4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
93 EXPORT_SYMBOL(mem_map);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
119 static int __init disable_randmaps(char *s)
121 randomize_va_space = 0;
124 __setup("norandmaps", disable_randmaps);
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
129 unsigned long highest_memmap_pfn __read_mostly;
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 static int __init init_zero_pfn(void)
136 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 core_initcall(init_zero_pfn);
142 #if defined(SPLIT_RSS_COUNTING)
144 void sync_mm_rss(struct mm_struct *mm)
148 for (i = 0; i < NR_MM_COUNTERS; i++) {
149 if (current->rss_stat.count[i]) {
150 add_mm_counter(mm, i, current->rss_stat.count[i]);
151 current->rss_stat.count[i] = 0;
154 current->rss_stat.events = 0;
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 struct task_struct *task = current;
161 if (likely(task->mm == mm))
162 task->rss_stat.count[member] += val;
164 add_mm_counter(mm, member, val);
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH (64)
171 static void check_sync_rss_stat(struct task_struct *task)
173 if (unlikely(task != current))
175 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176 sync_mm_rss(task->mm);
178 #else /* SPLIT_RSS_COUNTING */
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183 static void check_sync_rss_stat(struct task_struct *task)
187 #endif /* SPLIT_RSS_COUNTING */
189 #ifdef HAVE_GENERIC_MMU_GATHER
191 static bool tlb_next_batch(struct mmu_gather *tlb)
193 struct mmu_gather_batch *batch;
197 tlb->active = batch->next;
201 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
204 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
211 batch->max = MAX_GATHER_BATCH;
213 tlb->active->next = batch;
219 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220 unsigned long start, unsigned long end)
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238 __tlb_reset_range(tlb);
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 __tlb_reset_range(tlb);
251 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
253 struct mmu_gather_batch *batch;
255 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
256 tlb_table_flush(tlb);
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
262 tlb->active = &tlb->local;
265 void tlb_flush_mmu(struct mmu_gather *tlb)
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276 unsigned long start, unsigned long end, bool force)
278 struct mmu_gather_batch *batch, *next;
281 __tlb_adjust_range(tlb, start, end - start);
285 /* keep the page table cache within bounds */
288 for (batch = tlb->local.next; batch; batch = next) {
290 free_pages((unsigned long)batch, 0);
292 tlb->local.next = NULL;
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
302 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 struct mmu_gather_batch *batch;
306 VM_BUG_ON(!tlb->end);
307 VM_WARN_ON(tlb->page_size != page_size);
311 * Add the page and check if we are full. If so
314 batch->pages[batch->nr++] = page;
315 if (batch->nr == batch->max) {
316 if (!tlb_next_batch(tlb))
320 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
325 #endif /* HAVE_GENERIC_MMU_GATHER */
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
330 * See the comment near struct mmu_table_batch.
333 static void tlb_remove_table_smp_sync(void *arg)
335 /* Simply deliver the interrupt */
338 static void tlb_remove_table_one(void *table)
341 * This isn't an RCU grace period and hence the page-tables cannot be
342 * assumed to be actually RCU-freed.
344 * It is however sufficient for software page-table walkers that rely on
345 * IRQ disabling. See the comment near struct mmu_table_batch.
347 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
348 __tlb_remove_table(table);
351 static void tlb_remove_table_rcu(struct rcu_head *head)
353 struct mmu_table_batch *batch;
356 batch = container_of(head, struct mmu_table_batch, rcu);
358 for (i = 0; i < batch->nr; i++)
359 __tlb_remove_table(batch->tables[i]);
361 free_page((unsigned long)batch);
364 void tlb_table_flush(struct mmu_gather *tlb)
366 struct mmu_table_batch **batch = &tlb->batch;
369 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
374 void tlb_remove_table(struct mmu_gather *tlb, void *table)
376 struct mmu_table_batch **batch = &tlb->batch;
378 if (*batch == NULL) {
379 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
380 if (*batch == NULL) {
381 tlb_remove_table_one(table);
386 (*batch)->tables[(*batch)->nr++] = table;
387 if ((*batch)->nr == MAX_TABLE_BATCH)
388 tlb_table_flush(tlb);
391 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
394 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
395 * @tlb: the mmu_gather structure to initialize
396 * @mm: the mm_struct of the target address space
397 * @start: start of the region that will be removed from the page-table
398 * @end: end of the region that will be removed from the page-table
400 * Called to initialize an (on-stack) mmu_gather structure for page-table
401 * tear-down from @mm. The @start and @end are set to 0 and -1
402 * respectively when @mm is without users and we're going to destroy
403 * the full address space (exit/execve).
405 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
406 unsigned long start, unsigned long end)
408 arch_tlb_gather_mmu(tlb, mm, start, end);
409 inc_tlb_flush_pending(tlb->mm);
412 void tlb_finish_mmu(struct mmu_gather *tlb,
413 unsigned long start, unsigned long end)
416 * If there are parallel threads are doing PTE changes on same range
417 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
418 * flush by batching, a thread has stable TLB entry can fail to flush
419 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
420 * forcefully if we detect parallel PTE batching threads.
422 bool force = mm_tlb_flush_nested(tlb->mm);
424 arch_tlb_finish_mmu(tlb, start, end, force);
425 dec_tlb_flush_pending(tlb->mm);
429 * Note: this doesn't free the actual pages themselves. That
430 * has been handled earlier when unmapping all the memory regions.
432 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
435 pgtable_t token = pmd_pgtable(*pmd);
437 pte_free_tlb(tlb, token, addr);
438 mm_dec_nr_ptes(tlb->mm);
441 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
450 pmd = pmd_offset(pud, addr);
452 next = pmd_addr_end(addr, end);
453 if (pmd_none_or_clear_bad(pmd))
455 free_pte_range(tlb, pmd, addr);
456 } while (pmd++, addr = next, addr != end);
466 if (end - 1 > ceiling - 1)
469 pmd = pmd_offset(pud, start);
471 pmd_free_tlb(tlb, pmd, start);
472 mm_dec_nr_pmds(tlb->mm);
475 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
476 unsigned long addr, unsigned long end,
477 unsigned long floor, unsigned long ceiling)
484 pud = pud_offset(p4d, addr);
486 next = pud_addr_end(addr, end);
487 if (pud_none_or_clear_bad(pud))
489 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
490 } while (pud++, addr = next, addr != end);
500 if (end - 1 > ceiling - 1)
503 pud = pud_offset(p4d, start);
505 pud_free_tlb(tlb, pud, start);
506 mm_dec_nr_puds(tlb->mm);
509 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
510 unsigned long addr, unsigned long end,
511 unsigned long floor, unsigned long ceiling)
518 p4d = p4d_offset(pgd, addr);
520 next = p4d_addr_end(addr, end);
521 if (p4d_none_or_clear_bad(p4d))
523 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
524 } while (p4d++, addr = next, addr != end);
530 ceiling &= PGDIR_MASK;
534 if (end - 1 > ceiling - 1)
537 p4d = p4d_offset(pgd, start);
539 p4d_free_tlb(tlb, p4d, start);
543 * This function frees user-level page tables of a process.
545 void free_pgd_range(struct mmu_gather *tlb,
546 unsigned long addr, unsigned long end,
547 unsigned long floor, unsigned long ceiling)
553 * The next few lines have given us lots of grief...
555 * Why are we testing PMD* at this top level? Because often
556 * there will be no work to do at all, and we'd prefer not to
557 * go all the way down to the bottom just to discover that.
559 * Why all these "- 1"s? Because 0 represents both the bottom
560 * of the address space and the top of it (using -1 for the
561 * top wouldn't help much: the masks would do the wrong thing).
562 * The rule is that addr 0 and floor 0 refer to the bottom of
563 * the address space, but end 0 and ceiling 0 refer to the top
564 * Comparisons need to use "end - 1" and "ceiling - 1" (though
565 * that end 0 case should be mythical).
567 * Wherever addr is brought up or ceiling brought down, we must
568 * be careful to reject "the opposite 0" before it confuses the
569 * subsequent tests. But what about where end is brought down
570 * by PMD_SIZE below? no, end can't go down to 0 there.
572 * Whereas we round start (addr) and ceiling down, by different
573 * masks at different levels, in order to test whether a table
574 * now has no other vmas using it, so can be freed, we don't
575 * bother to round floor or end up - the tests don't need that.
589 if (end - 1 > ceiling - 1)
594 * We add page table cache pages with PAGE_SIZE,
595 * (see pte_free_tlb()), flush the tlb if we need
597 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
598 pgd = pgd_offset(tlb->mm, addr);
600 next = pgd_addr_end(addr, end);
601 if (pgd_none_or_clear_bad(pgd))
603 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
604 } while (pgd++, addr = next, addr != end);
607 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
608 unsigned long floor, unsigned long ceiling)
611 struct vm_area_struct *next = vma->vm_next;
612 unsigned long addr = vma->vm_start;
615 * Hide vma from rmap and truncate_pagecache before freeing
618 unlink_anon_vmas(vma);
619 unlink_file_vma(vma);
621 if (is_vm_hugetlb_page(vma)) {
622 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
623 floor, next ? next->vm_start : ceiling);
626 * Optimization: gather nearby vmas into one call down
628 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
629 && !is_vm_hugetlb_page(next)) {
632 unlink_anon_vmas(vma);
633 unlink_file_vma(vma);
635 free_pgd_range(tlb, addr, vma->vm_end,
636 floor, next ? next->vm_start : ceiling);
642 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
645 pgtable_t new = pte_alloc_one(mm, address);
650 * Ensure all pte setup (eg. pte page lock and page clearing) are
651 * visible before the pte is made visible to other CPUs by being
652 * put into page tables.
654 * The other side of the story is the pointer chasing in the page
655 * table walking code (when walking the page table without locking;
656 * ie. most of the time). Fortunately, these data accesses consist
657 * of a chain of data-dependent loads, meaning most CPUs (alpha
658 * being the notable exception) will already guarantee loads are
659 * seen in-order. See the alpha page table accessors for the
660 * smp_read_barrier_depends() barriers in page table walking code.
662 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
664 ptl = pmd_lock(mm, pmd);
665 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
667 pmd_populate(mm, pmd, new);
676 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
678 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
682 smp_wmb(); /* See comment in __pte_alloc */
684 spin_lock(&init_mm.page_table_lock);
685 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
686 pmd_populate_kernel(&init_mm, pmd, new);
689 spin_unlock(&init_mm.page_table_lock);
691 pte_free_kernel(&init_mm, new);
695 static inline void init_rss_vec(int *rss)
697 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
700 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
704 if (current->mm == mm)
706 for (i = 0; i < NR_MM_COUNTERS; i++)
708 add_mm_counter(mm, i, rss[i]);
712 * This function is called to print an error when a bad pte
713 * is found. For example, we might have a PFN-mapped pte in
714 * a region that doesn't allow it.
716 * The calling function must still handle the error.
718 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
719 pte_t pte, struct page *page)
721 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
722 p4d_t *p4d = p4d_offset(pgd, addr);
723 pud_t *pud = pud_offset(p4d, addr);
724 pmd_t *pmd = pmd_offset(pud, addr);
725 struct address_space *mapping;
727 static unsigned long resume;
728 static unsigned long nr_shown;
729 static unsigned long nr_unshown;
732 * Allow a burst of 60 reports, then keep quiet for that minute;
733 * or allow a steady drip of one report per second.
735 if (nr_shown == 60) {
736 if (time_before(jiffies, resume)) {
741 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
748 resume = jiffies + 60 * HZ;
750 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
751 index = linear_page_index(vma, addr);
753 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
755 (long long)pte_val(pte), (long long)pmd_val(*pmd));
757 dump_page(page, "bad pte");
758 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
759 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
760 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
762 vma->vm_ops ? vma->vm_ops->fault : NULL,
763 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
764 mapping ? mapping->a_ops->readpage : NULL);
766 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
770 * vm_normal_page -- This function gets the "struct page" associated with a pte.
772 * "Special" mappings do not wish to be associated with a "struct page" (either
773 * it doesn't exist, or it exists but they don't want to touch it). In this
774 * case, NULL is returned here. "Normal" mappings do have a struct page.
776 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
777 * pte bit, in which case this function is trivial. Secondly, an architecture
778 * may not have a spare pte bit, which requires a more complicated scheme,
781 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
782 * special mapping (even if there are underlying and valid "struct pages").
783 * COWed pages of a VM_PFNMAP are always normal.
785 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
786 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
787 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
788 * mapping will always honor the rule
790 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
792 * And for normal mappings this is false.
794 * This restricts such mappings to be a linear translation from virtual address
795 * to pfn. To get around this restriction, we allow arbitrary mappings so long
796 * as the vma is not a COW mapping; in that case, we know that all ptes are
797 * special (because none can have been COWed).
