thinkpad-acpi: fix initialization error paths
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / memory.c
blob19e0ae9beecb71062f4a8030f6ab8df14bfa9d51
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/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
55 #include <asm/pgalloc.h>
56 #include <asm/uaccess.h>
57 #include <asm/tlb.h>
58 #include <asm/tlbflush.h>
59 #include <asm/pgtable.h>
61 #include <linux/swapops.h>
62 #include <linux/elf.h>
64 #ifndef CONFIG_NEED_MULTIPLE_NODES
65 /* use the per-pgdat data instead for discontigmem - mbligh */
66 unsigned long max_mapnr;
67 struct page *mem_map;
69 EXPORT_SYMBOL(max_mapnr);
70 EXPORT_SYMBOL(mem_map);
71 #endif
73 unsigned long num_physpages;
75 * A number of key systems in x86 including ioremap() rely on the assumption
76 * that high_memory defines the upper bound on direct map memory, then end
77 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
78 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79 * and ZONE_HIGHMEM.
81 void * high_memory;
83 EXPORT_SYMBOL(num_physpages);
84 EXPORT_SYMBOL(high_memory);
87 * Randomize the address space (stacks, mmaps, brk, etc.).
89 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
90 * as ancient (libc5 based) binaries can segfault. )
92 int randomize_va_space __read_mostly =
93 #ifdef CONFIG_COMPAT_BRK
95 #else
97 #endif
99 static int __init disable_randmaps(char *s)
101 randomize_va_space = 0;
102 return 1;
104 __setup("norandmaps", disable_randmaps);
108 * If a p?d_bad entry is found while walking page tables, report
109 * the error, before resetting entry to p?d_none. Usually (but
110 * very seldom) called out from the p?d_none_or_clear_bad macros.
113 void pgd_clear_bad(pgd_t *pgd)
115 pgd_ERROR(*pgd);
116 pgd_clear(pgd);
119 void pud_clear_bad(pud_t *pud)
121 pud_ERROR(*pud);
122 pud_clear(pud);
125 void pmd_clear_bad(pmd_t *pmd)
127 pmd_ERROR(*pmd);
128 pmd_clear(pmd);
132 * Note: this doesn't free the actual pages themselves. That
133 * has been handled earlier when unmapping all the memory regions.
135 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
137 pgtable_t token = pmd_pgtable(*pmd);
138 pmd_clear(pmd);
139 pte_free_tlb(tlb, token);
140 tlb->mm->nr_ptes--;
143 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
144 unsigned long addr, unsigned long end,
145 unsigned long floor, unsigned long ceiling)
147 pmd_t *pmd;
148 unsigned long next;
149 unsigned long start;
151 start = addr;
152 pmd = pmd_offset(pud, addr);
153 do {
154 next = pmd_addr_end(addr, end);
155 if (pmd_none_or_clear_bad(pmd))
156 continue;
157 free_pte_range(tlb, pmd);
158 } while (pmd++, addr = next, addr != end);
160 start &= PUD_MASK;
161 if (start < floor)
162 return;
163 if (ceiling) {
164 ceiling &= PUD_MASK;
165 if (!ceiling)
166 return;
168 if (end - 1 > ceiling - 1)
169 return;
171 pmd = pmd_offset(pud, start);
172 pud_clear(pud);
173 pmd_free_tlb(tlb, pmd);
176 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
177 unsigned long addr, unsigned long end,
178 unsigned long floor, unsigned long ceiling)
180 pud_t *pud;
181 unsigned long next;
182 unsigned long start;
184 start = addr;
185 pud = pud_offset(pgd, addr);
186 do {
187 next = pud_addr_end(addr, end);
188 if (pud_none_or_clear_bad(pud))
189 continue;
190 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
191 } while (pud++, addr = next, addr != end);
193 start &= PGDIR_MASK;
194 if (start < floor)
195 return;
196 if (ceiling) {
197 ceiling &= PGDIR_MASK;
198 if (!ceiling)
199 return;
201 if (end - 1 > ceiling - 1)
202 return;
204 pud = pud_offset(pgd, start);
205 pgd_clear(pgd);
206 pud_free_tlb(tlb, pud);
210 * This function frees user-level page tables of a process.
212 * Must be called with pagetable lock held.
214 void free_pgd_range(struct mmu_gather **tlb,
215 unsigned long addr, unsigned long end,
216 unsigned long floor, unsigned long ceiling)
218 pgd_t *pgd;
219 unsigned long next;
220 unsigned long start;
223 * The next few lines have given us lots of grief...
225 * Why are we testing PMD* at this top level? Because often
226 * there will be no work to do at all, and we'd prefer not to
227 * go all the way down to the bottom just to discover that.
229 * Why all these "- 1"s? Because 0 represents both the bottom
230 * of the address space and the top of it (using -1 for the
231 * top wouldn't help much: the masks would do the wrong thing).
232 * The rule is that addr 0 and floor 0 refer to the bottom of
233 * the address space, but end 0 and ceiling 0 refer to the top
234 * Comparisons need to use "end - 1" and "ceiling - 1" (though
235 * that end 0 case should be mythical).
237 * Wherever addr is brought up or ceiling brought down, we must
238 * be careful to reject "the opposite 0" before it confuses the
239 * subsequent tests. But what about where end is brought down
240 * by PMD_SIZE below? no, end can't go down to 0 there.
242 * Whereas we round start (addr) and ceiling down, by different
243 * masks at different levels, in order to test whether a table
244 * now has no other vmas using it, so can be freed, we don't
245 * bother to round floor or end up - the tests don't need that.
248 addr &= PMD_MASK;
249 if (addr < floor) {
250 addr += PMD_SIZE;
251 if (!addr)
252 return;
254 if (ceiling) {
255 ceiling &= PMD_MASK;
256 if (!ceiling)
257 return;
259 if (end - 1 > ceiling - 1)
260 end -= PMD_SIZE;
261 if (addr > end - 1)
262 return;
264 start = addr;
265 pgd = pgd_offset((*tlb)->mm, addr);
266 do {
267 next = pgd_addr_end(addr, end);
268 if (pgd_none_or_clear_bad(pgd))
269 continue;
270 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
271 } while (pgd++, addr = next, addr != end);
274 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
275 unsigned long floor, unsigned long ceiling)
277 while (vma) {
278 struct vm_area_struct *next = vma->vm_next;
279 unsigned long addr = vma->vm_start;
282 * Hide vma from rmap and vmtruncate before freeing pgtables
284 anon_vma_unlink(vma);
285 unlink_file_vma(vma);
287 if (is_vm_hugetlb_page(vma)) {
288 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
289 floor, next? next->vm_start: ceiling);
290 } else {
292 * Optimization: gather nearby vmas into one call down
294 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
295 && !is_vm_hugetlb_page(next)) {
296 vma = next;
297 next = vma->vm_next;
298 anon_vma_unlink(vma);
299 unlink_file_vma(vma);
301 free_pgd_range(tlb, addr, vma->vm_end,
302 floor, next? next->vm_start: ceiling);
304 vma = next;
308 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
310 pgtable_t new = pte_alloc_one(mm, address);
311 if (!new)
312 return -ENOMEM;
315 * Ensure all pte setup (eg. pte page lock and page clearing) are
316 * visible before the pte is made visible to other CPUs by being
317 * put into page tables.
319 * The other side of the story is the pointer chasing in the page
320 * table walking code (when walking the page table without locking;
321 * ie. most of the time). Fortunately, these data accesses consist
322 * of a chain of data-dependent loads, meaning most CPUs (alpha
323 * being the notable exception) will already guarantee loads are
324 * seen in-order. See the alpha page table accessors for the
325 * smp_read_barrier_depends() barriers in page table walking code.
327 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
329 spin_lock(&mm->page_table_lock);
330 if (!pmd_present(*pmd)) { /* Has another populated it ? */
331 mm->nr_ptes++;
332 pmd_populate(mm, pmd, new);
333 new = NULL;
335 spin_unlock(&mm->page_table_lock);
336 if (new)
337 pte_free(mm, new);
338 return 0;
341 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
343 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
344 if (!new)
345 return -ENOMEM;
347 smp_wmb(); /* See comment in __pte_alloc */
349 spin_lock(&init_mm.page_table_lock);
350 if (!pmd_present(*pmd)) { /* Has another populated it ? */
351 pmd_populate_kernel(&init_mm, pmd, new);
352 new = NULL;
354 spin_unlock(&init_mm.page_table_lock);
355 if (new)
356 pte_free_kernel(&init_mm, new);
357 return 0;
360 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
362 if (file_rss)
363 add_mm_counter(mm, file_rss, file_rss);
364 if (anon_rss)
365 add_mm_counter(mm, anon_rss, anon_rss);
369 * This function is called to print an error when a bad pte
370 * is found. For example, we might have a PFN-mapped pte in
371 * a region that doesn't allow it.
373 * The calling function must still handle the error.
375 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
377 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
378 "vm_flags = %lx, vaddr = %lx\n",
379 (long long)pte_val(pte),
380 (vma->vm_mm == current->mm ? current->comm : "???"),
381 vma->vm_flags, vaddr);
382 dump_stack();
385 static inline int is_cow_mapping(unsigned int flags)
387 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
391 * vm_normal_page -- This function gets the "struct page" associated with a pte.
393 * "Special" mappings do not wish to be associated with a "struct page" (either
394 * it doesn't exist, or it exists but they don't want to touch it). In this
395 * case, NULL is returned here. "Normal" mappings do have a struct page.
397 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
398 * pte bit, in which case this function is trivial. Secondly, an architecture
399 * may not have a spare pte bit, which requires a more complicated scheme,
400 * described below.
402 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
403 * special mapping (even if there are underlying and valid "struct pages").
404 * COWed pages of a VM_PFNMAP are always normal.
406 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
407 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
408 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
409 * mapping will always honor the rule
411 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
413 * And for normal mappings this is false.
415 * This restricts such mappings to be a linear translation from virtual address
416 * to pfn. To get around this restriction, we allow arbitrary mappings so long
417 * as the vma is not a COW mapping; in that case, we know that all ptes are
418 * special (because none can have been COWed).
