4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
84 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
94 EXPORT_SYMBOL(mem_map);
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 EXPORT_SYMBOL(high_memory);
108 * Randomize the address space (stacks, mmaps, brk, etc.).
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
120 static int __init disable_randmaps(char *s)
122 randomize_va_space = 0;
125 __setup("norandmaps", disable_randmaps);
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
130 unsigned long highest_memmap_pfn __read_mostly;
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 static int __init init_zero_pfn(void)
137 zero_pfn = page_to_pfn(ZERO_PAGE(0));
140 core_initcall(init_zero_pfn);
143 #if defined(SPLIT_RSS_COUNTING)
145 void sync_mm_rss(struct mm_struct *mm)
149 for (i = 0; i < NR_MM_COUNTERS; i++) {
150 if (current->rss_stat.count[i]) {
151 add_mm_counter(mm, i, current->rss_stat.count[i]);
152 current->rss_stat.count[i] = 0;
155 current->rss_stat.events = 0;
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160 struct task_struct *task = current;
162 if (likely(task->mm == mm))
163 task->rss_stat.count[member] += val;
165 add_mm_counter(mm, member, val);
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH (64)
172 static void check_sync_rss_stat(struct task_struct *task)
174 if (unlikely(task != current))
176 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177 sync_mm_rss(task->mm);
179 #else /* SPLIT_RSS_COUNTING */
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184 static void check_sync_rss_stat(struct task_struct *task)
188 #endif /* SPLIT_RSS_COUNTING */
190 #ifdef HAVE_GENERIC_MMU_GATHER
192 static bool tlb_next_batch(struct mmu_gather *tlb)
194 struct mmu_gather_batch *batch;
198 tlb->active = batch->next;
202 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
205 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
212 batch->max = MAX_GATHER_BATCH;
214 tlb->active->next = batch;
220 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
221 unsigned long start, unsigned long end)
225 /* Is it from 0 to ~0? */
226 tlb->fullmm = !(start | (end+1));
227 tlb->need_flush_all = 0;
228 tlb->local.next = NULL;
230 tlb->local.max = ARRAY_SIZE(tlb->__pages);
231 tlb->active = &tlb->local;
232 tlb->batch_count = 0;
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 __tlb_reset_range(tlb);
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
248 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
249 __tlb_reset_range(tlb);
252 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
254 struct mmu_gather_batch *batch;
256 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
257 tlb_table_flush(tlb);
259 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
263 tlb->active = &tlb->local;
266 void tlb_flush_mmu(struct mmu_gather *tlb)
268 tlb_flush_mmu_tlbonly(tlb);
269 tlb_flush_mmu_free(tlb);
273 * Called at the end of the shootdown operation to free up any resources
274 * that were required.
276 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
277 unsigned long start, unsigned long end, bool force)
279 struct mmu_gather_batch *batch, *next;
282 __tlb_adjust_range(tlb, start, end - start);
286 /* keep the page table cache within bounds */
289 for (batch = tlb->local.next; batch; batch = next) {
291 free_pages((unsigned long)batch, 0);
293 tlb->local.next = NULL;
297 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298 * handling the additional races in SMP caused by other CPUs caching valid
299 * mappings in their TLBs. Returns the number of free page slots left.
300 * When out of page slots we must call tlb_flush_mmu().
301 *returns true if the caller should flush.
303 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
305 struct mmu_gather_batch *batch;
307 VM_BUG_ON(!tlb->end);
308 VM_WARN_ON(tlb->page_size != page_size);
312 * Add the page and check if we are full. If so
315 batch->pages[batch->nr++] = page;
316 if (batch->nr == batch->max) {
317 if (!tlb_next_batch(tlb))
321 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
326 #endif /* HAVE_GENERIC_MMU_GATHER */
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
331 * See the comment near struct mmu_table_batch.
335 * If we want tlb_remove_table() to imply TLB invalidates.
337 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
339 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
341 * Invalidate page-table caches used by hardware walkers. Then we still
342 * need to RCU-sched wait while freeing the pages because software
343 * walkers can still be in-flight.
345 tlb_flush_mmu_tlbonly(tlb);
349 static void tlb_remove_table_smp_sync(void *arg)
351 /* Simply deliver the interrupt */
354 static void tlb_remove_table_one(void *table)
357 * This isn't an RCU grace period and hence the page-tables cannot be
358 * assumed to be actually RCU-freed.
360 * It is however sufficient for software page-table walkers that rely on
361 * IRQ disabling. See the comment near struct mmu_table_batch.
363 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
364 __tlb_remove_table(table);
367 static void tlb_remove_table_rcu(struct rcu_head *head)
369 struct mmu_table_batch *batch;
372 batch = container_of(head, struct mmu_table_batch, rcu);
374 for (i = 0; i < batch->nr; i++)
375 __tlb_remove_table(batch->tables[i]);
377 free_page((unsigned long)batch);
380 void tlb_table_flush(struct mmu_gather *tlb)
382 struct mmu_table_batch **batch = &tlb->batch;
385 tlb_table_invalidate(tlb);
386 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
391 void tlb_remove_table(struct mmu_gather *tlb, void *table)
393 struct mmu_table_batch **batch = &tlb->batch;
395 if (*batch == NULL) {
396 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
397 if (*batch == NULL) {
398 tlb_table_invalidate(tlb);
399 tlb_remove_table_one(table);
405 (*batch)->tables[(*batch)->nr++] = table;
406 if ((*batch)->nr == MAX_TABLE_BATCH)
407 tlb_table_flush(tlb);
410 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
413 * Called to initialize an (on-stack) mmu_gather structure for page-table
414 * tear-down from @mm. The @fullmm argument is used when @mm is without
415 * users and we're going to destroy the full address space (exit/execve).
417 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
418 unsigned long start, unsigned long end)
420 arch_tlb_gather_mmu(tlb, mm, start, end);
421 inc_tlb_flush_pending(tlb->mm);
424 void tlb_finish_mmu(struct mmu_gather *tlb,
425 unsigned long start, unsigned long end)
428 * If there are parallel threads are doing PTE changes on same range
429 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
430 * flush by batching, a thread has stable TLB entry can fail to flush
431 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
432 * forcefully if we detect parallel PTE batching threads.
434 bool force = mm_tlb_flush_nested(tlb->mm);
436 arch_tlb_finish_mmu(tlb, start, end, force);
437 dec_tlb_flush_pending(tlb->mm);
441 * Note: this doesn't free the actual pages themselves. That
442 * has been handled earlier when unmapping all the memory regions.
444 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
447 pgtable_t token = pmd_pgtable(*pmd);
449 pte_free_tlb(tlb, token, addr);
450 atomic_long_dec(&tlb->mm->nr_ptes);
453 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
454 unsigned long addr, unsigned long end,
455 unsigned long floor, unsigned long ceiling)
462 pmd = pmd_offset(pud, addr);
464 next = pmd_addr_end(addr, end);
465 if (pmd_none_or_clear_bad(pmd))
467 free_pte_range(tlb, pmd, addr);
468 } while (pmd++, addr = next, addr != end);
478 if (end - 1 > ceiling - 1)
481 pmd = pmd_offset(pud, start);
483 pmd_free_tlb(tlb, pmd, start);
484 mm_dec_nr_pmds(tlb->mm);
487 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
488 unsigned long addr, unsigned long end,
489 unsigned long floor, unsigned long ceiling)
496 pud = pud_offset(p4d, addr);
498 next = pud_addr_end(addr, end);
499 if (pud_none_or_clear_bad(pud))
501 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
502 } while (pud++, addr = next, addr != end);
512 if (end - 1 > ceiling - 1)
515 pud = pud_offset(p4d, start);
517 pud_free_tlb(tlb, pud, start);
520 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
521 unsigned long addr, unsigned long end,
522 unsigned long floor, unsigned long ceiling)
529 p4d = p4d_offset(pgd, addr);
531 next = p4d_addr_end(addr, end);
532 if (p4d_none_or_clear_bad(p4d))
534 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
535 } while (p4d++, addr = next, addr != end);
541 ceiling &= PGDIR_MASK;
545 if (end - 1 > ceiling - 1)
548 p4d = p4d_offset(pgd, start);
550 p4d_free_tlb(tlb, p4d, start);
554 * This function frees user-level page tables of a process.
556 void free_pgd_range(struct mmu_gather *tlb,
557 unsigned long addr, unsigned long end,
558 unsigned long floor, unsigned long ceiling)
564 * The next few lines have given us lots of grief...
566 * Why are we testing PMD* at this top level? Because often
567 * there will be no work to do at all, and we'd prefer not to
568 * go all the way down to the bottom just to discover that.
570 * Why all these "- 1"s? Because 0 represents both the bottom
571 * of the address space and the top of it (using -1 for the
572 * top wouldn't help much: the masks would do the wrong thing).
573 * The rule is that addr 0 and floor 0 refer to the bottom of
574 * the address space, but end 0 and ceiling 0 refer to the top
575 * Comparisons need to use "end - 1" and "ceiling - 1" (though
576 * that end 0 case should be mythical).
578 * Wherever addr is brought up or ceiling brought down, we must
579 * be careful to reject "the opposite 0" before it confuses the
580 * subsequent tests. But what about where end is brought down
581 * by PMD_SIZE below? no, end can't go down to 0 there.
583 * Whereas we round start (addr) and ceiling down, by different
584 * masks at different levels, in order to test whether a table
585 * now has no other vmas using it, so can be freed, we don't
586 * bother to round floor or end up - the tests don't need that.
600 if (end - 1 > ceiling - 1)
605 * We add page table cache pages with PAGE_SIZE,
606 * (see pte_free_tlb()), flush the tlb if we need
608 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
609 pgd = pgd_offset(tlb->mm, addr);
611 next = pgd_addr_end(addr, end);
612 if (pgd_none_or_clear_bad(pgd))
614 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
615 } while (pgd++, addr = next, addr != end);
618 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
619 unsigned long floor, unsigned long ceiling)
622 struct vm_area_struct *next = vma->vm_next;
623 unsigned long addr = vma->vm_start;
626 * Hide vma from rmap and truncate_pagecache before freeing
629 unlink_anon_vmas(vma);
630 unlink_file_vma(vma);
632 if (is_vm_hugetlb_page(vma)) {
633 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
634 floor, next ? next->vm_start : ceiling);
637 * Optimization: gather nearby vmas into one call down
639 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
640 && !is_vm_hugetlb_page(next)) {
643 unlink_anon_vmas(vma);
644 unlink_file_vma(vma);
646 free_pgd_range(tlb, addr, vma->vm_end,
647 floor, next ? next->vm_start : ceiling);
653 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
656 pgtable_t new = pte_alloc_one(mm, address);
661 * Ensure all pte setup (eg. pte page lock and page clearing) are
662 * visible before the pte is made visible to other CPUs by being
663 * put into page tables.
665 * The other side of the story is the pointer chasing in the page
666 * table walking code (when walking the page table without locking;
667 * ie. most of the time). Fortunately, these data accesses consist
668 * of a chain of data-dependent loads, meaning most CPUs (alpha
669 * being the notable exception) will already guarantee loads are
670 * seen in-order. See the alpha page table accessors for the
671 * smp_read_barrier_depends() barriers in page table walking code.
673 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
675 ptl = pmd_lock(mm, pmd);
676 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
677 atomic_long_inc(&mm->nr_ptes);
678 pmd_populate(mm, pmd, new);
687 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
689 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
693 smp_wmb(); /* See comment in __pte_alloc */
695 spin_lock(&init_mm.page_table_lock);
696 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
697 pmd_populate_kernel(&init_mm, pmd, new);
700 spin_unlock(&init_mm.page_table_lock);
702 pte_free_kernel(&init_mm, new);
706 static inline void init_rss_vec(int *rss)
708 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
711 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
715 if (current->mm == mm)
717 for (i = 0; i < NR_MM_COUNTERS; i++)
719 add_mm_counter(mm, i, rss[i]);
723 * This function is called to print an error when a bad pte
724 * is found. For example, we might have a PFN-mapped pte in
725 * a region that doesn't allow it.
727 * The calling function must still handle the error.
729 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
730 pte_t pte, struct page *page)
732 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
733 p4d_t *p4d = p4d_offset(pgd, addr);
734 pud_t *pud = pud_offset(p4d, addr);
735 pmd_t *pmd = pmd_offset(pud, addr);
736 struct address_space *mapping;
738 static unsigned long resume;
739 static unsigned long nr_shown;
740 static unsigned long nr_unshown;
743 * Allow a burst of 60 reports, then keep quiet for that minute;
744 * or allow a steady drip of one report per second.
746 if (nr_shown == 60) {
747 if (time_before(jiffies, resume)) {
752 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
759 resume = jiffies + 60 * HZ;
761 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
762 index = linear_page_index(vma, addr);
764 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
766 (long long)pte_val(pte), (long long)pmd_val(*pmd));
768 dump_page(page, "bad pte");
769 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
770 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
772 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
774 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
776 vma->vm_ops ? vma->vm_ops->fault : NULL,
777 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
778 mapping ? mapping->a_ops->readpage : NULL);
780 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
784 * vm_normal_page -- This function gets the "struct page" associated with a pte.
