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.
334 * If we want tlb_remove_table() to imply TLB invalidates.
336 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
338 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
340 * Invalidate page-table caches used by hardware walkers. Then we still
341 * need to RCU-sched wait while freeing the pages because software
342 * walkers can still be in-flight.
344 tlb_flush_mmu_tlbonly(tlb);
348 static void tlb_remove_table_smp_sync(void *arg)
350 /* Simply deliver the interrupt */
353 static void tlb_remove_table_one(void *table)
356 * This isn't an RCU grace period and hence the page-tables cannot be
357 * assumed to be actually RCU-freed.
359 * It is however sufficient for software page-table walkers that rely on
360 * IRQ disabling. See the comment near struct mmu_table_batch.
362 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
363 __tlb_remove_table(table);
366 static void tlb_remove_table_rcu(struct rcu_head *head)
368 struct mmu_table_batch *batch;
371 batch = container_of(head, struct mmu_table_batch, rcu);
373 for (i = 0; i < batch->nr; i++)
374 __tlb_remove_table(batch->tables[i]);
376 free_page((unsigned long)batch);
379 void tlb_table_flush(struct mmu_gather *tlb)
381 struct mmu_table_batch **batch = &tlb->batch;
384 tlb_table_invalidate(tlb);
385 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
390 void tlb_remove_table(struct mmu_gather *tlb, void *table)
392 struct mmu_table_batch **batch = &tlb->batch;
394 if (*batch == NULL) {
395 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
396 if (*batch == NULL) {
397 tlb_table_invalidate(tlb);
398 tlb_remove_table_one(table);
404 (*batch)->tables[(*batch)->nr++] = table;
405 if ((*batch)->nr == MAX_TABLE_BATCH)
406 tlb_table_flush(tlb);
409 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
412 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
413 * @tlb: the mmu_gather structure to initialize
414 * @mm: the mm_struct of the target address space
415 * @start: start of the region that will be removed from the page-table
416 * @end: end of the region that will be removed from the page-table
418 * Called to initialize an (on-stack) mmu_gather structure for page-table
419 * tear-down from @mm. The @start and @end are set to 0 and -1
420 * respectively when @mm is without users and we're going to destroy
421 * the full address space (exit/execve).
423 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
424 unsigned long start, unsigned long end)
426 arch_tlb_gather_mmu(tlb, mm, start, end);
427 inc_tlb_flush_pending(tlb->mm);
430 void tlb_finish_mmu(struct mmu_gather *tlb,
431 unsigned long start, unsigned long end)
434 * If there are parallel threads are doing PTE changes on same range
435 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
436 * flush by batching, a thread has stable TLB entry can fail to flush
437 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
438 * forcefully if we detect parallel PTE batching threads.
440 bool force = mm_tlb_flush_nested(tlb->mm);
442 arch_tlb_finish_mmu(tlb, start, end, force);
443 dec_tlb_flush_pending(tlb->mm);
447 * Note: this doesn't free the actual pages themselves. That
448 * has been handled earlier when unmapping all the memory regions.
450 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
453 pgtable_t token = pmd_pgtable(*pmd);
455 pte_free_tlb(tlb, token, addr);
456 mm_dec_nr_ptes(tlb->mm);
459 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
460 unsigned long addr, unsigned long end,
461 unsigned long floor, unsigned long ceiling)
468 pmd = pmd_offset(pud, addr);
470 next = pmd_addr_end(addr, end);
471 if (pmd_none_or_clear_bad(pmd))
473 free_pte_range(tlb, pmd, addr);
474 } while (pmd++, addr = next, addr != end);
484 if (end - 1 > ceiling - 1)
487 pmd = pmd_offset(pud, start);
489 pmd_free_tlb(tlb, pmd, start);
490 mm_dec_nr_pmds(tlb->mm);
493 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
494 unsigned long addr, unsigned long end,
495 unsigned long floor, unsigned long ceiling)
502 pud = pud_offset(p4d, addr);
504 next = pud_addr_end(addr, end);
505 if (pud_none_or_clear_bad(pud))
507 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
508 } while (pud++, addr = next, addr != end);
518 if (end - 1 > ceiling - 1)
521 pud = pud_offset(p4d, start);
523 pud_free_tlb(tlb, pud, start);
524 mm_dec_nr_puds(tlb->mm);
527 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
528 unsigned long addr, unsigned long end,
529 unsigned long floor, unsigned long ceiling)
536 p4d = p4d_offset(pgd, addr);
538 next = p4d_addr_end(addr, end);
539 if (p4d_none_or_clear_bad(p4d))
541 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
542 } while (p4d++, addr = next, addr != end);
548 ceiling &= PGDIR_MASK;
552 if (end - 1 > ceiling - 1)
555 p4d = p4d_offset(pgd, start);
557 p4d_free_tlb(tlb, p4d, start);
561 * This function frees user-level page tables of a process.
563 void free_pgd_range(struct mmu_gather *tlb,
564 unsigned long addr, unsigned long end,
565 unsigned long floor, unsigned long ceiling)
571 * The next few lines have given us lots of grief...
573 * Why are we testing PMD* at this top level? Because often
574 * there will be no work to do at all, and we'd prefer not to
575 * go all the way down to the bottom just to discover that.
577 * Why all these "- 1"s? Because 0 represents both the bottom
578 * of the address space and the top of it (using -1 for the
579 * top wouldn't help much: the masks would do the wrong thing).
580 * The rule is that addr 0 and floor 0 refer to the bottom of
581 * the address space, but end 0 and ceiling 0 refer to the top
582 * Comparisons need to use "end - 1" and "ceiling - 1" (though
583 * that end 0 case should be mythical).
585 * Wherever addr is brought up or ceiling brought down, we must
586 * be careful to reject "the opposite 0" before it confuses the
587 * subsequent tests. But what about where end is brought down
588 * by PMD_SIZE below? no, end can't go down to 0 there.
590 * Whereas we round start (addr) and ceiling down, by different
591 * masks at different levels, in order to test whether a table
592 * now has no other vmas using it, so can be freed, we don't
593 * bother to round floor or end up - the tests don't need that.
607 if (end - 1 > ceiling - 1)
612 * We add page table cache pages with PAGE_SIZE,
613 * (see pte_free_tlb()), flush the tlb if we need
615 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
616 pgd = pgd_offset(tlb->mm, addr);
618 next = pgd_addr_end(addr, end);
619 if (pgd_none_or_clear_bad(pgd))
621 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
622 } while (pgd++, addr = next, addr != end);
625 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
626 unsigned long floor, unsigned long ceiling)
629 struct vm_area_struct *next = vma->vm_next;
630 unsigned long addr = vma->vm_start;
633 * Hide vma from rmap and truncate_pagecache before freeing
636 unlink_anon_vmas(vma);
637 unlink_file_vma(vma);
639 if (is_vm_hugetlb_page(vma)) {
640 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
641 floor, next ? next->vm_start : ceiling);
644 * Optimization: gather nearby vmas into one call down
646 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
647 && !is_vm_hugetlb_page(next)) {
650 unlink_anon_vmas(vma);
651 unlink_file_vma(vma);
653 free_pgd_range(tlb, addr, vma->vm_end,
654 floor, next ? next->vm_start : ceiling);
660 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
663 pgtable_t new = pte_alloc_one(mm, address);
668 * Ensure all pte setup (eg. pte page lock and page clearing) are
669 * visible before the pte is made visible to other CPUs by being
670 * put into page tables.
672 * The other side of the story is the pointer chasing in the page
673 * table walking code (when walking the page table without locking;
674 * ie. most of the time). Fortunately, these data accesses consist
675 * of a chain of data-dependent loads, meaning most CPUs (alpha
676 * being the notable exception) will already guarantee loads are
677 * seen in-order. See the alpha page table accessors for the
678 * smp_read_barrier_depends() barriers in page table walking code.
680 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
682 ptl = pmd_lock(mm, pmd);
683 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
685 pmd_populate(mm, pmd, new);
694 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
696 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
700 smp_wmb(); /* See comment in __pte_alloc */
702 spin_lock(&init_mm.page_table_lock);
703 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
704 pmd_populate_kernel(&init_mm, pmd, new);
707 spin_unlock(&init_mm.page_table_lock);
709 pte_free_kernel(&init_mm, new);
713 static inline void init_rss_vec(int *rss)
715 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
718 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
722 if (current->mm == mm)
724 for (i = 0; i < NR_MM_COUNTERS; i++)
726 add_mm_counter(mm, i, rss[i]);
730 * This function is called to print an error when a bad pte
731 * is found. For example, we might have a PFN-mapped pte in
732 * a region that doesn't allow it.
734 * The calling function must still handle the error.
736 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
737 pte_t pte, struct page *page)
739 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
740 p4d_t *p4d = p4d_offset(pgd, addr);
741 pud_t *pud = pud_offset(p4d, addr);
742 pmd_t *pmd = pmd_offset(pud, addr);
743 struct address_space *mapping;
745 static unsigned long resume;
746 static unsigned long nr_shown;
747 static unsigned long nr_unshown;
750 * Allow a burst of 60 reports, then keep quiet for that minute;
751 * or allow a steady drip of one report per second.
753 if (nr_shown == 60) {
754 if (time_before(jiffies, resume)) {
759 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
766 resume = jiffies + 60 * HZ;
768 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
769 index = linear_page_index(vma, addr);
771 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
773 (long long)pte_val(pte), (long long)pmd_val(*pmd));
775 dump_page(page, "bad pte");
776 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
777 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
778 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
780 vma->vm_ops ? vma->vm_ops->fault : NULL,
781 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
782 mapping ? mapping->a_ops->readpage : NULL);
784 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
788 * vm_normal_page -- This function gets the "struct page" associated with a pte.
790 * "Special" mappings do not wish to be associated with a "struct page" (either
791 * it doesn't exist, or it exists but they don't want to touch it). In this
792 * case, NULL is returned here. "Normal" mappings do have a struct page.
794 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
795 * pte bit, in which case this function is trivial. Secondly, an architecture
796 * may not have a spare pte bit, which requires a more complicated scheme,
799 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
800 * special mapping (even if there are underlying and valid "struct pages").
801 * COWed pages of a VM_PFNMAP are always normal.
803 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
804 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
805 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
806 * mapping will always honor the rule
808 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
810 * And for normal mappings this is false.
812 * This restricts such mappings to be a linear translation from virtual address
813 * to pfn. To get around this restriction, we allow arbitrary mappings so long
814 * as the vma is not a COW mapping; in that case, we know that all ptes are
815 * special (because none can have been COWed).
818 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
820 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
821 * page" backing, however the difference is that _all_ pages with a struct
822 * page (that is, those where pfn_valid is true) are refcounted and considered
823 * normal pages by the VM. The disadvantage is that pages are refcounted
824 * (which can be slower and simply not an option for some PFNMAP users). The
825 * advantage is that we don't have to follow the strict linearity rule of
826 * PFNMAP mappings in order to support COWable mappings.
829 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
830 pte_t pte, bool with_public_device)
832 unsigned long pfn = pte_pfn(pte);
834 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
835 if (likely(!pte_special(pte)))
837 if (vma->vm_ops && vma->vm_ops->find_special_page)
838 return vma->vm_ops->find_special_page(vma, addr);
839 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
841 if (is_zero_pfn(pfn))
845 * Device public pages are special pages (they are ZONE_DEVICE
846 * pages but different from persistent memory). They behave
847 * allmost like normal pages. The difference is that they are
848 * not on the lru and thus should never be involve with any-
849 * thing that involve lru manipulation (mlock, numa balancing,
852 * This is why we still want to return NULL for such page from
853 * vm_normal_page() so that we do not have to special case all
854 * call site of vm_normal_page().
856 if (likely(pfn <= highest_memmap_pfn)) {
857 struct page *page = pfn_to_page(pfn);
859 if (is_device_public_page(page)) {
860 if (with_public_device)
869 print_bad_pte(vma, addr, pte, NULL);
873 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
875 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
876 if (vma->vm_flags & VM_MIXEDMAP) {
882 off = (addr - vma->vm_start) >> PAGE_SHIFT;
883 if (pfn == vma->vm_pgoff + off)
885 if (!is_cow_mapping(vma->vm_flags))
890 if (is_zero_pfn(pfn))
894 if (unlikely(pfn > highest_memmap_pfn)) {
895 print_bad_pte(vma, addr, pte, NULL);
900 * NOTE! We still have PageReserved() pages in the page tables.
901 * eg. VDSO mappings can cause them to exist.
904 return pfn_to_page(pfn);
907 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
908 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
911 unsigned long pfn = pmd_pfn(pmd);
914 * There is no pmd_special() but there may be special pmds, e.g.
