Staging: vt6655: Integrate drivers/staging/vt6655 into build system.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / memory.c
blobd5d1653d60a6be51a4ba74de53fd997d8b0bc94c
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>
54 #include <linux/mmu_notifier.h>
55 #include <linux/kallsyms.h>
56 #include <linux/swapops.h>
57 #include <linux/elf.h>
59 #include <asm/pgalloc.h>
60 #include <asm/uaccess.h>
61 #include <asm/tlb.h>
62 #include <asm/tlbflush.h>
63 #include <asm/pgtable.h>
65 #include "internal.h"
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr;
70 struct page *mem_map;
72 EXPORT_SYMBOL(max_mapnr);
73 EXPORT_SYMBOL(mem_map);
74 #endif
76 unsigned long num_physpages;
78 * A number of key systems in x86 including ioremap() rely on the assumption
79 * that high_memory defines the upper bound on direct map memory, then end
80 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
81 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
82 * and ZONE_HIGHMEM.
84 void * high_memory;
86 EXPORT_SYMBOL(num_physpages);
87 EXPORT_SYMBOL(high_memory);
90 * Randomize the address space (stacks, mmaps, brk, etc.).
92 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93 * as ancient (libc5 based) binaries can segfault. )
95 int randomize_va_space __read_mostly =
96 #ifdef CONFIG_COMPAT_BRK
98 #else
100 #endif
102 static int __init disable_randmaps(char *s)
104 randomize_va_space = 0;
105 return 1;
107 __setup("norandmaps", disable_randmaps);
111 * If a p?d_bad entry is found while walking page tables, report
112 * the error, before resetting entry to p?d_none. Usually (but
113 * very seldom) called out from the p?d_none_or_clear_bad macros.
116 void pgd_clear_bad(pgd_t *pgd)
118 pgd_ERROR(*pgd);
119 pgd_clear(pgd);
122 void pud_clear_bad(pud_t *pud)
124 pud_ERROR(*pud);
125 pud_clear(pud);
128 void pmd_clear_bad(pmd_t *pmd)
130 pmd_ERROR(*pmd);
131 pmd_clear(pmd);
135 * Note: this doesn't free the actual pages themselves. That
136 * has been handled earlier when unmapping all the memory regions.
138 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
140 pgtable_t token = pmd_pgtable(*pmd);
141 pmd_clear(pmd);
142 pte_free_tlb(tlb, token);
143 tlb->mm->nr_ptes--;
146 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
147 unsigned long addr, unsigned long end,
148 unsigned long floor, unsigned long ceiling)
150 pmd_t *pmd;
151 unsigned long next;
152 unsigned long start;
154 start = addr;
155 pmd = pmd_offset(pud, addr);
156 do {
157 next = pmd_addr_end(addr, end);
158 if (pmd_none_or_clear_bad(pmd))
159 continue;
160 free_pte_range(tlb, pmd);
161 } while (pmd++, addr = next, addr != end);
163 start &= PUD_MASK;
164 if (start < floor)
165 return;
166 if (ceiling) {
167 ceiling &= PUD_MASK;
168 if (!ceiling)
169 return;
171 if (end - 1 > ceiling - 1)
172 return;
174 pmd = pmd_offset(pud, start);
175 pud_clear(pud);
176 pmd_free_tlb(tlb, pmd);
179 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
180 unsigned long addr, unsigned long end,
181 unsigned long floor, unsigned long ceiling)
183 pud_t *pud;
184 unsigned long next;
185 unsigned long start;
187 start = addr;
188 pud = pud_offset(pgd, addr);
189 do {
190 next = pud_addr_end(addr, end);
191 if (pud_none_or_clear_bad(pud))
192 continue;
193 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
194 } while (pud++, addr = next, addr != end);
196 start &= PGDIR_MASK;
197 if (start < floor)
198 return;
199 if (ceiling) {
200 ceiling &= PGDIR_MASK;
201 if (!ceiling)
202 return;
204 if (end - 1 > ceiling - 1)
205 return;
207 pud = pud_offset(pgd, start);
208 pgd_clear(pgd);
209 pud_free_tlb(tlb, pud);
213 * This function frees user-level page tables of a process.
215 * Must be called with pagetable lock held.
217 void free_pgd_range(struct mmu_gather *tlb,
218 unsigned long addr, unsigned long end,
219 unsigned long floor, unsigned long ceiling)
221 pgd_t *pgd;
222 unsigned long next;
223 unsigned long start;
226 * The next few lines have given us lots of grief...
228 * Why are we testing PMD* at this top level? Because often
229 * there will be no work to do at all, and we'd prefer not to
230 * go all the way down to the bottom just to discover that.
232 * Why all these "- 1"s? Because 0 represents both the bottom
233 * of the address space and the top of it (using -1 for the
234 * top wouldn't help much: the masks would do the wrong thing).
235 * The rule is that addr 0 and floor 0 refer to the bottom of
236 * the address space, but end 0 and ceiling 0 refer to the top
237 * Comparisons need to use "end - 1" and "ceiling - 1" (though
238 * that end 0 case should be mythical).
240 * Wherever addr is brought up or ceiling brought down, we must
241 * be careful to reject "the opposite 0" before it confuses the
242 * subsequent tests. But what about where end is brought down
243 * by PMD_SIZE below? no, end can't go down to 0 there.
245 * Whereas we round start (addr) and ceiling down, by different
246 * masks at different levels, in order to test whether a table
247 * now has no other vmas using it, so can be freed, we don't
248 * bother to round floor or end up - the tests don't need that.
251 addr &= PMD_MASK;
252 if (addr < floor) {
253 addr += PMD_SIZE;
254 if (!addr)
255 return;
257 if (ceiling) {
258 ceiling &= PMD_MASK;
259 if (!ceiling)
260 return;
262 if (end - 1 > ceiling - 1)
263 end -= PMD_SIZE;
264 if (addr > end - 1)
265 return;
267 start = addr;
268 pgd = pgd_offset(tlb->mm, addr);
269 do {
270 next = pgd_addr_end(addr, end);
271 if (pgd_none_or_clear_bad(pgd))
272 continue;
273 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
274 } while (pgd++, addr = next, addr != end);
277 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
278 unsigned long floor, unsigned long ceiling)
280 while (vma) {
281 struct vm_area_struct *next = vma->vm_next;
282 unsigned long addr = vma->vm_start;
285 * Hide vma from rmap and vmtruncate before freeing pgtables
287 anon_vma_unlink(vma);
288 unlink_file_vma(vma);
290 if (is_vm_hugetlb_page(vma)) {
291 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
292 floor, next? next->vm_start: ceiling);
293 } else {
295 * Optimization: gather nearby vmas into one call down
297 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
298 && !is_vm_hugetlb_page(next)) {
299 vma = next;
300 next = vma->vm_next;
301 anon_vma_unlink(vma);
302 unlink_file_vma(vma);
304 free_pgd_range(tlb, addr, vma->vm_end,
305 floor, next? next->vm_start: ceiling);
307 vma = next;
311 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
313 pgtable_t new = pte_alloc_one(mm, address);
314 if (!new)
315 return -ENOMEM;
318 * Ensure all pte setup (eg. pte page lock and page clearing) are
319 * visible before the pte is made visible to other CPUs by being
320 * put into page tables.
322 * The other side of the story is the pointer chasing in the page
323 * table walking code (when walking the page table without locking;
324 * ie. most of the time). Fortunately, these data accesses consist
325 * of a chain of data-dependent loads, meaning most CPUs (alpha
326 * being the notable exception) will already guarantee loads are
327 * seen in-order. See the alpha page table accessors for the
328 * smp_read_barrier_depends() barriers in page table walking code.
330 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
332 spin_lock(&mm->page_table_lock);
333 if (!pmd_present(*pmd)) { /* Has another populated it ? */
334 mm->nr_ptes++;
335 pmd_populate(mm, pmd, new);
336 new = NULL;
338 spin_unlock(&mm->page_table_lock);
339 if (new)
340 pte_free(mm, new);
341 return 0;
344 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
346 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
347 if (!new)
348 return -ENOMEM;
350 smp_wmb(); /* See comment in __pte_alloc */
352 spin_lock(&init_mm.page_table_lock);
353 if (!pmd_present(*pmd)) { /* Has another populated it ? */
354 pmd_populate_kernel(&init_mm, pmd, new);
355 new = NULL;
357 spin_unlock(&init_mm.page_table_lock);
358 if (new)
359 pte_free_kernel(&init_mm, new);
360 return 0;
363 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
365 if (file_rss)
366 add_mm_counter(mm, file_rss, file_rss);
367 if (anon_rss)
368 add_mm_counter(mm, anon_rss, anon_rss);
372 * This function is called to print an error when a bad pte
373 * is found. For example, we might have a PFN-mapped pte in
374 * a region that doesn't allow it.
376 * The calling function must still handle the error.
378 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
379 pte_t pte, struct page *page)
381 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
382 pud_t *pud = pud_offset(pgd, addr);
383 pmd_t *pmd = pmd_offset(pud, addr);
384 struct address_space *mapping;
385 pgoff_t index;
386 static unsigned long resume;
387 static unsigned long nr_shown;
388 static unsigned long nr_unshown;
391 * Allow a burst of 60 reports, then keep quiet for that minute;
392 * or allow a steady drip of one report per second.