800 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
802 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
803 * page" backing, however the difference is that _all_ pages with a struct
804 * page (that is, those where pfn_valid is true) are refcounted and considered
805 * normal pages by the VM. The disadvantage is that pages are refcounted
806 * (which can be slower and simply not an option for some PFNMAP users). The
807 * advantage is that we don't have to follow the strict linearity rule of
808 * PFNMAP mappings in order to support COWable mappings.
811 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
812 pte_t pte, bool with_public_device)
814 unsigned long pfn = pte_pfn(pte);
816 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
817 if (likely(!pte_special(pte)))
819 if (vma->vm_ops && vma->vm_ops->find_special_page)
820 return vma->vm_ops->find_special_page(vma, addr);
821 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
823 if (is_zero_pfn(pfn))
827 * Device public pages are special pages (they are ZONE_DEVICE
828 * pages but different from persistent memory). They behave
829 * allmost like normal pages. The difference is that they are
830 * not on the lru and thus should never be involve with any-
831 * thing that involve lru manipulation (mlock, numa balancing,
834 * This is why we still want to return NULL for such page from
835 * vm_normal_page() so that we do not have to special case all
836 * call site of vm_normal_page().
838 if (likely(pfn <= highest_memmap_pfn)) {
839 struct page *page = pfn_to_page(pfn);
841 if (is_device_public_page(page)) {
842 if (with_public_device)
851 print_bad_pte(vma, addr, pte, NULL);
855 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
857 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
858 if (vma->vm_flags & VM_MIXEDMAP) {
864 off = (addr - vma->vm_start) >> PAGE_SHIFT;
865 if (pfn == vma->vm_pgoff + off)
867 if (!is_cow_mapping(vma->vm_flags))
872 if (is_zero_pfn(pfn))
876 if (unlikely(pfn > highest_memmap_pfn)) {
877 print_bad_pte(vma, addr, pte, NULL);
882 * NOTE! We still have PageReserved() pages in the page tables.
883 * eg. VDSO mappings can cause them to exist.
886 return pfn_to_page(pfn);
889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
890 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
893 unsigned long pfn = pmd_pfn(pmd);
896 * There is no pmd_special() but there may be special pmds, e.g.
897 * in a direct-access (dax) mapping, so let's just replicate the
898 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
900 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
901 if (vma->vm_flags & VM_MIXEDMAP) {
907 off = (addr - vma->vm_start) >> PAGE_SHIFT;
908 if (pfn == vma->vm_pgoff + off)
910 if (!is_cow_mapping(vma->vm_flags))
917 if (is_zero_pfn(pfn))
919 if (unlikely(pfn > highest_memmap_pfn))
923 * NOTE! We still have PageReserved() pages in the page tables.
924 * eg. VDSO mappings can cause them to exist.
927 return pfn_to_page(pfn);
932 * copy one vm_area from one task to the other. Assumes the page tables
933 * already present in the new task to be cleared in the whole range
934 * covered by this vma.
937 static inline unsigned long
938 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
939 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
940 unsigned long addr, int *rss)
942 unsigned long vm_flags = vma->vm_flags;
943 pte_t pte = *src_pte;
946 /* pte contains position in swap or file, so copy. */
947 if (unlikely(!pte_present(pte))) {
948 swp_entry_t entry = pte_to_swp_entry(pte);
950 if (likely(!non_swap_entry(entry))) {
951 if (swap_duplicate(entry) < 0)
954 /* make sure dst_mm is on swapoff's mmlist. */
955 if (unlikely(list_empty(&dst_mm->mmlist))) {
956 spin_lock(&mmlist_lock);
957 if (list_empty(&dst_mm->mmlist))
958 list_add(&dst_mm->mmlist,
960 spin_unlock(&mmlist_lock);
963 } else if (is_migration_entry(entry)) {
964 page = migration_entry_to_page(entry);
966 rss[mm_counter(page)]++;
968 if (is_write_migration_entry(entry) &&
969 is_cow_mapping(vm_flags)) {
971 * COW mappings require pages in both
972 * parent and child to be set to read.
974 make_migration_entry_read(&entry);
975 pte = swp_entry_to_pte(entry);
976 if (pte_swp_soft_dirty(*src_pte))
977 pte = pte_swp_mksoft_dirty(pte);
978 set_pte_at(src_mm, addr, src_pte, pte);
980 } else if (is_device_private_entry(entry)) {
981 page = device_private_entry_to_page(entry);
984 * Update rss count even for unaddressable pages, as
985 * they should treated just like normal pages in this
988 * We will likely want to have some new rss counters
989 * for unaddressable pages, at some point. But for now
990 * keep things as they are.
993 rss[mm_counter(page)]++;
994 page_dup_rmap(page, false);
997 * We do not preserve soft-dirty information, because so
998 * far, checkpoint/restore is the only feature that
999 * requires that. And checkpoint/restore does not work
1000 * when a device driver is involved (you cannot easily
1001 * save and restore device driver state).
1003 if (is_write_device_private_entry(entry) &&
1004 is_cow_mapping(vm_flags)) {
1005 make_device_private_entry_read(&entry);
1006 pte = swp_entry_to_pte(entry);
1007 set_pte_at(src_mm, addr, src_pte, pte);
1014 * If it's a COW mapping, write protect it both
1015 * in the parent and the child
1017 if (is_cow_mapping(vm_flags)) {
1018 ptep_set_wrprotect(src_mm, addr, src_pte);
1019 pte = pte_wrprotect(pte);
1023 * If it's a shared mapping, mark it clean in
1026 if (vm_flags & VM_SHARED)
1027 pte = pte_mkclean(pte);
1028 pte = pte_mkold(pte);
1030 page = vm_normal_page(vma, addr, pte);
1033 page_dup_rmap(page, false);
1034 rss[mm_counter(page)]++;
1035 } else if (pte_devmap(pte)) {
1036 page = pte_page(pte);
1039 * Cache coherent device memory behave like regular page and
1040 * not like persistent memory page. For more informations see
1041 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1043 if (is_device_public_page(page)) {
1045 page_dup_rmap(page, false);
1046 rss[mm_counter(page)]++;
1051 set_pte_at(dst_mm, addr, dst_pte, pte);
1055 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1056 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1057 unsigned long addr, unsigned long end)
1059 pte_t *orig_src_pte, *orig_dst_pte;
1060 pte_t *src_pte, *dst_pte;
1061 spinlock_t *src_ptl, *dst_ptl;
1063 int rss[NR_MM_COUNTERS];
1064 swp_entry_t entry = (swp_entry_t){0};
1069 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1072 src_pte = pte_offset_map(src_pmd, addr);
1073 src_ptl = pte_lockptr(src_mm, src_pmd);
1074 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1075 orig_src_pte = src_pte;
1076 orig_dst_pte = dst_pte;
1077 arch_enter_lazy_mmu_mode();
1081 * We are holding two locks at this point - either of them
1082 * could generate latencies in another task on another CPU.
1084 if (progress >= 32) {
1086 if (need_resched() ||
1087 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1090 if (pte_none(*src_pte)) {
1094 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1099 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1101 arch_leave_lazy_mmu_mode();
1102 spin_unlock(src_ptl);
1103 pte_unmap(orig_src_pte);
1104 add_mm_rss_vec(dst_mm, rss);
1105 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1109 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1118 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1119 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1120 unsigned long addr, unsigned long end)
1122 pmd_t *src_pmd, *dst_pmd;
1125 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1128 src_pmd = pmd_offset(src_pud, addr);
1130 next = pmd_addr_end(addr, end);
1131 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1132 || pmd_devmap(*src_pmd)) {
1134 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1135 err = copy_huge_pmd(dst_mm, src_mm,
1136 dst_pmd, src_pmd, addr, vma);
1143 if (pmd_none_or_clear_bad(src_pmd))
1145 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1148 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1152 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1153 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1154 unsigned long addr, unsigned long end)
1156 pud_t *src_pud, *dst_pud;
1159 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1162 src_pud = pud_offset(src_p4d, addr);
1164 next = pud_addr_end(addr, end);
1165 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1168 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1169 err = copy_huge_pud(dst_mm, src_mm,
1170 dst_pud, src_pud, addr, vma);
1177 if (pud_none_or_clear_bad(src_pud))
1179 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1182 } while (dst_pud++, src_pud++, addr = next, addr != end);
1186 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1187 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1188 unsigned long addr, unsigned long end)
1190 p4d_t *src_p4d, *dst_p4d;
1193 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1196 src_p4d = p4d_offset(src_pgd, addr);
1198 next = p4d_addr_end(addr, end);
1199 if (p4d_none_or_clear_bad(src_p4d))
1201 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1204 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1208 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1209 struct vm_area_struct *vma)
1211 pgd_t *src_pgd, *dst_pgd;
1213 unsigned long addr = vma->vm_start;
1214 unsigned long end = vma->vm_end;
1215 unsigned long mmun_start; /* For mmu_notifiers */
1216 unsigned long mmun_end; /* For mmu_notifiers */
1221 * Don't copy ptes where a page fault will fill them correctly.
1222 * Fork becomes much lighter when there are big shared or private
1223 * readonly mappings. The tradeoff is that copy_page_range is more
1224 * efficient than faulting.
1226 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1230 if (is_vm_hugetlb_page(vma))
1231 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1233 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1235 * We do not free on error cases below as remove_vma
1236 * gets called on error from higher level routine
1238 ret = track_pfn_copy(vma);
1244 * We need to invalidate the secondary MMU mappings only when
1245 * there could be a permission downgrade on the ptes of the
1246 * parent mm. And a permission downgrade will only happen if
1247 * is_cow_mapping() returns true.
1249 is_cow = is_cow_mapping(vma->vm_flags);
1253 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1257 dst_pgd = pgd_offset(dst_mm, addr);
1258 src_pgd = pgd_offset(src_mm, addr);
1260 next = pgd_addr_end(addr, end);
1261 if (pgd_none_or_clear_bad(src_pgd))
1263 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1264 vma, addr, next))) {
1268 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1271 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1275 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1276 struct vm_area_struct *vma, pmd_t *pmd,
1277 unsigned long addr, unsigned long end,
1278 struct zap_details *details)
1280 struct mm_struct *mm = tlb->mm;
1281 int force_flush = 0;
1282 int rss[NR_MM_COUNTERS];
1288 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1291 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1293 flush_tlb_batched_pending(mm);
1294 arch_enter_lazy_mmu_mode();
1297 if (pte_none(ptent))
1300 if (pte_present(ptent)) {
1303 page = _vm_normal_page(vma, addr, ptent, true);
1304 if (unlikely(details) && page) {
1306 * unmap_shared_mapping_pages() wants to
1307 * invalidate cache without truncating:
1308 * unmap shared but keep private pages.
1310 if (details->check_mapping &&
1311 details->check_mapping != page_rmapping(page))
1314 ptent = ptep_get_and_clear_full(mm, addr, pte,
1316 tlb_remove_tlb_entry(tlb, pte, addr);
1317 if (unlikely(!page))
1320 if (!PageAnon(page)) {
1321 if (pte_dirty(ptent)) {
1323 set_page_dirty(page);
1325 if (pte_young(ptent) &&
1326 likely(!(vma->vm_flags & VM_SEQ_READ)))
1327 mark_page_accessed(page);
1329 rss[mm_counter(page)]--;
1330 page_remove_rmap(page, false);
1331 if (unlikely(page_mapcount(page) < 0))
1332 print_bad_pte(vma, addr, ptent, page);
1333 if (unlikely(__tlb_remove_page(tlb, page))) {
1341 entry = pte_to_swp_entry(ptent);
1342 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1343 struct page *page = device_private_entry_to_page(entry);
1345 if (unlikely(details && details->check_mapping)) {
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 page_rmapping(page))
1356 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1357 rss[mm_counter(page)]--;
1358 page_remove_rmap(page, false);
1363 /* If details->check_mapping, we leave swap entries. */
1364 if (unlikely(details))
1367 entry = pte_to_swp_entry(ptent);
1368 if (!non_swap_entry(entry))
1370 else if (is_migration_entry(entry)) {
1373 page = migration_entry_to_page(entry);
1374 rss[mm_counter(page)]--;
1376 if (unlikely(!free_swap_and_cache(entry)))
1377 print_bad_pte(vma, addr, ptent, NULL);
1378 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1379 } while (pte++, addr += PAGE_SIZE, addr != end);
1381 add_mm_rss_vec(mm, rss);
1382 arch_leave_lazy_mmu_mode();
1384 /* Do the actual TLB flush before dropping ptl */
1386 tlb_flush_mmu_tlbonly(tlb);
1387 pte_unmap_unlock(start_pte, ptl);
1390 * If we forced a TLB flush (either due to running out of
1391 * batch buffers or because we needed to flush dirty TLB
1392 * entries before releasing the ptl), free the batched
1393 * memory too. Restart if we didn't do everything.
