the
2371 * decremented mapcount is visible. And the old page
2372 * cannot be reused until after the decremented
2373 * mapcount is visible. So transitively, TLBs to
2374 * old page will be flushed before it can be reused.
2376 page_remove_rmap(old_page, false);
2379 /* Free the old page.. */
2380 new_page = old_page;
2383 mem_cgroup_cancel_charge(new_page, memcg, false);
2389 pte_unmap_unlock(vmf->pte, vmf->ptl);
2390 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2393 * Don't let another task, with possibly unlocked vma,
2394 * keep the mlocked page.
2396 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2397 lock_page(old_page); /* LRU manipulation */
2398 if (PageMlocked(old_page))
2399 munlock_vma_page(old_page);
2400 unlock_page(old_page);
2404 return page_copied ? VM_FAULT_WRITE : 0;
2410 return VM_FAULT_OOM;
2414 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2415 * writeable once the page is prepared
2417 * @vmf: structure describing the fault
2419 * This function handles all that is needed to finish a write page fault in a
2420 * shared mapping due to PTE being read-only once the mapped page is prepared.
2421 * It handles locking of PTE and modifying it. The function returns
2422 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2425 * The function expects the page to be locked or other protection against
2426 * concurrent faults / writeback (such as DAX radix tree locks).
2428 int finish_mkwrite_fault(struct vm_fault *vmf)
2430 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2431 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2434 * We might have raced with another page fault while we released the
2435 * pte_offset_map_lock.
2437 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2438 pte_unmap_unlock(vmf->pte, vmf->ptl);
2439 return VM_FAULT_NOPAGE;
2446 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2449 static int wp_pfn_shared(struct vm_fault *vmf)
2451 struct vm_area_struct *vma = vmf->vma;
2453 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2456 pte_unmap_unlock(vmf->pte, vmf->ptl);
2457 vmf->flags |= FAULT_FLAG_MKWRITE;
2458 ret = vma->vm_ops->pfn_mkwrite(vmf);
2459 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2461 return finish_mkwrite_fault(vmf);
2464 return VM_FAULT_WRITE;
2467 static int wp_page_shared(struct vm_fault *vmf)
2468 __releases(vmf->ptl)
2470 struct vm_area_struct *vma = vmf->vma;
2472 get_page(vmf->page);
2474 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2477 pte_unmap_unlock(vmf->pte, vmf->ptl);
2478 tmp = do_page_mkwrite(vmf);
2479 if (unlikely(!tmp || (tmp &
2480 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2481 put_page(vmf->page);
2484 tmp = finish_mkwrite_fault(vmf);
2485 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2486 unlock_page(vmf->page);
2487 put_page(vmf->page);
2492 lock_page(vmf->page);
2494 fault_dirty_shared_page(vma, vmf->page);
2495 put_page(vmf->page);
2497 return VM_FAULT_WRITE;
2501 * This routine handles present pages, when users try to write
2502 * to a shared page. It is done by copying the page to a new address
2503 * and decrementing the shared-page counter for the old page.
2505 * Note that this routine assumes that the protection checks have been
2506 * done by the caller (the low-level page fault routine in most cases).
2507 * Thus we can safely just mark it writable once we've done any necessary
2510 * We also mark the page dirty at this point even though the page will
2511 * change only once the write actually happens. This avoids a few races,
2512 * and potentially makes it more efficient.
2514 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2515 * but allow concurrent faults), with pte both mapped and locked.
2516 * We return with mmap_sem still held, but pte unmapped and unlocked.
2518 static int do_wp_page(struct vm_fault *vmf)
2519 __releases(vmf->ptl)
2521 struct vm_area_struct *vma = vmf->vma;
2523 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2526 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2529 * We should not cow pages in a shared writeable mapping.
2530 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2532 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2533 (VM_WRITE|VM_SHARED))
2534 return wp_pfn_shared(vmf);
2536 pte_unmap_unlock(vmf->pte, vmf->ptl);
2537 return wp_page_copy(vmf);
2541 * Take out anonymous pages first, anonymous shared vmas are
2542 * not dirty accountable.
2544 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2546 if (!trylock_page(vmf->page)) {
2547 get_page(vmf->page);
2548 pte_unmap_unlock(vmf->pte, vmf->ptl);
2549 lock_page(vmf->page);
2550 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2551 vmf->address, &vmf->ptl);
2552 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2553 unlock_page(vmf->page);
2554 pte_unmap_unlock(vmf->pte, vmf->ptl);
2555 put_page(vmf->page);
2558 put_page(vmf->page);
2560 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2561 if (total_mapcount == 1) {
2563 * The page is all ours. Move it to
2564 * our anon_vma so the rmap code will
2565 * not search our parent or siblings.
2566 * Protected against the rmap code by
2569 page_move_anon_rmap(vmf->page, vma);
2571 unlock_page(vmf->page);
2573 return VM_FAULT_WRITE;
2575 unlock_page(vmf->page);
2576 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2577 (VM_WRITE|VM_SHARED))) {
2578 return wp_page_shared(vmf);
2582 * Ok, we need to copy. Oh, well..
2584 get_page(vmf->page);
2586 pte_unmap_unlock(vmf->pte, vmf->ptl);
2587 return wp_page_copy(vmf);
2590 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2591 unsigned long start_addr, unsigned long end_addr,
2592 struct zap_details *details)
2594 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2597 static inline void unmap_mapping_range_tree(struct rb_root *root,
2598 struct zap_details *details)
2600 struct vm_area_struct *vma;
2601 pgoff_t vba, vea, zba, zea;
2603 vma_interval_tree_foreach(vma, root,
2604 details->first_index, details->last_index) {
2606 vba = vma->vm_pgoff;
2607 vea = vba + vma_pages(vma) - 1;
2608 zba = details->first_index;
2611 zea = details->last_index;
2615 unmap_mapping_range_vma(vma,
2616 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2617 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2623 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2624 * address_space corresponding to the specified page range in the underlying
2627 * @mapping: the address space containing mmaps to be unmapped.
