4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
93 EXPORT_SYMBOL(mem_map);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
119 #ifndef arch_faults_on_old_pte
120 static inline bool arch_faults_on_old_pte(void)
123 * Those arches which don't have hw access flag feature need to
124 * implement their own helper. By default, "true" means pagefault
125 * will be hit on old pte.
131 static int __init disable_randmaps(char *s)
133 randomize_va_space = 0;
136 __setup("norandmaps", disable_randmaps);
138 unsigned long zero_pfn __read_mostly;
139 EXPORT_SYMBOL(zero_pfn);
141 unsigned long highest_memmap_pfn __read_mostly;
144 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
146 static int __init init_zero_pfn(void)
148 zero_pfn = page_to_pfn(ZERO_PAGE(0));
151 early_initcall(init_zero_pfn);
154 #if defined(SPLIT_RSS_COUNTING)
156 void sync_mm_rss(struct mm_struct *mm)
160 for (i = 0; i < NR_MM_COUNTERS; i++) {
161 if (current->rss_stat.count[i]) {
162 add_mm_counter(mm, i, current->rss_stat.count[i]);
163 current->rss_stat.count[i] = 0;
166 current->rss_stat.events = 0;
169 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
171 struct task_struct *task = current;
173 if (likely(task->mm == mm))
174 task->rss_stat.count[member] += val;
176 add_mm_counter(mm, member, val);
178 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
179 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
181 /* sync counter once per 64 page faults */
182 #define TASK_RSS_EVENTS_THRESH (64)
183 static void check_sync_rss_stat(struct task_struct *task)
185 if (unlikely(task != current))
187 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
188 sync_mm_rss(task->mm);
190 #else /* SPLIT_RSS_COUNTING */
192 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
193 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
195 static void check_sync_rss_stat(struct task_struct *task)
199 #endif /* SPLIT_RSS_COUNTING */
201 #ifdef HAVE_GENERIC_MMU_GATHER
203 static bool tlb_next_batch(struct mmu_gather *tlb)
205 struct mmu_gather_batch *batch;
209 tlb->active = batch->next;
213 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
216 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
223 batch->max = MAX_GATHER_BATCH;
225 tlb->active->next = batch;
231 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
232 unsigned long start, unsigned long end)
236 /* Is it from 0 to ~0? */
237 tlb->fullmm = !(start | (end+1));
238 tlb->need_flush_all = 0;
239 tlb->local.next = NULL;
241 tlb->local.max = ARRAY_SIZE(tlb->__pages);
242 tlb->active = &tlb->local;
243 tlb->batch_count = 0;
245 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
250 __tlb_reset_range(tlb);
253 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255 struct mmu_gather_batch *batch;
257 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
258 tlb_table_flush(tlb);
260 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
261 free_pages_and_swap_cache(batch->pages, batch->nr);
264 tlb->active = &tlb->local;
267 void tlb_flush_mmu(struct mmu_gather *tlb)
269 tlb_flush_mmu_tlbonly(tlb);
270 tlb_flush_mmu_free(tlb);
274 * Called at the end of the shootdown operation to free up any resources
275 * that were required.
277 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
278 unsigned long start, unsigned long end, bool force)
280 struct mmu_gather_batch *batch, *next;
283 __tlb_adjust_range(tlb, start, end - start);
287 /* keep the page table cache within bounds */
290 for (batch = tlb->local.next; batch; batch = next) {
292 free_pages((unsigned long)batch, 0);
294 tlb->local.next = NULL;
298 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
299 * handling the additional races in SMP caused by other CPUs caching valid
300 * mappings in their TLBs. Returns the number of free page slots left.
301 * When out of page slots we must call tlb_flush_mmu().
302 *returns true if the caller should flush.
304 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
306 struct mmu_gather_batch *batch;
308 VM_BUG_ON(!tlb->end);
309 VM_WARN_ON(tlb->page_size != page_size);
313 * Add the page and check if we are full. If so
316 batch->pages[batch->nr++] = page;
317 if (batch->nr == batch->max) {
318 if (!tlb_next_batch(tlb))
322 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
327 #endif /* HAVE_GENERIC_MMU_GATHER */
329 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
332 * See the comment near struct mmu_table_batch.
336 * If we want tlb_remove_table() to imply TLB invalidates.
338 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
340 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
342 * Invalidate page-table caches used by hardware walkers. Then we still
343 * need to RCU-sched wait while freeing the pages because software
344 * walkers can still be in-flight.
346 tlb_flush_mmu_tlbonly(tlb);
350 static void tlb_remove_table_smp_sync(void *arg)
352 /* Simply deliver the interrupt */
355 static void tlb_remove_table_one(void *table)
358 * This isn't an RCU grace period and hence the page-tables cannot be
359 * assumed to be actually RCU-freed.
361 * It is however sufficient for software page-table walkers that rely on
362 * IRQ disabling. See the comment near struct mmu_table_batch.
364 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
365 __tlb_remove_table(table);
368 static void tlb_remove_table_rcu(struct rcu_head *head)
370 struct mmu_table_batch *batch;
373 batch = container_of(head, struct mmu_table_batch, rcu);
375 for (i = 0; i < batch->nr; i++)
376 __tlb_remove_table(batch->tables[i]);
378 free_page((unsigned long)batch);
381 void tlb_table_flush(struct mmu_gather *tlb)
383 struct mmu_table_batch **batch = &tlb->batch;
386 tlb_table_invalidate(tlb);
387 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
392 void tlb_remove_table(struct mmu_gather *tlb, void *table)
394 struct mmu_table_batch **batch = &tlb->batch;
396 if (*batch == NULL) {
397 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
398 if (*batch == NULL) {
399 tlb_table_invalidate(tlb);
400 tlb_remove_table_one(table);
406 (*batch)->tables[(*batch)->nr++] = table;
407 if ((*batch)->nr == MAX_TABLE_BATCH)
408 tlb_table_flush(tlb);
411 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
414 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
415 * @tlb: the mmu_gather structure to initialize
416 * @mm: the mm_struct of the target address space
417 * @start: start of the region that will be removed from the page-table
418 * @end: end of the region that will be removed from the page-table
420 * Called to initialize an (on-stack) mmu_gather structure for page-table
421 * tear-down from @mm. The @start and @end are set to 0 and -1
422 * respectively when @mm is without users and we're going to destroy
423 * the full address space (exit/execve).
425 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
426 unsigned long start, unsigned long end)
428 arch_tlb_gather_mmu(tlb, mm, start, end);
429 inc_tlb_flush_pending(tlb->mm);
432 void tlb_finish_mmu(struct mmu_gather *tlb,
433 unsigned long start, unsigned long end)
436 * If there are parallel threads are doing PTE changes on same range
437 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
438 * flush by batching, a thread has stable TLB entry can fail to flush
439 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
440 * forcefully if we detect parallel PTE batching threads.
442 bool force = mm_tlb_flush_nested(tlb->mm);
444 arch_tlb_finish_mmu(tlb, start, end, force);
445 dec_tlb_flush_pending(tlb->mm);
449 * Note: this doesn't free the actual pages themselves. That
450 * has been handled earlier when unmapping all the memory regions.
452 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
455 pgtable_t token = pmd_pgtable(*pmd);
457 pte_free_tlb(tlb, token, addr);
458 mm_dec_nr_ptes(tlb->mm);
461 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
462 unsigned long addr, unsigned long end,
463 unsigned long floor, unsigned long ceiling)
470 pmd = pmd_offset(pud, addr);
472 next = pmd_addr_end(addr, end);
473 if (pmd_none_or_clear_bad(pmd))
475 free_pte_range(tlb, pmd, addr);
476 } while (pmd++, addr = next, addr != end);
486 if (end - 1 > ceiling - 1)
489 pmd = pmd_offset(pud, start);
491 pmd_free_tlb(tlb, pmd, start);
492 mm_dec_nr_pmds(tlb->mm);
495 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
496 unsigned long addr, unsigned long end,
497 unsigned long floor, unsigned long ceiling)
504 pud = pud_offset(p4d, addr);
506 next = pud_addr_end(addr, end);
507 if (pud_none_or_clear_bad(pud))
509 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
510 } while (pud++, addr = next, addr != end);
520 if (end - 1 > ceiling - 1)
523 pud = pud_offset(p4d, start);
525 pud_free_tlb(tlb, pud, start);
526 mm_dec_nr_puds(tlb->mm);
529 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
530 unsigned long addr, unsigned long end,
531 unsigned long floor, unsigned long ceiling)
538 p4d = p4d_offset(pgd, addr);
540 next = p4d_addr_end(addr, end);
541 if (p4d_none_or_clear_bad(p4d))
543 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
544 } while (p4d++, addr = next, addr != end);
550 ceiling &= PGDIR_MASK;
554 if (end - 1 > ceiling - 1)
557 p4d = p4d_offset(pgd, start);
559 p4d_free_tlb(tlb, p4d, start);
563 * This function frees user-level page tables of a process.
565 void free_pgd_range(struct mmu_gather *tlb,
566 unsigned long addr, unsigned long end,
567 unsigned long floor, unsigned long ceiling)
573 * The next few lines have given us lots of grief...
575 * Why are we testing PMD* at this top level? Because often
576 * there will be no work to do at all, and we'd prefer not to
577 * go all the way down to the bottom just to discover that.
579 * Why all these "- 1"s? Because 0 represents both the bottom
580 * of the address space and the top of it (using -1 for the
581 * top wouldn't help much: the masks would do the wrong thing).
582 * The rule is that addr 0 and floor 0 refer to the bottom of
583 * the address space, but end 0 and ceiling 0 refer to the top
584 * Comparisons need to use "end - 1" and "ceiling - 1" (though
585 * that end 0 case should be mythical).
587 * Wherever addr is brought up or ceiling brought down, we must
588 * be careful to reject "the opposite 0" before it confuses the
589 * subsequent tests. But what about where end is brought down
590 * by PMD_SIZE below? no, end can't go down to 0 there.
592 * Whereas we round start (addr) and ceiling down, by different
593 * masks at different levels, in order to test whether a table
594 * now has no other vmas using it, so can be freed, we don't
595 * bother to round floor or end up - the tests don't need that.
609 if (end - 1 > ceiling - 1)
614 * We add page table cache pages with PAGE_SIZE,
615 * (see pte_free_tlb()), flush the tlb if we need
617 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
618 pgd = pgd_offset(tlb->mm, addr);
620 next = pgd_addr_end(addr, end);
621 if (pgd_none_or_clear_bad(pgd))
623 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
624 } while (pgd++, addr = next, addr != end);
627 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
628 unsigned long floor, unsigned long ceiling)
631 struct vm_area_struct *next = vma->vm_next;
632 unsigned long addr = vma->vm_start;
635 * Hide vma from rmap and truncate_pagecache before freeing
638 unlink_anon_vmas(vma);
639 unlink_file_vma(vma);
641 if (is_vm_hugetlb_page(vma)) {
642 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
643 floor, next ? next->vm_start : ceiling);
646 * Optimization: gather nearby vmas into one call down
648 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
649 && !is_vm_hugetlb_page(next)) {
652 unlink_anon_vmas(vma);
653 unlink_file_vma(vma);
655 free_pgd_range(tlb, addr, vma->vm_end,
656 floor, next ? next->vm_start : ceiling);
662 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
665 pgtable_t new = pte_alloc_one(mm, address);
670 * Ensure all pte setup (eg. pte page lock and page clearing) are
671 * visible before the pte is made visible to other CPUs by being
672 * put into page tables.
674 * The other side of the story is the pointer chasing in the page
675 * table walking code (when walking the page table without locking;
676 * ie. most of the time). Fortunately, these data accesses consist
677 * of a chain of data-dependent loads, meaning most CPUs (alpha
678 * being the notable exception) will already guarantee loads are
679 * seen in-order. See the alpha page table accessors for the
680 * smp_read_barrier_depends() barriers in page table walking code.
682 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
684 ptl = pmd_lock(mm, pmd);
685 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
687 pmd_populate(mm, pmd, new);
696 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
698 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
702 smp_wmb(); /* See comment in __pte_alloc */
704 spin_lock(&init_mm.page_table_lock);
705 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
706 pmd_populate_kernel(&init_mm, pmd, new);
709 spin_unlock(&init_mm.page_table_lock);
711 pte_free_kernel(&init_mm, new);
715 static inline void init_rss_vec(int *rss)
717 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
720 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
724 if (current->mm == mm)
726 for (i = 0; i < NR_MM_COUNTERS; i++)
728 add_mm_counter(mm, i, rss[i]);
732 * This function is called to print an error when a bad pte
733 * is found. For example, we might have a PFN-mapped pte in
734 * a region that doesn't allow it.
736 * The calling function must still handle the error.
738 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
739 pte_t pte, struct page *page)
741 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
742 p4d_t *p4d = p4d_offset(pgd, addr);
743 pud_t *pud = pud_offset(p4d, addr);
744 pmd_t *pmd = pmd_offset(pud, addr);
745 struct address_space *mapping;
747 static unsigned long resume;
748 static unsigned long nr_shown;
749 static unsigned long nr_unshown;
752 * Allow a burst of 60 reports, then keep quiet for that minute;
753 * or allow a steady drip of one report per second.
755 if (nr_shown == 60) {
756 if (time_before(jiffies, resume)) {
761 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
768 resume = jiffies + 60 * HZ;
770 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
771 index = linear_page_index(vma, addr);
773 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
775 (long long)pte_val(pte), (long long)pmd_val(*pmd));
777 dump_page(page, "bad pte");
778 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
779 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
780 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
782 vma->vm_ops ? vma->vm_ops->fault : NULL,
783 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
784 mapping ? mapping->a_ops->readpage : NULL);
786 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
790 * vm_normal_page -- This function gets the "struct page" associated with a pte.
