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>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
84 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr
;
91 EXPORT_SYMBOL(max_mapnr
);
94 EXPORT_SYMBOL(mem_map
);
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 EXPORT_SYMBOL(high_memory
);
108 * Randomize the address space (stacks, mmaps, brk, etc.).
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
113 int randomize_va_space __read_mostly
=
114 #ifdef CONFIG_COMPAT_BRK
120 static int __init
disable_randmaps(char *s
)
122 randomize_va_space
= 0;
125 __setup("norandmaps", disable_randmaps
);
127 unsigned long zero_pfn __read_mostly
;
128 EXPORT_SYMBOL(zero_pfn
);
130 unsigned long highest_memmap_pfn __read_mostly
;
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 static int __init
init_zero_pfn(void)
137 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
140 core_initcall(init_zero_pfn
);
143 #if defined(SPLIT_RSS_COUNTING)
145 void sync_mm_rss(struct mm_struct
*mm
)
149 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
150 if (current
->rss_stat
.count
[i
]) {
151 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
152 current
->rss_stat
.count
[i
] = 0;
155 current
->rss_stat
.events
= 0;
158 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
160 struct task_struct
*task
= current
;
162 if (likely(task
->mm
== mm
))
163 task
->rss_stat
.count
[member
] += val
;
165 add_mm_counter(mm
, member
, val
);
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH (64)
172 static void check_sync_rss_stat(struct task_struct
*task
)
174 if (unlikely(task
!= current
))
176 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
177 sync_mm_rss(task
->mm
);
179 #else /* SPLIT_RSS_COUNTING */
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184 static void check_sync_rss_stat(struct task_struct
*task
)
188 #endif /* SPLIT_RSS_COUNTING */
190 #ifdef HAVE_GENERIC_MMU_GATHER
192 static bool tlb_next_batch(struct mmu_gather
*tlb
)
194 struct mmu_gather_batch
*batch
;
198 tlb
->active
= batch
->next
;
202 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
205 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
212 batch
->max
= MAX_GATHER_BATCH
;
214 tlb
->active
->next
= batch
;
220 void arch_tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
221 unsigned long start
, unsigned long end
)
225 /* Is it from 0 to ~0? */
226 tlb
->fullmm
= !(start
| (end
+1));
227 tlb
->need_flush_all
= 0;
228 tlb
->local
.next
= NULL
;
230 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
231 tlb
->active
= &tlb
->local
;
232 tlb
->batch_count
= 0;
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 __tlb_reset_range(tlb
);
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
248 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
249 __tlb_reset_range(tlb
);
252 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
254 struct mmu_gather_batch
*batch
;
256 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
257 tlb_table_flush(tlb
);
259 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
260 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
263 tlb
->active
= &tlb
->local
;
266 void tlb_flush_mmu(struct mmu_gather
*tlb
)
268 tlb_flush_mmu_tlbonly(tlb
);
269 tlb_flush_mmu_free(tlb
);
273 * Called at the end of the shootdown operation to free up any resources
274 * that were required.
276 void arch_tlb_finish_mmu(struct mmu_gather
*tlb
,
277 unsigned long start
, unsigned long end
, bool force
)
279 struct mmu_gather_batch
*batch
, *next
;
282 __tlb_adjust_range(tlb
, start
, end
- start
);
286 /* keep the page table cache within bounds */
289 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
291 free_pages((unsigned long)batch
, 0);
293 tlb
->local
.next
= NULL
;
297 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298 * handling the additional races in SMP caused by other CPUs caching valid
299 * mappings in their TLBs. Returns the number of free page slots left.
300 * When out of page slots we must call tlb_flush_mmu().
301 *returns true if the caller should flush.
303 bool __tlb_remove_page_size(struct mmu_gather
*tlb
, struct page
*page
, int page_size
)
305 struct mmu_gather_batch
*batch
;
307 VM_BUG_ON(!tlb
->end
);
308 VM_WARN_ON(tlb
->page_size
!= page_size
);
312 * Add the page and check if we are full. If so
315 batch
->pages
[batch
->nr
++] = page
;
316 if (batch
->nr
== batch
->max
) {
317 if (!tlb_next_batch(tlb
))
321 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
326 #endif /* HAVE_GENERIC_MMU_GATHER */
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
331 * See the comment near struct mmu_table_batch.
335 * If we want tlb_remove_table() to imply TLB invalidates.
337 static inline void tlb_table_invalidate(struct mmu_gather
*tlb
)
339 #ifdef CONFIG_HAVE_RCU_TABLE_INVALIDATE
341 * Invalidate page-table caches used by hardware walkers. Then we still
342 * need to RCU-sched wait while freeing the pages because software
343 * walkers can still be in-flight.
345 tlb_flush_mmu_tlbonly(tlb
);
349 static void tlb_remove_table_smp_sync(void *arg
)
351 /* Simply deliver the interrupt */
354 static void tlb_remove_table_one(void *table
)
357 * This isn't an RCU grace period and hence the page-tables cannot be
358 * assumed to be actually RCU-freed.
360 * It is however sufficient for software page-table walkers that rely on
361 * IRQ disabling. See the comment near struct mmu_table_batch.
363 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
364 __tlb_remove_table(table
);
367 static void tlb_remove_table_rcu(struct rcu_head
*head
)
369 struct mmu_table_batch
*batch
;
372 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
374 for (i
= 0; i
< batch
->nr
; i
++)
375 __tlb_remove_table(batch
->tables
[i
]);
377 free_page((unsigned long)batch
);
380 void tlb_table_flush(struct mmu_gather
*tlb
)
382 struct mmu_table_batch
**batch
= &tlb
->batch
;
385 tlb_table_invalidate(tlb
);
386 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
391 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
393 struct mmu_table_batch
**batch
= &tlb
->batch
;
395 if (*batch
== NULL
) {
396 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
397 if (*batch
== NULL
) {
398 tlb_table_invalidate(tlb
);
399 tlb_remove_table_one(table
);
405 (*batch
)->tables
[(*batch
)->nr
++] = table
;
406 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
407 tlb_table_flush(tlb
);
410 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
413 * Called to initialize an (on-stack) mmu_gather structure for page-table
414 * tear-down from @mm. The @fullmm argument is used when @mm is without
415 * users and we're going to destroy the full address space (exit/execve).
417 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
418 unsigned long start
, unsigned long end
)
420 arch_tlb_gather_mmu(tlb
, mm
, start
, end
);
421 inc_tlb_flush_pending(tlb
->mm
);
424 void tlb_finish_mmu(struct mmu_gather
*tlb
,
425 unsigned long start
, unsigned long end
)
428 * If there are parallel threads are doing PTE changes on same range
429 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
430 * flush by batching, a thread has stable TLB entry can fail to flush
431 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
432 * forcefully if we detect parallel PTE batching threads.
434 bool force
= mm_tlb_flush_nested(tlb
->mm
);
436 arch_tlb_finish_mmu(tlb
, start
, end
, force
);
437 dec_tlb_flush_pending(tlb
->mm
);
441 * Note: this doesn't free the actual pages themselves. That
442 * has been handled earlier when unmapping all the memory regions.
444 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
447 pgtable_t token
= pmd_pgtable(*pmd
);
449 pte_free_tlb(tlb
, token
, addr
);
450 mm_dec_nr_ptes(tlb
->mm
);
453 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
454 unsigned long addr
, unsigned long end
,
455 unsigned long floor
, unsigned long ceiling
)
462 pmd
= pmd_offset(pud
, addr
);
464 next
= pmd_addr_end(addr
, end
);
465 if (pmd_none_or_clear_bad(pmd
))
467 free_pte_range(tlb
, pmd
, addr
);
468 } while (pmd
++, addr
= next
, addr
!= end
);
478 if (end
- 1 > ceiling
- 1)
481 pmd
= pmd_offset(pud
, start
);
483 pmd_free_tlb(tlb
, pmd
, start
);
484 mm_dec_nr_pmds(tlb
->mm
);
487 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
488 unsigned long addr
, unsigned long end
,
489 unsigned long floor
, unsigned long ceiling
)
496 pud
= pud_offset(p4d
, addr
);
498 next
= pud_addr_end(addr
, end
);
499 if (pud_none_or_clear_bad(pud
))
501 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
502 } while (pud
++, addr
= next
, addr
!= end
);
512 if (end
- 1 > ceiling
- 1)
515 pud
= pud_offset(p4d
, start
);
517 pud_free_tlb(tlb
, pud
, start
);
518 mm_dec_nr_puds(tlb
->mm
);
521 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
522 unsigned long addr
, unsigned long end
,
523 unsigned long floor
, unsigned long ceiling
)
530 p4d
= p4d_offset(pgd
, addr
);
532 next
= p4d_addr_end(addr
, end
);
533 if (p4d_none_or_clear_bad(p4d
))
535 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
536 } while (p4d
++, addr
= next
, addr
!= end
);
542 ceiling
&= PGDIR_MASK
;
546 if (end
- 1 > ceiling
- 1)
549 p4d
= p4d_offset(pgd
, start
);
551 p4d_free_tlb(tlb
, p4d
, start
);
555 * This function frees user-level page tables of a process.
557 void free_pgd_range(struct mmu_gather
*tlb
,
558 unsigned long addr
, unsigned long end
,
559 unsigned long floor
, unsigned long ceiling
)
565 * The next few lines have given us lots of grief...
567 * Why are we testing PMD* at this top level? Because often
568 * there will be no work to do at all, and we'd prefer not to
569 * go all the way down to the bottom just to discover that.
571 * Why all these "- 1"s? Because 0 represents both the bottom
572 * of the address space and the top of it (using -1 for the
573 * top wouldn't help much: the masks would do the wrong thing).
574 * The rule is that addr 0 and floor 0 refer to the bottom of
575 * the address space, but end 0 and ceiling 0 refer to the top
576 * Comparisons need to use "end - 1" and "ceiling - 1" (though
577 * that end 0 case should be mythical).
579 * Wherever addr is brought up or ceiling brought down, we must
580 * be careful to reject "the opposite 0" before it confuses the
581 * subsequent tests. But what about where end is brought down
582 * by PMD_SIZE below? no, end can't go down to 0 there.
584 * Whereas we round start (addr) and ceiling down, by different
585 * masks at different levels, in order to test whether a table
586 * now has no other vmas using it, so can be freed, we don't
587 * bother to round floor or end up - the tests don't need that.
601 if (end
- 1 > ceiling
- 1)
606 * We add page table cache pages with PAGE_SIZE,
607 * (see pte_free_tlb()), flush the tlb if we need
609 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
610 pgd
= pgd_offset(tlb
->mm
, addr
);
612 next
= pgd_addr_end(addr
, end
);
613 if (pgd_none_or_clear_bad(pgd
))
615 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
616 } while (pgd
++, addr
= next
, addr
!= end
);
619 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
620 unsigned long floor
, unsigned long ceiling
)
623 struct vm_area_struct
*next
= vma
->vm_next
;
624 unsigned long addr
= vma
->vm_start
;
627 * Hide vma from rmap and truncate_pagecache before freeing
630 unlink_anon_vmas(vma
);
631 unlink_file_vma(vma
);
633 if (is_vm_hugetlb_page(vma
)) {
634 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
635 floor
, next
? next
->vm_start
: ceiling
);
638 * Optimization: gather nearby vmas into one call down
640 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
641 && !is_vm_hugetlb_page(next
)) {
644 unlink_anon_vmas(vma
);
645 unlink_file_vma(vma
);
647 free_pgd_range(tlb
, addr
, vma
->vm_end
,
648 floor
, next
? next
->vm_start
: ceiling
);
654 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
657 pgtable_t
new = pte_alloc_one(mm
, address
);
662 * Ensure all pte setup (eg. pte page lock and page clearing) are
663 * visible before the pte is made visible to other CPUs by being
664 * put into page tables.
666 * The other side of the story is the pointer chasing in the page
667 * table walking code (when walking the page table without locking;
668 * ie. most of the time). Fortunately, these data accesses consist
669 * of a chain of data-dependent loads, meaning most CPUs (alpha
670 * being the notable exception) will already guarantee loads are
671 * seen in-order. See the alpha page table accessors for the
672 * smp_read_barrier_depends() barriers in page table walking code.
674 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
676 ptl
= pmd_lock(mm
, pmd
);
677 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
679 pmd_populate(mm
, pmd
, new);
688 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
690 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
694 smp_wmb(); /* See comment in __pte_alloc */
696 spin_lock(&init_mm
.page_table_lock
);
697 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
698 pmd_populate_kernel(&init_mm
, pmd
, new);
701 spin_unlock(&init_mm
.page_table_lock
);
703 pte_free_kernel(&init_mm
, new);
707 static inline void init_rss_vec(int *rss
)
709 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
712 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
716 if (current
->mm
== mm
)
718 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
720 add_mm_counter(mm
, i
, rss
[i
]);
724 * This function is called to print an error when a bad pte
725 * is found. For example, we might have a PFN-mapped pte in
726 * a region that doesn't allow it.
