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/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
74 #include <trace/events/kmem.h>
77 #include <asm/mmu_context.h>
78 #include <asm/pgalloc.h>
79 #include <linux/uaccess.h>
81 #include <asm/tlbflush.h>
82 #include <asm/pgtable.h>
86 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
87 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
90 #ifndef CONFIG_NEED_MULTIPLE_NODES
91 /* use the per-pgdat data instead for discontigmem - mbligh */
92 unsigned long max_mapnr
;
93 EXPORT_SYMBOL(max_mapnr
);
96 EXPORT_SYMBOL(mem_map
);
100 * A number of key systems in x86 including ioremap() rely on the assumption
101 * that high_memory defines the upper bound on direct map memory, then end
102 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
103 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
107 EXPORT_SYMBOL(high_memory
);
110 * Randomize the address space (stacks, mmaps, brk, etc.).
112 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
113 * as ancient (libc5 based) binaries can segfault. )
115 int randomize_va_space __read_mostly
=
116 #ifdef CONFIG_COMPAT_BRK
122 static int __init
disable_randmaps(char *s
)
124 randomize_va_space
= 0;
127 __setup("norandmaps", disable_randmaps
);
129 unsigned long zero_pfn __read_mostly
;
130 EXPORT_SYMBOL(zero_pfn
);
132 unsigned long highest_memmap_pfn __read_mostly
;
135 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
137 static int __init
init_zero_pfn(void)
139 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
142 core_initcall(init_zero_pfn
);
145 * This threshold is the boundary in the value space, that the counter has to
146 * advance before we trace it. Should be a power of 2. It is to reduce unwanted
147 * trace overhead. The counter is number of pages.
149 #define TRACE_MM_COUNTER_THRESHOLD 128
151 void mm_trace_rss_stat(int member
, long count
, long value
)
153 long thresh_mask
= ~(TRACE_MM_COUNTER_THRESHOLD
- 1);
155 /* Threshold roll-over, trace it */
156 if ((count
& thresh_mask
) != ((count
- value
) & thresh_mask
))
157 trace_rss_stat(member
, count
);
160 #if defined(SPLIT_RSS_COUNTING)
162 void sync_mm_rss(struct mm_struct
*mm
)
166 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
167 if (current
->rss_stat
.count
[i
]) {
168 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
169 current
->rss_stat
.count
[i
] = 0;
172 current
->rss_stat
.events
= 0;
175 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
177 struct task_struct
*task
= current
;
179 if (likely(task
->mm
== mm
))
180 task
->rss_stat
.count
[member
] += val
;
182 add_mm_counter(mm
, member
, val
);
184 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
185 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
187 /* sync counter once per 64 page faults */
188 #define TASK_RSS_EVENTS_THRESH (64)
189 static void check_sync_rss_stat(struct task_struct
*task
)
191 if (unlikely(task
!= current
))
193 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
194 sync_mm_rss(task
->mm
);
196 #else /* SPLIT_RSS_COUNTING */
198 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
199 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
201 static void check_sync_rss_stat(struct task_struct
*task
)
205 #endif /* SPLIT_RSS_COUNTING */
207 #ifdef HAVE_GENERIC_MMU_GATHER
209 static bool tlb_next_batch(struct mmu_gather
*tlb
)
211 struct mmu_gather_batch
*batch
;
215 tlb
->active
= batch
->next
;
219 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
222 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
229 batch
->max
= MAX_GATHER_BATCH
;
231 tlb
->active
->next
= batch
;
237 void arch_tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
238 unsigned long start
, unsigned long end
)
242 /* Is it from 0 to ~0? */
243 tlb
->fullmm
= !(start
| (end
+1));
244 tlb
->need_flush_all
= 0;
245 tlb
->local
.next
= NULL
;
247 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
248 tlb
->active
= &tlb
->local
;
249 tlb
->batch_count
= 0;
251 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
256 __tlb_reset_range(tlb
);
259 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
265 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
266 __tlb_reset_range(tlb
);
269 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
271 struct mmu_gather_batch
*batch
;
273 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
274 tlb_table_flush(tlb
);
276 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
277 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
280 tlb
->active
= &tlb
->local
;
283 void tlb_flush_mmu(struct mmu_gather
*tlb
)
285 tlb_flush_mmu_tlbonly(tlb
);
286 tlb_flush_mmu_free(tlb
);
290 * Called at the end of the shootdown operation to free up any resources
291 * that were required.
293 void arch_tlb_finish_mmu(struct mmu_gather
*tlb
,
294 unsigned long start
, unsigned long end
, bool force
)
296 struct mmu_gather_batch
*batch
, *next
;
299 __tlb_adjust_range(tlb
, start
, end
- start
);
303 /* keep the page table cache within bounds */
306 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
308 free_pages((unsigned long)batch
, 0);
310 tlb
->local
.next
= NULL
;
314 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
315 * handling the additional races in SMP caused by other CPUs caching valid
316 * mappings in their TLBs. Returns the number of free page slots left.
317 * When out of page slots we must call tlb_flush_mmu().
318 *returns true if the caller should flush.
320 bool __tlb_remove_page_size(struct mmu_gather
*tlb
, struct page
*page
, int page_size
)
322 struct mmu_gather_batch
*batch
;
324 VM_BUG_ON(!tlb
->end
);
325 VM_WARN_ON(tlb
->page_size
!= page_size
);
329 * Add the page and check if we are full. If so
332 batch
->pages
[batch
->nr
++] = page
;
333 if (batch
->nr
== batch
->max
) {
334 if (!tlb_next_batch(tlb
))
338 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
343 #endif /* HAVE_GENERIC_MMU_GATHER */
345 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
348 * See the comment near struct mmu_table_batch.
352 * If we want tlb_remove_table() to imply TLB invalidates.
354 static inline void tlb_table_invalidate(struct mmu_gather
*tlb
)
356 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
358 * Invalidate page-table caches used by hardware walkers. Then we still
359 * need to RCU-sched wait while freeing the pages because software
360 * walkers can still be in-flight.
362 tlb_flush_mmu_tlbonly(tlb
);
366 static void tlb_remove_table_smp_sync(void *arg
)
368 /* Simply deliver the interrupt */
371 static void tlb_remove_table_one(void *table
)
374 * This isn't an RCU grace period and hence the page-tables cannot be
375 * assumed to be actually RCU-freed.
377 * It is however sufficient for software page-table walkers that rely on
378 * IRQ disabling. See the comment near struct mmu_table_batch.
380 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
381 __tlb_remove_table(table
);
384 static void tlb_remove_table_rcu(struct rcu_head
*head
)
386 struct mmu_table_batch
*batch
;
389 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
391 for (i
= 0; i
< batch
->nr
; i
++)
392 __tlb_remove_table(batch
->tables
[i
]);
394 free_page((unsigned long)batch
);
397 void tlb_table_flush(struct mmu_gather
*tlb
)
399 struct mmu_table_batch
**batch
= &tlb
->batch
;
402 tlb_table_invalidate(tlb
);
403 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
408 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
410 struct mmu_table_batch
**batch
= &tlb
->batch
;
412 if (*batch
== NULL
) {
413 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
414 if (*batch
== NULL
) {
415 tlb_table_invalidate(tlb
);
416 tlb_remove_table_one(table
);
422 (*batch
)->tables
[(*batch
)->nr
++] = table
;
423 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
424 tlb_table_flush(tlb
);
427 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
430 * Called to initialize an (on-stack) mmu_gather structure for page-table
431 * tear-down from @mm. The @fullmm argument is used when @mm is without
432 * users and we're going to destroy the full address space (exit/execve).
434 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
435 unsigned long start
, unsigned long end
)
437 arch_tlb_gather_mmu(tlb
, mm
, start
, end
);
438 inc_tlb_flush_pending(tlb
->mm
);
441 void tlb_finish_mmu(struct mmu_gather
*tlb
,
442 unsigned long start
, unsigned long end
)
445 * If there are parallel threads are doing PTE changes on same range
446 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
447 * flush by batching, a thread has stable TLB entry can fail to flush
448 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
449 * forcefully if we detect parallel PTE batching threads.
451 bool force
= mm_tlb_flush_nested(tlb
->mm
);
453 arch_tlb_finish_mmu(tlb
, start
, end
, force
);
454 dec_tlb_flush_pending(tlb
->mm
);
458 * Note: this doesn't free the actual pages themselves. That
459 * has been handled earlier when unmapping all the memory regions.
461 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
464 pgtable_t token
= pmd_pgtable(*pmd
);
466 pte_free_tlb(tlb
, token
, addr
);
467 atomic_long_dec(&tlb
->mm
->nr_ptes
);
470 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
471 unsigned long addr
, unsigned long end
,
472 unsigned long floor
, unsigned long ceiling
)
479 pmd
= pmd_offset(pud
, addr
);
481 next
= pmd_addr_end(addr
, end
);
482 if (pmd_none_or_clear_bad(pmd
))
484 free_pte_range(tlb
, pmd
, addr
);
485 } while (pmd
++, addr
= next
, addr
!= end
);
495 if (end
- 1 > ceiling
- 1)
498 pmd
= pmd_offset(pud
, start
);
500 pmd_free_tlb(tlb
, pmd
, start
);
501 mm_dec_nr_pmds(tlb
->mm
);
504 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
505 unsigned long addr
, unsigned long end
,
506 unsigned long floor
, unsigned long ceiling
)
513 pud
= pud_offset(p4d
, addr
);
515 next
= pud_addr_end(addr
, end
);
516 if (pud_none_or_clear_bad(pud
))
518 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
519 } while (pud
++, addr
= next
, addr
!= end
);
529 if (end
- 1 > ceiling
- 1)
532 pud
= pud_offset(p4d
, start
);
534 pud_free_tlb(tlb
, pud
, start
);
537 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
538 unsigned long addr
, unsigned long end
,
539 unsigned long floor
, unsigned long ceiling
)
546 p4d
= p4d_offset(pgd
, addr
);
548 next
= p4d_addr_end(addr
, end
);
549 if (p4d_none_or_clear_bad(p4d
))
551 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
552 } while (p4d
++, addr
= next
, addr
!= end
);
558 ceiling
&= PGDIR_MASK
;
562 if (end
- 1 > ceiling
- 1)
565 p4d
= p4d_offset(pgd
, start
);
567 p4d_free_tlb(tlb
, p4d
, start
);
571 * This function frees user-level page tables of a process.
573 void free_pgd_range(struct mmu_gather
*tlb
,
574 unsigned long addr
, unsigned long end
,
575 unsigned long floor
, unsigned long ceiling
)
581 * The next few lines have given us lots of grief...
583 * Why are we testing PMD* at this top level? Because often
584 * there will be no work to do at all, and we'd prefer not to
585 * go all the way down to the bottom just to discover that.
587 * Why all these "- 1"s? Because 0 represents both the bottom
588 * of the address space and the top of it (using -1 for the
589 * top wouldn't help much: the masks would do the wrong thing).
590 * The rule is that addr 0 and floor 0 refer to the bottom of
591 * the address space, but end 0 and ceiling 0 refer to the top
592 * Comparisons need to use "end - 1" and "ceiling - 1" (though
593 * that end 0 case should be mythical).