915 * in a direct-access (dax) mapping, so let's just replicate the
916 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
918 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
919 if (vma->vm_flags & VM_MIXEDMAP) {
925 off = (addr - vma->vm_start) >> PAGE_SHIFT;
926 if (pfn == vma->vm_pgoff + off)
928 if (!is_cow_mapping(vma->vm_flags))
935 if (is_zero_pfn(pfn))
937 if (unlikely(pfn > highest_memmap_pfn))
941 * NOTE! We still have PageReserved() pages in the page tables.
942 * eg. VDSO mappings can cause them to exist.
945 return pfn_to_page(pfn);
950 * copy one vm_area from one task to the other. Assumes the page tables
951 * already present in the new task to be cleared in the whole range
952 * covered by this vma.
955 static inline unsigned long
956 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
957 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
958 unsigned long addr, int *rss)
960 unsigned long vm_flags = vma->vm_flags;
961 pte_t pte = *src_pte;
964 /* pte contains position in swap or file, so copy. */
965 if (unlikely(!pte_present(pte))) {
966 swp_entry_t entry = pte_to_swp_entry(pte);
968 if (likely(!non_swap_entry(entry))) {
969 if (swap_duplicate(entry) < 0)
972 /* make sure dst_mm is on swapoff's mmlist. */
973 if (unlikely(list_empty(&dst_mm->mmlist))) {
974 spin_lock(&mmlist_lock);
975 if (list_empty(&dst_mm->mmlist))
976 list_add(&dst_mm->mmlist,
978 spin_unlock(&mmlist_lock);
981 } else if (is_migration_entry(entry)) {
982 page = migration_entry_to_page(entry);
984 rss[mm_counter(page)]++;
986 if (is_write_migration_entry(entry) &&
987 is_cow_mapping(vm_flags)) {
989 * COW mappings require pages in both
990 * parent and child to be set to read.
992 make_migration_entry_read(&entry);
993 pte = swp_entry_to_pte(entry);
994 if (pte_swp_soft_dirty(*src_pte))
995 pte = pte_swp_mksoft_dirty(pte);
996 set_pte_at(src_mm, addr, src_pte, pte);
998 } else if (is_device_private_entry(entry)) {
999 page = device_private_entry_to_page(entry);
1002 * Update rss count even for unaddressable pages, as
1003 * they should treated just like normal pages in this
1006 * We will likely want to have some new rss counters
1007 * for unaddressable pages, at some point. But for now
1008 * keep things as they are.
1011 rss[mm_counter(page)]++;
1012 page_dup_rmap(page, false);
1015 * We do not preserve soft-dirty information, because so
1016 * far, checkpoint/restore is the only feature that
1017 * requires that. And checkpoint/restore does not work
1018 * when a device driver is involved (you cannot easily
1019 * save and restore device driver state).
1021 if (is_write_device_private_entry(entry) &&
1022 is_cow_mapping(vm_flags)) {
1023 make_device_private_entry_read(&entry);
1024 pte = swp_entry_to_pte(entry);
1025 set_pte_at(src_mm, addr, src_pte, pte);
1032 * If it's a COW mapping, write protect it both
1033 * in the parent and the child
1035 if (is_cow_mapping(vm_flags)) {
1036 ptep_set_wrprotect(src_mm, addr, src_pte);
1037 pte = pte_wrprotect(pte);
1041 * If it's a shared mapping, mark it clean in
1044 if (vm_flags & VM_SHARED)
1045 pte = pte_mkclean(pte);
1046 pte = pte_mkold(pte);
1048 page = vm_normal_page(vma, addr, pte);
1051 page_dup_rmap(page, false);
1052 rss[mm_counter(page)]++;
1053 } else if (pte_devmap(pte)) {
1054 page = pte_page(pte);
1057 * Cache coherent device memory behave like regular page and
1058 * not like persistent memory page. For more informations see
1059 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1061 if (is_device_public_page(page)) {
1063 page_dup_rmap(page, false);
1064 rss[mm_counter(page)]++;
1069 set_pte_at(dst_mm, addr, dst_pte, pte);
1073 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1074 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1075 unsigned long addr, unsigned long end)
1077 pte_t *orig_src_pte, *orig_dst_pte;
1078 pte_t *src_pte, *dst_pte;
1079 spinlock_t *src_ptl, *dst_ptl;
1081 int rss[NR_MM_COUNTERS];
1082 swp_entry_t entry = (swp_entry_t){0};
1087 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1090 src_pte = pte_offset_map(src_pmd, addr);
1091 src_ptl = pte_lockptr(src_mm, src_pmd);
1092 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1093 orig_src_pte = src_pte;
1094 orig_dst_pte = dst_pte;
1095 arch_enter_lazy_mmu_mode();
1099 * We are holding two locks at this point - either of them
1100 * could generate latencies in another task on another CPU.
1102 if (progress >= 32) {
1104 if (need_resched() ||
1105 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1108 if (pte_none(*src_pte)) {
1112 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1117 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1119 arch_leave_lazy_mmu_mode();
1120 spin_unlock(src_ptl);
1121 pte_unmap(orig_src_pte);
1122 add_mm_rss_vec(dst_mm, rss);
1123 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1127 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1136 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1137 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1138 unsigned long addr, unsigned long end)
1140 pmd_t *src_pmd, *dst_pmd;
1143 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1146 src_pmd = pmd_offset(src_pud, addr);
1148 next = pmd_addr_end(addr, end);
1149 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1150 || pmd_devmap(*src_pmd)) {
1152 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1153 err = copy_huge_pmd(dst_mm, src_mm,
1154 dst_pmd, src_pmd, addr, vma);
1161 if (pmd_none_or_clear_bad(src_pmd))
1163 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1166 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1170 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1171 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1172 unsigned long addr, unsigned long end)
1174 pud_t *src_pud, *dst_pud;
1177 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1180 src_pud = pud_offset(src_p4d, addr);
1182 next = pud_addr_end(addr, end);
1183 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1186 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1187 err = copy_huge_pud(dst_mm, src_mm,
1188 dst_pud, src_pud, addr, vma);
1195 if (pud_none_or_clear_bad(src_pud))
1197 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1200 } while (dst_pud++, src_pud++, addr = next, addr != end);
1204 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1205 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1206 unsigned long addr, unsigned long end)
1208 p4d_t *src_p4d, *dst_p4d;
1211 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1214 src_p4d = p4d_offset(src_pgd, addr);
1216 next = p4d_addr_end(addr, end);
1217 if (p4d_none_or_clear_bad(src_p4d))
1219 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1222 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1226 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1227 struct vm_area_struct *vma)
1229 pgd_t *src_pgd, *dst_pgd;
1231 unsigned long addr = vma->vm_start;
1232 unsigned long end = vma->vm_end;
1233 unsigned long mmun_start; /* For mmu_notifiers */
1234 unsigned long mmun_end; /* For mmu_notifiers */
1239 * Don't copy ptes where a page fault will fill them correctly.
1240 * Fork becomes much lighter when there are big shared or private
1241 * readonly mappings. The tradeoff is that copy_page_range is more
1242 * efficient than faulting.
1244 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1248 if (is_vm_hugetlb_page(vma))
1249 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1251 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1253 * We do not free on error cases below as remove_vma
1254 * gets called on error from higher level routine
1256 ret = track_pfn_copy(vma);
1262 * We need to invalidate the secondary MMU mappings only when
1263 * there could be a permission downgrade on the ptes of the
1264 * parent mm. And a permission downgrade will only happen if
1265 * is_cow_mapping() returns true.
1267 is_cow = is_cow_mapping(vma->vm_flags);
1271 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1275 dst_pgd = pgd_offset(dst_mm, addr);
1276 src_pgd = pgd_offset(src_mm, addr);
1278 next = pgd_addr_end(addr, end);
1279 if (pgd_none_or_clear_bad(src_pgd))
1281 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1282 vma, addr, next))) {
1286 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1289 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1293 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1294 struct vm_area_struct *vma, pmd_t *pmd,
1295 unsigned long addr, unsigned long end,
1296 struct zap_details *details)
1298 struct mm_struct *mm = tlb->mm;
1299 int force_flush = 0;
1300 int rss[NR_MM_COUNTERS];
1306 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1309 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1311 flush_tlb_batched_pending(mm);
1312 arch_enter_lazy_mmu_mode();
1315 if (pte_none(ptent))
1318 if (pte_present(ptent)) {
1321 page = _vm_normal_page(vma, addr, ptent, true);
1322 if (unlikely(details) && page) {
1324 * unmap_shared_mapping_pages() wants to
1325 * invalidate cache without truncating:
1326 * unmap shared but keep private pages.
1328 if (details->check_mapping &&
1329 details->check_mapping != page_rmapping(page))
1332 ptent = ptep_get_and_clear_full(mm, addr, pte,
1334 tlb_remove_tlb_entry(tlb, pte, addr);
1335 if (unlikely(!page))
1338 if (!PageAnon(page)) {
1339 if (pte_dirty(ptent)) {
1341 set_page_dirty(page);
1343 if (pte_young(ptent) &&
1344 likely(!(vma->vm_flags & VM_SEQ_READ)))
1345 mark_page_accessed(page);
1347 rss[mm_counter(page)]--;
1348 page_remove_rmap(page, false);
1349 if (unlikely(page_mapcount(page) < 0))
1350 print_bad_pte(vma, addr, ptent, page);
1351 if (unlikely(__tlb_remove_page(tlb, page))) {
1359 entry = pte_to_swp_entry(ptent);
1360 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1361 struct page *page = device_private_entry_to_page(entry);
1363 if (unlikely(details && details->check_mapping)) {
1365 * unmap_shared_mapping_pages() wants to
1366 * invalidate cache without truncating:
1367 * unmap shared but keep private pages.
1369 if (details->check_mapping !=
1370 page_rmapping(page))
1374 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1375 rss[mm_counter(page)]--;
1376 page_remove_rmap(page, false);
1381 /* If details->check_mapping, we leave swap entries. */
1382 if (unlikely(details))
1385 entry = pte_to_swp_entry(ptent);
1386 if (!non_swap_entry(entry))
1388 else if (is_migration_entry(entry)) {
1391 page = migration_entry_to_page(entry);
1392 rss[mm_counter(page)]--;
1394 if (unlikely(!free_swap_and_cache(entry)))
1395 print_bad_pte(vma, addr, ptent, NULL);
1396 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1397 } while (pte++, addr += PAGE_SIZE, addr != end);
1399 add_mm_rss_vec(mm, rss);
1400 arch_leave_lazy_mmu_mode();
1402 /* Do the actual TLB flush before dropping ptl */
1404 tlb_flush_mmu_tlbonly(tlb);
1405 pte_unmap_unlock(start_pte, ptl);
1408 * If we forced a TLB flush (either due to running out of
1409 * batch buffers or because we needed to flush dirty TLB
1410 * entries before releasing the ptl), free the batched
1411 * memory too. Restart if we didn't do everything.
1415 tlb_flush_mmu_free(tlb);
1423 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1424 struct vm_area_struct *vma, pud_t *pud,
1425 unsigned long addr, unsigned long end,
1426 struct zap_details *details)
1431 pmd = pmd_offset(pud, addr);
1433 next = pmd_addr_end(addr, end);
1434 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1435 if (next - addr != HPAGE_PMD_SIZE)
1436 __split_huge_pmd(vma, pmd, addr, false, NULL);
1437 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1442 * Here there can be other concurrent MADV_DONTNEED or
1443 * trans huge page faults running, and if the pmd is
1444 * none or trans huge it can change under us. This is
1445 * because MADV_DONTNEED holds the mmap_sem in read
1448 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1450 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1453 } while (pmd++, addr = next, addr != end);
1458 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1459 struct vm_area_struct *vma, p4d_t *p4d,
1460 unsigned long addr, unsigned long end,
1461 struct zap_details *details)
1466 pud = pud_offset(p4d, addr);
1468 next = pud_addr_end(addr, end);
1469 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1470 if (next - addr != HPAGE_PUD_SIZE) {
1471 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1472 split_huge_pud(vma, pud, addr);
1473 } else if (zap_huge_pud(tlb, vma, pud, addr))
1477 if (pud_none_or_clear_bad(pud))
1479 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1482 } while (pud++, addr = next, addr != end);
1487 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1488 struct vm_area_struct *vma, pgd_t *pgd,
1489 unsigned long addr, unsigned long end,
1490 struct zap_details *details)
1495 p4d = p4d_offset(pgd, addr);
1497 next = p4d_addr_end(addr, end);
1498 if (p4d_none_or_clear_bad(p4d))
1500 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1501 } while (p4d++, addr = next, addr != end);
1506 void unmap_page_range(struct mmu_gather *tlb,
1507 struct vm_area_struct *vma,
1508 unsigned long addr, unsigned long end,
1509 struct zap_details *details)
1514 BUG_ON(addr >= end);
1515 tlb_start_vma(tlb, vma);
1516 pgd = pgd_offset(vma->vm_mm, addr);
1518 next = pgd_addr_end(addr, end);
1519 if (pgd_none_or_clear_bad(pgd))
1521 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1522 } while (pgd++, addr = next, addr != end);
1523 tlb_end_vma(tlb, vma);
1527 static void unmap_single_vma(struct mmu_gather *tlb,
1528 struct vm_area_struct *vma, unsigned long start_addr,
1529 unsigned long end_addr,
1530 struct zap_details *details)
1532 unsigned long start = max(vma->vm_start, start_addr);
1535 if (start >= vma->vm_end)
1537 end = min(vma->vm_end, end_addr);
1538 if (end <= vma->vm_start)
1542 uprobe_munmap(vma, start, end);
1544 if (unlikely(vma->vm_flags & VM_PFNMAP))
1545 untrack_pfn(vma, 0, 0);
1548 if (unlikely(is_vm_hugetlb_page(vma))) {
1550 * It is undesirable to test vma->vm_file as it
1551 * should be non-null for valid hugetlb area.