728 * The calling function must still handle the error.
730 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
731 pte_t pte
, struct page
*page
)
733 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
734 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
735 pud_t
*pud
= pud_offset(p4d
, addr
);
736 pmd_t
*pmd
= pmd_offset(pud
, addr
);
737 struct address_space
*mapping
;
739 static unsigned long resume
;
740 static unsigned long nr_shown
;
741 static unsigned long nr_unshown
;
744 * Allow a burst of 60 reports, then keep quiet for that minute;
745 * or allow a steady drip of one report per second.
747 if (nr_shown
== 60) {
748 if (time_before(jiffies
, resume
)) {
753 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
760 resume
= jiffies
+ 60 * HZ
;
762 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
763 index
= linear_page_index(vma
, addr
);
765 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
767 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
769 dump_page(page
, "bad pte");
770 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
771 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
773 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
775 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
777 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
778 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
779 mapping
? mapping
->a_ops
->readpage
: NULL
);
781 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
785 * vm_normal_page -- This function gets the "struct page" associated with a pte.
787 * "Special" mappings do not wish to be associated with a "struct page" (either
788 * it doesn't exist, or it exists but they don't want to touch it). In this
789 * case, NULL is returned here. "Normal" mappings do have a struct page.
791 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
792 * pte bit, in which case this function is trivial. Secondly, an architecture
793 * may not have a spare pte bit, which requires a more complicated scheme,
796 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
797 * special mapping (even if there are underlying and valid "struct pages").
798 * COWed pages of a VM_PFNMAP are always normal.
800 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
801 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
802 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
803 * mapping will always honor the rule
805 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
807 * And for normal mappings this is false.
809 * This restricts such mappings to be a linear translation from virtual address
810 * to pfn. To get around this restriction, we allow arbitrary mappings so long
811 * as the vma is not a COW mapping; in that case, we know that all ptes are
812 * special (because none can have been COWed).
815 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
817 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
818 * page" backing, however the difference is that _all_ pages with a struct
819 * page (that is, those where pfn_valid is true) are refcounted and considered
820 * normal pages by the VM. The disadvantage is that pages are refcounted
821 * (which can be slower and simply not an option for some PFNMAP users). The
822 * advantage is that we don't have to follow the strict linearity rule of
823 * PFNMAP mappings in order to support COWable mappings.
826 #ifdef __HAVE_ARCH_PTE_SPECIAL
827 # define HAVE_PTE_SPECIAL 1
829 # define HAVE_PTE_SPECIAL 0
831 struct page
*_vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
832 pte_t pte
, bool with_public_device
)
834 unsigned long pfn
= pte_pfn(pte
);
836 if (HAVE_PTE_SPECIAL
) {
837 if (likely(!pte_special(pte
)))
839 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
840 return vma
->vm_ops
->find_special_page(vma
, addr
);
841 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
843 if (is_zero_pfn(pfn
))
847 * Device public pages are special pages (they are ZONE_DEVICE
848 * pages but different from persistent memory). They behave
849 * allmost like normal pages. The difference is that they are
850 * not on the lru and thus should never be involve with any-
851 * thing that involve lru manipulation (mlock, numa balancing,
854 * This is why we still want to return NULL for such page from
855 * vm_normal_page() so that we do not have to special case all
856 * call site of vm_normal_page().
858 if (likely(pfn
<= highest_memmap_pfn
)) {
859 struct page
*page
= pfn_to_page(pfn
);
861 if (is_device_public_page(page
)) {
862 if (with_public_device
)
867 print_bad_pte(vma
, addr
, pte
, NULL
);
871 /* !HAVE_PTE_SPECIAL case follows: */
873 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
874 if (vma
->vm_flags
& VM_MIXEDMAP
) {
880 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
881 if (pfn
== vma
->vm_pgoff
+ off
)
883 if (!is_cow_mapping(vma
->vm_flags
))
888 if (is_zero_pfn(pfn
))
891 if (unlikely(pfn
> highest_memmap_pfn
)) {
892 print_bad_pte(vma
, addr
, pte
, NULL
);
897 * NOTE! We still have PageReserved() pages in the page tables.
898 * eg. VDSO mappings can cause them to exist.
901 return pfn_to_page(pfn
);
904 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
905 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
908 unsigned long pfn
= pmd_pfn(pmd
);
911 * There is no pmd_special() but there may be special pmds, e.g.
912 * in a direct-access (dax) mapping, so let's just replicate the
913 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
915 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
916 if (vma
->vm_flags
& VM_MIXEDMAP
) {
922 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
923 if (pfn
== vma
->vm_pgoff
+ off
)
925 if (!is_cow_mapping(vma
->vm_flags
))
930 if (is_zero_pfn(pfn
))
932 if (unlikely(pfn
> highest_memmap_pfn
))
936 * NOTE! We still have PageReserved() pages in the page tables.
937 * eg. VDSO mappings can cause them to exist.
940 return pfn_to_page(pfn
);
945 * copy one vm_area from one task to the other. Assumes the page tables
946 * already present in the new task to be cleared in the whole range
947 * covered by this vma.
950 static inline unsigned long
951 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
952 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
953 unsigned long addr
, int *rss
)
955 unsigned long vm_flags
= vma
->vm_flags
;
956 pte_t pte
= *src_pte
;
959 /* pte contains position in swap or file, so copy. */
960 if (unlikely(!pte_present(pte
))) {
961 swp_entry_t entry
= pte_to_swp_entry(pte
);
963 if (likely(!non_swap_entry(entry
))) {
964 if (swap_duplicate(entry
) < 0)
967 /* make sure dst_mm is on swapoff's mmlist. */
968 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
969 spin_lock(&mmlist_lock
);
970 if (list_empty(&dst_mm
->mmlist
))
971 list_add(&dst_mm
->mmlist
,
973 spin_unlock(&mmlist_lock
);
976 } else if (is_migration_entry(entry
)) {
977 page
= migration_entry_to_page(entry
);
979 rss
[mm_counter(page
)]++;
981 if (is_write_migration_entry(entry
) &&
982 is_cow_mapping(vm_flags
)) {
984 * COW mappings require pages in both
985 * parent and child to be set to read.
987 make_migration_entry_read(&entry
);
988 pte
= swp_entry_to_pte(entry
);
989 if (pte_swp_soft_dirty(*src_pte
))
990 pte
= pte_swp_mksoft_dirty(pte
);
991 set_pte_at(src_mm
, addr
, src_pte
, pte
);
993 } else if (is_device_private_entry(entry
)) {
994 page
= device_private_entry_to_page(entry
);
997 * Update rss count even for unaddressable pages, as
998 * they should treated just like normal pages in this
1001 * We will likely want to have some new rss counters
1002 * for unaddressable pages, at some point. But for now
1003 * keep things as they are.
1006 rss
[mm_counter(page
)]++;
1007 page_dup_rmap(page
, false);
1010 * We do not preserve soft-dirty information, because so
1011 * far, checkpoint/restore is the only feature that
1012 * requires that. And checkpoint/restore does not work
1013 * when a device driver is involved (you cannot easily
1014 * save and restore device driver state).
1016 if (is_write_device_private_entry(entry
) &&
1017 is_cow_mapping(vm_flags
)) {
1018 make_device_private_entry_read(&entry
);
1019 pte
= swp_entry_to_pte(entry
);
1020 set_pte_at(src_mm
, addr
, src_pte
, pte
);
1027 * If it's a COW mapping, write protect it both
1028 * in the parent and the child
1030 if (is_cow_mapping(vm_flags
)) {
1031 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
1032 pte
= pte_wrprotect(pte
);
1036 * If it's a shared mapping, mark it clean in
1039 if (vm_flags
& VM_SHARED
)
1040 pte
= pte_mkclean(pte
);
1041 pte
= pte_mkold(pte
);
1043 page
= vm_normal_page(vma
, addr
, pte
);
1046 page_dup_rmap(page
, false);
1047 rss
[mm_counter(page
)]++;
1048 } else if (pte_devmap(pte
)) {
1049 page
= pte_page(pte
);
1052 * Cache coherent device memory behave like regular page and
1053 * not like persistent memory page. For more informations see
1054 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1056 if (is_device_public_page(page
)) {
1058 page_dup_rmap(page
, false);
1059 rss
[mm_counter(page
)]++;
1064 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
1068 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1069 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
1070 unsigned long addr
, unsigned long end
)
1072 pte_t
*orig_src_pte
, *orig_dst_pte
;
1073 pte_t
*src_pte
, *dst_pte
;
1074 spinlock_t
*src_ptl
, *dst_ptl
;
1076 int rss
[NR_MM_COUNTERS
];
1077 swp_entry_t entry
= (swp_entry_t
){0};
1082 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1085 src_pte
= pte_offset_map(src_pmd
, addr
);
1086 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
1087 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1088 orig_src_pte
= src_pte
;
1089 orig_dst_pte
= dst_pte
;
1090 arch_enter_lazy_mmu_mode();
1094 * We are holding two locks at this point - either of them
1095 * could generate latencies in another task on another CPU.
1097 if (progress
>= 32) {
1099 if (need_resched() ||
1100 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1103 if (pte_none(*src_pte
)) {
1107 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
1112 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1114 arch_leave_lazy_mmu_mode();
1115 spin_unlock(src_ptl
);
1116 pte_unmap(orig_src_pte
);
1117 add_mm_rss_vec(dst_mm
, rss
);
1118 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1122 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
1131 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1132 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
1133 unsigned long addr
, unsigned long end
)
1135 pmd_t
*src_pmd
, *dst_pmd
;
1138 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1141 src_pmd
= pmd_offset(src_pud
, addr
);
1143 next
= pmd_addr_end(addr
, end
);
1144 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1145 || pmd_devmap(*src_pmd
)) {
1147 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, vma
);
1148 err
= copy_huge_pmd(dst_mm
, src_mm
,
1149 dst_pmd
, src_pmd
, addr
, vma
);
1156 if (pmd_none_or_clear_bad(src_pmd
))
1158 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1161 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1165 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1166 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
1167 unsigned long addr
, unsigned long end
)
1169 pud_t
*src_pud
, *dst_pud
;
1172 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1175 src_pud
= pud_offset(src_p4d
, addr
);
1177 next
= pud_addr_end(addr
, end
);
1178 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1181 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
1182 err
= copy_huge_pud(dst_mm
, src_mm
,
1183 dst_pud
, src_pud
, addr
, vma
);
1190 if (pud_none_or_clear_bad(src_pud
))
1192 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1195 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1199 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1200 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1201 unsigned long addr
, unsigned long end
)
1203 p4d_t
*src_p4d
, *dst_p4d
;
1206 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1209 src_p4d
= p4d_offset(src_pgd
, addr
);
1211 next
= p4d_addr_end(addr
, end
);
1212 if (p4d_none_or_clear_bad(src_p4d
))
1214 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
1217 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1221 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1222 struct vm_area_struct
*vma
)
1224 pgd_t
*src_pgd
, *dst_pgd
;
1226 unsigned long addr
= vma
->vm_start
;
1227 unsigned long end
= vma
->vm_end
;
1228 unsigned long mmun_start
; /* For mmu_notifiers */
1229 unsigned long mmun_end
; /* For mmu_notifiers */
1234 * Don't copy ptes where a page fault will fill them correctly.
1235 * Fork becomes much lighter when there are big shared or private
1236 * readonly mappings. The tradeoff is that copy_page_range is more
1237 * efficient than faulting.
1239 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1243 if (is_vm_hugetlb_page(vma
))
1244 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1246 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1248 * We do not free on error cases below as remove_vma
1249 * gets called on error from higher level routine
1251 ret
= track_pfn_copy(vma
);
1257 * We need to invalidate the secondary MMU mappings only when
1258 * there could be a permission downgrade on the ptes of the
1259 * parent mm. And a permission downgrade will only happen if
1260 * is_cow_mapping() returns true.
1262 is_cow
= is_cow_mapping(vma
->vm_flags
);
1266 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1270 dst_pgd
= pgd_offset(dst_mm
, addr
);
1271 src_pgd
= pgd_offset(src_mm
, addr
);
1273 next
= pgd_addr_end(addr
, end
);
1274 if (pgd_none_or_clear_bad(src_pgd
))
1276 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1277 vma
, addr
, next
))) {
1281 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1284 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1288 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1289 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1290 unsigned long addr
, unsigned long end
,
1291 struct zap_details
*details
)
1293 struct mm_struct
*mm
= tlb
->mm
;
1294 int force_flush
= 0;
1295 int rss
[NR_MM_COUNTERS
];
1301 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
1304 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1306 flush_tlb_batched_pending(mm
);
1307 arch_enter_lazy_mmu_mode();
1310 if (pte_none(ptent
))
1313 if (pte_present(ptent
)) {
1316 page
= _vm_normal_page(vma
, addr
, ptent
, true);
1317 if (unlikely(details
) && page
) {
1319 * unmap_shared_mapping_pages() wants to
1320 * invalidate cache without truncating:
1321 * unmap shared but keep private pages.
