4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr
;
90 EXPORT_SYMBOL(max_mapnr
);
93 EXPORT_SYMBOL(mem_map
);
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
104 EXPORT_SYMBOL(high_memory
);
107 * Randomize the address space (stacks, mmaps, brk, etc.).
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
112 int randomize_va_space __read_mostly
=
113 #ifdef CONFIG_COMPAT_BRK
119 static int __init
disable_randmaps(char *s
)
121 randomize_va_space
= 0;
124 __setup("norandmaps", disable_randmaps
);
126 unsigned long zero_pfn __read_mostly
;
127 EXPORT_SYMBOL(zero_pfn
);
129 unsigned long highest_memmap_pfn __read_mostly
;
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 static int __init
init_zero_pfn(void)
136 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
139 core_initcall(init_zero_pfn
);
142 #if defined(SPLIT_RSS_COUNTING)
144 void sync_mm_rss(struct mm_struct
*mm
)
148 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
149 if (current
->rss_stat
.count
[i
]) {
150 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
151 current
->rss_stat
.count
[i
] = 0;
154 current
->rss_stat
.events
= 0;
157 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
159 struct task_struct
*task
= current
;
161 if (likely(task
->mm
== mm
))
162 task
->rss_stat
.count
[member
] += val
;
164 add_mm_counter(mm
, member
, val
);
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH (64)
171 static void check_sync_rss_stat(struct task_struct
*task
)
173 if (unlikely(task
!= current
))
175 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
176 sync_mm_rss(task
->mm
);
178 #else /* SPLIT_RSS_COUNTING */
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183 static void check_sync_rss_stat(struct task_struct
*task
)
187 #endif /* SPLIT_RSS_COUNTING */
189 #ifdef HAVE_GENERIC_MMU_GATHER
191 static bool tlb_next_batch(struct mmu_gather
*tlb
)
193 struct mmu_gather_batch
*batch
;
197 tlb
->active
= batch
->next
;
201 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
204 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
211 batch
->max
= MAX_GATHER_BATCH
;
213 tlb
->active
->next
= batch
;
219 void arch_tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
220 unsigned long start
, unsigned long end
)
224 /* Is it from 0 to ~0? */
225 tlb
->fullmm
= !(start
| (end
+1));
226 tlb
->need_flush_all
= 0;
227 tlb
->local
.next
= NULL
;
229 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
230 tlb
->active
= &tlb
->local
;
231 tlb
->batch_count
= 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238 __tlb_reset_range(tlb
);
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
247 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb
);
251 __tlb_reset_range(tlb
);
254 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
256 struct mmu_gather_batch
*batch
;
258 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
259 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
262 tlb
->active
= &tlb
->local
;
265 void tlb_flush_mmu(struct mmu_gather
*tlb
)
267 tlb_flush_mmu_tlbonly(tlb
);
268 tlb_flush_mmu_free(tlb
);
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void arch_tlb_finish_mmu(struct mmu_gather
*tlb
,
276 unsigned long start
, unsigned long end
, bool force
)
278 struct mmu_gather_batch
*batch
, *next
;
281 __tlb_adjust_range(tlb
, start
, end
- start
);
285 /* keep the page table cache within bounds */
288 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
290 free_pages((unsigned long)batch
, 0);
292 tlb
->local
.next
= NULL
;
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
302 bool __tlb_remove_page_size(struct mmu_gather
*tlb
, struct page
*page
, int page_size
)
304 struct mmu_gather_batch
*batch
;
306 VM_BUG_ON(!tlb
->end
);
307 VM_WARN_ON(tlb
->page_size
!= page_size
);
311 * Add the page and check if we are full. If so
314 batch
->pages
[batch
->nr
++] = page
;
315 if (batch
->nr
== batch
->max
) {
316 if (!tlb_next_batch(tlb
))
320 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
325 #endif /* HAVE_GENERIC_MMU_GATHER */
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
329 static void tlb_remove_table_smp_sync(void *arg
)
331 struct mm_struct __maybe_unused
*mm
= arg
;
333 * On most architectures this does nothing. Simply delivering the
334 * interrupt is enough to prevent races with software page table
335 * walking like that done in get_user_pages_fast.
337 * See the comment near struct mmu_table_batch.
339 tlb_flush_remove_tables_local(mm
);
342 static void tlb_remove_table_one(void *table
, struct mmu_gather
*tlb
)
345 * This isn't an RCU grace period and hence the page-tables cannot be
346 * assumed to be actually RCU-freed.
348 * It is however sufficient for software page-table walkers that rely on
349 * IRQ disabling. See the comment near struct mmu_table_batch.
351 smp_call_function(tlb_remove_table_smp_sync
, tlb
->mm
, 1);
352 __tlb_remove_table(table
);
355 static void tlb_remove_table_rcu(struct rcu_head
*head
)
357 struct mmu_table_batch
*batch
;
360 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
362 for (i
= 0; i
< batch
->nr
; i
++)
363 __tlb_remove_table(batch
->tables
[i
]);
365 free_page((unsigned long)batch
);
368 void tlb_table_flush(struct mmu_gather
*tlb
)
370 struct mmu_table_batch
**batch
= &tlb
->batch
;
372 tlb_flush_remove_tables(tlb
->mm
);
375 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
380 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
382 struct mmu_table_batch
**batch
= &tlb
->batch
;
385 * When there's less then two users of this mm there cannot be a
386 * concurrent page-table walk.
388 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
389 __tlb_remove_table(table
);
393 if (*batch
== NULL
) {
394 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
395 if (*batch
== NULL
) {
396 tlb_remove_table_one(table
, tlb
);
401 (*batch
)->tables
[(*batch
)->nr
++] = table
;
402 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
403 tlb_table_flush(tlb
);
406 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
409 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
410 * @tlb: the mmu_gather structure to initialize
411 * @mm: the mm_struct of the target address space
412 * @start: start of the region that will be removed from the page-table
413 * @end: end of the region that will be removed from the page-table
415 * Called to initialize an (on-stack) mmu_gather structure for page-table
416 * tear-down from @mm. The @start and @end are set to 0 and -1
417 * respectively when @mm is without users and we're going to destroy
418 * the full address space (exit/execve).
420 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
,
421 unsigned long start
, unsigned long end
)
423 arch_tlb_gather_mmu(tlb
, mm
, start
, end
);
424 inc_tlb_flush_pending(tlb
->mm
);
427 void tlb_finish_mmu(struct mmu_gather
*tlb
,
428 unsigned long start
, unsigned long end
)
431 * If there are parallel threads are doing PTE changes on same range
432 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
433 * flush by batching, a thread has stable TLB entry can fail to flush
434 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
435 * forcefully if we detect parallel PTE batching threads.
437 bool force
= mm_tlb_flush_nested(tlb
->mm
);
439 arch_tlb_finish_mmu(tlb
, start
, end
, force
);
440 dec_tlb_flush_pending(tlb
->mm
);
444 * Note: this doesn't free the actual pages themselves. That
445 * has been handled earlier when unmapping all the memory regions.
447 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
450 pgtable_t token
= pmd_pgtable(*pmd
);
452 pte_free_tlb(tlb
, token
, addr
);
453 mm_dec_nr_ptes(tlb
->mm
);
456 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
457 unsigned long addr
, unsigned long end
,
458 unsigned long floor
, unsigned long ceiling
)
465 pmd
= pmd_offset(pud
, addr
);
467 next
= pmd_addr_end(addr
, end
);
468 if (pmd_none_or_clear_bad(pmd
))
470 free_pte_range(tlb
, pmd
, addr
);
471 } while (pmd
++, addr
= next
, addr
!= end
);
481 if (end
- 1 > ceiling
- 1)
484 pmd
= pmd_offset(pud
, start
);
486 pmd_free_tlb(tlb
, pmd
, start
);
487 mm_dec_nr_pmds(tlb
->mm
);
490 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
491 unsigned long addr
, unsigned long end
,
492 unsigned long floor
, unsigned long ceiling
)
499 pud
= pud_offset(p4d
, addr
);
501 next
= pud_addr_end(addr
, end
);
502 if (pud_none_or_clear_bad(pud
))
504 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
505 } while (pud
++, addr
= next
, addr
!= end
);
515 if (end
- 1 > ceiling
- 1)
518 pud
= pud_offset(p4d
, start
);
520 pud_free_tlb(tlb
, pud
, start
);
521 mm_dec_nr_puds(tlb
->mm
);
524 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
525 unsigned long addr
, unsigned long end
,
526 unsigned long floor
, unsigned long ceiling
)
533 p4d
= p4d_offset(pgd
, addr
);
535 next
= p4d_addr_end(addr
, end
);
536 if (p4d_none_or_clear_bad(p4d
))
538 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
539 } while (p4d
++, addr
= next
, addr
!= end
);
545 ceiling
&= PGDIR_MASK
;
549 if (end
- 1 > ceiling
- 1)
552 p4d
= p4d_offset(pgd
, start
);
554 p4d_free_tlb(tlb
, p4d
, start
);
558 * This function frees user-level page tables of a process.
560 void free_pgd_range(struct mmu_gather
*tlb
,
561 unsigned long addr
, unsigned long end
,
562 unsigned long floor
, unsigned long ceiling
)
568 * The next few lines have given us lots of grief...
570 * Why are we testing PMD* at this top level? Because often
571 * there will be no work to do at all, and we'd prefer not to
572 * go all the way down to the bottom just to discover that.
574 * Why all these "- 1"s? Because 0 represents both the bottom
575 * of the address space and the top of it (using -1 for the
576 * top wouldn't help much: the masks would do the wrong thing).
577 * The rule is that addr 0 and floor 0 refer to the bottom of
578 * the address space, but end 0 and ceiling 0 refer to the top
579 * Comparisons need to use "end - 1" and "ceiling - 1" (though
580 * that end 0 case should be mythical).
582 * Wherever addr is brought up or ceiling brought down, we must
583 * be careful to reject "the opposite 0" before it confuses the
584 * subsequent tests. But what about where end is brought down
585 * by PMD_SIZE below? no, end can't go down to 0 there.
587 * Whereas we round start (addr) and ceiling down, by different
588 * masks at different levels, in order to test whether a table
589 * now has no other vmas using it, so can be freed, we don't
590 * bother to round floor or end up - the tests don't need that.
604 if (end
- 1 > ceiling
- 1)
609 * We add page table cache pages with PAGE_SIZE,
610 * (see pte_free_tlb()), flush the tlb if we need
612 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
613 pgd
= pgd_offset(tlb
->mm
, addr
);
615 next
= pgd_addr_end(addr
, end
);
616 if (pgd_none_or_clear_bad(pgd
))
618 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
619 } while (pgd
++, addr
= next
, addr
!= end
);
622 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
623 unsigned long floor
, unsigned long ceiling
)
626 struct vm_area_struct
*next
= vma
->vm_next
;
627 unsigned long addr
= vma
->vm_start
;
630 * Hide vma from rmap and truncate_pagecache before freeing
633 unlink_anon_vmas(vma
);
634 unlink_file_vma(vma
);
636 if (is_vm_hugetlb_page(vma
)) {
637 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
638 floor
, next
? next
->vm_start
: ceiling
);
641 * Optimization: gather nearby vmas into one call down
643 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
644 && !is_vm_hugetlb_page(next
)) {
647 unlink_anon_vmas(vma
);
648 unlink_file_vma(vma
);
650 free_pgd_range(tlb
, addr
, vma
->vm_end
,
651 floor
, next
? next
->vm_start
: ceiling
);
657 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
660 pgtable_t
new = pte_alloc_one(mm
, address
);
665 * Ensure all pte setup (eg. pte page lock and page clearing) are
666 * visible before the pte is made visible to other CPUs by being
667 * put into page tables.
669 * The other side of the story is the pointer chasing in the page
670 * table walking code (when walking the page table without locking;
671 * ie. most of the time). Fortunately, these data accesses consist
672 * of a chain of data-dependent loads, meaning most CPUs (alpha
673 * being the notable exception) will already guarantee loads are
674 * seen in-order. See the alpha page table accessors for the
675 * smp_read_barrier_depends() barriers in page table walking code.
677 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
679 ptl
= pmd_lock(mm
, pmd
);
680 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
682 pmd_populate(mm
, pmd
, new);
691 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
693 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
697 smp_wmb(); /* See comment in __pte_alloc */
699 spin_lock(&init_mm
.page_table_lock
);
700 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
701 pmd_populate_kernel(&init_mm
, pmd
, new);
704 spin_unlock(&init_mm
.page_table_lock
);
706 pte_free_kernel(&init_mm
, new);
710 static inline void init_rss_vec(int *rss
)
712 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
715 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
719 if (current
->mm
== mm
)
721 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
723 add_mm_counter(mm
, i
, rss
[i
]);
727 * This function is called to print an error when a bad pte
728 * is found. For example, we might have a PFN-mapped pte in
729 * a region that doesn't allow it.