595 * Wherever addr is brought up or ceiling brought down, we must
596 * be careful to reject "the opposite 0" before it confuses the
597 * subsequent tests. But what about where end is brought down
598 * by PMD_SIZE below? no, end can't go down to 0 there.
600 * Whereas we round start (addr) and ceiling down, by different
601 * masks at different levels, in order to test whether a table
602 * now has no other vmas using it, so can be freed, we don't
603 * bother to round floor or end up - the tests don't need that.
617 if (end
- 1 > ceiling
- 1)
622 * We add page table cache pages with PAGE_SIZE,
623 * (see pte_free_tlb()), flush the tlb if we need
625 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
626 pgd
= pgd_offset(tlb
->mm
, addr
);
628 next
= pgd_addr_end(addr
, end
);
629 if (pgd_none_or_clear_bad(pgd
))
631 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
632 } while (pgd
++, addr
= next
, addr
!= end
);
635 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
636 unsigned long floor
, unsigned long ceiling
)
639 struct vm_area_struct
*next
= vma
->vm_next
;
640 unsigned long addr
= vma
->vm_start
;
643 * Hide vma from rmap and truncate_pagecache before freeing
646 unlink_anon_vmas(vma
);
647 unlink_file_vma(vma
);
649 if (is_vm_hugetlb_page(vma
)) {
650 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
651 floor
, next
? next
->vm_start
: ceiling
);
654 * Optimization: gather nearby vmas into one call down
656 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
657 && !is_vm_hugetlb_page(next
)) {
660 unlink_anon_vmas(vma
);
661 unlink_file_vma(vma
);
663 free_pgd_range(tlb
, addr
, vma
->vm_end
,
664 floor
, next
? next
->vm_start
: ceiling
);
670 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
673 pgtable_t
new = pte_alloc_one(mm
, address
);
678 * Ensure all pte setup (eg. pte page lock and page clearing) are
679 * visible before the pte is made visible to other CPUs by being
680 * put into page tables.
682 * The other side of the story is the pointer chasing in the page
683 * table walking code (when walking the page table without locking;
684 * ie. most of the time). Fortunately, these data accesses consist
685 * of a chain of data-dependent loads, meaning most CPUs (alpha
686 * being the notable exception) will already guarantee loads are
687 * seen in-order. See the alpha page table accessors for the
688 * smp_read_barrier_depends() barriers in page table walking code.
690 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
692 ptl
= pmd_lock(mm
, pmd
);
693 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
694 atomic_long_inc(&mm
->nr_ptes
);
695 pmd_populate(mm
, pmd
, new);
704 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
706 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
710 smp_wmb(); /* See comment in __pte_alloc */
712 spin_lock(&init_mm
.page_table_lock
);
713 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
714 pmd_populate_kernel(&init_mm
, pmd
, new);
717 spin_unlock(&init_mm
.page_table_lock
);
719 pte_free_kernel(&init_mm
, new);
723 static inline void init_rss_vec(int *rss
)
725 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
728 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
732 if (current
->mm
== mm
)
734 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
736 add_mm_counter(mm
, i
, rss
[i
]);
740 * This function is called to print an error when a bad pte
741 * is found. For example, we might have a PFN-mapped pte in
742 * a region that doesn't allow it.
744 * The calling function must still handle the error.
746 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
747 pte_t pte
, struct page
*page
)
749 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
750 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
751 pud_t
*pud
= pud_offset(p4d
, addr
);
752 pmd_t
*pmd
= pmd_offset(pud
, addr
);
753 struct address_space
*mapping
;
755 static unsigned long resume
;
756 static unsigned long nr_shown
;
757 static unsigned long nr_unshown
;
760 * Allow a burst of 60 reports, then keep quiet for that minute;
761 * or allow a steady drip of one report per second.
763 if (nr_shown
== 60) {
764 if (time_before(jiffies
, resume
)) {
769 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
776 resume
= jiffies
+ 60 * HZ
;
778 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
779 index
= linear_page_index(vma
, addr
);
781 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
783 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
785 dump_page(page
, "bad pte");
786 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
787 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
789 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
791 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
793 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
794 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
795 mapping
? mapping
->a_ops
->readpage
: NULL
);
797 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
801 * vm_normal_page -- This function gets the "struct page" associated with a pte.
803 * "Special" mappings do not wish to be associated with a "struct page" (either
804 * it doesn't exist, or it exists but they don't want to touch it). In this
805 * case, NULL is returned here. "Normal" mappings do have a struct page.
807 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
808 * pte bit, in which case this function is trivial. Secondly, an architecture
809 * may not have a spare pte bit, which requires a more complicated scheme,
812 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
813 * special mapping (even if there are underlying and valid "struct pages").
814 * COWed pages of a VM_PFNMAP are always normal.
816 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
817 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
818 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
819 * mapping will always honor the rule
821 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
823 * And for normal mappings this is false.
825 * This restricts such mappings to be a linear translation from virtual address
826 * to pfn. To get around this restriction, we allow arbitrary mappings so long
827 * as the vma is not a COW mapping; in that case, we know that all ptes are
828 * special (because none can have been COWed).
831 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
833 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
834 * page" backing, however the difference is that _all_ pages with a struct
835 * page (that is, those where pfn_valid is true) are refcounted and considered
836 * normal pages by the VM. The disadvantage is that pages are refcounted
837 * (which can be slower and simply not an option for some PFNMAP users). The
838 * advantage is that we don't have to follow the strict linearity rule of
839 * PFNMAP mappings in order to support COWable mappings.
842 #ifdef __HAVE_ARCH_PTE_SPECIAL
843 # define HAVE_PTE_SPECIAL 1
845 # define HAVE_PTE_SPECIAL 0
847 struct page
*_vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
848 pte_t pte
, bool with_public_device
)
850 unsigned long pfn
= pte_pfn(pte
);
852 if (HAVE_PTE_SPECIAL
) {
853 if (likely(!pte_special(pte
)))
855 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
856 return vma
->vm_ops
->find_special_page(vma
, addr
);
857 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
859 if (is_zero_pfn(pfn
))
863 * Device public pages are special pages (they are ZONE_DEVICE
864 * pages but different from persistent memory). They behave
865 * allmost like normal pages. The difference is that they are
866 * not on the lru and thus should never be involve with any-
867 * thing that involve lru manipulation (mlock, numa balancing,
870 * This is why we still want to return NULL for such page from
871 * vm_normal_page() so that we do not have to special case all
872 * call site of vm_normal_page().
874 if (likely(pfn
<= highest_memmap_pfn
)) {
875 struct page
*page
= pfn_to_page(pfn
);
877 if (is_device_public_page(page
)) {
878 if (with_public_device
)
883 print_bad_pte(vma
, addr
, pte
, NULL
);
887 /* !HAVE_PTE_SPECIAL case follows: */
889 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
890 if (vma
->vm_flags
& VM_MIXEDMAP
) {
896 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
897 if (pfn
== vma
->vm_pgoff
+ off
)
899 if (!is_cow_mapping(vma
->vm_flags
))
904 if (is_zero_pfn(pfn
))
907 if (unlikely(pfn
> highest_memmap_pfn
)) {
908 print_bad_pte(vma
, addr
, pte
, NULL
);
913 * NOTE! We still have PageReserved() pages in the page tables.
914 * eg. VDSO mappings can cause them to exist.
917 return pfn_to_page(pfn
);
920 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
921 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
924 unsigned long pfn
= pmd_pfn(pmd
);
927 * There is no pmd_special() but there may be special pmds, e.g.
928 * in a direct-access (dax) mapping, so let's just replicate the
929 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
931 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
932 if (vma
->vm_flags
& VM_MIXEDMAP
) {
938 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
939 if (pfn
== vma
->vm_pgoff
+ off
)
941 if (!is_cow_mapping(vma
->vm_flags
))
946 if (is_zero_pfn(pfn
))
948 if (unlikely(pfn
> highest_memmap_pfn
))
952 * NOTE! We still have PageReserved() pages in the page tables.
953 * eg. VDSO mappings can cause them to exist.
956 return pfn_to_page(pfn
);
961 * copy one vm_area from one task to the other. Assumes the page tables
962 * already present in the new task to be cleared in the whole range
963 * covered by this vma.
966 static inline unsigned long
967 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
968 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
969 unsigned long addr
, int *rss
)
971 unsigned long vm_flags
= vma
->vm_flags
;
972 pte_t pte
= *src_pte
;
975 /* pte contains position in swap or file, so copy. */
976 if (unlikely(!pte_present(pte
))) {
977 swp_entry_t entry
= pte_to_swp_entry(pte
);
979 if (likely(!non_swap_entry(entry
))) {
980 if (swap_duplicate(entry
) < 0)
983 /* make sure dst_mm is on swapoff's mmlist. */
984 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
985 spin_lock(&mmlist_lock
);
986 if (list_empty(&dst_mm
->mmlist
))
987 list_add(&dst_mm
->mmlist
,
989 spin_unlock(&mmlist_lock
);
992 } else if (is_migration_entry(entry
)) {
993 page
= migration_entry_to_page(entry
);
995 rss
[mm_counter(page
)]++;
997 if (is_write_migration_entry(entry
) &&
998 is_cow_mapping(vm_flags
)) {
1000 * COW mappings require pages in both
1001 * parent and child to be set to read.
1003 make_migration_entry_read(&entry
);
1004 pte
= swp_entry_to_pte(entry
);
1005 if (pte_swp_soft_dirty(*src_pte
))
1006 pte
= pte_swp_mksoft_dirty(pte
);
1007 set_pte_at(src_mm
, addr
, src_pte
, pte
);
1009 } else if (is_device_private_entry(entry
)) {
1010 page
= device_private_entry_to_page(entry
);
1013 * Update rss count even for unaddressable pages, as
1014 * they should treated just like normal pages in this
1017 * We will likely want to have some new rss counters
1018 * for unaddressable pages, at some point. But for now
1019 * keep things as they are.
1022 rss
[mm_counter(page
)]++;
1023 page_dup_rmap(page
, false);
1026 * We do not preserve soft-dirty information, because so
1027 * far, checkpoint/restore is the only feature that
1028 * requires that. And checkpoint/restore does not work
1029 * when a device driver is involved (you cannot easily
1030 * save and restore device driver state).
1032 if (is_write_device_private_entry(entry
) &&
1033 is_cow_mapping(vm_flags
)) {
1034 make_device_private_entry_read(&entry
);
1035 pte
= swp_entry_to_pte(entry
);
1036 set_pte_at(src_mm
, addr
, src_pte
, pte
);
1043 * If it's a COW mapping, write protect it both
1044 * in the parent and the child
1046 if (is_cow_mapping(vm_flags
)) {
1047 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
1048 pte
= pte_wrprotect(pte
);
1052 * If it's a shared mapping, mark it clean in
1055 if (vm_flags
& VM_SHARED
)
1056 pte
= pte_mkclean(pte
);
1057 pte
= pte_mkold(pte
);
1059 page
= vm_normal_page(vma
, addr
, pte
);
1062 page_dup_rmap(page
, false);
1063 rss
[mm_counter(page
)]++;
1064 } else if (pte_devmap(pte
)) {
1065 page
= pte_page(pte
);
1068 * Cache coherent device memory behave like regular page and
1069 * not like persistent memory page. For more informations see
1070 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1072 if (is_device_public_page(page
)) {
1074 page_dup_rmap(page
, false);
1075 rss
[mm_counter(page
)]++;
1080 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
1084 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1085 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
1086 unsigned long addr
, unsigned long end
)
1088 pte_t
*orig_src_pte
, *orig_dst_pte
;
1089 pte_t
*src_pte
, *dst_pte
;
1090 spinlock_t
*src_ptl
, *dst_ptl
;
1092 int rss
[NR_MM_COUNTERS
];
1093 swp_entry_t entry
= (swp_entry_t
){0};
1098 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1101 src_pte
= pte_offset_map(src_pmd
, addr
);
1102 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
1103 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1104 orig_src_pte
= src_pte
;
1105 orig_dst_pte
= dst_pte
;
1106 arch_enter_lazy_mmu_mode();
1110 * We are holding two locks at this point - either of them
1111 * could generate latencies in another task on another CPU.