1397 tlb_flush_mmu_free(tlb);
1405 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1406 struct vm_area_struct *vma, pud_t *pud,
1407 unsigned long addr, unsigned long end,
1408 struct zap_details *details)
1413 pmd = pmd_offset(pud, addr);
1415 next = pmd_addr_end(addr, end);
1416 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1417 if (next - addr != HPAGE_PMD_SIZE)
1418 __split_huge_pmd(vma, pmd, addr, false, NULL);
1419 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1424 * Here there can be other concurrent MADV_DONTNEED or
1425 * trans huge page faults running, and if the pmd is
1426 * none or trans huge it can change under us. This is
1427 * because MADV_DONTNEED holds the mmap_sem in read
1430 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1432 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1435 } while (pmd++, addr = next, addr != end);
1440 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1441 struct vm_area_struct *vma, p4d_t *p4d,
1442 unsigned long addr, unsigned long end,
1443 struct zap_details *details)
1448 pud = pud_offset(p4d, addr);
1450 next = pud_addr_end(addr, end);
1451 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1452 if (next - addr != HPAGE_PUD_SIZE) {
1453 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1454 split_huge_pud(vma, pud, addr);
1455 } else if (zap_huge_pud(tlb, vma, pud, addr))
1459 if (pud_none_or_clear_bad(pud))
1461 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1464 } while (pud++, addr = next, addr != end);
1469 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1470 struct vm_area_struct *vma, pgd_t *pgd,
1471 unsigned long addr, unsigned long end,
1472 struct zap_details *details)
1477 p4d = p4d_offset(pgd, addr);
1479 next = p4d_addr_end(addr, end);
1480 if (p4d_none_or_clear_bad(p4d))
1482 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1483 } while (p4d++, addr = next, addr != end);
1488 void unmap_page_range(struct mmu_gather *tlb,
1489 struct vm_area_struct *vma,
1490 unsigned long addr, unsigned long end,
1491 struct zap_details *details)
1496 BUG_ON(addr >= end);
1497 tlb_start_vma(tlb, vma);
1498 pgd = pgd_offset(vma->vm_mm, addr);
1500 next = pgd_addr_end(addr, end);
1501 if (pgd_none_or_clear_bad(pgd))
1503 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1504 } while (pgd++, addr = next, addr != end);
1505 tlb_end_vma(tlb, vma);
1509 static void unmap_single_vma(struct mmu_gather *tlb,
1510 struct vm_area_struct *vma, unsigned long start_addr,
1511 unsigned long end_addr,
1512 struct zap_details *details)
1514 unsigned long start = max(vma->vm_start, start_addr);
1517 if (start >= vma->vm_end)
1519 end = min(vma->vm_end, end_addr);
1520 if (end <= vma->vm_start)
1524 uprobe_munmap(vma, start, end);
1526 if (unlikely(vma->vm_flags & VM_PFNMAP))
1527 untrack_pfn(vma, 0, 0);
1530 if (unlikely(is_vm_hugetlb_page(vma))) {
1532 * It is undesirable to test vma->vm_file as it
1533 * should be non-null for valid hugetlb area.
1534 * However, vm_file will be NULL in the error
1535 * cleanup path of mmap_region. When
1536 * hugetlbfs ->mmap method fails,
1537 * mmap_region() nullifies vma->vm_file
1538 * before calling this function to clean up.
1539 * Since no pte has actually been setup, it is
1540 * safe to do nothing in this case.
1543 i_mmap_lock_write(vma->vm_file->f_mapping);
1544 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1545 i_mmap_unlock_write(vma->vm_file->f_mapping);
1548 unmap_page_range(tlb, vma, start, end, details);
1553 * unmap_vmas - unmap a range of memory covered by a list of vma's
1554 * @tlb: address of the caller's struct mmu_gather
1555 * @vma: the starting vma
1556 * @start_addr: virtual address at which to start unmapping
1557 * @end_addr: virtual address at which to end unmapping
1559 * Unmap all pages in the vma list.
1561 * Only addresses between `start' and `end' will be unmapped.
1563 * The VMA list must be sorted in ascending virtual address order.
1565 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1566 * range after unmap_vmas() returns. So the only responsibility here is to
1567 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1568 * drops the lock and schedules.
1570 void unmap_vmas(struct mmu_gather *tlb,
1571 struct vm_area_struct *vma, unsigned long start_addr,
1572 unsigned long end_addr)
1574 struct mm_struct *mm = vma->vm_mm;
1576 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1577 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1578 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1579 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1583 * zap_page_range - remove user pages in a given range
1584 * @vma: vm_area_struct holding the applicable pages
1585 * @start: starting address of pages to zap
1586 * @size: number of bytes to zap
1588 * Caller must protect the VMA list
1590 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1593 struct mm_struct *mm = vma->vm_mm;
1594 struct mmu_gather tlb;
1595 unsigned long end = start + size;
1598 tlb_gather_mmu(&tlb, mm, start, end);
1599 update_hiwater_rss(mm);
1600 mmu_notifier_invalidate_range_start(mm, start, end);
1601 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1602 unmap_single_vma(&tlb, vma, start, end, NULL);
1603 mmu_notifier_invalidate_range_end(mm, start, end);
1604 tlb_finish_mmu(&tlb, start, end);
1608 * zap_page_range_single - remove user pages in a given range
1609 * @vma: vm_area_struct holding the applicable pages
1610 * @address: starting address of pages to zap
1611 * @size: number of bytes to zap
1612 * @details: details of shared cache invalidation
1614 * The range must fit into one VMA.
1616 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1617 unsigned long size, struct zap_details *details)
1619 struct mm_struct *mm = vma->vm_mm;
1620 struct mmu_gather tlb;
1621 unsigned long end = address + size;
1624 tlb_gather_mmu(&tlb, mm, address, end);
1625 update_hiwater_rss(mm);
1626 mmu_notifier_invalidate_range_start(mm, address, end);
1627 unmap_single_vma(&tlb, vma, address, end, details);
1628 mmu_notifier_invalidate_range_end(mm, address, end);
1629 tlb_finish_mmu(&tlb, address, end);
1633 * zap_vma_ptes - remove ptes mapping the vma
1634 * @vma: vm_area_struct holding ptes to be zapped
1635 * @address: starting address of pages to zap
1636 * @size: number of bytes to zap
1638 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1640 * The entire address range must be fully contained within the vma.
1643 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1646 if (address < vma->vm_start || address + size > vma->vm_end ||
1647 !(vma->vm_flags & VM_PFNMAP))
1650 zap_page_range_single(vma, address, size, NULL);
1652 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1654 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1662 pgd = pgd_offset(mm, addr);
1663 p4d = p4d_alloc(mm, pgd, addr);
1666 pud = pud_alloc(mm, p4d, addr);
1669 pmd = pmd_alloc(mm, pud, addr);
1673 VM_BUG_ON(pmd_trans_huge(*pmd));
1674 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1678 * This is the old fallback for page remapping.
1680 * For historical reasons, it only allows reserved pages. Only
1681 * old drivers should use this, and they needed to mark their
1682 * pages reserved for the old functions anyway.
1684 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1685 struct page *page, pgprot_t prot)
1687 struct mm_struct *mm = vma->vm_mm;
1696 flush_dcache_page(page);
1697 pte = get_locked_pte(mm, addr, &ptl);
1701 if (!pte_none(*pte))
1704 /* Ok, finally just insert the thing.. */
1706 inc_mm_counter_fast(mm, mm_counter_file(page));
1707 page_add_file_rmap(page, false);
1708 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1711 pte_unmap_unlock(pte, ptl);
1714 pte_unmap_unlock(pte, ptl);
1720 * vm_insert_page - insert single page into user vma
1721 * @vma: user vma to map to
1722 * @addr: target user address of this page
1723 * @page: source kernel page
1725 * This allows drivers to insert individual pages they've allocated
1728 * The page has to be a nice clean _individual_ kernel allocation.
1729 * If you allocate a compound page, you need to have marked it as
1730 * such (__GFP_COMP), or manually just split the page up yourself
1731 * (see split_page()).
1733 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1734 * took an arbitrary page protection parameter. This doesn't allow
1735 * that. Your vma protection will have to be set up correctly, which
1736 * means that if you want a shared writable mapping, you'd better
1737 * ask for a shared writable mapping!
1739 * The page does not need to be reserved.
1741 * Usually this function is called from f_op->mmap() handler
1742 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1743 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1744 * function from other places, for example from page-fault handler.
1746 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1749 if (addr < vma->vm_start || addr >= vma->vm_end)
1751 if (!page_count(page))
1753 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1754 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1755 BUG_ON(vma->vm_flags & VM_PFNMAP);
1756 vma->vm_flags |= VM_MIXEDMAP;
1758 return insert_page(vma, addr, page, vma->vm_page_prot);
1760 EXPORT_SYMBOL(vm_insert_page);
1762 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1763 pfn_t pfn, pgprot_t prot, bool mkwrite)
1765 struct mm_struct *mm = vma->vm_mm;
1771 pte = get_locked_pte(mm, addr, &ptl);
1775 if (!pte_none(*pte)) {
1778 * For read faults on private mappings the PFN passed
1779 * in may not match the PFN we have mapped if the
1780 * mapped PFN is a writeable COW page. In the mkwrite
1781 * case we are creating a writable PTE for a shared
1782 * mapping and we expect the PFNs to match.
1784 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1792 /* Ok, finally just insert the thing.. */
1793 if (pfn_t_devmap(pfn))
1794 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1796 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1800 entry = pte_mkyoung(entry);
1801 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1804 set_pte_at(mm, addr, pte, entry);
1805 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1809 pte_unmap_unlock(pte, ptl);
1815 * vm_insert_pfn - insert single pfn into user vma
1816 * @vma: user vma to map to
1817 * @addr: target user address of this page
1818 * @pfn: source kernel pfn
1820 * Similar to vm_insert_page, this allows drivers to insert individual pages
1821 * they've allocated into a user vma. Same comments apply.
1823 * This function should only be called from a vm_ops->fault handler, and
1824 * in that case the handler should return NULL.
1826 * vma cannot be a COW mapping.
1828 * As this is called only for pages that do not currently exist, we
1829 * do not need to flush old virtual caches or the TLB.
1831 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1834 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1836 EXPORT_SYMBOL(vm_insert_pfn);
1839 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1840 * @vma: user vma to map to
1841 * @addr: target user address of this page
1842 * @pfn: source kernel pfn
1843 * @pgprot: pgprot flags for the inserted page
1845 * This is exactly like vm_insert_pfn, except that it allows drivers to
1846 * to override pgprot on a per-page basis.
1848 * This only makes sense for IO mappings, and it makes no sense for
1849 * cow mappings. In general, using multiple vmas is preferable;
1850 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1853 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1854 unsigned long pfn, pgprot_t pgprot)
1858 * Technically, architectures with pte_special can avoid all these
1859 * restrictions (same for remap_pfn_range). However we would like
1860 * consistency in testing and feature parity among all, so we should
1861 * try to keep these invariants in place for everybody.
1863 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1864 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1865 (VM_PFNMAP|VM_MIXEDMAP));
1866 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1867 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1869 if (addr < vma->vm_start || addr >= vma->vm_end)
1872 if (!pfn_modify_allowed(pfn, pgprot))
1875 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1877 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1882 EXPORT_SYMBOL(vm_insert_pfn_prot);
1884 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1886 /* these checks mirror the abort conditions in vm_normal_page */
1887 if (vma->vm_flags & VM_MIXEDMAP)
1889 if (pfn_t_devmap(pfn))
1891 if (pfn_t_special(pfn))
1893 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1898 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1899 pfn_t pfn, bool mkwrite)
1901 pgprot_t pgprot = vma->vm_page_prot;
1903 BUG_ON(!vm_mixed_ok(vma, pfn));
1905 if (addr < vma->vm_start || addr >= vma->vm_end)
1908 track_pfn_insert(vma, &pgprot, pfn);
1910 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1914 * If we don't have pte special, then we have to use the pfn_valid()
1915 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1916 * refcount the page if pfn_valid is true (hence insert_page rather
1917 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1918 * without pte special, it would there be refcounted as a normal page.
1920 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1921 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1925 * At this point we are committed to insert_page()
1926 * regardless of whether the caller specified flags that
1927 * result in pfn_t_has_page() == false.
1929 page = pfn_to_page(pfn_t_to_pfn(pfn));
1930 return insert_page(vma, addr, page, pgprot);
1932 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1935 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1938 return __vm_insert_mixed(vma, addr, pfn, false);
1941 EXPORT_SYMBOL(vm_insert_mixed);
1944 * If the insertion of PTE failed because someone else already added a
1945 * different entry in the mean time, we treat that as success as we assume
1946 * the same entry was actually inserted.