731 * The calling function must still handle the error.
733 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
734 pte_t pte
, struct page
*page
)
736 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
737 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
738 pud_t
*pud
= pud_offset(p4d
, addr
);
739 pmd_t
*pmd
= pmd_offset(pud
, addr
);
740 struct address_space
*mapping
;
742 static unsigned long resume
;
743 static unsigned long nr_shown
;
744 static unsigned long nr_unshown
;
747 * Allow a burst of 60 reports, then keep quiet for that minute;
748 * or allow a steady drip of one report per second.
750 if (nr_shown
== 60) {
751 if (time_before(jiffies
, resume
)) {
756 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
763 resume
= jiffies
+ 60 * HZ
;
765 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
766 index
= linear_page_index(vma
, addr
);
768 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
770 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
772 dump_page(page
, "bad pte");
773 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
774 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
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 struct page
*_vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
827 pte_t pte
, bool with_public_device
)
829 unsigned long pfn
= pte_pfn(pte
);
831 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
832 if (likely(!pte_special(pte
)))
834 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
835 return vma
->vm_ops
->find_special_page(vma
, addr
);
836 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
838 if (is_zero_pfn(pfn
))
842 * Device public pages are special pages (they are ZONE_DEVICE
843 * pages but different from persistent memory). They behave
844 * allmost like normal pages. The difference is that they are
845 * not on the lru and thus should never be involve with any-
846 * thing that involve lru manipulation (mlock, numa balancing,
849 * This is why we still want to return NULL for such page from
850 * vm_normal_page() so that we do not have to special case all
851 * call site of vm_normal_page().
853 if (likely(pfn
<= highest_memmap_pfn
)) {
854 struct page
*page
= pfn_to_page(pfn
);
856 if (is_device_public_page(page
)) {
857 if (with_public_device
)
866 print_bad_pte(vma
, addr
, pte
, NULL
);
870 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
872 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
873 if (vma
->vm_flags
& VM_MIXEDMAP
) {
879 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
880 if (pfn
== vma
->vm_pgoff
+ off
)
882 if (!is_cow_mapping(vma
->vm_flags
))
887 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 * !CONFIG_ARCH_HAS_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
))
932 if (is_zero_pfn(pfn
))
934 if (unlikely(pfn
> highest_memmap_pfn
))
938 * NOTE! We still have PageReserved() pages in the page tables.
939 * eg. VDSO mappings can cause them to exist.
942 return pfn_to_page(pfn
);
947 * copy one vm_area from one task to the other. Assumes the page tables
948 * already present in the new task to be cleared in the whole range
949 * covered by this vma.
952 static inline unsigned long
953 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
954 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
955 unsigned long addr
, int *rss
)
957 unsigned long vm_flags
= vma
->vm_flags
;
958 pte_t pte
= *src_pte
;
961 /* pte contains position in swap or file, so copy. */
962 if (unlikely(!pte_present(pte
))) {
963 swp_entry_t entry
= pte_to_swp_entry(pte
);
965 if (likely(!non_swap_entry(entry
))) {
966 if (swap_duplicate(entry
) < 0)
969 /* make sure dst_mm is on swapoff's mmlist. */
970 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
971 spin_lock(&mmlist_lock
);
972 if (list_empty(&dst_mm
->mmlist
))
973 list_add(&dst_mm
->mmlist
,
975 spin_unlock(&mmlist_lock
);
978 } else if (is_migration_entry(entry
)) {
979 page
= migration_entry_to_page(entry
);
981 rss
[mm_counter(page
)]++;
983 if (is_write_migration_entry(entry
) &&
984 is_cow_mapping(vm_flags
)) {
986 * COW mappings require pages in both
987 * parent and child to be set to read.
989 make_migration_entry_read(&entry
);
990 pte
= swp_entry_to_pte(entry
);
991 if (pte_swp_soft_dirty(*src_pte
))
992 pte
= pte_swp_mksoft_dirty(pte
);
993 set_pte_at(src_mm
, addr
, src_pte
, pte
);
995 } else if (is_device_private_entry(entry
)) {
996 page
= device_private_entry_to_page(entry
);
999 * Update rss count even for unaddressable pages, as
1000 * they should treated just like normal pages in this
1003 * We will likely want to have some new rss counters
1004 * for unaddressable pages, at some point. But for now
1005 * keep things as they are.
1008 rss
[mm_counter(page
)]++;
1009 page_dup_rmap(page
, false);
1012 * We do not preserve soft-dirty information, because so
1013 * far, checkpoint/restore is the only feature that
1014 * requires that. And checkpoint/restore does not work
1015 * when a device driver is involved (you cannot easily
1016 * save and restore device driver state).
1018 if (is_write_device_private_entry(entry
) &&
1019 is_cow_mapping(vm_flags
)) {
1020 make_device_private_entry_read(&entry
);
1021 pte
= swp_entry_to_pte(entry
);
1022 set_pte_at(src_mm
, addr
, src_pte
, pte
);
1029 * If it's a COW mapping, write protect it both
1030 * in the parent and the child
1032 if (is_cow_mapping(vm_flags
)) {
1033 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
1034 pte
= pte_wrprotect(pte
);
1038 * If it's a shared mapping, mark it clean in
1041 if (vm_flags
& VM_SHARED
)
1042 pte
= pte_mkclean(pte
);
1043 pte
= pte_mkold(pte
);
1045 page
= vm_normal_page(vma
, addr
, pte
);
1048 page_dup_rmap(page
, false);
1049 rss
[mm_counter(page
)]++;
1050 } else if (pte_devmap(pte
)) {
1051 page
= pte_page(pte
);
1054 * Cache coherent device memory behave like regular page and
1055 * not like persistent memory page. For more informations see
1056 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1058 if (is_device_public_page(page
)) {
1060 page_dup_rmap(page
, false);
1061 rss
[mm_counter(page
)]++;
1066 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
1070 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1071 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
1072 unsigned long addr
, unsigned long end
)
1074 pte_t
*orig_src_pte
, *orig_dst_pte
;
1075 pte_t
*src_pte
, *dst_pte
;
1076 spinlock_t
*src_ptl
, *dst_ptl
;
1078 int rss
[NR_MM_COUNTERS
];
1079 swp_entry_t entry
= (swp_entry_t
){0};
1084 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1087 src_pte
= pte_offset_map(src_pmd
, addr
);
1088 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
1089 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1090 orig_src_pte
= src_pte
;
1091 orig_dst_pte
= dst_pte
;
1092 arch_enter_lazy_mmu_mode();
1096 * We are holding two locks at this point - either of them
1097 * could generate latencies in another task on another CPU.
1099 if (progress
>= 32) {
1101 if (need_resched() ||
1102 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1105 if (pte_none(*src_pte
)) {
1109 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
1114 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1116 arch_leave_lazy_mmu_mode();
1117 spin_unlock(src_ptl
);
1118 pte_unmap(orig_src_pte
);
1119 add_mm_rss_vec(dst_mm
, rss
);
1120 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1124 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
1133 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1134 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
1135 unsigned long addr
, unsigned long end
)
1137 pmd_t
*src_pmd
, *dst_pmd
;
1140 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1143 src_pmd
= pmd_offset(src_pud
, addr
);
1145 next
= pmd_addr_end(addr
, end
);
1146 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1147 || pmd_devmap(*src_pmd
)) {
1149 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, vma
);
1150 err
= copy_huge_pmd(dst_mm
, src_mm
,
1151 dst_pmd
, src_pmd
, addr
, vma
);
1158 if (pmd_none_or_clear_bad(src_pmd
))
1160 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1163 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1167 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1168 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
1169 unsigned long addr
, unsigned long end
)
1171 pud_t
*src_pud
, *dst_pud
;
1174 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1177 src_pud
= pud_offset(src_p4d
, addr
);
1179 next
= pud_addr_end(addr
, end
);
1180 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1183 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
1184 err
= copy_huge_pud(dst_mm
, src_mm
,
1185 dst_pud
, src_pud
, addr
, vma
);
1192 if (pud_none_or_clear_bad(src_pud
))
1194 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1197 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1201 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1202 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1203 unsigned long addr
, unsigned long end
)
1205 p4d_t
*src_p4d
, *dst_p4d
;
1208 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1211 src_p4d
= p4d_offset(src_pgd
, addr
);
1213 next
= p4d_addr_end(addr
, end
);
1214 if (p4d_none_or_clear_bad(src_p4d
))
1216 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
1219 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1223 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1224 struct vm_area_struct
*vma
)
1226 pgd_t
*src_pgd
, *dst_pgd
;
1228 unsigned long addr
= vma
->vm_start
;
1229 unsigned long end
= vma
->vm_end
;
1230 unsigned long mmun_start
; /* For mmu_notifiers */
1231 unsigned long mmun_end
; /* For mmu_notifiers */
1236 * Don't copy ptes where a page fault will fill them correctly.
1237 * Fork becomes much lighter when there are big shared or private
1238 * readonly mappings. The tradeoff is that copy_page_range is more
1239 * efficient than faulting.
1241 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1245 if (is_vm_hugetlb_page(vma
))
1246 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1248 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1250 * We do not free on error cases below as remove_vma
1251 * gets called on error from higher level routine
1253 ret
= track_pfn_copy(vma
);
1259 * We need to invalidate the secondary MMU mappings only when
1260 * there could be a permission downgrade on the ptes of the
1261 * parent mm. And a permission downgrade will only happen if
1262 * is_cow_mapping() returns true.
1264 is_cow
= is_cow_mapping(vma
->vm_flags
);
1268 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1272 dst_pgd
= pgd_offset(dst_mm
, addr
);
1273 src_pgd
= pgd_offset(src_mm
, addr
);
1275 next
= pgd_addr_end(addr
, end
);
1276 if (pgd_none_or_clear_bad(src_pgd
))
1278 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1279 vma
, addr
, next
))) {
1283 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1286 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1290 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1291 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1292 unsigned long addr
, unsigned long end
,
1293 struct zap_details
*details
)
1295 struct mm_struct
*mm
= tlb
->mm
;
1296 int force_flush
= 0;
1297 int rss
[NR_MM_COUNTERS
];
1303 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
1306 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1308 flush_tlb_batched_pending(mm
);
1309 arch_enter_lazy_mmu_mode();
1312 if (pte_none(ptent
))
1315 if (pte_present(ptent
)) {
1318 page
= _vm_normal_page(vma
, addr
, ptent
, true);
1319 if (unlikely(details
) && page
) {
1321 * unmap_shared_mapping_pages() wants to
1322 * invalidate cache without truncating:
1323 * unmap shared but keep private pages.
1325 if (details
->check_mapping
&&
1326 details
->check_mapping
!= page_rmapping(page
))
1329 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1331 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1332 if (unlikely(!page
))
1335 if (!PageAnon(page
)) {
1336 if (pte_dirty(ptent
)) {
1338 set_page_dirty(page
);
1340 if (pte_young(ptent
) &&
1341 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1342 mark_page_accessed(page
);
1344 rss
[mm_counter(page
)]--;
1345 page_remove_rmap(page
, false);
1346 if (unlikely(page_mapcount(page
) < 0))
1347 print_bad_pte(vma
, addr
, ptent
, page
);
1348 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1356 entry
= pte_to_swp_entry(ptent
);
1357 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1358 struct page
*page
= device_private_entry_to_page(entry
);
1360 if (unlikely(details
&& details
->check_mapping
)) {
1362 * unmap_shared_mapping_pages() wants to
1363 * invalidate cache without truncating:
1364 * unmap shared but keep private pages.
1366 if (details
->check_mapping
!=
1367 page_rmapping(page
))
1371 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1372 rss
[mm_counter(page
)]--;
1373 page_remove_rmap(page
, false);
1378 /* If details->check_mapping, we leave swap entries. */
1379 if (unlikely(details
))
1382 entry
= pte_to_swp_entry(ptent
);
1383 if (!non_swap_entry(entry
))
1385 else if (is_migration_entry(entry
)) {
1388 page
= migration_entry_to_page(entry
);
1389 rss
[mm_counter(page
)]--;
1391 if (unlikely(!free_swap_and_cache(entry
)))
1392 print_bad_pte(vma
, addr
, ptent
, NULL
);
1393 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1394 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1396 add_mm_rss_vec(mm
, rss
);
1397 arch_leave_lazy_mmu_mode();
1399 /* Do the actual TLB flush before dropping ptl */
1401 tlb_flush_mmu_tlbonly(tlb
);
1402 pte_unmap_unlock(start_pte
, ptl
);
1405 * If we forced a TLB flush (either due to running out of
1406 * batch buffers or because we needed to flush dirty TLB
1407 * entries before releasing the ptl), free the batched
1408 * memory too. Restart if we didn't do everything.