421 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
423 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
424 * page" backing, however the difference is that _all_ pages with a struct
425 * page (that is, those where pfn_valid is true) are refcounted and considered
426 * normal pages by the VM. The disadvantage is that pages are refcounted
427 * (which can be slower and simply not an option for some PFNMAP users). The
428 * advantage is that we don't have to follow the strict linearity rule of
429 * PFNMAP mappings in order to support COWable mappings.
432 #ifdef __HAVE_ARCH_PTE_SPECIAL
433 # define HAVE_PTE_SPECIAL 1
434 #else
435 # define HAVE_PTE_SPECIAL 0
436 #endif
437 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
438 pte_t pte)
440 unsigned long pfn;
442 if (HAVE_PTE_SPECIAL) {
443 if (likely(!pte_special(pte))) {
444 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
445 return pte_page(pte);
447 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
448 return NULL;
451 /* !HAVE_PTE_SPECIAL case follows: */
453 pfn = pte_pfn(pte);
455 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
456 if (vma->vm_flags & VM_MIXEDMAP) {
457 if (!pfn_valid(pfn))
458 return NULL;
459 goto out;
460 } else {
461 unsigned long off;
462 off = (addr - vma->vm_start) >> PAGE_SHIFT;
463 if (pfn == vma->vm_pgoff + off)
464 return NULL;
465 if (!is_cow_mapping(vma->vm_flags))
466 return NULL;
470 VM_BUG_ON(!pfn_valid(pfn));
473 * NOTE! We still have PageReserved() pages in the page tables.
475 * eg. VDSO mappings can cause them to exist.
477 out:
478 return pfn_to_page(pfn);
482 * copy one vm_area from one task to the other. Assumes the page tables
483 * already present in the new task to be cleared in the whole range
484 * covered by this vma.
487 static inline void
488 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
489 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
490 unsigned long addr, int *rss)
492 unsigned long vm_flags = vma->vm_flags;
493 pte_t pte = *src_pte;
494 struct page *page;
496 /* pte contains position in swap or file, so copy. */
497 if (unlikely(!pte_present(pte))) {
498 if (!pte_file(pte)) {
499 swp_entry_t entry = pte_to_swp_entry(pte);
501 swap_duplicate(entry);
502 /* make sure dst_mm is on swapoff's mmlist. */
503 if (unlikely(list_empty(&dst_mm->mmlist))) {
504 spin_lock(&mmlist_lock);
505 if (list_empty(&dst_mm->mmlist))
506 list_add(&dst_mm->mmlist,
507 &src_mm->mmlist);
508 spin_unlock(&mmlist_lock);
510 if (is_write_migration_entry(entry) &&
511 is_cow_mapping(vm_flags)) {
513 * COW mappings require pages in both parent
514 * and child to be set to read.
516 make_migration_entry_read(&entry);
517 pte = swp_entry_to_pte(entry);
518 set_pte_at(src_mm, addr, src_pte, pte);
521 goto out_set_pte;
525 * If it's a COW mapping, write protect it both
526 * in the parent and the child
528 if (is_cow_mapping(vm_flags)) {
529 ptep_set_wrprotect(src_mm, addr, src_pte);
530 pte = pte_wrprotect(pte);
534 * If it's a shared mapping, mark it clean in
535 * the child
537 if (vm_flags & VM_SHARED)
538 pte = pte_mkclean(pte);
539 pte = pte_mkold(pte);
541 page = vm_normal_page(vma, addr, pte);
542 if (page) {
543 get_page(page);
544 page_dup_rmap(page, vma, addr);
545 rss[!!PageAnon(page)]++;
548 out_set_pte:
549 set_pte_at(dst_mm, addr, dst_pte, pte);
552 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
553 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
554 unsigned long addr, unsigned long end)
556 pte_t *src_pte, *dst_pte;
557 spinlock_t *src_ptl, *dst_ptl;
558 int progress = 0;
559 int rss[2];
561 again:
562 rss[1] = rss[0] = 0;
563 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
564 if (!dst_pte)
565 return -ENOMEM;
566 src_pte = pte_offset_map_nested(src_pmd, addr);
567 src_ptl = pte_lockptr(src_mm, src_pmd);
568 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
569 arch_enter_lazy_mmu_mode();
571 do {
573 * We are holding two locks at this point - either of them
574 * could generate latencies in another task on another CPU.
576 if (progress >= 32) {
577 progress = 0;
578 if (need_resched() ||
579 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
580 break;
582 if (pte_none(*src_pte)) {
583 progress++;
584 continue;
586 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
587 progress += 8;
588 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
590 arch_leave_lazy_mmu_mode();
591 spin_unlock(src_ptl);
592 pte_unmap_nested(src_pte - 1);
593 add_mm_rss(dst_mm, rss[0], rss[1]);
594 pte_unmap_unlock(dst_pte - 1, dst_ptl);
595 cond_resched();
596 if (addr != end)
597 goto again;
598 return 0;
601 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
602 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
603 unsigned long addr, unsigned long end)
605 pmd_t *src_pmd, *dst_pmd;
606 unsigned long next;
608 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
609 if (!dst_pmd)
610 return -ENOMEM;
611 src_pmd = pmd_offset(src_pud, addr);
612 do {
613 next = pmd_addr_end(addr, end);
614 if (pmd_none_or_clear_bad(src_pmd))
615 continue;
616 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
617 vma, addr, next))
618 return -ENOMEM;
619 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
620 return 0;
623 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
624 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
625 unsigned long addr, unsigned long end)
627 pud_t *src_pud, *dst_pud;
628 unsigned long next;
630 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
631 if (!dst_pud)
632 return -ENOMEM;
633 src_pud = pud_offset(src_pgd, addr);
634 do {
635 next = pud_addr_end(addr, end);
636 if (pud_none_or_clear_bad(src_pud))
637 continue;
638 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
639 vma, addr, next))
640 return -ENOMEM;
641 } while (dst_pud++, src_pud++, addr = next, addr != end);
642 return 0;
645 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
646 struct vm_area_struct *vma)
648 pgd_t *src_pgd, *dst_pgd;
649 unsigned long next;
650 unsigned long addr = vma->vm_start;
651 unsigned long end = vma->vm_end;
654 * Don't copy ptes where a page fault will fill them correctly.
655 * Fork becomes much lighter when there are big shared or private
656 * readonly mappings. The tradeoff is that copy_page_range is more
657 * efficient than faulting.
659 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
660 if (!vma->anon_vma)
661 return 0;
664 if (is_vm_hugetlb_page(vma))
665 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
667 dst_pgd = pgd_offset(dst_mm, addr);
668 src_pgd = pgd_offset(src_mm, addr);
669 do {
670 next = pgd_addr_end(addr, end);
671 if (pgd_none_or_clear_bad(src_pgd))
672 continue;
673 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
674 vma, addr, next))
675 return -ENOMEM;
676 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
677 return 0;
680 static unsigned long zap_pte_range(struct mmu_gather *tlb,
681 struct vm_area_struct *vma, pmd_t *pmd,
682 unsigned long addr, unsigned long end,
683 long *zap_work, struct zap_details *details)
685 struct mm_struct *mm = tlb->mm;
686 pte_t *pte;
687 spinlock_t *ptl;
688 int file_rss = 0;
689 int anon_rss = 0;
691 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
692 arch_enter_lazy_mmu_mode();
693 do {
694 pte_t ptent = *pte;
695 if (pte_none(ptent)) {
696 (*zap_work)--;
697 continue;
700 (*zap_work) -= PAGE_SIZE;
702 if (pte_present(ptent)) {
703 struct page *page;
705 page = vm_normal_page(vma, addr, ptent);
706 if (unlikely(details) && page) {
708 * unmap_shared_mapping_pages() wants to
709 * invalidate cache without truncating:
710 * unmap shared but keep private pages.
712 if (details->check_mapping &&
713 details->check_mapping != page->mapping)
714 continue;
716 * Each page->index must be checked when
717 * invalidating or truncating nonlinear.
719 if (details->nonlinear_vma &&
720 (page->index < details->first_index ||
721 page->index > details->last_index))
722 continue;
724 ptent = ptep_get_and_clear_full(mm, addr, pte,
725 tlb->fullmm);
726 tlb_remove_tlb_entry(tlb, pte, addr);
727 if (unlikely(!page))
728 continue;
729 if (unlikely(details) && details->nonlinear_vma
730 && linear_page_index(details->nonlinear_vma,
731 addr) != page->index)
732 set_pte_at(mm, addr, pte,
733 pgoff_to_pte(page->index));
734 if (PageAnon(page))
735 anon_rss--;
736 else {
737 if (pte_dirty(ptent))
738 set_page_dirty(page);
739 if (pte_young(ptent))
740 SetPageReferenced(page);
741 file_rss--;
743 page_remove_rmap(page, vma);
744 tlb_remove_page(tlb, page);
745 continue;
748 * If details->check_mapping, we leave swap entries;
749 * if details->nonlinear_vma, we leave file entries.