1949 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1950 unsigned long addr, pfn_t pfn)
1954 err = __vm_insert_mixed(vma, addr, pfn, true);
1956 return VM_FAULT_OOM;
1957 if (err < 0 && err != -EBUSY)
1958 return VM_FAULT_SIGBUS;
1959 return VM_FAULT_NOPAGE;
1961 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1964 * maps a range of physical memory into the requested pages. the old
1965 * mappings are removed. any references to nonexistent pages results
1966 * in null mappings (currently treated as "copy-on-access")
1968 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1969 unsigned long addr, unsigned long end,
1970 unsigned long pfn, pgprot_t prot)
1976 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1979 arch_enter_lazy_mmu_mode();
1981 BUG_ON(!pte_none(*pte));
1982 if (!pfn_modify_allowed(pfn, prot)) {
1986 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1988 } while (pte++, addr += PAGE_SIZE, addr != end);
1989 arch_leave_lazy_mmu_mode();
1990 pte_unmap_unlock(pte - 1, ptl);
1994 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1995 unsigned long addr, unsigned long end,
1996 unsigned long pfn, pgprot_t prot)
2002 pfn -= addr >> PAGE_SHIFT;
2003 pmd = pmd_alloc(mm, pud, addr);
2006 VM_BUG_ON(pmd_trans_huge(*pmd));
2008 next = pmd_addr_end(addr, end);
2009 err = remap_pte_range(mm, pmd, addr, next,
2010 pfn + (addr >> PAGE_SHIFT), prot);
2013 } while (pmd++, addr = next, addr != end);
2017 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2018 unsigned long addr, unsigned long end,
2019 unsigned long pfn, pgprot_t prot)
2025 pfn -= addr >> PAGE_SHIFT;
2026 pud = pud_alloc(mm, p4d, addr);
2030 next = pud_addr_end(addr, end);
2031 err = remap_pmd_range(mm, pud, addr, next,
2032 pfn + (addr >> PAGE_SHIFT), prot);
2035 } while (pud++, addr = next, addr != end);
2039 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2040 unsigned long addr, unsigned long end,
2041 unsigned long pfn, pgprot_t prot)
2047 pfn -= addr >> PAGE_SHIFT;
2048 p4d = p4d_alloc(mm, pgd, addr);
2052 next = p4d_addr_end(addr, end);
2053 err = remap_pud_range(mm, p4d, addr, next,
2054 pfn + (addr >> PAGE_SHIFT), prot);
2057 } while (p4d++, addr = next, addr != end);
2062 * remap_pfn_range - remap kernel memory to userspace
2063 * @vma: user vma to map to
2064 * @addr: target user address to start at
2065 * @pfn: physical address of kernel memory
2066 * @size: size of map area
2067 * @prot: page protection flags for this mapping
2069 * Note: this is only safe if the mm semaphore is held when called.
2071 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2072 unsigned long pfn, unsigned long size, pgprot_t prot)
2076 unsigned long end = addr + PAGE_ALIGN(size);
2077 struct mm_struct *mm = vma->vm_mm;
2078 unsigned long remap_pfn = pfn;
2082 * Physically remapped pages are special. Tell the
2083 * rest of the world about it:
2084 * VM_IO tells people not to look at these pages
2085 * (accesses can have side effects).
2086 * VM_PFNMAP tells the core MM that the base pages are just
2087 * raw PFN mappings, and do not have a "struct page" associated
2090 * Disable vma merging and expanding with mremap().
2092 * Omit vma from core dump, even when VM_IO turned off.
2094 * There's a horrible special case to handle copy-on-write
2095 * behaviour that some programs depend on. We mark the "original"
2096 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2097 * See vm_normal_page() for details.
2099 if (is_cow_mapping(vma->vm_flags)) {
2100 if (addr != vma->vm_start || end != vma->vm_end)
2102 vma->vm_pgoff = pfn;
2105 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2109 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2111 BUG_ON(addr >= end);
2112 pfn -= addr >> PAGE_SHIFT;
2113 pgd = pgd_offset(mm, addr);
2114 flush_cache_range(vma, addr, end);
2116 next = pgd_addr_end(addr, end);
2117 err = remap_p4d_range(mm, pgd, addr, next,
2118 pfn + (addr >> PAGE_SHIFT), prot);
2121 } while (pgd++, addr = next, addr != end);
2124 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2128 EXPORT_SYMBOL(remap_pfn_range);
2131 * vm_iomap_memory - remap memory to userspace
2132 * @vma: user vma to map to
2133 * @start: start of area
2134 * @len: size of area
2136 * This is a simplified io_remap_pfn_range() for common driver use. The
2137 * driver just needs to give us the physical memory range to be mapped,
2138 * we'll figure out the rest from the vma information.
2140 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2141 * whatever write-combining details or similar.
2143 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2145 unsigned long vm_len, pfn, pages;
2147 /* Check that the physical memory area passed in looks valid */
2148 if (start + len < start)
2151 * You *really* shouldn't map things that aren't page-aligned,
2152 * but we've historically allowed it because IO memory might
2153 * just have smaller alignment.
2155 len += start & ~PAGE_MASK;
2156 pfn = start >> PAGE_SHIFT;
2157 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2158 if (pfn + pages < pfn)
2161 /* We start the mapping 'vm_pgoff' pages into the area */
2162 if (vma->vm_pgoff > pages)
2164 pfn += vma->vm_pgoff;
2165 pages -= vma->vm_pgoff;
2167 /* Can we fit all of the mapping? */
2168 vm_len = vma->vm_end - vma->vm_start;
2169 if (vm_len >> PAGE_SHIFT > pages)
2172 /* Ok, let it rip */
2173 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2175 EXPORT_SYMBOL(vm_iomap_memory);
2177 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2178 unsigned long addr, unsigned long end,
2179 pte_fn_t fn, void *data)
2184 spinlock_t *uninitialized_var(ptl);
2186 pte = (mm == &init_mm) ?
2187 pte_alloc_kernel(pmd, addr) :
2188 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2192 BUG_ON(pmd_huge(*pmd));
2194 arch_enter_lazy_mmu_mode();
2196 token = pmd_pgtable(*pmd);
2199 err = fn(pte++, token, addr, data);
2202 } while (addr += PAGE_SIZE, addr != end);
2204 arch_leave_lazy_mmu_mode();
2207 pte_unmap_unlock(pte-1, ptl);
2211 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2212 unsigned long addr, unsigned long end,
2213 pte_fn_t fn, void *data)
2219 BUG_ON(pud_huge(*pud));
2221 pmd = pmd_alloc(mm, pud, addr);
2225 next = pmd_addr_end(addr, end);
2226 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2229 } while (pmd++, addr = next, addr != end);
2233 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2234 unsigned long addr, unsigned long end,
2235 pte_fn_t fn, void *data)
2241 pud = pud_alloc(mm, p4d, addr);
2245 next = pud_addr_end(addr, end);
2246 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2249 } while (pud++, addr = next, addr != end);
2253 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2254 unsigned long addr, unsigned long end,
2255 pte_fn_t fn, void *data)
2261 p4d = p4d_alloc(mm, pgd, addr);
2265 next = p4d_addr_end(addr, end);
2266 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2269 } while (p4d++, addr = next, addr != end);
2274 * Scan a region of virtual memory, filling in page tables as necessary
2275 * and calling a provided function on each leaf page table.
2277 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2278 unsigned long size, pte_fn_t fn, void *data)
2282 unsigned long end = addr + size;
2285 if (WARN_ON(addr >= end))
2288 pgd = pgd_offset(mm, addr);
2290 next = pgd_addr_end(addr, end);
2291 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2294 } while (pgd++, addr = next, addr != end);
2298 EXPORT_SYMBOL_GPL(apply_to_page_range);
2301 * handle_pte_fault chooses page fault handler according to an entry which was
2302 * read non-atomically. Before making any commitment, on those architectures
2303 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2304 * parts, do_swap_page must check under lock before unmapping the pte and
2305 * proceeding (but do_wp_page is only called after already making such a check;
2306 * and do_anonymous_page can safely check later on).
2308 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2309 pte_t *page_table, pte_t orig_pte)
2312 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2313 if (sizeof(pte_t) > sizeof(unsigned long)) {
2314 spinlock_t *ptl = pte_lockptr(mm, pmd);
2316 same = pte_same(*page_table, orig_pte);
2320 pte_unmap(page_table);
2324 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2326 debug_dma_assert_idle(src);
2329 * If the source page was a PFN mapping, we don't have
2330 * a "struct page" for it. We do a best-effort copy by
2331 * just copying from the original user address. If that
2332 * fails, we just zero-fill it. Live with it.
2334 if (unlikely(!src)) {
2335 void *kaddr = kmap_atomic(dst);
2336 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2339 * This really shouldn't fail, because the page is there
2340 * in the page tables. But it might just be unreadable,
2341 * in which case we just give up and fill the result with
2344 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2346 kunmap_atomic(kaddr);
2347 flush_dcache_page(dst);
2349 copy_user_highpage(dst, src, va, vma);
2352 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2354 struct file *vm_file = vma->vm_file;
2357 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2360 * Special mappings (e.g. VDSO) do not have any file so fake
2361 * a default GFP_KERNEL for them.
2367 * Notify the address space that the page is about to become writable so that
2368 * it can prohibit this or wait for the page to get into an appropriate state.
2370 * We do this without the lock held, so that it can sleep if it needs to.
2372 static int do_page_mkwrite(struct vm_fault *vmf)
2375 struct page *page = vmf->page;
2376 unsigned int old_flags = vmf->flags;
2378 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2380 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2381 /* Restore original flags so that caller is not surprised */
2382 vmf->flags = old_flags;
2383 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2385 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2387 if (!page->mapping) {
2389 return 0; /* retry */
2391 ret |= VM_FAULT_LOCKED;
2393 VM_BUG_ON_PAGE(!PageLocked(page), page);
2398 * Handle dirtying of a page in shared file mapping on a write fault.
2400 * The function expects the page to be locked and unlocks it.
2402 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2405 struct address_space *mapping;
2407 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2409 dirtied = set_page_dirty(page);
2410 VM_BUG_ON_PAGE(PageAnon(page), page);
2412 * Take a local copy of the address_space - page.mapping may be zeroed
2413 * by truncate after unlock_page(). The address_space itself remains
2414 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2415 * release semantics to prevent the compiler from undoing this copying.
2417 mapping = page_rmapping(page);
2420 if ((dirtied || page_mkwrite) && mapping) {
2422 * Some device drivers do not set page.mapping
2423 * but still dirty their pages
2425 balance_dirty_pages_ratelimited(mapping);
2429 file_update_time(vma->vm_file);
2433 * Handle write page faults for pages that can be reused in the current vma
2435 * This can happen either due to the mapping being with the VM_SHARED flag,
2436 * or due to us being the last reference standing to the page. In either
2437 * case, all we need to do here is to mark the page as writable and update
2438 * any related book-keeping.
2440 static inline void wp_page_reuse(struct vm_fault *vmf)
2441 __releases(vmf->ptl)
2443 struct vm_area_struct *vma = vmf->vma;
2444 struct page *page = vmf->page;
2447 * Clear the pages cpupid information as the existing
2448 * information potentially belongs to a now completely
2449 * unrelated process.
2452 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2454 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2455 entry = pte_mkyoung(vmf->orig_pte);
2456 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2457 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2458 update_mmu_cache(vma, vmf->address, vmf->pte);
2459 pte_unmap_unlock(vmf->pte, vmf->ptl);
2463 * Handle the case of a page which we actually need to copy to a new page.
2465 * Called with mmap_sem locked and the old page referenced, but
2466 * without the ptl held.
2468 * High level logic flow:
2470 * - Allocate a page, copy the content of the old page to the new one.
2471 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2472 * - Take the PTL. If the pte changed, bail out and release the allocated page
2473 * - If the pte is still the way we remember it, update the page table and all
2474 * relevant references. This includes dropping the reference the page-table
2475 * held to the old page, as well as updating the rmap.
2476 * - In any case, unlock the PTL and drop the reference we took to the old page.
2478 static int wp_page_copy(struct vm_fault *vmf)
2480 struct vm_area_struct *vma = vmf->vma;
2481 struct mm_struct *mm = vma->vm_mm;
2482 struct page *old_page = vmf->page;
2483 struct page *new_page = NULL;
2485 int page_copied = 0;
2486 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2487 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2488 struct mem_cgroup *memcg;
2490 if (unlikely(anon_vma_prepare(vma)))
2493 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2494 new_page = alloc_zeroed_user_highpage_movable(vma,
2499 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2503 cow_user_page(new_page, old_page, vmf->address, vma);
2506 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2509 __SetPageUptodate(new_page);
2511 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2514 * Re-check the pte - we dropped the lock
2516 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2517 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2519 if (!PageAnon(old_page)) {
2520 dec_mm_counter_fast(mm,
2521 mm_counter_file(old_page));
2522 inc_mm_counter_fast(mm, MM_ANONPAGES);
2525 inc_mm_counter_fast(mm, MM_ANONPAGES);
2527 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2528 entry = mk_pte(new_page, vma->vm_page_prot);
2529 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2531 * Clear the pte entry and flush it first, before updating the
2532 * pte with the new entry. This will avoid a race condition
2533 * seen in the presence of one thread doing SMC and another
2536 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2537 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2538 mem_cgroup_commit_charge(new_page, memcg, false, false);
2539 lru_cache_add_active_or_unevictable(new_page, vma);
2541 * We call the notify macro here because, when using secondary
2542 * mmu page tables (such as kvm shadow page tables), we want the
2543 * new page to be mapped directly into the secondary page table.
