4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
84 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
94 EXPORT_SYMBOL(mem_map);
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 EXPORT_SYMBOL(high_memory);
108 * Randomize the address space (stacks, mmaps, brk, etc.).
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
120 static int __init disable_randmaps(char *s)
122 randomize_va_space = 0;
125 __setup("norandmaps", disable_randmaps);
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
130 unsigned long highest_memmap_pfn __read_mostly;
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 static int __init init_zero_pfn(void)
137 zero_pfn = page_to_pfn(ZERO_PAGE(0));
140 core_initcall(init_zero_pfn);
143 #if defined(SPLIT_RSS_COUNTING)
145 void sync_mm_rss(struct mm_struct *mm)
149 for (i = 0; i < NR_MM_COUNTERS; i++) {
150 if (current->rss_stat.count[i]) {
151 add_mm_counter(mm, i, current->rss_stat.count[i]);
152 current->rss_stat.count[i] = 0;
155 current->rss_stat.events = 0;
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160 struct task_struct *task = current;
162 if (likely(task->mm == mm))
163 task->rss_stat.count[member] += val;
165 add_mm_counter(mm, member, val);
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH (64)
172 static void check_sync_rss_stat(struct task_struct *task)
174 if (unlikely(task != current))
176 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177 sync_mm_rss(task->mm);
179 #else /* SPLIT_RSS_COUNTING */
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184 static void check_sync_rss_stat(struct task_struct *task)
188 #endif /* SPLIT_RSS_COUNTING */
190 #ifdef HAVE_GENERIC_MMU_GATHER
192 static bool tlb_next_batch(struct mmu_gather *tlb)
194 struct mmu_gather_batch *batch;
198 tlb->active = batch->next;
202 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
205 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
212 batch->max = MAX_GATHER_BATCH;
214 tlb->active->next = batch;
220 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
221 unsigned long start, unsigned long end)
225 /* Is it from 0 to ~0? */
226 tlb->fullmm = !(start | (end+1));
227 tlb->need_flush_all = 0;
228 tlb->local.next = NULL;
230 tlb->local.max = ARRAY_SIZE(tlb->__pages);
231 tlb->active = &tlb->local;
232 tlb->batch_count = 0;
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 __tlb_reset_range(tlb);
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
248 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
249 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
250 tlb_table_flush(tlb);
252 __tlb_reset_range(tlb);
255 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
257 struct mmu_gather_batch *batch;
259 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
263 tlb->active = &tlb->local;
266 void tlb_flush_mmu(struct mmu_gather *tlb)
268 tlb_flush_mmu_tlbonly(tlb);
269 tlb_flush_mmu_free(tlb);
273 * Called at the end of the shootdown operation to free up any resources
274 * that were required.
276 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
277 unsigned long start, unsigned long end, bool force)
279 struct mmu_gather_batch *batch, *next;
282 __tlb_adjust_range(tlb, start, end - start);
286 /* keep the page table cache within bounds */
289 for (batch = tlb->local.next; batch; batch = next) {
291 free_pages((unsigned long)batch, 0);
293 tlb->local.next = NULL;
297 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298 * handling the additional races in SMP caused by other CPUs caching valid
299 * mappings in their TLBs. Returns the number of free page slots left.
300 * When out of page slots we must call tlb_flush_mmu().
301 *returns true if the caller should flush.
303 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
305 struct mmu_gather_batch *batch;
307 VM_BUG_ON(!tlb->end);
308 VM_WARN_ON(tlb->page_size != page_size);
312 * Add the page and check if we are full. If so
315 batch->pages[batch->nr++] = page;
316 if (batch->nr == batch->max) {
317 if (!tlb_next_batch(tlb))
321 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
326 #endif /* HAVE_GENERIC_MMU_GATHER */
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
331 * See the comment near struct mmu_table_batch.
334 static void tlb_remove_table_smp_sync(void *arg)
336 /* Simply deliver the interrupt */
339 static void tlb_remove_table_one(void *table)
342 * This isn't an RCU grace period and hence the page-tables cannot be
343 * assumed to be actually RCU-freed.
345 * It is however sufficient for software page-table walkers that rely on
346 * IRQ disabling. See the comment near struct mmu_table_batch.
348 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
349 __tlb_remove_table(table);
352 static void tlb_remove_table_rcu(struct rcu_head *head)
354 struct mmu_table_batch *batch;
357 batch = container_of(head, struct mmu_table_batch, rcu);
359 for (i = 0; i < batch->nr; i++)
360 __tlb_remove_table(batch->tables[i]);
362 free_page((unsigned long)batch);
365 void tlb_table_flush(struct mmu_gather *tlb)
367 struct mmu_table_batch **batch = &tlb->batch;
370 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
375 void tlb_remove_table(struct mmu_gather *tlb, void *table)
377 struct mmu_table_batch **batch = &tlb->batch;
380 * When there's less then two users of this mm there cannot be a
381 * concurrent page-table walk.
383 if (atomic_read(&tlb->mm->mm_users) < 2) {
384 __tlb_remove_table(table);
388 if (*batch == NULL) {
389 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
390 if (*batch == NULL) {
391 tlb_remove_table_one(table);
396 (*batch)->tables[(*batch)->nr++] = table;
397 if ((*batch)->nr == MAX_TABLE_BATCH)
398 tlb_table_flush(tlb);
401 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
404 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
405 * @tlb: the mmu_gather structure to initialize
406 * @mm: the mm_struct of the target address space
407 * @start: start of the region that will be removed from the page-table
408 * @end: end of the region that will be removed from the page-table
410 * Called to initialize an (on-stack) mmu_gather structure for page-table
411 * tear-down from @mm. The @start and @end are set to 0 and -1
412 * respectively when @mm is without users and we're going to destroy
413 * the full address space (exit/execve).
415 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
416 unsigned long start, unsigned long end)
418 arch_tlb_gather_mmu(tlb, mm, start, end);
419 inc_tlb_flush_pending(tlb->mm);
422 void tlb_finish_mmu(struct mmu_gather *tlb,
423 unsigned long start, unsigned long end)
426 * If there are parallel threads are doing PTE changes on same range
427 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
428 * flush by batching, a thread has stable TLB entry can fail to flush
429 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
430 * forcefully if we detect parallel PTE batching threads.
432 bool force = mm_tlb_flush_nested(tlb->mm);
434 arch_tlb_finish_mmu(tlb, start, end, force);
435 dec_tlb_flush_pending(tlb->mm);
439 * Note: this doesn't free the actual pages themselves. That
440 * has been handled earlier when unmapping all the memory regions.
442 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
445 pgtable_t token = pmd_pgtable(*pmd);
447 pte_free_tlb(tlb, token, addr);
448 mm_dec_nr_ptes(tlb->mm);
451 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
452 unsigned long addr, unsigned long end,
453 unsigned long floor, unsigned long ceiling)
460 pmd = pmd_offset(pud, addr);
462 next = pmd_addr_end(addr, end);
463 if (pmd_none_or_clear_bad(pmd))
465 free_pte_range(tlb, pmd, addr);
466 } while (pmd++, addr = next, addr != end);
476 if (end - 1 > ceiling - 1)
479 pmd = pmd_offset(pud, start);
481 pmd_free_tlb(tlb, pmd, start);
482 mm_dec_nr_pmds(tlb->mm);
485 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
486 unsigned long addr, unsigned long end,
487 unsigned long floor, unsigned long ceiling)
494 pud = pud_offset(p4d, addr);
496 next = pud_addr_end(addr, end);
497 if (pud_none_or_clear_bad(pud))
499 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
500 } while (pud++, addr = next, addr != end);
510 if (end - 1 > ceiling - 1)
513 pud = pud_offset(p4d, start);
515 pud_free_tlb(tlb, pud, start);
516 mm_dec_nr_puds(tlb->mm);
519 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
520 unsigned long addr, unsigned long end,
521 unsigned long floor, unsigned long ceiling)
528 p4d = p4d_offset(pgd, addr);
530 next = p4d_addr_end(addr, end);
531 if (p4d_none_or_clear_bad(p4d))
533 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
534 } while (p4d++, addr = next, addr != end);
540 ceiling &= PGDIR_MASK;
544 if (end - 1 > ceiling - 1)
547 p4d = p4d_offset(pgd, start);
549 p4d_free_tlb(tlb, p4d, start);
553 * This function frees user-level page tables of a process.
555 void free_pgd_range(struct mmu_gather *tlb,
556 unsigned long addr, unsigned long end,
557 unsigned long floor, unsigned long ceiling)
563 * The next few lines have given us lots of grief...
565 * Why are we testing PMD* at this top level? Because often
566 * there will be no work to do at all, and we'd prefer not to
567 * go all the way down to the bottom just to discover that.
569 * Why all these "- 1"s? Because 0 represents both the bottom
570 * of the address space and the top of it (using -1 for the
571 * top wouldn't help much: the masks would do the wrong thing).
572 * The rule is that addr 0 and floor 0 refer to the bottom of
573 * the address space, but end 0 and ceiling 0 refer to the top
574 * Comparisons need to use "end - 1" and "ceiling - 1" (though
575 * that end 0 case should be mythical).
577 * Wherever addr is brought up or ceiling brought down, we must
578 * be careful to reject "the opposite 0" before it confuses the
579 * subsequent tests. But what about where end is brought down
580 * by PMD_SIZE below? no, end can't go down to 0 there.
582 * Whereas we round start (addr) and ceiling down, by different
583 * masks at different levels, in order to test whether a table
584 * now has no other vmas using it, so can be freed, we don't
585 * bother to round floor or end up - the tests don't need that.
599 if (end - 1 > ceiling - 1)
604 * We add page table cache pages with PAGE_SIZE,
605 * (see pte_free_tlb()), flush the tlb if we need
607 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
608 pgd = pgd_offset(tlb->mm, addr);
610 next = pgd_addr_end(addr, end);
611 if (pgd_none_or_clear_bad(pgd))
613 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
614 } while (pgd++, addr = next, addr != end);
617 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
618 unsigned long floor, unsigned long ceiling)
621 struct vm_area_struct *next = vma->vm_next;
622 unsigned long addr = vma->vm_start;
625 * Hide vma from rmap and truncate_pagecache before freeing
628 unlink_anon_vmas(vma);
629 unlink_file_vma(vma);
631 if (is_vm_hugetlb_page(vma)) {
632 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
633 floor, next ? next->vm_start : ceiling);
636 * Optimization: gather nearby vmas into one call down
638 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
639 && !is_vm_hugetlb_page(next)) {
642 unlink_anon_vmas(vma);
643 unlink_file_vma(vma);
645 free_pgd_range(tlb, addr, vma->vm_end,
646 floor, next ? next->vm_start : ceiling);
652 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
655 pgtable_t new = pte_alloc_one(mm, address);
660 * Ensure all pte setup (eg. pte page lock and page clearing) are
661 * visible before the pte is made visible to other CPUs by being
662 * put into page tables.
664 * The other side of the story is the pointer chasing in the page
665 * table walking code (when walking the page table without locking;
666 * ie. most of the time). Fortunately, these data accesses consist
667 * of a chain of data-dependent loads, meaning most CPUs (alpha
668 * being the notable exception) will already guarantee loads are
669 * seen in-order. See the alpha page table accessors for the
670 * smp_read_barrier_depends() barriers in page table walking code.
672 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
674 ptl = pmd_lock(mm, pmd);
675 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
677 pmd_populate(mm, pmd, new);
686 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
688 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
692 smp_wmb(); /* See comment in __pte_alloc */
694 spin_lock(&init_mm.page_table_lock);
695 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
696 pmd_populate_kernel(&init_mm, pmd, new);
699 spin_unlock(&init_mm.page_table_lock);
701 pte_free_kernel(&init_mm, new);
705 static inline void init_rss_vec(int *rss)
707 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
710 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
714 if (current->mm == mm)
716 for (i = 0; i < NR_MM_COUNTERS; i++)
718 add_mm_counter(mm, i, rss[i]);
722 * This function is called to print an error when a bad pte
723 * is found. For example, we might have a PFN-mapped pte in
724 * a region that doesn't allow it.
726 * The calling function must still handle the error.
728 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
729 pte_t pte, struct page *page)
731 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
732 p4d_t *p4d = p4d_offset(pgd, addr);
733 pud_t *pud = pud_offset(p4d, addr);
734 pmd_t *pmd = pmd_offset(pud, addr);
735 struct address_space *mapping;
737 static unsigned long resume;
738 static unsigned long nr_shown;
739 static unsigned long nr_unshown;
742 * Allow a burst of 60 reports, then keep quiet for that minute;
743 * or allow a steady drip of one report per second.
745 if (nr_shown == 60) {
746 if (time_before(jiffies, resume)) {
751 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
758 resume = jiffies + 60 * HZ;
760 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
761 index = linear_page_index(vma, addr);
763 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
765 (long long)pte_val(pte), (long long)pmd_val(*pmd));
767 dump_page(page, "bad pte");
768 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
769 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
771 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
773 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
775 vma->vm_ops ? vma->vm_ops->fault : NULL,
776 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
777 mapping ? mapping->a_ops->readpage : NULL);
779 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
783 * vm_normal_page -- This function gets the "struct page" associated with a pte.
