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/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
93 EXPORT_SYMBOL(mem_map);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
119 static int __init disable_randmaps(char *s)
121 randomize_va_space = 0;
124 __setup("norandmaps", disable_randmaps);
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
129 unsigned long highest_memmap_pfn __read_mostly;
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 static int __init init_zero_pfn(void)
136 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 core_initcall(init_zero_pfn);
142 #if defined(SPLIT_RSS_COUNTING)
144 void sync_mm_rss(struct mm_struct *mm)
148 for (i = 0; i < NR_MM_COUNTERS; i++) {
149 if (current->rss_stat.count[i]) {
150 add_mm_counter(mm, i, current->rss_stat.count[i]);
151 current->rss_stat.count[i] = 0;
154 current->rss_stat.events = 0;
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 struct task_struct *task = current;
161 if (likely(task->mm == mm))
162 task->rss_stat.count[member] += val;
164 add_mm_counter(mm, member, val);
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH (64)
171 static void check_sync_rss_stat(struct task_struct *task)
173 if (unlikely(task != current))
175 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176 sync_mm_rss(task->mm);
178 #else /* SPLIT_RSS_COUNTING */
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183 static void check_sync_rss_stat(struct task_struct *task)
187 #endif /* SPLIT_RSS_COUNTING */
189 #ifdef HAVE_GENERIC_MMU_GATHER
191 static bool tlb_next_batch(struct mmu_gather *tlb)
193 struct mmu_gather_batch *batch;
197 tlb->active = batch->next;
201 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
204 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
211 batch->max = MAX_GATHER_BATCH;
213 tlb->active->next = batch;
219 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220 unsigned long start, unsigned long end)
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238 __tlb_reset_range(tlb);
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
251 __tlb_reset_range(tlb);
254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
256 struct mmu_gather_batch *batch;
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
262 tlb->active = &tlb->local;
265 void tlb_flush_mmu(struct mmu_gather *tlb)
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276 unsigned long start, unsigned long end, bool force)
278 struct mmu_gather_batch *batch, *next;
281 __tlb_adjust_range(tlb, start, end - start);
285 /* keep the page table cache within bounds */
288 for (batch = tlb->local.next; batch; batch = next) {
290 free_pages((unsigned long)batch, 0);
292 tlb->local.next = NULL;
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
302 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 struct mmu_gather_batch *batch;
306 VM_BUG_ON(!tlb->end);
307 VM_WARN_ON(tlb->page_size != page_size);
311 * Add the page and check if we are full. If so
314 batch->pages[batch->nr++] = page;
315 if (batch->nr == batch->max) {
316 if (!tlb_next_batch(tlb))
320 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
325 #endif /* HAVE_GENERIC_MMU_GATHER */
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
329 static void tlb_remove_table_smp_sync(void *arg)
331 struct mm_struct __maybe_unused *mm = arg;
333 * On most architectures this does nothing. Simply delivering the
334 * interrupt is enough to prevent races with software page table
335 * walking like that done in get_user_pages_fast.
337 * See the comment near struct mmu_table_batch.
339 tlb_flush_remove_tables_local(mm);
342 static void tlb_remove_table_one(void *table, struct mmu_gather *tlb)
345 * This isn't an RCU grace period and hence the page-tables cannot be
346 * assumed to be actually RCU-freed.
348 * It is however sufficient for software page-table walkers that rely on
349 * IRQ disabling. See the comment near struct mmu_table_batch.
351 smp_call_function(tlb_remove_table_smp_sync, tlb->mm, 1);
352 __tlb_remove_table(table);
355 static void tlb_remove_table_rcu(struct rcu_head *head)
357 struct mmu_table_batch *batch;
360 batch = container_of(head, struct mmu_table_batch, rcu);
362 for (i = 0; i < batch->nr; i++)
363 __tlb_remove_table(batch->tables[i]);
365 free_page((unsigned long)batch);
368 void tlb_table_flush(struct mmu_gather *tlb)
370 struct mmu_table_batch **batch = &tlb->batch;
372 tlb_flush_remove_tables(tlb->mm);
375 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
380 void tlb_remove_table(struct mmu_gather *tlb, void *table)
382 struct mmu_table_batch **batch = &tlb->batch;
385 * When there's less then two users of this mm there cannot be a
386 * concurrent page-table walk.
388 if (atomic_read(&tlb->mm->mm_users) < 2) {
389 __tlb_remove_table(table);
393 if (*batch == NULL) {
394 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
395 if (*batch == NULL) {
396 tlb_remove_table_one(table, tlb);
401 (*batch)->tables[(*batch)->nr++] = table;
402 if ((*batch)->nr == MAX_TABLE_BATCH)
403 tlb_table_flush(tlb);
406 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
409 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
410 * @tlb: the mmu_gather structure to initialize
411 * @mm: the mm_struct of the target address space
412 * @start: start of the region that will be removed from the page-table
413 * @end: end of the region that will be removed from the page-table
415 * Called to initialize an (on-stack) mmu_gather structure for page-table
416 * tear-down from @mm. The @start and @end are set to 0 and -1
417 * respectively when @mm is without users and we're going to destroy
418 * the full address space (exit/execve).
420 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
421 unsigned long start, unsigned long end)
423 arch_tlb_gather_mmu(tlb, mm, start, end);
424 inc_tlb_flush_pending(tlb->mm);
427 void tlb_finish_mmu(struct mmu_gather *tlb,
428 unsigned long start, unsigned long end)
431 * If there are parallel threads are doing PTE changes on same range
432 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
433 * flush by batching, a thread has stable TLB entry can fail to flush
434 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
435 * forcefully if we detect parallel PTE batching threads.
437 bool force = mm_tlb_flush_nested(tlb->mm);
439 arch_tlb_finish_mmu(tlb, start, end, force);
440 dec_tlb_flush_pending(tlb->mm);
444 * Note: this doesn't free the actual pages themselves. That
445 * has been handled earlier when unmapping all the memory regions.
447 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
450 pgtable_t token = pmd_pgtable(*pmd);
452 pte_free_tlb(tlb, token, addr);
453 mm_dec_nr_ptes(tlb->mm);
456 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
457 unsigned long addr, unsigned long end,
458 unsigned long floor, unsigned long ceiling)
465 pmd = pmd_offset(pud, addr);
467 next = pmd_addr_end(addr, end);
468 if (pmd_none_or_clear_bad(pmd))
470 free_pte_range(tlb, pmd, addr);
471 } while (pmd++, addr = next, addr != end);
481 if (end - 1 > ceiling - 1)
484 pmd = pmd_offset(pud, start);
486 pmd_free_tlb(tlb, pmd, start);
487 mm_dec_nr_pmds(tlb->mm);
490 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
491 unsigned long addr, unsigned long end,
492 unsigned long floor, unsigned long ceiling)
499 pud = pud_offset(p4d, addr);
501 next = pud_addr_end(addr, end);
502 if (pud_none_or_clear_bad(pud))
504 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
505 } while (pud++, addr = next, addr != end);
515 if (end - 1 > ceiling - 1)
518 pud = pud_offset(p4d, start);
520 pud_free_tlb(tlb, pud, start);
521 mm_dec_nr_puds(tlb->mm);
524 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 unsigned long floor, unsigned long ceiling)
533 p4d = p4d_offset(pgd, addr);
535 next = p4d_addr_end(addr, end);
536 if (p4d_none_or_clear_bad(p4d))
538 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
539 } while (p4d++, addr = next, addr != end);
545 ceiling &= PGDIR_MASK;
549 if (end - 1 > ceiling - 1)
552 p4d = p4d_offset(pgd, start);
554 p4d_free_tlb(tlb, p4d, start);
558 * This function frees user-level page tables of a process.
560 void free_pgd_range(struct mmu_gather *tlb,
561 unsigned long addr, unsigned long end,
562 unsigned long floor, unsigned long ceiling)
568 * The next few lines have given us lots of grief...
570 * Why are we testing PMD* at this top level? Because often
571 * there will be no work to do at all, and we'd prefer not to
572 * go all the way down to the bottom just to discover that.
574 * Why all these "- 1"s? Because 0 represents both the bottom
575 * of the address space and the top of it (using -1 for the
576 * top wouldn't help much: the masks would do the wrong thing).
577 * The rule is that addr 0 and floor 0 refer to the bottom of
578 * the address space, but end 0 and ceiling 0 refer to the top
579 * Comparisons need to use "end - 1" and "ceiling - 1" (though
580 * that end 0 case should be mythical).
582 * Wherever addr is brought up or ceiling brought down, we must
583 * be careful to reject "the opposite 0" before it confuses the
584 * subsequent tests. But what about where end is brought down
585 * by PMD_SIZE below? no, end can't go down to 0 there.
587 * Whereas we round start (addr) and ceiling down, by different
588 * masks at different levels, in order to test whether a table
589 * now has no other vmas using it, so can be freed, we don't
590 * bother to round floor or end up - the tests don't need that.
604 if (end - 1 > ceiling - 1)
609 * We add page table cache pages with PAGE_SIZE,
610 * (see pte_free_tlb()), flush the tlb if we need
612 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
613 pgd = pgd_offset(tlb->mm, addr);
615 next = pgd_addr_end(addr, end);
616 if (pgd_none_or_clear_bad(pgd))
618 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
619 } while (pgd++, addr = next, addr != end);
622 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
623 unsigned long floor, unsigned long ceiling)
626 struct vm_area_struct *next = vma->vm_next;
627 unsigned long addr = vma->vm_start;
630 * Hide vma from rmap and truncate_pagecache before freeing
633 unlink_anon_vmas(vma);
634 unlink_file_vma(vma);
636 if (is_vm_hugetlb_page(vma)) {
637 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
638 floor, next ? next->vm_start : ceiling);
641 * Optimization: gather nearby vmas into one call down
643 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
644 && !is_vm_hugetlb_page(next)) {
647 unlink_anon_vmas(vma);
648 unlink_file_vma(vma);
650 free_pgd_range(tlb, addr, vma->vm_end,
651 floor, next ? next->vm_start : ceiling);
657 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
660 pgtable_t new = pte_alloc_one(mm, address);
665 * Ensure all pte setup (eg. pte page lock and page clearing) are
666 * visible before the pte is made visible to other CPUs by being
667 * put into page tables.
669 * The other side of the story is the pointer chasing in the page
670 * table walking code (when walking the page table without locking;
671 * ie. most of the time). Fortunately, these data accesses consist
672 * of a chain of data-dependent loads, meaning most CPUs (alpha
673 * being the notable exception) will already guarantee loads are
674 * seen in-order. See the alpha page table accessors for the
675 * smp_read_barrier_depends() barriers in page table walking code.
677 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
679 ptl = pmd_lock(mm, pmd);
680 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
682 pmd_populate(mm, pmd, new);
691 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
693 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
697 smp_wmb(); /* See comment in __pte_alloc */
699 spin_lock(&init_mm.page_table_lock);
700 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
701 pmd_populate_kernel(&init_mm, pmd, new);
704 spin_unlock(&init_mm.page_table_lock);
706 pte_free_kernel(&init_mm, new);
710 static inline void init_rss_vec(int *rss)
712 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
715 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
719 if (current->mm == mm)
721 for (i = 0; i < NR_MM_COUNTERS; i++)
723 add_mm_counter(mm, i, rss[i]);
727 * This function is called to print an error when a bad pte
728 * is found. For example, we might have a PFN-mapped pte in
729 * a region that doesn't allow it.
731 * The calling function must still handle the error.
733 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
734 pte_t pte, struct page *page)
736 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
737 p4d_t *p4d = p4d_offset(pgd, addr);
738 pud_t *pud = pud_offset(p4d, addr);
739 pmd_t *pmd = pmd_offset(pud, addr);
740 struct address_space *mapping;
742 static unsigned long resume;
743 static unsigned long nr_shown;
744 static unsigned long nr_unshown;
747 * Allow a burst of 60 reports, then keep quiet for that minute;
748 * or allow a steady drip of one report per second.
750 if (nr_shown == 60) {
751 if (time_before(jiffies, resume)) {
756 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
763 resume = jiffies + 60 * HZ;
765 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
766 index = linear_page_index(vma, addr);
768 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
770 (long long)pte_val(pte), (long long)pmd_val(*pmd));
772 dump_page(page, "bad pte");
773 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
774 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
775 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
777 vma->vm_ops ? vma->vm_ops->fault : NULL,
778 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
779 mapping ? mapping->a_ops->readpage : NULL);
781 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
785 * vm_normal_page -- This function gets the "struct page" associated with a pte.
787 * "Special" mappings do not wish to be associated with a "struct page" (either
788 * it doesn't exist, or it exists but they don't want to touch it). In this
789 * case, NULL is returned here. "Normal" mappings do have a struct page.
791 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
792 * pte bit, in which case this function is trivial. Secondly, an architecture
793 * may not have a spare pte bit, which requires a more complicated scheme,
796 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
797 * special mapping (even if there are underlying and valid "struct pages").
798 * COWed pages of a VM_PFNMAP are always normal.