1113 if (progress
>= 32) {
1115 if (need_resched() ||
1116 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1119 if (pte_none(*src_pte
)) {
1123 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
1128 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1130 arch_leave_lazy_mmu_mode();
1131 spin_unlock(src_ptl
);
1132 pte_unmap(orig_src_pte
);
1133 add_mm_rss_vec(dst_mm
, rss
);
1134 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1138 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
1147 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1148 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
1149 unsigned long addr
, unsigned long end
)
1151 pmd_t
*src_pmd
, *dst_pmd
;
1154 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1157 src_pmd
= pmd_offset(src_pud
, addr
);
1159 next
= pmd_addr_end(addr
, end
);
1160 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1161 || pmd_devmap(*src_pmd
)) {
1163 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, vma
);
1164 err
= copy_huge_pmd(dst_mm
, src_mm
,
1165 dst_pmd
, src_pmd
, addr
, vma
);
1172 if (pmd_none_or_clear_bad(src_pmd
))
1174 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1177 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1181 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1182 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
1183 unsigned long addr
, unsigned long end
)
1185 pud_t
*src_pud
, *dst_pud
;
1188 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1191 src_pud
= pud_offset(src_p4d
, addr
);
1193 next
= pud_addr_end(addr
, end
);
1194 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1197 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
1198 err
= copy_huge_pud(dst_mm
, src_mm
,
1199 dst_pud
, src_pud
, addr
, vma
);
1206 if (pud_none_or_clear_bad(src_pud
))
1208 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1211 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1215 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1216 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1217 unsigned long addr
, unsigned long end
)
1219 p4d_t
*src_p4d
, *dst_p4d
;
1222 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1225 src_p4d
= p4d_offset(src_pgd
, addr
);
1227 next
= p4d_addr_end(addr
, end
);
1228 if (p4d_none_or_clear_bad(src_p4d
))
1230 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
1233 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1237 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1238 struct vm_area_struct
*vma
)
1240 pgd_t
*src_pgd
, *dst_pgd
;
1242 unsigned long addr
= vma
->vm_start
;
1243 unsigned long end
= vma
->vm_end
;
1244 unsigned long mmun_start
; /* For mmu_notifiers */
1245 unsigned long mmun_end
; /* For mmu_notifiers */
1250 * Don't copy ptes where a page fault will fill them correctly.
1251 * Fork becomes much lighter when there are big shared or private
1252 * readonly mappings. The tradeoff is that copy_page_range is more
1253 * efficient than faulting.
1255 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1259 if (is_vm_hugetlb_page(vma
))
1260 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1262 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1264 * We do not free on error cases below as remove_vma
1265 * gets called on error from higher level routine
1267 ret
= track_pfn_copy(vma
);
1273 * We need to invalidate the secondary MMU mappings only when
1274 * there could be a permission downgrade on the ptes of the
1275 * parent mm. And a permission downgrade will only happen if
1276 * is_cow_mapping() returns true.
1278 is_cow
= is_cow_mapping(vma
->vm_flags
);
1282 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1286 dst_pgd
= pgd_offset(dst_mm
, addr
);
1287 src_pgd
= pgd_offset(src_mm
, addr
);
1289 next
= pgd_addr_end(addr
, end
);
1290 if (pgd_none_or_clear_bad(src_pgd
))
1292 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1293 vma
, addr
, next
))) {
1297 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1300 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1304 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1305 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1306 unsigned long addr
, unsigned long end
,
1307 struct zap_details
*details
)
1309 struct mm_struct
*mm
= tlb
->mm
;
1310 int force_flush
= 0;
1311 int rss
[NR_MM_COUNTERS
];
1317 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
1320 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1322 flush_tlb_batched_pending(mm
);
1323 arch_enter_lazy_mmu_mode();
1326 if (pte_none(ptent
))
1329 if (pte_present(ptent
)) {
1332 page
= _vm_normal_page(vma
, addr
, ptent
, true);
1333 if (unlikely(details
) && page
) {
1335 * unmap_shared_mapping_pages() wants to
1336 * invalidate cache without truncating:
1337 * unmap shared but keep private pages.
1339 if (details
->check_mapping
&&
1340 details
->check_mapping
!= page_rmapping(page
))
1343 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1345 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1346 if (unlikely(!page
))
1349 if (!PageAnon(page
)) {
1350 if (pte_dirty(ptent
)) {
1352 set_page_dirty(page
);
1354 if (pte_young(ptent
) &&
1355 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1356 mark_page_accessed(page
);
1358 rss
[mm_counter(page
)]--;
1359 page_remove_rmap(page
, false);
1360 if (unlikely(page_mapcount(page
) < 0))
1361 print_bad_pte(vma
, addr
, ptent
, page
);
1362 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1370 entry
= pte_to_swp_entry(ptent
);
1371 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1372 struct page
*page
= device_private_entry_to_page(entry
);
1374 if (unlikely(details
&& details
->check_mapping
)) {
1376 * unmap_shared_mapping_pages() wants to
1377 * invalidate cache without truncating:
1378 * unmap shared but keep private pages.
1380 if (details
->check_mapping
!=
1381 page_rmapping(page
))
1385 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1386 rss
[mm_counter(page
)]--;
1387 page_remove_rmap(page
, false);
1392 /* If details->check_mapping, we leave swap entries. */
1393 if (unlikely(details
))
1396 entry
= pte_to_swp_entry(ptent
);
1397 if (!non_swap_entry(entry
))
1399 else if (is_migration_entry(entry
)) {
1402 page
= migration_entry_to_page(entry
);
1403 rss
[mm_counter(page
)]--;
1405 if (unlikely(!free_swap_and_cache(entry
)))
1406 print_bad_pte(vma
, addr
, ptent
, NULL
);
1407 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1408 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1410 add_mm_rss_vec(mm
, rss
);
1411 arch_leave_lazy_mmu_mode();
1413 /* Do the actual TLB flush before dropping ptl */
1415 tlb_flush_mmu_tlbonly(tlb
);
1416 pte_unmap_unlock(start_pte
, ptl
);
1419 * If we forced a TLB flush (either due to running out of
1420 * batch buffers or because we needed to flush dirty TLB
1421 * entries before releasing the ptl), free the batched
1422 * memory too. Restart if we didn't do everything.
1426 tlb_flush_mmu_free(tlb
);
1434 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1435 struct vm_area_struct
*vma
, pud_t
*pud
,
1436 unsigned long addr
, unsigned long end
,
1437 struct zap_details
*details
)
1442 pmd
= pmd_offset(pud
, addr
);
1444 next
= pmd_addr_end(addr
, end
);
1445 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1446 if (next
- addr
!= HPAGE_PMD_SIZE
)
1447 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1448 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1453 * Here there can be other concurrent MADV_DONTNEED or
1454 * trans huge page faults running, and if the pmd is
1455 * none or trans huge it can change under us. This is
1456 * because MADV_DONTNEED holds the mmap_sem in read
1459 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1461 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1464 } while (pmd
++, addr
= next
, addr
!= end
);
1469 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1470 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1471 unsigned long addr
, unsigned long end
,
1472 struct zap_details
*details
)
1477 pud
= pud_offset(p4d
, addr
);
1479 next
= pud_addr_end(addr
, end
);
1480 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1481 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1482 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1483 split_huge_pud(vma
, pud
, addr
);
1484 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1488 if (pud_none_or_clear_bad(pud
))
1490 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1493 } while (pud
++, addr
= next
, addr
!= end
);
1498 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1499 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1500 unsigned long addr
, unsigned long end
,
1501 struct zap_details
*details
)
1506 p4d
= p4d_offset(pgd
, addr
);
1508 next
= p4d_addr_end(addr
, end
);
1509 if (p4d_none_or_clear_bad(p4d
))
1511 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1512 } while (p4d
++, addr
= next
, addr
!= end
);
1517 void unmap_page_range(struct mmu_gather
*tlb
,
1518 struct vm_area_struct
*vma
,
1519 unsigned long addr
, unsigned long end
,
1520 struct zap_details
*details
)
1525 BUG_ON(addr
>= end
);
1526 tlb_start_vma(tlb
, vma
);
1527 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1529 next
= pgd_addr_end(addr
, end
);
1530 if (pgd_none_or_clear_bad(pgd
))
1532 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1533 } while (pgd
++, addr
= next
, addr
!= end
);
1534 tlb_end_vma(tlb
, vma
);
1538 static void unmap_single_vma(struct mmu_gather
*tlb
,
1539 struct vm_area_struct
*vma
, unsigned long start_addr
,
1540 unsigned long end_addr
,
1541 struct zap_details
*details
)
1543 unsigned long start
= max(vma
->vm_start
, start_addr
);
1546 if (start
>= vma
->vm_end
)
1548 end
= min(vma
->vm_end
, end_addr
);
1549 if (end
<= vma
->vm_start
)
1553 uprobe_munmap(vma
, start
, end
);
1555 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1556 untrack_pfn(vma
, 0, 0);
1559 if (unlikely(is_vm_hugetlb_page(vma
))) {
1561 * It is undesirable to test vma->vm_file as it
1562 * should be non-null for valid hugetlb area.
1563 * However, vm_file will be NULL in the error
1564 * cleanup path of mmap_region. When
1565 * hugetlbfs ->mmap method fails,
1566 * mmap_region() nullifies vma->vm_file
1567 * before calling this function to clean up.
1568 * Since no pte has actually been setup, it is
1569 * safe to do nothing in this case.
1572 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1573 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1574 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1577 unmap_page_range(tlb
, vma
, start
, end
, details
);
1582 * unmap_vmas - unmap a range of memory covered by a list of vma's
1583 * @tlb: address of the caller's struct mmu_gather
1584 * @vma: the starting vma
1585 * @start_addr: virtual address at which to start unmapping
1586 * @end_addr: virtual address at which to end unmapping
1588 * Unmap all pages in the vma list.
1590 * Only addresses between `start' and `end' will be unmapped.
1592 * The VMA list must be sorted in ascending virtual address order.
1594 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1595 * range after unmap_vmas() returns. So the only responsibility here is to
1596 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1597 * drops the lock and schedules.
