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/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr;
83 EXPORT_SYMBOL(max_mapnr);
84 EXPORT_SYMBOL(mem_map);
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 EXPORT_SYMBOL(high_memory);
99 * Randomize the address space (stacks, mmaps, brk, etc.).
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
104 int randomize_va_space __read_mostly =
105 #ifdef CONFIG_COMPAT_BRK
111 static int __init disable_randmaps(char *s)
113 randomize_va_space = 0;
116 __setup("norandmaps", disable_randmaps);
118 unsigned long zero_pfn __read_mostly;
119 unsigned long highest_memmap_pfn __read_mostly;
121 EXPORT_SYMBOL(zero_pfn);
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126 static int __init init_zero_pfn(void)
128 zero_pfn = page_to_pfn(ZERO_PAGE(0));
131 core_initcall(init_zero_pfn);
134 #if defined(SPLIT_RSS_COUNTING)
136 void sync_mm_rss(struct mm_struct *mm)
140 for (i = 0; i < NR_MM_COUNTERS; i++) {
141 if (current->rss_stat.count[i]) {
142 add_mm_counter(mm, i, current->rss_stat.count[i]);
143 current->rss_stat.count[i] = 0;
146 current->rss_stat.events = 0;
149 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
151 struct task_struct *task = current;
153 if (likely(task->mm == mm))
154 task->rss_stat.count[member] += val;
156 add_mm_counter(mm, member, val);
158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161 /* sync counter once per 64 page faults */
162 #define TASK_RSS_EVENTS_THRESH (64)
163 static void check_sync_rss_stat(struct task_struct *task)
165 if (unlikely(task != current))
167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
168 sync_mm_rss(task->mm);
170 #else /* SPLIT_RSS_COUNTING */
172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175 static void check_sync_rss_stat(struct task_struct *task)
179 #endif /* SPLIT_RSS_COUNTING */
181 #ifdef HAVE_GENERIC_MMU_GATHER
183 static int tlb_next_batch(struct mmu_gather *tlb)
185 struct mmu_gather_batch *batch;
189 tlb->active = batch->next;
193 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
196 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
203 batch->max = MAX_GATHER_BATCH;
205 tlb->active->next = batch;
212 * Called to initialize an (on-stack) mmu_gather structure for page-table
213 * tear-down from @mm. The @fullmm argument is used when @mm is without
214 * users and we're going to destroy the full address space (exit/execve).
216 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
220 /* Is it from 0 to ~0? */
221 tlb->fullmm = !(start | (end+1));
222 tlb->need_flush_all = 0;
223 tlb->local.next = NULL;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
227 tlb->batch_count = 0;
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 __tlb_reset_range(tlb);
236 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
243 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb);
246 __tlb_reset_range(tlb);
249 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
251 struct mmu_gather_batch *batch;
253 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
254 free_pages_and_swap_cache(batch->pages, batch->nr);
257 tlb->active = &tlb->local;
260 void tlb_flush_mmu(struct mmu_gather *tlb)
262 tlb_flush_mmu_tlbonly(tlb);
263 tlb_flush_mmu_free(tlb);
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
272 struct mmu_gather_batch *batch, *next;
276 /* keep the page table cache within bounds */
279 for (batch = tlb->local.next; batch; batch = next) {
281 free_pages((unsigned long)batch, 0);
283 tlb->local.next = NULL;
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
294 struct mmu_gather_batch *batch;
296 VM_BUG_ON(!tlb->end);
299 batch->pages[batch->nr++] = page;
300 if (batch->nr == batch->max) {
301 if (!tlb_next_batch(tlb))
305 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
307 return batch->max - batch->nr;
310 #endif /* HAVE_GENERIC_MMU_GATHER */
312 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
315 * See the comment near struct mmu_table_batch.
318 static void tlb_remove_table_smp_sync(void *arg)
320 /* Simply deliver the interrupt */
323 static void tlb_remove_table_one(void *table)
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
329 * It is however sufficient for software page-table walkers that rely on
330 * IRQ disabling. See the comment near struct mmu_table_batch.
332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
333 __tlb_remove_table(table);
336 static void tlb_remove_table_rcu(struct rcu_head *head)
338 struct mmu_table_batch *batch;
341 batch = container_of(head, struct mmu_table_batch, rcu);
343 for (i = 0; i < batch->nr; i++)
344 __tlb_remove_table(batch->tables[i]);
346 free_page((unsigned long)batch);
349 void tlb_table_flush(struct mmu_gather *tlb)
351 struct mmu_table_batch **batch = &tlb->batch;
354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
359 void tlb_remove_table(struct mmu_gather *tlb, void *table)
361 struct mmu_table_batch **batch = &tlb->batch;
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
367 if (atomic_read(&tlb->mm->mm_users) < 2) {
368 __tlb_remove_table(table);
372 if (*batch == NULL) {
373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
374 if (*batch == NULL) {
375 tlb_remove_table_one(table);
380 (*batch)->tables[(*batch)->nr++] = table;
381 if ((*batch)->nr == MAX_TABLE_BATCH)
382 tlb_table_flush(tlb);
385 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
388 * Note: this doesn't free the actual pages themselves. That
389 * has been handled earlier when unmapping all the memory regions.
391 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
394 pgtable_t token = pmd_pgtable(*pmd);
396 pte_free_tlb(tlb, token, addr);
397 atomic_long_dec(&tlb->mm->nr_ptes);
400 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
401 unsigned long addr, unsigned long end,
402 unsigned long floor, unsigned long ceiling)
409 pmd = pmd_offset(pud, addr);
411 next = pmd_addr_end(addr, end);
412 if (pmd_none_or_clear_bad(pmd))
414 free_pte_range(tlb, pmd, addr);
415 } while (pmd++, addr = next, addr != end);
425 if (end - 1 > ceiling - 1)
428 pmd = pmd_offset(pud, start);
430 pmd_free_tlb(tlb, pmd, start);
433 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
434 unsigned long addr, unsigned long end,
435 unsigned long floor, unsigned long ceiling)
442 pud = pud_offset(pgd, addr);
444 next = pud_addr_end(addr, end);
445 if (pud_none_or_clear_bad(pud))
447 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
448 } while (pud++, addr = next, addr != end);
454 ceiling &= PGDIR_MASK;
458 if (end - 1 > ceiling - 1)
461 pud = pud_offset(pgd, start);
463 pud_free_tlb(tlb, pud, start);
467 * This function frees user-level page tables of a process.
469 void free_pgd_range(struct mmu_gather *tlb,
470 unsigned long addr, unsigned long end,
471 unsigned long floor, unsigned long ceiling)
477 * The next few lines have given us lots of grief...
479 * Why are we testing PMD* at this top level? Because often
480 * there will be no work to do at all, and we'd prefer not to
481 * go all the way down to the bottom just to discover that.
483 * Why all these "- 1"s? Because 0 represents both the bottom
484 * of the address space and the top of it (using -1 for the
485 * top wouldn't help much: the masks would do the wrong thing).
486 * The rule is that addr 0 and floor 0 refer to the bottom of
487 * the address space, but end 0 and ceiling 0 refer to the top
488 * Comparisons need to use "end - 1" and "ceiling - 1" (though
489 * that end 0 case should be mythical).
491 * Wherever addr is brought up or ceiling brought down, we must
492 * be careful to reject "the opposite 0" before it confuses the
493 * subsequent tests. But what about where end is brought down
494 * by PMD_SIZE below? no, end can't go down to 0 there.
496 * Whereas we round start (addr) and ceiling down, by different
497 * masks at different levels, in order to test whether a table
498 * now has no other vmas using it, so can be freed, we don't
499 * bother to round floor or end up - the tests don't need that.
513 if (end - 1 > ceiling - 1)
518 pgd = pgd_offset(tlb->mm, addr);
520 next = pgd_addr_end(addr, end);
521 if (pgd_none_or_clear_bad(pgd))
523 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
524 } while (pgd++, addr = next, addr != end);
527 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
528 unsigned long floor, unsigned long ceiling)
531 struct vm_area_struct *next = vma->vm_next;
532 unsigned long addr = vma->vm_start;
535 * Hide vma from rmap and truncate_pagecache before freeing
538 unlink_anon_vmas(vma);
539 unlink_file_vma(vma);
541 if (is_vm_hugetlb_page(vma)) {
542 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
543 floor, next? next->vm_start: ceiling);
546 * Optimization: gather nearby vmas into one call down
548 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
549 && !is_vm_hugetlb_page(next)) {
552 unlink_anon_vmas(vma);
553 unlink_file_vma(vma);
555 free_pgd_range(tlb, addr, vma->vm_end,
556 floor, next? next->vm_start: ceiling);
562 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
563 pmd_t *pmd, unsigned long address)
566 pgtable_t new = pte_alloc_one(mm, address);
567 int wait_split_huge_page;
572 * Ensure all pte setup (eg. pte page lock and page clearing) are
573 * visible before the pte is made visible to other CPUs by being
574 * put into page tables.
576 * The other side of the story is the pointer chasing in the page
577 * table walking code (when walking the page table without locking;
578 * ie. most of the time). Fortunately, these data accesses consist
579 * of a chain of data-dependent loads, meaning most CPUs (alpha
580 * being the notable exception) will already guarantee loads are
581 * seen in-order. See the alpha page table accessors for the
582 * smp_read_barrier_depends() barriers in page table walking code.
584 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
586 ptl = pmd_lock(mm, pmd);
587 wait_split_huge_page = 0;
588 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
589 atomic_long_inc(&mm->nr_ptes);
590 pmd_populate(mm, pmd, new);
592 } else if (unlikely(pmd_trans_splitting(*pmd)))
593 wait_split_huge_page = 1;
597 if (wait_split_huge_page)
598 wait_split_huge_page(vma->anon_vma, pmd);
602 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
604 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
608 smp_wmb(); /* See comment in __pte_alloc */
610 spin_lock(&init_mm.page_table_lock);
611 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
612 pmd_populate_kernel(&init_mm, pmd, new);
615 VM_BUG_ON(pmd_trans_splitting(*pmd));
616 spin_unlock(&init_mm.page_table_lock);
618 pte_free_kernel(&init_mm, new);
622 static inline void init_rss_vec(int *rss)
624 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
627 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
631 if (current->mm == mm)
633 for (i = 0; i < NR_MM_COUNTERS; i++)
635 add_mm_counter(mm, i, rss[i]);
639 * This function is called to print an error when a bad pte
640 * is found. For example, we might have a PFN-mapped pte in
641 * a region that doesn't allow it.
643 * The calling function must still handle the error.