2628 * @holebegin: byte in first page to unmap, relative to the start of
2629 * the underlying file. This will be rounded down to a PAGE_SIZE
2630 * boundary. Note that this is different from truncate_pagecache(), which
2631 * must keep the partial page. In contrast, we must get rid of
2633 * @holelen: size of prospective hole in bytes. This will be rounded
2634 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2636 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2637 * but 0 when invalidating pagecache, don't throw away private data.
2639 void unmap_mapping_range(struct address_space *mapping,
2640 loff_t const holebegin, loff_t const holelen, int even_cows)
2642 struct zap_details details = { };
2643 pgoff_t hba = holebegin >> PAGE_SHIFT;
2644 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2646 /* Check for overflow. */
2647 if (sizeof(holelen) > sizeof(hlen)) {
2649 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2650 if (holeend & ~(long long)ULONG_MAX)
2651 hlen = ULONG_MAX - hba + 1;
2654 details.check_mapping = even_cows ? NULL : mapping;
2655 details.first_index = hba;
2656 details.last_index = hba + hlen - 1;
2657 if (details.last_index < details.first_index)
2658 details.last_index = ULONG_MAX;
2660 i_mmap_lock_write(mapping);
2661 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2662 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2663 i_mmap_unlock_write(mapping);
2665 EXPORT_SYMBOL(unmap_mapping_range);
2668 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2669 * but allow concurrent faults), and pte mapped but not yet locked.
2670 * We return with pte unmapped and unlocked.
2672 * We return with the mmap_sem locked or unlocked in the same cases
2673 * as does filemap_fault().
2675 int do_swap_page(struct vm_fault *vmf)
2677 struct vm_area_struct *vma = vmf->vma;
2678 struct page *page, *swapcache;
2679 struct mem_cgroup *memcg;
2686 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2689 entry = pte_to_swp_entry(vmf->orig_pte);
2690 if (unlikely(non_swap_entry(entry))) {
2691 if (is_migration_entry(entry)) {
2692 migration_entry_wait(vma->vm_mm, vmf->pmd,
2694 } else if (is_hwpoison_entry(entry)) {
2695 ret = VM_FAULT_HWPOISON;
2697 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2698 ret = VM_FAULT_SIGBUS;
2702 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2703 page = lookup_swap_cache(entry);
2705 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2709 * Back out if somebody else faulted in this pte
2710 * while we released the pte lock.
2712 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2713 vmf->address, &vmf->ptl);
2714 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2716 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2720 /* Had to read the page from swap area: Major fault */
2721 ret = VM_FAULT_MAJOR;
2722 count_vm_event(PGMAJFAULT);
2723 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2724 } else if (PageHWPoison(page)) {
2726 * hwpoisoned dirty swapcache pages are kept for killing
2727 * owner processes (which may be unknown at hwpoison time)
2729 ret = VM_FAULT_HWPOISON;
2730 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2736 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2738 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2740 ret |= VM_FAULT_RETRY;
2745 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2746 * release the swapcache from under us. The page pin, and pte_same
2747 * test below, are not enough to exclude that. Even if it is still
2748 * swapcache, we need to check that the page's swap has not changed.
2750 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2753 page = ksm_might_need_to_copy(page, vma, vmf->address);
2754 if (unlikely(!page)) {
2760 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2767 * Back out if somebody else already faulted in this pte.
2769 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2771 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2774 if (unlikely(!PageUptodate(page))) {
2775 ret = VM_FAULT_SIGBUS;
2780 * The page isn't present yet, go ahead with the fault.
2782 * Be careful about the sequence of operations here.
2783 * To get its accounting right, reuse_swap_page() must be called
2784 * while the page is counted on swap but not yet in mapcount i.e.
2785 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2786 * must be called after the swap_free(), or it will never succeed.
2789 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2790 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2791 pte = mk_pte(page, vma->vm_page_prot);
2792 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2793 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2794 vmf->flags &= ~FAULT_FLAG_WRITE;
2795 ret |= VM_FAULT_WRITE;
2796 exclusive = RMAP_EXCLUSIVE;
2798 flush_icache_page(vma, page);
2799 if (pte_swp_soft_dirty(vmf->orig_pte))
2800 pte = pte_mksoft_dirty(pte);
2801 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2802 vmf->orig_pte = pte;
2803 if (page == swapcache) {
2804 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2805 mem_cgroup_commit_charge(page, memcg, true, false);
2806 activate_page(page);
2807 } else { /* ksm created a completely new copy */
2808 page_add_new_anon_rmap(page, vma, vmf->address, false);
2809 mem_cgroup_commit_charge(page, memcg, false, false);
2810 lru_cache_add_active_or_unevictable(page, vma);
2814 if (mem_cgroup_swap_full(page) ||
2815 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2816 try_to_free_swap(page);
2818 if (page != swapcache) {
2820 * Hold the lock to avoid the swap entry to be reused
2821 * until we take the PT lock for the pte_same() check
2822 * (to avoid false positives from pte_same). For
2823 * further safety release the lock after the swap_free
2824 * so that the swap count won't change under a
2825 * parallel locked swapcache.
2827 unlock_page(swapcache);
2828 put_page(swapcache);
2831 if (vmf->flags & FAULT_FLAG_WRITE) {
2832 ret |= do_wp_page(vmf);
2833 if (ret & VM_FAULT_ERROR)
2834 ret &= VM_FAULT_ERROR;
2838 /* No need to invalidate - it was non-present before */
2839 update_mmu_cache(vma, vmf->address, vmf->pte);
2841 pte_unmap_unlock(vmf->pte, vmf->ptl);
2845 mem_cgroup_cancel_charge(page, memcg, false);
2846 pte_unmap_unlock(vmf->pte, vmf->ptl);
2851 if (page != swapcache) {
2852 unlock_page(swapcache);
2853 put_page(swapcache);
2859 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2860 * but allow concurrent faults), and pte mapped but not yet locked.
