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;
218 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
219 unsigned long start, unsigned long end)
223 /* Is it from 0 to ~0? */
224 tlb->fullmm = !(start | (end+1));
225 tlb->need_flush_all = 0;
226 tlb->local.next = NULL;
228 tlb->local.max = ARRAY_SIZE(tlb->__pages);
229 tlb->active = &tlb->local;
230 tlb->batch_count = 0;
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237 __tlb_reset_range(tlb);
240 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
246 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
247 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
248 tlb_table_flush(tlb);
250 __tlb_reset_range(tlb);
253 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255 struct mmu_gather_batch *batch;
257 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
258 free_pages_and_swap_cache(batch->pages, batch->nr);
261 tlb->active = &tlb->local;
264 void tlb_flush_mmu(struct mmu_gather *tlb)
266 tlb_flush_mmu_tlbonly(tlb);
267 tlb_flush_mmu_free(tlb);
271 * Called at the end of the shootdown operation to free up any resources
272 * that were required.
274 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
275 unsigned long start, unsigned long end)
277 struct mmu_gather_batch *batch, *next;
281 /* keep the page table cache within bounds */
284 for (batch = tlb->local.next; batch; batch = next) {
286 free_pages((unsigned long)batch, 0);
288 tlb->local.next = NULL;
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
296 *returns true if the caller should flush.
298 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
300 struct mmu_gather_batch *batch;
302 VM_BUG_ON(!tlb->end);
303 VM_WARN_ON(tlb->page_size != page_size);
307 * Add the page and check if we are full. If so
310 batch->pages[batch->nr++] = page;
311 if (batch->nr == batch->max) {
312 if (!tlb_next_batch(tlb))
316 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
321 #endif /* HAVE_GENERIC_MMU_GATHER */
323 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
326 * See the comment near struct mmu_table_batch.
329 static void tlb_remove_table_smp_sync(void *arg)
331 /* Simply deliver the interrupt */
334 static void tlb_remove_table_one(void *table)
337 * This isn't an RCU grace period and hence the page-tables cannot be
338 * assumed to be actually RCU-freed.
340 * It is however sufficient for software page-table walkers that rely on
341 * IRQ disabling. See the comment near struct mmu_table_batch.
343 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
344 __tlb_remove_table(table);
347 static void tlb_remove_table_rcu(struct rcu_head *head)
349 struct mmu_table_batch *batch;
352 batch = container_of(head, struct mmu_table_batch, rcu);
354 for (i = 0; i < batch->nr; i++)
355 __tlb_remove_table(batch->tables[i]);
357 free_page((unsigned long)batch);
360 void tlb_table_flush(struct mmu_gather *tlb)
362 struct mmu_table_batch **batch = &tlb->batch;
365 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
370 void tlb_remove_table(struct mmu_gather *tlb, void *table)
372 struct mmu_table_batch **batch = &tlb->batch;
375 * When there's less then two users of this mm there cannot be a
376 * concurrent page-table walk.
378 if (atomic_read(&tlb->mm->mm_users) < 2) {
379 __tlb_remove_table(table);
383 if (*batch == NULL) {
384 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
385 if (*batch == NULL) {
386 tlb_remove_table_one(table);
391 (*batch)->tables[(*batch)->nr++] = table;
392 if ((*batch)->nr == MAX_TABLE_BATCH)
393 tlb_table_flush(tlb);
396 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
399 * Called to initialize an (on-stack) mmu_gather structure for page-table
400 * tear-down from @mm. The @fullmm argument is used when @mm is without
401 * users and we're going to destroy the full address space (exit/execve).
403 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
404 unsigned long start, unsigned long end)
406 arch_tlb_gather_mmu(tlb, mm, start, end);
409 void tlb_finish_mmu(struct mmu_gather *tlb,
410 unsigned long start, unsigned long end)
412 arch_tlb_finish_mmu(tlb, start, end);
416 * Note: this doesn't free the actual pages themselves. That
417 * has been handled earlier when unmapping all the memory regions.
419 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
422 pgtable_t token = pmd_pgtable(*pmd);
424 pte_free_tlb(tlb, token, addr);
425 atomic_long_dec(&tlb->mm->nr_ptes);
428 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
429 unsigned long addr, unsigned long end,
430 unsigned long floor, unsigned long ceiling)
437 pmd = pmd_offset(pud, addr);
439 next = pmd_addr_end(addr, end);
440 if (pmd_none_or_clear_bad(pmd))
442 free_pte_range(tlb, pmd, addr);
443 } while (pmd++, addr = next, addr != end);
453 if (end - 1 > ceiling - 1)
456 pmd = pmd_offset(pud, start);
458 pmd_free_tlb(tlb, pmd, start);
459 mm_dec_nr_pmds(tlb->mm);
462 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
463 unsigned long addr, unsigned long end,
464 unsigned long floor, unsigned long ceiling)
471 pud = pud_offset(p4d, addr);
473 next = pud_addr_end(addr, end);
474 if (pud_none_or_clear_bad(pud))
476 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
477 } while (pud++, addr = next, addr != end);
487 if (end - 1 > ceiling - 1)
490 pud = pud_offset(p4d, start);
492 pud_free_tlb(tlb, pud, start);
495 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
496 unsigned long addr, unsigned long end,
497 unsigned long floor, unsigned long ceiling)
504 p4d = p4d_offset(pgd, addr);
506 next = p4d_addr_end(addr, end);
507 if (p4d_none_or_clear_bad(p4d))
509 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
510 } while (p4d++, addr = next, addr != end);
516 ceiling &= PGDIR_MASK;
520 if (end - 1 > ceiling - 1)
523 p4d = p4d_offset(pgd, start);
525 p4d_free_tlb(tlb, p4d, start);
529 * This function frees user-level page tables of a process.
531 void free_pgd_range(struct mmu_gather *tlb,
532 unsigned long addr, unsigned long end,
533 unsigned long floor, unsigned long ceiling)
539 * The next few lines have given us lots of grief...
541 * Why are we testing PMD* at this top level? Because often
542 * there will be no work to do at all, and we'd prefer not to
543 * go all the way down to the bottom just to discover that.
545 * Why all these "- 1"s? Because 0 represents both the bottom
546 * of the address space and the top of it (using -1 for the
547 * top wouldn't help much: the masks would do the wrong thing).
548 * The rule is that addr 0 and floor 0 refer to the bottom of
549 * the address space, but end 0 and ceiling 0 refer to the top
550 * Comparisons need to use "end - 1" and "ceiling - 1" (though
551 * that end 0 case should be mythical).
553 * Wherever addr is brought up or ceiling brought down, we must
554 * be careful to reject "the opposite 0" before it confuses the
555 * subsequent tests. But what about where end is brought down
556 * by PMD_SIZE below? no, end can't go down to 0 there.
558 * Whereas we round start (addr) and ceiling down, by different
559 * masks at different levels, in order to test whether a table
560 * now has no other vmas using it, so can be freed, we don't
561 * bother to round floor or end up - the tests don't need that.
575 if (end - 1 > ceiling - 1)
580 * We add page table cache pages with PAGE_SIZE,
581 * (see pte_free_tlb()), flush the tlb if we need
583 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
584 pgd = pgd_offset(tlb->mm, addr);
586 next = pgd_addr_end(addr, end);
587 if (pgd_none_or_clear_bad(pgd))
589 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
590 } while (pgd++, addr = next, addr != end);
593 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
594 unsigned long floor, unsigned long ceiling)
597 struct vm_area_struct *next = vma->vm_next;
598 unsigned long addr = vma->vm_start;
601 * Hide vma from rmap and truncate_pagecache before freeing
604 unlink_anon_vmas(vma);
605 unlink_file_vma(vma);
607 if (is_vm_hugetlb_page(vma)) {
608 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
609 floor, next ? next->vm_start : ceiling);
612 * Optimization: gather nearby vmas into one call down
614 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
615 && !is_vm_hugetlb_page(next)) {
618 unlink_anon_vmas(vma);
619 unlink_file_vma(vma);
621 free_pgd_range(tlb, addr, vma->vm_end,
622 floor, next ? next->vm_start : ceiling);
628 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
631 pgtable_t new = pte_alloc_one(mm, address);
636 * Ensure all pte setup (eg. pte page lock and page clearing) are
637 * visible before the pte is made visible to other CPUs by being
638 * put into page tables.
640 * The other side of the story is the pointer chasing in the page
641 * table walking code (when walking the page table without locking;
642 * ie. most of the time). Fortunately, these data accesses consist
643 * of a chain of data-dependent loads, meaning most CPUs (alpha
644 * being the notable exception) will already guarantee loads are
645 * seen in-order. See the alpha page table accessors for the
646 * smp_read_barrier_depends() barriers in page table walking code.
648 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
650 ptl = pmd_lock(mm, pmd);
651 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
652 atomic_long_inc(&mm->nr_ptes);
653 pmd_populate(mm, pmd, new);
662 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
664 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
668 smp_wmb(); /* See comment in __pte_alloc */
670 spin_lock(&init_mm.page_table_lock);
671 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
672 pmd_populate_kernel(&init_mm, pmd, new);
675 spin_unlock(&init_mm.page_table_lock);
677 pte_free_kernel(&init_mm, new);
681 static inline void init_rss_vec(int *rss)
683 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
686 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
690 if (current->mm == mm)
692 for (i = 0; i < NR_MM_COUNTERS; i++)
694 add_mm_counter(mm, i, rss[i]);
698 * This function is called to print an error when a bad pte
699 * is found. For example, we might have a PFN-mapped pte in
700 * a region that doesn't allow it.
702 * The calling function must still handle the error.
704 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
705 pte_t pte, struct page *page)
707 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
708 p4d_t *p4d = p4d_offset(pgd, addr);
709 pud_t *pud = pud_offset(p4d, addr);
710 pmd_t *pmd = pmd_offset(pud, addr);
711 struct address_space *mapping;
713 static unsigned long resume;
714 static unsigned long nr_shown;
715 static unsigned long nr_unshown;
718 * Allow a burst of 60 reports, then keep quiet for that minute;
719 * or allow a steady drip of one report per second.
721 if (nr_shown == 60) {
722 if (time_before(jiffies, resume)) {
727 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
734 resume = jiffies + 60 * HZ;
736 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
737 index = linear_page_index(vma, addr);
739 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
741 (long long)pte_val(pte), (long long)pmd_val(*pmd));
743 dump_page(page, "bad pte");
744 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
745 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
747 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
749 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
751 vma->vm_ops ? vma->vm_ops->fault : NULL,
752 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
753 mapping ? mapping->a_ops->readpage : NULL);
755 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
759 * vm_normal_page -- This function gets the "struct page" associated with a pte.
761 * "Special" mappings do not wish to be associated with a "struct page" (either
762 * it doesn't exist, or it exists but they don't want to touch it). In this
763 * case, NULL is returned here. "Normal" mappings do have a struct page.
765 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
766 * pte bit, in which case this function is trivial. Secondly, an architecture
767 * may not have a spare pte bit, which requires a more complicated scheme,
770 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
771 * special mapping (even if there are underlying and valid "struct pages").
772 * COWed pages of a VM_PFNMAP are always normal.
774 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
775 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
776 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
777 * mapping will always honor the rule
779 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
781 * And for normal mappings this is false.
783 * This restricts such mappings to be a linear translation from virtual address
784 * to pfn. To get around this restriction, we allow arbitrary mappings so long
785 * as the vma is not a COW mapping; in that case, we know that all ptes are
786 * special (because none can have been COWed).
789 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
791 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
792 * page" backing, however the difference is that _all_ pages with a struct
793 * page (that is, those where pfn_valid is true) are refcounted and considered
794 * normal pages by the VM. The disadvantage is that pages are refcounted
795 * (which can be slower and simply not an option for some PFNMAP users). The
796 * advantage is that we don't have to follow the strict linearity rule of
797 * PFNMAP mappings in order to support COWable mappings.