800 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
801 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
802 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
803 * mapping will always honor the rule
805 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
807 * And for normal mappings this is false.
809 * This restricts such mappings to be a linear translation from virtual address
810 * to pfn. To get around this restriction, we allow arbitrary mappings so long
811 * as the vma is not a COW mapping; in that case, we know that all ptes are
812 * special (because none can have been COWed).
815 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
817 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
818 * page" backing, however the difference is that _all_ pages with a struct
819 * page (that is, those where pfn_valid is true) are refcounted and considered
820 * normal pages by the VM. The disadvantage is that pages are refcounted
821 * (which can be slower and simply not an option for some PFNMAP users). The
822 * advantage is that we don't have to follow the strict linearity rule of
823 * PFNMAP mappings in order to support COWable mappings.
826 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
827 pte_t pte, bool with_public_device)
829 unsigned long pfn = pte_pfn(pte);
831 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
832 if (likely(!pte_special(pte)))
834 if (vma->vm_ops && vma->vm_ops->find_special_page)
835 return vma->vm_ops->find_special_page(vma, addr);
836 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
838 if (is_zero_pfn(pfn))
842 * Device public pages are special pages (they are ZONE_DEVICE
843 * pages but different from persistent memory). They behave
844 * allmost like normal pages. The difference is that they are
845 * not on the lru and thus should never be involve with any-
846 * thing that involve lru manipulation (mlock, numa balancing,
849 * This is why we still want to return NULL for such page from
850 * vm_normal_page() so that we do not have to special case all
851 * call site of vm_normal_page().
853 if (likely(pfn <= highest_memmap_pfn)) {
854 struct page *page = pfn_to_page(pfn);
856 if (is_device_public_page(page)) {
857 if (with_public_device)
862 print_bad_pte(vma, addr, pte, NULL);
866 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
868 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
869 if (vma->vm_flags & VM_MIXEDMAP) {
875 off = (addr - vma->vm_start) >> PAGE_SHIFT;
876 if (pfn == vma->vm_pgoff + off)
878 if (!is_cow_mapping(vma->vm_flags))
883 if (is_zero_pfn(pfn))
887 if (unlikely(pfn > highest_memmap_pfn)) {
888 print_bad_pte(vma, addr, pte, NULL);
893 * NOTE! We still have PageReserved() pages in the page tables.
894 * eg. VDSO mappings can cause them to exist.
897 return pfn_to_page(pfn);
900 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
901 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
904 unsigned long pfn = pmd_pfn(pmd);
907 * There is no pmd_special() but there may be special pmds, e.g.
908 * in a direct-access (dax) mapping, so let's just replicate the
909 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
911 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
912 if (vma->vm_flags & VM_MIXEDMAP) {
918 off = (addr - vma->vm_start) >> PAGE_SHIFT;
919 if (pfn == vma->vm_pgoff + off)
921 if (!is_cow_mapping(vma->vm_flags))
926 if (is_zero_pfn(pfn))
928 if (unlikely(pfn > highest_memmap_pfn))
932 * NOTE! We still have PageReserved() pages in the page tables.
933 * eg. VDSO mappings can cause them to exist.
936 return pfn_to_page(pfn);
941 * copy one vm_area from one task to the other. Assumes the page tables
942 * already present in the new task to be cleared in the whole range
943 * covered by this vma.
946 static inline unsigned long
947 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
948 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
949 unsigned long addr, int *rss)
951 unsigned long vm_flags = vma->vm_flags;
952 pte_t pte = *src_pte;
955 /* pte contains position in swap or file, so copy. */
956 if (unlikely(!pte_present(pte))) {
957 swp_entry_t entry = pte_to_swp_entry(pte);
959 if (likely(!non_swap_entry(entry))) {
960 if (swap_duplicate(entry) < 0)
963 /* make sure dst_mm is on swapoff's mmlist. */
964 if (unlikely(list_empty(&dst_mm->mmlist))) {
965 spin_lock(&mmlist_lock);
966 if (list_empty(&dst_mm->mmlist))
967 list_add(&dst_mm->mmlist,
969 spin_unlock(&mmlist_lock);
972 } else if (is_migration_entry(entry)) {
973 page = migration_entry_to_page(entry);
975 rss[mm_counter(page)]++;
977 if (is_write_migration_entry(entry) &&
978 is_cow_mapping(vm_flags)) {
980 * COW mappings require pages in both
981 * parent and child to be set to read.
983 make_migration_entry_read(&entry);
984 pte = swp_entry_to_pte(entry);
985 if (pte_swp_soft_dirty(*src_pte))
986 pte = pte_swp_mksoft_dirty(pte);
987 set_pte_at(src_mm, addr, src_pte, pte);
989 } else if (is_device_private_entry(entry)) {
990 page = device_private_entry_to_page(entry);
993 * Update rss count even for unaddressable pages, as
994 * they should treated just like normal pages in this
997 * We will likely want to have some new rss counters
998 * for unaddressable pages, at some point. But for now
999 * keep things as they are.
1002 rss[mm_counter(page)]++;
1003 page_dup_rmap(page, false);
1006 * We do not preserve soft-dirty information, because so
1007 * far, checkpoint/restore is the only feature that
1008 * requires that. And checkpoint/restore does not work
1009 * when a device driver is involved (you cannot easily
1010 * save and restore device driver state).
1012 if (is_write_device_private_entry(entry) &&
1013 is_cow_mapping(vm_flags)) {
1014 make_device_private_entry_read(&entry);
1015 pte = swp_entry_to_pte(entry);
1016 set_pte_at(src_mm, addr, src_pte, pte);
1023 * If it's a COW mapping, write protect it both
1024 * in the parent and the child
1026 if (is_cow_mapping(vm_flags)) {
1027 ptep_set_wrprotect(src_mm, addr, src_pte);
1028 pte = pte_wrprotect(pte);
1032 * If it's a shared mapping, mark it clean in
1035 if (vm_flags & VM_SHARED)
1036 pte = pte_mkclean(pte);
1037 pte = pte_mkold(pte);
1039 page = vm_normal_page(vma, addr, pte);
1042 page_dup_rmap(page, false);
1043 rss[mm_counter(page)]++;
1044 } else if (pte_devmap(pte)) {
1045 page = pte_page(pte);
1048 * Cache coherent device memory behave like regular page and
1049 * not like persistent memory page. For more informations see
1050 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1052 if (is_device_public_page(page)) {
1054 page_dup_rmap(page, false);
1055 rss[mm_counter(page)]++;
1060 set_pte_at(dst_mm, addr, dst_pte, pte);
1064 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1065 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1066 unsigned long addr, unsigned long end)
1068 pte_t *orig_src_pte, *orig_dst_pte;
1069 pte_t *src_pte, *dst_pte;
1070 spinlock_t *src_ptl, *dst_ptl;
1072 int rss[NR_MM_COUNTERS];
1073 swp_entry_t entry = (swp_entry_t){0};
1078 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1081 src_pte = pte_offset_map(src_pmd, addr);
1082 src_ptl = pte_lockptr(src_mm, src_pmd);
1083 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1084 orig_src_pte = src_pte;
1085 orig_dst_pte = dst_pte;
1086 arch_enter_lazy_mmu_mode();
1090 * We are holding two locks at this point - either of them
1091 * could generate latencies in another task on another CPU.
1093 if (progress >= 32) {
1095 if (need_resched() ||
1096 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1099 if (pte_none(*src_pte)) {
1103 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1108 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1110 arch_leave_lazy_mmu_mode();
1111 spin_unlock(src_ptl);
1112 pte_unmap(orig_src_pte);
1113 add_mm_rss_vec(dst_mm, rss);
1114 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1118 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1127 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1128 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1129 unsigned long addr, unsigned long end)
1131 pmd_t *src_pmd, *dst_pmd;
1134 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1137 src_pmd = pmd_offset(src_pud, addr);
1139 next = pmd_addr_end(addr, end);
1140 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1141 || pmd_devmap(*src_pmd)) {
1143 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1144 err = copy_huge_pmd(dst_mm, src_mm,
1145 dst_pmd, src_pmd, addr, vma);
1152 if (pmd_none_or_clear_bad(src_pmd))
1154 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1157 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1161 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1162 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1163 unsigned long addr, unsigned long end)
1165 pud_t *src_pud, *dst_pud;
1168 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1171 src_pud = pud_offset(src_p4d, addr);
1173 next = pud_addr_end(addr, end);
1174 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1177 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1178 err = copy_huge_pud(dst_mm, src_mm,
1179 dst_pud, src_pud, addr, vma);
1186 if (pud_none_or_clear_bad(src_pud))
1188 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1191 } while (dst_pud++, src_pud++, addr = next, addr != end);
1195 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1196 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1197 unsigned long addr, unsigned long end)
1199 p4d_t *src_p4d, *dst_p4d;
1202 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1205 src_p4d = p4d_offset(src_pgd, addr);
1207 next = p4d_addr_end(addr, end);
1208 if (p4d_none_or_clear_bad(src_p4d))
1210 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1213 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1217 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1218 struct vm_area_struct *vma)
1220 pgd_t *src_pgd, *dst_pgd;
1222 unsigned long addr = vma->vm_start;
1223 unsigned long end = vma->vm_end;
1224 unsigned long mmun_start; /* For mmu_notifiers */
1225 unsigned long mmun_end; /* For mmu_notifiers */
1230 * Don't copy ptes where a page fault will fill them correctly.
1231 * Fork becomes much lighter when there are big shared or private
1232 * readonly mappings. The tradeoff is that copy_page_range is more
1233 * efficient than faulting.
1235 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1239 if (is_vm_hugetlb_page(vma))
1240 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1242 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1244 * We do not free on error cases below as remove_vma
1245 * gets called on error from higher level routine
1247 ret = track_pfn_copy(vma);
1253 * We need to invalidate the secondary MMU mappings only when
1254 * there could be a permission downgrade on the ptes of the
1255 * parent mm. And a permission downgrade will only happen if
1256 * is_cow_mapping() returns true.
1258 is_cow = is_cow_mapping(vma->vm_flags);
1262 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1266 dst_pgd = pgd_offset(dst_mm, addr);
1267 src_pgd = pgd_offset(src_mm, addr);
1269 next = pgd_addr_end(addr, end);
1270 if (pgd_none_or_clear_bad(src_pgd))
1272 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1273 vma, addr, next))) {
1277 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1280 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1284 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1285 struct vm_area_struct *vma, pmd_t *pmd,
1286 unsigned long addr, unsigned long end,
1287 struct zap_details *details)
1289 struct mm_struct *mm = tlb->mm;
1290 int force_flush = 0;
1291 int rss[NR_MM_COUNTERS];
1297 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1300 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1302 flush_tlb_batched_pending(mm);
1303 arch_enter_lazy_mmu_mode();
1306 if (pte_none(ptent))
1309 if (pte_present(ptent)) {
1312 page = _vm_normal_page(vma, addr, ptent, true);
1313 if (unlikely(details) && page) {
1315 * unmap_shared_mapping_pages() wants to
1316 * invalidate cache without truncating:
1317 * unmap shared but keep private pages.
1319 if (details->check_mapping &&
1320 details->check_mapping != page_rmapping(page))
1323 ptent = ptep_get_and_clear_full(mm, addr, pte,
1325 tlb_remove_tlb_entry(tlb, pte, addr);
1326 if (unlikely(!page))
1329 if (!PageAnon(page)) {
1330 if (pte_dirty(ptent)) {
1332 set_page_dirty(page);
1334 if (pte_young(ptent) &&
1335 likely(!(vma->vm_flags & VM_SEQ_READ)))
1336 mark_page_accessed(page);
1338 rss[mm_counter(page)]--;
1339 page_remove_rmap(page, false);
1340 if (unlikely(page_mapcount(page) < 0))
1341 print_bad_pte(vma, addr, ptent, page);
1342 if (unlikely(__tlb_remove_page(tlb, page))) {
1350 entry = pte_to_swp_entry(ptent);
1351 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1352 struct page *page = device_private_entry_to_page(entry);
1354 if (unlikely(details && details->check_mapping)) {
1356 * unmap_shared_mapping_pages() wants to
1357 * invalidate cache without truncating:
1358 * unmap shared but keep private pages.
1360 if (details->check_mapping !=
1361 page_rmapping(page))
1365 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1366 rss[mm_counter(page)]--;
1367 page_remove_rmap(page, false);
1372 /* If details->check_mapping, we leave swap entries. */
1373 if (unlikely(details))
1376 entry = pte_to_swp_entry(ptent);
1377 if (!non_swap_entry(entry))
1379 else if (is_migration_entry(entry)) {
1382 page = migration_entry_to_page(entry);
1383 rss[mm_counter(page)]--;
1385 if (unlikely(!free_swap_and_cache(entry)))
1386 print_bad_pte(vma, addr, ptent, NULL);
1387 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1388 } while (pte++, addr += PAGE_SIZE, addr != end);
1390 add_mm_rss_vec(mm, rss);
1391 arch_leave_lazy_mmu_mode();
1393 /* Do the actual TLB flush before dropping ptl */
1395 tlb_flush_mmu_tlbonly(tlb);
1396 pte_unmap_unlock(start_pte, ptl);
1399 * If we forced a TLB flush (either due to running out of
1400 * batch buffers or because we needed to flush dirty TLB
1401 * entries before releasing the ptl), free the batched
1402 * memory too. Restart if we didn't do everything.