645 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
646 pte_t pte, struct page *page)
648 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
649 pud_t *pud = pud_offset(pgd, addr);
650 pmd_t *pmd = pmd_offset(pud, addr);
651 struct address_space *mapping;
653 static unsigned long resume;
654 static unsigned long nr_shown;
655 static unsigned long nr_unshown;
658 * Allow a burst of 60 reports, then keep quiet for that minute;
659 * or allow a steady drip of one report per second.
661 if (nr_shown == 60) {
662 if (time_before(jiffies, resume)) {
668 "BUG: Bad page map: %lu messages suppressed\n",
675 resume = jiffies + 60 * HZ;
677 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
678 index = linear_page_index(vma, addr);
681 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
683 (long long)pte_val(pte), (long long)pmd_val(*pmd));
685 dump_page(page, "bad pte");
687 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
688 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
690 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
693 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
696 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
697 vma->vm_file->f_op->mmap);
699 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
703 * vm_normal_page -- This function gets the "struct page" associated with a pte.
705 * "Special" mappings do not wish to be associated with a "struct page" (either
706 * it doesn't exist, or it exists but they don't want to touch it). In this
707 * case, NULL is returned here. "Normal" mappings do have a struct page.
709 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
710 * pte bit, in which case this function is trivial. Secondly, an architecture
711 * may not have a spare pte bit, which requires a more complicated scheme,
714 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
715 * special mapping (even if there are underlying and valid "struct pages").
716 * COWed pages of a VM_PFNMAP are always normal.
718 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
719 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
720 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
721 * mapping will always honor the rule
723 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
725 * And for normal mappings this is false.
727 * This restricts such mappings to be a linear translation from virtual address
728 * to pfn. To get around this restriction, we allow arbitrary mappings so long
729 * as the vma is not a COW mapping; in that case, we know that all ptes are
730 * special (because none can have been COWed).
733 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
735 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
736 * page" backing, however the difference is that _all_ pages with a struct
737 * page (that is, those where pfn_valid is true) are refcounted and considered
738 * normal pages by the VM. The disadvantage is that pages are refcounted
739 * (which can be slower and simply not an option for some PFNMAP users). The
740 * advantage is that we don't have to follow the strict linearity rule of
741 * PFNMAP mappings in order to support COWable mappings.
744 #ifdef __HAVE_ARCH_PTE_SPECIAL
745 # define HAVE_PTE_SPECIAL 1
747 # define HAVE_PTE_SPECIAL 0
749 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
752 unsigned long pfn = pte_pfn(pte);
754 if (HAVE_PTE_SPECIAL) {
755 if (likely(!pte_special(pte)))
757 if (vma->vm_ops && vma->vm_ops->find_special_page)
758 return vma->vm_ops->find_special_page(vma, addr);
759 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
761 if (!is_zero_pfn(pfn))
762 print_bad_pte(vma, addr, pte, NULL);
766 /* !HAVE_PTE_SPECIAL case follows: */
768 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
769 if (vma->vm_flags & VM_MIXEDMAP) {
775 off = (addr - vma->vm_start) >> PAGE_SHIFT;
776 if (pfn == vma->vm_pgoff + off)
778 if (!is_cow_mapping(vma->vm_flags))
783 if (is_zero_pfn(pfn))
786 if (unlikely(pfn > highest_memmap_pfn)) {
787 print_bad_pte(vma, addr, pte, NULL);
792 * NOTE! We still have PageReserved() pages in the page tables.
793 * eg. VDSO mappings can cause them to exist.
796 return pfn_to_page(pfn);
800 * copy one vm_area from one task to the other. Assumes the page tables
801 * already present in the new task to be cleared in the whole range
802 * covered by this vma.
805 static inline unsigned long
806 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
807 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
808 unsigned long addr, int *rss)
810 unsigned long vm_flags = vma->vm_flags;
811 pte_t pte = *src_pte;
814 /* pte contains position in swap or file, so copy. */
815 if (unlikely(!pte_present(pte))) {
816 if (!pte_file(pte)) {
817 swp_entry_t entry = pte_to_swp_entry(pte);
819 if (likely(!non_swap_entry(entry))) {
820 if (swap_duplicate(entry) < 0)
823 /* make sure dst_mm is on swapoff's mmlist. */
824 if (unlikely(list_empty(&dst_mm->mmlist))) {
825 spin_lock(&mmlist_lock);
826 if (list_empty(&dst_mm->mmlist))
827 list_add(&dst_mm->mmlist,
829 spin_unlock(&mmlist_lock);
832 } else if (is_migration_entry(entry)) {
833 page = migration_entry_to_page(entry);
840 if (is_write_migration_entry(entry) &&
841 is_cow_mapping(vm_flags)) {
843 * COW mappings require pages in both
844 * parent and child to be set to read.
846 make_migration_entry_read(&entry);
847 pte = swp_entry_to_pte(entry);
848 if (pte_swp_soft_dirty(*src_pte))
849 pte = pte_swp_mksoft_dirty(pte);
850 set_pte_at(src_mm, addr, src_pte, pte);
858 * If it's a COW mapping, write protect it both
859 * in the parent and the child
861 if (is_cow_mapping(vm_flags)) {
862 ptep_set_wrprotect(src_mm, addr, src_pte);
863 pte = pte_wrprotect(pte);
867 * If it's a shared mapping, mark it clean in
870 if (vm_flags & VM_SHARED)
871 pte = pte_mkclean(pte);
872 pte = pte_mkold(pte);
874 page = vm_normal_page(vma, addr, pte);
885 set_pte_at(dst_mm, addr, dst_pte, pte);
889 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
890 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
891 unsigned long addr, unsigned long end)
893 pte_t *orig_src_pte, *orig_dst_pte;
894 pte_t *src_pte, *dst_pte;
895 spinlock_t *src_ptl, *dst_ptl;
897 int rss[NR_MM_COUNTERS];
898 swp_entry_t entry = (swp_entry_t){0};
903 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 src_pte = pte_offset_map(src_pmd, addr);
907 src_ptl = pte_lockptr(src_mm, src_pmd);
908 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
909 orig_src_pte = src_pte;
910 orig_dst_pte = dst_pte;
911 arch_enter_lazy_mmu_mode();
915 * We are holding two locks at this point - either of them
916 * could generate latencies in another task on another CPU.
918 if (progress >= 32) {
920 if (need_resched() ||
921 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 if (pte_none(*src_pte)) {
928 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
933 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
935 arch_leave_lazy_mmu_mode();
936 spin_unlock(src_ptl);
937 pte_unmap(orig_src_pte);
938 add_mm_rss_vec(dst_mm, rss);
939 pte_unmap_unlock(orig_dst_pte, dst_ptl);
943 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
952 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
953 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
954 unsigned long addr, unsigned long end)
956 pmd_t *src_pmd, *dst_pmd;
959 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 src_pmd = pmd_offset(src_pud, addr);
964 next = pmd_addr_end(addr, end);
965 if (pmd_trans_huge(*src_pmd)) {
967 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
968 err = copy_huge_pmd(dst_mm, src_mm,
969 dst_pmd, src_pmd, addr, vma);
976 if (pmd_none_or_clear_bad(src_pmd))
978 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
985 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
986 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
987 unsigned long addr, unsigned long end)
989 pud_t *src_pud, *dst_pud;
992 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 src_pud = pud_offset(src_pgd, addr);
997 next = pud_addr_end(addr, end);
998 if (pud_none_or_clear_bad(src_pud))
1000 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 } while (dst_pud++, src_pud++, addr = next, addr != end);
1007 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1008 struct vm_area_struct *vma)
1010 pgd_t *src_pgd, *dst_pgd;
1012 unsigned long addr = vma->vm_start;
1013 unsigned long end = vma->vm_end;
1014 unsigned long mmun_start; /* For mmu_notifiers */
1015 unsigned long mmun_end; /* For mmu_notifiers */
1020 * Don't copy ptes where a page fault will fill them correctly.
1021 * Fork becomes much lighter when there are big shared or private
1022 * readonly mappings. The tradeoff is that copy_page_range is more
1023 * efficient than faulting.
1025 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1026 VM_PFNMAP | VM_MIXEDMAP))) {
1031 if (is_vm_hugetlb_page(vma))
1032 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1034 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1036 * We do not free on error cases below as remove_vma
1037 * gets called on error from higher level routine
1039 ret = track_pfn_copy(vma);
1045 * We need to invalidate the secondary MMU mappings only when
1046 * there could be a permission downgrade on the ptes of the
1047 * parent mm. And a permission downgrade will only happen if
1048 * is_cow_mapping() returns true.
1050 is_cow = is_cow_mapping(vma->vm_flags);
1054 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1058 dst_pgd = pgd_offset(dst_mm, addr);
1059 src_pgd = pgd_offset(src_mm, addr);
1061 next = pgd_addr_end(addr, end);
1062 if (pgd_none_or_clear_bad(src_pgd))
1064 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1065 vma, addr, next))) {
1069 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1076 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1077 struct vm_area_struct *vma, pmd_t *pmd,
1078 unsigned long addr, unsigned long end,
1079 struct zap_details *details)
1081 struct mm_struct *mm = tlb->mm;
1082 int force_flush = 0;
1083 int rss[NR_MM_COUNTERS];
1090 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1092 arch_enter_lazy_mmu_mode();
1095 if (pte_none(ptent)) {
1099 if (pte_present(ptent)) {
1102 page = vm_normal_page(vma, addr, ptent);
1103 if (unlikely(details) && page) {
1105 * unmap_shared_mapping_pages() wants to
1106 * invalidate cache without truncating:
1107 * unmap shared but keep private pages.
1109 if (details->check_mapping &&
1110 details->check_mapping != page->mapping)
1113 * Each page->index must be checked when
1114 * invalidating or truncating nonlinear.
1116 if (details->nonlinear_vma &&
1117 (page->index < details->first_index ||
1118 page->index > details->last_index))
1121 ptent = ptep_get_and_clear_full(mm, addr, pte,
1123 tlb_remove_tlb_entry(tlb, pte, addr);
1124 if (unlikely(!page))
1126 if (unlikely(details) && details->nonlinear_vma
1127 && linear_page_index(details->nonlinear_vma,
1128 addr) != page->index) {
1129 pte_t ptfile = pgoff_to_pte(page->index);
1130 if (pte_soft_dirty(ptent))
1131 ptfile = pte_file_mksoft_dirty(ptfile);
1132 set_pte_at(mm, addr, pte, ptfile);
1135 rss[MM_ANONPAGES]--;
1137 if (pte_dirty(ptent)) {
1139 set_page_dirty(page);
1141 if (pte_young(ptent) &&
1142 likely(!(vma->vm_flags & VM_SEQ_READ)))
1143 mark_page_accessed(page);
1144 rss[MM_FILEPAGES]--;
1146 page_remove_rmap(page);
1147 if (unlikely(page_mapcount(page) < 0))
1148 print_bad_pte(vma, addr, ptent, page);
1149 if (unlikely(!__tlb_remove_page(tlb, page))) {
1157 * If details->check_mapping, we leave swap entries;
1158 * if details->nonlinear_vma, we leave file entries.
