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 void tlb_flush_pmd_range(struct mmu_gather *tlb, unsigned long address,
330 if (tlb->page_size != 0 && tlb->page_size != PMD_SIZE)
333 tlb->page_size = PMD_SIZE;
334 tlb->start = min(tlb->start, address);
335 tlb->end = max(tlb->end, address + size);
337 #endif /* HAVE_GENERIC_MMU_GATHER */
339 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
342 * See the comment near struct mmu_table_batch.
346 * If we want tlb_remove_table() to imply TLB invalidates.
348 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
350 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
352 * Invalidate page-table caches used by hardware walkers. Then we still
353 * need to RCU-sched wait while freeing the pages because software
354 * walkers can still be in-flight.
356 tlb_flush_mmu_tlbonly(tlb);
360 static void tlb_remove_table_smp_sync(void *arg)
362 /* Simply deliver the interrupt */
365 static void tlb_remove_table_one(void *table)
368 * This isn't an RCU grace period and hence the page-tables cannot be
369 * assumed to be actually RCU-freed.
371 * It is however sufficient for software page-table walkers that rely on
372 * IRQ disabling. See the comment near struct mmu_table_batch.
374 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
375 __tlb_remove_table(table);
378 static void tlb_remove_table_rcu(struct rcu_head *head)
380 struct mmu_table_batch *batch;
383 batch = container_of(head, struct mmu_table_batch, rcu);
385 for (i = 0; i < batch->nr; i++)
386 __tlb_remove_table(batch->tables[i]);
388 free_page((unsigned long)batch);
391 void tlb_table_flush(struct mmu_gather *tlb)
393 struct mmu_table_batch **batch = &tlb->batch;
396 tlb_table_invalidate(tlb);
397 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
402 void tlb_remove_table(struct mmu_gather *tlb, void *table)
404 struct mmu_table_batch **batch = &tlb->batch;
406 if (*batch == NULL) {
407 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
408 if (*batch == NULL) {
409 tlb_table_invalidate(tlb);
410 tlb_remove_table_one(table);
416 (*batch)->tables[(*batch)->nr++] = table;
417 if ((*batch)->nr == MAX_TABLE_BATCH)
418 tlb_table_flush(tlb);
421 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
424 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
425 * @tlb: the mmu_gather structure to initialize
426 * @mm: the mm_struct of the target address space
427 * @start: start of the region that will be removed from the page-table
428 * @end: end of the region that will be removed from the page-table
430 * Called to initialize an (on-stack) mmu_gather structure for page-table
431 * tear-down from @mm. The @start and @end are set to 0 and -1
432 * respectively when @mm is without users and we're going to destroy
433 * the full address space (exit/execve).
435 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
436 unsigned long start, unsigned long end)
438 arch_tlb_gather_mmu(tlb, mm, start, end);
439 inc_tlb_flush_pending(tlb->mm);
442 void tlb_finish_mmu(struct mmu_gather *tlb,
443 unsigned long start, unsigned long end)
446 * If there are parallel threads are doing PTE changes on same range
447 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
448 * flush by batching, a thread has stable TLB entry can fail to flush
449 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
450 * forcefully if we detect parallel PTE batching threads.
452 bool force = mm_tlb_flush_nested(tlb->mm);
454 arch_tlb_finish_mmu(tlb, start, end, force);
455 dec_tlb_flush_pending(tlb->mm);
459 * Note: this doesn't free the actual pages themselves. That
460 * has been handled earlier when unmapping all the memory regions.
462 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
465 pgtable_t token = pmd_pgtable(*pmd);
467 pte_free_tlb(tlb, token, addr);
468 mm_dec_nr_ptes(tlb->mm);
471 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
472 unsigned long addr, unsigned long end,
473 unsigned long floor, unsigned long ceiling)
480 pmd = pmd_offset(pud, addr);
482 next = pmd_addr_end(addr, end);
483 if (pmd_none_or_clear_bad(pmd))
485 free_pte_range(tlb, pmd, addr);
486 } while (pmd++, addr = next, addr != end);
496 if (end - 1 > ceiling - 1)
499 pmd = pmd_offset(pud, start);
501 pmd_free_tlb(tlb, pmd, start);
502 mm_dec_nr_pmds(tlb->mm);
505 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
506 unsigned long addr, unsigned long end,
507 unsigned long floor, unsigned long ceiling)
514 pud = pud_offset(p4d, addr);
516 next = pud_addr_end(addr, end);
517 if (pud_none_or_clear_bad(pud))
519 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
520 } while (pud++, addr = next, addr != end);
530 if (end - 1 > ceiling - 1)
533 pud = pud_offset(p4d, start);
535 pud_free_tlb(tlb, pud, start);
536 mm_dec_nr_puds(tlb->mm);
539 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
540 unsigned long addr, unsigned long end,
541 unsigned long floor, unsigned long ceiling)
548 p4d = p4d_offset(pgd, addr);
550 next = p4d_addr_end(addr, end);
551 if (p4d_none_or_clear_bad(p4d))
553 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
554 } while (p4d++, addr = next, addr != end);
560 ceiling &= PGDIR_MASK;
564 if (end - 1 > ceiling - 1)
567 p4d = p4d_offset(pgd, start);
569 p4d_free_tlb(tlb, p4d, start);
573 * This function frees user-level page tables of a process.
575 void free_pgd_range(struct mmu_gather *tlb,
576 unsigned long addr, unsigned long end,
577 unsigned long floor, unsigned long ceiling)
583 * The next few lines have given us lots of grief...
585 * Why are we testing PMD* at this top level? Because often
586 * there will be no work to do at all, and we'd prefer not to
587 * go all the way down to the bottom just to discover that.
589 * Why all these "- 1"s? Because 0 represents both the bottom
590 * of the address space and the top of it (using -1 for the
591 * top wouldn't help much: the masks would do the wrong thing).
592 * The rule is that addr 0 and floor 0 refer to the bottom of
593 * the address space, but end 0 and ceiling 0 refer to the top
594 * Comparisons need to use "end - 1" and "ceiling - 1" (though
595 * that end 0 case should be mythical).
597 * Wherever addr is brought up or ceiling brought down, we must
598 * be careful to reject "the opposite 0" before it confuses the
599 * subsequent tests. But what about where end is brought down
600 * by PMD_SIZE below? no, end can't go down to 0 there.
602 * Whereas we round start (addr) and ceiling down, by different
603 * masks at different levels, in order to test whether a table
604 * now has no other vmas using it, so can be freed, we don't
605 * bother to round floor or end up - the tests don't need that.
619 if (end - 1 > ceiling - 1)
624 * We add page table cache pages with PAGE_SIZE,
625 * (see pte_free_tlb()), flush the tlb if we need
627 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
628 pgd = pgd_offset(tlb->mm, addr);
630 next = pgd_addr_end(addr, end);
631 if (pgd_none_or_clear_bad(pgd))
633 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
634 } while (pgd++, addr = next, addr != end);
637 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
638 unsigned long floor, unsigned long ceiling)
641 struct vm_area_struct *next = vma->vm_next;
642 unsigned long addr = vma->vm_start;
645 * Hide vma from rmap and truncate_pagecache before freeing
648 unlink_anon_vmas(vma);
649 unlink_file_vma(vma);
651 if (is_vm_hugetlb_page(vma)) {
652 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
653 floor, next ? next->vm_start : ceiling);
656 * Optimization: gather nearby vmas into one call down
658 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
659 && !is_vm_hugetlb_page(next)) {
662 unlink_anon_vmas(vma);
663 unlink_file_vma(vma);
665 free_pgd_range(tlb, addr, vma->vm_end,
666 floor, next ? next->vm_start : ceiling);
672 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
675 pgtable_t new = pte_alloc_one(mm, address);
680 * Ensure all pte setup (eg. pte page lock and page clearing) are
681 * visible before the pte is made visible to other CPUs by being
682 * put into page tables.
684 * The other side of the story is the pointer chasing in the page
685 * table walking code (when walking the page table without locking;
686 * ie. most of the time). Fortunately, these data accesses consist
687 * of a chain of data-dependent loads, meaning most CPUs (alpha
688 * being the notable exception) will already guarantee loads are
689 * seen in-order. See the alpha page table accessors for the
690 * smp_read_barrier_depends() barriers in page table walking code.
692 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
694 ptl = pmd_lock(mm, pmd);
695 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
697 pmd_populate(mm, pmd, new);
706 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
708 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
712 smp_wmb(); /* See comment in __pte_alloc */
714 spin_lock(&init_mm.page_table_lock);
715 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
716 pmd_populate_kernel(&init_mm, pmd, new);
719 spin_unlock(&init_mm.page_table_lock);
721 pte_free_kernel(&init_mm, new);
725 static inline void init_rss_vec(int *rss)
727 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
730 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
734 if (current->mm == mm)
736 for (i = 0; i < NR_MM_COUNTERS; i++)
738 add_mm_counter(mm, i, rss[i]);
742 * This function is called to print an error when a bad pte
743 * is found. For example, we might have a PFN-mapped pte in
744 * a region that doesn't allow it.
746 * The calling function must still handle the error.
748 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
749 pte_t pte, struct page *page)
751 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
752 p4d_t *p4d = p4d_offset(pgd, addr);
753 pud_t *pud = pud_offset(p4d, addr);
754 pmd_t *pmd = pmd_offset(pud, addr);
755 struct address_space *mapping;
757 static unsigned long resume;
758 static unsigned long nr_shown;
759 static unsigned long nr_unshown;
762 * Allow a burst of 60 reports, then keep quiet for that minute;
763 * or allow a steady drip of one report per second.
765 if (nr_shown == 60) {
766 if (time_before(jiffies, resume)) {
771 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
778 resume = jiffies + 60 * HZ;
780 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
781 index = linear_page_index(vma, addr);
783 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
785 (long long)pte_val(pte), (long long)pmd_val(*pmd));
787 dump_page(page, "bad pte");
788 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
789 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
790 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
792 vma->vm_ops ? vma->vm_ops->fault : NULL,
793 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
794 mapping ? mapping->a_ops->readpage : NULL);
796 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
800 * vm_normal_page -- This function gets the "struct page" associated with a pte.
802 * "Special" mappings do not wish to be associated with a "struct page" (either
803 * it doesn't exist, or it exists but they don't want to touch it). In this
804 * case, NULL is returned here. "Normal" mappings do have a struct page.
806 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
807 * pte bit, in which case this function is trivial. Secondly, an architecture
808 * may not have a spare pte bit, which requires a more complicated scheme,
811 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
812 * special mapping (even if there are underlying and valid "struct pages").
813 * COWed pages of a VM_PFNMAP are always normal.
815 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
816 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
817 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
818 * mapping will always honor the rule
820 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
822 * And for normal mappings this is false.
824 * This restricts such mappings to be a linear translation from virtual address
825 * to pfn. To get around this restriction, we allow arbitrary mappings so long
826 * as the vma is not a COW mapping; in that case, we know that all ptes are
827 * special (because none can have been COWed).
830 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
832 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
833 * page" backing, however the difference is that _all_ pages with a struct
834 * page (that is, those where pfn_valid is true) are refcounted and considered
835 * normal pages by the VM. The disadvantage is that pages are refcounted
836 * (which can be slower and simply not an option for some PFNMAP users). The
837 * advantage is that we don't have to follow the strict linearity rule of
838 * PFNMAP mappings in order to support COWable mappings.
841 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
842 pte_t pte, bool with_public_device)
844 unsigned long pfn = pte_pfn(pte);
846 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
847 if (likely(!pte_special(pte)))
849 if (vma->vm_ops && vma->vm_ops->find_special_page)
850 return vma->vm_ops->find_special_page(vma, addr);
851 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
853 if (is_zero_pfn(pfn))
857 * Device public pages are special pages (they are ZONE_DEVICE
858 * pages but different from persistent memory). They behave
859 * allmost like normal pages. The difference is that they are
860 * not on the lru and thus should never be involve with any-
861 * thing that involve lru manipulation (mlock, numa balancing,
864 * This is why we still want to return NULL for such page from
865 * vm_normal_page() so that we do not have to special case all
866 * call site of vm_normal_page().
868 if (likely(pfn <= highest_memmap_pfn)) {
869 struct page *page = pfn_to_page(pfn);
871 if (is_device_public_page(page)) {
872 if (with_public_device)
881 print_bad_pte(vma, addr, pte, NULL);
885 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
887 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
888 if (vma->vm_flags & VM_MIXEDMAP) {
894 off = (addr - vma->vm_start) >> PAGE_SHIFT;
895 if (pfn == vma->vm_pgoff + off)
897 if (!is_cow_mapping(vma->vm_flags))
902 if (is_zero_pfn(pfn))
906 if (unlikely(pfn > highest_memmap_pfn)) {
907 print_bad_pte(vma, addr, pte, NULL);
912 * NOTE! We still have PageReserved() pages in the page tables.
913 * eg. VDSO mappings can cause them to exist.
916 return pfn_to_page(pfn);
919 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
920 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
923 unsigned long pfn = pmd_pfn(pmd);
926 * There is no pmd_special() but there may be special pmds, e.g.