394 if (nr_shown == 60) {
395 if (time_before(jiffies, resume)) {
396 nr_unshown++;
397 return;
399 if (nr_unshown) {
400 printk(KERN_ALERT
401 "BUG: Bad page map: %lu messages suppressed\n",
402 nr_unshown);
403 nr_unshown = 0;
405 nr_shown = 0;
407 if (nr_shown++ == 0)
408 resume = jiffies + 60 * HZ;
410 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
411 index = linear_page_index(vma, addr);
413 printk(KERN_ALERT
414 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
415 current->comm,
416 (long long)pte_val(pte), (long long)pmd_val(*pmd));
417 if (page) {
418 printk(KERN_ALERT
419 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
420 page, (void *)page->flags, page_count(page),
421 page_mapcount(page), page->mapping, page->index);
423 printk(KERN_ALERT
424 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
425 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
427 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
429 if (vma->vm_ops)
430 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
431 (unsigned long)vma->vm_ops->fault);
432 if (vma->vm_file && vma->vm_file->f_op)
433 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
434 (unsigned long)vma->vm_file->f_op->mmap);
435 dump_stack();
436 add_taint(TAINT_BAD_PAGE);
439 static inline int is_cow_mapping(unsigned int flags)
441 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
445 * vm_normal_page -- This function gets the "struct page" associated with a pte.
447 * "Special" mappings do not wish to be associated with a "struct page" (either
448 * it doesn't exist, or it exists but they don't want to touch it). In this
449 * case, NULL is returned here. "Normal" mappings do have a struct page.
451 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
452 * pte bit, in which case this function is trivial. Secondly, an architecture
453 * may not have a spare pte bit, which requires a more complicated scheme,
454 * described below.
456 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
457 * special mapping (even if there are underlying and valid "struct pages").
458 * COWed pages of a VM_PFNMAP are always normal.
460 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
461 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
462 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
463 * mapping will always honor the rule
465 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
467 * And for normal mappings this is false.
469 * This restricts such mappings to be a linear translation from virtual address
470 * to pfn. To get around this restriction, we allow arbitrary mappings so long
471 * as the vma is not a COW mapping; in that case, we know that all ptes are
472 * special (because none can have been COWed).
475 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
477 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
478 * page" backing, however the difference is that _all_ pages with a struct
479 * page (that is, those where pfn_valid is true) are refcounted and considered
480 * normal pages by the VM. The disadvantage is that pages are refcounted
481 * (which can be slower and simply not an option for some PFNMAP users). The
482 * advantage is that we don't have to follow the strict linearity rule of
483 * PFNMAP mappings in order to support COWable mappings.
486 #ifdef __HAVE_ARCH_PTE_SPECIAL
487 # define HAVE_PTE_SPECIAL 1
488 #else
489 # define HAVE_PTE_SPECIAL 0
490 #endif
491 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
492 pte_t pte)
494 unsigned long pfn = pte_pfn(pte);
496 if (HAVE_PTE_SPECIAL) {
497 if (likely(!pte_special(pte)))
498 goto check_pfn;
499 if (!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)))
500 print_bad_pte(vma, addr, pte, NULL);
501 return NULL;
504 /* !HAVE_PTE_SPECIAL case follows: */
506 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
507 if (vma->vm_flags & VM_MIXEDMAP) {
508 if (!pfn_valid(pfn))
509 return NULL;
510 goto out;
511 } else {
512 unsigned long off;
513 off = (addr - vma->vm_start) >> PAGE_SHIFT;
514 if (pfn == vma->vm_pgoff + off)
515 return NULL;
516 if (!is_cow_mapping(vma->vm_flags))
517 return NULL;
521 check_pfn:
522 if (unlikely(pfn > highest_memmap_pfn)) {
523 print_bad_pte(vma, addr, pte, NULL);
524 return NULL;
528 * NOTE! We still have PageReserved() pages in the page tables.
529 * eg. VDSO mappings can cause them to exist.
531 out:
532 return pfn_to_page(pfn);
536 * copy one vm_area from one task to the other. Assumes the page tables
537 * already present in the new task to be cleared in the whole range
538 * covered by this vma.
541 static inline void
542 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
543 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
544 unsigned long addr, int *rss)
546 unsigned long vm_flags = vma->vm_flags;
547 pte_t pte = *src_pte;
548 struct page *page;
550 /* pte contains position in swap or file, so copy. */
551 if (unlikely(!pte_present(pte))) {
552 if (!pte_file(pte)) {
553 swp_entry_t entry = pte_to_swp_entry(pte);
555 swap_duplicate(entry);
556 /* make sure dst_mm is on swapoff's mmlist. */
557 if (unlikely(list_empty(&dst_mm->mmlist))) {
558 spin_lock(&mmlist_lock);
559 if (list_empty(&dst_mm->mmlist))
560 list_add(&dst_mm->mmlist,
561 &src_mm->mmlist);
562 spin_unlock(&mmlist_lock);
564 if (is_write_migration_entry(entry) &&
565 is_cow_mapping(vm_flags)) {
567 * COW mappings require pages in both parent
568 * and child to be set to read.
570 make_migration_entry_read(&entry);
571 pte = swp_entry_to_pte(entry);
572 set_pte_at(src_mm, addr, src_pte, pte);
575 goto out_set_pte;
579 * If it's a COW mapping, write protect it both
580 * in the parent and the child
582 if (is_cow_mapping(vm_flags)) {
583 ptep_set_wrprotect(src_mm, addr, src_pte);
584 pte = pte_wrprotect(pte);
588 * If it's a shared mapping, mark it clean in
589 * the child
591 if (vm_flags & VM_SHARED)
592 pte = pte_mkclean(pte);
593 pte = pte_mkold(pte);
595 page = vm_normal_page(vma, addr, pte);
596 if (page) {
597 get_page(page);
598 page_dup_rmap(page, vma, addr);
599 rss[!!PageAnon(page)]++;
602 out_set_pte:
603 set_pte_at(dst_mm, addr, dst_pte, pte);
606 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
607 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
608 unsigned long addr, unsigned long end)
610 pte_t *src_pte, *dst_pte;
611 spinlock_t *src_ptl, *dst_ptl;
612 int progress = 0;
613 int rss[2];
615 again:
616 rss[1] = rss[0] = 0;
617 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
618 if (!dst_pte)
619 return -ENOMEM;
620 src_pte = pte_offset_map_nested(src_pmd, addr);
621 src_ptl = pte_lockptr(src_mm, src_pmd);
622 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
623 arch_enter_lazy_mmu_mode();
625 do {
627 * We are holding two locks at this point - either of them
628 * could generate latencies in another task on another CPU.
630 if (progress >= 32) {
631 progress = 0;
632 if (need_resched() ||
633 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
634 break;
636 if (pte_none(*src_pte)) {
637 progress++;
638 continue;
640 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
641 progress += 8;
642 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
644 arch_leave_lazy_mmu_mode();
645 spin_unlock(src_ptl);
646 pte_unmap_nested(src_pte - 1);
647 add_mm_rss(dst_mm, rss[0], rss[1]);
648 pte_unmap_unlock(dst_pte - 1, dst_ptl);
649 cond_resched();
650 if (addr != end)
651 goto again;
652 return 0;
655 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
656 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
657 unsigned long addr, unsigned long end)
659 pmd_t *src_pmd, *dst_pmd;
660 unsigned long next;
662 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
663 if (!dst_pmd)
664 return -ENOMEM;
665 src_pmd = pmd_offset(src_pud, addr);
666 do {
667 next = pmd_addr_end(addr, end);
668 if (pmd_none_or_clear_bad(src_pmd))
669 continue;
670 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
671 vma, addr, next))
672 return -ENOMEM;
673 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
674 return 0;
677 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
678 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
679 unsigned long addr, unsigned long end)
681 pud_t *src_pud, *dst_pud;
682 unsigned long next;
684 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
685 if (!dst_pud)
686 return -ENOMEM;
687 src_pud = pud_offset(src_pgd, addr);
688 do {
689 next = pud_addr_end(addr, end);
690 if (pud_none_or_clear_bad(src_pud))
691 continue;
692 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
693 vma, addr, next))
694 return -ENOMEM;
695 } while (dst_pud++, src_pud++, addr = next, addr != end);
696 return 0;
699 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700 struct vm_area_struct *vma)
702 pgd_t *src_pgd, *dst_pgd;
703 unsigned long next;
704 unsigned long addr = vma->vm_start;
705 unsigned long end = vma->vm_end;
706 int ret;
709 * Don't copy ptes where a page fault will fill them correctly.
710 * Fork becomes much lighter when there are big shared or private
711 * readonly mappings. The tradeoff is that copy_page_range is more
712 * efficient than faulting.
714 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
715 if (!vma->anon_vma)
716 return 0;
719 if (is_vm_hugetlb_page(vma))
720 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
722 if (unlikely(is_pfn_mapping(vma))) {
724 * We do not free on error cases below as remove_vma
725 * gets called on error from higher level routine
727 ret = track_pfn_vma_copy(vma);
728 if (ret)
729 return ret;
733 * We need to invalidate the secondary MMU mappings only when
734 * there could be a permission downgrade on the ptes of the
735 * parent mm. And a permission downgrade will only happen if
736 * is_cow_mapping() returns true.