785 * "Special" mappings do not wish to be associated with a "struct page" (either
786 * it doesn't exist, or it exists but they don't want to touch it). In this
787 * case, NULL is returned here. "Normal" mappings do have a struct page.
789 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
790 * pte bit, in which case this function is trivial. Secondly, an architecture
791 * may not have a spare pte bit, which requires a more complicated scheme,
794 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
795 * special mapping (even if there are underlying and valid "struct pages").
796 * COWed pages of a VM_PFNMAP are always normal.
798 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
799 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
800 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
801 * mapping will always honor the rule
803 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
805 * And for normal mappings this is false.
807 * This restricts such mappings to be a linear translation from virtual address
808 * to pfn. To get around this restriction, we allow arbitrary mappings so long
809 * as the vma is not a COW mapping; in that case, we know that all ptes are
810 * special (because none can have been COWed).
813 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
815 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
816 * page" backing, however the difference is that _all_ pages with a struct
817 * page (that is, those where pfn_valid is true) are refcounted and considered
818 * normal pages by the VM. The disadvantage is that pages are refcounted
819 * (which can be slower and simply not an option for some PFNMAP users). The
820 * advantage is that we don't have to follow the strict linearity rule of
821 * PFNMAP mappings in order to support COWable mappings.
824 #ifdef __HAVE_ARCH_PTE_SPECIAL
825 # define HAVE_PTE_SPECIAL 1
827 # define HAVE_PTE_SPECIAL 0
829 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
830 pte_t pte, bool with_public_device)
832 unsigned long pfn = pte_pfn(pte);
834 if (HAVE_PTE_SPECIAL) {
835 if (likely(!pte_special(pte)))
837 if (vma->vm_ops && vma->vm_ops->find_special_page)
838 return vma->vm_ops->find_special_page(vma, addr);
839 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
841 if (is_zero_pfn(pfn))
845 * Device public pages are special pages (they are ZONE_DEVICE
846 * pages but different from persistent memory). They behave
847 * allmost like normal pages. The difference is that they are
848 * not on the lru and thus should never be involve with any-
849 * thing that involve lru manipulation (mlock, numa balancing,
852 * This is why we still want to return NULL for such page from
853 * vm_normal_page() so that we do not have to special case all
854 * call site of vm_normal_page().
856 if (likely(pfn <= highest_memmap_pfn)) {
857 struct page *page = pfn_to_page(pfn);
859 if (is_device_public_page(page)) {
860 if (with_public_device)
865 print_bad_pte(vma, addr, pte, NULL);
869 /* !HAVE_PTE_SPECIAL case follows: */
871 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
872 if (vma->vm_flags & VM_MIXEDMAP) {
878 off = (addr - vma->vm_start) >> PAGE_SHIFT;
879 if (pfn == vma->vm_pgoff + off)
881 if (!is_cow_mapping(vma->vm_flags))
886 if (is_zero_pfn(pfn))
889 if (unlikely(pfn > highest_memmap_pfn)) {
890 print_bad_pte(vma, addr, pte, NULL);
895 * NOTE! We still have PageReserved() pages in the page tables.
896 * eg. VDSO mappings can cause them to exist.
899 return pfn_to_page(pfn);
902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
903 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
906 unsigned long pfn = pmd_pfn(pmd);
909 * There is no pmd_special() but there may be special pmds, e.g.
910 * in a direct-access (dax) mapping, so let's just replicate the
911 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
913 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
914 if (vma->vm_flags & VM_MIXEDMAP) {
920 off = (addr - vma->vm_start) >> PAGE_SHIFT;
921 if (pfn == vma->vm_pgoff + off)
923 if (!is_cow_mapping(vma->vm_flags))
928 if (is_zero_pfn(pfn))
930 if (unlikely(pfn > highest_memmap_pfn))
934 * NOTE! We still have PageReserved() pages in the page tables.
935 * eg. VDSO mappings can cause them to exist.
938 return pfn_to_page(pfn);
943 * copy one vm_area from one task to the other. Assumes the page tables
944 * already present in the new task to be cleared in the whole range
945 * covered by this vma.
948 static inline unsigned long
949 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
950 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
951 unsigned long addr, int *rss)
953 unsigned long vm_flags = vma->vm_flags;
954 pte_t pte = *src_pte;
957 /* pte contains position in swap or file, so copy. */
958 if (unlikely(!pte_present(pte))) {
959 swp_entry_t entry = pte_to_swp_entry(pte);
961 if (likely(!non_swap_entry(entry))) {
962 if (swap_duplicate(entry) < 0)
965 /* make sure dst_mm is on swapoff's mmlist. */
966 if (unlikely(list_empty(&dst_mm->mmlist))) {
967 spin_lock(&mmlist_lock);
968 if (list_empty(&dst_mm->mmlist))
969 list_add(&dst_mm->mmlist,
971 spin_unlock(&mmlist_lock);
974 } else if (is_migration_entry(entry)) {
975 page = migration_entry_to_page(entry);
977 rss[mm_counter(page)]++;
979 if (is_write_migration_entry(entry) &&
980 is_cow_mapping(vm_flags)) {
982 * COW mappings require pages in both
983 * parent and child to be set to read.
985 make_migration_entry_read(&entry);
986 pte = swp_entry_to_pte(entry);
987 if (pte_swp_soft_dirty(*src_pte))
988 pte = pte_swp_mksoft_dirty(pte);
989 set_pte_at(src_mm, addr, src_pte, pte);
991 } else if (is_device_private_entry(entry)) {
992 page = device_private_entry_to_page(entry);
995 * Update rss count even for unaddressable pages, as
996 * they should treated just like normal pages in this
999 * We will likely want to have some new rss counters
1000 * for unaddressable pages, at some point. But for now
1001 * keep things as they are.
1004 rss[mm_counter(page)]++;
1005 page_dup_rmap(page, false);
1008 * We do not preserve soft-dirty information, because so
1009 * far, checkpoint/restore is the only feature that
1010 * requires that. And checkpoint/restore does not work
1011 * when a device driver is involved (you cannot easily
1012 * save and restore device driver state).
1014 if (is_write_device_private_entry(entry) &&
1015 is_cow_mapping(vm_flags)) {
1016 make_device_private_entry_read(&entry);
1017 pte = swp_entry_to_pte(entry);
1018 set_pte_at(src_mm, addr, src_pte, pte);
1025 * If it's a COW mapping, write protect it both
1026 * in the parent and the child
1028 if (is_cow_mapping(vm_flags)) {
1029 ptep_set_wrprotect(src_mm, addr, src_pte);
1030 pte = pte_wrprotect(pte);
1034 * If it's a shared mapping, mark it clean in
1037 if (vm_flags & VM_SHARED)
1038 pte = pte_mkclean(pte);
1039 pte = pte_mkold(pte);
1041 page = vm_normal_page(vma, addr, pte);
1044 page_dup_rmap(page, false);
1045 rss[mm_counter(page)]++;
1046 } else if (pte_devmap(pte)) {
1047 page = pte_page(pte);
1050 * Cache coherent device memory behave like regular page and
1051 * not like persistent memory page. For more informations see
1052 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1054 if (is_device_public_page(page)) {
1056 page_dup_rmap(page, false);
1057 rss[mm_counter(page)]++;
1062 set_pte_at(dst_mm, addr, dst_pte, pte);
1066 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1067 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1068 unsigned long addr, unsigned long end)
1070 pte_t *orig_src_pte, *orig_dst_pte;
1071 pte_t *src_pte, *dst_pte;
1072 spinlock_t *src_ptl, *dst_ptl;
1074 int rss[NR_MM_COUNTERS];
1075 swp_entry_t entry = (swp_entry_t){0};
1080 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1083 src_pte = pte_offset_map(src_pmd, addr);
1084 src_ptl = pte_lockptr(src_mm, src_pmd);
1085 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1086 orig_src_pte = src_pte;
1087 orig_dst_pte = dst_pte;
1088 arch_enter_lazy_mmu_mode();
1092 * We are holding two locks at this point - either of them
1093 * could generate latencies in another task on another CPU.
1095 if (progress >= 32) {
1097 if (need_resched() ||
1098 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1101 if (pte_none(*src_pte)) {
1105 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1110 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1112 arch_leave_lazy_mmu_mode();
1113 spin_unlock(src_ptl);
1114 pte_unmap(orig_src_pte);
1115 add_mm_rss_vec(dst_mm, rss);
1116 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1120 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1129 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1130 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1131 unsigned long addr, unsigned long end)
1133 pmd_t *src_pmd, *dst_pmd;
1136 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1139 src_pmd = pmd_offset(src_pud, addr);
1141 next = pmd_addr_end(addr, end);
1142 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143 || pmd_devmap(*src_pmd)) {
1145 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1146 err = copy_huge_pmd(dst_mm, src_mm,
1147 dst_pmd, src_pmd, addr, vma);
1154 if (pmd_none_or_clear_bad(src_pmd))
1156 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1159 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1163 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1164 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1165 unsigned long addr, unsigned long end)
1167 pud_t *src_pud, *dst_pud;
1170 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1173 src_pud = pud_offset(src_p4d, addr);
1175 next = pud_addr_end(addr, end);
1176 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1179 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1180 err = copy_huge_pud(dst_mm, src_mm,
1181 dst_pud, src_pud, addr, vma);
1188 if (pud_none_or_clear_bad(src_pud))
1190 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1193 } while (dst_pud++, src_pud++, addr = next, addr != end);
1197 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1198 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1199 unsigned long addr, unsigned long end)
1201 p4d_t *src_p4d, *dst_p4d;
1204 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1207 src_p4d = p4d_offset(src_pgd, addr);
1209 next = p4d_addr_end(addr, end);
1210 if (p4d_none_or_clear_bad(src_p4d))
1212 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1215 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1219 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1220 struct vm_area_struct *vma)
1222 pgd_t *src_pgd, *dst_pgd;
1224 unsigned long addr = vma->vm_start;
1225 unsigned long end = vma->vm_end;
1226 unsigned long mmun_start; /* For mmu_notifiers */
1227 unsigned long mmun_end; /* For mmu_notifiers */
1232 * Don't copy ptes where a page fault will fill them correctly.
1233 * Fork becomes much lighter when there are big shared or private
1234 * readonly mappings. The tradeoff is that copy_page_range is more
1235 * efficient than faulting.
1237 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1241 if (is_vm_hugetlb_page(vma))
1242 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1244 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1246 * We do not free on error cases below as remove_vma
1247 * gets called on error from higher level routine
1249 ret = track_pfn_copy(vma);
1255 * We need to invalidate the secondary MMU mappings only when
1256 * there could be a permission downgrade on the ptes of the
1257 * parent mm. And a permission downgrade will only happen if
1258 * is_cow_mapping() returns true.
1260 is_cow = is_cow_mapping(vma->vm_flags);
1264 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1268 dst_pgd = pgd_offset(dst_mm, addr);
1269 src_pgd = pgd_offset(src_mm, addr);
1271 next = pgd_addr_end(addr, end);
1272 if (pgd_none_or_clear_bad(src_pgd))
1274 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1275 vma, addr, next))) {
1279 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1282 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1286 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1287 struct vm_area_struct *vma, pmd_t *pmd,
1288 unsigned long addr, unsigned long end,
1289 struct zap_details *details)
1291 struct mm_struct *mm = tlb->mm;
1292 int force_flush = 0;
1293 int rss[NR_MM_COUNTERS];
1299 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1302 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1304 flush_tlb_batched_pending(mm);
1305 arch_enter_lazy_mmu_mode();
1308 if (pte_none(ptent))
1311 if (pte_present(ptent)) {
1314 page = _vm_normal_page(vma, addr, ptent, true);
1315 if (unlikely(details) && page) {
1317 * unmap_shared_mapping_pages() wants to
1318 * invalidate cache without truncating:
1319 * unmap shared but keep private pages.
1321 if (details->check_mapping &&
1322 details->check_mapping != page_rmapping(page))
1325 ptent = ptep_get_and_clear_full(mm, addr, pte,
1327 tlb_remove_tlb_entry(tlb, pte, addr);
1328 if (unlikely(!page))
1331 if (!PageAnon(page)) {
1332 if (pte_dirty(ptent)) {
1334 set_page_dirty(page);
1336 if (pte_young(ptent) &&
1337 likely(!(vma->vm_flags & VM_SEQ_READ)))
1338 mark_page_accessed(page);
1340 rss[mm_counter(page)]--;
1341 page_remove_rmap(page, false);
1342 if (unlikely(page_mapcount(page) < 0))
1343 print_bad_pte(vma, addr, ptent, page);
1344 if (unlikely(__tlb_remove_page(tlb, page))) {
1352 entry = pte_to_swp_entry(ptent);
1353 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1354 struct page *page = device_private_entry_to_page(entry);
1356 if (unlikely(details && details->check_mapping)) {
1358 * unmap_shared_mapping_pages() wants to
1359 * invalidate cache without truncating:
1360 * unmap shared but keep private pages.