927 * in a direct-access (dax) mapping, so let's just replicate the
928 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
930 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
931 if (vma->vm_flags & VM_MIXEDMAP) {
937 off = (addr - vma->vm_start) >> PAGE_SHIFT;
938 if (pfn == vma->vm_pgoff + off)
940 if (!is_cow_mapping(vma->vm_flags))
947 if (is_zero_pfn(pfn))
949 if (unlikely(pfn > highest_memmap_pfn))
953 * NOTE! We still have PageReserved() pages in the page tables.
954 * eg. VDSO mappings can cause them to exist.
957 return pfn_to_page(pfn);
962 * copy one vm_area from one task to the other. Assumes the page tables
963 * already present in the new task to be cleared in the whole range
964 * covered by this vma.
967 static inline unsigned long
968 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
969 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
970 unsigned long addr, int *rss)
972 unsigned long vm_flags = vma->vm_flags;
973 pte_t pte = *src_pte;
976 /* pte contains position in swap or file, so copy. */
977 if (unlikely(!pte_present(pte))) {
978 swp_entry_t entry = pte_to_swp_entry(pte);
980 if (likely(!non_swap_entry(entry))) {
981 if (swap_duplicate(entry) < 0)
984 /* make sure dst_mm is on swapoff's mmlist. */
985 if (unlikely(list_empty(&dst_mm->mmlist))) {
986 spin_lock(&mmlist_lock);
987 if (list_empty(&dst_mm->mmlist))
988 list_add(&dst_mm->mmlist,
990 spin_unlock(&mmlist_lock);
993 } else if (is_migration_entry(entry)) {
994 page = migration_entry_to_page(entry);
996 rss[mm_counter(page)]++;
998 if (is_write_migration_entry(entry) &&
999 is_cow_mapping(vm_flags)) {
1001 * COW mappings require pages in both
1002 * parent and child to be set to read.
1004 make_migration_entry_read(&entry);
1005 pte = swp_entry_to_pte(entry);
1006 if (pte_swp_soft_dirty(*src_pte))
1007 pte = pte_swp_mksoft_dirty(pte);
1008 set_pte_at(src_mm, addr, src_pte, pte);
1010 } else if (is_device_private_entry(entry)) {
1011 page = device_private_entry_to_page(entry);
1014 * Update rss count even for unaddressable pages, as
1015 * they should treated just like normal pages in this
1018 * We will likely want to have some new rss counters
1019 * for unaddressable pages, at some point. But for now
1020 * keep things as they are.
1023 rss[mm_counter(page)]++;
1024 page_dup_rmap(page, false);
1027 * We do not preserve soft-dirty information, because so
1028 * far, checkpoint/restore is the only feature that
1029 * requires that. And checkpoint/restore does not work
1030 * when a device driver is involved (you cannot easily
1031 * save and restore device driver state).
1033 if (is_write_device_private_entry(entry) &&
1034 is_cow_mapping(vm_flags)) {
1035 make_device_private_entry_read(&entry);
1036 pte = swp_entry_to_pte(entry);
1037 set_pte_at(src_mm, addr, src_pte, pte);
1044 * If it's a COW mapping, write protect it both
1045 * in the parent and the child
1047 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
1048 ptep_set_wrprotect(src_mm, addr, src_pte);
1049 pte = pte_wrprotect(pte);
1053 * If it's a shared mapping, mark it clean in
1056 if (vm_flags & VM_SHARED)
1057 pte = pte_mkclean(pte);
1058 pte = pte_mkold(pte);
1060 page = vm_normal_page(vma, addr, pte);
1063 page_dup_rmap(page, false);
1064 rss[mm_counter(page)]++;
1065 } else if (pte_devmap(pte)) {
1066 page = pte_page(pte);
1069 * Cache coherent device memory behave like regular page and
1070 * not like persistent memory page. For more informations see
1071 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1073 if (is_device_public_page(page)) {
1075 page_dup_rmap(page, false);
1076 rss[mm_counter(page)]++;
1081 set_pte_at(dst_mm, addr, dst_pte, pte);
1085 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1086 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1087 unsigned long addr, unsigned long end)
1089 pte_t *orig_src_pte, *orig_dst_pte;
1090 pte_t *src_pte, *dst_pte;
1091 spinlock_t *src_ptl, *dst_ptl;
1093 int rss[NR_MM_COUNTERS];
1094 swp_entry_t entry = (swp_entry_t){0};
1099 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1102 src_pte = pte_offset_map(src_pmd, addr);
1103 src_ptl = pte_lockptr(src_mm, src_pmd);
1104 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1105 orig_src_pte = src_pte;
1106 orig_dst_pte = dst_pte;
1107 arch_enter_lazy_mmu_mode();
1111 * We are holding two locks at this point - either of them
1112 * could generate latencies in another task on another CPU.
1114 if (progress >= 32) {
1116 if (need_resched() ||
1117 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1120 if (pte_none(*src_pte)) {
1124 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1129 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1131 arch_leave_lazy_mmu_mode();
1132 spin_unlock(src_ptl);
1133 pte_unmap(orig_src_pte);
1134 add_mm_rss_vec(dst_mm, rss);
1135 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1139 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1148 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1149 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1150 unsigned long addr, unsigned long end)
1152 pmd_t *src_pmd, *dst_pmd;
1155 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1158 src_pmd = pmd_offset(src_pud, addr);
1160 next = pmd_addr_end(addr, end);
1161 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1162 || pmd_devmap(*src_pmd)) {
1164 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1165 err = copy_huge_pmd(dst_mm, src_mm,
1166 dst_pmd, src_pmd, addr, vma);
1173 if (pmd_none_or_clear_bad(src_pmd))
1175 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1178 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1182 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1183 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1184 unsigned long addr, unsigned long end)
1186 pud_t *src_pud, *dst_pud;
1189 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1192 src_pud = pud_offset(src_p4d, addr);
1194 next = pud_addr_end(addr, end);
1195 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1198 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1199 err = copy_huge_pud(dst_mm, src_mm,
1200 dst_pud, src_pud, addr, vma);
1207 if (pud_none_or_clear_bad(src_pud))
1209 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1212 } while (dst_pud++, src_pud++, addr = next, addr != end);
1216 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1217 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1218 unsigned long addr, unsigned long end)
1220 p4d_t *src_p4d, *dst_p4d;
1223 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1226 src_p4d = p4d_offset(src_pgd, addr);
1228 next = p4d_addr_end(addr, end);
1229 if (p4d_none_or_clear_bad(src_p4d))
1231 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1234 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1238 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1239 struct vm_area_struct *vma)
1241 pgd_t *src_pgd, *dst_pgd;
1243 unsigned long addr = vma->vm_start;
1244 unsigned long end = vma->vm_end;
1245 unsigned long mmun_start; /* For mmu_notifiers */
1246 unsigned long mmun_end; /* For mmu_notifiers */
1251 * Don't copy ptes where a page fault will fill them correctly.
1252 * Fork becomes much lighter when there are big shared or private
1253 * readonly mappings. The tradeoff is that copy_page_range is more
1254 * efficient than faulting.
1256 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1260 if (is_vm_hugetlb_page(vma))
1261 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1263 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1265 * We do not free on error cases below as remove_vma
1266 * gets called on error from higher level routine
1268 ret = track_pfn_copy(vma);
1274 * We need to invalidate the secondary MMU mappings only when
1275 * there could be a permission downgrade on the ptes of the
1276 * parent mm. And a permission downgrade will only happen if
1277 * is_cow_mapping() returns true.
1279 is_cow = is_cow_mapping(vma->vm_flags);
1283 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1287 dst_pgd = pgd_offset(dst_mm, addr);
1288 src_pgd = pgd_offset(src_mm, addr);
1290 next = pgd_addr_end(addr, end);
1291 if (pgd_none_or_clear_bad(src_pgd))
1293 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1294 vma, addr, next))) {
1298 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1301 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1305 /* Whether we should zap all COWed (private) pages too */
1306 static inline bool should_zap_cows(struct zap_details *details)
1308 /* By default, zap all pages */
1312 /* Or, we zap COWed pages only if the caller wants to */
1313 return !details->check_mapping;
1316 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1317 struct vm_area_struct *vma, pmd_t *pmd,
1318 unsigned long addr, unsigned long end,
1319 struct zap_details *details)
1321 struct mm_struct *mm = tlb->mm;
1322 int force_flush = 0;
1323 int rss[NR_MM_COUNTERS];
1329 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1332 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1334 flush_tlb_batched_pending(mm);
1335 arch_enter_lazy_mmu_mode();
1338 if (pte_none(ptent))
1341 if (pte_present(ptent)) {
1344 page = _vm_normal_page(vma, addr, ptent, true);
1345 if (unlikely(details) && page) {
1347 * unmap_shared_mapping_pages() wants to
1348 * invalidate cache without truncating:
1349 * unmap shared but keep private pages.
1351 if (details->check_mapping &&
1352 details->check_mapping != page_rmapping(page))
1355 ptent = ptep_get_and_clear_full(mm, addr, pte,
1357 tlb_remove_tlb_entry(tlb, pte, addr);
1358 if (unlikely(!page))
1361 if (!PageAnon(page)) {
1362 if (pte_dirty(ptent)) {
1364 set_page_dirty(page);
1366 if (pte_young(ptent) &&
1367 likely(!(vma->vm_flags & VM_SEQ_READ)))
1368 mark_page_accessed(page);
1370 rss[mm_counter(page)]--;
1371 page_remove_rmap(page, false);
1372 if (unlikely(page_mapcount(page) < 0))
1373 print_bad_pte(vma, addr, ptent, page);
1374 if (unlikely(__tlb_remove_page(tlb, page))) {
1382 entry = pte_to_swp_entry(ptent);
1383 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1384 struct page *page = device_private_entry_to_page(entry);
1386 if (unlikely(details && details->check_mapping)) {
1388 * unmap_shared_mapping_pages() wants to
1389 * invalidate cache without truncating:
1390 * unmap shared but keep private pages.
1392 if (details->check_mapping !=
1393 page_rmapping(page))
1397 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1398 rss[mm_counter(page)]--;
1399 page_remove_rmap(page, false);
1404 entry = pte_to_swp_entry(ptent);
1405 if (!non_swap_entry(entry)) {
1406 /* Genuine swap entry, hence a private anon page */
1407 if (!should_zap_cows(details))
1410 } else if (is_migration_entry(entry)) {
1413 page = migration_entry_to_page(entry);
1414 if (details && details->check_mapping &&
1415 details->check_mapping != page_rmapping(page))
1417 rss[mm_counter(page)]--;
1419 if (unlikely(!free_swap_and_cache(entry)))
1420 print_bad_pte(vma, addr, ptent, NULL);
1421 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1422 } while (pte++, addr += PAGE_SIZE, addr != end);
1424 add_mm_rss_vec(mm, rss);
1425 arch_leave_lazy_mmu_mode();
1427 /* Do the actual TLB flush before dropping ptl */
1429 tlb_flush_mmu_tlbonly(tlb);
1430 pte_unmap_unlock(start_pte, ptl);
1433 * If we forced a TLB flush (either due to running out of
1434 * batch buffers or because we needed to flush dirty TLB
1435 * entries before releasing the ptl), free the batched
1436 * memory too. Restart if we didn't do everything.
1440 tlb_flush_mmu_free(tlb);
1448 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1449 struct vm_area_struct *vma, pud_t *pud,
1450 unsigned long addr, unsigned long end,
1451 struct zap_details *details)
1456 pmd = pmd_offset(pud, addr);
1458 next = pmd_addr_end(addr, end);
1459 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1460 if (next - addr != HPAGE_PMD_SIZE)
1461 __split_huge_pmd(vma, pmd, addr, false, NULL);
1462 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1465 } else if (details && details->single_page &&
1466 PageTransCompound(details->single_page) &&
1467 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1468 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1470 * Take and drop THP pmd lock so that we cannot return
1471 * prematurely, while zap_huge_pmd() has cleared *pmd,
1472 * but not yet decremented compound_mapcount().
1478 * Here there can be other concurrent MADV_DONTNEED or
1479 * trans huge page faults running, and if the pmd is
1480 * none or trans huge it can change under us. This is
1481 * because MADV_DONTNEED holds the mmap_sem in read
1484 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1486 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1489 } while (pmd++, addr = next, addr != end);
1494 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1495 struct vm_area_struct *vma, p4d_t *p4d,
1496 unsigned long addr, unsigned long end,
1497 struct zap_details *details)
1502 pud = pud_offset(p4d, addr);
1504 next = pud_addr_end(addr, end);
1505 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1506 if (next - addr != HPAGE_PUD_SIZE) {
1507 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1508 split_huge_pud(vma, pud, addr);
1509 } else if (zap_huge_pud(tlb, vma, pud, addr))
1513 if (pud_none_or_clear_bad(pud))
1515 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1518 } while (pud++, addr = next, addr != end);
1523 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1524 struct vm_area_struct *vma, pgd_t *pgd,
1525 unsigned long addr, unsigned long end,
1526 struct zap_details *details)
1531 p4d = p4d_offset(pgd, addr);
1533 next = p4d_addr_end(addr, end);
1534 if (p4d_none_or_clear_bad(p4d))
1536 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1537 } while (p4d++, addr = next, addr != end);
1542 void unmap_page_range(struct mmu_gather *tlb,
1543 struct vm_area_struct *vma,
1544 unsigned long addr, unsigned long end,
1545 struct zap_details *details)
1550 BUG_ON(addr >= end);
1551 tlb_start_vma(tlb, vma);
1552 pgd = pgd_offset(vma->vm_mm, addr);
1554 next = pgd_addr_end(addr, end);
1555 if (pgd_none_or_clear_bad(pgd))
1557 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1558 } while (pgd++, addr = next, addr != end);
1559 tlb_end_vma(tlb, vma);
1563 static void unmap_single_vma(struct mmu_gather *tlb,
1564 struct vm_area_struct *vma, unsigned long start_addr,
1565 unsigned long end_addr,
1566 struct zap_details *details)
1568 unsigned long start = max(vma->vm_start, start_addr);
1571 if (start >= vma->vm_end)
1573 end = min(vma->vm_end, end_addr);
1574 if (end <= vma->vm_start)
1578 uprobe_munmap(vma, start, end);
1580 if (unlikely(vma->vm_flags & VM_PFNMAP))
1581 untrack_pfn(vma, 0, 0);
1584 if (unlikely(is_vm_hugetlb_page(vma))) {
1586 * It is undesirable to test vma->vm_file as it
1587 * should be non-null for valid hugetlb area.