786 * "Special" mappings do not wish to be associated with a "struct page" (either
787 * it doesn't exist, or it exists but they don't want to touch it). In this
788 * case, NULL is returned here. "Normal" mappings do have a struct page.
790 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
791 * pte bit, in which case this function is trivial. Secondly, an architecture
792 * may not have a spare pte bit, which requires a more complicated scheme,
795 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
796 * special mapping (even if there are underlying and valid "struct pages").
797 * COWed pages of a VM_PFNMAP are always normal.
799 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
800 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
801 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
802 * mapping will always honor the rule
804 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
806 * And for normal mappings this is false.
808 * This restricts such mappings to be a linear translation from virtual address
809 * to pfn. To get around this restriction, we allow arbitrary mappings so long
810 * as the vma is not a COW mapping; in that case, we know that all ptes are
811 * special (because none can have been COWed).
814 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
816 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
817 * page" backing, however the difference is that _all_ pages with a struct
818 * page (that is, those where pfn_valid is true) are refcounted and considered
819 * normal pages by the VM. The disadvantage is that pages are refcounted
820 * (which can be slower and simply not an option for some PFNMAP users). The
821 * advantage is that we don't have to follow the strict linearity rule of
822 * PFNMAP mappings in order to support COWable mappings.
825 #ifdef __HAVE_ARCH_PTE_SPECIAL
826 # define HAVE_PTE_SPECIAL 1
828 # define HAVE_PTE_SPECIAL 0
830 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
831 pte_t pte, bool with_public_device)
833 unsigned long pfn = pte_pfn(pte);
835 if (HAVE_PTE_SPECIAL) {
836 if (likely(!pte_special(pte)))
838 if (vma->vm_ops && vma->vm_ops->find_special_page)
839 return vma->vm_ops->find_special_page(vma, addr);
840 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
842 if (is_zero_pfn(pfn))
846 * Device public pages are special pages (they are ZONE_DEVICE
847 * pages but different from persistent memory). They behave
848 * allmost like normal pages. The difference is that they are
849 * not on the lru and thus should never be involve with any-
850 * thing that involve lru manipulation (mlock, numa balancing,
853 * This is why we still want to return NULL for such page from
854 * vm_normal_page() so that we do not have to special case all
855 * call site of vm_normal_page().
857 if (likely(pfn <= highest_memmap_pfn)) {
858 struct page *page = pfn_to_page(pfn);
860 if (is_device_public_page(page)) {
861 if (with_public_device)
866 print_bad_pte(vma, addr, pte, NULL);
870 /* !HAVE_PTE_SPECIAL case follows: */
872 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
873 if (vma->vm_flags & VM_MIXEDMAP) {
879 off = (addr - vma->vm_start) >> PAGE_SHIFT;
880 if (pfn == vma->vm_pgoff + off)
882 if (!is_cow_mapping(vma->vm_flags))
887 if (is_zero_pfn(pfn))
890 if (unlikely(pfn > highest_memmap_pfn)) {
891 print_bad_pte(vma, addr, pte, NULL);
896 * NOTE! We still have PageReserved() pages in the page tables.
897 * eg. VDSO mappings can cause them to exist.
900 return pfn_to_page(pfn);
903 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
904 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
907 unsigned long pfn = pmd_pfn(pmd);
910 * There is no pmd_special() but there may be special pmds, e.g.
911 * in a direct-access (dax) mapping, so let's just replicate the
912 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
914 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
915 if (vma->vm_flags & VM_MIXEDMAP) {
921 off = (addr - vma->vm_start) >> PAGE_SHIFT;
922 if (pfn == vma->vm_pgoff + off)
924 if (!is_cow_mapping(vma->vm_flags))
929 if (is_zero_pfn(pfn))
931 if (unlikely(pfn > highest_memmap_pfn))
935 * NOTE! We still have PageReserved() pages in the page tables.
936 * eg. VDSO mappings can cause them to exist.
939 return pfn_to_page(pfn);
944 * copy one vm_area from one task to the other. Assumes the page tables
945 * already present in the new task to be cleared in the whole range
946 * covered by this vma.
949 static inline unsigned long
950 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
951 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
952 unsigned long addr, int *rss)
954 unsigned long vm_flags = vma->vm_flags;
955 pte_t pte = *src_pte;
958 /* pte contains position in swap or file, so copy. */
959 if (unlikely(!pte_present(pte))) {
960 swp_entry_t entry = pte_to_swp_entry(pte);
962 if (likely(!non_swap_entry(entry))) {
963 if (swap_duplicate(entry) < 0)
966 /* make sure dst_mm is on swapoff's mmlist. */
967 if (unlikely(list_empty(&dst_mm->mmlist))) {
968 spin_lock(&mmlist_lock);
969 if (list_empty(&dst_mm->mmlist))
970 list_add(&dst_mm->mmlist,
972 spin_unlock(&mmlist_lock);
975 } else if (is_migration_entry(entry)) {
976 page = migration_entry_to_page(entry);
978 rss[mm_counter(page)]++;
980 if (is_write_migration_entry(entry) &&
981 is_cow_mapping(vm_flags)) {
983 * COW mappings require pages in both
984 * parent and child to be set to read.
986 make_migration_entry_read(&entry);
987 pte = swp_entry_to_pte(entry);
988 if (pte_swp_soft_dirty(*src_pte))
989 pte = pte_swp_mksoft_dirty(pte);
990 set_pte_at(src_mm, addr, src_pte, pte);
992 } else if (is_device_private_entry(entry)) {
993 page = device_private_entry_to_page(entry);
996 * Update rss count even for unaddressable pages, as
997 * they should treated just like normal pages in this
1000 * We will likely want to have some new rss counters
1001 * for unaddressable pages, at some point. But for now
1002 * keep things as they are.
1005 rss[mm_counter(page)]++;
1006 page_dup_rmap(page, false);
1009 * We do not preserve soft-dirty information, because so
1010 * far, checkpoint/restore is the only feature that
1011 * requires that. And checkpoint/restore does not work
1012 * when a device driver is involved (you cannot easily
1013 * save and restore device driver state).
1015 if (is_write_device_private_entry(entry) &&
1016 is_cow_mapping(vm_flags)) {
1017 make_device_private_entry_read(&entry);
1018 pte = swp_entry_to_pte(entry);
1019 set_pte_at(src_mm, addr, src_pte, pte);
1026 * If it's a COW mapping, write protect it both
1027 * in the parent and the child
1029 if (is_cow_mapping(vm_flags)) {
1030 ptep_set_wrprotect(src_mm, addr, src_pte);
1031 pte = pte_wrprotect(pte);
1035 * If it's a shared mapping, mark it clean in
1038 if (vm_flags & VM_SHARED)
1039 pte = pte_mkclean(pte);
1040 pte = pte_mkold(pte);
1042 page = vm_normal_page(vma, addr, pte);
1045 page_dup_rmap(page, false);
1046 rss[mm_counter(page)]++;
1047 } else if (pte_devmap(pte)) {
1048 page = pte_page(pte);
1051 * Cache coherent device memory behave like regular page and
1052 * not like persistent memory page. For more informations see
1053 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1055 if (is_device_public_page(page)) {
1057 page_dup_rmap(page, false);
1058 rss[mm_counter(page)]++;
1063 set_pte_at(dst_mm, addr, dst_pte, pte);
1067 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1068 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1069 unsigned long addr, unsigned long end)
1071 pte_t *orig_src_pte, *orig_dst_pte;
1072 pte_t *src_pte, *dst_pte;
1073 spinlock_t *src_ptl, *dst_ptl;
1075 int rss[NR_MM_COUNTERS];
1076 swp_entry_t entry = (swp_entry_t){0};
1081 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1084 src_pte = pte_offset_map(src_pmd, addr);
1085 src_ptl = pte_lockptr(src_mm, src_pmd);
1086 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1087 orig_src_pte = src_pte;
1088 orig_dst_pte = dst_pte;
1089 arch_enter_lazy_mmu_mode();
1093 * We are holding two locks at this point - either of them
1094 * could generate latencies in another task on another CPU.
1096 if (progress >= 32) {
1098 if (need_resched() ||
1099 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1102 if (pte_none(*src_pte)) {
1106 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1111 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1113 arch_leave_lazy_mmu_mode();
1114 spin_unlock(src_ptl);
1115 pte_unmap(orig_src_pte);
1116 add_mm_rss_vec(dst_mm, rss);
1117 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1121 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1130 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1131 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1132 unsigned long addr, unsigned long end)
1134 pmd_t *src_pmd, *dst_pmd;
1137 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1140 src_pmd = pmd_offset(src_pud, addr);
1142 next = pmd_addr_end(addr, end);
1143 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1144 || pmd_devmap(*src_pmd)) {
1146 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1147 err = copy_huge_pmd(dst_mm, src_mm,
1148 dst_pmd, src_pmd, addr, vma);
1155 if (pmd_none_or_clear_bad(src_pmd))
1157 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1160 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1164 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1165 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1166 unsigned long addr, unsigned long end)
1168 pud_t *src_pud, *dst_pud;
1171 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1174 src_pud = pud_offset(src_p4d, addr);
1176 next = pud_addr_end(addr, end);
1177 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1180 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1181 err = copy_huge_pud(dst_mm, src_mm,
1182 dst_pud, src_pud, addr, vma);
1189 if (pud_none_or_clear_bad(src_pud))
1191 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1194 } while (dst_pud++, src_pud++, addr = next, addr != end);
1198 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1199 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1200 unsigned long addr, unsigned long end)
1202 p4d_t *src_p4d, *dst_p4d;
1205 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1208 src_p4d = p4d_offset(src_pgd, addr);
1210 next = p4d_addr_end(addr, end);
1211 if (p4d_none_or_clear_bad(src_p4d))
1213 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1216 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1220 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1221 struct vm_area_struct *vma)
1223 pgd_t *src_pgd, *dst_pgd;
1225 unsigned long addr = vma->vm_start;
1226 unsigned long end = vma->vm_end;
1227 unsigned long mmun_start; /* For mmu_notifiers */
1228 unsigned long mmun_end; /* For mmu_notifiers */
1233 * Don't copy ptes where a page fault will fill them correctly.
1234 * Fork becomes much lighter when there are big shared or private
1235 * readonly mappings. The tradeoff is that copy_page_range is more
1236 * efficient than faulting.
1238 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1242 if (is_vm_hugetlb_page(vma))
1243 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1245 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1247 * We do not free on error cases below as remove_vma
1248 * gets called on error from higher level routine
1250 ret = track_pfn_copy(vma);
1256 * We need to invalidate the secondary MMU mappings only when
1257 * there could be a permission downgrade on the ptes of the
1258 * parent mm. And a permission downgrade will only happen if
1259 * is_cow_mapping() returns true.
1261 is_cow = is_cow_mapping(vma->vm_flags);
1265 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1269 dst_pgd = pgd_offset(dst_mm, addr);
1270 src_pgd = pgd_offset(src_mm, addr);
1272 next = pgd_addr_end(addr, end);
1273 if (pgd_none_or_clear_bad(src_pgd))
1275 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1276 vma, addr, next))) {
1280 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1283 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1287 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma, pmd_t *pmd,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1292 struct mm_struct *mm = tlb->mm;
1293 int force_flush = 0;
1294 int rss[NR_MM_COUNTERS];
1300 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1303 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1305 flush_tlb_batched_pending(mm);
1306 arch_enter_lazy_mmu_mode();
1309 if (pte_none(ptent))
1312 if (pte_present(ptent)) {
1315 page = _vm_normal_page(vma, addr, ptent, true);
1316 if (unlikely(details) && page) {
1318 * unmap_shared_mapping_pages() wants to
1319 * invalidate cache without truncating:
1320 * unmap shared but keep private pages.
1322 if (details->check_mapping &&
1323 details->check_mapping != page_rmapping(page))
1326 ptent = ptep_get_and_clear_full(mm, addr, pte,
1328 tlb_remove_tlb_entry(tlb, pte, addr);
1329 if (unlikely(!page))
1332 if (!PageAnon(page)) {
1333 if (pte_dirty(ptent)) {
1335 set_page_dirty(page);
1337 if (pte_young(ptent) &&
1338 likely(!(vma->vm_flags & VM_SEQ_READ)))
1339 mark_page_accessed(page);
1341 rss[mm_counter(page)]--;
1342 page_remove_rmap(page, false);
1343 if (unlikely(page_mapcount(page) < 0))
1344 print_bad_pte(vma, addr, ptent, page);
1345 if (unlikely(__tlb_remove_page(tlb, page))) {
1353 entry = pte_to_swp_entry(ptent);
1354 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1355 struct page *page = device_private_entry_to_page(entry);
1357 if (unlikely(details && details->check_mapping)) {
1359 * unmap_shared_mapping_pages() wants to
1360 * invalidate cache without truncating:
1361 * unmap shared but keep private pages.