1160 if (unlikely(details))
1162 if (pte_file(ptent)) {
1163 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1164 print_bad_pte(vma, addr, ptent, NULL);
1166 swp_entry_t entry = pte_to_swp_entry(ptent);
1168 if (!non_swap_entry(entry))
1170 else if (is_migration_entry(entry)) {
1173 page = migration_entry_to_page(entry);
1176 rss[MM_ANONPAGES]--;
1178 rss[MM_FILEPAGES]--;
1180 if (unlikely(!free_swap_and_cache(entry)))
1181 print_bad_pte(vma, addr, ptent, NULL);
1183 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1184 } while (pte++, addr += PAGE_SIZE, addr != end);
1186 add_mm_rss_vec(mm, rss);
1187 arch_leave_lazy_mmu_mode();
1189 /* Do the actual TLB flush before dropping ptl */
1191 tlb_flush_mmu_tlbonly(tlb);
1192 pte_unmap_unlock(start_pte, ptl);
1195 * If we forced a TLB flush (either due to running out of
1196 * batch buffers or because we needed to flush dirty TLB
1197 * entries before releasing the ptl), free the batched
1198 * memory too. Restart if we didn't do everything.
1202 tlb_flush_mmu_free(tlb);
1211 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1212 struct vm_area_struct *vma, pud_t *pud,
1213 unsigned long addr, unsigned long end,
1214 struct zap_details *details)
1219 pmd = pmd_offset(pud, addr);
1221 next = pmd_addr_end(addr, end);
1222 if (pmd_trans_huge(*pmd)) {
1223 if (next - addr != HPAGE_PMD_SIZE) {
1224 #ifdef CONFIG_DEBUG_VM
1225 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1226 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1227 __func__, addr, end,
1233 split_huge_page_pmd(vma, addr, pmd);
1234 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1239 * Here there can be other concurrent MADV_DONTNEED or
1240 * trans huge page faults running, and if the pmd is
1241 * none or trans huge it can change under us. This is
1242 * because MADV_DONTNEED holds the mmap_sem in read
1245 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1247 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1250 } while (pmd++, addr = next, addr != end);
1255 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1256 struct vm_area_struct *vma, pgd_t *pgd,
1257 unsigned long addr, unsigned long end,
1258 struct zap_details *details)
1263 pud = pud_offset(pgd, addr);
1265 next = pud_addr_end(addr, end);
1266 if (pud_none_or_clear_bad(pud))
1268 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1269 } while (pud++, addr = next, addr != end);
1274 static void unmap_page_range(struct mmu_gather *tlb,
1275 struct vm_area_struct *vma,
1276 unsigned long addr, unsigned long end,
1277 struct zap_details *details)
1282 if (details && !details->check_mapping && !details->nonlinear_vma)
1285 BUG_ON(addr >= end);
1286 tlb_start_vma(tlb, vma);
1287 pgd = pgd_offset(vma->vm_mm, addr);
1289 next = pgd_addr_end(addr, end);
1290 if (pgd_none_or_clear_bad(pgd))
1292 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1293 } while (pgd++, addr = next, addr != end);
1294 tlb_end_vma(tlb, vma);
1298 static void unmap_single_vma(struct mmu_gather *tlb,
1299 struct vm_area_struct *vma, unsigned long start_addr,
1300 unsigned long end_addr,
1301 struct zap_details *details)
1303 unsigned long start = max(vma->vm_start, start_addr);
1306 if (start >= vma->vm_end)
1308 end = min(vma->vm_end, end_addr);
1309 if (end <= vma->vm_start)
1313 uprobe_munmap(vma, start, end);
1315 if (unlikely(vma->vm_flags & VM_PFNMAP))
1316 untrack_pfn(vma, 0, 0);
1319 if (unlikely(is_vm_hugetlb_page(vma))) {
1321 * It is undesirable to test vma->vm_file as it
1322 * should be non-null for valid hugetlb area.
1323 * However, vm_file will be NULL in the error
1324 * cleanup path of mmap_region. When
1325 * hugetlbfs ->mmap method fails,
1326 * mmap_region() nullifies vma->vm_file
1327 * before calling this function to clean up.
1328 * Since no pte has actually been setup, it is
1329 * safe to do nothing in this case.
1332 i_mmap_lock_write(vma->vm_file->f_mapping);
1333 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1334 i_mmap_unlock_write(vma->vm_file->f_mapping);
1337 unmap_page_range(tlb, vma, start, end, details);
1342 * unmap_vmas - unmap a range of memory covered by a list of vma's
1343 * @tlb: address of the caller's struct mmu_gather
1344 * @vma: the starting vma
1345 * @start_addr: virtual address at which to start unmapping
1346 * @end_addr: virtual address at which to end unmapping
1348 * Unmap all pages in the vma list.
1350 * Only addresses between `start' and `end' will be unmapped.
1352 * The VMA list must be sorted in ascending virtual address order.
1354 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1355 * range after unmap_vmas() returns. So the only responsibility here is to
1356 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1357 * drops the lock and schedules.
1359 void unmap_vmas(struct mmu_gather *tlb,
1360 struct vm_area_struct *vma, unsigned long start_addr,
1361 unsigned long end_addr)
1363 struct mm_struct *mm = vma->vm_mm;
1365 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1366 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1367 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1368 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372 * zap_page_range - remove user pages in a given range
1373 * @vma: vm_area_struct holding the applicable pages
1374 * @start: starting address of pages to zap
1375 * @size: number of bytes to zap
1376 * @details: details of nonlinear truncation or shared cache invalidation
1378 * Caller must protect the VMA list
1380 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1381 unsigned long size, struct zap_details *details)
1383 struct mm_struct *mm = vma->vm_mm;
1384 struct mmu_gather tlb;
1385 unsigned long end = start + size;
1388 tlb_gather_mmu(&tlb, mm, start, end);
1389 update_hiwater_rss(mm);
1390 mmu_notifier_invalidate_range_start(mm, start, end);
1391 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1392 unmap_single_vma(&tlb, vma, start, end, details);
1393 mmu_notifier_invalidate_range_end(mm, start, end);
1394 tlb_finish_mmu(&tlb, start, end);
1398 * zap_page_range_single - remove user pages in a given range
1399 * @vma: vm_area_struct holding the applicable pages
1400 * @address: starting address of pages to zap
1401 * @size: number of bytes to zap
1402 * @details: details of nonlinear truncation or shared cache invalidation
1404 * The range must fit into one VMA.
1406 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1407 unsigned long size, struct zap_details *details)
1409 struct mm_struct *mm = vma->vm_mm;
1410 struct mmu_gather tlb;
1411 unsigned long end = address + size;
1414 tlb_gather_mmu(&tlb, mm, address, end);
1415 update_hiwater_rss(mm);
1416 mmu_notifier_invalidate_range_start(mm, address, end);
1417 unmap_single_vma(&tlb, vma, address, end, details);
1418 mmu_notifier_invalidate_range_end(mm, address, end);
1419 tlb_finish_mmu(&tlb, address, end);
1423 * zap_vma_ptes - remove ptes mapping the vma
1424 * @vma: vm_area_struct holding ptes to be zapped
1425 * @address: starting address of pages to zap
1426 * @size: number of bytes to zap
1428 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1430 * The entire address range must be fully contained within the vma.
1432 * Returns 0 if successful.
1434 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1437 if (address < vma->vm_start || address + size > vma->vm_end ||
1438 !(vma->vm_flags & VM_PFNMAP))
1440 zap_page_range_single(vma, address, size, NULL);
1443 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1445 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1448 pgd_t * pgd = pgd_offset(mm, addr);
1449 pud_t * pud = pud_alloc(mm, pgd, addr);
1451 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1453 VM_BUG_ON(pmd_trans_huge(*pmd));
1454 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1461 * This is the old fallback for page remapping.
1463 * For historical reasons, it only allows reserved pages. Only
1464 * old drivers should use this, and they needed to mark their
1465 * pages reserved for the old functions anyway.
1467 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1468 struct page *page, pgprot_t prot)
1470 struct mm_struct *mm = vma->vm_mm;
1479 flush_dcache_page(page);
1480 pte = get_locked_pte(mm, addr, &ptl);
1484 if (!pte_none(*pte))
1487 /* Ok, finally just insert the thing.. */
1489 inc_mm_counter_fast(mm, MM_FILEPAGES);
1490 page_add_file_rmap(page);
1491 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1494 pte_unmap_unlock(pte, ptl);
1497 pte_unmap_unlock(pte, ptl);
1503 * vm_insert_page - insert single page into user vma
1504 * @vma: user vma to map to
1505 * @addr: target user address of this page
1506 * @page: source kernel page
1508 * This allows drivers to insert individual pages they've allocated
1511 * The page has to be a nice clean _individual_ kernel allocation.
1512 * If you allocate a compound page, you need to have marked it as
1513 * such (__GFP_COMP), or manually just split the page up yourself
1514 * (see split_page()).
1516 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1517 * took an arbitrary page protection parameter. This doesn't allow
1518 * that. Your vma protection will have to be set up correctly, which
1519 * means that if you want a shared writable mapping, you'd better
1520 * ask for a shared writable mapping!
1522 * The page does not need to be reserved.
1524 * Usually this function is called from f_op->mmap() handler
1525 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1526 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1527 * function from other places, for example from page-fault handler.
1529 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1532 if (addr < vma->vm_start || addr >= vma->vm_end)
1534 if (!page_count(page))
1536 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1537 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1538 BUG_ON(vma->vm_flags & VM_PFNMAP);
1539 vma->vm_flags |= VM_MIXEDMAP;
1541 return insert_page(vma, addr, page, vma->vm_page_prot);
1543 EXPORT_SYMBOL(vm_insert_page);
1545 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1546 unsigned long pfn, pgprot_t prot)
1548 struct mm_struct *mm = vma->vm_mm;
1554 pte = get_locked_pte(mm, addr, &ptl);
1558 if (!pte_none(*pte))
1561 /* Ok, finally just insert the thing.. */
1562 entry = pte_mkspecial(pfn_pte(pfn, prot));
1563 set_pte_at(mm, addr, pte, entry);
1564 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1568 pte_unmap_unlock(pte, ptl);
1574 * vm_insert_pfn - insert single pfn into user vma
1575 * @vma: user vma to map to
1576 * @addr: target user address of this page
1577 * @pfn: source kernel pfn
1579 * Similar to vm_insert_page, this allows drivers to insert individual pages
1580 * they've allocated into a user vma. Same comments apply.
