4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
84 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
94 EXPORT_SYMBOL(mem_map);
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 EXPORT_SYMBOL(high_memory);
108 * Randomize the address space (stacks, mmaps, brk, etc.).
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
120 static int __init disable_randmaps(char *s)
122 randomize_va_space = 0;
125 __setup("norandmaps", disable_randmaps);
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
130 unsigned long highest_memmap_pfn __read_mostly;
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 static int __init init_zero_pfn(void)
137 zero_pfn = page_to_pfn(ZERO_PAGE(0));
140 core_initcall(init_zero_pfn);
143 #if defined(SPLIT_RSS_COUNTING)
145 void sync_mm_rss(struct mm_struct *mm)
149 for (i = 0; i < NR_MM_COUNTERS; i++) {
150 if (current->rss_stat.count[i]) {
151 add_mm_counter(mm, i, current->rss_stat.count[i]);
152 current->rss_stat.count[i] = 0;
155 current->rss_stat.events = 0;
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160 struct task_struct *task = current;
162 if (likely(task->mm == mm))
163 task->rss_stat.count[member] += val;
165 add_mm_counter(mm, member, val);
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH (64)
172 static void check_sync_rss_stat(struct task_struct *task)
174 if (unlikely(task != current))
176 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177 sync_mm_rss(task->mm);
179 #else /* SPLIT_RSS_COUNTING */
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184 static void check_sync_rss_stat(struct task_struct *task)
188 #endif /* SPLIT_RSS_COUNTING */
190 #ifdef HAVE_GENERIC_MMU_GATHER
192 static bool tlb_next_batch(struct mmu_gather *tlb)
194 struct mmu_gather_batch *batch;
198 tlb->active = batch->next;
202 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
205 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
212 batch->max = MAX_GATHER_BATCH;
214 tlb->active->next = batch;
220 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
221 unsigned long start, unsigned long end)
225 /* Is it from 0 to ~0? */
226 tlb->fullmm = !(start | (end+1));
227 tlb->need_flush_all = 0;
228 tlb->local.next = NULL;
230 tlb->local.max = ARRAY_SIZE(tlb->__pages);
231 tlb->active = &tlb->local;
232 tlb->batch_count = 0;
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 __tlb_reset_range(tlb);
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
248 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
249 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
250 tlb_table_flush(tlb);
252 __tlb_reset_range(tlb);
255 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
257 struct mmu_gather_batch *batch;
259 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
263 tlb->active = &tlb->local;
266 void tlb_flush_mmu(struct mmu_gather *tlb)
268 tlb_flush_mmu_tlbonly(tlb);
269 tlb_flush_mmu_free(tlb);
273 * Called at the end of the shootdown operation to free up any resources
274 * that were required.
276 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
277 unsigned long start, unsigned long end, bool force)
279 struct mmu_gather_batch *batch, *next;
282 __tlb_adjust_range(tlb, start, end - start);
286 /* keep the page table cache within bounds */
289 for (batch = tlb->local.next; batch; batch = next) {
291 free_pages((unsigned long)batch, 0);
293 tlb->local.next = NULL;
297 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298 * handling the additional races in SMP caused by other CPUs caching valid
299 * mappings in their TLBs. Returns the number of free page slots left.
300 * When out of page slots we must call tlb_flush_mmu().
301 *returns true if the caller should flush.
303 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
305 struct mmu_gather_batch *batch;
307 VM_BUG_ON(!tlb->end);
308 VM_WARN_ON(tlb->page_size != page_size);
312 * Add the page and check if we are full. If so
315 batch->pages[batch->nr++] = page;
316 if (batch->nr == batch->max) {
317 if (!tlb_next_batch(tlb))
321 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
326 #endif /* HAVE_GENERIC_MMU_GATHER */
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
331 * See the comment near struct mmu_table_batch.
334 static void tlb_remove_table_smp_sync(void *arg)
336 /* Simply deliver the interrupt */
339 static void tlb_remove_table_one(void *table)
342 * This isn't an RCU grace period and hence the page-tables cannot be
343 * assumed to be actually RCU-freed.
345 * It is however sufficient for software page-table walkers that rely on
346 * IRQ disabling. See the comment near struct mmu_table_batch.
348 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
349 __tlb_remove_table(table);
352 static void tlb_remove_table_rcu(struct rcu_head *head)
354 struct mmu_table_batch *batch;
357 batch = container_of(head, struct mmu_table_batch, rcu);
359 for (i = 0; i < batch->nr; i++)
360 __tlb_remove_table(batch->tables[i]);
362 free_page((unsigned long)batch);
365 void tlb_table_flush(struct mmu_gather *tlb)
367 struct mmu_table_batch **batch = &tlb->batch;
370 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
375 void tlb_remove_table(struct mmu_gather *tlb, void *table)
377 struct mmu_table_batch **batch = &tlb->batch;
380 * When there's less then two users of this mm there cannot be a
381 * concurrent page-table walk.
383 if (atomic_read(&tlb->mm->mm_users) < 2) {
384 __tlb_remove_table(table);
388 if (*batch == NULL) {
389 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
390 if (*batch == NULL) {
391 tlb_remove_table_one(table);
396 (*batch)->tables[(*batch)->nr++] = table;
397 if ((*batch)->nr == MAX_TABLE_BATCH)
398 tlb_table_flush(tlb);
401 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
404 * Called to initialize an (on-stack) mmu_gather structure for page-table
405 * tear-down from @mm. The @fullmm argument is used when @mm is without
406 * users and we're going to destroy the full address space (exit/execve).
408 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
409 unsigned long start, unsigned long end)
411 arch_tlb_gather_mmu(tlb, mm, start, end);
412 inc_tlb_flush_pending(tlb->mm);
415 void tlb_finish_mmu(struct mmu_gather *tlb,
416 unsigned long start, unsigned long end)
419 * If there are parallel threads are doing PTE changes on same range
420 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
421 * flush by batching, a thread has stable TLB entry can fail to flush
422 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
423 * forcefully if we detect parallel PTE batching threads.
425 bool force = mm_tlb_flush_nested(tlb->mm);
427 arch_tlb_finish_mmu(tlb, start, end, force);
428 dec_tlb_flush_pending(tlb->mm);
432 * Note: this doesn't free the actual pages themselves. That
433 * has been handled earlier when unmapping all the memory regions.
435 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
438 pgtable_t token = pmd_pgtable(*pmd);
440 pte_free_tlb(tlb, token, addr);
441 mm_dec_nr_ptes(tlb->mm);
444 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
445 unsigned long addr, unsigned long end,
446 unsigned long floor, unsigned long ceiling)
453 pmd = pmd_offset(pud, addr);
455 next = pmd_addr_end(addr, end);
456 if (pmd_none_or_clear_bad(pmd))
458 free_pte_range(tlb, pmd, addr);
459 } while (pmd++, addr = next, addr != end);
469 if (end - 1 > ceiling - 1)
472 pmd = pmd_offset(pud, start);
474 pmd_free_tlb(tlb, pmd, start);
475 mm_dec_nr_pmds(tlb->mm);
478 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
479 unsigned long addr, unsigned long end,
480 unsigned long floor, unsigned long ceiling)
487 pud = pud_offset(p4d, addr);
489 next = pud_addr_end(addr, end);
490 if (pud_none_or_clear_bad(pud))
492 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
493 } while (pud++, addr = next, addr != end);
503 if (end - 1 > ceiling - 1)
506 pud = pud_offset(p4d, start);
508 pud_free_tlb(tlb, pud, start);
509 mm_dec_nr_puds(tlb->mm);
512 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
513 unsigned long addr, unsigned long end,
514 unsigned long floor, unsigned long ceiling)
521 p4d = p4d_offset(pgd, addr);
523 next = p4d_addr_end(addr, end);
524 if (p4d_none_or_clear_bad(p4d))
526 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
527 } while (p4d++, addr = next, addr != end);
533 ceiling &= PGDIR_MASK;
537 if (end - 1 > ceiling - 1)
540 p4d = p4d_offset(pgd, start);
542 p4d_free_tlb(tlb, p4d, start);
546 * This function frees user-level page tables of a process.
548 void free_pgd_range(struct mmu_gather *tlb,
549 unsigned long addr, unsigned long end,
550 unsigned long floor, unsigned long ceiling)
556 * The next few lines have given us lots of grief...
558 * Why are we testing PMD* at this top level? Because often
559 * there will be no work to do at all, and we'd prefer not to
560 * go all the way down to the bottom just to discover that.
562 * Why all these "- 1"s? Because 0 represents both the bottom
563 * of the address space and the top of it (using -1 for the
564 * top wouldn't help much: the masks would do the wrong thing).
565 * The rule is that addr 0 and floor 0 refer to the bottom of
566 * the address space, but end 0 and ceiling 0 refer to the top
567 * Comparisons need to use "end - 1" and "ceiling - 1" (though
568 * that end 0 case should be mythical).
570 * Wherever addr is brought up or ceiling brought down, we must
571 * be careful to reject "the opposite 0" before it confuses the
572 * subsequent tests. But what about where end is brought down
573 * by PMD_SIZE below? no, end can't go down to 0 there.
575 * Whereas we round start (addr) and ceiling down, by different
576 * masks at different levels, in order to test whether a table
577 * now has no other vmas using it, so can be freed, we don't
578 * bother to round floor or end up - the tests don't need that.
592 if (end - 1 > ceiling - 1)
597 * We add page table cache pages with PAGE_SIZE,
598 * (see pte_free_tlb()), flush the tlb if we need
600 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
601 pgd = pgd_offset(tlb->mm, addr);
603 next = pgd_addr_end(addr, end);
604 if (pgd_none_or_clear_bad(pgd))
606 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
607 } while (pgd++, addr = next, addr != end);
610 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
611 unsigned long floor, unsigned long ceiling)
614 struct vm_area_struct *next = vma->vm_next;
615 unsigned long addr = vma->vm_start;
618 * Hide vma from rmap and truncate_pagecache before freeing
621 unlink_anon_vmas(vma);
622 unlink_file_vma(vma);
624 if (is_vm_hugetlb_page(vma)) {
625 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
626 floor, next ? next->vm_start : ceiling);
629 * Optimization: gather nearby vmas into one call down
631 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
632 && !is_vm_hugetlb_page(next)) {
635 unlink_anon_vmas(vma);
636 unlink_file_vma(vma);
638 free_pgd_range(tlb, addr, vma->vm_end,
639 floor, next ? next->vm_start : ceiling);
645 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
648 pgtable_t new = pte_alloc_one(mm, address);
653 * Ensure all pte setup (eg. pte page lock and page clearing) are
654 * visible before the pte is made visible to other CPUs by being
655 * put into page tables.
657 * The other side of the story is the pointer chasing in the page
658 * table walking code (when walking the page table without locking;
659 * ie. most of the time). Fortunately, these data accesses consist
660 * of a chain of data-dependent loads, meaning most CPUs (alpha
661 * being the notable exception) will already guarantee loads are
662 * seen in-order. See the alpha page table accessors for the
663 * smp_read_barrier_depends() barriers in page table walking code.
665 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
667 ptl = pmd_lock(mm, pmd);
668 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
670 pmd_populate(mm, pmd, new);
679 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
681 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
685 smp_wmb(); /* See comment in __pte_alloc */
687 spin_lock(&init_mm.page_table_lock);
688 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
689 pmd_populate_kernel(&init_mm, pmd, new);
692 spin_unlock(&init_mm.page_table_lock);
694 pte_free_kernel(&init_mm, new);
698 static inline void init_rss_vec(int *rss)
700 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
703 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
707 if (current->mm == mm)
709 for (i = 0; i < NR_MM_COUNTERS; i++)
711 add_mm_counter(mm, i, rss[i]);
715 * This function is called to print an error when a bad pte
716 * is found. For example, we might have a PFN-mapped pte in
717 * a region that doesn't allow it.
719 * The calling function must still handle the error.
721 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
722 pte_t pte, struct page *page)
724 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
725 p4d_t *p4d = p4d_offset(pgd, addr);
726 pud_t *pud = pud_offset(p4d, addr);
727 pmd_t *pmd = pmd_offset(pud, addr);
728 struct address_space *mapping;
730 static unsigned long resume;
731 static unsigned long nr_shown;
732 static unsigned long nr_unshown;
735 * Allow a burst of 60 reports, then keep quiet for that minute;
736 * or allow a steady drip of one report per second.
738 if (nr_shown == 60) {
739 if (time_before(jiffies, resume)) {
744 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
751 resume = jiffies + 60 * HZ;
753 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
754 index = linear_page_index(vma, addr);
756 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
758 (long long)pte_val(pte), (long long)pmd_val(*pmd));
760 dump_page(page, "bad pte");
761 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
762 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
764 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
766 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
768 vma->vm_ops ? vma->vm_ops->fault : NULL,
769 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
770 mapping ? mapping->a_ops->readpage : NULL);
772 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
776 * vm_normal_page -- This function gets the "struct page" associated with a pte.
