On Tue, Nov 06, 2007 at 02:33:53AM -0800, akpm@linux-foundation.org wrote:
[mmotm.git] / mm / memory.c
blob72a249420c11c76378ff5589a45ff46aa0d73bb1
1 /*
2 * linux/mm/memory.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
7 /*
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
18 * far as I could see.
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>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
60 #include <asm/io.h>
61 #include <asm/pgalloc.h>
62 #include <asm/uaccess.h>
63 #include <asm/tlb.h>
64 #include <asm/tlbflush.h>
65 #include <asm/pgtable.h>
67 #include "internal.h"
69 #ifndef CONFIG_NEED_MULTIPLE_NODES
70 /* use the per-pgdat data instead for discontigmem - mbligh */
71 unsigned long max_mapnr;
72 struct page *mem_map;
74 EXPORT_SYMBOL(max_mapnr);
75 EXPORT_SYMBOL(mem_map);
76 #endif
78 unsigned long num_physpages;
80 * A number of key systems in x86 including ioremap() rely on the assumption
81 * that high_memory defines the upper bound on direct map memory, then end
82 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
83 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
84 * and ZONE_HIGHMEM.
86 void * high_memory;
88 EXPORT_SYMBOL(num_physpages);
89 EXPORT_SYMBOL(high_memory);
92 * Randomize the address space (stacks, mmaps, brk, etc.).
94 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
95 * as ancient (libc5 based) binaries can segfault. )
97 int randomize_va_space __read_mostly =
98 #ifdef CONFIG_COMPAT_BRK
100 #else
102 #endif
104 static int __init disable_randmaps(char *s)
106 randomize_va_space = 0;
107 return 1;
109 __setup("norandmaps", disable_randmaps);
111 unsigned long zero_pfn __read_mostly;
112 unsigned long highest_memmap_pfn __read_mostly;
115 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 static int __init init_zero_pfn(void)
119 zero_pfn = page_to_pfn(ZERO_PAGE(0));
120 return 0;
122 core_initcall(init_zero_pfn);
125 * If a p?d_bad entry is found while walking page tables, report
126 * the error, before resetting entry to p?d_none. Usually (but
127 * very seldom) called out from the p?d_none_or_clear_bad macros.
130 void pgd_clear_bad(pgd_t *pgd)
132 pgd_ERROR(*pgd);
133 pgd_clear(pgd);
136 void pud_clear_bad(pud_t *pud)
138 pud_ERROR(*pud);
139 pud_clear(pud);
142 void pmd_clear_bad(pmd_t *pmd)
144 pmd_ERROR(*pmd);
145 pmd_clear(pmd);
149 * Note: this doesn't free the actual pages themselves. That
150 * has been handled earlier when unmapping all the memory regions.
152 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
153 unsigned long addr)
155 pgtable_t token = pmd_pgtable(*pmd);
156 pmd_clear(pmd);
157 pte_free_tlb(tlb, token, addr);
158 tlb->mm->nr_ptes--;
161 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
162 unsigned long addr, unsigned long end,
163 unsigned long floor, unsigned long ceiling)
165 pmd_t *pmd;
166 unsigned long next;
167 unsigned long start;
169 start = addr;
170 pmd = pmd_offset(pud, addr);
171 do {
172 next = pmd_addr_end(addr, end);
173 if (pmd_none_or_clear_bad(pmd))
174 continue;
175 free_pte_range(tlb, pmd, addr);
176 } while (pmd++, addr = next, addr != end);
178 start &= PUD_MASK;
179 if (start < floor)
180 return;
181 if (ceiling) {
182 ceiling &= PUD_MASK;
183 if (!ceiling)
184 return;
186 if (end - 1 > ceiling - 1)
187 return;
189 pmd = pmd_offset(pud, start);
190 pud_clear(pud);
191 pmd_free_tlb(tlb, pmd, start);
194 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
195 unsigned long addr, unsigned long end,
196 unsigned long floor, unsigned long ceiling)
198 pud_t *pud;
199 unsigned long next;
200 unsigned long start;
202 start = addr;
203 pud = pud_offset(pgd, addr);
204 do {
205 next = pud_addr_end(addr, end);
206 if (pud_none_or_clear_bad(pud))
207 continue;
208 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
209 } while (pud++, addr = next, addr != end);
211 start &= PGDIR_MASK;
212 if (start < floor)
213 return;
214 if (ceiling) {
215 ceiling &= PGDIR_MASK;
216 if (!ceiling)
217 return;
219 if (end - 1 > ceiling - 1)
220 return;
222 pud = pud_offset(pgd, start);
223 pgd_clear(pgd);
224 pud_free_tlb(tlb, pud, start);
228 * This function frees user-level page tables of a process.
230 * Must be called with pagetable lock held.
232 void free_pgd_range(struct mmu_gather *tlb,
233 unsigned long addr, unsigned long end,
234 unsigned long floor, unsigned long ceiling)
236 pgd_t *pgd;
237 unsigned long next;
238 unsigned long start;
241 * The next few lines have given us lots of grief...
243 * Why are we testing PMD* at this top level? Because often
244 * there will be no work to do at all, and we'd prefer not to
245 * go all the way down to the bottom just to discover that.
247 * Why all these "- 1"s? Because 0 represents both the bottom
248 * of the address space and the top of it (using -1 for the
249 * top wouldn't help much: the masks would do the wrong thing).
250 * The rule is that addr 0 and floor 0 refer to the bottom of
251 * the address space, but end 0 and ceiling 0 refer to the top
252 * Comparisons need to use "end - 1" and "ceiling - 1" (though
253 * that end 0 case should be mythical).
255 * Wherever addr is brought up or ceiling brought down, we must
256 * be careful to reject "the opposite 0" before it confuses the
257 * subsequent tests. But what about where end is brought down
258 * by PMD_SIZE below? no, end can't go down to 0 there.
260 * Whereas we round start (addr) and ceiling down, by different
261 * masks at different levels, in order to test whether a table
262 * now has no other vmas using it, so can be freed, we don't
263 * bother to round floor or end up - the tests don't need that.
266 addr &= PMD_MASK;
267 if (addr < floor) {
268 addr += PMD_SIZE;
269 if (!addr)
270 return;
272 if (ceiling) {
273 ceiling &= PMD_MASK;
274 if (!ceiling)
275 return;
277 if (end - 1 > ceiling - 1)
278 end -= PMD_SIZE;
279 if (addr > end - 1)
280 return;
282 start = addr;
283 pgd = pgd_offset(tlb->mm, addr);
284 do {
285 next = pgd_addr_end(addr, end);
286 if (pgd_none_or_clear_bad(pgd))
287 continue;
288 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
289 } while (pgd++, addr = next, addr != end);
292 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
293 unsigned long floor, unsigned long ceiling)
295 while (vma) {
296 struct vm_area_struct *next = vma->vm_next;
297 unsigned long addr = vma->vm_start;
300 * Hide vma from rmap and truncate_pagecache before freeing
301 * pgtables
303 anon_vma_unlink(vma);
304 unlink_file_vma(vma);
306 if (is_vm_hugetlb_page(vma)) {
307 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
308 floor, next? next->vm_start: ceiling);
309 } else {
311 * Optimization: gather nearby vmas into one call down
313 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
314 && !is_vm_hugetlb_page(next)) {
315 vma = next;
316 next = vma->vm_next;
317 anon_vma_unlink(vma);
318 unlink_file_vma(vma);
320 free_pgd_range(tlb, addr, vma->vm_end,
321 floor, next? next->vm_start: ceiling);
323 vma = next;
327 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
329 pgtable_t new = pte_alloc_one(mm, address);
330 if (!new)
331 return -ENOMEM;
334 * Ensure all pte setup (eg. pte page lock and page clearing) are
335 * visible before the pte is made visible to other CPUs by being
336 * put into page tables.
338 * The other side of the story is the pointer chasing in the page
339 * table walking code (when walking the page table without locking;
340 * ie. most of the time). Fortunately, these data accesses consist
341 * of a chain of data-dependent loads, meaning most CPUs (alpha
342 * being the notable exception) will already guarantee loads are
343 * seen in-order. See the alpha page table accessors for the
344 * smp_read_barrier_depends() barriers in page table walking code.
346 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
348 spin_lock(&mm->page_table_lock);
349 if (!pmd_present(*pmd)) { /* Has another populated it ? */
350 mm->nr_ptes++;
351 pmd_populate(mm, pmd, new);
352 new = NULL;
354 spin_unlock(&mm->page_table_lock);
355 if (new)
356 pte_free(mm, new);
357 return 0;
360 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
362 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
363 if (!new)
364 return -ENOMEM;
366 smp_wmb(); /* See comment in __pte_alloc */
368 spin_lock(&init_mm.page_table_lock);
369 if (!pmd_present(*pmd)) { /* Has another populated it ? */
370 pmd_populate_kernel(&init_mm, pmd, new);
371 new = NULL;
373 spin_unlock(&init_mm.page_table_lock);
374 if (new)
375 pte_free_kernel(&init_mm, new);
376 return 0;
379 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
381 if (file_rss)
382 add_mm_counter(mm, file_rss, file_rss);
383 if (anon_rss)
384 add_mm_counter(mm, anon_rss, anon_rss);
388 * This function is called to print an error when a bad pte
389 * is found. For example, we might have a PFN-mapped pte in
390 * a region that doesn't allow it.
392 * The calling function must still handle the error.
394 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
395 pte_t pte, struct page *page)
397 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
398 pud_t *pud = pud_offset(pgd, addr);
399 pmd_t *pmd = pmd_offset(pud, addr);
400 struct address_space *mapping;
401 pgoff_t index;
402 static unsigned long resume;
403 static unsigned long nr_shown;
404 static unsigned long nr_unshown;
407 * Allow a burst of 60 reports, then keep quiet for that minute;
408 * or allow a steady drip of one report per second.
410 if (nr_shown == 60) {
411 if (time_before(jiffies, resume)) {
412 nr_unshown++;
413 return;
415 if (nr_unshown) {
416 printk(KERN_ALERT
417 "BUG: Bad page map: %lu messages suppressed\n",
418 nr_unshown);
419 nr_unshown = 0;
421 nr_shown = 0;
423 if (nr_shown++ == 0)
424 resume = jiffies + 60 * HZ;
426 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
427 index = linear_page_index(vma, addr);
429 printk(KERN_ALERT
430 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
431 current->comm,
432 (long long)pte_val(pte), (long long)pmd_val(*pmd));
433 if (page) {
434 printk(KERN_ALERT
435 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
436 page, (void *)page->flags, page_count(page),
437 page_mapcount(page), page->mapping, page->index);
439 printk(KERN_ALERT
440 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
441 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
443 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
445 if (vma->vm_ops)
446 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
447 (unsigned long)vma->vm_ops->fault);
448 if (vma->vm_file && vma->vm_file->f_op)
449 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
450 (unsigned long)vma->vm_file->f_op->mmap);
451 dump_stack();
452 add_taint(TAINT_BAD_PAGE);
455 static inline int is_cow_mapping(unsigned int flags)
457 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
460 #ifndef is_zero_pfn
461 static inline int is_zero_pfn(unsigned long pfn)
463 return pfn == zero_pfn;
465 #endif
467 #ifndef my_zero_pfn
468 static inline unsigned long my_zero_pfn(unsigned long addr)
470 return zero_pfn;
472 #endif
475 * vm_normal_page -- This function gets the "struct page" associated with a pte.
477 * "Special" mappings do not wish to be associated with a "struct page" (either
478 * it doesn't exist, or it exists but they don't want to touch it). In this
479 * case, NULL is returned here. "Normal" mappings do have a struct page.
481 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
482 * pte bit, in which case this function is trivial. Secondly, an architecture
483 * may not have a spare pte bit, which requires a more complicated scheme,
484 * described below.
486 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
487 * special mapping (even if there are underlying and valid "struct pages").
488 * COWed pages of a VM_PFNMAP are always normal.
490 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
491 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
492 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
493 * mapping will always honor the rule
495 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
497 * And for normal mappings this is false.