751 if (unlikely(details))
752 continue;
753 if (!pte_file(ptent))
754 free_swap_and_cache(pte_to_swp_entry(ptent));
755 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
756 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
758 add_mm_rss(mm, file_rss, anon_rss);
759 arch_leave_lazy_mmu_mode();
760 pte_unmap_unlock(pte - 1, ptl);
762 return addr;
765 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
766 struct vm_area_struct *vma, pud_t *pud,
767 unsigned long addr, unsigned long end,
768 long *zap_work, struct zap_details *details)
770 pmd_t *pmd;
771 unsigned long next;
773 pmd = pmd_offset(pud, addr);
774 do {
775 next = pmd_addr_end(addr, end);
776 if (pmd_none_or_clear_bad(pmd)) {
777 (*zap_work)--;
778 continue;
780 next = zap_pte_range(tlb, vma, pmd, addr, next,
781 zap_work, details);
782 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
784 return addr;
787 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
788 struct vm_area_struct *vma, pgd_t *pgd,
789 unsigned long addr, unsigned long end,
790 long *zap_work, struct zap_details *details)
792 pud_t *pud;
793 unsigned long next;
795 pud = pud_offset(pgd, addr);
796 do {
797 next = pud_addr_end(addr, end);
798 if (pud_none_or_clear_bad(pud)) {
799 (*zap_work)--;
800 continue;
802 next = zap_pmd_range(tlb, vma, pud, addr, next,
803 zap_work, details);
804 } while (pud++, addr = next, (addr != end && *zap_work > 0));
806 return addr;
809 static unsigned long unmap_page_range(struct mmu_gather *tlb,
810 struct vm_area_struct *vma,
811 unsigned long addr, unsigned long end,
812 long *zap_work, struct zap_details *details)
814 pgd_t *pgd;
815 unsigned long next;
817 if (details && !details->check_mapping && !details->nonlinear_vma)
818 details = NULL;
820 BUG_ON(addr >= end);
821 tlb_start_vma(tlb, vma);
822 pgd = pgd_offset(vma->vm_mm, addr);
823 do {
824 next = pgd_addr_end(addr, end);
825 if (pgd_none_or_clear_bad(pgd)) {
826 (*zap_work)--;
827 continue;
829 next = zap_pud_range(tlb, vma, pgd, addr, next,
830 zap_work, details);
831 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
832 tlb_end_vma(tlb, vma);
834 return addr;
837 #ifdef CONFIG_PREEMPT
838 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
839 #else
840 /* No preempt: go for improved straight-line efficiency */
841 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
842 #endif
845 * unmap_vmas - unmap a range of memory covered by a list of vma's
846 * @tlbp: address of the caller's struct mmu_gather
847 * @vma: the starting vma
848 * @start_addr: virtual address at which to start unmapping
849 * @end_addr: virtual address at which to end unmapping
850 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
851 * @details: details of nonlinear truncation or shared cache invalidation
853 * Returns the end address of the unmapping (restart addr if interrupted).
855 * Unmap all pages in the vma list.
857 * We aim to not hold locks for too long (for scheduling latency reasons).
858 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
859 * return the ending mmu_gather to the caller.
861 * Only addresses between `start' and `end' will be unmapped.
863 * The VMA list must be sorted in ascending virtual address order.
865 * unmap_vmas() assumes that the caller will flush the whole unmapped address
866 * range after unmap_vmas() returns. So the only responsibility here is to
867 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
868 * drops the lock and schedules.
870 unsigned long unmap_vmas(struct mmu_gather **tlbp,
871 struct vm_area_struct *vma, unsigned long start_addr,
872 unsigned long end_addr, unsigned long *nr_accounted,
873 struct zap_details *details)
875 long zap_work = ZAP_BLOCK_SIZE;
876 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
877 int tlb_start_valid = 0;
878 unsigned long start = start_addr;
879 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
880 int fullmm = (*tlbp)->fullmm;
882 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
883 unsigned long end;
885 start = max(vma->vm_start, start_addr);
886 if (start >= vma->vm_end)
887 continue;
888 end = min(vma->vm_end, end_addr);
889 if (end <= vma->vm_start)
890 continue;
892 if (vma->vm_flags & VM_ACCOUNT)
893 *nr_accounted += (end - start) >> PAGE_SHIFT;
895 while (start != end) {
896 if (!tlb_start_valid) {
897 tlb_start = start;
898 tlb_start_valid = 1;
901 if (unlikely(is_vm_hugetlb_page(vma))) {
902 unmap_hugepage_range(vma, start, end);
903 zap_work -= (end - start) /
904 (HPAGE_SIZE / PAGE_SIZE);
905 start = end;
906 } else
907 start = unmap_page_range(*tlbp, vma,
908 start, end, &zap_work, details);
910 if (zap_work > 0) {
911 BUG_ON(start != end);
912 break;
915 tlb_finish_mmu(*tlbp, tlb_start, start);
917 if (need_resched() ||
918 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
919 if (i_mmap_lock) {
920 *tlbp = NULL;
921 goto out;
923 cond_resched();
926 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
927 tlb_start_valid = 0;
928 zap_work = ZAP_BLOCK_SIZE;
931 out:
932 return start; /* which is now the end (or restart) address */
936 * zap_page_range - remove user pages in a given range
937 * @vma: vm_area_struct holding the applicable pages
938 * @address: starting address of pages to zap
939 * @size: number of bytes to zap
940 * @details: details of nonlinear truncation or shared cache invalidation
942 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
943 unsigned long size, struct zap_details *details)
945 struct mm_struct *mm = vma->vm_mm;
946 struct mmu_gather *tlb;
947 unsigned long end = address + size;
948 unsigned long nr_accounted = 0;
950 lru_add_drain();
951 tlb = tlb_gather_mmu(mm, 0);
952 update_hiwater_rss(mm);
953 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
954 if (tlb)
955 tlb_finish_mmu(tlb, address, end);
956 return end;
960 * Do a quick page-table lookup for a single page.
962 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
963 unsigned int flags)
965 pgd_t *pgd;
966 pud_t *pud;
967 pmd_t *pmd;
968 pte_t *ptep, pte;
969 spinlock_t *ptl;
970 struct page *page;
971 struct mm_struct *mm = vma->vm_mm;
973 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
974 if (!IS_ERR(page)) {
975 BUG_ON(flags & FOLL_GET);
976 goto out;
979 page = NULL;
980 pgd = pgd_offset(mm, address);
981 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
982 goto no_page_table;
984 pud = pud_offset(pgd, address);
985 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
986 goto no_page_table;
988 pmd = pmd_offset(pud, address);
989 if (pmd_none(*pmd))
990 goto no_page_table;
992 if (pmd_huge(*pmd)) {
993 BUG_ON(flags & FOLL_GET);
994 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
995 goto out;
998 if (unlikely(pmd_bad(*pmd)))
999 goto no_page_table;
1001 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1002 if (!ptep)
1003 goto out;
1005 pte = *ptep;
1006 if (!pte_present(pte))
1007 goto unlock;
1008 if ((flags & FOLL_WRITE) && !pte_write(pte))
1009 goto unlock;
1010 page = vm_normal_page(vma, address, pte);
1011 if (unlikely(!page))
1012 goto unlock;
1014 if (flags & FOLL_GET)
1015 get_page(page);
1016 if (flags & FOLL_TOUCH) {
1017 if ((flags & FOLL_WRITE) &&
1018 !pte_dirty(pte) && !PageDirty(page))
1019 set_page_dirty(page);
1020 mark_page_accessed(page);
1022 unlock:
1023 pte_unmap_unlock(ptep, ptl);
1024 out:
1025 return page;
1027 no_page_table:
1029 * When core dumping an enormous anonymous area that nobody
1030 * has touched so far, we don't want to allocate page tables.
1032 if (flags & FOLL_ANON) {
1033 page = ZERO_PAGE(0);
1034 if (flags & FOLL_GET)
1035 get_page(page);
1036 BUG_ON(flags & FOLL_WRITE);
1038 return page;
1041 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1042 unsigned long start, int len, int write, int force,
1043 struct page **pages, struct vm_area_struct **vmas)
1045 int i;
1046 unsigned int vm_flags;
1048 if (len <= 0)
1049 return 0;
1051 * Require read or write permissions.
1052 * If 'force' is set, we only require the "MAY" flags.
1054 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1055 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1056 i = 0;
1058 do {
1059 struct vm_area_struct *vma;
1060 unsigned int foll_flags;
1062 vma = find_extend_vma(mm, start);
1063 if (!vma && in_gate_area(tsk, start)) {
1064 unsigned long pg = start & PAGE_MASK;
1065 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1066 pgd_t *pgd;
1067 pud_t *pud;
1068 pmd_t *pmd;
1069 pte_t *pte;
1070 if (write) /* user gate pages are read-only */
1071 return i ? : -EFAULT;
1072 if (pg > TASK_SIZE)
1073 pgd = pgd_offset_k(pg);
1074 else
1075 pgd = pgd_offset_gate(mm, pg);
1076 BUG_ON(pgd_none(*pgd));
1077 pud = pud_offset(pgd, pg);
1078 BUG_ON(pud_none(*pud));
1079 pmd = pmd_offset(pud, pg);
1080 if (pmd_none(*pmd))
1081 return i ? : -EFAULT;
1082 pte = pte_offset_map(pmd, pg);
1083 if (pte_none(*pte)) {
1084 pte_unmap(pte);
1085 return i ? : -EFAULT;
1087 if (pages) {
1088 struct page *page = vm_normal_page(gate_vma, start, *pte);
1089 pages[i] = page;
1090 if (page)
1091 get_page(page);
1093 pte_unmap(pte);
1094 if (vmas)
1095 vmas[i] = gate_vma;
1096 i++;
1097 start += PAGE_SIZE;
1098 len--;
1099 continue;
1102 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1103 || !(vm_flags & vma->vm_flags))
1104 return i ? : -EFAULT;
1106 if (is_vm_hugetlb_page(vma)) {
1107 i = follow_hugetlb_page(mm, vma, pages, vmas,
1108 &start, &len, i, write);
1109 continue;
1112 foll_flags = FOLL_TOUCH;
1113 if (pages)
1114 foll_flags |= FOLL_GET;
1115 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1116 (!vma->vm_ops || !vma->vm_ops->fault))
1117 foll_flags |= FOLL_ANON;
1119 do {
1120 struct page *page;
1123 * If tsk is ooming, cut off its access to large memory
1124 * allocations. It has a pending SIGKILL, but it can't
1125 * be processed until returning to user space.
1127 if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1128 return -ENOMEM;
1130 if (write)
1131 foll_flags |= FOLL_WRITE;
1133 cond_resched();
1134 while (!(page = follow_page(vma, start, foll_flags))) {
1135 int ret;
1136 ret = handle_mm_fault(mm, vma, start,
1137 foll_flags & FOLL_WRITE);
1138 if (ret & VM_FAULT_ERROR) {
1139 if (ret & VM_FAULT_OOM)
1140 return i ? i : -ENOMEM;
1141 else if (ret & VM_FAULT_SIGBUS)
1142 return i ? i : -EFAULT;
1143 BUG();
1145 if (ret & VM_FAULT_MAJOR)
1146 tsk->maj_flt++;
1147 else
1148 tsk->min_flt++;
1151 * The VM_FAULT_WRITE bit tells us that
1152 * do_wp_page has broken COW when necessary,
1153 * even if maybe_mkwrite decided not to set
1154 * pte_write. We can thus safely do subsequent
1155 * page lookups as if they were reads.