800 #ifdef __HAVE_ARCH_PTE_SPECIAL
801 # define HAVE_PTE_SPECIAL 1
803 # define HAVE_PTE_SPECIAL 0
805 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
808 unsigned long pfn = pte_pfn(pte);
810 if (HAVE_PTE_SPECIAL) {
811 if (likely(!pte_special(pte)))
813 if (vma->vm_ops && vma->vm_ops->find_special_page)
814 return vma->vm_ops->find_special_page(vma, addr);
815 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
817 if (!is_zero_pfn(pfn))
818 print_bad_pte(vma, addr, pte, NULL);
822 /* !HAVE_PTE_SPECIAL case follows: */
824 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
825 if (vma->vm_flags & VM_MIXEDMAP) {
831 off = (addr - vma->vm_start) >> PAGE_SHIFT;
832 if (pfn == vma->vm_pgoff + off)
834 if (!is_cow_mapping(vma->vm_flags))
839 if (is_zero_pfn(pfn))
842 if (unlikely(pfn > highest_memmap_pfn)) {
843 print_bad_pte(vma, addr, pte, NULL);
848 * NOTE! We still have PageReserved() pages in the page tables.
849 * eg. VDSO mappings can cause them to exist.
852 return pfn_to_page(pfn);
855 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
856 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
859 unsigned long pfn = pmd_pfn(pmd);
862 * There is no pmd_special() but there may be special pmds, e.g.
863 * in a direct-access (dax) mapping, so let's just replicate the
864 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
866 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
867 if (vma->vm_flags & VM_MIXEDMAP) {
873 off = (addr - vma->vm_start) >> PAGE_SHIFT;
874 if (pfn == vma->vm_pgoff + off)
876 if (!is_cow_mapping(vma->vm_flags))
881 if (is_zero_pfn(pfn))
883 if (unlikely(pfn > highest_memmap_pfn))
887 * NOTE! We still have PageReserved() pages in the page tables.
888 * eg. VDSO mappings can cause them to exist.
891 return pfn_to_page(pfn);
896 * copy one vm_area from one task to the other. Assumes the page tables
897 * already present in the new task to be cleared in the whole range
898 * covered by this vma.
901 static inline unsigned long
902 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
903 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
904 unsigned long addr, int *rss)
906 unsigned long vm_flags = vma->vm_flags;
907 pte_t pte = *src_pte;
910 /* pte contains position in swap or file, so copy. */
911 if (unlikely(!pte_present(pte))) {
912 swp_entry_t entry = pte_to_swp_entry(pte);
914 if (likely(!non_swap_entry(entry))) {
915 if (swap_duplicate(entry) < 0)
918 /* make sure dst_mm is on swapoff's mmlist. */
919 if (unlikely(list_empty(&dst_mm->mmlist))) {
920 spin_lock(&mmlist_lock);
921 if (list_empty(&dst_mm->mmlist))
922 list_add(&dst_mm->mmlist,
924 spin_unlock(&mmlist_lock);
927 } else if (is_migration_entry(entry)) {
928 page = migration_entry_to_page(entry);
930 rss[mm_counter(page)]++;
932 if (is_write_migration_entry(entry) &&
933 is_cow_mapping(vm_flags)) {
935 * COW mappings require pages in both
936 * parent and child to be set to read.
938 make_migration_entry_read(&entry);
939 pte = swp_entry_to_pte(entry);
940 if (pte_swp_soft_dirty(*src_pte))
941 pte = pte_swp_mksoft_dirty(pte);
942 set_pte_at(src_mm, addr, src_pte, pte);
949 * If it's a COW mapping, write protect it both
950 * in the parent and the child
952 if (is_cow_mapping(vm_flags)) {
953 ptep_set_wrprotect(src_mm, addr, src_pte);
954 pte = pte_wrprotect(pte);
958 * If it's a shared mapping, mark it clean in
961 if (vm_flags & VM_SHARED)
962 pte = pte_mkclean(pte);
963 pte = pte_mkold(pte);
965 page = vm_normal_page(vma, addr, pte);
968 page_dup_rmap(page, false);
969 rss[mm_counter(page)]++;
973 set_pte_at(dst_mm, addr, dst_pte, pte);
977 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
978 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
979 unsigned long addr, unsigned long end)
981 pte_t *orig_src_pte, *orig_dst_pte;
982 pte_t *src_pte, *dst_pte;
983 spinlock_t *src_ptl, *dst_ptl;
985 int rss[NR_MM_COUNTERS];
986 swp_entry_t entry = (swp_entry_t){0};
991 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
994 src_pte = pte_offset_map(src_pmd, addr);
995 src_ptl = pte_lockptr(src_mm, src_pmd);
996 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
997 orig_src_pte = src_pte;
998 orig_dst_pte = dst_pte;
999 arch_enter_lazy_mmu_mode();
1003 * We are holding two locks at this point - either of them
1004 * could generate latencies in another task on another CPU.
1006 if (progress >= 32) {
1008 if (need_resched() ||
1009 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1012 if (pte_none(*src_pte)) {
1016 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1021 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1023 arch_leave_lazy_mmu_mode();
1024 spin_unlock(src_ptl);
1025 pte_unmap(orig_src_pte);
1026 add_mm_rss_vec(dst_mm, rss);
1027 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1031 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1040 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1041 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1042 unsigned long addr, unsigned long end)
1044 pmd_t *src_pmd, *dst_pmd;
1047 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1050 src_pmd = pmd_offset(src_pud, addr);
1052 next = pmd_addr_end(addr, end);
1053 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1055 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1056 err = copy_huge_pmd(dst_mm, src_mm,
1057 dst_pmd, src_pmd, addr, vma);
1064 if (pmd_none_or_clear_bad(src_pmd))
1066 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1069 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1073 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1074 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1075 unsigned long addr, unsigned long end)
1077 pud_t *src_pud, *dst_pud;
1080 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1083 src_pud = pud_offset(src_p4d, addr);
1085 next = pud_addr_end(addr, end);
1086 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1089 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1090 err = copy_huge_pud(dst_mm, src_mm,
1091 dst_pud, src_pud, addr, vma);
1098 if (pud_none_or_clear_bad(src_pud))
1100 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1103 } while (dst_pud++, src_pud++, addr = next, addr != end);
1107 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1108 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1109 unsigned long addr, unsigned long end)
1111 p4d_t *src_p4d, *dst_p4d;
1114 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1117 src_p4d = p4d_offset(src_pgd, addr);
1119 next = p4d_addr_end(addr, end);
1120 if (p4d_none_or_clear_bad(src_p4d))
1122 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1125 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1129 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1130 struct vm_area_struct *vma)
1132 pgd_t *src_pgd, *dst_pgd;
1134 unsigned long addr = vma->vm_start;
1135 unsigned long end = vma->vm_end;
1136 unsigned long mmun_start; /* For mmu_notifiers */
1137 unsigned long mmun_end; /* For mmu_notifiers */
1142 * Don't copy ptes where a page fault will fill them correctly.
1143 * Fork becomes much lighter when there are big shared or private
1144 * readonly mappings. The tradeoff is that copy_page_range is more
1145 * efficient than faulting.
1147 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1151 if (is_vm_hugetlb_page(vma))
1152 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1154 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1156 * We do not free on error cases below as remove_vma
1157 * gets called on error from higher level routine
1159 ret = track_pfn_copy(vma);
1165 * We need to invalidate the secondary MMU mappings only when
1166 * there could be a permission downgrade on the ptes of the
1167 * parent mm. And a permission downgrade will only happen if
1168 * is_cow_mapping() returns true.
1170 is_cow = is_cow_mapping(vma->vm_flags);
1174 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1178 dst_pgd = pgd_offset(dst_mm, addr);
1179 src_pgd = pgd_offset(src_mm, addr);
1181 next = pgd_addr_end(addr, end);
1182 if (pgd_none_or_clear_bad(src_pgd))
1184 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1185 vma, addr, next))) {
1189 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1192 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1196 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1197 struct vm_area_struct *vma, pmd_t *pmd,
1198 unsigned long addr, unsigned long end,
1199 struct zap_details *details)
1201 struct mm_struct *mm = tlb->mm;
1202 int force_flush = 0;
1203 int rss[NR_MM_COUNTERS];
1209 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1212 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1214 flush_tlb_batched_pending(mm);
1215 arch_enter_lazy_mmu_mode();
1218 if (pte_none(ptent))
1221 if (pte_present(ptent)) {
1224 page = vm_normal_page(vma, addr, ptent);
1225 if (unlikely(details) && page) {
1227 * unmap_shared_mapping_pages() wants to
1228 * invalidate cache without truncating:
1229 * unmap shared but keep private pages.
1231 if (details->check_mapping &&
1232 details->check_mapping != page_rmapping(page))
1235 ptent = ptep_get_and_clear_full(mm, addr, pte,
1237 tlb_remove_tlb_entry(tlb, pte, addr);
1238 if (unlikely(!page))
1241 if (!PageAnon(page)) {
1242 if (pte_dirty(ptent)) {
1244 set_page_dirty(page);
1246 if (pte_young(ptent) &&
1247 likely(!(vma->vm_flags & VM_SEQ_READ)))
1248 mark_page_accessed(page);
1250 rss[mm_counter(page)]--;
1251 page_remove_rmap(page, false);
1252 if (unlikely(page_mapcount(page) < 0))
1253 print_bad_pte(vma, addr, ptent, page);
1254 if (unlikely(__tlb_remove_page(tlb, page))) {
1261 /* If details->check_mapping, we leave swap entries. */
1262 if (unlikely(details))
1265 entry = pte_to_swp_entry(ptent);
1266 if (!non_swap_entry(entry))
1268 else if (is_migration_entry(entry)) {
1271 page = migration_entry_to_page(entry);
1272 rss[mm_counter(page)]--;
1274 if (unlikely(!free_swap_and_cache(entry)))
1275 print_bad_pte(vma, addr, ptent, NULL);
1276 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1277 } while (pte++, addr += PAGE_SIZE, addr != end);
1279 add_mm_rss_vec(mm, rss);
1280 arch_leave_lazy_mmu_mode();
1282 /* Do the actual TLB flush before dropping ptl */
1284 tlb_flush_mmu_tlbonly(tlb);
1285 pte_unmap_unlock(start_pte, ptl);
1288 * If we forced a TLB flush (either due to running out of
1289 * batch buffers or because we needed to flush dirty TLB
1290 * entries before releasing the ptl), free the batched
1291 * memory too. Restart if we didn't do everything.