499 * This restricts such mappings to be a linear translation from virtual address
500 * to pfn. To get around this restriction, we allow arbitrary mappings so long
501 * as the vma is not a COW mapping; in that case, we know that all ptes are
502 * special (because none can have been COWed).
505 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
507 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
508 * page" backing, however the difference is that _all_ pages with a struct
509 * page (that is, those where pfn_valid is true) are refcounted and considered
510 * normal pages by the VM. The disadvantage is that pages are refcounted
511 * (which can be slower and simply not an option for some PFNMAP users). The
512 * advantage is that we don't have to follow the strict linearity rule of
513 * PFNMAP mappings in order to support COWable mappings.
516 #ifdef __HAVE_ARCH_PTE_SPECIAL
517 # define HAVE_PTE_SPECIAL 1
518 #else
519 # define HAVE_PTE_SPECIAL 0
520 #endif
521 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
522 pte_t pte)
524 unsigned long pfn = pte_pfn(pte);
526 if (HAVE_PTE_SPECIAL) {
527 if (likely(!pte_special(pte)))
528 goto check_pfn;
529 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
530 return NULL;
531 if (!is_zero_pfn(pfn))
532 print_bad_pte(vma, addr, pte, NULL);
533 return NULL;
536 /* !HAVE_PTE_SPECIAL case follows: */
538 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
539 if (vma->vm_flags & VM_MIXEDMAP) {
540 if (!pfn_valid(pfn))
541 return NULL;
542 goto out;
543 } else {
544 unsigned long off;
545 off = (addr - vma->vm_start) >> PAGE_SHIFT;
546 if (pfn == vma->vm_pgoff + off)
547 return NULL;
548 if (!is_cow_mapping(vma->vm_flags))
549 return NULL;
553 if (is_zero_pfn(pfn))
554 return NULL;
555 check_pfn:
556 if (unlikely(pfn > highest_memmap_pfn)) {
557 print_bad_pte(vma, addr, pte, NULL);
558 return NULL;
562 * NOTE! We still have PageReserved() pages in the page tables.
563 * eg. VDSO mappings can cause them to exist.
565 out:
566 return pfn_to_page(pfn);
570 * copy one vm_area from one task to the other. Assumes the page tables
571 * already present in the new task to be cleared in the whole range
572 * covered by this vma.
575 static inline void
576 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
577 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
578 unsigned long addr, int *rss)
580 unsigned long vm_flags = vma->vm_flags;
581 pte_t pte = *src_pte;
582 struct page *page;
584 /* pte contains position in swap or file, so copy. */
585 if (unlikely(!pte_present(pte))) {
586 if (!pte_file(pte)) {
587 swp_entry_t entry = pte_to_swp_entry(pte);
589 swap_duplicate(entry);
590 /* make sure dst_mm is on swapoff's mmlist. */
591 if (unlikely(list_empty(&dst_mm->mmlist))) {
592 spin_lock(&mmlist_lock);
593 if (list_empty(&dst_mm->mmlist))
594 list_add(&dst_mm->mmlist,
595 &src_mm->mmlist);
596 spin_unlock(&mmlist_lock);
598 if (is_write_migration_entry(entry) &&
599 is_cow_mapping(vm_flags)) {
601 * COW mappings require pages in both parent
602 * and child to be set to read.
604 make_migration_entry_read(&entry);
605 pte = swp_entry_to_pte(entry);
606 set_pte_at(src_mm, addr, src_pte, pte);
609 goto out_set_pte;
613 * If it's a COW mapping, write protect it both
614 * in the parent and the child
616 if (is_cow_mapping(vm_flags)) {
617 ptep_set_wrprotect(src_mm, addr, src_pte);
618 pte = pte_wrprotect(pte);
622 * If it's a shared mapping, mark it clean in
623 * the child
625 if (vm_flags & VM_SHARED)
626 pte = pte_mkclean(pte);
627 pte = pte_mkold(pte);
629 page = vm_normal_page(vma, addr, pte);
630 if (page) {
631 get_page(page);
632 page_dup_rmap(page);
633 rss[PageAnon(page)]++;
636 out_set_pte:
637 set_pte_at(dst_mm, addr, dst_pte, pte);
640 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
641 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
642 unsigned long addr, unsigned long end)
644 pte_t *src_pte, *dst_pte;
645 spinlock_t *src_ptl, *dst_ptl;
646 int progress = 0;
647 int rss[2];
649 again:
650 rss[1] = rss[0] = 0;
651 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
652 if (!dst_pte)
653 return -ENOMEM;
654 src_pte = pte_offset_map_nested(src_pmd, addr);
655 src_ptl = pte_lockptr(src_mm, src_pmd);
656 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
657 arch_enter_lazy_mmu_mode();
659 do {
661 * We are holding two locks at this point - either of them
662 * could generate latencies in another task on another CPU.
664 if (progress >= 32) {
665 progress = 0;
666 if (need_resched() ||
667 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
668 break;
670 if (pte_none(*src_pte)) {
671 progress++;
672 continue;
674 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
675 progress += 8;
676 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
678 arch_leave_lazy_mmu_mode();
679 spin_unlock(src_ptl);
680 pte_unmap_nested(src_pte - 1);
681 add_mm_rss(dst_mm, rss[0], rss[1]);
682 pte_unmap_unlock(dst_pte - 1, dst_ptl);
683 cond_resched();
684 if (addr != end)
685 goto again;
686 return 0;
689 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
690 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
691 unsigned long addr, unsigned long end)
693 pmd_t *src_pmd, *dst_pmd;
694 unsigned long next;
696 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
697 if (!dst_pmd)
698 return -ENOMEM;
699 src_pmd = pmd_offset(src_pud, addr);
700 do {
701 next = pmd_addr_end(addr, end);
702 if (pmd_none_or_clear_bad(src_pmd))
703 continue;
704 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
705 vma, addr, next))
706 return -ENOMEM;
707 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
708 return 0;
711 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
712 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
713 unsigned long addr, unsigned long end)
715 pud_t *src_pud, *dst_pud;
716 unsigned long next;
718 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
719 if (!dst_pud)
720 return -ENOMEM;
721 src_pud = pud_offset(src_pgd, addr);
722 do {
723 next = pud_addr_end(addr, end);
724 if (pud_none_or_clear_bad(src_pud))
725 continue;
726 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
727 vma, addr, next))
728 return -ENOMEM;
729 } while (dst_pud++, src_pud++, addr = next, addr != end);
730 return 0;
733 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
734 struct vm_area_struct *vma)
736 pgd_t *src_pgd, *dst_pgd;
737 unsigned long next;
738 unsigned long addr = vma->vm_start;
739 unsigned long end = vma->vm_end;
740 int ret;
743 * Don't copy ptes where a page fault will fill them correctly.
744 * Fork becomes much lighter when there are big shared or private
745 * readonly mappings. The tradeoff is that copy_page_range is more
746 * efficient than faulting.
748 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
749 if (!vma->anon_vma)
750 return 0;
753 if (is_vm_hugetlb_page(vma))
754 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
756 if (unlikely(is_pfn_mapping(vma))) {
758 * We do not free on error cases below as remove_vma
759 * gets called on error from higher level routine
761 ret = track_pfn_vma_copy(vma);
762 if (ret)
763 return ret;
767 * We need to invalidate the secondary MMU mappings only when
768 * there could be a permission downgrade on the ptes of the
769 * parent mm. And a permission downgrade will only happen if
770 * is_cow_mapping() returns true.
772 if (is_cow_mapping(vma->vm_flags))
773 mmu_notifier_invalidate_range_start(src_mm, addr, end);
775 ret = 0;
776 dst_pgd = pgd_offset(dst_mm, addr);
777 src_pgd = pgd_offset(src_mm, addr);
778 do {
779 next = pgd_addr_end(addr, end);
780 if (pgd_none_or_clear_bad(src_pgd))
781 continue;
782 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
783 vma, addr, next))) {
784 ret = -ENOMEM;
785 break;
787 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
789 if (is_cow_mapping(vma->vm_flags))
790 mmu_notifier_invalidate_range_end(src_mm,
791 vma->vm_start, end);
792 return ret;
795 static unsigned long zap_pte_range(struct mmu_gather *tlb,
796 struct vm_area_struct *vma, pmd_t *pmd,
797 unsigned long addr, unsigned long end,
798 long *zap_work, struct zap_details *details)
800 struct mm_struct *mm = tlb->mm;
801 pte_t *pte;
802 spinlock_t *ptl;
803 int file_rss = 0;
804 int anon_rss = 0;
806 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
807 arch_enter_lazy_mmu_mode();
808 do {
809 pte_t ptent = *pte;
810 if (pte_none(ptent)) {
811 (*zap_work)--;
812 continue;
815 (*zap_work) -= PAGE_SIZE;
817 if (pte_present(ptent)) {
818 struct page *page;
820 page = vm_normal_page(vma, addr, ptent);
821 if (unlikely(details) && page) {
823 * unmap_shared_mapping_pages() wants to
824 * invalidate cache without truncating:
825 * unmap shared but keep private pages.
827 if (details->check_mapping &&
828 details->check_mapping != page->mapping)
829 continue;
831 * Each page->index must be checked when
832 * invalidating or truncating nonlinear.
834 if (details->nonlinear_vma &&
835 (page->index < details->first_index ||
836 page->index > details->last_index))
837 continue;
839 ptent = ptep_get_and_clear_full(mm, addr, pte,
840 tlb->fullmm);
841 tlb_remove_tlb_entry(tlb, pte, addr);
842 if (unlikely(!page))
843 continue;
844 if (unlikely(details) && details->nonlinear_vma
845 && linear_page_index(details->nonlinear_vma,
846 addr) != page->index)
847 set_pte_at(mm, addr, pte,
848 pgoff_to_pte(page->index));
849 if (PageAnon(page))
850 anon_rss--;
851 else {
852 if (pte_dirty(ptent))
853 set_page_dirty(page);
854 if (pte_young(ptent) &&
855 likely(!VM_SequentialReadHint(vma)))
856 mark_page_accessed(page);
857 file_rss--;
859 page_remove_rmap(page);
860 if (unlikely(page_mapcount(page) < 0))
861 print_bad_pte(vma, addr, ptent, page);
862 tlb_remove_page(tlb, page);
863 continue;
866 * If details->check_mapping, we leave swap entries;
867 * if details->nonlinear_vma, we leave file entries.
869 if (unlikely(details))
870 continue;
871 if (pte_file(ptent)) {
872 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
873 print_bad_pte(vma, addr, ptent, NULL);
874 } else if
875 (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
876 print_bad_pte(vma, addr, ptent, NULL);
877 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
878 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
880 add_mm_rss(mm, file_rss, anon_rss);
881 arch_leave_lazy_mmu_mode();
882 pte_unmap_unlock(pte - 1, ptl);
884 return addr;
887 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
888 struct vm_area_struct *vma, pud_t *pud,
889 unsigned long addr, unsigned long end,
890 long *zap_work, struct zap_details *details)
892 pmd_t *pmd;
893 unsigned long next;
895 pmd = pmd_offset(pud, addr);
896 do {
897 next = pmd_addr_end(addr, end);
898 if (pmd_none_or_clear_bad(pmd)) {
899 (*zap_work)--;
900 continue;
902 next = zap_pte_range(tlb, vma, pmd, addr, next,
903 zap_work, details);
904 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
906 return addr;
909 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
910 struct vm_area_struct *vma, pgd_t *pgd,
911 unsigned long addr, unsigned long end,
912 long *zap_work, struct zap_details *details)
914 pud_t *pud;
915 unsigned long next;
917 pud = pud_offset(pgd, addr);
918 do {
919 next = pud_addr_end(addr, end);
920 if (pud_none_or_clear_bad(pud)) {
921 (*zap_work)--;
922 continue;
924 next = zap_pmd_range(tlb, vma, pud, addr, next,
925 zap_work, details);
926 } while (pud++, addr = next, (addr != end && *zap_work > 0));
928 return addr;
931 static unsigned long unmap_page_range(struct mmu_gather *tlb,
932 struct vm_area_struct *vma,
933 unsigned long addr, unsigned long end,
934 long *zap_work, struct zap_details *details)
936 pgd_t *pgd;
937 unsigned long next;
939 if (details && !details->check_mapping && !details->nonlinear_vma)
940 details = NULL;
942 BUG_ON(addr >= end);
943 mem_cgroup_uncharge_start();
944 tlb_start_vma(tlb, vma);
945 pgd = pgd_offset(vma->vm_mm, addr);
946 do {
947 next = pgd_addr_end(addr, end);
948 if (pgd_none_or_clear_bad(pgd)) {
949 (*zap_work)--;
950 continue;
952 next = zap_pud_range(tlb, vma, pgd, addr, next,
953 zap_work, details);
954 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
955 tlb_end_vma(tlb, vma);
956 mem_cgroup_uncharge_end();
958 return addr;
961 #ifdef CONFIG_PREEMPT
962 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
963 #else
964 /* No preempt: go for improved straight-line efficiency */
965 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
966 #endif
969 * unmap_vmas - unmap a range of memory covered by a list of vma's
970 * @tlbp: address of the caller's struct mmu_gather
971 * @vma: the starting vma
972 * @start_addr: virtual address at which to start unmapping
973 * @end_addr: virtual address at which to end unmapping
974 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
975 * @details: details of nonlinear truncation or shared cache invalidation
977 * Returns the end address of the unmapping (restart addr if interrupted).
