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_free(struct mmu_gather *tlb)
243 struct mmu_gather_batch *batch;
245 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
246 tlb_table_flush(tlb);
248 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
252 tlb->active = &tlb->local;
255 void tlb_flush_mmu(struct mmu_gather *tlb)
257 tlb_flush_mmu_tlbonly(tlb);
258 tlb_flush_mmu_free(tlb);
262 * Called at the end of the shootdown operation to free up any resources
263 * that were required.
265 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
266 unsigned long start, unsigned long end, bool force)
268 struct mmu_gather_batch *batch, *next;
271 __tlb_adjust_range(tlb, start, end - start);
275 /* keep the page table cache within bounds */
278 for (batch = tlb->local.next; batch; batch = next) {
280 free_pages((unsigned long)batch, 0);
282 tlb->local.next = NULL;
286 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
287 * handling the additional races in SMP caused by other CPUs caching valid
288 * mappings in their TLBs. Returns the number of free page slots left.
289 * When out of page slots we must call tlb_flush_mmu().
290 *returns true if the caller should flush.
292 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
294 struct mmu_gather_batch *batch;
296 VM_BUG_ON(!tlb->end);
297 VM_WARN_ON(tlb->page_size != page_size);
301 * Add the page and check if we are full. If so
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
310 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
315 #endif /* HAVE_GENERIC_MMU_GATHER */
317 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
320 * See the comment near struct mmu_table_batch.
324 * If we want tlb_remove_table() to imply TLB invalidates.
326 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
328 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
330 * Invalidate page-table caches used by hardware walkers. Then we still
331 * need to RCU-sched wait while freeing the pages because software
332 * walkers can still be in-flight.
334 tlb_flush_mmu_tlbonly(tlb);
338 static void tlb_remove_table_smp_sync(void *arg)
340 /* Simply deliver the interrupt */
343 static void tlb_remove_table_one(void *table)
346 * This isn't an RCU grace period and hence the page-tables cannot be
347 * assumed to be actually RCU-freed.
349 * It is however sufficient for software page-table walkers that rely on
350 * IRQ disabling. See the comment near struct mmu_table_batch.
352 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
353 __tlb_remove_table(table);
356 static void tlb_remove_table_rcu(struct rcu_head *head)
358 struct mmu_table_batch *batch;
361 batch = container_of(head, struct mmu_table_batch, rcu);
363 for (i = 0; i < batch->nr; i++)
364 __tlb_remove_table(batch->tables[i]);
366 free_page((unsigned long)batch);
369 void tlb_table_flush(struct mmu_gather *tlb)
371 struct mmu_table_batch **batch = &tlb->batch;
374 tlb_table_invalidate(tlb);
375 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
380 void tlb_remove_table(struct mmu_gather *tlb, void *table)
382 struct mmu_table_batch **batch = &tlb->batch;
384 if (*batch == NULL) {
385 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
386 if (*batch == NULL) {
387 tlb_table_invalidate(tlb);
388 tlb_remove_table_one(table);
394 (*batch)->tables[(*batch)->nr++] = table;
395 if ((*batch)->nr == MAX_TABLE_BATCH)
396 tlb_table_flush(tlb);
399 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
402 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
403 * @tlb: the mmu_gather structure to initialize
404 * @mm: the mm_struct of the target address space
405 * @start: start of the region that will be removed from the page-table
406 * @end: end of the region that will be removed from the page-table
408 * Called to initialize an (on-stack) mmu_gather structure for page-table
409 * tear-down from @mm. The @start and @end are set to 0 and -1
410 * respectively when @mm is without users and we're going to destroy
411 * the full address space (exit/execve).
413 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
414 unsigned long start, unsigned long end)
416 arch_tlb_gather_mmu(tlb, mm, start, end);
417 inc_tlb_flush_pending(tlb->mm);
420 void tlb_finish_mmu(struct mmu_gather *tlb,
421 unsigned long start, unsigned long end)
424 * If there are parallel threads are doing PTE changes on same range
425 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
426 * flush by batching, a thread has stable TLB entry can fail to flush
427 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
428 * forcefully if we detect parallel PTE batching threads.
430 bool force = mm_tlb_flush_nested(tlb->mm);
432 arch_tlb_finish_mmu(tlb, start, end, force);
433 dec_tlb_flush_pending(tlb->mm);
437 * Note: this doesn't free the actual pages themselves. That
438 * has been handled earlier when unmapping all the memory regions.
440 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
443 pgtable_t token = pmd_pgtable(*pmd);
445 pte_free_tlb(tlb, token, addr);
446 mm_dec_nr_ptes(tlb->mm);
449 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
450 unsigned long addr, unsigned long end,
451 unsigned long floor, unsigned long ceiling)
458 pmd = pmd_offset(pud, addr);
460 next = pmd_addr_end(addr, end);
461 if (pmd_none_or_clear_bad(pmd))
463 free_pte_range(tlb, pmd, addr);
464 } while (pmd++, addr = next, addr != end);
474 if (end - 1 > ceiling - 1)
477 pmd = pmd_offset(pud, start);
479 pmd_free_tlb(tlb, pmd, start);
480 mm_dec_nr_pmds(tlb->mm);
483 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
484 unsigned long addr, unsigned long end,
485 unsigned long floor, unsigned long ceiling)
492 pud = pud_offset(p4d, addr);
494 next = pud_addr_end(addr, end);
495 if (pud_none_or_clear_bad(pud))
497 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
498 } while (pud++, addr = next, addr != end);
508 if (end - 1 > ceiling - 1)
511 pud = pud_offset(p4d, start);
513 pud_free_tlb(tlb, pud, start);
514 mm_dec_nr_puds(tlb->mm);
517 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
518 unsigned long addr, unsigned long end,
519 unsigned long floor, unsigned long ceiling)
526 p4d = p4d_offset(pgd, addr);
528 next = p4d_addr_end(addr, end);
529 if (p4d_none_or_clear_bad(p4d))
531 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
532 } while (p4d++, addr = next, addr != end);
538 ceiling &= PGDIR_MASK;
542 if (end - 1 > ceiling - 1)
545 p4d = p4d_offset(pgd, start);
547 p4d_free_tlb(tlb, p4d, start);
551 * This function frees user-level page tables of a process.
553 void free_pgd_range(struct mmu_gather *tlb,
554 unsigned long addr, unsigned long end,
555 unsigned long floor, unsigned long ceiling)
561 * The next few lines have given us lots of grief...
563 * Why are we testing PMD* at this top level? Because often
564 * there will be no work to do at all, and we'd prefer not to
565 * go all the way down to the bottom just to discover that.
567 * Why all these "- 1"s? Because 0 represents both the bottom
568 * of the address space and the top of it (using -1 for the
569 * top wouldn't help much: the masks would do the wrong thing).
570 * The rule is that addr 0 and floor 0 refer to the bottom of
571 * the address space, but end 0 and ceiling 0 refer to the top
572 * Comparisons need to use "end - 1" and "ceiling - 1" (though
573 * that end 0 case should be mythical).
575 * Wherever addr is brought up or ceiling brought down, we must
576 * be careful to reject "the opposite 0" before it confuses the
577 * subsequent tests. But what about where end is brought down
578 * by PMD_SIZE below? no, end can't go down to 0 there.
580 * Whereas we round start (addr) and ceiling down, by different
581 * masks at different levels, in order to test whether a table
582 * now has no other vmas using it, so can be freed, we don't
583 * bother to round floor or end up - the tests don't need that.
597 if (end - 1 > ceiling - 1)
602 * We add page table cache pages with PAGE_SIZE,
603 * (see pte_free_tlb()), flush the tlb if we need
605 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
606 pgd = pgd_offset(tlb->mm, addr);
608 next = pgd_addr_end(addr, end);
609 if (pgd_none_or_clear_bad(pgd))
611 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
612 } while (pgd++, addr = next, addr != end);
615 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
616 unsigned long floor, unsigned long ceiling)
619 struct vm_area_struct *next = vma->vm_next;
620 unsigned long addr = vma->vm_start;
623 * Hide vma from rmap and truncate_pagecache before freeing
626 unlink_anon_vmas(vma);
627 unlink_file_vma(vma);
629 if (is_vm_hugetlb_page(vma)) {
630 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
631 floor, next ? next->vm_start : ceiling);
634 * Optimization: gather nearby vmas into one call down
636 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
637 && !is_vm_hugetlb_page(next)) {
640 unlink_anon_vmas(vma);
641 unlink_file_vma(vma);
643 free_pgd_range(tlb, addr, vma->vm_end,
644 floor, next ? next->vm_start : ceiling);
650 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
653 pgtable_t new = pte_alloc_one(mm, address);
658 * Ensure all pte setup (eg. pte page lock and page clearing) are
659 * visible before the pte is made visible to other CPUs by being
660 * put into page tables.
662 * The other side of the story is the pointer chasing in the page
663 * table walking code (when walking the page table without locking;
664 * ie. most of the time). Fortunately, these data accesses consist
665 * of a chain of data-dependent loads, meaning most CPUs (alpha
666 * being the notable exception) will already guarantee loads are
667 * seen in-order. See the alpha page table accessors for the
668 * smp_read_barrier_depends() barriers in page table walking code.
670 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
672 ptl = pmd_lock(mm, pmd);
673 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
675 pmd_populate(mm, pmd, new);
684 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
686 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
690 smp_wmb(); /* See comment in __pte_alloc */
692 spin_lock(&init_mm.page_table_lock);
693 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
694 pmd_populate_kernel(&init_mm, pmd, new);
697 spin_unlock(&init_mm.page_table_lock);
699 pte_free_kernel(&init_mm, new);
703 static inline void init_rss_vec(int *rss)
705 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
708 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
712 if (current->mm == mm)
714 for (i = 0; i < NR_MM_COUNTERS; i++)
716 add_mm_counter(mm, i, rss[i]);
720 * This function is called to print an error when a bad pte
721 * is found. For example, we might have a PFN-mapped pte in
722 * a region that doesn't allow it.
724 * The calling function must still handle the error.
726 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
727 pte_t pte, struct page *page)
729 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
730 p4d_t *p4d = p4d_offset(pgd, addr);
731 pud_t *pud = pud_offset(p4d, addr);
732 pmd_t *pmd = pmd_offset(pud, addr);
733 struct address_space *mapping;
735 static unsigned long resume;
736 static unsigned long nr_shown;
737 static unsigned long nr_unshown;
740 * Allow a burst of 60 reports, then keep quiet for that minute;
741 * or allow a steady drip of one report per second.
743 if (nr_shown == 60) {
744 if (time_before(jiffies, resume)) {
749 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
756 resume = jiffies + 60 * HZ;
758 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
759 index = linear_page_index(vma, addr);
761 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
763 (long long)pte_val(pte), (long long)pmd_val(*pmd));
765 dump_page(page, "bad pte");
766 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
767 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
768 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
770 vma->vm_ops ? vma->vm_ops->fault : NULL,
771 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
772 mapping ? mapping->a_ops->readpage : NULL);
774 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
778 * vm_normal_page -- This function gets the "struct page" associated with a pte.
780 * "Special" mappings do not wish to be associated with a "struct page" (either
781 * it doesn't exist, or it exists but they don't want to touch it). In this
782 * case, NULL is returned here. "Normal" mappings do have a struct page.
784 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
785 * pte bit, in which case this function is trivial. Secondly, an architecture
786 * may not have a spare pte bit, which requires a more complicated scheme,
789 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
790 * special mapping (even if there are underlying and valid "struct pages").
791 * COWed pages of a VM_PFNMAP are always normal.
793 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
794 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
795 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
796 * mapping will always honor the rule
798 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
800 * And for normal mappings this is false.
802 * This restricts such mappings to be a linear translation from virtual address
803 * to pfn. To get around this restriction, we allow arbitrary mappings so long
804 * as the vma is not a COW mapping; in that case, we know that all ptes are
805 * special (because none can have been COWed).
808 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
810 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
811 * page" backing, however the difference is that _all_ pages with a struct
812 * page (that is, those where pfn_valid is true) are refcounted and considered
813 * normal pages by the VM. The disadvantage is that pages are refcounted
814 * (which can be slower and simply not an option for some PFNMAP users). The
815 * advantage is that we don't have to follow the strict linearity rule of
816 * PFNMAP mappings in order to support COWable mappings.
819 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
820 pte_t pte, bool with_public_device)
822 unsigned long pfn = pte_pfn(pte);
824 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
825 if (likely(!pte_special(pte)))
827 if (vma->vm_ops && vma->vm_ops->find_special_page)
828 return vma->vm_ops->find_special_page(vma, addr);
829 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
831 if (is_zero_pfn(pfn))
835 * Device public pages are special pages (they are ZONE_DEVICE
836 * pages but different from persistent memory). They behave
837 * allmost like normal pages. The difference is that they are
838 * not on the lru and thus should never be involve with any-
839 * thing that involve lru manipulation (mlock, numa balancing,
842 * This is why we still want to return NULL for such page from
843 * vm_normal_page() so that we do not have to special case all
844 * call site of vm_normal_page().