1295 tlb_flush_mmu_free(tlb);
1303 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1304 struct vm_area_struct *vma, pud_t *pud,
1305 unsigned long addr, unsigned long end,
1306 struct zap_details *details)
1311 pmd = pmd_offset(pud, addr);
1313 next = pmd_addr_end(addr, end);
1314 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1315 if (next - addr != HPAGE_PMD_SIZE) {
1316 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1317 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1318 __split_huge_pmd(vma, pmd, addr, false, NULL);
1319 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1324 * Here there can be other concurrent MADV_DONTNEED or
1325 * trans huge page faults running, and if the pmd is
1326 * none or trans huge it can change under us. This is
1327 * because MADV_DONTNEED holds the mmap_sem in read
1330 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1332 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1335 } while (pmd++, addr = next, addr != end);
1340 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1341 struct vm_area_struct *vma, p4d_t *p4d,
1342 unsigned long addr, unsigned long end,
1343 struct zap_details *details)
1348 pud = pud_offset(p4d, addr);
1350 next = pud_addr_end(addr, end);
1351 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1352 if (next - addr != HPAGE_PUD_SIZE) {
1353 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1354 split_huge_pud(vma, pud, addr);
1355 } else if (zap_huge_pud(tlb, vma, pud, addr))
1359 if (pud_none_or_clear_bad(pud))
1361 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1364 } while (pud++, addr = next, addr != end);
1369 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1370 struct vm_area_struct *vma, pgd_t *pgd,
1371 unsigned long addr, unsigned long end,
1372 struct zap_details *details)
1377 p4d = p4d_offset(pgd, addr);
1379 next = p4d_addr_end(addr, end);
1380 if (p4d_none_or_clear_bad(p4d))
1382 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1383 } while (p4d++, addr = next, addr != end);
1388 void unmap_page_range(struct mmu_gather *tlb,
1389 struct vm_area_struct *vma,
1390 unsigned long addr, unsigned long end,
1391 struct zap_details *details)
1396 BUG_ON(addr >= end);
1397 tlb_start_vma(tlb, vma);
1398 pgd = pgd_offset(vma->vm_mm, addr);
1400 next = pgd_addr_end(addr, end);
1401 if (pgd_none_or_clear_bad(pgd))
1403 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1404 } while (pgd++, addr = next, addr != end);
1405 tlb_end_vma(tlb, vma);
1409 static void unmap_single_vma(struct mmu_gather *tlb,
1410 struct vm_area_struct *vma, unsigned long start_addr,
1411 unsigned long end_addr,
1412 struct zap_details *details)
1414 unsigned long start = max(vma->vm_start, start_addr);
1417 if (start >= vma->vm_end)
1419 end = min(vma->vm_end, end_addr);
1420 if (end <= vma->vm_start)
1424 uprobe_munmap(vma, start, end);
1426 if (unlikely(vma->vm_flags & VM_PFNMAP))
1427 untrack_pfn(vma, 0, 0);
1430 if (unlikely(is_vm_hugetlb_page(vma))) {
1432 * It is undesirable to test vma->vm_file as it
1433 * should be non-null for valid hugetlb area.
1434 * However, vm_file will be NULL in the error
1435 * cleanup path of mmap_region. When
1436 * hugetlbfs ->mmap method fails,
1437 * mmap_region() nullifies vma->vm_file
1438 * before calling this function to clean up.
1439 * Since no pte has actually been setup, it is
1440 * safe to do nothing in this case.
1443 i_mmap_lock_write(vma->vm_file->f_mapping);
1444 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1445 i_mmap_unlock_write(vma->vm_file->f_mapping);
1448 unmap_page_range(tlb, vma, start, end, details);
1453 * unmap_vmas - unmap a range of memory covered by a list of vma's
1454 * @tlb: address of the caller's struct mmu_gather
1455 * @vma: the starting vma
1456 * @start_addr: virtual address at which to start unmapping
1457 * @end_addr: virtual address at which to end unmapping
1459 * Unmap all pages in the vma list.
1461 * Only addresses between `start' and `end' will be unmapped.
1463 * The VMA list must be sorted in ascending virtual address order.
1465 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1466 * range after unmap_vmas() returns. So the only responsibility here is to
1467 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1468 * drops the lock and schedules.
1470 void unmap_vmas(struct mmu_gather *tlb,
1471 struct vm_area_struct *vma, unsigned long start_addr,
1472 unsigned long end_addr)
1474 struct mm_struct *mm = vma->vm_mm;
1476 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1477 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1478 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1479 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1483 * zap_page_range - remove user pages in a given range
1484 * @vma: vm_area_struct holding the applicable pages
1485 * @start: starting address of pages to zap
1486 * @size: number of bytes to zap
1488 * Caller must protect the VMA list
1490 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1493 struct mm_struct *mm = vma->vm_mm;
1494 struct mmu_gather tlb;
1495 unsigned long end = start + size;
1498 tlb_gather_mmu(&tlb, mm, start, end);
1499 update_hiwater_rss(mm);
1500 mmu_notifier_invalidate_range_start(mm, start, end);
1501 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1502 unmap_single_vma(&tlb, vma, start, end, NULL);
1503 mmu_notifier_invalidate_range_end(mm, start, end);
1504 tlb_finish_mmu(&tlb, start, end);
1508 * zap_page_range_single - remove user pages in a given range
1509 * @vma: vm_area_struct holding the applicable pages
1510 * @address: starting address of pages to zap
1511 * @size: number of bytes to zap
1512 * @details: details of shared cache invalidation
1514 * The range must fit into one VMA.
1516 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1517 unsigned long size, struct zap_details *details)
1519 struct mm_struct *mm = vma->vm_mm;
1520 struct mmu_gather tlb;
1521 unsigned long end = address + size;
1524 tlb_gather_mmu(&tlb, mm, address, end);
1525 update_hiwater_rss(mm);
1526 mmu_notifier_invalidate_range_start(mm, address, end);
1527 unmap_single_vma(&tlb, vma, address, end, details);
1528 mmu_notifier_invalidate_range_end(mm, address, end);
1529 tlb_finish_mmu(&tlb, address, end);
1533 * zap_vma_ptes - remove ptes mapping the vma
1534 * @vma: vm_area_struct holding ptes to be zapped
1535 * @address: starting address of pages to zap
1536 * @size: number of bytes to zap
1538 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1540 * The entire address range must be fully contained within the vma.
1542 * Returns 0 if successful.
1544 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1547 if (address < vma->vm_start || address + size > vma->vm_end ||
1548 !(vma->vm_flags & VM_PFNMAP))
1550 zap_page_range_single(vma, address, size, NULL);
1553 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1555 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1563 pgd = pgd_offset(mm, addr);
1564 p4d = p4d_alloc(mm, pgd, addr);
1567 pud = pud_alloc(mm, p4d, addr);
1570 pmd = pmd_alloc(mm, pud, addr);
1574 VM_BUG_ON(pmd_trans_huge(*pmd));
1575 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1579 * This is the old fallback for page remapping.
1581 * For historical reasons, it only allows reserved pages. Only
1582 * old drivers should use this, and they needed to mark their
1583 * pages reserved for the old functions anyway.
1585 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1586 struct page *page, pgprot_t prot)
1588 struct mm_struct *mm = vma->vm_mm;
1597 flush_dcache_page(page);
1598 pte = get_locked_pte(mm, addr, &ptl);
1602 if (!pte_none(*pte))
1605 /* Ok, finally just insert the thing.. */
1607 inc_mm_counter_fast(mm, mm_counter_file(page));
1608 page_add_file_rmap(page, false);
1609 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1612 pte_unmap_unlock(pte, ptl);
1615 pte_unmap_unlock(pte, ptl);
1621 * vm_insert_page - insert single page into user vma
1622 * @vma: user vma to map to
1623 * @addr: target user address of this page
1624 * @page: source kernel page
1626 * This allows drivers to insert individual pages they've allocated
1629 * The page has to be a nice clean _individual_ kernel allocation.
1630 * If you allocate a compound page, you need to have marked it as
1631 * such (__GFP_COMP), or manually just split the page up yourself
1632 * (see split_page()).
1634 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1635 * took an arbitrary page protection parameter. This doesn't allow
1636 * that. Your vma protection will have to be set up correctly, which
1637 * means that if you want a shared writable mapping, you'd better
1638 * ask for a shared writable mapping!
1640 * The page does not need to be reserved.
1642 * Usually this function is called from f_op->mmap() handler
1643 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1644 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1645 * function from other places, for example from page-fault handler.
1647 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1650 if (addr < vma->vm_start || addr >= vma->vm_end)
1652 if (!page_count(page))
1654 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1655 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1656 BUG_ON(vma->vm_flags & VM_PFNMAP);
1657 vma->vm_flags |= VM_MIXEDMAP;
1659 return insert_page(vma, addr, page, vma->vm_page_prot);
1661 EXPORT_SYMBOL(vm_insert_page);
1663 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1664 pfn_t pfn, pgprot_t prot)
1666 struct mm_struct *mm = vma->vm_mm;
1672 pte = get_locked_pte(mm, addr, &ptl);
1676 if (!pte_none(*pte))
1679 /* Ok, finally just insert the thing.. */
1680 if (pfn_t_devmap(pfn))
1681 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1683 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1684 set_pte_at(mm, addr, pte, entry);
1685 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1689 pte_unmap_unlock(pte, ptl);
1695 * vm_insert_pfn - insert single pfn into user vma
1696 * @vma: user vma to map to
1697 * @addr: target user address of this page
1698 * @pfn: source kernel pfn
1700 * Similar to vm_insert_page, this allows drivers to insert individual pages
1701 * they've allocated into a user vma. Same comments apply.
1703 * This function should only be called from a vm_ops->fault handler, and
1704 * in that case the handler should return NULL.
1706 * vma cannot be a COW mapping.
1708 * As this is called only for pages that do not currently exist, we
1709 * do not need to flush old virtual caches or the TLB.
1711 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1714 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1716 EXPORT_SYMBOL(vm_insert_pfn);
1719 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1720 * @vma: user vma to map to
1721 * @addr: target user address of this page
1722 * @pfn: source kernel pfn
1723 * @pgprot: pgprot flags for the inserted page
1725 * This is exactly like vm_insert_pfn, except that it allows drivers to
1726 * to override pgprot on a per-page basis.
1728 * This only makes sense for IO mappings, and it makes no sense for
1729 * cow mappings. In general, using multiple vmas is preferable;
1730 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1733 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1734 unsigned long pfn, pgprot_t pgprot)
1738 * Technically, architectures with pte_special can avoid all these
1739 * restrictions (same for remap_pfn_range). However we would like
1740 * consistency in testing and feature parity among all, so we should
1741 * try to keep these invariants in place for everybody.
1743 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1744 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1745 (VM_PFNMAP|VM_MIXEDMAP));
1746 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1747 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1749 if (addr < vma->vm_start || addr >= vma->vm_end)
1752 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1754 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1758 EXPORT_SYMBOL(vm_insert_pfn_prot);
1760 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1763 pgprot_t pgprot = vma->vm_page_prot;
1765 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1767 if (addr < vma->vm_start || addr >= vma->vm_end)
1770 track_pfn_insert(vma, &pgprot, pfn);
1773 * If we don't have pte special, then we have to use the pfn_valid()
1774 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1775 * refcount the page if pfn_valid is true (hence insert_page rather
1776 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1777 * without pte special, it would there be refcounted as a normal page.
1779 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1783 * At this point we are committed to insert_page()
1784 * regardless of whether the caller specified flags that
1785 * result in pfn_t_has_page() == false.
1787 page = pfn_to_page(pfn_t_to_pfn(pfn));
1788 return insert_page(vma, addr, page, pgprot);
1790 return insert_pfn(vma, addr, pfn, pgprot);
1792 EXPORT_SYMBOL(vm_insert_mixed);
1795 * maps a range of physical memory into the requested pages. the old
1796 * mappings are removed. any references to nonexistent pages results
1797 * in null mappings (currently treated as "copy-on-access")
1799 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1800 unsigned long addr, unsigned long end,
1801 unsigned long pfn, pgprot_t prot)
1806 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1809 arch_enter_lazy_mmu_mode();
1811 BUG_ON(!pte_none(*pte));
1812 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1814 } while (pte++, addr += PAGE_SIZE, addr != end);
1815 arch_leave_lazy_mmu_mode();
1816 pte_unmap_unlock(pte - 1, ptl);
1820 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1821 unsigned long addr, unsigned long end,
1822 unsigned long pfn, pgprot_t prot)
1827 pfn -= addr >> PAGE_SHIFT;
1828 pmd = pmd_alloc(mm, pud, addr);
1831 VM_BUG_ON(pmd_trans_huge(*pmd));
1833 next = pmd_addr_end(addr, end);
1834 if (remap_pte_range(mm, pmd, addr, next,
1835 pfn + (addr >> PAGE_SHIFT), prot))
1837 } while (pmd++, addr = next, addr != end);
1841 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1842 unsigned long addr, unsigned long end,
1843 unsigned long pfn, pgprot_t prot)
1848 pfn -= addr >> PAGE_SHIFT;
1849 pud = pud_alloc(mm, p4d, addr);
1853 next = pud_addr_end(addr, end);
1854 if (remap_pmd_range(mm, pud, addr, next,
1855 pfn + (addr >> PAGE_SHIFT), prot))
1857 } while (pud++, addr = next, addr != end);
1861 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1862 unsigned long addr, unsigned long end,
1863 unsigned long pfn, pgprot_t prot)
1868 pfn -= addr >> PAGE_SHIFT;
1869 p4d = p4d_alloc(mm, pgd, addr);
1873 next = p4d_addr_end(addr, end);
1874 if (remap_pud_range(mm, p4d, addr, next,
1875 pfn + (addr >> PAGE_SHIFT), prot))
1877 } while (p4d++, addr = next, addr != end);
1882 * remap_pfn_range - remap kernel memory to userspace
1883 * @vma: user vma to map to
1884 * @addr: target user address to start at
1885 * @pfn: physical address of kernel memory
1886 * @size: size of map area
1887 * @prot: page protection flags for this mapping
1889 * Note: this is only safe if the mm semaphore is held when called.
