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.
1663 * Returns 0 if successful.
1665 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1668 if (address < vma->vm_start || address + size > vma->vm_end ||
1669 !(vma->vm_flags & VM_PFNMAP))
1671 zap_page_range_single(vma, address, size, NULL);
1674 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1676 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1684 pgd = pgd_offset(mm, addr);
1685 p4d = p4d_alloc(mm, pgd, addr);
1688 pud = pud_alloc(mm, p4d, addr);
1691 pmd = pmd_alloc(mm, pud, addr);
1695 VM_BUG_ON(pmd_trans_huge(*pmd));
1696 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1700 * This is the old fallback for page remapping.
1702 * For historical reasons, it only allows reserved pages. Only
1703 * old drivers should use this, and they needed to mark their
1704 * pages reserved for the old functions anyway.
1706 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1707 struct page *page, pgprot_t prot)
1709 struct mm_struct *mm = vma->vm_mm;
1718 flush_dcache_page(page);
1719 pte = get_locked_pte(mm, addr, &ptl);
1723 if (!pte_none(*pte))
1726 /* Ok, finally just insert the thing.. */
1728 inc_mm_counter_fast(mm, mm_counter_file(page));
1729 page_add_file_rmap(page, false);
1730 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1733 pte_unmap_unlock(pte, ptl);
1736 pte_unmap_unlock(pte, ptl);
1742 * vm_insert_page - insert single page into user vma
1743 * @vma: user vma to map to
1744 * @addr: target user address of this page
1745 * @page: source kernel page
1747 * This allows drivers to insert individual pages they've allocated
1750 * The page has to be a nice clean _individual_ kernel allocation.
1751 * If you allocate a compound page, you need to have marked it as
1752 * such (__GFP_COMP), or manually just split the page up yourself
1753 * (see split_page()).
1755 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1756 * took an arbitrary page protection parameter. This doesn't allow
1757 * that. Your vma protection will have to be set up correctly, which
1758 * means that if you want a shared writable mapping, you'd better
1759 * ask for a shared writable mapping!
1761 * The page does not need to be reserved.
1763 * Usually this function is called from f_op->mmap() handler
1764 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1765 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1766 * function from other places, for example from page-fault handler.
1768 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1771 if (addr < vma->vm_start || addr >= vma->vm_end)
1773 if (!page_count(page))
1775 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1776 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1777 BUG_ON(vma->vm_flags & VM_PFNMAP);
1778 vma->vm_flags |= VM_MIXEDMAP;
1780 return insert_page(vma, addr, page, vma->vm_page_prot);
1782 EXPORT_SYMBOL(vm_insert_page);
1784 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1785 pfn_t pfn, pgprot_t prot, bool mkwrite)
1787 struct mm_struct *mm = vma->vm_mm;
1793 pte = get_locked_pte(mm, addr, &ptl);
1797 if (!pte_none(*pte)) {
1800 * For read faults on private mappings the PFN passed
1801 * in may not match the PFN we have mapped if the
1802 * mapped PFN is a writeable COW page. In the mkwrite
1803 * case we are creating a writable PTE for a shared
1804 * mapping and we expect the PFNs to match.
1806 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1814 /* Ok, finally just insert the thing.. */
1815 if (pfn_t_devmap(pfn))
1816 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1818 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1822 entry = pte_mkyoung(entry);
1823 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1826 set_pte_at(mm, addr, pte, entry);
1827 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1831 pte_unmap_unlock(pte, ptl);
1837 * vm_insert_pfn - insert single pfn into user vma
1838 * @vma: user vma to map to
1839 * @addr: target user address of this page
1840 * @pfn: source kernel pfn
1842 * Similar to vm_insert_page, this allows drivers to insert individual pages
1843 * they've allocated into a user vma. Same comments apply.
1845 * This function should only be called from a vm_ops->fault handler, and
1846 * in that case the handler should return NULL.
1848 * vma cannot be a COW mapping.
1850 * As this is called only for pages that do not currently exist, we
1851 * do not need to flush old virtual caches or the TLB.
1853 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1856 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1858 EXPORT_SYMBOL(vm_insert_pfn);
1861 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1862 * @vma: user vma to map to
1863 * @addr: target user address of this page
1864 * @pfn: source kernel pfn
1865 * @pgprot: pgprot flags for the inserted page
1867 * This is exactly like vm_insert_pfn, except that it allows drivers to
1868 * to override pgprot on a per-page basis.
1870 * This only makes sense for IO mappings, and it makes no sense for
1871 * cow mappings. In general, using multiple vmas is preferable;
1872 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1875 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1876 unsigned long pfn, pgprot_t pgprot)
1880 * Technically, architectures with pte_special can avoid all these
1881 * restrictions (same for remap_pfn_range). However we would like
1882 * consistency in testing and feature parity among all, so we should
1883 * try to keep these invariants in place for everybody.
1885 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1886 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1887 (VM_PFNMAP|VM_MIXEDMAP));
1888 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1889 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1891 if (addr < vma->vm_start || addr >= vma->vm_end)
1894 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1896 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1901 EXPORT_SYMBOL(vm_insert_pfn_prot);
1903 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1905 /* these checks mirror the abort conditions in vm_normal_page */
1906 if (vma->vm_flags & VM_MIXEDMAP)
1908 if (pfn_t_devmap(pfn))
1910 if (pfn_t_special(pfn))
1912 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1917 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918 pfn_t pfn, bool mkwrite)
1920 pgprot_t pgprot = vma->vm_page_prot;
1922 BUG_ON(!vm_mixed_ok(vma, pfn));
1924 if (addr < vma->vm_start || addr >= vma->vm_end)
1927 track_pfn_insert(vma, &pgprot, pfn);
1930 * If we don't have pte special, then we have to use the pfn_valid()
1931 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1932 * refcount the page if pfn_valid is true (hence insert_page rather
1933 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1934 * without pte special, it would there be refcounted as a normal page.
1936 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1940 * At this point we are committed to insert_page()
1941 * regardless of whether the caller specified flags that
1942 * result in pfn_t_has_page() == false.
1944 page = pfn_to_page(pfn_t_to_pfn(pfn));
1945 return insert_page(vma, addr, page, pgprot);
1947 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1950 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1953 return __vm_insert_mixed(vma, addr, pfn, false);
1956 EXPORT_SYMBOL(vm_insert_mixed);
1958 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1961 return __vm_insert_mixed(vma, addr, pfn, true);
1963 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1966 * maps a range of physical memory into the requested pages. the old
1967 * mappings are removed. any references to nonexistent pages results
1968 * in null mappings (currently treated as "copy-on-access")
1970 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1971 unsigned long addr, unsigned long end,
1972 unsigned long pfn, pgprot_t prot)
1977 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1980 arch_enter_lazy_mmu_mode();
1982 BUG_ON(!pte_none(*pte));
1983 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1985 } while (pte++, addr += PAGE_SIZE, addr != end);
1986 arch_leave_lazy_mmu_mode();
1987 pte_unmap_unlock(pte - 1, ptl);
1991 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1992 unsigned long addr, unsigned long end,
1993 unsigned long pfn, pgprot_t prot)
1998 pfn -= addr >> PAGE_SHIFT;
1999 pmd = pmd_alloc(mm, pud, addr);
2002 VM_BUG_ON(pmd_trans_huge(*pmd));
2004 next = pmd_addr_end(addr, end);
2005 if (remap_pte_range(mm, pmd, addr, next,
2006 pfn + (addr >> PAGE_SHIFT), prot))
2008 } while (pmd++, addr = next, addr != end);
2012 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2013 unsigned long addr, unsigned long end,
2014 unsigned long pfn, pgprot_t prot)
2019 pfn -= addr >> PAGE_SHIFT;
2020 pud = pud_alloc(mm, p4d, addr);
2024 next = pud_addr_end(addr, end);
2025 if (remap_pmd_range(mm, pud, addr, next,
2026 pfn + (addr >> PAGE_SHIFT), prot))
2028 } while (pud++, addr = next, addr != end);
2032 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2033 unsigned long addr, unsigned long end,
2034 unsigned long pfn, pgprot_t prot)
2039 pfn -= addr >> PAGE_SHIFT;
2040 p4d = p4d_alloc(mm, pgd, addr);
2044 next = p4d_addr_end(addr, end);
2045 if (remap_pud_range(mm, p4d, addr, next,
2046 pfn + (addr >> PAGE_SHIFT), prot))
2048 } while (p4d++, addr = next, addr != end);
2053 * remap_pfn_range - remap kernel memory to userspace
2054 * @vma: user vma to map to
2055 * @addr: target user address to start at
2056 * @pfn: physical address of kernel memory
2057 * @size: size of map area
2058 * @prot: page protection flags for this mapping
2060 * Note: this is only safe if the mm semaphore is held when called.