2861 * We return with mmap_sem still held, but pte unmapped and unlocked.
2863 static int do_anonymous_page(struct vm_fault *vmf)
2865 struct vm_area_struct *vma = vmf->vma;
2866 struct mem_cgroup *memcg;
2870 /* File mapping without ->vm_ops ? */
2871 if (vma->vm_flags & VM_SHARED)
2872 return VM_FAULT_SIGBUS;
2875 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2876 * pte_offset_map() on pmds where a huge pmd might be created
2877 * from a different thread.
2879 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2880 * parallel threads are excluded by other means.
2882 * Here we only have down_read(mmap_sem).
2884 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2885 return VM_FAULT_OOM;
2887 /* See the comment in pte_alloc_one_map() */
2888 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2891 /* Use the zero-page for reads */
2892 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2893 !mm_forbids_zeropage(vma->vm_mm)) {
2894 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2895 vma->vm_page_prot));
2896 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2897 vmf->address, &vmf->ptl);
2898 if (!pte_none(*vmf->pte))
2900 /* Deliver the page fault to userland, check inside PT lock */
2901 if (userfaultfd_missing(vma)) {
2902 pte_unmap_unlock(vmf->pte, vmf->ptl);
2903 return handle_userfault(vmf, VM_UFFD_MISSING);
2908 /* Allocate our own private page. */
2909 if (unlikely(anon_vma_prepare(vma)))
2911 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2915 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2919 * The memory barrier inside __SetPageUptodate makes sure that
2920 * preceeding stores to the page contents become visible before
2921 * the set_pte_at() write.
2923 __SetPageUptodate(page);
2925 entry = mk_pte(page, vma->vm_page_prot);
2926 if (vma->vm_flags & VM_WRITE)
2927 entry = pte_mkwrite(pte_mkdirty(entry));
2929 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2931 if (!pte_none(*vmf->pte))
2934 /* Deliver the page fault to userland, check inside PT lock */
2935 if (userfaultfd_missing(vma)) {
2936 pte_unmap_unlock(vmf->pte, vmf->ptl);
2937 mem_cgroup_cancel_charge(page, memcg, false);
2939 return handle_userfault(vmf, VM_UFFD_MISSING);
2942 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2943 page_add_new_anon_rmap(page, vma, vmf->address, false);
2944 mem_cgroup_commit_charge(page, memcg, false, false);
2945 lru_cache_add_active_or_unevictable(page, vma);
2947 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2949 /* No need to invalidate - it was non-present before */
2950 update_mmu_cache(vma, vmf->address, vmf->pte);
2952 pte_unmap_unlock(vmf->pte, vmf->ptl);
2955 mem_cgroup_cancel_charge(page, memcg, false);
2961 return VM_FAULT_OOM;
2965 * The mmap_sem must have been held on entry, and may have been
2966 * released depending on flags and vma->vm_ops->fault() return value.
2967 * See filemap_fault() and __lock_page_retry().
2969 static int __do_fault(struct vm_fault *vmf)
2971 struct vm_area_struct *vma = vmf->vma;
2974 ret = vma->vm_ops->fault(vmf);
2975 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2976 VM_FAULT_DONE_COW)))
2979 if (unlikely(PageHWPoison(vmf->page))) {
2980 if (ret & VM_FAULT_LOCKED)
2981 unlock_page(vmf->page);
2982 put_page(vmf->page);
2984 return VM_FAULT_HWPOISON;
2987 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2988 lock_page(vmf->page);
2990 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2996 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
2997 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
2998 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
2999 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3001 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3003 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3006 static int pte_alloc_one_map(struct vm_fault *vmf)
3008 struct vm_area_struct *vma = vmf->vma;
3010 if (!pmd_none(*vmf->pmd))
3012 if (vmf->prealloc_pte) {
3013 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3014 if (unlikely(!pmd_none(*vmf->pmd))) {
3015 spin_unlock(vmf->ptl);
3019 atomic_long_inc(&vma->vm_mm->nr_ptes);
3020 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3021 spin_unlock(vmf->ptl);
3022 vmf->prealloc_pte = NULL;
3023 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3024 return VM_FAULT_OOM;
3028 * If a huge pmd materialized under us just retry later. Use
3029 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3030 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3031 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3032 * running immediately after a huge pmd fault in a different thread of
3033 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3034 * All we have to ensure is that it is a regular pmd that we can walk
3035 * with pte_offset_map() and we can do that through an atomic read in
3036 * C, which is what pmd_trans_unstable() provides.
3038 if (pmd_devmap_trans_unstable(vmf->pmd))
3039 return VM_FAULT_NOPAGE;
3042 * At this point we know that our vmf->pmd points to a page of ptes
3043 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3044 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3045 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3046 * be valid and we will re-check to make sure the vmf->pte isn't
3047 * pte_none() under vmf->ptl protection when we return to
3050 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3055 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3057 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3058 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3059 unsigned long haddr)
3061 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3062 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3064 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3069 static void deposit_prealloc_pte(struct vm_fault *vmf)
3071 struct vm_area_struct *vma = vmf->vma;
3073 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3075 * We are going to consume the prealloc table,
3076 * count that as nr_ptes.
3078 atomic_long_inc(&vma->vm_mm->nr_ptes);
3079 vmf->prealloc_pte = NULL;
3082 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3084 struct vm_area_struct *vma = vmf->vma;
3085 bool write = vmf->flags & FAULT_FLAG_WRITE;
3086 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3090 if (!transhuge_vma_suitable(vma, haddr))
3091 return VM_FAULT_FALLBACK;
3093 ret = VM_FAULT_FALLBACK;
3094 page = compound_head(page);
3097 * Archs like ppc64 need additonal space to store information
3098 * related to pte entry. Use the preallocated table for that.
