4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
93 EXPORT_SYMBOL(mem_map);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
119 static int __init disable_randmaps(char *s)
121 randomize_va_space = 0;
124 __setup("norandmaps", disable_randmaps);
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
129 unsigned long highest_memmap_pfn __read_mostly;
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 static int __init init_zero_pfn(void)
136 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 core_initcall(init_zero_pfn);
142 #if defined(SPLIT_RSS_COUNTING)
144 void sync_mm_rss(struct mm_struct *mm)
148 for (i = 0; i < NR_MM_COUNTERS; i++) {
149 if (current->rss_stat.count[i]) {
150 add_mm_counter(mm, i, current->rss_stat.count[i]);
151 current->rss_stat.count[i] = 0;
154 current->rss_stat.events = 0;
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 struct task_struct *task = current;
161 if (likely(task->mm == mm))
162 task->rss_stat.count[member] += val;
164 add_mm_counter(mm, member, val);
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH (64)
171 static void check_sync_rss_stat(struct task_struct *task)
173 if (unlikely(task != current))
175 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176 sync_mm_rss(task->mm);
178 #else /* SPLIT_RSS_COUNTING */
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183 static void check_sync_rss_stat(struct task_struct *task)
187 #endif /* SPLIT_RSS_COUNTING */
189 #ifdef HAVE_GENERIC_MMU_GATHER
191 static bool tlb_next_batch(struct mmu_gather *tlb)
193 struct mmu_gather_batch *batch;
197 tlb->active = batch->next;
201 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
204 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
211 batch->max = MAX_GATHER_BATCH;
213 tlb->active->next = batch;
219 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220 unsigned long start, unsigned long end)
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238 __tlb_reset_range(tlb);
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
251 __tlb_reset_range(tlb);
254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
256 struct mmu_gather_batch *batch;
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
262 tlb->active = &tlb->local;
265 void tlb_flush_mmu(struct mmu_gather *tlb)
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276 unsigned long start, unsigned long end, bool force)
278 struct mmu_gather_batch *batch, *next;
281 __tlb_adjust_range(tlb, start, end - start);
285 /* keep the page table cache within bounds */
288 for (batch = tlb->local.next; batch; batch = next) {
290 free_pages((unsigned long)batch, 0);
292 tlb->local.next = NULL;
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
302 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 struct mmu_gather_batch *batch;
306 VM_BUG_ON(!tlb->end);
307 VM_WARN_ON(tlb->page_size != page_size);
311 * Add the page and check if we are full. If so
314 batch->pages[batch->nr++] = page;
315 if (batch->nr == batch->max) {
316 if (!tlb_next_batch(tlb))
320 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
325 #endif /* HAVE_GENERIC_MMU_GATHER */
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
330 * See the comment near struct mmu_table_batch.
333 static void tlb_remove_table_smp_sync(void *arg)
335 /* Simply deliver the interrupt */
338 static void tlb_remove_table_one(void *table)
341 * This isn't an RCU grace period and hence the page-tables cannot be
342 * assumed to be actually RCU-freed.
344 * It is however sufficient for software page-table walkers that rely on
345 * IRQ disabling. See the comment near struct mmu_table_batch.
347 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
348 __tlb_remove_table(table);
351 static void tlb_remove_table_rcu(struct rcu_head *head)
353 struct mmu_table_batch *batch;
356 batch = container_of(head, struct mmu_table_batch, rcu);
358 for (i = 0; i < batch->nr; i++)
359 __tlb_remove_table(batch->tables[i]);
361 free_page((unsigned long)batch);
364 void tlb_table_flush(struct mmu_gather *tlb)
366 struct mmu_table_batch **batch = &tlb->batch;
369 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
374 void tlb_remove_table(struct mmu_gather *tlb, void *table)
376 struct mmu_table_batch **batch = &tlb->batch;
379 * When there's less then two users of this mm there cannot be a
380 * concurrent page-table walk.
382 if (atomic_read(&tlb->mm->mm_users) < 2) {
383 __tlb_remove_table(table);
387 if (*batch == NULL) {
388 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
389 if (*batch == NULL) {
390 tlb_remove_table_one(table);
395 (*batch)->tables[(*batch)->nr++] = table;
396 if ((*batch)->nr == MAX_TABLE_BATCH)
397 tlb_table_flush(tlb);
400 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
403 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
404 * @tlb: the mmu_gather structure to initialize
405 * @mm: the mm_struct of the target address space
406 * @start: start of the region that will be removed from the page-table
407 * @end: end of the region that will be removed from the page-table
409 * Called to initialize an (on-stack) mmu_gather structure for page-table
410 * tear-down from @mm. The @start and @end are set to 0 and -1
411 * respectively when @mm is without users and we're going to destroy
412 * the full address space (exit/execve).
414 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
415 unsigned long start, unsigned long end)
417 arch_tlb_gather_mmu(tlb, mm, start, end);
418 inc_tlb_flush_pending(tlb->mm);
421 void tlb_finish_mmu(struct mmu_gather *tlb,
422 unsigned long start, unsigned long end)
425 * If there are parallel threads are doing PTE changes on same range
426 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
427 * flush by batching, a thread has stable TLB entry can fail to flush
428 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
429 * forcefully if we detect parallel PTE batching threads.
431 bool force = mm_tlb_flush_nested(tlb->mm);
433 arch_tlb_finish_mmu(tlb, start, end, force);
434 dec_tlb_flush_pending(tlb->mm);
438 * Note: this doesn't free the actual pages themselves. That
439 * has been handled earlier when unmapping all the memory regions.
441 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
444 pgtable_t token = pmd_pgtable(*pmd);
446 pte_free_tlb(tlb, token, addr);
447 mm_dec_nr_ptes(tlb->mm);
450 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
451 unsigned long addr, unsigned long end,
452 unsigned long floor, unsigned long ceiling)
459 pmd = pmd_offset(pud, addr);
461 next = pmd_addr_end(addr, end);
462 if (pmd_none_or_clear_bad(pmd))
464 free_pte_range(tlb, pmd, addr);
465 } while (pmd++, addr = next, addr != end);
475 if (end - 1 > ceiling - 1)
478 pmd = pmd_offset(pud, start);
480 pmd_free_tlb(tlb, pmd, start);
481 mm_dec_nr_pmds(tlb->mm);
484 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
493 pud = pud_offset(p4d, addr);
495 next = pud_addr_end(addr, end);
496 if (pud_none_or_clear_bad(pud))
498 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
499 } while (pud++, addr = next, addr != end);
509 if (end - 1 > ceiling - 1)
512 pud = pud_offset(p4d, start);
514 pud_free_tlb(tlb, pud, start);
515 mm_dec_nr_puds(tlb->mm);
518 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
519 unsigned long addr, unsigned long end,
520 unsigned long floor, unsigned long ceiling)
527 p4d = p4d_offset(pgd, addr);
529 next = p4d_addr_end(addr, end);
530 if (p4d_none_or_clear_bad(p4d))
532 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
533 } while (p4d++, addr = next, addr != end);
539 ceiling &= PGDIR_MASK;
543 if (end - 1 > ceiling - 1)
546 p4d = p4d_offset(pgd, start);
548 p4d_free_tlb(tlb, p4d, start);
552 * This function frees user-level page tables of a process.
554 void free_pgd_range(struct mmu_gather *tlb,
555 unsigned long addr, unsigned long end,
556 unsigned long floor, unsigned long ceiling)
562 * The next few lines have given us lots of grief...
564 * Why are we testing PMD* at this top level? Because often
565 * there will be no work to do at all, and we'd prefer not to
566 * go all the way down to the bottom just to discover that.
568 * Why all these "- 1"s? Because 0 represents both the bottom
569 * of the address space and the top of it (using -1 for the
570 * top wouldn't help much: the masks would do the wrong thing).
571 * The rule is that addr 0 and floor 0 refer to the bottom of
572 * the address space, but end 0 and ceiling 0 refer to the top
573 * Comparisons need to use "end - 1" and "ceiling - 1" (though
574 * that end 0 case should be mythical).
576 * Wherever addr is brought up or ceiling brought down, we must
577 * be careful to reject "the opposite 0" before it confuses the
578 * subsequent tests. But what about where end is brought down
579 * by PMD_SIZE below? no, end can't go down to 0 there.
581 * Whereas we round start (addr) and ceiling down, by different
582 * masks at different levels, in order to test whether a table
583 * now has no other vmas using it, so can be freed, we don't
584 * bother to round floor or end up - the tests don't need that.
598 if (end - 1 > ceiling - 1)
603 * We add page table cache pages with PAGE_SIZE,
604 * (see pte_free_tlb()), flush the tlb if we need
606 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
607 pgd = pgd_offset(tlb->mm, addr);
609 next = pgd_addr_end(addr, end);
610 if (pgd_none_or_clear_bad(pgd))
612 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
613 } while (pgd++, addr = next, addr != end);
616 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
617 unsigned long floor, unsigned long ceiling)
620 struct vm_area_struct *next = vma->vm_next;
621 unsigned long addr = vma->vm_start;
624 * Hide vma from rmap and truncate_pagecache before freeing
627 unlink_anon_vmas(vma);
628 unlink_file_vma(vma);
630 if (is_vm_hugetlb_page(vma)) {
631 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
632 floor, next ? next->vm_start : ceiling);
635 * Optimization: gather nearby vmas into one call down
637 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
638 && !is_vm_hugetlb_page(next)) {
641 unlink_anon_vmas(vma);
642 unlink_file_vma(vma);
644 free_pgd_range(tlb, addr, vma->vm_end,
645 floor, next ? next->vm_start : ceiling);
651 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
654 pgtable_t new = pte_alloc_one(mm, address);
659 * Ensure all pte setup (eg. pte page lock and page clearing) are
660 * visible before the pte is made visible to other CPUs by being
661 * put into page tables.
663 * The other side of the story is the pointer chasing in the page
664 * table walking code (when walking the page table without locking;
665 * ie. most of the time). Fortunately, these data accesses consist
666 * of a chain of data-dependent loads, meaning most CPUs (alpha
667 * being the notable exception) will already guarantee loads are
668 * seen in-order. See the alpha page table accessors for the
669 * smp_read_barrier_depends() barriers in page table walking code.
671 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
673 ptl = pmd_lock(mm, pmd);
674 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
676 pmd_populate(mm, pmd, new);
685 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
687 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
691 smp_wmb(); /* See comment in __pte_alloc */
693 spin_lock(&init_mm.page_table_lock);
694 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
695 pmd_populate_kernel(&init_mm, pmd, new);
698 spin_unlock(&init_mm.page_table_lock);
700 pte_free_kernel(&init_mm, new);
704 static inline void init_rss_vec(int *rss)
706 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
709 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
713 if (current->mm == mm)
715 for (i = 0; i < NR_MM_COUNTERS; i++)
717 add_mm_counter(mm, i, rss[i]);
721 * This function is called to print an error when a bad pte
722 * is found. For example, we might have a PFN-mapped pte in
723 * a region that doesn't allow it.
725 * The calling function must still handle the error.
727 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
728 pte_t pte, struct page *page)
730 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
731 p4d_t *p4d = p4d_offset(pgd, addr);
732 pud_t *pud = pud_offset(p4d, addr);
733 pmd_t *pmd = pmd_offset(pud, addr);
734 struct address_space *mapping;
736 static unsigned long resume;
737 static unsigned long nr_shown;
738 static unsigned long nr_unshown;
741 * Allow a burst of 60 reports, then keep quiet for that minute;
742 * or allow a steady drip of one report per second.
744 if (nr_shown == 60) {
745 if (time_before(jiffies, resume)) {
750 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
757 resume = jiffies + 60 * HZ;
759 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
760 index = linear_page_index(vma, addr);
762 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
764 (long long)pte_val(pte), (long long)pmd_val(*pmd));
766 dump_page(page, "bad pte");
767 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
768 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
769 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
771 vma->vm_ops ? vma->vm_ops->fault : NULL,
772 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
773 mapping ? mapping->a_ops->readpage : NULL);
775 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
779 * vm_normal_page -- This function gets the "struct page" associated with a pte.
781 * "Special" mappings do not wish to be associated with a "struct page" (either
782 * it doesn't exist, or it exists but they don't want to touch it). In this
783 * case, NULL is returned here. "Normal" mappings do have a struct page.
785 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
786 * pte bit, in which case this function is trivial. Secondly, an architecture
787 * may not have a spare pte bit, which requires a more complicated scheme,
790 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
791 * special mapping (even if there are underlying and valid "struct pages").
792 * COWed pages of a VM_PFNMAP are always normal.
794 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
795 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
796 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
797 * mapping will always honor the rule
799 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
801 * And for normal mappings this is false.
803 * This restricts such mappings to be a linear translation from virtual address
804 * to pfn. To get around this restriction, we allow arbitrary mappings so long
805 * as the vma is not a COW mapping; in that case, we know that all ptes are
806 * special (because none can have been COWed).