2062 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2063 unsigned long pfn, unsigned long size, pgprot_t prot)
2067 unsigned long end = addr + PAGE_ALIGN(size);
2068 struct mm_struct *mm = vma->vm_mm;
2069 unsigned long remap_pfn = pfn;
2073 * Physically remapped pages are special. Tell the
2074 * rest of the world about it:
2075 * VM_IO tells people not to look at these pages
2076 * (accesses can have side effects).
2077 * VM_PFNMAP tells the core MM that the base pages are just
2078 * raw PFN mappings, and do not have a "struct page" associated
2081 * Disable vma merging and expanding with mremap().
2083 * Omit vma from core dump, even when VM_IO turned off.
2085 * There's a horrible special case to handle copy-on-write
2086 * behaviour that some programs depend on. We mark the "original"
2087 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2088 * See vm_normal_page() for details.
2090 if (is_cow_mapping(vma->vm_flags)) {
2091 if (addr != vma->vm_start || end != vma->vm_end)
2093 vma->vm_pgoff = pfn;
2096 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2100 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2102 BUG_ON(addr >= end);
2103 pfn -= addr >> PAGE_SHIFT;
2104 pgd = pgd_offset(mm, addr);
2105 flush_cache_range(vma, addr, end);
2107 next = pgd_addr_end(addr, end);
2108 err = remap_p4d_range(mm, pgd, addr, next,
2109 pfn + (addr >> PAGE_SHIFT), prot);
2112 } while (pgd++, addr = next, addr != end);
2115 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2119 EXPORT_SYMBOL(remap_pfn_range);
2122 * vm_iomap_memory - remap memory to userspace
2123 * @vma: user vma to map to
2124 * @start: start of area
2125 * @len: size of area
2127 * This is a simplified io_remap_pfn_range() for common driver use. The
2128 * driver just needs to give us the physical memory range to be mapped,
2129 * we'll figure out the rest from the vma information.
2131 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2132 * whatever write-combining details or similar.
2134 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2136 unsigned long vm_len, pfn, pages;
2138 /* Check that the physical memory area passed in looks valid */
2139 if (start + len < start)
2142 * You *really* shouldn't map things that aren't page-aligned,
2143 * but we've historically allowed it because IO memory might
2144 * just have smaller alignment.
2146 len += start & ~PAGE_MASK;
2147 pfn = start >> PAGE_SHIFT;
2148 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2149 if (pfn + pages < pfn)
2152 /* We start the mapping 'vm_pgoff' pages into the area */
2153 if (vma->vm_pgoff > pages)
2155 pfn += vma->vm_pgoff;
2156 pages -= vma->vm_pgoff;
2158 /* Can we fit all of the mapping? */
2159 vm_len = vma->vm_end - vma->vm_start;
2160 if (vm_len >> PAGE_SHIFT > pages)
2163 /* Ok, let it rip */
2164 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2166 EXPORT_SYMBOL(vm_iomap_memory);
2168 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2169 unsigned long addr, unsigned long end,
2170 pte_fn_t fn, void *data)
2175 spinlock_t *uninitialized_var(ptl);
2177 pte = (mm == &init_mm) ?
2178 pte_alloc_kernel(pmd, addr) :
2179 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2183 BUG_ON(pmd_huge(*pmd));
2185 arch_enter_lazy_mmu_mode();
2187 token = pmd_pgtable(*pmd);
2190 err = fn(pte++, token, addr, data);
2193 } while (addr += PAGE_SIZE, addr != end);
2195 arch_leave_lazy_mmu_mode();
2198 pte_unmap_unlock(pte-1, ptl);
2202 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2203 unsigned long addr, unsigned long end,
2204 pte_fn_t fn, void *data)
2210 BUG_ON(pud_huge(*pud));
2212 pmd = pmd_alloc(mm, pud, addr);
2216 next = pmd_addr_end(addr, end);
2217 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2220 } while (pmd++, addr = next, addr != end);
2224 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2225 unsigned long addr, unsigned long end,
2226 pte_fn_t fn, void *data)
2232 pud = pud_alloc(mm, p4d, addr);
2236 next = pud_addr_end(addr, end);
2237 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2240 } while (pud++, addr = next, addr != end);
2244 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2245 unsigned long addr, unsigned long end,
2246 pte_fn_t fn, void *data)
2252 p4d = p4d_alloc(mm, pgd, addr);
2256 next = p4d_addr_end(addr, end);
2257 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2260 } while (p4d++, addr = next, addr != end);
2265 * Scan a region of virtual memory, filling in page tables as necessary
2266 * and calling a provided function on each leaf page table.
2268 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2269 unsigned long size, pte_fn_t fn, void *data)
2273 unsigned long end = addr + size;
2276 if (WARN_ON(addr >= end))
2279 pgd = pgd_offset(mm, addr);
2281 next = pgd_addr_end(addr, end);
2282 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2285 } while (pgd++, addr = next, addr != end);
2289 EXPORT_SYMBOL_GPL(apply_to_page_range);
2292 * handle_pte_fault chooses page fault handler according to an entry which was
2293 * read non-atomically. Before making any commitment, on those architectures
2294 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2295 * parts, do_swap_page must check under lock before unmapping the pte and
2296 * proceeding (but do_wp_page is only called after already making such a check;
2297 * and do_anonymous_page can safely check later on).
2299 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2300 pte_t *page_table, pte_t orig_pte)
2303 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2304 if (sizeof(pte_t) > sizeof(unsigned long)) {
2305 spinlock_t *ptl = pte_lockptr(mm, pmd);
2307 same = pte_same(*page_table, orig_pte);
2311 pte_unmap(page_table);
2315 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2317 debug_dma_assert_idle(src);
2320 * If the source page was a PFN mapping, we don't have
2321 * a "struct page" for it. We do a best-effort copy by
2322 * just copying from the original user address. If that
2323 * fails, we just zero-fill it. Live with it.
2325 if (unlikely(!src)) {
2326 void *kaddr = kmap_atomic(dst);
2327 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2330 * This really shouldn't fail, because the page is there
2331 * in the page tables. But it might just be unreadable,
2332 * in which case we just give up and fill the result with
2335 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2337 kunmap_atomic(kaddr);
2338 flush_dcache_page(dst);
2340 copy_user_highpage(dst, src, va, vma);
2343 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2345 struct file *vm_file = vma->vm_file;
2348 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2351 * Special mappings (e.g. VDSO) do not have any file so fake
2352 * a default GFP_KERNEL for them.
2358 * Notify the address space that the page is about to become writable so that
2359 * it can prohibit this or wait for the page to get into an appropriate state.
2361 * We do this without the lock held, so that it can sleep if it needs to.
2363 static int do_page_mkwrite(struct vm_fault *vmf)
2366 struct page *page = vmf->page;
2367 unsigned int old_flags = vmf->flags;
2369 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2371 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2372 /* Restore original flags so that caller is not surprised */
2373 vmf->flags = old_flags;
2374 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2376 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2378 if (!page->mapping) {
2380 return 0; /* retry */
2382 ret |= VM_FAULT_LOCKED;
2384 VM_BUG_ON_PAGE(!PageLocked(page), page);
2389 * Handle dirtying of a page in shared file mapping on a write fault.
2391 * The function expects the page to be locked and unlocks it.
2393 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2396 struct address_space *mapping;
2398 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2400 dirtied = set_page_dirty(page);
2401 VM_BUG_ON_PAGE(PageAnon(page), page);
2403 * Take a local copy of the address_space - page.mapping may be zeroed
2404 * by truncate after unlock_page(). The address_space itself remains
2405 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2406 * release semantics to prevent the compiler from undoing this copying.
2408 mapping = page_rmapping(page);
2411 if ((dirtied || page_mkwrite) && mapping) {
2413 * Some device drivers do not set page.mapping
2414 * but still dirty their pages
2416 balance_dirty_pages_ratelimited(mapping);
2420 file_update_time(vma->vm_file);
2424 * Handle write page faults for pages that can be reused in the current vma
2426 * This can happen either due to the mapping being with the VM_SHARED flag,
2427 * or due to us being the last reference standing to the page. In either
2428 * case, all we need to do here is to mark the page as writable and update
2429 * any related book-keeping.
2431 static inline void wp_page_reuse(struct vm_fault *vmf)
2432 __releases(vmf->ptl)
2434 struct vm_area_struct *vma = vmf->vma;
2435 struct page *page = vmf->page;
2438 * Clear the pages cpupid information as the existing
2439 * information potentially belongs to a now completely
2440 * unrelated process.
2443 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2445 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2446 entry = pte_mkyoung(vmf->orig_pte);
2447 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2448 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2449 update_mmu_cache(vma, vmf->address, vmf->pte);
2450 pte_unmap_unlock(vmf->pte, vmf->ptl);
2454 * Handle the case of a page which we actually need to copy to a new page.
2456 * Called with mmap_sem locked and the old page referenced, but
2457 * without the ptl held.
2459 * High level logic flow:
2461 * - Allocate a page, copy the content of the old page to the new one.