979 * Unmap all pages in the vma list.
981 * We aim to not hold locks for too long (for scheduling latency reasons).
982 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
983 * return the ending mmu_gather to the caller.
985 * Only addresses between `start' and `end' will be unmapped.
987 * The VMA list must be sorted in ascending virtual address order.
989 * unmap_vmas() assumes that the caller will flush the whole unmapped address
990 * range after unmap_vmas() returns. So the only responsibility here is to
991 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
992 * drops the lock and schedules.
994 unsigned long unmap_vmas(struct mmu_gather **tlbp,
995 struct vm_area_struct *vma, unsigned long start_addr,
996 unsigned long end_addr, unsigned long *nr_accounted,
997 struct zap_details *details)
999 long zap_work = ZAP_BLOCK_SIZE;
1000 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
1001 int tlb_start_valid = 0;
1002 unsigned long start = start_addr;
1003 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1004 int fullmm = (*tlbp)->fullmm;
1005 struct mm_struct *mm = vma->vm_mm;
1007 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1008 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1009 unsigned long end;
1011 start = max(vma->vm_start, start_addr);
1012 if (start >= vma->vm_end)
1013 continue;
1014 end = min(vma->vm_end, end_addr);
1015 if (end <= vma->vm_start)
1016 continue;
1018 if (vma->vm_flags & VM_ACCOUNT)
1019 *nr_accounted += (end - start) >> PAGE_SHIFT;
1021 if (unlikely(is_pfn_mapping(vma)))
1022 untrack_pfn_vma(vma, 0, 0);
1024 while (start != end) {
1025 if (!tlb_start_valid) {
1026 tlb_start = start;
1027 tlb_start_valid = 1;
1030 if (unlikely(is_vm_hugetlb_page(vma))) {
1032 * It is undesirable to test vma->vm_file as it
1033 * should be non-null for valid hugetlb area.
1034 * However, vm_file will be NULL in the error
1035 * cleanup path of do_mmap_pgoff. When
1036 * hugetlbfs ->mmap method fails,
1037 * do_mmap_pgoff() nullifies vma->vm_file
1038 * before calling this function to clean up.
1039 * Since no pte has actually been setup, it is
1040 * safe to do nothing in this case.
1042 if (vma->vm_file) {
1043 unmap_hugepage_range(vma, start, end, NULL);
1044 zap_work -= (end - start) /
1045 pages_per_huge_page(hstate_vma(vma));
1048 start = end;
1049 } else
1050 start = unmap_page_range(*tlbp, vma,
1051 start, end, &zap_work, details);
1053 if (zap_work > 0) {
1054 BUG_ON(start != end);
1055 break;
1058 tlb_finish_mmu(*tlbp, tlb_start, start);
1060 if (need_resched() ||
1061 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1062 if (i_mmap_lock) {
1063 *tlbp = NULL;
1064 goto out;
1066 cond_resched();
1069 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1070 tlb_start_valid = 0;
1071 zap_work = ZAP_BLOCK_SIZE;
1074 out:
1075 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1076 return start; /* which is now the end (or restart) address */
1080 * zap_page_range - remove user pages in a given range
1081 * @vma: vm_area_struct holding the applicable pages
1082 * @address: starting address of pages to zap
1083 * @size: number of bytes to zap
1084 * @details: details of nonlinear truncation or shared cache invalidation
1086 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1087 unsigned long size, struct zap_details *details)
1089 struct mm_struct *mm = vma->vm_mm;
1090 struct mmu_gather *tlb;
1091 unsigned long end = address + size;
1092 unsigned long nr_accounted = 0;
1094 lru_add_drain();
1095 tlb = tlb_gather_mmu(mm, 0);
1096 update_hiwater_rss(mm);
1097 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1098 if (tlb)
1099 tlb_finish_mmu(tlb, address, end);
1100 return end;
1104 * zap_vma_ptes - remove ptes mapping the vma
1105 * @vma: vm_area_struct holding ptes to be zapped
1106 * @address: starting address of pages to zap
1107 * @size: number of bytes to zap
1109 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1111 * The entire address range must be fully contained within the vma.
1113 * Returns 0 if successful.
1115 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1116 unsigned long size)
1118 if (address < vma->vm_start || address + size > vma->vm_end ||
1119 !(vma->vm_flags & VM_PFNMAP))
1120 return -1;
1121 zap_page_range(vma, address, size, NULL);
1122 return 0;
1124 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1127 * Do a quick page-table lookup for a single page.
1129 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1130 unsigned int flags)
1132 pgd_t *pgd;
1133 pud_t *pud;
1134 pmd_t *pmd;
1135 pte_t *ptep, pte;
1136 spinlock_t *ptl;
1137 struct page *page;
1138 struct mm_struct *mm = vma->vm_mm;
1140 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1141 if (!IS_ERR(page)) {
1142 BUG_ON(flags & FOLL_GET);
1143 goto out;
1146 page = NULL;
1147 pgd = pgd_offset(mm, address);
1148 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1149 goto no_page_table;
1151 pud = pud_offset(pgd, address);
1152 if (pud_none(*pud))
1153 goto no_page_table;
1154 if (pud_huge(*pud)) {
1155 BUG_ON(flags & FOLL_GET);
1156 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1157 goto out;
1159 if (unlikely(pud_bad(*pud)))
1160 goto no_page_table;
1162 pmd = pmd_offset(pud, address);
1163 if (pmd_none(*pmd))
1164 goto no_page_table;
1165 if (pmd_huge(*pmd)) {
1166 BUG_ON(flags & FOLL_GET);
1167 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1168 goto out;
1170 if (unlikely(pmd_bad(*pmd)))
1171 goto no_page_table;
1173 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1175 pte = *ptep;
1176 if (!pte_present(pte))
1177 goto no_page;
1178 if ((flags & FOLL_WRITE) && !pte_write(pte))
1179 goto unlock;
1181 page = vm_normal_page(vma, address, pte);
1182 if (unlikely(!page)) {
1183 if ((flags & FOLL_DUMP) ||
1184 !is_zero_pfn(pte_pfn(pte)))
1185 goto bad_page;
1186 page = pte_page(pte);
1189 if (flags & FOLL_GET)
1190 get_page(page);
1191 if (flags & FOLL_TOUCH) {
1192 if ((flags & FOLL_WRITE) &&
1193 !pte_dirty(pte) && !PageDirty(page))
1194 set_page_dirty(page);
1196 * pte_mkyoung() would be more correct here, but atomic care
1197 * is needed to avoid losing the dirty bit: it is easier to use
1198 * mark_page_accessed().
1200 mark_page_accessed(page);
1202 unlock:
1203 pte_unmap_unlock(ptep, ptl);
1204 out:
1205 return page;
1207 bad_page:
1208 pte_unmap_unlock(ptep, ptl);
1209 return ERR_PTR(-EFAULT);
1211 no_page:
1212 pte_unmap_unlock(ptep, ptl);
1213 if (!pte_none(pte))
1214 return page;
1216 no_page_table:
1218 * When core dumping an enormous anonymous area that nobody
1219 * has touched so far, we don't want to allocate unnecessary pages or
1220 * page tables. Return error instead of NULL to skip handle_mm_fault,
1221 * then get_dump_page() will return NULL to leave a hole in the dump.
1222 * But we can only make this optimization where a hole would surely
1223 * be zero-filled if handle_mm_fault() actually did handle it.
1225 if ((flags & FOLL_DUMP) &&
1226 (!vma->vm_ops || !vma->vm_ops->fault))
1227 return ERR_PTR(-EFAULT);
1228 return page;
1231 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1232 unsigned long start, int nr_pages, unsigned int gup_flags,
1233 struct page **pages, struct vm_area_struct **vmas)
1235 int i;
1236 unsigned long vm_flags;
1238 if (nr_pages <= 0)
1239 return 0;
1241 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1244 * Require read or write permissions.
1245 * If FOLL_FORCE is set, we only require the "MAY" flags.
1247 vm_flags = (gup_flags & FOLL_WRITE) ?
1248 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1249 vm_flags &= (gup_flags & FOLL_FORCE) ?
1250 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1251 i = 0;
1253 do {
1254 struct vm_area_struct *vma;
1256 vma = find_extend_vma(mm, start);
1257 if (!vma && in_gate_area(tsk, start)) {
1258 unsigned long pg = start & PAGE_MASK;
1259 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1260 pgd_t *pgd;
1261 pud_t *pud;
1262 pmd_t *pmd;
1263 pte_t *pte;
1265 /* user gate pages are read-only */
1266 if (gup_flags & FOLL_WRITE)
1267 return i ? : -EFAULT;
1268 if (pg > TASK_SIZE)
1269 pgd = pgd_offset_k(pg);
1270 else
1271 pgd = pgd_offset_gate(mm, pg);
1272 BUG_ON(pgd_none(*pgd));
1273 pud = pud_offset(pgd, pg);
1274 BUG_ON(pud_none(*pud));
1275 pmd = pmd_offset(pud, pg);
1276 if (pmd_none(*pmd))
1277 return i ? : -EFAULT;
1278 pte = pte_offset_map(pmd, pg);
1279 if (pte_none(*pte)) {
1280 pte_unmap(pte);
1281 return i ? : -EFAULT;
1283 if (pages) {
1284 struct page *page = vm_normal_page(gate_vma, start, *pte);
1285 pages[i] = page;
1286 if (page)
1287 get_page(page);
1289 pte_unmap(pte);
1290 if (vmas)
1291 vmas[i] = gate_vma;
1292 i++;
1293 start += PAGE_SIZE;
1294 nr_pages--;
1295 continue;
1298 if (!vma ||
1299 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1300 !(vm_flags & vma->vm_flags))
1301 return i ? : -EFAULT;
1303 if (is_vm_hugetlb_page(vma)) {
1304 i = follow_hugetlb_page(mm, vma, pages, vmas,
1305 &start, &nr_pages, i, gup_flags);
1306 continue;
1309 do {
1310 struct page *page;
1311 unsigned int foll_flags = gup_flags;
1314 * If we have a pending SIGKILL, don't keep faulting
1315 * pages and potentially allocating memory.
1317 if (unlikely(fatal_signal_pending(current)))
1318 return i ? i : -ERESTARTSYS;
1320 cond_resched();
1321 while (!(page = follow_page(vma, start, foll_flags))) {
1322 int ret;
1324 ret = handle_mm_fault(mm, vma, start,
1325 (foll_flags & FOLL_WRITE) ?
1326 FAULT_FLAG_WRITE : 0);
1328 if (ret & VM_FAULT_ERROR) {
1329 if (ret & VM_FAULT_OOM)
1330 return i ? i : -ENOMEM;
1331 if (ret &
1332 (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1333 return i ? i : -EFAULT;
1334 BUG();
1336 if (ret & VM_FAULT_MAJOR)
1337 tsk->maj_flt++;
1338 else
1339 tsk->min_flt++;
1342 * The VM_FAULT_WRITE bit tells us that
1343 * do_wp_page has broken COW when necessary,
1344 * even if maybe_mkwrite decided not to set
1345 * pte_write. We can thus safely do subsequent
1346 * page lookups as if they were reads. But only
1347 * do so when looping for pte_write is futile:
1348 * in some cases userspace may also be wanting
1349 * to write to the gotten user page, which a
1350 * read fault here might prevent (a readonly
1351 * page might get reCOWed by userspace write).