1157 if (ret & VM_FAULT_WRITE)
1158 foll_flags &= ~FOLL_WRITE;
1160 cond_resched();
1162 if (pages) {
1163 pages[i] = page;
1165 flush_anon_page(vma, page, start);
1166 flush_dcache_page(page);
1168 if (vmas)
1169 vmas[i] = vma;
1170 i++;
1171 start += PAGE_SIZE;
1172 len--;
1173 } while (len && start < vma->vm_end);
1174 } while (len);
1175 return i;
1177 EXPORT_SYMBOL(get_user_pages);
1179 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1180 spinlock_t **ptl)
1182 pgd_t * pgd = pgd_offset(mm, addr);
1183 pud_t * pud = pud_alloc(mm, pgd, addr);
1184 if (pud) {
1185 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1186 if (pmd)
1187 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1189 return NULL;
1193 * This is the old fallback for page remapping.
1195 * For historical reasons, it only allows reserved pages. Only
1196 * old drivers should use this, and they needed to mark their
1197 * pages reserved for the old functions anyway.
1199 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1200 struct page *page, pgprot_t prot)
1202 struct mm_struct *mm = vma->vm_mm;
1203 int retval;
1204 pte_t *pte;
1205 spinlock_t *ptl;
1207 retval = mem_cgroup_charge(page, mm, GFP_KERNEL);
1208 if (retval)
1209 goto out;
1211 retval = -EINVAL;
1212 if (PageAnon(page))
1213 goto out_uncharge;
1214 retval = -ENOMEM;
1215 flush_dcache_page(page);
1216 pte = get_locked_pte(mm, addr, &ptl);
1217 if (!pte)
1218 goto out_uncharge;
1219 retval = -EBUSY;
1220 if (!pte_none(*pte))
1221 goto out_unlock;
1223 /* Ok, finally just insert the thing.. */
1224 get_page(page);
1225 inc_mm_counter(mm, file_rss);
1226 page_add_file_rmap(page);
1227 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1229 retval = 0;
1230 pte_unmap_unlock(pte, ptl);
1231 return retval;
1232 out_unlock:
1233 pte_unmap_unlock(pte, ptl);
1234 out_uncharge:
1235 mem_cgroup_uncharge_page(page);
1236 out:
1237 return retval;
1241 * vm_insert_page - insert single page into user vma
1242 * @vma: user vma to map to
1243 * @addr: target user address of this page
1244 * @page: source kernel page
1246 * This allows drivers to insert individual pages they've allocated
1247 * into a user vma.
1249 * The page has to be a nice clean _individual_ kernel allocation.
1250 * If you allocate a compound page, you need to have marked it as
1251 * such (__GFP_COMP), or manually just split the page up yourself
1252 * (see split_page()).
1254 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1255 * took an arbitrary page protection parameter. This doesn't allow
1256 * that. Your vma protection will have to be set up correctly, which
1257 * means that if you want a shared writable mapping, you'd better
1258 * ask for a shared writable mapping!
1260 * The page does not need to be reserved.
1262 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1263 struct page *page)
1265 if (addr < vma->vm_start || addr >= vma->vm_end)
1266 return -EFAULT;
1267 if (!page_count(page))
1268 return -EINVAL;
1269 vma->vm_flags |= VM_INSERTPAGE;
1270 return insert_page(vma, addr, page, vma->vm_page_prot);
1272 EXPORT_SYMBOL(vm_insert_page);
1274 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1275 unsigned long pfn, pgprot_t prot)
1277 struct mm_struct *mm = vma->vm_mm;
1278 int retval;
1279 pte_t *pte, entry;
1280 spinlock_t *ptl;
1282 retval = -ENOMEM;
1283 pte = get_locked_pte(mm, addr, &ptl);
1284 if (!pte)
1285 goto out;
1286 retval = -EBUSY;
1287 if (!pte_none(*pte))
1288 goto out_unlock;
1290 /* Ok, finally just insert the thing.. */
1291 entry = pte_mkspecial(pfn_pte(pfn, prot));
1292 set_pte_at(mm, addr, pte, entry);
1293 update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1295 retval = 0;
1296 out_unlock:
1297 pte_unmap_unlock(pte, ptl);
1298 out:
1299 return retval;
1303 * vm_insert_pfn - insert single pfn into user vma
1304 * @vma: user vma to map to
1305 * @addr: target user address of this page
1306 * @pfn: source kernel pfn
1308 * Similar to vm_inert_page, this allows drivers to insert individual pages
1309 * they've allocated into a user vma. Same comments apply.
1311 * This function should only be called from a vm_ops->fault handler, and
1312 * in that case the handler should return NULL.
1314 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1315 unsigned long pfn)
1318 * Technically, architectures with pte_special can avoid all these
1319 * restrictions (same for remap_pfn_range). However we would like
1320 * consistency in testing and feature parity among all, so we should
1321 * try to keep these invariants in place for everybody.
1323 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1324 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1325 (VM_PFNMAP|VM_MIXEDMAP));
1326 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1327 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1329 if (addr < vma->vm_start || addr >= vma->vm_end)
1330 return -EFAULT;
1331 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1333 EXPORT_SYMBOL(vm_insert_pfn);
1335 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1336 unsigned long pfn)
1338 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1340 if (addr < vma->vm_start || addr >= vma->vm_end)
1341 return -EFAULT;
1344 * If we don't have pte special, then we have to use the pfn_valid()
1345 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1346 * refcount the page if pfn_valid is true (hence insert_page rather
1347 * than insert_pfn).
1349 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1350 struct page *page;
1352 page = pfn_to_page(pfn);
1353 return insert_page(vma, addr, page, vma->vm_page_prot);
1355 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1357 EXPORT_SYMBOL(vm_insert_mixed);
1360 * maps a range of physical memory into the requested pages. the old
1361 * mappings are removed. any references to nonexistent pages results
1362 * in null mappings (currently treated as "copy-on-access")
1364 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1365 unsigned long addr, unsigned long end,
1366 unsigned long pfn, pgprot_t prot)
1368 pte_t *pte;
1369 spinlock_t *ptl;
1371 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1372 if (!pte)
1373 return -ENOMEM;
1374 arch_enter_lazy_mmu_mode();
1375 do {
1376 BUG_ON(!pte_none(*pte));
1377 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1378 pfn++;
1379 } while (pte++, addr += PAGE_SIZE, addr != end);
1380 arch_leave_lazy_mmu_mode();
1381 pte_unmap_unlock(pte - 1, ptl);
1382 return 0;
1385 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1386 unsigned long addr, unsigned long end,
1387 unsigned long pfn, pgprot_t prot)
1389 pmd_t *pmd;
1390 unsigned long next;
1392 pfn -= addr >> PAGE_SHIFT;
1393 pmd = pmd_alloc(mm, pud, addr);
1394 if (!pmd)
1395 return -ENOMEM;
1396 do {
1397 next = pmd_addr_end(addr, end);
1398 if (remap_pte_range(mm, pmd, addr, next,
1399 pfn + (addr >> PAGE_SHIFT), prot))
1400 return -ENOMEM;
1401 } while (pmd++, addr = next, addr != end);
1402 return 0;
1405 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1406 unsigned long addr, unsigned long end,
1407 unsigned long pfn, pgprot_t prot)
1409 pud_t *pud;
1410 unsigned long next;
1412 pfn -= addr >> PAGE_SHIFT;
1413 pud = pud_alloc(mm, pgd, addr);
1414 if (!pud)
1415 return -ENOMEM;
1416 do {
1417 next = pud_addr_end(addr, end);
1418 if (remap_pmd_range(mm, pud, addr, next,
1419 pfn + (addr >> PAGE_SHIFT), prot))
1420 return -ENOMEM;
1421 } while (pud++, addr = next, addr != end);
1422 return 0;
1426 * remap_pfn_range - remap kernel memory to userspace
1427 * @vma: user vma to map to
1428 * @addr: target user address to start at
1429 * @pfn: physical address of kernel memory
1430 * @size: size of map area
1431 * @prot: page protection flags for this mapping
1433 * Note: this is only safe if the mm semaphore is held when called.
1435 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1436 unsigned long pfn, unsigned long size, pgprot_t prot)
1438 pgd_t *pgd;
1439 unsigned long next;
1440 unsigned long end = addr + PAGE_ALIGN(size);
1441 struct mm_struct *mm = vma->vm_mm;
1442 int err;
1445 * Physically remapped pages are special. Tell the
1446 * rest of the world about it:
1447 * VM_IO tells people not to look at these pages
1448 * (accesses can have side effects).
1449 * VM_RESERVED is specified all over the place, because
1450 * in 2.4 it kept swapout's vma scan off this vma; but
1451 * in 2.6 the LRU scan won't even find its pages, so this
1452 * flag means no more than count its pages in reserved_vm,
1453 * and omit it from core dump, even when VM_IO turned off.
1454 * VM_PFNMAP tells the core MM that the base pages are just
1455 * raw PFN mappings, and do not have a "struct page" associated
1456 * with them.
1458 * There's a horrible special case to handle copy-on-write
1459 * behaviour that some programs depend on. We mark the "original"
1460 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1462 if (is_cow_mapping(vma->vm_flags)) {
1463 if (addr != vma->vm_start || end != vma->vm_end)
1464 return -EINVAL;
1465 vma->vm_pgoff = pfn;
1468 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1470 BUG_ON(addr >= end);
1471 pfn -= addr >> PAGE_SHIFT;
1472 pgd = pgd_offset(mm, addr);
1473 flush_cache_range(vma, addr, end);
1474 do {
1475 next = pgd_addr_end(addr, end);
1476 err = remap_pud_range(mm, pgd, addr, next,
1477 pfn + (addr >> PAGE_SHIFT), prot);
1478 if (err)
1479 break;
1480 } while (pgd++, addr = next, addr != end);
1481 return err;
1483 EXPORT_SYMBOL(remap_pfn_range);
1485 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1486 unsigned long addr, unsigned long end,
1487 pte_fn_t fn, void *data)
1489 pte_t *pte;
1490 int err;
1491 pgtable_t token;
1492 spinlock_t *uninitialized_var(ptl);
1494 pte = (mm == &init_mm) ?