1412 tlb_flush_mmu_free(tlb
);
1420 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1421 struct vm_area_struct
*vma
, pud_t
*pud
,
1422 unsigned long addr
, unsigned long end
,
1423 struct zap_details
*details
)
1428 pmd
= pmd_offset(pud
, addr
);
1430 next
= pmd_addr_end(addr
, end
);
1431 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1432 if (next
- addr
!= HPAGE_PMD_SIZE
)
1433 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1434 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1439 * Here there can be other concurrent MADV_DONTNEED or
1440 * trans huge page faults running, and if the pmd is
1441 * none or trans huge it can change under us. This is
1442 * because MADV_DONTNEED holds the mmap_sem in read
1445 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1447 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1450 } while (pmd
++, addr
= next
, addr
!= end
);
1455 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1456 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1457 unsigned long addr
, unsigned long end
,
1458 struct zap_details
*details
)
1463 pud
= pud_offset(p4d
, addr
);
1465 next
= pud_addr_end(addr
, end
);
1466 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1467 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1468 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1469 split_huge_pud(vma
, pud
, addr
);
1470 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1474 if (pud_none_or_clear_bad(pud
))
1476 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1479 } while (pud
++, addr
= next
, addr
!= end
);
1484 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1485 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1486 unsigned long addr
, unsigned long end
,
1487 struct zap_details
*details
)
1492 p4d
= p4d_offset(pgd
, addr
);
1494 next
= p4d_addr_end(addr
, end
);
1495 if (p4d_none_or_clear_bad(p4d
))
1497 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1498 } while (p4d
++, addr
= next
, addr
!= end
);
1503 void unmap_page_range(struct mmu_gather
*tlb
,
1504 struct vm_area_struct
*vma
,
1505 unsigned long addr
, unsigned long end
,
1506 struct zap_details
*details
)
1511 BUG_ON(addr
>= end
);
1512 tlb_start_vma(tlb
, vma
);
1513 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1515 next
= pgd_addr_end(addr
, end
);
1516 if (pgd_none_or_clear_bad(pgd
))
1518 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1519 } while (pgd
++, addr
= next
, addr
!= end
);
1520 tlb_end_vma(tlb
, vma
);
1524 static void unmap_single_vma(struct mmu_gather
*tlb
,
1525 struct vm_area_struct
*vma
, unsigned long start_addr
,
1526 unsigned long end_addr
,
1527 struct zap_details
*details
)
1529 unsigned long start
= max(vma
->vm_start
, start_addr
);
1532 if (start
>= vma
->vm_end
)
1534 end
= min(vma
->vm_end
, end_addr
);
1535 if (end
<= vma
->vm_start
)
1539 uprobe_munmap(vma
, start
, end
);
1541 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1542 untrack_pfn(vma
, 0, 0);
1545 if (unlikely(is_vm_hugetlb_page(vma
))) {
1547 * It is undesirable to test vma->vm_file as it
1548 * should be non-null for valid hugetlb area.
1549 * However, vm_file will be NULL in the error
1550 * cleanup path of mmap_region. When
1551 * hugetlbfs ->mmap method fails,
1552 * mmap_region() nullifies vma->vm_file
1553 * before calling this function to clean up.
1554 * Since no pte has actually been setup, it is
1555 * safe to do nothing in this case.
1558 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1559 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1560 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1563 unmap_page_range(tlb
, vma
, start
, end
, details
);
1568 * unmap_vmas - unmap a range of memory covered by a list of vma's
1569 * @tlb: address of the caller's struct mmu_gather
1570 * @vma: the starting vma
1571 * @start_addr: virtual address at which to start unmapping
1572 * @end_addr: virtual address at which to end unmapping
1574 * Unmap all pages in the vma list.
1576 * Only addresses between `start' and `end' will be unmapped.
1578 * The VMA list must be sorted in ascending virtual address order.
1580 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1581 * range after unmap_vmas() returns. So the only responsibility here is to
1582 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1583 * drops the lock and schedules.
1585 void unmap_vmas(struct mmu_gather
*tlb
,
1586 struct vm_area_struct
*vma
, unsigned long start_addr
,
1587 unsigned long end_addr
)
1589 struct mm_struct
*mm
= vma
->vm_mm
;
1591 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1592 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1593 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1594 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1598 * zap_page_range - remove user pages in a given range
1599 * @vma: vm_area_struct holding the applicable pages
1600 * @start: starting address of pages to zap
1601 * @size: number of bytes to zap
1603 * Caller must protect the VMA list
1605 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1608 struct mm_struct
*mm
= vma
->vm_mm
;
1609 struct mmu_gather tlb
;
1610 unsigned long end
= start
+ size
;
1613 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1614 update_hiwater_rss(mm
);
1615 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1616 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1617 unmap_single_vma(&tlb
, vma
, start
, end
, NULL
);
1618 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1619 tlb_finish_mmu(&tlb
, start
, end
);
1623 * zap_page_range_single - remove user pages in a given range
1624 * @vma: vm_area_struct holding the applicable pages
1625 * @address: starting address of pages to zap
1626 * @size: number of bytes to zap
1627 * @details: details of shared cache invalidation
1629 * The range must fit into one VMA.
1631 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1632 unsigned long size
, struct zap_details
*details
)
1634 struct mm_struct
*mm
= vma
->vm_mm
;
1635 struct mmu_gather tlb
;
1636 unsigned long end
= address
+ size
;
1639 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1640 update_hiwater_rss(mm
);
1641 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1642 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1643 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1644 tlb_finish_mmu(&tlb
, address
, end
);
1648 * zap_vma_ptes - remove ptes mapping the vma
1649 * @vma: vm_area_struct holding ptes to be zapped
1650 * @address: starting address of pages to zap
1651 * @size: number of bytes to zap
1653 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1655 * The entire address range must be fully contained within the vma.
1658 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1661 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1662 !(vma
->vm_flags
& VM_PFNMAP
))
1665 zap_page_range_single(vma
, address
, size
, NULL
);
1667 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1669 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1677 pgd
= pgd_offset(mm
, addr
);
1678 p4d
= p4d_alloc(mm
, pgd
, addr
);
1681 pud
= pud_alloc(mm
, p4d
, addr
);
1684 pmd
= pmd_alloc(mm
, pud
, addr
);
1688 VM_BUG_ON(pmd_trans_huge(*pmd
));
1689 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1693 * This is the old fallback for page remapping.
1695 * For historical reasons, it only allows reserved pages. Only
1696 * old drivers should use this, and they needed to mark their
1697 * pages reserved for the old functions anyway.
1699 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1700 struct page
*page
, pgprot_t prot
)
1702 struct mm_struct
*mm
= vma
->vm_mm
;
1711 flush_dcache_page(page
);
1712 pte
= get_locked_pte(mm
, addr
, &ptl
);
1716 if (!pte_none(*pte
))
1719 /* Ok, finally just insert the thing.. */
1721 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1722 page_add_file_rmap(page
, false);
1723 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1726 pte_unmap_unlock(pte
, ptl
);
1729 pte_unmap_unlock(pte
, ptl
);
1735 * vm_insert_page - insert single page into user vma
1736 * @vma: user vma to map to
1737 * @addr: target user address of this page
1738 * @page: source kernel page
1740 * This allows drivers to insert individual pages they've allocated
1743 * The page has to be a nice clean _individual_ kernel allocation.
1744 * If you allocate a compound page, you need to have marked it as
1745 * such (__GFP_COMP), or manually just split the page up yourself
1746 * (see split_page()).
1748 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1749 * took an arbitrary page protection parameter. This doesn't allow
1750 * that. Your vma protection will have to be set up correctly, which
1751 * means that if you want a shared writable mapping, you'd better
1752 * ask for a shared writable mapping!
1754 * The page does not need to be reserved.
1756 * Usually this function is called from f_op->mmap() handler
1757 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1758 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1759 * function from other places, for example from page-fault handler.
1761 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1764 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1766 if (!page_count(page
))
1768 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1769 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1770 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1771 vma
->vm_flags
|= VM_MIXEDMAP
;
1773 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1775 EXPORT_SYMBOL(vm_insert_page
);
1777 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1778 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1780 struct mm_struct
*mm
= vma
->vm_mm
;
1786 pte
= get_locked_pte(mm
, addr
, &ptl
);
1790 if (!pte_none(*pte
)) {
1793 * For read faults on private mappings the PFN passed
1794 * in may not match the PFN we have mapped if the
1795 * mapped PFN is a writeable COW page. In the mkwrite
1796 * case we are creating a writable PTE for a shared
1797 * mapping and we expect the PFNs to match.
1799 if (WARN_ON_ONCE(pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)))
1807 /* Ok, finally just insert the thing.. */
1808 if (pfn_t_devmap(pfn
))
1809 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1811 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1815 entry
= pte_mkyoung(entry
);
1816 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1819 set_pte_at(mm
, addr
, pte
, entry
);
1820 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1824 pte_unmap_unlock(pte
, ptl
);
1830 * vm_insert_pfn - insert single pfn into user vma
1831 * @vma: user vma to map to
1832 * @addr: target user address of this page
1833 * @pfn: source kernel pfn
1835 * Similar to vm_insert_page, this allows drivers to insert individual pages
1836 * they've allocated into a user vma. Same comments apply.
1838 * This function should only be called from a vm_ops->fault handler, and
1839 * in that case the handler should return NULL.
1841 * vma cannot be a COW mapping.
1843 * As this is called only for pages that do not currently exist, we
1844 * do not need to flush old virtual caches or the TLB.
1846 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1849 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1851 EXPORT_SYMBOL(vm_insert_pfn
);
1854 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1855 * @vma: user vma to map to
1856 * @addr: target user address of this page
1857 * @pfn: source kernel pfn
1858 * @pgprot: pgprot flags for the inserted page
1860 * This is exactly like vm_insert_pfn, except that it allows drivers to
1861 * to override pgprot on a per-page basis.
1863 * This only makes sense for IO mappings, and it makes no sense for
1864 * cow mappings. In general, using multiple vmas is preferable;
1865 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1868 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1869 unsigned long pfn
, pgprot_t pgprot
)
1873 * Technically, architectures with pte_special can avoid all these
1874 * restrictions (same for remap_pfn_range). However we would like
1875 * consistency in testing and feature parity among all, so we should
1876 * try to keep these invariants in place for everybody.
1878 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1879 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1880 (VM_PFNMAP
|VM_MIXEDMAP
));
1881 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1882 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1884 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1887 if (!pfn_modify_allowed(pfn
, pgprot
))
1890 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1892 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1897 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1899 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
1901 /* these checks mirror the abort conditions in vm_normal_page */
1902 if (vma
->vm_flags
& VM_MIXEDMAP
)
1904 if (pfn_t_devmap(pfn
))
1906 if (pfn_t_special(pfn
))
1908 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
1913 static int __vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1914 pfn_t pfn
, bool mkwrite
)
1916 pgprot_t pgprot
= vma
->vm_page_prot
;
1918 BUG_ON(!vm_mixed_ok(vma
, pfn
));
1920 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1923 track_pfn_insert(vma
, &pgprot
, pfn
);
1925 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1929 * If we don't have pte special, then we have to use the pfn_valid()
1930 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1931 * refcount the page if pfn_valid is true (hence insert_page rather
1932 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1933 * without pte special, it would there be refcounted as a normal page.
1935 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
1936 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1940 * At this point we are committed to insert_page()
1941 * regardless of whether the caller specified flags that
1942 * result in pfn_t_has_page() == false.
1944 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1945 return insert_page(vma
, addr
, page
, pgprot
);
1947 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1950 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1953 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1956 EXPORT_SYMBOL(vm_insert_mixed
);
1959 * If the insertion of PTE failed because someone else already added a
1960 * different entry in the mean time, we treat that as success as we assume
1961 * the same entry was actually inserted.