1323 if (details
->check_mapping
&&
1324 details
->check_mapping
!= page_rmapping(page
))
1327 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1329 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1330 if (unlikely(!page
))
1333 if (!PageAnon(page
)) {
1334 if (pte_dirty(ptent
)) {
1336 set_page_dirty(page
);
1338 if (pte_young(ptent
) &&
1339 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1340 mark_page_accessed(page
);
1342 rss
[mm_counter(page
)]--;
1343 page_remove_rmap(page
, false);
1344 if (unlikely(page_mapcount(page
) < 0))
1345 print_bad_pte(vma
, addr
, ptent
, page
);
1346 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1354 entry
= pte_to_swp_entry(ptent
);
1355 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1356 struct page
*page
= device_private_entry_to_page(entry
);
1358 if (unlikely(details
&& details
->check_mapping
)) {
1360 * unmap_shared_mapping_pages() wants to
1361 * invalidate cache without truncating:
1362 * unmap shared but keep private pages.
1364 if (details
->check_mapping
!=
1365 page_rmapping(page
))
1369 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1370 rss
[mm_counter(page
)]--;
1371 page_remove_rmap(page
, false);
1376 /* If details->check_mapping, we leave swap entries. */
1377 if (unlikely(details
))
1380 entry
= pte_to_swp_entry(ptent
);
1381 if (!non_swap_entry(entry
))
1383 else if (is_migration_entry(entry
)) {
1386 page
= migration_entry_to_page(entry
);
1387 rss
[mm_counter(page
)]--;
1389 if (unlikely(!free_swap_and_cache(entry
)))
1390 print_bad_pte(vma
, addr
, ptent
, NULL
);
1391 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1392 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1394 add_mm_rss_vec(mm
, rss
);
1395 arch_leave_lazy_mmu_mode();
1397 /* Do the actual TLB flush before dropping ptl */
1399 tlb_flush_mmu_tlbonly(tlb
);
1400 pte_unmap_unlock(start_pte
, ptl
);
1403 * If we forced a TLB flush (either due to running out of
1404 * batch buffers or because we needed to flush dirty TLB
1405 * entries before releasing the ptl), free the batched
1406 * memory too. Restart if we didn't do everything.
1410 tlb_flush_mmu_free(tlb
);
1418 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1419 struct vm_area_struct
*vma
, pud_t
*pud
,
1420 unsigned long addr
, unsigned long end
,
1421 struct zap_details
*details
)
1426 pmd
= pmd_offset(pud
, addr
);
1428 next
= pmd_addr_end(addr
, end
);
1429 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1430 if (next
- addr
!= HPAGE_PMD_SIZE
)
1431 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1432 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1437 * Here there can be other concurrent MADV_DONTNEED or
1438 * trans huge page faults running, and if the pmd is
1439 * none or trans huge it can change under us. This is
1440 * because MADV_DONTNEED holds the mmap_sem in read
1443 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1445 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1448 } while (pmd
++, addr
= next
, addr
!= end
);
1453 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1454 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1455 unsigned long addr
, unsigned long end
,
1456 struct zap_details
*details
)
1461 pud
= pud_offset(p4d
, addr
);
1463 next
= pud_addr_end(addr
, end
);
1464 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1465 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1466 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1467 split_huge_pud(vma
, pud
, addr
);
1468 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1472 if (pud_none_or_clear_bad(pud
))
1474 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1477 } while (pud
++, addr
= next
, addr
!= end
);
1482 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1483 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1484 unsigned long addr
, unsigned long end
,
1485 struct zap_details
*details
)
1490 p4d
= p4d_offset(pgd
, addr
);
1492 next
= p4d_addr_end(addr
, end
);
1493 if (p4d_none_or_clear_bad(p4d
))
1495 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1496 } while (p4d
++, addr
= next
, addr
!= end
);
1501 void unmap_page_range(struct mmu_gather
*tlb
,
1502 struct vm_area_struct
*vma
,
1503 unsigned long addr
, unsigned long end
,
1504 struct zap_details
*details
)
1509 BUG_ON(addr
>= end
);
1510 tlb_start_vma(tlb
, vma
);
1511 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1513 next
= pgd_addr_end(addr
, end
);
1514 if (pgd_none_or_clear_bad(pgd
))
1516 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1517 } while (pgd
++, addr
= next
, addr
!= end
);
1518 tlb_end_vma(tlb
, vma
);
1522 static void unmap_single_vma(struct mmu_gather
*tlb
,
1523 struct vm_area_struct
*vma
, unsigned long start_addr
,
1524 unsigned long end_addr
,
1525 struct zap_details
*details
)
1527 unsigned long start
= max(vma
->vm_start
, start_addr
);
1530 if (start
>= vma
->vm_end
)
1532 end
= min(vma
->vm_end
, end_addr
);
1533 if (end
<= vma
->vm_start
)
1537 uprobe_munmap(vma
, start
, end
);
1539 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1540 untrack_pfn(vma
, 0, 0);
1543 if (unlikely(is_vm_hugetlb_page(vma
))) {
1545 * It is undesirable to test vma->vm_file as it
1546 * should be non-null for valid hugetlb area.
1547 * However, vm_file will be NULL in the error
1548 * cleanup path of mmap_region. When
1549 * hugetlbfs ->mmap method fails,
1550 * mmap_region() nullifies vma->vm_file
1551 * before calling this function to clean up.
1552 * Since no pte has actually been setup, it is
1553 * safe to do nothing in this case.
1556 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1557 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1558 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1561 unmap_page_range(tlb
, vma
, start
, end
, details
);
1566 * unmap_vmas - unmap a range of memory covered by a list of vma's
1567 * @tlb: address of the caller's struct mmu_gather
1568 * @vma: the starting vma
1569 * @start_addr: virtual address at which to start unmapping
1570 * @end_addr: virtual address at which to end unmapping
1572 * Unmap all pages in the vma list.
1574 * Only addresses between `start' and `end' will be unmapped.
1576 * The VMA list must be sorted in ascending virtual address order.
1578 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1579 * range after unmap_vmas() returns. So the only responsibility here is to
1580 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1581 * drops the lock and schedules.
1583 void unmap_vmas(struct mmu_gather
*tlb
,
1584 struct vm_area_struct
*vma
, unsigned long start_addr
,
1585 unsigned long end_addr
)
1587 struct mm_struct
*mm
= vma
->vm_mm
;
1589 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1590 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1591 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1592 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1596 * zap_page_range - remove user pages in a given range
1597 * @vma: vm_area_struct holding the applicable pages
1598 * @start: starting address of pages to zap
1599 * @size: number of bytes to zap
1601 * Caller must protect the VMA list
1603 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1606 struct mm_struct
*mm
= vma
->vm_mm
;
1607 struct mmu_gather tlb
;
1608 unsigned long end
= start
+ size
;
1611 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1612 update_hiwater_rss(mm
);
1613 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1614 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
1615 unmap_single_vma(&tlb
, vma
, start
, end
, NULL
);
1618 * zap_page_range does not specify whether mmap_sem should be
1619 * held for read or write. That allows parallel zap_page_range
1620 * operations to unmap a PTE and defer a flush meaning that
1621 * this call observes pte_none and fails to flush the TLB.
1622 * Rather than adding a complex API, ensure that no stale
1623 * TLB entries exist when this call returns.
1625 flush_tlb_range(vma
, start
, end
);
1628 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1629 tlb_finish_mmu(&tlb
, start
, end
);
1633 * zap_page_range_single - remove user pages in a given range
1634 * @vma: vm_area_struct holding the applicable pages
1635 * @address: starting address of pages to zap
1636 * @size: number of bytes to zap
1637 * @details: details of shared cache invalidation
1639 * The range must fit into one VMA.
1641 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1642 unsigned long size
, struct zap_details
*details
)
1644 struct mm_struct
*mm
= vma
->vm_mm
;
1645 struct mmu_gather tlb
;
1646 unsigned long end
= address
+ size
;
1649 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1650 update_hiwater_rss(mm
);
1651 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1652 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1653 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1654 tlb_finish_mmu(&tlb
, address
, end
);
1658 * zap_vma_ptes - remove ptes mapping the vma
1659 * @vma: vm_area_struct holding ptes to be zapped
1660 * @address: starting address of pages to zap
1661 * @size: number of bytes to zap
1663 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1665 * The entire address range must be fully contained within the vma.
1667 * Returns 0 if successful.
1669 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1672 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1673 !(vma
->vm_flags
& VM_PFNMAP
))
1675 zap_page_range_single(vma
, address
, size
, NULL
);
1678 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1680 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1688 pgd
= pgd_offset(mm
, addr
);
1689 p4d
= p4d_alloc(mm
, pgd
, addr
);
1692 pud
= pud_alloc(mm
, p4d
, addr
);
1695 pmd
= pmd_alloc(mm
, pud
, addr
);
1699 VM_BUG_ON(pmd_trans_huge(*pmd
));
1700 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1704 * This is the old fallback for page remapping.
1706 * For historical reasons, it only allows reserved pages. Only
1707 * old drivers should use this, and they needed to mark their
1708 * pages reserved for the old functions anyway.
1710 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1711 struct page
*page
, pgprot_t prot
)
1713 struct mm_struct
*mm
= vma
->vm_mm
;
1722 flush_dcache_page(page
);
1723 pte
= get_locked_pte(mm
, addr
, &ptl
);
1727 if (!pte_none(*pte
))
1730 /* Ok, finally just insert the thing.. */
1732 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1733 page_add_file_rmap(page
, false);
1734 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1737 pte_unmap_unlock(pte
, ptl
);
1740 pte_unmap_unlock(pte
, ptl
);
1746 * vm_insert_page - insert single page into user vma
1747 * @vma: user vma to map to
1748 * @addr: target user address of this page
1749 * @page: source kernel page
1751 * This allows drivers to insert individual pages they've allocated
1754 * The page has to be a nice clean _individual_ kernel allocation.
1755 * If you allocate a compound page, you need to have marked it as
1756 * such (__GFP_COMP), or manually just split the page up yourself
1757 * (see split_page()).
1759 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1760 * took an arbitrary page protection parameter. This doesn't allow
1761 * that. Your vma protection will have to be set up correctly, which
1762 * means that if you want a shared writable mapping, you'd better
1763 * ask for a shared writable mapping!
1765 * The page does not need to be reserved.
1767 * Usually this function is called from f_op->mmap() handler
1768 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1769 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1770 * function from other places, for example from page-fault handler.
1772 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1775 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1777 if (!page_count(page
))
1779 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1780 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1781 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1782 vma
->vm_flags
|= VM_MIXEDMAP
;
1784 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1786 EXPORT_SYMBOL(vm_insert_page
);
1788 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1789 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1791 struct mm_struct
*mm
= vma
->vm_mm
;
1797 pte
= get_locked_pte(mm
, addr
, &ptl
);
1801 if (!pte_none(*pte
)) {
1804 * For read faults on private mappings the PFN passed
1805 * in may not match the PFN we have mapped if the
1806 * mapped PFN is a writeable COW page. In the mkwrite
1807 * case we are creating a writable PTE for a shared
1808 * mapping and we expect the PFNs to match.
1810 if (WARN_ON_ONCE(pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)))
1818 /* Ok, finally just insert the thing.. */
1819 if (pfn_t_devmap(pfn
))
1820 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1822 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1826 entry
= pte_mkyoung(entry
);
1827 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1830 set_pte_at(mm
, addr
, pte
, entry
);
1831 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1835 pte_unmap_unlock(pte
, ptl
);
1841 * vm_insert_pfn - insert single pfn into user vma
1842 * @vma: user vma to map to
1843 * @addr: target user address of this page
1844 * @pfn: source kernel pfn
1846 * Similar to vm_insert_page, this allows drivers to insert individual pages
1847 * they've allocated into a user vma. Same comments apply.
1849 * This function should only be called from a vm_ops->fault handler, and
1850 * in that case the handler should return NULL.
1852 * vma cannot be a COW mapping.
1854 * As this is called only for pages that do not currently exist, we
1855 * do not need to flush old virtual caches or the TLB.
1857 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1860 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1862 EXPORT_SYMBOL(vm_insert_pfn
);
1865 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1866 * @vma: user vma to map to
1867 * @addr: target user address of this page
1868 * @pfn: source kernel pfn
1869 * @pgprot: pgprot flags for the inserted page
1871 * This is exactly like vm_insert_pfn, except that it allows drivers to
1872 * to override pgprot on a per-page basis.
1874 * This only makes sense for IO mappings, and it makes no sense for
1875 * cow mappings. In general, using multiple vmas is preferable;
1876 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1879 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1880 unsigned long pfn
, pgprot_t pgprot
)
1884 * Technically, architectures with pte_special can avoid all these
1885 * restrictions (same for remap_pfn_range). However we would like
1886 * consistency in testing and feature parity among all, so we should
1887 * try to keep these invariants in place for everybody.