1362 if (details->check_mapping !=
1363 page_rmapping(page))
1367 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1368 rss[mm_counter(page)]--;
1369 page_remove_rmap(page, false);
1374 /* If details->check_mapping, we leave swap entries. */
1375 if (unlikely(details))
1378 entry = pte_to_swp_entry(ptent);
1379 if (!non_swap_entry(entry))
1381 else if (is_migration_entry(entry)) {
1384 page = migration_entry_to_page(entry);
1385 rss[mm_counter(page)]--;
1387 if (unlikely(!free_swap_and_cache(entry)))
1388 print_bad_pte(vma, addr, ptent, NULL);
1389 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1390 } while (pte++, addr += PAGE_SIZE, addr != end);
1392 add_mm_rss_vec(mm, rss);
1393 arch_leave_lazy_mmu_mode();
1395 /* Do the actual TLB flush before dropping ptl */
1397 tlb_flush_mmu_tlbonly(tlb);
1398 pte_unmap_unlock(start_pte, ptl);
1401 * If we forced a TLB flush (either due to running out of
1402 * batch buffers or because we needed to flush dirty TLB
1403 * entries before releasing the ptl), free the batched
1404 * memory too. Restart if we didn't do everything.
1408 tlb_flush_mmu_free(tlb);
1416 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1417 struct vm_area_struct *vma, pud_t *pud,
1418 unsigned long addr, unsigned long end,
1419 struct zap_details *details)
1424 pmd = pmd_offset(pud, addr);
1426 next = pmd_addr_end(addr, end);
1427 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1428 if (next - addr != HPAGE_PMD_SIZE) {
1429 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1430 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1431 __split_huge_pmd(vma, pmd, addr, false, NULL);
1432 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1437 * Here there can be other concurrent MADV_DONTNEED or
1438 * trans huge page faults running, and if the pmd is
1439 * none or trans huge it can change under us. This is
1440 * because MADV_DONTNEED holds the mmap_sem in read
1443 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1445 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1448 } while (pmd++, addr = next, addr != end);
1453 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1454 struct vm_area_struct *vma, p4d_t *p4d,
1455 unsigned long addr, unsigned long end,
1456 struct zap_details *details)
1461 pud = pud_offset(p4d, addr);
1463 next = pud_addr_end(addr, end);
1464 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1465 if (next - addr != HPAGE_PUD_SIZE) {
1466 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1467 split_huge_pud(vma, pud, addr);
1468 } else if (zap_huge_pud(tlb, vma, pud, addr))
1472 if (pud_none_or_clear_bad(pud))
1474 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1477 } while (pud++, addr = next, addr != end);
1482 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1483 struct vm_area_struct *vma, pgd_t *pgd,
1484 unsigned long addr, unsigned long end,
1485 struct zap_details *details)
1490 p4d = p4d_offset(pgd, addr);
1492 next = p4d_addr_end(addr, end);
1493 if (p4d_none_or_clear_bad(p4d))
1495 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1496 } while (p4d++, addr = next, addr != end);
1501 void unmap_page_range(struct mmu_gather *tlb,
1502 struct vm_area_struct *vma,
1503 unsigned long addr, unsigned long end,
1504 struct zap_details *details)
1509 BUG_ON(addr >= end);
1510 tlb_start_vma(tlb, vma);
1511 pgd = pgd_offset(vma->vm_mm, addr);
1513 next = pgd_addr_end(addr, end);
1514 if (pgd_none_or_clear_bad(pgd))
1516 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1517 } while (pgd++, addr = next, addr != end);
1518 tlb_end_vma(tlb, vma);
1522 static void unmap_single_vma(struct mmu_gather *tlb,
1523 struct vm_area_struct *vma, unsigned long start_addr,
1524 unsigned long end_addr,
1525 struct zap_details *details)
1527 unsigned long start = max(vma->vm_start, start_addr);
1530 if (start >= vma->vm_end)
1532 end = min(vma->vm_end, end_addr);
1533 if (end <= vma->vm_start)
1537 uprobe_munmap(vma, start, end);
1539 if (unlikely(vma->vm_flags & VM_PFNMAP))
1540 untrack_pfn(vma, 0, 0);
1543 if (unlikely(is_vm_hugetlb_page(vma))) {
1545 * It is undesirable to test vma->vm_file as it
1546 * should be non-null for valid hugetlb area.
1547 * However, vm_file will be NULL in the error
1548 * cleanup path of mmap_region. When
1549 * hugetlbfs ->mmap method fails,
1550 * mmap_region() nullifies vma->vm_file
1551 * before calling this function to clean up.
1552 * Since no pte has actually been setup, it is
1553 * safe to do nothing in this case.
1556 i_mmap_lock_write(vma->vm_file->f_mapping);
1557 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1558 i_mmap_unlock_write(vma->vm_file->f_mapping);
1561 unmap_page_range(tlb, vma, start, end, details);
1566 * unmap_vmas - unmap a range of memory covered by a list of vma's
1567 * @tlb: address of the caller's struct mmu_gather
1568 * @vma: the starting vma
1569 * @start_addr: virtual address at which to start unmapping
1570 * @end_addr: virtual address at which to end unmapping
1572 * Unmap all pages in the vma list.
1574 * Only addresses between `start' and `end' will be unmapped.
1576 * The VMA list must be sorted in ascending virtual address order.
1578 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1579 * range after unmap_vmas() returns. So the only responsibility here is to
1580 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1581 * drops the lock and schedules.
1583 void unmap_vmas(struct mmu_gather *tlb,
1584 struct vm_area_struct *vma, unsigned long start_addr,
1585 unsigned long end_addr)
1587 struct mm_struct *mm = vma->vm_mm;
1589 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1590 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1591 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1592 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1596 * zap_page_range - remove user pages in a given range
1597 * @vma: vm_area_struct holding the applicable pages
1598 * @start: starting address of pages to zap
1599 * @size: number of bytes to zap
1601 * Caller must protect the VMA list
1603 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1606 struct mm_struct *mm = vma->vm_mm;
1607 struct mmu_gather tlb;
1608 unsigned long end = start + size;
1611 tlb_gather_mmu(&tlb, mm, start, end);
1612 update_hiwater_rss(mm);
1613 mmu_notifier_invalidate_range_start(mm, start, end);
1614 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1615 unmap_single_vma(&tlb, vma, start, end, NULL);
1618 * zap_page_range does not specify whether mmap_sem should be
1619 * held for read or write. That allows parallel zap_page_range
1620 * operations to unmap a PTE and defer a flush meaning that
1621 * this call observes pte_none and fails to flush the TLB.
1622 * Rather than adding a complex API, ensure that no stale
1623 * TLB entries exist when this call returns.
1625 flush_tlb_range(vma, start, end);
1628 mmu_notifier_invalidate_range_end(mm, start, end);
1629 tlb_finish_mmu(&tlb, start, end);
1633 * zap_page_range_single - remove user pages in a given range
1634 * @vma: vm_area_struct holding the applicable pages
1635 * @address: starting address of pages to zap
1636 * @size: number of bytes to zap
1637 * @details: details of shared cache invalidation
1639 * The range must fit into one VMA.
1641 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1642 unsigned long size, struct zap_details *details)
1644 struct mm_struct *mm = vma->vm_mm;
1645 struct mmu_gather tlb;
1646 unsigned long end = address + size;
1649 tlb_gather_mmu(&tlb, mm, address, end);
1650 update_hiwater_rss(mm);
1651 mmu_notifier_invalidate_range_start(mm, address, end);
1652 unmap_single_vma(&tlb, vma, address, end, details);
1653 mmu_notifier_invalidate_range_end(mm, address, end);
1654 tlb_finish_mmu(&tlb, address, end);
1658 * zap_vma_ptes - remove ptes mapping the vma
1659 * @vma: vm_area_struct holding ptes to be zapped
1660 * @address: starting address of pages to zap
1661 * @size: number of bytes to zap
1663 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1665 * The entire address range must be fully contained within the vma.
1667 * Returns 0 if successful.
1669 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1672 if (address < vma->vm_start || address + size > vma->vm_end ||
1673 !(vma->vm_flags & VM_PFNMAP))
1675 zap_page_range_single(vma, address, size, NULL);
1678 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1680 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1688 pgd = pgd_offset(mm, addr);
1689 p4d = p4d_alloc(mm, pgd, addr);
1692 pud = pud_alloc(mm, p4d, addr);
1695 pmd = pmd_alloc(mm, pud, addr);
1699 VM_BUG_ON(pmd_trans_huge(*pmd));
1700 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1704 * This is the old fallback for page remapping.
1706 * For historical reasons, it only allows reserved pages. Only
1707 * old drivers should use this, and they needed to mark their
1708 * pages reserved for the old functions anyway.
1710 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1711 struct page *page, pgprot_t prot)
1713 struct mm_struct *mm = vma->vm_mm;
1722 flush_dcache_page(page);
1723 pte = get_locked_pte(mm, addr, &ptl);
1727 if (!pte_none(*pte))
1730 /* Ok, finally just insert the thing.. */
1732 inc_mm_counter_fast(mm, mm_counter_file(page));
1733 page_add_file_rmap(page, false);
1734 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1737 pte_unmap_unlock(pte, ptl);
1740 pte_unmap_unlock(pte, ptl);
1746 * vm_insert_page - insert single page into user vma
1747 * @vma: user vma to map to
1748 * @addr: target user address of this page
1749 * @page: source kernel page
1751 * This allows drivers to insert individual pages they've allocated
1754 * The page has to be a nice clean _individual_ kernel allocation.
1755 * If you allocate a compound page, you need to have marked it as
1756 * such (__GFP_COMP), or manually just split the page up yourself
1757 * (see split_page()).
1759 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1760 * took an arbitrary page protection parameter. This doesn't allow
1761 * that. Your vma protection will have to be set up correctly, which
1762 * means that if you want a shared writable mapping, you'd better
1763 * ask for a shared writable mapping!
1765 * The page does not need to be reserved.
1767 * Usually this function is called from f_op->mmap() handler
1768 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1769 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1770 * function from other places, for example from page-fault handler.
1772 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1775 if (addr < vma->vm_start || addr >= vma->vm_end)
1777 if (!page_count(page))
1779 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1780 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1781 BUG_ON(vma->vm_flags & VM_PFNMAP);
1782 vma->vm_flags |= VM_MIXEDMAP;
1784 return insert_page(vma, addr, page, vma->vm_page_prot);
1786 EXPORT_SYMBOL(vm_insert_page);
1788 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1789 pfn_t pfn, pgprot_t prot, bool mkwrite)
1791 struct mm_struct *mm = vma->vm_mm;
1797 pte = get_locked_pte(mm, addr, &ptl);
1801 if (!pte_none(*pte)) {
1804 * For read faults on private mappings the PFN passed
1805 * in may not match the PFN we have mapped if the
1806 * mapped PFN is a writeable COW page. In the mkwrite
1807 * case we are creating a writable PTE for a shared
1808 * mapping and we expect the PFNs to match.
1810 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1818 /* Ok, finally just insert the thing.. */
1819 if (pfn_t_devmap(pfn))
1820 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1822 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1826 entry = pte_mkyoung(entry);
1827 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1830 set_pte_at(mm, addr, pte, entry);
1831 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1835 pte_unmap_unlock(pte, ptl);
1841 * vm_insert_pfn - insert single pfn into user vma
1842 * @vma: user vma to map to
1843 * @addr: target user address of this page
1844 * @pfn: source kernel pfn
1846 * Similar to vm_insert_page, this allows drivers to insert individual pages
1847 * they've allocated into a user vma. Same comments apply.
1849 * This function should only be called from a vm_ops->fault handler, and
1850 * in that case the handler should return NULL.
1852 * vma cannot be a COW mapping.
1854 * As this is called only for pages that do not currently exist, we
1855 * do not need to flush old virtual caches or the TLB.
1857 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1860 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1862 EXPORT_SYMBOL(vm_insert_pfn);
1865 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1866 * @vma: user vma to map to
1867 * @addr: target user address of this page
1868 * @pfn: source kernel pfn
1869 * @pgprot: pgprot flags for the inserted page
1871 * This is exactly like vm_insert_pfn, except that it allows drivers to
1872 * to override pgprot on a per-page basis.
1874 * This only makes sense for IO mappings, and it makes no sense for
1875 * cow mappings. In general, using multiple vmas is preferable;
1876 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1879 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1880 unsigned long pfn, pgprot_t pgprot)
1884 * Technically, architectures with pte_special can avoid all these
1885 * restrictions (same for remap_pfn_range). However we would like
1886 * consistency in testing and feature parity among all, so we should
1887 * try to keep these invariants in place for everybody.