1891 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1892 unsigned long pfn, unsigned long size, pgprot_t prot)
1896 unsigned long end = addr + PAGE_ALIGN(size);
1897 struct mm_struct *mm = vma->vm_mm;
1898 unsigned long remap_pfn = pfn;
1902 * Physically remapped pages are special. Tell the
1903 * rest of the world about it:
1904 * VM_IO tells people not to look at these pages
1905 * (accesses can have side effects).
1906 * VM_PFNMAP tells the core MM that the base pages are just
1907 * raw PFN mappings, and do not have a "struct page" associated
1910 * Disable vma merging and expanding with mremap().
1912 * Omit vma from core dump, even when VM_IO turned off.
1914 * There's a horrible special case to handle copy-on-write
1915 * behaviour that some programs depend on. We mark the "original"
1916 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1917 * See vm_normal_page() for details.
1919 if (is_cow_mapping(vma->vm_flags)) {
1920 if (addr != vma->vm_start || end != vma->vm_end)
1922 vma->vm_pgoff = pfn;
1925 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1929 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1931 BUG_ON(addr >= end);
1932 pfn -= addr >> PAGE_SHIFT;
1933 pgd = pgd_offset(mm, addr);
1934 flush_cache_range(vma, addr, end);
1936 next = pgd_addr_end(addr, end);
1937 err = remap_p4d_range(mm, pgd, addr, next,
1938 pfn + (addr >> PAGE_SHIFT), prot);
1941 } while (pgd++, addr = next, addr != end);
1944 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1948 EXPORT_SYMBOL(remap_pfn_range);
1951 * vm_iomap_memory - remap memory to userspace
1952 * @vma: user vma to map to
1953 * @start: start of area
1954 * @len: size of area
1956 * This is a simplified io_remap_pfn_range() for common driver use. The
1957 * driver just needs to give us the physical memory range to be mapped,
1958 * we'll figure out the rest from the vma information.
1960 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1961 * whatever write-combining details or similar.
1963 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1965 unsigned long vm_len, pfn, pages;
1967 /* Check that the physical memory area passed in looks valid */
1968 if (start + len < start)
1971 * You *really* shouldn't map things that aren't page-aligned,
1972 * but we've historically allowed it because IO memory might
1973 * just have smaller alignment.
1975 len += start & ~PAGE_MASK;
1976 pfn = start >> PAGE_SHIFT;
1977 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1978 if (pfn + pages < pfn)
1981 /* We start the mapping 'vm_pgoff' pages into the area */
1982 if (vma->vm_pgoff > pages)
1984 pfn += vma->vm_pgoff;
1985 pages -= vma->vm_pgoff;
1987 /* Can we fit all of the mapping? */
1988 vm_len = vma->vm_end - vma->vm_start;
1989 if (vm_len >> PAGE_SHIFT > pages)
1992 /* Ok, let it rip */
1993 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1995 EXPORT_SYMBOL(vm_iomap_memory);
1997 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1998 unsigned long addr, unsigned long end,
1999 pte_fn_t fn, void *data)
2004 spinlock_t *uninitialized_var(ptl);
2006 pte = (mm == &init_mm) ?
2007 pte_alloc_kernel(pmd, addr) :
2008 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2012 BUG_ON(pmd_huge(*pmd));
2014 arch_enter_lazy_mmu_mode();
2016 token = pmd_pgtable(*pmd);
2019 err = fn(pte++, token, addr, data);
2022 } while (addr += PAGE_SIZE, addr != end);
2024 arch_leave_lazy_mmu_mode();
2027 pte_unmap_unlock(pte-1, ptl);
2031 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2032 unsigned long addr, unsigned long end,
2033 pte_fn_t fn, void *data)
2039 BUG_ON(pud_huge(*pud));
2041 pmd = pmd_alloc(mm, pud, addr);
2045 next = pmd_addr_end(addr, end);
2046 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2049 } while (pmd++, addr = next, addr != end);
2053 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2054 unsigned long addr, unsigned long end,
2055 pte_fn_t fn, void *data)
2061 pud = pud_alloc(mm, p4d, addr);
2065 next = pud_addr_end(addr, end);
2066 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2069 } while (pud++, addr = next, addr != end);
2073 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2074 unsigned long addr, unsigned long end,
2075 pte_fn_t fn, void *data)
2081 p4d = p4d_alloc(mm, pgd, addr);
2085 next = p4d_addr_end(addr, end);
2086 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2089 } while (p4d++, addr = next, addr != end);
2094 * Scan a region of virtual memory, filling in page tables as necessary
2095 * and calling a provided function on each leaf page table.
2097 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2098 unsigned long size, pte_fn_t fn, void *data)
2102 unsigned long end = addr + size;
2105 if (WARN_ON(addr >= end))
2108 pgd = pgd_offset(mm, addr);
2110 next = pgd_addr_end(addr, end);
2111 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2114 } while (pgd++, addr = next, addr != end);
2118 EXPORT_SYMBOL_GPL(apply_to_page_range);
2121 * handle_pte_fault chooses page fault handler according to an entry which was
2122 * read non-atomically. Before making any commitment, on those architectures
2123 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2124 * parts, do_swap_page must check under lock before unmapping the pte and
2125 * proceeding (but do_wp_page is only called after already making such a check;
2126 * and do_anonymous_page can safely check later on).
2128 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2129 pte_t *page_table, pte_t orig_pte)
2132 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2133 if (sizeof(pte_t) > sizeof(unsigned long)) {
2134 spinlock_t *ptl = pte_lockptr(mm, pmd);
2136 same = pte_same(*page_table, orig_pte);
2140 pte_unmap(page_table);
2144 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2146 debug_dma_assert_idle(src);
2149 * If the source page was a PFN mapping, we don't have
2150 * a "struct page" for it. We do a best-effort copy by
2151 * just copying from the original user address. If that
2152 * fails, we just zero-fill it. Live with it.
2154 if (unlikely(!src)) {
2155 void *kaddr = kmap_atomic(dst);
2156 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2159 * This really shouldn't fail, because the page is there
2160 * in the page tables. But it might just be unreadable,
2161 * in which case we just give up and fill the result with
2164 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2166 kunmap_atomic(kaddr);
2167 flush_dcache_page(dst);
2169 copy_user_highpage(dst, src, va, vma);
2172 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2174 struct file *vm_file = vma->vm_file;
2177 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2180 * Special mappings (e.g. VDSO) do not have any file so fake
2181 * a default GFP_KERNEL for them.
2187 * Notify the address space that the page is about to become writable so that
2188 * it can prohibit this or wait for the page to get into an appropriate state.
2190 * We do this without the lock held, so that it can sleep if it needs to.
2192 static int do_page_mkwrite(struct vm_fault *vmf)
2195 struct page *page = vmf->page;
2196 unsigned int old_flags = vmf->flags;
2198 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2200 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2201 /* Restore original flags so that caller is not surprised */
2202 vmf->flags = old_flags;
2203 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2205 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2207 if (!page->mapping) {
2209 return 0; /* retry */
2211 ret |= VM_FAULT_LOCKED;
2213 VM_BUG_ON_PAGE(!PageLocked(page), page);
2218 * Handle dirtying of a page in shared file mapping on a write fault.
2220 * The function expects the page to be locked and unlocks it.
2222 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2225 struct address_space *mapping;
2227 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2229 dirtied = set_page_dirty(page);
2230 VM_BUG_ON_PAGE(PageAnon(page), page);
2232 * Take a local copy of the address_space - page.mapping may be zeroed
2233 * by truncate after unlock_page(). The address_space itself remains
2234 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2235 * release semantics to prevent the compiler from undoing this copying.
2237 mapping = page_rmapping(page);
2240 if ((dirtied || page_mkwrite) && mapping) {
2242 * Some device drivers do not set page.mapping
2243 * but still dirty their pages
2245 balance_dirty_pages_ratelimited(mapping);
2249 file_update_time(vma->vm_file);
2253 * Handle write page faults for pages that can be reused in the current vma
2255 * This can happen either due to the mapping being with the VM_SHARED flag,
2256 * or due to us being the last reference standing to the page. In either
2257 * case, all we need to do here is to mark the page as writable and update
2258 * any related book-keeping.
2260 static inline void wp_page_reuse(struct vm_fault *vmf)
2261 __releases(vmf->ptl)
2263 struct vm_area_struct *vma = vmf->vma;
2264 struct page *page = vmf->page;
2267 * Clear the pages cpupid information as the existing
2268 * information potentially belongs to a now completely
2269 * unrelated process.
2272 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2274 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2275 entry = pte_mkyoung(vmf->orig_pte);
2276 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2277 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2278 update_mmu_cache(vma, vmf->address, vmf->pte);
2279 pte_unmap_unlock(vmf->pte, vmf->ptl);
2283 * Handle the case of a page which we actually need to copy to a new page.
2285 * Called with mmap_sem locked and the old page referenced, but
2286 * without the ptl held.
2288 * High level logic flow:
2290 * - Allocate a page, copy the content of the old page to the new one.
2291 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2292 * - Take the PTL. If the pte changed, bail out and release the allocated page
2293 * - If the pte is still the way we remember it, update the page table and all
2294 * relevant references. This includes dropping the reference the page-table
2295 * held to the old page, as well as updating the rmap.
2296 * - In any case, unlock the PTL and drop the reference we took to the old page.
2298 static int wp_page_copy(struct vm_fault *vmf)
2300 struct vm_area_struct *vma = vmf->vma;
2301 struct mm_struct *mm = vma->vm_mm;
2302 struct page *old_page = vmf->page;
2303 struct page *new_page = NULL;
2305 int page_copied = 0;
2306 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2307 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2308 struct mem_cgroup *memcg;
2310 if (unlikely(anon_vma_prepare(vma)))
2313 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2314 new_page = alloc_zeroed_user_highpage_movable(vma,
2319 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2323 cow_user_page(new_page, old_page, vmf->address, vma);
2326 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2329 __SetPageUptodate(new_page);
2331 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2334 * Re-check the pte - we dropped the lock
2336 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2337 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2339 if (!PageAnon(old_page)) {
2340 dec_mm_counter_fast(mm,
2341 mm_counter_file(old_page));
2342 inc_mm_counter_fast(mm, MM_ANONPAGES);
2345 inc_mm_counter_fast(mm, MM_ANONPAGES);
2347 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2348 entry = mk_pte(new_page, vma->vm_page_prot);
2349 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2351 * Clear the pte entry and flush it first, before updating the
2352 * pte with the new entry. This will avoid a race condition
2353 * seen in the presence of one thread doing SMC and another
2356 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2357 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2358 mem_cgroup_commit_charge(new_page, memcg, false, false);
2359 lru_cache_add_active_or_unevictable(new_page, vma);
2361 * We call the notify macro here because, when using secondary
2362 * mmu page tables (such as kvm shadow page tables), we want the
2363 * new page to be mapped directly into the secondary page table.