1552 * However, vm_file will be NULL in the error
1553 * cleanup path of mmap_region. When
1554 * hugetlbfs ->mmap method fails,
1555 * mmap_region() nullifies vma->vm_file
1556 * before calling this function to clean up.
1557 * Since no pte has actually been setup, it is
1558 * safe to do nothing in this case.
1561 i_mmap_lock_write(vma->vm_file->f_mapping);
1562 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1563 i_mmap_unlock_write(vma->vm_file->f_mapping);
1566 unmap_page_range(tlb, vma, start, end, details);
1571 * unmap_vmas - unmap a range of memory covered by a list of vma's
1572 * @tlb: address of the caller's struct mmu_gather
1573 * @vma: the starting vma
1574 * @start_addr: virtual address at which to start unmapping
1575 * @end_addr: virtual address at which to end unmapping
1577 * Unmap all pages in the vma list.
1579 * Only addresses between `start' and `end' will be unmapped.
1581 * The VMA list must be sorted in ascending virtual address order.
1583 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1584 * range after unmap_vmas() returns. So the only responsibility here is to
1585 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1586 * drops the lock and schedules.
1588 void unmap_vmas(struct mmu_gather *tlb,
1589 struct vm_area_struct *vma, unsigned long start_addr,
1590 unsigned long end_addr)
1592 struct mm_struct *mm = vma->vm_mm;
1594 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1595 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1596 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1597 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1601 * zap_page_range - remove user pages in a given range
1602 * @vma: vm_area_struct holding the applicable pages
1603 * @start: starting address of pages to zap
1604 * @size: number of bytes to zap
1606 * Caller must protect the VMA list
1608 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1611 struct mm_struct *mm = vma->vm_mm;
1612 struct mmu_gather tlb;
1613 unsigned long end = start + size;
1616 tlb_gather_mmu(&tlb, mm, start, end);
1617 update_hiwater_rss(mm);
1618 mmu_notifier_invalidate_range_start(mm, start, end);
1619 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1620 unmap_single_vma(&tlb, vma, start, end, NULL);
1621 mmu_notifier_invalidate_range_end(mm, start, end);
1622 tlb_finish_mmu(&tlb, start, end);
1626 * zap_page_range_single - remove user pages in a given range
1627 * @vma: vm_area_struct holding the applicable pages
1628 * @address: starting address of pages to zap
1629 * @size: number of bytes to zap
1630 * @details: details of shared cache invalidation
1632 * The range must fit into one VMA.
1634 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1635 unsigned long size, struct zap_details *details)
1637 struct mm_struct *mm = vma->vm_mm;
1638 struct mmu_gather tlb;
1639 unsigned long end = address + size;
1642 tlb_gather_mmu(&tlb, mm, address, end);
1643 update_hiwater_rss(mm);
1644 mmu_notifier_invalidate_range_start(mm, address, end);
1645 unmap_single_vma(&tlb, vma, address, end, details);
1646 mmu_notifier_invalidate_range_end(mm, address, end);
1647 tlb_finish_mmu(&tlb, address, end);
1651 * zap_vma_ptes - remove ptes mapping the vma
1652 * @vma: vm_area_struct holding ptes to be zapped
1653 * @address: starting address of pages to zap
1654 * @size: number of bytes to zap
1656 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1658 * The entire address range must be fully contained within the vma.
1661 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1664 if (address < vma->vm_start || address + size > vma->vm_end ||
1665 !(vma->vm_flags & VM_PFNMAP))
1668 zap_page_range_single(vma, address, size, NULL);
1670 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1672 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1680 pgd = pgd_offset(mm, addr);
1681 p4d = p4d_alloc(mm, pgd, addr);
1684 pud = pud_alloc(mm, p4d, addr);
1687 pmd = pmd_alloc(mm, pud, addr);
1691 VM_BUG_ON(pmd_trans_huge(*pmd));
1692 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1696 * This is the old fallback for page remapping.
1698 * For historical reasons, it only allows reserved pages. Only
1699 * old drivers should use this, and they needed to mark their
1700 * pages reserved for the old functions anyway.
1702 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1703 struct page *page, pgprot_t prot)
1705 struct mm_struct *mm = vma->vm_mm;
1714 flush_dcache_page(page);
1715 pte = get_locked_pte(mm, addr, &ptl);
1719 if (!pte_none(*pte))
1722 /* Ok, finally just insert the thing.. */
1724 inc_mm_counter_fast(mm, mm_counter_file(page));
1725 page_add_file_rmap(page, false);
1726 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1729 pte_unmap_unlock(pte, ptl);
1732 pte_unmap_unlock(pte, ptl);
1738 * vm_insert_page - insert single page into user vma
1739 * @vma: user vma to map to
1740 * @addr: target user address of this page
1741 * @page: source kernel page
1743 * This allows drivers to insert individual pages they've allocated
1746 * The page has to be a nice clean _individual_ kernel allocation.
1747 * If you allocate a compound page, you need to have marked it as
1748 * such (__GFP_COMP), or manually just split the page up yourself
1749 * (see split_page()).
1751 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1752 * took an arbitrary page protection parameter. This doesn't allow
1753 * that. Your vma protection will have to be set up correctly, which
1754 * means that if you want a shared writable mapping, you'd better
1755 * ask for a shared writable mapping!
1757 * The page does not need to be reserved.
1759 * Usually this function is called from f_op->mmap() handler
1760 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1761 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1762 * function from other places, for example from page-fault handler.
1764 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1767 if (addr < vma->vm_start || addr >= vma->vm_end)
1769 if (!page_count(page))
1771 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1772 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1773 BUG_ON(vma->vm_flags & VM_PFNMAP);
1774 vma->vm_flags |= VM_MIXEDMAP;
1776 return insert_page(vma, addr, page, vma->vm_page_prot);
1778 EXPORT_SYMBOL(vm_insert_page);
1780 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1781 pfn_t pfn, pgprot_t prot, bool mkwrite)
1783 struct mm_struct *mm = vma->vm_mm;
1789 pte = get_locked_pte(mm, addr, &ptl);
1793 if (!pte_none(*pte)) {
1796 * For read faults on private mappings the PFN passed
1797 * in may not match the PFN we have mapped if the
1798 * mapped PFN is a writeable COW page. In the mkwrite
1799 * case we are creating a writable PTE for a shared
1800 * mapping and we expect the PFNs to match.
1802 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1810 /* Ok, finally just insert the thing.. */
1811 if (pfn_t_devmap(pfn))
1812 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1814 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1818 entry = pte_mkyoung(entry);
1819 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1822 set_pte_at(mm, addr, pte, entry);
1823 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1827 pte_unmap_unlock(pte, ptl);
1833 * vm_insert_pfn - insert single pfn into user vma
1834 * @vma: user vma to map to
1835 * @addr: target user address of this page
1836 * @pfn: source kernel pfn
1838 * Similar to vm_insert_page, this allows drivers to insert individual pages
1839 * they've allocated into a user vma. Same comments apply.
1841 * This function should only be called from a vm_ops->fault handler, and
1842 * in that case the handler should return NULL.
1844 * vma cannot be a COW mapping.
1846 * As this is called only for pages that do not currently exist, we
1847 * do not need to flush old virtual caches or the TLB.
1849 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1852 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1854 EXPORT_SYMBOL(vm_insert_pfn);
1857 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1858 * @vma: user vma to map to
1859 * @addr: target user address of this page
1860 * @pfn: source kernel pfn
1861 * @pgprot: pgprot flags for the inserted page
1863 * This is exactly like vm_insert_pfn, except that it allows drivers to
1864 * to override pgprot on a per-page basis.
1866 * This only makes sense for IO mappings, and it makes no sense for
1867 * cow mappings. In general, using multiple vmas is preferable;
1868 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1871 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1872 unsigned long pfn, pgprot_t pgprot)
1876 * Technically, architectures with pte_special can avoid all these
1877 * restrictions (same for remap_pfn_range). However we would like
1878 * consistency in testing and feature parity among all, so we should
1879 * try to keep these invariants in place for everybody.
1881 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1882 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1883 (VM_PFNMAP|VM_MIXEDMAP));
1884 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1885 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1887 if (addr < vma->vm_start || addr >= vma->vm_end)
1890 if (!pfn_modify_allowed(pfn, pgprot))
1893 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1895 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1900 EXPORT_SYMBOL(vm_insert_pfn_prot);
1902 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1904 /* these checks mirror the abort conditions in vm_normal_page */
1905 if (vma->vm_flags & VM_MIXEDMAP)
1907 if (pfn_t_devmap(pfn))
1909 if (pfn_t_special(pfn))
1911 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1916 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1917 pfn_t pfn, bool mkwrite)
1919 pgprot_t pgprot = vma->vm_page_prot;
1921 BUG_ON(!vm_mixed_ok(vma, pfn));
1923 if (addr < vma->vm_start || addr >= vma->vm_end)
1926 track_pfn_insert(vma, &pgprot, pfn);
1928 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1932 * If we don't have pte special, then we have to use the pfn_valid()
1933 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1934 * refcount the page if pfn_valid is true (hence insert_page rather
1935 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1936 * without pte special, it would there be refcounted as a normal page.
1938 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1939 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1943 * At this point we are committed to insert_page()
1944 * regardless of whether the caller specified flags that
1945 * result in pfn_t_has_page() == false.
1947 page = pfn_to_page(pfn_t_to_pfn(pfn));
1948 return insert_page(vma, addr, page, pgprot);
1950 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1953 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1956 return __vm_insert_mixed(vma, addr, pfn, false);
1959 EXPORT_SYMBOL(vm_insert_mixed);
1962 * If the insertion of PTE failed because someone else already added a
1963 * different entry in the mean time, we treat that as success as we assume
1964 * the same entry was actually inserted.
1967 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1968 unsigned long addr, pfn_t pfn)
1972 err = __vm_insert_mixed(vma, addr, pfn, true);
1974 return VM_FAULT_OOM;
1975 if (err < 0 && err != -EBUSY)
1976 return VM_FAULT_SIGBUS;
1977 return VM_FAULT_NOPAGE;
1979 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1982 * maps a range of physical memory into the requested pages. the old
1983 * mappings are removed. any references to nonexistent pages results
1984 * in null mappings (currently treated as "copy-on-access")
1986 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1987 unsigned long addr, unsigned long end,
1988 unsigned long pfn, pgprot_t prot)
1994 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1997 arch_enter_lazy_mmu_mode();
1999 BUG_ON(!pte_none(*pte));
2000 if (!pfn_modify_allowed(pfn, prot)) {
2004 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2006 } while (pte++, addr += PAGE_SIZE, addr != end);
2007 arch_leave_lazy_mmu_mode();
2008 pte_unmap_unlock(pte - 1, ptl);
2012 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2013 unsigned long addr, unsigned long end,
2014 unsigned long pfn, pgprot_t prot)
2020 pfn -= addr >> PAGE_SHIFT;
2021 pmd = pmd_alloc(mm, pud, addr);
2024 VM_BUG_ON(pmd_trans_huge(*pmd));
2026 next = pmd_addr_end(addr, end);
2027 err = remap_pte_range(mm, pmd, addr, next,
2028 pfn + (addr >> PAGE_SHIFT), prot);
2031 } while (pmd++, addr = next, addr != end);
2035 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2036 unsigned long addr, unsigned long end,
2037 unsigned long pfn, pgprot_t prot)
2043 pfn -= addr >> PAGE_SHIFT;
2044 pud = pud_alloc(mm, p4d, addr);
2048 next = pud_addr_end(addr, end);
2049 err = remap_pmd_range(mm, pud, addr, next,
2050 pfn + (addr >> PAGE_SHIFT), prot);
2053 } while (pud++, addr = next, addr != end);
2057 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2058 unsigned long addr, unsigned long end,
2059 unsigned long pfn, pgprot_t prot)
2065 pfn -= addr >> PAGE_SHIFT;
2066 p4d = p4d_alloc(mm, pgd, addr);
2070 next = p4d_addr_end(addr, end);
2071 err = remap_pud_range(mm, p4d, addr, next,
2072 pfn + (addr >> PAGE_SHIFT), prot);
2075 } while (p4d++, addr = next, addr != end);
2080 * remap_pfn_range - remap kernel memory to userspace
2081 * @vma: user vma to map to
2082 * @addr: target user address to start at
2083 * @pfn: physical address of kernel memory
2084 * @size: size of map area
2085 * @prot: page protection flags for this mapping
2087 * Note: this is only safe if the mm semaphore is held when called.