1582 * This function should only be called from a vm_ops->fault handler, and
1583 * in that case the handler should return NULL.
1585 * vma cannot be a COW mapping.
1587 * As this is called only for pages that do not currently exist, we
1588 * do not need to flush old virtual caches or the TLB.
1590 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1594 pgprot_t pgprot = vma->vm_page_prot;
1596 * Technically, architectures with pte_special can avoid all these
1597 * restrictions (same for remap_pfn_range). However we would like
1598 * consistency in testing and feature parity among all, so we should
1599 * try to keep these invariants in place for everybody.
1601 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1602 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1603 (VM_PFNMAP|VM_MIXEDMAP));
1604 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1605 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1607 if (addr < vma->vm_start || addr >= vma->vm_end)
1609 if (track_pfn_insert(vma, &pgprot, pfn))
1612 ret = insert_pfn(vma, addr, pfn, pgprot);
1616 EXPORT_SYMBOL(vm_insert_pfn);
1618 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1621 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1623 if (addr < vma->vm_start || addr >= vma->vm_end)
1627 * If we don't have pte special, then we have to use the pfn_valid()
1628 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1629 * refcount the page if pfn_valid is true (hence insert_page rather
1630 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1631 * without pte special, it would there be refcounted as a normal page.
1633 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1636 page = pfn_to_page(pfn);
1637 return insert_page(vma, addr, page, vma->vm_page_prot);
1639 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1641 EXPORT_SYMBOL(vm_insert_mixed);
1644 * maps a range of physical memory into the requested pages. the old
1645 * mappings are removed. any references to nonexistent pages results
1646 * in null mappings (currently treated as "copy-on-access")
1648 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1649 unsigned long addr, unsigned long end,
1650 unsigned long pfn, pgprot_t prot)
1655 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1658 arch_enter_lazy_mmu_mode();
1660 BUG_ON(!pte_none(*pte));
1661 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1663 } while (pte++, addr += PAGE_SIZE, addr != end);
1664 arch_leave_lazy_mmu_mode();
1665 pte_unmap_unlock(pte - 1, ptl);
1669 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1670 unsigned long addr, unsigned long end,
1671 unsigned long pfn, pgprot_t prot)
1676 pfn -= addr >> PAGE_SHIFT;
1677 pmd = pmd_alloc(mm, pud, addr);
1680 VM_BUG_ON(pmd_trans_huge(*pmd));
1682 next = pmd_addr_end(addr, end);
1683 if (remap_pte_range(mm, pmd, addr, next,
1684 pfn + (addr >> PAGE_SHIFT), prot))
1686 } while (pmd++, addr = next, addr != end);
1690 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1691 unsigned long addr, unsigned long end,
1692 unsigned long pfn, pgprot_t prot)
1697 pfn -= addr >> PAGE_SHIFT;
1698 pud = pud_alloc(mm, pgd, addr);
1702 next = pud_addr_end(addr, end);
1703 if (remap_pmd_range(mm, pud, addr, next,
1704 pfn + (addr >> PAGE_SHIFT), prot))
1706 } while (pud++, addr = next, addr != end);
1711 * remap_pfn_range - remap kernel memory to userspace
1712 * @vma: user vma to map to
1713 * @addr: target user address to start at
1714 * @pfn: physical address of kernel memory
1715 * @size: size of map area
1716 * @prot: page protection flags for this mapping
1718 * Note: this is only safe if the mm semaphore is held when called.
1720 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1721 unsigned long pfn, unsigned long size, pgprot_t prot)
1725 unsigned long end = addr + PAGE_ALIGN(size);
1726 struct mm_struct *mm = vma->vm_mm;
1730 * Physically remapped pages are special. Tell the
1731 * rest of the world about it:
1732 * VM_IO tells people not to look at these pages
1733 * (accesses can have side effects).
1734 * VM_PFNMAP tells the core MM that the base pages are just
1735 * raw PFN mappings, and do not have a "struct page" associated
1738 * Disable vma merging and expanding with mremap().
1740 * Omit vma from core dump, even when VM_IO turned off.
1742 * There's a horrible special case to handle copy-on-write
1743 * behaviour that some programs depend on. We mark the "original"
1744 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1745 * See vm_normal_page() for details.
1747 if (is_cow_mapping(vma->vm_flags)) {
1748 if (addr != vma->vm_start || end != vma->vm_end)
1750 vma->vm_pgoff = pfn;
1753 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1757 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1759 BUG_ON(addr >= end);
1760 pfn -= addr >> PAGE_SHIFT;
1761 pgd = pgd_offset(mm, addr);
1762 flush_cache_range(vma, addr, end);
1764 next = pgd_addr_end(addr, end);
1765 err = remap_pud_range(mm, pgd, addr, next,
1766 pfn + (addr >> PAGE_SHIFT), prot);
1769 } while (pgd++, addr = next, addr != end);
1772 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1776 EXPORT_SYMBOL(remap_pfn_range);
1779 * vm_iomap_memory - remap memory to userspace
1780 * @vma: user vma to map to
1781 * @start: start of area
1782 * @len: size of area
1784 * This is a simplified io_remap_pfn_range() for common driver use. The
1785 * driver just needs to give us the physical memory range to be mapped,
1786 * we'll figure out the rest from the vma information.
1788 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1789 * whatever write-combining details or similar.
1791 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1793 unsigned long vm_len, pfn, pages;
1795 /* Check that the physical memory area passed in looks valid */
1796 if (start + len < start)
1799 * You *really* shouldn't map things that aren't page-aligned,
1800 * but we've historically allowed it because IO memory might
1801 * just have smaller alignment.
1803 len += start & ~PAGE_MASK;
1804 pfn = start >> PAGE_SHIFT;
1805 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1806 if (pfn + pages < pfn)
1809 /* We start the mapping 'vm_pgoff' pages into the area */
1810 if (vma->vm_pgoff > pages)
1812 pfn += vma->vm_pgoff;
1813 pages -= vma->vm_pgoff;
1815 /* Can we fit all of the mapping? */
1816 vm_len = vma->vm_end - vma->vm_start;
1817 if (vm_len >> PAGE_SHIFT > pages)
1820 /* Ok, let it rip */
1821 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1823 EXPORT_SYMBOL(vm_iomap_memory);
1825 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1826 unsigned long addr, unsigned long end,
1827 pte_fn_t fn, void *data)
1832 spinlock_t *uninitialized_var(ptl);
1834 pte = (mm == &init_mm) ?
1835 pte_alloc_kernel(pmd, addr) :
1836 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1840 BUG_ON(pmd_huge(*pmd));
1842 arch_enter_lazy_mmu_mode();
1844 token = pmd_pgtable(*pmd);
1847 err = fn(pte++, token, addr, data);
1850 } while (addr += PAGE_SIZE, addr != end);
1852 arch_leave_lazy_mmu_mode();
1855 pte_unmap_unlock(pte-1, ptl);
1859 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1860 unsigned long addr, unsigned long end,
1861 pte_fn_t fn, void *data)
1867 BUG_ON(pud_huge(*pud));
1869 pmd = pmd_alloc(mm, pud, addr);
1873 next = pmd_addr_end(addr, end);
1874 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1877 } while (pmd++, addr = next, addr != end);
1881 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1882 unsigned long addr, unsigned long end,
1883 pte_fn_t fn, void *data)
1889 pud = pud_alloc(mm, pgd, addr);
1893 next = pud_addr_end(addr, end);
1894 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1897 } while (pud++, addr = next, addr != end);
1902 * Scan a region of virtual memory, filling in page tables as necessary
1903 * and calling a provided function on each leaf page table.
1905 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1906 unsigned long size, pte_fn_t fn, void *data)
1910 unsigned long end = addr + size;
1913 BUG_ON(addr >= end);
1914 pgd = pgd_offset(mm, addr);
1916 next = pgd_addr_end(addr, end);
1917 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1920 } while (pgd++, addr = next, addr != end);
1924 EXPORT_SYMBOL_GPL(apply_to_page_range);
1927 * handle_pte_fault chooses page fault handler according to an entry
1928 * which was read non-atomically. Before making any commitment, on
1929 * those architectures or configurations (e.g. i386 with PAE) which
1930 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
1931 * must check under lock before unmapping the pte and proceeding
1932 * (but do_wp_page is only called after already making such a check;
1933 * and do_anonymous_page can safely check later on).
1935 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1936 pte_t *page_table, pte_t orig_pte)
1939 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1940 if (sizeof(pte_t) > sizeof(unsigned long)) {
1941 spinlock_t *ptl = pte_lockptr(mm, pmd);
1943 same = pte_same(*page_table, orig_pte);
1947 pte_unmap(page_table);
1951 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1953 debug_dma_assert_idle(src);
1956 * If the source page was a PFN mapping, we don't have
1957 * a "struct page" for it. We do a best-effort copy by
1958 * just copying from the original user address. If that
1959 * fails, we just zero-fill it. Live with it.
1961 if (unlikely(!src)) {
1962 void *kaddr = kmap_atomic(dst);
1963 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1966 * This really shouldn't fail, because the page is there
1967 * in the page tables. But it might just be unreadable,
1968 * in which case we just give up and fill the result with
1971 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1973 kunmap_atomic(kaddr);
1974 flush_dcache_page(dst);
1976 copy_user_highpage(dst, src, va, vma);
1980 * Notify the address space that the page is about to become writable so that
1981 * it can prohibit this or wait for the page to get into an appropriate state.
1983 * We do this without the lock held, so that it can sleep if it needs to.
1985 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1986 unsigned long address)
1988 struct vm_fault vmf;
1991 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1992 vmf.pgoff = page->index;
1993 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1996 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1997 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1999 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2001 if (!page->mapping) {
2003 return 0; /* retry */
2005 ret |= VM_FAULT_LOCKED;
2007 VM_BUG_ON_PAGE(!PageLocked(page), page);
2012 * This routine handles present pages, when users try to write
2013 * to a shared page. It is done by copying the page to a new address
2014 * and decrementing the shared-page counter for the old page.
2016 * Note that this routine assumes that the protection checks have been
2017 * done by the caller (the low-level page fault routine in most cases).
2018 * Thus we can safely just mark it writable once we've done any necessary
2021 * We also mark the page dirty at this point even though the page will
2022 * change only once the write actually happens. This avoids a few races,
2023 * and potentially makes it more efficient.
2025 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2026 * but allow concurrent faults), with pte both mapped and locked.
2027 * We return with mmap_sem still held, but pte unmapped and unlocked.