2365 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2366 update_mmu_cache(vma, vmf->address, vmf->pte);
2369 * Only after switching the pte to the new page may
2370 * we remove the mapcount here. Otherwise another
2371 * process may come and find the rmap count decremented
2372 * before the pte is switched to the new page, and
2373 * "reuse" the old page writing into it while our pte
2374 * here still points into it and can be read by other
2377 * The critical issue is to order this
2378 * page_remove_rmap with the ptp_clear_flush above.
2379 * Those stores are ordered by (if nothing else,)
2380 * the barrier present in the atomic_add_negative
2381 * in page_remove_rmap.
2383 * Then the TLB flush in ptep_clear_flush ensures that
2384 * no process can access the old page before the
2385 * decremented mapcount is visible. And the old page
2386 * cannot be reused until after the decremented
2387 * mapcount is visible. So transitively, TLBs to
2388 * old page will be flushed before it can be reused.
2390 page_remove_rmap(old_page, false);
2393 /* Free the old page.. */
2394 new_page = old_page;
2397 mem_cgroup_cancel_charge(new_page, memcg, false);
2403 pte_unmap_unlock(vmf->pte, vmf->ptl);
2404 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2407 * Don't let another task, with possibly unlocked vma,
2408 * keep the mlocked page.
2410 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2411 lock_page(old_page); /* LRU manipulation */
2412 if (PageMlocked(old_page))
2413 munlock_vma_page(old_page);
2414 unlock_page(old_page);
2418 return page_copied ? VM_FAULT_WRITE : 0;
2424 return VM_FAULT_OOM;
2428 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2429 * writeable once the page is prepared
2431 * @vmf: structure describing the fault
2433 * This function handles all that is needed to finish a write page fault in a
2434 * shared mapping due to PTE being read-only once the mapped page is prepared.
2435 * It handles locking of PTE and modifying it. The function returns
2436 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2439 * The function expects the page to be locked or other protection against
2440 * concurrent faults / writeback (such as DAX radix tree locks).
2442 int finish_mkwrite_fault(struct vm_fault *vmf)
2444 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2445 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2448 * We might have raced with another page fault while we released the
2449 * pte_offset_map_lock.
2451 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2452 pte_unmap_unlock(vmf->pte, vmf->ptl);
2453 return VM_FAULT_NOPAGE;
2460 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2463 static int wp_pfn_shared(struct vm_fault *vmf)
2465 struct vm_area_struct *vma = vmf->vma;
2467 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2470 pte_unmap_unlock(vmf->pte, vmf->ptl);
2471 vmf->flags |= FAULT_FLAG_MKWRITE;
2472 ret = vma->vm_ops->pfn_mkwrite(vmf);
2473 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2475 return finish_mkwrite_fault(vmf);
2478 return VM_FAULT_WRITE;
2481 static int wp_page_shared(struct vm_fault *vmf)
2482 __releases(vmf->ptl)
2484 struct vm_area_struct *vma = vmf->vma;
2486 get_page(vmf->page);
2488 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2491 pte_unmap_unlock(vmf->pte, vmf->ptl);
2492 tmp = do_page_mkwrite(vmf);
2493 if (unlikely(!tmp || (tmp &
2494 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2495 put_page(vmf->page);
2498 tmp = finish_mkwrite_fault(vmf);
2499 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2500 unlock_page(vmf->page);
2501 put_page(vmf->page);
2506 lock_page(vmf->page);
2508 fault_dirty_shared_page(vma, vmf->page);
2509 put_page(vmf->page);
2511 return VM_FAULT_WRITE;
2515 * This routine handles present pages, when users try to write
2516 * to a shared page. It is done by copying the page to a new address
2517 * and decrementing the shared-page counter for the old page.
2519 * Note that this routine assumes that the protection checks have been
2520 * done by the caller (the low-level page fault routine in most cases).
2521 * Thus we can safely just mark it writable once we've done any necessary
2524 * We also mark the page dirty at this point even though the page will
2525 * change only once the write actually happens. This avoids a few races,
2526 * and potentially makes it more efficient.
2528 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2529 * but allow concurrent faults), with pte both mapped and locked.
2530 * We return with mmap_sem still held, but pte unmapped and unlocked.
2532 static int do_wp_page(struct vm_fault *vmf)
2533 __releases(vmf->ptl)
2535 struct vm_area_struct *vma = vmf->vma;
2537 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2540 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2543 * We should not cow pages in a shared writeable mapping.
2544 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2546 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2547 (VM_WRITE|VM_SHARED))
2548 return wp_pfn_shared(vmf);
2550 pte_unmap_unlock(vmf->pte, vmf->ptl);
2551 return wp_page_copy(vmf);
2555 * Take out anonymous pages first, anonymous shared vmas are
2556 * not dirty accountable.
2558 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2560 if (!trylock_page(vmf->page)) {
2561 get_page(vmf->page);
2562 pte_unmap_unlock(vmf->pte, vmf->ptl);
2563 lock_page(vmf->page);
2564 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2565 vmf->address, &vmf->ptl);
2566 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2567 unlock_page(vmf->page);
2568 pte_unmap_unlock(vmf->pte, vmf->ptl);
2569 put_page(vmf->page);
2572 put_page(vmf->page);
2574 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2575 if (total_mapcount == 1) {
2577 * The page is all ours. Move it to
2578 * our anon_vma so the rmap code will
2579 * not search our parent or siblings.
2580 * Protected against the rmap code by
2583 page_move_anon_rmap(vmf->page, vma);
2585 unlock_page(vmf->page);
2587 return VM_FAULT_WRITE;
2589 unlock_page(vmf->page);
2590 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2591 (VM_WRITE|VM_SHARED))) {
2592 return wp_page_shared(vmf);
2596 * Ok, we need to copy. Oh, well..
2598 get_page(vmf->page);
2600 pte_unmap_unlock(vmf->pte, vmf->ptl);
2601 return wp_page_copy(vmf);
2604 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2605 unsigned long start_addr, unsigned long end_addr,
2606 struct zap_details *details)
2608 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2611 static inline void unmap_mapping_range_tree(struct rb_root *root,
2612 struct zap_details *details)
2614 struct vm_area_struct *vma;
2615 pgoff_t vba, vea, zba, zea;
2617 vma_interval_tree_foreach(vma, root,
2618 details->first_index, details->last_index) {
2620 vba = vma->vm_pgoff;
2621 vea = vba + vma_pages(vma) - 1;
2622 zba = details->first_index;
2625 zea = details->last_index;
2629 unmap_mapping_range_vma(vma,
2630 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2631 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2637 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2638 * address_space corresponding to the specified page range in the underlying
2641 * @mapping: the address space containing mmaps to be unmapped.
2642 * @holebegin: byte in first page to unmap, relative to the start of
2643 * the underlying file. This will be rounded down to a PAGE_SIZE
2644 * boundary. Note that this is different from truncate_pagecache(), which
2645 * must keep the partial page. In contrast, we must get rid of
2647 * @holelen: size of prospective hole in bytes. This will be rounded
2648 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2650 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2651 * but 0 when invalidating pagecache, don't throw away private data.
2653 void unmap_mapping_range(struct address_space *mapping,
2654 loff_t const holebegin, loff_t const holelen, int even_cows)
2656 struct zap_details details = { };
2657 pgoff_t hba = holebegin >> PAGE_SHIFT;
2658 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2660 /* Check for overflow. */
2661 if (sizeof(holelen) > sizeof(hlen)) {
2663 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2664 if (holeend & ~(long long)ULONG_MAX)
2665 hlen = ULONG_MAX - hba + 1;
2668 details.check_mapping = even_cows ? NULL : mapping;
2669 details.first_index = hba;
2670 details.last_index = hba + hlen - 1;
2671 if (details.last_index < details.first_index)
2672 details.last_index = ULONG_MAX;
2674 i_mmap_lock_write(mapping);
2675 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2676 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2677 i_mmap_unlock_write(mapping);
2679 EXPORT_SYMBOL(unmap_mapping_range);
2682 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2683 * but allow concurrent faults), and pte mapped but not yet locked.
2684 * We return with pte unmapped and unlocked.
2686 * We return with the mmap_sem locked or unlocked in the same cases
2687 * as does filemap_fault().
2689 int do_swap_page(struct vm_fault *vmf)
2691 struct vm_area_struct *vma = vmf->vma;
2692 struct page *page, *swapcache;
2693 struct mem_cgroup *memcg;
2700 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2703 entry = pte_to_swp_entry(vmf->orig_pte);
2704 if (unlikely(non_swap_entry(entry))) {
2705 if (is_migration_entry(entry)) {
2706 migration_entry_wait(vma->vm_mm, vmf->pmd,
2708 } else if (is_hwpoison_entry(entry)) {
2709 ret = VM_FAULT_HWPOISON;
2711 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2712 ret = VM_FAULT_SIGBUS;
2716 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2717 page = lookup_swap_cache(entry);
2719 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2723 * Back out if somebody else faulted in this pte
2724 * while we released the pte lock.
2726 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2727 vmf->address, &vmf->ptl);
2728 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2730 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2734 /* Had to read the page from swap area: Major fault */
2735 ret = VM_FAULT_MAJOR;
2736 count_vm_event(PGMAJFAULT);
2737 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2738 } else if (PageHWPoison(page)) {
2740 * hwpoisoned dirty swapcache pages are kept for killing
2741 * owner processes (which may be unknown at hwpoison time)
2743 ret = VM_FAULT_HWPOISON;
2744 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2750 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2752 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2754 ret |= VM_FAULT_RETRY;
2759 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2760 * release the swapcache from under us. The page pin, and pte_same
2761 * test below, are not enough to exclude that. Even if it is still
2762 * swapcache, we need to check that the page's swap has not changed.
2764 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2767 page = ksm_might_need_to_copy(page, vma, vmf->address);
2768 if (unlikely(!page)) {
2774 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2781 * Back out if somebody else already faulted in this pte.
2783 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2785 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2788 if (unlikely(!PageUptodate(page))) {
2789 ret = VM_FAULT_SIGBUS;
2794 * The page isn't present yet, go ahead with the fault.
2796 * Be careful about the sequence of operations here.
2797 * To get its accounting right, reuse_swap_page() must be called
2798 * while the page is counted on swap but not yet in mapcount i.e.
2799 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2800 * must be called after the swap_free(), or it will never succeed.
2803 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2804 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2805 pte = mk_pte(page, vma->vm_page_prot);
2806 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2807 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2808 vmf->flags &= ~FAULT_FLAG_WRITE;
2809 ret |= VM_FAULT_WRITE;
2810 exclusive = RMAP_EXCLUSIVE;
2812 flush_icache_page(vma, page);
2813 if (pte_swp_soft_dirty(vmf->orig_pte))
2814 pte = pte_mksoft_dirty(pte);
2815 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2816 vmf->orig_pte = pte;
2817 if (page == swapcache) {
2818 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2819 mem_cgroup_commit_charge(page, memcg, true, false);
2820 activate_page(page);
2821 } else { /* ksm created a completely new copy */
2822 page_add_new_anon_rmap(page, vma, vmf->address, false);
2823 mem_cgroup_commit_charge(page, memcg, false, false);
2824 lru_cache_add_active_or_unevictable(page, vma);
2828 if (mem_cgroup_swap_full(page) ||
2829 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2830 try_to_free_swap(page);
2832 if (page != swapcache) {
2834 * Hold the lock to avoid the swap entry to be reused
2835 * until we take the PT lock for the pte_same() check
2836 * (to avoid false positives from pte_same). For
2837 * further safety release the lock after the swap_free
2838 * so that the swap count won't change under a
2839 * parallel locked swapcache.