738 if (is_cow_mapping(vma->vm_flags))
739 mmu_notifier_invalidate_range_start(src_mm, addr, end);
741 ret = 0;
742 dst_pgd = pgd_offset(dst_mm, addr);
743 src_pgd = pgd_offset(src_mm, addr);
744 do {
745 next = pgd_addr_end(addr, end);
746 if (pgd_none_or_clear_bad(src_pgd))
747 continue;
748 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
749 vma, addr, next))) {
750 ret = -ENOMEM;
751 break;
753 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
755 if (is_cow_mapping(vma->vm_flags))
756 mmu_notifier_invalidate_range_end(src_mm,
757 vma->vm_start, end);
758 return ret;
761 static unsigned long zap_pte_range(struct mmu_gather *tlb,
762 struct vm_area_struct *vma, pmd_t *pmd,
763 unsigned long addr, unsigned long end,
764 long *zap_work, struct zap_details *details)
766 struct mm_struct *mm = tlb->mm;
767 pte_t *pte;
768 spinlock_t *ptl;
769 int file_rss = 0;
770 int anon_rss = 0;
772 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
773 arch_enter_lazy_mmu_mode();
774 do {
775 pte_t ptent = *pte;
776 if (pte_none(ptent)) {
777 (*zap_work)--;
778 continue;
781 (*zap_work) -= PAGE_SIZE;
783 if (pte_present(ptent)) {
784 struct page *page;
786 page = vm_normal_page(vma, addr, ptent);
787 if (unlikely(details) && page) {
789 * unmap_shared_mapping_pages() wants to
790 * invalidate cache without truncating:
791 * unmap shared but keep private pages.
793 if (details->check_mapping &&
794 details->check_mapping != page->mapping)
795 continue;
797 * Each page->index must be checked when
798 * invalidating or truncating nonlinear.
800 if (details->nonlinear_vma &&
801 (page->index < details->first_index ||
802 page->index > details->last_index))
803 continue;
805 ptent = ptep_get_and_clear_full(mm, addr, pte,
806 tlb->fullmm);
807 tlb_remove_tlb_entry(tlb, pte, addr);
808 if (unlikely(!page))
809 continue;
810 if (unlikely(details) && details->nonlinear_vma
811 && linear_page_index(details->nonlinear_vma,
812 addr) != page->index)
813 set_pte_at(mm, addr, pte,
814 pgoff_to_pte(page->index));
815 if (PageAnon(page))
816 anon_rss--;
817 else {
818 if (pte_dirty(ptent))
819 set_page_dirty(page);
820 if (pte_young(ptent) &&
821 likely(!VM_SequentialReadHint(vma)))
822 mark_page_accessed(page);
823 file_rss--;
825 page_remove_rmap(page);
826 if (unlikely(page_mapcount(page) < 0))
827 print_bad_pte(vma, addr, ptent, page);
828 tlb_remove_page(tlb, page);
829 continue;
832 * If details->check_mapping, we leave swap entries;
833 * if details->nonlinear_vma, we leave file entries.
835 if (unlikely(details))
836 continue;
837 if (pte_file(ptent)) {
838 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
839 print_bad_pte(vma, addr, ptent, NULL);
840 } else if
841 (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
842 print_bad_pte(vma, addr, ptent, NULL);
843 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
844 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
846 add_mm_rss(mm, file_rss, anon_rss);
847 arch_leave_lazy_mmu_mode();
848 pte_unmap_unlock(pte - 1, ptl);
850 return addr;
853 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
854 struct vm_area_struct *vma, pud_t *pud,
855 unsigned long addr, unsigned long end,
856 long *zap_work, struct zap_details *details)
858 pmd_t *pmd;
859 unsigned long next;
861 pmd = pmd_offset(pud, addr);
862 do {
863 next = pmd_addr_end(addr, end);
864 if (pmd_none_or_clear_bad(pmd)) {
865 (*zap_work)--;
866 continue;
868 next = zap_pte_range(tlb, vma, pmd, addr, next,
869 zap_work, details);
870 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
872 return addr;
875 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
876 struct vm_area_struct *vma, pgd_t *pgd,
877 unsigned long addr, unsigned long end,
878 long *zap_work, struct zap_details *details)
880 pud_t *pud;
881 unsigned long next;
883 pud = pud_offset(pgd, addr);
884 do {
885 next = pud_addr_end(addr, end);
886 if (pud_none_or_clear_bad(pud)) {
887 (*zap_work)--;
888 continue;
890 next = zap_pmd_range(tlb, vma, pud, addr, next,
891 zap_work, details);
892 } while (pud++, addr = next, (addr != end && *zap_work > 0));
894 return addr;
897 static unsigned long unmap_page_range(struct mmu_gather *tlb,
898 struct vm_area_struct *vma,
899 unsigned long addr, unsigned long end,
900 long *zap_work, struct zap_details *details)
902 pgd_t *pgd;
903 unsigned long next;
905 if (details && !details->check_mapping && !details->nonlinear_vma)
906 details = NULL;
908 BUG_ON(addr >= end);
909 tlb_start_vma(tlb, vma);
910 pgd = pgd_offset(vma->vm_mm, addr);
911 do {
912 next = pgd_addr_end(addr, end);
913 if (pgd_none_or_clear_bad(pgd)) {
914 (*zap_work)--;
915 continue;
917 next = zap_pud_range(tlb, vma, pgd, addr, next,
918 zap_work, details);
919 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
920 tlb_end_vma(tlb, vma);
922 return addr;
925 #ifdef CONFIG_PREEMPT
926 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
927 #else
928 /* No preempt: go for improved straight-line efficiency */
929 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
930 #endif
933 * unmap_vmas - unmap a range of memory covered by a list of vma's
934 * @tlbp: address of the caller's struct mmu_gather
935 * @vma: the starting vma
936 * @start_addr: virtual address at which to start unmapping
937 * @end_addr: virtual address at which to end unmapping
938 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
939 * @details: details of nonlinear truncation or shared cache invalidation
941 * Returns the end address of the unmapping (restart addr if interrupted).
943 * Unmap all pages in the vma list.
945 * We aim to not hold locks for too long (for scheduling latency reasons).
946 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
947 * return the ending mmu_gather to the caller.
949 * Only addresses between `start' and `end' will be unmapped.
951 * The VMA list must be sorted in ascending virtual address order.
953 * unmap_vmas() assumes that the caller will flush the whole unmapped address
954 * range after unmap_vmas() returns. So the only responsibility here is to
955 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
956 * drops the lock and schedules.
958 unsigned long unmap_vmas(struct mmu_gather **tlbp,
959 struct vm_area_struct *vma, unsigned long start_addr,
960 unsigned long end_addr, unsigned long *nr_accounted,
961 struct zap_details *details)
963 long zap_work = ZAP_BLOCK_SIZE;
964 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
965 int tlb_start_valid = 0;
966 unsigned long start = start_addr;
967 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
968 int fullmm = (*tlbp)->fullmm;
969 struct mm_struct *mm = vma->vm_mm;
971 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
972 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
973 unsigned long end;
975 start = max(vma->vm_start, start_addr);
976 if (start >= vma->vm_end)
977 continue;
978 end = min(vma->vm_end, end_addr);
979 if (end <= vma->vm_start)
980 continue;
982 if (vma->vm_flags & VM_ACCOUNT)
983 *nr_accounted += (end - start) >> PAGE_SHIFT;
985 if (unlikely(is_pfn_mapping(vma)))
986 untrack_pfn_vma(vma, 0, 0);
988 while (start != end) {
989 if (!tlb_start_valid) {
990 tlb_start = start;
991 tlb_start_valid = 1;
994 if (unlikely(is_vm_hugetlb_page(vma))) {
996 * It is undesirable to test vma->vm_file as it
997 * should be non-null for valid hugetlb area.
998 * However, vm_file will be NULL in the error
999 * cleanup path of do_mmap_pgoff. When
1000 * hugetlbfs ->mmap method fails,
1001 * do_mmap_pgoff() nullifies vma->vm_file
1002 * before calling this function to clean up.
1003 * Since no pte has actually been setup, it is
1004 * safe to do nothing in this case.
1006 if (vma->vm_file) {
1007 unmap_hugepage_range(vma, start, end, NULL);
1008 zap_work -= (end - start) /
1009 pages_per_huge_page(hstate_vma(vma));
1012 start = end;
1013 } else
1014 start = unmap_page_range(*tlbp, vma,
1015 start, end, &zap_work, details);
1017 if (zap_work > 0) {
1018 BUG_ON(start != end);
1019 break;
1022 tlb_finish_mmu(*tlbp, tlb_start, start);
1024 if (need_resched() ||
1025 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1026 if (i_mmap_lock) {
1027 *tlbp = NULL;
1028 goto out;
1030 cond_resched();
1033 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1034 tlb_start_valid = 0;
1035 zap_work = ZAP_BLOCK_SIZE;
1038 out:
1039 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1040 return start; /* which is now the end (or restart) address */
1044 * zap_page_range - remove user pages in a given range
1045 * @vma: vm_area_struct holding the applicable pages
1046 * @address: starting address of pages to zap
1047 * @size: number of bytes to zap
1048 * @details: details of nonlinear truncation or shared cache invalidation
1050 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1051 unsigned long size, struct zap_details *details)
1053 struct mm_struct *mm = vma->vm_mm;
1054 struct mmu_gather *tlb;
1055 unsigned long end = address + size;
1056 unsigned long nr_accounted = 0;
1058 lru_add_drain();
1059 tlb = tlb_gather_mmu(mm, 0);
1060 update_hiwater_rss(mm);
1061 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1062 if (tlb)
1063 tlb_finish_mmu(tlb, address, end);
1064 return end;
1068 * zap_vma_ptes - remove ptes mapping the vma
1069 * @vma: vm_area_struct holding ptes to be zapped
1070 * @address: starting address of pages to zap
1071 * @size: number of bytes to zap
1073 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1075 * The entire address range must be fully contained within the vma.