2545 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2546 update_mmu_cache(vma, vmf->address, vmf->pte);
2549 * Only after switching the pte to the new page may
2550 * we remove the mapcount here. Otherwise another
2551 * process may come and find the rmap count decremented
2552 * before the pte is switched to the new page, and
2553 * "reuse" the old page writing into it while our pte
2554 * here still points into it and can be read by other
2557 * The critical issue is to order this
2558 * page_remove_rmap with the ptp_clear_flush above.
2559 * Those stores are ordered by (if nothing else,)
2560 * the barrier present in the atomic_add_negative
2561 * in page_remove_rmap.
2563 * Then the TLB flush in ptep_clear_flush ensures that
2564 * no process can access the old page before the
2565 * decremented mapcount is visible. And the old page
2566 * cannot be reused until after the decremented
2567 * mapcount is visible. So transitively, TLBs to
2568 * old page will be flushed before it can be reused.
2570 page_remove_rmap(old_page, false);
2573 /* Free the old page.. */
2574 new_page = old_page;
2577 mem_cgroup_cancel_charge(new_page, memcg, false);
2583 pte_unmap_unlock(vmf->pte, vmf->ptl);
2585 * No need to double call mmu_notifier->invalidate_range() callback as
2586 * the above ptep_clear_flush_notify() did already call it.
2588 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2591 * Don't let another task, with possibly unlocked vma,
2592 * keep the mlocked page.
2594 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2595 lock_page(old_page); /* LRU manipulation */
2596 if (PageMlocked(old_page))
2597 munlock_vma_page(old_page);
2598 unlock_page(old_page);
2602 return page_copied ? VM_FAULT_WRITE : 0;
2608 return VM_FAULT_OOM;
2612 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2613 * writeable once the page is prepared
2615 * @vmf: structure describing the fault
2617 * This function handles all that is needed to finish a write page fault in a
2618 * shared mapping due to PTE being read-only once the mapped page is prepared.
2619 * It handles locking of PTE and modifying it. The function returns
2620 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2623 * The function expects the page to be locked or other protection against
2624 * concurrent faults / writeback (such as DAX radix tree locks).
2626 int finish_mkwrite_fault(struct vm_fault *vmf)
2628 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2629 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2632 * We might have raced with another page fault while we released the
2633 * pte_offset_map_lock.
2635 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2636 pte_unmap_unlock(vmf->pte, vmf->ptl);
2637 return VM_FAULT_NOPAGE;
2644 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2647 static int wp_pfn_shared(struct vm_fault *vmf)
2649 struct vm_area_struct *vma = vmf->vma;
2651 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2654 pte_unmap_unlock(vmf->pte, vmf->ptl);
2655 vmf->flags |= FAULT_FLAG_MKWRITE;
2656 ret = vma->vm_ops->pfn_mkwrite(vmf);
2657 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2659 return finish_mkwrite_fault(vmf);
2662 return VM_FAULT_WRITE;
2665 static int wp_page_shared(struct vm_fault *vmf)
2666 __releases(vmf->ptl)
2668 struct vm_area_struct *vma = vmf->vma;
2670 get_page(vmf->page);
2672 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2675 pte_unmap_unlock(vmf->pte, vmf->ptl);
2676 tmp = do_page_mkwrite(vmf);
2677 if (unlikely(!tmp || (tmp &
2678 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2679 put_page(vmf->page);
2682 tmp = finish_mkwrite_fault(vmf);
2683 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2684 unlock_page(vmf->page);
2685 put_page(vmf->page);
2690 lock_page(vmf->page);
2692 fault_dirty_shared_page(vma, vmf->page);
2693 put_page(vmf->page);
2695 return VM_FAULT_WRITE;
2699 * This routine handles present pages, when users try to write
2700 * to a shared page. It is done by copying the page to a new address
2701 * and decrementing the shared-page counter for the old page.
2703 * Note that this routine assumes that the protection checks have been
2704 * done by the caller (the low-level page fault routine in most cases).
2705 * Thus we can safely just mark it writable once we've done any necessary
2708 * We also mark the page dirty at this point even though the page will
2709 * change only once the write actually happens. This avoids a few races,
2710 * and potentially makes it more efficient.
2712 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2713 * but allow concurrent faults), with pte both mapped and locked.
2714 * We return with mmap_sem still held, but pte unmapped and unlocked.
2716 static int do_wp_page(struct vm_fault *vmf)
2717 __releases(vmf->ptl)
2719 struct vm_area_struct *vma = vmf->vma;
2721 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2724 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2727 * We should not cow pages in a shared writeable mapping.
2728 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2730 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2731 (VM_WRITE|VM_SHARED))
2732 return wp_pfn_shared(vmf);
2734 pte_unmap_unlock(vmf->pte, vmf->ptl);
2735 return wp_page_copy(vmf);
2739 * Take out anonymous pages first, anonymous shared vmas are
2740 * not dirty accountable.
2742 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2743 int total_map_swapcount;
2744 if (!trylock_page(vmf->page)) {
2745 get_page(vmf->page);
2746 pte_unmap_unlock(vmf->pte, vmf->ptl);
2747 lock_page(vmf->page);
2748 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2749 vmf->address, &vmf->ptl);
2750 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2751 unlock_page(vmf->page);
2752 pte_unmap_unlock(vmf->pte, vmf->ptl);
2753 put_page(vmf->page);
2756 put_page(vmf->page);
2758 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2759 if (total_map_swapcount == 1) {
2761 * The page is all ours. Move it to
2762 * our anon_vma so the rmap code will
2763 * not search our parent or siblings.
2764 * Protected against the rmap code by
2767 page_move_anon_rmap(vmf->page, vma);
2769 unlock_page(vmf->page);
2771 return VM_FAULT_WRITE;
2773 unlock_page(vmf->page);
2774 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2775 (VM_WRITE|VM_SHARED))) {
2776 return wp_page_shared(vmf);
2780 * Ok, we need to copy. Oh, well..
2782 get_page(vmf->page);
2784 pte_unmap_unlock(vmf->pte, vmf->ptl);
2785 return wp_page_copy(vmf);
2788 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2789 unsigned long start_addr, unsigned long end_addr,
2790 struct zap_details *details)
2792 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2795 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2796 struct zap_details *details)
2798 struct vm_area_struct *vma;
2799 pgoff_t vba, vea, zba, zea;
2801 vma_interval_tree_foreach(vma, root,
2802 details->first_index, details->last_index) {
2804 vba = vma->vm_pgoff;
2805 vea = vba + vma_pages(vma) - 1;
2806 zba = details->first_index;
2809 zea = details->last_index;
2813 unmap_mapping_range_vma(vma,
2814 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2815 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2821 * unmap_mapping_pages() - Unmap pages from processes.
2822 * @mapping: The address space containing pages to be unmapped.
2823 * @start: Index of first page to be unmapped.
2824 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2825 * @even_cows: Whether to unmap even private COWed pages.
2827 * Unmap the pages in this address space from any userspace process which
2828 * has them mmaped. Generally, you want to remove COWed pages as well when
2829 * a file is being truncated, but not when invalidating pages from the page
2832 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2833 pgoff_t nr, bool even_cows)
2835 struct zap_details details = { };
2837 details.check_mapping = even_cows ? NULL : mapping;
2838 details.first_index = start;
2839 details.last_index = start + nr - 1;
2840 if (details.last_index < details.first_index)
2841 details.last_index = ULONG_MAX;
2843 i_mmap_lock_write(mapping);
2844 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2845 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2846 i_mmap_unlock_write(mapping);
2850 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2851 * address_space corresponding to the specified byte range in the underlying
2854 * @mapping: the address space containing mmaps to be unmapped.
2855 * @holebegin: byte in first page to unmap, relative to the start of
2856 * the underlying file. This will be rounded down to a PAGE_SIZE
2857 * boundary. Note that this is different from truncate_pagecache(), which
2858 * must keep the partial page. In contrast, we must get rid of
2860 * @holelen: size of prospective hole in bytes. This will be rounded
2861 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2863 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2864 * but 0 when invalidating pagecache, don't throw away private data.
2866 void unmap_mapping_range(struct address_space *mapping,
2867 loff_t const holebegin, loff_t const holelen, int even_cows)
2869 pgoff_t hba = holebegin >> PAGE_SHIFT;
2870 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2872 /* Check for overflow. */
2873 if (sizeof(holelen) > sizeof(hlen)) {
2875 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2876 if (holeend & ~(long long)ULONG_MAX)
2877 hlen = ULONG_MAX - hba + 1;
2880 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2882 EXPORT_SYMBOL(unmap_mapping_range);
2885 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2886 * but allow concurrent faults), and pte mapped but not yet locked.
2887 * We return with pte unmapped and unlocked.
2889 * We return with the mmap_sem locked or unlocked in the same cases
2890 * as does filemap_fault().
2892 int do_swap_page(struct vm_fault *vmf)
2894 struct vm_area_struct *vma = vmf->vma;
2895 struct page *page = NULL, *swapcache;
2896 struct mem_cgroup *memcg;
2903 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2906 entry = pte_to_swp_entry(vmf->orig_pte);
2907 if (unlikely(non_swap_entry(entry))) {
2908 if (is_migration_entry(entry)) {
2909 migration_entry_wait(vma->vm_mm, vmf->pmd,
2911 } else if (is_device_private_entry(entry)) {
2913 * For un-addressable device memory we call the pgmap
2914 * fault handler callback. The callback must migrate
2915 * the page back to some CPU accessible page.
2917 ret = device_private_entry_fault(vma, vmf->address, entry,
2918 vmf->flags, vmf->pmd);
2919 } else if (is_hwpoison_entry(entry)) {
2920 ret = VM_FAULT_HWPOISON;
2922 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2923 ret = VM_FAULT_SIGBUS;
2929 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2930 page = lookup_swap_cache(entry, vma, vmf->address);
2934 struct swap_info_struct *si = swp_swap_info(entry);
2936 if (si->flags & SWP_SYNCHRONOUS_IO &&
2937 __swap_count(si, entry) == 1) {
2938 /* skip swapcache */
2939 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2942 __SetPageLocked(page);
2943 __SetPageSwapBacked(page);
2944 set_page_private(page, entry.val);
2945 lru_cache_add_anon(page);
2946 swap_readpage(page, true);
2949 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2956 * Back out if somebody else faulted in this pte
2957 * while we released the pte lock.
2959 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2960 vmf->address, &vmf->ptl);
2961 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2963 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2967 /* Had to read the page from swap area: Major fault */
2968 ret = VM_FAULT_MAJOR;
2969 count_vm_event(PGMAJFAULT);
2970 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2971 } else if (PageHWPoison(page)) {
2973 * hwpoisoned dirty swapcache pages are kept for killing
2974 * owner processes (which may be unknown at hwpoison time)
2976 ret = VM_FAULT_HWPOISON;
2977 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2981 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2983 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2985 ret |= VM_FAULT_RETRY;
2990 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2991 * release the swapcache from under us. The page pin, and pte_same
2992 * test below, are not enough to exclude that. Even if it is still
2993 * swapcache, we need to check that the page's swap has not changed.
2995 if (unlikely((!PageSwapCache(page) ||
2996 page_private(page) != entry.val)) && swapcache)
2999 page = ksm_might_need_to_copy(page, vma, vmf->address);
3000 if (unlikely(!page)) {
3006 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3013 * Back out if somebody else already faulted in this pte.
3015 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3017 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3020 if (unlikely(!PageUptodate(page))) {
3021 ret = VM_FAULT_SIGBUS;
3026 * The page isn't present yet, go ahead with the fault.
3028 * Be careful about the sequence of operations here.
3029 * To get its accounting right, reuse_swap_page() must be called
3030 * while the page is counted on swap but not yet in mapcount i.e.
3031 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3032 * must be called after the swap_free(), or it will never succeed.
3035 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3036 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3037 pte = mk_pte(page, vma->vm_page_prot);
3038 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3039 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3040 vmf->flags &= ~FAULT_FLAG_WRITE;
3041 ret |= VM_FAULT_WRITE;
3042 exclusive = RMAP_EXCLUSIVE;
3044 flush_icache_page(vma, page);
3045 if (pte_swp_soft_dirty(vmf->orig_pte))
3046 pte = pte_mksoft_dirty(pte);
3047 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3048 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3049 vmf->orig_pte = pte;
3051 /* ksm created a completely new copy */
3052 if (unlikely(page != swapcache && swapcache)) {
3053 page_add_new_anon_rmap(page, vma, vmf->address, false);
3054 mem_cgroup_commit_charge(page, memcg, false, false);
3055 lru_cache_add_active_or_unevictable(page, vma);
3057 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3058 mem_cgroup_commit_charge(page, memcg, true, false);
3059 activate_page(page);
3063 if (mem_cgroup_swap_full(page) ||
3064 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3065 try_to_free_swap(page);
3067 if (page != swapcache && swapcache) {
3069 * Hold the lock to avoid the swap entry to be reused
3070 * until we take the PT lock for the pte_same() check
3071 * (to avoid false positives from pte_same). For
3072 * further safety release the lock after the swap_free
3073 * so that the swap count won't change under a
3074 * parallel locked swapcache.