809 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
811 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
812 * page" backing, however the difference is that _all_ pages with a struct
813 * page (that is, those where pfn_valid is true) are refcounted and considered
814 * normal pages by the VM. The disadvantage is that pages are refcounted
815 * (which can be slower and simply not an option for some PFNMAP users). The
816 * advantage is that we don't have to follow the strict linearity rule of
817 * PFNMAP mappings in order to support COWable mappings.
820 #ifdef __HAVE_ARCH_PTE_SPECIAL
821 # define HAVE_PTE_SPECIAL 1
823 # define HAVE_PTE_SPECIAL 0
825 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
826 pte_t pte, bool with_public_device)
828 unsigned long pfn = pte_pfn(pte);
830 if (HAVE_PTE_SPECIAL) {
831 if (likely(!pte_special(pte)))
833 if (vma->vm_ops && vma->vm_ops->find_special_page)
834 return vma->vm_ops->find_special_page(vma, addr);
835 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
837 if (is_zero_pfn(pfn))
841 * Device public pages are special pages (they are ZONE_DEVICE
842 * pages but different from persistent memory). They behave
843 * allmost like normal pages. The difference is that they are
844 * not on the lru and thus should never be involve with any-
845 * thing that involve lru manipulation (mlock, numa balancing,
848 * This is why we still want to return NULL for such page from
849 * vm_normal_page() so that we do not have to special case all
850 * call site of vm_normal_page().
852 if (likely(pfn <= highest_memmap_pfn)) {
853 struct page *page = pfn_to_page(pfn);
855 if (is_device_public_page(page)) {
856 if (with_public_device)
861 print_bad_pte(vma, addr, pte, NULL);
865 /* !HAVE_PTE_SPECIAL case follows: */
867 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
868 if (vma->vm_flags & VM_MIXEDMAP) {
874 off = (addr - vma->vm_start) >> PAGE_SHIFT;
875 if (pfn == vma->vm_pgoff + off)
877 if (!is_cow_mapping(vma->vm_flags))
882 if (is_zero_pfn(pfn))
885 if (unlikely(pfn > highest_memmap_pfn)) {
886 print_bad_pte(vma, addr, pte, NULL);
891 * NOTE! We still have PageReserved() pages in the page tables.
892 * eg. VDSO mappings can cause them to exist.
895 return pfn_to_page(pfn);
898 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
899 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
902 unsigned long pfn = pmd_pfn(pmd);
905 * There is no pmd_special() but there may be special pmds, e.g.
906 * in a direct-access (dax) mapping, so let's just replicate the
907 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
909 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
910 if (vma->vm_flags & VM_MIXEDMAP) {
916 off = (addr - vma->vm_start) >> PAGE_SHIFT;
917 if (pfn == vma->vm_pgoff + off)
919 if (!is_cow_mapping(vma->vm_flags))
924 if (is_zero_pfn(pfn))
926 if (unlikely(pfn > highest_memmap_pfn))
930 * NOTE! We still have PageReserved() pages in the page tables.
931 * eg. VDSO mappings can cause them to exist.
934 return pfn_to_page(pfn);
939 * copy one vm_area from one task to the other. Assumes the page tables
940 * already present in the new task to be cleared in the whole range
941 * covered by this vma.
944 static inline unsigned long
945 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
947 unsigned long addr, int *rss)
949 unsigned long vm_flags = vma->vm_flags;
950 pte_t pte = *src_pte;
953 /* pte contains position in swap or file, so copy. */
954 if (unlikely(!pte_present(pte))) {
955 swp_entry_t entry = pte_to_swp_entry(pte);
957 if (likely(!non_swap_entry(entry))) {
958 if (swap_duplicate(entry) < 0)
961 /* make sure dst_mm is on swapoff's mmlist. */
962 if (unlikely(list_empty(&dst_mm->mmlist))) {
963 spin_lock(&mmlist_lock);
964 if (list_empty(&dst_mm->mmlist))
965 list_add(&dst_mm->mmlist,
967 spin_unlock(&mmlist_lock);
970 } else if (is_migration_entry(entry)) {
971 page = migration_entry_to_page(entry);
973 rss[mm_counter(page)]++;
975 if (is_write_migration_entry(entry) &&
976 is_cow_mapping(vm_flags)) {
978 * COW mappings require pages in both
979 * parent and child to be set to read.
981 make_migration_entry_read(&entry);
982 pte = swp_entry_to_pte(entry);
983 if (pte_swp_soft_dirty(*src_pte))
984 pte = pte_swp_mksoft_dirty(pte);
985 set_pte_at(src_mm, addr, src_pte, pte);
987 } else if (is_device_private_entry(entry)) {
988 page = device_private_entry_to_page(entry);
991 * Update rss count even for unaddressable pages, as
992 * they should treated just like normal pages in this
995 * We will likely want to have some new rss counters
996 * for unaddressable pages, at some point. But for now
997 * keep things as they are.
1000 rss[mm_counter(page)]++;
1001 page_dup_rmap(page, false);
1004 * We do not preserve soft-dirty information, because so
1005 * far, checkpoint/restore is the only feature that
1006 * requires that. And checkpoint/restore does not work
1007 * when a device driver is involved (you cannot easily
1008 * save and restore device driver state).
1010 if (is_write_device_private_entry(entry) &&
1011 is_cow_mapping(vm_flags)) {
1012 make_device_private_entry_read(&entry);
1013 pte = swp_entry_to_pte(entry);
1014 set_pte_at(src_mm, addr, src_pte, pte);
1021 * If it's a COW mapping, write protect it both
1022 * in the parent and the child
1024 if (is_cow_mapping(vm_flags)) {
1025 ptep_set_wrprotect(src_mm, addr, src_pte);
1026 pte = pte_wrprotect(pte);
1030 * If it's a shared mapping, mark it clean in
1033 if (vm_flags & VM_SHARED)
1034 pte = pte_mkclean(pte);
1035 pte = pte_mkold(pte);
1037 page = vm_normal_page(vma, addr, pte);
1040 page_dup_rmap(page, false);
1041 rss[mm_counter(page)]++;
1042 } else if (pte_devmap(pte)) {
1043 page = pte_page(pte);
1046 * Cache coherent device memory behave like regular page and
1047 * not like persistent memory page. For more informations see
1048 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1050 if (is_device_public_page(page)) {
1052 page_dup_rmap(page, false);
1053 rss[mm_counter(page)]++;
1058 set_pte_at(dst_mm, addr, dst_pte, pte);
1062 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1063 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1064 unsigned long addr, unsigned long end)
1066 pte_t *orig_src_pte, *orig_dst_pte;
1067 pte_t *src_pte, *dst_pte;
1068 spinlock_t *src_ptl, *dst_ptl;
1070 int rss[NR_MM_COUNTERS];
1071 swp_entry_t entry = (swp_entry_t){0};
1076 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1079 src_pte = pte_offset_map(src_pmd, addr);
1080 src_ptl = pte_lockptr(src_mm, src_pmd);
1081 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1082 orig_src_pte = src_pte;
1083 orig_dst_pte = dst_pte;
1084 arch_enter_lazy_mmu_mode();
1088 * We are holding two locks at this point - either of them
1089 * could generate latencies in another task on another CPU.
1091 if (progress >= 32) {
1093 if (need_resched() ||
1094 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1097 if (pte_none(*src_pte)) {
1101 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1106 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1108 arch_leave_lazy_mmu_mode();
1109 spin_unlock(src_ptl);
1110 pte_unmap(orig_src_pte);
1111 add_mm_rss_vec(dst_mm, rss);
1112 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1116 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1125 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1126 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1127 unsigned long addr, unsigned long end)
1129 pmd_t *src_pmd, *dst_pmd;
1132 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1135 src_pmd = pmd_offset(src_pud, addr);
1137 next = pmd_addr_end(addr, end);
1138 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1139 || pmd_devmap(*src_pmd)) {
1141 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1142 err = copy_huge_pmd(dst_mm, src_mm,
1143 dst_pmd, src_pmd, addr, vma);
1150 if (pmd_none_or_clear_bad(src_pmd))
1152 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1155 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1159 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1160 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1161 unsigned long addr, unsigned long end)
1163 pud_t *src_pud, *dst_pud;
1166 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1169 src_pud = pud_offset(src_p4d, addr);
1171 next = pud_addr_end(addr, end);
1172 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1175 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1176 err = copy_huge_pud(dst_mm, src_mm,
1177 dst_pud, src_pud, addr, vma);
1184 if (pud_none_or_clear_bad(src_pud))
1186 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1189 } while (dst_pud++, src_pud++, addr = next, addr != end);
1193 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1194 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1195 unsigned long addr, unsigned long end)
1197 p4d_t *src_p4d, *dst_p4d;
1200 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1203 src_p4d = p4d_offset(src_pgd, addr);
1205 next = p4d_addr_end(addr, end);
1206 if (p4d_none_or_clear_bad(src_p4d))
1208 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1211 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1215 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1216 struct vm_area_struct *vma)
1218 pgd_t *src_pgd, *dst_pgd;
1220 unsigned long addr = vma->vm_start;
1221 unsigned long end = vma->vm_end;
1222 unsigned long mmun_start; /* For mmu_notifiers */
1223 unsigned long mmun_end; /* For mmu_notifiers */
1228 * Don't copy ptes where a page fault will fill them correctly.
1229 * Fork becomes much lighter when there are big shared or private
1230 * readonly mappings. The tradeoff is that copy_page_range is more
1231 * efficient than faulting.
1233 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1237 if (is_vm_hugetlb_page(vma))
1238 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1240 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1242 * We do not free on error cases below as remove_vma
1243 * gets called on error from higher level routine
1245 ret = track_pfn_copy(vma);
1251 * We need to invalidate the secondary MMU mappings only when
1252 * there could be a permission downgrade on the ptes of the
1253 * parent mm. And a permission downgrade will only happen if
1254 * is_cow_mapping() returns true.
1256 is_cow = is_cow_mapping(vma->vm_flags);
1260 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1264 dst_pgd = pgd_offset(dst_mm, addr);
1265 src_pgd = pgd_offset(src_mm, addr);
1267 next = pgd_addr_end(addr, end);
1268 if (pgd_none_or_clear_bad(src_pgd))
1270 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1271 vma, addr, next))) {
1275 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1278 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1282 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1283 struct vm_area_struct *vma, pmd_t *pmd,
1284 unsigned long addr, unsigned long end,
1285 struct zap_details *details)
1287 struct mm_struct *mm = tlb->mm;
1288 int force_flush = 0;
1289 int rss[NR_MM_COUNTERS];
1295 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1298 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1300 flush_tlb_batched_pending(mm);
1301 arch_enter_lazy_mmu_mode();
1304 if (pte_none(ptent))
1307 if (pte_present(ptent)) {
1310 page = _vm_normal_page(vma, addr, ptent, true);
1311 if (unlikely(details) && page) {
1313 * unmap_shared_mapping_pages() wants to
1314 * invalidate cache without truncating:
1315 * unmap shared but keep private pages.
1317 if (details->check_mapping &&
1318 details->check_mapping != page_rmapping(page))
1321 ptent = ptep_get_and_clear_full(mm, addr, pte,
1323 tlb_remove_tlb_entry(tlb, pte, addr);
1324 if (unlikely(!page))
1327 if (!PageAnon(page)) {
1328 if (pte_dirty(ptent)) {
1330 set_page_dirty(page);
1332 if (pte_young(ptent) &&
1333 likely(!(vma->vm_flags & VM_SEQ_READ)))
1334 mark_page_accessed(page);
1336 rss[mm_counter(page)]--;
1337 page_remove_rmap(page, false);
1338 if (unlikely(page_mapcount(page) < 0))
1339 print_bad_pte(vma, addr, ptent, page);
1340 if (unlikely(__tlb_remove_page(tlb, page))) {
1348 entry = pte_to_swp_entry(ptent);
1349 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1350 struct page *page = device_private_entry_to_page(entry);
1352 if (unlikely(details && details->check_mapping)) {
1354 * unmap_shared_mapping_pages() wants to
1355 * invalidate cache without truncating:
1356 * unmap shared but keep private pages.
1358 if (details->check_mapping !=
1359 page_rmapping(page))
1363 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1364 rss[mm_counter(page)]--;
1365 page_remove_rmap(page, false);
1370 /* If details->check_mapping, we leave swap entries. */
1371 if (unlikely(details))
1374 entry = pte_to_swp_entry(ptent);
1375 if (!non_swap_entry(entry))
1377 else if (is_migration_entry(entry)) {
1380 page = migration_entry_to_page(entry);
1381 rss[mm_counter(page)]--;
1383 if (unlikely(!free_swap_and_cache(entry)))
1384 print_bad_pte(vma, addr, ptent, NULL);
1385 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1386 } while (pte++, addr += PAGE_SIZE, addr != end);
1388 add_mm_rss_vec(mm, rss);
1389 arch_leave_lazy_mmu_mode();
1391 /* Do the actual TLB flush before dropping ptl */
1393 tlb_flush_mmu_tlbonly(tlb);
1394 pte_unmap_unlock(start_pte, ptl);
1397 * If we forced a TLB flush (either due to running out of
1398 * batch buffers or because we needed to flush dirty TLB
1399 * entries before releasing the ptl), free the batched
1400 * memory too. Restart if we didn't do everything.