846 if (likely(pfn <= highest_memmap_pfn)) {
847 struct page *page = pfn_to_page(pfn);
849 if (is_device_public_page(page)) {
850 if (with_public_device)
859 print_bad_pte(vma, addr, pte, NULL);
863 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
865 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
866 if (vma->vm_flags & VM_MIXEDMAP) {
872 off = (addr - vma->vm_start) >> PAGE_SHIFT;
873 if (pfn == vma->vm_pgoff + off)
875 if (!is_cow_mapping(vma->vm_flags))
880 if (is_zero_pfn(pfn))
884 if (unlikely(pfn > highest_memmap_pfn)) {
885 print_bad_pte(vma, addr, pte, NULL);
890 * NOTE! We still have PageReserved() pages in the page tables.
891 * eg. VDSO mappings can cause them to exist.
894 return pfn_to_page(pfn);
897 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
898 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
901 unsigned long pfn = pmd_pfn(pmd);
904 * There is no pmd_special() but there may be special pmds, e.g.
905 * in a direct-access (dax) mapping, so let's just replicate the
906 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
908 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
909 if (vma->vm_flags & VM_MIXEDMAP) {
915 off = (addr - vma->vm_start) >> PAGE_SHIFT;
916 if (pfn == vma->vm_pgoff + off)
918 if (!is_cow_mapping(vma->vm_flags))
925 if (is_zero_pfn(pfn))
927 if (unlikely(pfn > highest_memmap_pfn))
931 * NOTE! We still have PageReserved() pages in the page tables.
932 * eg. VDSO mappings can cause them to exist.
935 return pfn_to_page(pfn);
940 * copy one vm_area from one task to the other. Assumes the page tables
941 * already present in the new task to be cleared in the whole range
942 * covered by this vma.
945 static inline unsigned long
946 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
947 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
948 unsigned long addr, int *rss)
950 unsigned long vm_flags = vma->vm_flags;
951 pte_t pte = *src_pte;
954 /* pte contains position in swap or file, so copy. */
955 if (unlikely(!pte_present(pte))) {
956 swp_entry_t entry = pte_to_swp_entry(pte);
958 if (likely(!non_swap_entry(entry))) {
959 if (swap_duplicate(entry) < 0)
962 /* make sure dst_mm is on swapoff's mmlist. */
963 if (unlikely(list_empty(&dst_mm->mmlist))) {
964 spin_lock(&mmlist_lock);
965 if (list_empty(&dst_mm->mmlist))
966 list_add(&dst_mm->mmlist,
968 spin_unlock(&mmlist_lock);
971 } else if (is_migration_entry(entry)) {
972 page = migration_entry_to_page(entry);
974 rss[mm_counter(page)]++;
976 if (is_write_migration_entry(entry) &&
977 is_cow_mapping(vm_flags)) {
979 * COW mappings require pages in both
980 * parent and child to be set to read.
982 make_migration_entry_read(&entry);
983 pte = swp_entry_to_pte(entry);
984 if (pte_swp_soft_dirty(*src_pte))
985 pte = pte_swp_mksoft_dirty(pte);
986 set_pte_at(src_mm, addr, src_pte, pte);
988 } else if (is_device_private_entry(entry)) {
989 page = device_private_entry_to_page(entry);
992 * Update rss count even for unaddressable pages, as
993 * they should treated just like normal pages in this
996 * We will likely want to have some new rss counters
997 * for unaddressable pages, at some point. But for now
998 * keep things as they are.
1001 rss[mm_counter(page)]++;
1002 page_dup_rmap(page, false);
1005 * We do not preserve soft-dirty information, because so
1006 * far, checkpoint/restore is the only feature that
1007 * requires that. And checkpoint/restore does not work
1008 * when a device driver is involved (you cannot easily
1009 * save and restore device driver state).
1011 if (is_write_device_private_entry(entry) &&
1012 is_cow_mapping(vm_flags)) {
1013 make_device_private_entry_read(&entry);
1014 pte = swp_entry_to_pte(entry);
1015 set_pte_at(src_mm, addr, src_pte, pte);
1022 * If it's a COW mapping, write protect it both
1023 * in the parent and the child
1025 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
1026 ptep_set_wrprotect(src_mm, addr, src_pte);
1027 pte = pte_wrprotect(pte);
1031 * If it's a shared mapping, mark it clean in
1034 if (vm_flags & VM_SHARED)
1035 pte = pte_mkclean(pte);
1036 pte = pte_mkold(pte);
1038 page = vm_normal_page(vma, addr, pte);
1041 page_dup_rmap(page, false);
1042 rss[mm_counter(page)]++;
1043 } else if (pte_devmap(pte)) {
1044 page = pte_page(pte);
1047 * Cache coherent device memory behave like regular page and
1048 * not like persistent memory page. For more informations see
1049 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1051 if (is_device_public_page(page)) {
1053 page_dup_rmap(page, false);
1054 rss[mm_counter(page)]++;
1059 set_pte_at(dst_mm, addr, dst_pte, pte);
1063 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1064 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1065 unsigned long addr, unsigned long end)
1067 pte_t *orig_src_pte, *orig_dst_pte;
1068 pte_t *src_pte, *dst_pte;
1069 spinlock_t *src_ptl, *dst_ptl;
1071 int rss[NR_MM_COUNTERS];
1072 swp_entry_t entry = (swp_entry_t){0};
1077 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1080 src_pte = pte_offset_map(src_pmd, addr);
1081 src_ptl = pte_lockptr(src_mm, src_pmd);
1082 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1083 orig_src_pte = src_pte;
1084 orig_dst_pte = dst_pte;
1085 arch_enter_lazy_mmu_mode();
1089 * We are holding two locks at this point - either of them
1090 * could generate latencies in another task on another CPU.
1092 if (progress >= 32) {
1094 if (need_resched() ||
1095 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1098 if (pte_none(*src_pte)) {
1102 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1107 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1109 arch_leave_lazy_mmu_mode();
1110 spin_unlock(src_ptl);
1111 pte_unmap(orig_src_pte);
1112 add_mm_rss_vec(dst_mm, rss);
1113 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1117 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1126 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1127 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1128 unsigned long addr, unsigned long end)
1130 pmd_t *src_pmd, *dst_pmd;
1133 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1136 src_pmd = pmd_offset(src_pud, addr);
1138 next = pmd_addr_end(addr, end);
1139 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1140 || pmd_devmap(*src_pmd)) {
1142 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1143 err = copy_huge_pmd(dst_mm, src_mm,
1144 dst_pmd, src_pmd, addr, vma);
1151 if (pmd_none_or_clear_bad(src_pmd))
1153 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1156 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1160 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1161 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1162 unsigned long addr, unsigned long end)
1164 pud_t *src_pud, *dst_pud;
1167 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1170 src_pud = pud_offset(src_p4d, addr);
1172 next = pud_addr_end(addr, end);
1173 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1176 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1177 err = copy_huge_pud(dst_mm, src_mm,
1178 dst_pud, src_pud, addr, vma);
1185 if (pud_none_or_clear_bad(src_pud))
1187 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1190 } while (dst_pud++, src_pud++, addr = next, addr != end);
1194 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1195 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1196 unsigned long addr, unsigned long end)
1198 p4d_t *src_p4d, *dst_p4d;
1201 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1204 src_p4d = p4d_offset(src_pgd, addr);
1206 next = p4d_addr_end(addr, end);
1207 if (p4d_none_or_clear_bad(src_p4d))
1209 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1212 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1216 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1217 struct vm_area_struct *vma)
1219 pgd_t *src_pgd, *dst_pgd;
1221 unsigned long addr = vma->vm_start;
1222 unsigned long end = vma->vm_end;
1223 unsigned long mmun_start; /* For mmu_notifiers */
1224 unsigned long mmun_end; /* For mmu_notifiers */
1229 * Don't copy ptes where a page fault will fill them correctly.
1230 * Fork becomes much lighter when there are big shared or private
1231 * readonly mappings. The tradeoff is that copy_page_range is more
1232 * efficient than faulting.
1234 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1238 if (is_vm_hugetlb_page(vma))
1239 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1241 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1243 * We do not free on error cases below as remove_vma
1244 * gets called on error from higher level routine
1246 ret = track_pfn_copy(vma);
1252 * We need to invalidate the secondary MMU mappings only when
1253 * there could be a permission downgrade on the ptes of the
1254 * parent mm. And a permission downgrade will only happen if
1255 * is_cow_mapping() returns true.
1257 is_cow = is_cow_mapping(vma->vm_flags);
1261 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1265 dst_pgd = pgd_offset(dst_mm, addr);
1266 src_pgd = pgd_offset(src_mm, addr);
1268 next = pgd_addr_end(addr, end);
1269 if (pgd_none_or_clear_bad(src_pgd))
1271 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1272 vma, addr, next))) {
1276 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1279 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1283 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1284 struct vm_area_struct *vma, pmd_t *pmd,
1285 unsigned long addr, unsigned long end,
1286 struct zap_details *details)
1288 struct mm_struct *mm = tlb->mm;
1289 int force_flush = 0;
1290 int rss[NR_MM_COUNTERS];
1296 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1299 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1301 flush_tlb_batched_pending(mm);
1302 arch_enter_lazy_mmu_mode();
1305 if (pte_none(ptent))
1308 if (pte_present(ptent)) {
1311 page = _vm_normal_page(vma, addr, ptent, true);
1312 if (unlikely(details) && page) {
1314 * unmap_shared_mapping_pages() wants to
1315 * invalidate cache without truncating:
1316 * unmap shared but keep private pages.
1318 if (details->check_mapping &&
1319 details->check_mapping != page_rmapping(page))
1322 ptent = ptep_get_and_clear_full(mm, addr, pte,
1324 tlb_remove_tlb_entry(tlb, pte, addr);
1325 if (unlikely(!page))
1328 if (!PageAnon(page)) {
1329 if (pte_dirty(ptent)) {
1331 set_page_dirty(page);
1333 if (pte_young(ptent) &&
1334 likely(!(vma->vm_flags & VM_SEQ_READ)))
1335 mark_page_accessed(page);
1337 rss[mm_counter(page)]--;
1338 page_remove_rmap(page, false);
1339 if (unlikely(page_mapcount(page) < 0))
1340 print_bad_pte(vma, addr, ptent, page);
1341 if (unlikely(__tlb_remove_page(tlb, page))) {
1349 entry = pte_to_swp_entry(ptent);
1350 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1351 struct page *page = device_private_entry_to_page(entry);
1353 if (unlikely(details && details->check_mapping)) {
1355 * unmap_shared_mapping_pages() wants to
1356 * invalidate cache without truncating:
1357 * unmap shared but keep private pages.
1359 if (details->check_mapping !=
1360 page_rmapping(page))
1364 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1365 rss[mm_counter(page)]--;
1366 page_remove_rmap(page, false);
1371 /* If details->check_mapping, we leave swap entries. */
1372 if (unlikely(details))
1375 entry = pte_to_swp_entry(ptent);
1376 if (!non_swap_entry(entry))
1378 else if (is_migration_entry(entry)) {
1381 page = migration_entry_to_page(entry);
1382 rss[mm_counter(page)]--;
1384 if (unlikely(!free_swap_and_cache(entry)))
1385 print_bad_pte(vma, addr, ptent, NULL);
1386 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1387 } while (pte++, addr += PAGE_SIZE, addr != end);
1389 add_mm_rss_vec(mm, rss);
1390 arch_leave_lazy_mmu_mode();
1392 /* Do the actual TLB flush before dropping ptl */
1394 tlb_flush_mmu_tlbonly(tlb);
1395 pte_unmap_unlock(start_pte, ptl);
1398 * If we forced a TLB flush (either due to running out of
1399 * batch buffers or because we needed to flush dirty TLB
1400 * entries before releasing the ptl), free the batched
1401 * memory too. Restart if we didn't do everything.