3076 unlock_page(swapcache);
3077 put_page(swapcache);
3080 if (vmf->flags & FAULT_FLAG_WRITE) {
3081 ret |= do_wp_page(vmf);
3082 if (ret & VM_FAULT_ERROR)
3083 ret &= VM_FAULT_ERROR;
3087 /* No need to invalidate - it was non-present before */
3088 update_mmu_cache(vma, vmf->address, vmf->pte);
3090 pte_unmap_unlock(vmf->pte, vmf->ptl);
3094 mem_cgroup_cancel_charge(page, memcg, false);
3095 pte_unmap_unlock(vmf->pte, vmf->ptl);
3100 if (page != swapcache && swapcache) {
3101 unlock_page(swapcache);
3102 put_page(swapcache);
3108 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3109 * but allow concurrent faults), and pte mapped but not yet locked.
3110 * We return with mmap_sem still held, but pte unmapped and unlocked.
3112 static int do_anonymous_page(struct vm_fault *vmf)
3114 struct vm_area_struct *vma = vmf->vma;
3115 struct mem_cgroup *memcg;
3120 /* File mapping without ->vm_ops ? */
3121 if (vma->vm_flags & VM_SHARED)
3122 return VM_FAULT_SIGBUS;
3125 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3126 * pte_offset_map() on pmds where a huge pmd might be created
3127 * from a different thread.
3129 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3130 * parallel threads are excluded by other means.
3132 * Here we only have down_read(mmap_sem).
3134 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3135 return VM_FAULT_OOM;
3137 /* See the comment in pte_alloc_one_map() */
3138 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3141 /* Use the zero-page for reads */
3142 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3143 !mm_forbids_zeropage(vma->vm_mm)) {
3144 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3145 vma->vm_page_prot));
3146 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3147 vmf->address, &vmf->ptl);
3148 if (!pte_none(*vmf->pte))
3150 ret = check_stable_address_space(vma->vm_mm);
3153 /* Deliver the page fault to userland, check inside PT lock */
3154 if (userfaultfd_missing(vma)) {
3155 pte_unmap_unlock(vmf->pte, vmf->ptl);
3156 return handle_userfault(vmf, VM_UFFD_MISSING);
3161 /* Allocate our own private page. */
3162 if (unlikely(anon_vma_prepare(vma)))
3164 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3168 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3173 * The memory barrier inside __SetPageUptodate makes sure that
3174 * preceeding stores to the page contents become visible before
3175 * the set_pte_at() write.
3177 __SetPageUptodate(page);
3179 entry = mk_pte(page, vma->vm_page_prot);
3180 if (vma->vm_flags & VM_WRITE)
3181 entry = pte_mkwrite(pte_mkdirty(entry));
3183 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3185 if (!pte_none(*vmf->pte))
3188 ret = check_stable_address_space(vma->vm_mm);
3192 /* Deliver the page fault to userland, check inside PT lock */
3193 if (userfaultfd_missing(vma)) {
3194 pte_unmap_unlock(vmf->pte, vmf->ptl);
3195 mem_cgroup_cancel_charge(page, memcg, false);
3197 return handle_userfault(vmf, VM_UFFD_MISSING);
3200 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3201 page_add_new_anon_rmap(page, vma, vmf->address, false);
3202 mem_cgroup_commit_charge(page, memcg, false, false);
3203 lru_cache_add_active_or_unevictable(page, vma);
3205 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3207 /* No need to invalidate - it was non-present before */
3208 update_mmu_cache(vma, vmf->address, vmf->pte);
3210 pte_unmap_unlock(vmf->pte, vmf->ptl);
3213 mem_cgroup_cancel_charge(page, memcg, false);
3219 return VM_FAULT_OOM;
3223 * The mmap_sem must have been held on entry, and may have been
3224 * released depending on flags and vma->vm_ops->fault() return value.
3225 * See filemap_fault() and __lock_page_retry().
3227 static int __do_fault(struct vm_fault *vmf)
3229 struct vm_area_struct *vma = vmf->vma;
3232 ret = vma->vm_ops->fault(vmf);
3233 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3234 VM_FAULT_DONE_COW)))
3237 if (unlikely(PageHWPoison(vmf->page))) {
3238 if (ret & VM_FAULT_LOCKED)
3239 unlock_page(vmf->page);
3240 put_page(vmf->page);
3242 return VM_FAULT_HWPOISON;
3245 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3246 lock_page(vmf->page);
3248 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3254 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3255 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3256 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3257 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3259 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3261 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3264 static int pte_alloc_one_map(struct vm_fault *vmf)
3266 struct vm_area_struct *vma = vmf->vma;
3268 if (!pmd_none(*vmf->pmd))
3270 if (vmf->prealloc_pte) {
3271 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3272 if (unlikely(!pmd_none(*vmf->pmd))) {
3273 spin_unlock(vmf->ptl);
3277 mm_inc_nr_ptes(vma->vm_mm);
3278 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3279 spin_unlock(vmf->ptl);
3280 vmf->prealloc_pte = NULL;
3281 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3282 return VM_FAULT_OOM;
3286 * If a huge pmd materialized under us just retry later. Use
3287 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3288 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3289 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3290 * running immediately after a huge pmd fault in a different thread of
3291 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3292 * All we have to ensure is that it is a regular pmd that we can walk
3293 * with pte_offset_map() and we can do that through an atomic read in
3294 * C, which is what pmd_trans_unstable() provides.
3296 if (pmd_devmap_trans_unstable(vmf->pmd))
3297 return VM_FAULT_NOPAGE;
3300 * At this point we know that our vmf->pmd points to a page of ptes
3301 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3302 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3303 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3304 * be valid and we will re-check to make sure the vmf->pte isn't
3305 * pte_none() under vmf->ptl protection when we return to
3308 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3313 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3315 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3316 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3317 unsigned long haddr)
3319 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3320 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3322 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3327 static void deposit_prealloc_pte(struct vm_fault *vmf)
3329 struct vm_area_struct *vma = vmf->vma;
3331 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3333 * We are going to consume the prealloc table,
3334 * count that as nr_ptes.
3336 mm_inc_nr_ptes(vma->vm_mm);
3337 vmf->prealloc_pte = NULL;
3340 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3342 struct vm_area_struct *vma = vmf->vma;
3343 bool write = vmf->flags & FAULT_FLAG_WRITE;
3344 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3348 if (!transhuge_vma_suitable(vma, haddr))
3349 return VM_FAULT_FALLBACK;
3351 ret = VM_FAULT_FALLBACK;
3352 page = compound_head(page);
3355 * Archs like ppc64 need additonal space to store information
3356 * related to pte entry. Use the preallocated table for that.
3358 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3359 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3360 if (!vmf->prealloc_pte)
3361 return VM_FAULT_OOM;
3362 smp_wmb(); /* See comment in __pte_alloc() */
3365 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3366 if (unlikely(!pmd_none(*vmf->pmd)))
3369 for (i = 0; i < HPAGE_PMD_NR; i++)
3370 flush_icache_page(vma, page + i);
3372 entry = mk_huge_pmd(page, vma->vm_page_prot);
3374 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3376 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3377 page_add_file_rmap(page, true);
3379 * deposit and withdraw with pmd lock held
3381 if (arch_needs_pgtable_deposit())
3382 deposit_prealloc_pte(vmf);
3384 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3386 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3388 /* fault is handled */
3390 count_vm_event(THP_FILE_MAPPED);
3392 spin_unlock(vmf->ptl);
3396 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3404 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3405 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3407 * @vmf: fault environment
3408 * @memcg: memcg to charge page (only for private mappings)
3409 * @page: page to map
3411 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3414 * Target users are page handler itself and implementations of
3415 * vm_ops->map_pages.
3417 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3420 struct vm_area_struct *vma = vmf->vma;
3421 bool write = vmf->flags & FAULT_FLAG_WRITE;
3425 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3426 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3428 VM_BUG_ON_PAGE(memcg, page);
3430 ret = do_set_pmd(vmf, page);
3431 if (ret != VM_FAULT_FALLBACK)
3436 ret = pte_alloc_one_map(vmf);
3441 /* Re-check under ptl */
3442 if (unlikely(!pte_none(*vmf->pte)))
3443 return VM_FAULT_NOPAGE;
3445 flush_icache_page(vma, page);
3446 entry = mk_pte(page, vma->vm_page_prot);
3448 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3449 /* copy-on-write page */
3450 if (write && !(vma->vm_flags & VM_SHARED)) {
3451 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3452 page_add_new_anon_rmap(page, vma, vmf->address, false);
3453 mem_cgroup_commit_charge(page, memcg, false, false);
3454 lru_cache_add_active_or_unevictable(page, vma);
3456 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3457 page_add_file_rmap(page, false);
3459 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3461 /* no need to invalidate: a not-present page won't be cached */
3462 update_mmu_cache(vma, vmf->address, vmf->pte);
3469 * finish_fault - finish page fault once we have prepared the page to fault
3471 * @vmf: structure describing the fault
3473 * This function handles all that is needed to finish a page fault once the
3474 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3475 * given page, adds reverse page mapping, handles memcg charges and LRU
3476 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3479 * The function expects the page to be locked and on success it consumes a
3480 * reference of a page being mapped (for the PTE which maps it).
3482 int finish_fault(struct vm_fault *vmf)
3487 /* Did we COW the page? */
3488 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3489 !(vmf->vma->vm_flags & VM_SHARED))
3490 page = vmf->cow_page;
3495 * check even for read faults because we might have lost our CoWed
3498 if (!(vmf->vma->vm_flags & VM_SHARED))
3499 ret = check_stable_address_space(vmf->vma->vm_mm);
3501 ret = alloc_set_pte(vmf, vmf->memcg, page);
3503 pte_unmap_unlock(vmf->pte, vmf->ptl);
3507 static unsigned long fault_around_bytes __read_mostly =
3508 rounddown_pow_of_two(65536);
3510 #ifdef CONFIG_DEBUG_FS
3511 static int fault_around_bytes_get(void *data, u64 *val)
3513 *val = fault_around_bytes;
3518 * fault_around_bytes must be rounded down to the nearest page order as it's
3519 * what do_fault_around() expects to see.
3521 static int fault_around_bytes_set(void *data, u64 val)
3523 if (val / PAGE_SIZE > PTRS_PER_PTE)
3525 if (val > PAGE_SIZE)
3526 fault_around_bytes = rounddown_pow_of_two(val);
3528 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3531 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3532 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3534 static int __init fault_around_debugfs(void)
3538 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3539 &fault_around_bytes_fops);
3541 pr_warn("Failed to create fault_around_bytes in debugfs");
3544 late_initcall(fault_around_debugfs);
3548 * do_fault_around() tries to map few pages around the fault address. The hope
3549 * is that the pages will be needed soon and this will lower the number of
3552 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3553 * not ready to be mapped: not up-to-date, locked, etc.
3555 * This function is called with the page table lock taken. In the split ptlock
3556 * case the page table lock only protects only those entries which belong to
3557 * the page table corresponding to the fault address.
3559 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3562 * fault_around_bytes defines how many bytes we'll try to map.
3563 * do_fault_around() expects it to be set to a power of two less than or equal
3566 * The virtual address of the area that we map is naturally aligned to
3567 * fault_around_bytes rounded down to the machine page size
3568 * (and therefore to page order). This way it's easier to guarantee
3569 * that we don't cross page table boundaries.
3571 static int do_fault_around(struct vm_fault *vmf)
3573 unsigned long address = vmf->address, nr_pages, mask;
3574 pgoff_t start_pgoff = vmf->pgoff;
3578 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3579 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3581 vmf->address = max(address & mask, vmf->vma->vm_start);
3582 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3586 * end_pgoff is either the end of the page table, the end of
3587 * the vma or nr_pages from start_pgoff, depending what is nearest.