1964 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
1965 unsigned long addr
, pfn_t pfn
)
1969 err
= __vm_insert_mixed(vma
, addr
, pfn
, true);
1971 return VM_FAULT_OOM
;
1972 if (err
< 0 && err
!= -EBUSY
)
1973 return VM_FAULT_SIGBUS
;
1974 return VM_FAULT_NOPAGE
;
1976 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
1979 * maps a range of physical memory into the requested pages. the old
1980 * mappings are removed. any references to nonexistent pages results
1981 * in null mappings (currently treated as "copy-on-access")
1983 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1984 unsigned long addr
, unsigned long end
,
1985 unsigned long pfn
, pgprot_t prot
)
1991 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1994 arch_enter_lazy_mmu_mode();
1996 BUG_ON(!pte_none(*pte
));
1997 if (!pfn_modify_allowed(pfn
, prot
)) {
2001 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2003 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2004 arch_leave_lazy_mmu_mode();
2005 pte_unmap_unlock(pte
- 1, ptl
);
2009 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2010 unsigned long addr
, unsigned long end
,
2011 unsigned long pfn
, pgprot_t prot
)
2017 pfn
-= addr
>> PAGE_SHIFT
;
2018 pmd
= pmd_alloc(mm
, pud
, addr
);
2021 VM_BUG_ON(pmd_trans_huge(*pmd
));
2023 next
= pmd_addr_end(addr
, end
);
2024 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2025 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2028 } while (pmd
++, addr
= next
, addr
!= end
);
2032 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2033 unsigned long addr
, unsigned long end
,
2034 unsigned long pfn
, pgprot_t prot
)
2040 pfn
-= addr
>> PAGE_SHIFT
;
2041 pud
= pud_alloc(mm
, p4d
, addr
);
2045 next
= pud_addr_end(addr
, end
);
2046 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2047 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2050 } while (pud
++, addr
= next
, addr
!= end
);
2054 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2055 unsigned long addr
, unsigned long end
,
2056 unsigned long pfn
, pgprot_t prot
)
2062 pfn
-= addr
>> PAGE_SHIFT
;
2063 p4d
= p4d_alloc(mm
, pgd
, addr
);
2067 next
= p4d_addr_end(addr
, end
);
2068 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2069 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2072 } while (p4d
++, addr
= next
, addr
!= end
);
2077 * remap_pfn_range - remap kernel memory to userspace
2078 * @vma: user vma to map to
2079 * @addr: target user address to start at
2080 * @pfn: physical address of kernel memory
2081 * @size: size of map area
2082 * @prot: page protection flags for this mapping
2084 * Note: this is only safe if the mm semaphore is held when called.
2086 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2087 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2091 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2092 struct mm_struct
*mm
= vma
->vm_mm
;
2093 unsigned long remap_pfn
= pfn
;
2097 * Physically remapped pages are special. Tell the
2098 * rest of the world about it:
2099 * VM_IO tells people not to look at these pages
2100 * (accesses can have side effects).
2101 * VM_PFNMAP tells the core MM that the base pages are just
2102 * raw PFN mappings, and do not have a "struct page" associated
2105 * Disable vma merging and expanding with mremap().
2107 * Omit vma from core dump, even when VM_IO turned off.
2109 * There's a horrible special case to handle copy-on-write
2110 * behaviour that some programs depend on. We mark the "original"
2111 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2112 * See vm_normal_page() for details.
2114 if (is_cow_mapping(vma
->vm_flags
)) {
2115 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2117 vma
->vm_pgoff
= pfn
;
2120 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
2124 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2126 BUG_ON(addr
>= end
);
2127 pfn
-= addr
>> PAGE_SHIFT
;
2128 pgd
= pgd_offset(mm
, addr
);
2129 flush_cache_range(vma
, addr
, end
);
2131 next
= pgd_addr_end(addr
, end
);
2132 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2133 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2136 } while (pgd
++, addr
= next
, addr
!= end
);
2139 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
2143 EXPORT_SYMBOL(remap_pfn_range
);
2146 * vm_iomap_memory - remap memory to userspace
2147 * @vma: user vma to map to
2148 * @start: start of area
2149 * @len: size of area
2151 * This is a simplified io_remap_pfn_range() for common driver use. The
2152 * driver just needs to give us the physical memory range to be mapped,
2153 * we'll figure out the rest from the vma information.
2155 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2156 * whatever write-combining details or similar.
2158 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2160 unsigned long vm_len
, pfn
, pages
;
2162 /* Check that the physical memory area passed in looks valid */
2163 if (start
+ len
< start
)
2166 * You *really* shouldn't map things that aren't page-aligned,
2167 * but we've historically allowed it because IO memory might
2168 * just have smaller alignment.
2170 len
+= start
& ~PAGE_MASK
;
2171 pfn
= start
>> PAGE_SHIFT
;
2172 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2173 if (pfn
+ pages
< pfn
)
2176 /* We start the mapping 'vm_pgoff' pages into the area */
2177 if (vma
->vm_pgoff
> pages
)
2179 pfn
+= vma
->vm_pgoff
;
2180 pages
-= vma
->vm_pgoff
;
2182 /* Can we fit all of the mapping? */
2183 vm_len
= vma
->vm_end
- vma
->vm_start
;
2184 if (vm_len
>> PAGE_SHIFT
> pages
)
2187 /* Ok, let it rip */
2188 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2190 EXPORT_SYMBOL(vm_iomap_memory
);
2192 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2193 unsigned long addr
, unsigned long end
,
2194 pte_fn_t fn
, void *data
)
2199 spinlock_t
*uninitialized_var(ptl
);
2201 pte
= (mm
== &init_mm
) ?
2202 pte_alloc_kernel(pmd
, addr
) :
2203 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2207 BUG_ON(pmd_huge(*pmd
));
2209 arch_enter_lazy_mmu_mode();
2211 token
= pmd_pgtable(*pmd
);
2214 err
= fn(pte
++, token
, addr
, data
);
2217 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2219 arch_leave_lazy_mmu_mode();
2222 pte_unmap_unlock(pte
-1, ptl
);
2226 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2227 unsigned long addr
, unsigned long end
,
2228 pte_fn_t fn
, void *data
)
2234 BUG_ON(pud_huge(*pud
));
2236 pmd
= pmd_alloc(mm
, pud
, addr
);
2240 next
= pmd_addr_end(addr
, end
);
2241 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2244 } while (pmd
++, addr
= next
, addr
!= end
);
2248 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2249 unsigned long addr
, unsigned long end
,
2250 pte_fn_t fn
, void *data
)
2256 pud
= pud_alloc(mm
, p4d
, addr
);
2260 next
= pud_addr_end(addr
, end
);
2261 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2264 } while (pud
++, addr
= next
, addr
!= end
);
2268 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2269 unsigned long addr
, unsigned long end
,
2270 pte_fn_t fn
, void *data
)
2276 p4d
= p4d_alloc(mm
, pgd
, addr
);
2280 next
= p4d_addr_end(addr
, end
);
2281 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
);
2284 } while (p4d
++, addr
= next
, addr
!= end
);
2289 * Scan a region of virtual memory, filling in page tables as necessary
2290 * and calling a provided function on each leaf page table.
2292 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2293 unsigned long size
, pte_fn_t fn
, void *data
)
2297 unsigned long end
= addr
+ size
;
2300 if (WARN_ON(addr
>= end
))
2303 pgd
= pgd_offset(mm
, addr
);
2305 next
= pgd_addr_end(addr
, end
);
2306 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
);
2309 } while (pgd
++, addr
= next
, addr
!= end
);
2313 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2316 * handle_pte_fault chooses page fault handler according to an entry which was
2317 * read non-atomically. Before making any commitment, on those architectures
2318 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2319 * parts, do_swap_page must check under lock before unmapping the pte and
2320 * proceeding (but do_wp_page is only called after already making such a check;
2321 * and do_anonymous_page can safely check later on).
2323 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2324 pte_t
*page_table
, pte_t orig_pte
)
2327 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2328 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2329 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2331 same
= pte_same(*page_table
, orig_pte
);
2335 pte_unmap(page_table
);
2339 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2341 debug_dma_assert_idle(src
);
2344 * If the source page was a PFN mapping, we don't have
2345 * a "struct page" for it. We do a best-effort copy by
2346 * just copying from the original user address. If that
2347 * fails, we just zero-fill it. Live with it.
2349 if (unlikely(!src
)) {
2350 void *kaddr
= kmap_atomic(dst
);
2351 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2354 * This really shouldn't fail, because the page is there
2355 * in the page tables. But it might just be unreadable,
2356 * in which case we just give up and fill the result with
2359 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2361 kunmap_atomic(kaddr
);
2362 flush_dcache_page(dst
);
2364 copy_user_highpage(dst
, src
, va
, vma
);
2367 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2369 struct file
*vm_file
= vma
->vm_file
;
2372 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2375 * Special mappings (e.g. VDSO) do not have any file so fake
2376 * a default GFP_KERNEL for them.
2382 * Notify the address space that the page is about to become writable so that
2383 * it can prohibit this or wait for the page to get into an appropriate state.
2385 * We do this without the lock held, so that it can sleep if it needs to.
2387 static int do_page_mkwrite(struct vm_fault
*vmf
)
2390 struct page
*page
= vmf
->page
;
2391 unsigned int old_flags
= vmf
->flags
;
2393 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2395 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2396 /* Restore original flags so that caller is not surprised */
2397 vmf
->flags
= old_flags
;
2398 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2400 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2402 if (!page
->mapping
) {
2404 return 0; /* retry */
2406 ret
|= VM_FAULT_LOCKED
;
2408 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2413 * Handle dirtying of a page in shared file mapping on a write fault.
2415 * The function expects the page to be locked and unlocks it.
2417 static void fault_dirty_shared_page(struct vm_area_struct
*vma
,
2420 struct address_space
*mapping
;
2422 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2424 dirtied
= set_page_dirty(page
);
2425 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2427 * Take a local copy of the address_space - page.mapping may be zeroed
2428 * by truncate after unlock_page(). The address_space itself remains
2429 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2430 * release semantics to prevent the compiler from undoing this copying.
2432 mapping
= page_rmapping(page
);
2435 if ((dirtied
|| page_mkwrite
) && mapping
) {
2437 * Some device drivers do not set page.mapping
2438 * but still dirty their pages
2440 balance_dirty_pages_ratelimited(mapping
);
2444 file_update_time(vma
->vm_file
);
2448 * Handle write page faults for pages that can be reused in the current vma
2450 * This can happen either due to the mapping being with the VM_SHARED flag,
2451 * or due to us being the last reference standing to the page. In either
2452 * case, all we need to do here is to mark the page as writable and update
2453 * any related book-keeping.
2455 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2456 __releases(vmf
->ptl
)
2458 struct vm_area_struct
*vma
= vmf
->vma
;
2459 struct page
*page
= vmf
->page
;
2462 * Clear the pages cpupid information as the existing
2463 * information potentially belongs to a now completely
2464 * unrelated process.
2467 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2469 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2470 entry
= pte_mkyoung(vmf
->orig_pte
);
2471 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2472 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2473 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2474 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2478 * Handle the case of a page which we actually need to copy to a new page.
2480 * Called with mmap_sem locked and the old page referenced, but
2481 * without the ptl held.
2483 * High level logic flow:
2485 * - Allocate a page, copy the content of the old page to the new one.
2486 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2487 * - Take the PTL. If the pte changed, bail out and release the allocated page
2488 * - If the pte is still the way we remember it, update the page table and all
2489 * relevant references. This includes dropping the reference the page-table
2490 * held to the old page, as well as updating the rmap.
2491 * - In any case, unlock the PTL and drop the reference we took to the old page.
2493 static int wp_page_copy(struct vm_fault
*vmf
)
2495 struct vm_area_struct
*vma
= vmf
->vma
;
2496 struct mm_struct
*mm
= vma
->vm_mm
;
2497 struct page
*old_page
= vmf
->page
;
2498 struct page
*new_page
= NULL
;
2500 int page_copied
= 0;
2501 const unsigned long mmun_start
= vmf
->address
& PAGE_MASK
;
2502 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
;
2503 struct mem_cgroup
*memcg
;
2505 if (unlikely(anon_vma_prepare(vma
)))
2508 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2509 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2514 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2518 cow_user_page(new_page
, old_page
, vmf
->address
, vma
);
2521 if (mem_cgroup_try_charge_delay(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2524 __SetPageUptodate(new_page
);
2526 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2529 * Re-check the pte - we dropped the lock
2531 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2532 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2534 if (!PageAnon(old_page
)) {
2535 dec_mm_counter_fast(mm
,
2536 mm_counter_file(old_page
));
2537 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2540 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2542 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2543 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2544 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2546 * Clear the pte entry and flush it first, before updating the
2547 * pte with the new entry. This will avoid a race condition
2548 * seen in the presence of one thread doing SMC and another
2551 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2552 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2553 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2554 lru_cache_add_active_or_unevictable(new_page
, vma
);
2556 * We call the notify macro here because, when using secondary
2557 * mmu page tables (such as kvm shadow page tables), we want the
2558 * new page to be mapped directly into the secondary page table.