1363 if (details->check_mapping !=
1364 page_rmapping(page))
1368 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1369 rss[mm_counter(page)]--;
1370 page_remove_rmap(page, false);
1375 /* If details->check_mapping, we leave swap entries. */
1376 if (unlikely(details))
1379 entry = pte_to_swp_entry(ptent);
1380 if (!non_swap_entry(entry))
1382 else if (is_migration_entry(entry)) {
1385 page = migration_entry_to_page(entry);
1386 rss[mm_counter(page)]--;
1388 if (unlikely(!free_swap_and_cache(entry)))
1389 print_bad_pte(vma, addr, ptent, NULL);
1390 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1391 } while (pte++, addr += PAGE_SIZE, addr != end);
1393 add_mm_rss_vec(mm, rss);
1394 arch_leave_lazy_mmu_mode();
1396 /* Do the actual TLB flush before dropping ptl */
1398 tlb_flush_mmu_tlbonly(tlb);
1399 pte_unmap_unlock(start_pte, ptl);
1402 * If we forced a TLB flush (either due to running out of
1403 * batch buffers or because we needed to flush dirty TLB
1404 * entries before releasing the ptl), free the batched
1405 * memory too. Restart if we didn't do everything.
1409 tlb_flush_mmu_free(tlb);
1417 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1418 struct vm_area_struct *vma, pud_t *pud,
1419 unsigned long addr, unsigned long end,
1420 struct zap_details *details)
1425 pmd = pmd_offset(pud, addr);
1427 next = pmd_addr_end(addr, end);
1428 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1429 if (next - addr != HPAGE_PMD_SIZE)
1430 __split_huge_pmd(vma, pmd, addr, false, NULL);
1431 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1436 * Here there can be other concurrent MADV_DONTNEED or
1437 * trans huge page faults running, and if the pmd is
1438 * none or trans huge it can change under us. This is
1439 * because MADV_DONTNEED holds the mmap_sem in read
1442 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1444 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1447 } while (pmd++, addr = next, addr != end);
1452 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1453 struct vm_area_struct *vma, p4d_t *p4d,
1454 unsigned long addr, unsigned long end,
1455 struct zap_details *details)
1460 pud = pud_offset(p4d, addr);
1462 next = pud_addr_end(addr, end);
1463 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1464 if (next - addr != HPAGE_PUD_SIZE) {
1465 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1466 split_huge_pud(vma, pud, addr);
1467 } else if (zap_huge_pud(tlb, vma, pud, addr))
1471 if (pud_none_or_clear_bad(pud))
1473 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1476 } while (pud++, addr = next, addr != end);
1481 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1482 struct vm_area_struct *vma, pgd_t *pgd,
1483 unsigned long addr, unsigned long end,
1484 struct zap_details *details)
1489 p4d = p4d_offset(pgd, addr);
1491 next = p4d_addr_end(addr, end);
1492 if (p4d_none_or_clear_bad(p4d))
1494 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1495 } while (p4d++, addr = next, addr != end);
1500 void unmap_page_range(struct mmu_gather *tlb,
1501 struct vm_area_struct *vma,
1502 unsigned long addr, unsigned long end,
1503 struct zap_details *details)
1508 BUG_ON(addr >= end);
1509 tlb_start_vma(tlb, vma);
1510 pgd = pgd_offset(vma->vm_mm, addr);
1512 next = pgd_addr_end(addr, end);
1513 if (pgd_none_or_clear_bad(pgd))
1515 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1516 } while (pgd++, addr = next, addr != end);
1517 tlb_end_vma(tlb, vma);
1521 static void unmap_single_vma(struct mmu_gather *tlb,
1522 struct vm_area_struct *vma, unsigned long start_addr,
1523 unsigned long end_addr,
1524 struct zap_details *details)
1526 unsigned long start = max(vma->vm_start, start_addr);
1529 if (start >= vma->vm_end)
1531 end = min(vma->vm_end, end_addr);
1532 if (end <= vma->vm_start)
1536 uprobe_munmap(vma, start, end);
1538 if (unlikely(vma->vm_flags & VM_PFNMAP))
1539 untrack_pfn(vma, 0, 0);
1542 if (unlikely(is_vm_hugetlb_page(vma))) {
1544 * It is undesirable to test vma->vm_file as it
1545 * should be non-null for valid hugetlb area.
1546 * However, vm_file will be NULL in the error
1547 * cleanup path of mmap_region. When
1548 * hugetlbfs ->mmap method fails,
1549 * mmap_region() nullifies vma->vm_file
1550 * before calling this function to clean up.
1551 * Since no pte has actually been setup, it is
1552 * safe to do nothing in this case.
1555 i_mmap_lock_write(vma->vm_file->f_mapping);
1556 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1557 i_mmap_unlock_write(vma->vm_file->f_mapping);
1560 unmap_page_range(tlb, vma, start, end, details);
1565 * unmap_vmas - unmap a range of memory covered by a list of vma's
1566 * @tlb: address of the caller's struct mmu_gather
1567 * @vma: the starting vma
1568 * @start_addr: virtual address at which to start unmapping
1569 * @end_addr: virtual address at which to end unmapping
1571 * Unmap all pages in the vma list.
1573 * Only addresses between `start' and `end' will be unmapped.
1575 * The VMA list must be sorted in ascending virtual address order.
1577 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1578 * range after unmap_vmas() returns. So the only responsibility here is to
1579 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1580 * drops the lock and schedules.
1582 void unmap_vmas(struct mmu_gather *tlb,
1583 struct vm_area_struct *vma, unsigned long start_addr,
1584 unsigned long end_addr)
1586 struct mm_struct *mm = vma->vm_mm;
1588 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1589 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1590 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1591 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1595 * zap_page_range - remove user pages in a given range
1596 * @vma: vm_area_struct holding the applicable pages
1597 * @start: starting address of pages to zap
1598 * @size: number of bytes to zap
1600 * Caller must protect the VMA list
1602 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1605 struct mm_struct *mm = vma->vm_mm;
1606 struct mmu_gather tlb;
1607 unsigned long end = start + size;
1610 tlb_gather_mmu(&tlb, mm, start, end);
1611 update_hiwater_rss(mm);
1612 mmu_notifier_invalidate_range_start(mm, start, end);
1613 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1614 unmap_single_vma(&tlb, vma, start, end, NULL);
1617 * zap_page_range does not specify whether mmap_sem should be
1618 * held for read or write. That allows parallel zap_page_range
1619 * operations to unmap a PTE and defer a flush meaning that
1620 * this call observes pte_none and fails to flush the TLB.
1621 * Rather than adding a complex API, ensure that no stale
1622 * TLB entries exist when this call returns.
1624 flush_tlb_range(vma, start, end);
1627 mmu_notifier_invalidate_range_end(mm, start, end);
1628 tlb_finish_mmu(&tlb, start, end);
1632 * zap_page_range_single - remove user pages in a given range
1633 * @vma: vm_area_struct holding the applicable pages
1634 * @address: starting address of pages to zap
1635 * @size: number of bytes to zap
1636 * @details: details of shared cache invalidation
1638 * The range must fit into one VMA.
1640 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1641 unsigned long size, struct zap_details *details)
1643 struct mm_struct *mm = vma->vm_mm;
1644 struct mmu_gather tlb;
1645 unsigned long end = address + size;
1648 tlb_gather_mmu(&tlb, mm, address, end);
1649 update_hiwater_rss(mm);
1650 mmu_notifier_invalidate_range_start(mm, address, end);
1651 unmap_single_vma(&tlb, vma, address, end, details);
1652 mmu_notifier_invalidate_range_end(mm, address, end);
1653 tlb_finish_mmu(&tlb, address, end);
1657 * zap_vma_ptes - remove ptes mapping the vma
1658 * @vma: vm_area_struct holding ptes to be zapped
1659 * @address: starting address of pages to zap
1660 * @size: number of bytes to zap
1662 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1664 * The entire address range must be fully contained within the vma.
1666 * Returns 0 if successful.
1668 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1671 if (address < vma->vm_start || address + size > vma->vm_end ||
1672 !(vma->vm_flags & VM_PFNMAP))
1674 zap_page_range_single(vma, address, size, NULL);
1677 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1679 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1687 pgd = pgd_offset(mm, addr);
1688 p4d = p4d_alloc(mm, pgd, addr);
1691 pud = pud_alloc(mm, p4d, addr);
1694 pmd = pmd_alloc(mm, pud, addr);
1698 VM_BUG_ON(pmd_trans_huge(*pmd));
1699 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1703 * This is the old fallback for page remapping.
1705 * For historical reasons, it only allows reserved pages. Only
1706 * old drivers should use this, and they needed to mark their
1707 * pages reserved for the old functions anyway.
1709 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1710 struct page *page, pgprot_t prot)
1712 struct mm_struct *mm = vma->vm_mm;
1721 flush_dcache_page(page);
1722 pte = get_locked_pte(mm, addr, &ptl);
1726 if (!pte_none(*pte))
1729 /* Ok, finally just insert the thing.. */
1731 inc_mm_counter_fast(mm, mm_counter_file(page));
1732 page_add_file_rmap(page, false);
1733 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1736 pte_unmap_unlock(pte, ptl);
1739 pte_unmap_unlock(pte, ptl);
1745 * vm_insert_page - insert single page into user vma
1746 * @vma: user vma to map to
1747 * @addr: target user address of this page
1748 * @page: source kernel page
1750 * This allows drivers to insert individual pages they've allocated
1753 * The page has to be a nice clean _individual_ kernel allocation.
1754 * If you allocate a compound page, you need to have marked it as
1755 * such (__GFP_COMP), or manually just split the page up yourself
1756 * (see split_page()).
1758 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1759 * took an arbitrary page protection parameter. This doesn't allow
1760 * that. Your vma protection will have to be set up correctly, which
1761 * means that if you want a shared writable mapping, you'd better
1762 * ask for a shared writable mapping!
1764 * The page does not need to be reserved.
1766 * Usually this function is called from f_op->mmap() handler
1767 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1768 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1769 * function from other places, for example from page-fault handler.
1771 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1774 if (addr < vma->vm_start || addr >= vma->vm_end)
1776 if (!page_count(page))
1778 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1779 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1780 BUG_ON(vma->vm_flags & VM_PFNMAP);
1781 vma->vm_flags |= VM_MIXEDMAP;
1783 return insert_page(vma, addr, page, vma->vm_page_prot);
1785 EXPORT_SYMBOL(vm_insert_page);
1787 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1788 pfn_t pfn, pgprot_t prot, bool mkwrite)
1790 struct mm_struct *mm = vma->vm_mm;
1796 pte = get_locked_pte(mm, addr, &ptl);
1800 if (!pte_none(*pte)) {
1803 * For read faults on private mappings the PFN passed
1804 * in may not match the PFN we have mapped if the
1805 * mapped PFN is a writeable COW page. In the mkwrite
1806 * case we are creating a writable PTE for a shared
1807 * mapping and we expect the PFNs to match.
1809 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1817 /* Ok, finally just insert the thing.. */
1818 if (pfn_t_devmap(pfn))
1819 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1821 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1825 entry = pte_mkyoung(entry);
1826 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1829 set_pte_at(mm, addr, pte, entry);
1830 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1834 pte_unmap_unlock(pte, ptl);
1840 * vm_insert_pfn - insert single pfn into user vma
1841 * @vma: user vma to map to
1842 * @addr: target user address of this page
1843 * @pfn: source kernel pfn
1845 * Similar to vm_insert_page, this allows drivers to insert individual pages
1846 * they've allocated into a user vma. Same comments apply.
1848 * This function should only be called from a vm_ops->fault handler, and
1849 * in that case the handler should return NULL.
1851 * vma cannot be a COW mapping.
1853 * As this is called only for pages that do not currently exist, we
1854 * do not need to flush old virtual caches or the TLB.
1856 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1859 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1861 EXPORT_SYMBOL(vm_insert_pfn);
1864 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1865 * @vma: user vma to map to
1866 * @addr: target user address of this page
1867 * @pfn: source kernel pfn
1868 * @pgprot: pgprot flags for the inserted page
1870 * This is exactly like vm_insert_pfn, except that it allows drivers to
1871 * to override pgprot on a per-page basis.
1873 * This only makes sense for IO mappings, and it makes no sense for
1874 * cow mappings. In general, using multiple vmas is preferable;
1875 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1878 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1879 unsigned long pfn, pgprot_t pgprot)
1883 * Technically, architectures with pte_special can avoid all these
1884 * restrictions (same for remap_pfn_range). However we would like
1885 * consistency in testing and feature parity among all, so we should
1886 * try to keep these invariants in place for everybody.