1599 void unmap_vmas(struct mmu_gather
*tlb
,
1600 struct vm_area_struct
*vma
, unsigned long start_addr
,
1601 unsigned long end_addr
)
1603 struct mm_struct
*mm
= vma
->vm_mm
;
1605 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1606 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1607 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1608 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1612 * zap_page_range - remove user pages in a given range
1613 * @vma: vm_area_struct holding the applicable pages
1614 * @start: starting address of pages to zap
1615 * @size: number of bytes to zap
1617 * Caller must protect the VMA list
1619 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1622 struct mm_struct
*mm
= vma
->vm_mm
;
1623 struct mmu_gather tlb
;
1624 unsigned long end
= start
+ size
;
1627 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1628 update_hiwater_rss(mm
);
1629 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1630 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
1631 unmap_single_vma(&tlb
, vma
, start
, end
, NULL
);
1634 * zap_page_range does not specify whether mmap_sem should be
1635 * held for read or write. That allows parallel zap_page_range
1636 * operations to unmap a PTE and defer a flush meaning that
1637 * this call observes pte_none and fails to flush the TLB.
1638 * Rather than adding a complex API, ensure that no stale
1639 * TLB entries exist when this call returns.
1641 flush_tlb_range(vma
, start
, end
);
1644 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1645 tlb_finish_mmu(&tlb
, start
, end
);
1649 * zap_page_range_single - remove user pages in a given range
1650 * @vma: vm_area_struct holding the applicable pages
1651 * @address: starting address of pages to zap
1652 * @size: number of bytes to zap
1653 * @details: details of shared cache invalidation
1655 * The range must fit into one VMA.
1657 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1658 unsigned long size
, struct zap_details
*details
)
1660 struct mm_struct
*mm
= vma
->vm_mm
;
1661 struct mmu_gather tlb
;
1662 unsigned long end
= address
+ size
;
1665 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1666 update_hiwater_rss(mm
);
1667 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1668 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1669 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1670 tlb_finish_mmu(&tlb
, address
, end
);
1674 * zap_vma_ptes - remove ptes mapping the vma
1675 * @vma: vm_area_struct holding ptes to be zapped
1676 * @address: starting address of pages to zap
1677 * @size: number of bytes to zap
1679 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1681 * The entire address range must be fully contained within the vma.
1683 * Returns 0 if successful.
1685 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1688 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1689 !(vma
->vm_flags
& VM_PFNMAP
))
1691 zap_page_range_single(vma
, address
, size
, NULL
);
1694 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1696 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1704 pgd
= pgd_offset(mm
, addr
);
1705 p4d
= p4d_alloc(mm
, pgd
, addr
);
1708 pud
= pud_alloc(mm
, p4d
, addr
);
1711 pmd
= pmd_alloc(mm
, pud
, addr
);
1715 VM_BUG_ON(pmd_trans_huge(*pmd
));
1716 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1720 * This is the old fallback for page remapping.
1722 * For historical reasons, it only allows reserved pages. Only
1723 * old drivers should use this, and they needed to mark their
1724 * pages reserved for the old functions anyway.
1726 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1727 struct page
*page
, pgprot_t prot
)
1729 struct mm_struct
*mm
= vma
->vm_mm
;
1738 flush_dcache_page(page
);
1739 pte
= get_locked_pte(mm
, addr
, &ptl
);
1743 if (!pte_none(*pte
))
1746 /* Ok, finally just insert the thing.. */
1748 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1749 page_add_file_rmap(page
, false);
1750 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1753 pte_unmap_unlock(pte
, ptl
);
1756 pte_unmap_unlock(pte
, ptl
);
1762 * vm_insert_page - insert single page into user vma
1763 * @vma: user vma to map to
1764 * @addr: target user address of this page
1765 * @page: source kernel page
1767 * This allows drivers to insert individual pages they've allocated
1770 * The page has to be a nice clean _individual_ kernel allocation.
1771 * If you allocate a compound page, you need to have marked it as
1772 * such (__GFP_COMP), or manually just split the page up yourself
1773 * (see split_page()).
1775 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1776 * took an arbitrary page protection parameter. This doesn't allow
1777 * that. Your vma protection will have to be set up correctly, which
1778 * means that if you want a shared writable mapping, you'd better
1779 * ask for a shared writable mapping!
1781 * The page does not need to be reserved.
1783 * Usually this function is called from f_op->mmap() handler
1784 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1785 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1786 * function from other places, for example from page-fault handler.
1788 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1791 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1793 if (!page_count(page
))
1795 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1796 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1797 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1798 vma
->vm_flags
|= VM_MIXEDMAP
;
1800 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1802 EXPORT_SYMBOL(vm_insert_page
);
1804 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1805 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1807 struct mm_struct
*mm
= vma
->vm_mm
;
1813 pte
= get_locked_pte(mm
, addr
, &ptl
);
1817 if (!pte_none(*pte
)) {
1820 * For read faults on private mappings the PFN passed
1821 * in may not match the PFN we have mapped if the
1822 * mapped PFN is a writeable COW page. In the mkwrite
1823 * case we are creating a writable PTE for a shared
1824 * mapping and we expect the PFNs to match.
1826 if (WARN_ON_ONCE(pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)))
1834 /* Ok, finally just insert the thing.. */
1835 if (pfn_t_devmap(pfn
))
1836 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1838 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1842 entry
= pte_mkyoung(entry
);
1843 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1846 set_pte_at(mm
, addr
, pte
, entry
);
1847 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1851 pte_unmap_unlock(pte
, ptl
);
1857 * vm_insert_pfn - insert single pfn into user vma
1858 * @vma: user vma to map to
1859 * @addr: target user address of this page
1860 * @pfn: source kernel pfn
1862 * Similar to vm_insert_page, this allows drivers to insert individual pages
1863 * they've allocated into a user vma. Same comments apply.
1865 * This function should only be called from a vm_ops->fault handler, and
1866 * in that case the handler should return NULL.
1868 * vma cannot be a COW mapping.
1870 * As this is called only for pages that do not currently exist, we
1871 * do not need to flush old virtual caches or the TLB.
1873 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1876 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1878 EXPORT_SYMBOL(vm_insert_pfn
);
1881 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1882 * @vma: user vma to map to
1883 * @addr: target user address of this page
1884 * @pfn: source kernel pfn
1885 * @pgprot: pgprot flags for the inserted page
1887 * This is exactly like vm_insert_pfn, except that it allows drivers to
1888 * to override pgprot on a per-page basis.
1890 * This only makes sense for IO mappings, and it makes no sense for
1891 * cow mappings. In general, using multiple vmas is preferable;
1892 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1895 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1896 unsigned long pfn
, pgprot_t pgprot
)
1900 * Technically, architectures with pte_special can avoid all these
1901 * restrictions (same for remap_pfn_range). However we would like
1902 * consistency in testing and feature parity among all, so we should
1903 * try to keep these invariants in place for everybody.
1905 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1906 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1907 (VM_PFNMAP
|VM_MIXEDMAP
));
1908 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1909 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1911 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1914 if (!pfn_modify_allowed(pfn
, pgprot
))
1917 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1919 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1924 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1926 static int __vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1927 pfn_t pfn
, bool mkwrite
)
1929 pgprot_t pgprot
= vma
->vm_page_prot
;
1931 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1933 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1936 track_pfn_insert(vma
, &pgprot
, pfn
);
1938 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1942 * If we don't have pte special, then we have to use the pfn_valid()
1943 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1944 * refcount the page if pfn_valid is true (hence insert_page rather
1945 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1946 * without pte special, it would there be refcounted as a normal page.
1948 if (!HAVE_PTE_SPECIAL
&& !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1952 * At this point we are committed to insert_page()
1953 * regardless of whether the caller specified flags that
1954 * result in pfn_t_has_page() == false.
1956 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1957 return insert_page(vma
, addr
, page
, pgprot
);
1959 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1962 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1965 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1968 EXPORT_SYMBOL(vm_insert_mixed
);
1970 int vm_insert_mixed_mkwrite(struct vm_area_struct
*vma
, unsigned long addr
,
1973 return __vm_insert_mixed(vma
, addr
, pfn
, true);
1975 EXPORT_SYMBOL(vm_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 int 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 int 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(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
);
2598 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2601 * Don't let another task, with possibly unlocked vma,
2602 * keep the mlocked page.
2604 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2605 lock_page(old_page
); /* LRU manipulation */
2606 if (PageMlocked(old_page
))
2607 munlock_vma_page(old_page
);
2608 unlock_page(old_page
);
2612 return page_copied
? VM_FAULT_WRITE
: 0;
2618 return VM_FAULT_OOM
;
2622 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2623 * writeable once the page is prepared
2625 * @vmf: structure describing the fault
2627 * This function handles all that is needed to finish a write page fault in a
2628 * shared mapping due to PTE being read-only once the mapped page is prepared.
2629 * It handles locking of PTE and modifying it. The function returns
2630 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2633 * The function expects the page to be locked or other protection against
2634 * concurrent faults / writeback (such as DAX radix tree locks).
2636 int finish_mkwrite_fault(struct vm_fault
*vmf
)
2638 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2639 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2642 * We might have raced with another page fault while we released the
2643 * pte_offset_map_lock.
2645 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2646 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2647 return VM_FAULT_NOPAGE
;
2654 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2657 static int wp_pfn_shared(struct vm_fault
*vmf
)
2659 struct vm_area_struct
*vma
= vmf
->vma
;
2661 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2664 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2665 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2666 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2667 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2669 return finish_mkwrite_fault(vmf
);
2672 return VM_FAULT_WRITE
;
2675 static int wp_page_shared(struct vm_fault
*vmf
)
2676 __releases(vmf
->ptl
)
2678 struct vm_area_struct
*vma
= vmf
->vma
;
2680 get_page(vmf
->page
);
2682 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2685 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2686 tmp
= do_page_mkwrite(vmf
);
2687 if (unlikely(!tmp
|| (tmp
&
2688 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2689 put_page(vmf
->page
);
2692 tmp
= finish_mkwrite_fault(vmf
);
2693 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2694 unlock_page(vmf
->page
);
2695 put_page(vmf
->page
);
2700 lock_page(vmf
->page
);
2702 fault_dirty_shared_page(vma
, vmf
->page
);
2703 put_page(vmf
->page
);
2705 return VM_FAULT_WRITE
;
2709 * This routine handles present pages, when users try to write
2710 * to a shared page. It is done by copying the page to a new address
2711 * and decrementing the shared-page counter for the old page.
2713 * Note that this routine assumes that the protection checks have been
2714 * done by the caller (the low-level page fault routine in most cases).
2715 * Thus we can safely just mark it writable once we've done any necessary
2718 * We also mark the page dirty at this point even though the page will
2719 * change only once the write actually happens. This avoids a few races,
2720 * and potentially makes it more efficient.
2722 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2723 * but allow concurrent faults), with pte both mapped and locked.
2724 * We return with mmap_sem still held, but pte unmapped and unlocked.
2726 static int do_wp_page(struct vm_fault
*vmf
)
2727 __releases(vmf
->ptl
)
2729 struct vm_area_struct
*vma
= vmf
->vma
;
2731 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2734 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2737 * We should not cow pages in a shared writeable mapping.
2738 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2740 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2741 (VM_WRITE
|VM_SHARED
))
2742 return wp_pfn_shared(vmf
);
2744 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2745 return wp_page_copy(vmf
);
2749 * Take out anonymous pages first, anonymous shared vmas are
2750 * not dirty accountable.