1406 tlb_flush_mmu_free(tlb);
1414 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1415 struct vm_area_struct *vma, pud_t *pud,
1416 unsigned long addr, unsigned long end,
1417 struct zap_details *details)
1422 pmd = pmd_offset(pud, addr);
1424 next = pmd_addr_end(addr, end);
1425 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1426 if (next - addr != HPAGE_PMD_SIZE)
1427 __split_huge_pmd(vma, pmd, addr, false, NULL);
1428 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1433 * Here there can be other concurrent MADV_DONTNEED or
1434 * trans huge page faults running, and if the pmd is
1435 * none or trans huge it can change under us. This is
1436 * because MADV_DONTNEED holds the mmap_sem in read
1439 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1441 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1444 } while (pmd++, addr = next, addr != end);
1449 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1450 struct vm_area_struct *vma, p4d_t *p4d,
1451 unsigned long addr, unsigned long end,
1452 struct zap_details *details)
1457 pud = pud_offset(p4d, addr);
1459 next = pud_addr_end(addr, end);
1460 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1461 if (next - addr != HPAGE_PUD_SIZE) {
1462 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1463 split_huge_pud(vma, pud, addr);
1464 } else if (zap_huge_pud(tlb, vma, pud, addr))
1468 if (pud_none_or_clear_bad(pud))
1470 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1473 } while (pud++, addr = next, addr != end);
1478 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1479 struct vm_area_struct *vma, pgd_t *pgd,
1480 unsigned long addr, unsigned long end,
1481 struct zap_details *details)
1486 p4d = p4d_offset(pgd, addr);
1488 next = p4d_addr_end(addr, end);
1489 if (p4d_none_or_clear_bad(p4d))
1491 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1492 } while (p4d++, addr = next, addr != end);
1497 void unmap_page_range(struct mmu_gather *tlb,
1498 struct vm_area_struct *vma,
1499 unsigned long addr, unsigned long end,
1500 struct zap_details *details)
1505 BUG_ON(addr >= end);
1506 tlb_start_vma(tlb, vma);
1507 pgd = pgd_offset(vma->vm_mm, addr);
1509 next = pgd_addr_end(addr, end);
1510 if (pgd_none_or_clear_bad(pgd))
1512 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1513 } while (pgd++, addr = next, addr != end);
1514 tlb_end_vma(tlb, vma);
1518 static void unmap_single_vma(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, unsigned long start_addr,
1520 unsigned long end_addr,
1521 struct zap_details *details)
1523 unsigned long start = max(vma->vm_start, start_addr);
1526 if (start >= vma->vm_end)
1528 end = min(vma->vm_end, end_addr);
1529 if (end <= vma->vm_start)
1533 uprobe_munmap(vma, start, end);
1535 if (unlikely(vma->vm_flags & VM_PFNMAP))
1536 untrack_pfn(vma, 0, 0);
1539 if (unlikely(is_vm_hugetlb_page(vma))) {
1541 * It is undesirable to test vma->vm_file as it
1542 * should be non-null for valid hugetlb area.
1543 * However, vm_file will be NULL in the error
1544 * cleanup path of mmap_region. When
1545 * hugetlbfs ->mmap method fails,
1546 * mmap_region() nullifies vma->vm_file
1547 * before calling this function to clean up.
1548 * Since no pte has actually been setup, it is
1549 * safe to do nothing in this case.
1552 i_mmap_lock_write(vma->vm_file->f_mapping);
1553 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1554 i_mmap_unlock_write(vma->vm_file->f_mapping);
1557 unmap_page_range(tlb, vma, start, end, details);
1562 * unmap_vmas - unmap a range of memory covered by a list of vma's
1563 * @tlb: address of the caller's struct mmu_gather
1564 * @vma: the starting vma
1565 * @start_addr: virtual address at which to start unmapping
1566 * @end_addr: virtual address at which to end unmapping
1568 * Unmap all pages in the vma list.
1570 * Only addresses between `start' and `end' will be unmapped.
1572 * The VMA list must be sorted in ascending virtual address order.
1574 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1575 * range after unmap_vmas() returns. So the only responsibility here is to
1576 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1577 * drops the lock and schedules.
1579 void unmap_vmas(struct mmu_gather *tlb,
1580 struct vm_area_struct *vma, unsigned long start_addr,
1581 unsigned long end_addr)
1583 struct mm_struct *mm = vma->vm_mm;
1585 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1586 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1587 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1588 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1592 * zap_page_range - remove user pages in a given range
1593 * @vma: vm_area_struct holding the applicable pages
1594 * @start: starting address of pages to zap
1595 * @size: number of bytes to zap
1597 * Caller must protect the VMA list
1599 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1602 struct mm_struct *mm = vma->vm_mm;
1603 struct mmu_gather tlb;
1604 unsigned long end = start + size;
1607 tlb_gather_mmu(&tlb, mm, start, end);
1608 update_hiwater_rss(mm);
1609 mmu_notifier_invalidate_range_start(mm, start, end);
1610 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1611 unmap_single_vma(&tlb, vma, start, end, NULL);
1614 * zap_page_range does not specify whether mmap_sem should be
1615 * held for read or write. That allows parallel zap_page_range
1616 * operations to unmap a PTE and defer a flush meaning that
1617 * this call observes pte_none and fails to flush the TLB.
1618 * Rather than adding a complex API, ensure that no stale
1619 * TLB entries exist when this call returns.
1621 flush_tlb_range(vma, start, end);
1624 mmu_notifier_invalidate_range_end(mm, start, end);
1625 tlb_finish_mmu(&tlb, start, end);
1629 * zap_page_range_single - remove user pages in a given range
1630 * @vma: vm_area_struct holding the applicable pages
1631 * @address: starting address of pages to zap
1632 * @size: number of bytes to zap
1633 * @details: details of shared cache invalidation
1635 * The range must fit into one VMA.
1637 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1638 unsigned long size, struct zap_details *details)
1640 struct mm_struct *mm = vma->vm_mm;
1641 struct mmu_gather tlb;
1642 unsigned long end = address + size;
1645 tlb_gather_mmu(&tlb, mm, address, end);
1646 update_hiwater_rss(mm);
1647 mmu_notifier_invalidate_range_start(mm, address, end);
1648 unmap_single_vma(&tlb, vma, address, end, details);
1649 mmu_notifier_invalidate_range_end(mm, address, end);
1650 tlb_finish_mmu(&tlb, address, end);
1654 * zap_vma_ptes - remove ptes mapping the vma
1655 * @vma: vm_area_struct holding ptes to be zapped
1656 * @address: starting address of pages to zap
1657 * @size: number of bytes to zap
1659 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1661 * The entire address range must be fully contained within the vma.
1664 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1667 if (address < vma->vm_start || address + size > vma->vm_end ||
1668 !(vma->vm_flags & VM_PFNMAP))
1671 zap_page_range_single(vma, address, size, NULL);
1673 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1675 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1683 pgd = pgd_offset(mm, addr);
1684 p4d = p4d_alloc(mm, pgd, addr);
1687 pud = pud_alloc(mm, p4d, addr);
1690 pmd = pmd_alloc(mm, pud, addr);
1694 VM_BUG_ON(pmd_trans_huge(*pmd));
1695 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1699 * This is the old fallback for page remapping.
1701 * For historical reasons, it only allows reserved pages. Only
1702 * old drivers should use this, and they needed to mark their
1703 * pages reserved for the old functions anyway.
1705 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1706 struct page *page, pgprot_t prot)
1708 struct mm_struct *mm = vma->vm_mm;
1717 flush_dcache_page(page);
1718 pte = get_locked_pte(mm, addr, &ptl);
1722 if (!pte_none(*pte))
1725 /* Ok, finally just insert the thing.. */
1727 inc_mm_counter_fast(mm, mm_counter_file(page));
1728 page_add_file_rmap(page, false);
1729 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1732 pte_unmap_unlock(pte, ptl);
1735 pte_unmap_unlock(pte, ptl);
1741 * vm_insert_page - insert single page into user vma
1742 * @vma: user vma to map to
1743 * @addr: target user address of this page
1744 * @page: source kernel page
1746 * This allows drivers to insert individual pages they've allocated
1749 * The page has to be a nice clean _individual_ kernel allocation.
1750 * If you allocate a compound page, you need to have marked it as
1751 * such (__GFP_COMP), or manually just split the page up yourself
1752 * (see split_page()).
1754 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1755 * took an arbitrary page protection parameter. This doesn't allow
1756 * that. Your vma protection will have to be set up correctly, which
1757 * means that if you want a shared writable mapping, you'd better
1758 * ask for a shared writable mapping!
1760 * The page does not need to be reserved.
1762 * Usually this function is called from f_op->mmap() handler
1763 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1764 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1765 * function from other places, for example from page-fault handler.
1767 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1770 if (addr < vma->vm_start || addr >= vma->vm_end)
1772 if (!page_count(page))
1774 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1775 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1776 BUG_ON(vma->vm_flags & VM_PFNMAP);
1777 vma->vm_flags |= VM_MIXEDMAP;
1779 return insert_page(vma, addr, page, vma->vm_page_prot);
1781 EXPORT_SYMBOL(vm_insert_page);
1783 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1784 pfn_t pfn, pgprot_t prot, bool mkwrite)
1786 struct mm_struct *mm = vma->vm_mm;
1792 pte = get_locked_pte(mm, addr, &ptl);
1796 if (!pte_none(*pte)) {
1799 * For read faults on private mappings the PFN passed
1800 * in may not match the PFN we have mapped if the
1801 * mapped PFN is a writeable COW page. In the mkwrite
1802 * case we are creating a writable PTE for a shared
1803 * mapping and we expect the PFNs to match.
1805 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1813 /* Ok, finally just insert the thing.. */
1814 if (pfn_t_devmap(pfn))
1815 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1817 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1821 entry = pte_mkyoung(entry);
1822 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1825 set_pte_at(mm, addr, pte, entry);
1826 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1830 pte_unmap_unlock(pte, ptl);
1836 * vm_insert_pfn - insert single pfn into user vma
1837 * @vma: user vma to map to
1838 * @addr: target user address of this page
1839 * @pfn: source kernel pfn
1841 * Similar to vm_insert_page, this allows drivers to insert individual pages
1842 * they've allocated into a user vma. Same comments apply.
1844 * This function should only be called from a vm_ops->fault handler, and
1845 * in that case the handler should return NULL.
1847 * vma cannot be a COW mapping.
1849 * As this is called only for pages that do not currently exist, we
1850 * do not need to flush old virtual caches or the TLB.
1852 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1855 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1857 EXPORT_SYMBOL(vm_insert_pfn);
1860 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1861 * @vma: user vma to map to
1862 * @addr: target user address of this page
1863 * @pfn: source kernel pfn
1864 * @pgprot: pgprot flags for the inserted page
1866 * This is exactly like vm_insert_pfn, except that it allows drivers to
1867 * to override pgprot on a per-page basis.
1869 * This only makes sense for IO mappings, and it makes no sense for
1870 * cow mappings. In general, using multiple vmas is preferable;
1871 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1874 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1875 unsigned long pfn, pgprot_t pgprot)
1879 * Technically, architectures with pte_special can avoid all these
1880 * restrictions (same for remap_pfn_range). However we would like
1881 * consistency in testing and feature parity among all, so we should
1882 * try to keep these invariants in place for everybody.
1884 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1885 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1886 (VM_PFNMAP|VM_MIXEDMAP));
1887 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1888 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1890 if (addr < vma->vm_start || addr >= vma->vm_end)
1893 if (!pfn_modify_allowed(pfn, pgprot))
1896 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1898 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1903 EXPORT_SYMBOL(vm_insert_pfn_prot);
1905 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1907 /* these checks mirror the abort conditions in vm_normal_page */
1908 if (vma->vm_flags & VM_MIXEDMAP)
1910 if (pfn_t_devmap(pfn))
1912 if (pfn_t_special(pfn))
1914 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1919 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1920 pfn_t pfn, bool mkwrite)
1922 pgprot_t pgprot = vma->vm_page_prot;
1924 BUG_ON(!vm_mixed_ok(vma, pfn));
1926 if (addr < vma->vm_start || addr >= vma->vm_end)
1929 track_pfn_insert(vma, &pgprot, pfn);
1931 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1935 * If we don't have pte special, then we have to use the pfn_valid()
1936 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1937 * refcount the page if pfn_valid is true (hence insert_page rather
1938 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1939 * without pte special, it would there be refcounted as a normal page.
1941 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1942 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1946 * At this point we are committed to insert_page()
1947 * regardless of whether the caller specified flags that
1948 * result in pfn_t_has_page() == false.