778 * "Special" mappings do not wish to be associated with a "struct page" (either
779 * it doesn't exist, or it exists but they don't want to touch it). In this
780 * case, NULL is returned here. "Normal" mappings do have a struct page.
782 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
783 * pte bit, in which case this function is trivial. Secondly, an architecture
784 * may not have a spare pte bit, which requires a more complicated scheme,
787 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
788 * special mapping (even if there are underlying and valid "struct pages").
789 * COWed pages of a VM_PFNMAP are always normal.
791 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
792 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
793 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
794 * mapping will always honor the rule
796 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
798 * And for normal mappings this is false.
800 * This restricts such mappings to be a linear translation from virtual address
801 * to pfn. To get around this restriction, we allow arbitrary mappings so long
802 * as the vma is not a COW mapping; in that case, we know that all ptes are
803 * special (because none can have been COWed).
806 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
808 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
809 * page" backing, however the difference is that _all_ pages with a struct
810 * page (that is, those where pfn_valid is true) are refcounted and considered
811 * normal pages by the VM. The disadvantage is that pages are refcounted
812 * (which can be slower and simply not an option for some PFNMAP users). The
813 * advantage is that we don't have to follow the strict linearity rule of
814 * PFNMAP mappings in order to support COWable mappings.
817 #ifdef __HAVE_ARCH_PTE_SPECIAL
818 # define HAVE_PTE_SPECIAL 1
820 # define HAVE_PTE_SPECIAL 0
822 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
823 pte_t pte, bool with_public_device)
825 unsigned long pfn = pte_pfn(pte);
827 if (HAVE_PTE_SPECIAL) {
828 if (likely(!pte_special(pte)))
830 if (vma->vm_ops && vma->vm_ops->find_special_page)
831 return vma->vm_ops->find_special_page(vma, addr);
832 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
834 if (is_zero_pfn(pfn))
838 * Device public pages are special pages (they are ZONE_DEVICE
839 * pages but different from persistent memory). They behave
840 * allmost like normal pages. The difference is that they are
841 * not on the lru and thus should never be involve with any-
842 * thing that involve lru manipulation (mlock, numa balancing,
845 * This is why we still want to return NULL for such page from
846 * vm_normal_page() so that we do not have to special case all
847 * call site of vm_normal_page().
849 if (likely(pfn <= highest_memmap_pfn)) {
850 struct page *page = pfn_to_page(pfn);
852 if (is_device_public_page(page)) {
853 if (with_public_device)
858 print_bad_pte(vma, addr, pte, NULL);
862 /* !HAVE_PTE_SPECIAL case follows: */
864 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
865 if (vma->vm_flags & VM_MIXEDMAP) {
871 off = (addr - vma->vm_start) >> PAGE_SHIFT;
872 if (pfn == vma->vm_pgoff + off)
874 if (!is_cow_mapping(vma->vm_flags))
879 if (is_zero_pfn(pfn))
882 if (unlikely(pfn > highest_memmap_pfn)) {
883 print_bad_pte(vma, addr, pte, NULL);
888 * NOTE! We still have PageReserved() pages in the page tables.
889 * eg. VDSO mappings can cause them to exist.
892 return pfn_to_page(pfn);
895 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
896 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
899 unsigned long pfn = pmd_pfn(pmd);
902 * There is no pmd_special() but there may be special pmds, e.g.
903 * in a direct-access (dax) mapping, so let's just replicate the
904 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
906 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
907 if (vma->vm_flags & VM_MIXEDMAP) {
913 off = (addr - vma->vm_start) >> PAGE_SHIFT;
914 if (pfn == vma->vm_pgoff + off)
916 if (!is_cow_mapping(vma->vm_flags))
921 if (is_zero_pfn(pfn))
923 if (unlikely(pfn > highest_memmap_pfn))
927 * NOTE! We still have PageReserved() pages in the page tables.
928 * eg. VDSO mappings can cause them to exist.
931 return pfn_to_page(pfn);
936 * copy one vm_area from one task to the other. Assumes the page tables
937 * already present in the new task to be cleared in the whole range
938 * covered by this vma.
941 static inline unsigned long
942 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
943 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
944 unsigned long addr, int *rss)
946 unsigned long vm_flags = vma->vm_flags;
947 pte_t pte = *src_pte;
950 /* pte contains position in swap or file, so copy. */
951 if (unlikely(!pte_present(pte))) {
952 swp_entry_t entry = pte_to_swp_entry(pte);
954 if (likely(!non_swap_entry(entry))) {
955 if (swap_duplicate(entry) < 0)
958 /* make sure dst_mm is on swapoff's mmlist. */
959 if (unlikely(list_empty(&dst_mm->mmlist))) {
960 spin_lock(&mmlist_lock);
961 if (list_empty(&dst_mm->mmlist))
962 list_add(&dst_mm->mmlist,
964 spin_unlock(&mmlist_lock);
967 } else if (is_migration_entry(entry)) {
968 page = migration_entry_to_page(entry);
970 rss[mm_counter(page)]++;
972 if (is_write_migration_entry(entry) &&
973 is_cow_mapping(vm_flags)) {
975 * COW mappings require pages in both
976 * parent and child to be set to read.
978 make_migration_entry_read(&entry);
979 pte = swp_entry_to_pte(entry);
980 if (pte_swp_soft_dirty(*src_pte))
981 pte = pte_swp_mksoft_dirty(pte);
982 set_pte_at(src_mm, addr, src_pte, pte);
984 } else if (is_device_private_entry(entry)) {
985 page = device_private_entry_to_page(entry);
988 * Update rss count even for unaddressable pages, as
989 * they should treated just like normal pages in this
992 * We will likely want to have some new rss counters
993 * for unaddressable pages, at some point. But for now
994 * keep things as they are.
997 rss[mm_counter(page)]++;
998 page_dup_rmap(page, false);
1001 * We do not preserve soft-dirty information, because so
1002 * far, checkpoint/restore is the only feature that
1003 * requires that. And checkpoint/restore does not work
1004 * when a device driver is involved (you cannot easily
1005 * save and restore device driver state).
1007 if (is_write_device_private_entry(entry) &&
1008 is_cow_mapping(vm_flags)) {
1009 make_device_private_entry_read(&entry);
1010 pte = swp_entry_to_pte(entry);
1011 set_pte_at(src_mm, addr, src_pte, pte);
1018 * If it's a COW mapping, write protect it both
1019 * in the parent and the child
1021 if (is_cow_mapping(vm_flags)) {
1022 ptep_set_wrprotect(src_mm, addr, src_pte);
1023 pte = pte_wrprotect(pte);
1027 * If it's a shared mapping, mark it clean in
1030 if (vm_flags & VM_SHARED)
1031 pte = pte_mkclean(pte);
1032 pte = pte_mkold(pte);
1034 page = vm_normal_page(vma, addr, pte);
1037 page_dup_rmap(page, false);
1038 rss[mm_counter(page)]++;
1039 } else if (pte_devmap(pte)) {
1040 page = pte_page(pte);
1043 * Cache coherent device memory behave like regular page and
1044 * not like persistent memory page. For more informations see
1045 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1047 if (is_device_public_page(page)) {
1049 page_dup_rmap(page, false);
1050 rss[mm_counter(page)]++;
1055 set_pte_at(dst_mm, addr, dst_pte, pte);
1059 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1060 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1061 unsigned long addr, unsigned long end)
1063 pte_t *orig_src_pte, *orig_dst_pte;
1064 pte_t *src_pte, *dst_pte;
1065 spinlock_t *src_ptl, *dst_ptl;
1067 int rss[NR_MM_COUNTERS];
1068 swp_entry_t entry = (swp_entry_t){0};
1073 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1076 src_pte = pte_offset_map(src_pmd, addr);
1077 src_ptl = pte_lockptr(src_mm, src_pmd);
1078 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1079 orig_src_pte = src_pte;
1080 orig_dst_pte = dst_pte;
1081 arch_enter_lazy_mmu_mode();
1085 * We are holding two locks at this point - either of them
1086 * could generate latencies in another task on another CPU.
1088 if (progress >= 32) {
1090 if (need_resched() ||
1091 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1094 if (pte_none(*src_pte)) {
1098 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1103 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1105 arch_leave_lazy_mmu_mode();
1106 spin_unlock(src_ptl);
1107 pte_unmap(orig_src_pte);
1108 add_mm_rss_vec(dst_mm, rss);
1109 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1113 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1122 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1123 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1124 unsigned long addr, unsigned long end)
1126 pmd_t *src_pmd, *dst_pmd;
1129 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1132 src_pmd = pmd_offset(src_pud, addr);
1134 next = pmd_addr_end(addr, end);
1135 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1136 || pmd_devmap(*src_pmd)) {
1138 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1139 err = copy_huge_pmd(dst_mm, src_mm,
1140 dst_pmd, src_pmd, addr, vma);
1147 if (pmd_none_or_clear_bad(src_pmd))
1149 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1152 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1156 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1157 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1158 unsigned long addr, unsigned long end)
1160 pud_t *src_pud, *dst_pud;
1163 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1166 src_pud = pud_offset(src_p4d, addr);
1168 next = pud_addr_end(addr, end);
1169 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1172 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1173 err = copy_huge_pud(dst_mm, src_mm,
1174 dst_pud, src_pud, addr, vma);
1181 if (pud_none_or_clear_bad(src_pud))
1183 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1186 } while (dst_pud++, src_pud++, addr = next, addr != end);
1190 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1191 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1192 unsigned long addr, unsigned long end)
1194 p4d_t *src_p4d, *dst_p4d;
1197 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1200 src_p4d = p4d_offset(src_pgd, addr);
1202 next = p4d_addr_end(addr, end);
1203 if (p4d_none_or_clear_bad(src_p4d))
1205 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1208 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1212 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1213 struct vm_area_struct *vma)
1215 pgd_t *src_pgd, *dst_pgd;
1217 unsigned long addr = vma->vm_start;
1218 unsigned long end = vma->vm_end;
1219 unsigned long mmun_start; /* For mmu_notifiers */
1220 unsigned long mmun_end; /* For mmu_notifiers */
1225 * Don't copy ptes where a page fault will fill them correctly.
1226 * Fork becomes much lighter when there are big shared or private
1227 * readonly mappings. The tradeoff is that copy_page_range is more
1228 * efficient than faulting.
1230 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1234 if (is_vm_hugetlb_page(vma))
1235 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1237 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1239 * We do not free on error cases below as remove_vma
1240 * gets called on error from higher level routine
1242 ret = track_pfn_copy(vma);
1248 * We need to invalidate the secondary MMU mappings only when
1249 * there could be a permission downgrade on the ptes of the
1250 * parent mm. And a permission downgrade will only happen if
1251 * is_cow_mapping() returns true.
1253 is_cow = is_cow_mapping(vma->vm_flags);
1257 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1261 dst_pgd = pgd_offset(dst_mm, addr);
1262 src_pgd = pgd_offset(src_mm, addr);
1264 next = pgd_addr_end(addr, end);
1265 if (pgd_none_or_clear_bad(src_pgd))
1267 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1268 vma, addr, next))) {
1272 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1275 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1279 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1280 struct vm_area_struct *vma, pmd_t *pmd,
1281 unsigned long addr, unsigned long end,
1282 struct zap_details *details)
1284 struct mm_struct *mm = tlb->mm;
1285 int force_flush = 0;
1286 int rss[NR_MM_COUNTERS];
1292 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1295 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1297 flush_tlb_batched_pending(mm);
1298 arch_enter_lazy_mmu_mode();
1301 if (pte_none(ptent))
1304 if (pte_present(ptent)) {
1307 page = _vm_normal_page(vma, addr, ptent, true);
1308 if (unlikely(details) && page) {
1310 * unmap_shared_mapping_pages() wants to
1311 * invalidate cache without truncating:
1312 * unmap shared but keep private pages.
1314 if (details->check_mapping &&
1315 details->check_mapping != page_rmapping(page))
1318 ptent = ptep_get_and_clear_full(mm, addr, pte,
1320 tlb_remove_tlb_entry(tlb, pte, addr);
1321 if (unlikely(!page))
1324 if (!PageAnon(page)) {
1325 if (pte_dirty(ptent)) {
1327 set_page_dirty(page);
1329 if (pte_young(ptent) &&
1330 likely(!(vma->vm_flags & VM_SEQ_READ)))
1331 mark_page_accessed(page);
1333 rss[mm_counter(page)]--;
1334 page_remove_rmap(page, false);
1335 if (unlikely(page_mapcount(page) < 0))
1336 print_bad_pte(vma, addr, ptent, page);
1337 if (unlikely(__tlb_remove_page(tlb, page))) {
1345 entry = pte_to_swp_entry(ptent);
1346 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1347 struct page *page = device_private_entry_to_page(entry);
1349 if (unlikely(details && details->check_mapping)) {
1351 * unmap_shared_mapping_pages() wants to
1352 * invalidate cache without truncating:
1353 * unmap shared but keep private pages.