792 * "Special" mappings do not wish to be associated with a "struct page" (either
793 * it doesn't exist, or it exists but they don't want to touch it). In this
794 * case, NULL is returned here. "Normal" mappings do have a struct page.
796 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
797 * pte bit, in which case this function is trivial. Secondly, an architecture
798 * may not have a spare pte bit, which requires a more complicated scheme,
801 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
802 * special mapping (even if there are underlying and valid "struct pages").
803 * COWed pages of a VM_PFNMAP are always normal.
805 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
806 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
807 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
808 * mapping will always honor the rule
810 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
812 * And for normal mappings this is false.
814 * This restricts such mappings to be a linear translation from virtual address
815 * to pfn. To get around this restriction, we allow arbitrary mappings so long
816 * as the vma is not a COW mapping; in that case, we know that all ptes are
817 * special (because none can have been COWed).
820 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
822 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
823 * page" backing, however the difference is that _all_ pages with a struct
824 * page (that is, those where pfn_valid is true) are refcounted and considered
825 * normal pages by the VM. The disadvantage is that pages are refcounted
826 * (which can be slower and simply not an option for some PFNMAP users). The
827 * advantage is that we don't have to follow the strict linearity rule of
828 * PFNMAP mappings in order to support COWable mappings.
831 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
832 pte_t pte, bool with_public_device)
834 unsigned long pfn = pte_pfn(pte);
836 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
837 if (likely(!pte_special(pte)))
839 if (vma->vm_ops && vma->vm_ops->find_special_page)
840 return vma->vm_ops->find_special_page(vma, addr);
841 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
843 if (is_zero_pfn(pfn))
847 * Device public pages are special pages (they are ZONE_DEVICE
848 * pages but different from persistent memory). They behave
849 * allmost like normal pages. The difference is that they are
850 * not on the lru and thus should never be involve with any-
851 * thing that involve lru manipulation (mlock, numa balancing,
854 * This is why we still want to return NULL for such page from
855 * vm_normal_page() so that we do not have to special case all
856 * call site of vm_normal_page().
858 if (likely(pfn <= highest_memmap_pfn)) {
859 struct page *page = pfn_to_page(pfn);
861 if (is_device_public_page(page)) {
862 if (with_public_device)
871 print_bad_pte(vma, addr, pte, NULL);
875 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
877 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
878 if (vma->vm_flags & VM_MIXEDMAP) {
884 off = (addr - vma->vm_start) >> PAGE_SHIFT;
885 if (pfn == vma->vm_pgoff + off)
887 if (!is_cow_mapping(vma->vm_flags))
892 if (is_zero_pfn(pfn))
896 if (unlikely(pfn > highest_memmap_pfn)) {
897 print_bad_pte(vma, addr, pte, NULL);
902 * NOTE! We still have PageReserved() pages in the page tables.
903 * eg. VDSO mappings can cause them to exist.
906 return pfn_to_page(pfn);
909 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
910 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
913 unsigned long pfn = pmd_pfn(pmd);
916 * There is no pmd_special() but there may be special pmds, e.g.
917 * in a direct-access (dax) mapping, so let's just replicate the
918 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
920 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
921 if (vma->vm_flags & VM_MIXEDMAP) {
927 off = (addr - vma->vm_start) >> PAGE_SHIFT;
928 if (pfn == vma->vm_pgoff + off)
930 if (!is_cow_mapping(vma->vm_flags))
937 if (is_zero_pfn(pfn))
939 if (unlikely(pfn > highest_memmap_pfn))
943 * NOTE! We still have PageReserved() pages in the page tables.
944 * eg. VDSO mappings can cause them to exist.
947 return pfn_to_page(pfn);
952 * copy one vm_area from one task to the other. Assumes the page tables
953 * already present in the new task to be cleared in the whole range
954 * covered by this vma.
957 static inline unsigned long
958 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
959 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
960 unsigned long addr, int *rss)
962 unsigned long vm_flags = vma->vm_flags;
963 pte_t pte = *src_pte;
966 /* pte contains position in swap or file, so copy. */
967 if (unlikely(!pte_present(pte))) {
968 swp_entry_t entry = pte_to_swp_entry(pte);
970 if (likely(!non_swap_entry(entry))) {
971 if (swap_duplicate(entry) < 0)
974 /* make sure dst_mm is on swapoff's mmlist. */
975 if (unlikely(list_empty(&dst_mm->mmlist))) {
976 spin_lock(&mmlist_lock);
977 if (list_empty(&dst_mm->mmlist))
978 list_add(&dst_mm->mmlist,
980 spin_unlock(&mmlist_lock);
983 } else if (is_migration_entry(entry)) {
984 page = migration_entry_to_page(entry);
986 rss[mm_counter(page)]++;
988 if (is_write_migration_entry(entry) &&
989 is_cow_mapping(vm_flags)) {
991 * COW mappings require pages in both
992 * parent and child to be set to read.
994 make_migration_entry_read(&entry);
995 pte = swp_entry_to_pte(entry);
996 if (pte_swp_soft_dirty(*src_pte))
997 pte = pte_swp_mksoft_dirty(pte);
998 set_pte_at(src_mm, addr, src_pte, pte);
1000 } else if (is_device_private_entry(entry)) {
1001 page = device_private_entry_to_page(entry);
1004 * Update rss count even for unaddressable pages, as
1005 * they should treated just like normal pages in this
1008 * We will likely want to have some new rss counters
1009 * for unaddressable pages, at some point. But for now
1010 * keep things as they are.
1013 rss[mm_counter(page)]++;
1014 page_dup_rmap(page, false);
1017 * We do not preserve soft-dirty information, because so
1018 * far, checkpoint/restore is the only feature that
1019 * requires that. And checkpoint/restore does not work
1020 * when a device driver is involved (you cannot easily
1021 * save and restore device driver state).
1023 if (is_write_device_private_entry(entry) &&
1024 is_cow_mapping(vm_flags)) {
1025 make_device_private_entry_read(&entry);
1026 pte = swp_entry_to_pte(entry);
1027 set_pte_at(src_mm, addr, src_pte, pte);
1034 * If it's a COW mapping, write protect it both
1035 * in the parent and the child
1037 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
1038 ptep_set_wrprotect(src_mm, addr, src_pte);
1039 pte = pte_wrprotect(pte);
1043 * If it's a shared mapping, mark it clean in
1046 if (vm_flags & VM_SHARED)
1047 pte = pte_mkclean(pte);
1048 pte = pte_mkold(pte);
1050 page = vm_normal_page(vma, addr, pte);
1053 page_dup_rmap(page, false);
1054 rss[mm_counter(page)]++;
1055 } else if (pte_devmap(pte)) {
1056 page = pte_page(pte);
1059 * Cache coherent device memory behave like regular page and
1060 * not like persistent memory page. For more informations see
1061 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1063 if (is_device_public_page(page)) {
1065 page_dup_rmap(page, false);
1066 rss[mm_counter(page)]++;
1071 set_pte_at(dst_mm, addr, dst_pte, pte);
1075 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1076 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1077 unsigned long addr, unsigned long end)
1079 pte_t *orig_src_pte, *orig_dst_pte;
1080 pte_t *src_pte, *dst_pte;
1081 spinlock_t *src_ptl, *dst_ptl;
1083 int rss[NR_MM_COUNTERS];
1084 swp_entry_t entry = (swp_entry_t){0};
1089 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1092 src_pte = pte_offset_map(src_pmd, addr);
1093 src_ptl = pte_lockptr(src_mm, src_pmd);
1094 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1095 orig_src_pte = src_pte;
1096 orig_dst_pte = dst_pte;
1097 arch_enter_lazy_mmu_mode();
1101 * We are holding two locks at this point - either of them
1102 * could generate latencies in another task on another CPU.
1104 if (progress >= 32) {
1106 if (need_resched() ||
1107 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1110 if (pte_none(*src_pte)) {
1114 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1119 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1121 arch_leave_lazy_mmu_mode();
1122 spin_unlock(src_ptl);
1123 pte_unmap(orig_src_pte);
1124 add_mm_rss_vec(dst_mm, rss);
1125 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1129 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1138 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1139 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1140 unsigned long addr, unsigned long end)
1142 pmd_t *src_pmd, *dst_pmd;
1145 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1148 src_pmd = pmd_offset(src_pud, addr);
1150 next = pmd_addr_end(addr, end);
1151 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1152 || pmd_devmap(*src_pmd)) {
1154 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1155 err = copy_huge_pmd(dst_mm, src_mm,
1156 dst_pmd, src_pmd, addr, vma);
1163 if (pmd_none_or_clear_bad(src_pmd))
1165 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1168 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1172 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1173 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1174 unsigned long addr, unsigned long end)
1176 pud_t *src_pud, *dst_pud;
1179 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1182 src_pud = pud_offset(src_p4d, addr);
1184 next = pud_addr_end(addr, end);
1185 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1188 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1189 err = copy_huge_pud(dst_mm, src_mm,
1190 dst_pud, src_pud, addr, vma);
1197 if (pud_none_or_clear_bad(src_pud))
1199 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1202 } while (dst_pud++, src_pud++, addr = next, addr != end);
1206 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1207 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1208 unsigned long addr, unsigned long end)
1210 p4d_t *src_p4d, *dst_p4d;
1213 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1216 src_p4d = p4d_offset(src_pgd, addr);
1218 next = p4d_addr_end(addr, end);
1219 if (p4d_none_or_clear_bad(src_p4d))
1221 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1224 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1228 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1229 struct vm_area_struct *vma)
1231 pgd_t *src_pgd, *dst_pgd;
1233 unsigned long addr = vma->vm_start;
1234 unsigned long end = vma->vm_end;
1235 unsigned long mmun_start; /* For mmu_notifiers */
1236 unsigned long mmun_end; /* For mmu_notifiers */
1241 * Don't copy ptes where a page fault will fill them correctly.
1242 * Fork becomes much lighter when there are big shared or private
1243 * readonly mappings. The tradeoff is that copy_page_range is more
1244 * efficient than faulting.
1246 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1250 if (is_vm_hugetlb_page(vma))
1251 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1253 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1255 * We do not free on error cases below as remove_vma
1256 * gets called on error from higher level routine
1258 ret = track_pfn_copy(vma);
1264 * We need to invalidate the secondary MMU mappings only when
1265 * there could be a permission downgrade on the ptes of the
1266 * parent mm. And a permission downgrade will only happen if
1267 * is_cow_mapping() returns true.
1269 is_cow = is_cow_mapping(vma->vm_flags);
1273 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1277 dst_pgd = pgd_offset(dst_mm, addr);
1278 src_pgd = pgd_offset(src_mm, addr);
1280 next = pgd_addr_end(addr, end);
1281 if (pgd_none_or_clear_bad(src_pgd))
1283 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1284 vma, addr, next))) {
1288 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1291 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1295 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1296 struct vm_area_struct *vma, pmd_t *pmd,
1297 unsigned long addr, unsigned long end,
1298 struct zap_details *details)
1300 struct mm_struct *mm = tlb->mm;
1301 int force_flush = 0;
1302 int rss[NR_MM_COUNTERS];
1308 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1311 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1313 flush_tlb_batched_pending(mm);
1314 arch_enter_lazy_mmu_mode();
1317 if (pte_none(ptent))
1320 if (pte_present(ptent)) {
1323 page = _vm_normal_page(vma, addr, ptent, true);
1324 if (unlikely(details) && page) {
1326 * unmap_shared_mapping_pages() wants to
1327 * invalidate cache without truncating:
1328 * unmap shared but keep private pages.
1330 if (details->check_mapping &&
1331 details->check_mapping != page_rmapping(page))
1334 ptent = ptep_get_and_clear_full(mm, addr, pte,
1336 tlb_remove_tlb_entry(tlb, pte, addr);
1337 if (unlikely(!page))
1340 if (!PageAnon(page)) {
1341 if (pte_dirty(ptent)) {
1343 set_page_dirty(page);
1345 if (pte_young(ptent) &&
1346 likely(!(vma->vm_flags & VM_SEQ_READ)))
1347 mark_page_accessed(page);
1349 rss[mm_counter(page)]--;
1350 page_remove_rmap(page, false);
1351 if (unlikely(page_mapcount(page) < 0))
1352 print_bad_pte(vma, addr, ptent, page);
1353 if (unlikely(__tlb_remove_page(tlb, page))) {
1361 entry = pte_to_swp_entry(ptent);
1362 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1363 struct page *page = device_private_entry_to_page(entry);
1365 if (unlikely(details && details->check_mapping)) {
1367 * unmap_shared_mapping_pages() wants to
1368 * invalidate cache without truncating:
1369 * unmap shared but keep private pages.
1371 if (details->check_mapping !=
1372 page_rmapping(page))
1376 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1377 rss[mm_counter(page)]--;
1378 page_remove_rmap(page, false);
1383 /* If details->check_mapping, we leave swap entries. */
1384 if (unlikely(details))
1387 entry = pte_to_swp_entry(ptent);
1388 if (!non_swap_entry(entry))
1390 else if (is_migration_entry(entry)) {
1393 page = migration_entry_to_page(entry);
1394 rss[mm_counter(page)]--;
1396 if (unlikely(!free_swap_and_cache(entry)))
1397 print_bad_pte(vma, addr, ptent, NULL);
1398 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1399 } while (pte++, addr += PAGE_SIZE, addr != end);
1401 add_mm_rss_vec(mm, rss);
1402 arch_leave_lazy_mmu_mode();
1404 /* Do the actual TLB flush before dropping ptl */
1406 tlb_flush_mmu_tlbonly(tlb);
1407 pte_unmap_unlock(start_pte, ptl);
1410 * If we forced a TLB flush (either due to running out of
1411 * batch buffers or because we needed to flush dirty TLB
1412 * entries before releasing the ptl), free the batched
1413 * memory too. Restart if we didn't do everything.