1588 * However, vm_file will be NULL in the error
1589 * cleanup path of mmap_region. When
1590 * hugetlbfs ->mmap method fails,
1591 * mmap_region() nullifies vma->vm_file
1592 * before calling this function to clean up.
1593 * Since no pte has actually been setup, it is
1594 * safe to do nothing in this case.
1597 i_mmap_lock_write(vma->vm_file->f_mapping);
1598 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1599 i_mmap_unlock_write(vma->vm_file->f_mapping);
1602 unmap_page_range(tlb, vma, start, end, details);
1607 * unmap_vmas - unmap a range of memory covered by a list of vma's
1608 * @tlb: address of the caller's struct mmu_gather
1609 * @vma: the starting vma
1610 * @start_addr: virtual address at which to start unmapping
1611 * @end_addr: virtual address at which to end unmapping
1613 * Unmap all pages in the vma list.
1615 * Only addresses between `start' and `end' will be unmapped.
1617 * The VMA list must be sorted in ascending virtual address order.
1619 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1620 * range after unmap_vmas() returns. So the only responsibility here is to
1621 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1622 * drops the lock and schedules.
1624 void unmap_vmas(struct mmu_gather *tlb,
1625 struct vm_area_struct *vma, unsigned long start_addr,
1626 unsigned long end_addr)
1628 struct mm_struct *mm = vma->vm_mm;
1630 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1631 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1632 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1633 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1637 * zap_page_range - remove user pages in a given range
1638 * @vma: vm_area_struct holding the applicable pages
1639 * @start: starting address of pages to zap
1640 * @size: number of bytes to zap
1642 * Caller must protect the VMA list
1644 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1647 struct mm_struct *mm = vma->vm_mm;
1648 struct mmu_gather tlb;
1649 unsigned long end = start + size;
1652 tlb_gather_mmu(&tlb, mm, start, end);
1653 update_hiwater_rss(mm);
1654 mmu_notifier_invalidate_range_start(mm, start, end);
1655 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1656 unmap_single_vma(&tlb, vma, start, end, NULL);
1657 mmu_notifier_invalidate_range_end(mm, start, end);
1658 tlb_finish_mmu(&tlb, start, end);
1662 * zap_page_range_single - remove user pages in a given range
1663 * @vma: vm_area_struct holding the applicable pages
1664 * @address: starting address of pages to zap
1665 * @size: number of bytes to zap
1666 * @details: details of shared cache invalidation
1668 * The range must fit into one VMA.
1670 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1671 unsigned long size, struct zap_details *details)
1673 struct mm_struct *mm = vma->vm_mm;
1674 struct mmu_gather tlb;
1675 unsigned long end = address + size;
1678 tlb_gather_mmu(&tlb, mm, address, end);
1679 update_hiwater_rss(mm);
1680 mmu_notifier_invalidate_range_start(mm, address, end);
1681 unmap_single_vma(&tlb, vma, address, end, details);
1682 mmu_notifier_invalidate_range_end(mm, address, end);
1683 tlb_finish_mmu(&tlb, address, end);
1687 * zap_vma_ptes - remove ptes mapping the vma
1688 * @vma: vm_area_struct holding ptes to be zapped
1689 * @address: starting address of pages to zap
1690 * @size: number of bytes to zap
1692 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1694 * The entire address range must be fully contained within the vma.
1697 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1700 if (address < vma->vm_start || address + size > vma->vm_end ||
1701 !(vma->vm_flags & VM_PFNMAP))
1704 zap_page_range_single(vma, address, size, NULL);
1706 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1708 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1716 pgd = pgd_offset(mm, addr);
1717 p4d = p4d_alloc(mm, pgd, addr);
1720 pud = pud_alloc(mm, p4d, addr);
1723 pmd = pmd_alloc(mm, pud, addr);
1727 VM_BUG_ON(pmd_trans_huge(*pmd));
1728 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1732 * This is the old fallback for page remapping.
1734 * For historical reasons, it only allows reserved pages. Only
1735 * old drivers should use this, and they needed to mark their
1736 * pages reserved for the old functions anyway.
1738 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1739 struct page *page, pgprot_t prot)
1741 struct mm_struct *mm = vma->vm_mm;
1750 flush_dcache_page(page);
1751 pte = get_locked_pte(mm, addr, &ptl);
1755 if (!pte_none(*pte))
1758 /* Ok, finally just insert the thing.. */
1760 inc_mm_counter_fast(mm, mm_counter_file(page));
1761 page_add_file_rmap(page, false);
1762 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1765 pte_unmap_unlock(pte, ptl);
1768 pte_unmap_unlock(pte, ptl);
1774 * vm_insert_page - insert single page into user vma
1775 * @vma: user vma to map to
1776 * @addr: target user address of this page
1777 * @page: source kernel page
1779 * This allows drivers to insert individual pages they've allocated
1782 * The page has to be a nice clean _individual_ kernel allocation.
1783 * If you allocate a compound page, you need to have marked it as
1784 * such (__GFP_COMP), or manually just split the page up yourself
1785 * (see split_page()).
1787 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1788 * took an arbitrary page protection parameter. This doesn't allow
1789 * that. Your vma protection will have to be set up correctly, which
1790 * means that if you want a shared writable mapping, you'd better
1791 * ask for a shared writable mapping!
1793 * The page does not need to be reserved.
1795 * Usually this function is called from f_op->mmap() handler
1796 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1797 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1798 * function from other places, for example from page-fault handler.
1800 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1803 if (addr < vma->vm_start || addr >= vma->vm_end)
1805 if (!page_count(page))
1807 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1808 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1809 BUG_ON(vma->vm_flags & VM_PFNMAP);
1810 vma->vm_flags |= VM_MIXEDMAP;
1812 return insert_page(vma, addr, page, vma->vm_page_prot);
1814 EXPORT_SYMBOL(vm_insert_page);
1816 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1817 pfn_t pfn, pgprot_t prot, bool mkwrite)
1819 struct mm_struct *mm = vma->vm_mm;
1825 pte = get_locked_pte(mm, addr, &ptl);
1829 if (!pte_none(*pte)) {
1832 * For read faults on private mappings the PFN passed
1833 * in may not match the PFN we have mapped if the
1834 * mapped PFN is a writeable COW page. In the mkwrite
1835 * case we are creating a writable PTE for a shared
1836 * mapping and we expect the PFNs to match. If they
1837 * don't match, we are likely racing with block
1838 * allocation and mapping invalidation so just skip the
1841 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1842 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1845 entry = pte_mkyoung(*pte);
1846 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1847 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1848 update_mmu_cache(vma, addr, pte);
1853 /* Ok, finally just insert the thing.. */
1854 if (pfn_t_devmap(pfn))
1855 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1857 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1860 entry = pte_mkyoung(entry);
1861 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1864 set_pte_at(mm, addr, pte, entry);
1865 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1869 pte_unmap_unlock(pte, ptl);
1875 * vm_insert_pfn - insert single pfn into user vma
1876 * @vma: user vma to map to
1877 * @addr: target user address of this page
1878 * @pfn: source kernel pfn
1880 * Similar to vm_insert_page, this allows drivers to insert individual pages
1881 * they've allocated into a user vma. Same comments apply.
1883 * This function should only be called from a vm_ops->fault handler, and
1884 * in that case the handler should return NULL.
1886 * vma cannot be a COW mapping.
1888 * As this is called only for pages that do not currently exist, we
1889 * do not need to flush old virtual caches or the TLB.
1891 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1894 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1896 EXPORT_SYMBOL(vm_insert_pfn);
1899 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1900 * @vma: user vma to map to
1901 * @addr: target user address of this page
1902 * @pfn: source kernel pfn
1903 * @pgprot: pgprot flags for the inserted page
1905 * This is exactly like vm_insert_pfn, except that it allows drivers to
1906 * to override pgprot on a per-page basis.
1908 * This only makes sense for IO mappings, and it makes no sense for
1909 * cow mappings. In general, using multiple vmas is preferable;
1910 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1913 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1914 unsigned long pfn, pgprot_t pgprot)
1918 * Technically, architectures with pte_special can avoid all these
1919 * restrictions (same for remap_pfn_range). However we would like
1920 * consistency in testing and feature parity among all, so we should
1921 * try to keep these invariants in place for everybody.
1923 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1924 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1925 (VM_PFNMAP|VM_MIXEDMAP));
1926 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1927 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1929 if (addr < vma->vm_start || addr >= vma->vm_end)
1932 if (!pfn_modify_allowed(pfn, pgprot))
1935 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1937 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1942 EXPORT_SYMBOL(vm_insert_pfn_prot);
1944 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1946 /* these checks mirror the abort conditions in vm_normal_page */
1947 if (vma->vm_flags & VM_MIXEDMAP)
1949 if (pfn_t_devmap(pfn))
1951 if (pfn_t_special(pfn))
1953 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1958 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1959 pfn_t pfn, bool mkwrite)
1961 pgprot_t pgprot = vma->vm_page_prot;
1963 BUG_ON(!vm_mixed_ok(vma, pfn));
1965 if (addr < vma->vm_start || addr >= vma->vm_end)
1968 track_pfn_insert(vma, &pgprot, pfn);
1970 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1974 * If we don't have pte special, then we have to use the pfn_valid()
1975 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1976 * refcount the page if pfn_valid is true (hence insert_page rather
1977 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1978 * without pte special, it would there be refcounted as a normal page.
1980 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1981 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1985 * At this point we are committed to insert_page()
1986 * regardless of whether the caller specified flags that
1987 * result in pfn_t_has_page() == false.
1989 page = pfn_to_page(pfn_t_to_pfn(pfn));
1990 return insert_page(vma, addr, page, pgprot);
1992 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1995 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1998 return __vm_insert_mixed(vma, addr, pfn, false);
2001 EXPORT_SYMBOL(vm_insert_mixed);
2004 * If the insertion of PTE failed because someone else already added a
2005 * different entry in the mean time, we treat that as success as we assume
2006 * the same entry was actually inserted.
2009 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2010 unsigned long addr, pfn_t pfn)
2014 err = __vm_insert_mixed(vma, addr, pfn, true);
2016 return VM_FAULT_OOM;
2017 if (err < 0 && err != -EBUSY)
2018 return VM_FAULT_SIGBUS;
2019 return VM_FAULT_NOPAGE;
2021 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2024 * maps a range of physical memory into the requested pages. the old
2025 * mappings are removed. any references to nonexistent pages results
2026 * in null mappings (currently treated as "copy-on-access")
2028 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2029 unsigned long addr, unsigned long end,
2030 unsigned long pfn, pgprot_t prot)
2032 pte_t *pte, *mapped_pte;
2036 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2039 arch_enter_lazy_mmu_mode();
2041 BUG_ON(!pte_none(*pte));
2042 if (!pfn_modify_allowed(pfn, prot)) {
2046 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2048 } while (pte++, addr += PAGE_SIZE, addr != end);
2049 arch_leave_lazy_mmu_mode();
2050 pte_unmap_unlock(mapped_pte, ptl);
2054 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2055 unsigned long addr, unsigned long end,
2056 unsigned long pfn, pgprot_t prot)
2062 pfn -= addr >> PAGE_SHIFT;
2063 pmd = pmd_alloc(mm, pud, addr);
2066 VM_BUG_ON(pmd_trans_huge(*pmd));
2068 next = pmd_addr_end(addr, end);
2069 err = remap_pte_range(mm, pmd, addr, next,
2070 pfn + (addr >> PAGE_SHIFT), prot);
2073 } while (pmd++, addr = next, addr != end);
2077 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2078 unsigned long addr, unsigned long end,
2079 unsigned long pfn, pgprot_t prot)
2085 pfn -= addr >> PAGE_SHIFT;
2086 pud = pud_alloc(mm, p4d, addr);
2090 next = pud_addr_end(addr, end);
2091 err = remap_pmd_range(mm, pud, addr, next,
2092 pfn + (addr >> PAGE_SHIFT), prot);
2095 } while (pud++, addr = next, addr != end);
2099 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2100 unsigned long addr, unsigned long end,
2101 unsigned long pfn, pgprot_t prot)
2107 pfn -= addr >> PAGE_SHIFT;
2108 p4d = p4d_alloc(mm, pgd, addr);
2112 next = p4d_addr_end(addr, end);
2113 err = remap_pud_range(mm, p4d, addr, next,
2114 pfn + (addr >> PAGE_SHIFT), prot);
2117 } while (p4d++, addr = next, addr != end);
2122 * remap_pfn_range - remap kernel memory to userspace
2123 * @vma: user vma to map to
2124 * @addr: target user address to start at
2125 * @pfn: physical address of kernel memory
2126 * @size: size of map area
2127 * @prot: page protection flags for this mapping
2129 * Note: this is only safe if the mm semaphore is held when called.