1404 tlb_flush_mmu_free(tlb);
1412 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1413 struct vm_area_struct *vma, pud_t *pud,
1414 unsigned long addr, unsigned long end,
1415 struct zap_details *details)
1420 pmd = pmd_offset(pud, addr);
1422 next = pmd_addr_end(addr, end);
1423 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1424 if (next - addr != HPAGE_PMD_SIZE) {
1425 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1426 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1427 __split_huge_pmd(vma, pmd, addr, false, NULL);
1428 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1433 * Here there can be other concurrent MADV_DONTNEED or
1434 * trans huge page faults running, and if the pmd is
1435 * none or trans huge it can change under us. This is
1436 * because MADV_DONTNEED holds the mmap_sem in read
1439 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1441 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1444 } while (pmd++, addr = next, addr != end);
1449 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1450 struct vm_area_struct *vma, p4d_t *p4d,
1451 unsigned long addr, unsigned long end,
1452 struct zap_details *details)
1457 pud = pud_offset(p4d, addr);
1459 next = pud_addr_end(addr, end);
1460 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1461 if (next - addr != HPAGE_PUD_SIZE) {
1462 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1463 split_huge_pud(vma, pud, addr);
1464 } else if (zap_huge_pud(tlb, vma, pud, addr))
1468 if (pud_none_or_clear_bad(pud))
1470 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1473 } while (pud++, addr = next, addr != end);
1478 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1479 struct vm_area_struct *vma, pgd_t *pgd,
1480 unsigned long addr, unsigned long end,
1481 struct zap_details *details)
1486 p4d = p4d_offset(pgd, addr);
1488 next = p4d_addr_end(addr, end);
1489 if (p4d_none_or_clear_bad(p4d))
1491 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1492 } while (p4d++, addr = next, addr != end);
1497 void unmap_page_range(struct mmu_gather *tlb,
1498 struct vm_area_struct *vma,
1499 unsigned long addr, unsigned long end,
1500 struct zap_details *details)
1505 BUG_ON(addr >= end);
1506 tlb_start_vma(tlb, vma);
1507 pgd = pgd_offset(vma->vm_mm, addr);
1509 next = pgd_addr_end(addr, end);
1510 if (pgd_none_or_clear_bad(pgd))
1512 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1513 } while (pgd++, addr = next, addr != end);
1514 tlb_end_vma(tlb, vma);
1518 static void unmap_single_vma(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, unsigned long start_addr,
1520 unsigned long end_addr,
1521 struct zap_details *details)
1523 unsigned long start = max(vma->vm_start, start_addr);
1526 if (start >= vma->vm_end)
1528 end = min(vma->vm_end, end_addr);
1529 if (end <= vma->vm_start)
1533 uprobe_munmap(vma, start, end);
1535 if (unlikely(vma->vm_flags & VM_PFNMAP))
1536 untrack_pfn(vma, 0, 0);
1539 if (unlikely(is_vm_hugetlb_page(vma))) {
1541 * It is undesirable to test vma->vm_file as it
1542 * should be non-null for valid hugetlb area.
1543 * However, vm_file will be NULL in the error
1544 * cleanup path of mmap_region. When
1545 * hugetlbfs ->mmap method fails,
1546 * mmap_region() nullifies vma->vm_file
1547 * before calling this function to clean up.
1548 * Since no pte has actually been setup, it is
1549 * safe to do nothing in this case.
1552 i_mmap_lock_write(vma->vm_file->f_mapping);
1553 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1554 i_mmap_unlock_write(vma->vm_file->f_mapping);
1557 unmap_page_range(tlb, vma, start, end, details);
1562 * unmap_vmas - unmap a range of memory covered by a list of vma's
1563 * @tlb: address of the caller's struct mmu_gather
1564 * @vma: the starting vma
1565 * @start_addr: virtual address at which to start unmapping
1566 * @end_addr: virtual address at which to end unmapping
1568 * Unmap all pages in the vma list.
1570 * Only addresses between `start' and `end' will be unmapped.
1572 * The VMA list must be sorted in ascending virtual address order.
1574 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1575 * range after unmap_vmas() returns. So the only responsibility here is to
1576 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1577 * drops the lock and schedules.
1579 void unmap_vmas(struct mmu_gather *tlb,
1580 struct vm_area_struct *vma, unsigned long start_addr,
1581 unsigned long end_addr)
1583 struct mm_struct *mm = vma->vm_mm;
1585 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1586 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1587 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1588 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1592 * zap_page_range - remove user pages in a given range
1593 * @vma: vm_area_struct holding the applicable pages
1594 * @start: starting address of pages to zap
1595 * @size: number of bytes to zap
1597 * Caller must protect the VMA list
1599 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1602 struct mm_struct *mm = vma->vm_mm;
1603 struct mmu_gather tlb;
1604 unsigned long end = start + size;
1607 tlb_gather_mmu(&tlb, mm, start, end);
1608 update_hiwater_rss(mm);
1609 mmu_notifier_invalidate_range_start(mm, start, end);
1610 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1611 unmap_single_vma(&tlb, vma, start, end, NULL);
1614 * zap_page_range does not specify whether mmap_sem should be
1615 * held for read or write. That allows parallel zap_page_range
1616 * operations to unmap a PTE and defer a flush meaning that
1617 * this call observes pte_none and fails to flush the TLB.
1618 * Rather than adding a complex API, ensure that no stale
1619 * TLB entries exist when this call returns.
1621 flush_tlb_range(vma, start, end);
1624 mmu_notifier_invalidate_range_end(mm, start, end);
1625 tlb_finish_mmu(&tlb, start, end);
1629 * zap_page_range_single - remove user pages in a given range
1630 * @vma: vm_area_struct holding the applicable pages
1631 * @address: starting address of pages to zap
1632 * @size: number of bytes to zap
1633 * @details: details of shared cache invalidation
1635 * The range must fit into one VMA.
1637 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1638 unsigned long size, struct zap_details *details)
1640 struct mm_struct *mm = vma->vm_mm;
1641 struct mmu_gather tlb;
1642 unsigned long end = address + size;
1645 tlb_gather_mmu(&tlb, mm, address, end);
1646 update_hiwater_rss(mm);
1647 mmu_notifier_invalidate_range_start(mm, address, end);
1648 unmap_single_vma(&tlb, vma, address, end, details);
1649 mmu_notifier_invalidate_range_end(mm, address, end);
1650 tlb_finish_mmu(&tlb, address, end);
1654 * zap_vma_ptes - remove ptes mapping the vma
1655 * @vma: vm_area_struct holding ptes to be zapped
1656 * @address: starting address of pages to zap
1657 * @size: number of bytes to zap
1659 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1661 * The entire address range must be fully contained within the vma.
1664 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1667 if (address < vma->vm_start || address + size > vma->vm_end ||
1668 !(vma->vm_flags & VM_PFNMAP))
1671 zap_page_range_single(vma, address, size, NULL);
1673 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1675 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1683 pgd = pgd_offset(mm, addr);
1684 p4d = p4d_alloc(mm, pgd, addr);
1687 pud = pud_alloc(mm, p4d, addr);
1690 pmd = pmd_alloc(mm, pud, addr);
1694 VM_BUG_ON(pmd_trans_huge(*pmd));
1695 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1699 * This is the old fallback for page remapping.
1701 * For historical reasons, it only allows reserved pages. Only
1702 * old drivers should use this, and they needed to mark their
1703 * pages reserved for the old functions anyway.
1705 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1706 struct page *page, pgprot_t prot)
1708 struct mm_struct *mm = vma->vm_mm;
1717 flush_dcache_page(page);
1718 pte = get_locked_pte(mm, addr, &ptl);
1722 if (!pte_none(*pte))
1725 /* Ok, finally just insert the thing.. */
1727 inc_mm_counter_fast(mm, mm_counter_file(page));
1728 page_add_file_rmap(page, false);
1729 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1732 pte_unmap_unlock(pte, ptl);
1735 pte_unmap_unlock(pte, ptl);
1741 * vm_insert_page - insert single page into user vma
1742 * @vma: user vma to map to
1743 * @addr: target user address of this page
1744 * @page: source kernel page
1746 * This allows drivers to insert individual pages they've allocated
1749 * The page has to be a nice clean _individual_ kernel allocation.
1750 * If you allocate a compound page, you need to have marked it as
1751 * such (__GFP_COMP), or manually just split the page up yourself
1752 * (see split_page()).
1754 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1755 * took an arbitrary page protection parameter. This doesn't allow
1756 * that. Your vma protection will have to be set up correctly, which
1757 * means that if you want a shared writable mapping, you'd better
1758 * ask for a shared writable mapping!
1760 * The page does not need to be reserved.
1762 * Usually this function is called from f_op->mmap() handler
1763 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1764 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1765 * function from other places, for example from page-fault handler.
1767 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1770 if (addr < vma->vm_start || addr >= vma->vm_end)
1772 if (!page_count(page))
1774 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1775 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1776 BUG_ON(vma->vm_flags & VM_PFNMAP);
1777 vma->vm_flags |= VM_MIXEDMAP;
1779 return insert_page(vma, addr, page, vma->vm_page_prot);
1781 EXPORT_SYMBOL(vm_insert_page);
1783 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1784 pfn_t pfn, pgprot_t prot, bool mkwrite)
1786 struct mm_struct *mm = vma->vm_mm;
1792 pte = get_locked_pte(mm, addr, &ptl);
1796 if (!pte_none(*pte)) {
1799 * For read faults on private mappings the PFN passed
1800 * in may not match the PFN we have mapped if the
1801 * mapped PFN is a writeable COW page. In the mkwrite
1802 * case we are creating a writable PTE for a shared
1803 * mapping and we expect the PFNs to match.
1805 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1813 /* Ok, finally just insert the thing.. */
1814 if (pfn_t_devmap(pfn))
1815 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1817 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1821 entry = pte_mkyoung(entry);
1822 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1825 set_pte_at(mm, addr, pte, entry);
1826 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1830 pte_unmap_unlock(pte, ptl);
1836 * vm_insert_pfn - insert single pfn into user vma
1837 * @vma: user vma to map to
1838 * @addr: target user address of this page
1839 * @pfn: source kernel pfn
1841 * Similar to vm_insert_page, this allows drivers to insert individual pages
1842 * they've allocated into a user vma. Same comments apply.
1844 * This function should only be called from a vm_ops->fault handler, and
1845 * in that case the handler should return NULL.
1847 * vma cannot be a COW mapping.
1849 * As this is called only for pages that do not currently exist, we
1850 * do not need to flush old virtual caches or the TLB.
1852 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1855 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1857 EXPORT_SYMBOL(vm_insert_pfn);
1860 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1861 * @vma: user vma to map to
1862 * @addr: target user address of this page
1863 * @pfn: source kernel pfn
1864 * @pgprot: pgprot flags for the inserted page
1866 * This is exactly like vm_insert_pfn, except that it allows drivers to
1867 * to override pgprot on a per-page basis.
1869 * This only makes sense for IO mappings, and it makes no sense for
1870 * cow mappings. In general, using multiple vmas is preferable;
1871 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1874 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1875 unsigned long pfn, pgprot_t pgprot)
1879 * Technically, architectures with pte_special can avoid all these
1880 * restrictions (same for remap_pfn_range). However we would like
1881 * consistency in testing and feature parity among all, so we should
1882 * try to keep these invariants in place for everybody.
1884 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1885 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1886 (VM_PFNMAP|VM_MIXEDMAP));
1887 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1888 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1890 if (addr < vma->vm_start || addr >= vma->vm_end)
1893 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1895 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1900 EXPORT_SYMBOL(vm_insert_pfn_prot);
1902 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1904 /* these checks mirror the abort conditions in vm_normal_page */
1905 if (vma->vm_flags & VM_MIXEDMAP)
1907 if (pfn_t_devmap(pfn))
1909 if (pfn_t_special(pfn))
1911 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1916 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1917 pfn_t pfn, bool mkwrite)
1919 pgprot_t pgprot = vma->vm_page_prot;
1921 BUG_ON(!vm_mixed_ok(vma, pfn));
1923 if (addr < vma->vm_start || addr >= vma->vm_end)
1926 track_pfn_insert(vma, &pgprot, pfn);
1929 * If we don't have pte special, then we have to use the pfn_valid()
1930 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1931 * refcount the page if pfn_valid is true (hence insert_page rather
1932 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1933 * without pte special, it would there be refcounted as a normal page.