1889 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1890 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1891 (VM_PFNMAP
|VM_MIXEDMAP
));
1892 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1893 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1895 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1898 if (!pfn_modify_allowed(pfn
, pgprot
))
1901 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1903 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1908 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1910 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
1912 /* these checks mirror the abort conditions in vm_normal_page */
1913 if (vma
->vm_flags
& VM_MIXEDMAP
)
1915 if (pfn_t_devmap(pfn
))
1917 if (pfn_t_special(pfn
))
1919 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
1924 static int __vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1925 pfn_t pfn
, bool mkwrite
)
1927 pgprot_t pgprot
= vma
->vm_page_prot
;
1929 BUG_ON(!vm_mixed_ok(vma
, pfn
));
1931 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1934 track_pfn_insert(vma
, &pgprot
, pfn
);
1936 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1940 * If we don't have pte special, then we have to use the pfn_valid()
1941 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1942 * refcount the page if pfn_valid is true (hence insert_page rather
1943 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1944 * without pte special, it would there be refcounted as a normal page.
1946 if (!HAVE_PTE_SPECIAL
&& !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1950 * At this point we are committed to insert_page()
1951 * regardless of whether the caller specified flags that
1952 * result in pfn_t_has_page() == false.
1954 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1955 return insert_page(vma
, addr
, page
, pgprot
);
1957 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1960 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1963 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1966 EXPORT_SYMBOL(vm_insert_mixed
);
1968 int vm_insert_mixed_mkwrite(struct vm_area_struct
*vma
, unsigned long addr
,
1971 return __vm_insert_mixed(vma
, addr
, pfn
, true);
1973 EXPORT_SYMBOL(vm_insert_mixed_mkwrite
);
1976 * maps a range of physical memory into the requested pages. the old
1977 * mappings are removed. any references to nonexistent pages results
1978 * in null mappings (currently treated as "copy-on-access")
1980 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1981 unsigned long addr
, unsigned long end
,
1982 unsigned long pfn
, pgprot_t prot
)
1988 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1991 arch_enter_lazy_mmu_mode();
1993 BUG_ON(!pte_none(*pte
));
1994 if (!pfn_modify_allowed(pfn
, prot
)) {
1998 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2000 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2001 arch_leave_lazy_mmu_mode();
2002 pte_unmap_unlock(pte
- 1, ptl
);
2006 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2007 unsigned long addr
, unsigned long end
,
2008 unsigned long pfn
, pgprot_t prot
)
2014 pfn
-= addr
>> PAGE_SHIFT
;
2015 pmd
= pmd_alloc(mm
, pud
, addr
);
2018 VM_BUG_ON(pmd_trans_huge(*pmd
));
2020 next
= pmd_addr_end(addr
, end
);
2021 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2022 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2025 } while (pmd
++, addr
= next
, addr
!= end
);
2029 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2030 unsigned long addr
, unsigned long end
,
2031 unsigned long pfn
, pgprot_t prot
)
2037 pfn
-= addr
>> PAGE_SHIFT
;
2038 pud
= pud_alloc(mm
, p4d
, addr
);
2042 next
= pud_addr_end(addr
, end
);
2043 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2044 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2047 } while (pud
++, addr
= next
, addr
!= end
);
2051 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2052 unsigned long addr
, unsigned long end
,
2053 unsigned long pfn
, pgprot_t prot
)
2059 pfn
-= addr
>> PAGE_SHIFT
;
2060 p4d
= p4d_alloc(mm
, pgd
, addr
);
2064 next
= p4d_addr_end(addr
, end
);
2065 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2066 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2069 } while (p4d
++, addr
= next
, addr
!= end
);
2074 * remap_pfn_range - remap kernel memory to userspace
2075 * @vma: user vma to map to
2076 * @addr: target user address to start at
2077 * @pfn: physical address of kernel memory
2078 * @size: size of map area
2079 * @prot: page protection flags for this mapping
2081 * Note: this is only safe if the mm semaphore is held when called.
2083 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2084 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2088 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2089 struct mm_struct
*mm
= vma
->vm_mm
;
2090 unsigned long remap_pfn
= pfn
;
2094 * Physically remapped pages are special. Tell the
2095 * rest of the world about it:
2096 * VM_IO tells people not to look at these pages
2097 * (accesses can have side effects).
2098 * VM_PFNMAP tells the core MM that the base pages are just
2099 * raw PFN mappings, and do not have a "struct page" associated
2102 * Disable vma merging and expanding with mremap().
2104 * Omit vma from core dump, even when VM_IO turned off.
2106 * There's a horrible special case to handle copy-on-write
2107 * behaviour that some programs depend on. We mark the "original"
2108 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2109 * See vm_normal_page() for details.
2111 if (is_cow_mapping(vma
->vm_flags
)) {
2112 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2114 vma
->vm_pgoff
= pfn
;
2117 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
2121 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2123 BUG_ON(addr
>= end
);
2124 pfn
-= addr
>> PAGE_SHIFT
;
2125 pgd
= pgd_offset(mm
, addr
);
2126 flush_cache_range(vma
, addr
, end
);
2128 next
= pgd_addr_end(addr
, end
);
2129 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2130 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2133 } while (pgd
++, addr
= next
, addr
!= end
);
2136 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
2140 EXPORT_SYMBOL(remap_pfn_range
);
2143 * vm_iomap_memory - remap memory to userspace
2144 * @vma: user vma to map to
2145 * @start: start of area
2146 * @len: size of area
2148 * This is a simplified io_remap_pfn_range() for common driver use. The
2149 * driver just needs to give us the physical memory range to be mapped,
2150 * we'll figure out the rest from the vma information.
2152 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2153 * whatever write-combining details or similar.
2155 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2157 unsigned long vm_len
, pfn
, pages
;
2159 /* Check that the physical memory area passed in looks valid */
2160 if (start
+ len
< start
)
2163 * You *really* shouldn't map things that aren't page-aligned,
2164 * but we've historically allowed it because IO memory might
2165 * just have smaller alignment.
2167 len
+= start
& ~PAGE_MASK
;
2168 pfn
= start
>> PAGE_SHIFT
;
2169 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2170 if (pfn
+ pages
< pfn
)
2173 /* We start the mapping 'vm_pgoff' pages into the area */
2174 if (vma
->vm_pgoff
> pages
)
2176 pfn
+= vma
->vm_pgoff
;
2177 pages
-= vma
->vm_pgoff
;
2179 /* Can we fit all of the mapping? */
2180 vm_len
= vma
->vm_end
- vma
->vm_start
;
2181 if (vm_len
>> PAGE_SHIFT
> pages
)
2184 /* Ok, let it rip */
2185 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2187 EXPORT_SYMBOL(vm_iomap_memory
);
2189 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2190 unsigned long addr
, unsigned long end
,
2191 pte_fn_t fn
, void *data
)
2196 spinlock_t
*uninitialized_var(ptl
);
2198 pte
= (mm
== &init_mm
) ?
2199 pte_alloc_kernel(pmd
, addr
) :
2200 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2204 BUG_ON(pmd_huge(*pmd
));
2206 arch_enter_lazy_mmu_mode();
2208 token
= pmd_pgtable(*pmd
);
2211 err
= fn(pte
++, token
, addr
, data
);
2214 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2216 arch_leave_lazy_mmu_mode();
2219 pte_unmap_unlock(pte
-1, ptl
);
2223 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2224 unsigned long addr
, unsigned long end
,
2225 pte_fn_t fn
, void *data
)
2231 BUG_ON(pud_huge(*pud
));
2233 pmd
= pmd_alloc(mm
, pud
, addr
);
2237 next
= pmd_addr_end(addr
, end
);
2238 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2241 } while (pmd
++, addr
= next
, addr
!= end
);
2245 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2246 unsigned long addr
, unsigned long end
,
2247 pte_fn_t fn
, void *data
)
2253 pud
= pud_alloc(mm
, p4d
, addr
);
2257 next
= pud_addr_end(addr
, end
);
2258 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2261 } while (pud
++, addr
= next
, addr
!= end
);
2265 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2266 unsigned long addr
, unsigned long end
,
2267 pte_fn_t fn
, void *data
)
2273 p4d
= p4d_alloc(mm
, pgd
, addr
);
2277 next
= p4d_addr_end(addr
, end
);
2278 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
);
2281 } while (p4d
++, addr
= next
, addr
!= end
);
2286 * Scan a region of virtual memory, filling in page tables as necessary
2287 * and calling a provided function on each leaf page table.
2289 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2290 unsigned long size
, pte_fn_t fn
, void *data
)
2294 unsigned long end
= addr
+ size
;
2297 if (WARN_ON(addr
>= end
))
2300 pgd
= pgd_offset(mm
, addr
);
2302 next
= pgd_addr_end(addr
, end
);
2303 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
);
2306 } while (pgd
++, addr
= next
, addr
!= end
);
2310 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2313 * handle_pte_fault chooses page fault handler according to an entry which was
2314 * read non-atomically. Before making any commitment, on those architectures
2315 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2316 * parts, do_swap_page must check under lock before unmapping the pte and
2317 * proceeding (but do_wp_page is only called after already making such a check;
2318 * and do_anonymous_page can safely check later on).
2320 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2321 pte_t
*page_table
, pte_t orig_pte
)
2324 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2325 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2326 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2328 same
= pte_same(*page_table
, orig_pte
);
2332 pte_unmap(page_table
);
2336 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2338 debug_dma_assert_idle(src
);
2341 * If the source page was a PFN mapping, we don't have
2342 * a "struct page" for it. We do a best-effort copy by
2343 * just copying from the original user address. If that
2344 * fails, we just zero-fill it. Live with it.
2346 if (unlikely(!src
)) {
2347 void *kaddr
= kmap_atomic(dst
);
2348 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2351 * This really shouldn't fail, because the page is there
2352 * in the page tables. But it might just be unreadable,
2353 * in which case we just give up and fill the result with
2356 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2358 kunmap_atomic(kaddr
);
2359 flush_dcache_page(dst
);
2361 copy_user_highpage(dst
, src
, va
, vma
);
2364 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2366 struct file
*vm_file
= vma
->vm_file
;
2369 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2372 * Special mappings (e.g. VDSO) do not have any file so fake
2373 * a default GFP_KERNEL for them.
2379 * Notify the address space that the page is about to become writable so that
2380 * it can prohibit this or wait for the page to get into an appropriate state.
2382 * We do this without the lock held, so that it can sleep if it needs to.
2384 static int do_page_mkwrite(struct vm_fault
*vmf
)
2387 struct page
*page
= vmf
->page
;
2388 unsigned int old_flags
= vmf
->flags
;
2390 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2392 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2393 /* Restore original flags so that caller is not surprised */
2394 vmf
->flags
= old_flags
;
2395 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2397 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2399 if (!page
->mapping
) {
2401 return 0; /* retry */
2403 ret
|= VM_FAULT_LOCKED
;
2405 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2410 * Handle dirtying of a page in shared file mapping on a write fault.
2412 * The function expects the page to be locked and unlocks it.
2414 static void fault_dirty_shared_page(struct vm_area_struct
*vma
,
2417 struct address_space
*mapping
;
2419 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2421 dirtied
= set_page_dirty(page
);
2422 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2424 * Take a local copy of the address_space - page.mapping may be zeroed
2425 * by truncate after unlock_page(). The address_space itself remains
2426 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2427 * release semantics to prevent the compiler from undoing this copying.
2429 mapping
= page_rmapping(page
);
2432 if ((dirtied
|| page_mkwrite
) && mapping
) {
2434 * Some device drivers do not set page.mapping
2435 * but still dirty their pages
2437 balance_dirty_pages_ratelimited(mapping
);
2441 file_update_time(vma
->vm_file
);
2445 * Handle write page faults for pages that can be reused in the current vma
2447 * This can happen either due to the mapping being with the VM_SHARED flag,
2448 * or due to us being the last reference standing to the page. In either
2449 * case, all we need to do here is to mark the page as writable and update
2450 * any related book-keeping.
2452 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2453 __releases(vmf
->ptl
)
2455 struct vm_area_struct
*vma
= vmf
->vma
;
2456 struct page
*page
= vmf
->page
;
2459 * Clear the pages cpupid information as the existing
2460 * information potentially belongs to a now completely
2461 * unrelated process.
2464 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2466 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2467 entry
= pte_mkyoung(vmf
->orig_pte
);
2468 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2469 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2470 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2471 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2475 * Handle the case of a page which we actually need to copy to a new page.
2477 * Called with mmap_sem locked and the old page referenced, but
2478 * without the ptl held.
2480 * High level logic flow:
2482 * - Allocate a page, copy the content of the old page to the new one.
2483 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2484 * - Take the PTL. If the pte changed, bail out and release the allocated page
2485 * - If the pte is still the way we remember it, update the page table and all
2486 * relevant references. This includes dropping the reference the page-table
2487 * held to the old page, as well as updating the rmap.