1495 pte_alloc_kernel(pmd, addr) :
1496 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1497 if (!pte)
1498 return -ENOMEM;
1500 BUG_ON(pmd_huge(*pmd));
1502 token = pmd_pgtable(*pmd);
1504 do {
1505 err = fn(pte, token, addr, data);
1506 if (err)
1507 break;
1508 } while (pte++, addr += PAGE_SIZE, addr != end);
1510 if (mm != &init_mm)
1511 pte_unmap_unlock(pte-1, ptl);
1512 return err;
1515 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1516 unsigned long addr, unsigned long end,
1517 pte_fn_t fn, void *data)
1519 pmd_t *pmd;
1520 unsigned long next;
1521 int err;
1523 pmd = pmd_alloc(mm, pud, addr);
1524 if (!pmd)
1525 return -ENOMEM;
1526 do {
1527 next = pmd_addr_end(addr, end);
1528 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1529 if (err)
1530 break;
1531 } while (pmd++, addr = next, addr != end);
1532 return err;
1535 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1536 unsigned long addr, unsigned long end,
1537 pte_fn_t fn, void *data)
1539 pud_t *pud;
1540 unsigned long next;
1541 int err;
1543 pud = pud_alloc(mm, pgd, addr);
1544 if (!pud)
1545 return -ENOMEM;
1546 do {
1547 next = pud_addr_end(addr, end);
1548 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1549 if (err)
1550 break;
1551 } while (pud++, addr = next, addr != end);
1552 return err;
1556 * Scan a region of virtual memory, filling in page tables as necessary
1557 * and calling a provided function on each leaf page table.
1559 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1560 unsigned long size, pte_fn_t fn, void *data)
1562 pgd_t *pgd;
1563 unsigned long next;
1564 unsigned long end = addr + size;
1565 int err;
1567 BUG_ON(addr >= end);
1568 pgd = pgd_offset(mm, addr);
1569 do {
1570 next = pgd_addr_end(addr, end);
1571 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1572 if (err)
1573 break;
1574 } while (pgd++, addr = next, addr != end);
1575 return err;
1577 EXPORT_SYMBOL_GPL(apply_to_page_range);
1580 * handle_pte_fault chooses page fault handler according to an entry
1581 * which was read non-atomically. Before making any commitment, on
1582 * those architectures or configurations (e.g. i386 with PAE) which
1583 * might give a mix of unmatched parts, do_swap_page and do_file_page
1584 * must check under lock before unmapping the pte and proceeding
1585 * (but do_wp_page is only called after already making such a check;
1586 * and do_anonymous_page and do_no_page can safely check later on).
1588 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1589 pte_t *page_table, pte_t orig_pte)
1591 int same = 1;
1592 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1593 if (sizeof(pte_t) > sizeof(unsigned long)) {
1594 spinlock_t *ptl = pte_lockptr(mm, pmd);
1595 spin_lock(ptl);
1596 same = pte_same(*page_table, orig_pte);
1597 spin_unlock(ptl);
1599 #endif
1600 pte_unmap(page_table);
1601 return same;
1605 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1606 * servicing faults for write access. In the normal case, do always want
1607 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1608 * that do not have writing enabled, when used by access_process_vm.
1610 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1612 if (likely(vma->vm_flags & VM_WRITE))
1613 pte = pte_mkwrite(pte);
1614 return pte;
1617 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1620 * If the source page was a PFN mapping, we don't have
1621 * a "struct page" for it. We do a best-effort copy by
1622 * just copying from the original user address. If that
1623 * fails, we just zero-fill it. Live with it.
1625 if (unlikely(!src)) {
1626 void *kaddr = kmap_atomic(dst, KM_USER0);
1627 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1630 * This really shouldn't fail, because the page is there
1631 * in the page tables. But it might just be unreadable,
1632 * in which case we just give up and fill the result with
1633 * zeroes.
1635 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1636 memset(kaddr, 0, PAGE_SIZE);
1637 kunmap_atomic(kaddr, KM_USER0);
1638 flush_dcache_page(dst);
1639 } else
1640 copy_user_highpage(dst, src, va, vma);
1644 * This routine handles present pages, when users try to write
1645 * to a shared page. It is done by copying the page to a new address
1646 * and decrementing the shared-page counter for the old page.
1648 * Note that this routine assumes that the protection checks have been
1649 * done by the caller (the low-level page fault routine in most cases).
1650 * Thus we can safely just mark it writable once we've done any necessary
1651 * COW.
1653 * We also mark the page dirty at this point even though the page will
1654 * change only once the write actually happens. This avoids a few races,
1655 * and potentially makes it more efficient.
1657 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1658 * but allow concurrent faults), with pte both mapped and locked.
1659 * We return with mmap_sem still held, but pte unmapped and unlocked.
1661 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1662 unsigned long address, pte_t *page_table, pmd_t *pmd,
1663 spinlock_t *ptl, pte_t orig_pte)
1665 struct page *old_page, *new_page;
1666 pte_t entry;
1667 int reuse = 0, ret = 0;
1668 int page_mkwrite = 0;
1669 struct page *dirty_page = NULL;
1671 old_page = vm_normal_page(vma, address, orig_pte);
1672 if (!old_page)
1673 goto gotten;
1676 * Take out anonymous pages first, anonymous shared vmas are
1677 * not dirty accountable.
1679 if (PageAnon(old_page)) {
1680 if (!TestSetPageLocked(old_page)) {
1681 reuse = can_share_swap_page(old_page);
1682 unlock_page(old_page);
1684 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1685 (VM_WRITE|VM_SHARED))) {
1687 * Only catch write-faults on shared writable pages,
1688 * read-only shared pages can get COWed by
1689 * get_user_pages(.write=1, .force=1).
1691 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1693 * Notify the address space that the page is about to
1694 * become writable so that it can prohibit this or wait
1695 * for the page to get into an appropriate state.
1697 * We do this without the lock held, so that it can
1698 * sleep if it needs to.
1700 page_cache_get(old_page);
1701 pte_unmap_unlock(page_table, ptl);
1703 if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1704 goto unwritable_page;
1707 * Since we dropped the lock we need to revalidate
1708 * the PTE as someone else may have changed it. If
1709 * they did, we just return, as we can count on the
1710 * MMU to tell us if they didn't also make it writable.
1712 page_table = pte_offset_map_lock(mm, pmd, address,
1713 &ptl);
1714 page_cache_release(old_page);
1715 if (!pte_same(*page_table, orig_pte))
1716 goto unlock;
1718 page_mkwrite = 1;
1720 dirty_page = old_page;
1721 get_page(dirty_page);
1722 reuse = 1;
1725 if (reuse) {
1726 flush_cache_page(vma, address, pte_pfn(orig_pte));
1727 entry = pte_mkyoung(orig_pte);
1728 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1729 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1730 update_mmu_cache(vma, address, entry);
1731 ret |= VM_FAULT_WRITE;
1732 goto unlock;
1736 * Ok, we need to copy. Oh, well..
1738 page_cache_get(old_page);
1739 gotten:
1740 pte_unmap_unlock(page_table, ptl);
1742 if (unlikely(anon_vma_prepare(vma)))
1743 goto oom;
1744 VM_BUG_ON(old_page == ZERO_PAGE(0));
1745 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1746 if (!new_page)
1747 goto oom;
1748 cow_user_page(new_page, old_page, address, vma);
1749 __SetPageUptodate(new_page);
1751 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1752 goto oom_free_new;
1755 * Re-check the pte - we dropped the lock
1757 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1758 if (likely(pte_same(*page_table, orig_pte))) {
1759 if (old_page) {
1760 page_remove_rmap(old_page, vma);
1761 if (!PageAnon(old_page)) {
1762 dec_mm_counter(mm, file_rss);
1763 inc_mm_counter(mm, anon_rss);
1765 } else
1766 inc_mm_counter(mm, anon_rss);
1767 flush_cache_page(vma, address, pte_pfn(orig_pte));
1768 entry = mk_pte(new_page, vma->vm_page_prot);
1769 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1771 * Clear the pte entry and flush it first, before updating the
1772 * pte with the new entry. This will avoid a race condition
1773 * seen in the presence of one thread doing SMC and another
1774 * thread doing COW.
1776 ptep_clear_flush(vma, address, page_table);
1777 set_pte_at(mm, address, page_table, entry);
1778 update_mmu_cache(vma, address, entry);
1779 lru_cache_add_active(new_page);
1780 page_add_new_anon_rmap(new_page, vma, address);
1782 /* Free the old page.. */
1783 new_page = old_page;
1784 ret |= VM_FAULT_WRITE;
1785 } else
1786 mem_cgroup_uncharge_page(new_page);
1788 if (new_page)
1789 page_cache_release(new_page);
1790 if (old_page)
1791 page_cache_release(old_page);
1792 unlock:
1793 pte_unmap_unlock(page_table, ptl);
1794 if (dirty_page) {
1795 if (vma->vm_file)
1796 file_update_time(vma->vm_file);
1799 * Yes, Virginia, this is actually required to prevent a race
1800 * with clear_page_dirty_for_io() from clearing the page dirty
1801 * bit after it clear all dirty ptes, but before a racing
1802 * do_wp_page installs a dirty pte.
1804 * do_no_page is protected similarly.
1806 wait_on_page_locked(dirty_page);
1807 set_page_dirty_balance(dirty_page, page_mkwrite);
1808 put_page(dirty_page);
1810 return ret;
1811 oom_free_new:
1812 page_cache_release(new_page);
1813 oom:
1814 if (old_page)
1815 page_cache_release(old_page);
1816 return VM_FAULT_OOM;
1818 unwritable_page:
1819 page_cache_release(old_page);
1820 return VM_FAULT_SIGBUS;
1824 * Helper functions for unmap_mapping_range().