1888 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1889 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1890 (VM_PFNMAP|VM_MIXEDMAP));
1891 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1892 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1894 if (addr < vma->vm_start || addr >= vma->vm_end)
1897 if (!pfn_modify_allowed(pfn, pgprot))
1900 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1902 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1907 EXPORT_SYMBOL(vm_insert_pfn_prot);
1909 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1910 pfn_t pfn, bool mkwrite)
1912 pgprot_t pgprot = vma->vm_page_prot;
1914 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1916 if (addr < vma->vm_start || addr >= vma->vm_end)
1919 track_pfn_insert(vma, &pgprot, pfn);
1921 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1925 * If we don't have pte special, then we have to use the pfn_valid()
1926 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1927 * refcount the page if pfn_valid is true (hence insert_page rather
1928 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1929 * without pte special, it would there be refcounted as a normal page.
1931 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1935 * At this point we are committed to insert_page()
1936 * regardless of whether the caller specified flags that
1937 * result in pfn_t_has_page() == false.
1939 page = pfn_to_page(pfn_t_to_pfn(pfn));
1940 return insert_page(vma, addr, page, pgprot);
1942 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1945 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1948 return __vm_insert_mixed(vma, addr, pfn, false);
1951 EXPORT_SYMBOL(vm_insert_mixed);
1953 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1956 return __vm_insert_mixed(vma, addr, pfn, true);
1958 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1961 * maps a range of physical memory into the requested pages. the old
1962 * mappings are removed. any references to nonexistent pages results
1963 * in null mappings (currently treated as "copy-on-access")
1965 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1966 unsigned long addr, unsigned long end,
1967 unsigned long pfn, pgprot_t prot)
1973 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1976 arch_enter_lazy_mmu_mode();
1978 BUG_ON(!pte_none(*pte));
1979 if (!pfn_modify_allowed(pfn, prot)) {
1983 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1985 } while (pte++, addr += PAGE_SIZE, addr != end);
1986 arch_leave_lazy_mmu_mode();
1987 pte_unmap_unlock(pte - 1, ptl);
1991 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1992 unsigned long addr, unsigned long end,
1993 unsigned long pfn, pgprot_t prot)
1999 pfn -= addr >> PAGE_SHIFT;
2000 pmd = pmd_alloc(mm, pud, addr);
2003 VM_BUG_ON(pmd_trans_huge(*pmd));
2005 next = pmd_addr_end(addr, end);
2006 err = remap_pte_range(mm, pmd, addr, next,
2007 pfn + (addr >> PAGE_SHIFT), prot);
2010 } while (pmd++, addr = next, addr != end);
2014 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2015 unsigned long addr, unsigned long end,
2016 unsigned long pfn, pgprot_t prot)
2022 pfn -= addr >> PAGE_SHIFT;
2023 pud = pud_alloc(mm, p4d, addr);
2027 next = pud_addr_end(addr, end);
2028 err = remap_pmd_range(mm, pud, addr, next,
2029 pfn + (addr >> PAGE_SHIFT), prot);
2032 } while (pud++, addr = next, addr != end);
2036 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2037 unsigned long addr, unsigned long end,
2038 unsigned long pfn, pgprot_t prot)
2044 pfn -= addr >> PAGE_SHIFT;
2045 p4d = p4d_alloc(mm, pgd, addr);
2049 next = p4d_addr_end(addr, end);
2050 err = remap_pud_range(mm, p4d, addr, next,
2051 pfn + (addr >> PAGE_SHIFT), prot);
2054 } while (p4d++, addr = next, addr != end);
2059 * remap_pfn_range - remap kernel memory to userspace
2060 * @vma: user vma to map to
2061 * @addr: target user address to start at
2062 * @pfn: physical address of kernel memory
2063 * @size: size of map area
2064 * @prot: page protection flags for this mapping
2066 * Note: this is only safe if the mm semaphore is held when called.
2068 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2069 unsigned long pfn, unsigned long size, pgprot_t prot)
2073 unsigned long end = addr + PAGE_ALIGN(size);
2074 struct mm_struct *mm = vma->vm_mm;
2075 unsigned long remap_pfn = pfn;
2079 * Physically remapped pages are special. Tell the
2080 * rest of the world about it:
2081 * VM_IO tells people not to look at these pages
2082 * (accesses can have side effects).
2083 * VM_PFNMAP tells the core MM that the base pages are just
2084 * raw PFN mappings, and do not have a "struct page" associated
2087 * Disable vma merging and expanding with mremap().
2089 * Omit vma from core dump, even when VM_IO turned off.
2091 * There's a horrible special case to handle copy-on-write
2092 * behaviour that some programs depend on. We mark the "original"
2093 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2094 * See vm_normal_page() for details.
2096 if (is_cow_mapping(vma->vm_flags)) {
2097 if (addr != vma->vm_start || end != vma->vm_end)
2099 vma->vm_pgoff = pfn;
2102 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2106 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2108 BUG_ON(addr >= end);
2109 pfn -= addr >> PAGE_SHIFT;
2110 pgd = pgd_offset(mm, addr);
2111 flush_cache_range(vma, addr, end);
2113 next = pgd_addr_end(addr, end);
2114 err = remap_p4d_range(mm, pgd, addr, next,
2115 pfn + (addr >> PAGE_SHIFT), prot);
2118 } while (pgd++, addr = next, addr != end);
2121 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2125 EXPORT_SYMBOL(remap_pfn_range);
2128 * vm_iomap_memory - remap memory to userspace
2129 * @vma: user vma to map to
2130 * @start: start of area
2131 * @len: size of area
2133 * This is a simplified io_remap_pfn_range() for common driver use. The
2134 * driver just needs to give us the physical memory range to be mapped,
2135 * we'll figure out the rest from the vma information.
2137 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2138 * whatever write-combining details or similar.
2140 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2142 unsigned long vm_len, pfn, pages;
2144 /* Check that the physical memory area passed in looks valid */
2145 if (start + len < start)
2148 * You *really* shouldn't map things that aren't page-aligned,
2149 * but we've historically allowed it because IO memory might
2150 * just have smaller alignment.
2152 len += start & ~PAGE_MASK;
2153 pfn = start >> PAGE_SHIFT;
2154 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2155 if (pfn + pages < pfn)
2158 /* We start the mapping 'vm_pgoff' pages into the area */
2159 if (vma->vm_pgoff > pages)
2161 pfn += vma->vm_pgoff;
2162 pages -= vma->vm_pgoff;
2164 /* Can we fit all of the mapping? */
2165 vm_len = vma->vm_end - vma->vm_start;
2166 if (vm_len >> PAGE_SHIFT > pages)
2169 /* Ok, let it rip */
2170 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2172 EXPORT_SYMBOL(vm_iomap_memory);
2174 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2175 unsigned long addr, unsigned long end,
2176 pte_fn_t fn, void *data)
2181 spinlock_t *uninitialized_var(ptl);
2183 pte = (mm == &init_mm) ?
2184 pte_alloc_kernel(pmd, addr) :
2185 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2189 BUG_ON(pmd_huge(*pmd));
2191 arch_enter_lazy_mmu_mode();
2193 token = pmd_pgtable(*pmd);
2196 err = fn(pte++, token, addr, data);
2199 } while (addr += PAGE_SIZE, addr != end);
2201 arch_leave_lazy_mmu_mode();
2204 pte_unmap_unlock(pte-1, ptl);
2208 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2209 unsigned long addr, unsigned long end,
2210 pte_fn_t fn, void *data)
2216 BUG_ON(pud_huge(*pud));
2218 pmd = pmd_alloc(mm, pud, addr);
2222 next = pmd_addr_end(addr, end);
2223 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2226 } while (pmd++, addr = next, addr != end);
2230 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2231 unsigned long addr, unsigned long end,
2232 pte_fn_t fn, void *data)
2238 pud = pud_alloc(mm, p4d, addr);
2242 next = pud_addr_end(addr, end);
2243 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2246 } while (pud++, addr = next, addr != end);
2250 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2251 unsigned long addr, unsigned long end,
2252 pte_fn_t fn, void *data)
2258 p4d = p4d_alloc(mm, pgd, addr);
2262 next = p4d_addr_end(addr, end);
2263 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2266 } while (p4d++, addr = next, addr != end);
2271 * Scan a region of virtual memory, filling in page tables as necessary
2272 * and calling a provided function on each leaf page table.
2274 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2275 unsigned long size, pte_fn_t fn, void *data)
2279 unsigned long end = addr + size;
2282 if (WARN_ON(addr >= end))
2285 pgd = pgd_offset(mm, addr);
2287 next = pgd_addr_end(addr, end);
2288 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2291 } while (pgd++, addr = next, addr != end);
2295 EXPORT_SYMBOL_GPL(apply_to_page_range);
2298 * handle_pte_fault chooses page fault handler according to an entry which was
2299 * read non-atomically. Before making any commitment, on those architectures
2300 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2301 * parts, do_swap_page must check under lock before unmapping the pte and
2302 * proceeding (but do_wp_page is only called after already making such a check;
2303 * and do_anonymous_page can safely check later on).
2305 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2306 pte_t *page_table, pte_t orig_pte)
2309 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2310 if (sizeof(pte_t) > sizeof(unsigned long)) {
2311 spinlock_t *ptl = pte_lockptr(mm, pmd);
2313 same = pte_same(*page_table, orig_pte);
2317 pte_unmap(page_table);
2321 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2323 debug_dma_assert_idle(src);
2326 * If the source page was a PFN mapping, we don't have
2327 * a "struct page" for it. We do a best-effort copy by
2328 * just copying from the original user address. If that
2329 * fails, we just zero-fill it. Live with it.
2331 if (unlikely(!src)) {
2332 void *kaddr = kmap_atomic(dst);
2333 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2336 * This really shouldn't fail, because the page is there
2337 * in the page tables. But it might just be unreadable,
2338 * in which case we just give up and fill the result with
2341 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2343 kunmap_atomic(kaddr);
2344 flush_dcache_page(dst);
2346 copy_user_highpage(dst, src, va, vma);
2349 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2351 struct file *vm_file = vma->vm_file;
2354 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2357 * Special mappings (e.g. VDSO) do not have any file so fake
2358 * a default GFP_KERNEL for them.
2364 * Notify the address space that the page is about to become writable so that
2365 * it can prohibit this or wait for the page to get into an appropriate state.
2367 * We do this without the lock held, so that it can sleep if it needs to.
2369 static int do_page_mkwrite(struct vm_fault *vmf)
2372 struct page *page = vmf->page;
2373 unsigned int old_flags = vmf->flags;
2375 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2377 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2378 /* Restore original flags so that caller is not surprised */
2379 vmf->flags = old_flags;
2380 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2382 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2384 if (!page->mapping) {
2386 return 0; /* retry */
2388 ret |= VM_FAULT_LOCKED;
2390 VM_BUG_ON_PAGE(!PageLocked(page), page);
2395 * Handle dirtying of a page in shared file mapping on a write fault.
2397 * The function expects the page to be locked and unlocks it.
2399 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2402 struct address_space *mapping;
2404 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2406 dirtied = set_page_dirty(page);
2407 VM_BUG_ON_PAGE(PageAnon(page), page);
2409 * Take a local copy of the address_space - page.mapping may be zeroed
2410 * by truncate after unlock_page(). The address_space itself remains
2411 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2412 * release semantics to prevent the compiler from undoing this copying.
2414 mapping = page_rmapping(page);
2417 if ((dirtied || page_mkwrite) && mapping) {
2419 * Some device drivers do not set page.mapping
2420 * but still dirty their pages
2422 balance_dirty_pages_ratelimited(mapping);
2426 file_update_time(vma->vm_file);
2430 * Handle write page faults for pages that can be reused in the current vma
2432 * This can happen either due to the mapping being with the VM_SHARED flag,
2433 * or due to us being the last reference standing to the page. In either
2434 * case, all we need to do here is to mark the page as writable and update
2435 * any related book-keeping.
2437 static inline void wp_page_reuse(struct vm_fault *vmf)
2438 __releases(vmf->ptl)
2440 struct vm_area_struct *vma = vmf->vma;
2441 struct page *page = vmf->page;
2444 * Clear the pages cpupid information as the existing
2445 * information potentially belongs to a now completely
2446 * unrelated process.
2449 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2451 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2452 entry = pte_mkyoung(vmf->orig_pte);
2453 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2454 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2455 update_mmu_cache(vma, vmf->address, vmf->pte);
2456 pte_unmap_unlock(vmf->pte, vmf->ptl);
2460 * Handle the case of a page which we actually need to copy to a new page.
2462 * Called with mmap_sem locked and the old page referenced, but
2463 * without the ptl held.
2465 * High level logic flow:
2467 * - Allocate a page, copy the content of the old page to the new one.
2468 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2469 * - Take the PTL. If the pte changed, bail out and release the allocated page
2470 * - If the pte is still the way we remember it, update the page table and all
2471 * relevant references. This includes dropping the reference the page-table
2472 * held to the old page, as well as updating the rmap.
2473 * - In any case, unlock the PTL and drop the reference we took to the old page.