1417 tlb_flush_mmu_free(tlb);
1425 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1426 struct vm_area_struct *vma, pud_t *pud,
1427 unsigned long addr, unsigned long end,
1428 struct zap_details *details)
1433 pmd = pmd_offset(pud, addr);
1435 next = pmd_addr_end(addr, end);
1436 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1437 if (next - addr != HPAGE_PMD_SIZE)
1438 __split_huge_pmd(vma, pmd, addr, false, NULL);
1439 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1442 } else if (details && details->single_page &&
1443 PageTransCompound(details->single_page) &&
1444 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1445 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1447 * Take and drop THP pmd lock so that we cannot return
1448 * prematurely, while zap_huge_pmd() has cleared *pmd,
1449 * but not yet decremented compound_mapcount().
1455 * Here there can be other concurrent MADV_DONTNEED or
1456 * trans huge page faults running, and if the pmd is
1457 * none or trans huge it can change under us. This is
1458 * because MADV_DONTNEED holds the mmap_sem in read
1461 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1463 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1466 } while (pmd++, addr = next, addr != end);
1471 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1472 struct vm_area_struct *vma, p4d_t *p4d,
1473 unsigned long addr, unsigned long end,
1474 struct zap_details *details)
1479 pud = pud_offset(p4d, addr);
1481 next = pud_addr_end(addr, end);
1482 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1483 if (next - addr != HPAGE_PUD_SIZE) {
1484 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1485 split_huge_pud(vma, pud, addr);
1486 } else if (zap_huge_pud(tlb, vma, pud, addr))
1490 if (pud_none_or_clear_bad(pud))
1492 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1495 } while (pud++, addr = next, addr != end);
1500 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1501 struct vm_area_struct *vma, pgd_t *pgd,
1502 unsigned long addr, unsigned long end,
1503 struct zap_details *details)
1508 p4d = p4d_offset(pgd, addr);
1510 next = p4d_addr_end(addr, end);
1511 if (p4d_none_or_clear_bad(p4d))
1513 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1514 } while (p4d++, addr = next, addr != end);
1519 void unmap_page_range(struct mmu_gather *tlb,
1520 struct vm_area_struct *vma,
1521 unsigned long addr, unsigned long end,
1522 struct zap_details *details)
1527 BUG_ON(addr >= end);
1528 tlb_start_vma(tlb, vma);
1529 pgd = pgd_offset(vma->vm_mm, addr);
1531 next = pgd_addr_end(addr, end);
1532 if (pgd_none_or_clear_bad(pgd))
1534 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1535 } while (pgd++, addr = next, addr != end);
1536 tlb_end_vma(tlb, vma);
1540 static void unmap_single_vma(struct mmu_gather *tlb,
1541 struct vm_area_struct *vma, unsigned long start_addr,
1542 unsigned long end_addr,
1543 struct zap_details *details)
1545 unsigned long start = max(vma->vm_start, start_addr);
1548 if (start >= vma->vm_end)
1550 end = min(vma->vm_end, end_addr);
1551 if (end <= vma->vm_start)
1555 uprobe_munmap(vma, start, end);
1557 if (unlikely(vma->vm_flags & VM_PFNMAP))
1558 untrack_pfn(vma, 0, 0);
1561 if (unlikely(is_vm_hugetlb_page(vma))) {
1563 * It is undesirable to test vma->vm_file as it
1564 * should be non-null for valid hugetlb area.
1565 * However, vm_file will be NULL in the error
1566 * cleanup path of mmap_region. When
1567 * hugetlbfs ->mmap method fails,
1568 * mmap_region() nullifies vma->vm_file
1569 * before calling this function to clean up.
1570 * Since no pte has actually been setup, it is
1571 * safe to do nothing in this case.
1574 i_mmap_lock_write(vma->vm_file->f_mapping);
1575 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1576 i_mmap_unlock_write(vma->vm_file->f_mapping);
1579 unmap_page_range(tlb, vma, start, end, details);
1584 * unmap_vmas - unmap a range of memory covered by a list of vma's
1585 * @tlb: address of the caller's struct mmu_gather
1586 * @vma: the starting vma
1587 * @start_addr: virtual address at which to start unmapping
1588 * @end_addr: virtual address at which to end unmapping
1590 * Unmap all pages in the vma list.
1592 * Only addresses between `start' and `end' will be unmapped.
1594 * The VMA list must be sorted in ascending virtual address order.
1596 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1597 * range after unmap_vmas() returns. So the only responsibility here is to
1598 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1599 * drops the lock and schedules.
1601 void unmap_vmas(struct mmu_gather *tlb,
1602 struct vm_area_struct *vma, unsigned long start_addr,
1603 unsigned long end_addr)
1605 struct mm_struct *mm = vma->vm_mm;
1607 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1608 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1609 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1610 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1614 * zap_page_range - remove user pages in a given range
1615 * @vma: vm_area_struct holding the applicable pages
1616 * @start: starting address of pages to zap
1617 * @size: number of bytes to zap
1619 * Caller must protect the VMA list
1621 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1624 struct mm_struct *mm = vma->vm_mm;
1625 struct mmu_gather tlb;
1626 unsigned long end = start + size;
1629 tlb_gather_mmu(&tlb, mm, start, end);
1630 update_hiwater_rss(mm);
1631 mmu_notifier_invalidate_range_start(mm, start, end);
1632 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1633 unmap_single_vma(&tlb, vma, start, end, NULL);
1634 mmu_notifier_invalidate_range_end(mm, start, end);
1635 tlb_finish_mmu(&tlb, start, end);
1639 * zap_page_range_single - remove user pages in a given range
1640 * @vma: vm_area_struct holding the applicable pages
1641 * @address: starting address of pages to zap
1642 * @size: number of bytes to zap
1643 * @details: details of shared cache invalidation
1645 * The range must fit into one VMA.
1647 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1648 unsigned long size, struct zap_details *details)
1650 struct mm_struct *mm = vma->vm_mm;
1651 struct mmu_gather tlb;
1652 unsigned long end = address + size;
1655 tlb_gather_mmu(&tlb, mm, address, end);
1656 update_hiwater_rss(mm);
1657 mmu_notifier_invalidate_range_start(mm, address, end);
1658 unmap_single_vma(&tlb, vma, address, end, details);
1659 mmu_notifier_invalidate_range_end(mm, address, end);
1660 tlb_finish_mmu(&tlb, address, end);
1664 * zap_vma_ptes - remove ptes mapping the vma
1665 * @vma: vm_area_struct holding ptes to be zapped
1666 * @address: starting address of pages to zap
1667 * @size: number of bytes to zap
1669 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1671 * The entire address range must be fully contained within the vma.
1674 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1677 if (address < vma->vm_start || address + size > vma->vm_end ||
1678 !(vma->vm_flags & VM_PFNMAP))
1681 zap_page_range_single(vma, address, size, NULL);
1683 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1685 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1693 pgd = pgd_offset(mm, addr);
1694 p4d = p4d_alloc(mm, pgd, addr);
1697 pud = pud_alloc(mm, p4d, addr);
1700 pmd = pmd_alloc(mm, pud, addr);
1704 VM_BUG_ON(pmd_trans_huge(*pmd));
1705 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1709 * This is the old fallback for page remapping.
1711 * For historical reasons, it only allows reserved pages. Only
1712 * old drivers should use this, and they needed to mark their
1713 * pages reserved for the old functions anyway.
1715 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1716 struct page *page, pgprot_t prot)
1718 struct mm_struct *mm = vma->vm_mm;
1727 flush_dcache_page(page);
1728 pte = get_locked_pte(mm, addr, &ptl);
1732 if (!pte_none(*pte))
1735 /* Ok, finally just insert the thing.. */
1737 inc_mm_counter_fast(mm, mm_counter_file(page));
1738 page_add_file_rmap(page, false);
1739 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1742 pte_unmap_unlock(pte, ptl);
1745 pte_unmap_unlock(pte, ptl);
1751 * vm_insert_page - insert single page into user vma
1752 * @vma: user vma to map to
1753 * @addr: target user address of this page
1754 * @page: source kernel page
1756 * This allows drivers to insert individual pages they've allocated
1759 * The page has to be a nice clean _individual_ kernel allocation.
1760 * If you allocate a compound page, you need to have marked it as
1761 * such (__GFP_COMP), or manually just split the page up yourself
1762 * (see split_page()).
1764 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1765 * took an arbitrary page protection parameter. This doesn't allow
1766 * that. Your vma protection will have to be set up correctly, which
1767 * means that if you want a shared writable mapping, you'd better
1768 * ask for a shared writable mapping!
1770 * The page does not need to be reserved.
1772 * Usually this function is called from f_op->mmap() handler
1773 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1774 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1775 * function from other places, for example from page-fault handler.
1777 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1780 if (addr < vma->vm_start || addr >= vma->vm_end)
1782 if (!page_count(page))
1784 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1785 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1786 BUG_ON(vma->vm_flags & VM_PFNMAP);
1787 vma->vm_flags |= VM_MIXEDMAP;
1789 return insert_page(vma, addr, page, vma->vm_page_prot);
1791 EXPORT_SYMBOL(vm_insert_page);
1793 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1794 pfn_t pfn, pgprot_t prot, bool mkwrite)
1796 struct mm_struct *mm = vma->vm_mm;
1802 pte = get_locked_pte(mm, addr, &ptl);
1806 if (!pte_none(*pte)) {
1809 * For read faults on private mappings the PFN passed
1810 * in may not match the PFN we have mapped if the
1811 * mapped PFN is a writeable COW page. In the mkwrite
1812 * case we are creating a writable PTE for a shared
1813 * mapping and we expect the PFNs to match. If they
1814 * don't match, we are likely racing with block
1815 * allocation and mapping invalidation so just skip the
1818 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1819 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1822 entry = pte_mkyoung(*pte);
1823 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1824 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1825 update_mmu_cache(vma, addr, pte);
1830 /* Ok, finally just insert the thing.. */
1831 if (pfn_t_devmap(pfn))
1832 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1834 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1837 entry = pte_mkyoung(entry);
1838 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1841 set_pte_at(mm, addr, pte, entry);
1842 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1846 pte_unmap_unlock(pte, ptl);
1852 * vm_insert_pfn - insert single pfn into user vma
1853 * @vma: user vma to map to
1854 * @addr: target user address of this page
1855 * @pfn: source kernel pfn
1857 * Similar to vm_insert_page, this allows drivers to insert individual pages
1858 * they've allocated into a user vma. Same comments apply.
1860 * This function should only be called from a vm_ops->fault handler, and
1861 * in that case the handler should return NULL.
1863 * vma cannot be a COW mapping.
1865 * As this is called only for pages that do not currently exist, we
1866 * do not need to flush old virtual caches or the TLB.
1868 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1871 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1873 EXPORT_SYMBOL(vm_insert_pfn);
1876 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1877 * @vma: user vma to map to
1878 * @addr: target user address of this page
1879 * @pfn: source kernel pfn
1880 * @pgprot: pgprot flags for the inserted page
1882 * This is exactly like vm_insert_pfn, except that it allows drivers to
1883 * to override pgprot on a per-page basis.
1885 * This only makes sense for IO mappings, and it makes no sense for
1886 * cow mappings. In general, using multiple vmas is preferable;
1887 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1890 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1891 unsigned long pfn, pgprot_t pgprot)
1895 * Technically, architectures with pte_special can avoid all these
1896 * restrictions (same for remap_pfn_range). However we would like
1897 * consistency in testing and feature parity among all, so we should
1898 * try to keep these invariants in place for everybody.
1900 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1901 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1902 (VM_PFNMAP|VM_MIXEDMAP));
1903 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1904 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1906 if (addr < vma->vm_start || addr >= vma->vm_end)
1909 if (!pfn_modify_allowed(pfn, pgprot))
1912 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1914 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1919 EXPORT_SYMBOL(vm_insert_pfn_prot);
1921 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1923 /* these checks mirror the abort conditions in vm_normal_page */
1924 if (vma->vm_flags & VM_MIXEDMAP)
1926 if (pfn_t_devmap(pfn))
1928 if (pfn_t_special(pfn))
1930 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1935 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1936 pfn_t pfn, bool mkwrite)
1938 pgprot_t pgprot = vma->vm_page_prot;
1940 BUG_ON(!vm_mixed_ok(vma, pfn));
1942 if (addr < vma->vm_start || addr >= vma->vm_end)
1945 track_pfn_insert(vma, &pgprot, pfn);
1947 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1951 * If we don't have pte special, then we have to use the pfn_valid()
1952 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1953 * refcount the page if pfn_valid is true (hence insert_page rather
1954 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1955 * without pte special, it would there be refcounted as a normal page.
1957 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1958 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1962 * At this point we are committed to insert_page()
1963 * regardless of whether the caller specified flags that
1964 * result in pfn_t_has_page() == false.
1966 page = pfn_to_page(pfn_t_to_pfn(pfn));
1967 return insert_page(vma, addr, page, pgprot);
1969 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1972 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1975 return __vm_insert_mixed(vma, addr, pfn, false);
1978 EXPORT_SYMBOL(vm_insert_mixed);
1981 * If the insertion of PTE failed because someone else already added a
1982 * different entry in the mean time, we treat that as success as we assume
1983 * the same entry was actually inserted.