1826 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1828 * We have to restart searching the prio_tree whenever we drop the lock,
1829 * since the iterator is only valid while the lock is held, and anyway
1830 * a later vma might be split and reinserted earlier while lock dropped.
1832 * The list of nonlinear vmas could be handled more efficiently, using
1833 * a placeholder, but handle it in the same way until a need is shown.
1834 * It is important to search the prio_tree before nonlinear list: a vma
1835 * may become nonlinear and be shifted from prio_tree to nonlinear list
1836 * while the lock is dropped; but never shifted from list to prio_tree.
1838 * In order to make forward progress despite restarting the search,
1839 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1840 * quickly skip it next time around. Since the prio_tree search only
1841 * shows us those vmas affected by unmapping the range in question, we
1842 * can't efficiently keep all vmas in step with mapping->truncate_count:
1843 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1844 * mapping->truncate_count and vma->vm_truncate_count are protected by
1845 * i_mmap_lock.
1847 * In order to make forward progress despite repeatedly restarting some
1848 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1849 * and restart from that address when we reach that vma again. It might
1850 * have been split or merged, shrunk or extended, but never shifted: so
1851 * restart_addr remains valid so long as it remains in the vma's range.
1852 * unmap_mapping_range forces truncate_count to leap over page-aligned
1853 * values so we can save vma's restart_addr in its truncate_count field.
1855 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1857 static void reset_vma_truncate_counts(struct address_space *mapping)
1859 struct vm_area_struct *vma;
1860 struct prio_tree_iter iter;
1862 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1863 vma->vm_truncate_count = 0;
1864 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1865 vma->vm_truncate_count = 0;
1868 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1869 unsigned long start_addr, unsigned long end_addr,
1870 struct zap_details *details)
1872 unsigned long restart_addr;
1873 int need_break;
1876 * files that support invalidating or truncating portions of the
1877 * file from under mmaped areas must have their ->fault function
1878 * return a locked page (and set VM_FAULT_LOCKED in the return).
1879 * This provides synchronisation against concurrent unmapping here.
1882 again:
1883 restart_addr = vma->vm_truncate_count;
1884 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1885 start_addr = restart_addr;
1886 if (start_addr >= end_addr) {
1887 /* Top of vma has been split off since last time */
1888 vma->vm_truncate_count = details->truncate_count;
1889 return 0;
1893 restart_addr = zap_page_range(vma, start_addr,
1894 end_addr - start_addr, details);
1895 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1897 if (restart_addr >= end_addr) {
1898 /* We have now completed this vma: mark it so */
1899 vma->vm_truncate_count = details->truncate_count;
1900 if (!need_break)
1901 return 0;
1902 } else {
1903 /* Note restart_addr in vma's truncate_count field */
1904 vma->vm_truncate_count = restart_addr;
1905 if (!need_break)
1906 goto again;
1909 spin_unlock(details->i_mmap_lock);
1910 cond_resched();
1911 spin_lock(details->i_mmap_lock);
1912 return -EINTR;
1915 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1916 struct zap_details *details)
1918 struct vm_area_struct *vma;
1919 struct prio_tree_iter iter;
1920 pgoff_t vba, vea, zba, zea;
1922 restart:
1923 vma_prio_tree_foreach(vma, &iter, root,
1924 details->first_index, details->last_index) {
1925 /* Skip quickly over those we have already dealt with */
1926 if (vma->vm_truncate_count == details->truncate_count)
1927 continue;
1929 vba = vma->vm_pgoff;
1930 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1931 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1932 zba = details->first_index;
1933 if (zba < vba)
1934 zba = vba;
1935 zea = details->last_index;
1936 if (zea > vea)
1937 zea = vea;
1939 if (unmap_mapping_range_vma(vma,
1940 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1941 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1942 details) < 0)
1943 goto restart;
1947 static inline void unmap_mapping_range_list(struct list_head *head,
1948 struct zap_details *details)
1950 struct vm_area_struct *vma;
1953 * In nonlinear VMAs there is no correspondence between virtual address
1954 * offset and file offset. So we must perform an exhaustive search
1955 * across *all* the pages in each nonlinear VMA, not just the pages
1956 * whose virtual address lies outside the file truncation point.
1958 restart:
1959 list_for_each_entry(vma, head, shared.vm_set.list) {
1960 /* Skip quickly over those we have already dealt with */
1961 if (vma->vm_truncate_count == details->truncate_count)
1962 continue;
1963 details->nonlinear_vma = vma;
1964 if (unmap_mapping_range_vma(vma, vma->vm_start,
1965 vma->vm_end, details) < 0)
1966 goto restart;
1971 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1972 * @mapping: the address space containing mmaps to be unmapped.
1973 * @holebegin: byte in first page to unmap, relative to the start of
1974 * the underlying file. This will be rounded down to a PAGE_SIZE
1975 * boundary. Note that this is different from vmtruncate(), which
1976 * must keep the partial page. In contrast, we must get rid of
1977 * partial pages.
1978 * @holelen: size of prospective hole in bytes. This will be rounded
1979 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1980 * end of the file.
1981 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1982 * but 0 when invalidating pagecache, don't throw away private data.
1984 void unmap_mapping_range(struct address_space *mapping,
1985 loff_t const holebegin, loff_t const holelen, int even_cows)
1987 struct zap_details details;
1988 pgoff_t hba = holebegin >> PAGE_SHIFT;
1989 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1991 /* Check for overflow. */
1992 if (sizeof(holelen) > sizeof(hlen)) {
1993 long long holeend =
1994 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1995 if (holeend & ~(long long)ULONG_MAX)
1996 hlen = ULONG_MAX - hba + 1;
1999 details.check_mapping = even_cows? NULL: mapping;
2000 details.nonlinear_vma = NULL;
2001 details.first_index = hba;
2002 details.last_index = hba + hlen - 1;
2003 if (details.last_index < details.first_index)
2004 details.last_index = ULONG_MAX;
2005 details.i_mmap_lock = &mapping->i_mmap_lock;
2007 spin_lock(&mapping->i_mmap_lock);
2009 /* Protect against endless unmapping loops */
2010 mapping->truncate_count++;
2011 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2012 if (mapping->truncate_count == 0)
2013 reset_vma_truncate_counts(mapping);
2014 mapping->truncate_count++;
2016 details.truncate_count = mapping->truncate_count;
2018 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2019 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2020 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2021 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2022 spin_unlock(&mapping->i_mmap_lock);
2024 EXPORT_SYMBOL(unmap_mapping_range);
2027 * vmtruncate - unmap mappings "freed" by truncate() syscall
2028 * @inode: inode of the file used
2029 * @offset: file offset to start truncating
2031 * NOTE! We have to be ready to update the memory sharing
2032 * between the file and the memory map for a potential last
2033 * incomplete page. Ugly, but necessary.
2035 int vmtruncate(struct inode * inode, loff_t offset)
2037 if (inode->i_size < offset) {
2038 unsigned long limit;
2040 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2041 if (limit != RLIM_INFINITY && offset > limit)
2042 goto out_sig;
2043 if (offset > inode->i_sb->s_maxbytes)
2044 goto out_big;
2045 i_size_write(inode, offset);
2046 } else {
2047 struct address_space *mapping = inode->i_mapping;
2050 * truncation of in-use swapfiles is disallowed - it would
2051 * cause subsequent swapout to scribble on the now-freed
2052 * blocks.
2054 if (IS_SWAPFILE(inode))
2055 return -ETXTBSY;
2056 i_size_write(inode, offset);
2059 * unmap_mapping_range is called twice, first simply for
2060 * efficiency so that truncate_inode_pages does fewer
2061 * single-page unmaps. However after this first call, and
2062 * before truncate_inode_pages finishes, it is possible for
2063 * private pages to be COWed, which remain after
2064 * truncate_inode_pages finishes, hence the second
2065 * unmap_mapping_range call must be made for correctness.
2067 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2068 truncate_inode_pages(mapping, offset);
2069 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2072 if (inode->i_op && inode->i_op->truncate)
2073 inode->i_op->truncate(inode);
2074 return 0;
2076 out_sig:
2077 send_sig(SIGXFSZ, current, 0);
2078 out_big:
2079 return -EFBIG;
2081 EXPORT_SYMBOL(vmtruncate);
2083 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2085 struct address_space *mapping = inode->i_mapping;
2088 * If the underlying filesystem is not going to provide
2089 * a way to truncate a range of blocks (punch a hole) -
2090 * we should return failure right now.
2092 if (!inode->i_op || !inode->i_op->truncate_range)
2093 return -ENOSYS;
2095 mutex_lock(&inode->i_mutex);
2096 down_write(&inode->i_alloc_sem);
2097 unmap_mapping_range(mapping, offset, (end - offset), 1);
2098 truncate_inode_pages_range(mapping, offset, end);
2099 unmap_mapping_range(mapping, offset, (end - offset), 1);
2100 inode->i_op->truncate_range(inode, offset, end);
2101 up_write(&inode->i_alloc_sem);
2102 mutex_unlock(&inode->i_mutex);
2104 return 0;
2108 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2109 * but allow concurrent faults), and pte mapped but not yet locked.
2110 * We return with mmap_sem still held, but pte unmapped and unlocked.
2112 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2113 unsigned long address, pte_t *page_table, pmd_t *pmd,
2114 int write_access, pte_t orig_pte)
2116 spinlock_t *ptl;
2117 struct page *page;
2118 swp_entry_t entry;
2119 pte_t pte;
2120 int ret = 0;
2122 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2123 goto out;
2125 entry = pte_to_swp_entry(orig_pte);
2126 if (is_migration_entry(entry)) {
2127 migration_entry_wait(mm, pmd, address);
2128 goto out;
2130 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2131 page = lookup_swap_cache(entry);
2132 if (!page) {
2133 grab_swap_token(); /* Contend for token _before_ read-in */
2134 page = swapin_readahead(entry,
2135 GFP_HIGHUSER_MOVABLE, vma, address);
2136 if (!page) {
2138 * Back out if somebody else faulted in this pte
2139 * while we released the pte lock.