2029 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2030 unsigned long address, pte_t *page_table, pmd_t *pmd,
2031 spinlock_t *ptl, pte_t orig_pte)
2034 struct page *old_page, *new_page = NULL;
2037 int page_mkwrite = 0;
2038 struct page *dirty_page = NULL;
2039 unsigned long mmun_start = 0; /* For mmu_notifiers */
2040 unsigned long mmun_end = 0; /* For mmu_notifiers */
2041 struct mem_cgroup *memcg;
2043 old_page = vm_normal_page(vma, address, orig_pte);
2046 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2049 * We should not cow pages in a shared writeable mapping.
2050 * Just mark the pages writable as we can't do any dirty
2051 * accounting on raw pfn maps.
2053 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2054 (VM_WRITE|VM_SHARED))
2060 * Take out anonymous pages first, anonymous shared vmas are
2061 * not dirty accountable.
2063 if (PageAnon(old_page) && !PageKsm(old_page)) {
2064 if (!trylock_page(old_page)) {
2065 page_cache_get(old_page);
2066 pte_unmap_unlock(page_table, ptl);
2067 lock_page(old_page);
2068 page_table = pte_offset_map_lock(mm, pmd, address,
2070 if (!pte_same(*page_table, orig_pte)) {
2071 unlock_page(old_page);
2074 page_cache_release(old_page);
2076 if (reuse_swap_page(old_page)) {
2078 * The page is all ours. Move it to our anon_vma so
2079 * the rmap code will not search our parent or siblings.
2080 * Protected against the rmap code by the page lock.
2082 page_move_anon_rmap(old_page, vma, address);
2083 unlock_page(old_page);
2086 unlock_page(old_page);
2087 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2088 (VM_WRITE|VM_SHARED))) {
2090 * Only catch write-faults on shared writable pages,
2091 * read-only shared pages can get COWed by
2092 * get_user_pages(.write=1, .force=1).
2094 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2096 page_cache_get(old_page);
2097 pte_unmap_unlock(page_table, ptl);
2098 tmp = do_page_mkwrite(vma, old_page, address);
2099 if (unlikely(!tmp || (tmp &
2100 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2101 page_cache_release(old_page);
2105 * Since we dropped the lock we need to revalidate
2106 * the PTE as someone else may have changed it. If
2107 * they did, we just return, as we can count on the
2108 * MMU to tell us if they didn't also make it writable.
2110 page_table = pte_offset_map_lock(mm, pmd, address,
2112 if (!pte_same(*page_table, orig_pte)) {
2113 unlock_page(old_page);
2119 dirty_page = old_page;
2120 get_page(dirty_page);
2124 * Clear the pages cpupid information as the existing
2125 * information potentially belongs to a now completely
2126 * unrelated process.
2129 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2131 flush_cache_page(vma, address, pte_pfn(orig_pte));
2132 entry = pte_mkyoung(orig_pte);
2133 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2134 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2135 update_mmu_cache(vma, address, page_table);
2136 pte_unmap_unlock(page_table, ptl);
2137 ret |= VM_FAULT_WRITE;
2142 if (!page_mkwrite) {
2143 struct address_space *mapping;
2146 lock_page(dirty_page);
2147 dirtied = set_page_dirty(dirty_page);
2148 VM_BUG_ON_PAGE(PageAnon(dirty_page), dirty_page);
2149 mapping = dirty_page->mapping;
2150 unlock_page(dirty_page);
2152 if (dirtied && mapping) {
2154 * Some device drivers do not set page.mapping
2155 * but still dirty their pages
2157 balance_dirty_pages_ratelimited(mapping);
2160 /* file_update_time outside page_lock */
2162 file_update_time(vma->vm_file);
2164 put_page(dirty_page);
2166 struct address_space *mapping = dirty_page->mapping;
2168 set_page_dirty(dirty_page);
2169 unlock_page(dirty_page);
2170 page_cache_release(dirty_page);
2173 * Some device drivers do not set page.mapping
2174 * but still dirty their pages
2176 balance_dirty_pages_ratelimited(mapping);
2184 * Ok, we need to copy. Oh, well..
2186 page_cache_get(old_page);
2188 pte_unmap_unlock(page_table, ptl);
2190 if (unlikely(anon_vma_prepare(vma)))
2193 if (is_zero_pfn(pte_pfn(orig_pte))) {
2194 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2198 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2201 cow_user_page(new_page, old_page, address, vma);
2203 __SetPageUptodate(new_page);
2205 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2208 mmun_start = address & PAGE_MASK;
2209 mmun_end = mmun_start + PAGE_SIZE;
2210 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2213 * Re-check the pte - we dropped the lock
2215 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2216 if (likely(pte_same(*page_table, orig_pte))) {
2218 if (!PageAnon(old_page)) {
2219 dec_mm_counter_fast(mm, MM_FILEPAGES);
2220 inc_mm_counter_fast(mm, MM_ANONPAGES);
2223 inc_mm_counter_fast(mm, MM_ANONPAGES);
2224 flush_cache_page(vma, address, pte_pfn(orig_pte));
2225 entry = mk_pte(new_page, vma->vm_page_prot);
2226 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2228 * Clear the pte entry and flush it first, before updating the
2229 * pte with the new entry. This will avoid a race condition
2230 * seen in the presence of one thread doing SMC and another
2233 ptep_clear_flush_notify(vma, address, page_table);
2234 page_add_new_anon_rmap(new_page, vma, address);
2235 mem_cgroup_commit_charge(new_page, memcg, false);
2236 lru_cache_add_active_or_unevictable(new_page, vma);
2238 * We call the notify macro here because, when using secondary
2239 * mmu page tables (such as kvm shadow page tables), we want the
2240 * new page to be mapped directly into the secondary page table.
2242 set_pte_at_notify(mm, address, page_table, entry);
2243 update_mmu_cache(vma, address, page_table);
2246 * Only after switching the pte to the new page may
2247 * we remove the mapcount here. Otherwise another
2248 * process may come and find the rmap count decremented
2249 * before the pte is switched to the new page, and
2250 * "reuse" the old page writing into it while our pte
2251 * here still points into it and can be read by other
2254 * The critical issue is to order this
2255 * page_remove_rmap with the ptp_clear_flush above.
2256 * Those stores are ordered by (if nothing else,)
2257 * the barrier present in the atomic_add_negative
2258 * in page_remove_rmap.
2260 * Then the TLB flush in ptep_clear_flush ensures that
2261 * no process can access the old page before the
2262 * decremented mapcount is visible. And the old page
2263 * cannot be reused until after the decremented
2264 * mapcount is visible. So transitively, TLBs to
2265 * old page will be flushed before it can be reused.
2267 page_remove_rmap(old_page);
2270 /* Free the old page.. */
2271 new_page = old_page;
2272 ret |= VM_FAULT_WRITE;
2274 mem_cgroup_cancel_charge(new_page, memcg);
2277 page_cache_release(new_page);
2279 pte_unmap_unlock(page_table, ptl);
2280 if (mmun_end > mmun_start)
2281 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2284 * Don't let another task, with possibly unlocked vma,
2285 * keep the mlocked page.
2287 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2288 lock_page(old_page); /* LRU manipulation */
2289 munlock_vma_page(old_page);
2290 unlock_page(old_page);
2292 page_cache_release(old_page);
2296 page_cache_release(new_page);
2299 page_cache_release(old_page);
2300 return VM_FAULT_OOM;
2303 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2304 unsigned long start_addr, unsigned long end_addr,
2305 struct zap_details *details)
2307 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2310 static inline void unmap_mapping_range_tree(struct rb_root *root,
2311 struct zap_details *details)
2313 struct vm_area_struct *vma;
2314 pgoff_t vba, vea, zba, zea;
2316 vma_interval_tree_foreach(vma, root,
2317 details->first_index, details->last_index) {
2319 vba = vma->vm_pgoff;
2320 vea = vba + vma_pages(vma) - 1;
2321 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2322 zba = details->first_index;
2325 zea = details->last_index;
2329 unmap_mapping_range_vma(vma,
2330 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2331 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2336 static inline void unmap_mapping_range_list(struct list_head *head,
2337 struct zap_details *details)
2339 struct vm_area_struct *vma;
2342 * In nonlinear VMAs there is no correspondence between virtual address
2343 * offset and file offset. So we must perform an exhaustive search
2344 * across *all* the pages in each nonlinear VMA, not just the pages
2345 * whose virtual address lies outside the file truncation point.
2347 list_for_each_entry(vma, head, shared.nonlinear) {
2348 details->nonlinear_vma = vma;
2349 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2354 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2355 * @mapping: the address space containing mmaps to be unmapped.
2356 * @holebegin: byte in first page to unmap, relative to the start of
2357 * the underlying file. This will be rounded down to a PAGE_SIZE
2358 * boundary. Note that this is different from truncate_pagecache(), which
2359 * must keep the partial page. In contrast, we must get rid of
2361 * @holelen: size of prospective hole in bytes. This will be rounded
2362 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2364 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2365 * but 0 when invalidating pagecache, don't throw away private data.
2367 void unmap_mapping_range(struct address_space *mapping,
2368 loff_t const holebegin, loff_t const holelen, int even_cows)
2370 struct zap_details details;
2371 pgoff_t hba = holebegin >> PAGE_SHIFT;
2372 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2374 /* Check for overflow. */
2375 if (sizeof(holelen) > sizeof(hlen)) {
2377 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2378 if (holeend & ~(long long)ULONG_MAX)
2379 hlen = ULONG_MAX - hba + 1;
2382 details.check_mapping = even_cows? NULL: mapping;
2383 details.nonlinear_vma = NULL;
2384 details.first_index = hba;
2385 details.last_index = hba + hlen - 1;
2386 if (details.last_index < details.first_index)
2387 details.last_index = ULONG_MAX;
2390 i_mmap_lock_write(mapping);
2391 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2392 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2393 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2394 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2395 i_mmap_unlock_write(mapping);
2397 EXPORT_SYMBOL(unmap_mapping_range);
2400 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2401 * but allow concurrent faults), and pte mapped but not yet locked.
2402 * We return with pte unmapped and unlocked.
2404 * We return with the mmap_sem locked or unlocked in the same cases
2405 * as does filemap_fault().
2407 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2408 unsigned long address, pte_t *page_table, pmd_t *pmd,
2409 unsigned int flags, pte_t orig_pte)
2412 struct page *page, *swapcache;
2413 struct mem_cgroup *memcg;
2420 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2423 entry = pte_to_swp_entry(orig_pte);
2424 if (unlikely(non_swap_entry(entry))) {
2425 if (is_migration_entry(entry)) {
2426 migration_entry_wait(mm, pmd, address);
2427 } else if (is_hwpoison_entry(entry)) {
2428 ret = VM_FAULT_HWPOISON;
2430 print_bad_pte(vma, address, orig_pte, NULL);
2431 ret = VM_FAULT_SIGBUS;
2435 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2436 page = lookup_swap_cache(entry);
2438 page = swapin_readahead(entry,
2439 GFP_HIGHUSER_MOVABLE, vma, address);
2442 * Back out if somebody else faulted in this pte
2443 * while we released the pte lock.