1889 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1890 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1891 (VM_PFNMAP|VM_MIXEDMAP));
1892 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1893 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1895 if (addr < vma->vm_start || addr >= vma->vm_end)
1898 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1900 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1905 EXPORT_SYMBOL(vm_insert_pfn_prot);
1907 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1909 /* these checks mirror the abort conditions in vm_normal_page */
1910 if (vma->vm_flags & VM_MIXEDMAP)
1912 if (pfn_t_devmap(pfn))
1914 if (pfn_t_special(pfn))
1916 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1921 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1922 pfn_t pfn, bool mkwrite)
1924 pgprot_t pgprot = vma->vm_page_prot;
1926 BUG_ON(!vm_mixed_ok(vma, pfn));
1928 if (addr < vma->vm_start || addr >= vma->vm_end)
1931 track_pfn_insert(vma, &pgprot, pfn);
1934 * If we don't have pte special, then we have to use the pfn_valid()
1935 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1936 * refcount the page if pfn_valid is true (hence insert_page rather
1937 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1938 * without pte special, it would there be refcounted as a normal page.
1940 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1944 * At this point we are committed to insert_page()
1945 * regardless of whether the caller specified flags that
1946 * result in pfn_t_has_page() == false.
1948 page = pfn_to_page(pfn_t_to_pfn(pfn));
1949 return insert_page(vma, addr, page, pgprot);
1951 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1954 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1957 return __vm_insert_mixed(vma, addr, pfn, false);
1960 EXPORT_SYMBOL(vm_insert_mixed);
1962 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1965 return __vm_insert_mixed(vma, addr, pfn, true);
1967 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1970 * maps a range of physical memory into the requested pages. the old
1971 * mappings are removed. any references to nonexistent pages results
1972 * in null mappings (currently treated as "copy-on-access")
1974 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1975 unsigned long addr, unsigned long end,
1976 unsigned long pfn, pgprot_t prot)
1981 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1984 arch_enter_lazy_mmu_mode();
1986 BUG_ON(!pte_none(*pte));
1987 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1989 } while (pte++, addr += PAGE_SIZE, addr != end);
1990 arch_leave_lazy_mmu_mode();
1991 pte_unmap_unlock(pte - 1, ptl);
1995 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1996 unsigned long addr, unsigned long end,
1997 unsigned long pfn, pgprot_t prot)
2002 pfn -= addr >> PAGE_SHIFT;
2003 pmd = pmd_alloc(mm, pud, addr);
2006 VM_BUG_ON(pmd_trans_huge(*pmd));
2008 next = pmd_addr_end(addr, end);
2009 if (remap_pte_range(mm, pmd, addr, next,
2010 pfn + (addr >> PAGE_SHIFT), prot))
2012 } while (pmd++, addr = next, addr != end);
2016 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2017 unsigned long addr, unsigned long end,
2018 unsigned long pfn, pgprot_t prot)
2023 pfn -= addr >> PAGE_SHIFT;
2024 pud = pud_alloc(mm, p4d, addr);
2028 next = pud_addr_end(addr, end);
2029 if (remap_pmd_range(mm, pud, addr, next,
2030 pfn + (addr >> PAGE_SHIFT), prot))
2032 } while (pud++, addr = next, addr != end);
2036 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2037 unsigned long addr, unsigned long end,
2038 unsigned long pfn, pgprot_t prot)
2043 pfn -= addr >> PAGE_SHIFT;
2044 p4d = p4d_alloc(mm, pgd, addr);
2048 next = p4d_addr_end(addr, end);
2049 if (remap_pud_range(mm, p4d, addr, next,
2050 pfn + (addr >> PAGE_SHIFT), prot))
2052 } while (p4d++, addr = next, addr != end);
2057 * remap_pfn_range - remap kernel memory to userspace
2058 * @vma: user vma to map to
2059 * @addr: target user address to start at
2060 * @pfn: physical address of kernel memory
2061 * @size: size of map area
2062 * @prot: page protection flags for this mapping
2064 * Note: this is only safe if the mm semaphore is held when called.
2066 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2067 unsigned long pfn, unsigned long size, pgprot_t prot)
2071 unsigned long end = addr + PAGE_ALIGN(size);
2072 struct mm_struct *mm = vma->vm_mm;
2073 unsigned long remap_pfn = pfn;
2077 * Physically remapped pages are special. Tell the
2078 * rest of the world about it:
2079 * VM_IO tells people not to look at these pages
2080 * (accesses can have side effects).
2081 * VM_PFNMAP tells the core MM that the base pages are just
2082 * raw PFN mappings, and do not have a "struct page" associated
2085 * Disable vma merging and expanding with mremap().
2087 * Omit vma from core dump, even when VM_IO turned off.
2089 * There's a horrible special case to handle copy-on-write
2090 * behaviour that some programs depend on. We mark the "original"
2091 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2092 * See vm_normal_page() for details.
2094 if (is_cow_mapping(vma->vm_flags)) {
2095 if (addr != vma->vm_start || end != vma->vm_end)
2097 vma->vm_pgoff = pfn;
2100 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2104 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2106 BUG_ON(addr >= end);
2107 pfn -= addr >> PAGE_SHIFT;
2108 pgd = pgd_offset(mm, addr);
2109 flush_cache_range(vma, addr, end);
2111 next = pgd_addr_end(addr, end);
2112 err = remap_p4d_range(mm, pgd, addr, next,
2113 pfn + (addr >> PAGE_SHIFT), prot);
2116 } while (pgd++, addr = next, addr != end);
2119 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2123 EXPORT_SYMBOL(remap_pfn_range);
2126 * vm_iomap_memory - remap memory to userspace
2127 * @vma: user vma to map to
2128 * @start: start of area
2129 * @len: size of area
2131 * This is a simplified io_remap_pfn_range() for common driver use. The
2132 * driver just needs to give us the physical memory range to be mapped,
2133 * we'll figure out the rest from the vma information.
2135 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2136 * whatever write-combining details or similar.
2138 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2140 unsigned long vm_len, pfn, pages;
2142 /* Check that the physical memory area passed in looks valid */
2143 if (start + len < start)
2146 * You *really* shouldn't map things that aren't page-aligned,
2147 * but we've historically allowed it because IO memory might
2148 * just have smaller alignment.
2150 len += start & ~PAGE_MASK;
2151 pfn = start >> PAGE_SHIFT;
2152 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2153 if (pfn + pages < pfn)
2156 /* We start the mapping 'vm_pgoff' pages into the area */
2157 if (vma->vm_pgoff > pages)
2159 pfn += vma->vm_pgoff;
2160 pages -= vma->vm_pgoff;
2162 /* Can we fit all of the mapping? */
2163 vm_len = vma->vm_end - vma->vm_start;
2164 if (vm_len >> PAGE_SHIFT > pages)
2167 /* Ok, let it rip */
2168 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2170 EXPORT_SYMBOL(vm_iomap_memory);
2172 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2173 unsigned long addr, unsigned long end,
2174 pte_fn_t fn, void *data)
2179 spinlock_t *uninitialized_var(ptl);
2181 pte = (mm == &init_mm) ?
2182 pte_alloc_kernel(pmd, addr) :
2183 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2187 BUG_ON(pmd_huge(*pmd));
2189 arch_enter_lazy_mmu_mode();
2191 token = pmd_pgtable(*pmd);
2194 err = fn(pte++, token, addr, data);
2197 } while (addr += PAGE_SIZE, addr != end);
2199 arch_leave_lazy_mmu_mode();
2202 pte_unmap_unlock(pte-1, ptl);
2206 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2207 unsigned long addr, unsigned long end,
2208 pte_fn_t fn, void *data)
2214 BUG_ON(pud_huge(*pud));
2216 pmd = pmd_alloc(mm, pud, addr);
2220 next = pmd_addr_end(addr, end);
2221 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2224 } while (pmd++, addr = next, addr != end);
2228 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2229 unsigned long addr, unsigned long end,
2230 pte_fn_t fn, void *data)
2236 pud = pud_alloc(mm, p4d, addr);
2240 next = pud_addr_end(addr, end);
2241 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2244 } while (pud++, addr = next, addr != end);
2248 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2249 unsigned long addr, unsigned long end,
2250 pte_fn_t fn, void *data)
2256 p4d = p4d_alloc(mm, pgd, addr);
2260 next = p4d_addr_end(addr, end);
2261 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2264 } while (p4d++, addr = next, addr != end);
2269 * Scan a region of virtual memory, filling in page tables as necessary
2270 * and calling a provided function on each leaf page table.
2272 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2273 unsigned long size, pte_fn_t fn, void *data)
2277 unsigned long end = addr + size;
2280 if (WARN_ON(addr >= end))
2283 pgd = pgd_offset(mm, addr);
2285 next = pgd_addr_end(addr, end);
2286 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2289 } while (pgd++, addr = next, addr != end);
2293 EXPORT_SYMBOL_GPL(apply_to_page_range);
2296 * handle_pte_fault chooses page fault handler according to an entry which was
2297 * read non-atomically. Before making any commitment, on those architectures
2298 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2299 * parts, do_swap_page must check under lock before unmapping the pte and
2300 * proceeding (but do_wp_page is only called after already making such a check;
2301 * and do_anonymous_page can safely check later on).
2303 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2304 pte_t *page_table, pte_t orig_pte)
2307 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2308 if (sizeof(pte_t) > sizeof(unsigned long)) {
2309 spinlock_t *ptl = pte_lockptr(mm, pmd);
2311 same = pte_same(*page_table, orig_pte);
2315 pte_unmap(page_table);
2319 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2321 debug_dma_assert_idle(src);
2324 * If the source page was a PFN mapping, we don't have
2325 * a "struct page" for it. We do a best-effort copy by
2326 * just copying from the original user address. If that
2327 * fails, we just zero-fill it. Live with it.
2329 if (unlikely(!src)) {
2330 void *kaddr = kmap_atomic(dst);
2331 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2334 * This really shouldn't fail, because the page is there
2335 * in the page tables. But it might just be unreadable,
2336 * in which case we just give up and fill the result with
2339 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2341 kunmap_atomic(kaddr);
2342 flush_dcache_page(dst);
2344 copy_user_highpage(dst, src, va, vma);
2347 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2349 struct file *vm_file = vma->vm_file;
2352 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2355 * Special mappings (e.g. VDSO) do not have any file so fake
2356 * a default GFP_KERNEL for them.
2362 * Notify the address space that the page is about to become writable so that
2363 * it can prohibit this or wait for the page to get into an appropriate state.
2365 * We do this without the lock held, so that it can sleep if it needs to.
2367 static int do_page_mkwrite(struct vm_fault *vmf)
2370 struct page *page = vmf->page;
2371 unsigned int old_flags = vmf->flags;
2373 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2375 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2376 /* Restore original flags so that caller is not surprised */
2377 vmf->flags = old_flags;
2378 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2380 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2382 if (!page->mapping) {
2384 return 0; /* retry */
2386 ret |= VM_FAULT_LOCKED;
2388 VM_BUG_ON_PAGE(!PageLocked(page), page);
2393 * Handle dirtying of a page in shared file mapping on a write fault.
2395 * The function expects the page to be locked and unlocks it.
2397 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2400 struct address_space *mapping;
2402 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2404 dirtied = set_page_dirty(page);
2405 VM_BUG_ON_PAGE(PageAnon(page), page);
2407 * Take a local copy of the address_space - page.mapping may be zeroed
2408 * by truncate after unlock_page(). The address_space itself remains
2409 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2410 * release semantics to prevent the compiler from undoing this copying.
2412 mapping = page_rmapping(page);
2415 if ((dirtied || page_mkwrite) && mapping) {
2417 * Some device drivers do not set page.mapping
2418 * but still dirty their pages
2420 balance_dirty_pages_ratelimited(mapping);
2424 file_update_time(vma->vm_file);
2428 * Handle write page faults for pages that can be reused in the current vma
2430 * This can happen either due to the mapping being with the VM_SHARED flag,
2431 * or due to us being the last reference standing to the page. In either
2432 * case, all we need to do here is to mark the page as writable and update
2433 * any related book-keeping.
2435 static inline void wp_page_reuse(struct vm_fault *vmf)
2436 __releases(vmf->ptl)
2438 struct vm_area_struct *vma = vmf->vma;
2439 struct page *page = vmf->page;
2442 * Clear the pages cpupid information as the existing
2443 * information potentially belongs to a now completely
2444 * unrelated process.
2447 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2449 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2450 entry = pte_mkyoung(vmf->orig_pte);
2451 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2452 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2453 update_mmu_cache(vma, vmf->address, vmf->pte);
2454 pte_unmap_unlock(vmf->pte, vmf->ptl);
2458 * Handle the case of a page which we actually need to copy to a new page.
2460 * Called with mmap_sem locked and the old page referenced, but
2461 * without the ptl held.
2463 * High level logic flow:
2465 * - Allocate a page, copy the content of the old page to the new one.
2466 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2467 * - Take the PTL. If the pte changed, bail out and release the allocated page
2468 * - If the pte is still the way we remember it, update the page table and all
2469 * relevant references. This includes dropping the reference the page-table
2470 * held to the old page, as well as updating the rmap.
2471 * - In any case, unlock the PTL and drop the reference we took to the old page.