2462 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2463 * - Take the PTL. If the pte changed, bail out and release the allocated page
2464 * - If the pte is still the way we remember it, update the page table and all
2465 * relevant references. This includes dropping the reference the page-table
2466 * held to the old page, as well as updating the rmap.
2467 * - In any case, unlock the PTL and drop the reference we took to the old page.
2469 static int wp_page_copy(struct vm_fault *vmf)
2471 struct vm_area_struct *vma = vmf->vma;
2472 struct mm_struct *mm = vma->vm_mm;
2473 struct page *old_page = vmf->page;
2474 struct page *new_page = NULL;
2476 int page_copied = 0;
2477 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2478 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2479 struct mem_cgroup *memcg;
2481 if (unlikely(anon_vma_prepare(vma)))
2484 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2485 new_page = alloc_zeroed_user_highpage_movable(vma,
2490 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2494 cow_user_page(new_page, old_page, vmf->address, vma);
2497 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2500 __SetPageUptodate(new_page);
2502 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2505 * Re-check the pte - we dropped the lock
2507 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2508 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2510 if (!PageAnon(old_page)) {
2511 dec_mm_counter_fast(mm,
2512 mm_counter_file(old_page));
2513 inc_mm_counter_fast(mm, MM_ANONPAGES);
2516 inc_mm_counter_fast(mm, MM_ANONPAGES);
2518 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2519 entry = mk_pte(new_page, vma->vm_page_prot);
2520 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2522 * Clear the pte entry and flush it first, before updating the
2523 * pte with the new entry. This will avoid a race condition
2524 * seen in the presence of one thread doing SMC and another
2527 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2528 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2529 mem_cgroup_commit_charge(new_page, memcg, false, false);
2530 lru_cache_add_active_or_unevictable(new_page, vma);
2532 * We call the notify macro here because, when using secondary
2533 * mmu page tables (such as kvm shadow page tables), we want the
2534 * new page to be mapped directly into the secondary page table.
2536 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2537 update_mmu_cache(vma, vmf->address, vmf->pte);
2540 * Only after switching the pte to the new page may
2541 * we remove the mapcount here. Otherwise another
2542 * process may come and find the rmap count decremented
2543 * before the pte is switched to the new page, and
2544 * "reuse" the old page writing into it while our pte
2545 * here still points into it and can be read by other
2548 * The critical issue is to order this
2549 * page_remove_rmap with the ptp_clear_flush above.
2550 * Those stores are ordered by (if nothing else,)
2551 * the barrier present in the atomic_add_negative
2552 * in page_remove_rmap.
2554 * Then the TLB flush in ptep_clear_flush ensures that
2555 * no process can access the old page before the
2556 * decremented mapcount is visible. And the old page
2557 * cannot be reused until after the decremented
2558 * mapcount is visible. So transitively, TLBs to
2559 * old page will be flushed before it can be reused.
2561 page_remove_rmap(old_page, false);
2564 /* Free the old page.. */
2565 new_page = old_page;
2568 mem_cgroup_cancel_charge(new_page, memcg, false);
2574 pte_unmap_unlock(vmf->pte, vmf->ptl);
2576 * No need to double call mmu_notifier->invalidate_range() callback as
2577 * the above ptep_clear_flush_notify() did already call it.
2579 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2582 * Don't let another task, with possibly unlocked vma,
2583 * keep the mlocked page.
2585 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2586 lock_page(old_page); /* LRU manipulation */
2587 if (PageMlocked(old_page))
2588 munlock_vma_page(old_page);
2589 unlock_page(old_page);
2593 return page_copied ? VM_FAULT_WRITE : 0;
2599 return VM_FAULT_OOM;
2603 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2604 * writeable once the page is prepared
2606 * @vmf: structure describing the fault
2608 * This function handles all that is needed to finish a write page fault in a
2609 * shared mapping due to PTE being read-only once the mapped page is prepared.
2610 * It handles locking of PTE and modifying it. The function returns
2611 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2614 * The function expects the page to be locked or other protection against
2615 * concurrent faults / writeback (such as DAX radix tree locks).
2617 int finish_mkwrite_fault(struct vm_fault *vmf)
2619 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2620 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2623 * We might have raced with another page fault while we released the
2624 * pte_offset_map_lock.
2626 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2627 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628 return VM_FAULT_NOPAGE;
2635 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2638 static int wp_pfn_shared(struct vm_fault *vmf)
2640 struct vm_area_struct *vma = vmf->vma;
2642 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2645 pte_unmap_unlock(vmf->pte, vmf->ptl);
2646 vmf->flags |= FAULT_FLAG_MKWRITE;
2647 ret = vma->vm_ops->pfn_mkwrite(vmf);
2648 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2650 return finish_mkwrite_fault(vmf);
2653 return VM_FAULT_WRITE;
2656 static int wp_page_shared(struct vm_fault *vmf)
2657 __releases(vmf->ptl)
2659 struct vm_area_struct *vma = vmf->vma;
2661 get_page(vmf->page);
2663 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2666 pte_unmap_unlock(vmf->pte, vmf->ptl);
2667 tmp = do_page_mkwrite(vmf);
2668 if (unlikely(!tmp || (tmp &
2669 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2670 put_page(vmf->page);
2673 tmp = finish_mkwrite_fault(vmf);
2674 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2675 unlock_page(vmf->page);
2676 put_page(vmf->page);
2681 lock_page(vmf->page);
2683 fault_dirty_shared_page(vma, vmf->page);
2684 put_page(vmf->page);
2686 return VM_FAULT_WRITE;
2690 * This routine handles present pages, when users try to write
2691 * to a shared page. It is done by copying the page to a new address
2692 * and decrementing the shared-page counter for the old page.
2694 * Note that this routine assumes that the protection checks have been
2695 * done by the caller (the low-level page fault routine in most cases).
2696 * Thus we can safely just mark it writable once we've done any necessary
2699 * We also mark the page dirty at this point even though the page will
2700 * change only once the write actually happens. This avoids a few races,
2701 * and potentially makes it more efficient.
2703 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2704 * but allow concurrent faults), with pte both mapped and locked.
2705 * We return with mmap_sem still held, but pte unmapped and unlocked.
2707 static int do_wp_page(struct vm_fault *vmf)
2708 __releases(vmf->ptl)
2710 struct vm_area_struct *vma = vmf->vma;
2712 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2715 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2718 * We should not cow pages in a shared writeable mapping.
2719 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2721 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2722 (VM_WRITE|VM_SHARED))
2723 return wp_pfn_shared(vmf);
2725 pte_unmap_unlock(vmf->pte, vmf->ptl);
2726 return wp_page_copy(vmf);
2730 * Take out anonymous pages first, anonymous shared vmas are
2731 * not dirty accountable.
2733 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2734 int total_map_swapcount;
2735 if (!trylock_page(vmf->page)) {
2736 get_page(vmf->page);
2737 pte_unmap_unlock(vmf->pte, vmf->ptl);
2738 lock_page(vmf->page);
2739 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2740 vmf->address, &vmf->ptl);
2741 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2742 unlock_page(vmf->page);
2743 pte_unmap_unlock(vmf->pte, vmf->ptl);
2744 put_page(vmf->page);
2747 put_page(vmf->page);
2749 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2750 if (total_map_swapcount == 1) {
2752 * The page is all ours. Move it to
2753 * our anon_vma so the rmap code will
2754 * not search our parent or siblings.
2755 * Protected against the rmap code by
2758 page_move_anon_rmap(vmf->page, vma);
2760 unlock_page(vmf->page);
2762 return VM_FAULT_WRITE;
2764 unlock_page(vmf->page);
2765 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2766 (VM_WRITE|VM_SHARED))) {
2767 return wp_page_shared(vmf);
2771 * Ok, we need to copy. Oh, well..
2773 get_page(vmf->page);
2775 pte_unmap_unlock(vmf->pte, vmf->ptl);
2776 return wp_page_copy(vmf);
2779 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2780 unsigned long start_addr, unsigned long end_addr,
2781 struct zap_details *details)
2783 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2786 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2787 struct zap_details *details)
2789 struct vm_area_struct *vma;
2790 pgoff_t vba, vea, zba, zea;
2792 vma_interval_tree_foreach(vma, root,
2793 details->first_index, details->last_index) {
2795 vba = vma->vm_pgoff;
2796 vea = vba + vma_pages(vma) - 1;
2797 zba = details->first_index;
2800 zea = details->last_index;
2804 unmap_mapping_range_vma(vma,
2805 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2806 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2812 * unmap_mapping_pages() - Unmap pages from processes.
2813 * @mapping: The address space containing pages to be unmapped.
2814 * @start: Index of first page to be unmapped.
2815 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2816 * @even_cows: Whether to unmap even private COWed pages.