1077 * Returns 0 if successful.
1079 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1080 unsigned long size)
1082 if (address < vma->vm_start || address + size > vma->vm_end ||
1083 !(vma->vm_flags & VM_PFNMAP))
1084 return -1;
1085 zap_page_range(vma, address, size, NULL);
1086 return 0;
1088 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1091 * Do a quick page-table lookup for a single page.
1093 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1094 unsigned int flags)
1096 pgd_t *pgd;
1097 pud_t *pud;
1098 pmd_t *pmd;
1099 pte_t *ptep, pte;
1100 spinlock_t *ptl;
1101 struct page *page;
1102 struct mm_struct *mm = vma->vm_mm;
1104 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1105 if (!IS_ERR(page)) {
1106 BUG_ON(flags & FOLL_GET);
1107 goto out;
1110 page = NULL;
1111 pgd = pgd_offset(mm, address);
1112 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1113 goto no_page_table;
1115 pud = pud_offset(pgd, address);
1116 if (pud_none(*pud))
1117 goto no_page_table;
1118 if (pud_huge(*pud)) {
1119 BUG_ON(flags & FOLL_GET);
1120 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1121 goto out;
1123 if (unlikely(pud_bad(*pud)))
1124 goto no_page_table;
1126 pmd = pmd_offset(pud, address);
1127 if (pmd_none(*pmd))
1128 goto no_page_table;
1129 if (pmd_huge(*pmd)) {
1130 BUG_ON(flags & FOLL_GET);
1131 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1132 goto out;
1134 if (unlikely(pmd_bad(*pmd)))
1135 goto no_page_table;
1137 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1139 pte = *ptep;
1140 if (!pte_present(pte))
1141 goto no_page;
1142 if ((flags & FOLL_WRITE) && !pte_write(pte))
1143 goto unlock;
1144 page = vm_normal_page(vma, address, pte);
1145 if (unlikely(!page))
1146 goto bad_page;
1148 if (flags & FOLL_GET)
1149 get_page(page);
1150 if (flags & FOLL_TOUCH) {
1151 if ((flags & FOLL_WRITE) &&
1152 !pte_dirty(pte) && !PageDirty(page))
1153 set_page_dirty(page);
1155 * pte_mkyoung() would be more correct here, but atomic care
1156 * is needed to avoid losing the dirty bit: it is easier to use
1157 * mark_page_accessed().
1159 mark_page_accessed(page);
1161 unlock:
1162 pte_unmap_unlock(ptep, ptl);
1163 out:
1164 return page;
1166 bad_page:
1167 pte_unmap_unlock(ptep, ptl);
1168 return ERR_PTR(-EFAULT);
1170 no_page:
1171 pte_unmap_unlock(ptep, ptl);
1172 if (!pte_none(pte))
1173 return page;
1174 /* Fall through to ZERO_PAGE handling */
1175 no_page_table:
1177 * When core dumping an enormous anonymous area that nobody
1178 * has touched so far, we don't want to allocate page tables.
1180 if (flags & FOLL_ANON) {
1181 page = ZERO_PAGE(0);
1182 if (flags & FOLL_GET)
1183 get_page(page);
1184 BUG_ON(flags & FOLL_WRITE);
1186 return page;
1189 /* Can we do the FOLL_ANON optimization? */
1190 static inline int use_zero_page(struct vm_area_struct *vma)
1193 * We don't want to optimize FOLL_ANON for make_pages_present()
1194 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1195 * we want to get the page from the page tables to make sure
1196 * that we serialize and update with any other user of that
1197 * mapping.
1199 if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1200 return 0;
1202 * And if we have a fault routine, it's not an anonymous region.
1204 return !vma->vm_ops || !vma->vm_ops->fault;
1209 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1210 unsigned long start, int len, int flags,
1211 struct page **pages, struct vm_area_struct **vmas)
1213 int i;
1214 unsigned int vm_flags = 0;
1215 int write = !!(flags & GUP_FLAGS_WRITE);
1216 int force = !!(flags & GUP_FLAGS_FORCE);
1217 int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1218 int ignore_sigkill = !!(flags & GUP_FLAGS_IGNORE_SIGKILL);
1220 if (len <= 0)
1221 return 0;
1223 * Require read or write permissions.
1224 * If 'force' is set, we only require the "MAY" flags.
1226 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1227 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1228 i = 0;
1230 do {
1231 struct vm_area_struct *vma;
1232 unsigned int foll_flags;
1234 vma = find_extend_vma(mm, start);
1235 if (!vma && in_gate_area(tsk, start)) {
1236 unsigned long pg = start & PAGE_MASK;
1237 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1238 pgd_t *pgd;
1239 pud_t *pud;
1240 pmd_t *pmd;
1241 pte_t *pte;
1243 /* user gate pages are read-only */
1244 if (!ignore && write)
1245 return i ? : -EFAULT;
1246 if (pg > TASK_SIZE)
1247 pgd = pgd_offset_k(pg);
1248 else
1249 pgd = pgd_offset_gate(mm, pg);
1250 BUG_ON(pgd_none(*pgd));
1251 pud = pud_offset(pgd, pg);
1252 BUG_ON(pud_none(*pud));
1253 pmd = pmd_offset(pud, pg);
1254 if (pmd_none(*pmd))
1255 return i ? : -EFAULT;
1256 pte = pte_offset_map(pmd, pg);
1257 if (pte_none(*pte)) {
1258 pte_unmap(pte);
1259 return i ? : -EFAULT;
1261 if (pages) {
1262 struct page *page = vm_normal_page(gate_vma, start, *pte);
1263 pages[i] = page;
1264 if (page)
1265 get_page(page);
1267 pte_unmap(pte);
1268 if (vmas)
1269 vmas[i] = gate_vma;
1270 i++;
1271 start += PAGE_SIZE;
1272 len--;
1273 continue;
1276 if (!vma ||
1277 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1278 (!ignore && !(vm_flags & vma->vm_flags)))
1279 return i ? : -EFAULT;
1281 if (is_vm_hugetlb_page(vma)) {
1282 i = follow_hugetlb_page(mm, vma, pages, vmas,
1283 &start, &len, i, write);
1284 continue;
1287 foll_flags = FOLL_TOUCH;
1288 if (pages)
1289 foll_flags |= FOLL_GET;
1290 if (!write && use_zero_page(vma))
1291 foll_flags |= FOLL_ANON;
1293 do {
1294 struct page *page;
1297 * If we have a pending SIGKILL, don't keep faulting
1298 * pages and potentially allocating memory, unless
1299 * current is handling munlock--e.g., on exit. In
1300 * that case, we are not allocating memory. Rather,
1301 * we're only unlocking already resident/mapped pages.
1303 if (unlikely(!ignore_sigkill &&
1304 fatal_signal_pending(current)))
1305 return i ? i : -ERESTARTSYS;
1307 if (write)
1308 foll_flags |= FOLL_WRITE;
1310 cond_resched();
1311 while (!(page = follow_page(vma, start, foll_flags))) {
1312 int ret;
1313 ret = handle_mm_fault(mm, vma, start,
1314 foll_flags & FOLL_WRITE);
1315 if (ret & VM_FAULT_ERROR) {
1316 if (ret & VM_FAULT_OOM)
1317 return i ? i : -ENOMEM;
1318 else if (ret & VM_FAULT_SIGBUS)
1319 return i ? i : -EFAULT;
1320 BUG();
1322 if (ret & VM_FAULT_MAJOR)
1323 tsk->maj_flt++;
1324 else
1325 tsk->min_flt++;
1328 * The VM_FAULT_WRITE bit tells us that
1329 * do_wp_page has broken COW when necessary,
1330 * even if maybe_mkwrite decided not to set
1331 * pte_write. We can thus safely do subsequent
1332 * page lookups as if they were reads. But only
1333 * do so when looping for pte_write is futile:
1334 * in some cases userspace may also be wanting
1335 * to write to the gotten user page, which a
1336 * read fault here might prevent (a readonly
1337 * page might get reCOWed by userspace write).
1339 if ((ret & VM_FAULT_WRITE) &&
1340 !(vma->vm_flags & VM_WRITE))
1341 foll_flags &= ~FOLL_WRITE;
1343 cond_resched();
1345 if (IS_ERR(page))
1346 return i ? i : PTR_ERR(page);
1347 if (pages) {
1348 pages[i] = page;
1350 flush_anon_page(vma, page, start);
1351 flush_dcache_page(page);
1353 if (vmas)
1354 vmas[i] = vma;
1355 i++;
1356 start += PAGE_SIZE;
1357 len--;
1358 } while (len && start < vma->vm_end);
1359 } while (len);
1360 return i;
1364 * get_user_pages() - pin user pages in memory
1365 * @tsk: task_struct of target task
1366 * @mm: mm_struct of target mm
1367 * @start: starting user address
1368 * @len: number of pages from start to pin
1369 * @write: whether pages will be written to by the caller
1370 * @force: whether to force write access even if user mapping is
1371 * readonly. This will result in the page being COWed even
1372 * in MAP_SHARED mappings. You do not want this.