1353 if ((ret & VM_FAULT_WRITE) &&
1354 !(vma->vm_flags & VM_WRITE))
1355 foll_flags &= ~FOLL_WRITE;
1357 cond_resched();
1359 if (IS_ERR(page))
1360 return i ? i : PTR_ERR(page);
1361 if (pages) {
1362 pages[i] = page;
1364 flush_anon_page(vma, page, start);
1365 flush_dcache_page(page);
1367 if (vmas)
1368 vmas[i] = vma;
1369 i++;
1370 start += PAGE_SIZE;
1371 nr_pages--;
1372 } while (nr_pages && start < vma->vm_end);
1373 } while (nr_pages);
1374 return i;
1378 * get_user_pages() - pin user pages in memory
1379 * @tsk: task_struct of target task
1380 * @mm: mm_struct of target mm
1381 * @start: starting user address
1382 * @nr_pages: number of pages from start to pin
1383 * @write: whether pages will be written to by the caller
1384 * @force: whether to force write access even if user mapping is
1385 * readonly. This will result in the page being COWed even
1386 * in MAP_SHARED mappings. You do not want this.
1387 * @pages: array that receives pointers to the pages pinned.
1388 * Should be at least nr_pages long. Or NULL, if caller
1389 * only intends to ensure the pages are faulted in.
1390 * @vmas: array of pointers to vmas corresponding to each page.
1391 * Or NULL if the caller does not require them.
1393 * Returns number of pages pinned. This may be fewer than the number
1394 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1395 * were pinned, returns -errno. Each page returned must be released
1396 * with a put_page() call when it is finished with. vmas will only
1397 * remain valid while mmap_sem is held.
1399 * Must be called with mmap_sem held for read or write.
1401 * get_user_pages walks a process's page tables and takes a reference to
1402 * each struct page that each user address corresponds to at a given
1403 * instant. That is, it takes the page that would be accessed if a user
1404 * thread accesses the given user virtual address at that instant.
1406 * This does not guarantee that the page exists in the user mappings when
1407 * get_user_pages returns, and there may even be a completely different
1408 * page there in some cases (eg. if mmapped pagecache has been invalidated
1409 * and subsequently re faulted). However it does guarantee that the page
1410 * won't be freed completely. And mostly callers simply care that the page
1411 * contains data that was valid *at some point in time*. Typically, an IO
1412 * or similar operation cannot guarantee anything stronger anyway because
1413 * locks can't be held over the syscall boundary.
1415 * If write=0, the page must not be written to. If the page is written to,
1416 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1417 * after the page is finished with, and before put_page is called.
1419 * get_user_pages is typically used for fewer-copy IO operations, to get a
1420 * handle on the memory by some means other than accesses via the user virtual
1421 * addresses. The pages may be submitted for DMA to devices or accessed via
1422 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1423 * use the correct cache flushing APIs.
1425 * See also get_user_pages_fast, for performance critical applications.
1427 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1428 unsigned long start, int nr_pages, int write, int force,
1429 struct page **pages, struct vm_area_struct **vmas)
1431 int flags = FOLL_TOUCH;
1433 if (pages)
1434 flags |= FOLL_GET;
1435 if (write)
1436 flags |= FOLL_WRITE;
1437 if (force)
1438 flags |= FOLL_FORCE;
1440 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1442 EXPORT_SYMBOL(get_user_pages);
1445 * get_dump_page() - pin user page in memory while writing it to core dump
1446 * @addr: user address
1448 * Returns struct page pointer of user page pinned for dump,
1449 * to be freed afterwards by page_cache_release() or put_page().
1451 * Returns NULL on any kind of failure - a hole must then be inserted into
1452 * the corefile, to preserve alignment with its headers; and also returns
1453 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1454 * allowing a hole to be left in the corefile to save diskspace.
1456 * Called without mmap_sem, but after all other threads have been killed.
1458 #ifdef CONFIG_ELF_CORE
1459 struct page *get_dump_page(unsigned long addr)
1461 struct vm_area_struct *vma;
1462 struct page *page;
1464 if (__get_user_pages(current, current->mm, addr, 1,
1465 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1466 return NULL;
1467 flush_cache_page(vma, addr, page_to_pfn(page));
1468 return page;
1470 #endif /* CONFIG_ELF_CORE */
1472 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1473 spinlock_t **ptl)
1475 pgd_t * pgd = pgd_offset(mm, addr);
1476 pud_t * pud = pud_alloc(mm, pgd, addr);
1477 if (pud) {
1478 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1479 if (pmd)
1480 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1482 return NULL;
1486 * This is the old fallback for page remapping.
1488 * For historical reasons, it only allows reserved pages. Only
1489 * old drivers should use this, and they needed to mark their
1490 * pages reserved for the old functions anyway.
1492 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1493 struct page *page, pgprot_t prot)
1495 struct mm_struct *mm = vma->vm_mm;
1496 int retval;
1497 pte_t *pte;
1498 spinlock_t *ptl;
1500 retval = -EINVAL;
1501 if (PageAnon(page))
1502 goto out;
1503 retval = -ENOMEM;
1504 flush_dcache_page(page);
1505 pte = get_locked_pte(mm, addr, &ptl);
1506 if (!pte)
1507 goto out;
1508 retval = -EBUSY;
1509 if (!pte_none(*pte))
1510 goto out_unlock;
1512 /* Ok, finally just insert the thing.. */
1513 get_page(page);
1514 inc_mm_counter(mm, file_rss);
1515 page_add_file_rmap(page);
1516 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1518 retval = 0;
1519 pte_unmap_unlock(pte, ptl);
1520 return retval;
1521 out_unlock:
1522 pte_unmap_unlock(pte, ptl);
1523 out:
1524 return retval;
1528 * vm_insert_page - insert single page into user vma
1529 * @vma: user vma to map to
1530 * @addr: target user address of this page
1531 * @page: source kernel page
1533 * This allows drivers to insert individual pages they've allocated
1534 * into a user vma.
1536 * The page has to be a nice clean _individual_ kernel allocation.
1537 * If you allocate a compound page, you need to have marked it as
1538 * such (__GFP_COMP), or manually just split the page up yourself
1539 * (see split_page()).
1541 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1542 * took an arbitrary page protection parameter. This doesn't allow
1543 * that. Your vma protection will have to be set up correctly, which
1544 * means that if you want a shared writable mapping, you'd better
1545 * ask for a shared writable mapping!
1547 * The page does not need to be reserved.
1549 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1550 struct page *page)
1552 if (addr < vma->vm_start || addr >= vma->vm_end)
1553 return -EFAULT;
1554 if (!page_count(page))
1555 return -EINVAL;
1556 vma->vm_flags |= VM_INSERTPAGE;
1557 return insert_page(vma, addr, page, vma->vm_page_prot);
1559 EXPORT_SYMBOL(vm_insert_page);
1561 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1562 unsigned long pfn, pgprot_t prot)
1564 struct mm_struct *mm = vma->vm_mm;
1565 int retval;
1566 pte_t *pte, entry;
1567 spinlock_t *ptl;
1569 retval = -ENOMEM;
1570 pte = get_locked_pte(mm, addr, &ptl);
1571 if (!pte)
1572 goto out;
1573 retval = -EBUSY;
1574 if (!pte_none(*pte))
1575 goto out_unlock;
1577 /* Ok, finally just insert the thing.. */
1578 entry = pte_mkspecial(pfn_pte(pfn, prot));
1579 set_pte_at(mm, addr, pte, entry);
1580 update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1582 retval = 0;
1583 out_unlock:
1584 pte_unmap_unlock(pte, ptl);
1585 out:
1586 return retval;
1590 * vm_insert_pfn - insert single pfn into user vma
1591 * @vma: user vma to map to
1592 * @addr: target user address of this page
1593 * @pfn: source kernel pfn
1595 * Similar to vm_inert_page, this allows drivers to insert individual pages
1596 * they've allocated into a user vma. Same comments apply.
1598 * This function should only be called from a vm_ops->fault handler, and
1599 * in that case the handler should return NULL.
1601 * vma cannot be a COW mapping.
1603 * As this is called only for pages that do not currently exist, we
1604 * do not need to flush old virtual caches or the TLB.
1606 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1607 unsigned long pfn)
1609 int ret;
1610 pgprot_t pgprot = vma->vm_page_prot;
1612 * Technically, architectures with pte_special can avoid all these
1613 * restrictions (same for remap_pfn_range). However we would like
1614 * consistency in testing and feature parity among all, so we should
1615 * try to keep these invariants in place for everybody.
1617 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1618 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1619 (VM_PFNMAP|VM_MIXEDMAP));
1620 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1621 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1623 if (addr < vma->vm_start || addr >= vma->vm_end)
1624 return -EFAULT;
1625 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1626 return -EINVAL;
1628 ret = insert_pfn(vma, addr, pfn, pgprot);
1630 if (ret)
1631 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1633 return ret;
1635 EXPORT_SYMBOL(vm_insert_pfn);
1637 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1638 unsigned long pfn)
1640 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1642 if (addr < vma->vm_start || addr >= vma->vm_end)
1643 return -EFAULT;
1646 * If we don't have pte special, then we have to use the pfn_valid()
1647 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1648 * refcount the page if pfn_valid is true (hence insert_page rather
1649 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1650 * without pte special, it would there be refcounted as a normal page.
1652 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1653 struct page *page;
1655 page = pfn_to_page(pfn);
1656 return insert_page(vma, addr, page, vma->vm_page_prot);
1658 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1660 EXPORT_SYMBOL(vm_insert_mixed);
1663 * maps a range of physical memory into the requested pages. the old
1664 * mappings are removed. any references to nonexistent pages results
1665 * in null mappings (currently treated as "copy-on-access")
1667 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1668 unsigned long addr, unsigned long end,
1669 unsigned long pfn, pgprot_t prot)
1671 pte_t *pte;
1672 spinlock_t *ptl;
1674 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1675 if (!pte)
1676 return -ENOMEM;
1677 arch_enter_lazy_mmu_mode();
1678 do {
1679 BUG_ON(!pte_none(*pte));
1680 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1681 pfn++;
1682 } while (pte++, addr += PAGE_SIZE, addr != end);
1683 arch_leave_lazy_mmu_mode();
1684 pte_unmap_unlock(pte - 1, ptl);
1685 return 0;
1688 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1689 unsigned long addr, unsigned long end,
1690 unsigned long pfn, pgprot_t prot)
1692 pmd_t *pmd;
1693 unsigned long next;
1695 pfn -= addr >> PAGE_SHIFT;
1696 pmd = pmd_alloc(mm, pud, addr);
1697 if (!pmd)
1698 return -ENOMEM;
1699 do {
1700 next = pmd_addr_end(addr, end);
1701 if (remap_pte_range(mm, pmd, addr, next,
1702 pfn + (addr >> PAGE_SHIFT), prot))
1703 return -ENOMEM;
1704 } while (pmd++, addr = next, addr != end);
1705 return 0;
1708 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1709 unsigned long addr, unsigned long end,
1710 unsigned long pfn, pgprot_t prot)
1712 pud_t *pud;
1713 unsigned long next;
1715 pfn -= addr >> PAGE_SHIFT;
1716 pud = pud_alloc(mm, pgd, addr);
1717 if (!pud)
1718 return -ENOMEM;
1719 do {
1720 next = pud_addr_end(addr, end);
1721 if (remap_pmd_range(mm, pud, addr, next,
1722 pfn + (addr >> PAGE_SHIFT), prot))
1723 return -ENOMEM;
1724 } while (pud++, addr = next, addr != end);
1725 return 0;
1729 * remap_pfn_range - remap kernel memory to userspace
1730 * @vma: user vma to map to
1731 * @addr: target user address to start at
1732 * @pfn: physical address of kernel memory
1733 * @size: size of map area
1734 * @prot: page protection flags for this mapping
1736 * Note: this is only safe if the mm semaphore is held when called.