2818 * Unmap the pages in this address space from any userspace process which
2819 * has them mmaped. Generally, you want to remove COWed pages as well when
2820 * a file is being truncated, but not when invalidating pages from the page
2823 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2824 pgoff_t nr, bool even_cows)
2826 struct zap_details details = { };
2828 details.check_mapping = even_cows ? NULL : mapping;
2829 details.first_index = start;
2830 details.last_index = start + nr - 1;
2831 if (details.last_index < details.first_index)
2832 details.last_index = ULONG_MAX;
2834 i_mmap_lock_write(mapping);
2835 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2836 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2837 i_mmap_unlock_write(mapping);
2841 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2842 * address_space corresponding to the specified byte range in the underlying
2845 * @mapping: the address space containing mmaps to be unmapped.
2846 * @holebegin: byte in first page to unmap, relative to the start of
2847 * the underlying file. This will be rounded down to a PAGE_SIZE
2848 * boundary. Note that this is different from truncate_pagecache(), which
2849 * must keep the partial page. In contrast, we must get rid of
2851 * @holelen: size of prospective hole in bytes. This will be rounded
2852 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2854 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2855 * but 0 when invalidating pagecache, don't throw away private data.
2857 void unmap_mapping_range(struct address_space *mapping,
2858 loff_t const holebegin, loff_t const holelen, int even_cows)
2860 pgoff_t hba = holebegin >> PAGE_SHIFT;
2861 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2863 /* Check for overflow. */
2864 if (sizeof(holelen) > sizeof(hlen)) {
2866 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2867 if (holeend & ~(long long)ULONG_MAX)
2868 hlen = ULONG_MAX - hba + 1;
2871 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2873 EXPORT_SYMBOL(unmap_mapping_range);
2876 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2877 * but allow concurrent faults), and pte mapped but not yet locked.
2878 * We return with pte unmapped and unlocked.
2880 * We return with the mmap_sem locked or unlocked in the same cases
2881 * as does filemap_fault().
2883 int do_swap_page(struct vm_fault *vmf)
2885 struct vm_area_struct *vma = vmf->vma;
2886 struct page *page = NULL, *swapcache = NULL;
2887 struct mem_cgroup *memcg;
2888 struct vma_swap_readahead swap_ra;
2894 bool vma_readahead = swap_use_vma_readahead();
2896 if (vma_readahead) {
2897 page = swap_readahead_detect(vmf, &swap_ra);
2901 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2907 entry = pte_to_swp_entry(vmf->orig_pte);
2908 if (unlikely(non_swap_entry(entry))) {
2909 if (is_migration_entry(entry)) {
2910 migration_entry_wait(vma->vm_mm, vmf->pmd,
2912 } else if (is_device_private_entry(entry)) {
2914 * For un-addressable device memory we call the pgmap
2915 * fault handler callback. The callback must migrate
2916 * the page back to some CPU accessible page.
2918 ret = device_private_entry_fault(vma, vmf->address, entry,
2919 vmf->flags, vmf->pmd);
2920 } else if (is_hwpoison_entry(entry)) {
2921 ret = VM_FAULT_HWPOISON;
2923 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2924 ret = VM_FAULT_SIGBUS;
2930 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2932 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2938 struct swap_info_struct *si = swp_swap_info(entry);
2940 if (si->flags & SWP_SYNCHRONOUS_IO &&
2941 __swap_count(si, entry) == 1) {
2942 /* skip swapcache */
2943 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2945 __SetPageLocked(page);
2946 __SetPageSwapBacked(page);
2947 set_page_private(page, entry.val);
2948 lru_cache_add_anon(page);
2949 swap_readpage(page, true);
2953 page = do_swap_page_readahead(entry,
2954 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2956 page = swapin_readahead(entry,
2957 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2963 * Back out if somebody else faulted in this pte
2964 * while we released the pte lock.
2966 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2967 vmf->address, &vmf->ptl);
2968 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2970 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2974 /* Had to read the page from swap area: Major fault */
2975 ret = VM_FAULT_MAJOR;
2976 count_vm_event(PGMAJFAULT);
2977 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2978 } else if (PageHWPoison(page)) {
2980 * hwpoisoned dirty swapcache pages are kept for killing
2981 * owner processes (which may be unknown at hwpoison time)
2983 ret = VM_FAULT_HWPOISON;
2984 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2989 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2991 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2993 ret |= VM_FAULT_RETRY;
2998 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2999 * release the swapcache from under us. The page pin, and pte_same
3000 * test below, are not enough to exclude that. Even if it is still
3001 * swapcache, we need to check that the page's swap has not changed.
3003 if (unlikely((!PageSwapCache(page) ||
3004 page_private(page) != entry.val)) && swapcache)
3007 page = ksm_might_need_to_copy(page, vma, vmf->address);
3008 if (unlikely(!page)) {
3014 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3021 * Back out if somebody else already faulted in this pte.
3023 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3025 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3028 if (unlikely(!PageUptodate(page))) {
3029 ret = VM_FAULT_SIGBUS;
3034 * The page isn't present yet, go ahead with the fault.
3036 * Be careful about the sequence of operations here.
3037 * To get its accounting right, reuse_swap_page() must be called
3038 * while the page is counted on swap but not yet in mapcount i.e.
3039 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3040 * must be called after the swap_free(), or it will never succeed.
3043 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3044 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3045 pte = mk_pte(page, vma->vm_page_prot);
3046 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3047 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3048 vmf->flags &= ~FAULT_FLAG_WRITE;
3049 ret |= VM_FAULT_WRITE;
3050 exclusive = RMAP_EXCLUSIVE;
3052 flush_icache_page(vma, page);
3053 if (pte_swp_soft_dirty(vmf->orig_pte))
3054 pte = pte_mksoft_dirty(pte);
3055 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3056 vmf->orig_pte = pte;
3058 /* ksm created a completely new copy */
3059 if (unlikely(page != swapcache && swapcache)) {
3060 page_add_new_anon_rmap(page, vma, vmf->address, false);
3061 mem_cgroup_commit_charge(page, memcg, false, false);
3062 lru_cache_add_active_or_unevictable(page, vma);
3064 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3065 mem_cgroup_commit_charge(page, memcg, true, false);
3066 activate_page(page);
3070 if (mem_cgroup_swap_full(page) ||
3071 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3072 try_to_free_swap(page);
3074 if (page != swapcache && swapcache) {
3076 * Hold the lock to avoid the swap entry to be reused
3077 * until we take the PT lock for the pte_same() check
3078 * (to avoid false positives from pte_same). For
3079 * further safety release the lock after the swap_free
3080 * so that the swap count won't change under a
3081 * parallel locked swapcache.
3083 unlock_page(swapcache);
3084 put_page(swapcache);
3087 if (vmf->flags & FAULT_FLAG_WRITE) {
3088 ret |= do_wp_page(vmf);
3089 if (ret & VM_FAULT_ERROR)
3090 ret &= VM_FAULT_ERROR;
3094 /* No need to invalidate - it was non-present before */
3095 update_mmu_cache(vma, vmf->address, vmf->pte);
3097 pte_unmap_unlock(vmf->pte, vmf->ptl);
3101 mem_cgroup_cancel_charge(page, memcg, false);
3102 pte_unmap_unlock(vmf->pte, vmf->ptl);
3107 if (page != swapcache && swapcache) {
3108 unlock_page(swapcache);
3109 put_page(swapcache);
3115 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3116 * but allow concurrent faults), and pte mapped but not yet locked.
3117 * We return with mmap_sem still held, but pte unmapped and unlocked.
3119 static int do_anonymous_page(struct vm_fault *vmf)
3121 struct vm_area_struct *vma = vmf->vma;
3122 struct mem_cgroup *memcg;
3127 /* File mapping without ->vm_ops ? */
3128 if (vma->vm_flags & VM_SHARED)
3129 return VM_FAULT_SIGBUS;
3132 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3133 * pte_offset_map() on pmds where a huge pmd might be created
3134 * from a different thread.
3136 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3137 * parallel threads are excluded by other means.
3139 * Here we only have down_read(mmap_sem).
3141 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3142 return VM_FAULT_OOM;
3144 /* See the comment in pte_alloc_one_map() */
3145 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3148 /* Use the zero-page for reads */
3149 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3150 !mm_forbids_zeropage(vma->vm_mm)) {
3151 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3152 vma->vm_page_prot));
3153 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3154 vmf->address, &vmf->ptl);
3155 if (!pte_none(*vmf->pte))
3157 ret = check_stable_address_space(vma->vm_mm);
3160 /* Deliver the page fault to userland, check inside PT lock */
3161 if (userfaultfd_missing(vma)) {
3162 pte_unmap_unlock(vmf->pte, vmf->ptl);
3163 return handle_userfault(vmf, VM_UFFD_MISSING);
3168 /* Allocate our own private page. */
3169 if (unlikely(anon_vma_prepare(vma)))
3171 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3175 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3179 * The memory barrier inside __SetPageUptodate makes sure that
3180 * preceeding stores to the page contents become visible before
3181 * the set_pte_at() write.