1950 page = pfn_to_page(pfn_t_to_pfn(pfn));
1951 return insert_page(vma, addr, page, pgprot);
1953 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1956 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1959 return __vm_insert_mixed(vma, addr, pfn, false);
1962 EXPORT_SYMBOL(vm_insert_mixed);
1965 * If the insertion of PTE failed because someone else already added a
1966 * different entry in the mean time, we treat that as success as we assume
1967 * the same entry was actually inserted.
1970 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1971 unsigned long addr, pfn_t pfn)
1975 err = __vm_insert_mixed(vma, addr, pfn, true);
1977 return VM_FAULT_OOM;
1978 if (err < 0 && err != -EBUSY)
1979 return VM_FAULT_SIGBUS;
1980 return VM_FAULT_NOPAGE;
1982 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1985 * maps a range of physical memory into the requested pages. the old
1986 * mappings are removed. any references to nonexistent pages results
1987 * in null mappings (currently treated as "copy-on-access")
1989 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1990 unsigned long addr, unsigned long end,
1991 unsigned long pfn, pgprot_t prot)
1997 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2000 arch_enter_lazy_mmu_mode();
2002 BUG_ON(!pte_none(*pte));
2003 if (!pfn_modify_allowed(pfn, prot)) {
2007 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2009 } while (pte++, addr += PAGE_SIZE, addr != end);
2010 arch_leave_lazy_mmu_mode();
2011 pte_unmap_unlock(pte - 1, ptl);
2015 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2016 unsigned long addr, unsigned long end,
2017 unsigned long pfn, pgprot_t prot)
2023 pfn -= addr >> PAGE_SHIFT;
2024 pmd = pmd_alloc(mm, pud, addr);
2027 VM_BUG_ON(pmd_trans_huge(*pmd));
2029 next = pmd_addr_end(addr, end);
2030 err = remap_pte_range(mm, pmd, addr, next,
2031 pfn + (addr >> PAGE_SHIFT), prot);
2034 } while (pmd++, addr = next, addr != end);
2038 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2039 unsigned long addr, unsigned long end,
2040 unsigned long pfn, pgprot_t prot)
2046 pfn -= addr >> PAGE_SHIFT;
2047 pud = pud_alloc(mm, p4d, addr);
2051 next = pud_addr_end(addr, end);
2052 err = remap_pmd_range(mm, pud, addr, next,
2053 pfn + (addr >> PAGE_SHIFT), prot);
2056 } while (pud++, addr = next, addr != end);
2060 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2061 unsigned long addr, unsigned long end,
2062 unsigned long pfn, pgprot_t prot)
2068 pfn -= addr >> PAGE_SHIFT;
2069 p4d = p4d_alloc(mm, pgd, addr);
2073 next = p4d_addr_end(addr, end);
2074 err = remap_pud_range(mm, p4d, addr, next,
2075 pfn + (addr >> PAGE_SHIFT), prot);
2078 } while (p4d++, addr = next, addr != end);
2083 * remap_pfn_range - remap kernel memory to userspace
2084 * @vma: user vma to map to
2085 * @addr: target user address to start at
2086 * @pfn: physical address of kernel memory
2087 * @size: size of map area
2088 * @prot: page protection flags for this mapping
2090 * Note: this is only safe if the mm semaphore is held when called.
2092 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2093 unsigned long pfn, unsigned long size, pgprot_t prot)
2097 unsigned long end = addr + PAGE_ALIGN(size);
2098 struct mm_struct *mm = vma->vm_mm;
2099 unsigned long remap_pfn = pfn;
2103 * Physically remapped pages are special. Tell the
2104 * rest of the world about it:
2105 * VM_IO tells people not to look at these pages
2106 * (accesses can have side effects).
2107 * VM_PFNMAP tells the core MM that the base pages are just
2108 * raw PFN mappings, and do not have a "struct page" associated
2111 * Disable vma merging and expanding with mremap().
2113 * Omit vma from core dump, even when VM_IO turned off.
2115 * There's a horrible special case to handle copy-on-write
2116 * behaviour that some programs depend on. We mark the "original"
2117 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2118 * See vm_normal_page() for details.
2120 if (is_cow_mapping(vma->vm_flags)) {
2121 if (addr != vma->vm_start || end != vma->vm_end)
2123 vma->vm_pgoff = pfn;
2126 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2130 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2132 BUG_ON(addr >= end);
2133 pfn -= addr >> PAGE_SHIFT;
2134 pgd = pgd_offset(mm, addr);
2135 flush_cache_range(vma, addr, end);
2137 next = pgd_addr_end(addr, end);
2138 err = remap_p4d_range(mm, pgd, addr, next,
2139 pfn + (addr >> PAGE_SHIFT), prot);
2142 } while (pgd++, addr = next, addr != end);
2145 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2149 EXPORT_SYMBOL(remap_pfn_range);
2152 * vm_iomap_memory - remap memory to userspace
2153 * @vma: user vma to map to
2154 * @start: start of area
2155 * @len: size of area
2157 * This is a simplified io_remap_pfn_range() for common driver use. The
2158 * driver just needs to give us the physical memory range to be mapped,
2159 * we'll figure out the rest from the vma information.
2161 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2162 * whatever write-combining details or similar.
2164 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2166 unsigned long vm_len, pfn, pages;
2168 /* Check that the physical memory area passed in looks valid */
2169 if (start + len < start)
2172 * You *really* shouldn't map things that aren't page-aligned,
2173 * but we've historically allowed it because IO memory might
2174 * just have smaller alignment.
2176 len += start & ~PAGE_MASK;
2177 pfn = start >> PAGE_SHIFT;
2178 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2179 if (pfn + pages < pfn)
2182 /* We start the mapping 'vm_pgoff' pages into the area */
2183 if (vma->vm_pgoff > pages)
2185 pfn += vma->vm_pgoff;
2186 pages -= vma->vm_pgoff;
2188 /* Can we fit all of the mapping? */
2189 vm_len = vma->vm_end - vma->vm_start;
2190 if (vm_len >> PAGE_SHIFT > pages)
2193 /* Ok, let it rip */
2194 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2196 EXPORT_SYMBOL(vm_iomap_memory);
2198 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2199 unsigned long addr, unsigned long end,
2200 pte_fn_t fn, void *data)
2205 spinlock_t *uninitialized_var(ptl);
2207 pte = (mm == &init_mm) ?
2208 pte_alloc_kernel(pmd, addr) :
2209 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2213 BUG_ON(pmd_huge(*pmd));
2215 arch_enter_lazy_mmu_mode();
2217 token = pmd_pgtable(*pmd);
2220 err = fn(pte++, token, addr, data);
2223 } while (addr += PAGE_SIZE, addr != end);
2225 arch_leave_lazy_mmu_mode();
2228 pte_unmap_unlock(pte-1, ptl);
2232 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2233 unsigned long addr, unsigned long end,
2234 pte_fn_t fn, void *data)
2240 BUG_ON(pud_huge(*pud));
2242 pmd = pmd_alloc(mm, pud, addr);
2246 next = pmd_addr_end(addr, end);
2247 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2250 } while (pmd++, addr = next, addr != end);
2254 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2255 unsigned long addr, unsigned long end,
2256 pte_fn_t fn, void *data)
2262 pud = pud_alloc(mm, p4d, addr);
2266 next = pud_addr_end(addr, end);
2267 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2270 } while (pud++, addr = next, addr != end);
2274 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2275 unsigned long addr, unsigned long end,
2276 pte_fn_t fn, void *data)
2282 p4d = p4d_alloc(mm, pgd, addr);
2286 next = p4d_addr_end(addr, end);
2287 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2290 } while (p4d++, addr = next, addr != end);
2295 * Scan a region of virtual memory, filling in page tables as necessary
2296 * and calling a provided function on each leaf page table.
2298 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2299 unsigned long size, pte_fn_t fn, void *data)
2303 unsigned long end = addr + size;
2306 if (WARN_ON(addr >= end))
2309 pgd = pgd_offset(mm, addr);
2311 next = pgd_addr_end(addr, end);
2312 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2315 } while (pgd++, addr = next, addr != end);
2319 EXPORT_SYMBOL_GPL(apply_to_page_range);
2322 * handle_pte_fault chooses page fault handler according to an entry which was
2323 * read non-atomically. Before making any commitment, on those architectures
2324 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2325 * parts, do_swap_page must check under lock before unmapping the pte and
2326 * proceeding (but do_wp_page is only called after already making such a check;
2327 * and do_anonymous_page can safely check later on).
2329 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2330 pte_t *page_table, pte_t orig_pte)
2333 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2334 if (sizeof(pte_t) > sizeof(unsigned long)) {
2335 spinlock_t *ptl = pte_lockptr(mm, pmd);
2337 same = pte_same(*page_table, orig_pte);
2341 pte_unmap(page_table);
2345 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2347 debug_dma_assert_idle(src);
2350 * If the source page was a PFN mapping, we don't have
2351 * a "struct page" for it. We do a best-effort copy by
2352 * just copying from the original user address. If that
2353 * fails, we just zero-fill it. Live with it.
2355 if (unlikely(!src)) {
2356 void *kaddr = kmap_atomic(dst);
2357 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2360 * This really shouldn't fail, because the page is there
2361 * in the page tables. But it might just be unreadable,
2362 * in which case we just give up and fill the result with
2365 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2367 kunmap_atomic(kaddr);
2368 flush_dcache_page(dst);
2370 copy_user_highpage(dst, src, va, vma);
2373 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2375 struct file *vm_file = vma->vm_file;
2378 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2381 * Special mappings (e.g. VDSO) do not have any file so fake
2382 * a default GFP_KERNEL for them.
2388 * Notify the address space that the page is about to become writable so that
2389 * it can prohibit this or wait for the page to get into an appropriate state.
2391 * We do this without the lock held, so that it can sleep if it needs to.
2393 static int do_page_mkwrite(struct vm_fault *vmf)
2396 struct page *page = vmf->page;
2397 unsigned int old_flags = vmf->flags;
2399 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2401 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2402 /* Restore original flags so that caller is not surprised */
2403 vmf->flags = old_flags;
2404 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2406 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2408 if (!page->mapping) {
2410 return 0; /* retry */
2412 ret |= VM_FAULT_LOCKED;
2414 VM_BUG_ON_PAGE(!PageLocked(page), page);
2419 * Handle dirtying of a page in shared file mapping on a write fault.
2421 * The function expects the page to be locked and unlocks it.
2423 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2426 struct address_space *mapping;
2428 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2430 dirtied = set_page_dirty(page);
2431 VM_BUG_ON_PAGE(PageAnon(page), page);
2433 * Take a local copy of the address_space - page.mapping may be zeroed
2434 * by truncate after unlock_page(). The address_space itself remains
2435 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2436 * release semantics to prevent the compiler from undoing this copying.
2438 mapping = page_rmapping(page);
2441 if ((dirtied || page_mkwrite) && mapping) {
2443 * Some device drivers do not set page.mapping
2444 * but still dirty their pages
2446 balance_dirty_pages_ratelimited(mapping);
2450 file_update_time(vma->vm_file);
2454 * Handle write page faults for pages that can be reused in the current vma
2456 * This can happen either due to the mapping being with the VM_SHARED flag,
2457 * or due to us being the last reference standing to the page. In either
2458 * case, all we need to do here is to mark the page as writable and update
2459 * any related book-keeping.
2461 static inline void wp_page_reuse(struct vm_fault *vmf)
2462 __releases(vmf->ptl)
2464 struct vm_area_struct *vma = vmf->vma;
2465 struct page *page = vmf->page;
2468 * Clear the pages cpupid information as the existing
2469 * information potentially belongs to a now completely
2470 * unrelated process.
2473 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2475 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2476 entry = pte_mkyoung(vmf->orig_pte);
2477 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2478 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2479 update_mmu_cache(vma, vmf->address, vmf->pte);
2480 pte_unmap_unlock(vmf->pte, vmf->ptl);
2484 * Handle the case of a page which we actually need to copy to a new page.
2486 * Called with mmap_sem locked and the old page referenced, but
2487 * without the ptl held.
2489 * High level logic flow:
2491 * - Allocate a page, copy the content of the old page to the new one.
2492 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2493 * - Take the PTL. If the pte changed, bail out and release the allocated page
2494 * - If the pte is still the way we remember it, update the page table and all
2495 * relevant references. This includes dropping the reference the page-table
2496 * held to the old page, as well as updating the rmap.
2497 * - In any case, unlock the PTL and drop the reference we took to the old page.