1738 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1739 unsigned long pfn, unsigned long size, pgprot_t prot)
1741 pgd_t *pgd;
1742 unsigned long next;
1743 unsigned long end = addr + PAGE_ALIGN(size);
1744 struct mm_struct *mm = vma->vm_mm;
1745 int err;
1748 * Physically remapped pages are special. Tell the
1749 * rest of the world about it:
1750 * VM_IO tells people not to look at these pages
1751 * (accesses can have side effects).
1752 * VM_RESERVED is specified all over the place, because
1753 * in 2.4 it kept swapout's vma scan off this vma; but
1754 * in 2.6 the LRU scan won't even find its pages, so this
1755 * flag means no more than count its pages in reserved_vm,
1756 * and omit it from core dump, even when VM_IO turned off.
1757 * VM_PFNMAP tells the core MM that the base pages are just
1758 * raw PFN mappings, and do not have a "struct page" associated
1759 * with them.
1761 * There's a horrible special case to handle copy-on-write
1762 * behaviour that some programs depend on. We mark the "original"
1763 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1765 if (addr == vma->vm_start && end == vma->vm_end) {
1766 vma->vm_pgoff = pfn;
1767 vma->vm_flags |= VM_PFN_AT_MMAP;
1768 } else if (is_cow_mapping(vma->vm_flags))
1769 return -EINVAL;
1771 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1773 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1774 if (err) {
1776 * To indicate that track_pfn related cleanup is not
1777 * needed from higher level routine calling unmap_vmas
1779 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1780 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1781 return -EINVAL;
1784 BUG_ON(addr >= end);
1785 pfn -= addr >> PAGE_SHIFT;
1786 pgd = pgd_offset(mm, addr);
1787 flush_cache_range(vma, addr, end);
1788 do {
1789 next = pgd_addr_end(addr, end);
1790 err = remap_pud_range(mm, pgd, addr, next,
1791 pfn + (addr >> PAGE_SHIFT), prot);
1792 if (err)
1793 break;
1794 } while (pgd++, addr = next, addr != end);
1796 if (err)
1797 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1799 return err;
1801 EXPORT_SYMBOL(remap_pfn_range);
1803 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1804 unsigned long addr, unsigned long end,
1805 pte_fn_t fn, void *data)
1807 pte_t *pte;
1808 int err;
1809 pgtable_t token;
1810 spinlock_t *uninitialized_var(ptl);
1812 pte = (mm == &init_mm) ?
1813 pte_alloc_kernel(pmd, addr) :
1814 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1815 if (!pte)
1816 return -ENOMEM;
1818 BUG_ON(pmd_huge(*pmd));
1820 arch_enter_lazy_mmu_mode();
1822 token = pmd_pgtable(*pmd);
1824 do {
1825 err = fn(pte, token, addr, data);
1826 if (err)
1827 break;
1828 } while (pte++, addr += PAGE_SIZE, addr != end);
1830 arch_leave_lazy_mmu_mode();
1832 if (mm != &init_mm)
1833 pte_unmap_unlock(pte-1, ptl);
1834 return err;
1837 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1838 unsigned long addr, unsigned long end,
1839 pte_fn_t fn, void *data)
1841 pmd_t *pmd;
1842 unsigned long next;
1843 int err;
1845 BUG_ON(pud_huge(*pud));
1847 pmd = pmd_alloc(mm, pud, addr);
1848 if (!pmd)
1849 return -ENOMEM;
1850 do {
1851 next = pmd_addr_end(addr, end);
1852 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1853 if (err)
1854 break;
1855 } while (pmd++, addr = next, addr != end);
1856 return err;
1859 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1860 unsigned long addr, unsigned long end,
1861 pte_fn_t fn, void *data)
1863 pud_t *pud;
1864 unsigned long next;
1865 int err;
1867 pud = pud_alloc(mm, pgd, addr);
1868 if (!pud)
1869 return -ENOMEM;
1870 do {
1871 next = pud_addr_end(addr, end);
1872 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1873 if (err)
1874 break;
1875 } while (pud++, addr = next, addr != end);
1876 return err;
1880 * Scan a region of virtual memory, filling in page tables as necessary
1881 * and calling a provided function on each leaf page table.
1883 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1884 unsigned long size, pte_fn_t fn, void *data)
1886 pgd_t *pgd;
1887 unsigned long next;
1888 unsigned long start = addr, end = addr + size;
1889 int err;
1891 BUG_ON(addr >= end);
1892 mmu_notifier_invalidate_range_start(mm, start, end);
1893 pgd = pgd_offset(mm, addr);
1894 do {
1895 next = pgd_addr_end(addr, end);
1896 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1897 if (err)
1898 break;
1899 } while (pgd++, addr = next, addr != end);
1900 mmu_notifier_invalidate_range_end(mm, start, end);
1901 return err;
1903 EXPORT_SYMBOL_GPL(apply_to_page_range);
1906 * handle_pte_fault chooses page fault handler according to an entry
1907 * which was read non-atomically. Before making any commitment, on
1908 * those architectures or configurations (e.g. i386 with PAE) which
1909 * might give a mix of unmatched parts, do_swap_page and do_file_page
1910 * must check under lock before unmapping the pte and proceeding
1911 * (but do_wp_page is only called after already making such a check;
1912 * and do_anonymous_page and do_no_page can safely check later on).
1914 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1915 pte_t *page_table, pte_t orig_pte)
1917 int same = 1;
1918 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1919 if (sizeof(pte_t) > sizeof(unsigned long)) {
1920 spinlock_t *ptl = pte_lockptr(mm, pmd);
1921 spin_lock(ptl);
1922 same = pte_same(*page_table, orig_pte);
1923 spin_unlock(ptl);
1925 #endif
1926 pte_unmap(page_table);
1927 return same;
1931 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1932 * servicing faults for write access. In the normal case, do always want
1933 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1934 * that do not have writing enabled, when used by access_process_vm.
1936 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1938 if (likely(vma->vm_flags & VM_WRITE))
1939 pte = pte_mkwrite(pte);
1940 return pte;
1943 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1946 * If the source page was a PFN mapping, we don't have
1947 * a "struct page" for it. We do a best-effort copy by
1948 * just copying from the original user address. If that
1949 * fails, we just zero-fill it. Live with it.
1951 if (unlikely(!src)) {
1952 void *kaddr = kmap_atomic(dst, KM_USER0);
1953 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1956 * This really shouldn't fail, because the page is there
1957 * in the page tables. But it might just be unreadable,
1958 * in which case we just give up and fill the result with
1959 * zeroes.
1961 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1962 memset(kaddr, 0, PAGE_SIZE);
1963 kunmap_atomic(kaddr, KM_USER0);
1964 flush_dcache_page(dst);
1965 } else
1966 copy_user_highpage(dst, src, va, vma);
1970 * This routine handles present pages, when users try to write
1971 * to a shared page. It is done by copying the page to a new address
1972 * and decrementing the shared-page counter for the old page.
1974 * Note that this routine assumes that the protection checks have been
1975 * done by the caller (the low-level page fault routine in most cases).
1976 * Thus we can safely just mark it writable once we've done any necessary
1977 * COW.
1979 * We also mark the page dirty at this point even though the page will
1980 * change only once the write actually happens. This avoids a few races,
1981 * and potentially makes it more efficient.
1983 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1984 * but allow concurrent faults), with pte both mapped and locked.
1985 * We return with mmap_sem still held, but pte unmapped and unlocked.
1987 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1988 unsigned long address, pte_t *page_table, pmd_t *pmd,
1989 spinlock_t *ptl, pte_t orig_pte)
1991 struct page *old_page, *new_page;
1992 pte_t entry;
1993 int reuse = 0, ret = 0;
1994 int page_mkwrite = 0;
1995 struct page *dirty_page = NULL;
1997 old_page = vm_normal_page(vma, address, orig_pte);
1998 if (!old_page) {
2000 * VM_MIXEDMAP !pfn_valid() case
2002 * We should not cow pages in a shared writeable mapping.
2003 * Just mark the pages writable as we can't do any dirty
2004 * accounting on raw pfn maps.
2006 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2007 (VM_WRITE|VM_SHARED))
2008 goto reuse;
2009 goto gotten;
2013 * Take out anonymous pages first, anonymous shared vmas are
2014 * not dirty accountable.
2016 if (PageAnon(old_page) && !PageKsm(old_page)) {
2017 if (!trylock_page(old_page)) {
2018 page_cache_get(old_page);
2019 pte_unmap_unlock(page_table, ptl);
2020 lock_page(old_page);
2021 page_table = pte_offset_map_lock(mm, pmd, address,
2022 &ptl);
2023 if (!pte_same(*page_table, orig_pte)) {
2024 unlock_page(old_page);
2025 page_cache_release(old_page);
2026 goto unlock;
2028 page_cache_release(old_page);
2030 reuse = reuse_swap_page(old_page);
2031 unlock_page(old_page);
2032 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2033 (VM_WRITE|VM_SHARED))) {
2035 * Only catch write-faults on shared writable pages,
2036 * read-only shared pages can get COWed by
2037 * get_user_pages(.write=1, .force=1).
2039 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2040 struct vm_fault vmf;
2041 int tmp;
2043 vmf.virtual_address = (void __user *)(address &
2044 PAGE_MASK);
2045 vmf.pgoff = old_page->index;
2046 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2047 vmf.page = old_page;
2050 * Notify the address space that the page is about to
2051 * become writable so that it can prohibit this or wait
2052 * for the page to get into an appropriate state.
2054 * We do this without the lock held, so that it can
2055 * sleep if it needs to.
2057 page_cache_get(old_page);
2058 pte_unmap_unlock(page_table, ptl);
2060 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2061 if (unlikely(tmp &
2062 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2063 ret = tmp;
2064 goto unwritable_page;
2066 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2067 lock_page(old_page);
2068 if (!old_page->mapping) {
2069 ret = 0; /* retry the fault */
2070 unlock_page(old_page);
2071 goto unwritable_page;
2073 } else
2074 VM_BUG_ON(!PageLocked(old_page));
2077 * Since we dropped the lock we need to revalidate
2078 * the PTE as someone else may have changed it. If
2079 * they did, we just return, as we can count on the
2080 * MMU to tell us if they didn't also make it writable.
2082 page_table = pte_offset_map_lock(mm, pmd, address,
2083 &ptl);
2084 if (!pte_same(*page_table, orig_pte)) {
2085 unlock_page(old_page);
2086 page_cache_release(old_page);
2087 goto unlock;
2090 page_mkwrite = 1;
2092 dirty_page = old_page;
2093 get_page(dirty_page);
2094 reuse = 1;
2097 if (reuse) {
2098 reuse:
2099 flush_cache_page(vma, address, pte_pfn(orig_pte));
2100 entry = pte_mkyoung(orig_pte);
2101 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2102 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2103 update_mmu_cache(vma, address, entry);
2104 ret |= VM_FAULT_WRITE;
2105 goto unlock;
2109 * Ok, we need to copy. Oh, well..
2111 page_cache_get(old_page);
2112 gotten:
2113 pte_unmap_unlock(page_table, ptl);
2115 if (unlikely(anon_vma_prepare(vma)))
2116 goto oom;
2118 if (is_zero_pfn(pte_pfn(orig_pte))) {
2119 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2120 if (!new_page)
2121 goto oom;
2122 } else {
2123 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2124 if (!new_page)
2125 goto oom;
2126 cow_user_page(new_page, old_page, address, vma);
2128 __SetPageUptodate(new_page);
2131 * Don't let another task, with possibly unlocked vma,
2132 * keep the mlocked page.