1935 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1939 * At this point we are committed to insert_page()
1940 * regardless of whether the caller specified flags that
1941 * result in pfn_t_has_page() == false.
1943 page = pfn_to_page(pfn_t_to_pfn(pfn));
1944 return insert_page(vma, addr, page, pgprot);
1946 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1949 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1952 return __vm_insert_mixed(vma, addr, pfn, false);
1955 EXPORT_SYMBOL(vm_insert_mixed);
1957 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1960 return __vm_insert_mixed(vma, addr, pfn, true);
1962 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1965 * maps a range of physical memory into the requested pages. the old
1966 * mappings are removed. any references to nonexistent pages results
1967 * in null mappings (currently treated as "copy-on-access")
1969 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1970 unsigned long addr, unsigned long end,
1971 unsigned long pfn, pgprot_t prot)
1976 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1979 arch_enter_lazy_mmu_mode();
1981 BUG_ON(!pte_none(*pte));
1982 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1984 } while (pte++, addr += PAGE_SIZE, addr != end);
1985 arch_leave_lazy_mmu_mode();
1986 pte_unmap_unlock(pte - 1, ptl);
1990 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1991 unsigned long addr, unsigned long end,
1992 unsigned long pfn, pgprot_t prot)
1997 pfn -= addr >> PAGE_SHIFT;
1998 pmd = pmd_alloc(mm, pud, addr);
2001 VM_BUG_ON(pmd_trans_huge(*pmd));
2003 next = pmd_addr_end(addr, end);
2004 if (remap_pte_range(mm, pmd, addr, next,
2005 pfn + (addr >> PAGE_SHIFT), prot))
2007 } while (pmd++, addr = next, addr != end);
2011 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2012 unsigned long addr, unsigned long end,
2013 unsigned long pfn, pgprot_t prot)
2018 pfn -= addr >> PAGE_SHIFT;
2019 pud = pud_alloc(mm, p4d, addr);
2023 next = pud_addr_end(addr, end);
2024 if (remap_pmd_range(mm, pud, addr, next,
2025 pfn + (addr >> PAGE_SHIFT), prot))
2027 } while (pud++, addr = next, addr != end);
2031 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2032 unsigned long addr, unsigned long end,
2033 unsigned long pfn, pgprot_t prot)
2038 pfn -= addr >> PAGE_SHIFT;
2039 p4d = p4d_alloc(mm, pgd, addr);
2043 next = p4d_addr_end(addr, end);
2044 if (remap_pud_range(mm, p4d, addr, next,
2045 pfn + (addr >> PAGE_SHIFT), prot))
2047 } while (p4d++, addr = next, addr != end);
2052 * remap_pfn_range - remap kernel memory to userspace
2053 * @vma: user vma to map to
2054 * @addr: target user address to start at
2055 * @pfn: physical address of kernel memory
2056 * @size: size of map area
2057 * @prot: page protection flags for this mapping
2059 * Note: this is only safe if the mm semaphore is held when called.
2061 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2062 unsigned long pfn, unsigned long size, pgprot_t prot)
2066 unsigned long end = addr + PAGE_ALIGN(size);
2067 struct mm_struct *mm = vma->vm_mm;
2068 unsigned long remap_pfn = pfn;
2072 * Physically remapped pages are special. Tell the
2073 * rest of the world about it:
2074 * VM_IO tells people not to look at these pages
2075 * (accesses can have side effects).
2076 * VM_PFNMAP tells the core MM that the base pages are just
2077 * raw PFN mappings, and do not have a "struct page" associated
2080 * Disable vma merging and expanding with mremap().
2082 * Omit vma from core dump, even when VM_IO turned off.
2084 * There's a horrible special case to handle copy-on-write
2085 * behaviour that some programs depend on. We mark the "original"
2086 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2087 * See vm_normal_page() for details.
2089 if (is_cow_mapping(vma->vm_flags)) {
2090 if (addr != vma->vm_start || end != vma->vm_end)
2092 vma->vm_pgoff = pfn;
2095 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2099 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2101 BUG_ON(addr >= end);
2102 pfn -= addr >> PAGE_SHIFT;
2103 pgd = pgd_offset(mm, addr);
2104 flush_cache_range(vma, addr, end);
2106 next = pgd_addr_end(addr, end);
2107 err = remap_p4d_range(mm, pgd, addr, next,
2108 pfn + (addr >> PAGE_SHIFT), prot);
2111 } while (pgd++, addr = next, addr != end);
2114 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2118 EXPORT_SYMBOL(remap_pfn_range);
2121 * vm_iomap_memory - remap memory to userspace
2122 * @vma: user vma to map to
2123 * @start: start of area
2124 * @len: size of area
2126 * This is a simplified io_remap_pfn_range() for common driver use. The
2127 * driver just needs to give us the physical memory range to be mapped,
2128 * we'll figure out the rest from the vma information.
2130 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2131 * whatever write-combining details or similar.
2133 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2135 unsigned long vm_len, pfn, pages;
2137 /* Check that the physical memory area passed in looks valid */
2138 if (start + len < start)
2141 * You *really* shouldn't map things that aren't page-aligned,
2142 * but we've historically allowed it because IO memory might
2143 * just have smaller alignment.
2145 len += start & ~PAGE_MASK;
2146 pfn = start >> PAGE_SHIFT;
2147 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2148 if (pfn + pages < pfn)
2151 /* We start the mapping 'vm_pgoff' pages into the area */
2152 if (vma->vm_pgoff > pages)
2154 pfn += vma->vm_pgoff;
2155 pages -= vma->vm_pgoff;
2157 /* Can we fit all of the mapping? */
2158 vm_len = vma->vm_end - vma->vm_start;
2159 if (vm_len >> PAGE_SHIFT > pages)
2162 /* Ok, let it rip */
2163 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2165 EXPORT_SYMBOL(vm_iomap_memory);
2167 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2168 unsigned long addr, unsigned long end,
2169 pte_fn_t fn, void *data)
2174 spinlock_t *uninitialized_var(ptl);
2176 pte = (mm == &init_mm) ?
2177 pte_alloc_kernel(pmd, addr) :
2178 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2182 BUG_ON(pmd_huge(*pmd));
2184 arch_enter_lazy_mmu_mode();
2186 token = pmd_pgtable(*pmd);
2189 err = fn(pte++, token, addr, data);
2192 } while (addr += PAGE_SIZE, addr != end);
2194 arch_leave_lazy_mmu_mode();
2197 pte_unmap_unlock(pte-1, ptl);
2201 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2202 unsigned long addr, unsigned long end,
2203 pte_fn_t fn, void *data)
2209 BUG_ON(pud_huge(*pud));
2211 pmd = pmd_alloc(mm, pud, addr);
2215 next = pmd_addr_end(addr, end);
2216 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2219 } while (pmd++, addr = next, addr != end);
2223 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2224 unsigned long addr, unsigned long end,
2225 pte_fn_t fn, void *data)
2231 pud = pud_alloc(mm, p4d, addr);
2235 next = pud_addr_end(addr, end);
2236 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2239 } while (pud++, addr = next, addr != end);
2243 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2244 unsigned long addr, unsigned long end,
2245 pte_fn_t fn, void *data)
2251 p4d = p4d_alloc(mm, pgd, addr);
2255 next = p4d_addr_end(addr, end);
2256 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2259 } while (p4d++, addr = next, addr != end);
2264 * Scan a region of virtual memory, filling in page tables as necessary
2265 * and calling a provided function on each leaf page table.
2267 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2268 unsigned long size, pte_fn_t fn, void *data)
2272 unsigned long end = addr + size;
2275 if (WARN_ON(addr >= end))
2278 pgd = pgd_offset(mm, addr);
2280 next = pgd_addr_end(addr, end);
2281 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2284 } while (pgd++, addr = next, addr != end);
2288 EXPORT_SYMBOL_GPL(apply_to_page_range);
2291 * handle_pte_fault chooses page fault handler according to an entry which was
2292 * read non-atomically. Before making any commitment, on those architectures
2293 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2294 * parts, do_swap_page must check under lock before unmapping the pte and
2295 * proceeding (but do_wp_page is only called after already making such a check;
2296 * and do_anonymous_page can safely check later on).
2298 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2299 pte_t *page_table, pte_t orig_pte)
2302 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2303 if (sizeof(pte_t) > sizeof(unsigned long)) {
2304 spinlock_t *ptl = pte_lockptr(mm, pmd);
2306 same = pte_same(*page_table, orig_pte);
2310 pte_unmap(page_table);
2314 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2316 debug_dma_assert_idle(src);
2319 * If the source page was a PFN mapping, we don't have
2320 * a "struct page" for it. We do a best-effort copy by
2321 * just copying from the original user address. If that
2322 * fails, we just zero-fill it. Live with it.
2324 if (unlikely(!src)) {
2325 void *kaddr = kmap_atomic(dst);
2326 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2329 * This really shouldn't fail, because the page is there
2330 * in the page tables. But it might just be unreadable,
2331 * in which case we just give up and fill the result with
2334 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2336 kunmap_atomic(kaddr);
2337 flush_dcache_page(dst);
2339 copy_user_highpage(dst, src, va, vma);
2342 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2344 struct file *vm_file = vma->vm_file;
2347 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2350 * Special mappings (e.g. VDSO) do not have any file so fake
2351 * a default GFP_KERNEL for them.
2357 * Notify the address space that the page is about to become writable so that
2358 * it can prohibit this or wait for the page to get into an appropriate state.
2360 * We do this without the lock held, so that it can sleep if it needs to.
2362 static int do_page_mkwrite(struct vm_fault *vmf)
2365 struct page *page = vmf->page;
2366 unsigned int old_flags = vmf->flags;
2368 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2370 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2371 /* Restore original flags so that caller is not surprised */
2372 vmf->flags = old_flags;
2373 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2375 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2377 if (!page->mapping) {
2379 return 0; /* retry */
2381 ret |= VM_FAULT_LOCKED;
2383 VM_BUG_ON_PAGE(!PageLocked(page), page);
2388 * Handle dirtying of a page in shared file mapping on a write fault.
2390 * The function expects the page to be locked and unlocks it.
2392 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2395 struct address_space *mapping;
2397 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2399 dirtied = set_page_dirty(page);
2400 VM_BUG_ON_PAGE(PageAnon(page), page);
2402 * Take a local copy of the address_space - page.mapping may be zeroed
2403 * by truncate after unlock_page(). The address_space itself remains
2404 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2405 * release semantics to prevent the compiler from undoing this copying.
2407 mapping = page_rmapping(page);
2410 if ((dirtied || page_mkwrite) && mapping) {
2412 * Some device drivers do not set page.mapping
2413 * but still dirty their pages
2415 balance_dirty_pages_ratelimited(mapping);
2419 file_update_time(vma->vm_file);
2423 * Handle write page faults for pages that can be reused in the current vma
2425 * This can happen either due to the mapping being with the VM_SHARED flag,
2426 * or due to us being the last reference standing to the page. In either
2427 * case, all we need to do here is to mark the page as writable and update
2428 * any related book-keeping.
2430 static inline void wp_page_reuse(struct vm_fault *vmf)
2431 __releases(vmf->ptl)
2433 struct vm_area_struct *vma = vmf->vma;
2434 struct page *page = vmf->page;
2437 * Clear the pages cpupid information as the existing
2438 * information potentially belongs to a now completely
2439 * unrelated process.
2442 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2444 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2445 entry = pte_mkyoung(vmf->orig_pte);
2446 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2447 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2448 update_mmu_cache(vma, vmf->address, vmf->pte);
2449 pte_unmap_unlock(vmf->pte, vmf->ptl);
2453 * Handle the case of a page which we actually need to copy to a new page.
2455 * Called with mmap_sem locked and the old page referenced, but
2456 * without the ptl held.
2458 * High level logic flow:
2460 * - Allocate a page, copy the content of the old page to the new one.
2461 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2462 * - Take the PTL. If the pte changed, bail out and release the allocated page
2463 * - If the pte is still the way we remember it, update the page table and all
2464 * relevant references. This includes dropping the reference the page-table
2465 * held to the old page, as well as updating the rmap.
2466 * - In any case, unlock the PTL and drop the reference we took to the old page.