3183 __SetPageUptodate(page);
3185 entry = mk_pte(page, vma->vm_page_prot);
3186 if (vma->vm_flags & VM_WRITE)
3187 entry = pte_mkwrite(pte_mkdirty(entry));
3189 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3191 if (!pte_none(*vmf->pte))
3194 ret = check_stable_address_space(vma->vm_mm);
3198 /* Deliver the page fault to userland, check inside PT lock */
3199 if (userfaultfd_missing(vma)) {
3200 pte_unmap_unlock(vmf->pte, vmf->ptl);
3201 mem_cgroup_cancel_charge(page, memcg, false);
3203 return handle_userfault(vmf, VM_UFFD_MISSING);
3206 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3207 page_add_new_anon_rmap(page, vma, vmf->address, false);
3208 mem_cgroup_commit_charge(page, memcg, false, false);
3209 lru_cache_add_active_or_unevictable(page, vma);
3211 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3213 /* No need to invalidate - it was non-present before */
3214 update_mmu_cache(vma, vmf->address, vmf->pte);
3216 pte_unmap_unlock(vmf->pte, vmf->ptl);
3219 mem_cgroup_cancel_charge(page, memcg, false);
3225 return VM_FAULT_OOM;
3229 * The mmap_sem must have been held on entry, and may have been
3230 * released depending on flags and vma->vm_ops->fault() return value.
3231 * See filemap_fault() and __lock_page_retry().
3233 static int __do_fault(struct vm_fault *vmf)
3235 struct vm_area_struct *vma = vmf->vma;
3238 ret = vma->vm_ops->fault(vmf);
3239 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3240 VM_FAULT_DONE_COW)))
3243 if (unlikely(PageHWPoison(vmf->page))) {
3244 if (ret & VM_FAULT_LOCKED)
3245 unlock_page(vmf->page);
3246 put_page(vmf->page);
3248 return VM_FAULT_HWPOISON;
3251 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3252 lock_page(vmf->page);
3254 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3260 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3261 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3262 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3263 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3265 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3267 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3270 static int pte_alloc_one_map(struct vm_fault *vmf)
3272 struct vm_area_struct *vma = vmf->vma;
3274 if (!pmd_none(*vmf->pmd))
3276 if (vmf->prealloc_pte) {
3277 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3278 if (unlikely(!pmd_none(*vmf->pmd))) {
3279 spin_unlock(vmf->ptl);
3283 mm_inc_nr_ptes(vma->vm_mm);
3284 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3285 spin_unlock(vmf->ptl);
3286 vmf->prealloc_pte = NULL;
3287 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3288 return VM_FAULT_OOM;
3292 * If a huge pmd materialized under us just retry later. Use
3293 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3294 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3295 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3296 * running immediately after a huge pmd fault in a different thread of
3297 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3298 * All we have to ensure is that it is a regular pmd that we can walk
3299 * with pte_offset_map() and we can do that through an atomic read in
3300 * C, which is what pmd_trans_unstable() provides.
3302 if (pmd_devmap_trans_unstable(vmf->pmd))
3303 return VM_FAULT_NOPAGE;
3306 * At this point we know that our vmf->pmd points to a page of ptes
3307 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3308 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3309 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3310 * be valid and we will re-check to make sure the vmf->pte isn't
3311 * pte_none() under vmf->ptl protection when we return to
3314 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3319 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3321 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3322 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3323 unsigned long haddr)
3325 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3326 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3328 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3333 static void deposit_prealloc_pte(struct vm_fault *vmf)
3335 struct vm_area_struct *vma = vmf->vma;
3337 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3339 * We are going to consume the prealloc table,
3340 * count that as nr_ptes.
3342 mm_inc_nr_ptes(vma->vm_mm);
3343 vmf->prealloc_pte = NULL;
3346 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3348 struct vm_area_struct *vma = vmf->vma;
3349 bool write = vmf->flags & FAULT_FLAG_WRITE;
3350 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3354 if (!transhuge_vma_suitable(vma, haddr))
3355 return VM_FAULT_FALLBACK;
3357 ret = VM_FAULT_FALLBACK;
3358 page = compound_head(page);
3361 * Archs like ppc64 need additonal space to store information
3362 * related to pte entry. Use the preallocated table for that.
3364 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3365 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3366 if (!vmf->prealloc_pte)
3367 return VM_FAULT_OOM;
3368 smp_wmb(); /* See comment in __pte_alloc() */
3371 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3372 if (unlikely(!pmd_none(*vmf->pmd)))
3375 for (i = 0; i < HPAGE_PMD_NR; i++)
3376 flush_icache_page(vma, page + i);
3378 entry = mk_huge_pmd(page, vma->vm_page_prot);
3380 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3382 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3383 page_add_file_rmap(page, true);
3385 * deposit and withdraw with pmd lock held
3387 if (arch_needs_pgtable_deposit())
3388 deposit_prealloc_pte(vmf);
3390 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3392 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3394 /* fault is handled */
3396 count_vm_event(THP_FILE_MAPPED);
3398 spin_unlock(vmf->ptl);
3402 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3410 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3411 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3413 * @vmf: fault environment
3414 * @memcg: memcg to charge page (only for private mappings)
3415 * @page: page to map
3417 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3420 * Target users are page handler itself and implementations of
3421 * vm_ops->map_pages.
3423 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3426 struct vm_area_struct *vma = vmf->vma;
3427 bool write = vmf->flags & FAULT_FLAG_WRITE;
3431 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3432 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3434 VM_BUG_ON_PAGE(memcg, page);
3436 ret = do_set_pmd(vmf, page);
3437 if (ret != VM_FAULT_FALLBACK)
3442 ret = pte_alloc_one_map(vmf);
3447 /* Re-check under ptl */
3448 if (unlikely(!pte_none(*vmf->pte)))
3449 return VM_FAULT_NOPAGE;
3451 flush_icache_page(vma, page);
3452 entry = mk_pte(page, vma->vm_page_prot);
3454 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3455 /* copy-on-write page */
3456 if (write && !(vma->vm_flags & VM_SHARED)) {
3457 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3458 page_add_new_anon_rmap(page, vma, vmf->address, false);
3459 mem_cgroup_commit_charge(page, memcg, false, false);
3460 lru_cache_add_active_or_unevictable(page, vma);
3462 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3463 page_add_file_rmap(page, false);
3465 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3467 /* no need to invalidate: a not-present page won't be cached */
3468 update_mmu_cache(vma, vmf->address, vmf->pte);
3475 * finish_fault - finish page fault once we have prepared the page to fault
3477 * @vmf: structure describing the fault
3479 * This function handles all that is needed to finish a page fault once the
3480 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3481 * given page, adds reverse page mapping, handles memcg charges and LRU
3482 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3485 * The function expects the page to be locked and on success it consumes a
3486 * reference of a page being mapped (for the PTE which maps it).
3488 int finish_fault(struct vm_fault *vmf)
3493 /* Did we COW the page? */
3494 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3495 !(vmf->vma->vm_flags & VM_SHARED))
3496 page = vmf->cow_page;
3501 * check even for read faults because we might have lost our CoWed
3504 if (!(vmf->vma->vm_flags & VM_SHARED))
3505 ret = check_stable_address_space(vmf->vma->vm_mm);
3507 ret = alloc_set_pte(vmf, vmf->memcg, page);
3509 pte_unmap_unlock(vmf->pte, vmf->ptl);
3513 static unsigned long fault_around_bytes __read_mostly =
3514 rounddown_pow_of_two(65536);
3516 #ifdef CONFIG_DEBUG_FS
3517 static int fault_around_bytes_get(void *data, u64 *val)
3519 *val = fault_around_bytes;
3524 * fault_around_bytes must be rounded down to the nearest page order as it's
3525 * what do_fault_around() expects to see.
3527 static int fault_around_bytes_set(void *data, u64 val)
3529 if (val / PAGE_SIZE > PTRS_PER_PTE)
3531 if (val > PAGE_SIZE)
3532 fault_around_bytes = rounddown_pow_of_two(val);
3534 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3537 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3538 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3540 static int __init fault_around_debugfs(void)
3544 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3545 &fault_around_bytes_fops);
3547 pr_warn("Failed to create fault_around_bytes in debugfs");
3550 late_initcall(fault_around_debugfs);
3554 * do_fault_around() tries to map few pages around the fault address. The hope
3555 * is that the pages will be needed soon and this will lower the number of
3558 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3559 * not ready to be mapped: not up-to-date, locked, etc.
3561 * This function is called with the page table lock taken. In the split ptlock
3562 * case the page table lock only protects only those entries which belong to
3563 * the page table corresponding to the fault address.
3565 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3568 * fault_around_bytes defines how many bytes we'll try to map.