1355 if (details->check_mapping !=
1356 page_rmapping(page))
1360 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1361 rss[mm_counter(page)]--;
1362 page_remove_rmap(page, false);
1367 /* If details->check_mapping, we leave swap entries. */
1368 if (unlikely(details))
1371 entry = pte_to_swp_entry(ptent);
1372 if (!non_swap_entry(entry))
1374 else if (is_migration_entry(entry)) {
1377 page = migration_entry_to_page(entry);
1378 rss[mm_counter(page)]--;
1380 if (unlikely(!free_swap_and_cache(entry)))
1381 print_bad_pte(vma, addr, ptent, NULL);
1382 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1383 } while (pte++, addr += PAGE_SIZE, addr != end);
1385 add_mm_rss_vec(mm, rss);
1386 arch_leave_lazy_mmu_mode();
1388 /* Do the actual TLB flush before dropping ptl */
1390 tlb_flush_mmu_tlbonly(tlb);
1391 pte_unmap_unlock(start_pte, ptl);
1394 * If we forced a TLB flush (either due to running out of
1395 * batch buffers or because we needed to flush dirty TLB
1396 * entries before releasing the ptl), free the batched
1397 * memory too. Restart if we didn't do everything.
1401 tlb_flush_mmu_free(tlb);
1409 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1410 struct vm_area_struct *vma, pud_t *pud,
1411 unsigned long addr, unsigned long end,
1412 struct zap_details *details)
1417 pmd = pmd_offset(pud, addr);
1419 next = pmd_addr_end(addr, end);
1420 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1421 if (next - addr != HPAGE_PMD_SIZE) {
1422 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1423 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1424 __split_huge_pmd(vma, pmd, addr, false, NULL);
1425 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1430 * Here there can be other concurrent MADV_DONTNEED or
1431 * trans huge page faults running, and if the pmd is
1432 * none or trans huge it can change under us. This is
1433 * because MADV_DONTNEED holds the mmap_sem in read
1436 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1438 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1441 } while (pmd++, addr = next, addr != end);
1446 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1447 struct vm_area_struct *vma, p4d_t *p4d,
1448 unsigned long addr, unsigned long end,
1449 struct zap_details *details)
1454 pud = pud_offset(p4d, addr);
1456 next = pud_addr_end(addr, end);
1457 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1458 if (next - addr != HPAGE_PUD_SIZE) {
1459 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1460 split_huge_pud(vma, pud, addr);
1461 } else if (zap_huge_pud(tlb, vma, pud, addr))
1465 if (pud_none_or_clear_bad(pud))
1467 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1470 } while (pud++, addr = next, addr != end);
1475 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1476 struct vm_area_struct *vma, pgd_t *pgd,
1477 unsigned long addr, unsigned long end,
1478 struct zap_details *details)
1483 p4d = p4d_offset(pgd, addr);
1485 next = p4d_addr_end(addr, end);
1486 if (p4d_none_or_clear_bad(p4d))
1488 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1489 } while (p4d++, addr = next, addr != end);
1494 void unmap_page_range(struct mmu_gather *tlb,
1495 struct vm_area_struct *vma,
1496 unsigned long addr, unsigned long end,
1497 struct zap_details *details)
1502 BUG_ON(addr >= end);
1503 tlb_start_vma(tlb, vma);
1504 pgd = pgd_offset(vma->vm_mm, addr);
1506 next = pgd_addr_end(addr, end);
1507 if (pgd_none_or_clear_bad(pgd))
1509 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1510 } while (pgd++, addr = next, addr != end);
1511 tlb_end_vma(tlb, vma);
1515 static void unmap_single_vma(struct mmu_gather *tlb,
1516 struct vm_area_struct *vma, unsigned long start_addr,
1517 unsigned long end_addr,
1518 struct zap_details *details)
1520 unsigned long start = max(vma->vm_start, start_addr);
1523 if (start >= vma->vm_end)
1525 end = min(vma->vm_end, end_addr);
1526 if (end <= vma->vm_start)
1530 uprobe_munmap(vma, start, end);
1532 if (unlikely(vma->vm_flags & VM_PFNMAP))
1533 untrack_pfn(vma, 0, 0);
1536 if (unlikely(is_vm_hugetlb_page(vma))) {
1538 * It is undesirable to test vma->vm_file as it
1539 * should be non-null for valid hugetlb area.
1540 * However, vm_file will be NULL in the error
1541 * cleanup path of mmap_region. When
1542 * hugetlbfs ->mmap method fails,
1543 * mmap_region() nullifies vma->vm_file
1544 * before calling this function to clean up.
1545 * Since no pte has actually been setup, it is
1546 * safe to do nothing in this case.
1549 i_mmap_lock_write(vma->vm_file->f_mapping);
1550 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1551 i_mmap_unlock_write(vma->vm_file->f_mapping);
1554 unmap_page_range(tlb, vma, start, end, details);
1559 * unmap_vmas - unmap a range of memory covered by a list of vma's
1560 * @tlb: address of the caller's struct mmu_gather
1561 * @vma: the starting vma
1562 * @start_addr: virtual address at which to start unmapping
1563 * @end_addr: virtual address at which to end unmapping
1565 * Unmap all pages in the vma list.
1567 * Only addresses between `start' and `end' will be unmapped.
1569 * The VMA list must be sorted in ascending virtual address order.
1571 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1572 * range after unmap_vmas() returns. So the only responsibility here is to
1573 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1574 * drops the lock and schedules.
1576 void unmap_vmas(struct mmu_gather *tlb,
1577 struct vm_area_struct *vma, unsigned long start_addr,
1578 unsigned long end_addr)
1580 struct mm_struct *mm = vma->vm_mm;
1582 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1583 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1584 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1585 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1589 * zap_page_range - remove user pages in a given range
1590 * @vma: vm_area_struct holding the applicable pages
1591 * @start: starting address of pages to zap
1592 * @size: number of bytes to zap
1594 * Caller must protect the VMA list
1596 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1599 struct mm_struct *mm = vma->vm_mm;
1600 struct mmu_gather tlb;
1601 unsigned long end = start + size;
1604 tlb_gather_mmu(&tlb, mm, start, end);
1605 update_hiwater_rss(mm);
1606 mmu_notifier_invalidate_range_start(mm, start, end);
1607 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1608 unmap_single_vma(&tlb, vma, start, end, NULL);
1611 * zap_page_range does not specify whether mmap_sem should be
1612 * held for read or write. That allows parallel zap_page_range
1613 * operations to unmap a PTE and defer a flush meaning that
1614 * this call observes pte_none and fails to flush the TLB.
1615 * Rather than adding a complex API, ensure that no stale
1616 * TLB entries exist when this call returns.
1618 flush_tlb_range(vma, start, end);
1621 mmu_notifier_invalidate_range_end(mm, start, end);
1622 tlb_finish_mmu(&tlb, start, end);
1626 * zap_page_range_single - remove user pages in a given range
1627 * @vma: vm_area_struct holding the applicable pages
1628 * @address: starting address of pages to zap
1629 * @size: number of bytes to zap
1630 * @details: details of shared cache invalidation
1632 * The range must fit into one VMA.
1634 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1635 unsigned long size, struct zap_details *details)
1637 struct mm_struct *mm = vma->vm_mm;
1638 struct mmu_gather tlb;
1639 unsigned long end = address + size;
1642 tlb_gather_mmu(&tlb, mm, address, end);
1643 update_hiwater_rss(mm);
1644 mmu_notifier_invalidate_range_start(mm, address, end);
1645 unmap_single_vma(&tlb, vma, address, end, details);
1646 mmu_notifier_invalidate_range_end(mm, address, end);
1647 tlb_finish_mmu(&tlb, address, end);
1651 * zap_vma_ptes - remove ptes mapping the vma
1652 * @vma: vm_area_struct holding ptes to be zapped
1653 * @address: starting address of pages to zap
1654 * @size: number of bytes to zap
1656 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1658 * The entire address range must be fully contained within the vma.
1660 * Returns 0 if successful.
1662 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1665 if (address < vma->vm_start || address + size > vma->vm_end ||
1666 !(vma->vm_flags & VM_PFNMAP))
1668 zap_page_range_single(vma, address, size, NULL);
1671 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1673 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1681 pgd = pgd_offset(mm, addr);
1682 p4d = p4d_alloc(mm, pgd, addr);
1685 pud = pud_alloc(mm, p4d, addr);
1688 pmd = pmd_alloc(mm, pud, addr);
1692 VM_BUG_ON(pmd_trans_huge(*pmd));
1693 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1697 * This is the old fallback for page remapping.
1699 * For historical reasons, it only allows reserved pages. Only
1700 * old drivers should use this, and they needed to mark their
1701 * pages reserved for the old functions anyway.
1703 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1704 struct page *page, pgprot_t prot)
1706 struct mm_struct *mm = vma->vm_mm;
1715 flush_dcache_page(page);
1716 pte = get_locked_pte(mm, addr, &ptl);
1720 if (!pte_none(*pte))
1723 /* Ok, finally just insert the thing.. */
1725 inc_mm_counter_fast(mm, mm_counter_file(page));
1726 page_add_file_rmap(page, false);
1727 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1730 pte_unmap_unlock(pte, ptl);
1733 pte_unmap_unlock(pte, ptl);
1739 * vm_insert_page - insert single page into user vma
1740 * @vma: user vma to map to
1741 * @addr: target user address of this page
1742 * @page: source kernel page
1744 * This allows drivers to insert individual pages they've allocated
1747 * The page has to be a nice clean _individual_ kernel allocation.
1748 * If you allocate a compound page, you need to have marked it as
1749 * such (__GFP_COMP), or manually just split the page up yourself
1750 * (see split_page()).
1752 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1753 * took an arbitrary page protection parameter. This doesn't allow
1754 * that. Your vma protection will have to be set up correctly, which
1755 * means that if you want a shared writable mapping, you'd better
1756 * ask for a shared writable mapping!
1758 * The page does not need to be reserved.
1760 * Usually this function is called from f_op->mmap() handler
1761 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1762 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1763 * function from other places, for example from page-fault handler.
1765 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1768 if (addr < vma->vm_start || addr >= vma->vm_end)
1770 if (!page_count(page))
1772 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1773 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1774 BUG_ON(vma->vm_flags & VM_PFNMAP);
1775 vma->vm_flags |= VM_MIXEDMAP;
1777 return insert_page(vma, addr, page, vma->vm_page_prot);
1779 EXPORT_SYMBOL(vm_insert_page);
1781 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1782 pfn_t pfn, pgprot_t prot, bool mkwrite)
1784 struct mm_struct *mm = vma->vm_mm;
1790 pte = get_locked_pte(mm, addr, &ptl);
1794 if (!pte_none(*pte)) {
1797 * For read faults on private mappings the PFN passed
1798 * in may not match the PFN we have mapped if the
1799 * mapped PFN is a writeable COW page. In the mkwrite
1800 * case we are creating a writable PTE for a shared
1801 * mapping and we expect the PFNs to match.
1803 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1811 /* Ok, finally just insert the thing.. */
1812 if (pfn_t_devmap(pfn))
1813 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1815 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1819 entry = pte_mkyoung(entry);
1820 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1823 set_pte_at(mm, addr, pte, entry);
1824 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1828 pte_unmap_unlock(pte, ptl);
1834 * vm_insert_pfn - insert single pfn into user vma
1835 * @vma: user vma to map to
1836 * @addr: target user address of this page
1837 * @pfn: source kernel pfn
1839 * Similar to vm_insert_page, this allows drivers to insert individual pages
1840 * they've allocated into a user vma. Same comments apply.
1842 * This function should only be called from a vm_ops->fault handler, and
1843 * in that case the handler should return NULL.
1845 * vma cannot be a COW mapping.
1847 * As this is called only for pages that do not currently exist, we
1848 * do not need to flush old virtual caches or the TLB.
1850 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1853 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1855 EXPORT_SYMBOL(vm_insert_pfn);
1858 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1859 * @vma: user vma to map to
1860 * @addr: target user address of this page
1861 * @pfn: source kernel pfn
1862 * @pgprot: pgprot flags for the inserted page
1864 * This is exactly like vm_insert_pfn, except that it allows drivers to
1865 * to override pgprot on a per-page basis.
1867 * This only makes sense for IO mappings, and it makes no sense for
1868 * cow mappings. In general, using multiple vmas is preferable;
1869 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1872 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1873 unsigned long pfn, pgprot_t pgprot)
1877 * Technically, architectures with pte_special can avoid all these
1878 * restrictions (same for remap_pfn_range). However we would like
1879 * consistency in testing and feature parity among all, so we should
1880 * try to keep these invariants in place for everybody.