2131 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2132 unsigned long pfn, unsigned long size, pgprot_t prot)
2136 unsigned long end = addr + PAGE_ALIGN(size);
2137 struct mm_struct *mm = vma->vm_mm;
2138 unsigned long remap_pfn = pfn;
2142 * Physically remapped pages are special. Tell the
2143 * rest of the world about it:
2144 * VM_IO tells people not to look at these pages
2145 * (accesses can have side effects).
2146 * VM_PFNMAP tells the core MM that the base pages are just
2147 * raw PFN mappings, and do not have a "struct page" associated
2150 * Disable vma merging and expanding with mremap().
2152 * Omit vma from core dump, even when VM_IO turned off.
2154 * There's a horrible special case to handle copy-on-write
2155 * behaviour that some programs depend on. We mark the "original"
2156 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2157 * See vm_normal_page() for details.
2159 if (is_cow_mapping(vma->vm_flags)) {
2160 if (addr != vma->vm_start || end != vma->vm_end)
2162 vma->vm_pgoff = pfn;
2165 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2169 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2171 BUG_ON(addr >= end);
2172 pfn -= addr >> PAGE_SHIFT;
2173 pgd = pgd_offset(mm, addr);
2174 flush_cache_range(vma, addr, end);
2176 next = pgd_addr_end(addr, end);
2177 err = remap_p4d_range(mm, pgd, addr, next,
2178 pfn + (addr >> PAGE_SHIFT), prot);
2181 } while (pgd++, addr = next, addr != end);
2184 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2188 EXPORT_SYMBOL(remap_pfn_range);
2191 * vm_iomap_memory - remap memory to userspace
2192 * @vma: user vma to map to
2193 * @start: start of area
2194 * @len: size of area
2196 * This is a simplified io_remap_pfn_range() for common driver use. The
2197 * driver just needs to give us the physical memory range to be mapped,
2198 * we'll figure out the rest from the vma information.
2200 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2201 * whatever write-combining details or similar.
2203 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2205 unsigned long vm_len, pfn, pages;
2207 /* Check that the physical memory area passed in looks valid */
2208 if (start + len < start)
2211 * You *really* shouldn't map things that aren't page-aligned,
2212 * but we've historically allowed it because IO memory might
2213 * just have smaller alignment.
2215 len += start & ~PAGE_MASK;
2216 pfn = start >> PAGE_SHIFT;
2217 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2218 if (pfn + pages < pfn)
2221 /* We start the mapping 'vm_pgoff' pages into the area */
2222 if (vma->vm_pgoff > pages)
2224 pfn += vma->vm_pgoff;
2225 pages -= vma->vm_pgoff;
2227 /* Can we fit all of the mapping? */
2228 vm_len = vma->vm_end - vma->vm_start;
2229 if (vm_len >> PAGE_SHIFT > pages)
2232 /* Ok, let it rip */
2233 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2235 EXPORT_SYMBOL(vm_iomap_memory);
2237 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2238 unsigned long addr, unsigned long end,
2239 pte_fn_t fn, void *data)
2244 spinlock_t *uninitialized_var(ptl);
2246 pte = (mm == &init_mm) ?
2247 pte_alloc_kernel(pmd, addr) :
2248 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2252 BUG_ON(pmd_huge(*pmd));
2254 arch_enter_lazy_mmu_mode();
2256 token = pmd_pgtable(*pmd);
2259 err = fn(pte++, token, addr, data);
2262 } while (addr += PAGE_SIZE, addr != end);
2264 arch_leave_lazy_mmu_mode();
2267 pte_unmap_unlock(pte-1, ptl);
2271 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2272 unsigned long addr, unsigned long end,
2273 pte_fn_t fn, void *data)
2279 BUG_ON(pud_huge(*pud));
2281 pmd = pmd_alloc(mm, pud, addr);
2285 next = pmd_addr_end(addr, end);
2286 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2289 } while (pmd++, addr = next, addr != end);
2293 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2294 unsigned long addr, unsigned long end,
2295 pte_fn_t fn, void *data)
2301 pud = pud_alloc(mm, p4d, addr);
2305 next = pud_addr_end(addr, end);
2306 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2309 } while (pud++, addr = next, addr != end);
2313 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2314 unsigned long addr, unsigned long end,
2315 pte_fn_t fn, void *data)
2321 p4d = p4d_alloc(mm, pgd, addr);
2325 next = p4d_addr_end(addr, end);
2326 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2329 } while (p4d++, addr = next, addr != end);
2334 * Scan a region of virtual memory, filling in page tables as necessary
2335 * and calling a provided function on each leaf page table.
2337 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2338 unsigned long size, pte_fn_t fn, void *data)
2342 unsigned long end = addr + size;
2345 if (WARN_ON(addr >= end))
2348 pgd = pgd_offset(mm, addr);
2350 next = pgd_addr_end(addr, end);
2351 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2354 } while (pgd++, addr = next, addr != end);
2358 EXPORT_SYMBOL_GPL(apply_to_page_range);
2361 * handle_pte_fault chooses page fault handler according to an entry which was
2362 * read non-atomically. Before making any commitment, on those architectures
2363 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2364 * parts, do_swap_page must check under lock before unmapping the pte and
2365 * proceeding (but do_wp_page is only called after already making such a check;
2366 * and do_anonymous_page can safely check later on).
2368 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2369 pte_t *page_table, pte_t orig_pte)
2372 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2373 if (sizeof(pte_t) > sizeof(unsigned long)) {
2374 spinlock_t *ptl = pte_lockptr(mm, pmd);
2376 same = pte_same(*page_table, orig_pte);
2380 pte_unmap(page_table);
2384 static inline bool cow_user_page(struct page *dst, struct page *src,
2385 struct vm_fault *vmf)
2390 bool locked = false;
2391 struct vm_area_struct *vma = vmf->vma;
2392 struct mm_struct *mm = vma->vm_mm;
2393 unsigned long addr = vmf->address;
2395 debug_dma_assert_idle(src);
2398 copy_user_highpage(dst, src, addr, vma);
2403 * If the source page was a PFN mapping, we don't have
2404 * a "struct page" for it. We do a best-effort copy by
2405 * just copying from the original user address. If that
2406 * fails, we just zero-fill it. Live with it.
2408 kaddr = kmap_atomic(dst);
2409 uaddr = (void __user *)(addr & PAGE_MASK);
2412 * On architectures with software "accessed" bits, we would
2413 * take a double page fault, so mark it accessed here.
2415 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2418 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2420 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2422 * Other thread has already handled the fault
2423 * and we don't need to do anything. If it's
2424 * not the case, the fault will be triggered
2425 * again on the same address.
2431 entry = pte_mkyoung(vmf->orig_pte);
2432 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2433 update_mmu_cache(vma, addr, vmf->pte);
2437 * This really shouldn't fail, because the page is there
2438 * in the page tables. But it might just be unreadable,
2439 * in which case we just give up and fill the result with
2442 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2446 /* Re-validate under PTL if the page is still mapped */
2447 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2449 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2450 /* The PTE changed under us. Retry page fault. */
2456 * The same page can be mapped back since last copy attampt.
2457 * Try to copy again under PTL.
2459 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2461 * Give a warn in case there can be some obscure
2474 pte_unmap_unlock(vmf->pte, vmf->ptl);
2475 kunmap_atomic(kaddr);
2476 flush_dcache_page(dst);
2481 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2483 struct file *vm_file = vma->vm_file;
2486 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2489 * Special mappings (e.g. VDSO) do not have any file so fake
2490 * a default GFP_KERNEL for them.
2496 * Notify the address space that the page is about to become writable so that
2497 * it can prohibit this or wait for the page to get into an appropriate state.
2499 * We do this without the lock held, so that it can sleep if it needs to.
2501 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2504 struct page *page = vmf->page;
2505 unsigned int old_flags = vmf->flags;
2507 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2509 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2510 /* Restore original flags so that caller is not surprised */
2511 vmf->flags = old_flags;
2512 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2514 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2516 if (!page->mapping) {
2518 return 0; /* retry */
2520 ret |= VM_FAULT_LOCKED;
2522 VM_BUG_ON_PAGE(!PageLocked(page), page);
2527 * Handle dirtying of a page in shared file mapping on a write fault.
2529 * The function expects the page to be locked and unlocks it.
2531 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2534 struct address_space *mapping;
2536 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2538 dirtied = set_page_dirty(page);
2539 VM_BUG_ON_PAGE(PageAnon(page), page);
2541 * Take a local copy of the address_space - page.mapping may be zeroed
2542 * by truncate after unlock_page(). The address_space itself remains
2543 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2544 * release semantics to prevent the compiler from undoing this copying.
2546 mapping = page_rmapping(page);
2549 if ((dirtied || page_mkwrite) && mapping) {
2551 * Some device drivers do not set page.mapping
2552 * but still dirty their pages
2554 balance_dirty_pages_ratelimited(mapping);
2558 file_update_time(vma->vm_file);
2562 * Handle write page faults for pages that can be reused in the current vma
2564 * This can happen either due to the mapping being with the VM_SHARED flag,
2565 * or due to us being the last reference standing to the page. In either
2566 * case, all we need to do here is to mark the page as writable and update
2567 * any related book-keeping.
2569 static inline void wp_page_reuse(struct vm_fault *vmf)
2570 __releases(vmf->ptl)
2572 struct vm_area_struct *vma = vmf->vma;
2573 struct page *page = vmf->page;
2576 * Clear the pages cpupid information as the existing
2577 * information potentially belongs to a now completely
2578 * unrelated process.
2581 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2583 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2584 entry = pte_mkyoung(vmf->orig_pte);
2585 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2586 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2587 update_mmu_cache(vma, vmf->address, vmf->pte);
2588 pte_unmap_unlock(vmf->pte, vmf->ptl);
2592 * Handle the case of a page which we actually need to copy to a new page.
2594 * Called with mmap_sem locked and the old page referenced, but
2595 * without the ptl held.
2597 * High level logic flow:
2599 * - Allocate a page, copy the content of the old page to the new one.
2600 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2601 * - Take the PTL. If the pte changed, bail out and release the allocated page
2602 * - If the pte is still the way we remember it, update the page table and all
2603 * relevant references. This includes dropping the reference the page-table
2604 * held to the old page, as well as updating the rmap.
2605 * - In any case, unlock the PTL and drop the reference we took to the old page.
2607 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2609 struct vm_area_struct *vma = vmf->vma;
2610 struct mm_struct *mm = vma->vm_mm;
2611 struct page *old_page = vmf->page;
2612 struct page *new_page = NULL;
2614 int page_copied = 0;
2615 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2616 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2617 struct mem_cgroup *memcg;
2619 if (unlikely(anon_vma_prepare(vma)))
2622 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2623 new_page = alloc_zeroed_user_highpage_movable(vma,
2628 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2633 if (!cow_user_page(new_page, old_page, vmf)) {
2635 * COW failed, if the fault was solved by other,
2636 * it's fine. If not, userspace would re-fault on
2637 * the same address and we will handle the fault
2638 * from the second attempt.
2647 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2650 __SetPageUptodate(new_page);
2652 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2655 * Re-check the pte - we dropped the lock
2657 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2658 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2660 if (!PageAnon(old_page)) {
2661 dec_mm_counter_fast(mm,
2662 mm_counter_file(old_page));
2663 inc_mm_counter_fast(mm, MM_ANONPAGES);
2666 inc_mm_counter_fast(mm, MM_ANONPAGES);
2668 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2669 entry = mk_pte(new_page, vma->vm_page_prot);
2670 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2672 * Clear the pte entry and flush it first, before updating the
2673 * pte with the new entry. This will avoid a race condition
2674 * seen in the presence of one thread doing SMC and another
2677 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2678 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2679 mem_cgroup_commit_charge(new_page, memcg, false, false);
2680 lru_cache_add_active_or_unevictable(new_page, vma);
2682 * We call the notify macro here because, when using secondary
2683 * mmu page tables (such as kvm shadow page tables), we want the
2684 * new page to be mapped directly into the secondary page table.
2686 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2687 update_mmu_cache(vma, vmf->address, vmf->pte);
2690 * Only after switching the pte to the new page may
2691 * we remove the mapcount here. Otherwise another
2692 * process may come and find the rmap count decremented
2693 * before the pte is switched to the new page, and
2694 * "reuse" the old page writing into it while our pte
2695 * here still points into it and can be read by other
2698 * The critical issue is to order this
2699 * page_remove_rmap with the ptp_clear_flush above.