1405 tlb_flush_mmu_free(tlb);
1413 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1414 struct vm_area_struct *vma, pud_t *pud,
1415 unsigned long addr, unsigned long end,
1416 struct zap_details *details)
1421 pmd = pmd_offset(pud, addr);
1423 next = pmd_addr_end(addr, end);
1424 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1425 if (next - addr != HPAGE_PMD_SIZE)
1426 __split_huge_pmd(vma, pmd, addr, false, NULL);
1427 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1432 * Here there can be other concurrent MADV_DONTNEED or
1433 * trans huge page faults running, and if the pmd is
1434 * none or trans huge it can change under us. This is
1435 * because MADV_DONTNEED holds the mmap_sem in read
1438 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1440 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1443 } while (pmd++, addr = next, addr != end);
1448 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1449 struct vm_area_struct *vma, p4d_t *p4d,
1450 unsigned long addr, unsigned long end,
1451 struct zap_details *details)
1456 pud = pud_offset(p4d, addr);
1458 next = pud_addr_end(addr, end);
1459 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1460 if (next - addr != HPAGE_PUD_SIZE) {
1461 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1462 split_huge_pud(vma, pud, addr);
1463 } else if (zap_huge_pud(tlb, vma, pud, addr))
1467 if (pud_none_or_clear_bad(pud))
1469 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1472 } while (pud++, addr = next, addr != end);
1477 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1478 struct vm_area_struct *vma, pgd_t *pgd,
1479 unsigned long addr, unsigned long end,
1480 struct zap_details *details)
1485 p4d = p4d_offset(pgd, addr);
1487 next = p4d_addr_end(addr, end);
1488 if (p4d_none_or_clear_bad(p4d))
1490 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1491 } while (p4d++, addr = next, addr != end);
1496 void unmap_page_range(struct mmu_gather *tlb,
1497 struct vm_area_struct *vma,
1498 unsigned long addr, unsigned long end,
1499 struct zap_details *details)
1504 BUG_ON(addr >= end);
1505 tlb_start_vma(tlb, vma);
1506 pgd = pgd_offset(vma->vm_mm, addr);
1508 next = pgd_addr_end(addr, end);
1509 if (pgd_none_or_clear_bad(pgd))
1511 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1512 } while (pgd++, addr = next, addr != end);
1513 tlb_end_vma(tlb, vma);
1517 static void unmap_single_vma(struct mmu_gather *tlb,
1518 struct vm_area_struct *vma, unsigned long start_addr,
1519 unsigned long end_addr,
1520 struct zap_details *details)
1522 unsigned long start = max(vma->vm_start, start_addr);
1525 if (start >= vma->vm_end)
1527 end = min(vma->vm_end, end_addr);
1528 if (end <= vma->vm_start)
1532 uprobe_munmap(vma, start, end);
1534 if (unlikely(vma->vm_flags & VM_PFNMAP))
1535 untrack_pfn(vma, 0, 0);
1538 if (unlikely(is_vm_hugetlb_page(vma))) {
1540 * It is undesirable to test vma->vm_file as it
1541 * should be non-null for valid hugetlb area.
1542 * However, vm_file will be NULL in the error
1543 * cleanup path of mmap_region. When
1544 * hugetlbfs ->mmap method fails,
1545 * mmap_region() nullifies vma->vm_file
1546 * before calling this function to clean up.
1547 * Since no pte has actually been setup, it is
1548 * safe to do nothing in this case.
1551 i_mmap_lock_write(vma->vm_file->f_mapping);
1552 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1553 i_mmap_unlock_write(vma->vm_file->f_mapping);
1556 unmap_page_range(tlb, vma, start, end, details);
1561 * unmap_vmas - unmap a range of memory covered by a list of vma's
1562 * @tlb: address of the caller's struct mmu_gather
1563 * @vma: the starting vma
1564 * @start_addr: virtual address at which to start unmapping
1565 * @end_addr: virtual address at which to end unmapping
1567 * Unmap all pages in the vma list.
1569 * Only addresses between `start' and `end' will be unmapped.
1571 * The VMA list must be sorted in ascending virtual address order.
1573 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1574 * range after unmap_vmas() returns. So the only responsibility here is to
1575 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1576 * drops the lock and schedules.
1578 void unmap_vmas(struct mmu_gather *tlb,
1579 struct vm_area_struct *vma, unsigned long start_addr,
1580 unsigned long end_addr)
1582 struct mm_struct *mm = vma->vm_mm;
1584 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1585 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1586 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1587 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1591 * zap_page_range - remove user pages in a given range
1592 * @vma: vm_area_struct holding the applicable pages
1593 * @start: starting address of pages to zap
1594 * @size: number of bytes to zap
1596 * Caller must protect the VMA list
1598 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1601 struct mm_struct *mm = vma->vm_mm;
1602 struct mmu_gather tlb;
1603 unsigned long end = start + size;
1606 tlb_gather_mmu(&tlb, mm, start, end);
1607 update_hiwater_rss(mm);
1608 mmu_notifier_invalidate_range_start(mm, start, end);
1609 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1610 unmap_single_vma(&tlb, vma, start, end, NULL);
1611 mmu_notifier_invalidate_range_end(mm, start, end);
1612 tlb_finish_mmu(&tlb, start, end);
1616 * zap_page_range_single - remove user pages in a given range
1617 * @vma: vm_area_struct holding the applicable pages
1618 * @address: starting address of pages to zap
1619 * @size: number of bytes to zap
1620 * @details: details of shared cache invalidation
1622 * The range must fit into one VMA.
1624 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1625 unsigned long size, struct zap_details *details)
1627 struct mm_struct *mm = vma->vm_mm;
1628 struct mmu_gather tlb;
1629 unsigned long end = address + size;
1632 tlb_gather_mmu(&tlb, mm, address, end);
1633 update_hiwater_rss(mm);
1634 mmu_notifier_invalidate_range_start(mm, address, end);
1635 unmap_single_vma(&tlb, vma, address, end, details);
1636 mmu_notifier_invalidate_range_end(mm, address, end);
1637 tlb_finish_mmu(&tlb, address, end);
1641 * zap_vma_ptes - remove ptes mapping the vma
1642 * @vma: vm_area_struct holding ptes to be zapped
1643 * @address: starting address of pages to zap
1644 * @size: number of bytes to zap
1646 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1648 * The entire address range must be fully contained within the vma.
1651 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1654 if (address < vma->vm_start || address + size > vma->vm_end ||
1655 !(vma->vm_flags & VM_PFNMAP))
1658 zap_page_range_single(vma, address, size, NULL);
1660 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1662 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1670 pgd = pgd_offset(mm, addr);
1671 p4d = p4d_alloc(mm, pgd, addr);
1674 pud = pud_alloc(mm, p4d, addr);
1677 pmd = pmd_alloc(mm, pud, addr);
1681 VM_BUG_ON(pmd_trans_huge(*pmd));
1682 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1686 * This is the old fallback for page remapping.
1688 * For historical reasons, it only allows reserved pages. Only
1689 * old drivers should use this, and they needed to mark their
1690 * pages reserved for the old functions anyway.
1692 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1693 struct page *page, pgprot_t prot)
1695 struct mm_struct *mm = vma->vm_mm;
1704 flush_dcache_page(page);
1705 pte = get_locked_pte(mm, addr, &ptl);
1709 if (!pte_none(*pte))
1712 /* Ok, finally just insert the thing.. */
1714 inc_mm_counter_fast(mm, mm_counter_file(page));
1715 page_add_file_rmap(page, false);
1716 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1719 pte_unmap_unlock(pte, ptl);
1722 pte_unmap_unlock(pte, ptl);
1728 * vm_insert_page - insert single page into user vma
1729 * @vma: user vma to map to
1730 * @addr: target user address of this page
1731 * @page: source kernel page
1733 * This allows drivers to insert individual pages they've allocated
1736 * The page has to be a nice clean _individual_ kernel allocation.
1737 * If you allocate a compound page, you need to have marked it as
1738 * such (__GFP_COMP), or manually just split the page up yourself
1739 * (see split_page()).
1741 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1742 * took an arbitrary page protection parameter. This doesn't allow
1743 * that. Your vma protection will have to be set up correctly, which
1744 * means that if you want a shared writable mapping, you'd better
1745 * ask for a shared writable mapping!
1747 * The page does not need to be reserved.
1749 * Usually this function is called from f_op->mmap() handler
1750 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1751 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1752 * function from other places, for example from page-fault handler.
1754 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1757 if (addr < vma->vm_start || addr >= vma->vm_end)
1759 if (!page_count(page))
1761 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1762 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1763 BUG_ON(vma->vm_flags & VM_PFNMAP);
1764 vma->vm_flags |= VM_MIXEDMAP;
1766 return insert_page(vma, addr, page, vma->vm_page_prot);
1768 EXPORT_SYMBOL(vm_insert_page);
1770 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1771 pfn_t pfn, pgprot_t prot, bool mkwrite)
1773 struct mm_struct *mm = vma->vm_mm;
1779 pte = get_locked_pte(mm, addr, &ptl);
1783 if (!pte_none(*pte)) {
1786 * For read faults on private mappings the PFN passed
1787 * in may not match the PFN we have mapped if the
1788 * mapped PFN is a writeable COW page. In the mkwrite
1789 * case we are creating a writable PTE for a shared
1790 * mapping and we expect the PFNs to match. If they
1791 * don't match, we are likely racing with block
1792 * allocation and mapping invalidation so just skip the
1795 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1796 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1799 entry = pte_mkyoung(*pte);
1800 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1801 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1802 update_mmu_cache(vma, addr, pte);
1807 /* Ok, finally just insert the thing.. */
1808 if (pfn_t_devmap(pfn))
1809 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1811 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1814 entry = pte_mkyoung(entry);
1815 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1818 set_pte_at(mm, addr, pte, entry);
1819 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1823 pte_unmap_unlock(pte, ptl);
1829 * vm_insert_pfn - insert single pfn into user vma
1830 * @vma: user vma to map to
1831 * @addr: target user address of this page
1832 * @pfn: source kernel pfn
1834 * Similar to vm_insert_page, this allows drivers to insert individual pages
1835 * they've allocated into a user vma. Same comments apply.
1837 * This function should only be called from a vm_ops->fault handler, and
1838 * in that case the handler should return NULL.
1840 * vma cannot be a COW mapping.
1842 * As this is called only for pages that do not currently exist, we
1843 * do not need to flush old virtual caches or the TLB.
1845 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1848 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1850 EXPORT_SYMBOL(vm_insert_pfn);
1853 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1854 * @vma: user vma to map to
1855 * @addr: target user address of this page
1856 * @pfn: source kernel pfn
1857 * @pgprot: pgprot flags for the inserted page
1859 * This is exactly like vm_insert_pfn, except that it allows drivers to
1860 * to override pgprot on a per-page basis.
1862 * This only makes sense for IO mappings, and it makes no sense for
1863 * cow mappings. In general, using multiple vmas is preferable;
1864 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1867 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1868 unsigned long pfn, pgprot_t pgprot)
1872 * Technically, architectures with pte_special can avoid all these
1873 * restrictions (same for remap_pfn_range). However we would like
1874 * consistency in testing and feature parity among all, so we should
1875 * try to keep these invariants in place for everybody.
1877 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1878 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1879 (VM_PFNMAP|VM_MIXEDMAP));
1880 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1881 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1883 if (addr < vma->vm_start || addr >= vma->vm_end)
1886 if (!pfn_modify_allowed(pfn, pgprot))
1889 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1891 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1896 EXPORT_SYMBOL(vm_insert_pfn_prot);
1898 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1900 /* these checks mirror the abort conditions in vm_normal_page */
1901 if (vma->vm_flags & VM_MIXEDMAP)
1903 if (pfn_t_devmap(pfn))
1905 if (pfn_t_special(pfn))
1907 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1912 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1913 pfn_t pfn, bool mkwrite)
1915 pgprot_t pgprot = vma->vm_page_prot;
1917 BUG_ON(!vm_mixed_ok(vma, pfn));
1919 if (addr < vma->vm_start || addr >= vma->vm_end)
1922 track_pfn_insert(vma, &pgprot, pfn);
1924 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1928 * If we don't have pte special, then we have to use the pfn_valid()
1929 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1930 * refcount the page if pfn_valid is true (hence insert_page rather
1931 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1932 * without pte special, it would there be refcounted as a normal page.
1934 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1935 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1939 * At this point we are committed to insert_page()
1940 * regardless of whether the caller specified flags that
1941 * result in pfn_t_has_page() == false.
1943 page = pfn_to_page(pfn_t_to_pfn(pfn));
1944 return insert_page(vma, addr, page, pgprot);
1946 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1949 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1952 return __vm_insert_mixed(vma, addr, pfn, false);
1955 EXPORT_SYMBOL(vm_insert_mixed);
1958 * If the insertion of PTE failed because someone else already added a
1959 * different entry in the mean time, we treat that as success as we assume
1960 * the same entry was actually inserted.