2445 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2446 if (likely(pte_same(*page_table, orig_pte)))
2448 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2452 /* Had to read the page from swap area: Major fault */
2453 ret = VM_FAULT_MAJOR;
2454 count_vm_event(PGMAJFAULT);
2455 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2456 } else if (PageHWPoison(page)) {
2458 * hwpoisoned dirty swapcache pages are kept for killing
2459 * owner processes (which may be unknown at hwpoison time)
2461 ret = VM_FAULT_HWPOISON;
2462 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2468 locked = lock_page_or_retry(page, mm, flags);
2470 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2472 ret |= VM_FAULT_RETRY;
2477 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2478 * release the swapcache from under us. The page pin, and pte_same
2479 * test below, are not enough to exclude that. Even if it is still
2480 * swapcache, we need to check that the page's swap has not changed.
2482 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2485 page = ksm_might_need_to_copy(page, vma, address);
2486 if (unlikely(!page)) {
2492 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2498 * Back out if somebody else already faulted in this pte.
2500 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2501 if (unlikely(!pte_same(*page_table, orig_pte)))
2504 if (unlikely(!PageUptodate(page))) {
2505 ret = VM_FAULT_SIGBUS;
2510 * The page isn't present yet, go ahead with the fault.
2512 * Be careful about the sequence of operations here.
2513 * To get its accounting right, reuse_swap_page() must be called
2514 * while the page is counted on swap but not yet in mapcount i.e.
2515 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2516 * must be called after the swap_free(), or it will never succeed.
2519 inc_mm_counter_fast(mm, MM_ANONPAGES);
2520 dec_mm_counter_fast(mm, MM_SWAPENTS);
2521 pte = mk_pte(page, vma->vm_page_prot);
2522 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2523 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2524 flags &= ~FAULT_FLAG_WRITE;
2525 ret |= VM_FAULT_WRITE;
2528 flush_icache_page(vma, page);
2529 if (pte_swp_soft_dirty(orig_pte))
2530 pte = pte_mksoft_dirty(pte);
2531 set_pte_at(mm, address, page_table, pte);
2532 if (page == swapcache) {
2533 do_page_add_anon_rmap(page, vma, address, exclusive);
2534 mem_cgroup_commit_charge(page, memcg, true);
2535 } else { /* ksm created a completely new copy */
2536 page_add_new_anon_rmap(page, vma, address);
2537 mem_cgroup_commit_charge(page, memcg, false);
2538 lru_cache_add_active_or_unevictable(page, vma);
2542 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2543 try_to_free_swap(page);
2545 if (page != swapcache) {
2547 * Hold the lock to avoid the swap entry to be reused
2548 * until we take the PT lock for the pte_same() check
2549 * (to avoid false positives from pte_same). For
2550 * further safety release the lock after the swap_free
2551 * so that the swap count won't change under a
2552 * parallel locked swapcache.
2554 unlock_page(swapcache);
2555 page_cache_release(swapcache);
2558 if (flags & FAULT_FLAG_WRITE) {
2559 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2560 if (ret & VM_FAULT_ERROR)
2561 ret &= VM_FAULT_ERROR;
2565 /* No need to invalidate - it was non-present before */
2566 update_mmu_cache(vma, address, page_table);
2568 pte_unmap_unlock(page_table, ptl);
2572 mem_cgroup_cancel_charge(page, memcg);
2573 pte_unmap_unlock(page_table, ptl);
2577 page_cache_release(page);
2578 if (page != swapcache) {
2579 unlock_page(swapcache);
2580 page_cache_release(swapcache);
2586 * This is like a special single-page "expand_{down|up}wards()",
2587 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2588 * doesn't hit another vma.
2590 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2592 address &= PAGE_MASK;
2593 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2594 struct vm_area_struct *prev = vma->vm_prev;
2597 * Is there a mapping abutting this one below?
2599 * That's only ok if it's the same stack mapping
2600 * that has gotten split..
2602 if (prev && prev->vm_end == address)
2603 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2605 return expand_downwards(vma, address - PAGE_SIZE);
2607 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2608 struct vm_area_struct *next = vma->vm_next;
2610 /* As VM_GROWSDOWN but s/below/above/ */
2611 if (next && next->vm_start == address + PAGE_SIZE)
2612 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2614 return expand_upwards(vma, address + PAGE_SIZE);
2620 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2621 * but allow concurrent faults), and pte mapped but not yet locked.
2622 * We return with mmap_sem still held, but pte unmapped and unlocked.
2624 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2625 unsigned long address, pte_t *page_table, pmd_t *pmd,
2628 struct mem_cgroup *memcg;
2633 pte_unmap(page_table);
2635 /* Check if we need to add a guard page to the stack */
2636 if (check_stack_guard_page(vma, address) < 0)
2637 return VM_FAULT_SIGSEGV;
2639 /* Use the zero-page for reads */
2640 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2641 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2642 vma->vm_page_prot));
2643 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2644 if (!pte_none(*page_table))
2649 /* Allocate our own private page. */
2650 if (unlikely(anon_vma_prepare(vma)))
2652 page = alloc_zeroed_user_highpage_movable(vma, address);
2656 * The memory barrier inside __SetPageUptodate makes sure that
2657 * preceeding stores to the page contents become visible before
2658 * the set_pte_at() write.
2660 __SetPageUptodate(page);
2662 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2665 entry = mk_pte(page, vma->vm_page_prot);
2666 if (vma->vm_flags & VM_WRITE)
2667 entry = pte_mkwrite(pte_mkdirty(entry));
2669 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2670 if (!pte_none(*page_table))
2673 inc_mm_counter_fast(mm, MM_ANONPAGES);
2674 page_add_new_anon_rmap(page, vma, address);
2675 mem_cgroup_commit_charge(page, memcg, false);
2676 lru_cache_add_active_or_unevictable(page, vma);
2678 set_pte_at(mm, address, page_table, entry);
2680 /* No need to invalidate - it was non-present before */
2681 update_mmu_cache(vma, address, page_table);
2683 pte_unmap_unlock(page_table, ptl);
2686 mem_cgroup_cancel_charge(page, memcg);
2687 page_cache_release(page);
2690 page_cache_release(page);
2692 return VM_FAULT_OOM;
2696 * The mmap_sem must have been held on entry, and may have been
2697 * released depending on flags and vma->vm_ops->fault() return value.
2698 * See filemap_fault() and __lock_page_retry().
2700 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2701 pgoff_t pgoff, unsigned int flags, struct page **page)
2703 struct vm_fault vmf;
2706 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2711 ret = vma->vm_ops->fault(vma, &vmf);
2712 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2715 if (unlikely(PageHWPoison(vmf.page))) {
2716 if (ret & VM_FAULT_LOCKED)
2717 unlock_page(vmf.page);
2718 page_cache_release(vmf.page);
2719 return VM_FAULT_HWPOISON;
2722 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2723 lock_page(vmf.page);
2725 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2732 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2734 * @vma: virtual memory area
2735 * @address: user virtual address
2736 * @page: page to map
2737 * @pte: pointer to target page table entry
2738 * @write: true, if new entry is writable
2739 * @anon: true, if it's anonymous page
2741 * Caller must hold page table lock relevant for @pte.
2743 * Target users are page handler itself and implementations of
2744 * vm_ops->map_pages.
2746 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2747 struct page *page, pte_t *pte, bool write, bool anon)
2751 flush_icache_page(vma, page);
2752 entry = mk_pte(page, vma->vm_page_prot);
2754 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2755 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
2756 entry = pte_mksoft_dirty(entry);
2758 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2759 page_add_new_anon_rmap(page, vma, address);
2761 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2762 page_add_file_rmap(page);
2764 set_pte_at(vma->vm_mm, address, pte, entry);
2766 /* no need to invalidate: a not-present page won't be cached */
2767 update_mmu_cache(vma, address, pte);
2770 static unsigned long fault_around_bytes __read_mostly =
2771 rounddown_pow_of_two(65536);
2773 #ifdef CONFIG_DEBUG_FS
2774 static int fault_around_bytes_get(void *data, u64 *val)
2776 *val = fault_around_bytes;
2781 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2782 * rounded down to nearest page order. It's what do_fault_around() expects to
2785 static int fault_around_bytes_set(void *data, u64 val)
2787 if (val / PAGE_SIZE > PTRS_PER_PTE)
2789 if (val > PAGE_SIZE)
2790 fault_around_bytes = rounddown_pow_of_two(val);
2792 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2795 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2796 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2798 static int __init fault_around_debugfs(void)
2802 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2803 &fault_around_bytes_fops);
2805 pr_warn("Failed to create fault_around_bytes in debugfs");
2808 late_initcall(fault_around_debugfs);
2812 * do_fault_around() tries to map few pages around the fault address. The hope
2813 * is that the pages will be needed soon and this will lower the number of
2816 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2817 * not ready to be mapped: not up-to-date, locked, etc.
2819 * This function is called with the page table lock taken. In the split ptlock
2820 * case the page table lock only protects only those entries which belong to
2821 * the page table corresponding to the fault address.
2823 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2826 * fault_around_pages() defines how many pages we'll try to map.
2827 * do_fault_around() expects it to return a power of two less than or equal to
2830 * The virtual address of the area that we map is naturally aligned to the
2831 * fault_around_pages() value (and therefore to page order). This way it's
2832 * easier to guarantee that we don't cross page table boundaries.
2834 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2835 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2837 unsigned long start_addr, nr_pages, mask;
2839 struct vm_fault vmf;
2842 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2843 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2845 start_addr = max(address & mask, vma->vm_start);
2846 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2851 * max_pgoff is either end of page table or end of vma
2852 * or fault_around_pages() from pgoff, depending what is nearest.