2841 unlock_page(swapcache);
2842 put_page(swapcache);
2845 if (vmf->flags & FAULT_FLAG_WRITE) {
2846 ret |= do_wp_page(vmf);
2847 if (ret & VM_FAULT_ERROR)
2848 ret &= VM_FAULT_ERROR;
2852 /* No need to invalidate - it was non-present before */
2853 update_mmu_cache(vma, vmf->address, vmf->pte);
2855 pte_unmap_unlock(vmf->pte, vmf->ptl);
2859 mem_cgroup_cancel_charge(page, memcg, false);
2860 pte_unmap_unlock(vmf->pte, vmf->ptl);
2865 if (page != swapcache) {
2866 unlock_page(swapcache);
2867 put_page(swapcache);
2873 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2874 * but allow concurrent faults), and pte mapped but not yet locked.
2875 * We return with mmap_sem still held, but pte unmapped and unlocked.
2877 static int do_anonymous_page(struct vm_fault *vmf)
2879 struct vm_area_struct *vma = vmf->vma;
2880 struct mem_cgroup *memcg;
2884 /* File mapping without ->vm_ops ? */
2885 if (vma->vm_flags & VM_SHARED)
2886 return VM_FAULT_SIGBUS;
2889 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2890 * pte_offset_map() on pmds where a huge pmd might be created
2891 * from a different thread.
2893 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2894 * parallel threads are excluded by other means.
2896 * Here we only have down_read(mmap_sem).
2898 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2899 return VM_FAULT_OOM;
2901 /* See the comment in pte_alloc_one_map() */
2902 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2905 /* Use the zero-page for reads */
2906 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2907 !mm_forbids_zeropage(vma->vm_mm)) {
2908 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2909 vma->vm_page_prot));
2910 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2911 vmf->address, &vmf->ptl);
2912 if (!pte_none(*vmf->pte))
2914 /* Deliver the page fault to userland, check inside PT lock */
2915 if (userfaultfd_missing(vma)) {
2916 pte_unmap_unlock(vmf->pte, vmf->ptl);
2917 return handle_userfault(vmf, VM_UFFD_MISSING);
2922 /* Allocate our own private page. */
2923 if (unlikely(anon_vma_prepare(vma)))
2925 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2929 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2933 * The memory barrier inside __SetPageUptodate makes sure that
2934 * preceeding stores to the page contents become visible before
2935 * the set_pte_at() write.
2937 __SetPageUptodate(page);
2939 entry = mk_pte(page, vma->vm_page_prot);
2940 if (vma->vm_flags & VM_WRITE)
2941 entry = pte_mkwrite(pte_mkdirty(entry));
2943 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2945 if (!pte_none(*vmf->pte))
2948 /* Deliver the page fault to userland, check inside PT lock */
2949 if (userfaultfd_missing(vma)) {
2950 pte_unmap_unlock(vmf->pte, vmf->ptl);
2951 mem_cgroup_cancel_charge(page, memcg, false);
2953 return handle_userfault(vmf, VM_UFFD_MISSING);
2956 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2957 page_add_new_anon_rmap(page, vma, vmf->address, false);
2958 mem_cgroup_commit_charge(page, memcg, false, false);
2959 lru_cache_add_active_or_unevictable(page, vma);
2961 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2963 /* No need to invalidate - it was non-present before */
2964 update_mmu_cache(vma, vmf->address, vmf->pte);
2966 pte_unmap_unlock(vmf->pte, vmf->ptl);
2969 mem_cgroup_cancel_charge(page, memcg, false);
2975 return VM_FAULT_OOM;
2979 * The mmap_sem must have been held on entry, and may have been
2980 * released depending on flags and vma->vm_ops->fault() return value.
2981 * See filemap_fault() and __lock_page_retry().
2983 static int __do_fault(struct vm_fault *vmf)
2985 struct vm_area_struct *vma = vmf->vma;
2988 ret = vma->vm_ops->fault(vmf);
2989 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2990 VM_FAULT_DONE_COW)))
2993 if (unlikely(PageHWPoison(vmf->page))) {
2994 if (ret & VM_FAULT_LOCKED)
2995 unlock_page(vmf->page);
2996 put_page(vmf->page);
2998 return VM_FAULT_HWPOISON;
3001 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3002 lock_page(vmf->page);
3004 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3010 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3011 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3012 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3013 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3015 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3017 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3020 static int pte_alloc_one_map(struct vm_fault *vmf)
3022 struct vm_area_struct *vma = vmf->vma;
3024 if (!pmd_none(*vmf->pmd))
3026 if (vmf->prealloc_pte) {
3027 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3028 if (unlikely(!pmd_none(*vmf->pmd))) {
3029 spin_unlock(vmf->ptl);
3033 atomic_long_inc(&vma->vm_mm->nr_ptes);
3034 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3035 spin_unlock(vmf->ptl);
3036 vmf->prealloc_pte = NULL;
3037 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3038 return VM_FAULT_OOM;
3042 * If a huge pmd materialized under us just retry later. Use
3043 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3044 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3045 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3046 * running immediately after a huge pmd fault in a different thread of
3047 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3048 * All we have to ensure is that it is a regular pmd that we can walk
3049 * with pte_offset_map() and we can do that through an atomic read in
3050 * C, which is what pmd_trans_unstable() provides.
3052 if (pmd_devmap_trans_unstable(vmf->pmd))
3053 return VM_FAULT_NOPAGE;
3056 * At this point we know that our vmf->pmd points to a page of ptes
3057 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3058 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3059 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3060 * be valid and we will re-check to make sure the vmf->pte isn't
3061 * pte_none() under vmf->ptl protection when we return to
3064 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3069 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3071 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3072 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3073 unsigned long haddr)
3075 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3076 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3078 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3083 static void deposit_prealloc_pte(struct vm_fault *vmf)
3085 struct vm_area_struct *vma = vmf->vma;
3087 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3089 * We are going to consume the prealloc table,
3090 * count that as nr_ptes.
3092 atomic_long_inc(&vma->vm_mm->nr_ptes);
3093 vmf->prealloc_pte = NULL;
3096 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3098 struct vm_area_struct *vma = vmf->vma;
3099 bool write = vmf->flags & FAULT_FLAG_WRITE;
3100 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3104 if (!transhuge_vma_suitable(vma, haddr))
3105 return VM_FAULT_FALLBACK;
3107 ret = VM_FAULT_FALLBACK;
3108 page = compound_head(page);
3111 * Archs like ppc64 need additonal space to store information
3112 * related to pte entry. Use the preallocated table for that.
3114 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3115 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3116 if (!vmf->prealloc_pte)
3117 return VM_FAULT_OOM;
3118 smp_wmb(); /* See comment in __pte_alloc() */
3121 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3122 if (unlikely(!pmd_none(*vmf->pmd)))
3125 for (i = 0; i < HPAGE_PMD_NR; i++)
3126 flush_icache_page(vma, page + i);
3128 entry = mk_huge_pmd(page, vma->vm_page_prot);
3130 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3132 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3133 page_add_file_rmap(page, true);
3135 * deposit and withdraw with pmd lock held
3137 if (arch_needs_pgtable_deposit())
3138 deposit_prealloc_pte(vmf);
3140 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3142 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3144 /* fault is handled */
3146 count_vm_event(THP_FILE_MAPPED);
3148 spin_unlock(vmf->ptl);
3152 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3160 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3161 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3163 * @vmf: fault environment
3164 * @memcg: memcg to charge page (only for private mappings)
3165 * @page: page to map
3167 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3170 * Target users are page handler itself and implementations of
3171 * vm_ops->map_pages.
3173 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3176 struct vm_area_struct *vma = vmf->vma;
3177 bool write = vmf->flags & FAULT_FLAG_WRITE;
3181 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3182 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3184 VM_BUG_ON_PAGE(memcg, page);
3186 ret = do_set_pmd(vmf, page);
3187 if (ret != VM_FAULT_FALLBACK)
3192 ret = pte_alloc_one_map(vmf);
3197 /* Re-check under ptl */
3198 if (unlikely(!pte_none(*vmf->pte)))
3199 return VM_FAULT_NOPAGE;
3201 flush_icache_page(vma, page);
3202 entry = mk_pte(page, vma->vm_page_prot);
3204 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3205 /* copy-on-write page */
3206 if (write && !(vma->vm_flags & VM_SHARED)) {
3207 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3208 page_add_new_anon_rmap(page, vma, vmf->address, false);
3209 mem_cgroup_commit_charge(page, memcg, false, false);
3210 lru_cache_add_active_or_unevictable(page, vma);
3212 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3213 page_add_file_rmap(page, false);
3215 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3217 /* no need to invalidate: a not-present page won't be cached */
3218 update_mmu_cache(vma, vmf->address, vmf->pte);
3225 * finish_fault - finish page fault once we have prepared the page to fault
3227 * @vmf: structure describing the fault
3229 * This function handles all that is needed to finish a page fault once the
3230 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3231 * given page, adds reverse page mapping, handles memcg charges and LRU
3232 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3235 * The function expects the page to be locked and on success it consumes a
3236 * reference of a page being mapped (for the PTE which maps it).
3238 int finish_fault(struct vm_fault *vmf)
3243 /* Did we COW the page? */
3244 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3245 !(vmf->vma->vm_flags & VM_SHARED))
3246 page = vmf->cow_page;
3249 ret = alloc_set_pte(vmf, vmf->memcg, page);
3251 pte_unmap_unlock(vmf->pte, vmf->ptl);
3255 static unsigned long fault_around_bytes __read_mostly =
3256 rounddown_pow_of_two(65536);
3258 #ifdef CONFIG_DEBUG_FS
3259 static int fault_around_bytes_get(void *data, u64 *val)
3261 *val = fault_around_bytes;
3266 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3267 * rounded down to nearest page order. It's what do_fault_around() expects to
3270 static int fault_around_bytes_set(void *data, u64 val)
3272 if (val / PAGE_SIZE > PTRS_PER_PTE)
3274 if (val > PAGE_SIZE)
3275 fault_around_bytes = rounddown_pow_of_two(val);
3277 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3280 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3281 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3283 static int __init fault_around_debugfs(void)
3287 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3288 &fault_around_bytes_fops);
3290 pr_warn("Failed to create fault_around_bytes in debugfs");
3293 late_initcall(fault_around_debugfs);
3297 * do_fault_around() tries to map few pages around the fault address. The hope
3298 * is that the pages will be needed soon and this will lower the number of
3301 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3302 * not ready to be mapped: not up-to-date, locked, etc.
3304 * This function is called with the page table lock taken. In the split ptlock
3305 * case the page table lock only protects only those entries which belong to
3306 * the page table corresponding to the fault address.
3308 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3311 * fault_around_pages() defines how many pages we'll try to map.
3312 * do_fault_around() expects it to return a power of two less than or equal to
3315 * The virtual address of the area that we map is naturally aligned to the
3316 * fault_around_pages() value (and therefore to page order). This way it's
3317 * easier to guarantee that we don't cross page table boundaries.
3319 static int do_fault_around(struct vm_fault *vmf)
3321 unsigned long address = vmf->address, nr_pages, mask;
3322 pgoff_t start_pgoff = vmf->pgoff;
3326 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3327 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3329 vmf->address = max(address & mask, vmf->vma->vm_start);
3330 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3334 * end_pgoff is either end of page table or end of vma
3335 * or fault_around_pages() from start_pgoff, depending what is nearest.