2560 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2561 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2564 * Only after switching the pte to the new page may
2565 * we remove the mapcount here. Otherwise another
2566 * process may come and find the rmap count decremented
2567 * before the pte is switched to the new page, and
2568 * "reuse" the old page writing into it while our pte
2569 * here still points into it and can be read by other
2572 * The critical issue is to order this
2573 * page_remove_rmap with the ptp_clear_flush above.
2574 * Those stores are ordered by (if nothing else,)
2575 * the barrier present in the atomic_add_negative
2576 * in page_remove_rmap.
2578 * Then the TLB flush in ptep_clear_flush ensures that
2579 * no process can access the old page before the
2580 * decremented mapcount is visible. And the old page
2581 * cannot be reused until after the decremented
2582 * mapcount is visible. So transitively, TLBs to
2583 * old page will be flushed before it can be reused.
2585 page_remove_rmap(old_page
, false);
2588 /* Free the old page.. */
2589 new_page
= old_page
;
2592 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2598 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2600 * No need to double call mmu_notifier->invalidate_range() callback as
2601 * the above ptep_clear_flush_notify() did already call it.
2603 mmu_notifier_invalidate_range_only_end(mm
, mmun_start
, mmun_end
);
2606 * Don't let another task, with possibly unlocked vma,
2607 * keep the mlocked page.
2609 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2610 lock_page(old_page
); /* LRU manipulation */
2611 if (PageMlocked(old_page
))
2612 munlock_vma_page(old_page
);
2613 unlock_page(old_page
);
2617 return page_copied
? VM_FAULT_WRITE
: 0;
2623 return VM_FAULT_OOM
;
2627 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2628 * writeable once the page is prepared
2630 * @vmf: structure describing the fault
2632 * This function handles all that is needed to finish a write page fault in a
2633 * shared mapping due to PTE being read-only once the mapped page is prepared.
2634 * It handles locking of PTE and modifying it. The function returns
2635 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2638 * The function expects the page to be locked or other protection against
2639 * concurrent faults / writeback (such as DAX radix tree locks).
2641 int finish_mkwrite_fault(struct vm_fault
*vmf
)
2643 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2644 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2647 * We might have raced with another page fault while we released the
2648 * pte_offset_map_lock.
2650 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2651 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2652 return VM_FAULT_NOPAGE
;
2659 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2662 static int wp_pfn_shared(struct vm_fault
*vmf
)
2664 struct vm_area_struct
*vma
= vmf
->vma
;
2666 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2669 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2670 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2671 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2672 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2674 return finish_mkwrite_fault(vmf
);
2677 return VM_FAULT_WRITE
;
2680 static int wp_page_shared(struct vm_fault
*vmf
)
2681 __releases(vmf
->ptl
)
2683 struct vm_area_struct
*vma
= vmf
->vma
;
2685 get_page(vmf
->page
);
2687 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2690 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2691 tmp
= do_page_mkwrite(vmf
);
2692 if (unlikely(!tmp
|| (tmp
&
2693 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2694 put_page(vmf
->page
);
2697 tmp
= finish_mkwrite_fault(vmf
);
2698 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2699 unlock_page(vmf
->page
);
2700 put_page(vmf
->page
);
2705 lock_page(vmf
->page
);
2707 fault_dirty_shared_page(vma
, vmf
->page
);
2708 put_page(vmf
->page
);
2710 return VM_FAULT_WRITE
;
2714 * This routine handles present pages, when users try to write
2715 * to a shared page. It is done by copying the page to a new address
2716 * and decrementing the shared-page counter for the old page.
2718 * Note that this routine assumes that the protection checks have been
2719 * done by the caller (the low-level page fault routine in most cases).
2720 * Thus we can safely just mark it writable once we've done any necessary
2723 * We also mark the page dirty at this point even though the page will
2724 * change only once the write actually happens. This avoids a few races,
2725 * and potentially makes it more efficient.
2727 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2728 * but allow concurrent faults), with pte both mapped and locked.
2729 * We return with mmap_sem still held, but pte unmapped and unlocked.
2731 static int do_wp_page(struct vm_fault
*vmf
)
2732 __releases(vmf
->ptl
)
2734 struct vm_area_struct
*vma
= vmf
->vma
;
2736 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2739 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2742 * We should not cow pages in a shared writeable mapping.
2743 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2745 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2746 (VM_WRITE
|VM_SHARED
))
2747 return wp_pfn_shared(vmf
);
2749 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2750 return wp_page_copy(vmf
);
2754 * Take out anonymous pages first, anonymous shared vmas are
2755 * not dirty accountable.
2757 if (PageAnon(vmf
->page
) && !PageKsm(vmf
->page
)) {
2758 int total_map_swapcount
;
2759 if (!trylock_page(vmf
->page
)) {
2760 get_page(vmf
->page
);
2761 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2762 lock_page(vmf
->page
);
2763 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2764 vmf
->address
, &vmf
->ptl
);
2765 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2766 unlock_page(vmf
->page
);
2767 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2768 put_page(vmf
->page
);
2771 put_page(vmf
->page
);
2773 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2774 if (total_map_swapcount
== 1) {
2776 * The page is all ours. Move it to
2777 * our anon_vma so the rmap code will
2778 * not search our parent or siblings.
2779 * Protected against the rmap code by
2782 page_move_anon_rmap(vmf
->page
, vma
);
2784 unlock_page(vmf
->page
);
2786 return VM_FAULT_WRITE
;
2788 unlock_page(vmf
->page
);
2789 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2790 (VM_WRITE
|VM_SHARED
))) {
2791 return wp_page_shared(vmf
);
2795 * Ok, we need to copy. Oh, well..
2797 get_page(vmf
->page
);
2799 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2800 return wp_page_copy(vmf
);
2803 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2804 unsigned long start_addr
, unsigned long end_addr
,
2805 struct zap_details
*details
)
2807 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2810 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2811 struct zap_details
*details
)
2813 struct vm_area_struct
*vma
;
2814 pgoff_t vba
, vea
, zba
, zea
;
2816 vma_interval_tree_foreach(vma
, root
,
2817 details
->first_index
, details
->last_index
) {
2819 vba
= vma
->vm_pgoff
;
2820 vea
= vba
+ vma_pages(vma
) - 1;
2821 zba
= details
->first_index
;
2824 zea
= details
->last_index
;
2828 unmap_mapping_range_vma(vma
,
2829 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2830 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2836 * unmap_mapping_pages() - Unmap pages from processes.
2837 * @mapping: The address space containing pages to be unmapped.
2838 * @start: Index of first page to be unmapped.
2839 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2840 * @even_cows: Whether to unmap even private COWed pages.
2842 * Unmap the pages in this address space from any userspace process which
2843 * has them mmaped. Generally, you want to remove COWed pages as well when
2844 * a file is being truncated, but not when invalidating pages from the page
2847 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
2848 pgoff_t nr
, bool even_cows
)
2850 struct zap_details details
= { };
2852 details
.check_mapping
= even_cows
? NULL
: mapping
;
2853 details
.first_index
= start
;
2854 details
.last_index
= start
+ nr
- 1;
2855 if (details
.last_index
< details
.first_index
)
2856 details
.last_index
= ULONG_MAX
;
2858 i_mmap_lock_write(mapping
);
2859 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2860 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2861 i_mmap_unlock_write(mapping
);
2865 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2866 * address_space corresponding to the specified byte range in the underlying
2869 * @mapping: the address space containing mmaps to be unmapped.
2870 * @holebegin: byte in first page to unmap, relative to the start of
2871 * the underlying file. This will be rounded down to a PAGE_SIZE
2872 * boundary. Note that this is different from truncate_pagecache(), which
2873 * must keep the partial page. In contrast, we must get rid of
2875 * @holelen: size of prospective hole in bytes. This will be rounded
2876 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2878 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2879 * but 0 when invalidating pagecache, don't throw away private data.
2881 void unmap_mapping_range(struct address_space
*mapping
,
2882 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2884 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2885 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2887 /* Check for overflow. */
2888 if (sizeof(holelen
) > sizeof(hlen
)) {
2890 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2891 if (holeend
& ~(long long)ULONG_MAX
)
2892 hlen
= ULONG_MAX
- hba
+ 1;
2895 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
2897 EXPORT_SYMBOL(unmap_mapping_range
);
2900 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2901 * but allow concurrent faults), and pte mapped but not yet locked.
2902 * We return with pte unmapped and unlocked.
2904 * We return with the mmap_sem locked or unlocked in the same cases
2905 * as does filemap_fault().
2907 int do_swap_page(struct vm_fault
*vmf
)
2909 struct vm_area_struct
*vma
= vmf
->vma
;
2910 struct page
*page
= NULL
, *swapcache
;
2911 struct mem_cgroup
*memcg
;
2918 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
2921 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2922 if (unlikely(non_swap_entry(entry
))) {
2923 if (is_migration_entry(entry
)) {
2924 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2926 } else if (is_device_private_entry(entry
)) {
2928 * For un-addressable device memory we call the pgmap
2929 * fault handler callback. The callback must migrate
2930 * the page back to some CPU accessible page.
2932 ret
= device_private_entry_fault(vma
, vmf
->address
, entry
,
2933 vmf
->flags
, vmf
->pmd
);
2934 } else if (is_hwpoison_entry(entry
)) {
2935 ret
= VM_FAULT_HWPOISON
;
2937 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2938 ret
= VM_FAULT_SIGBUS
;
2944 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2945 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
2949 struct swap_info_struct
*si
= swp_swap_info(entry
);
2951 if (si
->flags
& SWP_SYNCHRONOUS_IO
&&
2952 __swap_count(si
, entry
) == 1) {
2953 /* skip swapcache */
2954 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2957 __SetPageLocked(page
);
2958 __SetPageSwapBacked(page
);
2959 set_page_private(page
, entry
.val
);
2960 lru_cache_add_anon(page
);
2961 swap_readpage(page
, true);
2964 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
2971 * Back out if somebody else faulted in this pte
2972 * while we released the pte lock.
2974 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2975 vmf
->address
, &vmf
->ptl
);
2976 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2978 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2982 /* Had to read the page from swap area: Major fault */
2983 ret
= VM_FAULT_MAJOR
;
2984 count_vm_event(PGMAJFAULT
);
2985 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2986 } else if (PageHWPoison(page
)) {
2988 * hwpoisoned dirty swapcache pages are kept for killing
2989 * owner processes (which may be unknown at hwpoison time)
2991 ret
= VM_FAULT_HWPOISON
;
2992 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2996 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2998 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3000 ret
|= VM_FAULT_RETRY
;
3005 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3006 * release the swapcache from under us. The page pin, and pte_same
3007 * test below, are not enough to exclude that. Even if it is still
3008 * swapcache, we need to check that the page's swap has not changed.
3010 if (unlikely((!PageSwapCache(page
) ||
3011 page_private(page
) != entry
.val
)) && swapcache
)
3014 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3015 if (unlikely(!page
)) {
3021 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
,
3028 * Back out if somebody else already faulted in this pte.
3030 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3032 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3035 if (unlikely(!PageUptodate(page
))) {
3036 ret
= VM_FAULT_SIGBUS
;
3041 * The page isn't present yet, go ahead with the fault.
3043 * Be careful about the sequence of operations here.
3044 * To get its accounting right, reuse_swap_page() must be called
3045 * while the page is counted on swap but not yet in mapcount i.e.
3046 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3047 * must be called after the swap_free(), or it will never succeed.