2468 static int wp_page_copy(struct vm_fault *vmf)
2470 struct vm_area_struct *vma = vmf->vma;
2471 struct mm_struct *mm = vma->vm_mm;
2472 struct page *old_page = vmf->page;
2473 struct page *new_page = NULL;
2475 int page_copied = 0;
2476 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2477 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2478 struct mem_cgroup *memcg;
2480 if (unlikely(anon_vma_prepare(vma)))
2483 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2484 new_page = alloc_zeroed_user_highpage_movable(vma,
2489 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2493 cow_user_page(new_page, old_page, vmf->address, vma);
2496 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2499 __SetPageUptodate(new_page);
2501 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2504 * Re-check the pte - we dropped the lock
2506 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2507 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2509 if (!PageAnon(old_page)) {
2510 dec_mm_counter_fast(mm,
2511 mm_counter_file(old_page));
2512 inc_mm_counter_fast(mm, MM_ANONPAGES);
2515 inc_mm_counter_fast(mm, MM_ANONPAGES);
2517 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2518 entry = mk_pte(new_page, vma->vm_page_prot);
2519 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2521 * Clear the pte entry and flush it first, before updating the
2522 * pte with the new entry. This will avoid a race condition
2523 * seen in the presence of one thread doing SMC and another
2526 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2527 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2528 mem_cgroup_commit_charge(new_page, memcg, false, false);
2529 lru_cache_add_active_or_unevictable(new_page, vma);
2531 * We call the notify macro here because, when using secondary
2532 * mmu page tables (such as kvm shadow page tables), we want the
2533 * new page to be mapped directly into the secondary page table.
2535 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2536 update_mmu_cache(vma, vmf->address, vmf->pte);
2539 * Only after switching the pte to the new page may
2540 * we remove the mapcount here. Otherwise another
2541 * process may come and find the rmap count decremented
2542 * before the pte is switched to the new page, and
2543 * "reuse" the old page writing into it while our pte
2544 * here still points into it and can be read by other
2547 * The critical issue is to order this
2548 * page_remove_rmap with the ptp_clear_flush above.
2549 * Those stores are ordered by (if nothing else,)
2550 * the barrier present in the atomic_add_negative
2551 * in page_remove_rmap.
2553 * Then the TLB flush in ptep_clear_flush ensures that
2554 * no process can access the old page before the
2555 * decremented mapcount is visible. And the old page
2556 * cannot be reused until after the decremented
2557 * mapcount is visible. So transitively, TLBs to
2558 * old page will be flushed before it can be reused.
2560 page_remove_rmap(old_page, false);
2563 /* Free the old page.. */
2564 new_page = old_page;
2567 mem_cgroup_cancel_charge(new_page, memcg, false);
2573 pte_unmap_unlock(vmf->pte, vmf->ptl);
2575 * No need to double call mmu_notifier->invalidate_range() callback as
2576 * the above ptep_clear_flush_notify() did already call it.
2578 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2581 * Don't let another task, with possibly unlocked vma,
2582 * keep the mlocked page.
2584 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2585 lock_page(old_page); /* LRU manipulation */
2586 if (PageMlocked(old_page))
2587 munlock_vma_page(old_page);
2588 unlock_page(old_page);
2592 return page_copied ? VM_FAULT_WRITE : 0;
2598 return VM_FAULT_OOM;
2602 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2603 * writeable once the page is prepared
2605 * @vmf: structure describing the fault
2607 * This function handles all that is needed to finish a write page fault in a
2608 * shared mapping due to PTE being read-only once the mapped page is prepared.
2609 * It handles locking of PTE and modifying it. The function returns
2610 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2613 * The function expects the page to be locked or other protection against
2614 * concurrent faults / writeback (such as DAX radix tree locks).
2616 int finish_mkwrite_fault(struct vm_fault *vmf)
2618 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2619 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2622 * We might have raced with another page fault while we released the
2623 * pte_offset_map_lock.
2625 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2626 pte_unmap_unlock(vmf->pte, vmf->ptl);
2627 return VM_FAULT_NOPAGE;
2634 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2637 static int wp_pfn_shared(struct vm_fault *vmf)
2639 struct vm_area_struct *vma = vmf->vma;
2641 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2644 pte_unmap_unlock(vmf->pte, vmf->ptl);
2645 vmf->flags |= FAULT_FLAG_MKWRITE;
2646 ret = vma->vm_ops->pfn_mkwrite(vmf);
2647 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2649 return finish_mkwrite_fault(vmf);
2652 return VM_FAULT_WRITE;
2655 static int wp_page_shared(struct vm_fault *vmf)
2656 __releases(vmf->ptl)
2658 struct vm_area_struct *vma = vmf->vma;
2660 get_page(vmf->page);
2662 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2665 pte_unmap_unlock(vmf->pte, vmf->ptl);
2666 tmp = do_page_mkwrite(vmf);
2667 if (unlikely(!tmp || (tmp &
2668 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2669 put_page(vmf->page);
2672 tmp = finish_mkwrite_fault(vmf);
2673 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2674 unlock_page(vmf->page);
2675 put_page(vmf->page);
2680 lock_page(vmf->page);
2682 fault_dirty_shared_page(vma, vmf->page);
2683 put_page(vmf->page);
2685 return VM_FAULT_WRITE;
2689 * This routine handles present pages, when users try to write
2690 * to a shared page. It is done by copying the page to a new address
2691 * and decrementing the shared-page counter for the old page.
2693 * Note that this routine assumes that the protection checks have been
2694 * done by the caller (the low-level page fault routine in most cases).
2695 * Thus we can safely just mark it writable once we've done any necessary
2698 * We also mark the page dirty at this point even though the page will
2699 * change only once the write actually happens. This avoids a few races,
2700 * and potentially makes it more efficient.
2702 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2703 * but allow concurrent faults), with pte both mapped and locked.
2704 * We return with mmap_sem still held, but pte unmapped and unlocked.
2706 static int do_wp_page(struct vm_fault *vmf)
2707 __releases(vmf->ptl)
2709 struct vm_area_struct *vma = vmf->vma;
2711 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2714 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2717 * We should not cow pages in a shared writeable mapping.
2718 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2720 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2721 (VM_WRITE|VM_SHARED))
2722 return wp_pfn_shared(vmf);
2724 pte_unmap_unlock(vmf->pte, vmf->ptl);
2725 return wp_page_copy(vmf);
2729 * Take out anonymous pages first, anonymous shared vmas are
2730 * not dirty accountable.
2732 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2733 int total_map_swapcount;
2734 if (!trylock_page(vmf->page)) {
2735 get_page(vmf->page);
2736 pte_unmap_unlock(vmf->pte, vmf->ptl);
2737 lock_page(vmf->page);
2738 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2739 vmf->address, &vmf->ptl);
2740 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2741 unlock_page(vmf->page);
2742 pte_unmap_unlock(vmf->pte, vmf->ptl);
2743 put_page(vmf->page);
2746 put_page(vmf->page);
2748 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2749 if (total_map_swapcount == 1) {
2751 * The page is all ours. Move it to
2752 * our anon_vma so the rmap code will
2753 * not search our parent or siblings.
2754 * Protected against the rmap code by
2757 page_move_anon_rmap(vmf->page, vma);
2759 unlock_page(vmf->page);
2761 return VM_FAULT_WRITE;
2763 unlock_page(vmf->page);
2764 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2765 (VM_WRITE|VM_SHARED))) {
2766 return wp_page_shared(vmf);
2770 * Ok, we need to copy. Oh, well..
2772 get_page(vmf->page);
2774 pte_unmap_unlock(vmf->pte, vmf->ptl);
2775 return wp_page_copy(vmf);
2778 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2779 unsigned long start_addr, unsigned long end_addr,
2780 struct zap_details *details)
2782 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2785 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2786 struct zap_details *details)
2788 struct vm_area_struct *vma;
2789 pgoff_t vba, vea, zba, zea;
2791 vma_interval_tree_foreach(vma, root,
2792 details->first_index, details->last_index) {
2794 vba = vma->vm_pgoff;
2795 vea = vba + vma_pages(vma) - 1;
2796 zba = details->first_index;
2799 zea = details->last_index;
2803 unmap_mapping_range_vma(vma,
2804 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2805 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2811 * unmap_mapping_pages() - Unmap pages from processes.
2812 * @mapping: The address space containing pages to be unmapped.
2813 * @start: Index of first page to be unmapped.
2814 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2815 * @even_cows: Whether to unmap even private COWed pages.
2817 * Unmap the pages in this address space from any userspace process which
2818 * has them mmaped. Generally, you want to remove COWed pages as well when
2819 * a file is being truncated, but not when invalidating pages from the page
2822 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2823 pgoff_t nr, bool even_cows)
2825 struct zap_details details = { };
2827 details.check_mapping = even_cows ? NULL : mapping;
2828 details.first_index = start;
2829 details.last_index = start + nr - 1;
2830 if (details.last_index < details.first_index)
2831 details.last_index = ULONG_MAX;
2833 i_mmap_lock_write(mapping);
2834 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2835 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2836 i_mmap_unlock_write(mapping);
2840 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2841 * address_space corresponding to the specified byte range in the underlying
2844 * @mapping: the address space containing mmaps to be unmapped.
2845 * @holebegin: byte in first page to unmap, relative to the start of
2846 * the underlying file. This will be rounded down to a PAGE_SIZE
2847 * boundary. Note that this is different from truncate_pagecache(), which
2848 * must keep the partial page. In contrast, we must get rid of
2850 * @holelen: size of prospective hole in bytes. This will be rounded
2851 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2853 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2854 * but 0 when invalidating pagecache, don't throw away private data.
2856 void unmap_mapping_range(struct address_space *mapping,
2857 loff_t const holebegin, loff_t const holelen, int even_cows)
2859 pgoff_t hba = holebegin >> PAGE_SHIFT;
2860 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862 /* Check for overflow. */
2863 if (sizeof(holelen) > sizeof(hlen)) {
2865 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2866 if (holeend & ~(long long)ULONG_MAX)
2867 hlen = ULONG_MAX - hba + 1;
2870 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2872 EXPORT_SYMBOL(unmap_mapping_range);
2875 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2876 * but allow concurrent faults), and pte mapped but not yet locked.
2877 * We return with pte unmapped and unlocked.
2879 * We return with the mmap_sem locked or unlocked in the same cases
2880 * as does filemap_fault().
2882 int do_swap_page(struct vm_fault *vmf)
2884 struct vm_area_struct *vma = vmf->vma;
2885 struct page *page = NULL, *swapcache;
2886 struct mem_cgroup *memcg;
2893 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2896 entry = pte_to_swp_entry(vmf->orig_pte);
2897 if (unlikely(non_swap_entry(entry))) {
2898 if (is_migration_entry(entry)) {
2899 migration_entry_wait(vma->vm_mm, vmf->pmd,
2901 } else if (is_device_private_entry(entry)) {
2903 * For un-addressable device memory we call the pgmap
2904 * fault handler callback. The callback must migrate
2905 * the page back to some CPU accessible page.
2907 ret = device_private_entry_fault(vma, vmf->address, entry,
2908 vmf->flags, vmf->pmd);
2909 } else if (is_hwpoison_entry(entry)) {
2910 ret = VM_FAULT_HWPOISON;
2912 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2913 ret = VM_FAULT_SIGBUS;
2919 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2920 page = lookup_swap_cache(entry, vma, vmf->address);
2924 struct swap_info_struct *si = swp_swap_info(entry);
2926 if (si->flags & SWP_SYNCHRONOUS_IO &&
2927 __swap_count(si, entry) == 1) {
2928 /* skip swapcache */
2929 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2932 __SetPageLocked(page);
2933 __SetPageSwapBacked(page);
2934 set_page_private(page, entry.val);
2935 lru_cache_add_anon(page);
2936 swap_readpage(page, true);
2939 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2946 * Back out if somebody else faulted in this pte
2947 * while we released the pte lock.
2949 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2950 vmf->address, &vmf->ptl);
2951 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2953 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2957 /* Had to read the page from swap area: Major fault */
2958 ret = VM_FAULT_MAJOR;
2959 count_vm_event(PGMAJFAULT);
2960 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2961 } else if (PageHWPoison(page)) {
2963 * hwpoisoned dirty swapcache pages are kept for killing
2964 * owner processes (which may be unknown at hwpoison time)
2966 ret = VM_FAULT_HWPOISON;
2967 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2971 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2973 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2975 ret |= VM_FAULT_RETRY;
2980 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2981 * release the swapcache from under us. The page pin, and pte_same
2982 * test below, are not enough to exclude that. Even if it is still
2983 * swapcache, we need to check that the page's swap has not changed.
2985 if (unlikely((!PageSwapCache(page) ||
2986 page_private(page) != entry.val)) && swapcache)
2989 page = ksm_might_need_to_copy(page, vma, vmf->address);
2990 if (unlikely(!page)) {
2996 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3003 * Back out if somebody else already faulted in this pte.
3005 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3007 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3010 if (unlikely(!PageUptodate(page))) {
3011 ret = VM_FAULT_SIGBUS;
3016 * The page isn't present yet, go ahead with the fault.
3018 * Be careful about the sequence of operations here.
3019 * To get its accounting right, reuse_swap_page() must be called
3020 * while the page is counted on swap but not yet in mapcount i.e.
3021 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3022 * must be called after the swap_free(), or it will never succeed.