1963 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1964 unsigned long addr, pfn_t pfn)
1968 err = __vm_insert_mixed(vma, addr, pfn, true);
1970 return VM_FAULT_OOM;
1971 if (err < 0 && err != -EBUSY)
1972 return VM_FAULT_SIGBUS;
1973 return VM_FAULT_NOPAGE;
1975 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1978 * maps a range of physical memory into the requested pages. the old
1979 * mappings are removed. any references to nonexistent pages results
1980 * in null mappings (currently treated as "copy-on-access")
1982 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1983 unsigned long addr, unsigned long end,
1984 unsigned long pfn, pgprot_t prot)
1990 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1993 arch_enter_lazy_mmu_mode();
1995 BUG_ON(!pte_none(*pte));
1996 if (!pfn_modify_allowed(pfn, prot)) {
2000 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2002 } while (pte++, addr += PAGE_SIZE, addr != end);
2003 arch_leave_lazy_mmu_mode();
2004 pte_unmap_unlock(pte - 1, ptl);
2008 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2009 unsigned long addr, unsigned long end,
2010 unsigned long pfn, pgprot_t prot)
2016 pfn -= addr >> PAGE_SHIFT;
2017 pmd = pmd_alloc(mm, pud, addr);
2020 VM_BUG_ON(pmd_trans_huge(*pmd));
2022 next = pmd_addr_end(addr, end);
2023 err = remap_pte_range(mm, pmd, addr, next,
2024 pfn + (addr >> PAGE_SHIFT), prot);
2027 } while (pmd++, addr = next, addr != end);
2031 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2032 unsigned long addr, unsigned long end,
2033 unsigned long pfn, pgprot_t prot)
2039 pfn -= addr >> PAGE_SHIFT;
2040 pud = pud_alloc(mm, p4d, addr);
2044 next = pud_addr_end(addr, end);
2045 err = remap_pmd_range(mm, pud, addr, next,
2046 pfn + (addr >> PAGE_SHIFT), prot);
2049 } while (pud++, addr = next, addr != end);
2053 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2054 unsigned long addr, unsigned long end,
2055 unsigned long pfn, pgprot_t prot)
2061 pfn -= addr >> PAGE_SHIFT;
2062 p4d = p4d_alloc(mm, pgd, addr);
2066 next = p4d_addr_end(addr, end);
2067 err = remap_pud_range(mm, p4d, addr, next,
2068 pfn + (addr >> PAGE_SHIFT), prot);
2071 } while (p4d++, addr = next, addr != end);
2076 * remap_pfn_range - remap kernel memory to userspace
2077 * @vma: user vma to map to
2078 * @addr: target user address to start at
2079 * @pfn: physical address of kernel memory
2080 * @size: size of map area
2081 * @prot: page protection flags for this mapping
2083 * Note: this is only safe if the mm semaphore is held when called.
2085 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2086 unsigned long pfn, unsigned long size, pgprot_t prot)
2090 unsigned long end = addr + PAGE_ALIGN(size);
2091 struct mm_struct *mm = vma->vm_mm;
2092 unsigned long remap_pfn = pfn;
2096 * Physically remapped pages are special. Tell the
2097 * rest of the world about it:
2098 * VM_IO tells people not to look at these pages
2099 * (accesses can have side effects).
2100 * VM_PFNMAP tells the core MM that the base pages are just
2101 * raw PFN mappings, and do not have a "struct page" associated
2104 * Disable vma merging and expanding with mremap().
2106 * Omit vma from core dump, even when VM_IO turned off.
2108 * There's a horrible special case to handle copy-on-write
2109 * behaviour that some programs depend on. We mark the "original"
2110 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2111 * See vm_normal_page() for details.
2113 if (is_cow_mapping(vma->vm_flags)) {
2114 if (addr != vma->vm_start || end != vma->vm_end)
2116 vma->vm_pgoff = pfn;
2119 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2123 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2125 BUG_ON(addr >= end);
2126 pfn -= addr >> PAGE_SHIFT;
2127 pgd = pgd_offset(mm, addr);
2128 flush_cache_range(vma, addr, end);
2130 next = pgd_addr_end(addr, end);
2131 err = remap_p4d_range(mm, pgd, addr, next,
2132 pfn + (addr >> PAGE_SHIFT), prot);
2135 } while (pgd++, addr = next, addr != end);
2138 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2142 EXPORT_SYMBOL(remap_pfn_range);
2145 * vm_iomap_memory - remap memory to userspace
2146 * @vma: user vma to map to
2147 * @start: start of area
2148 * @len: size of area
2150 * This is a simplified io_remap_pfn_range() for common driver use. The
2151 * driver just needs to give us the physical memory range to be mapped,
2152 * we'll figure out the rest from the vma information.
2154 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2155 * whatever write-combining details or similar.
2157 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2159 unsigned long vm_len, pfn, pages;
2161 /* Check that the physical memory area passed in looks valid */
2162 if (start + len < start)
2165 * You *really* shouldn't map things that aren't page-aligned,
2166 * but we've historically allowed it because IO memory might
2167 * just have smaller alignment.
2169 len += start & ~PAGE_MASK;
2170 pfn = start >> PAGE_SHIFT;
2171 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2172 if (pfn + pages < pfn)
2175 /* We start the mapping 'vm_pgoff' pages into the area */
2176 if (vma->vm_pgoff > pages)
2178 pfn += vma->vm_pgoff;
2179 pages -= vma->vm_pgoff;
2181 /* Can we fit all of the mapping? */
2182 vm_len = vma->vm_end - vma->vm_start;
2183 if (vm_len >> PAGE_SHIFT > pages)
2186 /* Ok, let it rip */
2187 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2189 EXPORT_SYMBOL(vm_iomap_memory);
2191 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2192 unsigned long addr, unsigned long end,
2193 pte_fn_t fn, void *data)
2198 spinlock_t *uninitialized_var(ptl);
2200 pte = (mm == &init_mm) ?
2201 pte_alloc_kernel(pmd, addr) :
2202 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2206 BUG_ON(pmd_huge(*pmd));
2208 arch_enter_lazy_mmu_mode();
2210 token = pmd_pgtable(*pmd);
2213 err = fn(pte++, token, addr, data);
2216 } while (addr += PAGE_SIZE, addr != end);
2218 arch_leave_lazy_mmu_mode();
2221 pte_unmap_unlock(pte-1, ptl);
2225 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2226 unsigned long addr, unsigned long end,
2227 pte_fn_t fn, void *data)
2233 BUG_ON(pud_huge(*pud));
2235 pmd = pmd_alloc(mm, pud, addr);
2239 next = pmd_addr_end(addr, end);
2240 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2243 } while (pmd++, addr = next, addr != end);
2247 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2248 unsigned long addr, unsigned long end,
2249 pte_fn_t fn, void *data)
2255 pud = pud_alloc(mm, p4d, addr);
2259 next = pud_addr_end(addr, end);
2260 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2263 } while (pud++, addr = next, addr != end);
2267 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2268 unsigned long addr, unsigned long end,
2269 pte_fn_t fn, void *data)
2275 p4d = p4d_alloc(mm, pgd, addr);
2279 next = p4d_addr_end(addr, end);
2280 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2283 } while (p4d++, addr = next, addr != end);
2288 * Scan a region of virtual memory, filling in page tables as necessary
2289 * and calling a provided function on each leaf page table.
2291 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2292 unsigned long size, pte_fn_t fn, void *data)
2296 unsigned long end = addr + size;
2299 if (WARN_ON(addr >= end))
2302 pgd = pgd_offset(mm, addr);
2304 next = pgd_addr_end(addr, end);
2305 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2308 } while (pgd++, addr = next, addr != end);
2312 EXPORT_SYMBOL_GPL(apply_to_page_range);
2315 * handle_pte_fault chooses page fault handler according to an entry which was
2316 * read non-atomically. Before making any commitment, on those architectures
2317 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2318 * parts, do_swap_page must check under lock before unmapping the pte and
2319 * proceeding (but do_wp_page is only called after already making such a check;
2320 * and do_anonymous_page can safely check later on).
2322 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2323 pte_t *page_table, pte_t orig_pte)
2326 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2327 if (sizeof(pte_t) > sizeof(unsigned long)) {
2328 spinlock_t *ptl = pte_lockptr(mm, pmd);
2330 same = pte_same(*page_table, orig_pte);
2334 pte_unmap(page_table);
2338 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2340 debug_dma_assert_idle(src);
2343 * If the source page was a PFN mapping, we don't have
2344 * a "struct page" for it. We do a best-effort copy by
2345 * just copying from the original user address. If that
2346 * fails, we just zero-fill it. Live with it.
2348 if (unlikely(!src)) {
2349 void *kaddr = kmap_atomic(dst);
2350 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2353 * This really shouldn't fail, because the page is there
2354 * in the page tables. But it might just be unreadable,
2355 * in which case we just give up and fill the result with
2358 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2360 kunmap_atomic(kaddr);
2361 flush_dcache_page(dst);
2363 copy_user_highpage(dst, src, va, vma);
2366 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2368 struct file *vm_file = vma->vm_file;
2371 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2374 * Special mappings (e.g. VDSO) do not have any file so fake
2375 * a default GFP_KERNEL for them.
2381 * Notify the address space that the page is about to become writable so that
2382 * it can prohibit this or wait for the page to get into an appropriate state.
2384 * We do this without the lock held, so that it can sleep if it needs to.
2386 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2389 struct page *page = vmf->page;
2390 unsigned int old_flags = vmf->flags;
2392 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2394 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2395 /* Restore original flags so that caller is not surprised */
2396 vmf->flags = old_flags;
2397 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2399 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2401 if (!page->mapping) {
2403 return 0; /* retry */
2405 ret |= VM_FAULT_LOCKED;
2407 VM_BUG_ON_PAGE(!PageLocked(page), page);
2412 * Handle dirtying of a page in shared file mapping on a write fault.
2414 * The function expects the page to be locked and unlocks it.
2416 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2419 struct address_space *mapping;
2421 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2423 dirtied = set_page_dirty(page);
2424 VM_BUG_ON_PAGE(PageAnon(page), page);
2426 * Take a local copy of the address_space - page.mapping may be zeroed
2427 * by truncate after unlock_page(). The address_space itself remains
2428 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2429 * release semantics to prevent the compiler from undoing this copying.
2431 mapping = page_rmapping(page);
2434 if ((dirtied || page_mkwrite) && mapping) {
2436 * Some device drivers do not set page.mapping
2437 * but still dirty their pages
2439 balance_dirty_pages_ratelimited(mapping);
2443 file_update_time(vma->vm_file);
2447 * Handle write page faults for pages that can be reused in the current vma
2449 * This can happen either due to the mapping being with the VM_SHARED flag,
2450 * or due to us being the last reference standing to the page. In either
2451 * case, all we need to do here is to mark the page as writable and update
2452 * any related book-keeping.
2454 static inline void wp_page_reuse(struct vm_fault *vmf)
2455 __releases(vmf->ptl)
2457 struct vm_area_struct *vma = vmf->vma;
2458 struct page *page = vmf->page;
2461 * Clear the pages cpupid information as the existing
2462 * information potentially belongs to a now completely
2463 * unrelated process.
2466 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2468 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2469 entry = pte_mkyoung(vmf->orig_pte);
2470 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2471 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2472 update_mmu_cache(vma, vmf->address, vmf->pte);
2473 pte_unmap_unlock(vmf->pte, vmf->ptl);
2477 * Handle the case of a page which we actually need to copy to a new page.
2479 * Called with mmap_sem locked and the old page referenced, but
2480 * without the ptl held.
2482 * High level logic flow:
2484 * - Allocate a page, copy the content of the old page to the new one.
2485 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2486 * - Take the PTL. If the pte changed, bail out and release the allocated page
2487 * - If the pte is still the way we remember it, update the page table and all
2488 * relevant references. This includes dropping the reference the page-table
2489 * held to the old page, as well as updating the rmap.
2490 * - In any case, unlock the PTL and drop the reference we took to the old page.
2492 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2494 struct vm_area_struct *vma = vmf->vma;
2495 struct mm_struct *mm = vma->vm_mm;
2496 struct page *old_page = vmf->page;
2497 struct page *new_page = NULL;
2499 int page_copied = 0;
2500 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2501 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2502 struct mem_cgroup *memcg;
2504 if (unlikely(anon_vma_prepare(vma)))
2507 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2508 new_page = alloc_zeroed_user_highpage_movable(vma,
2513 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2517 cow_user_page(new_page, old_page, vmf->address, vma);
2520 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2523 __SetPageUptodate(new_page);
2525 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2528 * Re-check the pte - we dropped the lock
2530 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2531 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2533 if (!PageAnon(old_page)) {
2534 dec_mm_counter_fast(mm,
2535 mm_counter_file(old_page));
2536 inc_mm_counter_fast(mm, MM_ANONPAGES);
2539 inc_mm_counter_fast(mm, MM_ANONPAGES);
2541 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2542 entry = mk_pte(new_page, vma->vm_page_prot);
2543 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2545 * Clear the pte entry and flush it first, before updating the
2546 * pte with the new entry. This will avoid a race condition
2547 * seen in the presence of one thread doing SMC and another
2550 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2551 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2552 mem_cgroup_commit_charge(new_page, memcg, false, false);
2553 lru_cache_add_active_or_unevictable(new_page, vma);
2555 * We call the notify macro here because, when using secondary
2556 * mmu page tables (such as kvm shadow page tables), we want the
2557 * new page to be mapped directly into the secondary page table.
2559 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2560 update_mmu_cache(vma, vmf->address, vmf->pte);
2563 * Only after switching the pte to the new page may
2564 * we remove the mapcount here. Otherwise another
2565 * process may come and find the rmap count decremented
2566 * before the pte is switched to the new page, and
2567 * "reuse" the old page writing into it while our pte
2568 * here still points into it and can be read by other
2571 * The critical issue is to order this
2572 * page_remove_rmap with the ptp_clear_flush above.