3337 end_pgoff = start_pgoff -
3338 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3340 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3341 start_pgoff + nr_pages - 1);
3343 if (pmd_none(*vmf->pmd)) {
3344 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3346 if (!vmf->prealloc_pte)
3348 smp_wmb(); /* See comment in __pte_alloc() */
3351 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3353 /* Huge page is mapped? Page fault is solved */
3354 if (pmd_trans_huge(*vmf->pmd)) {
3355 ret = VM_FAULT_NOPAGE;
3359 /* ->map_pages() haven't done anything useful. Cold page cache? */
3363 /* check if the page fault is solved */
3364 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3365 if (!pte_none(*vmf->pte))
3366 ret = VM_FAULT_NOPAGE;
3367 pte_unmap_unlock(vmf->pte, vmf->ptl);
3369 vmf->address = address;
3374 static int do_read_fault(struct vm_fault *vmf)
3376 struct vm_area_struct *vma = vmf->vma;
3380 * Let's call ->map_pages() first and use ->fault() as fallback
3381 * if page by the offset is not ready to be mapped (cold cache or
3384 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3385 ret = do_fault_around(vmf);
3390 ret = __do_fault(vmf);
3391 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3394 ret |= finish_fault(vmf);
3395 unlock_page(vmf->page);
3396 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3397 put_page(vmf->page);
3401 static int do_cow_fault(struct vm_fault *vmf)
3403 struct vm_area_struct *vma = vmf->vma;
3406 if (unlikely(anon_vma_prepare(vma)))
3407 return VM_FAULT_OOM;
3409 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3411 return VM_FAULT_OOM;
3413 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3414 &vmf->memcg, false)) {
3415 put_page(vmf->cow_page);
3416 return VM_FAULT_OOM;
3419 ret = __do_fault(vmf);
3420 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3422 if (ret & VM_FAULT_DONE_COW)
3425 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3426 __SetPageUptodate(vmf->cow_page);
3428 ret |= finish_fault(vmf);
3429 unlock_page(vmf->page);
3430 put_page(vmf->page);
3431 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3435 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3436 put_page(vmf->cow_page);
3440 static int do_shared_fault(struct vm_fault *vmf)
3442 struct vm_area_struct *vma = vmf->vma;
3445 ret = __do_fault(vmf);
3446 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3450 * Check if the backing address space wants to know that the page is
3451 * about to become writable
3453 if (vma->vm_ops->page_mkwrite) {
3454 unlock_page(vmf->page);
3455 tmp = do_page_mkwrite(vmf);
3456 if (unlikely(!tmp ||
3457 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3458 put_page(vmf->page);
3463 ret |= finish_fault(vmf);
3464 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3466 unlock_page(vmf->page);
3467 put_page(vmf->page);
3471 fault_dirty_shared_page(vma, vmf->page);
3476 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3477 * but allow concurrent faults).
3478 * The mmap_sem may have been released depending on flags and our
3479 * return value. See filemap_fault() and __lock_page_or_retry().
3481 static int do_fault(struct vm_fault *vmf)
3483 struct vm_area_struct *vma = vmf->vma;
3486 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3487 if (!vma->vm_ops->fault)
3488 ret = VM_FAULT_SIGBUS;
3489 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3490 ret = do_read_fault(vmf);
3491 else if (!(vma->vm_flags & VM_SHARED))
3492 ret = do_cow_fault(vmf);
3494 ret = do_shared_fault(vmf);
3496 /* preallocated pagetable is unused: free it */
3497 if (vmf->prealloc_pte) {
3498 pte_free(vma->vm_mm, vmf->prealloc_pte);
3499 vmf->prealloc_pte = NULL;
3504 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3505 unsigned long addr, int page_nid,
3510 count_vm_numa_event(NUMA_HINT_FAULTS);
3511 if (page_nid == numa_node_id()) {
3512 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3513 *flags |= TNF_FAULT_LOCAL;
3516 return mpol_misplaced(page, vma, addr);
3519 static int do_numa_page(struct vm_fault *vmf)
3521 struct vm_area_struct *vma = vmf->vma;
3522 struct page *page = NULL;
3526 bool migrated = false;
3528 bool was_writable = pte_savedwrite(vmf->orig_pte);
3532 * The "pte" at this point cannot be used safely without
3533 * validation through pte_unmap_same(). It's of NUMA type but
3534 * the pfn may be screwed if the read is non atomic.
3536 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3537 spin_lock(vmf->ptl);
3538 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3539 pte_unmap_unlock(vmf->pte, vmf->ptl);
3544 * Make it present again, Depending on how arch implementes non
3545 * accessible ptes, some can allow access by kernel mode.
3547 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3548 pte = pte_modify(pte, vma->vm_page_prot);
3549 pte = pte_mkyoung(pte);
3551 pte = pte_mkwrite(pte);
3552 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3553 update_mmu_cache(vma, vmf->address, vmf->pte);
3555 page = vm_normal_page(vma, vmf->address, pte);
3557 pte_unmap_unlock(vmf->pte, vmf->ptl);
3561 /* TODO: handle PTE-mapped THP */
3562 if (PageCompound(page)) {
3563 pte_unmap_unlock(vmf->pte, vmf->ptl);
3568 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3569 * much anyway since they can be in shared cache state. This misses
3570 * the case where a mapping is writable but the process never writes
3571 * to it but pte_write gets cleared during protection updates and
3572 * pte_dirty has unpredictable behaviour between PTE scan updates,
3573 * background writeback, dirty balancing and application behaviour.
3575 if (!pte_write(pte))
3576 flags |= TNF_NO_GROUP;
3579 * Flag if the page is shared between multiple address spaces. This
3580 * is later used when determining whether to group tasks together
3582 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3583 flags |= TNF_SHARED;
3585 last_cpupid = page_cpupid_last(page);
3586 page_nid = page_to_nid(page);
3587 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3589 pte_unmap_unlock(vmf->pte, vmf->ptl);
3590 if (target_nid == -1) {
3595 /* Migrate to the requested node */
3596 migrated = migrate_misplaced_page(page, vma, target_nid);
3598 page_nid = target_nid;
3599 flags |= TNF_MIGRATED;
3601 flags |= TNF_MIGRATE_FAIL;
3605 task_numa_fault(last_cpupid, page_nid, 1, flags);
3609 static inline int create_huge_pmd(struct vm_fault *vmf)
3611 if (vma_is_anonymous(vmf->vma))
3612 return do_huge_pmd_anonymous_page(vmf);
3613 if (vmf->vma->vm_ops->huge_fault)
3614 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3615 return VM_FAULT_FALLBACK;
3618 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3620 if (vma_is_anonymous(vmf->vma))
3621 return do_huge_pmd_wp_page(vmf, orig_pmd);
3622 if (vmf->vma->vm_ops->huge_fault)
3623 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3625 /* COW handled on pte level: split pmd */
3626 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3627 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3629 return VM_FAULT_FALLBACK;
3632 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3634 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3637 static int create_huge_pud(struct vm_fault *vmf)
3639 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3640 /* No support for anonymous transparent PUD pages yet */
3641 if (vma_is_anonymous(vmf->vma))
3642 return VM_FAULT_FALLBACK;
3643 if (vmf->vma->vm_ops->huge_fault)
3644 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3645 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3646 return VM_FAULT_FALLBACK;
3649 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3651 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3652 /* No support for anonymous transparent PUD pages yet */
3653 if (vma_is_anonymous(vmf->vma))
3654 return VM_FAULT_FALLBACK;
3655 if (vmf->vma->vm_ops->huge_fault)
3656 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3657 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3658 return VM_FAULT_FALLBACK;
3662 * These routines also need to handle stuff like marking pages dirty
3663 * and/or accessed for architectures that don't do it in hardware (most
3664 * RISC architectures). The early dirtying is also good on the i386.
3666 * There is also a hook called "update_mmu_cache()" that architectures
3667 * with external mmu caches can use to update those (ie the Sparc or
3668 * PowerPC hashed page tables that act as extended TLBs).
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3671 * concurrent faults).
3673 * The mmap_sem may have been released depending on flags and our return value.
3674 * See filemap_fault() and __lock_page_or_retry().
3676 static int handle_pte_fault(struct vm_fault *vmf)
3680 if (unlikely(pmd_none(*vmf->pmd))) {
3682 * Leave __pte_alloc() until later: because vm_ops->fault may
3683 * want to allocate huge page, and if we expose page table
3684 * for an instant, it will be difficult to retract from
3685 * concurrent faults and from rmap lookups.
3689 /* See comment in pte_alloc_one_map() */
3690 if (pmd_devmap_trans_unstable(vmf->pmd))
3693 * A regular pmd is established and it can't morph into a huge
3694 * pmd from under us anymore at this point because we hold the
3695 * mmap_sem read mode and khugepaged takes it in write mode.
3696 * So now it's safe to run pte_offset_map().
3698 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3699 vmf->orig_pte = *vmf->pte;
3702 * some architectures can have larger ptes than wordsize,
3703 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3704 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3705 * atomic accesses. The code below just needs a consistent
3706 * view for the ifs and we later double check anyway with the
3707 * ptl lock held. So here a barrier will do.
3710 if (pte_none(vmf->orig_pte)) {
3711 pte_unmap(vmf->pte);
3717 if (vma_is_anonymous(vmf->vma))
3718 return do_anonymous_page(vmf);
3720 return do_fault(vmf);
3723 if (!pte_present(vmf->orig_pte))
3724 return do_swap_page(vmf);
3726 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3727 return do_numa_page(vmf);
3729 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3730 spin_lock(vmf->ptl);
3731 entry = vmf->orig_pte;
3732 if (unlikely(!pte_same(*vmf->pte, entry)))
3734 if (vmf->flags & FAULT_FLAG_WRITE) {
3735 if (!pte_write(entry))
3736 return do_wp_page(vmf);
3737 entry = pte_mkdirty(entry);
3739 entry = pte_mkyoung(entry);
3740 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3741 vmf->flags & FAULT_FLAG_WRITE)) {
3742 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3745 * This is needed only for protection faults but the arch code
3746 * is not yet telling us if this is a protection fault or not.
3747 * This still avoids useless tlb flushes for .text page faults
3750 if (vmf->flags & FAULT_FLAG_WRITE)
3751 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3754 pte_unmap_unlock(vmf->pte, vmf->ptl);
3759 * By the time we get here, we already hold the mm semaphore
3761 * The mmap_sem may have been released depending on flags and our
3762 * return value. See filemap_fault() and __lock_page_or_retry().
3764 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3767 struct vm_fault vmf = {
3769 .address = address & PAGE_MASK,
3771 .pgoff = linear_page_index(vma, address),
3772 .gfp_mask = __get_fault_gfp_mask(vma),
3774 struct mm_struct *mm = vma->vm_mm;
3779 pgd = pgd_offset(mm, address);
3780 p4d = p4d_alloc(mm, pgd, address);
3782 return VM_FAULT_OOM;
3784 vmf.pud = pud_alloc(mm, p4d, address);
3786 return VM_FAULT_OOM;
3787 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3788 ret = create_huge_pud(&vmf);
3789 if (!(ret & VM_FAULT_FALLBACK))
3792 pud_t orig_pud = *vmf.pud;
3795 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3796 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3798 /* NUMA case for anonymous PUDs would go here */
3800 if (dirty && !pud_write(orig_pud)) {
3801 ret = wp_huge_pud(&vmf, orig_pud);
3802 if (!(ret & VM_FAULT_FALLBACK))
3805 huge_pud_set_accessed(&vmf, orig_pud);
3811 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3813 return VM_FAULT_OOM;
3814 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3815 ret = create_huge_pmd(&vmf);
3816 if (!(ret & VM_FAULT_FALLBACK))
3819 pmd_t orig_pmd = *vmf.pmd;
3822 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3823 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3824 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3826 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3827 !pmd_write(orig_pmd)) {
3828 ret = wp_huge_pmd(&vmf, orig_pmd);
3829 if (!(ret & VM_FAULT_FALLBACK))
3832 huge_pmd_set_accessed(&vmf, orig_pmd);
3838 return handle_pte_fault(&vmf);
3842 * By the time we get here, we already hold the mm semaphore
3844 * The mmap_sem may have been released depending on flags and our
3845 * return value. See filemap_fault() and __lock_page_or_retry().