1373 * @pages: array that receives pointers to the pages pinned.
1374 * Should be at least nr_pages long. Or NULL, if caller
1375 * only intends to ensure the pages are faulted in.
1376 * @vmas: array of pointers to vmas corresponding to each page.
1377 * Or NULL if the caller does not require them.
1379 * Returns number of pages pinned. This may be fewer than the number
1380 * requested. If len is 0 or negative, returns 0. If no pages
1381 * were pinned, returns -errno. Each page returned must be released
1382 * with a put_page() call when it is finished with. vmas will only
1383 * remain valid while mmap_sem is held.
1385 * Must be called with mmap_sem held for read or write.
1387 * get_user_pages walks a process's page tables and takes a reference to
1388 * each struct page that each user address corresponds to at a given
1389 * instant. That is, it takes the page that would be accessed if a user
1390 * thread accesses the given user virtual address at that instant.
1392 * This does not guarantee that the page exists in the user mappings when
1393 * get_user_pages returns, and there may even be a completely different
1394 * page there in some cases (eg. if mmapped pagecache has been invalidated
1395 * and subsequently re faulted). However it does guarantee that the page
1396 * won't be freed completely. And mostly callers simply care that the page
1397 * contains data that was valid *at some point in time*. Typically, an IO
1398 * or similar operation cannot guarantee anything stronger anyway because
1399 * locks can't be held over the syscall boundary.
1401 * If write=0, the page must not be written to. If the page is written to,
1402 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1403 * after the page is finished with, and before put_page is called.
1405 * get_user_pages is typically used for fewer-copy IO operations, to get a
1406 * handle on the memory by some means other than accesses via the user virtual
1407 * addresses. The pages may be submitted for DMA to devices or accessed via
1408 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1409 * use the correct cache flushing APIs.
1411 * See also get_user_pages_fast, for performance critical applications.
1413 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1414 unsigned long start, int len, int write, int force,
1415 struct page **pages, struct vm_area_struct **vmas)
1417 int flags = 0;
1419 if (write)
1420 flags |= GUP_FLAGS_WRITE;
1421 if (force)
1422 flags |= GUP_FLAGS_FORCE;
1424 return __get_user_pages(tsk, mm,
1425 start, len, flags,
1426 pages, vmas);
1429 EXPORT_SYMBOL(get_user_pages);
1431 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1432 spinlock_t **ptl)
1434 pgd_t * pgd = pgd_offset(mm, addr);
1435 pud_t * pud = pud_alloc(mm, pgd, addr);
1436 if (pud) {
1437 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1438 if (pmd)
1439 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1441 return NULL;
1445 * This is the old fallback for page remapping.
1447 * For historical reasons, it only allows reserved pages. Only
1448 * old drivers should use this, and they needed to mark their
1449 * pages reserved for the old functions anyway.
1451 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1452 struct page *page, pgprot_t prot)
1454 struct mm_struct *mm = vma->vm_mm;
1455 int retval;
1456 pte_t *pte;
1457 spinlock_t *ptl;
1459 retval = -EINVAL;
1460 if (PageAnon(page))
1461 goto out;
1462 retval = -ENOMEM;
1463 flush_dcache_page(page);
1464 pte = get_locked_pte(mm, addr, &ptl);
1465 if (!pte)
1466 goto out;
1467 retval = -EBUSY;
1468 if (!pte_none(*pte))
1469 goto out_unlock;
1471 /* Ok, finally just insert the thing.. */
1472 get_page(page);
1473 inc_mm_counter(mm, file_rss);
1474 page_add_file_rmap(page);
1475 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1477 retval = 0;
1478 pte_unmap_unlock(pte, ptl);
1479 return retval;
1480 out_unlock:
1481 pte_unmap_unlock(pte, ptl);
1482 out:
1483 return retval;
1487 * vm_insert_page - insert single page into user vma
1488 * @vma: user vma to map to
1489 * @addr: target user address of this page
1490 * @page: source kernel page
1492 * This allows drivers to insert individual pages they've allocated
1493 * into a user vma.
1495 * The page has to be a nice clean _individual_ kernel allocation.
1496 * If you allocate a compound page, you need to have marked it as
1497 * such (__GFP_COMP), or manually just split the page up yourself
1498 * (see split_page()).
1500 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1501 * took an arbitrary page protection parameter. This doesn't allow
1502 * that. Your vma protection will have to be set up correctly, which
1503 * means that if you want a shared writable mapping, you'd better
1504 * ask for a shared writable mapping!
1506 * The page does not need to be reserved.
1508 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1509 struct page *page)
1511 if (addr < vma->vm_start || addr >= vma->vm_end)
1512 return -EFAULT;
1513 if (!page_count(page))
1514 return -EINVAL;
1515 vma->vm_flags |= VM_INSERTPAGE;
1516 return insert_page(vma, addr, page, vma->vm_page_prot);
1518 EXPORT_SYMBOL(vm_insert_page);
1520 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1521 unsigned long pfn, pgprot_t prot)
1523 struct mm_struct *mm = vma->vm_mm;
1524 int retval;
1525 pte_t *pte, entry;
1526 spinlock_t *ptl;
1528 retval = -ENOMEM;
1529 pte = get_locked_pte(mm, addr, &ptl);
1530 if (!pte)
1531 goto out;
1532 retval = -EBUSY;
1533 if (!pte_none(*pte))
1534 goto out_unlock;
1536 /* Ok, finally just insert the thing.. */
1537 entry = pte_mkspecial(pfn_pte(pfn, prot));
1538 set_pte_at(mm, addr, pte, entry);
1539 update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1541 retval = 0;
1542 out_unlock:
1543 pte_unmap_unlock(pte, ptl);
1544 out:
1545 return retval;
1549 * vm_insert_pfn - insert single pfn into user vma
1550 * @vma: user vma to map to
1551 * @addr: target user address of this page
1552 * @pfn: source kernel pfn
1554 * Similar to vm_inert_page, this allows drivers to insert individual pages
1555 * they've allocated into a user vma. Same comments apply.
1557 * This function should only be called from a vm_ops->fault handler, and
1558 * in that case the handler should return NULL.
1560 * vma cannot be a COW mapping.
1562 * As this is called only for pages that do not currently exist, we
1563 * do not need to flush old virtual caches or the TLB.
1565 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1566 unsigned long pfn)
1568 int ret;
1569 pgprot_t pgprot = vma->vm_page_prot;
1571 * Technically, architectures with pte_special can avoid all these
1572 * restrictions (same for remap_pfn_range). However we would like
1573 * consistency in testing and feature parity among all, so we should
1574 * try to keep these invariants in place for everybody.
1576 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1577 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1578 (VM_PFNMAP|VM_MIXEDMAP));
1579 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1580 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1582 if (addr < vma->vm_start || addr >= vma->vm_end)
1583 return -EFAULT;
1584 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1585 return -EINVAL;
1587 ret = insert_pfn(vma, addr, pfn, pgprot);
1589 if (ret)
1590 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1592 return ret;
1594 EXPORT_SYMBOL(vm_insert_pfn);
1596 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1597 unsigned long pfn)
1599 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1601 if (addr < vma->vm_start || addr >= vma->vm_end)
1602 return -EFAULT;
1605 * If we don't have pte special, then we have to use the pfn_valid()
1606 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1607 * refcount the page if pfn_valid is true (hence insert_page rather
1608 * than insert_pfn).
1610 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1611 struct page *page;
1613 page = pfn_to_page(pfn);
1614 return insert_page(vma, addr, page, vma->vm_page_prot);
1616 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1618 EXPORT_SYMBOL(vm_insert_mixed);
1621 * maps a range of physical memory into the requested pages. the old
1622 * mappings are removed. any references to nonexistent pages results
1623 * in null mappings (currently treated as "copy-on-access")
1625 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1626 unsigned long addr, unsigned long end,
1627 unsigned long pfn, pgprot_t prot)
1629 pte_t *pte;
1630 spinlock_t *ptl;
1632 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1633 if (!pte)
1634 return -ENOMEM;
1635 arch_enter_lazy_mmu_mode();
1636 do {
1637 BUG_ON(!pte_none(*pte));
1638 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1639 pfn++;
1640 } while (pte++, addr += PAGE_SIZE, addr != end);
1641 arch_leave_lazy_mmu_mode();
1642 pte_unmap_unlock(pte - 1, ptl);
1643 return 0;
1646 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1647 unsigned long addr, unsigned long end,
1648 unsigned long pfn, pgprot_t prot)
1650 pmd_t *pmd;
1651 unsigned long next;
1653 pfn -= addr >> PAGE_SHIFT;
1654 pmd = pmd_alloc(mm, pud, addr);
1655 if (!pmd)
1656 return -ENOMEM;
1657 do {
1658 next = pmd_addr_end(addr, end);
1659 if (remap_pte_range(mm, pmd, addr, next,
1660 pfn + (addr >> PAGE_SHIFT), prot))
1661 return -ENOMEM;
1662 } while (pmd++, addr = next, addr != end);
1663 return 0;
1666 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1667 unsigned long addr, unsigned long end,
1668 unsigned long pfn, pgprot_t prot)
1670 pud_t *pud;
1671 unsigned long next;
1673 pfn -= addr >> PAGE_SHIFT;
1674 pud = pud_alloc(mm, pgd, addr);
1675 if (!pud)
1676 return -ENOMEM;
1677 do {
1678 next = pud_addr_end(addr, end);
1679 if (remap_pmd_range(mm, pud, addr, next,
1680 pfn + (addr >> PAGE_SHIFT), prot))
1681 return -ENOMEM;
1682 } while (pud++, addr = next, addr != end);
1683 return 0;
1687 * remap_pfn_range - remap kernel memory to userspace
1688 * @vma: user vma to map to
1689 * @addr: target user address to start at
1690 * @pfn: physical address of kernel memory
1691 * @size: size of map area
1692 * @prot: page protection flags for this mapping
1694 * Note: this is only safe if the mm semaphore is held when called.