2752 if (PageAnon(vmf
->page
) && !PageKsm(vmf
->page
)) {
2753 int total_map_swapcount
;
2754 if (!trylock_page(vmf
->page
)) {
2755 get_page(vmf
->page
);
2756 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2757 lock_page(vmf
->page
);
2758 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2759 vmf
->address
, &vmf
->ptl
);
2760 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2761 unlock_page(vmf
->page
);
2762 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2763 put_page(vmf
->page
);
2766 put_page(vmf
->page
);
2768 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2769 if (total_map_swapcount
== 1) {
2771 * The page is all ours. Move it to
2772 * our anon_vma so the rmap code will
2773 * not search our parent or siblings.
2774 * Protected against the rmap code by
2777 page_move_anon_rmap(vmf
->page
, vma
);
2779 unlock_page(vmf
->page
);
2781 return VM_FAULT_WRITE
;
2783 unlock_page(vmf
->page
);
2784 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2785 (VM_WRITE
|VM_SHARED
))) {
2786 return wp_page_shared(vmf
);
2790 * Ok, we need to copy. Oh, well..
2792 get_page(vmf
->page
);
2794 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2795 return wp_page_copy(vmf
);
2798 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2799 unsigned long start_addr
, unsigned long end_addr
,
2800 struct zap_details
*details
)
2802 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2805 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2806 struct zap_details
*details
)
2808 struct vm_area_struct
*vma
;
2809 pgoff_t vba
, vea
, zba
, zea
;
2811 vma_interval_tree_foreach(vma
, root
,
2812 details
->first_index
, details
->last_index
) {
2814 vba
= vma
->vm_pgoff
;
2815 vea
= vba
+ vma_pages(vma
) - 1;
2816 zba
= details
->first_index
;
2819 zea
= details
->last_index
;
2823 unmap_mapping_range_vma(vma
,
2824 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2825 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2831 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2832 * address_space corresponding to the specified page range in the underlying
2835 * @mapping: the address space containing mmaps to be unmapped.
2836 * @holebegin: byte in first page to unmap, relative to the start of
2837 * the underlying file. This will be rounded down to a PAGE_SIZE
2838 * boundary. Note that this is different from truncate_pagecache(), which
2839 * must keep the partial page. In contrast, we must get rid of
2841 * @holelen: size of prospective hole in bytes. This will be rounded
2842 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2844 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2845 * but 0 when invalidating pagecache, don't throw away private data.
2847 void unmap_mapping_range(struct address_space
*mapping
,
2848 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2850 struct zap_details details
= { };
2851 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2852 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2854 /* Check for overflow. */
2855 if (sizeof(holelen
) > sizeof(hlen
)) {
2857 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2858 if (holeend
& ~(long long)ULONG_MAX
)
2859 hlen
= ULONG_MAX
- hba
+ 1;
2862 details
.check_mapping
= even_cows
? NULL
: mapping
;
2863 details
.first_index
= hba
;
2864 details
.last_index
= hba
+ hlen
- 1;
2865 if (details
.last_index
< details
.first_index
)
2866 details
.last_index
= ULONG_MAX
;
2868 i_mmap_lock_write(mapping
);
2869 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2870 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2871 i_mmap_unlock_write(mapping
);
2873 EXPORT_SYMBOL(unmap_mapping_range
);
2876 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2877 * but allow concurrent faults), and pte mapped but not yet locked.
2878 * We return with pte unmapped and unlocked.
2880 * We return with the mmap_sem locked or unlocked in the same cases
2881 * as does filemap_fault().
2883 int do_swap_page(struct vm_fault
*vmf
)
2885 struct vm_area_struct
*vma
= vmf
->vma
;
2886 struct page
*page
= NULL
, *swapcache
;
2887 struct mem_cgroup
*memcg
;
2888 struct vma_swap_readahead swap_ra
;
2894 bool vma_readahead
= swap_use_vma_readahead();
2897 page
= swap_readahead_detect(vmf
, &swap_ra
);
2898 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
)) {
2904 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2905 if (unlikely(non_swap_entry(entry
))) {
2906 if (is_migration_entry(entry
)) {
2907 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2909 } else if (is_device_private_entry(entry
)) {
2911 * For un-addressable device memory we call the pgmap
2912 * fault handler callback. The callback must migrate
2913 * the page back to some CPU accessible page.
2915 ret
= device_private_entry_fault(vma
, vmf
->address
, entry
,
2916 vmf
->flags
, vmf
->pmd
);
2917 } else if (is_hwpoison_entry(entry
)) {
2918 ret
= VM_FAULT_HWPOISON
;
2920 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2921 ret
= VM_FAULT_SIGBUS
;
2925 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2927 page
= lookup_swap_cache(entry
, vma_readahead
? vma
: NULL
,
2931 page
= do_swap_page_readahead(entry
,
2932 GFP_HIGHUSER_MOVABLE
, vmf
, &swap_ra
);
2934 page
= swapin_readahead(entry
,
2935 GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
2938 * Back out if somebody else faulted in this pte
2939 * while we released the pte lock.
2941 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2942 vmf
->address
, &vmf
->ptl
);
2943 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2945 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2949 /* Had to read the page from swap area: Major fault */
2950 ret
= VM_FAULT_MAJOR
;
2951 count_vm_event(PGMAJFAULT
);
2952 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2953 } else if (PageHWPoison(page
)) {
2955 * hwpoisoned dirty swapcache pages are kept for killing
2956 * owner processes (which may be unknown at hwpoison time)
2958 ret
= VM_FAULT_HWPOISON
;
2959 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2965 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2967 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2969 ret
|= VM_FAULT_RETRY
;
2974 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2975 * release the swapcache from under us. The page pin, and pte_same
2976 * test below, are not enough to exclude that. Even if it is still
2977 * swapcache, we need to check that the page's swap has not changed.
2979 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2982 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
2983 if (unlikely(!page
)) {
2989 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
,
2996 * Back out if somebody else already faulted in this pte.
2998 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3000 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3003 if (unlikely(!PageUptodate(page
))) {
3004 ret
= VM_FAULT_SIGBUS
;
3009 * The page isn't present yet, go ahead with the fault.
3011 * Be careful about the sequence of operations here.
3012 * To get its accounting right, reuse_swap_page() must be called
3013 * while the page is counted on swap but not yet in mapcount i.e.
3014 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3015 * must be called after the swap_free(), or it will never succeed.
3018 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3019 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3020 pte
= mk_pte(page
, vma
->vm_page_prot
);
3021 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3022 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3023 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3024 ret
|= VM_FAULT_WRITE
;
3025 exclusive
= RMAP_EXCLUSIVE
;
3027 flush_icache_page(vma
, page
);
3028 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3029 pte
= pte_mksoft_dirty(pte
);
3030 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3031 vmf
->orig_pte
= pte
;
3032 if (page
== swapcache
) {
3033 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3034 mem_cgroup_commit_charge(page
, memcg
, true, false);
3035 activate_page(page
);
3036 } else { /* ksm created a completely new copy */
3037 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3038 mem_cgroup_commit_charge(page
, memcg
, false, false);
3039 lru_cache_add_active_or_unevictable(page
, vma
);
3043 if (mem_cgroup_swap_full(page
) ||
3044 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3045 try_to_free_swap(page
);
3047 if (page
!= swapcache
) {
3049 * Hold the lock to avoid the swap entry to be reused
3050 * until we take the PT lock for the pte_same() check
3051 * (to avoid false positives from pte_same). For
3052 * further safety release the lock after the swap_free
3053 * so that the swap count won't change under a
3054 * parallel locked swapcache.
3056 unlock_page(swapcache
);
3057 put_page(swapcache
);
3060 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3061 ret
|= do_wp_page(vmf
);
3062 if (ret
& VM_FAULT_ERROR
)
3063 ret
&= VM_FAULT_ERROR
;
3067 /* No need to invalidate - it was non-present before */
3068 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3070 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3072 return ret
| VM_FAULT_SWAP
;
3074 mem_cgroup_cancel_charge(page
, memcg
, false);
3075 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3080 if (page
!= swapcache
) {
3081 unlock_page(swapcache
);
3082 put_page(swapcache
);
3084 return ret
| VM_FAULT_SWAP
;
3088 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3089 * but allow concurrent faults), and pte mapped but not yet locked.
3090 * We return with mmap_sem still held, but pte unmapped and unlocked.
3092 static int do_anonymous_page(struct vm_fault
*vmf
)
3094 struct vm_area_struct
*vma
= vmf
->vma
;
3095 struct mem_cgroup
*memcg
;
3100 /* File mapping without ->vm_ops ? */
3101 if (vma
->vm_flags
& VM_SHARED
)
3102 return VM_FAULT_SIGBUS
;
3105 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3106 * pte_offset_map() on pmds where a huge pmd might be created
3107 * from a different thread.
3109 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3110 * parallel threads are excluded by other means.
3112 * Here we only have down_read(mmap_sem).
3114 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))
3115 return VM_FAULT_OOM
;
3117 /* See the comment in pte_alloc_one_map() */
3118 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3121 /* Use the zero-page for reads */
3122 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3123 !mm_forbids_zeropage(vma
->vm_mm
)) {
3124 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3125 vma
->vm_page_prot
));
3126 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3127 vmf
->address
, &vmf
->ptl
);
3128 if (!pte_none(*vmf
->pte
))
3130 ret
= check_stable_address_space(vma
->vm_mm
);
3133 /* Deliver the page fault to userland, check inside PT lock */
3134 if (userfaultfd_missing(vma
)) {
3135 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3136 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3141 /* Allocate our own private page. */
3142 if (unlikely(anon_vma_prepare(vma
)))
3144 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3148 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
3152 * The memory barrier inside __SetPageUptodate makes sure that
3153 * preceeding stores to the page contents become visible before
3154 * the set_pte_at() write.
3156 __SetPageUptodate(page
);
3158 entry
= mk_pte(page
, vma
->vm_page_prot
);
3159 if (vma
->vm_flags
& VM_WRITE
)
3160 entry
= pte_mkwrite(pte_mkdirty(entry
));
3162 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3164 if (!pte_none(*vmf
->pte
))
3167 ret
= check_stable_address_space(vma
->vm_mm
);
3171 /* Deliver the page fault to userland, check inside PT lock */
3172 if (userfaultfd_missing(vma
)) {
3173 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3174 mem_cgroup_cancel_charge(page
, memcg
, false);
3176 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3179 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3180 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3181 mem_cgroup_commit_charge(page
, memcg
, false, false);
3182 lru_cache_add_active_or_unevictable(page
, vma
);
3184 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3186 /* No need to invalidate - it was non-present before */
3187 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3189 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3192 mem_cgroup_cancel_charge(page
, memcg
, false);
3198 return VM_FAULT_OOM
;
3202 * The mmap_sem must have been held on entry, and may have been
3203 * released depending on flags and vma->vm_ops->fault() return value.
3204 * See filemap_fault() and __lock_page_retry().