1986 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1987 unsigned long addr, pfn_t pfn)
1991 err = __vm_insert_mixed(vma, addr, pfn, true);
1993 return VM_FAULT_OOM;
1994 if (err < 0 && err != -EBUSY)
1995 return VM_FAULT_SIGBUS;
1996 return VM_FAULT_NOPAGE;
1998 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2001 * maps a range of physical memory into the requested pages. the old
2002 * mappings are removed. any references to nonexistent pages results
2003 * in null mappings (currently treated as "copy-on-access")
2005 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2006 unsigned long addr, unsigned long end,
2007 unsigned long pfn, pgprot_t prot)
2009 pte_t *pte, *mapped_pte;
2013 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2016 arch_enter_lazy_mmu_mode();
2018 BUG_ON(!pte_none(*pte));
2019 if (!pfn_modify_allowed(pfn, prot)) {
2023 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2025 } while (pte++, addr += PAGE_SIZE, addr != end);
2026 arch_leave_lazy_mmu_mode();
2027 pte_unmap_unlock(mapped_pte, ptl);
2031 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2032 unsigned long addr, unsigned long end,
2033 unsigned long pfn, pgprot_t prot)
2039 pfn -= addr >> PAGE_SHIFT;
2040 pmd = pmd_alloc(mm, pud, addr);
2043 VM_BUG_ON(pmd_trans_huge(*pmd));
2045 next = pmd_addr_end(addr, end);
2046 err = remap_pte_range(mm, pmd, addr, next,
2047 pfn + (addr >> PAGE_SHIFT), prot);
2050 } while (pmd++, addr = next, addr != end);
2054 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2055 unsigned long addr, unsigned long end,
2056 unsigned long pfn, pgprot_t prot)
2062 pfn -= addr >> PAGE_SHIFT;
2063 pud = pud_alloc(mm, p4d, addr);
2067 next = pud_addr_end(addr, end);
2068 err = remap_pmd_range(mm, pud, addr, next,
2069 pfn + (addr >> PAGE_SHIFT), prot);
2072 } while (pud++, addr = next, addr != end);
2076 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2077 unsigned long addr, unsigned long end,
2078 unsigned long pfn, pgprot_t prot)
2084 pfn -= addr >> PAGE_SHIFT;
2085 p4d = p4d_alloc(mm, pgd, addr);
2089 next = p4d_addr_end(addr, end);
2090 err = remap_pud_range(mm, p4d, addr, next,
2091 pfn + (addr >> PAGE_SHIFT), prot);
2094 } while (p4d++, addr = next, addr != end);
2099 * remap_pfn_range - remap kernel memory to userspace
2100 * @vma: user vma to map to
2101 * @addr: target user address to start at
2102 * @pfn: physical address of kernel memory
2103 * @size: size of map area
2104 * @prot: page protection flags for this mapping
2106 * Note: this is only safe if the mm semaphore is held when called.
2108 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2109 unsigned long pfn, unsigned long size, pgprot_t prot)
2113 unsigned long end = addr + PAGE_ALIGN(size);
2114 struct mm_struct *mm = vma->vm_mm;
2115 unsigned long remap_pfn = pfn;
2119 * Physically remapped pages are special. Tell the
2120 * rest of the world about it:
2121 * VM_IO tells people not to look at these pages
2122 * (accesses can have side effects).
2123 * VM_PFNMAP tells the core MM that the base pages are just
2124 * raw PFN mappings, and do not have a "struct page" associated
2127 * Disable vma merging and expanding with mremap().
2129 * Omit vma from core dump, even when VM_IO turned off.
2131 * There's a horrible special case to handle copy-on-write
2132 * behaviour that some programs depend on. We mark the "original"
2133 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2134 * See vm_normal_page() for details.
2136 if (is_cow_mapping(vma->vm_flags)) {
2137 if (addr != vma->vm_start || end != vma->vm_end)
2139 vma->vm_pgoff = pfn;
2142 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2146 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2148 BUG_ON(addr >= end);
2149 pfn -= addr >> PAGE_SHIFT;
2150 pgd = pgd_offset(mm, addr);
2151 flush_cache_range(vma, addr, end);
2153 next = pgd_addr_end(addr, end);
2154 err = remap_p4d_range(mm, pgd, addr, next,
2155 pfn + (addr >> PAGE_SHIFT), prot);
2158 } while (pgd++, addr = next, addr != end);
2161 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2165 EXPORT_SYMBOL(remap_pfn_range);
2168 * vm_iomap_memory - remap memory to userspace
2169 * @vma: user vma to map to
2170 * @start: start of area
2171 * @len: size of area
2173 * This is a simplified io_remap_pfn_range() for common driver use. The
2174 * driver just needs to give us the physical memory range to be mapped,
2175 * we'll figure out the rest from the vma information.
2177 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2178 * whatever write-combining details or similar.
2180 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2182 unsigned long vm_len, pfn, pages;
2184 /* Check that the physical memory area passed in looks valid */
2185 if (start + len < start)
2188 * You *really* shouldn't map things that aren't page-aligned,
2189 * but we've historically allowed it because IO memory might
2190 * just have smaller alignment.
2192 len += start & ~PAGE_MASK;
2193 pfn = start >> PAGE_SHIFT;
2194 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2195 if (pfn + pages < pfn)
2198 /* We start the mapping 'vm_pgoff' pages into the area */
2199 if (vma->vm_pgoff > pages)
2201 pfn += vma->vm_pgoff;
2202 pages -= vma->vm_pgoff;
2204 /* Can we fit all of the mapping? */
2205 vm_len = vma->vm_end - vma->vm_start;
2206 if (vm_len >> PAGE_SHIFT > pages)
2209 /* Ok, let it rip */
2210 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2212 EXPORT_SYMBOL(vm_iomap_memory);
2214 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2215 unsigned long addr, unsigned long end,
2216 pte_fn_t fn, void *data)
2221 spinlock_t *uninitialized_var(ptl);
2223 pte = (mm == &init_mm) ?
2224 pte_alloc_kernel(pmd, addr) :
2225 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2229 BUG_ON(pmd_huge(*pmd));
2231 arch_enter_lazy_mmu_mode();
2233 token = pmd_pgtable(*pmd);
2236 err = fn(pte++, token, addr, data);
2239 } while (addr += PAGE_SIZE, addr != end);
2241 arch_leave_lazy_mmu_mode();
2244 pte_unmap_unlock(pte-1, ptl);
2248 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2249 unsigned long addr, unsigned long end,
2250 pte_fn_t fn, void *data)
2256 BUG_ON(pud_huge(*pud));
2258 pmd = pmd_alloc(mm, pud, addr);
2262 next = pmd_addr_end(addr, end);
2263 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2266 } while (pmd++, addr = next, addr != end);
2270 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2271 unsigned long addr, unsigned long end,
2272 pte_fn_t fn, void *data)
2278 pud = pud_alloc(mm, p4d, addr);
2282 next = pud_addr_end(addr, end);
2283 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2286 } while (pud++, addr = next, addr != end);
2290 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2291 unsigned long addr, unsigned long end,
2292 pte_fn_t fn, void *data)
2298 p4d = p4d_alloc(mm, pgd, addr);
2302 next = p4d_addr_end(addr, end);
2303 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2306 } while (p4d++, addr = next, addr != end);
2311 * Scan a region of virtual memory, filling in page tables as necessary
2312 * and calling a provided function on each leaf page table.
2314 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2315 unsigned long size, pte_fn_t fn, void *data)
2319 unsigned long end = addr + size;
2322 if (WARN_ON(addr >= end))
2325 pgd = pgd_offset(mm, addr);
2327 next = pgd_addr_end(addr, end);
2328 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2331 } while (pgd++, addr = next, addr != end);
2335 EXPORT_SYMBOL_GPL(apply_to_page_range);
2338 * handle_pte_fault chooses page fault handler according to an entry which was
2339 * read non-atomically. Before making any commitment, on those architectures
2340 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2341 * parts, do_swap_page must check under lock before unmapping the pte and
2342 * proceeding (but do_wp_page is only called after already making such a check;
2343 * and do_anonymous_page can safely check later on).
2345 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2346 pte_t *page_table, pte_t orig_pte)
2349 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2350 if (sizeof(pte_t) > sizeof(unsigned long)) {
2351 spinlock_t *ptl = pte_lockptr(mm, pmd);
2353 same = pte_same(*page_table, orig_pte);
2357 pte_unmap(page_table);
2361 static inline bool cow_user_page(struct page *dst, struct page *src,
2362 struct vm_fault *vmf)
2367 bool locked = false;
2368 struct vm_area_struct *vma = vmf->vma;
2369 struct mm_struct *mm = vma->vm_mm;
2370 unsigned long addr = vmf->address;
2372 debug_dma_assert_idle(src);
2375 copy_user_highpage(dst, src, addr, vma);
2380 * If the source page was a PFN mapping, we don't have
2381 * a "struct page" for it. We do a best-effort copy by
2382 * just copying from the original user address. If that
2383 * fails, we just zero-fill it. Live with it.
2385 kaddr = kmap_atomic(dst);
2386 uaddr = (void __user *)(addr & PAGE_MASK);
2389 * On architectures with software "accessed" bits, we would
2390 * take a double page fault, so mark it accessed here.
2392 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2395 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2397 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2399 * Other thread has already handled the fault
2400 * and we don't need to do anything. If it's
2401 * not the case, the fault will be triggered
2402 * again on the same address.
2408 entry = pte_mkyoung(vmf->orig_pte);
2409 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2410 update_mmu_cache(vma, addr, vmf->pte);
2414 * This really shouldn't fail, because the page is there
2415 * in the page tables. But it might just be unreadable,
2416 * in which case we just give up and fill the result with
2419 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2423 /* Re-validate under PTL if the page is still mapped */
2424 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2426 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2427 /* The PTE changed under us. Retry page fault. */
2433 * The same page can be mapped back since last copy attampt.
2434 * Try to copy again under PTL.
2436 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2438 * Give a warn in case there can be some obscure
2451 pte_unmap_unlock(vmf->pte, vmf->ptl);
2452 kunmap_atomic(kaddr);
2453 flush_dcache_page(dst);
2458 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2460 struct file *vm_file = vma->vm_file;
2463 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2466 * Special mappings (e.g. VDSO) do not have any file so fake
2467 * a default GFP_KERNEL for them.
2473 * Notify the address space that the page is about to become writable so that
2474 * it can prohibit this or wait for the page to get into an appropriate state.
2476 * We do this without the lock held, so that it can sleep if it needs to.
2478 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2481 struct page *page = vmf->page;
2482 unsigned int old_flags = vmf->flags;
2484 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2486 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2487 /* Restore original flags so that caller is not surprised */
2488 vmf->flags = old_flags;
2489 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2491 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2493 if (!page->mapping) {
2495 return 0; /* retry */
2497 ret |= VM_FAULT_LOCKED;
2499 VM_BUG_ON_PAGE(!PageLocked(page), page);
2504 * Handle dirtying of a page in shared file mapping on a write fault.
2506 * The function expects the page to be locked and unlocks it.
2508 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2511 struct address_space *mapping;
2513 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2515 dirtied = set_page_dirty(page);
2516 VM_BUG_ON_PAGE(PageAnon(page), page);
2518 * Take a local copy of the address_space - page.mapping may be zeroed
2519 * by truncate after unlock_page(). The address_space itself remains
2520 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2521 * release semantics to prevent the compiler from undoing this copying.
2523 mapping = page_rmapping(page);
2526 if ((dirtied || page_mkwrite) && mapping) {
2528 * Some device drivers do not set page.mapping
2529 * but still dirty their pages
2531 balance_dirty_pages_ratelimited(mapping);
2535 file_update_time(vma->vm_file);
2539 * Handle write page faults for pages that can be reused in the current vma
2541 * This can happen either due to the mapping being with the VM_SHARED flag,
2542 * or due to us being the last reference standing to the page. In either
2543 * case, all we need to do here is to mark the page as writable and update
2544 * any related book-keeping.
2546 static inline void wp_page_reuse(struct vm_fault *vmf)
2547 __releases(vmf->ptl)
2549 struct vm_area_struct *vma = vmf->vma;
2550 struct page *page = vmf->page;
2553 * Clear the pages cpupid information as the existing
2554 * information potentially belongs to a now completely
2555 * unrelated process.
2558 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2560 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2561 entry = pte_mkyoung(vmf->orig_pte);
2562 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2563 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2564 update_mmu_cache(vma, vmf->address, vmf->pte);
2565 pte_unmap_unlock(vmf->pte, vmf->ptl);
2569 * Handle the case of a page which we actually need to copy to a new page.
2571 * Called with mmap_sem locked and the old page referenced, but
2572 * without the ptl held.
2574 * High level logic flow:
2576 * - Allocate a page, copy the content of the old page to the new one.
2577 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2578 * - Take the PTL. If the pte changed, bail out and release the allocated page
2579 * - If the pte is still the way we remember it, update the page table and all
2580 * relevant references. This includes dropping the reference the page-table
2581 * held to the old page, as well as updating the rmap.
2582 * - In any case, unlock the PTL and drop the reference we took to the old page.
2584 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2586 struct vm_area_struct *vma = vmf->vma;
2587 struct mm_struct *mm = vma->vm_mm;
2588 struct page *old_page = vmf->page;
2589 struct page *new_page = NULL;
2591 int page_copied = 0;
2592 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2593 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2594 struct mem_cgroup *memcg;
2596 if (unlikely(anon_vma_prepare(vma)))
2599 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2600 new_page = alloc_zeroed_user_highpage_movable(vma,
2605 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2610 if (!cow_user_page(new_page, old_page, vmf)) {
2612 * COW failed, if the fault was solved by other,
2613 * it's fine. If not, userspace would re-fault on
2614 * the same address and we will handle the fault
2615 * from the second attempt.