2700 * Those stores are ordered by (if nothing else,)
2701 * the barrier present in the atomic_add_negative
2702 * in page_remove_rmap.
2704 * Then the TLB flush in ptep_clear_flush ensures that
2705 * no process can access the old page before the
2706 * decremented mapcount is visible. And the old page
2707 * cannot be reused until after the decremented
2708 * mapcount is visible. So transitively, TLBs to
2709 * old page will be flushed before it can be reused.
2711 page_remove_rmap(old_page, false);
2714 /* Free the old page.. */
2715 new_page = old_page;
2718 mem_cgroup_cancel_charge(new_page, memcg, false);
2724 pte_unmap_unlock(vmf->pte, vmf->ptl);
2726 * No need to double call mmu_notifier->invalidate_range() callback as
2727 * the above ptep_clear_flush_notify() did already call it.
2729 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2732 * Don't let another task, with possibly unlocked vma,
2733 * keep the mlocked page.
2735 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2736 lock_page(old_page); /* LRU manipulation */
2737 if (PageMlocked(old_page))
2738 munlock_vma_page(old_page);
2739 unlock_page(old_page);
2743 return page_copied ? VM_FAULT_WRITE : 0;
2749 return VM_FAULT_OOM;
2753 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2754 * writeable once the page is prepared
2756 * @vmf: structure describing the fault
2758 * This function handles all that is needed to finish a write page fault in a
2759 * shared mapping due to PTE being read-only once the mapped page is prepared.
2760 * It handles locking of PTE and modifying it. The function returns
2761 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2764 * The function expects the page to be locked or other protection against
2765 * concurrent faults / writeback (such as DAX radix tree locks).
2767 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2769 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2770 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2773 * We might have raced with another page fault while we released the
2774 * pte_offset_map_lock.
2776 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2777 pte_unmap_unlock(vmf->pte, vmf->ptl);
2778 return VM_FAULT_NOPAGE;
2785 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2788 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2790 struct vm_area_struct *vma = vmf->vma;
2792 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2795 pte_unmap_unlock(vmf->pte, vmf->ptl);
2796 vmf->flags |= FAULT_FLAG_MKWRITE;
2797 ret = vma->vm_ops->pfn_mkwrite(vmf);
2798 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2800 return finish_mkwrite_fault(vmf);
2803 return VM_FAULT_WRITE;
2806 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2807 __releases(vmf->ptl)
2809 struct vm_area_struct *vma = vmf->vma;
2811 get_page(vmf->page);
2813 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2816 pte_unmap_unlock(vmf->pte, vmf->ptl);
2817 tmp = do_page_mkwrite(vmf);
2818 if (unlikely(!tmp || (tmp &
2819 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2820 put_page(vmf->page);
2823 tmp = finish_mkwrite_fault(vmf);
2824 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2825 unlock_page(vmf->page);
2826 put_page(vmf->page);
2831 lock_page(vmf->page);
2833 fault_dirty_shared_page(vma, vmf->page);
2834 put_page(vmf->page);
2836 return VM_FAULT_WRITE;
2840 * This routine handles present pages, when users try to write
2841 * to a shared page. It is done by copying the page to a new address
2842 * and decrementing the shared-page counter for the old page.
2844 * Note that this routine assumes that the protection checks have been
2845 * done by the caller (the low-level page fault routine in most cases).
2846 * Thus we can safely just mark it writable once we've done any necessary
2849 * We also mark the page dirty at this point even though the page will
2850 * change only once the write actually happens. This avoids a few races,
2851 * and potentially makes it more efficient.
2853 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2854 * but allow concurrent faults), with pte both mapped and locked.
2855 * We return with mmap_sem still held, but pte unmapped and unlocked.
2857 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2858 __releases(vmf->ptl)
2860 struct vm_area_struct *vma = vmf->vma;
2862 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2865 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2868 * We should not cow pages in a shared writeable mapping.
2869 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2871 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2872 (VM_WRITE|VM_SHARED))
2873 return wp_pfn_shared(vmf);
2875 pte_unmap_unlock(vmf->pte, vmf->ptl);
2876 return wp_page_copy(vmf);
2880 * Take out anonymous pages first, anonymous shared vmas are
2881 * not dirty accountable.
2883 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2884 int total_map_swapcount;
2885 if (!trylock_page(vmf->page)) {
2886 get_page(vmf->page);
2887 pte_unmap_unlock(vmf->pte, vmf->ptl);
2888 lock_page(vmf->page);
2889 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2890 vmf->address, &vmf->ptl);
2891 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2892 unlock_page(vmf->page);
2893 pte_unmap_unlock(vmf->pte, vmf->ptl);
2894 put_page(vmf->page);
2897 put_page(vmf->page);
2899 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2900 if (total_map_swapcount == 1) {
2902 * The page is all ours. Move it to
2903 * our anon_vma so the rmap code will
2904 * not search our parent or siblings.
2905 * Protected against the rmap code by
2908 page_move_anon_rmap(vmf->page, vma);
2910 unlock_page(vmf->page);
2912 return VM_FAULT_WRITE;
2914 unlock_page(vmf->page);
2915 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2916 (VM_WRITE|VM_SHARED))) {
2917 return wp_page_shared(vmf);
2921 * Ok, we need to copy. Oh, well..
2923 get_page(vmf->page);
2925 pte_unmap_unlock(vmf->pte, vmf->ptl);
2926 return wp_page_copy(vmf);
2929 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2930 unsigned long start_addr, unsigned long end_addr,
2931 struct zap_details *details)
2933 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2936 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2937 struct zap_details *details)
2939 struct vm_area_struct *vma;
2940 pgoff_t vba, vea, zba, zea;
2942 vma_interval_tree_foreach(vma, root,
2943 details->first_index, details->last_index) {
2945 vba = vma->vm_pgoff;
2946 vea = vba + vma_pages(vma) - 1;
2947 zba = details->first_index;
2950 zea = details->last_index;
2954 unmap_mapping_range_vma(vma,
2955 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2956 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2962 * unmap_mapping_page() - Unmap single page from processes.
2963 * @page: The locked page to be unmapped.
2965 * Unmap this page from any userspace process which still has it mmaped.
2966 * Typically, for efficiency, the range of nearby pages has already been
2967 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
2968 * truncation or invalidation holds the lock on a page, it may find that
2969 * the page has been remapped again: and then uses unmap_mapping_page()
2970 * to unmap it finally.
2972 void unmap_mapping_page(struct page *page)
2974 struct address_space *mapping = page->mapping;
2975 struct zap_details details = { };
2977 VM_BUG_ON(!PageLocked(page));
2978 VM_BUG_ON(PageTail(page));
2980 details.check_mapping = mapping;
2981 details.first_index = page->index;
2982 details.last_index = page->index + hpage_nr_pages(page) - 1;
2983 details.single_page = page;
2985 i_mmap_lock_write(mapping);
2986 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2987 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2988 i_mmap_unlock_write(mapping);
2992 * unmap_mapping_pages() - Unmap pages from processes.
2993 * @mapping: The address space containing pages to be unmapped.
2994 * @start: Index of first page to be unmapped.
2995 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2996 * @even_cows: Whether to unmap even private COWed pages.
2998 * Unmap the pages in this address space from any userspace process which
2999 * has them mmaped. Generally, you want to remove COWed pages as well when
3000 * a file is being truncated, but not when invalidating pages from the page
3003 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3004 pgoff_t nr, bool even_cows)
3006 struct zap_details details = { };
3008 details.check_mapping = even_cows ? NULL : mapping;
3009 details.first_index = start;
3010 details.last_index = start + nr - 1;
3011 if (details.last_index < details.first_index)
3012 details.last_index = ULONG_MAX;
3014 i_mmap_lock_write(mapping);
3015 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3016 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3017 i_mmap_unlock_write(mapping);
3021 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3022 * address_space corresponding to the specified byte range in the underlying
3025 * @mapping: the address space containing mmaps to be unmapped.
3026 * @holebegin: byte in first page to unmap, relative to the start of
3027 * the underlying file. This will be rounded down to a PAGE_SIZE
3028 * boundary. Note that this is different from truncate_pagecache(), which
3029 * must keep the partial page. In contrast, we must get rid of
3031 * @holelen: size of prospective hole in bytes. This will be rounded
3032 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3034 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3035 * but 0 when invalidating pagecache, don't throw away private data.
3037 void unmap_mapping_range(struct address_space *mapping,
3038 loff_t const holebegin, loff_t const holelen, int even_cows)
3040 pgoff_t hba = holebegin >> PAGE_SHIFT;
3041 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3043 /* Check for overflow. */
3044 if (sizeof(holelen) > sizeof(hlen)) {
3046 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3047 if (holeend & ~(long long)ULONG_MAX)
3048 hlen = ULONG_MAX - hba + 1;
3051 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3053 EXPORT_SYMBOL(unmap_mapping_range);
3056 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3057 * but allow concurrent faults), and pte mapped but not yet locked.
3058 * We return with pte unmapped and unlocked.
3060 * We return with the mmap_sem locked or unlocked in the same cases
3061 * as does filemap_fault().
3063 vm_fault_t do_swap_page(struct vm_fault *vmf)
3065 struct vm_area_struct *vma = vmf->vma;
3066 struct page *page = NULL, *swapcache;
3067 struct mem_cgroup *memcg;
3074 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3077 entry = pte_to_swp_entry(vmf->orig_pte);
3078 if (unlikely(non_swap_entry(entry))) {
3079 if (is_migration_entry(entry)) {
3080 migration_entry_wait(vma->vm_mm, vmf->pmd,
3082 } else if (is_device_private_entry(entry)) {
3084 * For un-addressable device memory we call the pgmap
3085 * fault handler callback. The callback must migrate
3086 * the page back to some CPU accessible page.
3088 ret = device_private_entry_fault(vma, vmf->address, entry,
3089 vmf->flags, vmf->pmd);
3090 } else if (is_hwpoison_entry(entry)) {
3091 ret = VM_FAULT_HWPOISON;
3093 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3094 ret = VM_FAULT_SIGBUS;
3100 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3101 page = lookup_swap_cache(entry, vma, vmf->address);
3105 struct swap_info_struct *si = swp_swap_info(entry);
3107 if (si->flags & SWP_SYNCHRONOUS_IO &&
3108 __swap_count(si, entry) == 1) {
3109 /* skip swapcache */
3110 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3113 __SetPageLocked(page);
3114 __SetPageSwapBacked(page);
3115 set_page_private(page, entry.val);
3116 lru_cache_add_anon(page);
3117 swap_readpage(page, true);
3120 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3127 * Back out if somebody else faulted in this pte
3128 * while we released the pte lock.
3130 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3131 vmf->address, &vmf->ptl);
3132 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3134 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3138 /* Had to read the page from swap area: Major fault */
3139 ret = VM_FAULT_MAJOR;
3140 count_vm_event(PGMAJFAULT);
3141 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3142 } else if (PageHWPoison(page)) {
3144 * hwpoisoned dirty swapcache pages are kept for killing
3145 * owner processes (which may be unknown at hwpoison time)
3147 ret = VM_FAULT_HWPOISON;
3148 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3152 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3154 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3156 ret |= VM_FAULT_RETRY;
3161 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3162 * release the swapcache from under us. The page pin, and pte_same
3163 * test below, are not enough to exclude that. Even if it is still
3164 * swapcache, we need to check that the page's swap has not changed.
3166 if (unlikely((!PageSwapCache(page) ||
3167 page_private(page) != entry.val)) && swapcache)
3170 page = ksm_might_need_to_copy(page, vma, vmf->address);
3171 if (unlikely(!page)) {
3177 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3184 * Back out if somebody else already faulted in this pte.
3186 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3188 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3191 if (unlikely(!PageUptodate(page))) {
3192 ret = VM_FAULT_SIGBUS;
3197 * The page isn't present yet, go ahead with the fault.
3199 * Be careful about the sequence of operations here.
3200 * To get its accounting right, reuse_swap_page() must be called
3201 * while the page is counted on swap but not yet in mapcount i.e.
3202 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3203 * must be called after the swap_free(), or it will never succeed.
3206 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3207 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3208 pte = mk_pte(page, vma->vm_page_prot);
3209 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3210 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3211 vmf->flags &= ~FAULT_FLAG_WRITE;
3212 ret |= VM_FAULT_WRITE;
3213 exclusive = RMAP_EXCLUSIVE;
3215 flush_icache_page(vma, page);
3216 if (pte_swp_soft_dirty(vmf->orig_pte))
3217 pte = pte_mksoft_dirty(pte);
3218 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3219 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3220 vmf->orig_pte = pte;
3222 /* ksm created a completely new copy */
3223 if (unlikely(page != swapcache && swapcache)) {
3224 page_add_new_anon_rmap(page, vma, vmf->address, false);
3225 mem_cgroup_commit_charge(page, memcg, false, false);
3226 lru_cache_add_active_or_unevictable(page, vma);
3228 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3229 mem_cgroup_commit_charge(page, memcg, true, false);
3230 activate_page(page);
3234 if (mem_cgroup_swap_full(page) ||
3235 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3236 try_to_free_swap(page);
3238 if (page != swapcache && swapcache) {
3240 * Hold the lock to avoid the swap entry to be reused
3241 * until we take the PT lock for the pte_same() check
3242 * (to avoid false positives from pte_same). For
3243 * further safety release the lock after the swap_free
3244 * so that the swap count won't change under a
3245 * parallel locked swapcache.