1882 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1883 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1884 (VM_PFNMAP|VM_MIXEDMAP));
1885 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1886 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1888 if (addr < vma->vm_start || addr >= vma->vm_end)
1891 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1893 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1898 EXPORT_SYMBOL(vm_insert_pfn_prot);
1900 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1902 /* these checks mirror the abort conditions in vm_normal_page */
1903 if (vma->vm_flags & VM_MIXEDMAP)
1905 if (pfn_t_devmap(pfn))
1907 if (pfn_t_special(pfn))
1909 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1914 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1915 pfn_t pfn, bool mkwrite)
1917 pgprot_t pgprot = vma->vm_page_prot;
1919 BUG_ON(!vm_mixed_ok(vma, pfn));
1921 if (addr < vma->vm_start || addr >= vma->vm_end)
1924 track_pfn_insert(vma, &pgprot, pfn);
1927 * If we don't have pte special, then we have to use the pfn_valid()
1928 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1929 * refcount the page if pfn_valid is true (hence insert_page rather
1930 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1931 * without pte special, it would there be refcounted as a normal page.
1933 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1937 * At this point we are committed to insert_page()
1938 * regardless of whether the caller specified flags that
1939 * result in pfn_t_has_page() == false.
1941 page = pfn_to_page(pfn_t_to_pfn(pfn));
1942 return insert_page(vma, addr, page, pgprot);
1944 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1947 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1950 return __vm_insert_mixed(vma, addr, pfn, false);
1953 EXPORT_SYMBOL(vm_insert_mixed);
1955 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1958 return __vm_insert_mixed(vma, addr, pfn, true);
1960 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1963 * maps a range of physical memory into the requested pages. the old
1964 * mappings are removed. any references to nonexistent pages results
1965 * in null mappings (currently treated as "copy-on-access")
1967 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1968 unsigned long addr, unsigned long end,
1969 unsigned long pfn, pgprot_t prot)
1974 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1977 arch_enter_lazy_mmu_mode();
1979 BUG_ON(!pte_none(*pte));
1980 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1982 } while (pte++, addr += PAGE_SIZE, addr != end);
1983 arch_leave_lazy_mmu_mode();
1984 pte_unmap_unlock(pte - 1, ptl);
1988 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1989 unsigned long addr, unsigned long end,
1990 unsigned long pfn, pgprot_t prot)
1995 pfn -= addr >> PAGE_SHIFT;
1996 pmd = pmd_alloc(mm, pud, addr);
1999 VM_BUG_ON(pmd_trans_huge(*pmd));
2001 next = pmd_addr_end(addr, end);
2002 if (remap_pte_range(mm, pmd, addr, next,
2003 pfn + (addr >> PAGE_SHIFT), prot))
2005 } while (pmd++, addr = next, addr != end);
2009 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2010 unsigned long addr, unsigned long end,
2011 unsigned long pfn, pgprot_t prot)
2016 pfn -= addr >> PAGE_SHIFT;
2017 pud = pud_alloc(mm, p4d, addr);
2021 next = pud_addr_end(addr, end);
2022 if (remap_pmd_range(mm, pud, addr, next,
2023 pfn + (addr >> PAGE_SHIFT), prot))
2025 } while (pud++, addr = next, addr != end);
2029 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2030 unsigned long addr, unsigned long end,
2031 unsigned long pfn, pgprot_t prot)
2036 pfn -= addr >> PAGE_SHIFT;
2037 p4d = p4d_alloc(mm, pgd, addr);
2041 next = p4d_addr_end(addr, end);
2042 if (remap_pud_range(mm, p4d, addr, next,
2043 pfn + (addr >> PAGE_SHIFT), prot))
2045 } while (p4d++, addr = next, addr != end);
2050 * remap_pfn_range - remap kernel memory to userspace
2051 * @vma: user vma to map to
2052 * @addr: target user address to start at
2053 * @pfn: physical address of kernel memory
2054 * @size: size of map area
2055 * @prot: page protection flags for this mapping
2057 * Note: this is only safe if the mm semaphore is held when called.
2059 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2060 unsigned long pfn, unsigned long size, pgprot_t prot)
2064 unsigned long end = addr + PAGE_ALIGN(size);
2065 struct mm_struct *mm = vma->vm_mm;
2066 unsigned long remap_pfn = pfn;
2070 * Physically remapped pages are special. Tell the
2071 * rest of the world about it:
2072 * VM_IO tells people not to look at these pages
2073 * (accesses can have side effects).
2074 * VM_PFNMAP tells the core MM that the base pages are just
2075 * raw PFN mappings, and do not have a "struct page" associated
2078 * Disable vma merging and expanding with mremap().
2080 * Omit vma from core dump, even when VM_IO turned off.
2082 * There's a horrible special case to handle copy-on-write
2083 * behaviour that some programs depend on. We mark the "original"
2084 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2085 * See vm_normal_page() for details.
2087 if (is_cow_mapping(vma->vm_flags)) {
2088 if (addr != vma->vm_start || end != vma->vm_end)
2090 vma->vm_pgoff = pfn;
2093 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2097 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2099 BUG_ON(addr >= end);
2100 pfn -= addr >> PAGE_SHIFT;
2101 pgd = pgd_offset(mm, addr);
2102 flush_cache_range(vma, addr, end);
2104 next = pgd_addr_end(addr, end);
2105 err = remap_p4d_range(mm, pgd, addr, next,
2106 pfn + (addr >> PAGE_SHIFT), prot);
2109 } while (pgd++, addr = next, addr != end);
2112 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2116 EXPORT_SYMBOL(remap_pfn_range);
2119 * vm_iomap_memory - remap memory to userspace
2120 * @vma: user vma to map to
2121 * @start: start of area
2122 * @len: size of area
2124 * This is a simplified io_remap_pfn_range() for common driver use. The
2125 * driver just needs to give us the physical memory range to be mapped,
2126 * we'll figure out the rest from the vma information.
2128 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2129 * whatever write-combining details or similar.
2131 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2133 unsigned long vm_len, pfn, pages;
2135 /* Check that the physical memory area passed in looks valid */
2136 if (start + len < start)
2139 * You *really* shouldn't map things that aren't page-aligned,
2140 * but we've historically allowed it because IO memory might
2141 * just have smaller alignment.
2143 len += start & ~PAGE_MASK;
2144 pfn = start >> PAGE_SHIFT;
2145 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2146 if (pfn + pages < pfn)
2149 /* We start the mapping 'vm_pgoff' pages into the area */
2150 if (vma->vm_pgoff > pages)
2152 pfn += vma->vm_pgoff;
2153 pages -= vma->vm_pgoff;
2155 /* Can we fit all of the mapping? */
2156 vm_len = vma->vm_end - vma->vm_start;
2157 if (vm_len >> PAGE_SHIFT > pages)
2160 /* Ok, let it rip */
2161 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2163 EXPORT_SYMBOL(vm_iomap_memory);
2165 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2166 unsigned long addr, unsigned long end,
2167 pte_fn_t fn, void *data)
2172 spinlock_t *uninitialized_var(ptl);
2174 pte = (mm == &init_mm) ?
2175 pte_alloc_kernel(pmd, addr) :
2176 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2180 BUG_ON(pmd_huge(*pmd));
2182 arch_enter_lazy_mmu_mode();
2184 token = pmd_pgtable(*pmd);
2187 err = fn(pte++, token, addr, data);
2190 } while (addr += PAGE_SIZE, addr != end);
2192 arch_leave_lazy_mmu_mode();
2195 pte_unmap_unlock(pte-1, ptl);
2199 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2200 unsigned long addr, unsigned long end,
2201 pte_fn_t fn, void *data)
2207 BUG_ON(pud_huge(*pud));
2209 pmd = pmd_alloc(mm, pud, addr);
2213 next = pmd_addr_end(addr, end);
2214 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2217 } while (pmd++, addr = next, addr != end);
2221 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2222 unsigned long addr, unsigned long end,
2223 pte_fn_t fn, void *data)
2229 pud = pud_alloc(mm, p4d, addr);
2233 next = pud_addr_end(addr, end);
2234 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2237 } while (pud++, addr = next, addr != end);
2241 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2242 unsigned long addr, unsigned long end,
2243 pte_fn_t fn, void *data)
2249 p4d = p4d_alloc(mm, pgd, addr);
2253 next = p4d_addr_end(addr, end);
2254 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2257 } while (p4d++, addr = next, addr != end);
2262 * Scan a region of virtual memory, filling in page tables as necessary
2263 * and calling a provided function on each leaf page table.
2265 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2266 unsigned long size, pte_fn_t fn, void *data)
2270 unsigned long end = addr + size;
2273 if (WARN_ON(addr >= end))
2276 pgd = pgd_offset(mm, addr);
2278 next = pgd_addr_end(addr, end);
2279 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2282 } while (pgd++, addr = next, addr != end);
2286 EXPORT_SYMBOL_GPL(apply_to_page_range);
2289 * handle_pte_fault chooses page fault handler according to an entry which was
2290 * read non-atomically. Before making any commitment, on those architectures
2291 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2292 * parts, do_swap_page must check under lock before unmapping the pte and
2293 * proceeding (but do_wp_page is only called after already making such a check;
2294 * and do_anonymous_page can safely check later on).
2296 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2297 pte_t *page_table, pte_t orig_pte)
2300 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2301 if (sizeof(pte_t) > sizeof(unsigned long)) {
2302 spinlock_t *ptl = pte_lockptr(mm, pmd);
2304 same = pte_same(*page_table, orig_pte);
2308 pte_unmap(page_table);
2312 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2314 debug_dma_assert_idle(src);
2317 * If the source page was a PFN mapping, we don't have
2318 * a "struct page" for it. We do a best-effort copy by
2319 * just copying from the original user address. If that
2320 * fails, we just zero-fill it. Live with it.
2322 if (unlikely(!src)) {
2323 void *kaddr = kmap_atomic(dst);
2324 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2327 * This really shouldn't fail, because the page is there
2328 * in the page tables. But it might just be unreadable,
2329 * in which case we just give up and fill the result with
2332 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2334 kunmap_atomic(kaddr);
2335 flush_dcache_page(dst);
2337 copy_user_highpage(dst, src, va, vma);
2340 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2342 struct file *vm_file = vma->vm_file;
2345 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2348 * Special mappings (e.g. VDSO) do not have any file so fake
2349 * a default GFP_KERNEL for them.
2355 * Notify the address space that the page is about to become writable so that
2356 * it can prohibit this or wait for the page to get into an appropriate state.
2358 * We do this without the lock held, so that it can sleep if it needs to.
2360 static int do_page_mkwrite(struct vm_fault *vmf)
2363 struct page *page = vmf->page;
2364 unsigned int old_flags = vmf->flags;
2366 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2368 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2369 /* Restore original flags so that caller is not surprised */
2370 vmf->flags = old_flags;
2371 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2373 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2375 if (!page->mapping) {
2377 return 0; /* retry */
2379 ret |= VM_FAULT_LOCKED;
2381 VM_BUG_ON_PAGE(!PageLocked(page), page);
2386 * Handle dirtying of a page in shared file mapping on a write fault.
2388 * The function expects the page to be locked and unlocks it.
2390 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2393 struct address_space *mapping;
2395 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2397 dirtied = set_page_dirty(page);
2398 VM_BUG_ON_PAGE(PageAnon(page), page);
2400 * Take a local copy of the address_space - page.mapping may be zeroed
2401 * by truncate after unlock_page(). The address_space itself remains
2402 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2403 * release semantics to prevent the compiler from undoing this copying.
2405 mapping = page_rmapping(page);
2408 if ((dirtied || page_mkwrite) && mapping) {
2410 * Some device drivers do not set page.mapping
2411 * but still dirty their pages
2413 balance_dirty_pages_ratelimited(mapping);
2417 file_update_time(vma->vm_file);
2421 * Handle write page faults for pages that can be reused in the current vma
2423 * This can happen either due to the mapping being with the VM_SHARED flag,
2424 * or due to us being the last reference standing to the page. In either
2425 * case, all we need to do here is to mark the page as writable and update
2426 * any related book-keeping.
2428 static inline void wp_page_reuse(struct vm_fault *vmf)
2429 __releases(vmf->ptl)
2431 struct vm_area_struct *vma = vmf->vma;
2432 struct page *page = vmf->page;
2435 * Clear the pages cpupid information as the existing
2436 * information potentially belongs to a now completely
2437 * unrelated process.
2440 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2442 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2443 entry = pte_mkyoung(vmf->orig_pte);
2444 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2445 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2446 update_mmu_cache(vma, vmf->address, vmf->pte);
2447 pte_unmap_unlock(vmf->pte, vmf->ptl);
2451 * Handle the case of a page which we actually need to copy to a new page.
2453 * Called with mmap_sem locked and the old page referenced, but
2454 * without the ptl held.
2456 * High level logic flow:
2458 * - Allocate a page, copy the content of the old page to the new one.
2459 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2460 * - Take the PTL. If the pte changed, bail out and release the allocated page
2461 * - If the pte is still the way we remember it, update the page table and all
2462 * relevant references. This includes dropping the reference the page-table
2463 * held to the old page, as well as updating the rmap.
2464 * - In any case, unlock the PTL and drop the reference we took to the old page.