2475 static int wp_page_copy(struct vm_fault *vmf)
2477 struct vm_area_struct *vma = vmf->vma;
2478 struct mm_struct *mm = vma->vm_mm;
2479 struct page *old_page = vmf->page;
2480 struct page *new_page = NULL;
2482 int page_copied = 0;
2483 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2484 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2485 struct mem_cgroup *memcg;
2487 if (unlikely(anon_vma_prepare(vma)))
2490 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2491 new_page = alloc_zeroed_user_highpage_movable(vma,
2496 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2500 cow_user_page(new_page, old_page, vmf->address, vma);
2503 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2506 __SetPageUptodate(new_page);
2508 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2511 * Re-check the pte - we dropped the lock
2513 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2514 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2516 if (!PageAnon(old_page)) {
2517 dec_mm_counter_fast(mm,
2518 mm_counter_file(old_page));
2519 inc_mm_counter_fast(mm, MM_ANONPAGES);
2522 inc_mm_counter_fast(mm, MM_ANONPAGES);
2524 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2525 entry = mk_pte(new_page, vma->vm_page_prot);
2526 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2528 * Clear the pte entry and flush it first, before updating the
2529 * pte with the new entry. This will avoid a race condition
2530 * seen in the presence of one thread doing SMC and another
2533 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2534 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2535 mem_cgroup_commit_charge(new_page, memcg, false, false);
2536 lru_cache_add_active_or_unevictable(new_page, vma);
2538 * We call the notify macro here because, when using secondary
2539 * mmu page tables (such as kvm shadow page tables), we want the
2540 * new page to be mapped directly into the secondary page table.
2542 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2543 update_mmu_cache(vma, vmf->address, vmf->pte);
2546 * Only after switching the pte to the new page may
2547 * we remove the mapcount here. Otherwise another
2548 * process may come and find the rmap count decremented
2549 * before the pte is switched to the new page, and
2550 * "reuse" the old page writing into it while our pte
2551 * here still points into it and can be read by other
2554 * The critical issue is to order this
2555 * page_remove_rmap with the ptp_clear_flush above.
2556 * Those stores are ordered by (if nothing else,)
2557 * the barrier present in the atomic_add_negative
2558 * in page_remove_rmap.
2560 * Then the TLB flush in ptep_clear_flush ensures that
2561 * no process can access the old page before the
2562 * decremented mapcount is visible. And the old page
2563 * cannot be reused until after the decremented
2564 * mapcount is visible. So transitively, TLBs to
2565 * old page will be flushed before it can be reused.
2567 page_remove_rmap(old_page, false);
2570 /* Free the old page.. */
2571 new_page = old_page;
2574 mem_cgroup_cancel_charge(new_page, memcg, false);
2580 pte_unmap_unlock(vmf->pte, vmf->ptl);
2581 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2584 * Don't let another task, with possibly unlocked vma,
2585 * keep the mlocked page.
2587 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2588 lock_page(old_page); /* LRU manipulation */
2589 if (PageMlocked(old_page))
2590 munlock_vma_page(old_page);
2591 unlock_page(old_page);
2595 return page_copied ? VM_FAULT_WRITE : 0;
2601 return VM_FAULT_OOM;
2605 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2606 * writeable once the page is prepared
2608 * @vmf: structure describing the fault
2610 * This function handles all that is needed to finish a write page fault in a
2611 * shared mapping due to PTE being read-only once the mapped page is prepared.
2612 * It handles locking of PTE and modifying it. The function returns
2613 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2616 * The function expects the page to be locked or other protection against
2617 * concurrent faults / writeback (such as DAX radix tree locks).
2619 int finish_mkwrite_fault(struct vm_fault *vmf)
2621 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2622 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2625 * We might have raced with another page fault while we released the
2626 * pte_offset_map_lock.
2628 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2629 pte_unmap_unlock(vmf->pte, vmf->ptl);
2630 return VM_FAULT_NOPAGE;
2637 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2640 static int wp_pfn_shared(struct vm_fault *vmf)
2642 struct vm_area_struct *vma = vmf->vma;
2644 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2647 pte_unmap_unlock(vmf->pte, vmf->ptl);
2648 vmf->flags |= FAULT_FLAG_MKWRITE;
2649 ret = vma->vm_ops->pfn_mkwrite(vmf);
2650 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2652 return finish_mkwrite_fault(vmf);
2655 return VM_FAULT_WRITE;
2658 static int wp_page_shared(struct vm_fault *vmf)
2659 __releases(vmf->ptl)
2661 struct vm_area_struct *vma = vmf->vma;
2663 get_page(vmf->page);
2665 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2668 pte_unmap_unlock(vmf->pte, vmf->ptl);
2669 tmp = do_page_mkwrite(vmf);
2670 if (unlikely(!tmp || (tmp &
2671 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2672 put_page(vmf->page);
2675 tmp = finish_mkwrite_fault(vmf);
2676 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2677 unlock_page(vmf->page);
2678 put_page(vmf->page);
2683 lock_page(vmf->page);
2685 fault_dirty_shared_page(vma, vmf->page);
2686 put_page(vmf->page);
2688 return VM_FAULT_WRITE;
2692 * This routine handles present pages, when users try to write
2693 * to a shared page. It is done by copying the page to a new address
2694 * and decrementing the shared-page counter for the old page.
2696 * Note that this routine assumes that the protection checks have been
2697 * done by the caller (the low-level page fault routine in most cases).
2698 * Thus we can safely just mark it writable once we've done any necessary
2701 * We also mark the page dirty at this point even though the page will
2702 * change only once the write actually happens. This avoids a few races,
2703 * and potentially makes it more efficient.
2705 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2706 * but allow concurrent faults), with pte both mapped and locked.
2707 * We return with mmap_sem still held, but pte unmapped and unlocked.
2709 static int do_wp_page(struct vm_fault *vmf)
2710 __releases(vmf->ptl)
2712 struct vm_area_struct *vma = vmf->vma;
2714 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2717 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2720 * We should not cow pages in a shared writeable mapping.
2721 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2723 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2724 (VM_WRITE|VM_SHARED))
2725 return wp_pfn_shared(vmf);
2727 pte_unmap_unlock(vmf->pte, vmf->ptl);
2728 return wp_page_copy(vmf);
2732 * Take out anonymous pages first, anonymous shared vmas are
2733 * not dirty accountable.
2735 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2736 int total_map_swapcount;
2737 if (!trylock_page(vmf->page)) {
2738 get_page(vmf->page);
2739 pte_unmap_unlock(vmf->pte, vmf->ptl);
2740 lock_page(vmf->page);
2741 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2742 vmf->address, &vmf->ptl);
2743 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2744 unlock_page(vmf->page);
2745 pte_unmap_unlock(vmf->pte, vmf->ptl);
2746 put_page(vmf->page);
2749 put_page(vmf->page);
2751 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2752 if (total_map_swapcount == 1) {
2754 * The page is all ours. Move it to
2755 * our anon_vma so the rmap code will
2756 * not search our parent or siblings.
2757 * Protected against the rmap code by
2760 page_move_anon_rmap(vmf->page, vma);
2762 unlock_page(vmf->page);
2764 return VM_FAULT_WRITE;
2766 unlock_page(vmf->page);
2767 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2768 (VM_WRITE|VM_SHARED))) {
2769 return wp_page_shared(vmf);
2773 * Ok, we need to copy. Oh, well..
2775 get_page(vmf->page);
2777 pte_unmap_unlock(vmf->pte, vmf->ptl);
2778 return wp_page_copy(vmf);
2781 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2782 unsigned long start_addr, unsigned long end_addr,
2783 struct zap_details *details)
2785 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2788 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2789 struct zap_details *details)
2791 struct vm_area_struct *vma;
2792 pgoff_t vba, vea, zba, zea;
2794 vma_interval_tree_foreach(vma, root,
2795 details->first_index, details->last_index) {
2797 vba = vma->vm_pgoff;
2798 vea = vba + vma_pages(vma) - 1;
2799 zba = details->first_index;
2802 zea = details->last_index;
2806 unmap_mapping_range_vma(vma,
2807 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2808 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2814 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2815 * address_space corresponding to the specified page range in the underlying
2818 * @mapping: the address space containing mmaps to be unmapped.
2819 * @holebegin: byte in first page to unmap, relative to the start of
2820 * the underlying file. This will be rounded down to a PAGE_SIZE
2821 * boundary. Note that this is different from truncate_pagecache(), which
2822 * must keep the partial page. In contrast, we must get rid of
2824 * @holelen: size of prospective hole in bytes. This will be rounded
2825 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2827 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2828 * but 0 when invalidating pagecache, don't throw away private data.
2830 void unmap_mapping_range(struct address_space *mapping,
2831 loff_t const holebegin, loff_t const holelen, int even_cows)
2833 struct zap_details details = { };
2834 pgoff_t hba = holebegin >> PAGE_SHIFT;
2835 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2837 /* Check for overflow. */
2838 if (sizeof(holelen) > sizeof(hlen)) {
2840 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2841 if (holeend & ~(long long)ULONG_MAX)
2842 hlen = ULONG_MAX - hba + 1;
2845 details.check_mapping = even_cows ? NULL : mapping;
2846 details.first_index = hba;
2847 details.last_index = hba + hlen - 1;
2848 if (details.last_index < details.first_index)
2849 details.last_index = ULONG_MAX;
2851 i_mmap_lock_write(mapping);
2852 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2853 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2854 i_mmap_unlock_write(mapping);
2856 EXPORT_SYMBOL(unmap_mapping_range);
2859 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2860 * but allow concurrent faults), and pte mapped but not yet locked.
2861 * We return with pte unmapped and unlocked.
2863 * We return with the mmap_sem locked or unlocked in the same cases
2864 * as does filemap_fault().
2866 int do_swap_page(struct vm_fault *vmf)
2868 struct vm_area_struct *vma = vmf->vma;
2869 struct page *page = NULL, *swapcache;
2870 struct mem_cgroup *memcg;
2871 struct vma_swap_readahead swap_ra;
2877 bool vma_readahead = swap_use_vma_readahead();
2880 page = swap_readahead_detect(vmf, &swap_ra);
2881 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2887 entry = pte_to_swp_entry(vmf->orig_pte);
2888 if (unlikely(non_swap_entry(entry))) {
2889 if (is_migration_entry(entry)) {
2890 migration_entry_wait(vma->vm_mm, vmf->pmd,
2892 } else if (is_device_private_entry(entry)) {
2894 * For un-addressable device memory we call the pgmap
2895 * fault handler callback. The callback must migrate
2896 * the page back to some CPU accessible page.
2898 ret = device_private_entry_fault(vma, vmf->address, entry,
2899 vmf->flags, vmf->pmd);
2900 } else if (is_hwpoison_entry(entry)) {
2901 ret = VM_FAULT_HWPOISON;
2903 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2904 ret = VM_FAULT_SIGBUS;
2908 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2910 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2914 page = do_swap_page_readahead(entry,
2915 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2917 page = swapin_readahead(entry,
2918 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2921 * Back out if somebody else faulted in this pte
2922 * while we released the pte lock.
2924 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2925 vmf->address, &vmf->ptl);
2926 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2928 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2932 /* Had to read the page from swap area: Major fault */
2933 ret = VM_FAULT_MAJOR;
2934 count_vm_event(PGMAJFAULT);
2935 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2936 } else if (PageHWPoison(page)) {
2938 * hwpoisoned dirty swapcache pages are kept for killing
2939 * owner processes (which may be unknown at hwpoison time)
2941 ret = VM_FAULT_HWPOISON;
2942 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2948 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2950 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2952 ret |= VM_FAULT_RETRY;
2957 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2958 * release the swapcache from under us. The page pin, and pte_same
2959 * test below, are not enough to exclude that. Even if it is still
2960 * swapcache, we need to check that the page's swap has not changed.
2962 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2965 page = ksm_might_need_to_copy(page, vma, vmf->address);
2966 if (unlikely(!page)) {
2972 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2979 * Back out if somebody else already faulted in this pte.
2981 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2983 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2986 if (unlikely(!PageUptodate(page))) {
2987 ret = VM_FAULT_SIGBUS;
2992 * The page isn't present yet, go ahead with the fault.
2994 * Be careful about the sequence of operations here.
2995 * To get its accounting right, reuse_swap_page() must be called
2996 * while the page is counted on swap but not yet in mapcount i.e.
2997 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2998 * must be called after the swap_free(), or it will never succeed.
3001 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3002 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3003 pte = mk_pte(page, vma->vm_page_prot);
3004 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3005 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3006 vmf->flags &= ~FAULT_FLAG_WRITE;
3007 ret |= VM_FAULT_WRITE;
3008 exclusive = RMAP_EXCLUSIVE;
3010 flush_icache_page(vma, page);
3011 if (pte_swp_soft_dirty(vmf->orig_pte))
3012 pte = pte_mksoft_dirty(pte);
3013 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3014 vmf->orig_pte = pte;
3015 if (page == swapcache) {
3016 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3017 mem_cgroup_commit_charge(page, memcg, true, false);
3018 activate_page(page);
3019 } else { /* ksm created a completely new copy */
3020 page_add_new_anon_rmap(page, vma, vmf->address, false);
3021 mem_cgroup_commit_charge(page, memcg, false, false);
3022 lru_cache_add_active_or_unevictable(page, vma);
3026 if (mem_cgroup_swap_full(page) ||
3027 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3028 try_to_free_swap(page);
3030 if (page != swapcache) {
3032 * Hold the lock to avoid the swap entry to be reused
3033 * until we take the PT lock for the pte_same() check
3034 * (to avoid false positives from pte_same). For
3035 * further safety release the lock after the swap_free
3036 * so that the swap count won't change under a
3037 * parallel locked swapcache.