2573 * Those stores are ordered by (if nothing else,)
2574 * the barrier present in the atomic_add_negative
2575 * in page_remove_rmap.
2577 * Then the TLB flush in ptep_clear_flush ensures that
2578 * no process can access the old page before the
2579 * decremented mapcount is visible. And the old page
2580 * cannot be reused until after the decremented
2581 * mapcount is visible. So transitively, TLBs to
2582 * old page will be flushed before it can be reused.
2584 page_remove_rmap(old_page, false);
2587 /* Free the old page.. */
2588 new_page = old_page;
2591 mem_cgroup_cancel_charge(new_page, memcg, false);
2597 pte_unmap_unlock(vmf->pte, vmf->ptl);
2599 * No need to double call mmu_notifier->invalidate_range() callback as
2600 * the above ptep_clear_flush_notify() did already call it.
2602 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2605 * Don't let another task, with possibly unlocked vma,
2606 * keep the mlocked page.
2608 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2609 lock_page(old_page); /* LRU manipulation */
2610 if (PageMlocked(old_page))
2611 munlock_vma_page(old_page);
2612 unlock_page(old_page);
2616 return page_copied ? VM_FAULT_WRITE : 0;
2622 return VM_FAULT_OOM;
2626 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2627 * writeable once the page is prepared
2629 * @vmf: structure describing the fault
2631 * This function handles all that is needed to finish a write page fault in a
2632 * shared mapping due to PTE being read-only once the mapped page is prepared.
2633 * It handles locking of PTE and modifying it. The function returns
2634 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2637 * The function expects the page to be locked or other protection against
2638 * concurrent faults / writeback (such as DAX radix tree locks).
2640 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2642 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2643 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2646 * We might have raced with another page fault while we released the
2647 * pte_offset_map_lock.
2649 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2650 pte_unmap_unlock(vmf->pte, vmf->ptl);
2651 return VM_FAULT_NOPAGE;
2658 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2661 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2663 struct vm_area_struct *vma = vmf->vma;
2665 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2668 pte_unmap_unlock(vmf->pte, vmf->ptl);
2669 vmf->flags |= FAULT_FLAG_MKWRITE;
2670 ret = vma->vm_ops->pfn_mkwrite(vmf);
2671 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2673 return finish_mkwrite_fault(vmf);
2676 return VM_FAULT_WRITE;
2679 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2680 __releases(vmf->ptl)
2682 struct vm_area_struct *vma = vmf->vma;
2684 get_page(vmf->page);
2686 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2689 pte_unmap_unlock(vmf->pte, vmf->ptl);
2690 tmp = do_page_mkwrite(vmf);
2691 if (unlikely(!tmp || (tmp &
2692 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2693 put_page(vmf->page);
2696 tmp = finish_mkwrite_fault(vmf);
2697 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2698 unlock_page(vmf->page);
2699 put_page(vmf->page);
2704 lock_page(vmf->page);
2706 fault_dirty_shared_page(vma, vmf->page);
2707 put_page(vmf->page);
2709 return VM_FAULT_WRITE;
2713 * This routine handles present pages, when users try to write
2714 * to a shared page. It is done by copying the page to a new address
2715 * and decrementing the shared-page counter for the old page.
2717 * Note that this routine assumes that the protection checks have been
2718 * done by the caller (the low-level page fault routine in most cases).
2719 * Thus we can safely just mark it writable once we've done any necessary
2722 * We also mark the page dirty at this point even though the page will
2723 * change only once the write actually happens. This avoids a few races,
2724 * and potentially makes it more efficient.
2726 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2727 * but allow concurrent faults), with pte both mapped and locked.
2728 * We return with mmap_sem still held, but pte unmapped and unlocked.
2730 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2731 __releases(vmf->ptl)
2733 struct vm_area_struct *vma = vmf->vma;
2735 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2738 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2741 * We should not cow pages in a shared writeable mapping.
2742 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2744 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2745 (VM_WRITE|VM_SHARED))
2746 return wp_pfn_shared(vmf);
2748 pte_unmap_unlock(vmf->pte, vmf->ptl);
2749 return wp_page_copy(vmf);
2753 * Take out anonymous pages first, anonymous shared vmas are
2754 * not dirty accountable.
2756 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2757 int total_map_swapcount;
2758 if (!trylock_page(vmf->page)) {
2759 get_page(vmf->page);
2760 pte_unmap_unlock(vmf->pte, vmf->ptl);
2761 lock_page(vmf->page);
2762 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2763 vmf->address, &vmf->ptl);
2764 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2765 unlock_page(vmf->page);
2766 pte_unmap_unlock(vmf->pte, vmf->ptl);
2767 put_page(vmf->page);
2770 put_page(vmf->page);
2772 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2773 if (total_map_swapcount == 1) {
2775 * The page is all ours. Move it to
2776 * our anon_vma so the rmap code will
2777 * not search our parent or siblings.
2778 * Protected against the rmap code by
2781 page_move_anon_rmap(vmf->page, vma);
2783 unlock_page(vmf->page);
2785 return VM_FAULT_WRITE;
2787 unlock_page(vmf->page);
2788 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2789 (VM_WRITE|VM_SHARED))) {
2790 return wp_page_shared(vmf);
2794 * Ok, we need to copy. Oh, well..
2796 get_page(vmf->page);
2798 pte_unmap_unlock(vmf->pte, vmf->ptl);
2799 return wp_page_copy(vmf);
2802 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2803 unsigned long start_addr, unsigned long end_addr,
2804 struct zap_details *details)
2806 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2809 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2810 struct zap_details *details)
2812 struct vm_area_struct *vma;
2813 pgoff_t vba, vea, zba, zea;
2815 vma_interval_tree_foreach(vma, root,
2816 details->first_index, details->last_index) {
2818 vba = vma->vm_pgoff;
2819 vea = vba + vma_pages(vma) - 1;
2820 zba = details->first_index;
2823 zea = details->last_index;
2827 unmap_mapping_range_vma(vma,
2828 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2829 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2835 * unmap_mapping_pages() - Unmap pages from processes.
2836 * @mapping: The address space containing pages to be unmapped.
2837 * @start: Index of first page to be unmapped.
2838 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2839 * @even_cows: Whether to unmap even private COWed pages.
2841 * Unmap the pages in this address space from any userspace process which
2842 * has them mmaped. Generally, you want to remove COWed pages as well when
2843 * a file is being truncated, but not when invalidating pages from the page
2846 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2847 pgoff_t nr, bool even_cows)
2849 struct zap_details details = { };
2851 details.check_mapping = even_cows ? NULL : mapping;
2852 details.first_index = start;
2853 details.last_index = start + nr - 1;
2854 if (details.last_index < details.first_index)
2855 details.last_index = ULONG_MAX;
2857 i_mmap_lock_write(mapping);
2858 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2859 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2860 i_mmap_unlock_write(mapping);
2864 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2865 * address_space corresponding to the specified byte range in the underlying
2868 * @mapping: the address space containing mmaps to be unmapped.
2869 * @holebegin: byte in first page to unmap, relative to the start of
2870 * the underlying file. This will be rounded down to a PAGE_SIZE
2871 * boundary. Note that this is different from truncate_pagecache(), which
2872 * must keep the partial page. In contrast, we must get rid of
2874 * @holelen: size of prospective hole in bytes. This will be rounded
2875 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2877 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2878 * but 0 when invalidating pagecache, don't throw away private data.
2880 void unmap_mapping_range(struct address_space *mapping,
2881 loff_t const holebegin, loff_t const holelen, int even_cows)
2883 pgoff_t hba = holebegin >> PAGE_SHIFT;
2884 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2886 /* Check for overflow. */
2887 if (sizeof(holelen) > sizeof(hlen)) {
2889 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2890 if (holeend & ~(long long)ULONG_MAX)
2891 hlen = ULONG_MAX - hba + 1;
2894 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2896 EXPORT_SYMBOL(unmap_mapping_range);
2899 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2900 * but allow concurrent faults), and pte mapped but not yet locked.
2901 * We return with pte unmapped and unlocked.
2903 * We return with the mmap_sem locked or unlocked in the same cases
2904 * as does filemap_fault().
2906 vm_fault_t do_swap_page(struct vm_fault *vmf)
2908 struct vm_area_struct *vma = vmf->vma;
2909 struct page *page = NULL, *swapcache;
2910 struct mem_cgroup *memcg;
2917 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2920 entry = pte_to_swp_entry(vmf->orig_pte);
2921 if (unlikely(non_swap_entry(entry))) {
2922 if (is_migration_entry(entry)) {
2923 migration_entry_wait(vma->vm_mm, vmf->pmd,
2925 } else if (is_device_private_entry(entry)) {
2927 * For un-addressable device memory we call the pgmap
2928 * fault handler callback. The callback must migrate
2929 * the page back to some CPU accessible page.
2931 ret = device_private_entry_fault(vma, vmf->address, entry,
2932 vmf->flags, vmf->pmd);
2933 } else if (is_hwpoison_entry(entry)) {
2934 ret = VM_FAULT_HWPOISON;
2936 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2937 ret = VM_FAULT_SIGBUS;
2943 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2944 page = lookup_swap_cache(entry, vma, vmf->address);
2948 struct swap_info_struct *si = swp_swap_info(entry);
2950 if (si->flags & SWP_SYNCHRONOUS_IO &&
2951 __swap_count(si, entry) == 1) {
2952 /* skip swapcache */
2953 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2956 __SetPageLocked(page);
2957 __SetPageSwapBacked(page);
2958 set_page_private(page, entry.val);
2959 lru_cache_add_anon(page);
2960 swap_readpage(page, true);
2963 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2970 * Back out if somebody else faulted in this pte
2971 * while we released the pte lock.
2973 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2974 vmf->address, &vmf->ptl);
2975 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2977 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2981 /* Had to read the page from swap area: Major fault */
2982 ret = VM_FAULT_MAJOR;
2983 count_vm_event(PGMAJFAULT);
2984 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2985 } else if (PageHWPoison(page)) {
2987 * hwpoisoned dirty swapcache pages are kept for killing
2988 * owner processes (which may be unknown at hwpoison time)
2990 ret = VM_FAULT_HWPOISON;
2991 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2995 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2997 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2999 ret |= VM_FAULT_RETRY;
3004 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3005 * release the swapcache from under us. The page pin, and pte_same
3006 * test below, are not enough to exclude that. Even if it is still
3007 * swapcache, we need to check that the page's swap has not changed.
3009 if (unlikely((!PageSwapCache(page) ||
3010 page_private(page) != entry.val)) && swapcache)
3013 page = ksm_might_need_to_copy(page, vma, vmf->address);
3014 if (unlikely(!page)) {
3020 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3027 * Back out if somebody else already faulted in this pte.
3029 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3031 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3034 if (unlikely(!PageUptodate(page))) {
3035 ret = VM_FAULT_SIGBUS;
3040 * The page isn't present yet, go ahead with the fault.
3042 * Be careful about the sequence of operations here.
3043 * To get its accounting right, reuse_swap_page() must be called
3044 * while the page is counted on swap but not yet in mapcount i.e.
3045 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3046 * must be called after the swap_free(), or it will never succeed.
3049 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3050 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3051 pte = mk_pte(page, vma->vm_page_prot);
3052 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3053 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3054 vmf->flags &= ~FAULT_FLAG_WRITE;
3055 ret |= VM_FAULT_WRITE;
3056 exclusive = RMAP_EXCLUSIVE;
3058 flush_icache_page(vma, page);
3059 if (pte_swp_soft_dirty(vmf->orig_pte))
3060 pte = pte_mksoft_dirty(pte);
3061 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3062 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3063 vmf->orig_pte = pte;
3065 /* ksm created a completely new copy */
3066 if (unlikely(page != swapcache && swapcache)) {
3067 page_add_new_anon_rmap(page, vma, vmf->address, false);
3068 mem_cgroup_commit_charge(page, memcg, false, false);
3069 lru_cache_add_active_or_unevictable(page, vma);
3071 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3072 mem_cgroup_commit_charge(page, memcg, true, false);
3073 activate_page(page);
3077 if (mem_cgroup_swap_full(page) ||
3078 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3079 try_to_free_swap(page);
3081 if (page != swapcache && swapcache) {
3083 * Hold the lock to avoid the swap entry to be reused
3084 * until we take the PT lock for the pte_same() check
3085 * (to avoid false positives from pte_same). For
3086 * further safety release the lock after the swap_free
3087 * so that the swap count won't change under a
3088 * parallel locked swapcache.