2624 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2627 __SetPageUptodate(new_page);
2629 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2632 * Re-check the pte - we dropped the lock
2634 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2635 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2637 if (!PageAnon(old_page)) {
2638 dec_mm_counter_fast(mm,
2639 mm_counter_file(old_page));
2640 inc_mm_counter_fast(mm, MM_ANONPAGES);
2643 inc_mm_counter_fast(mm, MM_ANONPAGES);
2645 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2646 entry = mk_pte(new_page, vma->vm_page_prot);
2647 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2649 * Clear the pte entry and flush it first, before updating the
2650 * pte with the new entry. This will avoid a race condition
2651 * seen in the presence of one thread doing SMC and another
2654 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2655 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2656 mem_cgroup_commit_charge(new_page, memcg, false, false);
2657 lru_cache_add_active_or_unevictable(new_page, vma);
2659 * We call the notify macro here because, when using secondary
2660 * mmu page tables (such as kvm shadow page tables), we want the
2661 * new page to be mapped directly into the secondary page table.
2663 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2664 update_mmu_cache(vma, vmf->address, vmf->pte);
2667 * Only after switching the pte to the new page may
2668 * we remove the mapcount here. Otherwise another
2669 * process may come and find the rmap count decremented
2670 * before the pte is switched to the new page, and
2671 * "reuse" the old page writing into it while our pte
2672 * here still points into it and can be read by other
2675 * The critical issue is to order this
2676 * page_remove_rmap with the ptp_clear_flush above.
2677 * Those stores are ordered by (if nothing else,)
2678 * the barrier present in the atomic_add_negative
2679 * in page_remove_rmap.
2681 * Then the TLB flush in ptep_clear_flush ensures that
2682 * no process can access the old page before the
2683 * decremented mapcount is visible. And the old page
2684 * cannot be reused until after the decremented
2685 * mapcount is visible. So transitively, TLBs to
2686 * old page will be flushed before it can be reused.
2688 page_remove_rmap(old_page, false);
2691 /* Free the old page.. */
2692 new_page = old_page;
2695 mem_cgroup_cancel_charge(new_page, memcg, false);
2701 pte_unmap_unlock(vmf->pte, vmf->ptl);
2703 * No need to double call mmu_notifier->invalidate_range() callback as
2704 * the above ptep_clear_flush_notify() did already call it.
2706 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2709 * Don't let another task, with possibly unlocked vma,
2710 * keep the mlocked page.
2712 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2713 lock_page(old_page); /* LRU manipulation */
2714 if (PageMlocked(old_page))
2715 munlock_vma_page(old_page);
2716 unlock_page(old_page);
2720 return page_copied ? VM_FAULT_WRITE : 0;
2726 return VM_FAULT_OOM;
2730 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2731 * writeable once the page is prepared
2733 * @vmf: structure describing the fault
2735 * This function handles all that is needed to finish a write page fault in a
2736 * shared mapping due to PTE being read-only once the mapped page is prepared.
2737 * It handles locking of PTE and modifying it. The function returns
2738 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2741 * The function expects the page to be locked or other protection against
2742 * concurrent faults / writeback (such as DAX radix tree locks).
2744 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2746 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2747 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2750 * We might have raced with another page fault while we released the
2751 * pte_offset_map_lock.
2753 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2754 pte_unmap_unlock(vmf->pte, vmf->ptl);
2755 return VM_FAULT_NOPAGE;
2762 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2765 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2767 struct vm_area_struct *vma = vmf->vma;
2769 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2772 pte_unmap_unlock(vmf->pte, vmf->ptl);
2773 vmf->flags |= FAULT_FLAG_MKWRITE;
2774 ret = vma->vm_ops->pfn_mkwrite(vmf);
2775 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2777 return finish_mkwrite_fault(vmf);
2780 return VM_FAULT_WRITE;
2783 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2784 __releases(vmf->ptl)
2786 struct vm_area_struct *vma = vmf->vma;
2788 get_page(vmf->page);
2790 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2793 pte_unmap_unlock(vmf->pte, vmf->ptl);
2794 tmp = do_page_mkwrite(vmf);
2795 if (unlikely(!tmp || (tmp &
2796 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2797 put_page(vmf->page);
2800 tmp = finish_mkwrite_fault(vmf);
2801 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2802 unlock_page(vmf->page);
2803 put_page(vmf->page);
2808 lock_page(vmf->page);
2810 fault_dirty_shared_page(vma, vmf->page);
2811 put_page(vmf->page);
2813 return VM_FAULT_WRITE;
2817 * This routine handles present pages, when users try to write
2818 * to a shared page. It is done by copying the page to a new address
2819 * and decrementing the shared-page counter for the old page.
2821 * Note that this routine assumes that the protection checks have been
2822 * done by the caller (the low-level page fault routine in most cases).
2823 * Thus we can safely just mark it writable once we've done any necessary
2826 * We also mark the page dirty at this point even though the page will
2827 * change only once the write actually happens. This avoids a few races,
2828 * and potentially makes it more efficient.
2830 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2831 * but allow concurrent faults), with pte both mapped and locked.
2832 * We return with mmap_sem still held, but pte unmapped and unlocked.
2834 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2835 __releases(vmf->ptl)
2837 struct vm_area_struct *vma = vmf->vma;
2839 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2842 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2845 * We should not cow pages in a shared writeable mapping.
2846 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2848 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2849 (VM_WRITE|VM_SHARED))
2850 return wp_pfn_shared(vmf);
2852 pte_unmap_unlock(vmf->pte, vmf->ptl);
2853 return wp_page_copy(vmf);
2857 * Take out anonymous pages first, anonymous shared vmas are
2858 * not dirty accountable.
2860 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2861 int total_map_swapcount;
2862 if (!trylock_page(vmf->page)) {
2863 get_page(vmf->page);
2864 pte_unmap_unlock(vmf->pte, vmf->ptl);
2865 lock_page(vmf->page);
2866 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2867 vmf->address, &vmf->ptl);
2868 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2869 unlock_page(vmf->page);
2870 pte_unmap_unlock(vmf->pte, vmf->ptl);
2871 put_page(vmf->page);
2874 put_page(vmf->page);
2876 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2877 if (total_map_swapcount == 1) {
2879 * The page is all ours. Move it to
2880 * our anon_vma so the rmap code will
2881 * not search our parent or siblings.
2882 * Protected against the rmap code by
2885 page_move_anon_rmap(vmf->page, vma);
2887 unlock_page(vmf->page);
2889 return VM_FAULT_WRITE;
2891 unlock_page(vmf->page);
2892 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2893 (VM_WRITE|VM_SHARED))) {
2894 return wp_page_shared(vmf);
2898 * Ok, we need to copy. Oh, well..
2900 get_page(vmf->page);
2902 pte_unmap_unlock(vmf->pte, vmf->ptl);
2903 return wp_page_copy(vmf);
2906 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2907 unsigned long start_addr, unsigned long end_addr,
2908 struct zap_details *details)
2910 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2913 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2914 struct zap_details *details)
2916 struct vm_area_struct *vma;
2917 pgoff_t vba, vea, zba, zea;
2919 vma_interval_tree_foreach(vma, root,
2920 details->first_index, details->last_index) {
2922 vba = vma->vm_pgoff;
2923 vea = vba + vma_pages(vma) - 1;
2924 zba = details->first_index;
2927 zea = details->last_index;
2931 unmap_mapping_range_vma(vma,
2932 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2933 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2939 * unmap_mapping_page() - Unmap single page from processes.
2940 * @page: The locked page to be unmapped.
2942 * Unmap this page from any userspace process which still has it mmaped.
2943 * Typically, for efficiency, the range of nearby pages has already been
2944 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
2945 * truncation or invalidation holds the lock on a page, it may find that
2946 * the page has been remapped again: and then uses unmap_mapping_page()
2947 * to unmap it finally.
2949 void unmap_mapping_page(struct page *page)
2951 struct address_space *mapping = page->mapping;
2952 struct zap_details details = { };
2954 VM_BUG_ON(!PageLocked(page));
2955 VM_BUG_ON(PageTail(page));
2957 details.check_mapping = mapping;
2958 details.first_index = page->index;
2959 details.last_index = page->index + hpage_nr_pages(page) - 1;
2960 details.single_page = page;
2962 i_mmap_lock_write(mapping);
2963 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2964 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2965 i_mmap_unlock_write(mapping);
2969 * unmap_mapping_pages() - Unmap pages from processes.
2970 * @mapping: The address space containing pages to be unmapped.
2971 * @start: Index of first page to be unmapped.
2972 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2973 * @even_cows: Whether to unmap even private COWed pages.
2975 * Unmap the pages in this address space from any userspace process which
2976 * has them mmaped. Generally, you want to remove COWed pages as well when
2977 * a file is being truncated, but not when invalidating pages from the page
2980 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2981 pgoff_t nr, bool even_cows)
2983 struct zap_details details = { };
2985 details.check_mapping = even_cows ? NULL : mapping;
2986 details.first_index = start;
2987 details.last_index = start + nr - 1;
2988 if (details.last_index < details.first_index)
2989 details.last_index = ULONG_MAX;
2991 i_mmap_lock_write(mapping);
2992 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2993 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2994 i_mmap_unlock_write(mapping);
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2999 * address_space corresponding to the specified byte range in the underlying
3002 * @mapping: the address space containing mmaps to be unmapped.
3003 * @holebegin: byte in first page to unmap, relative to the start of
3004 * the underlying file. This will be rounded down to a PAGE_SIZE
3005 * boundary. Note that this is different from truncate_pagecache(), which
3006 * must keep the partial page. In contrast, we must get rid of
3008 * @holelen: size of prospective hole in bytes. This will be rounded
3009 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3011 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3012 * but 0 when invalidating pagecache, don't throw away private data.
3014 void unmap_mapping_range(struct address_space *mapping,
3015 loff_t const holebegin, loff_t const holelen, int even_cows)
3017 pgoff_t hba = holebegin >> PAGE_SHIFT;
3018 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3020 /* Check for overflow. */
3021 if (sizeof(holelen) > sizeof(hlen)) {
3023 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3024 if (holeend & ~(long long)ULONG_MAX)
3025 hlen = ULONG_MAX - hba + 1;
3028 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3030 EXPORT_SYMBOL(unmap_mapping_range);
3033 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3034 * but allow concurrent faults), and pte mapped but not yet locked.
3035 * We return with pte unmapped and unlocked.
3037 * We return with the mmap_sem locked or unlocked in the same cases
3038 * as does filemap_fault().
3040 vm_fault_t do_swap_page(struct vm_fault *vmf)
3042 struct vm_area_struct *vma = vmf->vma;
3043 struct page *page = NULL, *swapcache;
3044 struct mem_cgroup *memcg;
3051 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3054 entry = pte_to_swp_entry(vmf->orig_pte);
3055 if (unlikely(non_swap_entry(entry))) {
3056 if (is_migration_entry(entry)) {
3057 migration_entry_wait(vma->vm_mm, vmf->pmd,
3059 } else if (is_device_private_entry(entry)) {
3061 * For un-addressable device memory we call the pgmap
3062 * fault handler callback. The callback must migrate
3063 * the page back to some CPU accessible page.
3065 ret = device_private_entry_fault(vma, vmf->address, entry,
3066 vmf->flags, vmf->pmd);
3067 } else if (is_hwpoison_entry(entry)) {
3068 ret = VM_FAULT_HWPOISON;
3070 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3071 ret = VM_FAULT_SIGBUS;
3077 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3078 page = lookup_swap_cache(entry, vma, vmf->address);
3082 struct swap_info_struct *si = swp_swap_info(entry);
3084 if (si->flags & SWP_SYNCHRONOUS_IO &&
3085 __swap_count(si, entry) == 1) {
3086 /* skip swapcache */
3087 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3090 __SetPageLocked(page);
3091 __SetPageSwapBacked(page);
3092 set_page_private(page, entry.val);
3093 lru_cache_add_anon(page);
3094 swap_readpage(page, true);
3097 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3104 * Back out if somebody else faulted in this pte
3105 * while we released the pte lock.
3107 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3108 vmf->address, &vmf->ptl);
3109 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3115 /* Had to read the page from swap area: Major fault */
3116 ret = VM_FAULT_MAJOR;
3117 count_vm_event(PGMAJFAULT);
3118 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3119 } else if (PageHWPoison(page)) {
3121 * hwpoisoned dirty swapcache pages are kept for killing
3122 * owner processes (which may be unknown at hwpoison time)
3124 ret = VM_FAULT_HWPOISON;
3125 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3129 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3131 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3133 ret |= VM_FAULT_RETRY;
3138 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3139 * release the swapcache from under us. The page pin, and pte_same
3140 * test below, are not enough to exclude that. Even if it is still
3141 * swapcache, we need to check that the page's swap has not changed.
3143 if (unlikely((!PageSwapCache(page) ||
3144 page_private(page) != entry.val)) && swapcache)
3147 page = ksm_might_need_to_copy(page, vma, vmf->address);
3148 if (unlikely(!page)) {
3154 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3161 * Back out if somebody else already faulted in this pte.
3163 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3165 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3168 if (unlikely(!PageUptodate(page))) {
3169 ret = VM_FAULT_SIGBUS;
3174 * The page isn't present yet, go ahead with the fault.
3176 * Be careful about the sequence of operations here.