2134 if ((vma->vm_flags & VM_LOCKED) && old_page) {
2135 lock_page(old_page); /* for LRU manipulation */
2136 clear_page_mlock(old_page);
2137 unlock_page(old_page);
2140 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2141 goto oom_free_new;
2144 * Re-check the pte - we dropped the lock
2146 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2147 if (likely(pte_same(*page_table, orig_pte))) {
2148 if (old_page) {
2149 if (!PageAnon(old_page)) {
2150 dec_mm_counter(mm, file_rss);
2151 inc_mm_counter(mm, anon_rss);
2153 } else
2154 inc_mm_counter(mm, anon_rss);
2155 flush_cache_page(vma, address, pte_pfn(orig_pte));
2156 entry = mk_pte(new_page, vma->vm_page_prot);
2157 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2159 * Clear the pte entry and flush it first, before updating the
2160 * pte with the new entry. This will avoid a race condition
2161 * seen in the presence of one thread doing SMC and another
2162 * thread doing COW.
2164 ptep_clear_flush(vma, address, page_table);
2165 page_add_new_anon_rmap(new_page, vma, address);
2167 * We call the notify macro here because, when using secondary
2168 * mmu page tables (such as kvm shadow page tables), we want the
2169 * new page to be mapped directly into the secondary page table.
2171 set_pte_at_notify(mm, address, page_table, entry);
2172 update_mmu_cache(vma, address, entry);
2173 if (old_page) {
2175 * Only after switching the pte to the new page may
2176 * we remove the mapcount here. Otherwise another
2177 * process may come and find the rmap count decremented
2178 * before the pte is switched to the new page, and
2179 * "reuse" the old page writing into it while our pte
2180 * here still points into it and can be read by other
2181 * threads.
2183 * The critical issue is to order this
2184 * page_remove_rmap with the ptp_clear_flush above.
2185 * Those stores are ordered by (if nothing else,)
2186 * the barrier present in the atomic_add_negative
2187 * in page_remove_rmap.
2189 * Then the TLB flush in ptep_clear_flush ensures that
2190 * no process can access the old page before the
2191 * decremented mapcount is visible. And the old page
2192 * cannot be reused until after the decremented
2193 * mapcount is visible. So transitively, TLBs to
2194 * old page will be flushed before it can be reused.
2196 page_remove_rmap(old_page);
2199 /* Free the old page.. */
2200 new_page = old_page;
2201 ret |= VM_FAULT_WRITE;
2202 } else
2203 mem_cgroup_uncharge_page(new_page);
2205 if (new_page)
2206 page_cache_release(new_page);
2207 if (old_page)
2208 page_cache_release(old_page);
2209 unlock:
2210 pte_unmap_unlock(page_table, ptl);
2211 if (dirty_page) {
2213 * Yes, Virginia, this is actually required to prevent a race
2214 * with clear_page_dirty_for_io() from clearing the page dirty
2215 * bit after it clear all dirty ptes, but before a racing
2216 * do_wp_page installs a dirty pte.
2218 * do_no_page is protected similarly.
2220 if (!page_mkwrite) {
2221 wait_on_page_locked(dirty_page);
2222 set_page_dirty_balance(dirty_page, page_mkwrite);
2224 put_page(dirty_page);
2225 if (page_mkwrite) {
2226 struct address_space *mapping = dirty_page->mapping;
2228 set_page_dirty(dirty_page);
2229 unlock_page(dirty_page);
2230 page_cache_release(dirty_page);
2231 if (mapping) {
2233 * Some device drivers do not set page.mapping
2234 * but still dirty their pages
2236 balance_dirty_pages_ratelimited(mapping);
2240 /* file_update_time outside page_lock */
2241 if (vma->vm_file)
2242 file_update_time(vma->vm_file);
2244 return ret;
2245 oom_free_new:
2246 page_cache_release(new_page);
2247 oom:
2248 if (old_page) {
2249 if (page_mkwrite) {
2250 unlock_page(old_page);
2251 page_cache_release(old_page);
2253 page_cache_release(old_page);
2255 return VM_FAULT_OOM;
2257 unwritable_page:
2258 page_cache_release(old_page);
2259 return ret;
2263 * Helper functions for unmap_mapping_range().
2265 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2267 * We have to restart searching the prio_tree whenever we drop the lock,
2268 * since the iterator is only valid while the lock is held, and anyway
2269 * a later vma might be split and reinserted earlier while lock dropped.
2271 * The list of nonlinear vmas could be handled more efficiently, using
2272 * a placeholder, but handle it in the same way until a need is shown.
2273 * It is important to search the prio_tree before nonlinear list: a vma
2274 * may become nonlinear and be shifted from prio_tree to nonlinear list
2275 * while the lock is dropped; but never shifted from list to prio_tree.
2277 * In order to make forward progress despite restarting the search,
2278 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2279 * quickly skip it next time around. Since the prio_tree search only
2280 * shows us those vmas affected by unmapping the range in question, we
2281 * can't efficiently keep all vmas in step with mapping->truncate_count:
2282 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2283 * mapping->truncate_count and vma->vm_truncate_count are protected by
2284 * i_mmap_lock.
2286 * In order to make forward progress despite repeatedly restarting some
2287 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2288 * and restart from that address when we reach that vma again. It might
2289 * have been split or merged, shrunk or extended, but never shifted: so
2290 * restart_addr remains valid so long as it remains in the vma's range.
2291 * unmap_mapping_range forces truncate_count to leap over page-aligned
2292 * values so we can save vma's restart_addr in its truncate_count field.
2294 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2296 static void reset_vma_truncate_counts(struct address_space *mapping)
2298 struct vm_area_struct *vma;
2299 struct prio_tree_iter iter;
2301 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2302 vma->vm_truncate_count = 0;
2303 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2304 vma->vm_truncate_count = 0;
2307 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2308 unsigned long start_addr, unsigned long end_addr,
2309 struct zap_details *details)
2311 unsigned long restart_addr;
2312 int need_break;
2315 * files that support invalidating or truncating portions of the
2316 * file from under mmaped areas must have their ->fault function
2317 * return a locked page (and set VM_FAULT_LOCKED in the return).
2318 * This provides synchronisation against concurrent unmapping here.
2321 again:
2322 restart_addr = vma->vm_truncate_count;
2323 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2324 start_addr = restart_addr;
2325 if (start_addr >= end_addr) {
2326 /* Top of vma has been split off since last time */
2327 vma->vm_truncate_count = details->truncate_count;
2328 return 0;
2332 restart_addr = zap_page_range(vma, start_addr,
2333 end_addr - start_addr, details);
2334 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2336 if (restart_addr >= end_addr) {
2337 /* We have now completed this vma: mark it so */
2338 vma->vm_truncate_count = details->truncate_count;
2339 if (!need_break)
2340 return 0;
2341 } else {
2342 /* Note restart_addr in vma's truncate_count field */
2343 vma->vm_truncate_count = restart_addr;
2344 if (!need_break)
2345 goto again;
2348 spin_unlock(details->i_mmap_lock);
2349 cond_resched();
2350 spin_lock(details->i_mmap_lock);
2351 return -EINTR;
2354 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2355 struct zap_details *details)
2357 struct vm_area_struct *vma;
2358 struct prio_tree_iter iter;
2359 pgoff_t vba, vea, zba, zea;
2361 restart:
2362 vma_prio_tree_foreach(vma, &iter, root,
2363 details->first_index, details->last_index) {
2364 /* Skip quickly over those we have already dealt with */
2365 if (vma->vm_truncate_count == details->truncate_count)
2366 continue;
2368 vba = vma->vm_pgoff;
2369 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2370 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2371 zba = details->first_index;
2372 if (zba < vba)
2373 zba = vba;
2374 zea = details->last_index;
2375 if (zea > vea)
2376 zea = vea;
2378 if (unmap_mapping_range_vma(vma,
2379 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2380 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2381 details) < 0)
2382 goto restart;
2386 static inline void unmap_mapping_range_list(struct list_head *head,
2387 struct zap_details *details)
2389 struct vm_area_struct *vma;
2392 * In nonlinear VMAs there is no correspondence between virtual address
2393 * offset and file offset. So we must perform an exhaustive search
2394 * across *all* the pages in each nonlinear VMA, not just the pages
2395 * whose virtual address lies outside the file truncation point.
2397 restart:
2398 list_for_each_entry(vma, head, shared.vm_set.list) {
2399 /* Skip quickly over those we have already dealt with */
2400 if (vma->vm_truncate_count == details->truncate_count)
2401 continue;
2402 details->nonlinear_vma = vma;
2403 if (unmap_mapping_range_vma(vma, vma->vm_start,
2404 vma->vm_end, details) < 0)
2405 goto restart;
2410 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2411 * @mapping: the address space containing mmaps to be unmapped.
2412 * @holebegin: byte in first page to unmap, relative to the start of
2413 * the underlying file. This will be rounded down to a PAGE_SIZE
2414 * boundary. Note that this is different from truncate_pagecache(), which
2415 * must keep the partial page. In contrast, we must get rid of
2416 * partial pages.
2417 * @holelen: size of prospective hole in bytes. This will be rounded
2418 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2419 * end of the file.
2420 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2421 * but 0 when invalidating pagecache, don't throw away private data.
2423 void unmap_mapping_range(struct address_space *mapping,
2424 loff_t const holebegin, loff_t const holelen, int even_cows)
2426 struct zap_details details;
2427 pgoff_t hba = holebegin >> PAGE_SHIFT;
2428 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2430 /* Check for overflow. */
2431 if (sizeof(holelen) > sizeof(hlen)) {
2432 long long holeend =
2433 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2434 if (holeend & ~(long long)ULONG_MAX)
2435 hlen = ULONG_MAX - hba + 1;
2438 details.check_mapping = even_cows? NULL: mapping;
2439 details.nonlinear_vma = NULL;
2440 details.first_index = hba;
2441 details.last_index = hba + hlen - 1;
2442 if (details.last_index < details.first_index)
2443 details.last_index = ULONG_MAX;
2444 details.i_mmap_lock = &mapping->i_mmap_lock;
2446 spin_lock(&mapping->i_mmap_lock);
2448 /* Protect against endless unmapping loops */
2449 mapping->truncate_count++;
2450 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2451 if (mapping->truncate_count == 0)
2452 reset_vma_truncate_counts(mapping);
2453 mapping->truncate_count++;
2455 details.truncate_count = mapping->truncate_count;
2457 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2458 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2459 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2460 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2461 spin_unlock(&mapping->i_mmap_lock);
2463 EXPORT_SYMBOL(unmap_mapping_range);
2465 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2467 struct address_space *mapping = inode->i_mapping;
2470 * If the underlying filesystem is not going to provide
2471 * a way to truncate a range of blocks (punch a hole) -
2472 * we should return failure right now.
2474 if (!inode->i_op->truncate_range)
2475 return -ENOSYS;
2477 mutex_lock(&inode->i_mutex);
2478 down_write(&inode->i_alloc_sem);
2479 unmap_mapping_range(mapping, offset, (end - offset), 1);
2480 truncate_inode_pages_range(mapping, offset, end);
2481 unmap_mapping_range(mapping, offset, (end - offset), 1);
2482 inode->i_op->truncate_range(inode, offset, end);
2483 up_write(&inode->i_alloc_sem);
2484 mutex_unlock(&inode->i_mutex);
2486 return 0;
2490 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2491 * but allow concurrent faults), and pte mapped but not yet locked.
2492 * We return with mmap_sem still held, but pte unmapped and unlocked.
2494 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2495 unsigned long address, pte_t *page_table, pmd_t *pmd,
2496 unsigned int flags, pte_t orig_pte)
2498 spinlock_t *ptl;
2499 struct page *page;
2500 swp_entry_t entry;
2501 pte_t pte;
2502 struct mem_cgroup *ptr = NULL;
2503 int ret = 0;
2505 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2506 goto out;
2508 entry = pte_to_swp_entry(orig_pte);
2509 if (unlikely(non_swap_entry(entry))) {
2510 if (is_migration_entry(entry)) {
2511 migration_entry_wait(mm, pmd, address);
2512 } else if (is_hwpoison_entry(entry)) {
2513 ret = VM_FAULT_HWPOISON;
2514 } else {
2515 print_bad_pte(vma, address, orig_pte, NULL);
2516 ret = VM_FAULT_OOM;
2518 goto out;
2520 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2521 page = lookup_swap_cache(entry);
2522 if (!page) {
2523 grab_swap_token(mm); /* Contend for token _before_ read-in */
2524 page = swapin_readahead(entry,
2525 GFP_HIGHUSER_MOVABLE, vma, address);
2526 if (!page) {
2528 * Back out if somebody else faulted in this pte
2529 * while we released the pte lock.