2499 static int wp_page_copy(struct vm_fault *vmf)
2501 struct vm_area_struct *vma = vmf->vma;
2502 struct mm_struct *mm = vma->vm_mm;
2503 struct page *old_page = vmf->page;
2504 struct page *new_page = NULL;
2506 int page_copied = 0;
2507 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2508 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2509 struct mem_cgroup *memcg;
2511 if (unlikely(anon_vma_prepare(vma)))
2514 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2515 new_page = alloc_zeroed_user_highpage_movable(vma,
2520 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2524 cow_user_page(new_page, old_page, vmf->address, vma);
2527 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2530 __SetPageUptodate(new_page);
2532 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2535 * Re-check the pte - we dropped the lock
2537 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2538 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2540 if (!PageAnon(old_page)) {
2541 dec_mm_counter_fast(mm,
2542 mm_counter_file(old_page));
2543 inc_mm_counter_fast(mm, MM_ANONPAGES);
2546 inc_mm_counter_fast(mm, MM_ANONPAGES);
2548 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2549 entry = mk_pte(new_page, vma->vm_page_prot);
2550 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2552 * Clear the pte entry and flush it first, before updating the
2553 * pte with the new entry. This will avoid a race condition
2554 * seen in the presence of one thread doing SMC and another
2557 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2558 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2559 mem_cgroup_commit_charge(new_page, memcg, false, false);
2560 lru_cache_add_active_or_unevictable(new_page, vma);
2562 * We call the notify macro here because, when using secondary
2563 * mmu page tables (such as kvm shadow page tables), we want the
2564 * new page to be mapped directly into the secondary page table.
2566 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2567 update_mmu_cache(vma, vmf->address, vmf->pte);
2570 * Only after switching the pte to the new page may
2571 * we remove the mapcount here. Otherwise another
2572 * process may come and find the rmap count decremented
2573 * before the pte is switched to the new page, and
2574 * "reuse" the old page writing into it while our pte
2575 * here still points into it and can be read by other
2578 * The critical issue is to order this
2579 * page_remove_rmap with the ptp_clear_flush above.
2580 * Those stores are ordered by (if nothing else,)
2581 * the barrier present in the atomic_add_negative
2582 * in page_remove_rmap.
2584 * Then the TLB flush in ptep_clear_flush ensures that
2585 * no process can access the old page before the
2586 * decremented mapcount is visible. And the old page
2587 * cannot be reused until after the decremented
2588 * mapcount is visible. So transitively, TLBs to
2589 * old page will be flushed before it can be reused.
2591 page_remove_rmap(old_page, false);
2594 /* Free the old page.. */
2595 new_page = old_page;
2598 mem_cgroup_cancel_charge(new_page, memcg, false);
2604 pte_unmap_unlock(vmf->pte, vmf->ptl);
2606 * No need to double call mmu_notifier->invalidate_range() callback as
2607 * the above ptep_clear_flush_notify() did already call it.
2609 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2612 * Don't let another task, with possibly unlocked vma,
2613 * keep the mlocked page.
2615 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2616 lock_page(old_page); /* LRU manipulation */
2617 if (PageMlocked(old_page))
2618 munlock_vma_page(old_page);
2619 unlock_page(old_page);
2623 return page_copied ? VM_FAULT_WRITE : 0;
2629 return VM_FAULT_OOM;
2633 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2634 * writeable once the page is prepared
2636 * @vmf: structure describing the fault
2638 * This function handles all that is needed to finish a write page fault in a
2639 * shared mapping due to PTE being read-only once the mapped page is prepared.
2640 * It handles locking of PTE and modifying it. The function returns
2641 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2644 * The function expects the page to be locked or other protection against
2645 * concurrent faults / writeback (such as DAX radix tree locks).
2647 int finish_mkwrite_fault(struct vm_fault *vmf)
2649 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2650 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2653 * We might have raced with another page fault while we released the
2654 * pte_offset_map_lock.
2656 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2657 pte_unmap_unlock(vmf->pte, vmf->ptl);
2658 return VM_FAULT_NOPAGE;
2665 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2668 static int wp_pfn_shared(struct vm_fault *vmf)
2670 struct vm_area_struct *vma = vmf->vma;
2672 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2675 pte_unmap_unlock(vmf->pte, vmf->ptl);
2676 vmf->flags |= FAULT_FLAG_MKWRITE;
2677 ret = vma->vm_ops->pfn_mkwrite(vmf);
2678 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2680 return finish_mkwrite_fault(vmf);
2683 return VM_FAULT_WRITE;
2686 static int wp_page_shared(struct vm_fault *vmf)
2687 __releases(vmf->ptl)
2689 struct vm_area_struct *vma = vmf->vma;
2691 get_page(vmf->page);
2693 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2696 pte_unmap_unlock(vmf->pte, vmf->ptl);
2697 tmp = do_page_mkwrite(vmf);
2698 if (unlikely(!tmp || (tmp &
2699 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2700 put_page(vmf->page);
2703 tmp = finish_mkwrite_fault(vmf);
2704 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2705 unlock_page(vmf->page);
2706 put_page(vmf->page);
2711 lock_page(vmf->page);
2713 fault_dirty_shared_page(vma, vmf->page);
2714 put_page(vmf->page);
2716 return VM_FAULT_WRITE;
2720 * This routine handles present pages, when users try to write
2721 * to a shared page. It is done by copying the page to a new address
2722 * and decrementing the shared-page counter for the old page.
2724 * Note that this routine assumes that the protection checks have been
2725 * done by the caller (the low-level page fault routine in most cases).
2726 * Thus we can safely just mark it writable once we've done any necessary
2729 * We also mark the page dirty at this point even though the page will
2730 * change only once the write actually happens. This avoids a few races,
2731 * and potentially makes it more efficient.
2733 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2734 * but allow concurrent faults), with pte both mapped and locked.
2735 * We return with mmap_sem still held, but pte unmapped and unlocked.
2737 static int do_wp_page(struct vm_fault *vmf)
2738 __releases(vmf->ptl)
2740 struct vm_area_struct *vma = vmf->vma;
2742 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2745 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2748 * We should not cow pages in a shared writeable mapping.
2749 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2751 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2752 (VM_WRITE|VM_SHARED))
2753 return wp_pfn_shared(vmf);
2755 pte_unmap_unlock(vmf->pte, vmf->ptl);
2756 return wp_page_copy(vmf);
2760 * Take out anonymous pages first, anonymous shared vmas are
2761 * not dirty accountable.
2763 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2764 int total_map_swapcount;
2765 if (!trylock_page(vmf->page)) {
2766 get_page(vmf->page);
2767 pte_unmap_unlock(vmf->pte, vmf->ptl);
2768 lock_page(vmf->page);
2769 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2770 vmf->address, &vmf->ptl);
2771 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2772 unlock_page(vmf->page);
2773 pte_unmap_unlock(vmf->pte, vmf->ptl);
2774 put_page(vmf->page);
2777 put_page(vmf->page);
2779 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2780 if (total_map_swapcount == 1) {
2782 * The page is all ours. Move it to
2783 * our anon_vma so the rmap code will
2784 * not search our parent or siblings.
2785 * Protected against the rmap code by
2788 page_move_anon_rmap(vmf->page, vma);
2790 unlock_page(vmf->page);
2792 return VM_FAULT_WRITE;
2794 unlock_page(vmf->page);
2795 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2796 (VM_WRITE|VM_SHARED))) {
2797 return wp_page_shared(vmf);
2801 * Ok, we need to copy. Oh, well..
2803 get_page(vmf->page);
2805 pte_unmap_unlock(vmf->pte, vmf->ptl);
2806 return wp_page_copy(vmf);
2809 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2810 unsigned long start_addr, unsigned long end_addr,
2811 struct zap_details *details)
2813 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2816 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2817 struct zap_details *details)
2819 struct vm_area_struct *vma;
2820 pgoff_t vba, vea, zba, zea;
2822 vma_interval_tree_foreach(vma, root,
2823 details->first_index, details->last_index) {
2825 vba = vma->vm_pgoff;
2826 vea = vba + vma_pages(vma) - 1;
2827 zba = details->first_index;
2830 zea = details->last_index;
2834 unmap_mapping_range_vma(vma,
2835 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2836 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2842 * unmap_mapping_pages() - Unmap pages from processes.
2843 * @mapping: The address space containing pages to be unmapped.
2844 * @start: Index of first page to be unmapped.
2845 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2846 * @even_cows: Whether to unmap even private COWed pages.
2848 * Unmap the pages in this address space from any userspace process which
2849 * has them mmaped. Generally, you want to remove COWed pages as well when
2850 * a file is being truncated, but not when invalidating pages from the page
2853 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2854 pgoff_t nr, bool even_cows)
2856 struct zap_details details = { };
2858 details.check_mapping = even_cows ? NULL : mapping;
2859 details.first_index = start;
2860 details.last_index = start + nr - 1;
2861 if (details.last_index < details.first_index)
2862 details.last_index = ULONG_MAX;
2864 i_mmap_lock_write(mapping);
2865 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2866 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2867 i_mmap_unlock_write(mapping);
2871 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2872 * address_space corresponding to the specified byte range in the underlying
2875 * @mapping: the address space containing mmaps to be unmapped.
2876 * @holebegin: byte in first page to unmap, relative to the start of
2877 * the underlying file. This will be rounded down to a PAGE_SIZE
2878 * boundary. Note that this is different from truncate_pagecache(), which
2879 * must keep the partial page. In contrast, we must get rid of
2881 * @holelen: size of prospective hole in bytes. This will be rounded
2882 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2884 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2885 * but 0 when invalidating pagecache, don't throw away private data.
2887 void unmap_mapping_range(struct address_space *mapping,
2888 loff_t const holebegin, loff_t const holelen, int even_cows)
2890 pgoff_t hba = holebegin >> PAGE_SHIFT;
2891 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2893 /* Check for overflow. */
2894 if (sizeof(holelen) > sizeof(hlen)) {
2896 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2897 if (holeend & ~(long long)ULONG_MAX)
2898 hlen = ULONG_MAX - hba + 1;
2901 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2903 EXPORT_SYMBOL(unmap_mapping_range);
2906 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2907 * but allow concurrent faults), and pte mapped but not yet locked.
2908 * We return with pte unmapped and unlocked.
2910 * We return with the mmap_sem locked or unlocked in the same cases
2911 * as does filemap_fault().
2913 int do_swap_page(struct vm_fault *vmf)
2915 struct vm_area_struct *vma = vmf->vma;
2916 struct page *page = NULL, *swapcache;
2917 struct mem_cgroup *memcg;
2924 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2927 entry = pte_to_swp_entry(vmf->orig_pte);
2928 if (unlikely(non_swap_entry(entry))) {
2929 if (is_migration_entry(entry)) {
2930 migration_entry_wait(vma->vm_mm, vmf->pmd,
2932 } else if (is_device_private_entry(entry)) {
2934 * For un-addressable device memory we call the pgmap
2935 * fault handler callback. The callback must migrate
2936 * the page back to some CPU accessible page.
2938 ret = device_private_entry_fault(vma, vmf->address, entry,
2939 vmf->flags, vmf->pmd);
2940 } else if (is_hwpoison_entry(entry)) {
2941 ret = VM_FAULT_HWPOISON;
2943 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2944 ret = VM_FAULT_SIGBUS;
2950 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2951 page = lookup_swap_cache(entry, vma, vmf->address);
2955 struct swap_info_struct *si = swp_swap_info(entry);
2957 if (si->flags & SWP_SYNCHRONOUS_IO &&
2958 __swap_count(si, entry) == 1) {
2959 /* skip swapcache */
2960 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2963 __SetPageLocked(page);
2964 __SetPageSwapBacked(page);
2965 set_page_private(page, entry.val);
2966 lru_cache_add_anon(page);
2967 swap_readpage(page, true);
2970 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2977 * Back out if somebody else faulted in this pte
2978 * while we released the pte lock.
2980 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2981 vmf->address, &vmf->ptl);
2982 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2984 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2988 /* Had to read the page from swap area: Major fault */
2989 ret = VM_FAULT_MAJOR;
2990 count_vm_event(PGMAJFAULT);
2991 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2992 } else if (PageHWPoison(page)) {
2994 * hwpoisoned dirty swapcache pages are kept for killing
2995 * owner processes (which may be unknown at hwpoison time)
2997 ret = VM_FAULT_HWPOISON;
2998 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3002 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3004 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3006 ret |= VM_FAULT_RETRY;
3011 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3012 * release the swapcache from under us. The page pin, and pte_same
3013 * test below, are not enough to exclude that. Even if it is still
3014 * swapcache, we need to check that the page's swap has not changed.
3016 if (unlikely((!PageSwapCache(page) ||
3017 page_private(page) != entry.val)) && swapcache)
3020 page = ksm_might_need_to_copy(page, vma, vmf->address);
3021 if (unlikely(!page)) {
3027 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3034 * Back out if somebody else already faulted in this pte.
3036 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3038 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3041 if (unlikely(!PageUptodate(page))) {
3042 ret = VM_FAULT_SIGBUS;
3047 * The page isn't present yet, go ahead with the fault.
3049 * Be careful about the sequence of operations here.
3050 * To get its accounting right, reuse_swap_page() must be called
3051 * while the page is counted on swap but not yet in mapcount i.e.
3052 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3053 * must be called after the swap_free(), or it will never succeed.