3847 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3852 __set_current_state(TASK_RUNNING);
3854 count_vm_event(PGFAULT);
3855 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3857 /* do counter updates before entering really critical section. */
3858 check_sync_rss_stat(current);
3861 * Enable the memcg OOM handling for faults triggered in user
3862 * space. Kernel faults are handled more gracefully.
3864 if (flags & FAULT_FLAG_USER)
3865 mem_cgroup_oom_enable();
3867 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3868 flags & FAULT_FLAG_INSTRUCTION,
3869 flags & FAULT_FLAG_REMOTE))
3870 return VM_FAULT_SIGSEGV;
3872 if (unlikely(is_vm_hugetlb_page(vma)))
3873 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3875 ret = __handle_mm_fault(vma, address, flags);
3877 if (flags & FAULT_FLAG_USER) {
3878 mem_cgroup_oom_disable();
3880 * The task may have entered a memcg OOM situation but
3881 * if the allocation error was handled gracefully (no
3882 * VM_FAULT_OOM), there is no need to kill anything.
3883 * Just clean up the OOM state peacefully.
3885 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3886 mem_cgroup_oom_synchronize(false);
3890 * This mm has been already reaped by the oom reaper and so the
3891 * refault cannot be trusted in general. Anonymous refaults would
3892 * lose data and give a zero page instead e.g. This is especially
3893 * problem for use_mm() because regular tasks will just die and
3894 * the corrupted data will not be visible anywhere while kthread
3895 * will outlive the oom victim and potentially propagate the date
3898 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3899 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3900 ret = VM_FAULT_SIGBUS;
3904 EXPORT_SYMBOL_GPL(handle_mm_fault);
3906 #ifndef __PAGETABLE_P4D_FOLDED
3908 * Allocate p4d page table.
3909 * We've already handled the fast-path in-line.
3911 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3913 p4d_t *new = p4d_alloc_one(mm, address);
3917 smp_wmb(); /* See comment in __pte_alloc */
3919 spin_lock(&mm->page_table_lock);
3920 if (pgd_present(*pgd)) /* Another has populated it */
3923 pgd_populate(mm, pgd, new);
3924 spin_unlock(&mm->page_table_lock);
3927 #endif /* __PAGETABLE_P4D_FOLDED */
3929 #ifndef __PAGETABLE_PUD_FOLDED
3931 * Allocate page upper directory.
3932 * We've already handled the fast-path in-line.
3934 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
3936 pud_t *new = pud_alloc_one(mm, address);
3940 smp_wmb(); /* See comment in __pte_alloc */
3942 spin_lock(&mm->page_table_lock);
3943 #ifndef __ARCH_HAS_5LEVEL_HACK
3944 if (p4d_present(*p4d)) /* Another has populated it */
3947 p4d_populate(mm, p4d, new);
3949 if (pgd_present(*p4d)) /* Another has populated it */
3952 pgd_populate(mm, p4d, new);
3953 #endif /* __ARCH_HAS_5LEVEL_HACK */
3954 spin_unlock(&mm->page_table_lock);
3957 #endif /* __PAGETABLE_PUD_FOLDED */
3959 #ifndef __PAGETABLE_PMD_FOLDED
3961 * Allocate page middle directory.
3962 * We've already handled the fast-path in-line.
3964 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3967 pmd_t *new = pmd_alloc_one(mm, address);
3971 smp_wmb(); /* See comment in __pte_alloc */
3973 ptl = pud_lock(mm, pud);
3974 #ifndef __ARCH_HAS_4LEVEL_HACK
3975 if (!pud_present(*pud)) {
3977 pud_populate(mm, pud, new);
3978 } else /* Another has populated it */
3981 if (!pgd_present(*pud)) {
3983 pgd_populate(mm, pud, new);
3984 } else /* Another has populated it */
3986 #endif /* __ARCH_HAS_4LEVEL_HACK */
3990 #endif /* __PAGETABLE_PMD_FOLDED */
3992 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3993 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4001 pgd = pgd_offset(mm, address);
4002 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4005 p4d = p4d_offset(pgd, address);
4006 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4009 pud = pud_offset(p4d, address);
4010 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4013 pmd = pmd_offset(pud, address);
4014 VM_BUG_ON(pmd_trans_huge(*pmd));
4016 if (pmd_huge(*pmd)) {
4020 *ptlp = pmd_lock(mm, pmd);
4021 if (pmd_huge(*pmd)) {
4028 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4031 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4032 if (!pte_present(*ptep))
4037 pte_unmap_unlock(ptep, *ptlp);
4042 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4043 pte_t **ptepp, spinlock_t **ptlp)
4047 /* (void) is needed to make gcc happy */
4048 (void) __cond_lock(*ptlp,
4049 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
4054 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4055 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4059 /* (void) is needed to make gcc happy */
4060 (void) __cond_lock(*ptlp,
4061 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
4065 EXPORT_SYMBOL(follow_pte_pmd);
4068 * follow_pfn - look up PFN at a user virtual address
4069 * @vma: memory mapping
4070 * @address: user virtual address
4071 * @pfn: location to store found PFN
4073 * Only IO mappings and raw PFN mappings are allowed.
4075 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4077 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4084 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4087 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4090 *pfn = pte_pfn(*ptep);
4091 pte_unmap_unlock(ptep, ptl);
4094 EXPORT_SYMBOL(follow_pfn);
4096 #ifdef CONFIG_HAVE_IOREMAP_PROT
4097 int follow_phys(struct vm_area_struct *vma,
4098 unsigned long address, unsigned int flags,
4099 unsigned long *prot, resource_size_t *phys)
4105 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4108 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4112 if ((flags & FOLL_WRITE) && !pte_write(pte))
4115 *prot = pgprot_val(pte_pgprot(pte));
4116 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4120 pte_unmap_unlock(ptep, ptl);
4125 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4126 void *buf, int len, int write)
4128 resource_size_t phys_addr;
4129 unsigned long prot = 0;
4130 void __iomem *maddr;
4131 int offset = addr & (PAGE_SIZE-1);
4133 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4136 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4138 memcpy_toio(maddr + offset, buf, len);
4140 memcpy_fromio(buf, maddr + offset, len);
4145 EXPORT_SYMBOL_GPL(generic_access_phys);
4149 * Access another process' address space as given in mm. If non-NULL, use the
4150 * given task for page fault accounting.
4152 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4153 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4155 struct vm_area_struct *vma;
4156 void *old_buf = buf;
4157 int write = gup_flags & FOLL_WRITE;
4159 down_read(&mm->mmap_sem);
4160 /* ignore errors, just check how much was successfully transferred */
4162 int bytes, ret, offset;
4164 struct page *page = NULL;
4166 ret = get_user_pages_remote(tsk, mm, addr, 1,
4167 gup_flags, &page, &vma, NULL);
4169 #ifndef CONFIG_HAVE_IOREMAP_PROT
4173 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4174 * we can access using slightly different code.
4176 vma = find_vma(mm, addr);
4177 if (!vma || vma->vm_start > addr)
4179 if (vma->vm_ops && vma->vm_ops->access)
4180 ret = vma->vm_ops->access(vma, addr, buf,
4188 offset = addr & (PAGE_SIZE-1);
4189 if (bytes > PAGE_SIZE-offset)
4190 bytes = PAGE_SIZE-offset;
4194 copy_to_user_page(vma, page, addr,
4195 maddr + offset, buf, bytes);
4196 set_page_dirty_lock(page);
4198 copy_from_user_page(vma, page, addr,
4199 buf, maddr + offset, bytes);
4208 up_read(&mm->mmap_sem);
4210 return buf - old_buf;
4214 * access_remote_vm - access another process' address space
4215 * @mm: the mm_struct of the target address space
4216 * @addr: start address to access
4217 * @buf: source or destination buffer
4218 * @len: number of bytes to transfer
4219 * @gup_flags: flags modifying lookup behaviour
4221 * The caller must hold a reference on @mm.
4223 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4224 void *buf, int len, unsigned int gup_flags)
4226 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4230 * Access another process' address space.
4231 * Source/target buffer must be kernel space,
4232 * Do not walk the page table directly, use get_user_pages
4234 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4235 void *buf, int len, unsigned int gup_flags)
4237 struct mm_struct *mm;
4240 mm = get_task_mm(tsk);
4244 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4250 EXPORT_SYMBOL_GPL(access_process_vm);
4253 * Print the name of a VMA.
4255 void print_vma_addr(char *prefix, unsigned long ip)
4257 struct mm_struct *mm = current->mm;
4258 struct vm_area_struct *vma;
4261 * Do not print if we are in atomic
4262 * contexts (in exception stacks, etc.):
4264 if (preempt_count())
4267 down_read(&mm->mmap_sem);
4268 vma = find_vma(mm, ip);
4269 if (vma && vma->vm_file) {
4270 struct file *f = vma->vm_file;
4271 char *buf = (char *)__get_free_page(GFP_KERNEL);
4275 p = file_path(f, buf, PAGE_SIZE);
4278 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4280 vma->vm_end - vma->vm_start);
4281 free_page((unsigned long)buf);
4284 up_read(&mm->mmap_sem);
4287 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4288 void __might_fault(const char *file, int line)
4291 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4292 * holding the mmap_sem, this is safe because kernel memory doesn't
4293 * get paged out, therefore we'll never actually fault, and the
4294 * below annotations will generate false positives.
4296 if (uaccess_kernel())
4298 if (pagefault_disabled())
4300 __might_sleep(file, line, 0);
4301 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4303 might_lock_read(¤t->mm->mmap_sem);
4306 EXPORT_SYMBOL(__might_fault);
4309 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4310 static void clear_gigantic_page(struct page *page,
4312 unsigned int pages_per_huge_page)
4315 struct page *p = page;
4318 for (i = 0; i < pages_per_huge_page;
4319 i++, p = mem_map_next(p, page, i)) {
4321 clear_user_highpage(p, addr + i * PAGE_SIZE);
4324 void clear_huge_page(struct page *page,
4325 unsigned long addr, unsigned int pages_per_huge_page)
4329 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4330 clear_gigantic_page(page, addr, pages_per_huge_page);
4335 for (i = 0; i < pages_per_huge_page; i++) {
4337 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4341 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4343 struct vm_area_struct *vma,
4344 unsigned int pages_per_huge_page)
4347 struct page *dst_base = dst;
4348 struct page *src_base = src;
4350 for (i = 0; i < pages_per_huge_page; ) {
4352 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4355 dst = mem_map_next(dst, dst_base, i);
4356 src = mem_map_next(src, src_base, i);
4360 void copy_user_huge_page(struct page *dst, struct page *src,
4361 unsigned long addr, struct vm_area_struct *vma,
4362 unsigned int pages_per_huge_page)
4366 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4367 copy_user_gigantic_page(dst, src, addr, vma,
4368 pages_per_huge_page);
4373 for (i = 0; i < pages_per_huge_page; i++) {
4375 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4379 long copy_huge_page_from_user(struct page *dst_page,
4380 const void __user *usr_src,
4381 unsigned int pages_per_huge_page,
4382 bool allow_pagefault)
4384 void *src = (void *)usr_src;
4386 unsigned long i, rc = 0;
4387 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4389 for (i = 0; i < pages_per_huge_page; i++) {
4390 if (allow_pagefault)
4391 page_kaddr = kmap(dst_page + i);
4393 page_kaddr = kmap_atomic(dst_page + i);
4394 rc = copy_from_user(page_kaddr,
4395 (const void __user *)(src + i * PAGE_SIZE),
4397 if (allow_pagefault)
4398 kunmap(dst_page + i);
4400 kunmap_atomic(page_kaddr);
4402 ret_val -= (PAGE_SIZE - rc);
4410 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4412 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4414 static struct kmem_cache *page_ptl_cachep;
4416 void __init ptlock_cache_init(void)
4418 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4422 bool ptlock_alloc(struct page *page)
4426 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4433 void ptlock_free(struct page *page)
4435 kmem_cache_free(page_ptl_cachep, page->ptl);