3039 unlock_page(swapcache);
3040 put_page(swapcache);
3043 if (vmf->flags & FAULT_FLAG_WRITE) {
3044 ret |= do_wp_page(vmf);
3045 if (ret & VM_FAULT_ERROR)
3046 ret &= VM_FAULT_ERROR;
3050 /* No need to invalidate - it was non-present before */
3051 update_mmu_cache(vma, vmf->address, vmf->pte);
3053 pte_unmap_unlock(vmf->pte, vmf->ptl);
3057 mem_cgroup_cancel_charge(page, memcg, false);
3058 pte_unmap_unlock(vmf->pte, vmf->ptl);
3063 if (page != swapcache) {
3064 unlock_page(swapcache);
3065 put_page(swapcache);
3071 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3072 * but allow concurrent faults), and pte mapped but not yet locked.
3073 * We return with mmap_sem still held, but pte unmapped and unlocked.
3075 static int do_anonymous_page(struct vm_fault *vmf)
3077 struct vm_area_struct *vma = vmf->vma;
3078 struct mem_cgroup *memcg;
3083 /* File mapping without ->vm_ops ? */
3084 if (vma->vm_flags & VM_SHARED)
3085 return VM_FAULT_SIGBUS;
3088 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3089 * pte_offset_map() on pmds where a huge pmd might be created
3090 * from a different thread.
3092 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3093 * parallel threads are excluded by other means.
3095 * Here we only have down_read(mmap_sem).
3097 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3098 return VM_FAULT_OOM;
3100 /* See the comment in pte_alloc_one_map() */
3101 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3104 /* Use the zero-page for reads */
3105 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3106 !mm_forbids_zeropage(vma->vm_mm)) {
3107 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3108 vma->vm_page_prot));
3109 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3110 vmf->address, &vmf->ptl);
3111 if (!pte_none(*vmf->pte))
3113 ret = check_stable_address_space(vma->vm_mm);
3116 /* Deliver the page fault to userland, check inside PT lock */
3117 if (userfaultfd_missing(vma)) {
3118 pte_unmap_unlock(vmf->pte, vmf->ptl);
3119 return handle_userfault(vmf, VM_UFFD_MISSING);
3124 /* Allocate our own private page. */
3125 if (unlikely(anon_vma_prepare(vma)))
3127 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3131 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3135 * The memory barrier inside __SetPageUptodate makes sure that
3136 * preceeding stores to the page contents become visible before
3137 * the set_pte_at() write.
3139 __SetPageUptodate(page);
3141 entry = mk_pte(page, vma->vm_page_prot);
3142 if (vma->vm_flags & VM_WRITE)
3143 entry = pte_mkwrite(pte_mkdirty(entry));
3145 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3147 if (!pte_none(*vmf->pte))
3150 ret = check_stable_address_space(vma->vm_mm);
3154 /* Deliver the page fault to userland, check inside PT lock */
3155 if (userfaultfd_missing(vma)) {
3156 pte_unmap_unlock(vmf->pte, vmf->ptl);
3157 mem_cgroup_cancel_charge(page, memcg, false);
3159 return handle_userfault(vmf, VM_UFFD_MISSING);
3162 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3163 page_add_new_anon_rmap(page, vma, vmf->address, false);
3164 mem_cgroup_commit_charge(page, memcg, false, false);
3165 lru_cache_add_active_or_unevictable(page, vma);
3167 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3169 /* No need to invalidate - it was non-present before */
3170 update_mmu_cache(vma, vmf->address, vmf->pte);
3172 pte_unmap_unlock(vmf->pte, vmf->ptl);
3175 mem_cgroup_cancel_charge(page, memcg, false);
3181 return VM_FAULT_OOM;
3185 * The mmap_sem must have been held on entry, and may have been
3186 * released depending on flags and vma->vm_ops->fault() return value.
3187 * See filemap_fault() and __lock_page_retry().
3189 static int __do_fault(struct vm_fault *vmf)
3191 struct vm_area_struct *vma = vmf->vma;
3195 * Preallocate pte before we take page_lock because this might lead to
3196 * deadlocks for memcg reclaim which waits for pages under writeback:
3198 * SetPageWriteback(A)
3204 * wait_on_page_writeback(A)
3205 * SetPageWriteback(B)
3207 * # flush A, B to clear the writeback
3209 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3210 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3212 if (!vmf->prealloc_pte)
3213 return VM_FAULT_OOM;
3214 smp_wmb(); /* See comment in __pte_alloc() */
3217 ret = vma->vm_ops->fault(vmf);
3218 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3219 VM_FAULT_DONE_COW)))
3222 if (unlikely(PageHWPoison(vmf->page))) {
3223 if (ret & VM_FAULT_LOCKED)
3224 unlock_page(vmf->page);
3225 put_page(vmf->page);
3227 return VM_FAULT_HWPOISON;
3230 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3231 lock_page(vmf->page);
3233 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3239 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3240 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3241 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3242 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3244 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3246 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3249 static int pte_alloc_one_map(struct vm_fault *vmf)
3251 struct vm_area_struct *vma = vmf->vma;
3253 if (!pmd_none(*vmf->pmd))
3255 if (vmf->prealloc_pte) {
3256 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3257 if (unlikely(!pmd_none(*vmf->pmd))) {
3258 spin_unlock(vmf->ptl);
3262 atomic_long_inc(&vma->vm_mm->nr_ptes);
3263 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3264 spin_unlock(vmf->ptl);
3265 vmf->prealloc_pte = NULL;
3266 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3267 return VM_FAULT_OOM;
3271 * If a huge pmd materialized under us just retry later. Use
3272 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3273 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3274 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3275 * running immediately after a huge pmd fault in a different thread of
3276 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3277 * All we have to ensure is that it is a regular pmd that we can walk
3278 * with pte_offset_map() and we can do that through an atomic read in
3279 * C, which is what pmd_trans_unstable() provides.
3281 if (pmd_devmap_trans_unstable(vmf->pmd))
3282 return VM_FAULT_NOPAGE;
3285 * At this point we know that our vmf->pmd points to a page of ptes
3286 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3287 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3288 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3289 * be valid and we will re-check to make sure the vmf->pte isn't
3290 * pte_none() under vmf->ptl protection when we return to
3293 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3298 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3300 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3301 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3302 unsigned long haddr)
3304 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3305 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3307 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3312 static void deposit_prealloc_pte(struct vm_fault *vmf)
3314 struct vm_area_struct *vma = vmf->vma;
3316 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3318 * We are going to consume the prealloc table,
3319 * count that as nr_ptes.
3321 atomic_long_inc(&vma->vm_mm->nr_ptes);
3322 vmf->prealloc_pte = NULL;
3325 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3327 struct vm_area_struct *vma = vmf->vma;
3328 bool write = vmf->flags & FAULT_FLAG_WRITE;
3329 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3333 if (!transhuge_vma_suitable(vma, haddr))
3334 return VM_FAULT_FALLBACK;
3336 ret = VM_FAULT_FALLBACK;
3337 page = compound_head(page);
3340 * Archs like ppc64 need additonal space to store information
3341 * related to pte entry. Use the preallocated table for that.
3343 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3344 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3345 if (!vmf->prealloc_pte)
3346 return VM_FAULT_OOM;
3347 smp_wmb(); /* See comment in __pte_alloc() */
3350 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3351 if (unlikely(!pmd_none(*vmf->pmd)))
3354 for (i = 0; i < HPAGE_PMD_NR; i++)
3355 flush_icache_page(vma, page + i);
3357 entry = mk_huge_pmd(page, vma->vm_page_prot);
3359 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3361 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3362 page_add_file_rmap(page, true);
3364 * deposit and withdraw with pmd lock held
3366 if (arch_needs_pgtable_deposit())
3367 deposit_prealloc_pte(vmf);
3369 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3371 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3373 /* fault is handled */
3375 count_vm_event(THP_FILE_MAPPED);
3377 spin_unlock(vmf->ptl);
3381 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3389 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3390 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3392 * @vmf: fault environment
3393 * @memcg: memcg to charge page (only for private mappings)
3394 * @page: page to map
3396 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3399 * Target users are page handler itself and implementations of
3400 * vm_ops->map_pages.
3402 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3405 struct vm_area_struct *vma = vmf->vma;
3406 bool write = vmf->flags & FAULT_FLAG_WRITE;
3410 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3411 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3413 VM_BUG_ON_PAGE(memcg, page);
3415 ret = do_set_pmd(vmf, page);
3416 if (ret != VM_FAULT_FALLBACK)
3421 ret = pte_alloc_one_map(vmf);
3426 /* Re-check under ptl */
3427 if (unlikely(!pte_none(*vmf->pte)))
3428 return VM_FAULT_NOPAGE;
3430 flush_icache_page(vma, page);
3431 entry = mk_pte(page, vma->vm_page_prot);
3433 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3434 /* copy-on-write page */
3435 if (write && !(vma->vm_flags & VM_SHARED)) {
3436 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3437 page_add_new_anon_rmap(page, vma, vmf->address, false);
3438 mem_cgroup_commit_charge(page, memcg, false, false);
3439 lru_cache_add_active_or_unevictable(page, vma);
3441 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3442 page_add_file_rmap(page, false);
3444 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3446 /* no need to invalidate: a not-present page won't be cached */
3447 update_mmu_cache(vma, vmf->address, vmf->pte);
3454 * finish_fault - finish page fault once we have prepared the page to fault
3456 * @vmf: structure describing the fault
3458 * This function handles all that is needed to finish a page fault once the
3459 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3460 * given page, adds reverse page mapping, handles memcg charges and LRU
3461 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3464 * The function expects the page to be locked and on success it consumes a
3465 * reference of a page being mapped (for the PTE which maps it).
3467 int finish_fault(struct vm_fault *vmf)
3472 /* Did we COW the page? */
3473 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3474 !(vmf->vma->vm_flags & VM_SHARED))
3475 page = vmf->cow_page;
3480 * check even for read faults because we might have lost our CoWed
3483 if (!(vmf->vma->vm_flags & VM_SHARED))
3484 ret = check_stable_address_space(vmf->vma->vm_mm);
3486 ret = alloc_set_pte(vmf, vmf->memcg, page);
3488 pte_unmap_unlock(vmf->pte, vmf->ptl);
3492 static unsigned long fault_around_bytes __read_mostly =
3493 rounddown_pow_of_two(65536);
3495 #ifdef CONFIG_DEBUG_FS
3496 static int fault_around_bytes_get(void *data, u64 *val)
3498 *val = fault_around_bytes;
3503 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3504 * rounded down to nearest page order. It's what do_fault_around() expects to
3507 static int fault_around_bytes_set(void *data, u64 val)
3509 if (val / PAGE_SIZE > PTRS_PER_PTE)
3511 if (val > PAGE_SIZE)
3512 fault_around_bytes = rounddown_pow_of_two(val);
3514 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3517 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3518 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3520 static int __init fault_around_debugfs(void)
3524 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3525 &fault_around_bytes_fops);
3527 pr_warn("Failed to create fault_around_bytes in debugfs");
3530 late_initcall(fault_around_debugfs);
3534 * do_fault_around() tries to map few pages around the fault address. The hope
3535 * is that the pages will be needed soon and this will lower the number of
3538 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3539 * not ready to be mapped: not up-to-date, locked, etc.
3541 * This function is called with the page table lock taken. In the split ptlock
3542 * case the page table lock only protects only those entries which belong to
3543 * the page table corresponding to the fault address.
3545 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3548 * fault_around_pages() defines how many pages we'll try to map.
3549 * do_fault_around() expects it to return a power of two less than or equal to
3552 * The virtual address of the area that we map is naturally aligned to the
3553 * fault_around_pages() value (and therefore to page order). This way it's
3554 * easier to guarantee that we don't cross page table boundaries.
3556 static int do_fault_around(struct vm_fault *vmf)
3558 unsigned long address = vmf->address, nr_pages, mask;
3559 pgoff_t start_pgoff = vmf->pgoff;
3563 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3564 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3566 vmf->address = max(address & mask, vmf->vma->vm_start);
3567 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3571 * end_pgoff is either end of page table or end of vma
3572 * or fault_around_pages() from start_pgoff, depending what is nearest.