2466 static int wp_page_copy(struct vm_fault *vmf)
2468 struct vm_area_struct *vma = vmf->vma;
2469 struct mm_struct *mm = vma->vm_mm;
2470 struct page *old_page = vmf->page;
2471 struct page *new_page = NULL;
2473 int page_copied = 0;
2474 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2475 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2476 struct mem_cgroup *memcg;
2478 if (unlikely(anon_vma_prepare(vma)))
2481 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2482 new_page = alloc_zeroed_user_highpage_movable(vma,
2487 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2491 cow_user_page(new_page, old_page, vmf->address, vma);
2494 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2497 __SetPageUptodate(new_page);
2499 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2502 * Re-check the pte - we dropped the lock
2504 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2505 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2507 if (!PageAnon(old_page)) {
2508 dec_mm_counter_fast(mm,
2509 mm_counter_file(old_page));
2510 inc_mm_counter_fast(mm, MM_ANONPAGES);
2513 inc_mm_counter_fast(mm, MM_ANONPAGES);
2515 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2516 entry = mk_pte(new_page, vma->vm_page_prot);
2517 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2519 * Clear the pte entry and flush it first, before updating the
2520 * pte with the new entry. This will avoid a race condition
2521 * seen in the presence of one thread doing SMC and another
2524 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2525 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2526 mem_cgroup_commit_charge(new_page, memcg, false, false);
2527 lru_cache_add_active_or_unevictable(new_page, vma);
2529 * We call the notify macro here because, when using secondary
2530 * mmu page tables (such as kvm shadow page tables), we want the
2531 * new page to be mapped directly into the secondary page table.
2533 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2534 update_mmu_cache(vma, vmf->address, vmf->pte);
2537 * Only after switching the pte to the new page may
2538 * we remove the mapcount here. Otherwise another
2539 * process may come and find the rmap count decremented
2540 * before the pte is switched to the new page, and
2541 * "reuse" the old page writing into it while our pte
2542 * here still points into it and can be read by other
2545 * The critical issue is to order this
2546 * page_remove_rmap with the ptp_clear_flush above.
2547 * Those stores are ordered by (if nothing else,)
2548 * the barrier present in the atomic_add_negative
2549 * in page_remove_rmap.
2551 * Then the TLB flush in ptep_clear_flush ensures that
2552 * no process can access the old page before the
2553 * decremented mapcount is visible. And the old page
2554 * cannot be reused until after the decremented
2555 * mapcount is visible. So transitively, TLBs to
2556 * old page will be flushed before it can be reused.
2558 page_remove_rmap(old_page, false);
2561 /* Free the old page.. */
2562 new_page = old_page;
2565 mem_cgroup_cancel_charge(new_page, memcg, false);
2571 pte_unmap_unlock(vmf->pte, vmf->ptl);
2573 * No need to double call mmu_notifier->invalidate_range() callback as
2574 * the above ptep_clear_flush_notify() did already call it.
2576 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2579 * Don't let another task, with possibly unlocked vma,
2580 * keep the mlocked page.
2582 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2583 lock_page(old_page); /* LRU manipulation */
2584 if (PageMlocked(old_page))
2585 munlock_vma_page(old_page);
2586 unlock_page(old_page);
2590 return page_copied ? VM_FAULT_WRITE : 0;
2596 return VM_FAULT_OOM;
2600 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2601 * writeable once the page is prepared
2603 * @vmf: structure describing the fault
2605 * This function handles all that is needed to finish a write page fault in a
2606 * shared mapping due to PTE being read-only once the mapped page is prepared.
2607 * It handles locking of PTE and modifying it. The function returns
2608 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2611 * The function expects the page to be locked or other protection against
2612 * concurrent faults / writeback (such as DAX radix tree locks).
2614 int finish_mkwrite_fault(struct vm_fault *vmf)
2616 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2617 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2620 * We might have raced with another page fault while we released the
2621 * pte_offset_map_lock.
2623 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2624 pte_unmap_unlock(vmf->pte, vmf->ptl);
2625 return VM_FAULT_NOPAGE;
2632 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2635 static int wp_pfn_shared(struct vm_fault *vmf)
2637 struct vm_area_struct *vma = vmf->vma;
2639 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2642 pte_unmap_unlock(vmf->pte, vmf->ptl);
2643 vmf->flags |= FAULT_FLAG_MKWRITE;
2644 ret = vma->vm_ops->pfn_mkwrite(vmf);
2645 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2647 return finish_mkwrite_fault(vmf);
2650 return VM_FAULT_WRITE;
2653 static int wp_page_shared(struct vm_fault *vmf)
2654 __releases(vmf->ptl)
2656 struct vm_area_struct *vma = vmf->vma;
2658 get_page(vmf->page);
2660 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2663 pte_unmap_unlock(vmf->pte, vmf->ptl);
2664 tmp = do_page_mkwrite(vmf);
2665 if (unlikely(!tmp || (tmp &
2666 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2667 put_page(vmf->page);
2670 tmp = finish_mkwrite_fault(vmf);
2671 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2672 unlock_page(vmf->page);
2673 put_page(vmf->page);
2678 lock_page(vmf->page);
2680 fault_dirty_shared_page(vma, vmf->page);
2681 put_page(vmf->page);
2683 return VM_FAULT_WRITE;
2687 * This routine handles present pages, when users try to write
2688 * to a shared page. It is done by copying the page to a new address
2689 * and decrementing the shared-page counter for the old page.
2691 * Note that this routine assumes that the protection checks have been
2692 * done by the caller (the low-level page fault routine in most cases).
2693 * Thus we can safely just mark it writable once we've done any necessary
2696 * We also mark the page dirty at this point even though the page will
2697 * change only once the write actually happens. This avoids a few races,
2698 * and potentially makes it more efficient.
2700 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2701 * but allow concurrent faults), with pte both mapped and locked.
2702 * We return with mmap_sem still held, but pte unmapped and unlocked.
2704 static int do_wp_page(struct vm_fault *vmf)
2705 __releases(vmf->ptl)
2707 struct vm_area_struct *vma = vmf->vma;
2709 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2712 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2715 * We should not cow pages in a shared writeable mapping.
2716 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2718 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2719 (VM_WRITE|VM_SHARED))
2720 return wp_pfn_shared(vmf);
2722 pte_unmap_unlock(vmf->pte, vmf->ptl);
2723 return wp_page_copy(vmf);
2727 * Take out anonymous pages first, anonymous shared vmas are
2728 * not dirty accountable.
2730 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2731 int total_map_swapcount;
2732 if (!trylock_page(vmf->page)) {
2733 get_page(vmf->page);
2734 pte_unmap_unlock(vmf->pte, vmf->ptl);
2735 lock_page(vmf->page);
2736 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2737 vmf->address, &vmf->ptl);
2738 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2739 unlock_page(vmf->page);
2740 pte_unmap_unlock(vmf->pte, vmf->ptl);
2741 put_page(vmf->page);
2744 put_page(vmf->page);
2746 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2747 if (total_map_swapcount == 1) {
2749 * The page is all ours. Move it to
2750 * our anon_vma so the rmap code will
2751 * not search our parent or siblings.
2752 * Protected against the rmap code by
2755 page_move_anon_rmap(vmf->page, vma);
2757 unlock_page(vmf->page);
2759 return VM_FAULT_WRITE;
2761 unlock_page(vmf->page);
2762 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2763 (VM_WRITE|VM_SHARED))) {
2764 return wp_page_shared(vmf);
2768 * Ok, we need to copy. Oh, well..
2770 get_page(vmf->page);
2772 pte_unmap_unlock(vmf->pte, vmf->ptl);
2773 return wp_page_copy(vmf);
2776 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2777 unsigned long start_addr, unsigned long end_addr,
2778 struct zap_details *details)
2780 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2783 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2784 struct zap_details *details)
2786 struct vm_area_struct *vma;
2787 pgoff_t vba, vea, zba, zea;
2789 vma_interval_tree_foreach(vma, root,
2790 details->first_index, details->last_index) {
2792 vba = vma->vm_pgoff;
2793 vea = vba + vma_pages(vma) - 1;
2794 zba = details->first_index;
2797 zea = details->last_index;
2801 unmap_mapping_range_vma(vma,
2802 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2803 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2809 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2810 * address_space corresponding to the specified page range in the underlying
2813 * @mapping: the address space containing mmaps to be unmapped.
2814 * @holebegin: byte in first page to unmap, relative to the start of
2815 * the underlying file. This will be rounded down to a PAGE_SIZE
2816 * boundary. Note that this is different from truncate_pagecache(), which
2817 * must keep the partial page. In contrast, we must get rid of
2819 * @holelen: size of prospective hole in bytes. This will be rounded
2820 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2822 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2823 * but 0 when invalidating pagecache, don't throw away private data.
2825 void unmap_mapping_range(struct address_space *mapping,
2826 loff_t const holebegin, loff_t const holelen, int even_cows)
2828 struct zap_details details = { };
2829 pgoff_t hba = holebegin >> PAGE_SHIFT;
2830 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2832 /* Check for overflow. */
2833 if (sizeof(holelen) > sizeof(hlen)) {
2835 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2836 if (holeend & ~(long long)ULONG_MAX)
2837 hlen = ULONG_MAX - hba + 1;
2840 details.check_mapping = even_cows ? NULL : mapping;
2841 details.first_index = hba;
2842 details.last_index = hba + hlen - 1;
2843 if (details.last_index < details.first_index)
2844 details.last_index = ULONG_MAX;
2846 i_mmap_lock_write(mapping);
2847 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2848 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2849 i_mmap_unlock_write(mapping);
2851 EXPORT_SYMBOL(unmap_mapping_range);
2854 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2855 * but allow concurrent faults), and pte mapped but not yet locked.
2856 * We return with pte unmapped and unlocked.
2858 * We return with the mmap_sem locked or unlocked in the same cases
2859 * as does filemap_fault().
2861 int do_swap_page(struct vm_fault *vmf)
2863 struct vm_area_struct *vma = vmf->vma;
2864 struct page *page = NULL, *swapcache = NULL;
2865 struct mem_cgroup *memcg;
2866 struct vma_swap_readahead swap_ra;
2872 bool vma_readahead = swap_use_vma_readahead();
2875 page = swap_readahead_detect(vmf, &swap_ra);
2876 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2882 entry = pte_to_swp_entry(vmf->orig_pte);
2883 if (unlikely(non_swap_entry(entry))) {
2884 if (is_migration_entry(entry)) {
2885 migration_entry_wait(vma->vm_mm, vmf->pmd,
2887 } else if (is_device_private_entry(entry)) {
2889 * For un-addressable device memory we call the pgmap
2890 * fault handler callback. The callback must migrate
2891 * the page back to some CPU accessible page.
2893 ret = device_private_entry_fault(vma, vmf->address, entry,
2894 vmf->flags, vmf->pmd);
2895 } else if (is_hwpoison_entry(entry)) {
2896 ret = VM_FAULT_HWPOISON;
2898 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2899 ret = VM_FAULT_SIGBUS;
2905 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2907 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2910 struct swap_info_struct *si = swp_swap_info(entry);
2912 if (si->flags & SWP_SYNCHRONOUS_IO &&
2913 __swap_count(si, entry) == 1) {
2914 /* skip swapcache */
2915 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2917 __SetPageLocked(page);
2918 __SetPageSwapBacked(page);
2919 set_page_private(page, entry.val);
2920 lru_cache_add_anon(page);
2921 swap_readpage(page, true);
2925 page = do_swap_page_readahead(entry,
2926 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2928 page = swapin_readahead(entry,
2929 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2935 * Back out if somebody else faulted in this pte
2936 * while we released the pte lock.
2938 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2939 vmf->address, &vmf->ptl);
2940 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2942 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2946 /* Had to read the page from swap area: Major fault */
2947 ret = VM_FAULT_MAJOR;
2948 count_vm_event(PGMAJFAULT);
2949 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2950 } else if (PageHWPoison(page)) {
2952 * hwpoisoned dirty swapcache pages are kept for killing
2953 * owner processes (which may be unknown at hwpoison time)
2955 ret = VM_FAULT_HWPOISON;
2956 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2961 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2963 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2965 ret |= VM_FAULT_RETRY;
2970 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2971 * release the swapcache from under us. The page pin, and pte_same
2972 * test below, are not enough to exclude that. Even if it is still
2973 * swapcache, we need to check that the page's swap has not changed.
2975 if (unlikely((!PageSwapCache(page) ||
2976 page_private(page) != entry.val)) && swapcache)
2979 page = ksm_might_need_to_copy(page, vma, vmf->address);
2980 if (unlikely(!page)) {
2986 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2993 * Back out if somebody else already faulted in this pte.
2995 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2997 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3000 if (unlikely(!PageUptodate(page))) {
3001 ret = VM_FAULT_SIGBUS;
3006 * The page isn't present yet, go ahead with the fault.
3008 * Be careful about the sequence of operations here.
3009 * To get its accounting right, reuse_swap_page() must be called
3010 * while the page is counted on swap but not yet in mapcount i.e.