2089 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2090 unsigned long pfn, unsigned long size, pgprot_t prot)
2094 unsigned long end = addr + PAGE_ALIGN(size);
2095 struct mm_struct *mm = vma->vm_mm;
2096 unsigned long remap_pfn = pfn;
2100 * Physically remapped pages are special. Tell the
2101 * rest of the world about it:
2102 * VM_IO tells people not to look at these pages
2103 * (accesses can have side effects).
2104 * VM_PFNMAP tells the core MM that the base pages are just
2105 * raw PFN mappings, and do not have a "struct page" associated
2108 * Disable vma merging and expanding with mremap().
2110 * Omit vma from core dump, even when VM_IO turned off.
2112 * There's a horrible special case to handle copy-on-write
2113 * behaviour that some programs depend on. We mark the "original"
2114 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2115 * See vm_normal_page() for details.
2117 if (is_cow_mapping(vma->vm_flags)) {
2118 if (addr != vma->vm_start || end != vma->vm_end)
2120 vma->vm_pgoff = pfn;
2123 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2127 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2129 BUG_ON(addr >= end);
2130 pfn -= addr >> PAGE_SHIFT;
2131 pgd = pgd_offset(mm, addr);
2132 flush_cache_range(vma, addr, end);
2134 next = pgd_addr_end(addr, end);
2135 err = remap_p4d_range(mm, pgd, addr, next,
2136 pfn + (addr >> PAGE_SHIFT), prot);
2139 } while (pgd++, addr = next, addr != end);
2142 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2146 EXPORT_SYMBOL(remap_pfn_range);
2149 * vm_iomap_memory - remap memory to userspace
2150 * @vma: user vma to map to
2151 * @start: start of area
2152 * @len: size of area
2154 * This is a simplified io_remap_pfn_range() for common driver use. The
2155 * driver just needs to give us the physical memory range to be mapped,
2156 * we'll figure out the rest from the vma information.
2158 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2159 * whatever write-combining details or similar.
2161 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2163 unsigned long vm_len, pfn, pages;
2165 /* Check that the physical memory area passed in looks valid */
2166 if (start + len < start)
2169 * You *really* shouldn't map things that aren't page-aligned,
2170 * but we've historically allowed it because IO memory might
2171 * just have smaller alignment.
2173 len += start & ~PAGE_MASK;
2174 pfn = start >> PAGE_SHIFT;
2175 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2176 if (pfn + pages < pfn)
2179 /* We start the mapping 'vm_pgoff' pages into the area */
2180 if (vma->vm_pgoff > pages)
2182 pfn += vma->vm_pgoff;
2183 pages -= vma->vm_pgoff;
2185 /* Can we fit all of the mapping? */
2186 vm_len = vma->vm_end - vma->vm_start;
2187 if (vm_len >> PAGE_SHIFT > pages)
2190 /* Ok, let it rip */
2191 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2193 EXPORT_SYMBOL(vm_iomap_memory);
2195 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2196 unsigned long addr, unsigned long end,
2197 pte_fn_t fn, void *data)
2202 spinlock_t *uninitialized_var(ptl);
2204 pte = (mm == &init_mm) ?
2205 pte_alloc_kernel(pmd, addr) :
2206 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2210 BUG_ON(pmd_huge(*pmd));
2212 arch_enter_lazy_mmu_mode();
2214 token = pmd_pgtable(*pmd);
2217 err = fn(pte++, token, addr, data);
2220 } while (addr += PAGE_SIZE, addr != end);
2222 arch_leave_lazy_mmu_mode();
2225 pte_unmap_unlock(pte-1, ptl);
2229 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2230 unsigned long addr, unsigned long end,
2231 pte_fn_t fn, void *data)
2237 BUG_ON(pud_huge(*pud));
2239 pmd = pmd_alloc(mm, pud, addr);
2243 next = pmd_addr_end(addr, end);
2244 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2247 } while (pmd++, addr = next, addr != end);
2251 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2252 unsigned long addr, unsigned long end,
2253 pte_fn_t fn, void *data)
2259 pud = pud_alloc(mm, p4d, addr);
2263 next = pud_addr_end(addr, end);
2264 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2267 } while (pud++, addr = next, addr != end);
2271 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2272 unsigned long addr, unsigned long end,
2273 pte_fn_t fn, void *data)
2279 p4d = p4d_alloc(mm, pgd, addr);
2283 next = p4d_addr_end(addr, end);
2284 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2287 } while (p4d++, addr = next, addr != end);
2292 * Scan a region of virtual memory, filling in page tables as necessary
2293 * and calling a provided function on each leaf page table.
2295 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2296 unsigned long size, pte_fn_t fn, void *data)
2300 unsigned long end = addr + size;
2303 if (WARN_ON(addr >= end))
2306 pgd = pgd_offset(mm, addr);
2308 next = pgd_addr_end(addr, end);
2309 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2312 } while (pgd++, addr = next, addr != end);
2316 EXPORT_SYMBOL_GPL(apply_to_page_range);
2319 * handle_pte_fault chooses page fault handler according to an entry which was
2320 * read non-atomically. Before making any commitment, on those architectures
2321 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2322 * parts, do_swap_page must check under lock before unmapping the pte and
2323 * proceeding (but do_wp_page is only called after already making such a check;
2324 * and do_anonymous_page can safely check later on).
2326 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2327 pte_t *page_table, pte_t orig_pte)
2330 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2331 if (sizeof(pte_t) > sizeof(unsigned long)) {
2332 spinlock_t *ptl = pte_lockptr(mm, pmd);
2334 same = pte_same(*page_table, orig_pte);
2338 pte_unmap(page_table);
2342 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2344 debug_dma_assert_idle(src);
2347 * If the source page was a PFN mapping, we don't have
2348 * a "struct page" for it. We do a best-effort copy by
2349 * just copying from the original user address. If that
2350 * fails, we just zero-fill it. Live with it.
2352 if (unlikely(!src)) {
2353 void *kaddr = kmap_atomic(dst);
2354 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2357 * This really shouldn't fail, because the page is there
2358 * in the page tables. But it might just be unreadable,
2359 * in which case we just give up and fill the result with
2362 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2364 kunmap_atomic(kaddr);
2365 flush_dcache_page(dst);
2367 copy_user_highpage(dst, src, va, vma);
2370 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2372 struct file *vm_file = vma->vm_file;
2375 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2378 * Special mappings (e.g. VDSO) do not have any file so fake
2379 * a default GFP_KERNEL for them.
2385 * Notify the address space that the page is about to become writable so that
2386 * it can prohibit this or wait for the page to get into an appropriate state.
2388 * We do this without the lock held, so that it can sleep if it needs to.
2390 static int do_page_mkwrite(struct vm_fault *vmf)
2393 struct page *page = vmf->page;
2394 unsigned int old_flags = vmf->flags;
2396 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2398 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2399 /* Restore original flags so that caller is not surprised */
2400 vmf->flags = old_flags;
2401 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2403 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2405 if (!page->mapping) {
2407 return 0; /* retry */
2409 ret |= VM_FAULT_LOCKED;
2411 VM_BUG_ON_PAGE(!PageLocked(page), page);
2416 * Handle dirtying of a page in shared file mapping on a write fault.
2418 * The function expects the page to be locked and unlocks it.
2420 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2423 struct address_space *mapping;
2425 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2427 dirtied = set_page_dirty(page);
2428 VM_BUG_ON_PAGE(PageAnon(page), page);
2430 * Take a local copy of the address_space - page.mapping may be zeroed
2431 * by truncate after unlock_page(). The address_space itself remains
2432 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2433 * release semantics to prevent the compiler from undoing this copying.
2435 mapping = page_rmapping(page);
2438 if ((dirtied || page_mkwrite) && mapping) {
2440 * Some device drivers do not set page.mapping
2441 * but still dirty their pages
2443 balance_dirty_pages_ratelimited(mapping);
2447 file_update_time(vma->vm_file);
2451 * Handle write page faults for pages that can be reused in the current vma
2453 * This can happen either due to the mapping being with the VM_SHARED flag,
2454 * or due to us being the last reference standing to the page. In either
2455 * case, all we need to do here is to mark the page as writable and update
2456 * any related book-keeping.
2458 static inline void wp_page_reuse(struct vm_fault *vmf)
2459 __releases(vmf->ptl)
2461 struct vm_area_struct *vma = vmf->vma;
2462 struct page *page = vmf->page;
2465 * Clear the pages cpupid information as the existing
2466 * information potentially belongs to a now completely
2467 * unrelated process.
2470 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2472 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2473 entry = pte_mkyoung(vmf->orig_pte);
2474 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2475 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2476 update_mmu_cache(vma, vmf->address, vmf->pte);
2477 pte_unmap_unlock(vmf->pte, vmf->ptl);
2481 * Handle the case of a page which we actually need to copy to a new page.
2483 * Called with mmap_sem locked and the old page referenced, but
2484 * without the ptl held.
2486 * High level logic flow:
2488 * - Allocate a page, copy the content of the old page to the new one.
2489 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2490 * - Take the PTL. If the pte changed, bail out and release the allocated page
2491 * - If the pte is still the way we remember it, update the page table and all
2492 * relevant references. This includes dropping the reference the page-table
2493 * held to the old page, as well as updating the rmap.
2494 * - In any case, unlock the PTL and drop the reference we took to the old page.
2496 static int wp_page_copy(struct vm_fault *vmf)
2498 struct vm_area_struct *vma = vmf->vma;
2499 struct mm_struct *mm = vma->vm_mm;
2500 struct page *old_page = vmf->page;
2501 struct page *new_page = NULL;
2503 int page_copied = 0;
2504 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2505 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2506 struct mem_cgroup *memcg;
2508 if (unlikely(anon_vma_prepare(vma)))
2511 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2512 new_page = alloc_zeroed_user_highpage_movable(vma,
2517 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2521 cow_user_page(new_page, old_page, vmf->address, vma);
2524 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2527 __SetPageUptodate(new_page);
2529 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2532 * Re-check the pte - we dropped the lock
2534 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2535 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2537 if (!PageAnon(old_page)) {
2538 dec_mm_counter_fast(mm,
2539 mm_counter_file(old_page));
2540 inc_mm_counter_fast(mm, MM_ANONPAGES);
2543 inc_mm_counter_fast(mm, MM_ANONPAGES);
2545 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2546 entry = mk_pte(new_page, vma->vm_page_prot);
2547 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2549 * Clear the pte entry and flush it first, before updating the
2550 * pte with the new entry. This will avoid a race condition
2551 * seen in the presence of one thread doing SMC and another
2554 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2555 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2556 mem_cgroup_commit_charge(new_page, memcg, false, false);
2557 lru_cache_add_active_or_unevictable(new_page, vma);
2559 * We call the notify macro here because, when using secondary
2560 * mmu page tables (such as kvm shadow page tables), we want the
2561 * new page to be mapped directly into the secondary page table.
2563 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2564 update_mmu_cache(vma, vmf->address, vmf->pte);
2567 * Only after switching the pte to the new page may
2568 * we remove the mapcount here. Otherwise another
2569 * process may come and find the rmap count decremented
2570 * before the pte is switched to the new page, and
2571 * "reuse" the old page writing into it while our pte
2572 * here still points into it and can be read by other
2575 * The critical issue is to order this
2576 * page_remove_rmap with the ptp_clear_flush above.
2577 * Those stores are ordered by (if nothing else,)
2578 * the barrier present in the atomic_add_negative
2579 * in page_remove_rmap.