3247 unlock_page(swapcache);
3248 put_page(swapcache);
3251 if (vmf->flags & FAULT_FLAG_WRITE) {
3252 ret |= do_wp_page(vmf);
3253 if (ret & VM_FAULT_ERROR)
3254 ret &= VM_FAULT_ERROR;
3258 /* No need to invalidate - it was non-present before */
3259 update_mmu_cache(vma, vmf->address, vmf->pte);
3261 pte_unmap_unlock(vmf->pte, vmf->ptl);
3265 mem_cgroup_cancel_charge(page, memcg, false);
3266 pte_unmap_unlock(vmf->pte, vmf->ptl);
3271 if (page != swapcache && swapcache) {
3272 unlock_page(swapcache);
3273 put_page(swapcache);
3279 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3280 * but allow concurrent faults), and pte mapped but not yet locked.
3281 * We return with mmap_sem still held, but pte unmapped and unlocked.
3283 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3285 struct vm_area_struct *vma = vmf->vma;
3286 struct mem_cgroup *memcg;
3291 /* File mapping without ->vm_ops ? */
3292 if (vma->vm_flags & VM_SHARED)
3293 return VM_FAULT_SIGBUS;
3296 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3297 * pte_offset_map() on pmds where a huge pmd might be created
3298 * from a different thread.
3300 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3301 * parallel threads are excluded by other means.
3303 * Here we only have down_read(mmap_sem).
3305 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3306 return VM_FAULT_OOM;
3308 /* See the comment in pte_alloc_one_map() */
3309 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3312 /* Use the zero-page for reads */
3313 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3314 !mm_forbids_zeropage(vma->vm_mm)) {
3315 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3316 vma->vm_page_prot));
3317 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3318 vmf->address, &vmf->ptl);
3319 if (!pte_none(*vmf->pte))
3321 ret = check_stable_address_space(vma->vm_mm);
3324 /* Deliver the page fault to userland, check inside PT lock */
3325 if (userfaultfd_missing(vma)) {
3326 pte_unmap_unlock(vmf->pte, vmf->ptl);
3327 return handle_userfault(vmf, VM_UFFD_MISSING);
3332 /* Allocate our own private page. */
3333 if (unlikely(anon_vma_prepare(vma)))
3335 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3339 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3344 * The memory barrier inside __SetPageUptodate makes sure that
3345 * preceeding stores to the page contents become visible before
3346 * the set_pte_at() write.
3348 __SetPageUptodate(page);
3350 entry = mk_pte(page, vma->vm_page_prot);
3351 if (vma->vm_flags & VM_WRITE)
3352 entry = pte_mkwrite(pte_mkdirty(entry));
3354 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3356 if (!pte_none(*vmf->pte))
3359 ret = check_stable_address_space(vma->vm_mm);
3363 /* Deliver the page fault to userland, check inside PT lock */
3364 if (userfaultfd_missing(vma)) {
3365 pte_unmap_unlock(vmf->pte, vmf->ptl);
3366 mem_cgroup_cancel_charge(page, memcg, false);
3368 return handle_userfault(vmf, VM_UFFD_MISSING);
3371 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3372 page_add_new_anon_rmap(page, vma, vmf->address, false);
3373 mem_cgroup_commit_charge(page, memcg, false, false);
3374 lru_cache_add_active_or_unevictable(page, vma);
3376 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3378 /* No need to invalidate - it was non-present before */
3379 update_mmu_cache(vma, vmf->address, vmf->pte);
3381 pte_unmap_unlock(vmf->pte, vmf->ptl);
3384 mem_cgroup_cancel_charge(page, memcg, false);
3390 return VM_FAULT_OOM;
3394 * The mmap_sem must have been held on entry, and may have been
3395 * released depending on flags and vma->vm_ops->fault() return value.
3396 * See filemap_fault() and __lock_page_retry().
3398 static vm_fault_t __do_fault(struct vm_fault *vmf)
3400 struct vm_area_struct *vma = vmf->vma;
3404 * Preallocate pte before we take page_lock because this might lead to
3405 * deadlocks for memcg reclaim which waits for pages under writeback:
3407 * SetPageWriteback(A)
3413 * wait_on_page_writeback(A)
3414 * SetPageWriteback(B)
3416 * # flush A, B to clear the writeback
3418 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3419 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3421 if (!vmf->prealloc_pte)
3422 return VM_FAULT_OOM;
3423 smp_wmb(); /* See comment in __pte_alloc() */
3426 ret = vma->vm_ops->fault(vmf);
3427 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3428 VM_FAULT_DONE_COW)))
3431 if (unlikely(PageHWPoison(vmf->page))) {
3432 struct page *page = vmf->page;
3433 vm_fault_t poisonret = VM_FAULT_HWPOISON;
3434 if (ret & VM_FAULT_LOCKED) {
3435 if (page_mapped(page))
3436 unmap_mapping_pages(page_mapping(page),
3437 page->index, 1, false);
3438 /* Retry if a clean page was removed from the cache. */
3439 if (invalidate_inode_page(page))
3440 poisonret = VM_FAULT_NOPAGE;
3448 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3449 lock_page(vmf->page);
3451 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3457 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3458 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3459 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3460 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3462 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3464 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3467 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3469 struct vm_area_struct *vma = vmf->vma;
3471 if (!pmd_none(*vmf->pmd))
3473 if (vmf->prealloc_pte) {
3474 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3475 if (unlikely(!pmd_none(*vmf->pmd))) {
3476 spin_unlock(vmf->ptl);
3480 mm_inc_nr_ptes(vma->vm_mm);
3481 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3482 spin_unlock(vmf->ptl);
3483 vmf->prealloc_pte = NULL;
3484 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3485 return VM_FAULT_OOM;
3489 * If a huge pmd materialized under us just retry later. Use
3490 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3491 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3492 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3493 * running immediately after a huge pmd fault in a different thread of
3494 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3495 * All we have to ensure is that it is a regular pmd that we can walk
3496 * with pte_offset_map() and we can do that through an atomic read in
3497 * C, which is what pmd_trans_unstable() provides.
3499 if (pmd_devmap_trans_unstable(vmf->pmd))
3500 return VM_FAULT_NOPAGE;
3503 * At this point we know that our vmf->pmd points to a page of ptes
3504 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3505 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3506 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3507 * be valid and we will re-check to make sure the vmf->pte isn't
3508 * pte_none() under vmf->ptl protection when we return to
3511 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3516 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3518 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3519 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3520 unsigned long haddr)
3522 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3523 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3525 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3530 static void deposit_prealloc_pte(struct vm_fault *vmf)
3532 struct vm_area_struct *vma = vmf->vma;
3534 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3536 * We are going to consume the prealloc table,
3537 * count that as nr_ptes.
3539 mm_inc_nr_ptes(vma->vm_mm);
3540 vmf->prealloc_pte = NULL;
3543 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3545 struct vm_area_struct *vma = vmf->vma;
3546 bool write = vmf->flags & FAULT_FLAG_WRITE;
3547 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3552 if (!transhuge_vma_suitable(vma, haddr))
3553 return VM_FAULT_FALLBACK;
3555 ret = VM_FAULT_FALLBACK;
3556 page = compound_head(page);
3559 * Archs like ppc64 need additonal space to store information
3560 * related to pte entry. Use the preallocated table for that.
3562 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3563 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3564 if (!vmf->prealloc_pte)
3565 return VM_FAULT_OOM;
3566 smp_wmb(); /* See comment in __pte_alloc() */
3569 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3570 if (unlikely(!pmd_none(*vmf->pmd)))
3573 for (i = 0; i < HPAGE_PMD_NR; i++)
3574 flush_icache_page(vma, page + i);
3576 entry = mk_huge_pmd(page, vma->vm_page_prot);
3578 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3580 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3581 page_add_file_rmap(page, true);
3583 * deposit and withdraw with pmd lock held
3585 if (arch_needs_pgtable_deposit())
3586 deposit_prealloc_pte(vmf);
3588 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3590 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3592 /* fault is handled */
3594 count_vm_event(THP_FILE_MAPPED);
3596 spin_unlock(vmf->ptl);
3600 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3608 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3609 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3611 * @vmf: fault environment
3612 * @memcg: memcg to charge page (only for private mappings)
3613 * @page: page to map
3615 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3618 * Target users are page handler itself and implementations of
3619 * vm_ops->map_pages.
3621 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3624 struct vm_area_struct *vma = vmf->vma;
3625 bool write = vmf->flags & FAULT_FLAG_WRITE;
3629 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3630 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3632 VM_BUG_ON_PAGE(memcg, page);
3634 ret = do_set_pmd(vmf, page);
3635 if (ret != VM_FAULT_FALLBACK)
3640 ret = pte_alloc_one_map(vmf);
3645 /* Re-check under ptl */
3646 if (unlikely(!pte_none(*vmf->pte)))
3647 return VM_FAULT_NOPAGE;
3649 flush_icache_page(vma, page);
3650 entry = mk_pte(page, vma->vm_page_prot);
3652 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3653 /* copy-on-write page */
3654 if (write && !(vma->vm_flags & VM_SHARED)) {
3655 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3656 page_add_new_anon_rmap(page, vma, vmf->address, false);
3657 mem_cgroup_commit_charge(page, memcg, false, false);
3658 lru_cache_add_active_or_unevictable(page, vma);
3660 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3661 page_add_file_rmap(page, false);
3663 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3665 /* no need to invalidate: a not-present page won't be cached */
3666 update_mmu_cache(vma, vmf->address, vmf->pte);
3673 * finish_fault - finish page fault once we have prepared the page to fault
3675 * @vmf: structure describing the fault
3677 * This function handles all that is needed to finish a page fault once the
3678 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3679 * given page, adds reverse page mapping, handles memcg charges and LRU
3680 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3683 * The function expects the page to be locked and on success it consumes a
3684 * reference of a page being mapped (for the PTE which maps it).
3686 vm_fault_t finish_fault(struct vm_fault *vmf)
3691 /* Did we COW the page? */
3692 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3693 !(vmf->vma->vm_flags & VM_SHARED))
3694 page = vmf->cow_page;
3699 * check even for read faults because we might have lost our CoWed
3702 if (!(vmf->vma->vm_flags & VM_SHARED))
3703 ret = check_stable_address_space(vmf->vma->vm_mm);
3705 ret = alloc_set_pte(vmf, vmf->memcg, page);
3707 pte_unmap_unlock(vmf->pte, vmf->ptl);
3711 static unsigned long fault_around_bytes __read_mostly =
3712 rounddown_pow_of_two(65536);
3714 #ifdef CONFIG_DEBUG_FS
3715 static int fault_around_bytes_get(void *data, u64 *val)
3717 *val = fault_around_bytes;
3722 * fault_around_bytes must be rounded down to the nearest page order as it's
3723 * what do_fault_around() expects to see.
3725 static int fault_around_bytes_set(void *data, u64 val)
3727 if (val / PAGE_SIZE > PTRS_PER_PTE)
3729 if (val > PAGE_SIZE)
3730 fault_around_bytes = rounddown_pow_of_two(val);
3732 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3735 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3736 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3738 static int __init fault_around_debugfs(void)
3742 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3743 &fault_around_bytes_fops);
3745 pr_warn("Failed to create fault_around_bytes in debugfs");
3748 late_initcall(fault_around_debugfs);
3752 * do_fault_around() tries to map few pages around the fault address. The hope
3753 * is that the pages will be needed soon and this will lower the number of
3756 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3757 * not ready to be mapped: not up-to-date, locked, etc.
3759 * This function is called with the page table lock taken. In the split ptlock
3760 * case the page table lock only protects only those entries which belong to
3761 * the page table corresponding to the fault address.
3763 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3766 * fault_around_bytes defines how many bytes we'll try to map.
3767 * do_fault_around() expects it to be set to a power of two less than or equal
3770 * The virtual address of the area that we map is naturally aligned to
3771 * fault_around_bytes rounded down to the machine page size
3772 * (and therefore to page order). This way it's easier to guarantee
3773 * that we don't cross page table boundaries.
3775 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3777 unsigned long address = vmf->address, nr_pages, mask;
3778 pgoff_t start_pgoff = vmf->pgoff;
3783 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3784 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3786 vmf->address = max(address & mask, vmf->vma->vm_start);
3787 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3791 * end_pgoff is either the end of the page table, the end of
3792 * the vma or nr_pages from start_pgoff, depending what is nearest.