1696 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1697 unsigned long pfn, unsigned long size, pgprot_t prot)
1699 pgd_t *pgd;
1700 unsigned long next;
1701 unsigned long end = addr + PAGE_ALIGN(size);
1702 struct mm_struct *mm = vma->vm_mm;
1703 int err;
1706 * Physically remapped pages are special. Tell the
1707 * rest of the world about it:
1708 * VM_IO tells people not to look at these pages
1709 * (accesses can have side effects).
1710 * VM_RESERVED is specified all over the place, because
1711 * in 2.4 it kept swapout's vma scan off this vma; but
1712 * in 2.6 the LRU scan won't even find its pages, so this
1713 * flag means no more than count its pages in reserved_vm,
1714 * and omit it from core dump, even when VM_IO turned off.
1715 * VM_PFNMAP tells the core MM that the base pages are just
1716 * raw PFN mappings, and do not have a "struct page" associated
1717 * with them.
1719 * There's a horrible special case to handle copy-on-write
1720 * behaviour that some programs depend on. We mark the "original"
1721 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1723 if (addr == vma->vm_start && end == vma->vm_end) {
1724 vma->vm_pgoff = pfn;
1725 vma->vm_flags |= VM_PFN_AT_MMAP;
1726 } else if (is_cow_mapping(vma->vm_flags))
1727 return -EINVAL;
1729 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1731 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1732 if (err) {
1734 * To indicate that track_pfn related cleanup is not
1735 * needed from higher level routine calling unmap_vmas
1737 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1738 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1739 return -EINVAL;
1742 BUG_ON(addr >= end);
1743 pfn -= addr >> PAGE_SHIFT;
1744 pgd = pgd_offset(mm, addr);
1745 flush_cache_range(vma, addr, end);
1746 do {
1747 next = pgd_addr_end(addr, end);
1748 err = remap_pud_range(mm, pgd, addr, next,
1749 pfn + (addr >> PAGE_SHIFT), prot);
1750 if (err)
1751 break;
1752 } while (pgd++, addr = next, addr != end);
1754 if (err)
1755 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1757 return err;
1759 EXPORT_SYMBOL(remap_pfn_range);
1761 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1762 unsigned long addr, unsigned long end,
1763 pte_fn_t fn, void *data)
1765 pte_t *pte;
1766 int err;
1767 pgtable_t token;
1768 spinlock_t *uninitialized_var(ptl);
1770 pte = (mm == &init_mm) ?
1771 pte_alloc_kernel(pmd, addr) :
1772 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1773 if (!pte)
1774 return -ENOMEM;
1776 BUG_ON(pmd_huge(*pmd));
1778 arch_enter_lazy_mmu_mode();
1780 token = pmd_pgtable(*pmd);
1782 do {
1783 err = fn(pte, token, addr, data);
1784 if (err)
1785 break;
1786 } while (pte++, addr += PAGE_SIZE, addr != end);
1788 arch_leave_lazy_mmu_mode();
1790 if (mm != &init_mm)
1791 pte_unmap_unlock(pte-1, ptl);
1792 return err;
1795 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1796 unsigned long addr, unsigned long end,
1797 pte_fn_t fn, void *data)
1799 pmd_t *pmd;
1800 unsigned long next;
1801 int err;
1803 BUG_ON(pud_huge(*pud));
1805 pmd = pmd_alloc(mm, pud, addr);
1806 if (!pmd)
1807 return -ENOMEM;
1808 do {
1809 next = pmd_addr_end(addr, end);
1810 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1811 if (err)
1812 break;
1813 } while (pmd++, addr = next, addr != end);
1814 return err;
1817 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1818 unsigned long addr, unsigned long end,
1819 pte_fn_t fn, void *data)
1821 pud_t *pud;
1822 unsigned long next;
1823 int err;
1825 pud = pud_alloc(mm, pgd, addr);
1826 if (!pud)
1827 return -ENOMEM;
1828 do {
1829 next = pud_addr_end(addr, end);
1830 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1831 if (err)
1832 break;
1833 } while (pud++, addr = next, addr != end);
1834 return err;
1838 * Scan a region of virtual memory, filling in page tables as necessary
1839 * and calling a provided function on each leaf page table.
1841 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1842 unsigned long size, pte_fn_t fn, void *data)
1844 pgd_t *pgd;
1845 unsigned long next;
1846 unsigned long start = addr, end = addr + size;
1847 int err;
1849 BUG_ON(addr >= end);
1850 mmu_notifier_invalidate_range_start(mm, start, end);
1851 pgd = pgd_offset(mm, addr);
1852 do {
1853 next = pgd_addr_end(addr, end);
1854 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1855 if (err)
1856 break;
1857 } while (pgd++, addr = next, addr != end);
1858 mmu_notifier_invalidate_range_end(mm, start, end);
1859 return err;
1861 EXPORT_SYMBOL_GPL(apply_to_page_range);
1864 * handle_pte_fault chooses page fault handler according to an entry
1865 * which was read non-atomically. Before making any commitment, on
1866 * those architectures or configurations (e.g. i386 with PAE) which
1867 * might give a mix of unmatched parts, do_swap_page and do_file_page
1868 * must check under lock before unmapping the pte and proceeding
1869 * (but do_wp_page is only called after already making such a check;
1870 * and do_anonymous_page and do_no_page can safely check later on).
1872 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1873 pte_t *page_table, pte_t orig_pte)
1875 int same = 1;
1876 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1877 if (sizeof(pte_t) > sizeof(unsigned long)) {
1878 spinlock_t *ptl = pte_lockptr(mm, pmd);
1879 spin_lock(ptl);
1880 same = pte_same(*page_table, orig_pte);
1881 spin_unlock(ptl);
1883 #endif
1884 pte_unmap(page_table);
1885 return same;
1889 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1890 * servicing faults for write access. In the normal case, do always want
1891 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1892 * that do not have writing enabled, when used by access_process_vm.
1894 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1896 if (likely(vma->vm_flags & VM_WRITE))
1897 pte = pte_mkwrite(pte);
1898 return pte;
1901 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1904 * If the source page was a PFN mapping, we don't have
1905 * a "struct page" for it. We do a best-effort copy by
1906 * just copying from the original user address. If that
1907 * fails, we just zero-fill it. Live with it.
1909 if (unlikely(!src)) {
1910 void *kaddr = kmap_atomic(dst, KM_USER0);
1911 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1914 * This really shouldn't fail, because the page is there
1915 * in the page tables. But it might just be unreadable,
1916 * in which case we just give up and fill the result with
1917 * zeroes.
1919 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1920 memset(kaddr, 0, PAGE_SIZE);
1921 kunmap_atomic(kaddr, KM_USER0);
1922 flush_dcache_page(dst);
1923 } else
1924 copy_user_highpage(dst, src, va, vma);
1928 * This routine handles present pages, when users try to write
1929 * to a shared page. It is done by copying the page to a new address
1930 * and decrementing the shared-page counter for the old page.
1932 * Note that this routine assumes that the protection checks have been
1933 * done by the caller (the low-level page fault routine in most cases).
1934 * Thus we can safely just mark it writable once we've done any necessary
1935 * COW.
1937 * We also mark the page dirty at this point even though the page will
1938 * change only once the write actually happens. This avoids a few races,
1939 * and potentially makes it more efficient.
1941 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1942 * but allow concurrent faults), with pte both mapped and locked.
1943 * We return with mmap_sem still held, but pte unmapped and unlocked.
1945 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1946 unsigned long address, pte_t *page_table, pmd_t *pmd,
1947 spinlock_t *ptl, pte_t orig_pte)
1949 struct page *old_page, *new_page;
1950 pte_t entry;
1951 int reuse = 0, ret = 0;
1952 int page_mkwrite = 0;
1953 struct page *dirty_page = NULL;
1955 old_page = vm_normal_page(vma, address, orig_pte);
1956 if (!old_page) {
1958 * VM_MIXEDMAP !pfn_valid() case
1960 * We should not cow pages in a shared writeable mapping.
1961 * Just mark the pages writable as we can't do any dirty
1962 * accounting on raw pfn maps.
1964 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1965 (VM_WRITE|VM_SHARED))
1966 goto reuse;
1967 goto gotten;
1971 * Take out anonymous pages first, anonymous shared vmas are
1972 * not dirty accountable.