3011 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3012 * must be called after the swap_free(), or it will never succeed.
3015 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3016 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3017 pte = mk_pte(page, vma->vm_page_prot);
3018 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3019 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3020 vmf->flags &= ~FAULT_FLAG_WRITE;
3021 ret |= VM_FAULT_WRITE;
3022 exclusive = RMAP_EXCLUSIVE;
3024 flush_icache_page(vma, page);
3025 if (pte_swp_soft_dirty(vmf->orig_pte))
3026 pte = pte_mksoft_dirty(pte);
3027 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3028 vmf->orig_pte = pte;
3030 /* ksm created a completely new copy */
3031 if (unlikely(page != swapcache && swapcache)) {
3032 page_add_new_anon_rmap(page, vma, vmf->address, false);
3033 mem_cgroup_commit_charge(page, memcg, false, false);
3034 lru_cache_add_active_or_unevictable(page, vma);
3036 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3037 mem_cgroup_commit_charge(page, memcg, true, false);
3038 activate_page(page);
3042 if (mem_cgroup_swap_full(page) ||
3043 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3044 try_to_free_swap(page);
3046 if (page != swapcache && swapcache) {
3048 * Hold the lock to avoid the swap entry to be reused
3049 * until we take the PT lock for the pte_same() check
3050 * (to avoid false positives from pte_same). For
3051 * further safety release the lock after the swap_free
3052 * so that the swap count won't change under a
3053 * parallel locked swapcache.
3055 unlock_page(swapcache);
3056 put_page(swapcache);
3059 if (vmf->flags & FAULT_FLAG_WRITE) {
3060 ret |= do_wp_page(vmf);
3061 if (ret & VM_FAULT_ERROR)
3062 ret &= VM_FAULT_ERROR;
3066 /* No need to invalidate - it was non-present before */
3067 update_mmu_cache(vma, vmf->address, vmf->pte);
3069 pte_unmap_unlock(vmf->pte, vmf->ptl);
3073 mem_cgroup_cancel_charge(page, memcg, false);
3074 pte_unmap_unlock(vmf->pte, vmf->ptl);
3079 if (page != swapcache && swapcache) {
3080 unlock_page(swapcache);
3081 put_page(swapcache);
3087 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3088 * but allow concurrent faults), and pte mapped but not yet locked.
3089 * We return with mmap_sem still held, but pte unmapped and unlocked.
3091 static int do_anonymous_page(struct vm_fault *vmf)
3093 struct vm_area_struct *vma = vmf->vma;
3094 struct mem_cgroup *memcg;
3099 /* File mapping without ->vm_ops ? */
3100 if (vma->vm_flags & VM_SHARED)
3101 return VM_FAULT_SIGBUS;
3104 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3105 * pte_offset_map() on pmds where a huge pmd might be created
3106 * from a different thread.
3108 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3109 * parallel threads are excluded by other means.
3111 * Here we only have down_read(mmap_sem).
3113 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3114 return VM_FAULT_OOM;
3116 /* See the comment in pte_alloc_one_map() */
3117 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3120 /* Use the zero-page for reads */
3121 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3122 !mm_forbids_zeropage(vma->vm_mm)) {
3123 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3124 vma->vm_page_prot));
3125 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3126 vmf->address, &vmf->ptl);
3127 if (!pte_none(*vmf->pte))
3129 ret = check_stable_address_space(vma->vm_mm);
3132 /* Deliver the page fault to userland, check inside PT lock */
3133 if (userfaultfd_missing(vma)) {
3134 pte_unmap_unlock(vmf->pte, vmf->ptl);
3135 return handle_userfault(vmf, VM_UFFD_MISSING);
3140 /* Allocate our own private page. */
3141 if (unlikely(anon_vma_prepare(vma)))
3143 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3147 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3151 * The memory barrier inside __SetPageUptodate makes sure that
3152 * preceeding stores to the page contents become visible before
3153 * the set_pte_at() write.
3155 __SetPageUptodate(page);
3157 entry = mk_pte(page, vma->vm_page_prot);
3158 if (vma->vm_flags & VM_WRITE)
3159 entry = pte_mkwrite(pte_mkdirty(entry));
3161 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3163 if (!pte_none(*vmf->pte))
3166 ret = check_stable_address_space(vma->vm_mm);
3170 /* Deliver the page fault to userland, check inside PT lock */
3171 if (userfaultfd_missing(vma)) {
3172 pte_unmap_unlock(vmf->pte, vmf->ptl);
3173 mem_cgroup_cancel_charge(page, memcg, false);
3175 return handle_userfault(vmf, VM_UFFD_MISSING);
3178 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3179 page_add_new_anon_rmap(page, vma, vmf->address, false);
3180 mem_cgroup_commit_charge(page, memcg, false, false);
3181 lru_cache_add_active_or_unevictable(page, vma);
3183 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3185 /* No need to invalidate - it was non-present before */
3186 update_mmu_cache(vma, vmf->address, vmf->pte);
3188 pte_unmap_unlock(vmf->pte, vmf->ptl);
3191 mem_cgroup_cancel_charge(page, memcg, false);
3197 return VM_FAULT_OOM;
3201 * The mmap_sem must have been held on entry, and may have been
3202 * released depending on flags and vma->vm_ops->fault() return value.
3203 * See filemap_fault() and __lock_page_retry().
3205 static int __do_fault(struct vm_fault *vmf)
3207 struct vm_area_struct *vma = vmf->vma;
3210 ret = vma->vm_ops->fault(vmf);
3211 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3212 VM_FAULT_DONE_COW)))
3215 if (unlikely(PageHWPoison(vmf->page))) {
3216 if (ret & VM_FAULT_LOCKED)
3217 unlock_page(vmf->page);
3218 put_page(vmf->page);
3220 return VM_FAULT_HWPOISON;
3223 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3224 lock_page(vmf->page);
3226 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3232 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3233 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3234 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3235 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3237 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3239 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3242 static int pte_alloc_one_map(struct vm_fault *vmf)
3244 struct vm_area_struct *vma = vmf->vma;
3246 if (!pmd_none(*vmf->pmd))
3248 if (vmf->prealloc_pte) {
3249 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3250 if (unlikely(!pmd_none(*vmf->pmd))) {
3251 spin_unlock(vmf->ptl);
3255 mm_inc_nr_ptes(vma->vm_mm);
3256 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3257 spin_unlock(vmf->ptl);
3258 vmf->prealloc_pte = NULL;
3259 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3260 return VM_FAULT_OOM;
3264 * If a huge pmd materialized under us just retry later. Use
3265 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3266 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3267 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3268 * running immediately after a huge pmd fault in a different thread of
3269 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3270 * All we have to ensure is that it is a regular pmd that we can walk
3271 * with pte_offset_map() and we can do that through an atomic read in
3272 * C, which is what pmd_trans_unstable() provides.
3274 if (pmd_devmap_trans_unstable(vmf->pmd))
3275 return VM_FAULT_NOPAGE;
3278 * At this point we know that our vmf->pmd points to a page of ptes
3279 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3280 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3281 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3282 * be valid and we will re-check to make sure the vmf->pte isn't
3283 * pte_none() under vmf->ptl protection when we return to
3286 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3291 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3293 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3294 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3295 unsigned long haddr)
3297 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3298 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3300 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3305 static void deposit_prealloc_pte(struct vm_fault *vmf)
3307 struct vm_area_struct *vma = vmf->vma;
3309 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3311 * We are going to consume the prealloc table,
3312 * count that as nr_ptes.
3314 mm_inc_nr_ptes(vma->vm_mm);
3315 vmf->prealloc_pte = NULL;
3318 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3320 struct vm_area_struct *vma = vmf->vma;
3321 bool write = vmf->flags & FAULT_FLAG_WRITE;
3322 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3326 if (!transhuge_vma_suitable(vma, haddr))
3327 return VM_FAULT_FALLBACK;
3329 ret = VM_FAULT_FALLBACK;
3330 page = compound_head(page);
3333 * Archs like ppc64 need additonal space to store information
3334 * related to pte entry. Use the preallocated table for that.
3336 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3337 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3338 if (!vmf->prealloc_pte)
3339 return VM_FAULT_OOM;
3340 smp_wmb(); /* See comment in __pte_alloc() */
3343 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3344 if (unlikely(!pmd_none(*vmf->pmd)))
3347 for (i = 0; i < HPAGE_PMD_NR; i++)
3348 flush_icache_page(vma, page + i);
3350 entry = mk_huge_pmd(page, vma->vm_page_prot);
3352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3354 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3355 page_add_file_rmap(page, true);
3357 * deposit and withdraw with pmd lock held
3359 if (arch_needs_pgtable_deposit())
3360 deposit_prealloc_pte(vmf);
3362 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3364 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3366 /* fault is handled */
3368 count_vm_event(THP_FILE_MAPPED);
3370 spin_unlock(vmf->ptl);
3374 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3382 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3383 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3385 * @vmf: fault environment
3386 * @memcg: memcg to charge page (only for private mappings)
3387 * @page: page to map
3389 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3392 * Target users are page handler itself and implementations of
3393 * vm_ops->map_pages.
3395 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3398 struct vm_area_struct *vma = vmf->vma;
3399 bool write = vmf->flags & FAULT_FLAG_WRITE;
3403 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3404 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3406 VM_BUG_ON_PAGE(memcg, page);
3408 ret = do_set_pmd(vmf, page);
3409 if (ret != VM_FAULT_FALLBACK)
3414 ret = pte_alloc_one_map(vmf);
3419 /* Re-check under ptl */
3420 if (unlikely(!pte_none(*vmf->pte)))
3421 return VM_FAULT_NOPAGE;
3423 flush_icache_page(vma, page);
3424 entry = mk_pte(page, vma->vm_page_prot);
3426 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3427 /* copy-on-write page */
3428 if (write && !(vma->vm_flags & VM_SHARED)) {
3429 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3430 page_add_new_anon_rmap(page, vma, vmf->address, false);
3431 mem_cgroup_commit_charge(page, memcg, false, false);
3432 lru_cache_add_active_or_unevictable(page, vma);
3434 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3435 page_add_file_rmap(page, false);
3437 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3439 /* no need to invalidate: a not-present page won't be cached */
3440 update_mmu_cache(vma, vmf->address, vmf->pte);
3447 * finish_fault - finish page fault once we have prepared the page to fault
3449 * @vmf: structure describing the fault
3451 * This function handles all that is needed to finish a page fault once the
3452 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3453 * given page, adds reverse page mapping, handles memcg charges and LRU
3454 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3457 * The function expects the page to be locked and on success it consumes a
3458 * reference of a page being mapped (for the PTE which maps it).
3460 int finish_fault(struct vm_fault *vmf)
3465 /* Did we COW the page? */
3466 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3467 !(vmf->vma->vm_flags & VM_SHARED))
3468 page = vmf->cow_page;
3473 * check even for read faults because we might have lost our CoWed
3476 if (!(vmf->vma->vm_flags & VM_SHARED))
3477 ret = check_stable_address_space(vmf->vma->vm_mm);
3479 ret = alloc_set_pte(vmf, vmf->memcg, page);
3481 pte_unmap_unlock(vmf->pte, vmf->ptl);
3485 static unsigned long fault_around_bytes __read_mostly =
3486 rounddown_pow_of_two(65536);
3488 #ifdef CONFIG_DEBUG_FS
3489 static int fault_around_bytes_get(void *data, u64 *val)
3491 *val = fault_around_bytes;
3496 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3497 * rounded down to nearest page order. It's what do_fault_around() expects to
3500 static int fault_around_bytes_set(void *data, u64 val)
3502 if (val / PAGE_SIZE > PTRS_PER_PTE)
3504 if (val > PAGE_SIZE)
3505 fault_around_bytes = rounddown_pow_of_two(val);
3507 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3510 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3511 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3513 static int __init fault_around_debugfs(void)
3517 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3518 &fault_around_bytes_fops);
3520 pr_warn("Failed to create fault_around_bytes in debugfs");
3523 late_initcall(fault_around_debugfs);
3527 * do_fault_around() tries to map few pages around the fault address. The hope
3528 * is that the pages will be needed soon and this will lower the number of
3531 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3532 * not ready to be mapped: not up-to-date, locked, etc.
3534 * This function is called with the page table lock taken. In the split ptlock
3535 * case the page table lock only protects only those entries which belong to
3536 * the page table corresponding to the fault address.
3538 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3541 * fault_around_pages() defines how many pages we'll try to map.
3542 * do_fault_around() expects it to return a power of two less than or equal to
3545 * The virtual address of the area that we map is naturally aligned to the
3546 * fault_around_pages() value (and therefore to page order). This way it's
3547 * easier to guarantee that we don't cross page table boundaries.
3549 static int do_fault_around(struct vm_fault *vmf)
3551 unsigned long address = vmf->address, nr_pages, mask;
3552 pgoff_t start_pgoff = vmf->pgoff;
3556 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3557 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3559 vmf->address = max(address & mask, vmf->vma->vm_start);
3560 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3564 * end_pgoff is either end of page table or end of vma
3565 * or fault_around_pages() from start_pgoff, depending what is nearest.