2488 * - In any case, unlock the PTL and drop the reference we took to the old page.
2490 static int wp_page_copy(struct vm_fault
*vmf
)
2492 struct vm_area_struct
*vma
= vmf
->vma
;
2493 struct mm_struct
*mm
= vma
->vm_mm
;
2494 struct page
*old_page
= vmf
->page
;
2495 struct page
*new_page
= NULL
;
2497 int page_copied
= 0;
2498 const unsigned long mmun_start
= vmf
->address
& PAGE_MASK
;
2499 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
;
2500 struct mem_cgroup
*memcg
;
2502 if (unlikely(anon_vma_prepare(vma
)))
2505 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2506 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2511 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2515 cow_user_page(new_page
, old_page
, vmf
->address
, vma
);
2518 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2521 __SetPageUptodate(new_page
);
2523 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2526 * Re-check the pte - we dropped the lock
2528 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2529 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2531 if (!PageAnon(old_page
)) {
2532 dec_mm_counter_fast(mm
,
2533 mm_counter_file(old_page
));
2534 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2537 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2539 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2540 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2541 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2543 * Clear the pte entry and flush it first, before updating the
2544 * pte with the new entry. This will avoid a race condition
2545 * seen in the presence of one thread doing SMC and another
2548 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2549 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2550 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2551 lru_cache_add_active_or_unevictable(new_page
, vma
);
2553 * We call the notify macro here because, when using secondary
2554 * mmu page tables (such as kvm shadow page tables), we want the
2555 * new page to be mapped directly into the secondary page table.
2557 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2558 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2561 * Only after switching the pte to the new page may
2562 * we remove the mapcount here. Otherwise another
2563 * process may come and find the rmap count decremented
2564 * before the pte is switched to the new page, and
2565 * "reuse" the old page writing into it while our pte
2566 * here still points into it and can be read by other
2569 * The critical issue is to order this
2570 * page_remove_rmap with the ptp_clear_flush above.
2571 * Those stores are ordered by (if nothing else,)
2572 * the barrier present in the atomic_add_negative
2573 * in page_remove_rmap.
2575 * Then the TLB flush in ptep_clear_flush ensures that
2576 * no process can access the old page before the
2577 * decremented mapcount is visible. And the old page
2578 * cannot be reused until after the decremented
2579 * mapcount is visible. So transitively, TLBs to
2580 * old page will be flushed before it can be reused.
2582 page_remove_rmap(old_page
, false);
2585 /* Free the old page.. */
2586 new_page
= old_page
;
2589 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2595 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2597 * No need to double call mmu_notifier->invalidate_range() callback as
2598 * the above ptep_clear_flush_notify() did already call it.
2600 mmu_notifier_invalidate_range_only_end(mm
, mmun_start
, mmun_end
);
2603 * Don't let another task, with possibly unlocked vma,
2604 * keep the mlocked page.
2606 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2607 lock_page(old_page
); /* LRU manipulation */
2608 if (PageMlocked(old_page
))
2609 munlock_vma_page(old_page
);
2610 unlock_page(old_page
);
2614 return page_copied
? VM_FAULT_WRITE
: 0;
2620 return VM_FAULT_OOM
;
2624 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2625 * writeable once the page is prepared
2627 * @vmf: structure describing the fault
2629 * This function handles all that is needed to finish a write page fault in a
2630 * shared mapping due to PTE being read-only once the mapped page is prepared.
2631 * It handles locking of PTE and modifying it. The function returns
2632 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2635 * The function expects the page to be locked or other protection against
2636 * concurrent faults / writeback (such as DAX radix tree locks).
2638 int finish_mkwrite_fault(struct vm_fault
*vmf
)
2640 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2641 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2644 * We might have raced with another page fault while we released the
2645 * pte_offset_map_lock.
2647 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2648 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2649 return VM_FAULT_NOPAGE
;
2656 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2659 static int wp_pfn_shared(struct vm_fault
*vmf
)
2661 struct vm_area_struct
*vma
= vmf
->vma
;
2663 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2666 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2667 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2668 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2669 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2671 return finish_mkwrite_fault(vmf
);
2674 return VM_FAULT_WRITE
;
2677 static int wp_page_shared(struct vm_fault
*vmf
)
2678 __releases(vmf
->ptl
)
2680 struct vm_area_struct
*vma
= vmf
->vma
;
2682 get_page(vmf
->page
);
2684 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2687 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2688 tmp
= do_page_mkwrite(vmf
);
2689 if (unlikely(!tmp
|| (tmp
&
2690 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2691 put_page(vmf
->page
);
2694 tmp
= finish_mkwrite_fault(vmf
);
2695 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2696 unlock_page(vmf
->page
);
2697 put_page(vmf
->page
);
2702 lock_page(vmf
->page
);
2704 fault_dirty_shared_page(vma
, vmf
->page
);
2705 put_page(vmf
->page
);
2707 return VM_FAULT_WRITE
;
2711 * This routine handles present pages, when users try to write
2712 * to a shared page. It is done by copying the page to a new address
2713 * and decrementing the shared-page counter for the old page.
2715 * Note that this routine assumes that the protection checks have been
2716 * done by the caller (the low-level page fault routine in most cases).
2717 * Thus we can safely just mark it writable once we've done any necessary
2720 * We also mark the page dirty at this point even though the page will
2721 * change only once the write actually happens. This avoids a few races,
2722 * and potentially makes it more efficient.
2724 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2725 * but allow concurrent faults), with pte both mapped and locked.
2726 * We return with mmap_sem still held, but pte unmapped and unlocked.
2728 static int do_wp_page(struct vm_fault
*vmf
)
2729 __releases(vmf
->ptl
)
2731 struct vm_area_struct
*vma
= vmf
->vma
;
2733 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2736 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2739 * We should not cow pages in a shared writeable mapping.
2740 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2742 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2743 (VM_WRITE
|VM_SHARED
))
2744 return wp_pfn_shared(vmf
);
2746 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2747 return wp_page_copy(vmf
);
2751 * Take out anonymous pages first, anonymous shared vmas are
2752 * not dirty accountable.
2754 if (PageAnon(vmf
->page
) && !PageKsm(vmf
->page
)) {
2755 int total_map_swapcount
;
2756 if (!trylock_page(vmf
->page
)) {
2757 get_page(vmf
->page
);
2758 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2759 lock_page(vmf
->page
);
2760 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2761 vmf
->address
, &vmf
->ptl
);
2762 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2763 unlock_page(vmf
->page
);
2764 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2765 put_page(vmf
->page
);
2768 put_page(vmf
->page
);
2770 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2771 if (total_map_swapcount
== 1) {
2773 * The page is all ours. Move it to
2774 * our anon_vma so the rmap code will
2775 * not search our parent or siblings.
2776 * Protected against the rmap code by
2779 page_move_anon_rmap(vmf
->page
, vma
);
2781 unlock_page(vmf
->page
);
2783 return VM_FAULT_WRITE
;
2785 unlock_page(vmf
->page
);
2786 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2787 (VM_WRITE
|VM_SHARED
))) {
2788 return wp_page_shared(vmf
);
2792 * Ok, we need to copy. Oh, well..
2794 get_page(vmf
->page
);
2796 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2797 return wp_page_copy(vmf
);
2800 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2801 unsigned long start_addr
, unsigned long end_addr
,
2802 struct zap_details
*details
)
2804 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2807 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2808 struct zap_details
*details
)
2810 struct vm_area_struct
*vma
;
2811 pgoff_t vba
, vea
, zba
, zea
;
2813 vma_interval_tree_foreach(vma
, root
,
2814 details
->first_index
, details
->last_index
) {
2816 vba
= vma
->vm_pgoff
;
2817 vea
= vba
+ vma_pages(vma
) - 1;
2818 zba
= details
->first_index
;
2821 zea
= details
->last_index
;
2825 unmap_mapping_range_vma(vma
,
2826 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2827 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2833 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2834 * address_space corresponding to the specified page range in the underlying
2837 * @mapping: the address space containing mmaps to be unmapped.
2838 * @holebegin: byte in first page to unmap, relative to the start of
2839 * the underlying file. This will be rounded down to a PAGE_SIZE
2840 * boundary. Note that this is different from truncate_pagecache(), which
2841 * must keep the partial page. In contrast, we must get rid of
2843 * @holelen: size of prospective hole in bytes. This will be rounded
2844 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2846 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2847 * but 0 when invalidating pagecache, don't throw away private data.
2849 void unmap_mapping_range(struct address_space
*mapping
,
2850 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2852 struct zap_details details
= { };
2853 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2854 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2856 /* Check for overflow. */
2857 if (sizeof(holelen
) > sizeof(hlen
)) {
2859 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2860 if (holeend
& ~(long long)ULONG_MAX
)
2861 hlen
= ULONG_MAX
- hba
+ 1;
2864 details
.check_mapping
= even_cows
? NULL
: mapping
;
2865 details
.first_index
= hba
;
2866 details
.last_index
= hba
+ hlen
- 1;
2867 if (details
.last_index
< details
.first_index
)
2868 details
.last_index
= ULONG_MAX
;
2870 i_mmap_lock_write(mapping
);
2871 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2872 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2873 i_mmap_unlock_write(mapping
);
2875 EXPORT_SYMBOL(unmap_mapping_range
);
2878 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2879 * but allow concurrent faults), and pte mapped but not yet locked.
2880 * We return with pte unmapped and unlocked.
2882 * We return with the mmap_sem locked or unlocked in the same cases
2883 * as does filemap_fault().
2885 int do_swap_page(struct vm_fault
*vmf
)
2887 struct vm_area_struct
*vma
= vmf
->vma
;
2888 struct page
*page
= NULL
, *swapcache
= NULL
;
2889 struct mem_cgroup
*memcg
;
2890 struct vma_swap_readahead swap_ra
;
2896 bool vma_readahead
= swap_use_vma_readahead();
2898 if (vma_readahead
) {
2899 page
= swap_readahead_detect(vmf
, &swap_ra
);
2903 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
)) {
2909 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2910 if (unlikely(non_swap_entry(entry
))) {
2911 if (is_migration_entry(entry
)) {
2912 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2914 } else if (is_device_private_entry(entry
)) {
2916 * For un-addressable device memory we call the pgmap
2917 * fault handler callback. The callback must migrate
2918 * the page back to some CPU accessible page.
2920 ret
= device_private_entry_fault(vma
, vmf
->address
, entry
,
2921 vmf
->flags
, vmf
->pmd
);
2922 } else if (is_hwpoison_entry(entry
)) {
2923 ret
= VM_FAULT_HWPOISON
;
2925 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2926 ret
= VM_FAULT_SIGBUS
;
2932 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2934 page
= lookup_swap_cache(entry
, vma_readahead
? vma
: NULL
,
2940 struct swap_info_struct
*si
= swp_swap_info(entry
);
2942 if (si
->flags
& SWP_SYNCHRONOUS_IO
&&
2943 __swap_count(si
, entry
) == 1) {
2944 /* skip swapcache */
2945 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
2947 __SetPageLocked(page
);
2948 __SetPageSwapBacked(page
);
2949 set_page_private(page
, entry
.val
);
2950 lru_cache_add_anon(page
);
2951 swap_readpage(page
, true);
2955 page
= do_swap_page_readahead(entry
,
2956 GFP_HIGHUSER_MOVABLE
, vmf
, &swap_ra
);
2958 page
= swapin_readahead(entry
,
2959 GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
2965 * Back out if somebody else faulted in this pte
2966 * while we released the pte lock.
2968 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2969 vmf
->address
, &vmf
->ptl
);
2970 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2972 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2976 /* Had to read the page from swap area: Major fault */
2977 ret
= VM_FAULT_MAJOR
;
2978 count_vm_event(PGMAJFAULT
);
2979 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2980 } else if (PageHWPoison(page
)) {
2982 * hwpoisoned dirty swapcache pages are kept for killing
2983 * owner processes (which may be unknown at hwpoison time)
2985 ret
= VM_FAULT_HWPOISON
;
2986 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2991 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2993 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2995 ret
|= VM_FAULT_RETRY
;
3000 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3001 * release the swapcache from under us. The page pin, and pte_same
3002 * test below, are not enough to exclude that. Even if it is still
3003 * swapcache, we need to check that the page's swap has not changed.
3005 if (unlikely((!PageSwapCache(page
) ||
3006 page_private(page
) != entry
.val
)) && swapcache
)
3009 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3010 if (unlikely(!page
)) {
3016 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
,
3023 * Back out if somebody else already faulted in this pte.
3025 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3027 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3030 if (unlikely(!PageUptodate(page
))) {
3031 ret
= VM_FAULT_SIGBUS
;
3036 * The page isn't present yet, go ahead with the fault.
3038 * Be careful about the sequence of operations here.
3039 * To get its accounting right, reuse_swap_page() must be called
3040 * while the page is counted on swap but not yet in mapcount i.e.