3056 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3057 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3058 pte = mk_pte(page, vma->vm_page_prot);
3059 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3060 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3061 vmf->flags &= ~FAULT_FLAG_WRITE;
3062 ret |= VM_FAULT_WRITE;
3063 exclusive = RMAP_EXCLUSIVE;
3065 flush_icache_page(vma, page);
3066 if (pte_swp_soft_dirty(vmf->orig_pte))
3067 pte = pte_mksoft_dirty(pte);
3068 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3069 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3070 vmf->orig_pte = pte;
3072 /* ksm created a completely new copy */
3073 if (unlikely(page != swapcache && swapcache)) {
3074 page_add_new_anon_rmap(page, vma, vmf->address, false);
3075 mem_cgroup_commit_charge(page, memcg, false, false);
3076 lru_cache_add_active_or_unevictable(page, vma);
3078 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3079 mem_cgroup_commit_charge(page, memcg, true, false);
3080 activate_page(page);
3084 if (mem_cgroup_swap_full(page) ||
3085 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3086 try_to_free_swap(page);
3088 if (page != swapcache && swapcache) {
3090 * Hold the lock to avoid the swap entry to be reused
3091 * until we take the PT lock for the pte_same() check
3092 * (to avoid false positives from pte_same). For
3093 * further safety release the lock after the swap_free
3094 * so that the swap count won't change under a
3095 * parallel locked swapcache.
3097 unlock_page(swapcache);
3098 put_page(swapcache);
3101 if (vmf->flags & FAULT_FLAG_WRITE) {
3102 ret |= do_wp_page(vmf);
3103 if (ret & VM_FAULT_ERROR)
3104 ret &= VM_FAULT_ERROR;
3108 /* No need to invalidate - it was non-present before */
3109 update_mmu_cache(vma, vmf->address, vmf->pte);
3111 pte_unmap_unlock(vmf->pte, vmf->ptl);
3115 mem_cgroup_cancel_charge(page, memcg, false);
3116 pte_unmap_unlock(vmf->pte, vmf->ptl);
3121 if (page != swapcache && swapcache) {
3122 unlock_page(swapcache);
3123 put_page(swapcache);
3129 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3130 * but allow concurrent faults), and pte mapped but not yet locked.
3131 * We return with mmap_sem still held, but pte unmapped and unlocked.
3133 static int do_anonymous_page(struct vm_fault *vmf)
3135 struct vm_area_struct *vma = vmf->vma;
3136 struct mem_cgroup *memcg;
3141 /* File mapping without ->vm_ops ? */
3142 if (vma->vm_flags & VM_SHARED)
3143 return VM_FAULT_SIGBUS;
3146 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3147 * pte_offset_map() on pmds where a huge pmd might be created
3148 * from a different thread.
3150 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3151 * parallel threads are excluded by other means.
3153 * Here we only have down_read(mmap_sem).
3155 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3156 return VM_FAULT_OOM;
3158 /* See the comment in pte_alloc_one_map() */
3159 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3162 /* Use the zero-page for reads */
3163 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3164 !mm_forbids_zeropage(vma->vm_mm)) {
3165 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3166 vma->vm_page_prot));
3167 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3168 vmf->address, &vmf->ptl);
3169 if (!pte_none(*vmf->pte))
3171 ret = check_stable_address_space(vma->vm_mm);
3174 /* Deliver the page fault to userland, check inside PT lock */
3175 if (userfaultfd_missing(vma)) {
3176 pte_unmap_unlock(vmf->pte, vmf->ptl);
3177 return handle_userfault(vmf, VM_UFFD_MISSING);
3182 /* Allocate our own private page. */
3183 if (unlikely(anon_vma_prepare(vma)))
3185 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3189 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3194 * The memory barrier inside __SetPageUptodate makes sure that
3195 * preceeding stores to the page contents become visible before
3196 * the set_pte_at() write.
3198 __SetPageUptodate(page);
3200 entry = mk_pte(page, vma->vm_page_prot);
3201 if (vma->vm_flags & VM_WRITE)
3202 entry = pte_mkwrite(pte_mkdirty(entry));
3204 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3206 if (!pte_none(*vmf->pte))
3209 ret = check_stable_address_space(vma->vm_mm);
3213 /* Deliver the page fault to userland, check inside PT lock */
3214 if (userfaultfd_missing(vma)) {
3215 pte_unmap_unlock(vmf->pte, vmf->ptl);
3216 mem_cgroup_cancel_charge(page, memcg, false);
3218 return handle_userfault(vmf, VM_UFFD_MISSING);
3221 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3222 page_add_new_anon_rmap(page, vma, vmf->address, false);
3223 mem_cgroup_commit_charge(page, memcg, false, false);
3224 lru_cache_add_active_or_unevictable(page, vma);
3226 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3228 /* No need to invalidate - it was non-present before */
3229 update_mmu_cache(vma, vmf->address, vmf->pte);
3231 pte_unmap_unlock(vmf->pte, vmf->ptl);
3234 mem_cgroup_cancel_charge(page, memcg, false);
3240 return VM_FAULT_OOM;
3244 * The mmap_sem must have been held on entry, and may have been
3245 * released depending on flags and vma->vm_ops->fault() return value.
3246 * See filemap_fault() and __lock_page_retry().
3248 static int __do_fault(struct vm_fault *vmf)
3250 struct vm_area_struct *vma = vmf->vma;
3253 ret = vma->vm_ops->fault(vmf);
3254 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3255 VM_FAULT_DONE_COW)))
3258 if (unlikely(PageHWPoison(vmf->page))) {
3259 if (ret & VM_FAULT_LOCKED)
3260 unlock_page(vmf->page);
3261 put_page(vmf->page);
3263 return VM_FAULT_HWPOISON;
3266 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3267 lock_page(vmf->page);
3269 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3275 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3276 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3277 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3278 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3280 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3282 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3285 static int pte_alloc_one_map(struct vm_fault *vmf)
3287 struct vm_area_struct *vma = vmf->vma;
3289 if (!pmd_none(*vmf->pmd))
3291 if (vmf->prealloc_pte) {
3292 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3293 if (unlikely(!pmd_none(*vmf->pmd))) {
3294 spin_unlock(vmf->ptl);
3298 mm_inc_nr_ptes(vma->vm_mm);
3299 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3300 spin_unlock(vmf->ptl);
3301 vmf->prealloc_pte = NULL;
3302 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3303 return VM_FAULT_OOM;
3307 * If a huge pmd materialized under us just retry later. Use
3308 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3309 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3310 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3311 * running immediately after a huge pmd fault in a different thread of
3312 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3313 * All we have to ensure is that it is a regular pmd that we can walk
3314 * with pte_offset_map() and we can do that through an atomic read in
3315 * C, which is what pmd_trans_unstable() provides.
3317 if (pmd_devmap_trans_unstable(vmf->pmd))
3318 return VM_FAULT_NOPAGE;
3321 * At this point we know that our vmf->pmd points to a page of ptes
3322 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3323 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3324 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3325 * be valid and we will re-check to make sure the vmf->pte isn't
3326 * pte_none() under vmf->ptl protection when we return to
3329 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3334 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3336 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3337 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3338 unsigned long haddr)
3340 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3341 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3343 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3348 static void deposit_prealloc_pte(struct vm_fault *vmf)
3350 struct vm_area_struct *vma = vmf->vma;
3352 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3354 * We are going to consume the prealloc table,
3355 * count that as nr_ptes.
3357 mm_inc_nr_ptes(vma->vm_mm);
3358 vmf->prealloc_pte = NULL;
3361 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3363 struct vm_area_struct *vma = vmf->vma;
3364 bool write = vmf->flags & FAULT_FLAG_WRITE;
3365 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3369 if (!transhuge_vma_suitable(vma, haddr))
3370 return VM_FAULT_FALLBACK;
3372 ret = VM_FAULT_FALLBACK;
3373 page = compound_head(page);
3376 * Archs like ppc64 need additonal space to store information
3377 * related to pte entry. Use the preallocated table for that.
3379 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3380 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3381 if (!vmf->prealloc_pte)
3382 return VM_FAULT_OOM;
3383 smp_wmb(); /* See comment in __pte_alloc() */
3386 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3387 if (unlikely(!pmd_none(*vmf->pmd)))
3390 for (i = 0; i < HPAGE_PMD_NR; i++)
3391 flush_icache_page(vma, page + i);
3393 entry = mk_huge_pmd(page, vma->vm_page_prot);
3395 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3397 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3398 page_add_file_rmap(page, true);
3400 * deposit and withdraw with pmd lock held
3402 if (arch_needs_pgtable_deposit())
3403 deposit_prealloc_pte(vmf);
3405 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3407 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3409 /* fault is handled */
3411 count_vm_event(THP_FILE_MAPPED);
3413 spin_unlock(vmf->ptl);
3417 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3425 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3426 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3428 * @vmf: fault environment
3429 * @memcg: memcg to charge page (only for private mappings)
3430 * @page: page to map
3432 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3435 * Target users are page handler itself and implementations of
3436 * vm_ops->map_pages.
3438 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3441 struct vm_area_struct *vma = vmf->vma;
3442 bool write = vmf->flags & FAULT_FLAG_WRITE;
3446 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3447 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3449 VM_BUG_ON_PAGE(memcg, page);
3451 ret = do_set_pmd(vmf, page);
3452 if (ret != VM_FAULT_FALLBACK)
3457 ret = pte_alloc_one_map(vmf);
3462 /* Re-check under ptl */
3463 if (unlikely(!pte_none(*vmf->pte)))
3464 return VM_FAULT_NOPAGE;
3466 flush_icache_page(vma, page);
3467 entry = mk_pte(page, vma->vm_page_prot);
3469 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3470 /* copy-on-write page */
3471 if (write && !(vma->vm_flags & VM_SHARED)) {
3472 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3473 page_add_new_anon_rmap(page, vma, vmf->address, false);
3474 mem_cgroup_commit_charge(page, memcg, false, false);
3475 lru_cache_add_active_or_unevictable(page, vma);
3477 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3478 page_add_file_rmap(page, false);
3480 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3482 /* no need to invalidate: a not-present page won't be cached */
3483 update_mmu_cache(vma, vmf->address, vmf->pte);
3490 * finish_fault - finish page fault once we have prepared the page to fault
3492 * @vmf: structure describing the fault
3494 * This function handles all that is needed to finish a page fault once the
3495 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3496 * given page, adds reverse page mapping, handles memcg charges and LRU
3497 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3500 * The function expects the page to be locked and on success it consumes a
3501 * reference of a page being mapped (for the PTE which maps it).
3503 int finish_fault(struct vm_fault *vmf)
3508 /* Did we COW the page? */
3509 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3510 !(vmf->vma->vm_flags & VM_SHARED))
3511 page = vmf->cow_page;
3516 * check even for read faults because we might have lost our CoWed
3519 if (!(vmf->vma->vm_flags & VM_SHARED))
3520 ret = check_stable_address_space(vmf->vma->vm_mm);
3522 ret = alloc_set_pte(vmf, vmf->memcg, page);
3524 pte_unmap_unlock(vmf->pte, vmf->ptl);
3528 static unsigned long fault_around_bytes __read_mostly =
3529 rounddown_pow_of_two(65536);
3531 #ifdef CONFIG_DEBUG_FS
3532 static int fault_around_bytes_get(void *data, u64 *val)
3534 *val = fault_around_bytes;
3539 * fault_around_bytes must be rounded down to the nearest page order as it's
3540 * what do_fault_around() expects to see.
3542 static int fault_around_bytes_set(void *data, u64 val)
3544 if (val / PAGE_SIZE > PTRS_PER_PTE)
3546 if (val > PAGE_SIZE)
3547 fault_around_bytes = rounddown_pow_of_two(val);
3549 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3552 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3553 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3555 static int __init fault_around_debugfs(void)
3559 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3560 &fault_around_bytes_fops);
3562 pr_warn("Failed to create fault_around_bytes in debugfs");
3565 late_initcall(fault_around_debugfs);
3569 * do_fault_around() tries to map few pages around the fault address. The hope
3570 * is that the pages will be needed soon and this will lower the number of
3573 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3574 * not ready to be mapped: not up-to-date, locked, etc.
3576 * This function is called with the page table lock taken. In the split ptlock
3577 * case the page table lock only protects only those entries which belong to
3578 * the page table corresponding to the fault address.
3580 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3583 * fault_around_bytes defines how many bytes we'll try to map.
3584 * do_fault_around() expects it to be set to a power of two less than or equal
3587 * The virtual address of the area that we map is naturally aligned to
3588 * fault_around_bytes rounded down to the machine page size
3589 * (and therefore to page order). This way it's easier to guarantee
3590 * that we don't cross page table boundaries.
3592 static int do_fault_around(struct vm_fault *vmf)
3594 unsigned long address = vmf->address, nr_pages, mask;
3595 pgoff_t start_pgoff = vmf->pgoff;
3599 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3600 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3602 vmf->address = max(address & mask, vmf->vma->vm_start);
3603 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3607 * end_pgoff is either the end of the page table, the end of
3608 * the vma or nr_pages from start_pgoff, depending what is nearest.