2473 static int wp_page_copy(struct vm_fault *vmf)
2475 struct vm_area_struct *vma = vmf->vma;
2476 struct mm_struct *mm = vma->vm_mm;
2477 struct page *old_page = vmf->page;
2478 struct page *new_page = NULL;
2480 int page_copied = 0;
2481 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2482 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2483 struct mem_cgroup *memcg;
2485 if (unlikely(anon_vma_prepare(vma)))
2488 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2489 new_page = alloc_zeroed_user_highpage_movable(vma,
2494 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2498 cow_user_page(new_page, old_page, vmf->address, vma);
2501 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2504 __SetPageUptodate(new_page);
2506 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2509 * Re-check the pte - we dropped the lock
2511 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2512 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2514 if (!PageAnon(old_page)) {
2515 dec_mm_counter_fast(mm,
2516 mm_counter_file(old_page));
2517 inc_mm_counter_fast(mm, MM_ANONPAGES);
2520 inc_mm_counter_fast(mm, MM_ANONPAGES);
2522 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2523 entry = mk_pte(new_page, vma->vm_page_prot);
2524 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2526 * Clear the pte entry and flush it first, before updating the
2527 * pte with the new entry. This will avoid a race condition
2528 * seen in the presence of one thread doing SMC and another
2531 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2532 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2533 mem_cgroup_commit_charge(new_page, memcg, false, false);
2534 lru_cache_add_active_or_unevictable(new_page, vma);
2536 * We call the notify macro here because, when using secondary
2537 * mmu page tables (such as kvm shadow page tables), we want the
2538 * new page to be mapped directly into the secondary page table.
2540 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2541 update_mmu_cache(vma, vmf->address, vmf->pte);
2544 * Only after switching the pte to the new page may
2545 * we remove the mapcount here. Otherwise another
2546 * process may come and find the rmap count decremented
2547 * before the pte is switched to the new page, and
2548 * "reuse" the old page writing into it while our pte
2549 * here still points into it and can be read by other
2552 * The critical issue is to order this
2553 * page_remove_rmap with the ptp_clear_flush above.
2554 * Those stores are ordered by (if nothing else,)
2555 * the barrier present in the atomic_add_negative
2556 * in page_remove_rmap.
2558 * Then the TLB flush in ptep_clear_flush ensures that
2559 * no process can access the old page before the
2560 * decremented mapcount is visible. And the old page
2561 * cannot be reused until after the decremented
2562 * mapcount is visible. So transitively, TLBs to
2563 * old page will be flushed before it can be reused.
2565 page_remove_rmap(old_page, false);
2568 /* Free the old page.. */
2569 new_page = old_page;
2572 mem_cgroup_cancel_charge(new_page, memcg, false);
2578 pte_unmap_unlock(vmf->pte, vmf->ptl);
2580 * No need to double call mmu_notifier->invalidate_range() callback as
2581 * the above ptep_clear_flush_notify() did already call it.
2583 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2586 * Don't let another task, with possibly unlocked vma,
2587 * keep the mlocked page.
2589 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2590 lock_page(old_page); /* LRU manipulation */
2591 if (PageMlocked(old_page))
2592 munlock_vma_page(old_page);
2593 unlock_page(old_page);
2597 return page_copied ? VM_FAULT_WRITE : 0;
2603 return VM_FAULT_OOM;
2607 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2608 * writeable once the page is prepared
2610 * @vmf: structure describing the fault
2612 * This function handles all that is needed to finish a write page fault in a
2613 * shared mapping due to PTE being read-only once the mapped page is prepared.
2614 * It handles locking of PTE and modifying it. The function returns
2615 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2618 * The function expects the page to be locked or other protection against
2619 * concurrent faults / writeback (such as DAX radix tree locks).
2621 int finish_mkwrite_fault(struct vm_fault *vmf)
2623 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2624 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2627 * We might have raced with another page fault while we released the
2628 * pte_offset_map_lock.
2630 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2631 pte_unmap_unlock(vmf->pte, vmf->ptl);
2632 return VM_FAULT_NOPAGE;
2639 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2642 static int wp_pfn_shared(struct vm_fault *vmf)
2644 struct vm_area_struct *vma = vmf->vma;
2646 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2649 pte_unmap_unlock(vmf->pte, vmf->ptl);
2650 vmf->flags |= FAULT_FLAG_MKWRITE;
2651 ret = vma->vm_ops->pfn_mkwrite(vmf);
2652 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2654 return finish_mkwrite_fault(vmf);
2657 return VM_FAULT_WRITE;
2660 static int wp_page_shared(struct vm_fault *vmf)
2661 __releases(vmf->ptl)
2663 struct vm_area_struct *vma = vmf->vma;
2665 get_page(vmf->page);
2667 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2670 pte_unmap_unlock(vmf->pte, vmf->ptl);
2671 tmp = do_page_mkwrite(vmf);
2672 if (unlikely(!tmp || (tmp &
2673 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2674 put_page(vmf->page);
2677 tmp = finish_mkwrite_fault(vmf);
2678 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2679 unlock_page(vmf->page);
2680 put_page(vmf->page);
2685 lock_page(vmf->page);
2687 fault_dirty_shared_page(vma, vmf->page);
2688 put_page(vmf->page);
2690 return VM_FAULT_WRITE;
2694 * This routine handles present pages, when users try to write
2695 * to a shared page. It is done by copying the page to a new address
2696 * and decrementing the shared-page counter for the old page.
2698 * Note that this routine assumes that the protection checks have been
2699 * done by the caller (the low-level page fault routine in most cases).
2700 * Thus we can safely just mark it writable once we've done any necessary
2703 * We also mark the page dirty at this point even though the page will
2704 * change only once the write actually happens. This avoids a few races,
2705 * and potentially makes it more efficient.
2707 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2708 * but allow concurrent faults), with pte both mapped and locked.
2709 * We return with mmap_sem still held, but pte unmapped and unlocked.
2711 static int do_wp_page(struct vm_fault *vmf)
2712 __releases(vmf->ptl)
2714 struct vm_area_struct *vma = vmf->vma;
2716 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2719 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2722 * We should not cow pages in a shared writeable mapping.
2723 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2725 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2726 (VM_WRITE|VM_SHARED))
2727 return wp_pfn_shared(vmf);
2729 pte_unmap_unlock(vmf->pte, vmf->ptl);
2730 return wp_page_copy(vmf);
2734 * Take out anonymous pages first, anonymous shared vmas are
2735 * not dirty accountable.
2737 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2738 int total_map_swapcount;
2739 if (!trylock_page(vmf->page)) {
2740 get_page(vmf->page);
2741 pte_unmap_unlock(vmf->pte, vmf->ptl);
2742 lock_page(vmf->page);
2743 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2744 vmf->address, &vmf->ptl);
2745 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2746 unlock_page(vmf->page);
2747 pte_unmap_unlock(vmf->pte, vmf->ptl);
2748 put_page(vmf->page);
2751 put_page(vmf->page);
2753 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2754 if (total_map_swapcount == 1) {
2756 * The page is all ours. Move it to
2757 * our anon_vma so the rmap code will
2758 * not search our parent or siblings.
2759 * Protected against the rmap code by
2762 page_move_anon_rmap(vmf->page, vma);
2764 unlock_page(vmf->page);
2766 return VM_FAULT_WRITE;
2768 unlock_page(vmf->page);
2769 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2770 (VM_WRITE|VM_SHARED))) {
2771 return wp_page_shared(vmf);
2775 * Ok, we need to copy. Oh, well..
2777 get_page(vmf->page);
2779 pte_unmap_unlock(vmf->pte, vmf->ptl);
2780 return wp_page_copy(vmf);
2783 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2784 unsigned long start_addr, unsigned long end_addr,
2785 struct zap_details *details)
2787 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2790 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2791 struct zap_details *details)
2793 struct vm_area_struct *vma;
2794 pgoff_t vba, vea, zba, zea;
2796 vma_interval_tree_foreach(vma, root,
2797 details->first_index, details->last_index) {
2799 vba = vma->vm_pgoff;
2800 vea = vba + vma_pages(vma) - 1;
2801 zba = details->first_index;
2804 zea = details->last_index;
2808 unmap_mapping_range_vma(vma,
2809 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2810 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2816 * unmap_mapping_pages() - Unmap pages from processes.
2817 * @mapping: The address space containing pages to be unmapped.
2818 * @start: Index of first page to be unmapped.
2819 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2820 * @even_cows: Whether to unmap even private COWed pages.
2822 * Unmap the pages in this address space from any userspace process which
2823 * has them mmaped. Generally, you want to remove COWed pages as well when
2824 * a file is being truncated, but not when invalidating pages from the page
2827 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2828 pgoff_t nr, bool even_cows)
2830 struct zap_details details = { };
2832 details.check_mapping = even_cows ? NULL : mapping;
2833 details.first_index = start;
2834 details.last_index = start + nr - 1;
2835 if (details.last_index < details.first_index)
2836 details.last_index = ULONG_MAX;
2838 i_mmap_lock_write(mapping);
2839 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2840 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2841 i_mmap_unlock_write(mapping);
2845 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2846 * address_space corresponding to the specified byte range in the underlying
2849 * @mapping: the address space containing mmaps to be unmapped.
2850 * @holebegin: byte in first page to unmap, relative to the start of
2851 * the underlying file. This will be rounded down to a PAGE_SIZE
2852 * boundary. Note that this is different from truncate_pagecache(), which
2853 * must keep the partial page. In contrast, we must get rid of
2855 * @holelen: size of prospective hole in bytes. This will be rounded
2856 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2858 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2859 * but 0 when invalidating pagecache, don't throw away private data.
2861 void unmap_mapping_range(struct address_space *mapping,
2862 loff_t const holebegin, loff_t const holelen, int even_cows)
2864 pgoff_t hba = holebegin >> PAGE_SHIFT;
2865 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2867 /* Check for overflow. */
2868 if (sizeof(holelen) > sizeof(hlen)) {
2870 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2871 if (holeend & ~(long long)ULONG_MAX)
2872 hlen = ULONG_MAX - hba + 1;
2875 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2877 EXPORT_SYMBOL(unmap_mapping_range);
2880 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2881 * but allow concurrent faults), and pte mapped but not yet locked.
2882 * We return with pte unmapped and unlocked.
2884 * We return with the mmap_sem locked or unlocked in the same cases
2885 * as does filemap_fault().
2887 int do_swap_page(struct vm_fault *vmf)
2889 struct vm_area_struct *vma = vmf->vma;
2890 struct page *page = NULL, *swapcache = NULL;
2891 struct mem_cgroup *memcg;
2892 struct vma_swap_readahead swap_ra;
2898 bool vma_readahead = swap_use_vma_readahead();
2900 if (vma_readahead) {
2901 page = swap_readahead_detect(vmf, &swap_ra);
2905 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2911 entry = pte_to_swp_entry(vmf->orig_pte);
2912 if (unlikely(non_swap_entry(entry))) {
2913 if (is_migration_entry(entry)) {
2914 migration_entry_wait(vma->vm_mm, vmf->pmd,
2916 } else if (is_device_private_entry(entry)) {
2918 * For un-addressable device memory we call the pgmap
2919 * fault handler callback. The callback must migrate
2920 * the page back to some CPU accessible page.
2922 ret = device_private_entry_fault(vma, vmf->address, entry,
2923 vmf->flags, vmf->pmd);
2924 } else if (is_hwpoison_entry(entry)) {
2925 ret = VM_FAULT_HWPOISON;
2927 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2928 ret = VM_FAULT_SIGBUS;
2934 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2936 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2942 struct swap_info_struct *si = swp_swap_info(entry);
2944 if (si->flags & SWP_SYNCHRONOUS_IO &&
2945 __swap_count(si, entry) == 1) {
2946 /* skip swapcache */
2947 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2949 __SetPageLocked(page);
2950 __SetPageSwapBacked(page);
2951 set_page_private(page, entry.val);
2952 lru_cache_add_anon(page);
2953 swap_readpage(page, true);
2957 page = do_swap_page_readahead(entry,
2958 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2960 page = swapin_readahead(entry,
2961 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2967 * Back out if somebody else faulted in this pte
2968 * while we released the pte lock.
2970 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2971 vmf->address, &vmf->ptl);
2972 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2974 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2978 /* Had to read the page from swap area: Major fault */
2979 ret = VM_FAULT_MAJOR;
2980 count_vm_event(PGMAJFAULT);
2981 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2982 } else if (PageHWPoison(page)) {
2984 * hwpoisoned dirty swapcache pages are kept for killing
2985 * owner processes (which may be unknown at hwpoison time)
2987 ret = VM_FAULT_HWPOISON;
2988 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2993 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2995 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2997 ret |= VM_FAULT_RETRY;
3002 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3003 * release the swapcache from under us. The page pin, and pte_same
3004 * test below, are not enough to exclude that. Even if it is still
3005 * swapcache, we need to check that the page's swap has not changed.
3007 if (unlikely((!PageSwapCache(page) ||
3008 page_private(page) != entry.val)) && swapcache)
3011 page = ksm_might_need_to_copy(page, vma, vmf->address);
3012 if (unlikely(!page)) {
3018 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3025 * Back out if somebody else already faulted in this pte.
3027 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3029 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3032 if (unlikely(!PageUptodate(page))) {
3033 ret = VM_FAULT_SIGBUS;
3038 * The page isn't present yet, go ahead with the fault.