3025 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3026 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3027 pte = mk_pte(page, vma->vm_page_prot);
3028 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3029 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3030 vmf->flags &= ~FAULT_FLAG_WRITE;
3031 ret |= VM_FAULT_WRITE;
3032 exclusive = RMAP_EXCLUSIVE;
3034 flush_icache_page(vma, page);
3035 if (pte_swp_soft_dirty(vmf->orig_pte))
3036 pte = pte_mksoft_dirty(pte);
3037 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3038 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3039 vmf->orig_pte = pte;
3041 /* ksm created a completely new copy */
3042 if (unlikely(page != swapcache && swapcache)) {
3043 page_add_new_anon_rmap(page, vma, vmf->address, false);
3044 mem_cgroup_commit_charge(page, memcg, false, false);
3045 lru_cache_add_active_or_unevictable(page, vma);
3047 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3048 mem_cgroup_commit_charge(page, memcg, true, false);
3049 activate_page(page);
3053 if (mem_cgroup_swap_full(page) ||
3054 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3055 try_to_free_swap(page);
3057 if (page != swapcache && swapcache) {
3059 * Hold the lock to avoid the swap entry to be reused
3060 * until we take the PT lock for the pte_same() check
3061 * (to avoid false positives from pte_same). For
3062 * further safety release the lock after the swap_free
3063 * so that the swap count won't change under a
3064 * parallel locked swapcache.
3066 unlock_page(swapcache);
3067 put_page(swapcache);
3070 if (vmf->flags & FAULT_FLAG_WRITE) {
3071 ret |= do_wp_page(vmf);
3072 if (ret & VM_FAULT_ERROR)
3073 ret &= VM_FAULT_ERROR;
3077 /* No need to invalidate - it was non-present before */
3078 update_mmu_cache(vma, vmf->address, vmf->pte);
3080 pte_unmap_unlock(vmf->pte, vmf->ptl);
3084 mem_cgroup_cancel_charge(page, memcg, false);
3085 pte_unmap_unlock(vmf->pte, vmf->ptl);
3090 if (page != swapcache && swapcache) {
3091 unlock_page(swapcache);
3092 put_page(swapcache);
3098 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3099 * but allow concurrent faults), and pte mapped but not yet locked.
3100 * We return with mmap_sem still held, but pte unmapped and unlocked.
3102 static int do_anonymous_page(struct vm_fault *vmf)
3104 struct vm_area_struct *vma = vmf->vma;
3105 struct mem_cgroup *memcg;
3110 /* File mapping without ->vm_ops ? */
3111 if (vma->vm_flags & VM_SHARED)
3112 return VM_FAULT_SIGBUS;
3115 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3116 * pte_offset_map() on pmds where a huge pmd might be created
3117 * from a different thread.
3119 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3120 * parallel threads are excluded by other means.
3122 * Here we only have down_read(mmap_sem).
3124 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3125 return VM_FAULT_OOM;
3127 /* See the comment in pte_alloc_one_map() */
3128 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3131 /* Use the zero-page for reads */
3132 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3133 !mm_forbids_zeropage(vma->vm_mm)) {
3134 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3135 vma->vm_page_prot));
3136 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3137 vmf->address, &vmf->ptl);
3138 if (!pte_none(*vmf->pte))
3140 ret = check_stable_address_space(vma->vm_mm);
3143 /* Deliver the page fault to userland, check inside PT lock */
3144 if (userfaultfd_missing(vma)) {
3145 pte_unmap_unlock(vmf->pte, vmf->ptl);
3146 return handle_userfault(vmf, VM_UFFD_MISSING);
3151 /* Allocate our own private page. */
3152 if (unlikely(anon_vma_prepare(vma)))
3154 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3158 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3162 * The memory barrier inside __SetPageUptodate makes sure that
3163 * preceeding stores to the page contents become visible before
3164 * the set_pte_at() write.
3166 __SetPageUptodate(page);
3168 entry = mk_pte(page, vma->vm_page_prot);
3169 if (vma->vm_flags & VM_WRITE)
3170 entry = pte_mkwrite(pte_mkdirty(entry));
3172 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3174 if (!pte_none(*vmf->pte))
3177 ret = check_stable_address_space(vma->vm_mm);
3181 /* Deliver the page fault to userland, check inside PT lock */
3182 if (userfaultfd_missing(vma)) {
3183 pte_unmap_unlock(vmf->pte, vmf->ptl);
3184 mem_cgroup_cancel_charge(page, memcg, false);
3186 return handle_userfault(vmf, VM_UFFD_MISSING);
3189 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3190 page_add_new_anon_rmap(page, vma, vmf->address, false);
3191 mem_cgroup_commit_charge(page, memcg, false, false);
3192 lru_cache_add_active_or_unevictable(page, vma);
3194 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3196 /* No need to invalidate - it was non-present before */
3197 update_mmu_cache(vma, vmf->address, vmf->pte);
3199 pte_unmap_unlock(vmf->pte, vmf->ptl);
3202 mem_cgroup_cancel_charge(page, memcg, false);
3208 return VM_FAULT_OOM;
3212 * The mmap_sem must have been held on entry, and may have been
3213 * released depending on flags and vma->vm_ops->fault() return value.
3214 * See filemap_fault() and __lock_page_retry().
3216 static int __do_fault(struct vm_fault *vmf)
3218 struct vm_area_struct *vma = vmf->vma;
3221 ret = vma->vm_ops->fault(vmf);
3222 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3223 VM_FAULT_DONE_COW)))
3226 if (unlikely(PageHWPoison(vmf->page))) {
3227 if (ret & VM_FAULT_LOCKED)
3228 unlock_page(vmf->page);
3229 put_page(vmf->page);
3231 return VM_FAULT_HWPOISON;
3234 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3235 lock_page(vmf->page);
3237 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3243 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3244 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3245 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3246 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3248 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3250 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3253 static int pte_alloc_one_map(struct vm_fault *vmf)
3255 struct vm_area_struct *vma = vmf->vma;
3257 if (!pmd_none(*vmf->pmd))
3259 if (vmf->prealloc_pte) {
3260 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3261 if (unlikely(!pmd_none(*vmf->pmd))) {
3262 spin_unlock(vmf->ptl);
3266 mm_inc_nr_ptes(vma->vm_mm);
3267 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3268 spin_unlock(vmf->ptl);
3269 vmf->prealloc_pte = NULL;
3270 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3271 return VM_FAULT_OOM;
3275 * If a huge pmd materialized under us just retry later. Use
3276 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3277 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3278 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3279 * running immediately after a huge pmd fault in a different thread of
3280 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3281 * All we have to ensure is that it is a regular pmd that we can walk
3282 * with pte_offset_map() and we can do that through an atomic read in
3283 * C, which is what pmd_trans_unstable() provides.
3285 if (pmd_devmap_trans_unstable(vmf->pmd))
3286 return VM_FAULT_NOPAGE;
3289 * At this point we know that our vmf->pmd points to a page of ptes
3290 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3291 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3292 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3293 * be valid and we will re-check to make sure the vmf->pte isn't
3294 * pte_none() under vmf->ptl protection when we return to
3297 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3302 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3304 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3305 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3306 unsigned long haddr)
3308 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3309 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3311 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3316 static void deposit_prealloc_pte(struct vm_fault *vmf)
3318 struct vm_area_struct *vma = vmf->vma;
3320 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3322 * We are going to consume the prealloc table,
3323 * count that as nr_ptes.
3325 mm_inc_nr_ptes(vma->vm_mm);
3326 vmf->prealloc_pte = NULL;
3329 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3331 struct vm_area_struct *vma = vmf->vma;
3332 bool write = vmf->flags & FAULT_FLAG_WRITE;
3333 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3337 if (!transhuge_vma_suitable(vma, haddr))
3338 return VM_FAULT_FALLBACK;
3340 ret = VM_FAULT_FALLBACK;
3341 page = compound_head(page);
3344 * Archs like ppc64 need additonal space to store information
3345 * related to pte entry. Use the preallocated table for that.
3347 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3348 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3349 if (!vmf->prealloc_pte)
3350 return VM_FAULT_OOM;
3351 smp_wmb(); /* See comment in __pte_alloc() */
3354 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3355 if (unlikely(!pmd_none(*vmf->pmd)))
3358 for (i = 0; i < HPAGE_PMD_NR; i++)
3359 flush_icache_page(vma, page + i);
3361 entry = mk_huge_pmd(page, vma->vm_page_prot);
3363 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3365 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3366 page_add_file_rmap(page, true);
3368 * deposit and withdraw with pmd lock held
3370 if (arch_needs_pgtable_deposit())
3371 deposit_prealloc_pte(vmf);
3373 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3375 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3377 /* fault is handled */
3379 count_vm_event(THP_FILE_MAPPED);
3381 spin_unlock(vmf->ptl);
3385 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3393 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3394 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3396 * @vmf: fault environment
3397 * @memcg: memcg to charge page (only for private mappings)
3398 * @page: page to map
3400 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3403 * Target users are page handler itself and implementations of
3404 * vm_ops->map_pages.
3406 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3409 struct vm_area_struct *vma = vmf->vma;
3410 bool write = vmf->flags & FAULT_FLAG_WRITE;
3414 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3415 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3417 VM_BUG_ON_PAGE(memcg, page);
3419 ret = do_set_pmd(vmf, page);
3420 if (ret != VM_FAULT_FALLBACK)
3425 ret = pte_alloc_one_map(vmf);
3430 /* Re-check under ptl */
3431 if (unlikely(!pte_none(*vmf->pte)))
3432 return VM_FAULT_NOPAGE;
3434 flush_icache_page(vma, page);
3435 entry = mk_pte(page, vma->vm_page_prot);
3437 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3438 /* copy-on-write page */
3439 if (write && !(vma->vm_flags & VM_SHARED)) {
3440 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3441 page_add_new_anon_rmap(page, vma, vmf->address, false);
3442 mem_cgroup_commit_charge(page, memcg, false, false);
3443 lru_cache_add_active_or_unevictable(page, vma);
3445 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3446 page_add_file_rmap(page, false);
3448 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3450 /* no need to invalidate: a not-present page won't be cached */
3451 update_mmu_cache(vma, vmf->address, vmf->pte);
3458 * finish_fault - finish page fault once we have prepared the page to fault
3460 * @vmf: structure describing the fault
3462 * This function handles all that is needed to finish a page fault once the
3463 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3464 * given page, adds reverse page mapping, handles memcg charges and LRU
3465 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3468 * The function expects the page to be locked and on success it consumes a
3469 * reference of a page being mapped (for the PTE which maps it).
3471 int finish_fault(struct vm_fault *vmf)
3476 /* Did we COW the page? */
3477 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3478 !(vmf->vma->vm_flags & VM_SHARED))
3479 page = vmf->cow_page;
3484 * check even for read faults because we might have lost our CoWed
3487 if (!(vmf->vma->vm_flags & VM_SHARED))
3488 ret = check_stable_address_space(vmf->vma->vm_mm);
3490 ret = alloc_set_pte(vmf, vmf->memcg, page);
3492 pte_unmap_unlock(vmf->pte, vmf->ptl);
3496 static unsigned long fault_around_bytes __read_mostly =
3497 rounddown_pow_of_two(65536);
3499 #ifdef CONFIG_DEBUG_FS
3500 static int fault_around_bytes_get(void *data, u64 *val)
3502 *val = fault_around_bytes;
3507 * fault_around_bytes must be rounded down to the nearest page order as it's
3508 * what do_fault_around() expects to see.
3510 static int fault_around_bytes_set(void *data, u64 val)
3512 if (val / PAGE_SIZE > PTRS_PER_PTE)
3514 if (val > PAGE_SIZE)
3515 fault_around_bytes = rounddown_pow_of_two(val);
3517 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3520 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3521 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3523 static int __init fault_around_debugfs(void)
3527 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3528 &fault_around_bytes_fops);
3530 pr_warn("Failed to create fault_around_bytes in debugfs");
3533 late_initcall(fault_around_debugfs);
3537 * do_fault_around() tries to map few pages around the fault address. The hope
3538 * is that the pages will be needed soon and this will lower the number of
3541 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3542 * not ready to be mapped: not up-to-date, locked, etc.
3544 * This function is called with the page table lock taken. In the split ptlock
3545 * case the page table lock only protects only those entries which belong to
3546 * the page table corresponding to the fault address.
3548 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3551 * fault_around_bytes defines how many bytes we'll try to map.
3552 * do_fault_around() expects it to be set to a power of two less than or equal
3555 * The virtual address of the area that we map is naturally aligned to
3556 * fault_around_bytes rounded down to the machine page size
3557 * (and therefore to page order). This way it's easier to guarantee
3558 * that we don't cross page table boundaries.
3560 static int do_fault_around(struct vm_fault *vmf)
3562 unsigned long address = vmf->address, nr_pages, mask;
3563 pgoff_t start_pgoff = vmf->pgoff;
3567 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3568 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3570 vmf->address = max(address & mask, vmf->vma->vm_start);
3571 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3575 * end_pgoff is either the end of the page table, the end of
3576 * the vma or nr_pages from start_pgoff, depending what is nearest.