3050 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3051 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3052 pte
= mk_pte(page
, vma
->vm_page_prot
);
3053 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3054 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3055 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3056 ret
|= VM_FAULT_WRITE
;
3057 exclusive
= RMAP_EXCLUSIVE
;
3059 flush_icache_page(vma
, page
);
3060 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3061 pte
= pte_mksoft_dirty(pte
);
3062 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3063 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3064 vmf
->orig_pte
= pte
;
3066 /* ksm created a completely new copy */
3067 if (unlikely(page
!= swapcache
&& swapcache
)) {
3068 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3069 mem_cgroup_commit_charge(page
, memcg
, false, false);
3070 lru_cache_add_active_or_unevictable(page
, vma
);
3072 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3073 mem_cgroup_commit_charge(page
, memcg
, true, false);
3074 activate_page(page
);
3078 if (mem_cgroup_swap_full(page
) ||
3079 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3080 try_to_free_swap(page
);
3082 if (page
!= swapcache
&& swapcache
) {
3084 * Hold the lock to avoid the swap entry to be reused
3085 * until we take the PT lock for the pte_same() check
3086 * (to avoid false positives from pte_same). For
3087 * further safety release the lock after the swap_free
3088 * so that the swap count won't change under a
3089 * parallel locked swapcache.
3091 unlock_page(swapcache
);
3092 put_page(swapcache
);
3095 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3096 ret
|= do_wp_page(vmf
);
3097 if (ret
& VM_FAULT_ERROR
)
3098 ret
&= VM_FAULT_ERROR
;
3102 /* No need to invalidate - it was non-present before */
3103 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3105 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3109 mem_cgroup_cancel_charge(page
, memcg
, false);
3110 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3115 if (page
!= swapcache
&& swapcache
) {
3116 unlock_page(swapcache
);
3117 put_page(swapcache
);
3123 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3124 * but allow concurrent faults), and pte mapped but not yet locked.
3125 * We return with mmap_sem still held, but pte unmapped and unlocked.
3127 static int do_anonymous_page(struct vm_fault
*vmf
)
3129 struct vm_area_struct
*vma
= vmf
->vma
;
3130 struct mem_cgroup
*memcg
;
3135 /* File mapping without ->vm_ops ? */
3136 if (vma
->vm_flags
& VM_SHARED
)
3137 return VM_FAULT_SIGBUS
;
3140 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3141 * pte_offset_map() on pmds where a huge pmd might be created
3142 * from a different thread.
3144 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3145 * parallel threads are excluded by other means.
3147 * Here we only have down_read(mmap_sem).
3149 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))
3150 return VM_FAULT_OOM
;
3152 /* See the comment in pte_alloc_one_map() */
3153 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3156 /* Use the zero-page for reads */
3157 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3158 !mm_forbids_zeropage(vma
->vm_mm
)) {
3159 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3160 vma
->vm_page_prot
));
3161 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3162 vmf
->address
, &vmf
->ptl
);
3163 if (!pte_none(*vmf
->pte
))
3165 ret
= check_stable_address_space(vma
->vm_mm
);
3168 /* Deliver the page fault to userland, check inside PT lock */
3169 if (userfaultfd_missing(vma
)) {
3170 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3171 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3176 /* Allocate our own private page. */
3177 if (unlikely(anon_vma_prepare(vma
)))
3179 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3183 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
,
3188 * The memory barrier inside __SetPageUptodate makes sure that
3189 * preceeding stores to the page contents become visible before
3190 * the set_pte_at() write.
3192 __SetPageUptodate(page
);
3194 entry
= mk_pte(page
, vma
->vm_page_prot
);
3195 if (vma
->vm_flags
& VM_WRITE
)
3196 entry
= pte_mkwrite(pte_mkdirty(entry
));
3198 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3200 if (!pte_none(*vmf
->pte
))
3203 ret
= check_stable_address_space(vma
->vm_mm
);
3207 /* Deliver the page fault to userland, check inside PT lock */
3208 if (userfaultfd_missing(vma
)) {
3209 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3210 mem_cgroup_cancel_charge(page
, memcg
, false);
3212 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3215 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3216 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3217 mem_cgroup_commit_charge(page
, memcg
, false, false);
3218 lru_cache_add_active_or_unevictable(page
, vma
);
3220 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3222 /* No need to invalidate - it was non-present before */
3223 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3225 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3228 mem_cgroup_cancel_charge(page
, memcg
, false);
3234 return VM_FAULT_OOM
;
3238 * The mmap_sem must have been held on entry, and may have been
3239 * released depending on flags and vma->vm_ops->fault() return value.
3240 * See filemap_fault() and __lock_page_retry().
3242 static int __do_fault(struct vm_fault
*vmf
)
3244 struct vm_area_struct
*vma
= vmf
->vma
;
3247 ret
= vma
->vm_ops
->fault(vmf
);
3248 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3249 VM_FAULT_DONE_COW
)))
3252 if (unlikely(PageHWPoison(vmf
->page
))) {
3253 if (ret
& VM_FAULT_LOCKED
)
3254 unlock_page(vmf
->page
);
3255 put_page(vmf
->page
);
3257 return VM_FAULT_HWPOISON
;
3260 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3261 lock_page(vmf
->page
);
3263 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3269 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3270 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3271 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3272 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3274 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3276 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3279 static int pte_alloc_one_map(struct vm_fault
*vmf
)
3281 struct vm_area_struct
*vma
= vmf
->vma
;
3283 if (!pmd_none(*vmf
->pmd
))
3285 if (vmf
->prealloc_pte
) {
3286 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3287 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3288 spin_unlock(vmf
->ptl
);
3292 mm_inc_nr_ptes(vma
->vm_mm
);
3293 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3294 spin_unlock(vmf
->ptl
);
3295 vmf
->prealloc_pte
= NULL
;
3296 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
, vmf
->address
))) {
3297 return VM_FAULT_OOM
;
3301 * If a huge pmd materialized under us just retry later. Use
3302 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3303 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3304 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3305 * running immediately after a huge pmd fault in a different thread of
3306 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3307 * All we have to ensure is that it is a regular pmd that we can walk
3308 * with pte_offset_map() and we can do that through an atomic read in
3309 * C, which is what pmd_trans_unstable() provides.
3311 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3312 return VM_FAULT_NOPAGE
;
3315 * At this point we know that our vmf->pmd points to a page of ptes
3316 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3317 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3318 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3319 * be valid and we will re-check to make sure the vmf->pte isn't
3320 * pte_none() under vmf->ptl protection when we return to
3323 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3328 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3330 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3331 static inline bool transhuge_vma_suitable(struct vm_area_struct
*vma
,
3332 unsigned long haddr
)
3334 if (((vma
->vm_start
>> PAGE_SHIFT
) & HPAGE_CACHE_INDEX_MASK
) !=
3335 (vma
->vm_pgoff
& HPAGE_CACHE_INDEX_MASK
))
3337 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
3342 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3344 struct vm_area_struct
*vma
= vmf
->vma
;
3346 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3348 * We are going to consume the prealloc table,
3349 * count that as nr_ptes.
3351 mm_inc_nr_ptes(vma
->vm_mm
);
3352 vmf
->prealloc_pte
= NULL
;
3355 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3357 struct vm_area_struct
*vma
= vmf
->vma
;
3358 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3359 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3363 if (!transhuge_vma_suitable(vma
, haddr
))
3364 return VM_FAULT_FALLBACK
;
3366 ret
= VM_FAULT_FALLBACK
;
3367 page
= compound_head(page
);
3370 * Archs like ppc64 need additonal space to store information
3371 * related to pte entry. Use the preallocated table for that.
3373 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3374 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
, vmf
->address
);
3375 if (!vmf
->prealloc_pte
)
3376 return VM_FAULT_OOM
;
3377 smp_wmb(); /* See comment in __pte_alloc() */
3380 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3381 if (unlikely(!pmd_none(*vmf
->pmd
)))
3384 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3385 flush_icache_page(vma
, page
+ i
);
3387 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3389 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3391 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3392 page_add_file_rmap(page
, true);
3394 * deposit and withdraw with pmd lock held
3396 if (arch_needs_pgtable_deposit())
3397 deposit_prealloc_pte(vmf
);
3399 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3401 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3403 /* fault is handled */
3405 count_vm_event(THP_FILE_MAPPED
);
3407 spin_unlock(vmf
->ptl
);
3411 static int do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3419 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3420 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3422 * @vmf: fault environment
3423 * @memcg: memcg to charge page (only for private mappings)
3424 * @page: page to map
3426 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3429 * Target users are page handler itself and implementations of
3430 * vm_ops->map_pages.
3432 int alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3435 struct vm_area_struct
*vma
= vmf
->vma
;
3436 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3440 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3441 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3443 VM_BUG_ON_PAGE(memcg
, page
);
3445 ret
= do_set_pmd(vmf
, page
);
3446 if (ret
!= VM_FAULT_FALLBACK
)
3451 ret
= pte_alloc_one_map(vmf
);
3456 /* Re-check under ptl */
3457 if (unlikely(!pte_none(*vmf
->pte
)))
3458 return VM_FAULT_NOPAGE
;
3460 flush_icache_page(vma
, page
);
3461 entry
= mk_pte(page
, vma
->vm_page_prot
);
3463 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3464 /* copy-on-write page */
3465 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3466 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3467 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3468 mem_cgroup_commit_charge(page
, memcg
, false, false);
3469 lru_cache_add_active_or_unevictable(page
, vma
);
3471 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3472 page_add_file_rmap(page
, false);
3474 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3476 /* no need to invalidate: a not-present page won't be cached */
3477 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3484 * finish_fault - finish page fault once we have prepared the page to fault
3486 * @vmf: structure describing the fault
3488 * This function handles all that is needed to finish a page fault once the
3489 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3490 * given page, adds reverse page mapping, handles memcg charges and LRU
3491 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3494 * The function expects the page to be locked and on success it consumes a
3495 * reference of a page being mapped (for the PTE which maps it).
3497 int finish_fault(struct vm_fault
*vmf
)
3502 /* Did we COW the page? */
3503 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3504 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3505 page
= vmf
->cow_page
;
3510 * check even for read faults because we might have lost our CoWed
3513 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3514 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3516 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3518 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3522 static unsigned long fault_around_bytes __read_mostly
=
3523 rounddown_pow_of_two(65536);
3525 #ifdef CONFIG_DEBUG_FS
3526 static int fault_around_bytes_get(void *data
, u64
*val
)
3528 *val
= fault_around_bytes
;
3533 * fault_around_bytes must be rounded down to the nearest page order as it's
3534 * what do_fault_around() expects to see.
3536 static int fault_around_bytes_set(void *data
, u64 val
)
3538 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3540 if (val
> PAGE_SIZE
)
3541 fault_around_bytes
= rounddown_pow_of_two(val
);
3543 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3546 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3547 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3549 static int __init
fault_around_debugfs(void)
3553 ret
= debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3554 &fault_around_bytes_fops
);
3556 pr_warn("Failed to create fault_around_bytes in debugfs");
3559 late_initcall(fault_around_debugfs
);
3563 * do_fault_around() tries to map few pages around the fault address. The hope
3564 * is that the pages will be needed soon and this will lower the number of
3567 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3568 * not ready to be mapped: not up-to-date, locked, etc.
3570 * This function is called with the page table lock taken. In the split ptlock
3571 * case the page table lock only protects only those entries which belong to
3572 * the page table corresponding to the fault address.
3574 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3577 * fault_around_bytes defines how many bytes we'll try to map.
3578 * do_fault_around() expects it to be set to a power of two less than or equal
3581 * The virtual address of the area that we map is naturally aligned to
3582 * fault_around_bytes rounded down to the machine page size
3583 * (and therefore to page order). This way it's easier to guarantee
3584 * that we don't cross page table boundaries.
3586 static int do_fault_around(struct vm_fault
*vmf
)
3588 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3589 pgoff_t start_pgoff
= vmf
->pgoff
;
3593 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3594 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3596 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3597 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3601 * end_pgoff is either the end of the page table, the end of
3602 * the vma or nr_pages from start_pgoff, depending what is nearest.