3090 unlock_page(swapcache);
3091 put_page(swapcache);
3094 if (vmf->flags & FAULT_FLAG_WRITE) {
3095 ret |= do_wp_page(vmf);
3096 if (ret & VM_FAULT_ERROR)
3097 ret &= VM_FAULT_ERROR;
3101 /* No need to invalidate - it was non-present before */
3102 update_mmu_cache(vma, vmf->address, vmf->pte);
3104 pte_unmap_unlock(vmf->pte, vmf->ptl);
3108 mem_cgroup_cancel_charge(page, memcg, false);
3109 pte_unmap_unlock(vmf->pte, vmf->ptl);
3114 if (page != swapcache && swapcache) {
3115 unlock_page(swapcache);
3116 put_page(swapcache);
3122 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3123 * but allow concurrent faults), and pte mapped but not yet locked.
3124 * We return with mmap_sem still held, but pte unmapped and unlocked.
3126 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3128 struct vm_area_struct *vma = vmf->vma;
3129 struct mem_cgroup *memcg;
3134 /* File mapping without ->vm_ops ? */
3135 if (vma->vm_flags & VM_SHARED)
3136 return VM_FAULT_SIGBUS;
3139 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3140 * pte_offset_map() on pmds where a huge pmd might be created
3141 * from a different thread.
3143 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3144 * parallel threads are excluded by other means.
3146 * Here we only have down_read(mmap_sem).
3148 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3149 return VM_FAULT_OOM;
3151 /* See the comment in pte_alloc_one_map() */
3152 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3155 /* Use the zero-page for reads */
3156 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3157 !mm_forbids_zeropage(vma->vm_mm)) {
3158 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3159 vma->vm_page_prot));
3160 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3161 vmf->address, &vmf->ptl);
3162 if (!pte_none(*vmf->pte))
3164 ret = check_stable_address_space(vma->vm_mm);
3167 /* Deliver the page fault to userland, check inside PT lock */
3168 if (userfaultfd_missing(vma)) {
3169 pte_unmap_unlock(vmf->pte, vmf->ptl);
3170 return handle_userfault(vmf, VM_UFFD_MISSING);
3175 /* Allocate our own private page. */
3176 if (unlikely(anon_vma_prepare(vma)))
3178 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3182 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3187 * The memory barrier inside __SetPageUptodate makes sure that
3188 * preceeding stores to the page contents become visible before
3189 * the set_pte_at() write.
3191 __SetPageUptodate(page);
3193 entry = mk_pte(page, vma->vm_page_prot);
3194 if (vma->vm_flags & VM_WRITE)
3195 entry = pte_mkwrite(pte_mkdirty(entry));
3197 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3199 if (!pte_none(*vmf->pte))
3202 ret = check_stable_address_space(vma->vm_mm);
3206 /* Deliver the page fault to userland, check inside PT lock */
3207 if (userfaultfd_missing(vma)) {
3208 pte_unmap_unlock(vmf->pte, vmf->ptl);
3209 mem_cgroup_cancel_charge(page, memcg, false);
3211 return handle_userfault(vmf, VM_UFFD_MISSING);
3214 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3215 page_add_new_anon_rmap(page, vma, vmf->address, false);
3216 mem_cgroup_commit_charge(page, memcg, false, false);
3217 lru_cache_add_active_or_unevictable(page, vma);
3219 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3221 /* No need to invalidate - it was non-present before */
3222 update_mmu_cache(vma, vmf->address, vmf->pte);
3224 pte_unmap_unlock(vmf->pte, vmf->ptl);
3227 mem_cgroup_cancel_charge(page, memcg, false);
3233 return VM_FAULT_OOM;
3237 * The mmap_sem must have been held on entry, and may have been
3238 * released depending on flags and vma->vm_ops->fault() return value.
3239 * See filemap_fault() and __lock_page_retry().
3241 static vm_fault_t __do_fault(struct vm_fault *vmf)
3243 struct vm_area_struct *vma = vmf->vma;
3247 * Preallocate pte before we take page_lock because this might lead to
3248 * deadlocks for memcg reclaim which waits for pages under writeback:
3250 * SetPageWriteback(A)
3256 * wait_on_page_writeback(A)
3257 * SetPageWriteback(B)
3259 * # flush A, B to clear the writeback
3261 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3262 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3264 if (!vmf->prealloc_pte)
3265 return VM_FAULT_OOM;
3266 smp_wmb(); /* See comment in __pte_alloc() */
3269 ret = vma->vm_ops->fault(vmf);
3270 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3271 VM_FAULT_DONE_COW)))
3274 if (unlikely(PageHWPoison(vmf->page))) {
3275 if (ret & VM_FAULT_LOCKED)
3276 unlock_page(vmf->page);
3277 put_page(vmf->page);
3279 return VM_FAULT_HWPOISON;
3282 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3283 lock_page(vmf->page);
3285 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3291 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3292 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3293 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3294 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3296 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3298 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3301 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3303 struct vm_area_struct *vma = vmf->vma;
3305 if (!pmd_none(*vmf->pmd))
3307 if (vmf->prealloc_pte) {
3308 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3309 if (unlikely(!pmd_none(*vmf->pmd))) {
3310 spin_unlock(vmf->ptl);
3314 mm_inc_nr_ptes(vma->vm_mm);
3315 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3316 spin_unlock(vmf->ptl);
3317 vmf->prealloc_pte = NULL;
3318 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3319 return VM_FAULT_OOM;
3323 * If a huge pmd materialized under us just retry later. Use
3324 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3325 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3326 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3327 * running immediately after a huge pmd fault in a different thread of
3328 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3329 * All we have to ensure is that it is a regular pmd that we can walk
3330 * with pte_offset_map() and we can do that through an atomic read in
3331 * C, which is what pmd_trans_unstable() provides.
3333 if (pmd_devmap_trans_unstable(vmf->pmd))
3334 return VM_FAULT_NOPAGE;
3337 * At this point we know that our vmf->pmd points to a page of ptes
3338 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3339 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3340 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3341 * be valid and we will re-check to make sure the vmf->pte isn't
3342 * pte_none() under vmf->ptl protection when we return to
3345 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3350 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3352 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3353 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3354 unsigned long haddr)
3356 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3357 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3359 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3364 static void deposit_prealloc_pte(struct vm_fault *vmf)
3366 struct vm_area_struct *vma = vmf->vma;
3368 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3370 * We are going to consume the prealloc table,
3371 * count that as nr_ptes.
3373 mm_inc_nr_ptes(vma->vm_mm);
3374 vmf->prealloc_pte = NULL;
3377 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3379 struct vm_area_struct *vma = vmf->vma;
3380 bool write = vmf->flags & FAULT_FLAG_WRITE;
3381 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3386 if (!transhuge_vma_suitable(vma, haddr))
3387 return VM_FAULT_FALLBACK;
3389 ret = VM_FAULT_FALLBACK;
3390 page = compound_head(page);
3393 * Archs like ppc64 need additonal space to store information
3394 * related to pte entry. Use the preallocated table for that.
3396 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3397 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3398 if (!vmf->prealloc_pte)
3399 return VM_FAULT_OOM;
3400 smp_wmb(); /* See comment in __pte_alloc() */
3403 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3404 if (unlikely(!pmd_none(*vmf->pmd)))
3407 for (i = 0; i < HPAGE_PMD_NR; i++)
3408 flush_icache_page(vma, page + i);
3410 entry = mk_huge_pmd(page, vma->vm_page_prot);
3412 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3414 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3415 page_add_file_rmap(page, true);
3417 * deposit and withdraw with pmd lock held
3419 if (arch_needs_pgtable_deposit())
3420 deposit_prealloc_pte(vmf);
3422 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3424 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3426 /* fault is handled */
3428 count_vm_event(THP_FILE_MAPPED);
3430 spin_unlock(vmf->ptl);
3434 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3442 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3443 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3445 * @vmf: fault environment
3446 * @memcg: memcg to charge page (only for private mappings)
3447 * @page: page to map
3449 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3452 * Target users are page handler itself and implementations of
3453 * vm_ops->map_pages.
3455 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3458 struct vm_area_struct *vma = vmf->vma;
3459 bool write = vmf->flags & FAULT_FLAG_WRITE;
3463 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3464 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3466 VM_BUG_ON_PAGE(memcg, page);
3468 ret = do_set_pmd(vmf, page);
3469 if (ret != VM_FAULT_FALLBACK)
3474 ret = pte_alloc_one_map(vmf);
3479 /* Re-check under ptl */
3480 if (unlikely(!pte_none(*vmf->pte)))
3481 return VM_FAULT_NOPAGE;
3483 flush_icache_page(vma, page);
3484 entry = mk_pte(page, vma->vm_page_prot);
3486 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3487 /* copy-on-write page */
3488 if (write && !(vma->vm_flags & VM_SHARED)) {
3489 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3490 page_add_new_anon_rmap(page, vma, vmf->address, false);
3491 mem_cgroup_commit_charge(page, memcg, false, false);
3492 lru_cache_add_active_or_unevictable(page, vma);
3494 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3495 page_add_file_rmap(page, false);
3497 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3499 /* no need to invalidate: a not-present page won't be cached */
3500 update_mmu_cache(vma, vmf->address, vmf->pte);
3507 * finish_fault - finish page fault once we have prepared the page to fault
3509 * @vmf: structure describing the fault
3511 * This function handles all that is needed to finish a page fault once the
3512 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3513 * given page, adds reverse page mapping, handles memcg charges and LRU
3514 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3517 * The function expects the page to be locked and on success it consumes a
3518 * reference of a page being mapped (for the PTE which maps it).
3520 vm_fault_t finish_fault(struct vm_fault *vmf)
3525 /* Did we COW the page? */
3526 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3527 !(vmf->vma->vm_flags & VM_SHARED))
3528 page = vmf->cow_page;
3533 * check even for read faults because we might have lost our CoWed
3536 if (!(vmf->vma->vm_flags & VM_SHARED))
3537 ret = check_stable_address_space(vmf->vma->vm_mm);
3539 ret = alloc_set_pte(vmf, vmf->memcg, page);
3541 pte_unmap_unlock(vmf->pte, vmf->ptl);
3545 static unsigned long fault_around_bytes __read_mostly =
3546 rounddown_pow_of_two(65536);
3548 #ifdef CONFIG_DEBUG_FS
3549 static int fault_around_bytes_get(void *data, u64 *val)
3551 *val = fault_around_bytes;
3556 * fault_around_bytes must be rounded down to the nearest page order as it's
3557 * what do_fault_around() expects to see.
3559 static int fault_around_bytes_set(void *data, u64 val)
3561 if (val / PAGE_SIZE > PTRS_PER_PTE)
3563 if (val > PAGE_SIZE)
3564 fault_around_bytes = rounddown_pow_of_two(val);
3566 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3569 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3570 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3572 static int __init fault_around_debugfs(void)
3576 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3577 &fault_around_bytes_fops);
3579 pr_warn("Failed to create fault_around_bytes in debugfs");
3582 late_initcall(fault_around_debugfs);
3586 * do_fault_around() tries to map few pages around the fault address. The hope
3587 * is that the pages will be needed soon and this will lower the number of
3590 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3591 * not ready to be mapped: not up-to-date, locked, etc.
3593 * This function is called with the page table lock taken. In the split ptlock
3594 * case the page table lock only protects only those entries which belong to
3595 * the page table corresponding to the fault address.
3597 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3600 * fault_around_bytes defines how many bytes we'll try to map.
3601 * do_fault_around() expects it to be set to a power of two less than or equal
3604 * The virtual address of the area that we map is naturally aligned to
3605 * fault_around_bytes rounded down to the machine page size
3606 * (and therefore to page order). This way it's easier to guarantee
3607 * that we don't cross page table boundaries.
3609 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3611 unsigned long address = vmf->address, nr_pages, mask;
3612 pgoff_t start_pgoff = vmf->pgoff;
3617 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3618 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3620 vmf->address = max(address & mask, vmf->vma->vm_start);
3621 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3625 * end_pgoff is either the end of the page table, the end of
3626 * the vma or nr_pages from start_pgoff, depending what is nearest.