2141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2142 if (likely(pte_same(*page_table, orig_pte)))
2143 ret = VM_FAULT_OOM;
2144 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2145 goto unlock;
2148 /* Had to read the page from swap area: Major fault */
2149 ret = VM_FAULT_MAJOR;
2150 count_vm_event(PGMAJFAULT);
2153 if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2154 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2155 ret = VM_FAULT_OOM;
2156 goto out;
2159 mark_page_accessed(page);
2160 lock_page(page);
2161 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2164 * Back out if somebody else already faulted in this pte.
2166 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2167 if (unlikely(!pte_same(*page_table, orig_pte)))
2168 goto out_nomap;
2170 if (unlikely(!PageUptodate(page))) {
2171 ret = VM_FAULT_SIGBUS;
2172 goto out_nomap;
2175 /* The page isn't present yet, go ahead with the fault. */
2177 inc_mm_counter(mm, anon_rss);
2178 pte = mk_pte(page, vma->vm_page_prot);
2179 if (write_access && can_share_swap_page(page)) {
2180 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2181 write_access = 0;
2184 flush_icache_page(vma, page);
2185 set_pte_at(mm, address, page_table, pte);
2186 page_add_anon_rmap(page, vma, address);
2188 swap_free(entry);
2189 if (vm_swap_full())
2190 remove_exclusive_swap_page(page);
2191 unlock_page(page);
2193 if (write_access) {
2194 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2195 if (ret & VM_FAULT_ERROR)
2196 ret &= VM_FAULT_ERROR;
2197 goto out;
2200 /* No need to invalidate - it was non-present before */
2201 update_mmu_cache(vma, address, pte);
2202 unlock:
2203 pte_unmap_unlock(page_table, ptl);
2204 out:
2205 return ret;
2206 out_nomap:
2207 mem_cgroup_uncharge_page(page);
2208 pte_unmap_unlock(page_table, ptl);
2209 unlock_page(page);
2210 page_cache_release(page);
2211 return ret;
2215 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2216 * but allow concurrent faults), and pte mapped but not yet locked.
2217 * We return with mmap_sem still held, but pte unmapped and unlocked.
2219 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2220 unsigned long address, pte_t *page_table, pmd_t *pmd,
2221 int write_access)
2223 struct page *page;
2224 spinlock_t *ptl;
2225 pte_t entry;
2227 /* Allocate our own private page. */
2228 pte_unmap(page_table);
2230 if (unlikely(anon_vma_prepare(vma)))
2231 goto oom;
2232 page = alloc_zeroed_user_highpage_movable(vma, address);
2233 if (!page)
2234 goto oom;
2235 __SetPageUptodate(page);
2237 if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2238 goto oom_free_page;
2240 entry = mk_pte(page, vma->vm_page_prot);
2241 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2243 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2244 if (!pte_none(*page_table))
2245 goto release;
2246 inc_mm_counter(mm, anon_rss);
2247 lru_cache_add_active(page);
2248 page_add_new_anon_rmap(page, vma, address);
2249 set_pte_at(mm, address, page_table, entry);
2251 /* No need to invalidate - it was non-present before */
2252 update_mmu_cache(vma, address, entry);
2253 unlock:
2254 pte_unmap_unlock(page_table, ptl);
2255 return 0;
2256 release:
2257 mem_cgroup_uncharge_page(page);
2258 page_cache_release(page);
2259 goto unlock;
2260 oom_free_page:
2261 page_cache_release(page);
2262 oom:
2263 return VM_FAULT_OOM;
2267 * __do_fault() tries to create a new page mapping. It aggressively
2268 * tries to share with existing pages, but makes a separate copy if
2269 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2270 * the next page fault.
2272 * As this is called only for pages that do not currently exist, we
2273 * do not need to flush old virtual caches or the TLB.
2275 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2276 * but allow concurrent faults), and pte neither mapped nor locked.
2277 * We return with mmap_sem still held, but pte unmapped and unlocked.
2279 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2280 unsigned long address, pmd_t *pmd,
2281 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2283 pte_t *page_table;
2284 spinlock_t *ptl;
2285 struct page *page;
2286 pte_t entry;
2287 int anon = 0;
2288 struct page *dirty_page = NULL;
2289 struct vm_fault vmf;
2290 int ret;
2291 int page_mkwrite = 0;
2293 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2294 vmf.pgoff = pgoff;
2295 vmf.flags = flags;
2296 vmf.page = NULL;
2298 ret = vma->vm_ops->fault(vma, &vmf);
2299 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2300 return ret;
2303 * For consistency in subsequent calls, make the faulted page always
2304 * locked.
2306 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2307 lock_page(vmf.page);
2308 else
2309 VM_BUG_ON(!PageLocked(vmf.page));
2312 * Should we do an early C-O-W break?
2314 page = vmf.page;
2315 if (flags & FAULT_FLAG_WRITE) {
2316 if (!(vma->vm_flags & VM_SHARED)) {
2317 anon = 1;
2318 if (unlikely(anon_vma_prepare(vma))) {
2319 ret = VM_FAULT_OOM;
2320 goto out;
2322 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2323 vma, address);
2324 if (!page) {
2325 ret = VM_FAULT_OOM;
2326 goto out;
2328 copy_user_highpage(page, vmf.page, address, vma);
2329 __SetPageUptodate(page);
2330 } else {
2332 * If the page will be shareable, see if the backing
2333 * address space wants to know that the page is about
2334 * to become writable
2336 if (vma->vm_ops->page_mkwrite) {
2337 unlock_page(page);
2338 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2339 ret = VM_FAULT_SIGBUS;
2340 anon = 1; /* no anon but release vmf.page */
2341 goto out_unlocked;
2343 lock_page(page);
2345 * XXX: this is not quite right (racy vs
2346 * invalidate) to unlock and relock the page
2347 * like this, however a better fix requires
2348 * reworking page_mkwrite locking API, which
2349 * is better done later.
2351 if (!page->mapping) {
2352 ret = 0;
2353 anon = 1; /* no anon but release vmf.page */
2354 goto out;
2356 page_mkwrite = 1;
2362 if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2363 ret = VM_FAULT_OOM;
2364 goto out;
2367 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2370 * This silly early PAGE_DIRTY setting removes a race
2371 * due to the bad i386 page protection. But it's valid
2372 * for other architectures too.
2374 * Note that if write_access is true, we either now have
2375 * an exclusive copy of the page, or this is a shared mapping,
2376 * so we can make it writable and dirty to avoid having to
2377 * handle that later.
2379 /* Only go through if we didn't race with anybody else... */
2380 if (likely(pte_same(*page_table, orig_pte))) {
2381 flush_icache_page(vma, page);
2382 entry = mk_pte(page, vma->vm_page_prot);
2383 if (flags & FAULT_FLAG_WRITE)
2384 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2385 set_pte_at(mm, address, page_table, entry);
2386 if (anon) {
2387 inc_mm_counter(mm, anon_rss);
2388 lru_cache_add_active(page);
2389 page_add_new_anon_rmap(page, vma, address);
2390 } else {
2391 inc_mm_counter(mm, file_rss);
2392 page_add_file_rmap(page);
2393 if (flags & FAULT_FLAG_WRITE) {
2394 dirty_page = page;
2395 get_page(dirty_page);
2399 /* no need to invalidate: a not-present page won't be cached */
2400 update_mmu_cache(vma, address, entry);
2401 } else {
2402 mem_cgroup_uncharge_page(page);
2403 if (anon)
2404 page_cache_release(page);
2405 else
2406 anon = 1; /* no anon but release faulted_page */
2409 pte_unmap_unlock(page_table, ptl);
2411 out:
2412 unlock_page(vmf.page);
2413 out_unlocked:
2414 if (anon)
2415 page_cache_release(vmf.page);
2416 else if (dirty_page) {
2417 if (vma->vm_file)
2418 file_update_time(vma->vm_file);
2420 set_page_dirty_balance(dirty_page, page_mkwrite);
2421 put_page(dirty_page);
2424 return ret;
2427 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2428 unsigned long address, pte_t *page_table, pmd_t *pmd,
2429 int write_access, pte_t orig_pte)
2431 pgoff_t pgoff = (((address & PAGE_MASK)
2432 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2433 unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2435 pte_unmap(page_table);
2436 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2441 * do_no_pfn() tries to create a new page mapping for a page without
2442 * a struct_page backing it
2444 * As this is called only for pages that do not currently exist, we
2445 * do not need to flush old virtual caches or the TLB.
2447 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2448 * but allow concurrent faults), and pte mapped but not yet locked.
2449 * We return with mmap_sem still held, but pte unmapped and unlocked.
2451 * It is expected that the ->nopfn handler always returns the same pfn
2452 * for a given virtual mapping.
2454 * Mark this `noinline' to prevent it from bloating the main pagefault code.
2456 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2457 unsigned long address, pte_t *page_table, pmd_t *pmd,
2458 int write_access)
2460 spinlock_t *ptl;
2461 pte_t entry;
2462 unsigned long pfn;
2464 pte_unmap(page_table);
2465 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2466 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2468 pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2470 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2472 if (unlikely(pfn == NOPFN_OOM))
2473 return VM_FAULT_OOM;
2474 else if (unlikely(pfn == NOPFN_SIGBUS))
2475 return VM_FAULT_SIGBUS;
2476 else if (unlikely(pfn == NOPFN_REFAULT))
2477 return 0;
2479 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2481 /* Only go through if we didn't race with anybody else... */
2482 if (pte_none(*page_table)) {
2483 entry = pfn_pte(pfn, vma->vm_page_prot);
2484 if (write_access)
2485 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2486 set_pte_at(mm, address, page_table, entry);
2488 pte_unmap_unlock(page_table, ptl);
2489 return 0;
2493 * Fault of a previously existing named mapping. Repopulate the pte
2494 * from the encoded file_pte if possible. This enables swappable
2495 * nonlinear vmas.
2497 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2498 * but allow concurrent faults), and pte mapped but not yet locked.