3794 end_pgoff = start_pgoff -
3795 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3797 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3798 start_pgoff + nr_pages - 1);
3800 if (pmd_none(*vmf->pmd)) {
3801 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3803 if (!vmf->prealloc_pte)
3805 smp_wmb(); /* See comment in __pte_alloc() */
3808 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3810 /* Huge page is mapped? Page fault is solved */
3811 if (pmd_trans_huge(*vmf->pmd)) {
3812 ret = VM_FAULT_NOPAGE;
3816 /* ->map_pages() haven't done anything useful. Cold page cache? */
3820 /* check if the page fault is solved */
3821 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3822 if (!pte_none(*vmf->pte))
3823 ret = VM_FAULT_NOPAGE;
3824 pte_unmap_unlock(vmf->pte, vmf->ptl);
3826 vmf->address = address;
3831 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3833 struct vm_area_struct *vma = vmf->vma;
3837 * Let's call ->map_pages() first and use ->fault() as fallback
3838 * if page by the offset is not ready to be mapped (cold cache or
3841 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3842 ret = do_fault_around(vmf);
3847 ret = __do_fault(vmf);
3848 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3851 ret |= finish_fault(vmf);
3852 unlock_page(vmf->page);
3853 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3854 put_page(vmf->page);
3858 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3860 struct vm_area_struct *vma = vmf->vma;
3863 if (unlikely(anon_vma_prepare(vma)))
3864 return VM_FAULT_OOM;
3866 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3868 return VM_FAULT_OOM;
3870 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3871 &vmf->memcg, false)) {
3872 put_page(vmf->cow_page);
3873 return VM_FAULT_OOM;
3876 ret = __do_fault(vmf);
3877 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3879 if (ret & VM_FAULT_DONE_COW)
3882 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3883 __SetPageUptodate(vmf->cow_page);
3885 ret |= finish_fault(vmf);
3886 unlock_page(vmf->page);
3887 put_page(vmf->page);
3888 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3892 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3893 put_page(vmf->cow_page);
3897 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3899 struct vm_area_struct *vma = vmf->vma;
3900 vm_fault_t ret, tmp;
3902 ret = __do_fault(vmf);
3903 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3907 * Check if the backing address space wants to know that the page is
3908 * about to become writable
3910 if (vma->vm_ops->page_mkwrite) {
3911 unlock_page(vmf->page);
3912 tmp = do_page_mkwrite(vmf);
3913 if (unlikely(!tmp ||
3914 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3915 put_page(vmf->page);
3920 ret |= finish_fault(vmf);
3921 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3923 unlock_page(vmf->page);
3924 put_page(vmf->page);
3928 fault_dirty_shared_page(vma, vmf->page);
3933 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3934 * but allow concurrent faults).
3935 * The mmap_sem may have been released depending on flags and our
3936 * return value. See filemap_fault() and __lock_page_or_retry().
3937 * If mmap_sem is released, vma may become invalid (for example
3938 * by other thread calling munmap()).
3940 static vm_fault_t do_fault(struct vm_fault *vmf)
3942 struct vm_area_struct *vma = vmf->vma;
3943 struct mm_struct *vm_mm = vma->vm_mm;
3947 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3949 if (!vma->vm_ops->fault) {
3951 * If we find a migration pmd entry or a none pmd entry, which
3952 * should never happen, return SIGBUS
3954 if (unlikely(!pmd_present(*vmf->pmd)))
3955 ret = VM_FAULT_SIGBUS;
3957 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3962 * Make sure this is not a temporary clearing of pte
3963 * by holding ptl and checking again. A R/M/W update
3964 * of pte involves: take ptl, clearing the pte so that
3965 * we don't have concurrent modification by hardware
3966 * followed by an update.
3968 if (unlikely(pte_none(*vmf->pte)))
3969 ret = VM_FAULT_SIGBUS;
3971 ret = VM_FAULT_NOPAGE;
3973 pte_unmap_unlock(vmf->pte, vmf->ptl);
3975 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3976 ret = do_read_fault(vmf);
3977 else if (!(vma->vm_flags & VM_SHARED))
3978 ret = do_cow_fault(vmf);
3980 ret = do_shared_fault(vmf);
3982 /* preallocated pagetable is unused: free it */
3983 if (vmf->prealloc_pte) {
3984 pte_free(vm_mm, vmf->prealloc_pte);
3985 vmf->prealloc_pte = NULL;
3990 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3991 unsigned long addr, int page_nid,
3996 count_vm_numa_event(NUMA_HINT_FAULTS);
3997 if (page_nid == numa_node_id()) {
3998 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3999 *flags |= TNF_FAULT_LOCAL;
4002 return mpol_misplaced(page, vma, addr);
4005 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4007 struct vm_area_struct *vma = vmf->vma;
4008 struct page *page = NULL;
4012 bool migrated = false;
4014 bool was_writable = pte_savedwrite(vmf->orig_pte);
4018 * The "pte" at this point cannot be used safely without
4019 * validation through pte_unmap_same(). It's of NUMA type but
4020 * the pfn may be screwed if the read is non atomic.
4022 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4023 spin_lock(vmf->ptl);
4024 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4025 pte_unmap_unlock(vmf->pte, vmf->ptl);
4030 * Make it present again, Depending on how arch implementes non
4031 * accessible ptes, some can allow access by kernel mode.
4033 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
4034 pte = pte_modify(pte, vma->vm_page_prot);
4035 pte = pte_mkyoung(pte);
4037 pte = pte_mkwrite(pte);
4038 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
4039 update_mmu_cache(vma, vmf->address, vmf->pte);
4041 page = vm_normal_page(vma, vmf->address, pte);
4043 pte_unmap_unlock(vmf->pte, vmf->ptl);
4047 /* TODO: handle PTE-mapped THP */
4048 if (PageCompound(page)) {
4049 pte_unmap_unlock(vmf->pte, vmf->ptl);
4054 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4055 * much anyway since they can be in shared cache state. This misses
4056 * the case where a mapping is writable but the process never writes
4057 * to it but pte_write gets cleared during protection updates and
4058 * pte_dirty has unpredictable behaviour between PTE scan updates,
4059 * background writeback, dirty balancing and application behaviour.
4061 if (!pte_write(pte))
4062 flags |= TNF_NO_GROUP;
4065 * Flag if the page is shared between multiple address spaces. This
4066 * is later used when determining whether to group tasks together
4068 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4069 flags |= TNF_SHARED;
4071 last_cpupid = page_cpupid_last(page);
4072 page_nid = page_to_nid(page);
4073 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4075 pte_unmap_unlock(vmf->pte, vmf->ptl);
4076 if (target_nid == -1) {
4081 /* Migrate to the requested node */
4082 migrated = migrate_misplaced_page(page, vma, target_nid);
4084 page_nid = target_nid;
4085 flags |= TNF_MIGRATED;
4087 flags |= TNF_MIGRATE_FAIL;
4091 task_numa_fault(last_cpupid, page_nid, 1, flags);
4095 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4097 if (vma_is_anonymous(vmf->vma))
4098 return do_huge_pmd_anonymous_page(vmf);
4099 if (vmf->vma->vm_ops->huge_fault)
4100 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4101 return VM_FAULT_FALLBACK;
4104 /* `inline' is required to avoid gcc 4.1.2 build error */
4105 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4107 if (vma_is_anonymous(vmf->vma))
4108 return do_huge_pmd_wp_page(vmf, orig_pmd);
4109 if (vmf->vma->vm_ops->huge_fault)
4110 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4112 /* COW handled on pte level: split pmd */
4113 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
4114 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4116 return VM_FAULT_FALLBACK;
4119 static inline bool vma_is_accessible(struct vm_area_struct *vma)
4121 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
4124 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4126 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4127 /* No support for anonymous transparent PUD pages yet */
4128 if (vma_is_anonymous(vmf->vma))
4129 return VM_FAULT_FALLBACK;
4130 if (vmf->vma->vm_ops->huge_fault)
4131 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4132 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4133 return VM_FAULT_FALLBACK;
4136 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4138 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4139 /* No support for anonymous transparent PUD pages yet */
4140 if (vma_is_anonymous(vmf->vma))
4141 return VM_FAULT_FALLBACK;
4142 if (vmf->vma->vm_ops->huge_fault)
4143 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4144 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4145 return VM_FAULT_FALLBACK;
4149 * These routines also need to handle stuff like marking pages dirty
4150 * and/or accessed for architectures that don't do it in hardware (most
4151 * RISC architectures). The early dirtying is also good on the i386.
4153 * There is also a hook called "update_mmu_cache()" that architectures
4154 * with external mmu caches can use to update those (ie the Sparc or
4155 * PowerPC hashed page tables that act as extended TLBs).
4157 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
4158 * concurrent faults).
4160 * The mmap_sem may have been released depending on flags and our return value.
4161 * See filemap_fault() and __lock_page_or_retry().
4163 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4167 if (unlikely(pmd_none(*vmf->pmd))) {
4169 * Leave __pte_alloc() until later: because vm_ops->fault may
4170 * want to allocate huge page, and if we expose page table
4171 * for an instant, it will be difficult to retract from
4172 * concurrent faults and from rmap lookups.
4176 /* See comment in pte_alloc_one_map() */
4177 if (pmd_devmap_trans_unstable(vmf->pmd))
4180 * A regular pmd is established and it can't morph into a huge
4181 * pmd from under us anymore at this point because we hold the
4182 * mmap_sem read mode and khugepaged takes it in write mode.
4183 * So now it's safe to run pte_offset_map().
4185 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4186 vmf->orig_pte = *vmf->pte;
4189 * some architectures can have larger ptes than wordsize,
4190 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4191 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4192 * accesses. The code below just needs a consistent view
4193 * for the ifs and we later double check anyway with the
4194 * ptl lock held. So here a barrier will do.
4197 if (pte_none(vmf->orig_pte)) {
4198 pte_unmap(vmf->pte);
4204 if (vma_is_anonymous(vmf->vma))
4205 return do_anonymous_page(vmf);
4207 return do_fault(vmf);
4210 if (!pte_present(vmf->orig_pte))
4211 return do_swap_page(vmf);
4213 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4214 return do_numa_page(vmf);
4216 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4217 spin_lock(vmf->ptl);
4218 entry = vmf->orig_pte;
4219 if (unlikely(!pte_same(*vmf->pte, entry)))
4221 if (vmf->flags & FAULT_FLAG_WRITE) {
4222 if (!pte_write(entry))
4223 return do_wp_page(vmf);
4224 entry = pte_mkdirty(entry);
4226 entry = pte_mkyoung(entry);
4227 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4228 vmf->flags & FAULT_FLAG_WRITE)) {
4229 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4232 * This is needed only for protection faults but the arch code
4233 * is not yet telling us if this is a protection fault or not.
4234 * This still avoids useless tlb flushes for .text page faults
4237 if (vmf->flags & FAULT_FLAG_WRITE)
4238 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4241 pte_unmap_unlock(vmf->pte, vmf->ptl);
4246 * By the time we get here, we already hold the mm semaphore
4248 * The mmap_sem may have been released depending on flags and our
4249 * return value. See filemap_fault() and __lock_page_or_retry().
4251 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4252 unsigned long address, unsigned int flags)
4254 struct vm_fault vmf = {
4256 .address = address & PAGE_MASK,
4258 .pgoff = linear_page_index(vma, address),
4259 .gfp_mask = __get_fault_gfp_mask(vma),
4261 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4262 struct mm_struct *mm = vma->vm_mm;
4267 pgd = pgd_offset(mm, address);
4268 p4d = p4d_alloc(mm, pgd, address);
4270 return VM_FAULT_OOM;
4272 vmf.pud = pud_alloc(mm, p4d, address);
4274 return VM_FAULT_OOM;
4275 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4276 ret = create_huge_pud(&vmf);
4277 if (!(ret & VM_FAULT_FALLBACK))
4280 pud_t orig_pud = *vmf.pud;
4283 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4285 /* NUMA case for anonymous PUDs would go here */
4287 if (dirty && !pud_write(orig_pud)) {
4288 ret = wp_huge_pud(&vmf, orig_pud);
4289 if (!(ret & VM_FAULT_FALLBACK))
4292 huge_pud_set_accessed(&vmf, orig_pud);
4298 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4300 return VM_FAULT_OOM;
4301 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4302 ret = create_huge_pmd(&vmf);
4303 if (!(ret & VM_FAULT_FALLBACK))
4306 pmd_t orig_pmd = *vmf.pmd;
4309 if (unlikely(is_swap_pmd(orig_pmd))) {
4310 VM_BUG_ON(thp_migration_supported() &&
4311 !is_pmd_migration_entry(orig_pmd));
4312 if (is_pmd_migration_entry(orig_pmd))
4313 pmd_migration_entry_wait(mm, vmf.pmd);
4316 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4317 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4318 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4320 if (dirty && !pmd_write(orig_pmd)) {
4321 ret = wp_huge_pmd(&vmf, orig_pmd);
4322 if (!(ret & VM_FAULT_FALLBACK))
4325 huge_pmd_set_accessed(&vmf, orig_pmd);
4331 return handle_pte_fault(&vmf);
4335 * By the time we get here, we already hold the mm semaphore
4337 * The mmap_sem may have been released depending on flags and our
4338 * return value. See filemap_fault() and __lock_page_or_retry().
4340 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4345 __set_current_state(TASK_RUNNING);
4347 count_vm_event(PGFAULT);
4348 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4350 /* do counter updates before entering really critical section. */
4351 check_sync_rss_stat(current);
4353 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4354 flags & FAULT_FLAG_INSTRUCTION,
4355 flags & FAULT_FLAG_REMOTE))
4356 return VM_FAULT_SIGSEGV;
4359 * Enable the memcg OOM handling for faults triggered in user
4360 * space. Kernel faults are handled more gracefully.