3041 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3042 * must be called after the swap_free(), or it will never succeed.
3045 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3046 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3047 pte
= mk_pte(page
, vma
->vm_page_prot
);
3048 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3049 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3050 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3051 ret
|= VM_FAULT_WRITE
;
3052 exclusive
= RMAP_EXCLUSIVE
;
3054 flush_icache_page(vma
, page
);
3055 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3056 pte
= pte_mksoft_dirty(pte
);
3057 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3058 vmf
->orig_pte
= pte
;
3060 /* ksm created a completely new copy */
3061 if (unlikely(page
!= swapcache
&& swapcache
)) {
3062 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3063 mem_cgroup_commit_charge(page
, memcg
, false, false);
3064 lru_cache_add_active_or_unevictable(page
, vma
);
3066 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3067 mem_cgroup_commit_charge(page
, memcg
, true, false);
3068 activate_page(page
);
3072 if (mem_cgroup_swap_full(page
) ||
3073 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3074 try_to_free_swap(page
);
3076 if (page
!= swapcache
&& swapcache
) {
3078 * Hold the lock to avoid the swap entry to be reused
3079 * until we take the PT lock for the pte_same() check
3080 * (to avoid false positives from pte_same). For
3081 * further safety release the lock after the swap_free
3082 * so that the swap count won't change under a
3083 * parallel locked swapcache.
3085 unlock_page(swapcache
);
3086 put_page(swapcache
);
3089 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3090 ret
|= do_wp_page(vmf
);
3091 if (ret
& VM_FAULT_ERROR
)
3092 ret
&= VM_FAULT_ERROR
;
3096 /* No need to invalidate - it was non-present before */
3097 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3099 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3103 mem_cgroup_cancel_charge(page
, memcg
, false);
3104 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3109 if (page
!= swapcache
&& swapcache
) {
3110 unlock_page(swapcache
);
3111 put_page(swapcache
);
3117 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3118 * but allow concurrent faults), and pte mapped but not yet locked.
3119 * We return with mmap_sem still held, but pte unmapped and unlocked.
3121 static int do_anonymous_page(struct vm_fault
*vmf
)
3123 struct vm_area_struct
*vma
= vmf
->vma
;
3124 struct mem_cgroup
*memcg
;
3129 /* File mapping without ->vm_ops ? */
3130 if (vma
->vm_flags
& VM_SHARED
)
3131 return VM_FAULT_SIGBUS
;
3134 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3135 * pte_offset_map() on pmds where a huge pmd might be created
3136 * from a different thread.
3138 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3139 * parallel threads are excluded by other means.
3141 * Here we only have down_read(mmap_sem).
3143 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))
3144 return VM_FAULT_OOM
;
3146 /* See the comment in pte_alloc_one_map() */
3147 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3150 /* Use the zero-page for reads */
3151 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3152 !mm_forbids_zeropage(vma
->vm_mm
)) {
3153 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3154 vma
->vm_page_prot
));
3155 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3156 vmf
->address
, &vmf
->ptl
);
3157 if (!pte_none(*vmf
->pte
))
3159 ret
= check_stable_address_space(vma
->vm_mm
);
3162 /* Deliver the page fault to userland, check inside PT lock */
3163 if (userfaultfd_missing(vma
)) {
3164 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3165 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3170 /* Allocate our own private page. */
3171 if (unlikely(anon_vma_prepare(vma
)))
3173 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3177 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
3181 * The memory barrier inside __SetPageUptodate makes sure that
3182 * preceeding stores to the page contents become visible before
3183 * the set_pte_at() write.
3185 __SetPageUptodate(page
);
3187 entry
= mk_pte(page
, vma
->vm_page_prot
);
3188 if (vma
->vm_flags
& VM_WRITE
)
3189 entry
= pte_mkwrite(pte_mkdirty(entry
));
3191 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3193 if (!pte_none(*vmf
->pte
))
3196 ret
= check_stable_address_space(vma
->vm_mm
);
3200 /* Deliver the page fault to userland, check inside PT lock */
3201 if (userfaultfd_missing(vma
)) {
3202 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3203 mem_cgroup_cancel_charge(page
, memcg
, false);
3205 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3208 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3209 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3210 mem_cgroup_commit_charge(page
, memcg
, false, false);
3211 lru_cache_add_active_or_unevictable(page
, vma
);
3213 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3215 /* No need to invalidate - it was non-present before */
3216 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3218 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3221 mem_cgroup_cancel_charge(page
, memcg
, false);
3227 return VM_FAULT_OOM
;
3231 * The mmap_sem must have been held on entry, and may have been
3232 * released depending on flags and vma->vm_ops->fault() return value.
3233 * See filemap_fault() and __lock_page_retry().
3235 static int __do_fault(struct vm_fault
*vmf
)
3237 struct vm_area_struct
*vma
= vmf
->vma
;
3241 * Preallocate pte before we take page_lock because this might lead to
3242 * deadlocks for memcg reclaim which waits for pages under writeback:
3244 * SetPageWriteback(A)
3250 * wait_on_page_writeback(A)
3251 * SetPageWriteback(B)
3253 * # flush A, B to clear the writeback
3255 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3256 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
,
3258 if (!vmf
->prealloc_pte
)
3259 return VM_FAULT_OOM
;
3260 smp_wmb(); /* See comment in __pte_alloc() */
3263 ret
= vma
->vm_ops
->fault(vmf
);
3264 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3265 VM_FAULT_DONE_COW
)))
3268 if (unlikely(PageHWPoison(vmf
->page
))) {
3269 if (ret
& VM_FAULT_LOCKED
)
3270 unlock_page(vmf
->page
);
3271 put_page(vmf
->page
);
3273 return VM_FAULT_HWPOISON
;
3276 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3277 lock_page(vmf
->page
);
3279 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3285 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3286 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3287 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3288 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3290 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3292 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3295 static int pte_alloc_one_map(struct vm_fault
*vmf
)
3297 struct vm_area_struct
*vma
= vmf
->vma
;
3299 if (!pmd_none(*vmf
->pmd
))
3301 if (vmf
->prealloc_pte
) {
3302 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3303 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3304 spin_unlock(vmf
->ptl
);
3308 mm_inc_nr_ptes(vma
->vm_mm
);
3309 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3310 spin_unlock(vmf
->ptl
);
3311 vmf
->prealloc_pte
= NULL
;
3312 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))) {
3313 return VM_FAULT_OOM
;
3317 * If a huge pmd materialized under us just retry later. Use
3318 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3319 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3320 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3321 * running immediately after a huge pmd fault in a different thread of
3322 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3323 * All we have to ensure is that it is a regular pmd that we can walk
3324 * with pte_offset_map() and we can do that through an atomic read in
3325 * C, which is what pmd_trans_unstable() provides.
3327 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3328 return VM_FAULT_NOPAGE
;
3331 * At this point we know that our vmf->pmd points to a page of ptes
3332 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3333 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3334 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3335 * be valid and we will re-check to make sure the vmf->pte isn't
3336 * pte_none() under vmf->ptl protection when we return to
3339 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3344 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3346 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3347 static inline bool transhuge_vma_suitable(struct vm_area_struct
*vma
,
3348 unsigned long haddr
)
3350 if (((vma
->vm_start
>> PAGE_SHIFT
) & HPAGE_CACHE_INDEX_MASK
) !=
3351 (vma
->vm_pgoff
& HPAGE_CACHE_INDEX_MASK
))
3353 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
3358 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3360 struct vm_area_struct
*vma
= vmf
->vma
;
3362 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3364 * We are going to consume the prealloc table,
3365 * count that as nr_ptes.
3367 mm_inc_nr_ptes(vma
->vm_mm
);
3368 vmf
->prealloc_pte
= NULL
;
3371 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3373 struct vm_area_struct
*vma
= vmf
->vma
;
3374 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3375 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3379 if (!transhuge_vma_suitable(vma
, haddr
))
3380 return VM_FAULT_FALLBACK
;
3382 ret
= VM_FAULT_FALLBACK
;
3383 page
= compound_head(page
);
3386 * Archs like ppc64 need additonal space to store information
3387 * related to pte entry. Use the preallocated table for that.
3389 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3390 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
, vmf
->address
);
3391 if (!vmf
->prealloc_pte
)
3392 return VM_FAULT_OOM
;
3393 smp_wmb(); /* See comment in __pte_alloc() */
3396 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3397 if (unlikely(!pmd_none(*vmf
->pmd
)))
3400 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3401 flush_icache_page(vma
, page
+ i
);
3403 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3405 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3407 add_mm_counter(vma
->vm_mm
, MM_FILEPAGES
, HPAGE_PMD_NR
);
3408 page_add_file_rmap(page
, true);
3410 * deposit and withdraw with pmd lock held
3412 if (arch_needs_pgtable_deposit())
3413 deposit_prealloc_pte(vmf
);
3415 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3417 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3419 /* fault is handled */
3421 count_vm_event(THP_FILE_MAPPED
);
3423 spin_unlock(vmf
->ptl
);
3427 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3435 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3436 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3438 * @vmf: fault environment
3439 * @memcg: memcg to charge page (only for private mappings)
3440 * @page: page to map
3442 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3445 * Target users are page handler itself and implementations of
3446 * vm_ops->map_pages.
3448 int alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3451 struct vm_area_struct
*vma
= vmf
->vma
;
3452 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3456 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3457 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3459 VM_BUG_ON_PAGE(memcg
, page
);
3461 ret
= do_set_pmd(vmf
, page
);
3462 if (ret
!= VM_FAULT_FALLBACK
)
3467 ret
= pte_alloc_one_map(vmf
);
3472 /* Re-check under ptl */
3473 if (unlikely(!pte_none(*vmf
->pte
)))
3474 return VM_FAULT_NOPAGE
;
3476 flush_icache_page(vma
, page
);
3477 entry
= mk_pte(page
, vma
->vm_page_prot
);
3479 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3480 /* copy-on-write page */
3481 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3482 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3483 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3484 mem_cgroup_commit_charge(page
, memcg
, false, false);
3485 lru_cache_add_active_or_unevictable(page
, vma
);
3487 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3488 page_add_file_rmap(page
, false);
3490 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3492 /* no need to invalidate: a not-present page won't be cached */
3493 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3500 * finish_fault - finish page fault once we have prepared the page to fault
3502 * @vmf: structure describing the fault
3504 * This function handles all that is needed to finish a page fault once the
3505 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3506 * given page, adds reverse page mapping, handles memcg charges and LRU
3507 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3510 * The function expects the page to be locked and on success it consumes a
3511 * reference of a page being mapped (for the PTE which maps it).
3513 int finish_fault(struct vm_fault
*vmf
)
3518 /* Did we COW the page? */
3519 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3520 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3521 page
= vmf
->cow_page
;
3526 * check even for read faults because we might have lost our CoWed
3529 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3530 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3532 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3534 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3538 static unsigned long fault_around_bytes __read_mostly
=
3539 rounddown_pow_of_two(65536);
3541 #ifdef CONFIG_DEBUG_FS
3542 static int fault_around_bytes_get(void *data
, u64
*val
)
3544 *val
= fault_around_bytes
;
3549 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3550 * rounded down to nearest page order. It's what do_fault_around() expects to
3553 static int fault_around_bytes_set(void *data
, u64 val
)
3555 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3557 if (val
> PAGE_SIZE
)
3558 fault_around_bytes
= rounddown_pow_of_two(val
);
3560 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3563 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3564 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3566 static int __init
fault_around_debugfs(void)
3570 ret
= debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3571 &fault_around_bytes_fops
);
3573 pr_warn("Failed to create fault_around_bytes in debugfs");
3576 late_initcall(fault_around_debugfs
);
3580 * do_fault_around() tries to map few pages around the fault address. The hope
3581 * is that the pages will be needed soon and this will lower the number of
3584 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3585 * not ready to be mapped: not up-to-date, locked, etc.
3587 * This function is called with the page table lock taken. In the split ptlock
3588 * case the page table lock only protects only those entries which belong to
3589 * the page table corresponding to the fault address.
3591 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3594 * fault_around_pages() defines how many pages we'll try to map.
3595 * do_fault_around() expects it to return a power of two less than or equal to
3598 * The virtual address of the area that we map is naturally aligned to the
3599 * fault_around_pages() value (and therefore to page order). This way it's
3600 * easier to guarantee that we don't cross page table boundaries.
3602 static int do_fault_around(struct vm_fault
*vmf
)
3604 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3605 pgoff_t start_pgoff
= vmf
->pgoff
;
3609 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3610 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3612 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3613 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3617 * end_pgoff is either end of page table or end of vma
3618 * or fault_around_pages() from start_pgoff, depending what is nearest.