2581 * Then the TLB flush in ptep_clear_flush ensures that
2582 * no process can access the old page before the
2583 * decremented mapcount is visible. And the old page
2584 * cannot be reused until after the decremented
2585 * mapcount is visible. So transitively, TLBs to
2586 * old page will be flushed before it can be reused.
2588 page_remove_rmap(old_page, false);
2591 /* Free the old page.. */
2592 new_page = old_page;
2595 mem_cgroup_cancel_charge(new_page, memcg, false);
2601 pte_unmap_unlock(vmf->pte, vmf->ptl);
2603 * No need to double call mmu_notifier->invalidate_range() callback as
2604 * the above ptep_clear_flush_notify() did already call it.
2606 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2609 * Don't let another task, with possibly unlocked vma,
2610 * keep the mlocked page.
2612 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2613 lock_page(old_page); /* LRU manipulation */
2614 if (PageMlocked(old_page))
2615 munlock_vma_page(old_page);
2616 unlock_page(old_page);
2620 return page_copied ? VM_FAULT_WRITE : 0;
2626 return VM_FAULT_OOM;
2630 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2631 * writeable once the page is prepared
2633 * @vmf: structure describing the fault
2635 * This function handles all that is needed to finish a write page fault in a
2636 * shared mapping due to PTE being read-only once the mapped page is prepared.
2637 * It handles locking of PTE and modifying it. The function returns
2638 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2641 * The function expects the page to be locked or other protection against
2642 * concurrent faults / writeback (such as DAX radix tree locks).
2644 int finish_mkwrite_fault(struct vm_fault *vmf)
2646 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2647 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2650 * We might have raced with another page fault while we released the
2651 * pte_offset_map_lock.
2653 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2654 pte_unmap_unlock(vmf->pte, vmf->ptl);
2655 return VM_FAULT_NOPAGE;
2662 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2665 static int wp_pfn_shared(struct vm_fault *vmf)
2667 struct vm_area_struct *vma = vmf->vma;
2669 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2672 pte_unmap_unlock(vmf->pte, vmf->ptl);
2673 vmf->flags |= FAULT_FLAG_MKWRITE;
2674 ret = vma->vm_ops->pfn_mkwrite(vmf);
2675 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2677 return finish_mkwrite_fault(vmf);
2680 return VM_FAULT_WRITE;
2683 static int wp_page_shared(struct vm_fault *vmf)
2684 __releases(vmf->ptl)
2686 struct vm_area_struct *vma = vmf->vma;
2688 get_page(vmf->page);
2690 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2693 pte_unmap_unlock(vmf->pte, vmf->ptl);
2694 tmp = do_page_mkwrite(vmf);
2695 if (unlikely(!tmp || (tmp &
2696 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2697 put_page(vmf->page);
2700 tmp = finish_mkwrite_fault(vmf);
2701 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2702 unlock_page(vmf->page);
2703 put_page(vmf->page);
2708 lock_page(vmf->page);
2710 fault_dirty_shared_page(vma, vmf->page);
2711 put_page(vmf->page);
2713 return VM_FAULT_WRITE;
2717 * This routine handles present pages, when users try to write
2718 * to a shared page. It is done by copying the page to a new address
2719 * and decrementing the shared-page counter for the old page.
2721 * Note that this routine assumes that the protection checks have been
2722 * done by the caller (the low-level page fault routine in most cases).
2723 * Thus we can safely just mark it writable once we've done any necessary
2726 * We also mark the page dirty at this point even though the page will
2727 * change only once the write actually happens. This avoids a few races,
2728 * and potentially makes it more efficient.
2730 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2731 * but allow concurrent faults), with pte both mapped and locked.
2732 * We return with mmap_sem still held, but pte unmapped and unlocked.
2734 static int do_wp_page(struct vm_fault *vmf)
2735 __releases(vmf->ptl)
2737 struct vm_area_struct *vma = vmf->vma;
2739 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2742 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2745 * We should not cow pages in a shared writeable mapping.
2746 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2748 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2749 (VM_WRITE|VM_SHARED))
2750 return wp_pfn_shared(vmf);
2752 pte_unmap_unlock(vmf->pte, vmf->ptl);
2753 return wp_page_copy(vmf);
2757 * Take out anonymous pages first, anonymous shared vmas are
2758 * not dirty accountable.
2760 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2761 int total_map_swapcount;
2762 if (!trylock_page(vmf->page)) {
2763 get_page(vmf->page);
2764 pte_unmap_unlock(vmf->pte, vmf->ptl);
2765 lock_page(vmf->page);
2766 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2767 vmf->address, &vmf->ptl);
2768 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2769 unlock_page(vmf->page);
2770 pte_unmap_unlock(vmf->pte, vmf->ptl);
2771 put_page(vmf->page);
2774 put_page(vmf->page);
2776 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2777 if (total_map_swapcount == 1) {
2779 * The page is all ours. Move it to
2780 * our anon_vma so the rmap code will
2781 * not search our parent or siblings.
2782 * Protected against the rmap code by
2785 page_move_anon_rmap(vmf->page, vma);
2787 unlock_page(vmf->page);
2789 return VM_FAULT_WRITE;
2791 unlock_page(vmf->page);
2792 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2793 (VM_WRITE|VM_SHARED))) {
2794 return wp_page_shared(vmf);
2798 * Ok, we need to copy. Oh, well..
2800 get_page(vmf->page);
2802 pte_unmap_unlock(vmf->pte, vmf->ptl);
2803 return wp_page_copy(vmf);
2806 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2807 unsigned long start_addr, unsigned long end_addr,
2808 struct zap_details *details)
2810 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2813 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2814 struct zap_details *details)
2816 struct vm_area_struct *vma;
2817 pgoff_t vba, vea, zba, zea;
2819 vma_interval_tree_foreach(vma, root,
2820 details->first_index, details->last_index) {
2822 vba = vma->vm_pgoff;
2823 vea = vba + vma_pages(vma) - 1;
2824 zba = details->first_index;
2827 zea = details->last_index;
2831 unmap_mapping_range_vma(vma,
2832 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2833 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2839 * unmap_mapping_pages() - Unmap pages from processes.
2840 * @mapping: The address space containing pages to be unmapped.
2841 * @start: Index of first page to be unmapped.
2842 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2843 * @even_cows: Whether to unmap even private COWed pages.
2845 * Unmap the pages in this address space from any userspace process which
2846 * has them mmaped. Generally, you want to remove COWed pages as well when
2847 * a file is being truncated, but not when invalidating pages from the page
2850 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2851 pgoff_t nr, bool even_cows)
2853 struct zap_details details = { };
2855 details.check_mapping = even_cows ? NULL : mapping;
2856 details.first_index = start;
2857 details.last_index = start + nr - 1;
2858 if (details.last_index < details.first_index)
2859 details.last_index = ULONG_MAX;
2861 i_mmap_lock_write(mapping);
2862 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2863 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2864 i_mmap_unlock_write(mapping);
2868 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2869 * address_space corresponding to the specified byte range in the underlying
2872 * @mapping: the address space containing mmaps to be unmapped.
2873 * @holebegin: byte in first page to unmap, relative to the start of
2874 * the underlying file. This will be rounded down to a PAGE_SIZE
2875 * boundary. Note that this is different from truncate_pagecache(), which
2876 * must keep the partial page. In contrast, we must get rid of
2878 * @holelen: size of prospective hole in bytes. This will be rounded
2879 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2881 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2882 * but 0 when invalidating pagecache, don't throw away private data.
2884 void unmap_mapping_range(struct address_space *mapping,
2885 loff_t const holebegin, loff_t const holelen, int even_cows)
2887 pgoff_t hba = holebegin >> PAGE_SHIFT;
2888 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2890 /* Check for overflow. */
2891 if (sizeof(holelen) > sizeof(hlen)) {
2893 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2894 if (holeend & ~(long long)ULONG_MAX)
2895 hlen = ULONG_MAX - hba + 1;
2898 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2900 EXPORT_SYMBOL(unmap_mapping_range);
2903 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2904 * but allow concurrent faults), and pte mapped but not yet locked.
2905 * We return with pte unmapped and unlocked.
2907 * We return with the mmap_sem locked or unlocked in the same cases
2908 * as does filemap_fault().
2910 int do_swap_page(struct vm_fault *vmf)
2912 struct vm_area_struct *vma = vmf->vma;
2913 struct page *page = NULL, *swapcache;
2914 struct mem_cgroup *memcg;
2921 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2924 entry = pte_to_swp_entry(vmf->orig_pte);
2925 if (unlikely(non_swap_entry(entry))) {
2926 if (is_migration_entry(entry)) {
2927 migration_entry_wait(vma->vm_mm, vmf->pmd,
2929 } else if (is_device_private_entry(entry)) {
2931 * For un-addressable device memory we call the pgmap
2932 * fault handler callback. The callback must migrate
2933 * the page back to some CPU accessible page.
2935 ret = device_private_entry_fault(vma, vmf->address, entry,
2936 vmf->flags, vmf->pmd);
2937 } else if (is_hwpoison_entry(entry)) {
2938 ret = VM_FAULT_HWPOISON;
2940 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2941 ret = VM_FAULT_SIGBUS;
2947 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2948 page = lookup_swap_cache(entry, vma, vmf->address);
2952 struct swap_info_struct *si = swp_swap_info(entry);
2954 if (si->flags & SWP_SYNCHRONOUS_IO &&
2955 __swap_count(si, entry) == 1) {
2956 /* skip swapcache */
2957 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2960 __SetPageLocked(page);
2961 __SetPageSwapBacked(page);
2962 set_page_private(page, entry.val);
2963 lru_cache_add_anon(page);
2964 swap_readpage(page, true);
2967 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2974 * Back out if somebody else faulted in this pte
2975 * while we released the pte lock.
2977 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2978 vmf->address, &vmf->ptl);
2979 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2981 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2985 /* Had to read the page from swap area: Major fault */
2986 ret = VM_FAULT_MAJOR;
2987 count_vm_event(PGMAJFAULT);
2988 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2989 } else if (PageHWPoison(page)) {
2991 * hwpoisoned dirty swapcache pages are kept for killing
2992 * owner processes (which may be unknown at hwpoison time)
2994 ret = VM_FAULT_HWPOISON;
2995 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2999 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3001 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3003 ret |= VM_FAULT_RETRY;
3008 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3009 * release the swapcache from under us. The page pin, and pte_same
3010 * test below, are not enough to exclude that. Even if it is still
3011 * swapcache, we need to check that the page's swap has not changed.
3013 if (unlikely((!PageSwapCache(page) ||
3014 page_private(page) != entry.val)) && swapcache)
3017 page = ksm_might_need_to_copy(page, vma, vmf->address);
3018 if (unlikely(!page)) {
3024 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3031 * Back out if somebody else already faulted in this pte.
3033 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3035 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3038 if (unlikely(!PageUptodate(page))) {
3039 ret = VM_FAULT_SIGBUS;
3044 * The page isn't present yet, go ahead with the fault.
3046 * Be careful about the sequence of operations here.
3047 * To get its accounting right, reuse_swap_page() must be called
3048 * while the page is counted on swap but not yet in mapcount i.e.
3049 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3050 * must be called after the swap_free(), or it will never succeed.
3053 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3054 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3055 pte = mk_pte(page, vma->vm_page_prot);
3056 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3057 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3058 vmf->flags &= ~FAULT_FLAG_WRITE;
3059 ret |= VM_FAULT_WRITE;
3060 exclusive = RMAP_EXCLUSIVE;
3062 flush_icache_page(vma, page);
3063 if (pte_swp_soft_dirty(vmf->orig_pte))
3064 pte = pte_mksoft_dirty(pte);
3065 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3066 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3067 vmf->orig_pte = pte;
3069 /* ksm created a completely new copy */
3070 if (unlikely(page != swapcache && swapcache)) {
3071 page_add_new_anon_rmap(page, vma, vmf->address, false);
3072 mem_cgroup_commit_charge(page, memcg, false, false);
3073 lru_cache_add_active_or_unevictable(page, vma);
3075 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3076 mem_cgroup_commit_charge(page, memcg, true, false);
3077 activate_page(page);
3081 if (mem_cgroup_swap_full(page) ||
3082 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3083 try_to_free_swap(page);
3085 if (page != swapcache && swapcache) {
3087 * Hold the lock to avoid the swap entry to be reused
3088 * until we take the PT lock for the pte_same() check
3089 * (to avoid false positives from pte_same). For
3090 * further safety release the lock after the swap_free
3091 * so that the swap count won't change under a
3092 * parallel locked swapcache.