3100 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3101 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3102 if (!vmf->prealloc_pte)
3103 return VM_FAULT_OOM;
3104 smp_wmb(); /* See comment in __pte_alloc() */
3107 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3108 if (unlikely(!pmd_none(*vmf->pmd)))
3111 for (i = 0; i < HPAGE_PMD_NR; i++)
3112 flush_icache_page(vma, page + i);
3114 entry = mk_huge_pmd(page, vma->vm_page_prot);
3116 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3118 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3119 page_add_file_rmap(page, true);
3121 * deposit and withdraw with pmd lock held
3123 if (arch_needs_pgtable_deposit())
3124 deposit_prealloc_pte(vmf);
3126 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3128 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3130 /* fault is handled */
3132 count_vm_event(THP_FILE_MAPPED);
3134 spin_unlock(vmf->ptl);
3138 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3146 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3147 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3149 * @vmf: fault environment
3150 * @memcg: memcg to charge page (only for private mappings)
3151 * @page: page to map
3153 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3156 * Target users are page handler itself and implementations of
3157 * vm_ops->map_pages.
3159 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3162 struct vm_area_struct *vma = vmf->vma;
3163 bool write = vmf->flags & FAULT_FLAG_WRITE;
3167 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3168 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3170 VM_BUG_ON_PAGE(memcg, page);
3172 ret = do_set_pmd(vmf, page);
3173 if (ret != VM_FAULT_FALLBACK)
3178 ret = pte_alloc_one_map(vmf);
3183 /* Re-check under ptl */
3184 if (unlikely(!pte_none(*vmf->pte)))
3185 return VM_FAULT_NOPAGE;
3187 flush_icache_page(vma, page);
3188 entry = mk_pte(page, vma->vm_page_prot);
3190 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3191 /* copy-on-write page */
3192 if (write && !(vma->vm_flags & VM_SHARED)) {
3193 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3194 page_add_new_anon_rmap(page, vma, vmf->address, false);
3195 mem_cgroup_commit_charge(page, memcg, false, false);
3196 lru_cache_add_active_or_unevictable(page, vma);
3198 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3199 page_add_file_rmap(page, false);
3201 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3203 /* no need to invalidate: a not-present page won't be cached */
3204 update_mmu_cache(vma, vmf->address, vmf->pte);
3211 * finish_fault - finish page fault once we have prepared the page to fault
3213 * @vmf: structure describing the fault
3215 * This function handles all that is needed to finish a page fault once the
3216 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3217 * given page, adds reverse page mapping, handles memcg charges and LRU
3218 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3221 * The function expects the page to be locked and on success it consumes a
3222 * reference of a page being mapped (for the PTE which maps it).
3224 int finish_fault(struct vm_fault *vmf)
3229 /* Did we COW the page? */
3230 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3231 !(vmf->vma->vm_flags & VM_SHARED))
3232 page = vmf->cow_page;
3235 ret = alloc_set_pte(vmf, vmf->memcg, page);
3237 pte_unmap_unlock(vmf->pte, vmf->ptl);
3241 static unsigned long fault_around_bytes __read_mostly =
3242 rounddown_pow_of_two(65536);
3244 #ifdef CONFIG_DEBUG_FS
3245 static int fault_around_bytes_get(void *data, u64 *val)
3247 *val = fault_around_bytes;
3252 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3253 * rounded down to nearest page order. It's what do_fault_around() expects to
3256 static int fault_around_bytes_set(void *data, u64 val)
3258 if (val / PAGE_SIZE > PTRS_PER_PTE)
3260 if (val > PAGE_SIZE)
3261 fault_around_bytes = rounddown_pow_of_two(val);
3263 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3266 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3267 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3269 static int __init fault_around_debugfs(void)
3273 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3274 &fault_around_bytes_fops);
3276 pr_warn("Failed to create fault_around_bytes in debugfs");
3279 late_initcall(fault_around_debugfs);
3283 * do_fault_around() tries to map few pages around the fault address. The hope
3284 * is that the pages will be needed soon and this will lower the number of
3287 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3288 * not ready to be mapped: not up-to-date, locked, etc.
3290 * This function is called with the page table lock taken. In the split ptlock
3291 * case the page table lock only protects only those entries which belong to
3292 * the page table corresponding to the fault address.
3294 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3297 * fault_around_pages() defines how many pages we'll try to map.
3298 * do_fault_around() expects it to return a power of two less than or equal to
3301 * The virtual address of the area that we map is naturally aligned to the
3302 * fault_around_pages() value (and therefore to page order). This way it's
3303 * easier to guarantee that we don't cross page table boundaries.
3305 static int do_fault_around(struct vm_fault *vmf)
3307 unsigned long address = vmf->address, nr_pages, mask;
3308 pgoff_t start_pgoff = vmf->pgoff;
3312 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3313 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3315 vmf->address = max(address & mask, vmf->vma->vm_start);
3316 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3320 * end_pgoff is either end of page table or end of vma
3321 * or fault_around_pages() from start_pgoff, depending what is nearest.