3206 static int __do_fault(struct vm_fault
*vmf
)
3208 struct vm_area_struct
*vma
= vmf
->vma
;
3212 * Preallocate pte before we take page_lock because this might lead to
3213 * deadlocks for memcg reclaim which waits for pages under writeback:
3215 * SetPageWriteback(A)
3221 * wait_on_page_writeback(A)
3222 * SetPageWriteback(B)
3224 * # flush A, B to clear the writeback
3226 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3227 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
,
3229 if (!vmf
->prealloc_pte
)
3230 return VM_FAULT_OOM
;
3231 smp_wmb(); /* See comment in __pte_alloc() */
3234 ret
= vma
->vm_ops
->fault(vmf
);
3235 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3236 VM_FAULT_DONE_COW
)))
3239 if (unlikely(PageHWPoison(vmf
->page
))) {
3240 if (ret
& VM_FAULT_LOCKED
)
3241 unlock_page(vmf
->page
);
3242 put_page(vmf
->page
);
3244 return VM_FAULT_HWPOISON
;
3247 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3248 lock_page(vmf
->page
);
3250 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3256 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3257 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3258 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3259 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3261 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3263 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3266 static int pte_alloc_one_map(struct vm_fault
*vmf
)
3268 struct vm_area_struct
*vma
= vmf
->vma
;
3270 if (!pmd_none(*vmf
->pmd
))
3272 if (vmf
->prealloc_pte
) {
3273 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3274 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3275 spin_unlock(vmf
->ptl
);
3279 atomic_long_inc(&vma
->vm_mm
->nr_ptes
);
3280 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3281 spin_unlock(vmf
->ptl
);
3282 vmf
->prealloc_pte
= NULL
;
3283 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))) {
3284 return VM_FAULT_OOM
;
3288 * If a huge pmd materialized under us just retry later. Use
3289 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3290 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3291 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3292 * running immediately after a huge pmd fault in a different thread of
3293 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3294 * All we have to ensure is that it is a regular pmd that we can walk
3295 * with pte_offset_map() and we can do that through an atomic read in
3296 * C, which is what pmd_trans_unstable() provides.
3298 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3299 return VM_FAULT_NOPAGE
;
3302 * At this point we know that our vmf->pmd points to a page of ptes
3303 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3304 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3305 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3306 * be valid and we will re-check to make sure the vmf->pte isn't
3307 * pte_none() under vmf->ptl protection when we return to
3310 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3315 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3317 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3318 static inline bool transhuge_vma_suitable(struct vm_area_struct
*vma
,
3319 unsigned long haddr
)
3321 if (((vma
->vm_start
>> PAGE_SHIFT
) & HPAGE_CACHE_INDEX_MASK
) !=
3322 (vma
->vm_pgoff
& HPAGE_CACHE_INDEX_MASK
))
3324 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
3329 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3331 struct vm_area_struct
*vma
= vmf
->vma
;
3333 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3335 * We are going to consume the prealloc table,
3336 * count that as nr_ptes.
3338 atomic_long_inc(&vma
->vm_mm
->nr_ptes
);
3339 vmf
->prealloc_pte
= NULL
;
3342 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3344 struct vm_area_struct
*vma
= vmf
->vma
;
3345 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3346 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3350 if (!transhuge_vma_suitable(vma
, haddr
))
3351 return VM_FAULT_FALLBACK
;
3353 ret
= VM_FAULT_FALLBACK
;
3354 page
= compound_head(page
);
3357 * Archs like ppc64 need additonal space to store information
3358 * related to pte entry. Use the preallocated table for that.
3360 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3361 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
, vmf
->address
);
3362 if (!vmf
->prealloc_pte
)
3363 return VM_FAULT_OOM
;
3364 smp_wmb(); /* See comment in __pte_alloc() */
3367 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3368 if (unlikely(!pmd_none(*vmf
->pmd
)))
3371 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3372 flush_icache_page(vma
, page
+ i
);
3374 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3376 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3378 add_mm_counter(vma
->vm_mm
, MM_FILEPAGES
, HPAGE_PMD_NR
);
3379 page_add_file_rmap(page
, true);
3381 * deposit and withdraw with pmd lock held
3383 if (arch_needs_pgtable_deposit())
3384 deposit_prealloc_pte(vmf
);
3386 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3388 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3390 /* fault is handled */
3392 count_vm_event(THP_FILE_MAPPED
);
3394 spin_unlock(vmf
->ptl
);
3398 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3406 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3407 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3409 * @vmf: fault environment
3410 * @memcg: memcg to charge page (only for private mappings)
3411 * @page: page to map
3413 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3416 * Target users are page handler itself and implementations of
3417 * vm_ops->map_pages.
3419 int alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3422 struct vm_area_struct
*vma
= vmf
->vma
;
3423 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3427 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3428 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3430 VM_BUG_ON_PAGE(memcg
, page
);
3432 ret
= do_set_pmd(vmf
, page
);
3433 if (ret
!= VM_FAULT_FALLBACK
)
3438 ret
= pte_alloc_one_map(vmf
);
3443 /* Re-check under ptl */
3444 if (unlikely(!pte_none(*vmf
->pte
)))
3445 return VM_FAULT_NOPAGE
;
3447 flush_icache_page(vma
, page
);
3448 entry
= mk_pte(page
, vma
->vm_page_prot
);
3450 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3451 /* copy-on-write page */
3452 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3453 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3454 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3455 mem_cgroup_commit_charge(page
, memcg
, false, false);
3456 lru_cache_add_active_or_unevictable(page
, vma
);
3458 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3459 page_add_file_rmap(page
, false);
3461 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3463 /* no need to invalidate: a not-present page won't be cached */
3464 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3471 * finish_fault - finish page fault once we have prepared the page to fault
3473 * @vmf: structure describing the fault
3475 * This function handles all that is needed to finish a page fault once the
3476 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3477 * given page, adds reverse page mapping, handles memcg charges and LRU
3478 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3481 * The function expects the page to be locked and on success it consumes a
3482 * reference of a page being mapped (for the PTE which maps it).
3484 int finish_fault(struct vm_fault
*vmf
)
3489 /* Did we COW the page? */
3490 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3491 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3492 page
= vmf
->cow_page
;
3497 * check even for read faults because we might have lost our CoWed
3500 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3501 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3503 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3505 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3509 static unsigned long fault_around_bytes __read_mostly
=
3510 rounddown_pow_of_two(65536);
3512 #ifdef CONFIG_DEBUG_FS
3513 static int fault_around_bytes_get(void *data
, u64
*val
)
3515 *val
= fault_around_bytes
;
3520 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3521 * rounded down to nearest page order. It's what do_fault_around() expects to
3524 static int fault_around_bytes_set(void *data
, u64 val
)
3526 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3528 if (val
> PAGE_SIZE
)
3529 fault_around_bytes
= rounddown_pow_of_two(val
);
3531 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3534 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3535 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3537 static int __init
fault_around_debugfs(void)
3541 ret
= debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3542 &fault_around_bytes_fops
);
3544 pr_warn("Failed to create fault_around_bytes in debugfs");
3547 late_initcall(fault_around_debugfs
);
3551 * do_fault_around() tries to map few pages around the fault address. The hope
3552 * is that the pages will be needed soon and this will lower the number of
3555 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3556 * not ready to be mapped: not up-to-date, locked, etc.
3558 * This function is called with the page table lock taken. In the split ptlock
3559 * case the page table lock only protects only those entries which belong to
3560 * the page table corresponding to the fault address.
3562 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3565 * fault_around_pages() defines how many pages we'll try to map.
3566 * do_fault_around() expects it to return a power of two less than or equal to
3569 * The virtual address of the area that we map is naturally aligned to the
3570 * fault_around_pages() value (and therefore to page order). This way it's
3571 * easier to guarantee that we don't cross page table boundaries.
3573 static int do_fault_around(struct vm_fault
*vmf
)
3575 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3576 pgoff_t start_pgoff
= vmf
->pgoff
;
3580 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3581 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3583 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3584 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3588 * end_pgoff is either end of page table or end of vma
3589 * or fault_around_pages() from start_pgoff, depending what is nearest.
3591 end_pgoff
= start_pgoff
-
3592 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3594 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3595 start_pgoff
+ nr_pages
- 1);
3597 if (pmd_none(*vmf
->pmd
)) {
3598 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
,
3600 if (!vmf
->prealloc_pte
)
3602 smp_wmb(); /* See comment in __pte_alloc() */
3605 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3607 /* Huge page is mapped? Page fault is solved */
3608 if (pmd_trans_huge(*vmf
->pmd
)) {
3609 ret
= VM_FAULT_NOPAGE
;
3613 /* ->map_pages() haven't done anything useful. Cold page cache? */
3617 /* check if the page fault is solved */
3618 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3619 if (!pte_none(*vmf
->pte
))
3620 ret
= VM_FAULT_NOPAGE
;
3621 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3623 vmf
->address
= address
;
3628 static int do_read_fault(struct vm_fault
*vmf
)
3630 struct vm_area_struct
*vma
= vmf
->vma
;
3634 * Let's call ->map_pages() first and use ->fault() as fallback
3635 * if page by the offset is not ready to be mapped (cold cache or
3638 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3639 ret
= do_fault_around(vmf
);
3644 ret
= __do_fault(vmf
);
3645 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3648 ret
|= finish_fault(vmf
);
3649 unlock_page(vmf
->page
);
3650 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3651 put_page(vmf
->page
);
3655 static int do_cow_fault(struct vm_fault
*vmf
)
3657 struct vm_area_struct
*vma
= vmf
->vma
;
3660 if (unlikely(anon_vma_prepare(vma
)))
3661 return VM_FAULT_OOM
;
3663 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3665 return VM_FAULT_OOM
;
3667 if (mem_cgroup_try_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3668 &vmf
->memcg
, false)) {
3669 put_page(vmf
->cow_page
);
3670 return VM_FAULT_OOM
;
3673 ret
= __do_fault(vmf
);
3674 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3676 if (ret
& VM_FAULT_DONE_COW
)
3679 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3680 __SetPageUptodate(vmf
->cow_page
);
3682 ret
|= finish_fault(vmf
);
3683 unlock_page(vmf
->page
);
3684 put_page(vmf
->page
);
3685 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3689 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3690 put_page(vmf
->cow_page
);
3694 static int do_shared_fault(struct vm_fault
*vmf
)
3696 struct vm_area_struct
*vma
= vmf
->vma
;
3699 ret
= __do_fault(vmf
);
3700 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3704 * Check if the backing address space wants to know that the page is
3705 * about to become writable
3707 if (vma
->vm_ops
->page_mkwrite
) {
3708 unlock_page(vmf
->page
);
3709 tmp
= do_page_mkwrite(vmf
);
3710 if (unlikely(!tmp
||
3711 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3712 put_page(vmf
->page
);
3717 ret
|= finish_fault(vmf
);
3718 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3720 unlock_page(vmf
->page
);
3721 put_page(vmf
->page
);
3725 fault_dirty_shared_page(vma
, vmf
->page
);
3730 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3731 * but allow concurrent faults).
3732 * The mmap_sem may have been released depending on flags and our
3733 * return value. See filemap_fault() and __lock_page_or_retry().
3735 static int do_fault(struct vm_fault
*vmf
)
3737 struct vm_area_struct
*vma
= vmf
->vma
;
3741 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3743 if (!vma
->vm_ops
->fault
) {
3745 * If we find a migration pmd entry or a none pmd entry, which
3746 * should never happen, return SIGBUS
3748 if (unlikely(!pmd_present(*vmf
->pmd
)))
3749 ret
= VM_FAULT_SIGBUS
;
3751 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
3756 * Make sure this is not a temporary clearing of pte
3757 * by holding ptl and checking again. A R/M/W update
3758 * of pte involves: take ptl, clearing the pte so that
3759 * we don't have concurrent modification by hardware
3760 * followed by an update.