3574 end_pgoff = start_pgoff -
3575 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3577 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3578 start_pgoff + nr_pages - 1);
3580 if (pmd_none(*vmf->pmd)) {
3581 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3583 if (!vmf->prealloc_pte)
3585 smp_wmb(); /* See comment in __pte_alloc() */
3588 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3590 /* Huge page is mapped? Page fault is solved */
3591 if (pmd_trans_huge(*vmf->pmd)) {
3592 ret = VM_FAULT_NOPAGE;
3596 /* ->map_pages() haven't done anything useful. Cold page cache? */
3600 /* check if the page fault is solved */
3601 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3602 if (!pte_none(*vmf->pte))
3603 ret = VM_FAULT_NOPAGE;
3604 pte_unmap_unlock(vmf->pte, vmf->ptl);
3606 vmf->address = address;
3611 static int do_read_fault(struct vm_fault *vmf)
3613 struct vm_area_struct *vma = vmf->vma;
3617 * Let's call ->map_pages() first and use ->fault() as fallback
3618 * if page by the offset is not ready to be mapped (cold cache or
3621 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3622 ret = do_fault_around(vmf);
3627 ret = __do_fault(vmf);
3628 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3631 ret |= finish_fault(vmf);
3632 unlock_page(vmf->page);
3633 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3634 put_page(vmf->page);
3638 static int do_cow_fault(struct vm_fault *vmf)
3640 struct vm_area_struct *vma = vmf->vma;
3643 if (unlikely(anon_vma_prepare(vma)))
3644 return VM_FAULT_OOM;
3646 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3648 return VM_FAULT_OOM;
3650 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3651 &vmf->memcg, false)) {
3652 put_page(vmf->cow_page);
3653 return VM_FAULT_OOM;
3656 ret = __do_fault(vmf);
3657 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3659 if (ret & VM_FAULT_DONE_COW)
3662 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3663 __SetPageUptodate(vmf->cow_page);
3665 ret |= finish_fault(vmf);
3666 unlock_page(vmf->page);
3667 put_page(vmf->page);
3668 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3672 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3673 put_page(vmf->cow_page);
3677 static int do_shared_fault(struct vm_fault *vmf)
3679 struct vm_area_struct *vma = vmf->vma;
3682 ret = __do_fault(vmf);
3683 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3687 * Check if the backing address space wants to know that the page is
3688 * about to become writable
3690 if (vma->vm_ops->page_mkwrite) {
3691 unlock_page(vmf->page);
3692 tmp = do_page_mkwrite(vmf);
3693 if (unlikely(!tmp ||
3694 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3695 put_page(vmf->page);
3700 ret |= finish_fault(vmf);
3701 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3703 unlock_page(vmf->page);
3704 put_page(vmf->page);
3708 fault_dirty_shared_page(vma, vmf->page);
3713 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3714 * but allow concurrent faults).
3715 * The mmap_sem may have been released depending on flags and our
3716 * return value. See filemap_fault() and __lock_page_or_retry().
3718 static int do_fault(struct vm_fault *vmf)
3720 struct vm_area_struct *vma = vmf->vma;
3724 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3726 if (!vma->vm_ops->fault) {
3728 * If we find a migration pmd entry or a none pmd entry, which
3729 * should never happen, return SIGBUS
3731 if (unlikely(!pmd_present(*vmf->pmd)))
3732 ret = VM_FAULT_SIGBUS;
3734 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3739 * Make sure this is not a temporary clearing of pte
3740 * by holding ptl and checking again. A R/M/W update
3741 * of pte involves: take ptl, clearing the pte so that
3742 * we don't have concurrent modification by hardware
3743 * followed by an update.
3745 if (unlikely(pte_none(*vmf->pte)))
3746 ret = VM_FAULT_SIGBUS;
3748 ret = VM_FAULT_NOPAGE;
3750 pte_unmap_unlock(vmf->pte, vmf->ptl);
3752 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3753 ret = do_read_fault(vmf);
3754 else if (!(vma->vm_flags & VM_SHARED))
3755 ret = do_cow_fault(vmf);
3757 ret = do_shared_fault(vmf);
3759 /* preallocated pagetable is unused: free it */
3760 if (vmf->prealloc_pte) {
3761 pte_free(vma->vm_mm, vmf->prealloc_pte);
3762 vmf->prealloc_pte = NULL;
3767 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3768 unsigned long addr, int page_nid,
3773 count_vm_numa_event(NUMA_HINT_FAULTS);
3774 if (page_nid == numa_node_id()) {
3775 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3776 *flags |= TNF_FAULT_LOCAL;
3779 return mpol_misplaced(page, vma, addr);
3782 static int do_numa_page(struct vm_fault *vmf)
3784 struct vm_area_struct *vma = vmf->vma;
3785 struct page *page = NULL;
3789 bool migrated = false;
3791 bool was_writable = pte_savedwrite(vmf->orig_pte);
3795 * The "pte" at this point cannot be used safely without
3796 * validation through pte_unmap_same(). It's of NUMA type but
3797 * the pfn may be screwed if the read is non atomic.
3799 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3800 spin_lock(vmf->ptl);
3801 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3802 pte_unmap_unlock(vmf->pte, vmf->ptl);
3807 * Make it present again, Depending on how arch implementes non
3808 * accessible ptes, some can allow access by kernel mode.
3810 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3811 pte = pte_modify(pte, vma->vm_page_prot);
3812 pte = pte_mkyoung(pte);
3814 pte = pte_mkwrite(pte);
3815 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3816 update_mmu_cache(vma, vmf->address, vmf->pte);
3818 page = vm_normal_page(vma, vmf->address, pte);
3820 pte_unmap_unlock(vmf->pte, vmf->ptl);
3824 /* TODO: handle PTE-mapped THP */
3825 if (PageCompound(page)) {
3826 pte_unmap_unlock(vmf->pte, vmf->ptl);
3831 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3832 * much anyway since they can be in shared cache state. This misses
3833 * the case where a mapping is writable but the process never writes
3834 * to it but pte_write gets cleared during protection updates and
3835 * pte_dirty has unpredictable behaviour between PTE scan updates,
3836 * background writeback, dirty balancing and application behaviour.
3838 if (!pte_write(pte))
3839 flags |= TNF_NO_GROUP;
3842 * Flag if the page is shared between multiple address spaces. This
3843 * is later used when determining whether to group tasks together
3845 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3846 flags |= TNF_SHARED;
3848 last_cpupid = page_cpupid_last(page);
3849 page_nid = page_to_nid(page);
3850 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3852 pte_unmap_unlock(vmf->pte, vmf->ptl);
3853 if (target_nid == -1) {
3858 /* Migrate to the requested node */
3859 migrated = migrate_misplaced_page(page, vma, target_nid);
3861 page_nid = target_nid;
3862 flags |= TNF_MIGRATED;
3864 flags |= TNF_MIGRATE_FAIL;
3868 task_numa_fault(last_cpupid, page_nid, 1, flags);
3872 static inline int create_huge_pmd(struct vm_fault *vmf)
3874 if (vma_is_anonymous(vmf->vma))
3875 return do_huge_pmd_anonymous_page(vmf);
3876 if (vmf->vma->vm_ops->huge_fault)
3877 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3878 return VM_FAULT_FALLBACK;
3881 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3883 if (vma_is_anonymous(vmf->vma))
3884 return do_huge_pmd_wp_page(vmf, orig_pmd);
3885 if (vmf->vma->vm_ops->huge_fault)
3886 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3888 /* COW handled on pte level: split pmd */
3889 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3890 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3892 return VM_FAULT_FALLBACK;
3895 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3897 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3900 static int create_huge_pud(struct vm_fault *vmf)
3902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3903 /* No support for anonymous transparent PUD pages yet */
3904 if (vma_is_anonymous(vmf->vma))
3905 return VM_FAULT_FALLBACK;
3906 if (vmf->vma->vm_ops->huge_fault)
3907 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3908 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3909 return VM_FAULT_FALLBACK;
3912 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3914 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3915 /* No support for anonymous transparent PUD pages yet */
3916 if (vma_is_anonymous(vmf->vma))
3917 return VM_FAULT_FALLBACK;
3918 if (vmf->vma->vm_ops->huge_fault)
3919 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3920 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3921 return VM_FAULT_FALLBACK;
3925 * These routines also need to handle stuff like marking pages dirty
3926 * and/or accessed for architectures that don't do it in hardware (most
3927 * RISC architectures). The early dirtying is also good on the i386.
3929 * There is also a hook called "update_mmu_cache()" that architectures
3930 * with external mmu caches can use to update those (ie the Sparc or
3931 * PowerPC hashed page tables that act as extended TLBs).
3933 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3934 * concurrent faults).
3936 * The mmap_sem may have been released depending on flags and our return value.
3937 * See filemap_fault() and __lock_page_or_retry().
3939 static int handle_pte_fault(struct vm_fault *vmf)
3943 if (unlikely(pmd_none(*vmf->pmd))) {
3945 * Leave __pte_alloc() until later: because vm_ops->fault may
3946 * want to allocate huge page, and if we expose page table
3947 * for an instant, it will be difficult to retract from
3948 * concurrent faults and from rmap lookups.
3952 /* See comment in pte_alloc_one_map() */
3953 if (pmd_devmap_trans_unstable(vmf->pmd))
3956 * A regular pmd is established and it can't morph into a huge
3957 * pmd from under us anymore at this point because we hold the
3958 * mmap_sem read mode and khugepaged takes it in write mode.
3959 * So now it's safe to run pte_offset_map().
3961 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3962 vmf->orig_pte = *vmf->pte;
3965 * some architectures can have larger ptes than wordsize,
3966 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3967 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3968 * atomic accesses. The code below just needs a consistent
3969 * view for the ifs and we later double check anyway with the
3970 * ptl lock held. So here a barrier will do.
3973 if (pte_none(vmf->orig_pte)) {
3974 pte_unmap(vmf->pte);
3980 if (vma_is_anonymous(vmf->vma))
3981 return do_anonymous_page(vmf);
3983 return do_fault(vmf);
3986 if (!pte_present(vmf->orig_pte))
3987 return do_swap_page(vmf);
3989 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3990 return do_numa_page(vmf);
3992 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3993 spin_lock(vmf->ptl);
3994 entry = vmf->orig_pte;
3995 if (unlikely(!pte_same(*vmf->pte, entry)))
3997 if (vmf->flags & FAULT_FLAG_WRITE) {
3998 if (!pte_write(entry))
3999 return do_wp_page(vmf);
4000 entry = pte_mkdirty(entry);
4002 entry = pte_mkyoung(entry);
4003 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4004 vmf->flags & FAULT_FLAG_WRITE)) {
4005 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4008 * This is needed only for protection faults but the arch code
4009 * is not yet telling us if this is a protection fault or not.
4010 * This still avoids useless tlb flushes for .text page faults
4013 if (vmf->flags & FAULT_FLAG_WRITE)
4014 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4017 pte_unmap_unlock(vmf->pte, vmf->ptl);
4022 * By the time we get here, we already hold the mm semaphore
4024 * The mmap_sem may have been released depending on flags and our
4025 * return value. See filemap_fault() and __lock_page_or_retry().
4027 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4030 struct vm_fault vmf = {
4032 .address = address & PAGE_MASK,
4034 .pgoff = linear_page_index(vma, address),
4035 .gfp_mask = __get_fault_gfp_mask(vma),
4037 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4038 struct mm_struct *mm = vma->vm_mm;
4043 pgd = pgd_offset(mm, address);
4044 p4d = p4d_alloc(mm, pgd, address);
4046 return VM_FAULT_OOM;
4048 vmf.pud = pud_alloc(mm, p4d, address);
4050 return VM_FAULT_OOM;
4051 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4052 ret = create_huge_pud(&vmf);
4053 if (!(ret & VM_FAULT_FALLBACK))
4056 pud_t orig_pud = *vmf.pud;
4059 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4061 /* NUMA case for anonymous PUDs would go here */
4063 if (dirty && !pud_write(orig_pud)) {
4064 ret = wp_huge_pud(&vmf, orig_pud);
4065 if (!(ret & VM_FAULT_FALLBACK))
4068 huge_pud_set_accessed(&vmf, orig_pud);
4074 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4076 return VM_FAULT_OOM;
4077 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4078 ret = create_huge_pmd(&vmf);
4079 if (!(ret & VM_FAULT_FALLBACK))
4082 pmd_t orig_pmd = *vmf.pmd;
4085 if (unlikely(is_swap_pmd(orig_pmd))) {
4086 VM_BUG_ON(thp_migration_supported() &&
4087 !is_pmd_migration_entry(orig_pmd));
4088 if (is_pmd_migration_entry(orig_pmd))
4089 pmd_migration_entry_wait(mm, vmf.pmd);
4092 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4093 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4094 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4096 if (dirty && !pmd_write(orig_pmd)) {
4097 ret = wp_huge_pmd(&vmf, orig_pmd);
4098 if (!(ret & VM_FAULT_FALLBACK))
4101 huge_pmd_set_accessed(&vmf, orig_pmd);
4107 return handle_pte_fault(&vmf);
4111 * By the time we get here, we already hold the mm semaphore
4113 * The mmap_sem may have been released depending on flags and our
4114 * return value. See filemap_fault() and __lock_page_or_retry().