3323 end_pgoff = start_pgoff -
3324 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3326 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3327 start_pgoff + nr_pages - 1);
3329 if (pmd_none(*vmf->pmd)) {
3330 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3332 if (!vmf->prealloc_pte)
3334 smp_wmb(); /* See comment in __pte_alloc() */
3337 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3339 /* Huge page is mapped? Page fault is solved */
3340 if (pmd_trans_huge(*vmf->pmd)) {
3341 ret = VM_FAULT_NOPAGE;
3345 /* ->map_pages() haven't done anything useful. Cold page cache? */
3349 /* check if the page fault is solved */
3350 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3351 if (!pte_none(*vmf->pte))
3352 ret = VM_FAULT_NOPAGE;
3353 pte_unmap_unlock(vmf->pte, vmf->ptl);
3355 vmf->address = address;
3360 static int do_read_fault(struct vm_fault *vmf)
3362 struct vm_area_struct *vma = vmf->vma;
3366 * Let's call ->map_pages() first and use ->fault() as fallback
3367 * if page by the offset is not ready to be mapped (cold cache or
3370 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3371 ret = do_fault_around(vmf);
3376 ret = __do_fault(vmf);
3377 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3380 ret |= finish_fault(vmf);
3381 unlock_page(vmf->page);
3382 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3383 put_page(vmf->page);
3387 static int do_cow_fault(struct vm_fault *vmf)
3389 struct vm_area_struct *vma = vmf->vma;
3392 if (unlikely(anon_vma_prepare(vma)))
3393 return VM_FAULT_OOM;
3395 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3397 return VM_FAULT_OOM;
3399 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3400 &vmf->memcg, false)) {
3401 put_page(vmf->cow_page);
3402 return VM_FAULT_OOM;
3405 ret = __do_fault(vmf);
3406 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3408 if (ret & VM_FAULT_DONE_COW)
3411 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3412 __SetPageUptodate(vmf->cow_page);
3414 ret |= finish_fault(vmf);
3415 unlock_page(vmf->page);
3416 put_page(vmf->page);
3417 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3421 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3422 put_page(vmf->cow_page);
3426 static int do_shared_fault(struct vm_fault *vmf)
3428 struct vm_area_struct *vma = vmf->vma;
3431 ret = __do_fault(vmf);
3432 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3436 * Check if the backing address space wants to know that the page is
3437 * about to become writable
3439 if (vma->vm_ops->page_mkwrite) {
3440 unlock_page(vmf->page);
3441 tmp = do_page_mkwrite(vmf);
3442 if (unlikely(!tmp ||
3443 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3444 put_page(vmf->page);
3449 ret |= finish_fault(vmf);
3450 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3452 unlock_page(vmf->page);
3453 put_page(vmf->page);
3457 fault_dirty_shared_page(vma, vmf->page);
3462 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3463 * but allow concurrent faults).
3464 * The mmap_sem may have been released depending on flags and our
3465 * return value. See filemap_fault() and __lock_page_or_retry().
3467 static int do_fault(struct vm_fault *vmf)
3469 struct vm_area_struct *vma = vmf->vma;
3472 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3473 if (!vma->vm_ops->fault)
3474 ret = VM_FAULT_SIGBUS;
3475 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3476 ret = do_read_fault(vmf);
3477 else if (!(vma->vm_flags & VM_SHARED))
3478 ret = do_cow_fault(vmf);
3480 ret = do_shared_fault(vmf);
3482 /* preallocated pagetable is unused: free it */
3483 if (vmf->prealloc_pte) {
3484 pte_free(vma->vm_mm, vmf->prealloc_pte);
3485 vmf->prealloc_pte = NULL;
3490 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3491 unsigned long addr, int page_nid,
3496 count_vm_numa_event(NUMA_HINT_FAULTS);
3497 if (page_nid == numa_node_id()) {
3498 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3499 *flags |= TNF_FAULT_LOCAL;
3502 return mpol_misplaced(page, vma, addr);
3505 static int do_numa_page(struct vm_fault *vmf)
3507 struct vm_area_struct *vma = vmf->vma;
3508 struct page *page = NULL;
3512 bool migrated = false;
3514 bool was_writable = pte_savedwrite(vmf->orig_pte);
3518 * The "pte" at this point cannot be used safely without
3519 * validation through pte_unmap_same(). It's of NUMA type but
3520 * the pfn may be screwed if the read is non atomic.
3522 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3523 spin_lock(vmf->ptl);
3524 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3525 pte_unmap_unlock(vmf->pte, vmf->ptl);
3530 * Make it present again, Depending on how arch implementes non
3531 * accessible ptes, some can allow access by kernel mode.
3533 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3534 pte = pte_modify(pte, vma->vm_page_prot);
3535 pte = pte_mkyoung(pte);
3537 pte = pte_mkwrite(pte);
3538 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3539 update_mmu_cache(vma, vmf->address, vmf->pte);
3541 page = vm_normal_page(vma, vmf->address, pte);
3543 pte_unmap_unlock(vmf->pte, vmf->ptl);
3547 /* TODO: handle PTE-mapped THP */
3548 if (PageCompound(page)) {
3549 pte_unmap_unlock(vmf->pte, vmf->ptl);
3554 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3555 * much anyway since they can be in shared cache state. This misses
3556 * the case where a mapping is writable but the process never writes
3557 * to it but pte_write gets cleared during protection updates and
3558 * pte_dirty has unpredictable behaviour between PTE scan updates,
3559 * background writeback, dirty balancing and application behaviour.
3561 if (!pte_write(pte))
3562 flags |= TNF_NO_GROUP;
3565 * Flag if the page is shared between multiple address spaces. This
3566 * is later used when determining whether to group tasks together
3568 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3569 flags |= TNF_SHARED;
3571 last_cpupid = page_cpupid_last(page);
3572 page_nid = page_to_nid(page);
3573 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3575 pte_unmap_unlock(vmf->pte, vmf->ptl);
3576 if (target_nid == -1) {
3581 /* Migrate to the requested node */
3582 migrated = migrate_misplaced_page(page, vma, target_nid);
3584 page_nid = target_nid;
3585 flags |= TNF_MIGRATED;
3587 flags |= TNF_MIGRATE_FAIL;
3591 task_numa_fault(last_cpupid, page_nid, 1, flags);
3595 static inline int create_huge_pmd(struct vm_fault *vmf)
3597 if (vma_is_anonymous(vmf->vma))
3598 return do_huge_pmd_anonymous_page(vmf);
3599 if (vmf->vma->vm_ops->huge_fault)
3600 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3601 return VM_FAULT_FALLBACK;
3604 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3606 if (vma_is_anonymous(vmf->vma))
3607 return do_huge_pmd_wp_page(vmf, orig_pmd);
3608 if (vmf->vma->vm_ops->huge_fault)
3609 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3611 /* COW handled on pte level: split pmd */
3612 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3613 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3615 return VM_FAULT_FALLBACK;
3618 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3620 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3623 static int create_huge_pud(struct vm_fault *vmf)
3625 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3626 /* No support for anonymous transparent PUD pages yet */
3627 if (vma_is_anonymous(vmf->vma))
3628 return VM_FAULT_FALLBACK;
3629 if (vmf->vma->vm_ops->huge_fault)
3630 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3631 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3632 return VM_FAULT_FALLBACK;
3635 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3637 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3638 /* No support for anonymous transparent PUD pages yet */
3639 if (vma_is_anonymous(vmf->vma))
3640 return VM_FAULT_FALLBACK;
3641 if (vmf->vma->vm_ops->huge_fault)
3642 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3643 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3644 return VM_FAULT_FALLBACK;
3648 * These routines also need to handle stuff like marking pages dirty
3649 * and/or accessed for architectures that don't do it in hardware (most
3650 * RISC architectures). The early dirtying is also good on the i386.