3578 end_pgoff = start_pgoff -
3579 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3581 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3582 start_pgoff + nr_pages - 1);
3584 if (pmd_none(*vmf->pmd)) {
3585 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3587 if (!vmf->prealloc_pte)
3589 smp_wmb(); /* See comment in __pte_alloc() */
3592 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3594 /* Huge page is mapped? Page fault is solved */
3595 if (pmd_trans_huge(*vmf->pmd)) {
3596 ret = VM_FAULT_NOPAGE;
3600 /* ->map_pages() haven't done anything useful. Cold page cache? */
3604 /* check if the page fault is solved */
3605 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3606 if (!pte_none(*vmf->pte))
3607 ret = VM_FAULT_NOPAGE;
3608 pte_unmap_unlock(vmf->pte, vmf->ptl);
3610 vmf->address = address;
3615 static int do_read_fault(struct vm_fault *vmf)
3617 struct vm_area_struct *vma = vmf->vma;
3621 * Let's call ->map_pages() first and use ->fault() as fallback
3622 * if page by the offset is not ready to be mapped (cold cache or
3625 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3626 ret = do_fault_around(vmf);
3631 ret = __do_fault(vmf);
3632 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3635 ret |= finish_fault(vmf);
3636 unlock_page(vmf->page);
3637 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3638 put_page(vmf->page);
3642 static int do_cow_fault(struct vm_fault *vmf)
3644 struct vm_area_struct *vma = vmf->vma;
3647 if (unlikely(anon_vma_prepare(vma)))
3648 return VM_FAULT_OOM;
3650 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3652 return VM_FAULT_OOM;
3654 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3655 &vmf->memcg, false)) {
3656 put_page(vmf->cow_page);
3657 return VM_FAULT_OOM;
3660 ret = __do_fault(vmf);
3661 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3663 if (ret & VM_FAULT_DONE_COW)
3666 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3667 __SetPageUptodate(vmf->cow_page);
3669 ret |= finish_fault(vmf);
3670 unlock_page(vmf->page);
3671 put_page(vmf->page);
3672 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3676 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3677 put_page(vmf->cow_page);
3681 static int do_shared_fault(struct vm_fault *vmf)
3683 struct vm_area_struct *vma = vmf->vma;
3686 ret = __do_fault(vmf);
3687 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3691 * Check if the backing address space wants to know that the page is
3692 * about to become writable
3694 if (vma->vm_ops->page_mkwrite) {
3695 unlock_page(vmf->page);
3696 tmp = do_page_mkwrite(vmf);
3697 if (unlikely(!tmp ||
3698 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3699 put_page(vmf->page);
3704 ret |= finish_fault(vmf);
3705 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3707 unlock_page(vmf->page);
3708 put_page(vmf->page);
3712 fault_dirty_shared_page(vma, vmf->page);
3717 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3718 * but allow concurrent faults).
3719 * The mmap_sem may have been released depending on flags and our
3720 * return value. See filemap_fault() and __lock_page_or_retry().
3722 static int do_fault(struct vm_fault *vmf)
3724 struct vm_area_struct *vma = vmf->vma;
3727 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3728 if (!vma->vm_ops->fault)
3729 ret = VM_FAULT_SIGBUS;
3730 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3731 ret = do_read_fault(vmf);
3732 else if (!(vma->vm_flags & VM_SHARED))
3733 ret = do_cow_fault(vmf);
3735 ret = do_shared_fault(vmf);
3737 /* preallocated pagetable is unused: free it */
3738 if (vmf->prealloc_pte) {
3739 pte_free(vma->vm_mm, vmf->prealloc_pte);
3740 vmf->prealloc_pte = NULL;
3745 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3746 unsigned long addr, int page_nid,
3751 count_vm_numa_event(NUMA_HINT_FAULTS);
3752 if (page_nid == numa_node_id()) {
3753 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3754 *flags |= TNF_FAULT_LOCAL;
3757 return mpol_misplaced(page, vma, addr);
3760 static int do_numa_page(struct vm_fault *vmf)
3762 struct vm_area_struct *vma = vmf->vma;
3763 struct page *page = NULL;
3767 bool migrated = false;
3769 bool was_writable = pte_savedwrite(vmf->orig_pte);
3773 * The "pte" at this point cannot be used safely without
3774 * validation through pte_unmap_same(). It's of NUMA type but
3775 * the pfn may be screwed if the read is non atomic.
3777 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3778 spin_lock(vmf->ptl);
3779 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3780 pte_unmap_unlock(vmf->pte, vmf->ptl);
3785 * Make it present again, Depending on how arch implementes non
3786 * accessible ptes, some can allow access by kernel mode.
3788 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3789 pte = pte_modify(pte, vma->vm_page_prot);
3790 pte = pte_mkyoung(pte);
3792 pte = pte_mkwrite(pte);
3793 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3794 update_mmu_cache(vma, vmf->address, vmf->pte);
3796 page = vm_normal_page(vma, vmf->address, pte);
3798 pte_unmap_unlock(vmf->pte, vmf->ptl);
3802 /* TODO: handle PTE-mapped THP */
3803 if (PageCompound(page)) {
3804 pte_unmap_unlock(vmf->pte, vmf->ptl);
3809 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3810 * much anyway since they can be in shared cache state. This misses
3811 * the case where a mapping is writable but the process never writes
3812 * to it but pte_write gets cleared during protection updates and
3813 * pte_dirty has unpredictable behaviour between PTE scan updates,
3814 * background writeback, dirty balancing and application behaviour.
3816 if (!pte_write(pte))
3817 flags |= TNF_NO_GROUP;
3820 * Flag if the page is shared between multiple address spaces. This
3821 * is later used when determining whether to group tasks together
3823 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3824 flags |= TNF_SHARED;
3826 last_cpupid = page_cpupid_last(page);
3827 page_nid = page_to_nid(page);
3828 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3830 pte_unmap_unlock(vmf->pte, vmf->ptl);
3831 if (target_nid == -1) {
3836 /* Migrate to the requested node */
3837 migrated = migrate_misplaced_page(page, vma, target_nid);
3839 page_nid = target_nid;
3840 flags |= TNF_MIGRATED;
3842 flags |= TNF_MIGRATE_FAIL;
3846 task_numa_fault(last_cpupid, page_nid, 1, flags);
3850 static inline int create_huge_pmd(struct vm_fault *vmf)
3852 if (vma_is_anonymous(vmf->vma))
3853 return do_huge_pmd_anonymous_page(vmf);
3854 if (vmf->vma->vm_ops->huge_fault)
3855 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3856 return VM_FAULT_FALLBACK;
3859 /* `inline' is required to avoid gcc 4.1.2 build error */
3860 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3862 if (vma_is_anonymous(vmf->vma))
3863 return do_huge_pmd_wp_page(vmf, orig_pmd);
3864 if (vmf->vma->vm_ops->huge_fault)
3865 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3867 /* COW handled on pte level: split pmd */
3868 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3869 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3871 return VM_FAULT_FALLBACK;
3874 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3876 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3879 static int create_huge_pud(struct vm_fault *vmf)
3881 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3882 /* No support for anonymous transparent PUD pages yet */
3883 if (vma_is_anonymous(vmf->vma))
3884 return VM_FAULT_FALLBACK;
3885 if (vmf->vma->vm_ops->huge_fault)
3886 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3887 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3888 return VM_FAULT_FALLBACK;
3891 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3893 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3894 /* No support for anonymous transparent PUD pages yet */
3895 if (vma_is_anonymous(vmf->vma))
3896 return VM_FAULT_FALLBACK;
3897 if (vmf->vma->vm_ops->huge_fault)
3898 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3899 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3900 return VM_FAULT_FALLBACK;
3904 * These routines also need to handle stuff like marking pages dirty
3905 * and/or accessed for architectures that don't do it in hardware (most
3906 * RISC architectures). The early dirtying is also good on the i386.
3908 * There is also a hook called "update_mmu_cache()" that architectures
3909 * with external mmu caches can use to update those (ie the Sparc or
3910 * PowerPC hashed page tables that act as extended TLBs).
3912 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3913 * concurrent faults).
3915 * The mmap_sem may have been released depending on flags and our return value.
3916 * See filemap_fault() and __lock_page_or_retry().
3918 static int handle_pte_fault(struct vm_fault *vmf)
3922 if (unlikely(pmd_none(*vmf->pmd))) {
3924 * Leave __pte_alloc() until later: because vm_ops->fault may
3925 * want to allocate huge page, and if we expose page table
3926 * for an instant, it will be difficult to retract from
3927 * concurrent faults and from rmap lookups.
3931 /* See comment in pte_alloc_one_map() */
3932 if (pmd_devmap_trans_unstable(vmf->pmd))
3935 * A regular pmd is established and it can't morph into a huge
3936 * pmd from under us anymore at this point because we hold the
3937 * mmap_sem read mode and khugepaged takes it in write mode.
3938 * So now it's safe to run pte_offset_map().
3940 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3941 vmf->orig_pte = *vmf->pte;
3944 * some architectures can have larger ptes than wordsize,
3945 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3946 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3947 * accesses. The code below just needs a consistent view
3948 * for the ifs and we later double check anyway with the
3949 * ptl lock held. So here a barrier will do.
3952 if (pte_none(vmf->orig_pte)) {
3953 pte_unmap(vmf->pte);
3959 if (vma_is_anonymous(vmf->vma))
3960 return do_anonymous_page(vmf);
3962 return do_fault(vmf);
3965 if (!pte_present(vmf->orig_pte))
3966 return do_swap_page(vmf);
3968 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3969 return do_numa_page(vmf);
3971 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3972 spin_lock(vmf->ptl);
3973 entry = vmf->orig_pte;
3974 if (unlikely(!pte_same(*vmf->pte, entry)))
3976 if (vmf->flags & FAULT_FLAG_WRITE) {
3977 if (!pte_write(entry))
3978 return do_wp_page(vmf);
3979 entry = pte_mkdirty(entry);
3981 entry = pte_mkyoung(entry);
3982 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3983 vmf->flags & FAULT_FLAG_WRITE)) {
3984 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3987 * This is needed only for protection faults but the arch code
3988 * is not yet telling us if this is a protection fault or not.
3989 * This still avoids useless tlb flushes for .text page faults
3992 if (vmf->flags & FAULT_FLAG_WRITE)
3993 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3996 pte_unmap_unlock(vmf->pte, vmf->ptl);
4001 * By the time we get here, we already hold the mm semaphore
4003 * The mmap_sem may have been released depending on flags and our
4004 * return value. See filemap_fault() and __lock_page_or_retry().
4006 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4009 struct vm_fault vmf = {
4011 .address = address & PAGE_MASK,
4013 .pgoff = linear_page_index(vma, address),
4014 .gfp_mask = __get_fault_gfp_mask(vma),
4016 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4017 struct mm_struct *mm = vma->vm_mm;
4022 pgd = pgd_offset(mm, address);
4023 p4d = p4d_alloc(mm, pgd, address);
4025 return VM_FAULT_OOM;
4027 vmf.pud = pud_alloc(mm, p4d, address);
4029 return VM_FAULT_OOM;
4030 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4031 ret = create_huge_pud(&vmf);
4032 if (!(ret & VM_FAULT_FALLBACK))
4035 pud_t orig_pud = *vmf.pud;
4038 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4040 /* NUMA case for anonymous PUDs would go here */
4042 if (dirty && !pud_write(orig_pud)) {
4043 ret = wp_huge_pud(&vmf, orig_pud);
4044 if (!(ret & VM_FAULT_FALLBACK))
4047 huge_pud_set_accessed(&vmf, orig_pud);
4053 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4055 return VM_FAULT_OOM;
4056 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4057 ret = create_huge_pmd(&vmf);
4058 if (!(ret & VM_FAULT_FALLBACK))
4061 pmd_t orig_pmd = *vmf.pmd;
4064 if (unlikely(is_swap_pmd(orig_pmd))) {
4065 VM_BUG_ON(thp_migration_supported() &&
4066 !is_pmd_migration_entry(orig_pmd));
4067 if (is_pmd_migration_entry(orig_pmd))
4068 pmd_migration_entry_wait(mm, vmf.pmd);
4071 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4072 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4073 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4075 if (dirty && !pmd_write(orig_pmd)) {
4076 ret = wp_huge_pmd(&vmf, orig_pmd);
4077 if (!(ret & VM_FAULT_FALLBACK))
4080 huge_pmd_set_accessed(&vmf, orig_pmd);
4086 return handle_pte_fault(&vmf);
4090 * By the time we get here, we already hold the mm semaphore
4092 * The mmap_sem may have been released depending on flags and our
4093 * return value. See filemap_fault() and __lock_page_or_retry().
4095 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4100 __set_current_state(TASK_RUNNING);
4102 count_vm_event(PGFAULT);
4103 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4105 /* do counter updates before entering really critical section. */
4106 check_sync_rss_stat(current);
4108 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4109 flags & FAULT_FLAG_INSTRUCTION,
4110 flags & FAULT_FLAG_REMOTE))
4111 return VM_FAULT_SIGSEGV;
4114 * Enable the memcg OOM handling for faults triggered in user
4115 * space. Kernel faults are handled more gracefully.