3610 end_pgoff = start_pgoff -
3611 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3613 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3614 start_pgoff + nr_pages - 1);
3616 if (pmd_none(*vmf->pmd)) {
3617 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3619 if (!vmf->prealloc_pte)
3621 smp_wmb(); /* See comment in __pte_alloc() */
3624 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3626 /* Huge page is mapped? Page fault is solved */
3627 if (pmd_trans_huge(*vmf->pmd)) {
3628 ret = VM_FAULT_NOPAGE;
3632 /* ->map_pages() haven't done anything useful. Cold page cache? */
3636 /* check if the page fault is solved */
3637 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3638 if (!pte_none(*vmf->pte))
3639 ret = VM_FAULT_NOPAGE;
3640 pte_unmap_unlock(vmf->pte, vmf->ptl);
3642 vmf->address = address;
3647 static int do_read_fault(struct vm_fault *vmf)
3649 struct vm_area_struct *vma = vmf->vma;
3653 * Let's call ->map_pages() first and use ->fault() as fallback
3654 * if page by the offset is not ready to be mapped (cold cache or
3657 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3658 ret = do_fault_around(vmf);
3663 ret = __do_fault(vmf);
3664 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3667 ret |= finish_fault(vmf);
3668 unlock_page(vmf->page);
3669 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3670 put_page(vmf->page);
3674 static int do_cow_fault(struct vm_fault *vmf)
3676 struct vm_area_struct *vma = vmf->vma;
3679 if (unlikely(anon_vma_prepare(vma)))
3680 return VM_FAULT_OOM;
3682 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3684 return VM_FAULT_OOM;
3686 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3687 &vmf->memcg, false)) {
3688 put_page(vmf->cow_page);
3689 return VM_FAULT_OOM;
3692 ret = __do_fault(vmf);
3693 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3695 if (ret & VM_FAULT_DONE_COW)
3698 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3699 __SetPageUptodate(vmf->cow_page);
3701 ret |= finish_fault(vmf);
3702 unlock_page(vmf->page);
3703 put_page(vmf->page);
3704 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3708 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3709 put_page(vmf->cow_page);
3713 static int do_shared_fault(struct vm_fault *vmf)
3715 struct vm_area_struct *vma = vmf->vma;
3718 ret = __do_fault(vmf);
3719 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3723 * Check if the backing address space wants to know that the page is
3724 * about to become writable
3726 if (vma->vm_ops->page_mkwrite) {
3727 unlock_page(vmf->page);
3728 tmp = do_page_mkwrite(vmf);
3729 if (unlikely(!tmp ||
3730 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3731 put_page(vmf->page);
3736 ret |= finish_fault(vmf);
3737 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3739 unlock_page(vmf->page);
3740 put_page(vmf->page);
3744 fault_dirty_shared_page(vma, vmf->page);
3749 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3750 * but allow concurrent faults).
3751 * The mmap_sem may have been released depending on flags and our
3752 * return value. See filemap_fault() and __lock_page_or_retry().
3754 static int do_fault(struct vm_fault *vmf)
3756 struct vm_area_struct *vma = vmf->vma;
3759 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3760 if (!vma->vm_ops->fault)
3761 ret = VM_FAULT_SIGBUS;
3762 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3763 ret = do_read_fault(vmf);
3764 else if (!(vma->vm_flags & VM_SHARED))
3765 ret = do_cow_fault(vmf);
3767 ret = do_shared_fault(vmf);
3769 /* preallocated pagetable is unused: free it */
3770 if (vmf->prealloc_pte) {
3771 pte_free(vma->vm_mm, vmf->prealloc_pte);
3772 vmf->prealloc_pte = NULL;
3777 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3778 unsigned long addr, int page_nid,
3783 count_vm_numa_event(NUMA_HINT_FAULTS);
3784 if (page_nid == numa_node_id()) {
3785 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3786 *flags |= TNF_FAULT_LOCAL;
3789 return mpol_misplaced(page, vma, addr);
3792 static int do_numa_page(struct vm_fault *vmf)
3794 struct vm_area_struct *vma = vmf->vma;
3795 struct page *page = NULL;
3799 bool migrated = false;
3801 bool was_writable = pte_savedwrite(vmf->orig_pte);
3805 * The "pte" at this point cannot be used safely without
3806 * validation through pte_unmap_same(). It's of NUMA type but
3807 * the pfn may be screwed if the read is non atomic.
3809 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3810 spin_lock(vmf->ptl);
3811 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3812 pte_unmap_unlock(vmf->pte, vmf->ptl);
3817 * Make it present again, Depending on how arch implementes non
3818 * accessible ptes, some can allow access by kernel mode.
3820 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3821 pte = pte_modify(pte, vma->vm_page_prot);
3822 pte = pte_mkyoung(pte);
3824 pte = pte_mkwrite(pte);
3825 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3826 update_mmu_cache(vma, vmf->address, vmf->pte);
3828 page = vm_normal_page(vma, vmf->address, pte);
3830 pte_unmap_unlock(vmf->pte, vmf->ptl);
3834 /* TODO: handle PTE-mapped THP */
3835 if (PageCompound(page)) {
3836 pte_unmap_unlock(vmf->pte, vmf->ptl);
3841 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3842 * much anyway since they can be in shared cache state. This misses
3843 * the case where a mapping is writable but the process never writes
3844 * to it but pte_write gets cleared during protection updates and
3845 * pte_dirty has unpredictable behaviour between PTE scan updates,
3846 * background writeback, dirty balancing and application behaviour.
3848 if (!pte_write(pte))
3849 flags |= TNF_NO_GROUP;
3852 * Flag if the page is shared between multiple address spaces. This
3853 * is later used when determining whether to group tasks together
3855 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3856 flags |= TNF_SHARED;
3858 last_cpupid = page_cpupid_last(page);
3859 page_nid = page_to_nid(page);
3860 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3862 pte_unmap_unlock(vmf->pte, vmf->ptl);
3863 if (target_nid == -1) {
3868 /* Migrate to the requested node */
3869 migrated = migrate_misplaced_page(page, vma, target_nid);
3871 page_nid = target_nid;
3872 flags |= TNF_MIGRATED;
3874 flags |= TNF_MIGRATE_FAIL;
3878 task_numa_fault(last_cpupid, page_nid, 1, flags);
3882 static inline int create_huge_pmd(struct vm_fault *vmf)
3884 if (vma_is_anonymous(vmf->vma))
3885 return do_huge_pmd_anonymous_page(vmf);
3886 if (vmf->vma->vm_ops->huge_fault)
3887 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3888 return VM_FAULT_FALLBACK;
3891 /* `inline' is required to avoid gcc 4.1.2 build error */
3892 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3894 if (vma_is_anonymous(vmf->vma))
3895 return do_huge_pmd_wp_page(vmf, orig_pmd);
3896 if (vmf->vma->vm_ops->huge_fault)
3897 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3899 /* COW handled on pte level: split pmd */
3900 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3901 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3903 return VM_FAULT_FALLBACK;
3906 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3908 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3911 static int create_huge_pud(struct vm_fault *vmf)
3913 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3914 /* No support for anonymous transparent PUD pages yet */
3915 if (vma_is_anonymous(vmf->vma))
3916 return VM_FAULT_FALLBACK;
3917 if (vmf->vma->vm_ops->huge_fault)
3918 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3919 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3920 return VM_FAULT_FALLBACK;
3923 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3925 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3926 /* No support for anonymous transparent PUD pages yet */
3927 if (vma_is_anonymous(vmf->vma))
3928 return VM_FAULT_FALLBACK;
3929 if (vmf->vma->vm_ops->huge_fault)
3930 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3931 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3932 return VM_FAULT_FALLBACK;
3936 * These routines also need to handle stuff like marking pages dirty
3937 * and/or accessed for architectures that don't do it in hardware (most
3938 * RISC architectures). The early dirtying is also good on the i386.
3940 * There is also a hook called "update_mmu_cache()" that architectures
3941 * with external mmu caches can use to update those (ie the Sparc or
3942 * PowerPC hashed page tables that act as extended TLBs).
3944 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3945 * concurrent faults).
3947 * The mmap_sem may have been released depending on flags and our return value.
3948 * See filemap_fault() and __lock_page_or_retry().
3950 static int handle_pte_fault(struct vm_fault *vmf)
3954 if (unlikely(pmd_none(*vmf->pmd))) {
3956 * Leave __pte_alloc() until later: because vm_ops->fault may
3957 * want to allocate huge page, and if we expose page table
3958 * for an instant, it will be difficult to retract from
3959 * concurrent faults and from rmap lookups.
3963 /* See comment in pte_alloc_one_map() */
3964 if (pmd_devmap_trans_unstable(vmf->pmd))
3967 * A regular pmd is established and it can't morph into a huge
3968 * pmd from under us anymore at this point because we hold the
3969 * mmap_sem read mode and khugepaged takes it in write mode.
3970 * So now it's safe to run pte_offset_map().
3972 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3973 vmf->orig_pte = *vmf->pte;
3976 * some architectures can have larger ptes than wordsize,
3977 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3978 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3979 * accesses. The code below just needs a consistent view
3980 * for the ifs and we later double check anyway with the
3981 * ptl lock held. So here a barrier will do.
3984 if (pte_none(vmf->orig_pte)) {
3985 pte_unmap(vmf->pte);
3991 if (vma_is_anonymous(vmf->vma))
3992 return do_anonymous_page(vmf);
3994 return do_fault(vmf);
3997 if (!pte_present(vmf->orig_pte))
3998 return do_swap_page(vmf);
4000 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4001 return do_numa_page(vmf);
4003 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4004 spin_lock(vmf->ptl);
4005 entry = vmf->orig_pte;
4006 if (unlikely(!pte_same(*vmf->pte, entry)))
4008 if (vmf->flags & FAULT_FLAG_WRITE) {
4009 if (!pte_write(entry))
4010 return do_wp_page(vmf);
4011 entry = pte_mkdirty(entry);
4013 entry = pte_mkyoung(entry);
4014 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4015 vmf->flags & FAULT_FLAG_WRITE)) {
4016 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4019 * This is needed only for protection faults but the arch code
4020 * is not yet telling us if this is a protection fault or not.
4021 * This still avoids useless tlb flushes for .text page faults
4024 if (vmf->flags & FAULT_FLAG_WRITE)
4025 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4028 pte_unmap_unlock(vmf->pte, vmf->ptl);
4033 * By the time we get here, we already hold the mm semaphore
4035 * The mmap_sem may have been released depending on flags and our
4036 * return value. See filemap_fault() and __lock_page_or_retry().
4038 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4041 struct vm_fault vmf = {
4043 .address = address & PAGE_MASK,
4045 .pgoff = linear_page_index(vma, address),
4046 .gfp_mask = __get_fault_gfp_mask(vma),
4048 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4049 struct mm_struct *mm = vma->vm_mm;
4054 pgd = pgd_offset(mm, address);
4055 p4d = p4d_alloc(mm, pgd, address);
4057 return VM_FAULT_OOM;
4059 vmf.pud = pud_alloc(mm, p4d, address);
4061 return VM_FAULT_OOM;
4062 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4063 ret = create_huge_pud(&vmf);
4064 if (!(ret & VM_FAULT_FALLBACK))
4067 pud_t orig_pud = *vmf.pud;
4070 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4072 /* NUMA case for anonymous PUDs would go here */
4074 if (dirty && !pud_write(orig_pud)) {
4075 ret = wp_huge_pud(&vmf, orig_pud);
4076 if (!(ret & VM_FAULT_FALLBACK))
4079 huge_pud_set_accessed(&vmf, orig_pud);
4085 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4087 return VM_FAULT_OOM;
4088 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4089 ret = create_huge_pmd(&vmf);
4090 if (!(ret & VM_FAULT_FALLBACK))
4093 pmd_t orig_pmd = *vmf.pmd;
4096 if (unlikely(is_swap_pmd(orig_pmd))) {
4097 VM_BUG_ON(thp_migration_supported() &&
4098 !is_pmd_migration_entry(orig_pmd));
4099 if (is_pmd_migration_entry(orig_pmd))
4100 pmd_migration_entry_wait(mm, vmf.pmd);
4103 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4104 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4105 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4107 if (dirty && !pmd_write(orig_pmd)) {
4108 ret = wp_huge_pmd(&vmf, orig_pmd);
4109 if (!(ret & VM_FAULT_FALLBACK))
4112 huge_pmd_set_accessed(&vmf, orig_pmd);
4118 return handle_pte_fault(&vmf);
4122 * By the time we get here, we already hold the mm semaphore
4124 * The mmap_sem may have been released depending on flags and our
4125 * return value. See filemap_fault() and __lock_page_or_retry().