3620 end_pgoff
= start_pgoff
-
3621 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3623 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3624 start_pgoff
+ nr_pages
- 1);
3626 if (pmd_none(*vmf
->pmd
)) {
3627 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
,
3629 if (!vmf
->prealloc_pte
)
3631 smp_wmb(); /* See comment in __pte_alloc() */
3634 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3636 /* Huge page is mapped? Page fault is solved */
3637 if (pmd_trans_huge(*vmf
->pmd
)) {
3638 ret
= VM_FAULT_NOPAGE
;
3642 /* ->map_pages() haven't done anything useful. Cold page cache? */
3646 /* check if the page fault is solved */
3647 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3648 if (!pte_none(*vmf
->pte
))
3649 ret
= VM_FAULT_NOPAGE
;
3650 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3652 vmf
->address
= address
;
3657 static int do_read_fault(struct vm_fault
*vmf
)
3659 struct vm_area_struct
*vma
= vmf
->vma
;
3663 * Let's call ->map_pages() first and use ->fault() as fallback
3664 * if page by the offset is not ready to be mapped (cold cache or
3667 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3668 ret
= do_fault_around(vmf
);
3673 ret
= __do_fault(vmf
);
3674 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3677 ret
|= finish_fault(vmf
);
3678 unlock_page(vmf
->page
);
3679 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3680 put_page(vmf
->page
);
3684 static int do_cow_fault(struct vm_fault
*vmf
)
3686 struct vm_area_struct
*vma
= vmf
->vma
;
3689 if (unlikely(anon_vma_prepare(vma
)))
3690 return VM_FAULT_OOM
;
3692 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3694 return VM_FAULT_OOM
;
3696 if (mem_cgroup_try_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3697 &vmf
->memcg
, false)) {
3698 put_page(vmf
->cow_page
);
3699 return VM_FAULT_OOM
;
3702 ret
= __do_fault(vmf
);
3703 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3705 if (ret
& VM_FAULT_DONE_COW
)
3708 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3709 __SetPageUptodate(vmf
->cow_page
);
3711 ret
|= finish_fault(vmf
);
3712 unlock_page(vmf
->page
);
3713 put_page(vmf
->page
);
3714 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3718 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3719 put_page(vmf
->cow_page
);
3723 static int do_shared_fault(struct vm_fault
*vmf
)
3725 struct vm_area_struct
*vma
= vmf
->vma
;
3728 ret
= __do_fault(vmf
);
3729 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3733 * Check if the backing address space wants to know that the page is
3734 * about to become writable
3736 if (vma
->vm_ops
->page_mkwrite
) {
3737 unlock_page(vmf
->page
);
3738 tmp
= do_page_mkwrite(vmf
);
3739 if (unlikely(!tmp
||
3740 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3741 put_page(vmf
->page
);
3746 ret
|= finish_fault(vmf
);
3747 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3749 unlock_page(vmf
->page
);
3750 put_page(vmf
->page
);
3754 fault_dirty_shared_page(vma
, vmf
->page
);
3759 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3760 * but allow concurrent faults).
3761 * The mmap_sem may have been released depending on flags and our
3762 * return value. See filemap_fault() and __lock_page_or_retry().
3764 static int do_fault(struct vm_fault
*vmf
)
3766 struct vm_area_struct
*vma
= vmf
->vma
;
3770 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3772 if (!vma
->vm_ops
->fault
) {
3774 * If we find a migration pmd entry or a none pmd entry, which
3775 * should never happen, return SIGBUS
3777 if (unlikely(!pmd_present(*vmf
->pmd
)))
3778 ret
= VM_FAULT_SIGBUS
;
3780 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
3785 * Make sure this is not a temporary clearing of pte
3786 * by holding ptl and checking again. A R/M/W update
3787 * of pte involves: take ptl, clearing the pte so that
3788 * we don't have concurrent modification by hardware
3789 * followed by an update.
3791 if (unlikely(pte_none(*vmf
->pte
)))
3792 ret
= VM_FAULT_SIGBUS
;
3794 ret
= VM_FAULT_NOPAGE
;
3796 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3798 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3799 ret
= do_read_fault(vmf
);
3800 else if (!(vma
->vm_flags
& VM_SHARED
))
3801 ret
= do_cow_fault(vmf
);
3803 ret
= do_shared_fault(vmf
);
3805 /* preallocated pagetable is unused: free it */
3806 if (vmf
->prealloc_pte
) {
3807 pte_free(vma
->vm_mm
, vmf
->prealloc_pte
);
3808 vmf
->prealloc_pte
= NULL
;
3813 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3814 unsigned long addr
, int page_nid
,
3819 count_vm_numa_event(NUMA_HINT_FAULTS
);
3820 if (page_nid
== numa_node_id()) {
3821 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3822 *flags
|= TNF_FAULT_LOCAL
;
3825 return mpol_misplaced(page
, vma
, addr
);
3828 static int do_numa_page(struct vm_fault
*vmf
)
3830 struct vm_area_struct
*vma
= vmf
->vma
;
3831 struct page
*page
= NULL
;
3835 bool migrated
= false;
3837 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3841 * The "pte" at this point cannot be used safely without
3842 * validation through pte_unmap_same(). It's of NUMA type but
3843 * the pfn may be screwed if the read is non atomic.
3845 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3846 spin_lock(vmf
->ptl
);
3847 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3848 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3853 * Make it present again, Depending on how arch implementes non
3854 * accessible ptes, some can allow access by kernel mode.
3856 pte
= ptep_modify_prot_start(vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3857 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3858 pte
= pte_mkyoung(pte
);
3860 pte
= pte_mkwrite(pte
);
3861 ptep_modify_prot_commit(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3862 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3864 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3866 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3870 /* TODO: handle PTE-mapped THP */
3871 if (PageCompound(page
)) {
3872 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3877 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3878 * much anyway since they can be in shared cache state. This misses
3879 * the case where a mapping is writable but the process never writes
3880 * to it but pte_write gets cleared during protection updates and
3881 * pte_dirty has unpredictable behaviour between PTE scan updates,
3882 * background writeback, dirty balancing and application behaviour.
3884 if (!pte_write(pte
))
3885 flags
|= TNF_NO_GROUP
;
3888 * Flag if the page is shared between multiple address spaces. This
3889 * is later used when determining whether to group tasks together
3891 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3892 flags
|= TNF_SHARED
;
3894 last_cpupid
= page_cpupid_last(page
);
3895 page_nid
= page_to_nid(page
);
3896 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3898 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3899 if (target_nid
== -1) {
3904 /* Migrate to the requested node */
3905 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3907 page_nid
= target_nid
;
3908 flags
|= TNF_MIGRATED
;
3910 flags
|= TNF_MIGRATE_FAIL
;
3914 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3918 static inline int create_huge_pmd(struct vm_fault
*vmf
)
3920 if (vma_is_anonymous(vmf
->vma
))
3921 return do_huge_pmd_anonymous_page(vmf
);
3922 if (vmf
->vma
->vm_ops
->huge_fault
)
3923 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3924 return VM_FAULT_FALLBACK
;
3927 /* `inline' is required to avoid gcc 4.1.2 build error */
3928 static inline int wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3930 if (vma_is_anonymous(vmf
->vma
))
3931 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3932 if (vmf
->vma
->vm_ops
->huge_fault
)
3933 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3935 /* COW handled on pte level: split pmd */
3936 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3937 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3939 return VM_FAULT_FALLBACK
;
3942 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3944 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3947 static int create_huge_pud(struct vm_fault
*vmf
)
3949 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3950 /* No support for anonymous transparent PUD pages yet */
3951 if (vma_is_anonymous(vmf
->vma
))
3952 return VM_FAULT_FALLBACK
;
3953 if (vmf
->vma
->vm_ops
->huge_fault
)
3954 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3955 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3956 return VM_FAULT_FALLBACK
;
3959 static int wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3961 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3962 /* No support for anonymous transparent PUD pages yet */
3963 if (vma_is_anonymous(vmf
->vma
))
3964 return VM_FAULT_FALLBACK
;
3965 if (vmf
->vma
->vm_ops
->huge_fault
)
3966 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3967 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3968 return VM_FAULT_FALLBACK
;
3972 * These routines also need to handle stuff like marking pages dirty
3973 * and/or accessed for architectures that don't do it in hardware (most
3974 * RISC architectures). The early dirtying is also good on the i386.
3976 * There is also a hook called "update_mmu_cache()" that architectures
3977 * with external mmu caches can use to update those (ie the Sparc or
3978 * PowerPC hashed page tables that act as extended TLBs).
3980 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3981 * concurrent faults).
3983 * The mmap_sem may have been released depending on flags and our return value.
3984 * See filemap_fault() and __lock_page_or_retry().
3986 static int handle_pte_fault(struct vm_fault
*vmf
)
3990 if (unlikely(pmd_none(*vmf
->pmd
))) {
3992 * Leave __pte_alloc() until later: because vm_ops->fault may
3993 * want to allocate huge page, and if we expose page table
3994 * for an instant, it will be difficult to retract from
3995 * concurrent faults and from rmap lookups.
3999 /* See comment in pte_alloc_one_map() */
4000 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4003 * A regular pmd is established and it can't morph into a huge
4004 * pmd from under us anymore at this point because we hold the
4005 * mmap_sem read mode and khugepaged takes it in write mode.
4006 * So now it's safe to run pte_offset_map().
4008 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4009 vmf
->orig_pte
= *vmf
->pte
;
4012 * some architectures can have larger ptes than wordsize,
4013 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4014 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4015 * accesses. The code below just needs a consistent view
4016 * for the ifs and we later double check anyway with the
4017 * ptl lock held. So here a barrier will do.
4020 if (pte_none(vmf
->orig_pte
)) {
4021 pte_unmap(vmf
->pte
);
4027 if (vma_is_anonymous(vmf
->vma
))
4028 return do_anonymous_page(vmf
);
4030 return do_fault(vmf
);
4033 if (!pte_present(vmf
->orig_pte
))
4034 return do_swap_page(vmf
);
4036 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4037 return do_numa_page(vmf
);
4039 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4040 spin_lock(vmf
->ptl
);
4041 entry
= vmf
->orig_pte
;
4042 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
4044 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4045 if (!pte_write(entry
))
4046 return do_wp_page(vmf
);
4047 entry
= pte_mkdirty(entry
);
4049 entry
= pte_mkyoung(entry
);
4050 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4051 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4052 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4055 * This is needed only for protection faults but the arch code
4056 * is not yet telling us if this is a protection fault or not.
4057 * This still avoids useless tlb flushes for .text page faults
4060 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4061 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4064 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4069 * By the time we get here, we already hold the mm semaphore
4071 * The mmap_sem may have been released depending on flags and our
4072 * return value. See filemap_fault() and __lock_page_or_retry().
4074 static int __handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4077 struct vm_fault vmf
= {
4079 .address
= address
& PAGE_MASK
,
4081 .pgoff
= linear_page_index(vma
, address
),
4082 .gfp_mask
= __get_fault_gfp_mask(vma
),
4084 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4085 struct mm_struct
*mm
= vma
->vm_mm
;
4090 pgd
= pgd_offset(mm
, address
);
4091 p4d
= p4d_alloc(mm
, pgd
, address
);
4093 return VM_FAULT_OOM
;
4095 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4097 return VM_FAULT_OOM
;
4098 if (pud_none(*vmf
.pud
) && transparent_hugepage_enabled(vma
)) {
4099 ret
= create_huge_pud(&vmf
);
4100 if (!(ret
& VM_FAULT_FALLBACK
))
4103 pud_t orig_pud
= *vmf
.pud
;
4106 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4108 /* NUMA case for anonymous PUDs would go here */
4110 if (dirty
&& !pud_write(orig_pud
)) {
4111 ret
= wp_huge_pud(&vmf
, orig_pud
);
4112 if (!(ret
& VM_FAULT_FALLBACK
))
4115 huge_pud_set_accessed(&vmf
, orig_pud
);
4121 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4123 return VM_FAULT_OOM
;
4124 if (pmd_none(*vmf
.pmd
) && transparent_hugepage_enabled(vma
)) {
4125 ret
= create_huge_pmd(&vmf
);
4126 if (!(ret
& VM_FAULT_FALLBACK
))
4129 pmd_t orig_pmd
= *vmf
.pmd
;
4132 if (unlikely(is_swap_pmd(orig_pmd
))) {
4133 VM_BUG_ON(thp_migration_supported() &&
4134 !is_pmd_migration_entry(orig_pmd
));
4135 if (is_pmd_migration_entry(orig_pmd
))
4136 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4139 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4140 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4141 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4143 if (dirty
&& !pmd_write(orig_pmd
)) {
4144 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4145 if (!(ret
& VM_FAULT_FALLBACK
))
4148 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4154 return handle_pte_fault(&vmf
);
4158 * By the time we get here, we already hold the mm semaphore
4160 * The mmap_sem may have been released depending on flags and our
4161 * return value. See filemap_fault() and __lock_page_or_retry().