2499 * We return with mmap_sem still held, but pte unmapped and unlocked.
2501 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2502 unsigned long address, pte_t *page_table, pmd_t *pmd,
2503 int write_access, pte_t orig_pte)
2505 unsigned int flags = FAULT_FLAG_NONLINEAR |
2506 (write_access ? FAULT_FLAG_WRITE : 0);
2507 pgoff_t pgoff;
2509 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2510 return 0;
2512 if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2513 !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2515 * Page table corrupted: show pte and kill process.
2517 print_bad_pte(vma, orig_pte, address);
2518 return VM_FAULT_OOM;
2521 pgoff = pte_to_pgoff(orig_pte);
2522 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2526 * These routines also need to handle stuff like marking pages dirty
2527 * and/or accessed for architectures that don't do it in hardware (most
2528 * RISC architectures). The early dirtying is also good on the i386.
2530 * There is also a hook called "update_mmu_cache()" that architectures
2531 * with external mmu caches can use to update those (ie the Sparc or
2532 * PowerPC hashed page tables that act as extended TLBs).
2534 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2535 * but allow concurrent faults), and pte mapped but not yet locked.
2536 * We return with mmap_sem still held, but pte unmapped and unlocked.
2538 static inline int handle_pte_fault(struct mm_struct *mm,
2539 struct vm_area_struct *vma, unsigned long address,
2540 pte_t *pte, pmd_t *pmd, int write_access)
2542 pte_t entry;
2543 spinlock_t *ptl;
2545 entry = *pte;
2546 if (!pte_present(entry)) {
2547 if (pte_none(entry)) {
2548 if (vma->vm_ops) {
2549 if (likely(vma->vm_ops->fault))
2550 return do_linear_fault(mm, vma, address,
2551 pte, pmd, write_access, entry);
2552 if (unlikely(vma->vm_ops->nopfn))
2553 return do_no_pfn(mm, vma, address, pte,
2554 pmd, write_access);
2556 return do_anonymous_page(mm, vma, address,
2557 pte, pmd, write_access);
2559 if (pte_file(entry))
2560 return do_nonlinear_fault(mm, vma, address,
2561 pte, pmd, write_access, entry);
2562 return do_swap_page(mm, vma, address,
2563 pte, pmd, write_access, entry);
2566 ptl = pte_lockptr(mm, pmd);
2567 spin_lock(ptl);
2568 if (unlikely(!pte_same(*pte, entry)))
2569 goto unlock;
2570 if (write_access) {
2571 if (!pte_write(entry))
2572 return do_wp_page(mm, vma, address,
2573 pte, pmd, ptl, entry);
2574 entry = pte_mkdirty(entry);
2576 entry = pte_mkyoung(entry);
2577 if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2578 update_mmu_cache(vma, address, entry);
2579 } else {
2581 * This is needed only for protection faults but the arch code
2582 * is not yet telling us if this is a protection fault or not.
2583 * This still avoids useless tlb flushes for .text page faults
2584 * with threads.
2586 if (write_access)
2587 flush_tlb_page(vma, address);
2589 unlock:
2590 pte_unmap_unlock(pte, ptl);
2591 return 0;
2595 * By the time we get here, we already hold the mm semaphore
2597 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2598 unsigned long address, int write_access)
2600 pgd_t *pgd;
2601 pud_t *pud;
2602 pmd_t *pmd;
2603 pte_t *pte;
2605 __set_current_state(TASK_RUNNING);
2607 count_vm_event(PGFAULT);
2609 if (unlikely(is_vm_hugetlb_page(vma)))
2610 return hugetlb_fault(mm, vma, address, write_access);
2612 pgd = pgd_offset(mm, address);
2613 pud = pud_alloc(mm, pgd, address);
2614 if (!pud)
2615 return VM_FAULT_OOM;
2616 pmd = pmd_alloc(mm, pud, address);
2617 if (!pmd)
2618 return VM_FAULT_OOM;
2619 pte = pte_alloc_map(mm, pmd, address);
2620 if (!pte)
2621 return VM_FAULT_OOM;
2623 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2626 #ifndef __PAGETABLE_PUD_FOLDED
2628 * Allocate page upper directory.
2629 * We've already handled the fast-path in-line.
2631 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2633 pud_t *new = pud_alloc_one(mm, address);
2634 if (!new)
2635 return -ENOMEM;
2637 smp_wmb(); /* See comment in __pte_alloc */
2639 spin_lock(&mm->page_table_lock);
2640 if (pgd_present(*pgd)) /* Another has populated it */
2641 pud_free(mm, new);
2642 else
2643 pgd_populate(mm, pgd, new);
2644 spin_unlock(&mm->page_table_lock);
2645 return 0;
2647 #endif /* __PAGETABLE_PUD_FOLDED */
2649 #ifndef __PAGETABLE_PMD_FOLDED
2651 * Allocate page middle directory.
2652 * We've already handled the fast-path in-line.
2654 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2656 pmd_t *new = pmd_alloc_one(mm, address);
2657 if (!new)
2658 return -ENOMEM;
2660 smp_wmb(); /* See comment in __pte_alloc */
2662 spin_lock(&mm->page_table_lock);
2663 #ifndef __ARCH_HAS_4LEVEL_HACK
2664 if (pud_present(*pud)) /* Another has populated it */
2665 pmd_free(mm, new);
2666 else
2667 pud_populate(mm, pud, new);
2668 #else
2669 if (pgd_present(*pud)) /* Another has populated it */
2670 pmd_free(mm, new);
2671 else
2672 pgd_populate(mm, pud, new);
2673 #endif /* __ARCH_HAS_4LEVEL_HACK */
2674 spin_unlock(&mm->page_table_lock);
2675 return 0;
2677 #endif /* __PAGETABLE_PMD_FOLDED */
2679 int make_pages_present(unsigned long addr, unsigned long end)
2681 int ret, len, write;
2682 struct vm_area_struct * vma;
2684 vma = find_vma(current->mm, addr);
2685 if (!vma)
2686 return -1;
2687 write = (vma->vm_flags & VM_WRITE) != 0;
2688 BUG_ON(addr >= end);
2689 BUG_ON(end > vma->vm_end);
2690 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2691 ret = get_user_pages(current, current->mm, addr,
2692 len, write, 0, NULL, NULL);
2693 if (ret < 0)
2694 return ret;
2695 return ret == len ? 0 : -1;
2698 #if !defined(__HAVE_ARCH_GATE_AREA)
2700 #if defined(AT_SYSINFO_EHDR)
2701 static struct vm_area_struct gate_vma;
2703 static int __init gate_vma_init(void)
2705 gate_vma.vm_mm = NULL;
2706 gate_vma.vm_start = FIXADDR_USER_START;
2707 gate_vma.vm_end = FIXADDR_USER_END;
2708 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2709 gate_vma.vm_page_prot = __P101;
2711 * Make sure the vDSO gets into every core dump.
2712 * Dumping its contents makes post-mortem fully interpretable later
2713 * without matching up the same kernel and hardware config to see
2714 * what PC values meant.
2716 gate_vma.vm_flags |= VM_ALWAYSDUMP;
2717 return 0;
2719 __initcall(gate_vma_init);
2720 #endif
2722 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2724 #ifdef AT_SYSINFO_EHDR
2725 return &gate_vma;
2726 #else
2727 return NULL;
2728 #endif
2731 int in_gate_area_no_task(unsigned long addr)
2733 #ifdef AT_SYSINFO_EHDR
2734 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2735 return 1;
2736 #endif
2737 return 0;
2740 #endif /* __HAVE_ARCH_GATE_AREA */
2743 * Access another process' address space.
2744 * Source/target buffer must be kernel space,
2745 * Do not walk the page table directly, use get_user_pages
2747 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2749 struct mm_struct *mm;
2750 struct vm_area_struct *vma;
2751 struct page *page;
2752 void *old_buf = buf;
2754 mm = get_task_mm(tsk);
2755 if (!mm)
2756 return 0;
2758 down_read(&mm->mmap_sem);
2759 /* ignore errors, just check how much was successfully transferred */
2760 while (len) {
2761 int bytes, ret, offset;
2762 void *maddr;
2764 ret = get_user_pages(tsk, mm, addr, 1,
2765 write, 1, &page, &vma);
2766 if (ret <= 0)
2767 break;
2769 bytes = len;
2770 offset = addr & (PAGE_SIZE-1);
2771 if (bytes > PAGE_SIZE-offset)
2772 bytes = PAGE_SIZE-offset;
2774 maddr = kmap(page);
2775 if (write) {
2776 copy_to_user_page(vma, page, addr,
2777 maddr + offset, buf, bytes);
2778 set_page_dirty_lock(page);
2779 } else {
2780 copy_from_user_page(vma, page, addr,
2781 buf, maddr + offset, bytes);
2783 kunmap(page);
2784 page_cache_release(page);
2785 len -= bytes;
2786 buf += bytes;
2787 addr += bytes;
2789 up_read(&mm->mmap_sem);
2790 mmput(mm);
2792 return buf - old_buf;
2796 * Print the name of a VMA.
2798 void print_vma_addr(char *prefix, unsigned long ip)
2800 struct mm_struct *mm = current->mm;
2801 struct vm_area_struct *vma;
2804 * Do not print if we are in atomic
2805 * contexts (in exception stacks, etc.):
2807 if (preempt_count())
2808 return;
2810 down_read(&mm->mmap_sem);
2811 vma = find_vma(mm, ip);
2812 if (vma && vma->vm_file) {
2813 struct file *f = vma->vm_file;
2814 char *buf = (char *)__get_free_page(GFP_KERNEL);
2815 if (buf) {
2816 char *p, *s;
2818 p = d_path(&f->f_path, buf, PAGE_SIZE);
2819 if (IS_ERR(p))
2820 p = "?";
2821 s = strrchr(p, '/');
2822 if (s)
2823 p = s+1;
2824 printk("%s%s[%lx+%lx]", prefix, p,
2825 vma->vm_start,
2826 vma->vm_end - vma->vm_start);
2827 free_page((unsigned long)buf);
2830 up_read(&current->mm->mmap_sem);