3567 end_pgoff = start_pgoff -
3568 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3570 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3571 start_pgoff + nr_pages - 1);
3573 if (pmd_none(*vmf->pmd)) {
3574 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3576 if (!vmf->prealloc_pte)
3578 smp_wmb(); /* See comment in __pte_alloc() */
3581 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3583 /* Huge page is mapped? Page fault is solved */
3584 if (pmd_trans_huge(*vmf->pmd)) {
3585 ret = VM_FAULT_NOPAGE;
3589 /* ->map_pages() haven't done anything useful. Cold page cache? */
3593 /* check if the page fault is solved */
3594 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3595 if (!pte_none(*vmf->pte))
3596 ret = VM_FAULT_NOPAGE;
3597 pte_unmap_unlock(vmf->pte, vmf->ptl);
3599 vmf->address = address;
3604 static int do_read_fault(struct vm_fault *vmf)
3606 struct vm_area_struct *vma = vmf->vma;
3610 * Let's call ->map_pages() first and use ->fault() as fallback
3611 * if page by the offset is not ready to be mapped (cold cache or
3614 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3615 ret = do_fault_around(vmf);
3620 ret = __do_fault(vmf);
3621 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3624 ret |= finish_fault(vmf);
3625 unlock_page(vmf->page);
3626 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3627 put_page(vmf->page);
3631 static int do_cow_fault(struct vm_fault *vmf)
3633 struct vm_area_struct *vma = vmf->vma;
3636 if (unlikely(anon_vma_prepare(vma)))
3637 return VM_FAULT_OOM;
3639 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3641 return VM_FAULT_OOM;
3643 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3644 &vmf->memcg, false)) {
3645 put_page(vmf->cow_page);
3646 return VM_FAULT_OOM;
3649 ret = __do_fault(vmf);
3650 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3652 if (ret & VM_FAULT_DONE_COW)
3655 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3656 __SetPageUptodate(vmf->cow_page);
3658 ret |= finish_fault(vmf);
3659 unlock_page(vmf->page);
3660 put_page(vmf->page);
3661 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3665 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3666 put_page(vmf->cow_page);
3670 static int do_shared_fault(struct vm_fault *vmf)
3672 struct vm_area_struct *vma = vmf->vma;
3675 ret = __do_fault(vmf);
3676 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3680 * Check if the backing address space wants to know that the page is
3681 * about to become writable
3683 if (vma->vm_ops->page_mkwrite) {
3684 unlock_page(vmf->page);
3685 tmp = do_page_mkwrite(vmf);
3686 if (unlikely(!tmp ||
3687 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3688 put_page(vmf->page);
3693 ret |= finish_fault(vmf);
3694 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3696 unlock_page(vmf->page);
3697 put_page(vmf->page);
3701 fault_dirty_shared_page(vma, vmf->page);
3706 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3707 * but allow concurrent faults).
3708 * The mmap_sem may have been released depending on flags and our
3709 * return value. See filemap_fault() and __lock_page_or_retry().
3711 static int do_fault(struct vm_fault *vmf)
3713 struct vm_area_struct *vma = vmf->vma;
3716 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3717 if (!vma->vm_ops->fault)
3718 ret = VM_FAULT_SIGBUS;
3719 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3720 ret = do_read_fault(vmf);
3721 else if (!(vma->vm_flags & VM_SHARED))
3722 ret = do_cow_fault(vmf);
3724 ret = do_shared_fault(vmf);
3726 /* preallocated pagetable is unused: free it */
3727 if (vmf->prealloc_pte) {
3728 pte_free(vma->vm_mm, vmf->prealloc_pte);
3729 vmf->prealloc_pte = NULL;
3734 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3735 unsigned long addr, int page_nid,
3740 count_vm_numa_event(NUMA_HINT_FAULTS);
3741 if (page_nid == numa_node_id()) {
3742 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3743 *flags |= TNF_FAULT_LOCAL;
3746 return mpol_misplaced(page, vma, addr);
3749 static int do_numa_page(struct vm_fault *vmf)
3751 struct vm_area_struct *vma = vmf->vma;
3752 struct page *page = NULL;
3756 bool migrated = false;
3758 bool was_writable = pte_savedwrite(vmf->orig_pte);
3762 * The "pte" at this point cannot be used safely without
3763 * validation through pte_unmap_same(). It's of NUMA type but
3764 * the pfn may be screwed if the read is non atomic.
3766 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3767 spin_lock(vmf->ptl);
3768 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3769 pte_unmap_unlock(vmf->pte, vmf->ptl);
3774 * Make it present again, Depending on how arch implementes non
3775 * accessible ptes, some can allow access by kernel mode.
3777 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3778 pte = pte_modify(pte, vma->vm_page_prot);
3779 pte = pte_mkyoung(pte);
3781 pte = pte_mkwrite(pte);
3782 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3783 update_mmu_cache(vma, vmf->address, vmf->pte);
3785 page = vm_normal_page(vma, vmf->address, pte);
3787 pte_unmap_unlock(vmf->pte, vmf->ptl);
3791 /* TODO: handle PTE-mapped THP */
3792 if (PageCompound(page)) {
3793 pte_unmap_unlock(vmf->pte, vmf->ptl);
3798 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3799 * much anyway since they can be in shared cache state. This misses
3800 * the case where a mapping is writable but the process never writes
3801 * to it but pte_write gets cleared during protection updates and
3802 * pte_dirty has unpredictable behaviour between PTE scan updates,
3803 * background writeback, dirty balancing and application behaviour.
3805 if (!pte_write(pte))
3806 flags |= TNF_NO_GROUP;
3809 * Flag if the page is shared between multiple address spaces. This
3810 * is later used when determining whether to group tasks together
3812 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3813 flags |= TNF_SHARED;
3815 last_cpupid = page_cpupid_last(page);
3816 page_nid = page_to_nid(page);
3817 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3819 pte_unmap_unlock(vmf->pte, vmf->ptl);
3820 if (target_nid == -1) {
3825 /* Migrate to the requested node */
3826 migrated = migrate_misplaced_page(page, vma, target_nid);
3828 page_nid = target_nid;
3829 flags |= TNF_MIGRATED;
3831 flags |= TNF_MIGRATE_FAIL;
3835 task_numa_fault(last_cpupid, page_nid, 1, flags);
3839 static inline int create_huge_pmd(struct vm_fault *vmf)
3841 if (vma_is_anonymous(vmf->vma))
3842 return do_huge_pmd_anonymous_page(vmf);
3843 if (vmf->vma->vm_ops->huge_fault)
3844 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3845 return VM_FAULT_FALLBACK;
3848 /* `inline' is required to avoid gcc 4.1.2 build error */
3849 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3851 if (vma_is_anonymous(vmf->vma))
3852 return do_huge_pmd_wp_page(vmf, orig_pmd);
3853 if (vmf->vma->vm_ops->huge_fault)
3854 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3856 /* COW handled on pte level: split pmd */
3857 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3858 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3860 return VM_FAULT_FALLBACK;
3863 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3865 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3868 static int create_huge_pud(struct vm_fault *vmf)
3870 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3871 /* No support for anonymous transparent PUD pages yet */
3872 if (vma_is_anonymous(vmf->vma))
3873 return VM_FAULT_FALLBACK;
3874 if (vmf->vma->vm_ops->huge_fault)
3875 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3876 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3877 return VM_FAULT_FALLBACK;
3880 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3882 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3883 /* No support for anonymous transparent PUD pages yet */
3884 if (vma_is_anonymous(vmf->vma))
3885 return VM_FAULT_FALLBACK;
3886 if (vmf->vma->vm_ops->huge_fault)
3887 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3888 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3889 return VM_FAULT_FALLBACK;
3893 * These routines also need to handle stuff like marking pages dirty
3894 * and/or accessed for architectures that don't do it in hardware (most
3895 * RISC architectures). The early dirtying is also good on the i386.
3897 * There is also a hook called "update_mmu_cache()" that architectures
3898 * with external mmu caches can use to update those (ie the Sparc or
3899 * PowerPC hashed page tables that act as extended TLBs).
3901 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3902 * concurrent faults).
3904 * The mmap_sem may have been released depending on flags and our return value.
3905 * See filemap_fault() and __lock_page_or_retry().
3907 static int handle_pte_fault(struct vm_fault *vmf)
3911 if (unlikely(pmd_none(*vmf->pmd))) {
3913 * Leave __pte_alloc() until later: because vm_ops->fault may
3914 * want to allocate huge page, and if we expose page table
3915 * for an instant, it will be difficult to retract from
3916 * concurrent faults and from rmap lookups.
3920 /* See comment in pte_alloc_one_map() */
3921 if (pmd_devmap_trans_unstable(vmf->pmd))
3924 * A regular pmd is established and it can't morph into a huge
3925 * pmd from under us anymore at this point because we hold the
3926 * mmap_sem read mode and khugepaged takes it in write mode.
3927 * So now it's safe to run pte_offset_map().
3929 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3930 vmf->orig_pte = *vmf->pte;
3933 * some architectures can have larger ptes than wordsize,
3934 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3935 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3936 * accesses. The code below just needs a consistent view
3937 * for the ifs and we later double check anyway with the
3938 * ptl lock held. So here a barrier will do.
3941 if (pte_none(vmf->orig_pte)) {
3942 pte_unmap(vmf->pte);
3948 if (vma_is_anonymous(vmf->vma))
3949 return do_anonymous_page(vmf);
3951 return do_fault(vmf);
3954 if (!pte_present(vmf->orig_pte))
3955 return do_swap_page(vmf);
3957 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3958 return do_numa_page(vmf);
3960 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3961 spin_lock(vmf->ptl);
3962 entry = vmf->orig_pte;
3963 if (unlikely(!pte_same(*vmf->pte, entry)))
3965 if (vmf->flags & FAULT_FLAG_WRITE) {
3966 if (!pte_write(entry))
3967 return do_wp_page(vmf);
3968 entry = pte_mkdirty(entry);
3970 entry = pte_mkyoung(entry);
3971 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3972 vmf->flags & FAULT_FLAG_WRITE)) {
3973 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3976 * This is needed only for protection faults but the arch code
3977 * is not yet telling us if this is a protection fault or not.
3978 * This still avoids useless tlb flushes for .text page faults
3981 if (vmf->flags & FAULT_FLAG_WRITE)
3982 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3985 pte_unmap_unlock(vmf->pte, vmf->ptl);
3990 * By the time we get here, we already hold the mm semaphore
3992 * The mmap_sem may have been released depending on flags and our
3993 * return value. See filemap_fault() and __lock_page_or_retry().
3995 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3998 struct vm_fault vmf = {
4000 .address = address & PAGE_MASK,
4002 .pgoff = linear_page_index(vma, address),
4003 .gfp_mask = __get_fault_gfp_mask(vma),
4005 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4006 struct mm_struct *mm = vma->vm_mm;
4011 pgd = pgd_offset(mm, address);
4012 p4d = p4d_alloc(mm, pgd, address);
4014 return VM_FAULT_OOM;
4016 vmf.pud = pud_alloc(mm, p4d, address);
4018 return VM_FAULT_OOM;
4019 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4020 ret = create_huge_pud(&vmf);
4021 if (!(ret & VM_FAULT_FALLBACK))
4024 pud_t orig_pud = *vmf.pud;
4027 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4029 /* NUMA case for anonymous PUDs would go here */
4031 if (dirty && !pud_write(orig_pud)) {
4032 ret = wp_huge_pud(&vmf, orig_pud);
4033 if (!(ret & VM_FAULT_FALLBACK))
4036 huge_pud_set_accessed(&vmf, orig_pud);
4042 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4044 return VM_FAULT_OOM;
4045 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4046 ret = create_huge_pmd(&vmf);
4047 if (!(ret & VM_FAULT_FALLBACK))
4050 pmd_t orig_pmd = *vmf.pmd;
4053 if (unlikely(is_swap_pmd(orig_pmd))) {
4054 VM_BUG_ON(thp_migration_supported() &&
4055 !is_pmd_migration_entry(orig_pmd));
4056 if (is_pmd_migration_entry(orig_pmd))
4057 pmd_migration_entry_wait(mm, vmf.pmd);
4060 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4061 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4062 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4064 if (dirty && !pmd_write(orig_pmd)) {
4065 ret = wp_huge_pmd(&vmf, orig_pmd);
4066 if (!(ret & VM_FAULT_FALLBACK))
4069 huge_pmd_set_accessed(&vmf, orig_pmd);
4075 return handle_pte_fault(&vmf);
4079 * By the time we get here, we already hold the mm semaphore
4081 * The mmap_sem may have been released depending on flags and our
4082 * return value. See filemap_fault() and __lock_page_or_retry().