3040 * Be careful about the sequence of operations here.
3041 * To get its accounting right, reuse_swap_page() must be called
3042 * while the page is counted on swap but not yet in mapcount i.e.
3043 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3044 * must be called after the swap_free(), or it will never succeed.
3047 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3048 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3049 pte = mk_pte(page, vma->vm_page_prot);
3050 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3051 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3052 vmf->flags &= ~FAULT_FLAG_WRITE;
3053 ret |= VM_FAULT_WRITE;
3054 exclusive = RMAP_EXCLUSIVE;
3056 flush_icache_page(vma, page);
3057 if (pte_swp_soft_dirty(vmf->orig_pte))
3058 pte = pte_mksoft_dirty(pte);
3059 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3060 vmf->orig_pte = pte;
3062 /* ksm created a completely new copy */
3063 if (unlikely(page != swapcache && swapcache)) {
3064 page_add_new_anon_rmap(page, vma, vmf->address, false);
3065 mem_cgroup_commit_charge(page, memcg, false, false);
3066 lru_cache_add_active_or_unevictable(page, vma);
3068 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3069 mem_cgroup_commit_charge(page, memcg, true, false);
3070 activate_page(page);
3074 if (mem_cgroup_swap_full(page) ||
3075 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3076 try_to_free_swap(page);
3078 if (page != swapcache && swapcache) {
3080 * Hold the lock to avoid the swap entry to be reused
3081 * until we take the PT lock for the pte_same() check
3082 * (to avoid false positives from pte_same). For
3083 * further safety release the lock after the swap_free
3084 * so that the swap count won't change under a
3085 * parallel locked swapcache.
3087 unlock_page(swapcache);
3088 put_page(swapcache);
3091 if (vmf->flags & FAULT_FLAG_WRITE) {
3092 ret |= do_wp_page(vmf);
3093 if (ret & VM_FAULT_ERROR)
3094 ret &= VM_FAULT_ERROR;
3098 /* No need to invalidate - it was non-present before */
3099 update_mmu_cache(vma, vmf->address, vmf->pte);
3101 pte_unmap_unlock(vmf->pte, vmf->ptl);
3105 mem_cgroup_cancel_charge(page, memcg, false);
3106 pte_unmap_unlock(vmf->pte, vmf->ptl);
3111 if (page != swapcache && swapcache) {
3112 unlock_page(swapcache);
3113 put_page(swapcache);
3119 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3120 * but allow concurrent faults), and pte mapped but not yet locked.
3121 * We return with mmap_sem still held, but pte unmapped and unlocked.
3123 static int do_anonymous_page(struct vm_fault *vmf)
3125 struct vm_area_struct *vma = vmf->vma;
3126 struct mem_cgroup *memcg;
3131 /* File mapping without ->vm_ops ? */
3132 if (vma->vm_flags & VM_SHARED)
3133 return VM_FAULT_SIGBUS;
3136 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3137 * pte_offset_map() on pmds where a huge pmd might be created
3138 * from a different thread.
3140 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3141 * parallel threads are excluded by other means.
3143 * Here we only have down_read(mmap_sem).
3145 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3146 return VM_FAULT_OOM;
3148 /* See the comment in pte_alloc_one_map() */
3149 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3152 /* Use the zero-page for reads */
3153 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3154 !mm_forbids_zeropage(vma->vm_mm)) {
3155 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3156 vma->vm_page_prot));
3157 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3158 vmf->address, &vmf->ptl);
3159 if (!pte_none(*vmf->pte))
3161 ret = check_stable_address_space(vma->vm_mm);
3164 /* Deliver the page fault to userland, check inside PT lock */
3165 if (userfaultfd_missing(vma)) {
3166 pte_unmap_unlock(vmf->pte, vmf->ptl);
3167 return handle_userfault(vmf, VM_UFFD_MISSING);
3172 /* Allocate our own private page. */
3173 if (unlikely(anon_vma_prepare(vma)))
3175 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3179 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3183 * The memory barrier inside __SetPageUptodate makes sure that
3184 * preceeding stores to the page contents become visible before
3185 * the set_pte_at() write.
3187 __SetPageUptodate(page);
3189 entry = mk_pte(page, vma->vm_page_prot);
3190 if (vma->vm_flags & VM_WRITE)
3191 entry = pte_mkwrite(pte_mkdirty(entry));
3193 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3195 if (!pte_none(*vmf->pte))
3198 ret = check_stable_address_space(vma->vm_mm);
3202 /* Deliver the page fault to userland, check inside PT lock */
3203 if (userfaultfd_missing(vma)) {
3204 pte_unmap_unlock(vmf->pte, vmf->ptl);
3205 mem_cgroup_cancel_charge(page, memcg, false);
3207 return handle_userfault(vmf, VM_UFFD_MISSING);
3210 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3211 page_add_new_anon_rmap(page, vma, vmf->address, false);
3212 mem_cgroup_commit_charge(page, memcg, false, false);
3213 lru_cache_add_active_or_unevictable(page, vma);
3215 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3217 /* No need to invalidate - it was non-present before */
3218 update_mmu_cache(vma, vmf->address, vmf->pte);
3220 pte_unmap_unlock(vmf->pte, vmf->ptl);
3223 mem_cgroup_cancel_charge(page, memcg, false);
3229 return VM_FAULT_OOM;
3233 * The mmap_sem must have been held on entry, and may have been
3234 * released depending on flags and vma->vm_ops->fault() return value.
3235 * See filemap_fault() and __lock_page_retry().
3237 static int __do_fault(struct vm_fault *vmf)
3239 struct vm_area_struct *vma = vmf->vma;
3242 ret = vma->vm_ops->fault(vmf);
3243 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3244 VM_FAULT_DONE_COW)))
3247 if (unlikely(PageHWPoison(vmf->page))) {
3248 if (ret & VM_FAULT_LOCKED)
3249 unlock_page(vmf->page);
3250 put_page(vmf->page);
3252 return VM_FAULT_HWPOISON;
3255 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3256 lock_page(vmf->page);
3258 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3264 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3265 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3266 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3267 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3269 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3271 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3274 static int pte_alloc_one_map(struct vm_fault *vmf)
3276 struct vm_area_struct *vma = vmf->vma;
3278 if (!pmd_none(*vmf->pmd))
3280 if (vmf->prealloc_pte) {
3281 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3282 if (unlikely(!pmd_none(*vmf->pmd))) {
3283 spin_unlock(vmf->ptl);
3287 mm_inc_nr_ptes(vma->vm_mm);
3288 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3289 spin_unlock(vmf->ptl);
3290 vmf->prealloc_pte = NULL;
3291 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3292 return VM_FAULT_OOM;
3296 * If a huge pmd materialized under us just retry later. Use
3297 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3298 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3299 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3300 * running immediately after a huge pmd fault in a different thread of
3301 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3302 * All we have to ensure is that it is a regular pmd that we can walk
3303 * with pte_offset_map() and we can do that through an atomic read in
3304 * C, which is what pmd_trans_unstable() provides.
3306 if (pmd_devmap_trans_unstable(vmf->pmd))
3307 return VM_FAULT_NOPAGE;
3310 * At this point we know that our vmf->pmd points to a page of ptes
3311 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3312 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3313 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3314 * be valid and we will re-check to make sure the vmf->pte isn't
3315 * pte_none() under vmf->ptl protection when we return to
3318 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3323 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3325 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3326 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3327 unsigned long haddr)
3329 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3330 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3332 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3337 static void deposit_prealloc_pte(struct vm_fault *vmf)
3339 struct vm_area_struct *vma = vmf->vma;
3341 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3343 * We are going to consume the prealloc table,
3344 * count that as nr_ptes.
3346 mm_inc_nr_ptes(vma->vm_mm);
3347 vmf->prealloc_pte = NULL;
3350 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3352 struct vm_area_struct *vma = vmf->vma;
3353 bool write = vmf->flags & FAULT_FLAG_WRITE;
3354 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3358 if (!transhuge_vma_suitable(vma, haddr))
3359 return VM_FAULT_FALLBACK;
3361 ret = VM_FAULT_FALLBACK;
3362 page = compound_head(page);
3365 * Archs like ppc64 need additonal space to store information
3366 * related to pte entry. Use the preallocated table for that.
3368 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3369 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3370 if (!vmf->prealloc_pte)
3371 return VM_FAULT_OOM;
3372 smp_wmb(); /* See comment in __pte_alloc() */
3375 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3376 if (unlikely(!pmd_none(*vmf->pmd)))
3379 for (i = 0; i < HPAGE_PMD_NR; i++)
3380 flush_icache_page(vma, page + i);
3382 entry = mk_huge_pmd(page, vma->vm_page_prot);
3384 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3386 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3387 page_add_file_rmap(page, true);
3389 * deposit and withdraw with pmd lock held
3391 if (arch_needs_pgtable_deposit())
3392 deposit_prealloc_pte(vmf);
3394 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3396 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3398 /* fault is handled */
3400 count_vm_event(THP_FILE_MAPPED);
3402 spin_unlock(vmf->ptl);
3406 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3414 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3415 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3417 * @vmf: fault environment
3418 * @memcg: memcg to charge page (only for private mappings)
3419 * @page: page to map
3421 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3424 * Target users are page handler itself and implementations of
3425 * vm_ops->map_pages.
3427 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3430 struct vm_area_struct *vma = vmf->vma;
3431 bool write = vmf->flags & FAULT_FLAG_WRITE;
3435 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3436 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3438 VM_BUG_ON_PAGE(memcg, page);
3440 ret = do_set_pmd(vmf, page);
3441 if (ret != VM_FAULT_FALLBACK)
3446 ret = pte_alloc_one_map(vmf);
3451 /* Re-check under ptl */
3452 if (unlikely(!pte_none(*vmf->pte)))
3453 return VM_FAULT_NOPAGE;
3455 flush_icache_page(vma, page);
3456 entry = mk_pte(page, vma->vm_page_prot);
3458 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3459 /* copy-on-write page */
3460 if (write && !(vma->vm_flags & VM_SHARED)) {
3461 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3462 page_add_new_anon_rmap(page, vma, vmf->address, false);
3463 mem_cgroup_commit_charge(page, memcg, false, false);
3464 lru_cache_add_active_or_unevictable(page, vma);
3466 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3467 page_add_file_rmap(page, false);
3469 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3471 /* no need to invalidate: a not-present page won't be cached */
3472 update_mmu_cache(vma, vmf->address, vmf->pte);
3479 * finish_fault - finish page fault once we have prepared the page to fault
3481 * @vmf: structure describing the fault
3483 * This function handles all that is needed to finish a page fault once the
3484 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3485 * given page, adds reverse page mapping, handles memcg charges and LRU
3486 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3489 * The function expects the page to be locked and on success it consumes a
3490 * reference of a page being mapped (for the PTE which maps it).
3492 int finish_fault(struct vm_fault *vmf)
3497 /* Did we COW the page? */
3498 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3499 !(vmf->vma->vm_flags & VM_SHARED))
3500 page = vmf->cow_page;
3505 * check even for read faults because we might have lost our CoWed
3508 if (!(vmf->vma->vm_flags & VM_SHARED))
3509 ret = check_stable_address_space(vmf->vma->vm_mm);
3511 ret = alloc_set_pte(vmf, vmf->memcg, page);
3513 pte_unmap_unlock(vmf->pte, vmf->ptl);
3517 static unsigned long fault_around_bytes __read_mostly =
3518 rounddown_pow_of_two(65536);
3520 #ifdef CONFIG_DEBUG_FS
3521 static int fault_around_bytes_get(void *data, u64 *val)
3523 *val = fault_around_bytes;
3528 * fault_around_bytes must be rounded down to the nearest page order as it's
3529 * what do_fault_around() expects to see.
3531 static int fault_around_bytes_set(void *data, u64 val)
3533 if (val / PAGE_SIZE > PTRS_PER_PTE)
3535 if (val > PAGE_SIZE)
3536 fault_around_bytes = rounddown_pow_of_two(val);
3538 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3541 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3542 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3544 static int __init fault_around_debugfs(void)
3548 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3549 &fault_around_bytes_fops);
3551 pr_warn("Failed to create fault_around_bytes in debugfs");
3554 late_initcall(fault_around_debugfs);
3558 * do_fault_around() tries to map few pages around the fault address. The hope
3559 * is that the pages will be needed soon and this will lower the number of
3562 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3563 * not ready to be mapped: not up-to-date, locked, etc.
3565 * This function is called with the page table lock taken. In the split ptlock
3566 * case the page table lock only protects only those entries which belong to
3567 * the page table corresponding to the fault address.
3569 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3572 * fault_around_bytes defines how many bytes we'll try to map.
3573 * do_fault_around() expects it to be set to a power of two less than or equal
3576 * The virtual address of the area that we map is naturally aligned to
3577 * fault_around_bytes rounded down to the machine page size
3578 * (and therefore to page order). This way it's easier to guarantee
3579 * that we don't cross page table boundaries.