3177 * To get its accounting right, reuse_swap_page() must be called
3178 * while the page is counted on swap but not yet in mapcount i.e.
3179 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3180 * must be called after the swap_free(), or it will never succeed.
3183 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3184 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3185 pte = mk_pte(page, vma->vm_page_prot);
3186 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3187 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3188 vmf->flags &= ~FAULT_FLAG_WRITE;
3189 ret |= VM_FAULT_WRITE;
3190 exclusive = RMAP_EXCLUSIVE;
3192 flush_icache_page(vma, page);
3193 if (pte_swp_soft_dirty(vmf->orig_pte))
3194 pte = pte_mksoft_dirty(pte);
3195 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3196 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3197 vmf->orig_pte = pte;
3199 /* ksm created a completely new copy */
3200 if (unlikely(page != swapcache && swapcache)) {
3201 page_add_new_anon_rmap(page, vma, vmf->address, false);
3202 mem_cgroup_commit_charge(page, memcg, false, false);
3203 lru_cache_add_active_or_unevictable(page, vma);
3205 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3206 mem_cgroup_commit_charge(page, memcg, true, false);
3207 activate_page(page);
3211 if (mem_cgroup_swap_full(page) ||
3212 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3213 try_to_free_swap(page);
3215 if (page != swapcache && swapcache) {
3217 * Hold the lock to avoid the swap entry to be reused
3218 * until we take the PT lock for the pte_same() check
3219 * (to avoid false positives from pte_same). For
3220 * further safety release the lock after the swap_free
3221 * so that the swap count won't change under a
3222 * parallel locked swapcache.
3224 unlock_page(swapcache);
3225 put_page(swapcache);
3228 if (vmf->flags & FAULT_FLAG_WRITE) {
3229 ret |= do_wp_page(vmf);
3230 if (ret & VM_FAULT_ERROR)
3231 ret &= VM_FAULT_ERROR;
3235 /* No need to invalidate - it was non-present before */
3236 update_mmu_cache(vma, vmf->address, vmf->pte);
3238 pte_unmap_unlock(vmf->pte, vmf->ptl);
3242 mem_cgroup_cancel_charge(page, memcg, false);
3243 pte_unmap_unlock(vmf->pte, vmf->ptl);
3248 if (page != swapcache && swapcache) {
3249 unlock_page(swapcache);
3250 put_page(swapcache);
3256 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3257 * but allow concurrent faults), and pte mapped but not yet locked.
3258 * We return with mmap_sem still held, but pte unmapped and unlocked.
3260 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3262 struct vm_area_struct *vma = vmf->vma;
3263 struct mem_cgroup *memcg;
3268 /* File mapping without ->vm_ops ? */
3269 if (vma->vm_flags & VM_SHARED)
3270 return VM_FAULT_SIGBUS;
3273 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3274 * pte_offset_map() on pmds where a huge pmd might be created
3275 * from a different thread.
3277 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3278 * parallel threads are excluded by other means.
3280 * Here we only have down_read(mmap_sem).
3282 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3283 return VM_FAULT_OOM;
3285 /* See the comment in pte_alloc_one_map() */
3286 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3289 /* Use the zero-page for reads */
3290 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3291 !mm_forbids_zeropage(vma->vm_mm)) {
3292 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3293 vma->vm_page_prot));
3294 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3295 vmf->address, &vmf->ptl);
3296 if (!pte_none(*vmf->pte))
3298 ret = check_stable_address_space(vma->vm_mm);
3301 /* Deliver the page fault to userland, check inside PT lock */
3302 if (userfaultfd_missing(vma)) {
3303 pte_unmap_unlock(vmf->pte, vmf->ptl);
3304 return handle_userfault(vmf, VM_UFFD_MISSING);
3309 /* Allocate our own private page. */
3310 if (unlikely(anon_vma_prepare(vma)))
3312 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3316 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3321 * The memory barrier inside __SetPageUptodate makes sure that
3322 * preceeding stores to the page contents become visible before
3323 * the set_pte_at() write.
3325 __SetPageUptodate(page);
3327 entry = mk_pte(page, vma->vm_page_prot);
3328 if (vma->vm_flags & VM_WRITE)
3329 entry = pte_mkwrite(pte_mkdirty(entry));
3331 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3333 if (!pte_none(*vmf->pte))
3336 ret = check_stable_address_space(vma->vm_mm);
3340 /* Deliver the page fault to userland, check inside PT lock */
3341 if (userfaultfd_missing(vma)) {
3342 pte_unmap_unlock(vmf->pte, vmf->ptl);
3343 mem_cgroup_cancel_charge(page, memcg, false);
3345 return handle_userfault(vmf, VM_UFFD_MISSING);
3348 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3349 page_add_new_anon_rmap(page, vma, vmf->address, false);
3350 mem_cgroup_commit_charge(page, memcg, false, false);
3351 lru_cache_add_active_or_unevictable(page, vma);
3353 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3355 /* No need to invalidate - it was non-present before */
3356 update_mmu_cache(vma, vmf->address, vmf->pte);
3358 pte_unmap_unlock(vmf->pte, vmf->ptl);
3361 mem_cgroup_cancel_charge(page, memcg, false);
3367 return VM_FAULT_OOM;
3371 * The mmap_sem must have been held on entry, and may have been
3372 * released depending on flags and vma->vm_ops->fault() return value.
3373 * See filemap_fault() and __lock_page_retry().
3375 static vm_fault_t __do_fault(struct vm_fault *vmf)
3377 struct vm_area_struct *vma = vmf->vma;
3381 * Preallocate pte before we take page_lock because this might lead to
3382 * deadlocks for memcg reclaim which waits for pages under writeback:
3384 * SetPageWriteback(A)
3390 * wait_on_page_writeback(A)
3391 * SetPageWriteback(B)
3393 * # flush A, B to clear the writeback
3395 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3396 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3398 if (!vmf->prealloc_pte)
3399 return VM_FAULT_OOM;
3400 smp_wmb(); /* See comment in __pte_alloc() */
3403 ret = vma->vm_ops->fault(vmf);
3404 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3405 VM_FAULT_DONE_COW)))
3408 if (unlikely(PageHWPoison(vmf->page))) {
3409 if (ret & VM_FAULT_LOCKED)
3410 unlock_page(vmf->page);
3411 put_page(vmf->page);
3413 return VM_FAULT_HWPOISON;
3416 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3417 lock_page(vmf->page);
3419 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3425 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3426 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3427 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3428 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3430 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3432 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3435 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3437 struct vm_area_struct *vma = vmf->vma;
3439 if (!pmd_none(*vmf->pmd))
3441 if (vmf->prealloc_pte) {
3442 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3443 if (unlikely(!pmd_none(*vmf->pmd))) {
3444 spin_unlock(vmf->ptl);
3448 mm_inc_nr_ptes(vma->vm_mm);
3449 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3450 spin_unlock(vmf->ptl);
3451 vmf->prealloc_pte = NULL;
3452 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3453 return VM_FAULT_OOM;
3457 * If a huge pmd materialized under us just retry later. Use
3458 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3459 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3460 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3461 * running immediately after a huge pmd fault in a different thread of
3462 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3463 * All we have to ensure is that it is a regular pmd that we can walk
3464 * with pte_offset_map() and we can do that through an atomic read in
3465 * C, which is what pmd_trans_unstable() provides.
3467 if (pmd_devmap_trans_unstable(vmf->pmd))
3468 return VM_FAULT_NOPAGE;
3471 * At this point we know that our vmf->pmd points to a page of ptes
3472 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3473 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3474 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3475 * be valid and we will re-check to make sure the vmf->pte isn't
3476 * pte_none() under vmf->ptl protection when we return to
3479 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3484 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3486 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3487 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3488 unsigned long haddr)
3490 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3491 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3493 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3498 static void deposit_prealloc_pte(struct vm_fault *vmf)
3500 struct vm_area_struct *vma = vmf->vma;
3502 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3504 * We are going to consume the prealloc table,
3505 * count that as nr_ptes.
3507 mm_inc_nr_ptes(vma->vm_mm);
3508 vmf->prealloc_pte = NULL;
3511 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3513 struct vm_area_struct *vma = vmf->vma;
3514 bool write = vmf->flags & FAULT_FLAG_WRITE;
3515 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3520 if (!transhuge_vma_suitable(vma, haddr))
3521 return VM_FAULT_FALLBACK;
3523 ret = VM_FAULT_FALLBACK;
3524 page = compound_head(page);
3527 * Archs like ppc64 need additonal space to store information
3528 * related to pte entry. Use the preallocated table for that.
3530 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3531 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3532 if (!vmf->prealloc_pte)
3533 return VM_FAULT_OOM;
3534 smp_wmb(); /* See comment in __pte_alloc() */
3537 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3538 if (unlikely(!pmd_none(*vmf->pmd)))
3541 for (i = 0; i < HPAGE_PMD_NR; i++)
3542 flush_icache_page(vma, page + i);
3544 entry = mk_huge_pmd(page, vma->vm_page_prot);
3546 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3548 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3549 page_add_file_rmap(page, true);
3551 * deposit and withdraw with pmd lock held
3553 if (arch_needs_pgtable_deposit())
3554 deposit_prealloc_pte(vmf);
3556 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3558 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3560 /* fault is handled */
3562 count_vm_event(THP_FILE_MAPPED);
3564 spin_unlock(vmf->ptl);
3568 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3576 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3577 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3579 * @vmf: fault environment
3580 * @memcg: memcg to charge page (only for private mappings)
3581 * @page: page to map
3583 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3586 * Target users are page handler itself and implementations of
3587 * vm_ops->map_pages.
3589 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3592 struct vm_area_struct *vma = vmf->vma;
3593 bool write = vmf->flags & FAULT_FLAG_WRITE;
3597 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3598 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3600 VM_BUG_ON_PAGE(memcg, page);
3602 ret = do_set_pmd(vmf, page);
3603 if (ret != VM_FAULT_FALLBACK)
3608 ret = pte_alloc_one_map(vmf);
3613 /* Re-check under ptl */
3614 if (unlikely(!pte_none(*vmf->pte)))
3615 return VM_FAULT_NOPAGE;
3617 flush_icache_page(vma, page);
3618 entry = mk_pte(page, vma->vm_page_prot);
3620 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3621 /* copy-on-write page */
3622 if (write && !(vma->vm_flags & VM_SHARED)) {
3623 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3624 page_add_new_anon_rmap(page, vma, vmf->address, false);
3625 mem_cgroup_commit_charge(page, memcg, false, false);
3626 lru_cache_add_active_or_unevictable(page, vma);
3628 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3629 page_add_file_rmap(page, false);
3631 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3633 /* no need to invalidate: a not-present page won't be cached */
3634 update_mmu_cache(vma, vmf->address, vmf->pte);
3641 * finish_fault - finish page fault once we have prepared the page to fault
3643 * @vmf: structure describing the fault
3645 * This function handles all that is needed to finish a page fault once the
3646 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3647 * given page, adds reverse page mapping, handles memcg charges and LRU
3648 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3651 * The function expects the page to be locked and on success it consumes a
3652 * reference of a page being mapped (for the PTE which maps it).
3654 vm_fault_t finish_fault(struct vm_fault *vmf)
3659 /* Did we COW the page? */
3660 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3661 !(vmf->vma->vm_flags & VM_SHARED))
3662 page = vmf->cow_page;
3667 * check even for read faults because we might have lost our CoWed
3670 if (!(vmf->vma->vm_flags & VM_SHARED))
3671 ret = check_stable_address_space(vmf->vma->vm_mm);
3673 ret = alloc_set_pte(vmf, vmf->memcg, page);
3675 pte_unmap_unlock(vmf->pte, vmf->ptl);
3679 static unsigned long fault_around_bytes __read_mostly =
3680 rounddown_pow_of_two(65536);
3682 #ifdef CONFIG_DEBUG_FS
3683 static int fault_around_bytes_get(void *data, u64 *val)
3685 *val = fault_around_bytes;
3690 * fault_around_bytes must be rounded down to the nearest page order as it's
3691 * what do_fault_around() expects to see.
3693 static int fault_around_bytes_set(void *data, u64 val)
3695 if (val / PAGE_SIZE > PTRS_PER_PTE)
3697 if (val > PAGE_SIZE)
3698 fault_around_bytes = rounddown_pow_of_two(val);
3700 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3703 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3704 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3706 static int __init fault_around_debugfs(void)
3710 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3711 &fault_around_bytes_fops);
3713 pr_warn("Failed to create fault_around_bytes in debugfs");
3716 late_initcall(fault_around_debugfs);
3720 * do_fault_around() tries to map few pages around the fault address. The hope
3721 * is that the pages will be needed soon and this will lower the number of
3724 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3725 * not ready to be mapped: not up-to-date, locked, etc.
3727 * This function is called with the page table lock taken. In the split ptlock
3728 * case the page table lock only protects only those entries which belong to
3729 * the page table corresponding to the fault address.
3731 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3734 * fault_around_bytes defines how many bytes we'll try to map.
3735 * do_fault_around() expects it to be set to a power of two less than or equal
3738 * The virtual address of the area that we map is naturally aligned to
3739 * fault_around_bytes rounded down to the machine page size
3740 * (and therefore to page order). This way it's easier to guarantee
3741 * that we don't cross page table boundaries.
3743 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3745 unsigned long address = vmf->address, nr_pages, mask;
3746 pgoff_t start_pgoff = vmf->pgoff;
3751 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3752 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3754 vmf->address = max(address & mask, vmf->vma->vm_start);
3755 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3759 * end_pgoff is either the end of the page table, the end of
3760 * the vma or nr_pages from start_pgoff, depending what is nearest.