3652 * There is also a hook called "update_mmu_cache()" that architectures
3653 * with external mmu caches can use to update those (ie the Sparc or
3654 * PowerPC hashed page tables that act as extended TLBs).
3656 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3657 * concurrent faults).
3659 * The mmap_sem may have been released depending on flags and our return value.
3660 * See filemap_fault() and __lock_page_or_retry().
3662 static int handle_pte_fault(struct vm_fault *vmf)
3666 if (unlikely(pmd_none(*vmf->pmd))) {
3668 * Leave __pte_alloc() until later: because vm_ops->fault may
3669 * want to allocate huge page, and if we expose page table
3670 * for an instant, it will be difficult to retract from
3671 * concurrent faults and from rmap lookups.
3675 /* See comment in pte_alloc_one_map() */
3676 if (pmd_devmap_trans_unstable(vmf->pmd))
3679 * A regular pmd is established and it can't morph into a huge
3680 * pmd from under us anymore at this point because we hold the
3681 * mmap_sem read mode and khugepaged takes it in write mode.
3682 * So now it's safe to run pte_offset_map().
3684 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3685 vmf->orig_pte = *vmf->pte;
3688 * some architectures can have larger ptes than wordsize,
3689 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3690 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3691 * atomic accesses. The code below just needs a consistent
3692 * view for the ifs and we later double check anyway with the
3693 * ptl lock held. So here a barrier will do.
3696 if (pte_none(vmf->orig_pte)) {
3697 pte_unmap(vmf->pte);
3703 if (vma_is_anonymous(vmf->vma))
3704 return do_anonymous_page(vmf);
3706 return do_fault(vmf);
3709 if (!pte_present(vmf->orig_pte))
3710 return do_swap_page(vmf);
3712 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3713 return do_numa_page(vmf);
3715 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3716 spin_lock(vmf->ptl);
3717 entry = vmf->orig_pte;
3718 if (unlikely(!pte_same(*vmf->pte, entry)))
3720 if (vmf->flags & FAULT_FLAG_WRITE) {
3721 if (!pte_write(entry))
3722 return do_wp_page(vmf);
3723 entry = pte_mkdirty(entry);
3725 entry = pte_mkyoung(entry);
3726 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3727 vmf->flags & FAULT_FLAG_WRITE)) {
3728 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3731 * This is needed only for protection faults but the arch code
3732 * is not yet telling us if this is a protection fault or not.
3733 * This still avoids useless tlb flushes for .text page faults
3736 if (vmf->flags & FAULT_FLAG_WRITE)
3737 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3740 pte_unmap_unlock(vmf->pte, vmf->ptl);
3745 * By the time we get here, we already hold the mm semaphore
3747 * The mmap_sem may have been released depending on flags and our
3748 * return value. See filemap_fault() and __lock_page_or_retry().
3750 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3753 struct vm_fault vmf = {
3755 .address = address & PAGE_MASK,
3757 .pgoff = linear_page_index(vma, address),
3758 .gfp_mask = __get_fault_gfp_mask(vma),
3760 struct mm_struct *mm = vma->vm_mm;
3765 pgd = pgd_offset(mm, address);
3766 p4d = p4d_alloc(mm, pgd, address);
3768 return VM_FAULT_OOM;
3770 vmf.pud = pud_alloc(mm, p4d, address);
3772 return VM_FAULT_OOM;
3773 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3774 ret = create_huge_pud(&vmf);
3775 if (!(ret & VM_FAULT_FALLBACK))
3778 pud_t orig_pud = *vmf.pud;
3781 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3782 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3784 /* NUMA case for anonymous PUDs would go here */
3786 if (dirty && !pud_write(orig_pud)) {
3787 ret = wp_huge_pud(&vmf, orig_pud);
3788 if (!(ret & VM_FAULT_FALLBACK))
3791 huge_pud_set_accessed(&vmf, orig_pud);
3797 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3799 return VM_FAULT_OOM;
3800 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3801 ret = create_huge_pmd(&vmf);
3802 if (!(ret & VM_FAULT_FALLBACK))
3805 pmd_t orig_pmd = *vmf.pmd;
3808 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3809 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3810 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3812 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3813 !pmd_write(orig_pmd)) {
3814 ret = wp_huge_pmd(&vmf, orig_pmd);
3815 if (!(ret & VM_FAULT_FALLBACK))
3818 huge_pmd_set_accessed(&vmf, orig_pmd);
3824 return handle_pte_fault(&vmf);
3828 * By the time we get here, we already hold the mm semaphore
3830 * The mmap_sem may have been released depending on flags and our
3831 * return value. See filemap_fault() and __lock_page_or_retry().
3833 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3838 __set_current_state(TASK_RUNNING);
3840 count_vm_event(PGFAULT);
3841 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3843 /* do counter updates before entering really critical section. */
3844 check_sync_rss_stat(current);
3847 * Enable the memcg OOM handling for faults triggered in user
3848 * space. Kernel faults are handled more gracefully.