2854 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2856 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2857 pgoff + nr_pages - 1);
2859 /* Check if it makes any sense to call ->map_pages */
2860 while (!pte_none(*pte)) {
2861 if (++pgoff > max_pgoff)
2863 start_addr += PAGE_SIZE;
2864 if (start_addr >= vma->vm_end)
2869 vmf.virtual_address = (void __user *) start_addr;
2872 vmf.max_pgoff = max_pgoff;
2874 vma->vm_ops->map_pages(vma, &vmf);
2877 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2878 unsigned long address, pmd_t *pmd,
2879 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2881 struct page *fault_page;
2887 * Let's call ->map_pages() first and use ->fault() as fallback
2888 * if page by the offset is not ready to be mapped (cold cache or
2891 if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
2892 fault_around_bytes >> PAGE_SHIFT > 1) {
2893 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2894 do_fault_around(vma, address, pte, pgoff, flags);
2895 if (!pte_same(*pte, orig_pte))
2897 pte_unmap_unlock(pte, ptl);
2900 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2901 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2904 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2905 if (unlikely(!pte_same(*pte, orig_pte))) {
2906 pte_unmap_unlock(pte, ptl);
2907 unlock_page(fault_page);
2908 page_cache_release(fault_page);
2911 do_set_pte(vma, address, fault_page, pte, false, false);
2912 unlock_page(fault_page);
2914 pte_unmap_unlock(pte, ptl);
2918 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2919 unsigned long address, pmd_t *pmd,
2920 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2922 struct page *fault_page, *new_page;
2923 struct mem_cgroup *memcg;
2928 if (unlikely(anon_vma_prepare(vma)))
2929 return VM_FAULT_OOM;
2931 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2933 return VM_FAULT_OOM;
2935 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
2936 page_cache_release(new_page);
2937 return VM_FAULT_OOM;
2940 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2941 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2944 copy_user_highpage(new_page, fault_page, address, vma);
2945 __SetPageUptodate(new_page);
2947 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2948 if (unlikely(!pte_same(*pte, orig_pte))) {
2949 pte_unmap_unlock(pte, ptl);
2950 unlock_page(fault_page);
2951 page_cache_release(fault_page);
2954 do_set_pte(vma, address, new_page, pte, true, true);
2955 mem_cgroup_commit_charge(new_page, memcg, false);
2956 lru_cache_add_active_or_unevictable(new_page, vma);
2957 pte_unmap_unlock(pte, ptl);
2958 unlock_page(fault_page);
2959 page_cache_release(fault_page);
2962 mem_cgroup_cancel_charge(new_page, memcg);
2963 page_cache_release(new_page);
2967 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2968 unsigned long address, pmd_t *pmd,
2969 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2971 struct page *fault_page;
2972 struct address_space *mapping;
2978 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2979 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2983 * Check if the backing address space wants to know that the page is
2984 * about to become writable
2986 if (vma->vm_ops->page_mkwrite) {
2987 unlock_page(fault_page);
2988 tmp = do_page_mkwrite(vma, fault_page, address);
2989 if (unlikely(!tmp ||
2990 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2991 page_cache_release(fault_page);
2996 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2997 if (unlikely(!pte_same(*pte, orig_pte))) {
2998 pte_unmap_unlock(pte, ptl);
2999 unlock_page(fault_page);
3000 page_cache_release(fault_page);
3003 do_set_pte(vma, address, fault_page, pte, true, false);
3004 pte_unmap_unlock(pte, ptl);
3006 if (set_page_dirty(fault_page))
3009 * Take a local copy of the address_space - page.mapping may be zeroed
3010 * by truncate after unlock_page(). The address_space itself remains
3011 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3012 * release semantics to prevent the compiler from undoing this copying.
3014 mapping = fault_page->mapping;
3015 unlock_page(fault_page);
3016 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3018 * Some device drivers do not set page.mapping but still
3021 balance_dirty_pages_ratelimited(mapping);
3024 /* file_update_time outside page_lock */
3025 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3026 file_update_time(vma->vm_file);
3032 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3033 * but allow concurrent faults).
3034 * The mmap_sem may have been released depending on flags and our
3035 * return value. See filemap_fault() and __lock_page_or_retry().
3037 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3038 unsigned long address, pte_t *page_table, pmd_t *pmd,
3039 unsigned int flags, pte_t orig_pte)
3041 pgoff_t pgoff = (((address & PAGE_MASK)
3042 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3044 pte_unmap(page_table);
3045 if (!(flags & FAULT_FLAG_WRITE))
3046 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3048 if (!(vma->vm_flags & VM_SHARED))
3049 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3051 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3055 * Fault of a previously existing named mapping. Repopulate the pte
3056 * from the encoded file_pte if possible. This enables swappable
3059 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3060 * but allow concurrent faults), and pte mapped but not yet locked.
3061 * We return with pte unmapped and unlocked.
3062 * The mmap_sem may have been released depending on flags and our
3063 * return value. See filemap_fault() and __lock_page_or_retry().
3065 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3066 unsigned long address, pte_t *page_table, pmd_t *pmd,
3067 unsigned int flags, pte_t orig_pte)
3071 flags |= FAULT_FLAG_NONLINEAR;
3073 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3076 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3078 * Page table corrupted: show pte and kill process.
3080 print_bad_pte(vma, address, orig_pte, NULL);
3081 return VM_FAULT_SIGBUS;
3084 pgoff = pte_to_pgoff(orig_pte);
3085 if (!(flags & FAULT_FLAG_WRITE))
3086 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3088 if (!(vma->vm_flags & VM_SHARED))
3089 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3091 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3094 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3095 unsigned long addr, int page_nid,
3100 count_vm_numa_event(NUMA_HINT_FAULTS);
3101 if (page_nid == numa_node_id()) {
3102 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3103 *flags |= TNF_FAULT_LOCAL;
3106 return mpol_misplaced(page, vma, addr);
3109 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3110 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3112 struct page *page = NULL;
3117 bool migrated = false;
3121 * The "pte" at this point cannot be used safely without
3122 * validation through pte_unmap_same(). It's of NUMA type but
3123 * the pfn may be screwed if the read is non atomic.
3125 * ptep_modify_prot_start is not called as this is clearing
3126 * the _PAGE_NUMA bit and it is not really expected that there
3127 * would be concurrent hardware modifications to the PTE.
3129 ptl = pte_lockptr(mm, pmd);
3131 if (unlikely(!pte_same(*ptep, pte))) {
3132 pte_unmap_unlock(ptep, ptl);
3136 pte = pte_mknonnuma(pte);
3137 set_pte_at(mm, addr, ptep, pte);
3138 update_mmu_cache(vma, addr, ptep);
3140 page = vm_normal_page(vma, addr, pte);
3142 pte_unmap_unlock(ptep, ptl);
3145 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3148 * Avoid grouping on DSO/COW pages in specific and RO pages
3149 * in general, RO pages shouldn't hurt as much anyway since
3150 * they can be in shared cache state.
3152 if (!pte_write(pte))
3153 flags |= TNF_NO_GROUP;
3156 * Flag if the page is shared between multiple address spaces. This
3157 * is later used when determining whether to group tasks together
3159 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3160 flags |= TNF_SHARED;
3162 last_cpupid = page_cpupid_last(page);
3163 page_nid = page_to_nid(page);
3164 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3165 pte_unmap_unlock(ptep, ptl);
3166 if (target_nid == -1) {
3171 /* Migrate to the requested node */
3172 migrated = migrate_misplaced_page(page, vma, target_nid);
3174 page_nid = target_nid;
3175 flags |= TNF_MIGRATED;
3180 task_numa_fault(last_cpupid, page_nid, 1, flags);
3185 * These routines also need to handle stuff like marking pages dirty
3186 * and/or accessed for architectures that don't do it in hardware (most
3187 * RISC architectures). The early dirtying is also good on the i386.
3189 * There is also a hook called "update_mmu_cache()" that architectures
3190 * with external mmu caches can use to update those (ie the Sparc or
3191 * PowerPC hashed page tables that act as extended TLBs).
3193 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3194 * but allow concurrent faults), and pte mapped but not yet locked.
3195 * We return with pte unmapped and unlocked.
3197 * The mmap_sem may have been released depending on flags and our
3198 * return value. See filemap_fault() and __lock_page_or_retry().
3200 static int handle_pte_fault(struct mm_struct *mm,
3201 struct vm_area_struct *vma, unsigned long address,
3202 pte_t *pte, pmd_t *pmd, unsigned int flags)
3208 * some architectures can have larger ptes than wordsize,
3209 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3210 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3211 * The code below just needs a consistent view for the ifs and
3212 * we later double check anyway with the ptl lock held. So here
3213 * a barrier will do.
3217 if (!pte_present(entry)) {
3218 if (pte_none(entry)) {
3220 if (likely(vma->vm_ops->fault))
3221 return do_linear_fault(mm, vma, address,
3222 pte, pmd, flags, entry);
3224 return do_anonymous_page(mm, vma, address,
3227 if (pte_file(entry))
3228 return do_nonlinear_fault(mm, vma, address,
3229 pte, pmd, flags, entry);
3230 return do_swap_page(mm, vma, address,
3231 pte, pmd, flags, entry);
3234 if (pte_numa(entry))
3235 return do_numa_page(mm, vma, address, entry, pte, pmd);
3237 ptl = pte_lockptr(mm, pmd);
3239 if (unlikely(!pte_same(*pte, entry)))
3241 if (flags & FAULT_FLAG_WRITE) {
3242 if (!pte_write(entry))
3243 return do_wp_page(mm, vma, address,
3244 pte, pmd, ptl, entry);
3245 entry = pte_mkdirty(entry);
3247 entry = pte_mkyoung(entry);
3248 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3249 update_mmu_cache(vma, address, pte);
3252 * This is needed only for protection faults but the arch code
3253 * is not yet telling us if this is a protection fault or not.
3254 * This still avoids useless tlb flushes for .text page faults
3257 if (flags & FAULT_FLAG_WRITE)
3258 flush_tlb_fix_spurious_fault(vma, address);
3261 pte_unmap_unlock(pte, ptl);
3266 * By the time we get here, we already hold the mm semaphore
3268 * The mmap_sem may have been released depending on flags and our
3269 * return value. See filemap_fault() and __lock_page_or_retry().
3271 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3272 unsigned long address, unsigned int flags)
3279 if (unlikely(is_vm_hugetlb_page(vma)))
3280 return hugetlb_fault(mm, vma, address, flags);
3282 pgd = pgd_offset(mm, address);
3283 pud = pud_alloc(mm, pgd, address);
3285 return VM_FAULT_OOM;
3286 pmd = pmd_alloc(mm, pud, address);
3288 return VM_FAULT_OOM;
3289 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3290 int ret = VM_FAULT_FALLBACK;
3292 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3294 if (!(ret & VM_FAULT_FALLBACK))
3297 pmd_t orig_pmd = *pmd;
3301 if (pmd_trans_huge(orig_pmd)) {
3302 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3305 * If the pmd is splitting, return and retry the
3306 * the fault. Alternative: wait until the split
3307 * is done, and goto retry.