4127 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4132 __set_current_state(TASK_RUNNING);
4134 count_vm_event(PGFAULT);
4135 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4137 /* do counter updates before entering really critical section. */
4138 check_sync_rss_stat(current);
4140 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4141 flags & FAULT_FLAG_INSTRUCTION,
4142 flags & FAULT_FLAG_REMOTE))
4143 return VM_FAULT_SIGSEGV;
4146 * Enable the memcg OOM handling for faults triggered in user
4147 * space. Kernel faults are handled more gracefully.
4149 if (flags & FAULT_FLAG_USER)
4150 mem_cgroup_oom_enable();
4152 if (unlikely(is_vm_hugetlb_page(vma)))
4153 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4155 ret = __handle_mm_fault(vma, address, flags);
4157 if (flags & FAULT_FLAG_USER) {
4158 mem_cgroup_oom_disable();
4160 * The task may have entered a memcg OOM situation but
4161 * if the allocation error was handled gracefully (no
4162 * VM_FAULT_OOM), there is no need to kill anything.
4163 * Just clean up the OOM state peacefully.
4165 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4166 mem_cgroup_oom_synchronize(false);
4171 EXPORT_SYMBOL_GPL(handle_mm_fault);
4173 #ifndef __PAGETABLE_P4D_FOLDED
4175 * Allocate p4d page table.
4176 * We've already handled the fast-path in-line.
4178 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4180 p4d_t *new = p4d_alloc_one(mm, address);
4184 smp_wmb(); /* See comment in __pte_alloc */
4186 spin_lock(&mm->page_table_lock);
4187 if (pgd_present(*pgd)) /* Another has populated it */
4190 pgd_populate(mm, pgd, new);
4191 spin_unlock(&mm->page_table_lock);
4194 #endif /* __PAGETABLE_P4D_FOLDED */
4196 #ifndef __PAGETABLE_PUD_FOLDED
4198 * Allocate page upper directory.
4199 * We've already handled the fast-path in-line.
4201 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4203 pud_t *new = pud_alloc_one(mm, address);
4207 smp_wmb(); /* See comment in __pte_alloc */
4209 spin_lock(&mm->page_table_lock);
4210 #ifndef __ARCH_HAS_5LEVEL_HACK
4211 if (!p4d_present(*p4d)) {
4213 p4d_populate(mm, p4d, new);
4214 } else /* Another has populated it */
4217 if (!pgd_present(*p4d)) {
4219 pgd_populate(mm, p4d, new);
4220 } else /* Another has populated it */
4222 #endif /* __ARCH_HAS_5LEVEL_HACK */
4223 spin_unlock(&mm->page_table_lock);
4226 #endif /* __PAGETABLE_PUD_FOLDED */
4228 #ifndef __PAGETABLE_PMD_FOLDED
4230 * Allocate page middle directory.
4231 * We've already handled the fast-path in-line.
4233 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4236 pmd_t *new = pmd_alloc_one(mm, address);
4240 smp_wmb(); /* See comment in __pte_alloc */
4242 ptl = pud_lock(mm, pud);
4243 #ifndef __ARCH_HAS_4LEVEL_HACK
4244 if (!pud_present(*pud)) {
4246 pud_populate(mm, pud, new);
4247 } else /* Another has populated it */
4250 if (!pgd_present(*pud)) {
4252 pgd_populate(mm, pud, new);
4253 } else /* Another has populated it */
4255 #endif /* __ARCH_HAS_4LEVEL_HACK */
4259 #endif /* __PAGETABLE_PMD_FOLDED */
4261 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4262 unsigned long *start, unsigned long *end,
4263 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4271 pgd = pgd_offset(mm, address);
4272 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4275 p4d = p4d_offset(pgd, address);
4276 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4279 pud = pud_offset(p4d, address);
4280 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4283 pmd = pmd_offset(pud, address);
4284 VM_BUG_ON(pmd_trans_huge(*pmd));
4286 if (pmd_huge(*pmd)) {
4291 *start = address & PMD_MASK;
4292 *end = *start + PMD_SIZE;
4293 mmu_notifier_invalidate_range_start(mm, *start, *end);
4295 *ptlp = pmd_lock(mm, pmd);
4296 if (pmd_huge(*pmd)) {
4302 mmu_notifier_invalidate_range_end(mm, *start, *end);
4305 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4309 *start = address & PAGE_MASK;
4310 *end = *start + PAGE_SIZE;
4311 mmu_notifier_invalidate_range_start(mm, *start, *end);
4313 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4314 if (!pte_present(*ptep))
4319 pte_unmap_unlock(ptep, *ptlp);
4321 mmu_notifier_invalidate_range_end(mm, *start, *end);
4326 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4327 pte_t **ptepp, spinlock_t **ptlp)
4331 /* (void) is needed to make gcc happy */
4332 (void) __cond_lock(*ptlp,
4333 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4334 ptepp, NULL, ptlp)));
4338 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4339 unsigned long *start, unsigned long *end,
4340 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4344 /* (void) is needed to make gcc happy */
4345 (void) __cond_lock(*ptlp,
4346 !(res = __follow_pte_pmd(mm, address, start, end,
4347 ptepp, pmdpp, ptlp)));
4350 EXPORT_SYMBOL(follow_pte_pmd);
4353 * follow_pfn - look up PFN at a user virtual address
4354 * @vma: memory mapping
4355 * @address: user virtual address
4356 * @pfn: location to store found PFN
4358 * Only IO mappings and raw PFN mappings are allowed.
4360 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4362 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4369 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4372 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4375 *pfn = pte_pfn(*ptep);
4376 pte_unmap_unlock(ptep, ptl);
4379 EXPORT_SYMBOL(follow_pfn);
4381 #ifdef CONFIG_HAVE_IOREMAP_PROT
4382 int follow_phys(struct vm_area_struct *vma,
4383 unsigned long address, unsigned int flags,
4384 unsigned long *prot, resource_size_t *phys)
4390 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4393 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4397 if ((flags & FOLL_WRITE) && !pte_write(pte))
4400 *prot = pgprot_val(pte_pgprot(pte));
4401 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4405 pte_unmap_unlock(ptep, ptl);
4410 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4411 void *buf, int len, int write)
4413 resource_size_t phys_addr;
4414 unsigned long prot = 0;
4415 void __iomem *maddr;
4416 int offset = addr & (PAGE_SIZE-1);
4418 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4421 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4426 memcpy_toio(maddr + offset, buf, len);
4428 memcpy_fromio(buf, maddr + offset, len);
4433 EXPORT_SYMBOL_GPL(generic_access_phys);
4437 * Access another process' address space as given in mm. If non-NULL, use the
4438 * given task for page fault accounting.
4440 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4441 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4443 struct vm_area_struct *vma;
4444 void *old_buf = buf;
4445 int write = gup_flags & FOLL_WRITE;
4447 down_read(&mm->mmap_sem);
4448 /* ignore errors, just check how much was successfully transferred */
4450 int bytes, ret, offset;
4452 struct page *page = NULL;
4454 ret = get_user_pages_remote(tsk, mm, addr, 1,
4455 gup_flags, &page, &vma, NULL);
4457 #ifndef CONFIG_HAVE_IOREMAP_PROT
4461 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4462 * we can access using slightly different code.
4464 vma = find_vma(mm, addr);
4465 if (!vma || vma->vm_start > addr)
4467 if (vma->vm_ops && vma->vm_ops->access)
4468 ret = vma->vm_ops->access(vma, addr, buf,
4476 offset = addr & (PAGE_SIZE-1);
4477 if (bytes > PAGE_SIZE-offset)
4478 bytes = PAGE_SIZE-offset;
4482 copy_to_user_page(vma, page, addr,
4483 maddr + offset, buf, bytes);
4484 set_page_dirty_lock(page);
4486 copy_from_user_page(vma, page, addr,
4487 buf, maddr + offset, bytes);
4496 up_read(&mm->mmap_sem);
4498 return buf - old_buf;
4502 * access_remote_vm - access another process' address space
4503 * @mm: the mm_struct of the target address space
4504 * @addr: start address to access
4505 * @buf: source or destination buffer
4506 * @len: number of bytes to transfer
4507 * @gup_flags: flags modifying lookup behaviour
4509 * The caller must hold a reference on @mm.
4511 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4512 void *buf, int len, unsigned int gup_flags)
4514 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4518 * Access another process' address space.
4519 * Source/target buffer must be kernel space,
4520 * Do not walk the page table directly, use get_user_pages
4522 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4523 void *buf, int len, unsigned int gup_flags)
4525 struct mm_struct *mm;
4528 mm = get_task_mm(tsk);
4532 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4538 EXPORT_SYMBOL_GPL(access_process_vm);
4541 * Print the name of a VMA.
4543 void print_vma_addr(char *prefix, unsigned long ip)
4545 struct mm_struct *mm = current->mm;
4546 struct vm_area_struct *vma;
4549 * we might be running from an atomic context so we cannot sleep
4551 if (!down_read_trylock(&mm->mmap_sem))
4554 vma = find_vma(mm, ip);
4555 if (vma && vma->vm_file) {
4556 struct file *f = vma->vm_file;
4557 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4561 p = file_path(f, buf, PAGE_SIZE);
4564 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4566 vma->vm_end - vma->vm_start);
4567 free_page((unsigned long)buf);
4570 up_read(&mm->mmap_sem);
4573 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4574 void __might_fault(const char *file, int line)
4577 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4578 * holding the mmap_sem, this is safe because kernel memory doesn't
4579 * get paged out, therefore we'll never actually fault, and the
4580 * below annotations will generate false positives.
4582 if (uaccess_kernel())
4584 if (pagefault_disabled())
4586 __might_sleep(file, line, 0);
4587 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4589 might_lock_read(¤t->mm->mmap_sem);
4592 EXPORT_SYMBOL(__might_fault);
4595 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4596 static void clear_gigantic_page(struct page *page,
4598 unsigned int pages_per_huge_page)
4601 struct page *p = page;
4604 for (i = 0; i < pages_per_huge_page;
4605 i++, p = mem_map_next(p, page, i)) {
4607 clear_user_highpage(p, addr + i * PAGE_SIZE);
4610 void clear_huge_page(struct page *page,
4611 unsigned long addr_hint, unsigned int pages_per_huge_page)
4614 unsigned long addr = addr_hint &
4615 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4617 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4618 clear_gigantic_page(page, addr, pages_per_huge_page);
4622 /* Clear sub-page to access last to keep its cache lines hot */
4624 n = (addr_hint - addr) / PAGE_SIZE;
4625 if (2 * n <= pages_per_huge_page) {
4626 /* If sub-page to access in first half of huge page */
4629 /* Clear sub-pages at the end of huge page */
4630 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4632 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4635 /* If sub-page to access in second half of huge page */
4636 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4637 l = pages_per_huge_page - n;
4638 /* Clear sub-pages at the begin of huge page */
4639 for (i = 0; i < base; i++) {
4641 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4645 * Clear remaining sub-pages in left-right-left-right pattern
4646 * towards the sub-page to access
4648 for (i = 0; i < l; i++) {
4649 int left_idx = base + i;
4650 int right_idx = base + 2 * l - 1 - i;
4653 clear_user_highpage(page + left_idx,
4654 addr + left_idx * PAGE_SIZE);
4656 clear_user_highpage(page + right_idx,
4657 addr + right_idx * PAGE_SIZE);
4661 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4663 struct vm_area_struct *vma,
4664 unsigned int pages_per_huge_page)
4667 struct page *dst_base = dst;
4668 struct page *src_base = src;
4670 for (i = 0; i < pages_per_huge_page; ) {
4672 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4675 dst = mem_map_next(dst, dst_base, i);
4676 src = mem_map_next(src, src_base, i);
4680 void copy_user_huge_page(struct page *dst, struct page *src,
4681 unsigned long addr, struct vm_area_struct *vma,
4682 unsigned int pages_per_huge_page)
4686 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4687 copy_user_gigantic_page(dst, src, addr, vma,
4688 pages_per_huge_page);
4693 for (i = 0; i < pages_per_huge_page; i++) {
4695 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4699 long copy_huge_page_from_user(struct page *dst_page,
4700 const void __user *usr_src,
4701 unsigned int pages_per_huge_page,
4702 bool allow_pagefault)
4704 void *src = (void *)usr_src;
4706 unsigned long i, rc = 0;
4707 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4709 for (i = 0; i < pages_per_huge_page; i++) {
4710 if (allow_pagefault)
4711 page_kaddr = kmap(dst_page + i);
4713 page_kaddr = kmap_atomic(dst_page + i);
4714 rc = copy_from_user(page_kaddr,
4715 (const void __user *)(src + i * PAGE_SIZE),
4717 if (allow_pagefault)
4718 kunmap(dst_page + i);
4720 kunmap_atomic(page_kaddr);
4722 ret_val -= (PAGE_SIZE - rc);
4730 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4732 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4734 static struct kmem_cache *page_ptl_cachep;
4736 void __init ptlock_cache_init(void)
4738 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4742 bool ptlock_alloc(struct page *page)
4746 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4753 void ptlock_free(struct page *page)
4755 kmem_cache_free(page_ptl_cachep, page->ptl);