3628 end_pgoff = start_pgoff -
3629 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3631 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3632 start_pgoff + nr_pages - 1);
3634 if (pmd_none(*vmf->pmd)) {
3635 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3637 if (!vmf->prealloc_pte)
3639 smp_wmb(); /* See comment in __pte_alloc() */
3642 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3644 /* Huge page is mapped? Page fault is solved */
3645 if (pmd_trans_huge(*vmf->pmd)) {
3646 ret = VM_FAULT_NOPAGE;
3650 /* ->map_pages() haven't done anything useful. Cold page cache? */
3654 /* check if the page fault is solved */
3655 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3656 if (!pte_none(*vmf->pte))
3657 ret = VM_FAULT_NOPAGE;
3658 pte_unmap_unlock(vmf->pte, vmf->ptl);
3660 vmf->address = address;
3665 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3667 struct vm_area_struct *vma = vmf->vma;
3671 * Let's call ->map_pages() first and use ->fault() as fallback
3672 * if page by the offset is not ready to be mapped (cold cache or
3675 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3676 ret = do_fault_around(vmf);
3681 ret = __do_fault(vmf);
3682 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3685 ret |= finish_fault(vmf);
3686 unlock_page(vmf->page);
3687 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3688 put_page(vmf->page);
3692 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3694 struct vm_area_struct *vma = vmf->vma;
3697 if (unlikely(anon_vma_prepare(vma)))
3698 return VM_FAULT_OOM;
3700 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3702 return VM_FAULT_OOM;
3704 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3705 &vmf->memcg, false)) {
3706 put_page(vmf->cow_page);
3707 return VM_FAULT_OOM;
3710 ret = __do_fault(vmf);
3711 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3713 if (ret & VM_FAULT_DONE_COW)
3716 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3717 __SetPageUptodate(vmf->cow_page);
3719 ret |= finish_fault(vmf);
3720 unlock_page(vmf->page);
3721 put_page(vmf->page);
3722 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3726 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3727 put_page(vmf->cow_page);
3731 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3733 struct vm_area_struct *vma = vmf->vma;
3734 vm_fault_t ret, tmp;
3736 ret = __do_fault(vmf);
3737 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3741 * Check if the backing address space wants to know that the page is
3742 * about to become writable
3744 if (vma->vm_ops->page_mkwrite) {
3745 unlock_page(vmf->page);
3746 tmp = do_page_mkwrite(vmf);
3747 if (unlikely(!tmp ||
3748 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3749 put_page(vmf->page);
3754 ret |= finish_fault(vmf);
3755 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3757 unlock_page(vmf->page);
3758 put_page(vmf->page);
3762 fault_dirty_shared_page(vma, vmf->page);
3767 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3768 * but allow concurrent faults).
3769 * The mmap_sem may have been released depending on flags and our
3770 * return value. See filemap_fault() and __lock_page_or_retry().
3771 * If mmap_sem is released, vma may become invalid (for example
3772 * by other thread calling munmap()).
3774 static vm_fault_t do_fault(struct vm_fault *vmf)
3776 struct vm_area_struct *vma = vmf->vma;
3777 struct mm_struct *vm_mm = vma->vm_mm;
3781 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3783 if (!vma->vm_ops->fault) {
3785 * If we find a migration pmd entry or a none pmd entry, which
3786 * should never happen, return SIGBUS
3788 if (unlikely(!pmd_present(*vmf->pmd)))
3789 ret = VM_FAULT_SIGBUS;
3791 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3796 * Make sure this is not a temporary clearing of pte
3797 * by holding ptl and checking again. A R/M/W update
3798 * of pte involves: take ptl, clearing the pte so that
3799 * we don't have concurrent modification by hardware
3800 * followed by an update.
3802 if (unlikely(pte_none(*vmf->pte)))
3803 ret = VM_FAULT_SIGBUS;
3805 ret = VM_FAULT_NOPAGE;
3807 pte_unmap_unlock(vmf->pte, vmf->ptl);
3809 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3810 ret = do_read_fault(vmf);
3811 else if (!(vma->vm_flags & VM_SHARED))
3812 ret = do_cow_fault(vmf);
3814 ret = do_shared_fault(vmf);
3816 /* preallocated pagetable is unused: free it */
3817 if (vmf->prealloc_pte) {
3818 pte_free(vm_mm, vmf->prealloc_pte);
3819 vmf->prealloc_pte = NULL;
3824 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3825 unsigned long addr, int page_nid,
3830 count_vm_numa_event(NUMA_HINT_FAULTS);
3831 if (page_nid == numa_node_id()) {
3832 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3833 *flags |= TNF_FAULT_LOCAL;
3836 return mpol_misplaced(page, vma, addr);
3839 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3841 struct vm_area_struct *vma = vmf->vma;
3842 struct page *page = NULL;
3846 bool migrated = false;
3848 bool was_writable = pte_savedwrite(vmf->orig_pte);
3852 * The "pte" at this point cannot be used safely without
3853 * validation through pte_unmap_same(). It's of NUMA type but
3854 * the pfn may be screwed if the read is non atomic.
3856 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3857 spin_lock(vmf->ptl);
3858 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3859 pte_unmap_unlock(vmf->pte, vmf->ptl);
3864 * Make it present again, Depending on how arch implementes non
3865 * accessible ptes, some can allow access by kernel mode.
3867 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3868 pte = pte_modify(pte, vma->vm_page_prot);
3869 pte = pte_mkyoung(pte);
3871 pte = pte_mkwrite(pte);
3872 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3873 update_mmu_cache(vma, vmf->address, vmf->pte);
3875 page = vm_normal_page(vma, vmf->address, pte);
3877 pte_unmap_unlock(vmf->pte, vmf->ptl);
3881 /* TODO: handle PTE-mapped THP */
3882 if (PageCompound(page)) {
3883 pte_unmap_unlock(vmf->pte, vmf->ptl);
3888 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3889 * much anyway since they can be in shared cache state. This misses
3890 * the case where a mapping is writable but the process never writes
3891 * to it but pte_write gets cleared during protection updates and
3892 * pte_dirty has unpredictable behaviour between PTE scan updates,
3893 * background writeback, dirty balancing and application behaviour.
3895 if (!pte_write(pte))
3896 flags |= TNF_NO_GROUP;
3899 * Flag if the page is shared between multiple address spaces. This
3900 * is later used when determining whether to group tasks together
3902 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3903 flags |= TNF_SHARED;
3905 last_cpupid = page_cpupid_last(page);
3906 page_nid = page_to_nid(page);
3907 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3909 pte_unmap_unlock(vmf->pte, vmf->ptl);
3910 if (target_nid == -1) {
3915 /* Migrate to the requested node */
3916 migrated = migrate_misplaced_page(page, vma, target_nid);
3918 page_nid = target_nid;
3919 flags |= TNF_MIGRATED;
3921 flags |= TNF_MIGRATE_FAIL;
3925 task_numa_fault(last_cpupid, page_nid, 1, flags);
3929 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3931 if (vma_is_anonymous(vmf->vma))
3932 return do_huge_pmd_anonymous_page(vmf);
3933 if (vmf->vma->vm_ops->huge_fault)
3934 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3935 return VM_FAULT_FALLBACK;
3938 /* `inline' is required to avoid gcc 4.1.2 build error */
3939 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3941 if (vma_is_anonymous(vmf->vma))
3942 return do_huge_pmd_wp_page(vmf, orig_pmd);
3943 if (vmf->vma->vm_ops->huge_fault)
3944 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3946 /* COW handled on pte level: split pmd */
3947 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3948 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3950 return VM_FAULT_FALLBACK;
3953 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3955 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3958 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3960 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3961 /* No support for anonymous transparent PUD pages yet */
3962 if (vma_is_anonymous(vmf->vma))
3963 return VM_FAULT_FALLBACK;
3964 if (vmf->vma->vm_ops->huge_fault)
3965 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3966 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3967 return VM_FAULT_FALLBACK;
3970 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3972 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3973 /* No support for anonymous transparent PUD pages yet */
3974 if (vma_is_anonymous(vmf->vma))
3975 return VM_FAULT_FALLBACK;
3976 if (vmf->vma->vm_ops->huge_fault)
3977 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3978 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3979 return VM_FAULT_FALLBACK;
3983 * These routines also need to handle stuff like marking pages dirty
3984 * and/or accessed for architectures that don't do it in hardware (most
3985 * RISC architectures). The early dirtying is also good on the i386.
3987 * There is also a hook called "update_mmu_cache()" that architectures
3988 * with external mmu caches can use to update those (ie the Sparc or
3989 * PowerPC hashed page tables that act as extended TLBs).
3991 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3992 * concurrent faults).
3994 * The mmap_sem may have been released depending on flags and our return value.
3995 * See filemap_fault() and __lock_page_or_retry().
3997 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4001 if (unlikely(pmd_none(*vmf->pmd))) {
4003 * Leave __pte_alloc() until later: because vm_ops->fault may
4004 * want to allocate huge page, and if we expose page table
4005 * for an instant, it will be difficult to retract from
4006 * concurrent faults and from rmap lookups.
4010 /* See comment in pte_alloc_one_map() */
4011 if (pmd_devmap_trans_unstable(vmf->pmd))
4014 * A regular pmd is established and it can't morph into a huge
4015 * pmd from under us anymore at this point because we hold the
4016 * mmap_sem read mode and khugepaged takes it in write mode.
4017 * So now it's safe to run pte_offset_map().
4019 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4020 vmf->orig_pte = *vmf->pte;
4023 * some architectures can have larger ptes than wordsize,
4024 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4025 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4026 * accesses. The code below just needs a consistent view
4027 * for the ifs and we later double check anyway with the
4028 * ptl lock held. So here a barrier will do.
4031 if (pte_none(vmf->orig_pte)) {
4032 pte_unmap(vmf->pte);
4038 if (vma_is_anonymous(vmf->vma))
4039 return do_anonymous_page(vmf);
4041 return do_fault(vmf);
4044 if (!pte_present(vmf->orig_pte))
4045 return do_swap_page(vmf);
4047 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4048 return do_numa_page(vmf);
4050 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4051 spin_lock(vmf->ptl);
4052 entry = vmf->orig_pte;
4053 if (unlikely(!pte_same(*vmf->pte, entry)))
4055 if (vmf->flags & FAULT_FLAG_WRITE) {
4056 if (!pte_write(entry))
4057 return do_wp_page(vmf);
4058 entry = pte_mkdirty(entry);
4060 entry = pte_mkyoung(entry);
4061 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4062 vmf->flags & FAULT_FLAG_WRITE)) {
4063 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4066 * This is needed only for protection faults but the arch code
4067 * is not yet telling us if this is a protection fault or not.
4068 * This still avoids useless tlb flushes for .text page faults
4071 if (vmf->flags & FAULT_FLAG_WRITE)
4072 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4075 pte_unmap_unlock(vmf->pte, vmf->ptl);
4080 * By the time we get here, we already hold the mm semaphore
4082 * The mmap_sem may have been released depending on flags and our
4083 * return value. See filemap_fault() and __lock_page_or_retry().
4085 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4086 unsigned long address, unsigned int flags)
4088 struct vm_fault vmf = {
4090 .address = address & PAGE_MASK,
4092 .pgoff = linear_page_index(vma, address),
4093 .gfp_mask = __get_fault_gfp_mask(vma),
4095 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4096 struct mm_struct *mm = vma->vm_mm;
4101 pgd = pgd_offset(mm, address);
4102 p4d = p4d_alloc(mm, pgd, address);
4104 return VM_FAULT_OOM;
4106 vmf.pud = pud_alloc(mm, p4d, address);
4108 return VM_FAULT_OOM;
4109 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4110 ret = create_huge_pud(&vmf);
4111 if (!(ret & VM_FAULT_FALLBACK))
4114 pud_t orig_pud = *vmf.pud;
4117 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4119 /* NUMA case for anonymous PUDs would go here */
4121 if (dirty && !pud_write(orig_pud)) {
4122 ret = wp_huge_pud(&vmf, orig_pud);
4123 if (!(ret & VM_FAULT_FALLBACK))
4126 huge_pud_set_accessed(&vmf, orig_pud);
4132 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4134 return VM_FAULT_OOM;
4135 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4136 ret = create_huge_pmd(&vmf);
4137 if (!(ret & VM_FAULT_FALLBACK))
4140 pmd_t orig_pmd = *vmf.pmd;
4143 if (unlikely(is_swap_pmd(orig_pmd))) {
4144 VM_BUG_ON(thp_migration_supported() &&
4145 !is_pmd_migration_entry(orig_pmd));
4146 if (is_pmd_migration_entry(orig_pmd))
4147 pmd_migration_entry_wait(mm, vmf.pmd);
4150 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4151 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4152 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4154 if (dirty && !pmd_write(orig_pmd)) {
4155 ret = wp_huge_pmd(&vmf, orig_pmd);
4156 if (!(ret & VM_FAULT_FALLBACK))
4159 huge_pmd_set_accessed(&vmf, orig_pmd);
4165 return handle_pte_fault(&vmf);
4169 * By the time we get here, we already hold the mm semaphore
4171 * The mmap_sem may have been released depending on flags and our
4172 * return value. See filemap_fault() and __lock_page_or_retry().
4174 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4179 __set_current_state(TASK_RUNNING);
4181 count_vm_event(PGFAULT);
4182 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4184 /* do counter updates before entering really critical section. */
4185 check_sync_rss_stat(current);
4187 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4188 flags & FAULT_FLAG_INSTRUCTION,
4189 flags & FAULT_FLAG_REMOTE))
4190 return VM_FAULT_SIGSEGV;
4193 * Enable the memcg OOM handling for faults triggered in user
4194 * space. Kernel faults are handled more gracefully.
4196 if (flags & FAULT_FLAG_USER)
4197 mem_cgroup_enter_user_fault();
4199 if (unlikely(is_vm_hugetlb_page(vma)))
4200 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4202 ret = __handle_mm_fault(vma, address, flags);
4204 if (flags & FAULT_FLAG_USER) {
4205 mem_cgroup_exit_user_fault();
4207 * The task may have entered a memcg OOM situation but
4208 * if the allocation error was handled gracefully (no
4209 * VM_FAULT_OOM), there is no need to kill anything.