3569 * do_fault_around() expects it to be set to a power of two less than or equal
3572 * The virtual address of the area that we map is naturally aligned to
3573 * fault_around_bytes rounded down to the machine page size
3574 * (and therefore to page order). This way it's easier to guarantee
3575 * that we don't cross page table boundaries.
3577 static int do_fault_around(struct vm_fault *vmf)
3579 unsigned long address = vmf->address, nr_pages, mask;
3580 pgoff_t start_pgoff = vmf->pgoff;
3584 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3585 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3587 vmf->address = max(address & mask, vmf->vma->vm_start);
3588 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3592 * end_pgoff is either the end of the page table, the end of
3593 * the vma or nr_pages from start_pgoff, depending what is nearest.
3595 end_pgoff = start_pgoff -
3596 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3598 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3599 start_pgoff + nr_pages - 1);
3601 if (pmd_none(*vmf->pmd)) {
3602 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3604 if (!vmf->prealloc_pte)
3606 smp_wmb(); /* See comment in __pte_alloc() */
3609 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3611 /* Huge page is mapped? Page fault is solved */
3612 if (pmd_trans_huge(*vmf->pmd)) {
3613 ret = VM_FAULT_NOPAGE;
3617 /* ->map_pages() haven't done anything useful. Cold page cache? */
3621 /* check if the page fault is solved */
3622 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3623 if (!pte_none(*vmf->pte))
3624 ret = VM_FAULT_NOPAGE;
3625 pte_unmap_unlock(vmf->pte, vmf->ptl);
3627 vmf->address = address;
3632 static int do_read_fault(struct vm_fault *vmf)
3634 struct vm_area_struct *vma = vmf->vma;
3638 * Let's call ->map_pages() first and use ->fault() as fallback
3639 * if page by the offset is not ready to be mapped (cold cache or
3642 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3643 ret = do_fault_around(vmf);
3648 ret = __do_fault(vmf);
3649 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3652 ret |= finish_fault(vmf);
3653 unlock_page(vmf->page);
3654 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3655 put_page(vmf->page);
3659 static int do_cow_fault(struct vm_fault *vmf)
3661 struct vm_area_struct *vma = vmf->vma;
3664 if (unlikely(anon_vma_prepare(vma)))
3665 return VM_FAULT_OOM;
3667 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3669 return VM_FAULT_OOM;
3671 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3672 &vmf->memcg, false)) {
3673 put_page(vmf->cow_page);
3674 return VM_FAULT_OOM;
3677 ret = __do_fault(vmf);
3678 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3680 if (ret & VM_FAULT_DONE_COW)
3683 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3684 __SetPageUptodate(vmf->cow_page);
3686 ret |= finish_fault(vmf);
3687 unlock_page(vmf->page);
3688 put_page(vmf->page);
3689 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3693 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3694 put_page(vmf->cow_page);
3698 static int do_shared_fault(struct vm_fault *vmf)
3700 struct vm_area_struct *vma = vmf->vma;
3703 ret = __do_fault(vmf);
3704 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3708 * Check if the backing address space wants to know that the page is
3709 * about to become writable
3711 if (vma->vm_ops->page_mkwrite) {
3712 unlock_page(vmf->page);
3713 tmp = do_page_mkwrite(vmf);
3714 if (unlikely(!tmp ||
3715 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3716 put_page(vmf->page);
3721 ret |= finish_fault(vmf);
3722 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3724 unlock_page(vmf->page);
3725 put_page(vmf->page);
3729 fault_dirty_shared_page(vma, vmf->page);
3734 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3735 * but allow concurrent faults).
3736 * The mmap_sem may have been released depending on flags and our
3737 * return value. See filemap_fault() and __lock_page_or_retry().
3739 static int do_fault(struct vm_fault *vmf)
3741 struct vm_area_struct *vma = vmf->vma;
3744 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3745 if (!vma->vm_ops->fault)
3746 ret = VM_FAULT_SIGBUS;
3747 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3748 ret = do_read_fault(vmf);
3749 else if (!(vma->vm_flags & VM_SHARED))
3750 ret = do_cow_fault(vmf);
3752 ret = do_shared_fault(vmf);
3754 /* preallocated pagetable is unused: free it */
3755 if (vmf->prealloc_pte) {
3756 pte_free(vma->vm_mm, vmf->prealloc_pte);
3757 vmf->prealloc_pte = NULL;
3762 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3763 unsigned long addr, int page_nid,
3768 count_vm_numa_event(NUMA_HINT_FAULTS);
3769 if (page_nid == numa_node_id()) {
3770 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3771 *flags |= TNF_FAULT_LOCAL;
3774 return mpol_misplaced(page, vma, addr);
3777 static int do_numa_page(struct vm_fault *vmf)
3779 struct vm_area_struct *vma = vmf->vma;
3780 struct page *page = NULL;
3784 bool migrated = false;
3786 bool was_writable = pte_savedwrite(vmf->orig_pte);
3790 * The "pte" at this point cannot be used safely without
3791 * validation through pte_unmap_same(). It's of NUMA type but
3792 * the pfn may be screwed if the read is non atomic.
3794 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3795 spin_lock(vmf->ptl);
3796 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3797 pte_unmap_unlock(vmf->pte, vmf->ptl);
3802 * Make it present again, Depending on how arch implementes non
3803 * accessible ptes, some can allow access by kernel mode.
3805 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3806 pte = pte_modify(pte, vma->vm_page_prot);
3807 pte = pte_mkyoung(pte);
3809 pte = pte_mkwrite(pte);
3810 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3811 update_mmu_cache(vma, vmf->address, vmf->pte);
3813 page = vm_normal_page(vma, vmf->address, pte);
3815 pte_unmap_unlock(vmf->pte, vmf->ptl);
3819 /* TODO: handle PTE-mapped THP */
3820 if (PageCompound(page)) {
3821 pte_unmap_unlock(vmf->pte, vmf->ptl);
3826 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3827 * much anyway since they can be in shared cache state. This misses
3828 * the case where a mapping is writable but the process never writes
3829 * to it but pte_write gets cleared during protection updates and
3830 * pte_dirty has unpredictable behaviour between PTE scan updates,
3831 * background writeback, dirty balancing and application behaviour.
3833 if (!pte_write(pte))
3834 flags |= TNF_NO_GROUP;
3837 * Flag if the page is shared between multiple address spaces. This
3838 * is later used when determining whether to group tasks together
3840 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3841 flags |= TNF_SHARED;
3843 last_cpupid = page_cpupid_last(page);
3844 page_nid = page_to_nid(page);
3845 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3847 pte_unmap_unlock(vmf->pte, vmf->ptl);
3848 if (target_nid == -1) {
3853 /* Migrate to the requested node */
3854 migrated = migrate_misplaced_page(page, vma, target_nid);
3856 page_nid = target_nid;
3857 flags |= TNF_MIGRATED;
3859 flags |= TNF_MIGRATE_FAIL;
3863 task_numa_fault(last_cpupid, page_nid, 1, flags);
3867 static inline int create_huge_pmd(struct vm_fault *vmf)
3869 if (vma_is_anonymous(vmf->vma))
3870 return do_huge_pmd_anonymous_page(vmf);
3871 if (vmf->vma->vm_ops->huge_fault)
3872 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3873 return VM_FAULT_FALLBACK;
3876 /* `inline' is required to avoid gcc 4.1.2 build error */
3877 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3879 if (vma_is_anonymous(vmf->vma))
3880 return do_huge_pmd_wp_page(vmf, orig_pmd);
3881 if (vmf->vma->vm_ops->huge_fault)
3882 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3884 /* COW handled on pte level: split pmd */
3885 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3886 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3888 return VM_FAULT_FALLBACK;
3891 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3893 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3896 static int create_huge_pud(struct vm_fault *vmf)
3898 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3899 /* No support for anonymous transparent PUD pages yet */
3900 if (vma_is_anonymous(vmf->vma))
3901 return VM_FAULT_FALLBACK;
3902 if (vmf->vma->vm_ops->huge_fault)
3903 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3904 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3905 return VM_FAULT_FALLBACK;
3908 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3910 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3911 /* No support for anonymous transparent PUD pages yet */
3912 if (vma_is_anonymous(vmf->vma))
3913 return VM_FAULT_FALLBACK;
3914 if (vmf->vma->vm_ops->huge_fault)
3915 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3916 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3917 return VM_FAULT_FALLBACK;
3921 * These routines also need to handle stuff like marking pages dirty
3922 * and/or accessed for architectures that don't do it in hardware (most
3923 * RISC architectures). The early dirtying is also good on the i386.
3925 * There is also a hook called "update_mmu_cache()" that architectures
3926 * with external mmu caches can use to update those (ie the Sparc or
3927 * PowerPC hashed page tables that act as extended TLBs).
3929 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3930 * concurrent faults).
3932 * The mmap_sem may have been released depending on flags and our return value.
3933 * See filemap_fault() and __lock_page_or_retry().