3581 static int do_fault_around(struct vm_fault *vmf)
3583 unsigned long address = vmf->address, nr_pages, mask;
3584 pgoff_t start_pgoff = vmf->pgoff;
3588 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3589 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3591 vmf->address = max(address & mask, vmf->vma->vm_start);
3592 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3596 * end_pgoff is either the end of the page table, the end of
3597 * the vma or nr_pages from start_pgoff, depending what is nearest.
3599 end_pgoff = start_pgoff -
3600 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3602 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3603 start_pgoff + nr_pages - 1);
3605 if (pmd_none(*vmf->pmd)) {
3606 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3608 if (!vmf->prealloc_pte)
3610 smp_wmb(); /* See comment in __pte_alloc() */
3613 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3615 /* Huge page is mapped? Page fault is solved */
3616 if (pmd_trans_huge(*vmf->pmd)) {
3617 ret = VM_FAULT_NOPAGE;
3621 /* ->map_pages() haven't done anything useful. Cold page cache? */
3625 /* check if the page fault is solved */
3626 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3627 if (!pte_none(*vmf->pte))
3628 ret = VM_FAULT_NOPAGE;
3629 pte_unmap_unlock(vmf->pte, vmf->ptl);
3631 vmf->address = address;
3636 static int do_read_fault(struct vm_fault *vmf)
3638 struct vm_area_struct *vma = vmf->vma;
3642 * Let's call ->map_pages() first and use ->fault() as fallback
3643 * if page by the offset is not ready to be mapped (cold cache or
3646 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3647 ret = do_fault_around(vmf);
3652 ret = __do_fault(vmf);
3653 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3656 ret |= finish_fault(vmf);
3657 unlock_page(vmf->page);
3658 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3659 put_page(vmf->page);
3663 static int do_cow_fault(struct vm_fault *vmf)
3665 struct vm_area_struct *vma = vmf->vma;
3668 if (unlikely(anon_vma_prepare(vma)))
3669 return VM_FAULT_OOM;
3671 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3673 return VM_FAULT_OOM;
3675 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3676 &vmf->memcg, false)) {
3677 put_page(vmf->cow_page);
3678 return VM_FAULT_OOM;
3681 ret = __do_fault(vmf);
3682 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3684 if (ret & VM_FAULT_DONE_COW)
3687 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3688 __SetPageUptodate(vmf->cow_page);
3690 ret |= finish_fault(vmf);
3691 unlock_page(vmf->page);
3692 put_page(vmf->page);
3693 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3697 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3698 put_page(vmf->cow_page);
3702 static int do_shared_fault(struct vm_fault *vmf)
3704 struct vm_area_struct *vma = vmf->vma;
3707 ret = __do_fault(vmf);
3708 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3712 * Check if the backing address space wants to know that the page is
3713 * about to become writable
3715 if (vma->vm_ops->page_mkwrite) {
3716 unlock_page(vmf->page);
3717 tmp = do_page_mkwrite(vmf);
3718 if (unlikely(!tmp ||
3719 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3720 put_page(vmf->page);
3725 ret |= finish_fault(vmf);
3726 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3728 unlock_page(vmf->page);
3729 put_page(vmf->page);
3733 fault_dirty_shared_page(vma, vmf->page);
3738 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3739 * but allow concurrent faults).
3740 * The mmap_sem may have been released depending on flags and our
3741 * return value. See filemap_fault() and __lock_page_or_retry().
3743 static int do_fault(struct vm_fault *vmf)
3745 struct vm_area_struct *vma = vmf->vma;
3748 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3749 if (!vma->vm_ops->fault)
3750 ret = VM_FAULT_SIGBUS;
3751 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3752 ret = do_read_fault(vmf);
3753 else if (!(vma->vm_flags & VM_SHARED))
3754 ret = do_cow_fault(vmf);
3756 ret = do_shared_fault(vmf);
3758 /* preallocated pagetable is unused: free it */
3759 if (vmf->prealloc_pte) {
3760 pte_free(vma->vm_mm, vmf->prealloc_pte);
3761 vmf->prealloc_pte = NULL;
3766 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3767 unsigned long addr, int page_nid,
3772 count_vm_numa_event(NUMA_HINT_FAULTS);
3773 if (page_nid == numa_node_id()) {
3774 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3775 *flags |= TNF_FAULT_LOCAL;
3778 return mpol_misplaced(page, vma, addr);
3781 static int do_numa_page(struct vm_fault *vmf)
3783 struct vm_area_struct *vma = vmf->vma;
3784 struct page *page = NULL;
3788 bool migrated = false;
3790 bool was_writable = pte_savedwrite(vmf->orig_pte);
3794 * The "pte" at this point cannot be used safely without
3795 * validation through pte_unmap_same(). It's of NUMA type but
3796 * the pfn may be screwed if the read is non atomic.
3798 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3799 spin_lock(vmf->ptl);
3800 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3801 pte_unmap_unlock(vmf->pte, vmf->ptl);
3806 * Make it present again, Depending on how arch implementes non
3807 * accessible ptes, some can allow access by kernel mode.
3809 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3810 pte = pte_modify(pte, vma->vm_page_prot);
3811 pte = pte_mkyoung(pte);
3813 pte = pte_mkwrite(pte);
3814 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3815 update_mmu_cache(vma, vmf->address, vmf->pte);
3817 page = vm_normal_page(vma, vmf->address, pte);
3819 pte_unmap_unlock(vmf->pte, vmf->ptl);
3823 /* TODO: handle PTE-mapped THP */
3824 if (PageCompound(page)) {
3825 pte_unmap_unlock(vmf->pte, vmf->ptl);
3830 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3831 * much anyway since they can be in shared cache state. This misses
3832 * the case where a mapping is writable but the process never writes
3833 * to it but pte_write gets cleared during protection updates and
3834 * pte_dirty has unpredictable behaviour between PTE scan updates,
3835 * background writeback, dirty balancing and application behaviour.
3837 if (!pte_write(pte))
3838 flags |= TNF_NO_GROUP;
3841 * Flag if the page is shared between multiple address spaces. This
3842 * is later used when determining whether to group tasks together
3844 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3845 flags |= TNF_SHARED;
3847 last_cpupid = page_cpupid_last(page);
3848 page_nid = page_to_nid(page);
3849 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3851 pte_unmap_unlock(vmf->pte, vmf->ptl);
3852 if (target_nid == -1) {
3857 /* Migrate to the requested node */
3858 migrated = migrate_misplaced_page(page, vma, target_nid);
3860 page_nid = target_nid;
3861 flags |= TNF_MIGRATED;
3863 flags |= TNF_MIGRATE_FAIL;
3867 task_numa_fault(last_cpupid, page_nid, 1, flags);
3871 static inline int create_huge_pmd(struct vm_fault *vmf)
3873 if (vma_is_anonymous(vmf->vma))
3874 return do_huge_pmd_anonymous_page(vmf);
3875 if (vmf->vma->vm_ops->huge_fault)
3876 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3877 return VM_FAULT_FALLBACK;
3880 /* `inline' is required to avoid gcc 4.1.2 build error */
3881 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3883 if (vma_is_anonymous(vmf->vma))
3884 return do_huge_pmd_wp_page(vmf, orig_pmd);
3885 if (vmf->vma->vm_ops->huge_fault)
3886 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3888 /* COW handled on pte level: split pmd */
3889 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3890 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3892 return VM_FAULT_FALLBACK;
3895 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3897 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3900 static int create_huge_pud(struct vm_fault *vmf)
3902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3903 /* No support for anonymous transparent PUD pages yet */
3904 if (vma_is_anonymous(vmf->vma))
3905 return VM_FAULT_FALLBACK;
3906 if (vmf->vma->vm_ops->huge_fault)
3907 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3908 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3909 return VM_FAULT_FALLBACK;
3912 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3914 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3915 /* No support for anonymous transparent PUD pages yet */
3916 if (vma_is_anonymous(vmf->vma))
3917 return VM_FAULT_FALLBACK;
3918 if (vmf->vma->vm_ops->huge_fault)
3919 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3920 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3921 return VM_FAULT_FALLBACK;
3925 * These routines also need to handle stuff like marking pages dirty
3926 * and/or accessed for architectures that don't do it in hardware (most
3927 * RISC architectures). The early dirtying is also good on the i386.
3929 * There is also a hook called "update_mmu_cache()" that architectures
3930 * with external mmu caches can use to update those (ie the Sparc or
3931 * PowerPC hashed page tables that act as extended TLBs).
3933 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3934 * concurrent faults).
3936 * The mmap_sem may have been released depending on flags and our return value.
3937 * See filemap_fault() and __lock_page_or_retry().
3939 static int handle_pte_fault(struct vm_fault *vmf)
3943 if (unlikely(pmd_none(*vmf->pmd))) {
3945 * Leave __pte_alloc() until later: because vm_ops->fault may
3946 * want to allocate huge page, and if we expose page table
3947 * for an instant, it will be difficult to retract from
3948 * concurrent faults and from rmap lookups.
3952 /* See comment in pte_alloc_one_map() */
3953 if (pmd_devmap_trans_unstable(vmf->pmd))
3956 * A regular pmd is established and it can't morph into a huge
3957 * pmd from under us anymore at this point because we hold the
3958 * mmap_sem read mode and khugepaged takes it in write mode.
3959 * So now it's safe to run pte_offset_map().
3961 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3962 vmf->orig_pte = *vmf->pte;
3965 * some architectures can have larger ptes than wordsize,
3966 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3967 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3968 * accesses. The code below just needs a consistent view
3969 * for the ifs and we later double check anyway with the
3970 * ptl lock held. So here a barrier will do.
3973 if (pte_none(vmf->orig_pte)) {
3974 pte_unmap(vmf->pte);
3980 if (vma_is_anonymous(vmf->vma))
3981 return do_anonymous_page(vmf);
3983 return do_fault(vmf);
3986 if (!pte_present(vmf->orig_pte))
3987 return do_swap_page(vmf);
3989 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3990 return do_numa_page(vmf);
3992 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3993 spin_lock(vmf->ptl);
3994 entry = vmf->orig_pte;
3995 if (unlikely(!pte_same(*vmf->pte, entry)))
3997 if (vmf->flags & FAULT_FLAG_WRITE) {
3998 if (!pte_write(entry))
3999 return do_wp_page(vmf);
4000 entry = pte_mkdirty(entry);
4002 entry = pte_mkyoung(entry);
4003 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4004 vmf->flags & FAULT_FLAG_WRITE)) {
4005 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4008 * This is needed only for protection faults but the arch code
4009 * is not yet telling us if this is a protection fault or not.
4010 * This still avoids useless tlb flushes for .text page faults
4013 if (vmf->flags & FAULT_FLAG_WRITE)
4014 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4017 pte_unmap_unlock(vmf->pte, vmf->ptl);
4022 * By the time we get here, we already hold the mm semaphore
4024 * The mmap_sem may have been released depending on flags and our
4025 * return value. See filemap_fault() and __lock_page_or_retry().
4027 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4030 struct vm_fault vmf = {
4032 .address = address & PAGE_MASK,
4034 .pgoff = linear_page_index(vma, address),
4035 .gfp_mask = __get_fault_gfp_mask(vma),
4037 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4038 struct mm_struct *mm = vma->vm_mm;
4043 pgd = pgd_offset(mm, address);
4044 p4d = p4d_alloc(mm, pgd, address);
4046 return VM_FAULT_OOM;
4048 vmf.pud = pud_alloc(mm, p4d, address);
4050 return VM_FAULT_OOM;
4051 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4052 ret = create_huge_pud(&vmf);
4053 if (!(ret & VM_FAULT_FALLBACK))
4056 pud_t orig_pud = *vmf.pud;
4059 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4061 /* NUMA case for anonymous PUDs would go here */
4063 if (dirty && !pud_write(orig_pud)) {
4064 ret = wp_huge_pud(&vmf, orig_pud);
4065 if (!(ret & VM_FAULT_FALLBACK))
4068 huge_pud_set_accessed(&vmf, orig_pud);
4074 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4076 return VM_FAULT_OOM;
4077 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4078 ret = create_huge_pmd(&vmf);
4079 if (!(ret & VM_FAULT_FALLBACK))
4082 pmd_t orig_pmd = *vmf.pmd;
4085 if (unlikely(is_swap_pmd(orig_pmd))) {
4086 VM_BUG_ON(thp_migration_supported() &&
4087 !is_pmd_migration_entry(orig_pmd));
4088 if (is_pmd_migration_entry(orig_pmd))
4089 pmd_migration_entry_wait(mm, vmf.pmd);
4092 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4093 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4094 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4096 if (dirty && !pmd_write(orig_pmd)) {
4097 ret = wp_huge_pmd(&vmf, orig_pmd);
4098 if (!(ret & VM_FAULT_FALLBACK))
4101 huge_pmd_set_accessed(&vmf, orig_pmd);
4107 return handle_pte_fault(&vmf);
4111 * By the time we get here, we already hold the mm semaphore
4113 * The mmap_sem may have been released depending on flags and our
4114 * return value. See filemap_fault() and __lock_page_or_retry().