4084 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4089 __set_current_state(TASK_RUNNING);
4091 count_vm_event(PGFAULT);
4092 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4094 /* do counter updates before entering really critical section. */
4095 check_sync_rss_stat(current);
4097 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4098 flags & FAULT_FLAG_INSTRUCTION,
4099 flags & FAULT_FLAG_REMOTE))
4100 return VM_FAULT_SIGSEGV;
4103 * Enable the memcg OOM handling for faults triggered in user
4104 * space. Kernel faults are handled more gracefully.
4106 if (flags & FAULT_FLAG_USER)
4107 mem_cgroup_oom_enable();
4109 if (unlikely(is_vm_hugetlb_page(vma)))
4110 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4112 ret = __handle_mm_fault(vma, address, flags);
4114 if (flags & FAULT_FLAG_USER) {
4115 mem_cgroup_oom_disable();
4117 * The task may have entered a memcg OOM situation but
4118 * if the allocation error was handled gracefully (no
4119 * VM_FAULT_OOM), there is no need to kill anything.
4120 * Just clean up the OOM state peacefully.
4122 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4123 mem_cgroup_oom_synchronize(false);
4128 EXPORT_SYMBOL_GPL(handle_mm_fault);
4130 #ifndef __PAGETABLE_P4D_FOLDED
4132 * Allocate p4d page table.
4133 * We've already handled the fast-path in-line.
4135 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4137 p4d_t *new = p4d_alloc_one(mm, address);
4141 smp_wmb(); /* See comment in __pte_alloc */
4143 spin_lock(&mm->page_table_lock);
4144 if (pgd_present(*pgd)) /* Another has populated it */
4147 pgd_populate(mm, pgd, new);
4148 spin_unlock(&mm->page_table_lock);
4151 #endif /* __PAGETABLE_P4D_FOLDED */
4153 #ifndef __PAGETABLE_PUD_FOLDED
4155 * Allocate page upper directory.
4156 * We've already handled the fast-path in-line.
4158 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4160 pud_t *new = pud_alloc_one(mm, address);
4164 smp_wmb(); /* See comment in __pte_alloc */
4166 spin_lock(&mm->page_table_lock);
4167 #ifndef __ARCH_HAS_5LEVEL_HACK
4168 if (!p4d_present(*p4d)) {
4170 p4d_populate(mm, p4d, new);
4171 } else /* Another has populated it */
4174 if (!pgd_present(*p4d)) {
4176 pgd_populate(mm, p4d, new);
4177 } else /* Another has populated it */
4179 #endif /* __ARCH_HAS_5LEVEL_HACK */
4180 spin_unlock(&mm->page_table_lock);
4183 #endif /* __PAGETABLE_PUD_FOLDED */
4185 #ifndef __PAGETABLE_PMD_FOLDED
4187 * Allocate page middle directory.
4188 * We've already handled the fast-path in-line.
4190 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4193 pmd_t *new = pmd_alloc_one(mm, address);
4197 smp_wmb(); /* See comment in __pte_alloc */
4199 ptl = pud_lock(mm, pud);
4200 #ifndef __ARCH_HAS_4LEVEL_HACK
4201 if (!pud_present(*pud)) {
4203 pud_populate(mm, pud, new);
4204 } else /* Another has populated it */
4207 if (!pgd_present(*pud)) {
4209 pgd_populate(mm, pud, new);
4210 } else /* Another has populated it */
4212 #endif /* __ARCH_HAS_4LEVEL_HACK */
4216 #endif /* __PAGETABLE_PMD_FOLDED */
4218 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4219 unsigned long *start, unsigned long *end,
4220 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4228 pgd = pgd_offset(mm, address);
4229 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4232 p4d = p4d_offset(pgd, address);
4233 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4236 pud = pud_offset(p4d, address);
4237 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4240 pmd = pmd_offset(pud, address);
4241 VM_BUG_ON(pmd_trans_huge(*pmd));
4243 if (pmd_huge(*pmd)) {
4248 *start = address & PMD_MASK;
4249 *end = *start + PMD_SIZE;
4250 mmu_notifier_invalidate_range_start(mm, *start, *end);
4252 *ptlp = pmd_lock(mm, pmd);
4253 if (pmd_huge(*pmd)) {
4259 mmu_notifier_invalidate_range_end(mm, *start, *end);
4262 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4266 *start = address & PAGE_MASK;
4267 *end = *start + PAGE_SIZE;
4268 mmu_notifier_invalidate_range_start(mm, *start, *end);
4270 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4271 if (!pte_present(*ptep))
4276 pte_unmap_unlock(ptep, *ptlp);
4278 mmu_notifier_invalidate_range_end(mm, *start, *end);
4283 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4284 pte_t **ptepp, spinlock_t **ptlp)
4288 /* (void) is needed to make gcc happy */
4289 (void) __cond_lock(*ptlp,
4290 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4291 ptepp, NULL, ptlp)));
4295 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4296 unsigned long *start, unsigned long *end,
4297 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4301 /* (void) is needed to make gcc happy */
4302 (void) __cond_lock(*ptlp,
4303 !(res = __follow_pte_pmd(mm, address, start, end,
4304 ptepp, pmdpp, ptlp)));
4307 EXPORT_SYMBOL(follow_pte_pmd);
4310 * follow_pfn - look up PFN at a user virtual address
4311 * @vma: memory mapping
4312 * @address: user virtual address
4313 * @pfn: location to store found PFN
4315 * Only IO mappings and raw PFN mappings are allowed.
4317 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4319 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4326 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4329 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4332 *pfn = pte_pfn(*ptep);
4333 pte_unmap_unlock(ptep, ptl);
4336 EXPORT_SYMBOL(follow_pfn);
4338 #ifdef CONFIG_HAVE_IOREMAP_PROT
4339 int follow_phys(struct vm_area_struct *vma,
4340 unsigned long address, unsigned int flags,
4341 unsigned long *prot, resource_size_t *phys)
4347 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4350 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4354 if ((flags & FOLL_WRITE) && !pte_write(pte))
4357 *prot = pgprot_val(pte_pgprot(pte));
4358 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4362 pte_unmap_unlock(ptep, ptl);
4367 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4368 void *buf, int len, int write)
4370 resource_size_t phys_addr;
4371 unsigned long prot = 0;
4372 void __iomem *maddr;
4373 int offset = addr & (PAGE_SIZE-1);
4375 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4378 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4380 memcpy_toio(maddr + offset, buf, len);
4382 memcpy_fromio(buf, maddr + offset, len);
4387 EXPORT_SYMBOL_GPL(generic_access_phys);
4391 * Access another process' address space as given in mm. If non-NULL, use the
4392 * given task for page fault accounting.
4394 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4395 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4397 struct vm_area_struct *vma;
4398 void *old_buf = buf;
4399 int write = gup_flags & FOLL_WRITE;
4401 down_read(&mm->mmap_sem);
4402 /* ignore errors, just check how much was successfully transferred */
4404 int bytes, ret, offset;
4406 struct page *page = NULL;
4408 ret = get_user_pages_remote(tsk, mm, addr, 1,
4409 gup_flags, &page, &vma, NULL);
4411 #ifndef CONFIG_HAVE_IOREMAP_PROT
4415 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4416 * we can access using slightly different code.
4418 vma = find_vma(mm, addr);
4419 if (!vma || vma->vm_start > addr)
4421 if (vma->vm_ops && vma->vm_ops->access)
4422 ret = vma->vm_ops->access(vma, addr, buf,
4430 offset = addr & (PAGE_SIZE-1);
4431 if (bytes > PAGE_SIZE-offset)
4432 bytes = PAGE_SIZE-offset;
4436 copy_to_user_page(vma, page, addr,
4437 maddr + offset, buf, bytes);
4438 set_page_dirty_lock(page);
4440 copy_from_user_page(vma, page, addr,
4441 buf, maddr + offset, bytes);
4450 up_read(&mm->mmap_sem);
4452 return buf - old_buf;
4456 * access_remote_vm - access another process' address space
4457 * @mm: the mm_struct of the target address space
4458 * @addr: start address to access
4459 * @buf: source or destination buffer
4460 * @len: number of bytes to transfer
4461 * @gup_flags: flags modifying lookup behaviour
4463 * The caller must hold a reference on @mm.
4465 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4466 void *buf, int len, unsigned int gup_flags)
4468 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4472 * Access another process' address space.
4473 * Source/target buffer must be kernel space,
4474 * Do not walk the page table directly, use get_user_pages
4476 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4477 void *buf, int len, unsigned int gup_flags)
4479 struct mm_struct *mm;
4482 mm = get_task_mm(tsk);
4486 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4492 EXPORT_SYMBOL_GPL(access_process_vm);
4495 * Print the name of a VMA.
4497 void print_vma_addr(char *prefix, unsigned long ip)
4499 struct mm_struct *mm = current->mm;
4500 struct vm_area_struct *vma;
4503 * we might be running from an atomic context so we cannot sleep
4505 if (!down_read_trylock(&mm->mmap_sem))
4508 vma = find_vma(mm, ip);
4509 if (vma && vma->vm_file) {
4510 struct file *f = vma->vm_file;
4511 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4515 p = file_path(f, buf, PAGE_SIZE);
4518 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4520 vma->vm_end - vma->vm_start);
4521 free_page((unsigned long)buf);
4524 up_read(&mm->mmap_sem);
4527 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4528 void __might_fault(const char *file, int line)
4531 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4532 * holding the mmap_sem, this is safe because kernel memory doesn't
4533 * get paged out, therefore we'll never actually fault, and the
4534 * below annotations will generate false positives.
4536 if (uaccess_kernel())
4538 if (pagefault_disabled())
4540 __might_sleep(file, line, 0);
4541 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4543 might_lock_read(¤t->mm->mmap_sem);
4546 EXPORT_SYMBOL(__might_fault);
4549 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4550 static void clear_gigantic_page(struct page *page,
4552 unsigned int pages_per_huge_page)
4555 struct page *p = page;
4558 for (i = 0; i < pages_per_huge_page;
4559 i++, p = mem_map_next(p, page, i)) {
4561 clear_user_highpage(p, addr + i * PAGE_SIZE);
4564 void clear_huge_page(struct page *page,
4565 unsigned long addr_hint, unsigned int pages_per_huge_page)
4568 unsigned long addr = addr_hint &
4569 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4571 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4572 clear_gigantic_page(page, addr, pages_per_huge_page);
4576 /* Clear sub-page to access last to keep its cache lines hot */
4578 n = (addr_hint - addr) / PAGE_SIZE;
4579 if (2 * n <= pages_per_huge_page) {
4580 /* If sub-page to access in first half of huge page */
4583 /* Clear sub-pages at the end of huge page */
4584 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4586 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4589 /* If sub-page to access in second half of huge page */
4590 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4591 l = pages_per_huge_page - n;
4592 /* Clear sub-pages at the begin of huge page */
4593 for (i = 0; i < base; i++) {
4595 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4599 * Clear remaining sub-pages in left-right-left-right pattern
4600 * towards the sub-page to access
4602 for (i = 0; i < l; i++) {
4603 int left_idx = base + i;
4604 int right_idx = base + 2 * l - 1 - i;
4607 clear_user_highpage(page + left_idx,
4608 addr + left_idx * PAGE_SIZE);
4610 clear_user_highpage(page + right_idx,
4611 addr + right_idx * PAGE_SIZE);
4615 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4617 struct vm_area_struct *vma,
4618 unsigned int pages_per_huge_page)
4621 struct page *dst_base = dst;
4622 struct page *src_base = src;
4624 for (i = 0; i < pages_per_huge_page; ) {
4626 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4629 dst = mem_map_next(dst, dst_base, i);
4630 src = mem_map_next(src, src_base, i);
4634 void copy_user_huge_page(struct page *dst, struct page *src,
4635 unsigned long addr, struct vm_area_struct *vma,
4636 unsigned int pages_per_huge_page)
4640 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4641 copy_user_gigantic_page(dst, src, addr, vma,
4642 pages_per_huge_page);
4647 for (i = 0; i < pages_per_huge_page; i++) {
4649 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4653 long copy_huge_page_from_user(struct page *dst_page,
4654 const void __user *usr_src,
4655 unsigned int pages_per_huge_page,
4656 bool allow_pagefault)
4658 void *src = (void *)usr_src;
4660 unsigned long i, rc = 0;
4661 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4663 for (i = 0; i < pages_per_huge_page; i++) {
4664 if (allow_pagefault)
4665 page_kaddr = kmap(dst_page + i);
4667 page_kaddr = kmap_atomic(dst_page + i);
4668 rc = copy_from_user(page_kaddr,
4669 (const void __user *)(src + i * PAGE_SIZE),
4671 if (allow_pagefault)
4672 kunmap(dst_page + i);
4674 kunmap_atomic(page_kaddr);
4676 ret_val -= (PAGE_SIZE - rc);
4684 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4686 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4688 static struct kmem_cache *page_ptl_cachep;
4690 void __init ptlock_cache_init(void)
4692 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4696 bool ptlock_alloc(struct page *page)
4700 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4707 void ptlock_free(struct page *page)
4709 kmem_cache_free(page_ptl_cachep, page->ptl);