1974 if (PageAnon(old_page)) {
1975 if (!trylock_page(old_page)) {
1976 page_cache_get(old_page);
1977 pte_unmap_unlock(page_table, ptl);
1978 lock_page(old_page);
1979 page_table = pte_offset_map_lock(mm, pmd, address,
1980 &ptl);
1981 if (!pte_same(*page_table, orig_pte)) {
1982 unlock_page(old_page);
1983 page_cache_release(old_page);
1984 goto unlock;
1986 page_cache_release(old_page);
1988 reuse = reuse_swap_page(old_page);
1989 unlock_page(old_page);
1990 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1991 (VM_WRITE|VM_SHARED))) {
1993 * Only catch write-faults on shared writable pages,
1994 * read-only shared pages can get COWed by
1995 * get_user_pages(.write=1, .force=1).
1997 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1998 struct vm_fault vmf;
1999 int tmp;
2001 vmf.virtual_address = (void __user *)(address &
2002 PAGE_MASK);
2003 vmf.pgoff = old_page->index;
2004 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2005 vmf.page = old_page;
2008 * Notify the address space that the page is about to
2009 * become writable so that it can prohibit this or wait
2010 * for the page to get into an appropriate state.
2012 * We do this without the lock held, so that it can
2013 * sleep if it needs to.
2015 page_cache_get(old_page);
2016 pte_unmap_unlock(page_table, ptl);
2018 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2019 if (unlikely(tmp &
2020 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2021 ret = tmp;
2022 goto unwritable_page;
2024 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2025 lock_page(old_page);
2026 if (!old_page->mapping) {
2027 ret = 0; /* retry the fault */
2028 unlock_page(old_page);
2029 goto unwritable_page;
2031 } else
2032 VM_BUG_ON(!PageLocked(old_page));
2035 * Since we dropped the lock we need to revalidate
2036 * the PTE as someone else may have changed it. If
2037 * they did, we just return, as we can count on the
2038 * MMU to tell us if they didn't also make it writable.
2040 page_table = pte_offset_map_lock(mm, pmd, address,
2041 &ptl);
2042 if (!pte_same(*page_table, orig_pte)) {
2043 unlock_page(old_page);
2044 page_cache_release(old_page);
2045 goto unlock;
2048 page_mkwrite = 1;
2050 dirty_page = old_page;
2051 get_page(dirty_page);
2052 reuse = 1;
2055 if (reuse) {
2056 reuse:
2057 flush_cache_page(vma, address, pte_pfn(orig_pte));
2058 entry = pte_mkyoung(orig_pte);
2059 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2060 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2061 update_mmu_cache(vma, address, entry);
2062 ret |= VM_FAULT_WRITE;
2063 goto unlock;
2067 * Ok, we need to copy. Oh, well..
2069 page_cache_get(old_page);
2070 gotten:
2071 pte_unmap_unlock(page_table, ptl);
2073 if (unlikely(anon_vma_prepare(vma)))
2074 goto oom;
2075 VM_BUG_ON(old_page == ZERO_PAGE(0));
2076 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2077 if (!new_page)
2078 goto oom;
2080 * Don't let another task, with possibly unlocked vma,
2081 * keep the mlocked page.
2083 if ((vma->vm_flags & VM_LOCKED) && old_page) {
2084 lock_page(old_page); /* for LRU manipulation */
2085 clear_page_mlock(old_page);
2086 unlock_page(old_page);
2088 cow_user_page(new_page, old_page, address, vma);
2089 __SetPageUptodate(new_page);
2091 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2092 goto oom_free_new;
2095 * Re-check the pte - we dropped the lock
2097 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2098 if (likely(pte_same(*page_table, orig_pte))) {
2099 if (old_page) {
2100 if (!PageAnon(old_page)) {
2101 dec_mm_counter(mm, file_rss);
2102 inc_mm_counter(mm, anon_rss);
2104 } else
2105 inc_mm_counter(mm, anon_rss);
2106 flush_cache_page(vma, address, pte_pfn(orig_pte));
2107 entry = mk_pte(new_page, vma->vm_page_prot);
2108 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2110 * Clear the pte entry and flush it first, before updating the
2111 * pte with the new entry. This will avoid a race condition
2112 * seen in the presence of one thread doing SMC and another
2113 * thread doing COW.
2115 ptep_clear_flush_notify(vma, address, page_table);
2116 page_add_new_anon_rmap(new_page, vma, address);
2117 set_pte_at(mm, address, page_table, entry);
2118 update_mmu_cache(vma, address, entry);
2119 if (old_page) {
2121 * Only after switching the pte to the new page may
2122 * we remove the mapcount here. Otherwise another
2123 * process may come and find the rmap count decremented
2124 * before the pte is switched to the new page, and
2125 * "reuse" the old page writing into it while our pte
2126 * here still points into it and can be read by other
2127 * threads.
2129 * The critical issue is to order this
2130 * page_remove_rmap with the ptp_clear_flush above.
2131 * Those stores are ordered by (if nothing else,)
2132 * the barrier present in the atomic_add_negative
2133 * in page_remove_rmap.
2135 * Then the TLB flush in ptep_clear_flush ensures that
2136 * no process can access the old page before the
2137 * decremented mapcount is visible. And the old page
2138 * cannot be reused until after the decremented
2139 * mapcount is visible. So transitively, TLBs to
2140 * old page will be flushed before it can be reused.
2142 page_remove_rmap(old_page);
2145 /* Free the old page.. */
2146 new_page = old_page;
2147 ret |= VM_FAULT_WRITE;
2148 } else
2149 mem_cgroup_uncharge_page(new_page);
2151 if (new_page)
2152 page_cache_release(new_page);
2153 if (old_page)
2154 page_cache_release(old_page);
2155 unlock:
2156 pte_unmap_unlock(page_table, ptl);
2157 if (dirty_page) {
2159 * Yes, Virginia, this is actually required to prevent a race
2160 * with clear_page_dirty_for_io() from clearing the page dirty
2161 * bit after it clear all dirty ptes, but before a racing
2162 * do_wp_page installs a dirty pte.
2164 * do_no_page is protected similarly.
2166 if (!page_mkwrite) {
2167 wait_on_page_locked(dirty_page);
2168 set_page_dirty_balance(dirty_page, page_mkwrite);
2170 put_page(dirty_page);
2171 if (page_mkwrite) {
2172 struct address_space *mapping = dirty_page->mapping;
2174 set_page_dirty(dirty_page);
2175 unlock_page(dirty_page);
2176 page_cache_release(dirty_page);
2177 if (mapping) {
2179 * Some device drivers do not set page.mapping
2180 * but still dirty their pages
2182 balance_dirty_pages_ratelimited(mapping);
2186 /* file_update_time outside page_lock */
2187 if (vma->vm_file)
2188 file_update_time(vma->vm_file);
2190 return ret;
2191 oom_free_new:
2192 page_cache_release(new_page);
2193 oom:
2194 if (old_page) {
2195 if (page_mkwrite) {
2196 unlock_page(old_page);
2197 page_cache_release(old_page);
2199 page_cache_release(old_page);
2201 return VM_FAULT_OOM;
2203 unwritable_page:
2204 page_cache_release(old_page);
2205 return ret;
2209 * Helper functions for unmap_mapping_range().
2211 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2213 * We have to restart searching the prio_tree whenever we drop the lock,
2214 * since the iterator is only valid while the lock is held, and anyway
2215 * a later vma might be split and reinserted earlier while lock dropped.
2217 * The list of nonlinear vmas could be handled more efficiently, using
2218 * a placeholder, but handle it in the same way until a need is shown.
2219 * It is important to search the prio_tree before nonlinear list: a vma
2220 * may become nonlinear and be shifted from prio_tree to nonlinear list
2221 * while the lock is dropped; but never shifted from list to prio_tree.
2223 * In order to make forward progress despite restarting the search,
2224 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2225 * quickly skip it next time around. Since the prio_tree search only
2226 * shows us those vmas affected by unmapping the range in question, we
2227 * can't efficiently keep all vmas in step with mapping->truncate_count:
2228 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2229 * mapping->truncate_count and vma->vm_truncate_count are protected by
2230 * i_mmap_lock.
2232 * In order to make forward progress despite repeatedly restarting some
2233 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2234 * and restart from that address when we reach that vma again. It might
2235 * have been split or merged, shrunk or extended, but never shifted: so
2236 * restart_addr remains valid so long as it remains in the vma's range.
2237 * unmap_mapping_range forces truncate_count to leap over page-aligned
2238 * values so we can save vma's restart_addr in its truncate_count field.
2240 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2242 static void reset_vma_truncate_counts(struct address_space *mapping)
2244 struct vm_area_struct *vma;
2245 struct prio_tree_iter iter;
2247 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2248 vma->vm_truncate_count = 0;
2249 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2250 vma->vm_truncate_count = 0;
2253 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2254 unsigned long start_addr, unsigned long end_addr,
2255 struct zap_details *details)
2257 unsigned long restart_addr;
2258 int need_break;
2261 * files that support invalidating or truncating portions of the
2262 * file from under mmaped areas must have their ->fault function
2263 * return a locked page (and set VM_FAULT_LOCKED in the return).
2264 * This provides synchronisation against concurrent unmapping here.