3589 end_pgoff = start_pgoff -
3590 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3592 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3593 start_pgoff + nr_pages - 1);
3595 if (pmd_none(*vmf->pmd)) {
3596 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3598 if (!vmf->prealloc_pte)
3600 smp_wmb(); /* See comment in __pte_alloc() */
3603 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3605 /* Huge page is mapped? Page fault is solved */
3606 if (pmd_trans_huge(*vmf->pmd)) {
3607 ret = VM_FAULT_NOPAGE;
3611 /* ->map_pages() haven't done anything useful. Cold page cache? */
3615 /* check if the page fault is solved */
3616 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3617 if (!pte_none(*vmf->pte))
3618 ret = VM_FAULT_NOPAGE;
3619 pte_unmap_unlock(vmf->pte, vmf->ptl);
3621 vmf->address = address;
3626 static int do_read_fault(struct vm_fault *vmf)
3628 struct vm_area_struct *vma = vmf->vma;
3632 * Let's call ->map_pages() first and use ->fault() as fallback
3633 * if page by the offset is not ready to be mapped (cold cache or
3636 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3637 ret = do_fault_around(vmf);
3642 ret = __do_fault(vmf);
3643 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3646 ret |= finish_fault(vmf);
3647 unlock_page(vmf->page);
3648 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3649 put_page(vmf->page);
3653 static int do_cow_fault(struct vm_fault *vmf)
3655 struct vm_area_struct *vma = vmf->vma;
3658 if (unlikely(anon_vma_prepare(vma)))
3659 return VM_FAULT_OOM;
3661 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3663 return VM_FAULT_OOM;
3665 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3666 &vmf->memcg, false)) {
3667 put_page(vmf->cow_page);
3668 return VM_FAULT_OOM;
3671 ret = __do_fault(vmf);
3672 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3674 if (ret & VM_FAULT_DONE_COW)
3677 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3678 __SetPageUptodate(vmf->cow_page);
3680 ret |= finish_fault(vmf);
3681 unlock_page(vmf->page);
3682 put_page(vmf->page);
3683 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3687 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3688 put_page(vmf->cow_page);
3692 static int do_shared_fault(struct vm_fault *vmf)
3694 struct vm_area_struct *vma = vmf->vma;
3697 ret = __do_fault(vmf);
3698 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3702 * Check if the backing address space wants to know that the page is
3703 * about to become writable
3705 if (vma->vm_ops->page_mkwrite) {
3706 unlock_page(vmf->page);
3707 tmp = do_page_mkwrite(vmf);
3708 if (unlikely(!tmp ||
3709 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3710 put_page(vmf->page);
3715 ret |= finish_fault(vmf);
3716 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3718 unlock_page(vmf->page);
3719 put_page(vmf->page);
3723 fault_dirty_shared_page(vma, vmf->page);
3728 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3729 * but allow concurrent faults).
3730 * The mmap_sem may have been released depending on flags and our
3731 * return value. See filemap_fault() and __lock_page_or_retry().
3733 static int do_fault(struct vm_fault *vmf)
3735 struct vm_area_struct *vma = vmf->vma;
3738 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3739 if (!vma->vm_ops->fault)
3740 ret = VM_FAULT_SIGBUS;
3741 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3742 ret = do_read_fault(vmf);
3743 else if (!(vma->vm_flags & VM_SHARED))
3744 ret = do_cow_fault(vmf);
3746 ret = do_shared_fault(vmf);
3748 /* preallocated pagetable is unused: free it */
3749 if (vmf->prealloc_pte) {
3750 pte_free(vma->vm_mm, vmf->prealloc_pte);
3751 vmf->prealloc_pte = NULL;
3756 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3757 unsigned long addr, int page_nid,
3762 count_vm_numa_event(NUMA_HINT_FAULTS);
3763 if (page_nid == numa_node_id()) {
3764 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3765 *flags |= TNF_FAULT_LOCAL;
3768 return mpol_misplaced(page, vma, addr);
3771 static int do_numa_page(struct vm_fault *vmf)
3773 struct vm_area_struct *vma = vmf->vma;
3774 struct page *page = NULL;
3778 bool migrated = false;
3780 bool was_writable = pte_savedwrite(vmf->orig_pte);
3784 * The "pte" at this point cannot be used safely without
3785 * validation through pte_unmap_same(). It's of NUMA type but
3786 * the pfn may be screwed if the read is non atomic.
3788 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3789 spin_lock(vmf->ptl);
3790 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3791 pte_unmap_unlock(vmf->pte, vmf->ptl);
3796 * Make it present again, Depending on how arch implementes non
3797 * accessible ptes, some can allow access by kernel mode.
3799 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3800 pte = pte_modify(pte, vma->vm_page_prot);
3801 pte = pte_mkyoung(pte);
3803 pte = pte_mkwrite(pte);
3804 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3805 update_mmu_cache(vma, vmf->address, vmf->pte);
3807 page = vm_normal_page(vma, vmf->address, pte);
3809 pte_unmap_unlock(vmf->pte, vmf->ptl);
3813 /* TODO: handle PTE-mapped THP */
3814 if (PageCompound(page)) {
3815 pte_unmap_unlock(vmf->pte, vmf->ptl);
3820 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3821 * much anyway since they can be in shared cache state. This misses
3822 * the case where a mapping is writable but the process never writes
3823 * to it but pte_write gets cleared during protection updates and
3824 * pte_dirty has unpredictable behaviour between PTE scan updates,
3825 * background writeback, dirty balancing and application behaviour.
3827 if (!pte_write(pte))
3828 flags |= TNF_NO_GROUP;
3831 * Flag if the page is shared between multiple address spaces. This
3832 * is later used when determining whether to group tasks together
3834 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3835 flags |= TNF_SHARED;
3837 last_cpupid = page_cpupid_last(page);
3838 page_nid = page_to_nid(page);
3839 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3841 pte_unmap_unlock(vmf->pte, vmf->ptl);
3842 if (target_nid == -1) {
3847 /* Migrate to the requested node */
3848 migrated = migrate_misplaced_page(page, vma, target_nid);
3850 page_nid = target_nid;
3851 flags |= TNF_MIGRATED;
3853 flags |= TNF_MIGRATE_FAIL;
3857 task_numa_fault(last_cpupid, page_nid, 1, flags);
3861 static inline int create_huge_pmd(struct vm_fault *vmf)
3863 if (vma_is_anonymous(vmf->vma))
3864 return do_huge_pmd_anonymous_page(vmf);
3865 if (vmf->vma->vm_ops->huge_fault)
3866 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3867 return VM_FAULT_FALLBACK;
3870 /* `inline' is required to avoid gcc 4.1.2 build error */
3871 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3873 if (vma_is_anonymous(vmf->vma))
3874 return do_huge_pmd_wp_page(vmf, orig_pmd);
3875 if (vmf->vma->vm_ops->huge_fault)
3876 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3878 /* COW handled on pte level: split pmd */
3879 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3880 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3882 return VM_FAULT_FALLBACK;
3885 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3887 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3890 static int create_huge_pud(struct vm_fault *vmf)
3892 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3893 /* No support for anonymous transparent PUD pages yet */
3894 if (vma_is_anonymous(vmf->vma))
3895 return VM_FAULT_FALLBACK;
3896 if (vmf->vma->vm_ops->huge_fault)
3897 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3898 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3899 return VM_FAULT_FALLBACK;
3902 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3904 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3905 /* No support for anonymous transparent PUD pages yet */
3906 if (vma_is_anonymous(vmf->vma))
3907 return VM_FAULT_FALLBACK;
3908 if (vmf->vma->vm_ops->huge_fault)
3909 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3910 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3911 return VM_FAULT_FALLBACK;
3915 * These routines also need to handle stuff like marking pages dirty
3916 * and/or accessed for architectures that don't do it in hardware (most
3917 * RISC architectures). The early dirtying is also good on the i386.
3919 * There is also a hook called "update_mmu_cache()" that architectures
3920 * with external mmu caches can use to update those (ie the Sparc or
3921 * PowerPC hashed page tables that act as extended TLBs).
3923 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3924 * concurrent faults).
3926 * The mmap_sem may have been released depending on flags and our return value.
3927 * See filemap_fault() and __lock_page_or_retry().
3929 static int handle_pte_fault(struct vm_fault *vmf)
3933 if (unlikely(pmd_none(*vmf->pmd))) {
3935 * Leave __pte_alloc() until later: because vm_ops->fault may
3936 * want to allocate huge page, and if we expose page table
3937 * for an instant, it will be difficult to retract from
3938 * concurrent faults and from rmap lookups.
3942 /* See comment in pte_alloc_one_map() */
3943 if (pmd_devmap_trans_unstable(vmf->pmd))
3946 * A regular pmd is established and it can't morph into a huge
3947 * pmd from under us anymore at this point because we hold the
3948 * mmap_sem read mode and khugepaged takes it in write mode.
3949 * So now it's safe to run pte_offset_map().
3951 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3952 vmf->orig_pte = *vmf->pte;
3955 * some architectures can have larger ptes than wordsize,
3956 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3957 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3958 * accesses. The code below just needs a consistent view
3959 * for the ifs and we later double check anyway with the
3960 * ptl lock held. So here a barrier will do.
3963 if (pte_none(vmf->orig_pte)) {
3964 pte_unmap(vmf->pte);
3970 if (vma_is_anonymous(vmf->vma))
3971 return do_anonymous_page(vmf);
3973 return do_fault(vmf);
3976 if (!pte_present(vmf->orig_pte))
3977 return do_swap_page(vmf);
3979 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3980 return do_numa_page(vmf);
3982 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3983 spin_lock(vmf->ptl);
3984 entry = vmf->orig_pte;
3985 if (unlikely(!pte_same(*vmf->pte, entry)))
3987 if (vmf->flags & FAULT_FLAG_WRITE) {
3988 if (!pte_write(entry))
3989 return do_wp_page(vmf);
3990 entry = pte_mkdirty(entry);
3992 entry = pte_mkyoung(entry);
3993 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3994 vmf->flags & FAULT_FLAG_WRITE)) {
3995 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3998 * This is needed only for protection faults but the arch code
3999 * is not yet telling us if this is a protection fault or not.
4000 * This still avoids useless tlb flushes for .text page faults
4003 if (vmf->flags & FAULT_FLAG_WRITE)
4004 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4007 pte_unmap_unlock(vmf->pte, vmf->ptl);
4012 * By the time we get here, we already hold the mm semaphore
4014 * The mmap_sem may have been released depending on flags and our
4015 * return value. See filemap_fault() and __lock_page_or_retry().
4017 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4020 struct vm_fault vmf = {
4022 .address = address & PAGE_MASK,
4024 .pgoff = linear_page_index(vma, address),
4025 .gfp_mask = __get_fault_gfp_mask(vma),
4027 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4028 struct mm_struct *mm = vma->vm_mm;
4033 pgd = pgd_offset(mm, address);
4034 p4d = p4d_alloc(mm, pgd, address);
4036 return VM_FAULT_OOM;
4038 vmf.pud = pud_alloc(mm, p4d, address);
4040 return VM_FAULT_OOM;
4041 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4042 ret = create_huge_pud(&vmf);
4043 if (!(ret & VM_FAULT_FALLBACK))
4046 pud_t orig_pud = *vmf.pud;
4049 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4051 /* NUMA case for anonymous PUDs would go here */
4053 if (dirty && !pud_write(orig_pud)) {
4054 ret = wp_huge_pud(&vmf, orig_pud);
4055 if (!(ret & VM_FAULT_FALLBACK))
4058 huge_pud_set_accessed(&vmf, orig_pud);
4064 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4066 return VM_FAULT_OOM;
4067 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4068 ret = create_huge_pmd(&vmf);
4069 if (!(ret & VM_FAULT_FALLBACK))
4072 pmd_t orig_pmd = *vmf.pmd;
4075 if (unlikely(is_swap_pmd(orig_pmd))) {
4076 VM_BUG_ON(thp_migration_supported() &&
4077 !is_pmd_migration_entry(orig_pmd));
4078 if (is_pmd_migration_entry(orig_pmd))
4079 pmd_migration_entry_wait(mm, vmf.pmd);
4082 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4083 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4084 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4086 if (dirty && !pmd_write(orig_pmd)) {
4087 ret = wp_huge_pmd(&vmf, orig_pmd);
4088 if (!(ret & VM_FAULT_FALLBACK))
4091 huge_pmd_set_accessed(&vmf, orig_pmd);
4097 return handle_pte_fault(&vmf);
4101 * By the time we get here, we already hold the mm semaphore
4103 * The mmap_sem may have been released depending on flags and our
4104 * return value. See filemap_fault() and __lock_page_or_retry().
4106 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4111 __set_current_state(TASK_RUNNING);
4113 count_vm_event(PGFAULT);
4114 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4116 /* do counter updates before entering really critical section. */
4117 check_sync_rss_stat(current);
4119 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4120 flags & FAULT_FLAG_INSTRUCTION,
4121 flags & FAULT_FLAG_REMOTE))
4122 return VM_FAULT_SIGSEGV;
4125 * Enable the memcg OOM handling for faults triggered in user
4126 * space. Kernel faults are handled more gracefully.
4128 if (flags & FAULT_FLAG_USER)
4129 mem_cgroup_enter_user_fault();
4131 if (unlikely(is_vm_hugetlb_page(vma)))
4132 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4134 ret = __handle_mm_fault(vma, address, flags);
4136 if (flags & FAULT_FLAG_USER) {
4137 mem_cgroup_exit_user_fault();
4139 * The task may have entered a memcg OOM situation but
4140 * if the allocation error was handled gracefully (no
4141 * VM_FAULT_OOM), there is no need to kill anything.