4362 if (flags & FAULT_FLAG_USER)
4363 mem_cgroup_enter_user_fault();
4365 if (unlikely(is_vm_hugetlb_page(vma)))
4366 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4368 ret = __handle_mm_fault(vma, address, flags);
4370 if (flags & FAULT_FLAG_USER) {
4371 mem_cgroup_exit_user_fault();
4373 * The task may have entered a memcg OOM situation but
4374 * if the allocation error was handled gracefully (no
4375 * VM_FAULT_OOM), there is no need to kill anything.
4376 * Just clean up the OOM state peacefully.
4378 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4379 mem_cgroup_oom_synchronize(false);
4384 EXPORT_SYMBOL_GPL(handle_mm_fault);
4386 #ifndef __PAGETABLE_P4D_FOLDED
4388 * Allocate p4d page table.
4389 * We've already handled the fast-path in-line.
4391 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4393 p4d_t *new = p4d_alloc_one(mm, address);
4397 smp_wmb(); /* See comment in __pte_alloc */
4399 spin_lock(&mm->page_table_lock);
4400 if (pgd_present(*pgd)) /* Another has populated it */
4403 pgd_populate(mm, pgd, new);
4404 spin_unlock(&mm->page_table_lock);
4407 #endif /* __PAGETABLE_P4D_FOLDED */
4409 #ifndef __PAGETABLE_PUD_FOLDED
4411 * Allocate page upper directory.
4412 * We've already handled the fast-path in-line.
4414 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4416 pud_t *new = pud_alloc_one(mm, address);
4420 smp_wmb(); /* See comment in __pte_alloc */
4422 spin_lock(&mm->page_table_lock);
4423 #ifndef __ARCH_HAS_5LEVEL_HACK
4424 if (!p4d_present(*p4d)) {
4426 p4d_populate(mm, p4d, new);
4427 } else /* Another has populated it */
4430 if (!pgd_present(*p4d)) {
4432 pgd_populate(mm, p4d, new);
4433 } else /* Another has populated it */
4435 #endif /* __ARCH_HAS_5LEVEL_HACK */
4436 spin_unlock(&mm->page_table_lock);
4439 #endif /* __PAGETABLE_PUD_FOLDED */
4441 #ifndef __PAGETABLE_PMD_FOLDED
4443 * Allocate page middle directory.
4444 * We've already handled the fast-path in-line.
4446 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4449 pmd_t *new = pmd_alloc_one(mm, address);
4453 smp_wmb(); /* See comment in __pte_alloc */
4455 ptl = pud_lock(mm, pud);
4456 #ifndef __ARCH_HAS_4LEVEL_HACK
4457 if (!pud_present(*pud)) {
4459 pud_populate(mm, pud, new);
4460 } else /* Another has populated it */
4463 if (!pgd_present(*pud)) {
4465 pgd_populate(mm, pud, new);
4466 } else /* Another has populated it */
4468 #endif /* __ARCH_HAS_4LEVEL_HACK */
4472 #endif /* __PAGETABLE_PMD_FOLDED */
4474 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4475 unsigned long *start, unsigned long *end,
4476 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4484 pgd = pgd_offset(mm, address);
4485 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4488 p4d = p4d_offset(pgd, address);
4489 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4492 pud = pud_offset(p4d, address);
4493 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4496 pmd = pmd_offset(pud, address);
4497 VM_BUG_ON(pmd_trans_huge(*pmd));
4499 if (pmd_huge(*pmd)) {
4504 *start = address & PMD_MASK;
4505 *end = *start + PMD_SIZE;
4506 mmu_notifier_invalidate_range_start(mm, *start, *end);
4508 *ptlp = pmd_lock(mm, pmd);
4509 if (pmd_huge(*pmd)) {
4515 mmu_notifier_invalidate_range_end(mm, *start, *end);
4518 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4522 *start = address & PAGE_MASK;
4523 *end = *start + PAGE_SIZE;
4524 mmu_notifier_invalidate_range_start(mm, *start, *end);
4526 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4527 if (!pte_present(*ptep))
4532 pte_unmap_unlock(ptep, *ptlp);
4534 mmu_notifier_invalidate_range_end(mm, *start, *end);
4539 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4540 pte_t **ptepp, spinlock_t **ptlp)
4544 /* (void) is needed to make gcc happy */
4545 (void) __cond_lock(*ptlp,
4546 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4547 ptepp, NULL, ptlp)));
4551 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4552 unsigned long *start, unsigned long *end,
4553 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4557 /* (void) is needed to make gcc happy */
4558 (void) __cond_lock(*ptlp,
4559 !(res = __follow_pte_pmd(mm, address, start, end,
4560 ptepp, pmdpp, ptlp)));
4563 EXPORT_SYMBOL(follow_pte_pmd);
4566 * follow_pfn - look up PFN at a user virtual address
4567 * @vma: memory mapping
4568 * @address: user virtual address
4569 * @pfn: location to store found PFN
4571 * Only IO mappings and raw PFN mappings are allowed.
4573 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4575 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4582 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4585 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4588 *pfn = pte_pfn(*ptep);
4589 pte_unmap_unlock(ptep, ptl);
4592 EXPORT_SYMBOL(follow_pfn);
4594 #ifdef CONFIG_HAVE_IOREMAP_PROT
4595 int follow_phys(struct vm_area_struct *vma,
4596 unsigned long address, unsigned int flags,
4597 unsigned long *prot, resource_size_t *phys)
4603 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4606 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4610 if ((flags & FOLL_WRITE) && !pte_write(pte))
4613 *prot = pgprot_val(pte_pgprot(pte));
4614 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4618 pte_unmap_unlock(ptep, ptl);
4623 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4624 void *buf, int len, int write)
4626 resource_size_t phys_addr;
4627 unsigned long prot = 0;
4628 void __iomem *maddr;
4629 int offset = addr & (PAGE_SIZE-1);
4631 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4634 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4639 memcpy_toio(maddr + offset, buf, len);
4641 memcpy_fromio(buf, maddr + offset, len);
4646 EXPORT_SYMBOL_GPL(generic_access_phys);
4650 * Access another process' address space as given in mm. If non-NULL, use the
4651 * given task for page fault accounting.
4653 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4654 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4656 struct vm_area_struct *vma;
4657 void *old_buf = buf;
4658 int write = gup_flags & FOLL_WRITE;
4660 if (down_read_killable(&mm->mmap_sem))
4663 /* ignore errors, just check how much was successfully transferred */
4665 int bytes, ret, offset;
4667 struct page *page = NULL;
4669 ret = get_user_pages_remote(tsk, mm, addr, 1,
4670 gup_flags, &page, &vma, NULL);
4672 #ifndef CONFIG_HAVE_IOREMAP_PROT
4676 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4677 * we can access using slightly different code.
4679 vma = find_vma(mm, addr);
4680 if (!vma || vma->vm_start > addr)
4682 if (vma->vm_ops && vma->vm_ops->access)
4683 ret = vma->vm_ops->access(vma, addr, buf,
4691 offset = addr & (PAGE_SIZE-1);
4692 if (bytes > PAGE_SIZE-offset)
4693 bytes = PAGE_SIZE-offset;
4697 copy_to_user_page(vma, page, addr,
4698 maddr + offset, buf, bytes);
4699 set_page_dirty_lock(page);
4701 copy_from_user_page(vma, page, addr,
4702 buf, maddr + offset, bytes);
4711 up_read(&mm->mmap_sem);
4713 return buf - old_buf;
4717 * access_remote_vm - access another process' address space
4718 * @mm: the mm_struct of the target address space
4719 * @addr: start address to access
4720 * @buf: source or destination buffer
4721 * @len: number of bytes to transfer
4722 * @gup_flags: flags modifying lookup behaviour
4724 * The caller must hold a reference on @mm.
4726 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4727 void *buf, int len, unsigned int gup_flags)
4729 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4733 * Access another process' address space.
4734 * Source/target buffer must be kernel space,
4735 * Do not walk the page table directly, use get_user_pages
4737 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4738 void *buf, int len, unsigned int gup_flags)
4740 struct mm_struct *mm;
4743 mm = get_task_mm(tsk);
4747 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4753 EXPORT_SYMBOL_GPL(access_process_vm);
4756 * Print the name of a VMA.
4758 void print_vma_addr(char *prefix, unsigned long ip)
4760 struct mm_struct *mm = current->mm;
4761 struct vm_area_struct *vma;
4764 * we might be running from an atomic context so we cannot sleep
4766 if (!down_read_trylock(&mm->mmap_sem))
4769 vma = find_vma(mm, ip);
4770 if (vma && vma->vm_file) {
4771 struct file *f = vma->vm_file;
4772 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4776 p = file_path(f, buf, PAGE_SIZE);
4779 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4781 vma->vm_end - vma->vm_start);
4782 free_page((unsigned long)buf);
4785 up_read(&mm->mmap_sem);
4788 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4789 void __might_fault(const char *file, int line)
4792 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4793 * holding the mmap_sem, this is safe because kernel memory doesn't
4794 * get paged out, therefore we'll never actually fault, and the
4795 * below annotations will generate false positives.
4797 if (uaccess_kernel())
4799 if (pagefault_disabled())
4801 __might_sleep(file, line, 0);
4802 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4804 might_lock_read(¤t->mm->mmap_sem);
4807 EXPORT_SYMBOL(__might_fault);
4810 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4812 * Process all subpages of the specified huge page with the specified
4813 * operation. The target subpage will be processed last to keep its
4816 static inline void process_huge_page(
4817 unsigned long addr_hint, unsigned int pages_per_huge_page,
4818 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4822 unsigned long addr = addr_hint &
4823 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4825 /* Process target subpage last to keep its cache lines hot */
4827 n = (addr_hint - addr) / PAGE_SIZE;
4828 if (2 * n <= pages_per_huge_page) {
4829 /* If target subpage in first half of huge page */
4832 /* Process subpages at the end of huge page */
4833 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4835 process_subpage(addr + i * PAGE_SIZE, i, arg);
4838 /* If target subpage in second half of huge page */
4839 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4840 l = pages_per_huge_page - n;
4841 /* Process subpages at the begin of huge page */
4842 for (i = 0; i < base; i++) {
4844 process_subpage(addr + i * PAGE_SIZE, i, arg);
4848 * Process remaining subpages in left-right-left-right pattern
4849 * towards the target subpage
4851 for (i = 0; i < l; i++) {
4852 int left_idx = base + i;
4853 int right_idx = base + 2 * l - 1 - i;
4856 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4858 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4862 static void clear_gigantic_page(struct page *page,
4864 unsigned int pages_per_huge_page)
4867 struct page *p = page;
4870 for (i = 0; i < pages_per_huge_page;
4871 i++, p = mem_map_next(p, page, i)) {
4873 clear_user_highpage(p, addr + i * PAGE_SIZE);
4877 static void clear_subpage(unsigned long addr, int idx, void *arg)
4879 struct page *page = arg;
4881 clear_user_highpage(page + idx, addr);
4884 void clear_huge_page(struct page *page,
4885 unsigned long addr_hint, unsigned int pages_per_huge_page)
4887 unsigned long addr = addr_hint &
4888 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4890 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4891 clear_gigantic_page(page, addr, pages_per_huge_page);
4895 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4898 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4900 struct vm_area_struct *vma,
4901 unsigned int pages_per_huge_page)
4904 struct page *dst_base = dst;
4905 struct page *src_base = src;
4907 for (i = 0; i < pages_per_huge_page; ) {
4909 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4912 dst = mem_map_next(dst, dst_base, i);
4913 src = mem_map_next(src, src_base, i);
4917 struct copy_subpage_arg {
4920 struct vm_area_struct *vma;
4923 static void copy_subpage(unsigned long addr, int idx, void *arg)
4925 struct copy_subpage_arg *copy_arg = arg;
4927 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4928 addr, copy_arg->vma);
4931 void copy_user_huge_page(struct page *dst, struct page *src,
4932 unsigned long addr_hint, struct vm_area_struct *vma,
4933 unsigned int pages_per_huge_page)
4935 unsigned long addr = addr_hint &
4936 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4937 struct copy_subpage_arg arg = {
4943 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4944 copy_user_gigantic_page(dst, src, addr, vma,
4945 pages_per_huge_page);
4949 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4952 long copy_huge_page_from_user(struct page *dst_page,
4953 const void __user *usr_src,
4954 unsigned int pages_per_huge_page,
4955 bool allow_pagefault)
4957 void *src = (void *)usr_src;
4959 unsigned long i, rc = 0;
4960 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4961 struct page *subpage = dst_page;
4963 for (i = 0; i < pages_per_huge_page;
4964 i++, subpage = mem_map_next(subpage, dst_page, i)) {
4965 if (allow_pagefault)
4966 page_kaddr = kmap(subpage);
4968 page_kaddr = kmap_atomic(subpage);
4969 rc = copy_from_user(page_kaddr,
4970 (const void __user *)(src + i * PAGE_SIZE),
4972 if (allow_pagefault)
4975 kunmap_atomic(page_kaddr);
4977 ret_val -= (PAGE_SIZE - rc);
4985 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4987 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4989 static struct kmem_cache *page_ptl_cachep;
4991 void __init ptlock_cache_init(void)
4993 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4997 bool ptlock_alloc(struct page *page)
5001 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5008 void ptlock_free(struct page *page)
5010 kmem_cache_free(page_ptl_cachep, page->ptl);