3850 if (flags & FAULT_FLAG_USER)
3851 mem_cgroup_oom_enable();
3853 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3854 flags & FAULT_FLAG_INSTRUCTION,
3855 flags & FAULT_FLAG_REMOTE))
3856 return VM_FAULT_SIGSEGV;
3858 if (unlikely(is_vm_hugetlb_page(vma)))
3859 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3861 ret = __handle_mm_fault(vma, address, flags);
3863 if (flags & FAULT_FLAG_USER) {
3864 mem_cgroup_oom_disable();
3866 * The task may have entered a memcg OOM situation but
3867 * if the allocation error was handled gracefully (no
3868 * VM_FAULT_OOM), there is no need to kill anything.
3869 * Just clean up the OOM state peacefully.
3871 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3872 mem_cgroup_oom_synchronize(false);
3876 * This mm has been already reaped by the oom reaper and so the
3877 * refault cannot be trusted in general. Anonymous refaults would
3878 * lose data and give a zero page instead e.g. This is especially
3879 * problem for use_mm() because regular tasks will just die and
3880 * the corrupted data will not be visible anywhere while kthread
3881 * will outlive the oom victim and potentially propagate the date
3884 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3885 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3886 ret = VM_FAULT_SIGBUS;
3890 EXPORT_SYMBOL_GPL(handle_mm_fault);
3892 #ifndef __PAGETABLE_P4D_FOLDED
3894 * Allocate p4d page table.
3895 * We've already handled the fast-path in-line.
3897 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3899 p4d_t *new = p4d_alloc_one(mm, address);
3903 smp_wmb(); /* See comment in __pte_alloc */
3905 spin_lock(&mm->page_table_lock);
3906 if (pgd_present(*pgd)) /* Another has populated it */
3909 pgd_populate(mm, pgd, new);
3910 spin_unlock(&mm->page_table_lock);
3913 #endif /* __PAGETABLE_P4D_FOLDED */
3915 #ifndef __PAGETABLE_PUD_FOLDED
3917 * Allocate page upper directory.
3918 * We've already handled the fast-path in-line.
3920 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
3922 pud_t *new = pud_alloc_one(mm, address);
3926 smp_wmb(); /* See comment in __pte_alloc */
3928 spin_lock(&mm->page_table_lock);
3929 #ifndef __ARCH_HAS_5LEVEL_HACK
3930 if (p4d_present(*p4d)) /* Another has populated it */
3933 p4d_populate(mm, p4d, new);
3935 if (pgd_present(*p4d)) /* Another has populated it */
3938 pgd_populate(mm, p4d, new);
3939 #endif /* __ARCH_HAS_5LEVEL_HACK */
3940 spin_unlock(&mm->page_table_lock);
3943 #endif /* __PAGETABLE_PUD_FOLDED */
3945 #ifndef __PAGETABLE_PMD_FOLDED
3947 * Allocate page middle directory.
3948 * We've already handled the fast-path in-line.
3950 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3953 pmd_t *new = pmd_alloc_one(mm, address);
3957 smp_wmb(); /* See comment in __pte_alloc */
3959 ptl = pud_lock(mm, pud);
3960 #ifndef __ARCH_HAS_4LEVEL_HACK
3961 if (!pud_present(*pud)) {
3963 pud_populate(mm, pud, new);
3964 } else /* Another has populated it */
3967 if (!pgd_present(*pud)) {
3969 pgd_populate(mm, pud, new);
3970 } else /* Another has populated it */
3972 #endif /* __ARCH_HAS_4LEVEL_HACK */
3976 #endif /* __PAGETABLE_PMD_FOLDED */
3978 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3979 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3987 pgd = pgd_offset(mm, address);
3988 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3991 p4d = p4d_offset(pgd, address);
3992 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
3995 pud = pud_offset(p4d, address);
3996 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3999 pmd = pmd_offset(pud, address);
4000 VM_BUG_ON(pmd_trans_huge(*pmd));
4002 if (pmd_huge(*pmd)) {
4006 *ptlp = pmd_lock(mm, pmd);
4007 if (pmd_huge(*pmd)) {
4014 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4017 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4018 if (!pte_present(*ptep))
4023 pte_unmap_unlock(ptep, *ptlp);
4028 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4029 pte_t **ptepp, spinlock_t **ptlp)
4033 /* (void) is needed to make gcc happy */
4034 (void) __cond_lock(*ptlp,
4035 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
4040 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4041 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4045 /* (void) is needed to make gcc happy */
4046 (void) __cond_lock(*ptlp,
4047 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
4051 EXPORT_SYMBOL(follow_pte_pmd);
4054 * follow_pfn - look up PFN at a user virtual address
4055 * @vma: memory mapping
4056 * @address: user virtual address
4057 * @pfn: location to store found PFN
4059 * Only IO mappings and raw PFN mappings are allowed.
4061 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4063 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4070 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4073 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4076 *pfn = pte_pfn(*ptep);
4077 pte_unmap_unlock(ptep, ptl);
4080 EXPORT_SYMBOL(follow_pfn);
4082 #ifdef CONFIG_HAVE_IOREMAP_PROT
4083 int follow_phys(struct vm_area_struct *vma,
4084 unsigned long address, unsigned int flags,
4085 unsigned long *prot, resource_size_t *phys)
4091 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4094 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4098 if ((flags & FOLL_WRITE) && !pte_write(pte))
4101 *prot = pgprot_val(pte_pgprot(pte));
4102 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4106 pte_unmap_unlock(ptep, ptl);
4111 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4112 void *buf, int len, int write)
4114 resource_size_t phys_addr;
4115 unsigned long prot = 0;
4116 void __iomem *maddr;
4117 int offset = addr & (PAGE_SIZE-1);
4119 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4122 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4124 memcpy_toio(maddr + offset, buf, len);
4126 memcpy_fromio(buf, maddr + offset, len);
4131 EXPORT_SYMBOL_GPL(generic_access_phys);
4135 * Access another process' address space as given in mm. If non-NULL, use the
4136 * given task for page fault accounting.