4142 * Just clean up the OOM state peacefully.
4144 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4145 mem_cgroup_oom_synchronize(false);
4150 EXPORT_SYMBOL_GPL(handle_mm_fault);
4152 #ifndef __PAGETABLE_P4D_FOLDED
4154 * Allocate p4d page table.
4155 * We've already handled the fast-path in-line.
4157 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4159 p4d_t *new = p4d_alloc_one(mm, address);
4163 smp_wmb(); /* See comment in __pte_alloc */
4165 spin_lock(&mm->page_table_lock);
4166 if (pgd_present(*pgd)) /* Another has populated it */
4169 pgd_populate(mm, pgd, new);
4170 spin_unlock(&mm->page_table_lock);
4173 #endif /* __PAGETABLE_P4D_FOLDED */
4175 #ifndef __PAGETABLE_PUD_FOLDED
4177 * Allocate page upper directory.
4178 * We've already handled the fast-path in-line.
4180 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4182 pud_t *new = pud_alloc_one(mm, address);
4186 smp_wmb(); /* See comment in __pte_alloc */
4188 spin_lock(&mm->page_table_lock);
4189 #ifndef __ARCH_HAS_5LEVEL_HACK
4190 if (!p4d_present(*p4d)) {
4192 p4d_populate(mm, p4d, new);
4193 } else /* Another has populated it */
4196 if (!pgd_present(*p4d)) {
4198 pgd_populate(mm, p4d, new);
4199 } else /* Another has populated it */
4201 #endif /* __ARCH_HAS_5LEVEL_HACK */
4202 spin_unlock(&mm->page_table_lock);
4205 #endif /* __PAGETABLE_PUD_FOLDED */
4207 #ifndef __PAGETABLE_PMD_FOLDED
4209 * Allocate page middle directory.
4210 * We've already handled the fast-path in-line.
4212 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4215 pmd_t *new = pmd_alloc_one(mm, address);
4219 smp_wmb(); /* See comment in __pte_alloc */
4221 ptl = pud_lock(mm, pud);
4222 #ifndef __ARCH_HAS_4LEVEL_HACK
4223 if (!pud_present(*pud)) {
4225 pud_populate(mm, pud, new);
4226 } else /* Another has populated it */
4229 if (!pgd_present(*pud)) {
4231 pgd_populate(mm, pud, new);
4232 } else /* Another has populated it */
4234 #endif /* __ARCH_HAS_4LEVEL_HACK */
4238 #endif /* __PAGETABLE_PMD_FOLDED */
4240 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4241 unsigned long *start, unsigned long *end,
4242 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4250 pgd = pgd_offset(mm, address);
4251 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4254 p4d = p4d_offset(pgd, address);
4255 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4258 pud = pud_offset(p4d, address);
4259 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4262 pmd = pmd_offset(pud, address);
4263 VM_BUG_ON(pmd_trans_huge(*pmd));
4265 if (pmd_huge(*pmd)) {
4270 *start = address & PMD_MASK;
4271 *end = *start + PMD_SIZE;
4272 mmu_notifier_invalidate_range_start(mm, *start, *end);
4274 *ptlp = pmd_lock(mm, pmd);
4275 if (pmd_huge(*pmd)) {
4281 mmu_notifier_invalidate_range_end(mm, *start, *end);
4284 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4288 *start = address & PAGE_MASK;
4289 *end = *start + PAGE_SIZE;
4290 mmu_notifier_invalidate_range_start(mm, *start, *end);
4292 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4293 if (!pte_present(*ptep))
4298 pte_unmap_unlock(ptep, *ptlp);
4300 mmu_notifier_invalidate_range_end(mm, *start, *end);
4305 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4306 pte_t **ptepp, spinlock_t **ptlp)
4310 /* (void) is needed to make gcc happy */
4311 (void) __cond_lock(*ptlp,
4312 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4313 ptepp, NULL, ptlp)));
4317 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4318 unsigned long *start, unsigned long *end,
4319 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4323 /* (void) is needed to make gcc happy */
4324 (void) __cond_lock(*ptlp,
4325 !(res = __follow_pte_pmd(mm, address, start, end,
4326 ptepp, pmdpp, ptlp)));
4329 EXPORT_SYMBOL(follow_pte_pmd);
4332 * follow_pfn - look up PFN at a user virtual address
4333 * @vma: memory mapping
4334 * @address: user virtual address
4335 * @pfn: location to store found PFN
4337 * Only IO mappings and raw PFN mappings are allowed.
4339 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4341 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4348 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4351 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4354 *pfn = pte_pfn(*ptep);
4355 pte_unmap_unlock(ptep, ptl);
4358 EXPORT_SYMBOL(follow_pfn);
4360 #ifdef CONFIG_HAVE_IOREMAP_PROT
4361 int follow_phys(struct vm_area_struct *vma,
4362 unsigned long address, unsigned int flags,
4363 unsigned long *prot, resource_size_t *phys)
4369 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4372 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4376 if ((flags & FOLL_WRITE) && !pte_write(pte))
4379 *prot = pgprot_val(pte_pgprot(pte));
4380 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4384 pte_unmap_unlock(ptep, ptl);
4389 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4390 void *buf, int len, int write)
4392 resource_size_t phys_addr;
4393 unsigned long prot = 0;
4394 void __iomem *maddr;
4395 int offset = addr & (PAGE_SIZE-1);
4397 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4400 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4405 memcpy_toio(maddr + offset, buf, len);
4407 memcpy_fromio(buf, maddr + offset, len);
4412 EXPORT_SYMBOL_GPL(generic_access_phys);
4416 * Access another process' address space as given in mm. If non-NULL, use the
4417 * given task for page fault accounting.
4419 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4420 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4422 struct vm_area_struct *vma;
4423 void *old_buf = buf;
4424 int write = gup_flags & FOLL_WRITE;
4426 down_read(&mm->mmap_sem);
4427 /* ignore errors, just check how much was successfully transferred */
4429 int bytes, ret, offset;
4431 struct page *page = NULL;
4433 ret = get_user_pages_remote(tsk, mm, addr, 1,
4434 gup_flags, &page, &vma, NULL);
4436 #ifndef CONFIG_HAVE_IOREMAP_PROT
4440 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4441 * we can access using slightly different code.
4443 vma = find_vma(mm, addr);
4444 if (!vma || vma->vm_start > addr)
4446 if (vma->vm_ops && vma->vm_ops->access)
4447 ret = vma->vm_ops->access(vma, addr, buf,
4455 offset = addr & (PAGE_SIZE-1);
4456 if (bytes > PAGE_SIZE-offset)
4457 bytes = PAGE_SIZE-offset;
4461 copy_to_user_page(vma, page, addr,
4462 maddr + offset, buf, bytes);
4463 set_page_dirty_lock(page);
4465 copy_from_user_page(vma, page, addr,
4466 buf, maddr + offset, bytes);
4475 up_read(&mm->mmap_sem);
4477 return buf - old_buf;
4481 * access_remote_vm - access another process' address space
4482 * @mm: the mm_struct of the target address space
4483 * @addr: start address to access
4484 * @buf: source or destination buffer
4485 * @len: number of bytes to transfer
4486 * @gup_flags: flags modifying lookup behaviour
4488 * The caller must hold a reference on @mm.
4490 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4491 void *buf, int len, unsigned int gup_flags)
4493 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4497 * Access another process' address space.
4498 * Source/target buffer must be kernel space,
4499 * Do not walk the page table directly, use get_user_pages
4501 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4502 void *buf, int len, unsigned int gup_flags)
4504 struct mm_struct *mm;
4507 mm = get_task_mm(tsk);
4511 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4517 EXPORT_SYMBOL_GPL(access_process_vm);
4520 * Print the name of a VMA.
4522 void print_vma_addr(char *prefix, unsigned long ip)
4524 struct mm_struct *mm = current->mm;
4525 struct vm_area_struct *vma;
4528 * we might be running from an atomic context so we cannot sleep
4530 if (!down_read_trylock(&mm->mmap_sem))
4533 vma = find_vma(mm, ip);
4534 if (vma && vma->vm_file) {
4535 struct file *f = vma->vm_file;
4536 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4540 p = file_path(f, buf, PAGE_SIZE);
4543 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4545 vma->vm_end - vma->vm_start);
4546 free_page((unsigned long)buf);
4549 up_read(&mm->mmap_sem);
4552 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4553 void __might_fault(const char *file, int line)
4556 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4557 * holding the mmap_sem, this is safe because kernel memory doesn't
4558 * get paged out, therefore we'll never actually fault, and the
4559 * below annotations will generate false positives.
4561 if (uaccess_kernel())
4563 if (pagefault_disabled())
4565 __might_sleep(file, line, 0);
4566 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4568 might_lock_read(¤t->mm->mmap_sem);
4571 EXPORT_SYMBOL(__might_fault);
4574 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4576 * Process all subpages of the specified huge page with the specified
4577 * operation. The target subpage will be processed last to keep its
4580 static inline void process_huge_page(
4581 unsigned long addr_hint, unsigned int pages_per_huge_page,
4582 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4586 unsigned long addr = addr_hint &
4587 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4589 /* Process target subpage last to keep its cache lines hot */
4591 n = (addr_hint - addr) / PAGE_SIZE;
4592 if (2 * n <= pages_per_huge_page) {
4593 /* If target subpage in first half of huge page */
4596 /* Process subpages at the end of huge page */
4597 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4599 process_subpage(addr + i * PAGE_SIZE, i, arg);
4602 /* If target subpage in second half of huge page */
4603 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4604 l = pages_per_huge_page - n;
4605 /* Process subpages at the begin of huge page */
4606 for (i = 0; i < base; i++) {
4608 process_subpage(addr + i * PAGE_SIZE, i, arg);
4612 * Process remaining subpages in left-right-left-right pattern
4613 * towards the target subpage
4615 for (i = 0; i < l; i++) {
4616 int left_idx = base + i;
4617 int right_idx = base + 2 * l - 1 - i;
4620 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4622 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4626 static void clear_gigantic_page(struct page *page,
4628 unsigned int pages_per_huge_page)
4631 struct page *p = page;
4634 for (i = 0; i < pages_per_huge_page;
4635 i++, p = mem_map_next(p, page, i)) {
4637 clear_user_highpage(p, addr + i * PAGE_SIZE);
4641 static void clear_subpage(unsigned long addr, int idx, void *arg)
4643 struct page *page = arg;
4645 clear_user_highpage(page + idx, addr);
4648 void clear_huge_page(struct page *page,
4649 unsigned long addr_hint, unsigned int pages_per_huge_page)
4651 unsigned long addr = addr_hint &
4652 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4654 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4655 clear_gigantic_page(page, addr, pages_per_huge_page);
4659 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4662 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4664 struct vm_area_struct *vma,
4665 unsigned int pages_per_huge_page)
4668 struct page *dst_base = dst;
4669 struct page *src_base = src;
4671 for (i = 0; i < pages_per_huge_page; ) {
4673 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4676 dst = mem_map_next(dst, dst_base, i);
4677 src = mem_map_next(src, src_base, i);
4681 struct copy_subpage_arg {
4684 struct vm_area_struct *vma;
4687 static void copy_subpage(unsigned long addr, int idx, void *arg)
4689 struct copy_subpage_arg *copy_arg = arg;
4691 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4692 addr, copy_arg->vma);
4695 void copy_user_huge_page(struct page *dst, struct page *src,
4696 unsigned long addr_hint, struct vm_area_struct *vma,
4697 unsigned int pages_per_huge_page)
4699 unsigned long addr = addr_hint &
4700 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4701 struct copy_subpage_arg arg = {
4707 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4708 copy_user_gigantic_page(dst, src, addr, vma,
4709 pages_per_huge_page);
4713 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4716 long copy_huge_page_from_user(struct page *dst_page,
4717 const void __user *usr_src,
4718 unsigned int pages_per_huge_page,
4719 bool allow_pagefault)
4721 void *src = (void *)usr_src;
4723 unsigned long i, rc = 0;
4724 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4726 for (i = 0; i < pages_per_huge_page; i++) {
4727 if (allow_pagefault)
4728 page_kaddr = kmap(dst_page + i);
4730 page_kaddr = kmap_atomic(dst_page + i);
4731 rc = copy_from_user(page_kaddr,
4732 (const void __user *)(src + i * PAGE_SIZE),
4734 if (allow_pagefault)
4735 kunmap(dst_page + i);
4737 kunmap_atomic(page_kaddr);
4739 ret_val -= (PAGE_SIZE - rc);
4747 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4749 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4751 static struct kmem_cache *page_ptl_cachep;
4753 void __init ptlock_cache_init(void)
4755 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4759 bool ptlock_alloc(struct page *page)
4763 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4770 void ptlock_free(struct page *page)
4772 kmem_cache_free(page_ptl_cachep, page->ptl);