3935 static int handle_pte_fault(struct vm_fault *vmf)
3939 if (unlikely(pmd_none(*vmf->pmd))) {
3941 * Leave __pte_alloc() until later: because vm_ops->fault may
3942 * want to allocate huge page, and if we expose page table
3943 * for an instant, it will be difficult to retract from
3944 * concurrent faults and from rmap lookups.
3948 /* See comment in pte_alloc_one_map() */
3949 if (pmd_devmap_trans_unstable(vmf->pmd))
3952 * A regular pmd is established and it can't morph into a huge
3953 * pmd from under us anymore at this point because we hold the
3954 * mmap_sem read mode and khugepaged takes it in write mode.
3955 * So now it's safe to run pte_offset_map().
3957 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3958 vmf->orig_pte = *vmf->pte;
3961 * some architectures can have larger ptes than wordsize,
3962 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3963 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3964 * accesses. The code below just needs a consistent view
3965 * for the ifs and we later double check anyway with the
3966 * ptl lock held. So here a barrier will do.
3969 if (pte_none(vmf->orig_pte)) {
3970 pte_unmap(vmf->pte);
3976 if (vma_is_anonymous(vmf->vma))
3977 return do_anonymous_page(vmf);
3979 return do_fault(vmf);
3982 if (!pte_present(vmf->orig_pte))
3983 return do_swap_page(vmf);
3985 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3986 return do_numa_page(vmf);
3988 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3989 spin_lock(vmf->ptl);
3990 entry = vmf->orig_pte;
3991 if (unlikely(!pte_same(*vmf->pte, entry)))
3993 if (vmf->flags & FAULT_FLAG_WRITE) {
3994 if (!pte_write(entry))
3995 return do_wp_page(vmf);
3996 entry = pte_mkdirty(entry);
3998 entry = pte_mkyoung(entry);
3999 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4000 vmf->flags & FAULT_FLAG_WRITE)) {
4001 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4004 * This is needed only for protection faults but the arch code
4005 * is not yet telling us if this is a protection fault or not.
4006 * This still avoids useless tlb flushes for .text page faults
4009 if (vmf->flags & FAULT_FLAG_WRITE)
4010 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4013 pte_unmap_unlock(vmf->pte, vmf->ptl);
4018 * By the time we get here, we already hold the mm semaphore
4020 * The mmap_sem may have been released depending on flags and our
4021 * return value. See filemap_fault() and __lock_page_or_retry().
4023 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4026 struct vm_fault vmf = {
4028 .address = address & PAGE_MASK,
4030 .pgoff = linear_page_index(vma, address),
4031 .gfp_mask = __get_fault_gfp_mask(vma),
4033 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4034 struct mm_struct *mm = vma->vm_mm;
4039 pgd = pgd_offset(mm, address);
4040 p4d = p4d_alloc(mm, pgd, address);
4042 return VM_FAULT_OOM;
4044 vmf.pud = pud_alloc(mm, p4d, address);
4046 return VM_FAULT_OOM;
4047 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4048 ret = create_huge_pud(&vmf);
4049 if (!(ret & VM_FAULT_FALLBACK))
4052 pud_t orig_pud = *vmf.pud;
4055 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4057 /* NUMA case for anonymous PUDs would go here */
4059 if (dirty && !pud_write(orig_pud)) {
4060 ret = wp_huge_pud(&vmf, orig_pud);
4061 if (!(ret & VM_FAULT_FALLBACK))
4064 huge_pud_set_accessed(&vmf, orig_pud);
4070 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4072 return VM_FAULT_OOM;
4073 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4074 ret = create_huge_pmd(&vmf);
4075 if (!(ret & VM_FAULT_FALLBACK))
4078 pmd_t orig_pmd = *vmf.pmd;
4081 if (unlikely(is_swap_pmd(orig_pmd))) {
4082 VM_BUG_ON(thp_migration_supported() &&
4083 !is_pmd_migration_entry(orig_pmd));
4084 if (is_pmd_migration_entry(orig_pmd))
4085 pmd_migration_entry_wait(mm, vmf.pmd);
4088 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4089 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4090 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4092 if (dirty && !pmd_write(orig_pmd)) {
4093 ret = wp_huge_pmd(&vmf, orig_pmd);
4094 if (!(ret & VM_FAULT_FALLBACK))
4097 huge_pmd_set_accessed(&vmf, orig_pmd);
4103 return handle_pte_fault(&vmf);
4107 * By the time we get here, we already hold the mm semaphore
4109 * The mmap_sem may have been released depending on flags and our
4110 * return value. See filemap_fault() and __lock_page_or_retry().
4112 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4117 __set_current_state(TASK_RUNNING);
4119 count_vm_event(PGFAULT);
4120 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4122 /* do counter updates before entering really critical section. */
4123 check_sync_rss_stat(current);
4125 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4126 flags & FAULT_FLAG_INSTRUCTION,
4127 flags & FAULT_FLAG_REMOTE))
4128 return VM_FAULT_SIGSEGV;
4131 * Enable the memcg OOM handling for faults triggered in user
4132 * space. Kernel faults are handled more gracefully.
4134 if (flags & FAULT_FLAG_USER)
4135 mem_cgroup_oom_enable();
4137 if (unlikely(is_vm_hugetlb_page(vma)))
4138 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4140 ret = __handle_mm_fault(vma, address, flags);
4142 if (flags & FAULT_FLAG_USER) {
4143 mem_cgroup_oom_disable();
4145 * The task may have entered a memcg OOM situation but
4146 * if the allocation error was handled gracefully (no
4147 * VM_FAULT_OOM), there is no need to kill anything.
4148 * Just clean up the OOM state peacefully.
4150 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4151 mem_cgroup_oom_synchronize(false);
4156 EXPORT_SYMBOL_GPL(handle_mm_fault);
4158 #ifndef __PAGETABLE_P4D_FOLDED
4160 * Allocate p4d page table.
4161 * We've already handled the fast-path in-line.
4163 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4165 p4d_t *new = p4d_alloc_one(mm, address);
4169 smp_wmb(); /* See comment in __pte_alloc */
4171 spin_lock(&mm->page_table_lock);
4172 if (pgd_present(*pgd)) /* Another has populated it */
4175 pgd_populate(mm, pgd, new);
4176 spin_unlock(&mm->page_table_lock);
4179 #endif /* __PAGETABLE_P4D_FOLDED */
4181 #ifndef __PAGETABLE_PUD_FOLDED
4183 * Allocate page upper directory.
4184 * We've already handled the fast-path in-line.
4186 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4188 pud_t *new = pud_alloc_one(mm, address);
4192 smp_wmb(); /* See comment in __pte_alloc */
4194 spin_lock(&mm->page_table_lock);
4195 #ifndef __ARCH_HAS_5LEVEL_HACK
4196 if (!p4d_present(*p4d)) {
4198 p4d_populate(mm, p4d, new);
4199 } else /* Another has populated it */
4202 if (!pgd_present(*p4d)) {
4204 pgd_populate(mm, p4d, new);
4205 } else /* Another has populated it */
4207 #endif /* __ARCH_HAS_5LEVEL_HACK */
4208 spin_unlock(&mm->page_table_lock);
4211 #endif /* __PAGETABLE_PUD_FOLDED */
4213 #ifndef __PAGETABLE_PMD_FOLDED
4215 * Allocate page middle directory.
4216 * We've already handled the fast-path in-line.
4218 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4221 pmd_t *new = pmd_alloc_one(mm, address);
4225 smp_wmb(); /* See comment in __pte_alloc */
4227 ptl = pud_lock(mm, pud);
4228 #ifndef __ARCH_HAS_4LEVEL_HACK
4229 if (!pud_present(*pud)) {
4231 pud_populate(mm, pud, new);
4232 } else /* Another has populated it */
4235 if (!pgd_present(*pud)) {
4237 pgd_populate(mm, pud, new);
4238 } else /* Another has populated it */
4240 #endif /* __ARCH_HAS_4LEVEL_HACK */
4244 #endif /* __PAGETABLE_PMD_FOLDED */
4246 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4247 unsigned long *start, unsigned long *end,
4248 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4256 pgd = pgd_offset(mm, address);
4257 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4260 p4d = p4d_offset(pgd, address);
4261 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4264 pud = pud_offset(p4d, address);
4265 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4268 pmd = pmd_offset(pud, address);
4269 VM_BUG_ON(pmd_trans_huge(*pmd));
4271 if (pmd_huge(*pmd)) {
4276 *start = address & PMD_MASK;
4277 *end = *start + PMD_SIZE;
4278 mmu_notifier_invalidate_range_start(mm, *start, *end);
4280 *ptlp = pmd_lock(mm, pmd);
4281 if (pmd_huge(*pmd)) {
4287 mmu_notifier_invalidate_range_end(mm, *start, *end);
4290 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4294 *start = address & PAGE_MASK;
4295 *end = *start + PAGE_SIZE;
4296 mmu_notifier_invalidate_range_start(mm, *start, *end);
4298 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4299 if (!pte_present(*ptep))
4304 pte_unmap_unlock(ptep, *ptlp);
4306 mmu_notifier_invalidate_range_end(mm, *start, *end);
4311 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4312 pte_t **ptepp, spinlock_t **ptlp)
4316 /* (void) is needed to make gcc happy */
4317 (void) __cond_lock(*ptlp,
4318 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4319 ptepp, NULL, ptlp)));
4323 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4324 unsigned long *start, unsigned long *end,
4325 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4329 /* (void) is needed to make gcc happy */
4330 (void) __cond_lock(*ptlp,
4331 !(res = __follow_pte_pmd(mm, address, start, end,
4332 ptepp, pmdpp, ptlp)));
4335 EXPORT_SYMBOL(follow_pte_pmd);
4338 * follow_pfn - look up PFN at a user virtual address
4339 * @vma: memory mapping
4340 * @address: user virtual address
4341 * @pfn: location to store found PFN
4343 * Only IO mappings and raw PFN mappings are allowed.