3762 if (unlikely(pte_none(*vmf
->pte
)))
3763 ret
= VM_FAULT_SIGBUS
;
3765 ret
= VM_FAULT_NOPAGE
;
3767 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3769 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3770 ret
= do_read_fault(vmf
);
3771 else if (!(vma
->vm_flags
& VM_SHARED
))
3772 ret
= do_cow_fault(vmf
);
3774 ret
= do_shared_fault(vmf
);
3776 /* preallocated pagetable is unused: free it */
3777 if (vmf
->prealloc_pte
) {
3778 pte_free(vma
->vm_mm
, vmf
->prealloc_pte
);
3779 vmf
->prealloc_pte
= NULL
;
3784 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3785 unsigned long addr
, int page_nid
,
3790 count_vm_numa_event(NUMA_HINT_FAULTS
);
3791 if (page_nid
== numa_node_id()) {
3792 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3793 *flags
|= TNF_FAULT_LOCAL
;
3796 return mpol_misplaced(page
, vma
, addr
);
3799 static int do_numa_page(struct vm_fault
*vmf
)
3801 struct vm_area_struct
*vma
= vmf
->vma
;
3802 struct page
*page
= NULL
;
3806 bool migrated
= false;
3808 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3812 * The "pte" at this point cannot be used safely without
3813 * validation through pte_unmap_same(). It's of NUMA type but
3814 * the pfn may be screwed if the read is non atomic.
3816 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3817 spin_lock(vmf
->ptl
);
3818 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3819 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3824 * Make it present again, Depending on how arch implementes non
3825 * accessible ptes, some can allow access by kernel mode.
3827 pte
= ptep_modify_prot_start(vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3828 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3829 pte
= pte_mkyoung(pte
);
3831 pte
= pte_mkwrite(pte
);
3832 ptep_modify_prot_commit(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3833 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3835 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3837 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3841 /* TODO: handle PTE-mapped THP */
3842 if (PageCompound(page
)) {
3843 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3848 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3849 * much anyway since they can be in shared cache state. This misses
3850 * the case where a mapping is writable but the process never writes
3851 * to it but pte_write gets cleared during protection updates and
3852 * pte_dirty has unpredictable behaviour between PTE scan updates,
3853 * background writeback, dirty balancing and application behaviour.
3855 if (!pte_write(pte
))
3856 flags
|= TNF_NO_GROUP
;
3859 * Flag if the page is shared between multiple address spaces. This
3860 * is later used when determining whether to group tasks together
3862 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3863 flags
|= TNF_SHARED
;
3865 last_cpupid
= page_cpupid_last(page
);
3866 page_nid
= page_to_nid(page
);
3867 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3869 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3870 if (target_nid
== -1) {
3875 /* Migrate to the requested node */
3876 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3878 page_nid
= target_nid
;
3879 flags
|= TNF_MIGRATED
;
3881 flags
|= TNF_MIGRATE_FAIL
;
3885 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3889 static inline int create_huge_pmd(struct vm_fault
*vmf
)
3891 if (vma_is_anonymous(vmf
->vma
))
3892 return do_huge_pmd_anonymous_page(vmf
);
3893 if (vmf
->vma
->vm_ops
->huge_fault
)
3894 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3895 return VM_FAULT_FALLBACK
;
3898 static int wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3900 if (vma_is_anonymous(vmf
->vma
))
3901 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3902 if (vmf
->vma
->vm_ops
->huge_fault
)
3903 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3905 /* COW handled on pte level: split pmd */
3906 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3907 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3909 return VM_FAULT_FALLBACK
;
3912 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3914 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3917 static int create_huge_pud(struct vm_fault
*vmf
)
3919 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3920 /* No support for anonymous transparent PUD pages yet */
3921 if (vma_is_anonymous(vmf
->vma
))
3922 return VM_FAULT_FALLBACK
;
3923 if (vmf
->vma
->vm_ops
->huge_fault
)
3924 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3925 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3926 return VM_FAULT_FALLBACK
;
3929 static int wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3931 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3932 /* No support for anonymous transparent PUD pages yet */
3933 if (vma_is_anonymous(vmf
->vma
))
3934 return VM_FAULT_FALLBACK
;
3935 if (vmf
->vma
->vm_ops
->huge_fault
)
3936 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3937 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3938 return VM_FAULT_FALLBACK
;
3942 * These routines also need to handle stuff like marking pages dirty
3943 * and/or accessed for architectures that don't do it in hardware (most
3944 * RISC architectures). The early dirtying is also good on the i386.
3946 * There is also a hook called "update_mmu_cache()" that architectures
3947 * with external mmu caches can use to update those (ie the Sparc or
3948 * PowerPC hashed page tables that act as extended TLBs).
3950 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3951 * concurrent faults).
3953 * The mmap_sem may have been released depending on flags and our return value.
3954 * See filemap_fault() and __lock_page_or_retry().
3956 static int handle_pte_fault(struct vm_fault
*vmf
)
3960 if (unlikely(pmd_none(*vmf
->pmd
))) {
3962 * Leave __pte_alloc() until later: because vm_ops->fault may
3963 * want to allocate huge page, and if we expose page table
3964 * for an instant, it will be difficult to retract from
3965 * concurrent faults and from rmap lookups.
3969 /* See comment in pte_alloc_one_map() */
3970 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3973 * A regular pmd is established and it can't morph into a huge
3974 * pmd from under us anymore at this point because we hold the
3975 * mmap_sem read mode and khugepaged takes it in write mode.
3976 * So now it's safe to run pte_offset_map().
3978 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
3979 vmf
->orig_pte
= *vmf
->pte
;
3982 * some architectures can have larger ptes than wordsize,
3983 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3984 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3985 * atomic accesses. The code below just needs a consistent
3986 * view for the ifs and we later double check anyway with the
3987 * ptl lock held. So here a barrier will do.
3990 if (pte_none(vmf
->orig_pte
)) {
3991 pte_unmap(vmf
->pte
);
3997 if (vma_is_anonymous(vmf
->vma
))
3998 return do_anonymous_page(vmf
);
4000 return do_fault(vmf
);
4003 if (!pte_present(vmf
->orig_pte
))
4004 return do_swap_page(vmf
);
4006 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4007 return do_numa_page(vmf
);
4009 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4010 spin_lock(vmf
->ptl
);
4011 entry
= vmf
->orig_pte
;
4012 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
4014 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4015 if (!pte_write(entry
))
4016 return do_wp_page(vmf
);
4017 entry
= pte_mkdirty(entry
);
4019 entry
= pte_mkyoung(entry
);
4020 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4021 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4022 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4025 * This is needed only for protection faults but the arch code
4026 * is not yet telling us if this is a protection fault or not.
4027 * This still avoids useless tlb flushes for .text page faults
4030 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4031 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4034 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4039 * By the time we get here, we already hold the mm semaphore
4041 * The mmap_sem may have been released depending on flags and our
4042 * return value. See filemap_fault() and __lock_page_or_retry().
4044 static int __handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4047 struct vm_fault vmf
= {
4049 .address
= address
& PAGE_MASK
,
4051 .pgoff
= linear_page_index(vma
, address
),
4052 .gfp_mask
= __get_fault_gfp_mask(vma
),
4054 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4055 struct mm_struct
*mm
= vma
->vm_mm
;
4060 pgd
= pgd_offset(mm
, address
);
4061 p4d
= p4d_alloc(mm
, pgd
, address
);
4063 return VM_FAULT_OOM
;
4065 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4067 return VM_FAULT_OOM
;
4068 if (pud_none(*vmf
.pud
) && transparent_hugepage_enabled(vma
)) {
4069 ret
= create_huge_pud(&vmf
);
4070 if (!(ret
& VM_FAULT_FALLBACK
))
4073 pud_t orig_pud
= *vmf
.pud
;
4076 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4078 /* NUMA case for anonymous PUDs would go here */
4080 if (dirty
&& !pud_write(orig_pud
)) {
4081 ret
= wp_huge_pud(&vmf
, orig_pud
);
4082 if (!(ret
& VM_FAULT_FALLBACK
))
4085 huge_pud_set_accessed(&vmf
, orig_pud
);
4091 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4093 return VM_FAULT_OOM
;
4094 if (pmd_none(*vmf
.pmd
) && transparent_hugepage_enabled(vma
)) {
4095 ret
= create_huge_pmd(&vmf
);
4096 if (!(ret
& VM_FAULT_FALLBACK
))
4099 pmd_t orig_pmd
= *vmf
.pmd
;
4102 if (unlikely(is_swap_pmd(orig_pmd
))) {
4103 VM_BUG_ON(thp_migration_supported() &&
4104 !is_pmd_migration_entry(orig_pmd
));
4105 if (is_pmd_migration_entry(orig_pmd
))
4106 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4109 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4110 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4111 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4113 if (dirty
&& !pmd_write(orig_pmd
)) {
4114 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4115 if (!(ret
& VM_FAULT_FALLBACK
))
4118 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4124 return handle_pte_fault(&vmf
);
4128 * By the time we get here, we already hold the mm semaphore
4130 * The mmap_sem may have been released depending on flags and our
4131 * return value. See filemap_fault() and __lock_page_or_retry().
4133 int handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4138 __set_current_state(TASK_RUNNING
);
4140 count_vm_event(PGFAULT
);
4141 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4143 /* do counter updates before entering really critical section. */
4144 check_sync_rss_stat(current
);
4146 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4147 flags
& FAULT_FLAG_INSTRUCTION
,
4148 flags
& FAULT_FLAG_REMOTE
))
4149 return VM_FAULT_SIGSEGV
;
4152 * Enable the memcg OOM handling for faults triggered in user
4153 * space. Kernel faults are handled more gracefully.
4155 if (flags
& FAULT_FLAG_USER
)
4156 mem_cgroup_oom_enable();
4158 if (unlikely(is_vm_hugetlb_page(vma
)))
4159 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4161 ret
= __handle_mm_fault(vma
, address
, flags
);
4163 if (flags
& FAULT_FLAG_USER
) {
4164 mem_cgroup_oom_disable();
4166 * The task may have entered a memcg OOM situation but
4167 * if the allocation error was handled gracefully (no
4168 * VM_FAULT_OOM), there is no need to kill anything.
4169 * Just clean up the OOM state peacefully.
4171 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4172 mem_cgroup_oom_synchronize(false);
4177 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4179 #ifndef __PAGETABLE_P4D_FOLDED
4181 * Allocate p4d page table.
4182 * We've already handled the fast-path in-line.
4184 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4186 p4d_t
*new = p4d_alloc_one(mm
, address
);
4190 smp_wmb(); /* See comment in __pte_alloc */
4192 spin_lock(&mm
->page_table_lock
);
4193 if (pgd_present(*pgd
)) /* Another has populated it */
4196 pgd_populate(mm
, pgd
, new);
4197 spin_unlock(&mm
->page_table_lock
);
4200 #endif /* __PAGETABLE_P4D_FOLDED */
4202 #ifndef __PAGETABLE_PUD_FOLDED
4204 * Allocate page upper directory.