3604 end_pgoff
= start_pgoff
-
3605 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3607 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3608 start_pgoff
+ nr_pages
- 1);
3610 if (pmd_none(*vmf
->pmd
)) {
3611 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
,
3613 if (!vmf
->prealloc_pte
)
3615 smp_wmb(); /* See comment in __pte_alloc() */
3618 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3620 /* Huge page is mapped? Page fault is solved */
3621 if (pmd_trans_huge(*vmf
->pmd
)) {
3622 ret
= VM_FAULT_NOPAGE
;
3626 /* ->map_pages() haven't done anything useful. Cold page cache? */
3630 /* check if the page fault is solved */
3631 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3632 if (!pte_none(*vmf
->pte
))
3633 ret
= VM_FAULT_NOPAGE
;
3634 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3636 vmf
->address
= address
;
3641 static int do_read_fault(struct vm_fault
*vmf
)
3643 struct vm_area_struct
*vma
= vmf
->vma
;
3647 * Let's call ->map_pages() first and use ->fault() as fallback
3648 * if page by the offset is not ready to be mapped (cold cache or
3651 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3652 ret
= do_fault_around(vmf
);
3657 ret
= __do_fault(vmf
);
3658 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3661 ret
|= finish_fault(vmf
);
3662 unlock_page(vmf
->page
);
3663 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3664 put_page(vmf
->page
);
3668 static int do_cow_fault(struct vm_fault
*vmf
)
3670 struct vm_area_struct
*vma
= vmf
->vma
;
3673 if (unlikely(anon_vma_prepare(vma
)))
3674 return VM_FAULT_OOM
;
3676 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3678 return VM_FAULT_OOM
;
3680 if (mem_cgroup_try_charge_delay(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3681 &vmf
->memcg
, false)) {
3682 put_page(vmf
->cow_page
);
3683 return VM_FAULT_OOM
;
3686 ret
= __do_fault(vmf
);
3687 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3689 if (ret
& VM_FAULT_DONE_COW
)
3692 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3693 __SetPageUptodate(vmf
->cow_page
);
3695 ret
|= finish_fault(vmf
);
3696 unlock_page(vmf
->page
);
3697 put_page(vmf
->page
);
3698 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3702 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3703 put_page(vmf
->cow_page
);
3707 static int do_shared_fault(struct vm_fault
*vmf
)
3709 struct vm_area_struct
*vma
= vmf
->vma
;
3712 ret
= __do_fault(vmf
);
3713 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3717 * Check if the backing address space wants to know that the page is
3718 * about to become writable
3720 if (vma
->vm_ops
->page_mkwrite
) {
3721 unlock_page(vmf
->page
);
3722 tmp
= do_page_mkwrite(vmf
);
3723 if (unlikely(!tmp
||
3724 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3725 put_page(vmf
->page
);
3730 ret
|= finish_fault(vmf
);
3731 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3733 unlock_page(vmf
->page
);
3734 put_page(vmf
->page
);
3738 fault_dirty_shared_page(vma
, vmf
->page
);
3743 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3744 * but allow concurrent faults).
3745 * The mmap_sem may have been released depending on flags and our
3746 * return value. See filemap_fault() and __lock_page_or_retry().
3748 static int do_fault(struct vm_fault
*vmf
)
3750 struct vm_area_struct
*vma
= vmf
->vma
;
3753 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3754 if (!vma
->vm_ops
->fault
)
3755 ret
= VM_FAULT_SIGBUS
;
3756 else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3757 ret
= do_read_fault(vmf
);
3758 else if (!(vma
->vm_flags
& VM_SHARED
))
3759 ret
= do_cow_fault(vmf
);
3761 ret
= do_shared_fault(vmf
);
3763 /* preallocated pagetable is unused: free it */
3764 if (vmf
->prealloc_pte
) {
3765 pte_free(vma
->vm_mm
, vmf
->prealloc_pte
);
3766 vmf
->prealloc_pte
= NULL
;
3771 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3772 unsigned long addr
, int page_nid
,
3777 count_vm_numa_event(NUMA_HINT_FAULTS
);
3778 if (page_nid
== numa_node_id()) {
3779 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3780 *flags
|= TNF_FAULT_LOCAL
;
3783 return mpol_misplaced(page
, vma
, addr
);
3786 static int do_numa_page(struct vm_fault
*vmf
)
3788 struct vm_area_struct
*vma
= vmf
->vma
;
3789 struct page
*page
= NULL
;
3793 bool migrated
= false;
3795 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3799 * The "pte" at this point cannot be used safely without
3800 * validation through pte_unmap_same(). It's of NUMA type but
3801 * the pfn may be screwed if the read is non atomic.
3803 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3804 spin_lock(vmf
->ptl
);
3805 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3806 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3811 * Make it present again, Depending on how arch implementes non
3812 * accessible ptes, some can allow access by kernel mode.
3814 pte
= ptep_modify_prot_start(vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3815 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3816 pte
= pte_mkyoung(pte
);
3818 pte
= pte_mkwrite(pte
);
3819 ptep_modify_prot_commit(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3820 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3822 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3824 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3828 /* TODO: handle PTE-mapped THP */
3829 if (PageCompound(page
)) {
3830 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3835 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3836 * much anyway since they can be in shared cache state. This misses
3837 * the case where a mapping is writable but the process never writes
3838 * to it but pte_write gets cleared during protection updates and
3839 * pte_dirty has unpredictable behaviour between PTE scan updates,
3840 * background writeback, dirty balancing and application behaviour.
3842 if (!pte_write(pte
))
3843 flags
|= TNF_NO_GROUP
;
3846 * Flag if the page is shared between multiple address spaces. This
3847 * is later used when determining whether to group tasks together
3849 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3850 flags
|= TNF_SHARED
;
3852 last_cpupid
= page_cpupid_last(page
);
3853 page_nid
= page_to_nid(page
);
3854 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3856 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3857 if (target_nid
== -1) {
3862 /* Migrate to the requested node */
3863 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3865 page_nid
= target_nid
;
3866 flags
|= TNF_MIGRATED
;
3868 flags
|= TNF_MIGRATE_FAIL
;
3872 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3876 static inline int create_huge_pmd(struct vm_fault
*vmf
)
3878 if (vma_is_anonymous(vmf
->vma
))
3879 return do_huge_pmd_anonymous_page(vmf
);
3880 if (vmf
->vma
->vm_ops
->huge_fault
)
3881 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3882 return VM_FAULT_FALLBACK
;
3885 /* `inline' is required to avoid gcc 4.1.2 build error */
3886 static inline int wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3888 if (vma_is_anonymous(vmf
->vma
))
3889 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3890 if (vmf
->vma
->vm_ops
->huge_fault
)
3891 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3893 /* COW handled on pte level: split pmd */
3894 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3895 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3897 return VM_FAULT_FALLBACK
;
3900 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3902 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3905 static int create_huge_pud(struct vm_fault
*vmf
)
3907 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3908 /* No support for anonymous transparent PUD pages yet */
3909 if (vma_is_anonymous(vmf
->vma
))
3910 return VM_FAULT_FALLBACK
;
3911 if (vmf
->vma
->vm_ops
->huge_fault
)
3912 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3913 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3914 return VM_FAULT_FALLBACK
;
3917 static int wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3919 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3920 /* No support for anonymous transparent PUD pages yet */
3921 if (vma_is_anonymous(vmf
->vma
))
3922 return VM_FAULT_FALLBACK
;
3923 if (vmf
->vma
->vm_ops
->huge_fault
)
3924 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3925 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3926 return VM_FAULT_FALLBACK
;
3930 * These routines also need to handle stuff like marking pages dirty
3931 * and/or accessed for architectures that don't do it in hardware (most
3932 * RISC architectures). The early dirtying is also good on the i386.
3934 * There is also a hook called "update_mmu_cache()" that architectures
3935 * with external mmu caches can use to update those (ie the Sparc or
3936 * PowerPC hashed page tables that act as extended TLBs).
3938 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3939 * concurrent faults).
3941 * The mmap_sem may have been released depending on flags and our return value.
3942 * See filemap_fault() and __lock_page_or_retry().
3944 static int handle_pte_fault(struct vm_fault
*vmf
)
3948 if (unlikely(pmd_none(*vmf
->pmd
))) {
3950 * Leave __pte_alloc() until later: because vm_ops->fault may
3951 * want to allocate huge page, and if we expose page table
3952 * for an instant, it will be difficult to retract from
3953 * concurrent faults and from rmap lookups.
3957 /* See comment in pte_alloc_one_map() */
3958 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3961 * A regular pmd is established and it can't morph into a huge
3962 * pmd from under us anymore at this point because we hold the
3963 * mmap_sem read mode and khugepaged takes it in write mode.
3964 * So now it's safe to run pte_offset_map().
3966 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
3967 vmf
->orig_pte
= *vmf
->pte
;
3970 * some architectures can have larger ptes than wordsize,
3971 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3972 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3973 * accesses. The code below just needs a consistent view
3974 * for the ifs and we later double check anyway with the
3975 * ptl lock held. So here a barrier will do.
3978 if (pte_none(vmf
->orig_pte
)) {
3979 pte_unmap(vmf
->pte
);
3985 if (vma_is_anonymous(vmf
->vma
))
3986 return do_anonymous_page(vmf
);
3988 return do_fault(vmf
);
3991 if (!pte_present(vmf
->orig_pte
))
3992 return do_swap_page(vmf
);
3994 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
3995 return do_numa_page(vmf
);
3997 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
3998 spin_lock(vmf
->ptl
);
3999 entry
= vmf
->orig_pte
;
4000 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
4002 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4003 if (!pte_write(entry
))
4004 return do_wp_page(vmf
);
4005 entry
= pte_mkdirty(entry
);
4007 entry
= pte_mkyoung(entry
);
4008 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4009 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4010 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4013 * This is needed only for protection faults but the arch code
4014 * is not yet telling us if this is a protection fault or not.
4015 * This still avoids useless tlb flushes for .text page faults
4018 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4019 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4022 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4027 * By the time we get here, we already hold the mm semaphore
4029 * The mmap_sem may have been released depending on flags and our
4030 * return value. See filemap_fault() and __lock_page_or_retry().
4032 static int __handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4035 struct vm_fault vmf
= {
4037 .address
= address
& PAGE_MASK
,
4039 .pgoff
= linear_page_index(vma
, address
),
4040 .gfp_mask
= __get_fault_gfp_mask(vma
),
4042 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4043 struct mm_struct
*mm
= vma
->vm_mm
;
4048 pgd
= pgd_offset(mm
, address
);
4049 p4d
= p4d_alloc(mm
, pgd
, address
);
4051 return VM_FAULT_OOM
;
4053 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4055 return VM_FAULT_OOM
;
4056 if (pud_none(*vmf
.pud
) && transparent_hugepage_enabled(vma
)) {
4057 ret
= create_huge_pud(&vmf
);
4058 if (!(ret
& VM_FAULT_FALLBACK
))
4061 pud_t orig_pud
= *vmf
.pud
;
4064 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4066 /* NUMA case for anonymous PUDs would go here */
4068 if (dirty
&& !pud_write(orig_pud
)) {
4069 ret
= wp_huge_pud(&vmf
, orig_pud
);
4070 if (!(ret
& VM_FAULT_FALLBACK
))
4073 huge_pud_set_accessed(&vmf
, orig_pud
);
4079 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4081 return VM_FAULT_OOM
;
4082 if (pmd_none(*vmf
.pmd
) && transparent_hugepage_enabled(vma
)) {
4083 ret
= create_huge_pmd(&vmf
);
4084 if (!(ret
& VM_FAULT_FALLBACK
))
4087 pmd_t orig_pmd
= *vmf
.pmd
;
4090 if (unlikely(is_swap_pmd(orig_pmd
))) {
4091 VM_BUG_ON(thp_migration_supported() &&
4092 !is_pmd_migration_entry(orig_pmd
));
4093 if (is_pmd_migration_entry(orig_pmd
))
4094 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4097 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4098 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4099 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4101 if (dirty
&& !pmd_write(orig_pmd
)) {
4102 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4103 if (!(ret
& VM_FAULT_FALLBACK
))
4106 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4112 return handle_pte_fault(&vmf
);
4116 * By the time we get here, we already hold the mm semaphore
4118 * The mmap_sem may have been released depending on flags and our
4119 * return value. See filemap_fault() and __lock_page_or_retry().
4121 int handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4126 __set_current_state(TASK_RUNNING
);
4128 count_vm_event(PGFAULT
);
4129 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4131 /* do counter updates before entering really critical section. */
4132 check_sync_rss_stat(current
);
4134 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4135 flags
& FAULT_FLAG_INSTRUCTION
,
4136 flags
& FAULT_FLAG_REMOTE
))
4137 return VM_FAULT_SIGSEGV
;
4140 * Enable the memcg OOM handling for faults triggered in user
4141 * space. Kernel faults are handled more gracefully.
4143 if (flags
& FAULT_FLAG_USER
)
4144 mem_cgroup_enter_user_fault();
4146 if (unlikely(is_vm_hugetlb_page(vma
)))
4147 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4149 ret
= __handle_mm_fault(vma
, address
, flags
);
4151 if (flags
& FAULT_FLAG_USER
) {
4152 mem_cgroup_exit_user_fault();
4154 * The task may have entered a memcg OOM situation but
4155 * if the allocation error was handled gracefully (no
4156 * VM_FAULT_OOM), there is no need to kill anything.
4157 * Just clean up the OOM state peacefully.