4345 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4347 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4354 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4357 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4360 *pfn = pte_pfn(*ptep);
4361 pte_unmap_unlock(ptep, ptl);
4364 EXPORT_SYMBOL(follow_pfn);
4366 #ifdef CONFIG_HAVE_IOREMAP_PROT
4367 int follow_phys(struct vm_area_struct *vma,
4368 unsigned long address, unsigned int flags,
4369 unsigned long *prot, resource_size_t *phys)
4375 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4378 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4382 if ((flags & FOLL_WRITE) && !pte_write(pte))
4385 *prot = pgprot_val(pte_pgprot(pte));
4386 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4390 pte_unmap_unlock(ptep, ptl);
4395 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4396 void *buf, int len, int write)
4398 resource_size_t phys_addr;
4399 unsigned long prot = 0;
4400 void __iomem *maddr;
4401 int offset = addr & (PAGE_SIZE-1);
4403 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4406 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4408 memcpy_toio(maddr + offset, buf, len);
4410 memcpy_fromio(buf, maddr + offset, len);
4415 EXPORT_SYMBOL_GPL(generic_access_phys);
4419 * Access another process' address space as given in mm. If non-NULL, use the
4420 * given task for page fault accounting.
4422 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4423 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4425 struct vm_area_struct *vma;
4426 void *old_buf = buf;
4427 int write = gup_flags & FOLL_WRITE;
4429 down_read(&mm->mmap_sem);
4430 /* ignore errors, just check how much was successfully transferred */
4432 int bytes, ret, offset;
4434 struct page *page = NULL;
4436 ret = get_user_pages_remote(tsk, mm, addr, 1,
4437 gup_flags, &page, &vma, NULL);
4439 #ifndef CONFIG_HAVE_IOREMAP_PROT
4443 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4444 * we can access using slightly different code.
4446 vma = find_vma(mm, addr);
4447 if (!vma || vma->vm_start > addr)
4449 if (vma->vm_ops && vma->vm_ops->access)
4450 ret = vma->vm_ops->access(vma, addr, buf,
4458 offset = addr & (PAGE_SIZE-1);
4459 if (bytes > PAGE_SIZE-offset)
4460 bytes = PAGE_SIZE-offset;
4464 copy_to_user_page(vma, page, addr,
4465 maddr + offset, buf, bytes);
4466 set_page_dirty_lock(page);
4468 copy_from_user_page(vma, page, addr,
4469 buf, maddr + offset, bytes);
4478 up_read(&mm->mmap_sem);
4480 return buf - old_buf;
4484 * access_remote_vm - access another process' address space
4485 * @mm: the mm_struct of the target address space
4486 * @addr: start address to access
4487 * @buf: source or destination buffer
4488 * @len: number of bytes to transfer
4489 * @gup_flags: flags modifying lookup behaviour
4491 * The caller must hold a reference on @mm.
4493 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4494 void *buf, int len, unsigned int gup_flags)
4496 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4500 * Access another process' address space.
4501 * Source/target buffer must be kernel space,
4502 * Do not walk the page table directly, use get_user_pages
4504 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4505 void *buf, int len, unsigned int gup_flags)
4507 struct mm_struct *mm;
4510 mm = get_task_mm(tsk);
4514 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4520 EXPORT_SYMBOL_GPL(access_process_vm);
4523 * Print the name of a VMA.
4525 void print_vma_addr(char *prefix, unsigned long ip)
4527 struct mm_struct *mm = current->mm;
4528 struct vm_area_struct *vma;
4531 * we might be running from an atomic context so we cannot sleep
4533 if (!down_read_trylock(&mm->mmap_sem))
4536 vma = find_vma(mm, ip);
4537 if (vma && vma->vm_file) {
4538 struct file *f = vma->vm_file;
4539 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4543 p = file_path(f, buf, PAGE_SIZE);
4546 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4548 vma->vm_end - vma->vm_start);
4549 free_page((unsigned long)buf);
4552 up_read(&mm->mmap_sem);
4555 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4556 void __might_fault(const char *file, int line)
4559 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4560 * holding the mmap_sem, this is safe because kernel memory doesn't
4561 * get paged out, therefore we'll never actually fault, and the
4562 * below annotations will generate false positives.
4564 if (uaccess_kernel())
4566 if (pagefault_disabled())
4568 __might_sleep(file, line, 0);
4569 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4571 might_lock_read(¤t->mm->mmap_sem);
4574 EXPORT_SYMBOL(__might_fault);
4577 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4578 static void clear_gigantic_page(struct page *page,
4580 unsigned int pages_per_huge_page)
4583 struct page *p = page;
4586 for (i = 0; i < pages_per_huge_page;
4587 i++, p = mem_map_next(p, page, i)) {
4589 clear_user_highpage(p, addr + i * PAGE_SIZE);
4592 void clear_huge_page(struct page *page,
4593 unsigned long addr_hint, unsigned int pages_per_huge_page)
4596 unsigned long addr = addr_hint &
4597 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4599 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4600 clear_gigantic_page(page, addr, pages_per_huge_page);
4604 /* Clear sub-page to access last to keep its cache lines hot */
4606 n = (addr_hint - addr) / PAGE_SIZE;
4607 if (2 * n <= pages_per_huge_page) {
4608 /* If sub-page to access in first half of huge page */
4611 /* Clear sub-pages at the end of huge page */
4612 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4614 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4617 /* If sub-page to access in second half of huge page */
4618 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4619 l = pages_per_huge_page - n;
4620 /* Clear sub-pages at the begin of huge page */
4621 for (i = 0; i < base; i++) {
4623 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4627 * Clear remaining sub-pages in left-right-left-right pattern
4628 * towards the sub-page to access
4630 for (i = 0; i < l; i++) {
4631 int left_idx = base + i;
4632 int right_idx = base + 2 * l - 1 - i;
4635 clear_user_highpage(page + left_idx,
4636 addr + left_idx * PAGE_SIZE);
4638 clear_user_highpage(page + right_idx,
4639 addr + right_idx * PAGE_SIZE);
4643 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4645 struct vm_area_struct *vma,
4646 unsigned int pages_per_huge_page)
4649 struct page *dst_base = dst;
4650 struct page *src_base = src;
4652 for (i = 0; i < pages_per_huge_page; ) {
4654 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4657 dst = mem_map_next(dst, dst_base, i);
4658 src = mem_map_next(src, src_base, i);
4662 void copy_user_huge_page(struct page *dst, struct page *src,
4663 unsigned long addr, struct vm_area_struct *vma,
4664 unsigned int pages_per_huge_page)
4668 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4669 copy_user_gigantic_page(dst, src, addr, vma,
4670 pages_per_huge_page);
4675 for (i = 0; i < pages_per_huge_page; i++) {
4677 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4681 long copy_huge_page_from_user(struct page *dst_page,
4682 const void __user *usr_src,
4683 unsigned int pages_per_huge_page,
4684 bool allow_pagefault)
4686 void *src = (void *)usr_src;
4688 unsigned long i, rc = 0;
4689 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4691 for (i = 0; i < pages_per_huge_page; i++) {
4692 if (allow_pagefault)
4693 page_kaddr = kmap(dst_page + i);
4695 page_kaddr = kmap_atomic(dst_page + i);
4696 rc = copy_from_user(page_kaddr,
4697 (const void __user *)(src + i * PAGE_SIZE),
4699 if (allow_pagefault)
4700 kunmap(dst_page + i);
4702 kunmap_atomic(page_kaddr);
4704 ret_val -= (PAGE_SIZE - rc);
4712 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4714 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4716 static struct kmem_cache *page_ptl_cachep;
4718 void __init ptlock_cache_init(void)
4720 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4724 bool ptlock_alloc(struct page *page)
4728 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4735 void ptlock_free(struct page *page)
4737 kmem_cache_free(page_ptl_cachep, page->ptl);