2531 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2532 if (likely(pte_same(*page_table, orig_pte)))
2533 ret = VM_FAULT_OOM;
2534 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2535 goto unlock;
2538 /* Had to read the page from swap area: Major fault */
2539 ret = VM_FAULT_MAJOR;
2540 count_vm_event(PGMAJFAULT);
2541 } else if (PageHWPoison(page)) {
2542 ret = VM_FAULT_HWPOISON;
2543 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2544 goto out;
2547 lock_page(page);
2548 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2550 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2551 ret = VM_FAULT_OOM;
2552 goto out_page;
2556 * Back out if somebody else already faulted in this pte.
2558 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2559 if (unlikely(!pte_same(*page_table, orig_pte)))
2560 goto out_nomap;
2562 if (unlikely(!PageUptodate(page))) {
2563 ret = VM_FAULT_SIGBUS;
2564 goto out_nomap;
2568 * The page isn't present yet, go ahead with the fault.
2570 * Be careful about the sequence of operations here.
2571 * To get its accounting right, reuse_swap_page() must be called
2572 * while the page is counted on swap but not yet in mapcount i.e.
2573 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2574 * must be called after the swap_free(), or it will never succeed.
2575 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2576 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2577 * in page->private. In this case, a record in swap_cgroup is silently
2578 * discarded at swap_free().
2581 inc_mm_counter(mm, anon_rss);
2582 pte = mk_pte(page, vma->vm_page_prot);
2583 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2584 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2585 flags &= ~FAULT_FLAG_WRITE;
2587 flush_icache_page(vma, page);
2588 set_pte_at(mm, address, page_table, pte);
2589 page_add_anon_rmap(page, vma, address);
2590 /* It's better to call commit-charge after rmap is established */
2591 mem_cgroup_commit_charge_swapin(page, ptr);
2593 swap_free(entry);
2594 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2595 try_to_free_swap(page);
2596 unlock_page(page);
2598 if (flags & FAULT_FLAG_WRITE) {
2599 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2600 if (ret & VM_FAULT_ERROR)
2601 ret &= VM_FAULT_ERROR;
2602 goto out;
2605 /* No need to invalidate - it was non-present before */
2606 update_mmu_cache(vma, address, pte);
2607 unlock:
2608 pte_unmap_unlock(page_table, ptl);
2609 out:
2610 return ret;
2611 out_nomap:
2612 mem_cgroup_cancel_charge_swapin(ptr);
2613 pte_unmap_unlock(page_table, ptl);
2614 out_page:
2615 unlock_page(page);
2616 page_cache_release(page);
2617 return ret;
2621 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2622 * but allow concurrent faults), and pte mapped but not yet locked.
2623 * We return with mmap_sem still held, but pte unmapped and unlocked.
2625 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2626 unsigned long address, pte_t *page_table, pmd_t *pmd,
2627 unsigned int flags)
2629 struct page *page;
2630 spinlock_t *ptl;
2631 pte_t entry;
2633 if (!(flags & FAULT_FLAG_WRITE)) {
2634 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2635 vma->vm_page_prot));
2636 ptl = pte_lockptr(mm, pmd);
2637 spin_lock(ptl);
2638 if (!pte_none(*page_table))
2639 goto unlock;
2640 goto setpte;
2643 /* Allocate our own private page. */
2644 pte_unmap(page_table);
2646 if (unlikely(anon_vma_prepare(vma)))
2647 goto oom;
2648 page = alloc_zeroed_user_highpage_movable(vma, address);
2649 if (!page)
2650 goto oom;
2651 __SetPageUptodate(page);
2653 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2654 goto oom_free_page;
2656 entry = mk_pte(page, vma->vm_page_prot);
2657 if (vma->vm_flags & VM_WRITE)
2658 entry = pte_mkwrite(pte_mkdirty(entry));
2660 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2661 if (!pte_none(*page_table))
2662 goto release;
2664 inc_mm_counter(mm, anon_rss);
2665 page_add_new_anon_rmap(page, vma, address);
2666 setpte:
2667 set_pte_at(mm, address, page_table, entry);
2669 /* No need to invalidate - it was non-present before */
2670 update_mmu_cache(vma, address, entry);
2671 unlock:
2672 pte_unmap_unlock(page_table, ptl);
2673 return 0;
2674 release:
2675 mem_cgroup_uncharge_page(page);
2676 page_cache_release(page);
2677 goto unlock;
2678 oom_free_page:
2679 page_cache_release(page);
2680 oom:
2681 return VM_FAULT_OOM;
2685 * __do_fault() tries to create a new page mapping. It aggressively
2686 * tries to share with existing pages, but makes a separate copy if
2687 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2688 * the next page fault.
2690 * As this is called only for pages that do not currently exist, we
2691 * do not need to flush old virtual caches or the TLB.
2693 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2694 * but allow concurrent faults), and pte neither mapped nor locked.
2695 * We return with mmap_sem still held, but pte unmapped and unlocked.
2697 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2698 unsigned long address, pmd_t *pmd,
2699 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2701 pte_t *page_table;
2702 spinlock_t *ptl;
2703 struct page *page;
2704 pte_t entry;
2705 int anon = 0;
2706 int charged = 0;
2707 struct page *dirty_page = NULL;
2708 struct vm_fault vmf;
2709 int ret;
2710 int page_mkwrite = 0;
2712 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2713 vmf.pgoff = pgoff;
2714 vmf.flags = flags;
2715 vmf.page = NULL;
2717 ret = vma->vm_ops->fault(vma, &vmf);
2718 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2719 return ret;
2721 if (unlikely(PageHWPoison(vmf.page))) {
2722 if (ret & VM_FAULT_LOCKED)
2723 unlock_page(vmf.page);
2724 return VM_FAULT_HWPOISON;
2728 * For consistency in subsequent calls, make the faulted page always
2729 * locked.
2731 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2732 lock_page(vmf.page);
2733 else
2734 VM_BUG_ON(!PageLocked(vmf.page));
2737 * Should we do an early C-O-W break?
2739 page = vmf.page;
2740 if (flags & FAULT_FLAG_WRITE) {
2741 if (!(vma->vm_flags & VM_SHARED)) {
2742 anon = 1;
2743 if (unlikely(anon_vma_prepare(vma))) {
2744 ret = VM_FAULT_OOM;
2745 goto out;
2747 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2748 vma, address);
2749 if (!page) {
2750 ret = VM_FAULT_OOM;
2751 goto out;
2753 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2754 ret = VM_FAULT_OOM;
2755 page_cache_release(page);
2756 goto out;
2758 charged = 1;
2760 * Don't let another task, with possibly unlocked vma,
2761 * keep the mlocked page.
2763 if (vma->vm_flags & VM_LOCKED)
2764 clear_page_mlock(vmf.page);
2765 copy_user_highpage(page, vmf.page, address, vma);
2766 __SetPageUptodate(page);
2767 } else {
2769 * If the page will be shareable, see if the backing
2770 * address space wants to know that the page is about
2771 * to become writable
2773 if (vma->vm_ops->page_mkwrite) {
2774 int tmp;
2776 unlock_page(page);
2777 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2778 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2779 if (unlikely(tmp &
2780 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2781 ret = tmp;
2782 goto unwritable_page;
2784 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2785 lock_page(page);
2786 if (!page->mapping) {
2787 ret = 0; /* retry the fault */
2788 unlock_page(page);
2789 goto unwritable_page;
2791 } else
2792 VM_BUG_ON(!PageLocked(page));
2793 page_mkwrite = 1;
2799 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2802 * This silly early PAGE_DIRTY setting removes a race
2803 * due to the bad i386 page protection. But it's valid
2804 * for other architectures too.
2806 * Note that if FAULT_FLAG_WRITE is set, we either now have
2807 * an exclusive copy of the page, or this is a shared mapping,
2808 * so we can make it writable and dirty to avoid having to
2809 * handle that later.
2811 /* Only go through if we didn't race with anybody else... */
2812 if (likely(pte_same(*page_table, orig_pte))) {
2813 flush_icache_page(vma, page);
2814 entry = mk_pte(page, vma->vm_page_prot);
2815 if (flags & FAULT_FLAG_WRITE)
2816 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2817 if (anon) {
2818 inc_mm_counter(mm, anon_rss);
2819 page_add_new_anon_rmap(page, vma, address);
2820 } else {
2821 inc_mm_counter(mm, file_rss);
2822 page_add_file_rmap(page);
2823 if (flags & FAULT_FLAG_WRITE) {
2824 dirty_page = page;
2825 get_page(dirty_page);
2828 set_pte_at(mm, address, page_table, entry);
2830 /* no need to invalidate: a not-present page won't be cached */
2831 update_mmu_cache(vma, address, entry);
2832 } else {
2833 if (charged)
2834 mem_cgroup_uncharge_page(page);
2835 if (anon)
2836 page_cache_release(page);
2837 else
2838 anon = 1; /* no anon but release faulted_page */
2841 pte_unmap_unlock(page_table, ptl);
2843 out:
2844 if (dirty_page) {
2845 struct address_space *mapping = page->mapping;
2847 if (set_page_dirty(dirty_page))
2848 page_mkwrite = 1;
2849 unlock_page(dirty_page);
2850 put_page(dirty_page);
2851 if (page_mkwrite && mapping) {
2853 * Some device drivers do not set page.mapping but still
2854 * dirty their pages
2856 balance_dirty_pages_ratelimited(mapping);
2859 /* file_update_time outside page_lock */
2860 if (vma->vm_file)
2861 file_update_time(vma->vm_file);
2862 } else {
2863 unlock_page(vmf.page);
2864 if (anon)
2865 page_cache_release(vmf.page);
2868 return ret;
2870 unwritable_page:
2871 page_cache_release(page);
2872 return ret;
2875 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2876 unsigned long address, pte_t *page_table, pmd_t *pmd,
2877 unsigned int flags, pte_t orig_pte)
2879 pgoff_t pgoff = (((address & PAGE_MASK)
2880 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2882 pte_unmap(page_table);
2883 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2887 * Fault of a previously existing named mapping. Repopulate the pte
2888 * from the encoded file_pte if possible. This enables swappable
2889 * nonlinear vmas.
2891 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2892 * but allow concurrent faults), and pte mapped but not yet locked.
2893 * We return with mmap_sem still held, but pte unmapped and unlocked.
2895 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2896 unsigned long address, pte_t *page_table, pmd_t *pmd,
2897 unsigned int flags, pte_t orig_pte)
2899 pgoff_t pgoff;
2901 flags |= FAULT_FLAG_NONLINEAR;
2903 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2904 return 0;
2906 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2908 * Page table corrupted: show pte and kill process.
2910 print_bad_pte(vma, address, orig_pte, NULL);
2911 return VM_FAULT_OOM;
2914 pgoff = pte_to_pgoff(orig_pte);
2915 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2919 * These routines also need to handle stuff like marking pages dirty
2920 * and/or accessed for architectures that don't do it in hardware (most
2921 * RISC architectures). The early dirtying is also good on the i386.
2923 * There is also a hook called "update_mmu_cache()" that architectures
2924 * with external mmu caches can use to update those (ie the Sparc or
2925 * PowerPC hashed page tables that act as extended TLBs).
2927 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2928 * but allow concurrent faults), and pte mapped but not yet locked.
2929 * We return with mmap_sem still held, but pte unmapped and unlocked.