4117 if (flags & FAULT_FLAG_USER)
4118 mem_cgroup_oom_enable();
4120 if (unlikely(is_vm_hugetlb_page(vma)))
4121 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4123 ret = __handle_mm_fault(vma, address, flags);
4125 if (flags & FAULT_FLAG_USER) {
4126 mem_cgroup_oom_disable();
4128 * The task may have entered a memcg OOM situation but
4129 * if the allocation error was handled gracefully (no
4130 * VM_FAULT_OOM), there is no need to kill anything.
4131 * Just clean up the OOM state peacefully.
4133 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4134 mem_cgroup_oom_synchronize(false);
4139 EXPORT_SYMBOL_GPL(handle_mm_fault);
4141 #ifndef __PAGETABLE_P4D_FOLDED
4143 * Allocate p4d page table.
4144 * We've already handled the fast-path in-line.
4146 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4148 p4d_t *new = p4d_alloc_one(mm, address);
4152 smp_wmb(); /* See comment in __pte_alloc */
4154 spin_lock(&mm->page_table_lock);
4155 if (pgd_present(*pgd)) /* Another has populated it */
4158 pgd_populate(mm, pgd, new);
4159 spin_unlock(&mm->page_table_lock);
4162 #endif /* __PAGETABLE_P4D_FOLDED */
4164 #ifndef __PAGETABLE_PUD_FOLDED
4166 * Allocate page upper directory.
4167 * We've already handled the fast-path in-line.
4169 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4171 pud_t *new = pud_alloc_one(mm, address);
4175 smp_wmb(); /* See comment in __pte_alloc */
4177 spin_lock(&mm->page_table_lock);
4178 #ifndef __ARCH_HAS_5LEVEL_HACK
4179 if (!p4d_present(*p4d)) {
4181 p4d_populate(mm, p4d, new);
4182 } else /* Another has populated it */
4185 if (!pgd_present(*p4d)) {
4187 pgd_populate(mm, p4d, new);
4188 } else /* Another has populated it */
4190 #endif /* __ARCH_HAS_5LEVEL_HACK */
4191 spin_unlock(&mm->page_table_lock);
4194 #endif /* __PAGETABLE_PUD_FOLDED */
4196 #ifndef __PAGETABLE_PMD_FOLDED
4198 * Allocate page middle directory.
4199 * We've already handled the fast-path in-line.
4201 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4204 pmd_t *new = pmd_alloc_one(mm, address);
4208 smp_wmb(); /* See comment in __pte_alloc */
4210 ptl = pud_lock(mm, pud);
4211 #ifndef __ARCH_HAS_4LEVEL_HACK
4212 if (!pud_present(*pud)) {
4214 pud_populate(mm, pud, new);
4215 } else /* Another has populated it */
4218 if (!pgd_present(*pud)) {
4220 pgd_populate(mm, pud, new);
4221 } else /* Another has populated it */
4223 #endif /* __ARCH_HAS_4LEVEL_HACK */
4227 #endif /* __PAGETABLE_PMD_FOLDED */
4229 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4230 unsigned long *start, unsigned long *end,
4231 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4239 pgd = pgd_offset(mm, address);
4240 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4243 p4d = p4d_offset(pgd, address);
4244 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4247 pud = pud_offset(p4d, address);
4248 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4251 pmd = pmd_offset(pud, address);
4252 VM_BUG_ON(pmd_trans_huge(*pmd));
4254 if (pmd_huge(*pmd)) {
4259 *start = address & PMD_MASK;
4260 *end = *start + PMD_SIZE;
4261 mmu_notifier_invalidate_range_start(mm, *start, *end);
4263 *ptlp = pmd_lock(mm, pmd);
4264 if (pmd_huge(*pmd)) {
4270 mmu_notifier_invalidate_range_end(mm, *start, *end);
4273 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4277 *start = address & PAGE_MASK;
4278 *end = *start + PAGE_SIZE;
4279 mmu_notifier_invalidate_range_start(mm, *start, *end);
4281 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4282 if (!pte_present(*ptep))
4287 pte_unmap_unlock(ptep, *ptlp);
4289 mmu_notifier_invalidate_range_end(mm, *start, *end);
4294 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4295 pte_t **ptepp, spinlock_t **ptlp)
4299 /* (void) is needed to make gcc happy */
4300 (void) __cond_lock(*ptlp,
4301 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4302 ptepp, NULL, ptlp)));
4306 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4307 unsigned long *start, unsigned long *end,
4308 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4312 /* (void) is needed to make gcc happy */
4313 (void) __cond_lock(*ptlp,
4314 !(res = __follow_pte_pmd(mm, address, start, end,
4315 ptepp, pmdpp, ptlp)));
4318 EXPORT_SYMBOL(follow_pte_pmd);
4321 * follow_pfn - look up PFN at a user virtual address
4322 * @vma: memory mapping
4323 * @address: user virtual address
4324 * @pfn: location to store found PFN
4326 * Only IO mappings and raw PFN mappings are allowed.
4328 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4330 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4337 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4340 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4343 *pfn = pte_pfn(*ptep);
4344 pte_unmap_unlock(ptep, ptl);
4347 EXPORT_SYMBOL(follow_pfn);
4349 #ifdef CONFIG_HAVE_IOREMAP_PROT
4350 int follow_phys(struct vm_area_struct *vma,
4351 unsigned long address, unsigned int flags,
4352 unsigned long *prot, resource_size_t *phys)
4358 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4361 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4365 if ((flags & FOLL_WRITE) && !pte_write(pte))
4368 *prot = pgprot_val(pte_pgprot(pte));
4369 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4373 pte_unmap_unlock(ptep, ptl);
4378 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4379 void *buf, int len, int write)
4381 resource_size_t phys_addr;
4382 unsigned long prot = 0;
4383 void __iomem *maddr;
4384 int offset = addr & (PAGE_SIZE-1);
4386 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4389 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4391 memcpy_toio(maddr + offset, buf, len);
4393 memcpy_fromio(buf, maddr + offset, len);
4398 EXPORT_SYMBOL_GPL(generic_access_phys);
4402 * Access another process' address space as given in mm. If non-NULL, use the
4403 * given task for page fault accounting.
4405 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4406 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4408 struct vm_area_struct *vma;
4409 void *old_buf = buf;
4410 int write = gup_flags & FOLL_WRITE;
4412 down_read(&mm->mmap_sem);
4413 /* ignore errors, just check how much was successfully transferred */
4415 int bytes, ret, offset;
4417 struct page *page = NULL;
4419 ret = get_user_pages_remote(tsk, mm, addr, 1,
4420 gup_flags, &page, &vma, NULL);
4422 #ifndef CONFIG_HAVE_IOREMAP_PROT
4426 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4427 * we can access using slightly different code.
4429 vma = find_vma(mm, addr);
4430 if (!vma || vma->vm_start > addr)
4432 if (vma->vm_ops && vma->vm_ops->access)
4433 ret = vma->vm_ops->access(vma, addr, buf,
4441 offset = addr & (PAGE_SIZE-1);
4442 if (bytes > PAGE_SIZE-offset)
4443 bytes = PAGE_SIZE-offset;
4447 copy_to_user_page(vma, page, addr,
4448 maddr + offset, buf, bytes);
4449 set_page_dirty_lock(page);
4451 copy_from_user_page(vma, page, addr,
4452 buf, maddr + offset, bytes);
4461 up_read(&mm->mmap_sem);
4463 return buf - old_buf;
4467 * access_remote_vm - access another process' address space
4468 * @mm: the mm_struct of the target address space
4469 * @addr: start address to access
4470 * @buf: source or destination buffer
4471 * @len: number of bytes to transfer
4472 * @gup_flags: flags modifying lookup behaviour
4474 * The caller must hold a reference on @mm.
4476 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4477 void *buf, int len, unsigned int gup_flags)
4479 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4483 * Access another process' address space.
4484 * Source/target buffer must be kernel space,
4485 * Do not walk the page table directly, use get_user_pages
4487 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4488 void *buf, int len, unsigned int gup_flags)
4490 struct mm_struct *mm;
4493 mm = get_task_mm(tsk);
4497 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4503 EXPORT_SYMBOL_GPL(access_process_vm);
4506 * Print the name of a VMA.
4508 void print_vma_addr(char *prefix, unsigned long ip)
4510 struct mm_struct *mm = current->mm;
4511 struct vm_area_struct *vma;
4514 * we might be running from an atomic context so we cannot sleep
4516 if (!down_read_trylock(&mm->mmap_sem))
4519 vma = find_vma(mm, ip);
4520 if (vma && vma->vm_file) {
4521 struct file *f = vma->vm_file;
4522 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4526 p = file_path(f, buf, PAGE_SIZE);
4529 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4531 vma->vm_end - vma->vm_start);
4532 free_page((unsigned long)buf);
4535 up_read(&mm->mmap_sem);
4538 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4539 void __might_fault(const char *file, int line)
4542 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4543 * holding the mmap_sem, this is safe because kernel memory doesn't
4544 * get paged out, therefore we'll never actually fault, and the
4545 * below annotations will generate false positives.
4547 if (uaccess_kernel())
4549 if (pagefault_disabled())
4551 __might_sleep(file, line, 0);
4552 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4554 might_lock_read(¤t->mm->mmap_sem);
4557 EXPORT_SYMBOL(__might_fault);
4560 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4561 static void clear_gigantic_page(struct page *page,
4563 unsigned int pages_per_huge_page)
4566 struct page *p = page;
4569 for (i = 0; i < pages_per_huge_page;
4570 i++, p = mem_map_next(p, page, i)) {
4572 clear_user_highpage(p, addr + i * PAGE_SIZE);
4575 void clear_huge_page(struct page *page,
4576 unsigned long addr_hint, unsigned int pages_per_huge_page)
4579 unsigned long addr = addr_hint &
4580 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4582 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4583 clear_gigantic_page(page, addr, pages_per_huge_page);
4587 /* Clear sub-page to access last to keep its cache lines hot */
4589 n = (addr_hint - addr) / PAGE_SIZE;
4590 if (2 * n <= pages_per_huge_page) {
4591 /* If sub-page to access in first half of huge page */
4594 /* Clear sub-pages at the end of huge page */
4595 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4597 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4600 /* If sub-page to access in second half of huge page */
4601 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4602 l = pages_per_huge_page - n;
4603 /* Clear sub-pages at the begin of huge page */
4604 for (i = 0; i < base; i++) {
4606 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4610 * Clear remaining sub-pages in left-right-left-right pattern
4611 * towards the sub-page to access
4613 for (i = 0; i < l; i++) {
4614 int left_idx = base + i;
4615 int right_idx = base + 2 * l - 1 - i;
4618 clear_user_highpage(page + left_idx,
4619 addr + left_idx * PAGE_SIZE);
4621 clear_user_highpage(page + right_idx,
4622 addr + right_idx * PAGE_SIZE);
4626 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4628 struct vm_area_struct *vma,
4629 unsigned int pages_per_huge_page)
4632 struct page *dst_base = dst;
4633 struct page *src_base = src;
4635 for (i = 0; i < pages_per_huge_page; ) {
4637 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4640 dst = mem_map_next(dst, dst_base, i);
4641 src = mem_map_next(src, src_base, i);
4645 void copy_user_huge_page(struct page *dst, struct page *src,
4646 unsigned long addr, struct vm_area_struct *vma,
4647 unsigned int pages_per_huge_page)
4651 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4652 copy_user_gigantic_page(dst, src, addr, vma,
4653 pages_per_huge_page);
4658 for (i = 0; i < pages_per_huge_page; i++) {
4660 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4664 long copy_huge_page_from_user(struct page *dst_page,
4665 const void __user *usr_src,
4666 unsigned int pages_per_huge_page,
4667 bool allow_pagefault)
4669 void *src = (void *)usr_src;
4671 unsigned long i, rc = 0;
4672 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4674 for (i = 0; i < pages_per_huge_page; i++) {
4675 if (allow_pagefault)
4676 page_kaddr = kmap(dst_page + i);
4678 page_kaddr = kmap_atomic(dst_page + i);
4679 rc = copy_from_user(page_kaddr,
4680 (const void __user *)(src + i * PAGE_SIZE),
4682 if (allow_pagefault)
4683 kunmap(dst_page + i);
4685 kunmap_atomic(page_kaddr);
4687 ret_val -= (PAGE_SIZE - rc);
4695 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4697 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4699 static struct kmem_cache *page_ptl_cachep;
4701 void __init ptlock_cache_init(void)
4703 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4707 bool ptlock_alloc(struct page *page)
4711 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4718 void ptlock_free(struct page *page)
4720 kmem_cache_free(page_ptl_cachep, page->ptl);