4163 int handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4168 __set_current_state(TASK_RUNNING
);
4170 count_vm_event(PGFAULT
);
4171 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4173 /* do counter updates before entering really critical section. */
4174 check_sync_rss_stat(current
);
4176 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4177 flags
& FAULT_FLAG_INSTRUCTION
,
4178 flags
& FAULT_FLAG_REMOTE
))
4179 return VM_FAULT_SIGSEGV
;
4182 * Enable the memcg OOM handling for faults triggered in user
4183 * space. Kernel faults are handled more gracefully.
4185 if (flags
& FAULT_FLAG_USER
)
4186 mem_cgroup_oom_enable();
4188 if (unlikely(is_vm_hugetlb_page(vma
)))
4189 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4191 ret
= __handle_mm_fault(vma
, address
, flags
);
4193 if (flags
& FAULT_FLAG_USER
) {
4194 mem_cgroup_oom_disable();
4196 * The task may have entered a memcg OOM situation but
4197 * if the allocation error was handled gracefully (no
4198 * VM_FAULT_OOM), there is no need to kill anything.
4199 * Just clean up the OOM state peacefully.
4201 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4202 mem_cgroup_oom_synchronize(false);
4207 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4209 #ifndef __PAGETABLE_P4D_FOLDED
4211 * Allocate p4d page table.
4212 * We've already handled the fast-path in-line.
4214 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4216 p4d_t
*new = p4d_alloc_one(mm
, address
);
4220 smp_wmb(); /* See comment in __pte_alloc */
4222 spin_lock(&mm
->page_table_lock
);
4223 if (pgd_present(*pgd
)) /* Another has populated it */
4226 pgd_populate(mm
, pgd
, new);
4227 spin_unlock(&mm
->page_table_lock
);
4230 #endif /* __PAGETABLE_P4D_FOLDED */
4232 #ifndef __PAGETABLE_PUD_FOLDED
4234 * Allocate page upper directory.
4235 * We've already handled the fast-path in-line.
4237 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4239 pud_t
*new = pud_alloc_one(mm
, address
);
4243 smp_wmb(); /* See comment in __pte_alloc */
4245 spin_lock(&mm
->page_table_lock
);
4246 #ifndef __ARCH_HAS_5LEVEL_HACK
4247 if (!p4d_present(*p4d
)) {
4249 p4d_populate(mm
, p4d
, new);
4250 } else /* Another has populated it */
4253 if (!pgd_present(*p4d
)) {
4255 pgd_populate(mm
, p4d
, new);
4256 } else /* Another has populated it */
4258 #endif /* __ARCH_HAS_5LEVEL_HACK */
4259 spin_unlock(&mm
->page_table_lock
);
4262 #endif /* __PAGETABLE_PUD_FOLDED */
4264 #ifndef __PAGETABLE_PMD_FOLDED
4266 * Allocate page middle directory.
4267 * We've already handled the fast-path in-line.
4269 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4272 pmd_t
*new = pmd_alloc_one(mm
, address
);
4276 smp_wmb(); /* See comment in __pte_alloc */
4278 ptl
= pud_lock(mm
, pud
);
4279 #ifndef __ARCH_HAS_4LEVEL_HACK
4280 if (!pud_present(*pud
)) {
4282 pud_populate(mm
, pud
, new);
4283 } else /* Another has populated it */
4286 if (!pgd_present(*pud
)) {
4288 pgd_populate(mm
, pud
, new);
4289 } else /* Another has populated it */
4291 #endif /* __ARCH_HAS_4LEVEL_HACK */
4295 #endif /* __PAGETABLE_PMD_FOLDED */
4297 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4298 unsigned long *start
, unsigned long *end
,
4299 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4307 pgd
= pgd_offset(mm
, address
);
4308 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4311 p4d
= p4d_offset(pgd
, address
);
4312 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4315 pud
= pud_offset(p4d
, address
);
4316 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4319 pmd
= pmd_offset(pud
, address
);
4320 VM_BUG_ON(pmd_trans_huge(*pmd
));
4322 if (pmd_huge(*pmd
)) {
4327 *start
= address
& PMD_MASK
;
4328 *end
= *start
+ PMD_SIZE
;
4329 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4331 *ptlp
= pmd_lock(mm
, pmd
);
4332 if (pmd_huge(*pmd
)) {
4338 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4341 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4345 *start
= address
& PAGE_MASK
;
4346 *end
= *start
+ PAGE_SIZE
;
4347 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4349 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4350 if (!pte_present(*ptep
))
4355 pte_unmap_unlock(ptep
, *ptlp
);
4357 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4362 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4363 pte_t
**ptepp
, spinlock_t
**ptlp
)
4367 /* (void) is needed to make gcc happy */
4368 (void) __cond_lock(*ptlp
,
4369 !(res
= __follow_pte_pmd(mm
, address
, NULL
, NULL
,
4370 ptepp
, NULL
, ptlp
)));
4374 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4375 unsigned long *start
, unsigned long *end
,
4376 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4380 /* (void) is needed to make gcc happy */
4381 (void) __cond_lock(*ptlp
,
4382 !(res
= __follow_pte_pmd(mm
, address
, start
, end
,
4383 ptepp
, pmdpp
, ptlp
)));
4386 EXPORT_SYMBOL(follow_pte_pmd
);
4389 * follow_pfn - look up PFN at a user virtual address
4390 * @vma: memory mapping
4391 * @address: user virtual address
4392 * @pfn: location to store found PFN
4394 * Only IO mappings and raw PFN mappings are allowed.
4396 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4398 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4405 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4408 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4411 *pfn
= pte_pfn(*ptep
);
4412 pte_unmap_unlock(ptep
, ptl
);
4415 EXPORT_SYMBOL(follow_pfn
);
4417 #ifdef CONFIG_HAVE_IOREMAP_PROT
4418 int follow_phys(struct vm_area_struct
*vma
,
4419 unsigned long address
, unsigned int flags
,
4420 unsigned long *prot
, resource_size_t
*phys
)
4426 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4429 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4433 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4436 *prot
= pgprot_val(pte_pgprot(pte
));
4437 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4441 pte_unmap_unlock(ptep
, ptl
);
4446 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4447 void *buf
, int len
, int write
)
4449 resource_size_t phys_addr
;
4450 unsigned long prot
= 0;
4451 void __iomem
*maddr
;
4452 int offset
= addr
& (PAGE_SIZE
-1);
4454 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4457 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4462 memcpy_toio(maddr
+ offset
, buf
, len
);
4464 memcpy_fromio(buf
, maddr
+ offset
, len
);
4469 EXPORT_SYMBOL_GPL(generic_access_phys
);
4473 * Access another process' address space as given in mm. If non-NULL, use the
4474 * given task for page fault accounting.
4476 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4477 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4479 struct vm_area_struct
*vma
;
4480 void *old_buf
= buf
;
4481 int write
= gup_flags
& FOLL_WRITE
;
4483 down_read(&mm
->mmap_sem
);
4484 /* ignore errors, just check how much was successfully transferred */
4486 int bytes
, ret
, offset
;
4488 struct page
*page
= NULL
;
4490 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4491 gup_flags
, &page
, &vma
, NULL
);
4493 #ifndef CONFIG_HAVE_IOREMAP_PROT
4497 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4498 * we can access using slightly different code.
4500 vma
= find_vma(mm
, addr
);
4501 if (!vma
|| vma
->vm_start
> addr
)
4503 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4504 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4512 offset
= addr
& (PAGE_SIZE
-1);
4513 if (bytes
> PAGE_SIZE
-offset
)
4514 bytes
= PAGE_SIZE
-offset
;
4518 copy_to_user_page(vma
, page
, addr
,
4519 maddr
+ offset
, buf
, bytes
);
4520 set_page_dirty_lock(page
);
4522 copy_from_user_page(vma
, page
, addr
,
4523 buf
, maddr
+ offset
, bytes
);
4532 up_read(&mm
->mmap_sem
);
4534 return buf
- old_buf
;
4538 * access_remote_vm - access another process' address space
4539 * @mm: the mm_struct of the target address space
4540 * @addr: start address to access
4541 * @buf: source or destination buffer
4542 * @len: number of bytes to transfer
4543 * @gup_flags: flags modifying lookup behaviour
4545 * The caller must hold a reference on @mm.
4547 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4548 void *buf
, int len
, unsigned int gup_flags
)
4550 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4554 * Access another process' address space.
4555 * Source/target buffer must be kernel space,
4556 * Do not walk the page table directly, use get_user_pages
4558 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4559 void *buf
, int len
, unsigned int gup_flags
)
4561 struct mm_struct
*mm
;
4564 mm
= get_task_mm(tsk
);
4568 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4574 EXPORT_SYMBOL_GPL(access_process_vm
);
4577 * Print the name of a VMA.
4579 void print_vma_addr(char *prefix
, unsigned long ip
)
4581 struct mm_struct
*mm
= current
->mm
;
4582 struct vm_area_struct
*vma
;
4585 * we might be running from an atomic context so we cannot sleep
4587 if (!down_read_trylock(&mm
->mmap_sem
))
4590 vma
= find_vma(mm
, ip
);
4591 if (vma
&& vma
->vm_file
) {
4592 struct file
*f
= vma
->vm_file
;
4593 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4597 p
= file_path(f
, buf
, PAGE_SIZE
);
4600 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4602 vma
->vm_end
- vma
->vm_start
);
4603 free_page((unsigned long)buf
);
4606 up_read(&mm
->mmap_sem
);
4609 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4610 void __might_fault(const char *file
, int line
)
4613 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4614 * holding the mmap_sem, this is safe because kernel memory doesn't
4615 * get paged out, therefore we'll never actually fault, and the
4616 * below annotations will generate false positives.
4618 if (uaccess_kernel())
4620 if (pagefault_disabled())
4622 __might_sleep(file
, line
, 0);
4623 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4625 might_lock_read(¤t
->mm
->mmap_sem
);
4628 EXPORT_SYMBOL(__might_fault
);
4631 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4632 static void clear_gigantic_page(struct page
*page
,
4634 unsigned int pages_per_huge_page
)
4637 struct page
*p
= page
;
4640 for (i
= 0; i
< pages_per_huge_page
;
4641 i
++, p
= mem_map_next(p
, page
, i
)) {
4643 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4646 void clear_huge_page(struct page
*page
,
4647 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4650 unsigned long addr
= addr_hint
&
4651 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4653 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4654 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4658 /* Clear sub-page to access last to keep its cache lines hot */
4660 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4661 if (2 * n
<= pages_per_huge_page
) {
4662 /* If sub-page to access in first half of huge page */
4665 /* Clear sub-pages at the end of huge page */
4666 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4668 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4671 /* If sub-page to access in second half of huge page */
4672 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4673 l
= pages_per_huge_page
- n
;
4674 /* Clear sub-pages at the begin of huge page */
4675 for (i
= 0; i
< base
; i
++) {
4677 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4681 * Clear remaining sub-pages in left-right-left-right pattern
4682 * towards the sub-page to access
4684 for (i
= 0; i
< l
; i
++) {
4685 int left_idx
= base
+ i
;
4686 int right_idx
= base
+ 2 * l
- 1 - i
;
4689 clear_user_highpage(page
+ left_idx
,
4690 addr
+ left_idx
* PAGE_SIZE
);
4692 clear_user_highpage(page
+ right_idx
,
4693 addr
+ right_idx
* PAGE_SIZE
);
4697 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4699 struct vm_area_struct
*vma
,
4700 unsigned int pages_per_huge_page
)
4703 struct page
*dst_base
= dst
;
4704 struct page
*src_base
= src
;
4706 for (i
= 0; i
< pages_per_huge_page
; ) {
4708 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4711 dst
= mem_map_next(dst
, dst_base
, i
);
4712 src
= mem_map_next(src
, src_base
, i
);
4716 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4717 unsigned long addr
, struct vm_area_struct
*vma
,
4718 unsigned int pages_per_huge_page
)
4722 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4723 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4724 pages_per_huge_page
);
4729 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4731 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4735 long copy_huge_page_from_user(struct page
*dst_page
,
4736 const void __user
*usr_src
,
4737 unsigned int pages_per_huge_page
,
4738 bool allow_pagefault
)
4740 void *src
= (void *)usr_src
;
4742 unsigned long i
, rc
= 0;
4743 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4745 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4746 if (allow_pagefault
)
4747 page_kaddr
= kmap(dst_page
+ i
);
4749 page_kaddr
= kmap_atomic(dst_page
+ i
);
4750 rc
= copy_from_user(page_kaddr
,
4751 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4753 if (allow_pagefault
)
4754 kunmap(dst_page
+ i
);
4756 kunmap_atomic(page_kaddr
);
4758 ret_val
-= (PAGE_SIZE
- rc
);
4766 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4768 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4770 static struct kmem_cache
*page_ptl_cachep
;
4772 void __init
ptlock_cache_init(void)
4774 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4778 bool ptlock_alloc(struct page
*page
)
4782 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4789 void ptlock_free(struct page
*page
)
4791 kmem_cache_free(page_ptl_cachep
, page
->ptl
);