4159 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4160 mem_cgroup_oom_synchronize(false);
4165 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4167 #ifndef __PAGETABLE_P4D_FOLDED
4169 * Allocate p4d page table.
4170 * We've already handled the fast-path in-line.
4172 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4174 p4d_t
*new = p4d_alloc_one(mm
, address
);
4178 smp_wmb(); /* See comment in __pte_alloc */
4180 spin_lock(&mm
->page_table_lock
);
4181 if (pgd_present(*pgd
)) /* Another has populated it */
4184 pgd_populate(mm
, pgd
, new);
4185 spin_unlock(&mm
->page_table_lock
);
4188 #endif /* __PAGETABLE_P4D_FOLDED */
4190 #ifndef __PAGETABLE_PUD_FOLDED
4192 * Allocate page upper directory.
4193 * We've already handled the fast-path in-line.
4195 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4197 pud_t
*new = pud_alloc_one(mm
, address
);
4201 smp_wmb(); /* See comment in __pte_alloc */
4203 spin_lock(&mm
->page_table_lock
);
4204 #ifndef __ARCH_HAS_5LEVEL_HACK
4205 if (!p4d_present(*p4d
)) {
4207 p4d_populate(mm
, p4d
, new);
4208 } else /* Another has populated it */
4211 if (!pgd_present(*p4d
)) {
4213 pgd_populate(mm
, p4d
, new);
4214 } else /* Another has populated it */
4216 #endif /* __ARCH_HAS_5LEVEL_HACK */
4217 spin_unlock(&mm
->page_table_lock
);
4220 #endif /* __PAGETABLE_PUD_FOLDED */
4222 #ifndef __PAGETABLE_PMD_FOLDED
4224 * Allocate page middle directory.
4225 * We've already handled the fast-path in-line.
4227 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4230 pmd_t
*new = pmd_alloc_one(mm
, address
);
4234 smp_wmb(); /* See comment in __pte_alloc */
4236 ptl
= pud_lock(mm
, pud
);
4237 #ifndef __ARCH_HAS_4LEVEL_HACK
4238 if (!pud_present(*pud
)) {
4240 pud_populate(mm
, pud
, new);
4241 } else /* Another has populated it */
4244 if (!pgd_present(*pud
)) {
4246 pgd_populate(mm
, pud
, new);
4247 } else /* Another has populated it */
4249 #endif /* __ARCH_HAS_4LEVEL_HACK */
4253 #endif /* __PAGETABLE_PMD_FOLDED */
4255 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4256 unsigned long *start
, unsigned long *end
,
4257 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4265 pgd
= pgd_offset(mm
, address
);
4266 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4269 p4d
= p4d_offset(pgd
, address
);
4270 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4273 pud
= pud_offset(p4d
, address
);
4274 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4277 pmd
= pmd_offset(pud
, address
);
4278 VM_BUG_ON(pmd_trans_huge(*pmd
));
4280 if (pmd_huge(*pmd
)) {
4285 *start
= address
& PMD_MASK
;
4286 *end
= *start
+ PMD_SIZE
;
4287 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4289 *ptlp
= pmd_lock(mm
, pmd
);
4290 if (pmd_huge(*pmd
)) {
4296 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4299 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4303 *start
= address
& PAGE_MASK
;
4304 *end
= *start
+ PAGE_SIZE
;
4305 mmu_notifier_invalidate_range_start(mm
, *start
, *end
);
4307 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4308 if (!pte_present(*ptep
))
4313 pte_unmap_unlock(ptep
, *ptlp
);
4315 mmu_notifier_invalidate_range_end(mm
, *start
, *end
);
4320 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4321 pte_t
**ptepp
, spinlock_t
**ptlp
)
4325 /* (void) is needed to make gcc happy */
4326 (void) __cond_lock(*ptlp
,
4327 !(res
= __follow_pte_pmd(mm
, address
, NULL
, NULL
,
4328 ptepp
, NULL
, ptlp
)));
4332 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4333 unsigned long *start
, unsigned long *end
,
4334 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4338 /* (void) is needed to make gcc happy */
4339 (void) __cond_lock(*ptlp
,
4340 !(res
= __follow_pte_pmd(mm
, address
, start
, end
,
4341 ptepp
, pmdpp
, ptlp
)));
4344 EXPORT_SYMBOL(follow_pte_pmd
);
4347 * follow_pfn - look up PFN at a user virtual address
4348 * @vma: memory mapping
4349 * @address: user virtual address
4350 * @pfn: location to store found PFN
4352 * Only IO mappings and raw PFN mappings are allowed.
4354 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4356 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4363 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4366 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4369 *pfn
= pte_pfn(*ptep
);
4370 pte_unmap_unlock(ptep
, ptl
);
4373 EXPORT_SYMBOL(follow_pfn
);
4375 #ifdef CONFIG_HAVE_IOREMAP_PROT
4376 int follow_phys(struct vm_area_struct
*vma
,
4377 unsigned long address
, unsigned int flags
,
4378 unsigned long *prot
, resource_size_t
*phys
)
4384 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4387 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4391 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4394 *prot
= pgprot_val(pte_pgprot(pte
));
4395 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4399 pte_unmap_unlock(ptep
, ptl
);
4404 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4405 void *buf
, int len
, int write
)
4407 resource_size_t phys_addr
;
4408 unsigned long prot
= 0;
4409 void __iomem
*maddr
;
4410 int offset
= addr
& (PAGE_SIZE
-1);
4412 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4415 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4420 memcpy_toio(maddr
+ offset
, buf
, len
);
4422 memcpy_fromio(buf
, maddr
+ offset
, len
);
4427 EXPORT_SYMBOL_GPL(generic_access_phys
);
4431 * Access another process' address space as given in mm. If non-NULL, use the
4432 * given task for page fault accounting.
4434 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4435 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4437 struct vm_area_struct
*vma
;
4438 void *old_buf
= buf
;
4439 int write
= gup_flags
& FOLL_WRITE
;
4441 down_read(&mm
->mmap_sem
);
4442 /* ignore errors, just check how much was successfully transferred */
4444 int bytes
, ret
, offset
;
4446 struct page
*page
= NULL
;
4448 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4449 gup_flags
, &page
, &vma
, NULL
);
4451 #ifndef CONFIG_HAVE_IOREMAP_PROT
4455 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4456 * we can access using slightly different code.
4458 vma
= find_vma(mm
, addr
);
4459 if (!vma
|| vma
->vm_start
> addr
)
4461 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4462 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4470 offset
= addr
& (PAGE_SIZE
-1);
4471 if (bytes
> PAGE_SIZE
-offset
)
4472 bytes
= PAGE_SIZE
-offset
;
4476 copy_to_user_page(vma
, page
, addr
,
4477 maddr
+ offset
, buf
, bytes
);
4478 set_page_dirty_lock(page
);
4480 copy_from_user_page(vma
, page
, addr
,
4481 buf
, maddr
+ offset
, bytes
);
4490 up_read(&mm
->mmap_sem
);
4492 return buf
- old_buf
;
4496 * access_remote_vm - access another process' address space
4497 * @mm: the mm_struct of the target address space
4498 * @addr: start address to access
4499 * @buf: source or destination buffer
4500 * @len: number of bytes to transfer
4501 * @gup_flags: flags modifying lookup behaviour
4503 * The caller must hold a reference on @mm.
4505 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4506 void *buf
, int len
, unsigned int gup_flags
)
4508 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4512 * Access another process' address space.
4513 * Source/target buffer must be kernel space,
4514 * Do not walk the page table directly, use get_user_pages
4516 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4517 void *buf
, int len
, unsigned int gup_flags
)
4519 struct mm_struct
*mm
;
4522 mm
= get_task_mm(tsk
);
4526 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4532 EXPORT_SYMBOL_GPL(access_process_vm
);
4535 * Print the name of a VMA.
4537 void print_vma_addr(char *prefix
, unsigned long ip
)
4539 struct mm_struct
*mm
= current
->mm
;
4540 struct vm_area_struct
*vma
;
4543 * we might be running from an atomic context so we cannot sleep
4545 if (!down_read_trylock(&mm
->mmap_sem
))
4548 vma
= find_vma(mm
, ip
);
4549 if (vma
&& vma
->vm_file
) {
4550 struct file
*f
= vma
->vm_file
;
4551 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4555 p
= file_path(f
, buf
, PAGE_SIZE
);
4558 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4560 vma
->vm_end
- vma
->vm_start
);
4561 free_page((unsigned long)buf
);
4564 up_read(&mm
->mmap_sem
);
4567 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4568 void __might_fault(const char *file
, int line
)
4571 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4572 * holding the mmap_sem, this is safe because kernel memory doesn't
4573 * get paged out, therefore we'll never actually fault, and the
4574 * below annotations will generate false positives.
4576 if (uaccess_kernel())
4578 if (pagefault_disabled())
4580 __might_sleep(file
, line
, 0);
4581 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4583 might_lock_read(¤t
->mm
->mmap_sem
);
4586 EXPORT_SYMBOL(__might_fault
);
4589 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4591 * Process all subpages of the specified huge page with the specified
4592 * operation. The target subpage will be processed last to keep its
4595 static inline void process_huge_page(
4596 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
4597 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
4601 unsigned long addr
= addr_hint
&
4602 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4604 /* Process target subpage last to keep its cache lines hot */
4606 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4607 if (2 * n
<= pages_per_huge_page
) {
4608 /* If target subpage in first half of huge page */
4611 /* Process subpages at the end of huge page */
4612 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4614 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4617 /* If target subpage in second half of huge page */
4618 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4619 l
= pages_per_huge_page
- n
;
4620 /* Process subpages at the begin of huge page */
4621 for (i
= 0; i
< base
; i
++) {
4623 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4627 * Process remaining subpages in left-right-left-right pattern
4628 * towards the target subpage
4630 for (i
= 0; i
< l
; i
++) {
4631 int left_idx
= base
+ i
;
4632 int right_idx
= base
+ 2 * l
- 1 - i
;
4635 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
4637 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
4641 static void clear_gigantic_page(struct page
*page
,
4643 unsigned int pages_per_huge_page
)
4646 struct page
*p
= page
;
4649 for (i
= 0; i
< pages_per_huge_page
;
4650 i
++, p
= mem_map_next(p
, page
, i
)) {
4652 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4656 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
4658 struct page
*page
= arg
;
4660 clear_user_highpage(page
+ idx
, addr
);
4663 void clear_huge_page(struct page
*page
,
4664 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4666 unsigned long addr
= addr_hint
&
4667 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4669 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4670 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4674 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
4677 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4679 struct vm_area_struct
*vma
,
4680 unsigned int pages_per_huge_page
)
4683 struct page
*dst_base
= dst
;
4684 struct page
*src_base
= src
;
4686 for (i
= 0; i
< pages_per_huge_page
; ) {
4688 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4691 dst
= mem_map_next(dst
, dst_base
, i
);
4692 src
= mem_map_next(src
, src_base
, i
);
4696 struct copy_subpage_arg
{
4699 struct vm_area_struct
*vma
;
4702 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
4704 struct copy_subpage_arg
*copy_arg
= arg
;
4706 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
4707 addr
, copy_arg
->vma
);
4710 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4711 unsigned long addr_hint
, struct vm_area_struct
*vma
,
4712 unsigned int pages_per_huge_page
)
4714 unsigned long addr
= addr_hint
&
4715 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4716 struct copy_subpage_arg arg
= {
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
);
4728 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
4731 long copy_huge_page_from_user(struct page
*dst_page
,
4732 const void __user
*usr_src
,
4733 unsigned int pages_per_huge_page
,
4734 bool allow_pagefault
)
4736 void *src
= (void *)usr_src
;
4738 unsigned long i
, rc
= 0;
4739 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4741 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4742 if (allow_pagefault
)
4743 page_kaddr
= kmap(dst_page
+ i
);
4745 page_kaddr
= kmap_atomic(dst_page
+ i
);
4746 rc
= copy_from_user(page_kaddr
,
4747 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4749 if (allow_pagefault
)
4750 kunmap(dst_page
+ i
);
4752 kunmap_atomic(page_kaddr
);
4754 ret_val
-= (PAGE_SIZE
- rc
);
4762 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4764 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4766 static struct kmem_cache
*page_ptl_cachep
;
4768 void __init
ptlock_cache_init(void)
4770 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4774 bool ptlock_alloc(struct page
*page
)
4778 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4785 void ptlock_free(struct page
*page
)
4787 kmem_cache_free(page_ptl_cachep
, page
->ptl
);