3762 end_pgoff = start_pgoff -
3763 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3765 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3766 start_pgoff + nr_pages - 1);
3768 if (pmd_none(*vmf->pmd)) {
3769 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3771 if (!vmf->prealloc_pte)
3773 smp_wmb(); /* See comment in __pte_alloc() */
3776 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3778 /* Huge page is mapped? Page fault is solved */
3779 if (pmd_trans_huge(*vmf->pmd)) {
3780 ret = VM_FAULT_NOPAGE;
3784 /* ->map_pages() haven't done anything useful. Cold page cache? */
3788 /* check if the page fault is solved */
3789 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3790 if (!pte_none(*vmf->pte))
3791 ret = VM_FAULT_NOPAGE;
3792 pte_unmap_unlock(vmf->pte, vmf->ptl);
3794 vmf->address = address;
3799 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3801 struct vm_area_struct *vma = vmf->vma;
3805 * Let's call ->map_pages() first and use ->fault() as fallback
3806 * if page by the offset is not ready to be mapped (cold cache or
3809 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3810 ret = do_fault_around(vmf);
3815 ret = __do_fault(vmf);
3816 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3819 ret |= finish_fault(vmf);
3820 unlock_page(vmf->page);
3821 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3822 put_page(vmf->page);
3826 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3828 struct vm_area_struct *vma = vmf->vma;
3831 if (unlikely(anon_vma_prepare(vma)))
3832 return VM_FAULT_OOM;
3834 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3836 return VM_FAULT_OOM;
3838 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3839 &vmf->memcg, false)) {
3840 put_page(vmf->cow_page);
3841 return VM_FAULT_OOM;
3844 ret = __do_fault(vmf);
3845 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3847 if (ret & VM_FAULT_DONE_COW)
3850 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3851 __SetPageUptodate(vmf->cow_page);
3853 ret |= finish_fault(vmf);
3854 unlock_page(vmf->page);
3855 put_page(vmf->page);
3856 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3860 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3861 put_page(vmf->cow_page);
3865 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3867 struct vm_area_struct *vma = vmf->vma;
3868 vm_fault_t ret, tmp;
3870 ret = __do_fault(vmf);
3871 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3875 * Check if the backing address space wants to know that the page is
3876 * about to become writable
3878 if (vma->vm_ops->page_mkwrite) {
3879 unlock_page(vmf->page);
3880 tmp = do_page_mkwrite(vmf);
3881 if (unlikely(!tmp ||
3882 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3883 put_page(vmf->page);
3888 ret |= finish_fault(vmf);
3889 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3891 unlock_page(vmf->page);
3892 put_page(vmf->page);
3896 fault_dirty_shared_page(vma, vmf->page);
3901 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3902 * but allow concurrent faults).
3903 * The mmap_sem may have been released depending on flags and our
3904 * return value. See filemap_fault() and __lock_page_or_retry().
3905 * If mmap_sem is released, vma may become invalid (for example
3906 * by other thread calling munmap()).
3908 static vm_fault_t do_fault(struct vm_fault *vmf)
3910 struct vm_area_struct *vma = vmf->vma;
3911 struct mm_struct *vm_mm = vma->vm_mm;
3915 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3917 if (!vma->vm_ops->fault) {
3919 * If we find a migration pmd entry or a none pmd entry, which
3920 * should never happen, return SIGBUS
3922 if (unlikely(!pmd_present(*vmf->pmd)))
3923 ret = VM_FAULT_SIGBUS;
3925 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3930 * Make sure this is not a temporary clearing of pte
3931 * by holding ptl and checking again. A R/M/W update
3932 * of pte involves: take ptl, clearing the pte so that
3933 * we don't have concurrent modification by hardware
3934 * followed by an update.
3936 if (unlikely(pte_none(*vmf->pte)))
3937 ret = VM_FAULT_SIGBUS;
3939 ret = VM_FAULT_NOPAGE;
3941 pte_unmap_unlock(vmf->pte, vmf->ptl);
3943 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3944 ret = do_read_fault(vmf);
3945 else if (!(vma->vm_flags & VM_SHARED))
3946 ret = do_cow_fault(vmf);
3948 ret = do_shared_fault(vmf);
3950 /* preallocated pagetable is unused: free it */
3951 if (vmf->prealloc_pte) {
3952 pte_free(vm_mm, vmf->prealloc_pte);
3953 vmf->prealloc_pte = NULL;
3958 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3959 unsigned long addr, int page_nid,
3964 count_vm_numa_event(NUMA_HINT_FAULTS);
3965 if (page_nid == numa_node_id()) {
3966 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3967 *flags |= TNF_FAULT_LOCAL;
3970 return mpol_misplaced(page, vma, addr);
3973 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3975 struct vm_area_struct *vma = vmf->vma;
3976 struct page *page = NULL;
3980 bool migrated = false;
3982 bool was_writable = pte_savedwrite(vmf->orig_pte);
3986 * The "pte" at this point cannot be used safely without
3987 * validation through pte_unmap_same(). It's of NUMA type but
3988 * the pfn may be screwed if the read is non atomic.
3990 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3991 spin_lock(vmf->ptl);
3992 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3993 pte_unmap_unlock(vmf->pte, vmf->ptl);
3998 * Make it present again, Depending on how arch implementes non
3999 * accessible ptes, some can allow access by kernel mode.
4001 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
4002 pte = pte_modify(pte, vma->vm_page_prot);
4003 pte = pte_mkyoung(pte);
4005 pte = pte_mkwrite(pte);
4006 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
4007 update_mmu_cache(vma, vmf->address, vmf->pte);
4009 page = vm_normal_page(vma, vmf->address, pte);
4011 pte_unmap_unlock(vmf->pte, vmf->ptl);
4015 /* TODO: handle PTE-mapped THP */
4016 if (PageCompound(page)) {
4017 pte_unmap_unlock(vmf->pte, vmf->ptl);
4022 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4023 * much anyway since they can be in shared cache state. This misses
4024 * the case where a mapping is writable but the process never writes
4025 * to it but pte_write gets cleared during protection updates and
4026 * pte_dirty has unpredictable behaviour between PTE scan updates,
4027 * background writeback, dirty balancing and application behaviour.
4029 if (!pte_write(pte))
4030 flags |= TNF_NO_GROUP;
4033 * Flag if the page is shared between multiple address spaces. This
4034 * is later used when determining whether to group tasks together
4036 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4037 flags |= TNF_SHARED;
4039 last_cpupid = page_cpupid_last(page);
4040 page_nid = page_to_nid(page);
4041 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4043 pte_unmap_unlock(vmf->pte, vmf->ptl);
4044 if (target_nid == -1) {
4049 /* Migrate to the requested node */
4050 migrated = migrate_misplaced_page(page, vma, target_nid);
4052 page_nid = target_nid;
4053 flags |= TNF_MIGRATED;
4055 flags |= TNF_MIGRATE_FAIL;
4059 task_numa_fault(last_cpupid, page_nid, 1, flags);
4063 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4065 if (vma_is_anonymous(vmf->vma))
4066 return do_huge_pmd_anonymous_page(vmf);
4067 if (vmf->vma->vm_ops->huge_fault)
4068 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4069 return VM_FAULT_FALLBACK;
4072 /* `inline' is required to avoid gcc 4.1.2 build error */
4073 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4075 if (vma_is_anonymous(vmf->vma))
4076 return do_huge_pmd_wp_page(vmf, orig_pmd);
4077 if (vmf->vma->vm_ops->huge_fault)
4078 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4080 /* COW handled on pte level: split pmd */
4081 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
4082 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4084 return VM_FAULT_FALLBACK;
4087 static inline bool vma_is_accessible(struct vm_area_struct *vma)
4089 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
4092 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4094 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4095 /* No support for anonymous transparent PUD pages yet */
4096 if (vma_is_anonymous(vmf->vma))
4097 return VM_FAULT_FALLBACK;
4098 if (vmf->vma->vm_ops->huge_fault)
4099 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4100 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4101 return VM_FAULT_FALLBACK;
4104 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4106 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4107 /* No support for anonymous transparent PUD pages yet */
4108 if (vma_is_anonymous(vmf->vma))
4109 return VM_FAULT_FALLBACK;
4110 if (vmf->vma->vm_ops->huge_fault)
4111 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4112 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4113 return VM_FAULT_FALLBACK;
4117 * These routines also need to handle stuff like marking pages dirty
4118 * and/or accessed for architectures that don't do it in hardware (most
4119 * RISC architectures). The early dirtying is also good on the i386.
4121 * There is also a hook called "update_mmu_cache()" that architectures
4122 * with external mmu caches can use to update those (ie the Sparc or
4123 * PowerPC hashed page tables that act as extended TLBs).
4125 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
4126 * concurrent faults).
4128 * The mmap_sem may have been released depending on flags and our return value.
4129 * See filemap_fault() and __lock_page_or_retry().
4131 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4135 if (unlikely(pmd_none(*vmf->pmd))) {
4137 * Leave __pte_alloc() until later: because vm_ops->fault may
4138 * want to allocate huge page, and if we expose page table
4139 * for an instant, it will be difficult to retract from
4140 * concurrent faults and from rmap lookups.
4144 /* See comment in pte_alloc_one_map() */
4145 if (pmd_devmap_trans_unstable(vmf->pmd))
4148 * A regular pmd is established and it can't morph into a huge
4149 * pmd from under us anymore at this point because we hold the
4150 * mmap_sem read mode and khugepaged takes it in write mode.
4151 * So now it's safe to run pte_offset_map().
4153 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4154 vmf->orig_pte = *vmf->pte;
4157 * some architectures can have larger ptes than wordsize,
4158 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4159 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4160 * accesses. The code below just needs a consistent view
4161 * for the ifs and we later double check anyway with the
4162 * ptl lock held. So here a barrier will do.
4165 if (pte_none(vmf->orig_pte)) {
4166 pte_unmap(vmf->pte);
4172 if (vma_is_anonymous(vmf->vma))
4173 return do_anonymous_page(vmf);
4175 return do_fault(vmf);
4178 if (!pte_present(vmf->orig_pte))
4179 return do_swap_page(vmf);
4181 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4182 return do_numa_page(vmf);
4184 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4185 spin_lock(vmf->ptl);
4186 entry = vmf->orig_pte;
4187 if (unlikely(!pte_same(*vmf->pte, entry)))
4189 if (vmf->flags & FAULT_FLAG_WRITE) {
4190 if (!pte_write(entry))
4191 return do_wp_page(vmf);
4192 entry = pte_mkdirty(entry);
4194 entry = pte_mkyoung(entry);
4195 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4196 vmf->flags & FAULT_FLAG_WRITE)) {
4197 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4200 * This is needed only for protection faults but the arch code
4201 * is not yet telling us if this is a protection fault or not.
4202 * This still avoids useless tlb flushes for .text page faults
4205 if (vmf->flags & FAULT_FLAG_WRITE)
4206 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4209 pte_unmap_unlock(vmf->pte, vmf->ptl);
4214 * By the time we get here, we already hold the mm semaphore
4216 * The mmap_sem may have been released depending on flags and our
4217 * return value. See filemap_fault() and __lock_page_or_retry().
4219 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4220 unsigned long address, unsigned int flags)
4222 struct vm_fault vmf = {
4224 .address = address & PAGE_MASK,
4226 .pgoff = linear_page_index(vma, address),
4227 .gfp_mask = __get_fault_gfp_mask(vma),
4229 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4230 struct mm_struct *mm = vma->vm_mm;
4235 pgd = pgd_offset(mm, address);
4236 p4d = p4d_alloc(mm, pgd, address);
4238 return VM_FAULT_OOM;
4240 vmf.pud = pud_alloc(mm, p4d, address);
4242 return VM_FAULT_OOM;
4243 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4244 ret = create_huge_pud(&vmf);
4245 if (!(ret & VM_FAULT_FALLBACK))
4248 pud_t orig_pud = *vmf.pud;
4251 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4253 /* NUMA case for anonymous PUDs would go here */
4255 if (dirty && !pud_write(orig_pud)) {
4256 ret = wp_huge_pud(&vmf, orig_pud);
4257 if (!(ret & VM_FAULT_FALLBACK))
4260 huge_pud_set_accessed(&vmf, orig_pud);
4266 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4268 return VM_FAULT_OOM;
4269 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4270 ret = create_huge_pmd(&vmf);
4271 if (!(ret & VM_FAULT_FALLBACK))
4274 pmd_t orig_pmd = *vmf.pmd;
4277 if (unlikely(is_swap_pmd(orig_pmd))) {
4278 VM_BUG_ON(thp_migration_supported() &&
4279 !is_pmd_migration_entry(orig_pmd));
4280 if (is_pmd_migration_entry(orig_pmd))
4281 pmd_migration_entry_wait(mm, vmf.pmd);
4284 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4285 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4286 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4288 if (dirty && !pmd_write(orig_pmd)) {
4289 ret = wp_huge_pmd(&vmf, orig_pmd);
4290 if (!(ret & VM_FAULT_FALLBACK))
4293 huge_pmd_set_accessed(&vmf, orig_pmd);
4299 return handle_pte_fault(&vmf);
4303 * By the time we get here, we already hold the mm semaphore
4305 * The mmap_sem may have been released depending on flags and our
4306 * return value. See filemap_fault() and __lock_page_or_retry().
4308 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4313 __set_current_state(TASK_RUNNING);
4315 count_vm_event(PGFAULT);
4316 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4318 /* do counter updates before entering really critical section. */
4319 check_sync_rss_stat(current);
4321 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4322 flags & FAULT_FLAG_INSTRUCTION,
4323 flags & FAULT_FLAG_REMOTE))
4324 return VM_FAULT_SIGSEGV;
4327 * Enable the memcg OOM handling for faults triggered in user
4328 * space. Kernel faults are handled more gracefully.