3094 unlock_page(swapcache);
3095 put_page(swapcache);
3098 if (vmf->flags & FAULT_FLAG_WRITE) {
3099 ret |= do_wp_page(vmf);
3100 if (ret & VM_FAULT_ERROR)
3101 ret &= VM_FAULT_ERROR;
3105 /* No need to invalidate - it was non-present before */
3106 update_mmu_cache(vma, vmf->address, vmf->pte);
3108 pte_unmap_unlock(vmf->pte, vmf->ptl);
3112 mem_cgroup_cancel_charge(page, memcg, false);
3113 pte_unmap_unlock(vmf->pte, vmf->ptl);
3118 if (page != swapcache && swapcache) {
3119 unlock_page(swapcache);
3120 put_page(swapcache);
3126 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3127 * but allow concurrent faults), and pte mapped but not yet locked.
3128 * We return with mmap_sem still held, but pte unmapped and unlocked.
3130 static int do_anonymous_page(struct vm_fault *vmf)
3132 struct vm_area_struct *vma = vmf->vma;
3133 struct mem_cgroup *memcg;
3138 /* File mapping without ->vm_ops ? */
3139 if (vma->vm_flags & VM_SHARED)
3140 return VM_FAULT_SIGBUS;
3143 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3144 * pte_offset_map() on pmds where a huge pmd might be created
3145 * from a different thread.
3147 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3148 * parallel threads are excluded by other means.
3150 * Here we only have down_read(mmap_sem).
3152 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3153 return VM_FAULT_OOM;
3155 /* See the comment in pte_alloc_one_map() */
3156 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3159 /* Use the zero-page for reads */
3160 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3161 !mm_forbids_zeropage(vma->vm_mm)) {
3162 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3163 vma->vm_page_prot));
3164 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3165 vmf->address, &vmf->ptl);
3166 if (!pte_none(*vmf->pte))
3168 ret = check_stable_address_space(vma->vm_mm);
3171 /* Deliver the page fault to userland, check inside PT lock */
3172 if (userfaultfd_missing(vma)) {
3173 pte_unmap_unlock(vmf->pte, vmf->ptl);
3174 return handle_userfault(vmf, VM_UFFD_MISSING);
3179 /* Allocate our own private page. */
3180 if (unlikely(anon_vma_prepare(vma)))
3182 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3186 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3191 * The memory barrier inside __SetPageUptodate makes sure that
3192 * preceeding stores to the page contents become visible before
3193 * the set_pte_at() write.
3195 __SetPageUptodate(page);
3197 entry = mk_pte(page, vma->vm_page_prot);
3198 if (vma->vm_flags & VM_WRITE)
3199 entry = pte_mkwrite(pte_mkdirty(entry));
3201 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3203 if (!pte_none(*vmf->pte))
3206 ret = check_stable_address_space(vma->vm_mm);
3210 /* Deliver the page fault to userland, check inside PT lock */
3211 if (userfaultfd_missing(vma)) {
3212 pte_unmap_unlock(vmf->pte, vmf->ptl);
3213 mem_cgroup_cancel_charge(page, memcg, false);
3215 return handle_userfault(vmf, VM_UFFD_MISSING);
3218 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3219 page_add_new_anon_rmap(page, vma, vmf->address, false);
3220 mem_cgroup_commit_charge(page, memcg, false, false);
3221 lru_cache_add_active_or_unevictable(page, vma);
3223 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3225 /* No need to invalidate - it was non-present before */
3226 update_mmu_cache(vma, vmf->address, vmf->pte);
3228 pte_unmap_unlock(vmf->pte, vmf->ptl);
3231 mem_cgroup_cancel_charge(page, memcg, false);
3237 return VM_FAULT_OOM;
3241 * The mmap_sem must have been held on entry, and may have been
3242 * released depending on flags and vma->vm_ops->fault() return value.
3243 * See filemap_fault() and __lock_page_retry().
3245 static int __do_fault(struct vm_fault *vmf)
3247 struct vm_area_struct *vma = vmf->vma;
3250 ret = vma->vm_ops->fault(vmf);
3251 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3252 VM_FAULT_DONE_COW)))
3255 if (unlikely(PageHWPoison(vmf->page))) {
3256 if (ret & VM_FAULT_LOCKED)
3257 unlock_page(vmf->page);
3258 put_page(vmf->page);
3260 return VM_FAULT_HWPOISON;
3263 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3264 lock_page(vmf->page);
3266 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3272 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3273 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3274 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3275 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3277 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3279 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3282 static int pte_alloc_one_map(struct vm_fault *vmf)
3284 struct vm_area_struct *vma = vmf->vma;
3286 if (!pmd_none(*vmf->pmd))
3288 if (vmf->prealloc_pte) {
3289 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3290 if (unlikely(!pmd_none(*vmf->pmd))) {
3291 spin_unlock(vmf->ptl);
3295 mm_inc_nr_ptes(vma->vm_mm);
3296 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3297 spin_unlock(vmf->ptl);
3298 vmf->prealloc_pte = NULL;
3299 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3300 return VM_FAULT_OOM;
3304 * If a huge pmd materialized under us just retry later. Use
3305 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3306 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3307 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3308 * running immediately after a huge pmd fault in a different thread of
3309 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3310 * All we have to ensure is that it is a regular pmd that we can walk
3311 * with pte_offset_map() and we can do that through an atomic read in
3312 * C, which is what pmd_trans_unstable() provides.
3314 if (pmd_devmap_trans_unstable(vmf->pmd))
3315 return VM_FAULT_NOPAGE;
3318 * At this point we know that our vmf->pmd points to a page of ptes
3319 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3320 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3321 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3322 * be valid and we will re-check to make sure the vmf->pte isn't
3323 * pte_none() under vmf->ptl protection when we return to
3326 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3331 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3333 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3334 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3335 unsigned long haddr)
3337 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3338 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3340 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3345 static void deposit_prealloc_pte(struct vm_fault *vmf)
3347 struct vm_area_struct *vma = vmf->vma;
3349 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3351 * We are going to consume the prealloc table,
3352 * count that as nr_ptes.
3354 mm_inc_nr_ptes(vma->vm_mm);
3355 vmf->prealloc_pte = NULL;
3358 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3360 struct vm_area_struct *vma = vmf->vma;
3361 bool write = vmf->flags & FAULT_FLAG_WRITE;
3362 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3366 if (!transhuge_vma_suitable(vma, haddr))
3367 return VM_FAULT_FALLBACK;
3369 ret = VM_FAULT_FALLBACK;
3370 page = compound_head(page);
3373 * Archs like ppc64 need additonal space to store information
3374 * related to pte entry. Use the preallocated table for that.
3376 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3377 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3378 if (!vmf->prealloc_pte)
3379 return VM_FAULT_OOM;
3380 smp_wmb(); /* See comment in __pte_alloc() */
3383 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3384 if (unlikely(!pmd_none(*vmf->pmd)))
3387 for (i = 0; i < HPAGE_PMD_NR; i++)
3388 flush_icache_page(vma, page + i);
3390 entry = mk_huge_pmd(page, vma->vm_page_prot);
3392 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3394 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3395 page_add_file_rmap(page, true);
3397 * deposit and withdraw with pmd lock held
3399 if (arch_needs_pgtable_deposit())
3400 deposit_prealloc_pte(vmf);
3402 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3404 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3406 /* fault is handled */
3408 count_vm_event(THP_FILE_MAPPED);
3410 spin_unlock(vmf->ptl);
3414 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3422 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3423 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3425 * @vmf: fault environment
3426 * @memcg: memcg to charge page (only for private mappings)
3427 * @page: page to map
3429 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3432 * Target users are page handler itself and implementations of
3433 * vm_ops->map_pages.
3435 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3438 struct vm_area_struct *vma = vmf->vma;
3439 bool write = vmf->flags & FAULT_FLAG_WRITE;
3443 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3444 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3446 VM_BUG_ON_PAGE(memcg, page);
3448 ret = do_set_pmd(vmf, page);
3449 if (ret != VM_FAULT_FALLBACK)
3454 ret = pte_alloc_one_map(vmf);
3459 /* Re-check under ptl */
3460 if (unlikely(!pte_none(*vmf->pte)))
3461 return VM_FAULT_NOPAGE;
3463 flush_icache_page(vma, page);
3464 entry = mk_pte(page, vma->vm_page_prot);
3466 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3467 /* copy-on-write page */
3468 if (write && !(vma->vm_flags & VM_SHARED)) {
3469 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3470 page_add_new_anon_rmap(page, vma, vmf->address, false);
3471 mem_cgroup_commit_charge(page, memcg, false, false);
3472 lru_cache_add_active_or_unevictable(page, vma);
3474 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3475 page_add_file_rmap(page, false);
3477 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3479 /* no need to invalidate: a not-present page won't be cached */
3480 update_mmu_cache(vma, vmf->address, vmf->pte);
3487 * finish_fault - finish page fault once we have prepared the page to fault
3489 * @vmf: structure describing the fault
3491 * This function handles all that is needed to finish a page fault once the
3492 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3493 * given page, adds reverse page mapping, handles memcg charges and LRU
3494 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3497 * The function expects the page to be locked and on success it consumes a
3498 * reference of a page being mapped (for the PTE which maps it).
3500 int finish_fault(struct vm_fault *vmf)
3505 /* Did we COW the page? */
3506 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3507 !(vmf->vma->vm_flags & VM_SHARED))
3508 page = vmf->cow_page;
3513 * check even for read faults because we might have lost our CoWed
3516 if (!(vmf->vma->vm_flags & VM_SHARED))
3517 ret = check_stable_address_space(vmf->vma->vm_mm);
3519 ret = alloc_set_pte(vmf, vmf->memcg, page);
3521 pte_unmap_unlock(vmf->pte, vmf->ptl);
3525 static unsigned long fault_around_bytes __read_mostly =
3526 rounddown_pow_of_two(65536);
3528 #ifdef CONFIG_DEBUG_FS
3529 static int fault_around_bytes_get(void *data, u64 *val)
3531 *val = fault_around_bytes;
3536 * fault_around_bytes must be rounded down to the nearest page order as it's
3537 * what do_fault_around() expects to see.
3539 static int fault_around_bytes_set(void *data, u64 val)
3541 if (val / PAGE_SIZE > PTRS_PER_PTE)
3543 if (val > PAGE_SIZE)
3544 fault_around_bytes = rounddown_pow_of_two(val);
3546 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3549 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3550 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3552 static int __init fault_around_debugfs(void)
3556 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3557 &fault_around_bytes_fops);
3559 pr_warn("Failed to create fault_around_bytes in debugfs");
3562 late_initcall(fault_around_debugfs);
3566 * do_fault_around() tries to map few pages around the fault address. The hope
3567 * is that the pages will be needed soon and this will lower the number of
3570 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3571 * not ready to be mapped: not up-to-date, locked, etc.
3573 * This function is called with the page table lock taken. In the split ptlock
3574 * case the page table lock only protects only those entries which belong to
3575 * the page table corresponding to the fault address.
3577 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3580 * fault_around_bytes defines how many bytes we'll try to map.
3581 * do_fault_around() expects it to be set to a power of two less than or equal
3584 * The virtual address of the area that we map is naturally aligned to
3585 * fault_around_bytes rounded down to the machine page size
3586 * (and therefore to page order). This way it's easier to guarantee
3587 * that we don't cross page table boundaries.
3589 static int do_fault_around(struct vm_fault *vmf)
3591 unsigned long address = vmf->address, nr_pages, mask;
3592 pgoff_t start_pgoff = vmf->pgoff;
3596 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3597 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3599 vmf->address = max(address & mask, vmf->vma->vm_start);
3600 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3604 * end_pgoff is either the end of the page table, the end of
3605 * the vma or nr_pages from start_pgoff, depending what is nearest.