4116 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4121 __set_current_state(TASK_RUNNING);
4123 count_vm_event(PGFAULT);
4124 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4126 /* do counter updates before entering really critical section. */
4127 check_sync_rss_stat(current);
4129 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4130 flags & FAULT_FLAG_INSTRUCTION,
4131 flags & FAULT_FLAG_REMOTE))
4132 return VM_FAULT_SIGSEGV;
4135 * Enable the memcg OOM handling for faults triggered in user
4136 * space. Kernel faults are handled more gracefully.
4138 if (flags & FAULT_FLAG_USER)
4139 mem_cgroup_oom_enable();
4141 if (unlikely(is_vm_hugetlb_page(vma)))
4142 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4144 ret = __handle_mm_fault(vma, address, flags);
4146 if (flags & FAULT_FLAG_USER) {
4147 mem_cgroup_oom_disable();
4149 * The task may have entered a memcg OOM situation but
4150 * if the allocation error was handled gracefully (no
4151 * VM_FAULT_OOM), there is no need to kill anything.
4152 * Just clean up the OOM state peacefully.
4154 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4155 mem_cgroup_oom_synchronize(false);
4160 EXPORT_SYMBOL_GPL(handle_mm_fault);
4162 #ifndef __PAGETABLE_P4D_FOLDED
4164 * Allocate p4d page table.
4165 * We've already handled the fast-path in-line.
4167 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4169 p4d_t *new = p4d_alloc_one(mm, address);
4173 smp_wmb(); /* See comment in __pte_alloc */
4175 spin_lock(&mm->page_table_lock);
4176 if (pgd_present(*pgd)) /* Another has populated it */
4179 pgd_populate(mm, pgd, new);
4180 spin_unlock(&mm->page_table_lock);
4183 #endif /* __PAGETABLE_P4D_FOLDED */
4185 #ifndef __PAGETABLE_PUD_FOLDED
4187 * Allocate page upper directory.
4188 * We've already handled the fast-path in-line.
4190 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4192 pud_t *new = pud_alloc_one(mm, address);
4196 smp_wmb(); /* See comment in __pte_alloc */
4198 spin_lock(&mm->page_table_lock);
4199 #ifndef __ARCH_HAS_5LEVEL_HACK
4200 if (p4d_present(*p4d)) /* Another has populated it */
4203 p4d_populate(mm, p4d, new);
4205 if (pgd_present(*p4d)) /* Another has populated it */
4208 pgd_populate(mm, p4d, new);
4209 #endif /* __ARCH_HAS_5LEVEL_HACK */
4210 spin_unlock(&mm->page_table_lock);
4213 #endif /* __PAGETABLE_PUD_FOLDED */
4215 #ifndef __PAGETABLE_PMD_FOLDED
4217 * Allocate page middle directory.
4218 * We've already handled the fast-path in-line.
4220 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4223 pmd_t *new = pmd_alloc_one(mm, address);
4227 smp_wmb(); /* See comment in __pte_alloc */
4229 ptl = pud_lock(mm, pud);
4230 #ifndef __ARCH_HAS_4LEVEL_HACK
4231 if (!pud_present(*pud)) {
4233 pud_populate(mm, pud, new);
4234 } else /* Another has populated it */
4237 if (!pgd_present(*pud)) {
4239 pgd_populate(mm, pud, new);
4240 } else /* Another has populated it */
4242 #endif /* __ARCH_HAS_4LEVEL_HACK */
4246 #endif /* __PAGETABLE_PMD_FOLDED */
4248 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4249 unsigned long *start, unsigned long *end,
4250 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4258 pgd = pgd_offset(mm, address);
4259 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4262 p4d = p4d_offset(pgd, address);
4263 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4266 pud = pud_offset(p4d, address);
4267 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4270 pmd = pmd_offset(pud, address);
4271 VM_BUG_ON(pmd_trans_huge(*pmd));
4273 if (pmd_huge(*pmd)) {
4278 *start = address & PMD_MASK;
4279 *end = *start + PMD_SIZE;
4280 mmu_notifier_invalidate_range_start(mm, *start, *end);
4282 *ptlp = pmd_lock(mm, pmd);
4283 if (pmd_huge(*pmd)) {
4289 mmu_notifier_invalidate_range_end(mm, *start, *end);
4292 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4296 *start = address & PAGE_MASK;
4297 *end = *start + PAGE_SIZE;
4298 mmu_notifier_invalidate_range_start(mm, *start, *end);
4300 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4301 if (!pte_present(*ptep))
4306 pte_unmap_unlock(ptep, *ptlp);
4308 mmu_notifier_invalidate_range_end(mm, *start, *end);
4313 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4314 pte_t **ptepp, spinlock_t **ptlp)
4318 /* (void) is needed to make gcc happy */
4319 (void) __cond_lock(*ptlp,
4320 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4321 ptepp, NULL, ptlp)));
4325 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4326 unsigned long *start, unsigned long *end,
4327 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4331 /* (void) is needed to make gcc happy */
4332 (void) __cond_lock(*ptlp,
4333 !(res = __follow_pte_pmd(mm, address, start, end,
4334 ptepp, pmdpp, ptlp)));
4337 EXPORT_SYMBOL(follow_pte_pmd);
4340 * follow_pfn - look up PFN at a user virtual address
4341 * @vma: memory mapping
4342 * @address: user virtual address
4343 * @pfn: location to store found PFN
4345 * Only IO mappings and raw PFN mappings are allowed.
4347 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4349 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4356 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4359 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4362 *pfn = pte_pfn(*ptep);
4363 pte_unmap_unlock(ptep, ptl);
4366 EXPORT_SYMBOL(follow_pfn);
4368 #ifdef CONFIG_HAVE_IOREMAP_PROT
4369 int follow_phys(struct vm_area_struct *vma,
4370 unsigned long address, unsigned int flags,
4371 unsigned long *prot, resource_size_t *phys)
4377 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4380 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4384 if ((flags & FOLL_WRITE) && !pte_write(pte))
4387 *prot = pgprot_val(pte_pgprot(pte));
4388 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4392 pte_unmap_unlock(ptep, ptl);
4397 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4398 void *buf, int len, int write)
4400 resource_size_t phys_addr;
4401 unsigned long prot = 0;
4402 void __iomem *maddr;
4403 int offset = addr & (PAGE_SIZE-1);
4405 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4408 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4413 memcpy_toio(maddr + offset, buf, len);
4415 memcpy_fromio(buf, maddr + offset, len);
4420 EXPORT_SYMBOL_GPL(generic_access_phys);
4424 * Access another process' address space as given in mm. If non-NULL, use the
4425 * given task for page fault accounting.
4427 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4428 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4430 struct vm_area_struct *vma;
4431 void *old_buf = buf;
4432 int write = gup_flags & FOLL_WRITE;
4434 down_read(&mm->mmap_sem);
4435 /* ignore errors, just check how much was successfully transferred */
4437 int bytes, ret, offset;
4439 struct page *page = NULL;
4441 ret = get_user_pages_remote(tsk, mm, addr, 1,
4442 gup_flags, &page, &vma, NULL);
4444 #ifndef CONFIG_HAVE_IOREMAP_PROT
4448 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4449 * we can access using slightly different code.
4451 vma = find_vma(mm, addr);
4452 if (!vma || vma->vm_start > addr)
4454 if (vma->vm_ops && vma->vm_ops->access)
4455 ret = vma->vm_ops->access(vma, addr, buf,
4463 offset = addr & (PAGE_SIZE-1);
4464 if (bytes > PAGE_SIZE-offset)
4465 bytes = PAGE_SIZE-offset;
4469 copy_to_user_page(vma, page, addr,
4470 maddr + offset, buf, bytes);
4471 set_page_dirty_lock(page);
4473 copy_from_user_page(vma, page, addr,
4474 buf, maddr + offset, bytes);
4483 up_read(&mm->mmap_sem);
4485 return buf - old_buf;
4489 * access_remote_vm - access another process' address space
4490 * @mm: the mm_struct of the target address space
4491 * @addr: start address to access
4492 * @buf: source or destination buffer
4493 * @len: number of bytes to transfer
4494 * @gup_flags: flags modifying lookup behaviour
4496 * The caller must hold a reference on @mm.
4498 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4499 void *buf, int len, unsigned int gup_flags)
4501 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4505 * Access another process' address space.
4506 * Source/target buffer must be kernel space,
4507 * Do not walk the page table directly, use get_user_pages
4509 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4510 void *buf, int len, unsigned int gup_flags)
4512 struct mm_struct *mm;
4515 mm = get_task_mm(tsk);
4519 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4525 EXPORT_SYMBOL_GPL(access_process_vm);
4528 * Print the name of a VMA.
4530 void print_vma_addr(char *prefix, unsigned long ip)
4532 struct mm_struct *mm = current->mm;
4533 struct vm_area_struct *vma;
4536 * Do not print if we are in atomic
4537 * contexts (in exception stacks, etc.):
4539 if (preempt_count())
4542 down_read(&mm->mmap_sem);
4543 vma = find_vma(mm, ip);
4544 if (vma && vma->vm_file) {
4545 struct file *f = vma->vm_file;
4546 char *buf = (char *)__get_free_page(GFP_KERNEL);
4550 p = file_path(f, buf, PAGE_SIZE);
4553 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4555 vma->vm_end - vma->vm_start);
4556 free_page((unsigned long)buf);
4559 up_read(&mm->mmap_sem);
4562 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4563 void __might_fault(const char *file, int line)
4566 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4567 * holding the mmap_sem, this is safe because kernel memory doesn't
4568 * get paged out, therefore we'll never actually fault, and the
4569 * below annotations will generate false positives.
4571 if (uaccess_kernel())
4573 if (pagefault_disabled())
4575 __might_sleep(file, line, 0);
4576 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4578 might_lock_read(¤t->mm->mmap_sem);
4581 EXPORT_SYMBOL(__might_fault);
4584 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4585 static void clear_gigantic_page(struct page *page,
4587 unsigned int pages_per_huge_page)
4590 struct page *p = page;
4593 for (i = 0; i < pages_per_huge_page;
4594 i++, p = mem_map_next(p, page, i)) {
4596 clear_user_highpage(p, addr + i * PAGE_SIZE);
4599 void clear_huge_page(struct page *page,
4600 unsigned long addr_hint, unsigned int pages_per_huge_page)
4603 unsigned long addr = addr_hint &
4604 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4606 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4607 clear_gigantic_page(page, addr, pages_per_huge_page);
4611 /* Clear sub-page to access last to keep its cache lines hot */
4613 n = (addr_hint - addr) / PAGE_SIZE;
4614 if (2 * n <= pages_per_huge_page) {
4615 /* If sub-page to access in first half of huge page */
4618 /* Clear sub-pages at the end of huge page */
4619 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4621 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4624 /* If sub-page to access in second half of huge page */
4625 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4626 l = pages_per_huge_page - n;
4627 /* Clear sub-pages at the begin of huge page */
4628 for (i = 0; i < base; i++) {
4630 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4634 * Clear remaining sub-pages in left-right-left-right pattern
4635 * towards the sub-page to access
4637 for (i = 0; i < l; i++) {
4638 int left_idx = base + i;
4639 int right_idx = base + 2 * l - 1 - i;
4642 clear_user_highpage(page + left_idx,
4643 addr + left_idx * PAGE_SIZE);
4645 clear_user_highpage(page + right_idx,
4646 addr + right_idx * PAGE_SIZE);
4650 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4652 struct vm_area_struct *vma,
4653 unsigned int pages_per_huge_page)
4656 struct page *dst_base = dst;
4657 struct page *src_base = src;
4659 for (i = 0; i < pages_per_huge_page; ) {
4661 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4664 dst = mem_map_next(dst, dst_base, i);
4665 src = mem_map_next(src, src_base, i);
4669 void copy_user_huge_page(struct page *dst, struct page *src,
4670 unsigned long addr, struct vm_area_struct *vma,
4671 unsigned int pages_per_huge_page)
4675 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4676 copy_user_gigantic_page(dst, src, addr, vma,
4677 pages_per_huge_page);
4682 for (i = 0; i < pages_per_huge_page; i++) {
4684 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4688 long copy_huge_page_from_user(struct page *dst_page,
4689 const void __user *usr_src,
4690 unsigned int pages_per_huge_page,
4691 bool allow_pagefault)
4693 void *src = (void *)usr_src;
4695 unsigned long i, rc = 0;
4696 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4698 for (i = 0; i < pages_per_huge_page; i++) {
4699 if (allow_pagefault)
4700 page_kaddr = kmap(dst_page + i);
4702 page_kaddr = kmap_atomic(dst_page + i);
4703 rc = copy_from_user(page_kaddr,
4704 (const void __user *)(src + i * PAGE_SIZE),
4706 if (allow_pagefault)
4707 kunmap(dst_page + i);
4709 kunmap_atomic(page_kaddr);
4711 ret_val -= (PAGE_SIZE - rc);
4719 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4721 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4723 static struct kmem_cache *page_ptl_cachep;
4725 void __init ptlock_cache_init(void)
4727 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4731 bool ptlock_alloc(struct page *page)
4735 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4742 void ptlock_free(struct page *page)
4744 kmem_cache_free(page_ptl_cachep, page->ptl);