4330 if (flags & FAULT_FLAG_USER)
4331 mem_cgroup_enter_user_fault();
4333 if (unlikely(is_vm_hugetlb_page(vma)))
4334 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4336 ret = __handle_mm_fault(vma, address, flags);
4338 if (flags & FAULT_FLAG_USER) {
4339 mem_cgroup_exit_user_fault();
4341 * The task may have entered a memcg OOM situation but
4342 * if the allocation error was handled gracefully (no
4343 * VM_FAULT_OOM), there is no need to kill anything.
4344 * Just clean up the OOM state peacefully.
4346 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4347 mem_cgroup_oom_synchronize(false);
4352 EXPORT_SYMBOL_GPL(handle_mm_fault);
4354 #ifndef __PAGETABLE_P4D_FOLDED
4356 * Allocate p4d page table.
4357 * We've already handled the fast-path in-line.
4359 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4361 p4d_t *new = p4d_alloc_one(mm, address);
4365 smp_wmb(); /* See comment in __pte_alloc */
4367 spin_lock(&mm->page_table_lock);
4368 if (pgd_present(*pgd)) /* Another has populated it */
4371 pgd_populate(mm, pgd, new);
4372 spin_unlock(&mm->page_table_lock);
4375 #endif /* __PAGETABLE_P4D_FOLDED */
4377 #ifndef __PAGETABLE_PUD_FOLDED
4379 * Allocate page upper directory.
4380 * We've already handled the fast-path in-line.
4382 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4384 pud_t *new = pud_alloc_one(mm, address);
4388 smp_wmb(); /* See comment in __pte_alloc */
4390 spin_lock(&mm->page_table_lock);
4391 #ifndef __ARCH_HAS_5LEVEL_HACK
4392 if (!p4d_present(*p4d)) {
4394 p4d_populate(mm, p4d, new);
4395 } else /* Another has populated it */
4398 if (!pgd_present(*p4d)) {
4400 pgd_populate(mm, p4d, new);
4401 } else /* Another has populated it */
4403 #endif /* __ARCH_HAS_5LEVEL_HACK */
4404 spin_unlock(&mm->page_table_lock);
4407 #endif /* __PAGETABLE_PUD_FOLDED */
4409 #ifndef __PAGETABLE_PMD_FOLDED
4411 * Allocate page middle directory.
4412 * We've already handled the fast-path in-line.
4414 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4417 pmd_t *new = pmd_alloc_one(mm, address);
4421 smp_wmb(); /* See comment in __pte_alloc */
4423 ptl = pud_lock(mm, pud);
4424 #ifndef __ARCH_HAS_4LEVEL_HACK
4425 if (!pud_present(*pud)) {
4427 pud_populate(mm, pud, new);
4428 } else /* Another has populated it */
4431 if (!pgd_present(*pud)) {
4433 pgd_populate(mm, pud, new);
4434 } else /* Another has populated it */
4436 #endif /* __ARCH_HAS_4LEVEL_HACK */
4440 #endif /* __PAGETABLE_PMD_FOLDED */
4442 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4443 unsigned long *start, unsigned long *end,
4444 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4452 pgd = pgd_offset(mm, address);
4453 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4456 p4d = p4d_offset(pgd, address);
4457 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4460 pud = pud_offset(p4d, address);
4461 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4464 pmd = pmd_offset(pud, address);
4465 VM_BUG_ON(pmd_trans_huge(*pmd));
4467 if (pmd_huge(*pmd)) {
4472 *start = address & PMD_MASK;
4473 *end = *start + PMD_SIZE;
4474 mmu_notifier_invalidate_range_start(mm, *start, *end);
4476 *ptlp = pmd_lock(mm, pmd);
4477 if (pmd_huge(*pmd)) {
4483 mmu_notifier_invalidate_range_end(mm, *start, *end);
4486 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4490 *start = address & PAGE_MASK;
4491 *end = *start + PAGE_SIZE;
4492 mmu_notifier_invalidate_range_start(mm, *start, *end);
4494 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4495 if (!pte_present(*ptep))
4500 pte_unmap_unlock(ptep, *ptlp);
4502 mmu_notifier_invalidate_range_end(mm, *start, *end);
4507 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4508 pte_t **ptepp, spinlock_t **ptlp)
4512 /* (void) is needed to make gcc happy */
4513 (void) __cond_lock(*ptlp,
4514 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4515 ptepp, NULL, ptlp)));
4519 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4520 unsigned long *start, unsigned long *end,
4521 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4525 /* (void) is needed to make gcc happy */
4526 (void) __cond_lock(*ptlp,
4527 !(res = __follow_pte_pmd(mm, address, start, end,
4528 ptepp, pmdpp, ptlp)));
4531 EXPORT_SYMBOL(follow_pte_pmd);
4534 * follow_pfn - look up PFN at a user virtual address
4535 * @vma: memory mapping
4536 * @address: user virtual address
4537 * @pfn: location to store found PFN
4539 * Only IO mappings and raw PFN mappings are allowed.
4541 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4543 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4550 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4553 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4556 *pfn = pte_pfn(*ptep);
4557 pte_unmap_unlock(ptep, ptl);
4560 EXPORT_SYMBOL(follow_pfn);
4562 #ifdef CONFIG_HAVE_IOREMAP_PROT
4563 int follow_phys(struct vm_area_struct *vma,
4564 unsigned long address, unsigned int flags,
4565 unsigned long *prot, resource_size_t *phys)
4571 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4574 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4578 if ((flags & FOLL_WRITE) && !pte_write(pte))
4581 *prot = pgprot_val(pte_pgprot(pte));
4582 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4586 pte_unmap_unlock(ptep, ptl);
4591 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4592 void *buf, int len, int write)
4594 resource_size_t phys_addr;
4595 unsigned long prot = 0;
4596 void __iomem *maddr;
4597 int offset = addr & (PAGE_SIZE-1);
4599 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4602 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4607 memcpy_toio(maddr + offset, buf, len);
4609 memcpy_fromio(buf, maddr + offset, len);
4614 EXPORT_SYMBOL_GPL(generic_access_phys);
4618 * Access another process' address space as given in mm. If non-NULL, use the
4619 * given task for page fault accounting.
4621 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4622 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4624 struct vm_area_struct *vma;
4625 void *old_buf = buf;
4626 int write = gup_flags & FOLL_WRITE;
4628 if (down_read_killable(&mm->mmap_sem))
4631 /* ignore errors, just check how much was successfully transferred */
4633 int bytes, ret, offset;
4635 struct page *page = NULL;
4637 ret = get_user_pages_remote(tsk, mm, addr, 1,
4638 gup_flags, &page, &vma, NULL);
4640 #ifndef CONFIG_HAVE_IOREMAP_PROT
4644 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4645 * we can access using slightly different code.
4647 vma = find_vma(mm, addr);
4648 if (!vma || vma->vm_start > addr)
4650 if (vma->vm_ops && vma->vm_ops->access)
4651 ret = vma->vm_ops->access(vma, addr, buf,
4659 offset = addr & (PAGE_SIZE-1);
4660 if (bytes > PAGE_SIZE-offset)
4661 bytes = PAGE_SIZE-offset;
4665 copy_to_user_page(vma, page, addr,
4666 maddr + offset, buf, bytes);
4667 set_page_dirty_lock(page);
4669 copy_from_user_page(vma, page, addr,
4670 buf, maddr + offset, bytes);
4679 up_read(&mm->mmap_sem);
4681 return buf - old_buf;
4685 * access_remote_vm - access another process' address space
4686 * @mm: the mm_struct of the target address space
4687 * @addr: start address to access
4688 * @buf: source or destination buffer
4689 * @len: number of bytes to transfer
4690 * @gup_flags: flags modifying lookup behaviour
4692 * The caller must hold a reference on @mm.
4694 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4695 void *buf, int len, unsigned int gup_flags)
4697 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4701 * Access another process' address space.
4702 * Source/target buffer must be kernel space,
4703 * Do not walk the page table directly, use get_user_pages
4705 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4706 void *buf, int len, unsigned int gup_flags)
4708 struct mm_struct *mm;
4711 mm = get_task_mm(tsk);
4715 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4721 EXPORT_SYMBOL_GPL(access_process_vm);
4724 * Print the name of a VMA.
4726 void print_vma_addr(char *prefix, unsigned long ip)
4728 struct mm_struct *mm = current->mm;
4729 struct vm_area_struct *vma;
4732 * we might be running from an atomic context so we cannot sleep
4734 if (!down_read_trylock(&mm->mmap_sem))
4737 vma = find_vma(mm, ip);
4738 if (vma && vma->vm_file) {
4739 struct file *f = vma->vm_file;
4740 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4744 p = file_path(f, buf, PAGE_SIZE);
4747 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4749 vma->vm_end - vma->vm_start);
4750 free_page((unsigned long)buf);
4753 up_read(&mm->mmap_sem);
4756 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4757 void __might_fault(const char *file, int line)
4760 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4761 * holding the mmap_sem, this is safe because kernel memory doesn't
4762 * get paged out, therefore we'll never actually fault, and the
4763 * below annotations will generate false positives.
4765 if (uaccess_kernel())
4767 if (pagefault_disabled())
4769 __might_sleep(file, line, 0);
4770 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4772 might_lock_read(¤t->mm->mmap_sem);
4775 EXPORT_SYMBOL(__might_fault);
4778 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4780 * Process all subpages of the specified huge page with the specified
4781 * operation. The target subpage will be processed last to keep its
4784 static inline void process_huge_page(
4785 unsigned long addr_hint, unsigned int pages_per_huge_page,
4786 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4790 unsigned long addr = addr_hint &
4791 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4793 /* Process target subpage last to keep its cache lines hot */
4795 n = (addr_hint - addr) / PAGE_SIZE;
4796 if (2 * n <= pages_per_huge_page) {
4797 /* If target subpage in first half of huge page */
4800 /* Process subpages at the end of huge page */
4801 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4803 process_subpage(addr + i * PAGE_SIZE, i, arg);
4806 /* If target subpage in second half of huge page */
4807 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4808 l = pages_per_huge_page - n;
4809 /* Process subpages at the begin of huge page */
4810 for (i = 0; i < base; i++) {
4812 process_subpage(addr + i * PAGE_SIZE, i, arg);
4816 * Process remaining subpages in left-right-left-right pattern
4817 * towards the target subpage
4819 for (i = 0; i < l; i++) {
4820 int left_idx = base + i;
4821 int right_idx = base + 2 * l - 1 - i;
4824 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4826 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4830 static void clear_gigantic_page(struct page *page,
4832 unsigned int pages_per_huge_page)
4835 struct page *p = page;
4838 for (i = 0; i < pages_per_huge_page;
4839 i++, p = mem_map_next(p, page, i)) {
4841 clear_user_highpage(p, addr + i * PAGE_SIZE);
4845 static void clear_subpage(unsigned long addr, int idx, void *arg)
4847 struct page *page = arg;
4849 clear_user_highpage(page + idx, addr);
4852 void clear_huge_page(struct page *page,
4853 unsigned long addr_hint, unsigned int pages_per_huge_page)
4855 unsigned long addr = addr_hint &
4856 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4858 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4859 clear_gigantic_page(page, addr, pages_per_huge_page);
4863 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4866 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4868 struct vm_area_struct *vma,
4869 unsigned int pages_per_huge_page)
4872 struct page *dst_base = dst;
4873 struct page *src_base = src;
4875 for (i = 0; i < pages_per_huge_page; ) {
4877 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4880 dst = mem_map_next(dst, dst_base, i);
4881 src = mem_map_next(src, src_base, i);
4885 struct copy_subpage_arg {
4888 struct vm_area_struct *vma;
4891 static void copy_subpage(unsigned long addr, int idx, void *arg)
4893 struct copy_subpage_arg *copy_arg = arg;
4895 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4896 addr, copy_arg->vma);
4899 void copy_user_huge_page(struct page *dst, struct page *src,
4900 unsigned long addr_hint, struct vm_area_struct *vma,
4901 unsigned int pages_per_huge_page)
4903 unsigned long addr = addr_hint &
4904 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4905 struct copy_subpage_arg arg = {
4911 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4912 copy_user_gigantic_page(dst, src, addr, vma,
4913 pages_per_huge_page);
4917 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4920 long copy_huge_page_from_user(struct page *dst_page,
4921 const void __user *usr_src,
4922 unsigned int pages_per_huge_page,
4923 bool allow_pagefault)
4925 void *src = (void *)usr_src;
4927 unsigned long i, rc = 0;
4928 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4929 struct page *subpage = dst_page;
4931 for (i = 0; i < pages_per_huge_page;
4932 i++, subpage = mem_map_next(subpage, dst_page, i)) {
4933 if (allow_pagefault)
4934 page_kaddr = kmap(subpage);
4936 page_kaddr = kmap_atomic(subpage);
4937 rc = copy_from_user(page_kaddr,
4938 (const void __user *)(src + i * PAGE_SIZE),
4940 if (allow_pagefault)
4943 kunmap_atomic(page_kaddr);
4945 ret_val -= (PAGE_SIZE - rc);
4953 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4955 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4957 static struct kmem_cache *page_ptl_cachep;
4959 void __init ptlock_cache_init(void)
4961 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4965 bool ptlock_alloc(struct page *page)
4969 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4976 void ptlock_free(struct page *page)
4978 kmem_cache_free(page_ptl_cachep, page->ptl);