4205 * We've already handled the fast-path in-line.
4207 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4209 pud_t
*new = pud_alloc_one(mm
, address
);
4213 smp_wmb(); /* See comment in __pte_alloc */
4215 spin_lock(&mm
->page_table_lock
);
4216 #ifndef __ARCH_HAS_5LEVEL_HACK
4217 if (p4d_present(*p4d
)) /* Another has populated it */
4220 p4d_populate(mm
, p4d
, new);
4222 if (pgd_present(*p4d
)) /* Another has populated it */
4225 pgd_populate(mm
, p4d
, new);
4226 #endif /* __ARCH_HAS_5LEVEL_HACK */
4227 spin_unlock(&mm
->page_table_lock
);
4230 #endif /* __PAGETABLE_PUD_FOLDED */
4232 #ifndef __PAGETABLE_PMD_FOLDED
4234 * Allocate page middle directory.
4235 * We've already handled the fast-path in-line.
4237 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4240 pmd_t
*new = pmd_alloc_one(mm
, address
);
4244 smp_wmb(); /* See comment in __pte_alloc */
4246 ptl
= pud_lock(mm
, pud
);
4247 #ifndef __ARCH_HAS_4LEVEL_HACK
4248 if (!pud_present(*pud
)) {
4250 pud_populate(mm
, pud
, new);
4251 } else /* Another has populated it */
4254 if (!pgd_present(*pud
)) {
4256 pgd_populate(mm
, pud
, new);
4257 } else /* Another has populated it */
4259 #endif /* __ARCH_HAS_4LEVEL_HACK */
4263 #endif /* __PAGETABLE_PMD_FOLDED */
4265 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4266 unsigned long *start
, unsigned long *end
,
4267 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4275 pgd
= pgd_offset(mm
, address
);
4276 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4279 p4d
= p4d_offset(pgd
, address
);
4280 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4283 pud
= pud_offset(p4d
, address
);
4284 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4287 pmd
= pmd_offset(pud
, address
);
4288 VM_BUG_ON(pmd_trans_huge(*pmd
));
4290 if (pmd_huge(*pmd
)) {
4295 *start
= address
& PMD_MASK
;
4296 *end
= *start
+ PMD_SIZE
;
4297 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4299 *ptlp
= pmd_lock(mm
, pmd
);
4300 if (pmd_huge(*pmd
)) {
4306 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4309 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4313 *start
= address
& PAGE_MASK
;
4314 *end
= *start
+ PAGE_SIZE
;
4315 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4317 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4318 if (!pte_present(*ptep
))
4323 pte_unmap_unlock(ptep
, *ptlp
);
4325 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4330 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4331 pte_t
**ptepp
, spinlock_t
**ptlp
)
4335 /* (void) is needed to make gcc happy */
4336 (void) __cond_lock(*ptlp
,
4337 !(res
= __follow_pte_pmd(mm
, address
, NULL
, NULL
,
4338 ptepp
, NULL
, ptlp
)));
4342 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4343 unsigned long *start
, unsigned long *end
,
4344 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4348 /* (void) is needed to make gcc happy */
4349 (void) __cond_lock(*ptlp
,
4350 !(res
= __follow_pte_pmd(mm
, address
, start
, end
,
4351 ptepp
, pmdpp
, ptlp
)));
4354 EXPORT_SYMBOL(follow_pte_pmd
);
4357 * follow_pfn - look up PFN at a user virtual address
4358 * @vma: memory mapping
4359 * @address: user virtual address
4360 * @pfn: location to store found PFN
4362 * Only IO mappings and raw PFN mappings are allowed.
4364 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4366 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4373 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4376 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4379 *pfn
= pte_pfn(*ptep
);
4380 pte_unmap_unlock(ptep
, ptl
);
4383 EXPORT_SYMBOL(follow_pfn
);
4385 #ifdef CONFIG_HAVE_IOREMAP_PROT
4386 int follow_phys(struct vm_area_struct
*vma
,
4387 unsigned long address
, unsigned int flags
,
4388 unsigned long *prot
, resource_size_t
*phys
)
4394 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4397 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4401 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4404 *prot
= pgprot_val(pte_pgprot(pte
));
4405 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4409 pte_unmap_unlock(ptep
, ptl
);
4414 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4415 void *buf
, int len
, int write
)
4417 resource_size_t phys_addr
;
4418 unsigned long prot
= 0;
4419 void __iomem
*maddr
;
4420 int offset
= addr
& (PAGE_SIZE
-1);
4422 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4425 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4430 memcpy_toio(maddr
+ offset
, buf
, len
);
4432 memcpy_fromio(buf
, maddr
+ offset
, len
);
4437 EXPORT_SYMBOL_GPL(generic_access_phys
);
4441 * Access another process' address space as given in mm. If non-NULL, use the
4442 * given task for page fault accounting.
4444 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4445 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4447 struct vm_area_struct
*vma
;
4448 void *old_buf
= buf
;
4449 int write
= gup_flags
& FOLL_WRITE
;
4451 down_read(&mm
->mmap_sem
);
4452 /* ignore errors, just check how much was successfully transferred */
4454 int bytes
, ret
, offset
;
4456 struct page
*page
= NULL
;
4458 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4459 gup_flags
, &page
, &vma
, NULL
);
4461 #ifndef CONFIG_HAVE_IOREMAP_PROT
4465 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4466 * we can access using slightly different code.
4468 vma
= find_vma(mm
, addr
);
4469 if (!vma
|| vma
->vm_start
> addr
)
4471 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4472 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4480 offset
= addr
& (PAGE_SIZE
-1);
4481 if (bytes
> PAGE_SIZE
-offset
)
4482 bytes
= PAGE_SIZE
-offset
;
4486 copy_to_user_page(vma
, page
, addr
,
4487 maddr
+ offset
, buf
, bytes
);
4488 set_page_dirty_lock(page
);
4490 copy_from_user_page(vma
, page
, addr
,
4491 buf
, maddr
+ offset
, bytes
);
4500 up_read(&mm
->mmap_sem
);
4502 return buf
- old_buf
;
4506 * access_remote_vm - access another process' address space
4507 * @mm: the mm_struct of the target address space
4508 * @addr: start address to access
4509 * @buf: source or destination buffer
4510 * @len: number of bytes to transfer
4511 * @gup_flags: flags modifying lookup behaviour
4513 * The caller must hold a reference on @mm.
4515 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4516 void *buf
, int len
, unsigned int gup_flags
)
4518 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4522 * Access another process' address space.
4523 * Source/target buffer must be kernel space,
4524 * Do not walk the page table directly, use get_user_pages
4526 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4527 void *buf
, int len
, unsigned int gup_flags
)
4529 struct mm_struct
*mm
;
4532 mm
= get_task_mm(tsk
);
4536 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4542 EXPORT_SYMBOL_GPL(access_process_vm
);
4545 * Print the name of a VMA.
4547 void print_vma_addr(char *prefix
, unsigned long ip
)
4549 struct mm_struct
*mm
= current
->mm
;
4550 struct vm_area_struct
*vma
;
4553 * Do not print if we are in atomic
4554 * contexts (in exception stacks, etc.):
4556 if (preempt_count())
4559 down_read(&mm
->mmap_sem
);
4560 vma
= find_vma(mm
, ip
);
4561 if (vma
&& vma
->vm_file
) {
4562 struct file
*f
= vma
->vm_file
;
4563 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4567 p
= file_path(f
, buf
, PAGE_SIZE
);
4570 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4572 vma
->vm_end
- vma
->vm_start
);
4573 free_page((unsigned long)buf
);
4576 up_read(&mm
->mmap_sem
);
4579 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4580 void __might_fault(const char *file
, int line
)
4583 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4584 * holding the mmap_sem, this is safe because kernel memory doesn't
4585 * get paged out, therefore we'll never actually fault, and the
4586 * below annotations will generate false positives.
4588 if (uaccess_kernel())
4590 if (pagefault_disabled())
4592 __might_sleep(file
, line
, 0);
4593 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4595 might_lock_read(¤t
->mm
->mmap_sem
);
4598 EXPORT_SYMBOL(__might_fault
);
4601 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4602 static void clear_gigantic_page(struct page
*page
,
4604 unsigned int pages_per_huge_page
)
4607 struct page
*p
= page
;
4610 for (i
= 0; i
< pages_per_huge_page
;
4611 i
++, p
= mem_map_next(p
, page
, i
)) {
4613 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4616 void clear_huge_page(struct page
*page
,
4617 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4620 unsigned long addr
= addr_hint
&
4621 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4623 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4624 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4628 /* Clear sub-page to access last to keep its cache lines hot */
4630 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4631 if (2 * n
<= pages_per_huge_page
) {
4632 /* If sub-page to access in first half of huge page */
4635 /* Clear sub-pages at the end of huge page */
4636 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4638 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4641 /* If sub-page to access in second half of huge page */
4642 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4643 l
= pages_per_huge_page
- n
;
4644 /* Clear sub-pages at the begin of huge page */
4645 for (i
= 0; i
< base
; i
++) {
4647 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4651 * Clear remaining sub-pages in left-right-left-right pattern
4652 * towards the sub-page to access
4654 for (i
= 0; i
< l
; i
++) {
4655 int left_idx
= base
+ i
;
4656 int right_idx
= base
+ 2 * l
- 1 - i
;
4659 clear_user_highpage(page
+ left_idx
,
4660 addr
+ left_idx
* PAGE_SIZE
);
4662 clear_user_highpage(page
+ right_idx
,
4663 addr
+ right_idx
* PAGE_SIZE
);
4667 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4669 struct vm_area_struct
*vma
,
4670 unsigned int pages_per_huge_page
)
4673 struct page
*dst_base
= dst
;
4674 struct page
*src_base
= src
;
4676 for (i
= 0; i
< pages_per_huge_page
; ) {
4678 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4681 dst
= mem_map_next(dst
, dst_base
, i
);
4682 src
= mem_map_next(src
, src_base
, i
);
4686 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4687 unsigned long addr
, struct vm_area_struct
*vma
,
4688 unsigned int pages_per_huge_page
)
4692 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4693 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4694 pages_per_huge_page
);
4699 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4701 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4705 long copy_huge_page_from_user(struct page
*dst_page
,
4706 const void __user
*usr_src
,
4707 unsigned int pages_per_huge_page
,
4708 bool allow_pagefault
)
4710 void *src
= (void *)usr_src
;
4712 unsigned long i
, rc
= 0;
4713 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4715 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4716 if (allow_pagefault
)
4717 page_kaddr
= kmap(dst_page
+ i
);
4719 page_kaddr
= kmap_atomic(dst_page
+ i
);
4720 rc
= copy_from_user(page_kaddr
,
4721 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4723 if (allow_pagefault
)
4724 kunmap(dst_page
+ i
);
4726 kunmap_atomic(page_kaddr
);
4728 ret_val
-= (PAGE_SIZE
- rc
);
4736 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4738 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4740 static struct kmem_cache
*page_ptl_cachep
;
4742 void __init
ptlock_cache_init(void)
4744 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4748 bool ptlock_alloc(struct page
*page
)
4752 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4759 void ptlock_free(struct page
*page
)
4761 kmem_cache_free(page_ptl_cachep
, page
->ptl
);