4138 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4139 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4141 struct vm_area_struct *vma;
4142 void *old_buf = buf;
4143 int write = gup_flags & FOLL_WRITE;
4145 down_read(&mm->mmap_sem);
4146 /* ignore errors, just check how much was successfully transferred */
4148 int bytes, ret, offset;
4150 struct page *page = NULL;
4152 ret = get_user_pages_remote(tsk, mm, addr, 1,
4153 gup_flags, &page, &vma, NULL);
4155 #ifndef CONFIG_HAVE_IOREMAP_PROT
4159 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4160 * we can access using slightly different code.
4162 vma = find_vma(mm, addr);
4163 if (!vma || vma->vm_start > addr)
4165 if (vma->vm_ops && vma->vm_ops->access)
4166 ret = vma->vm_ops->access(vma, addr, buf,
4174 offset = addr & (PAGE_SIZE-1);
4175 if (bytes > PAGE_SIZE-offset)
4176 bytes = PAGE_SIZE-offset;
4180 copy_to_user_page(vma, page, addr,
4181 maddr + offset, buf, bytes);
4182 set_page_dirty_lock(page);
4184 copy_from_user_page(vma, page, addr,
4185 buf, maddr + offset, bytes);
4194 up_read(&mm->mmap_sem);
4196 return buf - old_buf;
4200 * access_remote_vm - access another process' address space
4201 * @mm: the mm_struct of the target address space
4202 * @addr: start address to access
4203 * @buf: source or destination buffer
4204 * @len: number of bytes to transfer
4205 * @gup_flags: flags modifying lookup behaviour
4207 * The caller must hold a reference on @mm.
4209 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4210 void *buf, int len, unsigned int gup_flags)
4212 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4216 * Access another process' address space.
4217 * Source/target buffer must be kernel space,
4218 * Do not walk the page table directly, use get_user_pages
4220 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4221 void *buf, int len, unsigned int gup_flags)
4223 struct mm_struct *mm;
4226 mm = get_task_mm(tsk);
4230 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4236 EXPORT_SYMBOL_GPL(access_process_vm);
4239 * Print the name of a VMA.
4241 void print_vma_addr(char *prefix, unsigned long ip)
4243 struct mm_struct *mm = current->mm;
4244 struct vm_area_struct *vma;
4247 * Do not print if we are in atomic
4248 * contexts (in exception stacks, etc.):
4250 if (preempt_count())
4253 down_read(&mm->mmap_sem);
4254 vma = find_vma(mm, ip);
4255 if (vma && vma->vm_file) {
4256 struct file *f = vma->vm_file;
4257 char *buf = (char *)__get_free_page(GFP_KERNEL);
4261 p = file_path(f, buf, PAGE_SIZE);
4264 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4266 vma->vm_end - vma->vm_start);
4267 free_page((unsigned long)buf);
4270 up_read(&mm->mmap_sem);
4273 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4274 void __might_fault(const char *file, int line)
4277 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4278 * holding the mmap_sem, this is safe because kernel memory doesn't
4279 * get paged out, therefore we'll never actually fault, and the
4280 * below annotations will generate false positives.
4282 if (uaccess_kernel())
4284 if (pagefault_disabled())
4286 __might_sleep(file, line, 0);
4287 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4289 might_lock_read(¤t->mm->mmap_sem);
4292 EXPORT_SYMBOL(__might_fault);
4295 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4296 static void clear_gigantic_page(struct page *page,
4298 unsigned int pages_per_huge_page)
4301 struct page *p = page;
4304 for (i = 0; i < pages_per_huge_page;
4305 i++, p = mem_map_next(p, page, i)) {
4307 clear_user_highpage(p, addr + i * PAGE_SIZE);
4310 void clear_huge_page(struct page *page,
4311 unsigned long addr, unsigned int pages_per_huge_page)
4315 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4316 clear_gigantic_page(page, addr, pages_per_huge_page);
4321 for (i = 0; i < pages_per_huge_page; i++) {
4323 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4327 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4329 struct vm_area_struct *vma,
4330 unsigned int pages_per_huge_page)
4333 struct page *dst_base = dst;
4334 struct page *src_base = src;
4336 for (i = 0; i < pages_per_huge_page; ) {
4338 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4341 dst = mem_map_next(dst, dst_base, i);
4342 src = mem_map_next(src, src_base, i);
4346 void copy_user_huge_page(struct page *dst, struct page *src,
4347 unsigned long addr, struct vm_area_struct *vma,
4348 unsigned int pages_per_huge_page)
4352 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4353 copy_user_gigantic_page(dst, src, addr, vma,
4354 pages_per_huge_page);
4359 for (i = 0; i < pages_per_huge_page; i++) {
4361 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4365 long copy_huge_page_from_user(struct page *dst_page,
4366 const void __user *usr_src,
4367 unsigned int pages_per_huge_page,
4368 bool allow_pagefault)
4370 void *src = (void *)usr_src;
4372 unsigned long i, rc = 0;
4373 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4375 for (i = 0; i < pages_per_huge_page; i++) {
4376 if (allow_pagefault)
4377 page_kaddr = kmap(dst_page + i);
4379 page_kaddr = kmap_atomic(dst_page + i);
4380 rc = copy_from_user(page_kaddr,
4381 (const void __user *)(src + i * PAGE_SIZE),
4383 if (allow_pagefault)
4384 kunmap(dst_page + i);
4386 kunmap_atomic(page_kaddr);
4388 ret_val -= (PAGE_SIZE - rc);
4396 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4398 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4400 static struct kmem_cache *page_ptl_cachep;
4402 void __init ptlock_cache_init(void)
4404 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4408 bool ptlock_alloc(struct page *page)
4412 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4419 void ptlock_free(struct page *page)
4421 kmem_cache_free(page_ptl_cachep, page->ptl);