3607 end_pgoff = start_pgoff -
3608 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3610 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3611 start_pgoff + nr_pages - 1);
3613 if (pmd_none(*vmf->pmd)) {
3614 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3616 if (!vmf->prealloc_pte)
3618 smp_wmb(); /* See comment in __pte_alloc() */
3621 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3623 /* Huge page is mapped? Page fault is solved */
3624 if (pmd_trans_huge(*vmf->pmd)) {
3625 ret = VM_FAULT_NOPAGE;
3629 /* ->map_pages() haven't done anything useful. Cold page cache? */
3633 /* check if the page fault is solved */
3634 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3635 if (!pte_none(*vmf->pte))
3636 ret = VM_FAULT_NOPAGE;
3637 pte_unmap_unlock(vmf->pte, vmf->ptl);
3639 vmf->address = address;
3644 static int do_read_fault(struct vm_fault *vmf)
3646 struct vm_area_struct *vma = vmf->vma;
3650 * Let's call ->map_pages() first and use ->fault() as fallback
3651 * if page by the offset is not ready to be mapped (cold cache or
3654 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3655 ret = do_fault_around(vmf);
3660 ret = __do_fault(vmf);
3661 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3664 ret |= finish_fault(vmf);
3665 unlock_page(vmf->page);
3666 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3667 put_page(vmf->page);
3671 static int do_cow_fault(struct vm_fault *vmf)
3673 struct vm_area_struct *vma = vmf->vma;
3676 if (unlikely(anon_vma_prepare(vma)))
3677 return VM_FAULT_OOM;
3679 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3681 return VM_FAULT_OOM;
3683 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3684 &vmf->memcg, false)) {
3685 put_page(vmf->cow_page);
3686 return VM_FAULT_OOM;
3689 ret = __do_fault(vmf);
3690 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3692 if (ret & VM_FAULT_DONE_COW)
3695 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3696 __SetPageUptodate(vmf->cow_page);
3698 ret |= finish_fault(vmf);
3699 unlock_page(vmf->page);
3700 put_page(vmf->page);
3701 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3705 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3706 put_page(vmf->cow_page);
3710 static int do_shared_fault(struct vm_fault *vmf)
3712 struct vm_area_struct *vma = vmf->vma;
3715 ret = __do_fault(vmf);
3716 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3720 * Check if the backing address space wants to know that the page is
3721 * about to become writable
3723 if (vma->vm_ops->page_mkwrite) {
3724 unlock_page(vmf->page);
3725 tmp = do_page_mkwrite(vmf);
3726 if (unlikely(!tmp ||
3727 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3728 put_page(vmf->page);
3733 ret |= finish_fault(vmf);
3734 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3736 unlock_page(vmf->page);
3737 put_page(vmf->page);
3741 fault_dirty_shared_page(vma, vmf->page);
3746 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3747 * but allow concurrent faults).
3748 * The mmap_sem may have been released depending on flags and our
3749 * return value. See filemap_fault() and __lock_page_or_retry().
3751 static int do_fault(struct vm_fault *vmf)
3753 struct vm_area_struct *vma = vmf->vma;
3756 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3757 if (!vma->vm_ops->fault)
3758 ret = VM_FAULT_SIGBUS;
3759 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3760 ret = do_read_fault(vmf);
3761 else if (!(vma->vm_flags & VM_SHARED))
3762 ret = do_cow_fault(vmf);
3764 ret = do_shared_fault(vmf);
3766 /* preallocated pagetable is unused: free it */
3767 if (vmf->prealloc_pte) {
3768 pte_free(vma->vm_mm, vmf->prealloc_pte);
3769 vmf->prealloc_pte = NULL;
3774 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3775 unsigned long addr, int page_nid,
3780 count_vm_numa_event(NUMA_HINT_FAULTS);
3781 if (page_nid == numa_node_id()) {
3782 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3783 *flags |= TNF_FAULT_LOCAL;
3786 return mpol_misplaced(page, vma, addr);
3789 static int do_numa_page(struct vm_fault *vmf)
3791 struct vm_area_struct *vma = vmf->vma;
3792 struct page *page = NULL;
3796 bool migrated = false;
3798 bool was_writable = pte_savedwrite(vmf->orig_pte);
3802 * The "pte" at this point cannot be used safely without
3803 * validation through pte_unmap_same(). It's of NUMA type but
3804 * the pfn may be screwed if the read is non atomic.
3806 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3807 spin_lock(vmf->ptl);
3808 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3809 pte_unmap_unlock(vmf->pte, vmf->ptl);
3814 * Make it present again, Depending on how arch implementes non
3815 * accessible ptes, some can allow access by kernel mode.
3817 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3818 pte = pte_modify(pte, vma->vm_page_prot);
3819 pte = pte_mkyoung(pte);
3821 pte = pte_mkwrite(pte);
3822 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3823 update_mmu_cache(vma, vmf->address, vmf->pte);
3825 page = vm_normal_page(vma, vmf->address, pte);
3827 pte_unmap_unlock(vmf->pte, vmf->ptl);
3831 /* TODO: handle PTE-mapped THP */
3832 if (PageCompound(page)) {
3833 pte_unmap_unlock(vmf->pte, vmf->ptl);
3838 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3839 * much anyway since they can be in shared cache state. This misses
3840 * the case where a mapping is writable but the process never writes
3841 * to it but pte_write gets cleared during protection updates and
3842 * pte_dirty has unpredictable behaviour between PTE scan updates,
3843 * background writeback, dirty balancing and application behaviour.
3845 if (!pte_write(pte))
3846 flags |= TNF_NO_GROUP;
3849 * Flag if the page is shared between multiple address spaces. This
3850 * is later used when determining whether to group tasks together
3852 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3853 flags |= TNF_SHARED;
3855 last_cpupid = page_cpupid_last(page);
3856 page_nid = page_to_nid(page);
3857 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3859 pte_unmap_unlock(vmf->pte, vmf->ptl);
3860 if (target_nid == -1) {
3865 /* Migrate to the requested node */
3866 migrated = migrate_misplaced_page(page, vma, target_nid);
3868 page_nid = target_nid;
3869 flags |= TNF_MIGRATED;
3871 flags |= TNF_MIGRATE_FAIL;
3875 task_numa_fault(last_cpupid, page_nid, 1, flags);
3879 static inline int create_huge_pmd(struct vm_fault *vmf)
3881 if (vma_is_anonymous(vmf->vma))
3882 return do_huge_pmd_anonymous_page(vmf);
3883 if (vmf->vma->vm_ops->huge_fault)
3884 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3885 return VM_FAULT_FALLBACK;
3888 /* `inline' is required to avoid gcc 4.1.2 build error */
3889 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3891 if (vma_is_anonymous(vmf->vma))
3892 return do_huge_pmd_wp_page(vmf, orig_pmd);
3893 if (vmf->vma->vm_ops->huge_fault)
3894 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3896 /* COW handled on pte level: split pmd */
3897 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3898 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3900 return VM_FAULT_FALLBACK;
3903 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3905 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3908 static int create_huge_pud(struct vm_fault *vmf)
3910 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3911 /* No support for anonymous transparent PUD pages yet */
3912 if (vma_is_anonymous(vmf->vma))
3913 return VM_FAULT_FALLBACK;
3914 if (vmf->vma->vm_ops->huge_fault)
3915 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3916 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3917 return VM_FAULT_FALLBACK;
3920 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3922 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3923 /* No support for anonymous transparent PUD pages yet */
3924 if (vma_is_anonymous(vmf->vma))
3925 return VM_FAULT_FALLBACK;
3926 if (vmf->vma->vm_ops->huge_fault)
3927 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3928 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3929 return VM_FAULT_FALLBACK;
3933 * These routines also need to handle stuff like marking pages dirty
3934 * and/or accessed for architectures that don't do it in hardware (most
3935 * RISC architectures). The early dirtying is also good on the i386.
3937 * There is also a hook called "update_mmu_cache()" that architectures
3938 * with external mmu caches can use to update those (ie the Sparc or
3939 * PowerPC hashed page tables that act as extended TLBs).
3941 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3942 * concurrent faults).
3944 * The mmap_sem may have been released depending on flags and our return value.
3945 * See filemap_fault() and __lock_page_or_retry().
3947 static int handle_pte_fault(struct vm_fault *vmf)
3951 if (unlikely(pmd_none(*vmf->pmd))) {
3953 * Leave __pte_alloc() until later: because vm_ops->fault may
3954 * want to allocate huge page, and if we expose page table
3955 * for an instant, it will be difficult to retract from
3956 * concurrent faults and from rmap lookups.
3960 /* See comment in pte_alloc_one_map() */
3961 if (pmd_devmap_trans_unstable(vmf->pmd))
3964 * A regular pmd is established and it can't morph into a huge
3965 * pmd from under us anymore at this point because we hold the
3966 * mmap_sem read mode and khugepaged takes it in write mode.
3967 * So now it's safe to run pte_offset_map().
3969 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3970 vmf->orig_pte = *vmf->pte;
3973 * some architectures can have larger ptes than wordsize,
3974 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3975 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3976 * accesses. The code below just needs a consistent view
3977 * for the ifs and we later double check anyway with the
3978 * ptl lock held. So here a barrier will do.
3981 if (pte_none(vmf->orig_pte)) {
3982 pte_unmap(vmf->pte);
3988 if (vma_is_anonymous(vmf->vma))
3989 return do_anonymous_page(vmf);
3991 return do_fault(vmf);
3994 if (!pte_present(vmf->orig_pte))
3995 return do_swap_page(vmf);
3997 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3998 return do_numa_page(vmf);
4000 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4001 spin_lock(vmf->ptl);
4002 entry = vmf->orig_pte;
4003 if (unlikely(!pte_same(*vmf->pte, entry)))
4005 if (vmf->flags & FAULT_FLAG_WRITE) {
4006 if (!pte_write(entry))
4007 return do_wp_page(vmf);
4008 entry = pte_mkdirty(entry);
4010 entry = pte_mkyoung(entry);
4011 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4012 vmf->flags & FAULT_FLAG_WRITE)) {
4013 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4016 * This is needed only for protection faults but the arch code
4017 * is not yet telling us if this is a protection fault or not.
4018 * This still avoids useless tlb flushes for .text page faults
4021 if (vmf->flags & FAULT_FLAG_WRITE)
4022 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4025 pte_unmap_unlock(vmf->pte, vmf->ptl);
4030 * By the time we get here, we already hold the mm semaphore
4032 * The mmap_sem may have been released depending on flags and our
4033 * return value. See filemap_fault() and __lock_page_or_retry().
4035 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4038 struct vm_fault vmf = {
4040 .address = address & PAGE_MASK,
4042 .pgoff = linear_page_index(vma, address),
4043 .gfp_mask = __get_fault_gfp_mask(vma),
4045 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4046 struct mm_struct *mm = vma->vm_mm;
4051 pgd = pgd_offset(mm, address);
4052 p4d = p4d_alloc(mm, pgd, address);
4054 return VM_FAULT_OOM;
4056 vmf.pud = pud_alloc(mm, p4d, address);
4058 return VM_FAULT_OOM;
4059 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4060 ret = create_huge_pud(&vmf);
4061 if (!(ret & VM_FAULT_FALLBACK))
4064 pud_t orig_pud = *vmf.pud;
4067 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4069 /* NUMA case for anonymous PUDs would go here */
4071 if (dirty && !pud_write(orig_pud)) {
4072 ret = wp_huge_pud(&vmf, orig_pud);
4073 if (!(ret & VM_FAULT_FALLBACK))
4076 huge_pud_set_accessed(&vmf, orig_pud);
4082 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4084 return VM_FAULT_OOM;
4085 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4086 ret = create_huge_pmd(&vmf);
4087 if (!(ret & VM_FAULT_FALLBACK))
4090 pmd_t orig_pmd = *vmf.pmd;
4093 if (unlikely(is_swap_pmd(orig_pmd))) {
4094 VM_BUG_ON(thp_migration_supported() &&
4095 !is_pmd_migration_entry(orig_pmd));
4096 if (is_pmd_migration_entry(orig_pmd))
4097 pmd_migration_entry_wait(mm, vmf.pmd);
4100 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4101 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4102 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4104 if (dirty && !pmd_write(orig_pmd)) {
4105 ret = wp_huge_pmd(&vmf, orig_pmd);
4106 if (!(ret & VM_FAULT_FALLBACK))
4109 huge_pmd_set_accessed(&vmf, orig_pmd);
4115 return handle_pte_fault(&vmf);
4119 * By the time we get here, we already hold the mm semaphore
4121 * The mmap_sem may have been released depending on flags and our
4122 * return value. See filemap_fault() and __lock_page_or_retry().
4124 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4129 __set_current_state(TASK_RUNNING);
4131 count_vm_event(PGFAULT);
4132 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4134 /* do counter updates before entering really critical section. */
4135 check_sync_rss_stat(current);
4137 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4138 flags & FAULT_FLAG_INSTRUCTION,
4139 flags & FAULT_FLAG_REMOTE))
4140 return VM_FAULT_SIGSEGV;
4143 * Enable the memcg OOM handling for faults triggered in user
4144 * space. Kernel faults are handled more gracefully.
4146 if (flags & FAULT_FLAG_USER)
4147 mem_cgroup_enter_user_fault();
4149 if (unlikely(is_vm_hugetlb_page(vma)))
4150 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4152 ret = __handle_mm_fault(vma, address, flags);
4154 if (flags & FAULT_FLAG_USER) {
4155 mem_cgroup_exit_user_fault();
4157 * The task may have entered a memcg OOM situation but
4158 * if the allocation error was handled gracefully (no
4159 * VM_FAULT_OOM), there is no need to kill anything.
4160 * Just clean up the OOM state peacefully.