4210 * Just clean up the OOM state peacefully.
4212 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4213 mem_cgroup_oom_synchronize(false);
4218 EXPORT_SYMBOL_GPL(handle_mm_fault);
4220 #ifndef __PAGETABLE_P4D_FOLDED
4222 * Allocate p4d page table.
4223 * We've already handled the fast-path in-line.
4225 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4227 p4d_t *new = p4d_alloc_one(mm, address);
4231 smp_wmb(); /* See comment in __pte_alloc */
4233 spin_lock(&mm->page_table_lock);
4234 if (pgd_present(*pgd)) /* Another has populated it */
4237 pgd_populate(mm, pgd, new);
4238 spin_unlock(&mm->page_table_lock);
4241 #endif /* __PAGETABLE_P4D_FOLDED */
4243 #ifndef __PAGETABLE_PUD_FOLDED
4245 * Allocate page upper directory.
4246 * We've already handled the fast-path in-line.
4248 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4250 pud_t *new = pud_alloc_one(mm, address);
4254 smp_wmb(); /* See comment in __pte_alloc */
4256 spin_lock(&mm->page_table_lock);
4257 #ifndef __ARCH_HAS_5LEVEL_HACK
4258 if (!p4d_present(*p4d)) {
4260 p4d_populate(mm, p4d, new);
4261 } else /* Another has populated it */
4264 if (!pgd_present(*p4d)) {
4266 pgd_populate(mm, p4d, new);
4267 } else /* Another has populated it */
4269 #endif /* __ARCH_HAS_5LEVEL_HACK */
4270 spin_unlock(&mm->page_table_lock);
4273 #endif /* __PAGETABLE_PUD_FOLDED */
4275 #ifndef __PAGETABLE_PMD_FOLDED
4277 * Allocate page middle directory.
4278 * We've already handled the fast-path in-line.
4280 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4283 pmd_t *new = pmd_alloc_one(mm, address);
4287 smp_wmb(); /* See comment in __pte_alloc */
4289 ptl = pud_lock(mm, pud);
4290 #ifndef __ARCH_HAS_4LEVEL_HACK
4291 if (!pud_present(*pud)) {
4293 pud_populate(mm, pud, new);
4294 } else /* Another has populated it */
4297 if (!pgd_present(*pud)) {
4299 pgd_populate(mm, pud, new);
4300 } else /* Another has populated it */
4302 #endif /* __ARCH_HAS_4LEVEL_HACK */
4306 #endif /* __PAGETABLE_PMD_FOLDED */
4308 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4309 unsigned long *start, unsigned long *end,
4310 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4318 pgd = pgd_offset(mm, address);
4319 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4322 p4d = p4d_offset(pgd, address);
4323 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4326 pud = pud_offset(p4d, address);
4327 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4330 pmd = pmd_offset(pud, address);
4331 VM_BUG_ON(pmd_trans_huge(*pmd));
4333 if (pmd_huge(*pmd)) {
4338 *start = address & PMD_MASK;
4339 *end = *start + PMD_SIZE;
4340 mmu_notifier_invalidate_range_start(mm, *start, *end);
4342 *ptlp = pmd_lock(mm, pmd);
4343 if (pmd_huge(*pmd)) {
4349 mmu_notifier_invalidate_range_end(mm, *start, *end);
4352 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4356 *start = address & PAGE_MASK;
4357 *end = *start + PAGE_SIZE;
4358 mmu_notifier_invalidate_range_start(mm, *start, *end);
4360 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4361 if (!pte_present(*ptep))
4366 pte_unmap_unlock(ptep, *ptlp);
4368 mmu_notifier_invalidate_range_end(mm, *start, *end);
4373 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4374 pte_t **ptepp, spinlock_t **ptlp)
4378 /* (void) is needed to make gcc happy */
4379 (void) __cond_lock(*ptlp,
4380 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4381 ptepp, NULL, ptlp)));
4385 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4386 unsigned long *start, unsigned long *end,
4387 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4391 /* (void) is needed to make gcc happy */
4392 (void) __cond_lock(*ptlp,
4393 !(res = __follow_pte_pmd(mm, address, start, end,
4394 ptepp, pmdpp, ptlp)));
4397 EXPORT_SYMBOL(follow_pte_pmd);
4400 * follow_pfn - look up PFN at a user virtual address
4401 * @vma: memory mapping
4402 * @address: user virtual address
4403 * @pfn: location to store found PFN
4405 * Only IO mappings and raw PFN mappings are allowed.
4407 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4409 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4416 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4419 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4422 *pfn = pte_pfn(*ptep);
4423 pte_unmap_unlock(ptep, ptl);
4426 EXPORT_SYMBOL(follow_pfn);
4428 #ifdef CONFIG_HAVE_IOREMAP_PROT
4429 int follow_phys(struct vm_area_struct *vma,
4430 unsigned long address, unsigned int flags,
4431 unsigned long *prot, resource_size_t *phys)
4437 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4440 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4444 if ((flags & FOLL_WRITE) && !pte_write(pte))
4447 *prot = pgprot_val(pte_pgprot(pte));
4448 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4452 pte_unmap_unlock(ptep, ptl);
4457 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4458 void *buf, int len, int write)
4460 resource_size_t phys_addr;
4461 unsigned long prot = 0;
4462 void __iomem *maddr;
4463 int offset = addr & (PAGE_SIZE-1);
4465 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4468 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4473 memcpy_toio(maddr + offset, buf, len);
4475 memcpy_fromio(buf, maddr + offset, len);
4480 EXPORT_SYMBOL_GPL(generic_access_phys);
4484 * Access another process' address space as given in mm. If non-NULL, use the
4485 * given task for page fault accounting.
4487 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4488 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4490 struct vm_area_struct *vma;
4491 void *old_buf = buf;
4492 int write = gup_flags & FOLL_WRITE;
4494 if (down_read_killable(&mm->mmap_sem))
4497 /* ignore errors, just check how much was successfully transferred */
4499 int bytes, ret, offset;
4501 struct page *page = NULL;
4503 ret = get_user_pages_remote(tsk, mm, addr, 1,
4504 gup_flags, &page, &vma, NULL);
4506 #ifndef CONFIG_HAVE_IOREMAP_PROT
4510 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4511 * we can access using slightly different code.
4513 vma = find_vma(mm, addr);
4514 if (!vma || vma->vm_start > addr)
4516 if (vma->vm_ops && vma->vm_ops->access)
4517 ret = vma->vm_ops->access(vma, addr, buf,
4525 offset = addr & (PAGE_SIZE-1);
4526 if (bytes > PAGE_SIZE-offset)
4527 bytes = PAGE_SIZE-offset;
4531 copy_to_user_page(vma, page, addr,
4532 maddr + offset, buf, bytes);
4533 set_page_dirty_lock(page);
4535 copy_from_user_page(vma, page, addr,
4536 buf, maddr + offset, bytes);
4545 up_read(&mm->mmap_sem);
4547 return buf - old_buf;
4551 * access_remote_vm - access another process' address space
4552 * @mm: the mm_struct of the target address space
4553 * @addr: start address to access
4554 * @buf: source or destination buffer
4555 * @len: number of bytes to transfer
4556 * @gup_flags: flags modifying lookup behaviour
4558 * The caller must hold a reference on @mm.
4560 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4561 void *buf, int len, unsigned int gup_flags)
4563 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4567 * Access another process' address space.
4568 * Source/target buffer must be kernel space,
4569 * Do not walk the page table directly, use get_user_pages
4571 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4572 void *buf, int len, unsigned int gup_flags)
4574 struct mm_struct *mm;
4577 mm = get_task_mm(tsk);
4581 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4587 EXPORT_SYMBOL_GPL(access_process_vm);
4590 * Print the name of a VMA.
4592 void print_vma_addr(char *prefix, unsigned long ip)
4594 struct mm_struct *mm = current->mm;
4595 struct vm_area_struct *vma;
4598 * we might be running from an atomic context so we cannot sleep
4600 if (!down_read_trylock(&mm->mmap_sem))
4603 vma = find_vma(mm, ip);
4604 if (vma && vma->vm_file) {
4605 struct file *f = vma->vm_file;
4606 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4610 p = file_path(f, buf, PAGE_SIZE);
4613 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4615 vma->vm_end - vma->vm_start);
4616 free_page((unsigned long)buf);
4619 up_read(&mm->mmap_sem);
4622 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4623 void __might_fault(const char *file, int line)
4626 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4627 * holding the mmap_sem, this is safe because kernel memory doesn't
4628 * get paged out, therefore we'll never actually fault, and the
4629 * below annotations will generate false positives.
4631 if (uaccess_kernel())
4633 if (pagefault_disabled())
4635 __might_sleep(file, line, 0);
4636 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4638 might_lock_read(¤t->mm->mmap_sem);
4641 EXPORT_SYMBOL(__might_fault);
4644 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4646 * Process all subpages of the specified huge page with the specified
4647 * operation. The target subpage will be processed last to keep its
4650 static inline void process_huge_page(
4651 unsigned long addr_hint, unsigned int pages_per_huge_page,
4652 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4656 unsigned long addr = addr_hint &
4657 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4659 /* Process target subpage last to keep its cache lines hot */
4661 n = (addr_hint - addr) / PAGE_SIZE;
4662 if (2 * n <= pages_per_huge_page) {
4663 /* If target subpage in first half of huge page */
4666 /* Process subpages at the end of huge page */
4667 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4669 process_subpage(addr + i * PAGE_SIZE, i, arg);
4672 /* If target subpage in second half of huge page */
4673 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4674 l = pages_per_huge_page - n;
4675 /* Process subpages at the begin of huge page */
4676 for (i = 0; i < base; i++) {
4678 process_subpage(addr + i * PAGE_SIZE, i, arg);
4682 * Process remaining subpages in left-right-left-right pattern
4683 * towards the target subpage
4685 for (i = 0; i < l; i++) {
4686 int left_idx = base + i;
4687 int right_idx = base + 2 * l - 1 - i;
4690 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4692 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4696 static void clear_gigantic_page(struct page *page,
4698 unsigned int pages_per_huge_page)
4701 struct page *p = page;
4704 for (i = 0; i < pages_per_huge_page;
4705 i++, p = mem_map_next(p, page, i)) {
4707 clear_user_highpage(p, addr + i * PAGE_SIZE);
4711 static void clear_subpage(unsigned long addr, int idx, void *arg)
4713 struct page *page = arg;
4715 clear_user_highpage(page + idx, addr);
4718 void clear_huge_page(struct page *page,
4719 unsigned long addr_hint, unsigned int pages_per_huge_page)
4721 unsigned long addr = addr_hint &
4722 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4724 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4725 clear_gigantic_page(page, addr, pages_per_huge_page);
4729 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4732 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4734 struct vm_area_struct *vma,
4735 unsigned int pages_per_huge_page)
4738 struct page *dst_base = dst;
4739 struct page *src_base = src;
4741 for (i = 0; i < pages_per_huge_page; ) {
4743 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4746 dst = mem_map_next(dst, dst_base, i);
4747 src = mem_map_next(src, src_base, i);
4751 struct copy_subpage_arg {
4754 struct vm_area_struct *vma;
4757 static void copy_subpage(unsigned long addr, int idx, void *arg)
4759 struct copy_subpage_arg *copy_arg = arg;
4761 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4762 addr, copy_arg->vma);
4765 void copy_user_huge_page(struct page *dst, struct page *src,
4766 unsigned long addr_hint, struct vm_area_struct *vma,
4767 unsigned int pages_per_huge_page)
4769 unsigned long addr = addr_hint &
4770 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4771 struct copy_subpage_arg arg = {
4777 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4778 copy_user_gigantic_page(dst, src, addr, vma,
4779 pages_per_huge_page);
4783 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4786 long copy_huge_page_from_user(struct page *dst_page,
4787 const void __user *usr_src,
4788 unsigned int pages_per_huge_page,
4789 bool allow_pagefault)
4791 void *src = (void *)usr_src;
4793 unsigned long i, rc = 0;
4794 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4796 for (i = 0; i < pages_per_huge_page; i++) {
4797 if (allow_pagefault)
4798 page_kaddr = kmap(dst_page + i);
4800 page_kaddr = kmap_atomic(dst_page + i);
4801 rc = copy_from_user(page_kaddr,
4802 (const void __user *)(src + i * PAGE_SIZE),
4804 if (allow_pagefault)
4805 kunmap(dst_page + i);
4807 kunmap_atomic(page_kaddr);
4809 ret_val -= (PAGE_SIZE - rc);
4817 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4819 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4821 static struct kmem_cache *page_ptl_cachep;
4823 void __init ptlock_cache_init(void)
4825 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4829 bool ptlock_alloc(struct page *page)
4833 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4840 void ptlock_free(struct page *page)
4842 kmem_cache_free(page_ptl_cachep, page->ptl);