4116 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4121 __set_current_state(TASK_RUNNING);
4123 count_vm_event(PGFAULT);
4124 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4126 /* do counter updates before entering really critical section. */
4127 check_sync_rss_stat(current);
4129 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4130 flags & FAULT_FLAG_INSTRUCTION,
4131 flags & FAULT_FLAG_REMOTE))
4132 return VM_FAULT_SIGSEGV;
4135 * Enable the memcg OOM handling for faults triggered in user
4136 * space. Kernel faults are handled more gracefully.
4138 if (flags & FAULT_FLAG_USER)
4139 mem_cgroup_oom_enable();
4141 if (unlikely(is_vm_hugetlb_page(vma)))
4142 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4144 ret = __handle_mm_fault(vma, address, flags);
4146 if (flags & FAULT_FLAG_USER) {
4147 mem_cgroup_oom_disable();
4149 * The task may have entered a memcg OOM situation but
4150 * if the allocation error was handled gracefully (no
4151 * VM_FAULT_OOM), there is no need to kill anything.
4152 * Just clean up the OOM state peacefully.
4154 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4155 mem_cgroup_oom_synchronize(false);
4160 EXPORT_SYMBOL_GPL(handle_mm_fault);
4162 #ifndef __PAGETABLE_P4D_FOLDED
4164 * Allocate p4d page table.
4165 * We've already handled the fast-path in-line.
4167 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4169 p4d_t *new = p4d_alloc_one(mm, address);
4173 smp_wmb(); /* See comment in __pte_alloc */
4175 spin_lock(&mm->page_table_lock);
4176 if (pgd_present(*pgd)) /* Another has populated it */
4179 pgd_populate(mm, pgd, new);
4180 spin_unlock(&mm->page_table_lock);
4183 #endif /* __PAGETABLE_P4D_FOLDED */
4185 #ifndef __PAGETABLE_PUD_FOLDED
4187 * Allocate page upper directory.
4188 * We've already handled the fast-path in-line.
4190 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4192 pud_t *new = pud_alloc_one(mm, address);
4196 smp_wmb(); /* See comment in __pte_alloc */
4198 spin_lock(&mm->page_table_lock);
4199 #ifndef __ARCH_HAS_5LEVEL_HACK
4200 if (!p4d_present(*p4d)) {
4202 p4d_populate(mm, p4d, new);
4203 } else /* Another has populated it */
4206 if (!pgd_present(*p4d)) {
4208 pgd_populate(mm, p4d, new);
4209 } else /* Another has populated it */
4211 #endif /* __ARCH_HAS_5LEVEL_HACK */
4212 spin_unlock(&mm->page_table_lock);
4215 #endif /* __PAGETABLE_PUD_FOLDED */
4217 #ifndef __PAGETABLE_PMD_FOLDED
4219 * Allocate page middle directory.
4220 * We've already handled the fast-path in-line.
4222 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4225 pmd_t *new = pmd_alloc_one(mm, address);
4229 smp_wmb(); /* See comment in __pte_alloc */
4231 ptl = pud_lock(mm, pud);
4232 #ifndef __ARCH_HAS_4LEVEL_HACK
4233 if (!pud_present(*pud)) {
4235 pud_populate(mm, pud, new);
4236 } else /* Another has populated it */
4239 if (!pgd_present(*pud)) {
4241 pgd_populate(mm, pud, new);
4242 } else /* Another has populated it */
4244 #endif /* __ARCH_HAS_4LEVEL_HACK */
4248 #endif /* __PAGETABLE_PMD_FOLDED */
4250 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4251 unsigned long *start, unsigned long *end,
4252 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4260 pgd = pgd_offset(mm, address);
4261 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4264 p4d = p4d_offset(pgd, address);
4265 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4268 pud = pud_offset(p4d, address);
4269 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4272 pmd = pmd_offset(pud, address);
4273 VM_BUG_ON(pmd_trans_huge(*pmd));
4275 if (pmd_huge(*pmd)) {
4280 *start = address & PMD_MASK;
4281 *end = *start + PMD_SIZE;
4282 mmu_notifier_invalidate_range_start(mm, *start, *end);
4284 *ptlp = pmd_lock(mm, pmd);
4285 if (pmd_huge(*pmd)) {
4291 mmu_notifier_invalidate_range_end(mm, *start, *end);
4294 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4298 *start = address & PAGE_MASK;
4299 *end = *start + PAGE_SIZE;
4300 mmu_notifier_invalidate_range_start(mm, *start, *end);
4302 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4303 if (!pte_present(*ptep))
4308 pte_unmap_unlock(ptep, *ptlp);
4310 mmu_notifier_invalidate_range_end(mm, *start, *end);
4315 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4316 pte_t **ptepp, spinlock_t **ptlp)
4320 /* (void) is needed to make gcc happy */
4321 (void) __cond_lock(*ptlp,
4322 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4323 ptepp, NULL, ptlp)));
4327 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4328 unsigned long *start, unsigned long *end,
4329 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4333 /* (void) is needed to make gcc happy */
4334 (void) __cond_lock(*ptlp,
4335 !(res = __follow_pte_pmd(mm, address, start, end,
4336 ptepp, pmdpp, ptlp)));
4339 EXPORT_SYMBOL(follow_pte_pmd);
4342 * follow_pfn - look up PFN at a user virtual address
4343 * @vma: memory mapping
4344 * @address: user virtual address
4345 * @pfn: location to store found PFN
4347 * Only IO mappings and raw PFN mappings are allowed.
4349 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4351 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4358 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4361 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4364 *pfn = pte_pfn(*ptep);
4365 pte_unmap_unlock(ptep, ptl);
4368 EXPORT_SYMBOL(follow_pfn);
4370 #ifdef CONFIG_HAVE_IOREMAP_PROT
4371 int follow_phys(struct vm_area_struct *vma,
4372 unsigned long address, unsigned int flags,
4373 unsigned long *prot, resource_size_t *phys)
4379 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4382 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4386 if ((flags & FOLL_WRITE) && !pte_write(pte))
4389 *prot = pgprot_val(pte_pgprot(pte));
4390 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4394 pte_unmap_unlock(ptep, ptl);
4399 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4400 void *buf, int len, int write)
4402 resource_size_t phys_addr;
4403 unsigned long prot = 0;
4404 void __iomem *maddr;
4405 int offset = addr & (PAGE_SIZE-1);
4407 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4410 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4412 memcpy_toio(maddr + offset, buf, len);
4414 memcpy_fromio(buf, maddr + offset, len);
4419 EXPORT_SYMBOL_GPL(generic_access_phys);
4423 * Access another process' address space as given in mm. If non-NULL, use the
4424 * given task for page fault accounting.
4426 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4427 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4429 struct vm_area_struct *vma;
4430 void *old_buf = buf;
4431 int write = gup_flags & FOLL_WRITE;
4433 down_read(&mm->mmap_sem);
4434 /* ignore errors, just check how much was successfully transferred */
4436 int bytes, ret, offset;
4438 struct page *page = NULL;
4440 ret = get_user_pages_remote(tsk, mm, addr, 1,
4441 gup_flags, &page, &vma, NULL);
4443 #ifndef CONFIG_HAVE_IOREMAP_PROT
4447 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4448 * we can access using slightly different code.
4450 vma = find_vma(mm, addr);
4451 if (!vma || vma->vm_start > addr)
4453 if (vma->vm_ops && vma->vm_ops->access)
4454 ret = vma->vm_ops->access(vma, addr, buf,
4462 offset = addr & (PAGE_SIZE-1);
4463 if (bytes > PAGE_SIZE-offset)
4464 bytes = PAGE_SIZE-offset;
4468 copy_to_user_page(vma, page, addr,
4469 maddr + offset, buf, bytes);
4470 set_page_dirty_lock(page);
4472 copy_from_user_page(vma, page, addr,
4473 buf, maddr + offset, bytes);
4482 up_read(&mm->mmap_sem);
4484 return buf - old_buf;
4488 * access_remote_vm - access another process' address space
4489 * @mm: the mm_struct of the target address space
4490 * @addr: start address to access
4491 * @buf: source or destination buffer
4492 * @len: number of bytes to transfer
4493 * @gup_flags: flags modifying lookup behaviour
4495 * The caller must hold a reference on @mm.
4497 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4498 void *buf, int len, unsigned int gup_flags)
4500 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4504 * Access another process' address space.
4505 * Source/target buffer must be kernel space,
4506 * Do not walk the page table directly, use get_user_pages
4508 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4509 void *buf, int len, unsigned int gup_flags)
4511 struct mm_struct *mm;
4514 mm = get_task_mm(tsk);
4518 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4524 EXPORT_SYMBOL_GPL(access_process_vm);
4527 * Print the name of a VMA.
4529 void print_vma_addr(char *prefix, unsigned long ip)
4531 struct mm_struct *mm = current->mm;
4532 struct vm_area_struct *vma;
4535 * we might be running from an atomic context so we cannot sleep
4537 if (!down_read_trylock(&mm->mmap_sem))
4540 vma = find_vma(mm, ip);
4541 if (vma && vma->vm_file) {
4542 struct file *f = vma->vm_file;
4543 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4547 p = file_path(f, buf, PAGE_SIZE);
4550 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4552 vma->vm_end - vma->vm_start);
4553 free_page((unsigned long)buf);
4556 up_read(&mm->mmap_sem);
4559 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4560 void __might_fault(const char *file, int line)
4563 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4564 * holding the mmap_sem, this is safe because kernel memory doesn't
4565 * get paged out, therefore we'll never actually fault, and the
4566 * below annotations will generate false positives.
4568 if (uaccess_kernel())
4570 if (pagefault_disabled())
4572 __might_sleep(file, line, 0);
4573 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4575 might_lock_read(¤t->mm->mmap_sem);
4578 EXPORT_SYMBOL(__might_fault);
4581 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4582 static void clear_gigantic_page(struct page *page,
4584 unsigned int pages_per_huge_page)
4587 struct page *p = page;
4590 for (i = 0; i < pages_per_huge_page;
4591 i++, p = mem_map_next(p, page, i)) {
4593 clear_user_highpage(p, addr + i * PAGE_SIZE);
4596 void clear_huge_page(struct page *page,
4597 unsigned long addr_hint, unsigned int pages_per_huge_page)
4600 unsigned long addr = addr_hint &
4601 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4603 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4604 clear_gigantic_page(page, addr, pages_per_huge_page);
4608 /* Clear sub-page to access last to keep its cache lines hot */
4610 n = (addr_hint - addr) / PAGE_SIZE;
4611 if (2 * n <= pages_per_huge_page) {
4612 /* If sub-page to access in first half of huge page */
4615 /* Clear sub-pages at the end of huge page */
4616 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4618 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4621 /* If sub-page to access in second half of huge page */
4622 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4623 l = pages_per_huge_page - n;
4624 /* Clear sub-pages at the begin of huge page */
4625 for (i = 0; i < base; i++) {
4627 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4631 * Clear remaining sub-pages in left-right-left-right pattern
4632 * towards the sub-page to access
4634 for (i = 0; i < l; i++) {
4635 int left_idx = base + i;
4636 int right_idx = base + 2 * l - 1 - i;
4639 clear_user_highpage(page + left_idx,
4640 addr + left_idx * PAGE_SIZE);
4642 clear_user_highpage(page + right_idx,
4643 addr + right_idx * PAGE_SIZE);
4647 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4649 struct vm_area_struct *vma,
4650 unsigned int pages_per_huge_page)
4653 struct page *dst_base = dst;
4654 struct page *src_base = src;
4656 for (i = 0; i < pages_per_huge_page; ) {
4658 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4661 dst = mem_map_next(dst, dst_base, i);
4662 src = mem_map_next(src, src_base, i);
4666 void copy_user_huge_page(struct page *dst, struct page *src,
4667 unsigned long addr, struct vm_area_struct *vma,
4668 unsigned int pages_per_huge_page)
4672 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4673 copy_user_gigantic_page(dst, src, addr, vma,
4674 pages_per_huge_page);
4679 for (i = 0; i < pages_per_huge_page; i++) {
4681 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4685 long copy_huge_page_from_user(struct page *dst_page,
4686 const void __user *usr_src,
4687 unsigned int pages_per_huge_page,
4688 bool allow_pagefault)
4690 void *src = (void *)usr_src;
4692 unsigned long i, rc = 0;
4693 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4695 for (i = 0; i < pages_per_huge_page; i++) {
4696 if (allow_pagefault)
4697 page_kaddr = kmap(dst_page + i);
4699 page_kaddr = kmap_atomic(dst_page + i);
4700 rc = copy_from_user(page_kaddr,
4701 (const void __user *)(src + i * PAGE_SIZE),
4703 if (allow_pagefault)
4704 kunmap(dst_page + i);
4706 kunmap_atomic(page_kaddr);
4708 ret_val -= (PAGE_SIZE - rc);
4716 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4718 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4720 static struct kmem_cache *page_ptl_cachep;
4722 void __init ptlock_cache_init(void)
4724 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4728 bool ptlock_alloc(struct page *page)
4732 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4739 void ptlock_free(struct page *page)
4741 kmem_cache_free(page_ptl_cachep, page->ptl);