4162 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4163 mem_cgroup_oom_synchronize(false);
4168 EXPORT_SYMBOL_GPL(handle_mm_fault);
4170 #ifndef __PAGETABLE_P4D_FOLDED
4172 * Allocate p4d page table.
4173 * We've already handled the fast-path in-line.
4175 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4177 p4d_t *new = p4d_alloc_one(mm, address);
4181 smp_wmb(); /* See comment in __pte_alloc */
4183 spin_lock(&mm->page_table_lock);
4184 if (pgd_present(*pgd)) /* Another has populated it */
4187 pgd_populate(mm, pgd, new);
4188 spin_unlock(&mm->page_table_lock);
4191 #endif /* __PAGETABLE_P4D_FOLDED */
4193 #ifndef __PAGETABLE_PUD_FOLDED
4195 * Allocate page upper directory.
4196 * We've already handled the fast-path in-line.
4198 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4200 pud_t *new = pud_alloc_one(mm, address);
4204 smp_wmb(); /* See comment in __pte_alloc */
4206 spin_lock(&mm->page_table_lock);
4207 #ifndef __ARCH_HAS_5LEVEL_HACK
4208 if (!p4d_present(*p4d)) {
4210 p4d_populate(mm, p4d, new);
4211 } else /* Another has populated it */
4214 if (!pgd_present(*p4d)) {
4216 pgd_populate(mm, p4d, new);
4217 } else /* Another has populated it */
4219 #endif /* __ARCH_HAS_5LEVEL_HACK */
4220 spin_unlock(&mm->page_table_lock);
4223 #endif /* __PAGETABLE_PUD_FOLDED */
4225 #ifndef __PAGETABLE_PMD_FOLDED
4227 * Allocate page middle directory.
4228 * We've already handled the fast-path in-line.
4230 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4233 pmd_t *new = pmd_alloc_one(mm, address);
4237 smp_wmb(); /* See comment in __pte_alloc */
4239 ptl = pud_lock(mm, pud);
4240 #ifndef __ARCH_HAS_4LEVEL_HACK
4241 if (!pud_present(*pud)) {
4243 pud_populate(mm, pud, new);
4244 } else /* Another has populated it */
4247 if (!pgd_present(*pud)) {
4249 pgd_populate(mm, pud, new);
4250 } else /* Another has populated it */
4252 #endif /* __ARCH_HAS_4LEVEL_HACK */
4256 #endif /* __PAGETABLE_PMD_FOLDED */
4258 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4259 unsigned long *start, unsigned long *end,
4260 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4268 pgd = pgd_offset(mm, address);
4269 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4272 p4d = p4d_offset(pgd, address);
4273 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4276 pud = pud_offset(p4d, address);
4277 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4280 pmd = pmd_offset(pud, address);
4281 VM_BUG_ON(pmd_trans_huge(*pmd));
4283 if (pmd_huge(*pmd)) {
4288 *start = address & PMD_MASK;
4289 *end = *start + PMD_SIZE;
4290 mmu_notifier_invalidate_range_start(mm, *start, *end);
4292 *ptlp = pmd_lock(mm, pmd);
4293 if (pmd_huge(*pmd)) {
4299 mmu_notifier_invalidate_range_end(mm, *start, *end);
4302 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4306 *start = address & PAGE_MASK;
4307 *end = *start + PAGE_SIZE;
4308 mmu_notifier_invalidate_range_start(mm, *start, *end);
4310 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4311 if (!pte_present(*ptep))
4316 pte_unmap_unlock(ptep, *ptlp);
4318 mmu_notifier_invalidate_range_end(mm, *start, *end);
4323 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4324 pte_t **ptepp, spinlock_t **ptlp)
4328 /* (void) is needed to make gcc happy */
4329 (void) __cond_lock(*ptlp,
4330 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4331 ptepp, NULL, ptlp)));
4335 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4336 unsigned long *start, unsigned long *end,
4337 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4341 /* (void) is needed to make gcc happy */
4342 (void) __cond_lock(*ptlp,
4343 !(res = __follow_pte_pmd(mm, address, start, end,
4344 ptepp, pmdpp, ptlp)));
4347 EXPORT_SYMBOL(follow_pte_pmd);
4350 * follow_pfn - look up PFN at a user virtual address
4351 * @vma: memory mapping
4352 * @address: user virtual address
4353 * @pfn: location to store found PFN
4355 * Only IO mappings and raw PFN mappings are allowed.
4357 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4359 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4366 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4369 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4372 *pfn = pte_pfn(*ptep);
4373 pte_unmap_unlock(ptep, ptl);
4376 EXPORT_SYMBOL(follow_pfn);
4378 #ifdef CONFIG_HAVE_IOREMAP_PROT
4379 int follow_phys(struct vm_area_struct *vma,
4380 unsigned long address, unsigned int flags,
4381 unsigned long *prot, resource_size_t *phys)
4387 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4390 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4394 if ((flags & FOLL_WRITE) && !pte_write(pte))
4397 *prot = pgprot_val(pte_pgprot(pte));
4398 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4402 pte_unmap_unlock(ptep, ptl);
4407 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4408 void *buf, int len, int write)
4410 resource_size_t phys_addr;
4411 unsigned long prot = 0;
4412 void __iomem *maddr;
4413 int offset = addr & (PAGE_SIZE-1);
4415 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4418 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4423 memcpy_toio(maddr + offset, buf, len);
4425 memcpy_fromio(buf, maddr + offset, len);
4430 EXPORT_SYMBOL_GPL(generic_access_phys);
4434 * Access another process' address space as given in mm. If non-NULL, use the
4435 * given task for page fault accounting.
4437 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4438 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4440 struct vm_area_struct *vma;
4441 void *old_buf = buf;
4442 int write = gup_flags & FOLL_WRITE;
4444 down_read(&mm->mmap_sem);
4445 /* ignore errors, just check how much was successfully transferred */
4447 int bytes, ret, offset;
4449 struct page *page = NULL;
4451 ret = get_user_pages_remote(tsk, mm, addr, 1,
4452 gup_flags, &page, &vma, NULL);
4454 #ifndef CONFIG_HAVE_IOREMAP_PROT
4458 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4459 * we can access using slightly different code.
4461 vma = find_vma(mm, addr);
4462 if (!vma || vma->vm_start > addr)
4464 if (vma->vm_ops && vma->vm_ops->access)
4465 ret = vma->vm_ops->access(vma, addr, buf,
4473 offset = addr & (PAGE_SIZE-1);
4474 if (bytes > PAGE_SIZE-offset)
4475 bytes = PAGE_SIZE-offset;
4479 copy_to_user_page(vma, page, addr,
4480 maddr + offset, buf, bytes);
4481 set_page_dirty_lock(page);
4483 copy_from_user_page(vma, page, addr,
4484 buf, maddr + offset, bytes);
4493 up_read(&mm->mmap_sem);
4495 return buf - old_buf;
4499 * access_remote_vm - access another process' address space
4500 * @mm: the mm_struct of the target address space
4501 * @addr: start address to access
4502 * @buf: source or destination buffer
4503 * @len: number of bytes to transfer
4504 * @gup_flags: flags modifying lookup behaviour
4506 * The caller must hold a reference on @mm.
4508 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4509 void *buf, int len, unsigned int gup_flags)
4511 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4515 * Access another process' address space.
4516 * Source/target buffer must be kernel space,
4517 * Do not walk the page table directly, use get_user_pages
4519 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4520 void *buf, int len, unsigned int gup_flags)
4522 struct mm_struct *mm;
4525 mm = get_task_mm(tsk);
4529 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4535 EXPORT_SYMBOL_GPL(access_process_vm);
4538 * Print the name of a VMA.
4540 void print_vma_addr(char *prefix, unsigned long ip)
4542 struct mm_struct *mm = current->mm;
4543 struct vm_area_struct *vma;
4546 * we might be running from an atomic context so we cannot sleep
4548 if (!down_read_trylock(&mm->mmap_sem))
4551 vma = find_vma(mm, ip);
4552 if (vma && vma->vm_file) {
4553 struct file *f = vma->vm_file;
4554 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4558 p = file_path(f, buf, PAGE_SIZE);
4561 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4563 vma->vm_end - vma->vm_start);
4564 free_page((unsigned long)buf);
4567 up_read(&mm->mmap_sem);
4570 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4571 void __might_fault(const char *file, int line)
4574 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4575 * holding the mmap_sem, this is safe because kernel memory doesn't
4576 * get paged out, therefore we'll never actually fault, and the
4577 * below annotations will generate false positives.
4579 if (uaccess_kernel())
4581 if (pagefault_disabled())
4583 __might_sleep(file, line, 0);
4584 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4586 might_lock_read(¤t->mm->mmap_sem);
4589 EXPORT_SYMBOL(__might_fault);
4592 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4594 * Process all subpages of the specified huge page with the specified
4595 * operation. The target subpage will be processed last to keep its
4598 static inline void process_huge_page(
4599 unsigned long addr_hint, unsigned int pages_per_huge_page,
4600 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4604 unsigned long addr = addr_hint &
4605 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4607 /* Process target subpage last to keep its cache lines hot */
4609 n = (addr_hint - addr) / PAGE_SIZE;
4610 if (2 * n <= pages_per_huge_page) {
4611 /* If target subpage in first half of huge page */
4614 /* Process subpages at the end of huge page */
4615 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4617 process_subpage(addr + i * PAGE_SIZE, i, arg);
4620 /* If target subpage in second half of huge page */
4621 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4622 l = pages_per_huge_page - n;
4623 /* Process subpages at the begin of huge page */
4624 for (i = 0; i < base; i++) {
4626 process_subpage(addr + i * PAGE_SIZE, i, arg);
4630 * Process remaining subpages in left-right-left-right pattern
4631 * towards the target subpage
4633 for (i = 0; i < l; i++) {
4634 int left_idx = base + i;
4635 int right_idx = base + 2 * l - 1 - i;
4638 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4640 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4644 static void clear_gigantic_page(struct page *page,
4646 unsigned int pages_per_huge_page)
4649 struct page *p = page;
4652 for (i = 0; i < pages_per_huge_page;
4653 i++, p = mem_map_next(p, page, i)) {
4655 clear_user_highpage(p, addr + i * PAGE_SIZE);
4659 static void clear_subpage(unsigned long addr, int idx, void *arg)
4661 struct page *page = arg;
4663 clear_user_highpage(page + idx, addr);
4666 void clear_huge_page(struct page *page,
4667 unsigned long addr_hint, unsigned int pages_per_huge_page)
4669 unsigned long addr = addr_hint &
4670 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4672 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4673 clear_gigantic_page(page, addr, pages_per_huge_page);
4677 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4680 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4682 struct vm_area_struct *vma,
4683 unsigned int pages_per_huge_page)
4686 struct page *dst_base = dst;
4687 struct page *src_base = src;
4689 for (i = 0; i < pages_per_huge_page; ) {
4691 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4694 dst = mem_map_next(dst, dst_base, i);
4695 src = mem_map_next(src, src_base, i);
4699 struct copy_subpage_arg {
4702 struct vm_area_struct *vma;
4705 static void copy_subpage(unsigned long addr, int idx, void *arg)
4707 struct copy_subpage_arg *copy_arg = arg;
4709 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4710 addr, copy_arg->vma);
4713 void copy_user_huge_page(struct page *dst, struct page *src,
4714 unsigned long addr_hint, struct vm_area_struct *vma,
4715 unsigned int pages_per_huge_page)
4717 unsigned long addr = addr_hint &
4718 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4719 struct copy_subpage_arg arg = {
4725 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4726 copy_user_gigantic_page(dst, src, addr, vma,
4727 pages_per_huge_page);
4731 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4734 long copy_huge_page_from_user(struct page *dst_page,
4735 const void __user *usr_src,
4736 unsigned int pages_per_huge_page,
4737 bool allow_pagefault)
4739 void *src = (void *)usr_src;
4741 unsigned long i, rc = 0;
4742 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4744 for (i = 0; i < pages_per_huge_page; i++) {
4745 if (allow_pagefault)
4746 page_kaddr = kmap(dst_page + i);
4748 page_kaddr = kmap_atomic(dst_page + i);
4749 rc = copy_from_user(page_kaddr,
4750 (const void __user *)(src + i * PAGE_SIZE),
4752 if (allow_pagefault)
4753 kunmap(dst_page + i);
4755 kunmap_atomic(page_kaddr);
4757 ret_val -= (PAGE_SIZE - rc);
4765 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4767 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4769 static struct kmem_cache *page_ptl_cachep;
4771 void __init ptlock_cache_init(void)
4773 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4777 bool ptlock_alloc(struct page *page)
4781 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4788 void ptlock_free(struct page *page)
4790 kmem_cache_free(page_ptl_cachep, page->ptl);