2931 static inline int handle_pte_fault(struct mm_struct *mm,
2932 struct vm_area_struct *vma, unsigned long address,
2933 pte_t *pte, pmd_t *pmd, unsigned int flags)
2935 pte_t entry;
2936 spinlock_t *ptl;
2938 entry = *pte;
2939 if (!pte_present(entry)) {
2940 if (pte_none(entry)) {
2941 if (vma->vm_ops) {
2942 if (likely(vma->vm_ops->fault))
2943 return do_linear_fault(mm, vma, address,
2944 pte, pmd, flags, entry);
2946 return do_anonymous_page(mm, vma, address,
2947 pte, pmd, flags);
2949 if (pte_file(entry))
2950 return do_nonlinear_fault(mm, vma, address,
2951 pte, pmd, flags, entry);
2952 return do_swap_page(mm, vma, address,
2953 pte, pmd, flags, entry);
2956 ptl = pte_lockptr(mm, pmd);
2957 spin_lock(ptl);
2958 if (unlikely(!pte_same(*pte, entry)))
2959 goto unlock;
2960 if (flags & FAULT_FLAG_WRITE) {
2961 if (!pte_write(entry))
2962 return do_wp_page(mm, vma, address,
2963 pte, pmd, ptl, entry);
2964 entry = pte_mkdirty(entry);
2966 entry = pte_mkyoung(entry);
2967 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
2968 update_mmu_cache(vma, address, entry);
2969 } else {
2971 * This is needed only for protection faults but the arch code
2972 * is not yet telling us if this is a protection fault or not.
2973 * This still avoids useless tlb flushes for .text page faults
2974 * with threads.
2976 if (flags & FAULT_FLAG_WRITE)
2977 flush_tlb_page(vma, address);
2979 unlock:
2980 pte_unmap_unlock(pte, ptl);
2981 return 0;
2985 * By the time we get here, we already hold the mm semaphore
2987 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2988 unsigned long address, unsigned int flags)
2990 pgd_t *pgd;
2991 pud_t *pud;
2992 pmd_t *pmd;
2993 pte_t *pte;
2995 __set_current_state(TASK_RUNNING);
2997 count_vm_event(PGFAULT);
2999 if (unlikely(is_vm_hugetlb_page(vma)))
3000 return hugetlb_fault(mm, vma, address, flags);
3002 pgd = pgd_offset(mm, address);
3003 pud = pud_alloc(mm, pgd, address);
3004 if (!pud)
3005 return VM_FAULT_OOM;
3006 pmd = pmd_alloc(mm, pud, address);
3007 if (!pmd)
3008 return VM_FAULT_OOM;
3009 pte = pte_alloc_map(mm, pmd, address);
3010 if (!pte)
3011 return VM_FAULT_OOM;
3013 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3016 #ifndef __PAGETABLE_PUD_FOLDED
3018 * Allocate page upper directory.
3019 * We've already handled the fast-path in-line.
3021 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3023 pud_t *new = pud_alloc_one(mm, address);
3024 if (!new)
3025 return -ENOMEM;
3027 smp_wmb(); /* See comment in __pte_alloc */
3029 spin_lock(&mm->page_table_lock);
3030 if (pgd_present(*pgd)) /* Another has populated it */
3031 pud_free(mm, new);
3032 else
3033 pgd_populate(mm, pgd, new);
3034 spin_unlock(&mm->page_table_lock);
3035 return 0;
3037 #endif /* __PAGETABLE_PUD_FOLDED */
3039 #ifndef __PAGETABLE_PMD_FOLDED
3041 * Allocate page middle directory.
3042 * We've already handled the fast-path in-line.
3044 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3046 pmd_t *new = pmd_alloc_one(mm, address);
3047 if (!new)
3048 return -ENOMEM;
3050 smp_wmb(); /* See comment in __pte_alloc */
3052 spin_lock(&mm->page_table_lock);
3053 #ifndef __ARCH_HAS_4LEVEL_HACK
3054 if (pud_present(*pud)) /* Another has populated it */
3055 pmd_free(mm, new);
3056 else
3057 pud_populate(mm, pud, new);
3058 #else
3059 if (pgd_present(*pud)) /* Another has populated it */
3060 pmd_free(mm, new);
3061 else
3062 pgd_populate(mm, pud, new);
3063 #endif /* __ARCH_HAS_4LEVEL_HACK */
3064 spin_unlock(&mm->page_table_lock);
3065 return 0;
3067 #endif /* __PAGETABLE_PMD_FOLDED */
3069 int make_pages_present(unsigned long addr, unsigned long end)
3071 int ret, len, write;
3072 struct vm_area_struct * vma;
3074 vma = find_vma(current->mm, addr);
3075 if (!vma)
3076 return -ENOMEM;
3077 write = (vma->vm_flags & VM_WRITE) != 0;
3078 BUG_ON(addr >= end);
3079 BUG_ON(end > vma->vm_end);
3080 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3081 ret = get_user_pages(current, current->mm, addr,
3082 len, write, 0, NULL, NULL);
3083 if (ret < 0)
3084 return ret;
3085 return ret == len ? 0 : -EFAULT;
3088 #if !defined(__HAVE_ARCH_GATE_AREA)
3090 #if defined(AT_SYSINFO_EHDR)
3091 static struct vm_area_struct gate_vma;
3093 static int __init gate_vma_init(void)
3095 gate_vma.vm_mm = NULL;
3096 gate_vma.vm_start = FIXADDR_USER_START;
3097 gate_vma.vm_end = FIXADDR_USER_END;
3098 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3099 gate_vma.vm_page_prot = __P101;
3101 * Make sure the vDSO gets into every core dump.
3102 * Dumping its contents makes post-mortem fully interpretable later
3103 * without matching up the same kernel and hardware config to see
3104 * what PC values meant.
3106 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3107 return 0;
3109 __initcall(gate_vma_init);
3110 #endif
3112 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3114 #ifdef AT_SYSINFO_EHDR
3115 return &gate_vma;
3116 #else
3117 return NULL;
3118 #endif
3121 int in_gate_area_no_task(unsigned long addr)
3123 #ifdef AT_SYSINFO_EHDR
3124 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3125 return 1;
3126 #endif
3127 return 0;
3130 #endif /* __HAVE_ARCH_GATE_AREA */
3132 static int follow_pte(struct mm_struct *mm, unsigned long address,
3133 pte_t **ptepp, spinlock_t **ptlp)
3135 pgd_t *pgd;
3136 pud_t *pud;
3137 pmd_t *pmd;
3138 pte_t *ptep;
3140 pgd = pgd_offset(mm, address);
3141 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3142 goto out;
3144 pud = pud_offset(pgd, address);
3145 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3146 goto out;
3148 pmd = pmd_offset(pud, address);
3149 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3150 goto out;
3152 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3153 if (pmd_huge(*pmd))
3154 goto out;
3156 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3157 if (!ptep)
3158 goto out;
3159 if (!pte_present(*ptep))
3160 goto unlock;
3161 *ptepp = ptep;
3162 return 0;
3163 unlock:
3164 pte_unmap_unlock(ptep, *ptlp);
3165 out:
3166 return -EINVAL;
3170 * follow_pfn - look up PFN at a user virtual address
3171 * @vma: memory mapping
3172 * @address: user virtual address
3173 * @pfn: location to store found PFN
3175 * Only IO mappings and raw PFN mappings are allowed.
3177 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3179 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3180 unsigned long *pfn)
3182 int ret = -EINVAL;
3183 spinlock_t *ptl;
3184 pte_t *ptep;
3186 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3187 return ret;
3189 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3190 if (ret)
3191 return ret;
3192 *pfn = pte_pfn(*ptep);
3193 pte_unmap_unlock(ptep, ptl);
3194 return 0;
3196 EXPORT_SYMBOL(follow_pfn);
3198 #ifdef CONFIG_HAVE_IOREMAP_PROT
3199 int follow_phys(struct vm_area_struct *vma,
3200 unsigned long address, unsigned int flags,
3201 unsigned long *prot, resource_size_t *phys)
3203 int ret = -EINVAL;
3204 pte_t *ptep, pte;
3205 spinlock_t *ptl;
3207 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3208 goto out;
3210 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3211 goto out;
3212 pte = *ptep;
3214 if ((flags & FOLL_WRITE) && !pte_write(pte))
3215 goto unlock;
3217 *prot = pgprot_val(pte_pgprot(pte));
3218 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3220 ret = 0;
3221 unlock:
3222 pte_unmap_unlock(ptep, ptl);
3223 out:
3224 return ret;
3227 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3228 void *buf, int len, int write)
3230 resource_size_t phys_addr;
3231 unsigned long prot = 0;
3232 void __iomem *maddr;
3233 int offset = addr & (PAGE_SIZE-1);
3235 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3236 return -EINVAL;
3238 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3239 if (write)
3240 memcpy_toio(maddr + offset, buf, len);
3241 else
3242 memcpy_fromio(buf, maddr + offset, len);
3243 iounmap(maddr);
3245 return len;
3247 #endif
3250 * Access another process' address space.
3251 * Source/target buffer must be kernel space,
3252 * Do not walk the page table directly, use get_user_pages
3254 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3256 struct mm_struct *mm;
3257 struct vm_area_struct *vma;
3258 void *old_buf = buf;
3260 mm = get_task_mm(tsk);
3261 if (!mm)
3262 return 0;
3264 down_read(&mm->mmap_sem);
3265 /* ignore errors, just check how much was successfully transferred */
3266 while (len) {
3267 int bytes, ret, offset;
3268 void *maddr;
3269 struct page *page = NULL;
3271 ret = get_user_pages(tsk, mm, addr, 1,
3272 write, 1, &page, &vma);
3273 if (ret <= 0) {
3275 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3276 * we can access using slightly different code.
3278 #ifdef CONFIG_HAVE_IOREMAP_PROT
3279 vma = find_vma(mm, addr);
3280 if (!vma)
3281 break;
3282 if (vma->vm_ops && vma->vm_ops->access)
3283 ret = vma->vm_ops->access(vma, addr, buf,
3284 len, write);
3285 if (ret <= 0)
3286 #endif
3287 break;
3288 bytes = ret;
3289 } else {
3290 bytes = len;
3291 offset = addr & (PAGE_SIZE-1);
3292 if (bytes > PAGE_SIZE-offset)
3293 bytes = PAGE_SIZE-offset;
3295 maddr = kmap(page);
3296 if (write) {
3297 copy_to_user_page(vma, page, addr,
3298 maddr + offset, buf, bytes);
3299 set_page_dirty_lock(page);
3300 } else {
3301 copy_from_user_page(vma, page, addr,
3302 buf, maddr + offset, bytes);
3304 kunmap(page);
3305 page_cache_release(page);
3307 len -= bytes;
3308 buf += bytes;
3309 addr += bytes;
3311 up_read(&mm->mmap_sem);
3312 mmput(mm);
3314 return buf - old_buf;
3318 * Print the name of a VMA.
3320 void print_vma_addr(char *prefix, unsigned long ip)
3322 struct mm_struct *mm = current->mm;
3323 struct vm_area_struct *vma;
3326 * Do not print if we are in atomic
3327 * contexts (in exception stacks, etc.):
3329 if (preempt_count())
3330 return;
3332 down_read(&mm->mmap_sem);
3333 vma = find_vma(mm, ip);
3334 if (vma && vma->vm_file) {
3335 struct file *f = vma->vm_file;
3336 char *buf = (char *)__get_free_page(GFP_KERNEL);
3337 if (buf) {
3338 char *p, *s;
3340 p = d_path(&f->f_path, buf, PAGE_SIZE);
3341 if (IS_ERR(p))
3342 p = "?";
3343 s = strrchr(p, '/');
3344 if (s)
3345 p = s+1;
3346 printk("%s%s[%lx+%lx]", prefix, p,
3347 vma->vm_start,
3348 vma->vm_end - vma->vm_start);
3349 free_page((unsigned long)buf);
3352 up_read(&current->mm->mmap_sem);
3355 #ifdef CONFIG_PROVE_LOCKING
3356 void might_fault(void)
3359 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3360 * holding the mmap_sem, this is safe because kernel memory doesn't
3361 * get paged out, therefore we'll never actually fault, and the
3362 * below annotations will generate false positives.
3364 if (segment_eq(get_fs(), KERNEL_DS))
3365 return;
3367 might_sleep();
3369 * it would be nicer only to annotate paths which are not under
3370 * pagefault_disable, however that requires a larger audit and
3371 * providing helpers like get_user_atomic.
3373 if (!in_atomic() && current->mm)
3374 might_lock_read(&current->mm->mmap_sem);
3376 EXPORT_SYMBOL(might_fault);
3377 #endif