3309 if (pmd_trans_splitting(orig_pmd))
3312 if (pmd_numa(orig_pmd))
3313 return do_huge_pmd_numa_page(mm, vma, address,
3316 if (dirty && !pmd_write(orig_pmd)) {
3317 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3319 if (!(ret & VM_FAULT_FALLBACK))
3322 huge_pmd_set_accessed(mm, vma, address, pmd,
3330 * Use __pte_alloc instead of pte_alloc_map, because we can't
3331 * run pte_offset_map on the pmd, if an huge pmd could
3332 * materialize from under us from a different thread.
3334 if (unlikely(pmd_none(*pmd)) &&
3335 unlikely(__pte_alloc(mm, vma, pmd, address)))
3336 return VM_FAULT_OOM;
3337 /* if an huge pmd materialized from under us just retry later */
3338 if (unlikely(pmd_trans_huge(*pmd)))
3341 * A regular pmd is established and it can't morph into a huge pmd
3342 * from under us anymore at this point because we hold the mmap_sem
3343 * read mode and khugepaged takes it in write mode. So now it's
3344 * safe to run pte_offset_map().
3346 pte = pte_offset_map(pmd, address);
3348 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3352 * By the time we get here, we already hold the mm semaphore
3354 * The mmap_sem may have been released depending on flags and our
3355 * return value. See filemap_fault() and __lock_page_or_retry().
3357 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3358 unsigned long address, unsigned int flags)
3362 __set_current_state(TASK_RUNNING);
3364 count_vm_event(PGFAULT);
3365 mem_cgroup_count_vm_event(mm, PGFAULT);
3367 /* do counter updates before entering really critical section. */
3368 check_sync_rss_stat(current);
3371 * Enable the memcg OOM handling for faults triggered in user
3372 * space. Kernel faults are handled more gracefully.
3374 if (flags & FAULT_FLAG_USER)
3375 mem_cgroup_oom_enable();
3377 ret = __handle_mm_fault(mm, vma, address, flags);
3379 if (flags & FAULT_FLAG_USER) {
3380 mem_cgroup_oom_disable();
3382 * The task may have entered a memcg OOM situation but
3383 * if the allocation error was handled gracefully (no
3384 * VM_FAULT_OOM), there is no need to kill anything.
3385 * Just clean up the OOM state peacefully.
3387 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3388 mem_cgroup_oom_synchronize(false);
3393 EXPORT_SYMBOL_GPL(handle_mm_fault);
3395 #ifndef __PAGETABLE_PUD_FOLDED
3397 * Allocate page upper directory.
3398 * We've already handled the fast-path in-line.
3400 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3402 pud_t *new = pud_alloc_one(mm, address);
3406 smp_wmb(); /* See comment in __pte_alloc */
3408 spin_lock(&mm->page_table_lock);
3409 if (pgd_present(*pgd)) /* Another has populated it */
3412 pgd_populate(mm, pgd, new);
3413 spin_unlock(&mm->page_table_lock);
3416 #endif /* __PAGETABLE_PUD_FOLDED */
3418 #ifndef __PAGETABLE_PMD_FOLDED
3420 * Allocate page middle directory.
3421 * We've already handled the fast-path in-line.
3423 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3425 pmd_t *new = pmd_alloc_one(mm, address);
3429 smp_wmb(); /* See comment in __pte_alloc */
3431 spin_lock(&mm->page_table_lock);
3432 #ifndef __ARCH_HAS_4LEVEL_HACK
3433 if (pud_present(*pud)) /* Another has populated it */
3436 pud_populate(mm, pud, new);
3438 if (pgd_present(*pud)) /* Another has populated it */
3441 pgd_populate(mm, pud, new);
3442 #endif /* __ARCH_HAS_4LEVEL_HACK */
3443 spin_unlock(&mm->page_table_lock);
3446 #endif /* __PAGETABLE_PMD_FOLDED */
3448 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3449 pte_t **ptepp, spinlock_t **ptlp)
3456 pgd = pgd_offset(mm, address);
3457 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3460 pud = pud_offset(pgd, address);
3461 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3464 pmd = pmd_offset(pud, address);
3465 VM_BUG_ON(pmd_trans_huge(*pmd));
3466 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3469 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3473 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3476 if (!pte_present(*ptep))
3481 pte_unmap_unlock(ptep, *ptlp);
3486 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3487 pte_t **ptepp, spinlock_t **ptlp)
3491 /* (void) is needed to make gcc happy */
3492 (void) __cond_lock(*ptlp,
3493 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3498 * follow_pfn - look up PFN at a user virtual address
3499 * @vma: memory mapping
3500 * @address: user virtual address
3501 * @pfn: location to store found PFN
3503 * Only IO mappings and raw PFN mappings are allowed.
3505 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3507 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3514 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3517 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3520 *pfn = pte_pfn(*ptep);
3521 pte_unmap_unlock(ptep, ptl);
3524 EXPORT_SYMBOL(follow_pfn);
3526 #ifdef CONFIG_HAVE_IOREMAP_PROT
3527 int follow_phys(struct vm_area_struct *vma,
3528 unsigned long address, unsigned int flags,
3529 unsigned long *prot, resource_size_t *phys)
3535 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3538 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3542 if ((flags & FOLL_WRITE) && !pte_write(pte))
3545 *prot = pgprot_val(pte_pgprot(pte));
3546 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3550 pte_unmap_unlock(ptep, ptl);
3555 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3556 void *buf, int len, int write)
3558 resource_size_t phys_addr;
3559 unsigned long prot = 0;
3560 void __iomem *maddr;
3561 int offset = addr & (PAGE_SIZE-1);
3563 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3566 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3568 memcpy_toio(maddr + offset, buf, len);
3570 memcpy_fromio(buf, maddr + offset, len);
3575 EXPORT_SYMBOL_GPL(generic_access_phys);
3579 * Access another process' address space as given in mm. If non-NULL, use the
3580 * given task for page fault accounting.
3582 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3583 unsigned long addr, void *buf, int len, int write)
3585 struct vm_area_struct *vma;
3586 void *old_buf = buf;
3588 down_read(&mm->mmap_sem);
3589 /* ignore errors, just check how much was successfully transferred */
3591 int bytes, ret, offset;
3593 struct page *page = NULL;
3595 ret = get_user_pages(tsk, mm, addr, 1,
3596 write, 1, &page, &vma);
3598 #ifndef CONFIG_HAVE_IOREMAP_PROT
3602 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3603 * we can access using slightly different code.
3605 vma = find_vma(mm, addr);
3606 if (!vma || vma->vm_start > addr)
3608 if (vma->vm_ops && vma->vm_ops->access)
3609 ret = vma->vm_ops->access(vma, addr, buf,
3617 offset = addr & (PAGE_SIZE-1);
3618 if (bytes > PAGE_SIZE-offset)
3619 bytes = PAGE_SIZE-offset;
3623 copy_to_user_page(vma, page, addr,
3624 maddr + offset, buf, bytes);
3625 set_page_dirty_lock(page);
3627 copy_from_user_page(vma, page, addr,
3628 buf, maddr + offset, bytes);
3631 page_cache_release(page);
3637 up_read(&mm->mmap_sem);
3639 return buf - old_buf;
3643 * access_remote_vm - access another process' address space
3644 * @mm: the mm_struct of the target address space
3645 * @addr: start address to access
3646 * @buf: source or destination buffer
3647 * @len: number of bytes to transfer
3648 * @write: whether the access is a write
3650 * The caller must hold a reference on @mm.
3652 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3653 void *buf, int len, int write)
3655 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3659 * Access another process' address space.
3660 * Source/target buffer must be kernel space,
3661 * Do not walk the page table directly, use get_user_pages
3663 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3664 void *buf, int len, int write)
3666 struct mm_struct *mm;
3669 mm = get_task_mm(tsk);
3673 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3680 * Print the name of a VMA.
3682 void print_vma_addr(char *prefix, unsigned long ip)
3684 struct mm_struct *mm = current->mm;
3685 struct vm_area_struct *vma;
3688 * Do not print if we are in atomic
3689 * contexts (in exception stacks, etc.):
3691 if (preempt_count())
3694 down_read(&mm->mmap_sem);
3695 vma = find_vma(mm, ip);
3696 if (vma && vma->vm_file) {
3697 struct file *f = vma->vm_file;
3698 char *buf = (char *)__get_free_page(GFP_KERNEL);
3702 p = d_path(&f->f_path, buf, PAGE_SIZE);
3705 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3707 vma->vm_end - vma->vm_start);
3708 free_page((unsigned long)buf);
3711 up_read(&mm->mmap_sem);
3714 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3715 void might_fault(void)
3718 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3719 * holding the mmap_sem, this is safe because kernel memory doesn't
3720 * get paged out, therefore we'll never actually fault, and the
3721 * below annotations will generate false positives.
3723 if (segment_eq(get_fs(), KERNEL_DS))
3727 * it would be nicer only to annotate paths which are not under
3728 * pagefault_disable, however that requires a larger audit and
3729 * providing helpers like get_user_atomic.
3734 __might_sleep(__FILE__, __LINE__, 0);
3737 might_lock_read(¤t->mm->mmap_sem);
3739 EXPORT_SYMBOL(might_fault);
3742 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3743 static void clear_gigantic_page(struct page *page,
3745 unsigned int pages_per_huge_page)
3748 struct page *p = page;
3751 for (i = 0; i < pages_per_huge_page;
3752 i++, p = mem_map_next(p, page, i)) {
3754 clear_user_highpage(p, addr + i * PAGE_SIZE);
3757 void clear_huge_page(struct page *page,
3758 unsigned long addr, unsigned int pages_per_huge_page)
3762 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3763 clear_gigantic_page(page, addr, pages_per_huge_page);
3768 for (i = 0; i < pages_per_huge_page; i++) {
3770 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3774 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3776 struct vm_area_struct *vma,
3777 unsigned int pages_per_huge_page)
3780 struct page *dst_base = dst;
3781 struct page *src_base = src;
3783 for (i = 0; i < pages_per_huge_page; ) {
3785 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3788 dst = mem_map_next(dst, dst_base, i);
3789 src = mem_map_next(src, src_base, i);
3793 void copy_user_huge_page(struct page *dst, struct page *src,
3794 unsigned long addr, struct vm_area_struct *vma,
3795 unsigned int pages_per_huge_page)
3799 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3800 copy_user_gigantic_page(dst, src, addr, vma,
3801 pages_per_huge_page);
3806 for (i = 0; i < pages_per_huge_page; i++) {
3808 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3811 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3813 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3815 static struct kmem_cache *page_ptl_cachep;
3817 void __init ptlock_cache_init(void)
3819 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3823 bool ptlock_alloc(struct page *page)
3827 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3834 void ptlock_free(struct page *page)
3836 kmem_cache_free(page_ptl_cachep, page->ptl);