2267 again:
2268 restart_addr = vma->vm_truncate_count;
2269 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2270 start_addr = restart_addr;
2271 if (start_addr >= end_addr) {
2272 /* Top of vma has been split off since last time */
2273 vma->vm_truncate_count = details->truncate_count;
2274 return 0;
2278 restart_addr = zap_page_range(vma, start_addr,
2279 end_addr - start_addr, details);
2280 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2282 if (restart_addr >= end_addr) {
2283 /* We have now completed this vma: mark it so */
2284 vma->vm_truncate_count = details->truncate_count;
2285 if (!need_break)
2286 return 0;
2287 } else {
2288 /* Note restart_addr in vma's truncate_count field */
2289 vma->vm_truncate_count = restart_addr;
2290 if (!need_break)
2291 goto again;
2294 spin_unlock(details->i_mmap_lock);
2295 cond_resched();
2296 spin_lock(details->i_mmap_lock);
2297 return -EINTR;
2300 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2301 struct zap_details *details)
2303 struct vm_area_struct *vma;
2304 struct prio_tree_iter iter;
2305 pgoff_t vba, vea, zba, zea;
2307 restart:
2308 vma_prio_tree_foreach(vma, &iter, root,
2309 details->first_index, details->last_index) {
2310 /* Skip quickly over those we have already dealt with */
2311 if (vma->vm_truncate_count == details->truncate_count)
2312 continue;
2314 vba = vma->vm_pgoff;
2315 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2316 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2317 zba = details->first_index;
2318 if (zba < vba)
2319 zba = vba;
2320 zea = details->last_index;
2321 if (zea > vea)
2322 zea = vea;
2324 if (unmap_mapping_range_vma(vma,
2325 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2326 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2327 details) < 0)
2328 goto restart;
2332 static inline void unmap_mapping_range_list(struct list_head *head,
2333 struct zap_details *details)
2335 struct vm_area_struct *vma;
2338 * In nonlinear VMAs there is no correspondence between virtual address
2339 * offset and file offset. So we must perform an exhaustive search
2340 * across *all* the pages in each nonlinear VMA, not just the pages
2341 * whose virtual address lies outside the file truncation point.
2343 restart:
2344 list_for_each_entry(vma, head, shared.vm_set.list) {
2345 /* Skip quickly over those we have already dealt with */
2346 if (vma->vm_truncate_count == details->truncate_count)
2347 continue;
2348 details->nonlinear_vma = vma;
2349 if (unmap_mapping_range_vma(vma, vma->vm_start,
2350 vma->vm_end, details) < 0)
2351 goto restart;
2356 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2357 * @mapping: the address space containing mmaps to be unmapped.
2358 * @holebegin: byte in first page to unmap, relative to the start of
2359 * the underlying file. This will be rounded down to a PAGE_SIZE
2360 * boundary. Note that this is different from vmtruncate(), which
2361 * must keep the partial page. In contrast, we must get rid of
2362 * partial pages.
2363 * @holelen: size of prospective hole in bytes. This will be rounded
2364 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2365 * end of the file.
2366 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2367 * but 0 when invalidating pagecache, don't throw away private data.
2369 void unmap_mapping_range(struct address_space *mapping,
2370 loff_t const holebegin, loff_t const holelen, int even_cows)
2372 struct zap_details details;
2373 pgoff_t hba = holebegin >> PAGE_SHIFT;
2374 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2376 /* Check for overflow. */
2377 if (sizeof(holelen) > sizeof(hlen)) {
2378 long long holeend =
2379 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2380 if (holeend & ~(long long)ULONG_MAX)
2381 hlen = ULONG_MAX - hba + 1;
2384 details.check_mapping = even_cows? NULL: mapping;
2385 details.nonlinear_vma = NULL;
2386 details.first_index = hba;
2387 details.last_index = hba + hlen - 1;
2388 if (details.last_index < details.first_index)
2389 details.last_index = ULONG_MAX;
2390 details.i_mmap_lock = &mapping->i_mmap_lock;
2392 spin_lock(&mapping->i_mmap_lock);
2394 /* Protect against endless unmapping loops */
2395 mapping->truncate_count++;
2396 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2397 if (mapping->truncate_count == 0)
2398 reset_vma_truncate_counts(mapping);
2399 mapping->truncate_count++;
2401 details.truncate_count = mapping->truncate_count;
2403 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2404 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2405 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2406 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2407 spin_unlock(&mapping->i_mmap_lock);
2409 EXPORT_SYMBOL(unmap_mapping_range);
2412 * vmtruncate - unmap mappings "freed" by truncate() syscall
2413 * @inode: inode of the file used
2414 * @offset: file offset to start truncating
2416 * NOTE! We have to be ready to update the memory sharing
2417 * between the file and the memory map for a potential last
2418 * incomplete page. Ugly, but necessary.
2420 int vmtruncate(struct inode * inode, loff_t offset)
2422 if (inode->i_size < offset) {
2423 unsigned long limit;
2425 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2426 if (limit != RLIM_INFINITY && offset > limit)
2427 goto out_sig;
2428 if (offset > inode->i_sb->s_maxbytes)
2429 goto out_big;
2430 i_size_write(inode, offset);
2431 } else {
2432 struct address_space *mapping = inode->i_mapping;
2435 * truncation of in-use swapfiles is disallowed - it would
2436 * cause subsequent swapout to scribble on the now-freed
2437 * blocks.
2439 if (IS_SWAPFILE(inode))
2440 return -ETXTBSY;
2441 i_size_write(inode, offset);
2444 * unmap_mapping_range is called twice, first simply for
2445 * efficiency so that truncate_inode_pages does fewer
2446 * single-page unmaps. However after this first call, and
2447 * before truncate_inode_pages finishes, it is possible for
2448 * private pages to be COWed, which remain after
2449 * truncate_inode_pages finishes, hence the second
2450 * unmap_mapping_range call must be made for correctness.
2452 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2453 truncate_inode_pages(mapping, offset);
2454 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2457 if (inode->i_op->truncate)
2458 inode->i_op->truncate(inode);
2459 return 0;
2461 out_sig:
2462 send_sig(SIGXFSZ, current, 0);
2463 out_big:
2464 return -EFBIG;
2466 EXPORT_SYMBOL(vmtruncate);
2468 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2470 struct address_space *mapping = inode->i_mapping;
2473 * If the underlying filesystem is not going to provide
2474 * a way to truncate a range of blocks (punch a hole) -
2475 * we should return failure right now.
2477 if (!inode->i_op->truncate_range)
2478 return -ENOSYS;
2480 mutex_lock(&inode->i_mutex);
2481 down_write(&inode->i_alloc_sem);
2482 unmap_mapping_range(mapping, offset, (end - offset), 1);
2483 truncate_inode_pages_range(mapping, offset, end);
2484 unmap_mapping_range(mapping, offset, (end - offset), 1);
2485 inode->i_op->truncate_range(inode, offset, end);
2486 up_write(&inode->i_alloc_sem);
2487 mutex_unlock(&inode->i_mutex);
2489 return 0;
2493 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2494 * but allow concurrent faults), and pte mapped but not yet locked.
2495 * We return with mmap_sem still held, but pte unmapped and unlocked.
2497 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2498 unsigned long address, pte_t *page_table, pmd_t *pmd,
2499 int write_access, pte_t orig_pte)
2501 spinlock_t *ptl;
2502 struct page *page;
2503 swp_entry_t entry;
2504 pte_t pte;
2505 struct mem_cgroup *ptr = NULL;
2506 int ret = 0;
2508 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2509 goto out;
2511 entry = pte_to_swp_entry(orig_pte);
2512 if (is_migration_entry(entry)) {
2513 migration_entry_wait(mm, pmd, address);
2514 goto out;
2516 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2517 page = lookup_swap_cache(entry);
2518 if (!page) {
2519 grab_swap_token(); /* Contend for token _before_ read-in */
2520 page = swapin_readahead(entry,
2521 GFP_HIGHUSER_MOVABLE, vma, address);
2522 if (!page) {
2524 * Back out if somebody else faulted in this pte
2525 * while we released the pte lock.
2527 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2528 if (likely(pte_same(*page_table, orig_pte)))
2529 ret = VM_FAULT_OOM;
2530 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2531 goto unlock;
2534 /* Had to read the page from swap area: Major fault */
2535 ret = VM_FAULT_MAJOR;
2536 count_vm_event(PGMAJFAULT);
2539 lock_page(page);
2540 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2542 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2543 ret = VM_FAULT_OOM;
2544 goto out_page;
2548 * Back out if somebody else already faulted in this pte.
2550 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2551 if (unlikely(!pte_same(*page_table, orig_pte)))
2552 goto out_nomap;
2554 if (unlikely(!PageUptodate(page))) {
2555 ret = VM_FAULT_SIGBUS;
2556 goto out_nomap;
2560 * The page isn't present yet, go ahead with the fault.
2562 * Be careful about the sequence of operations here.
2563 * To get its accounting right, reuse_swap_page() must be called
2564 * while the page is counted on swap but not yet in mapcount i.e.
2565 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2566 * must be called after the swap_free(), or it will never succeed.
2567 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2568 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2569 * in page->private. In this case, a record in swap_cgroup is silently
2570 * discarded at swap_free().
2573 inc_mm_counter(mm, anon_rss);
2574 pte = mk_pte(page, vma->vm_page_prot);
2575 if (write_access && reuse_swap_page(page)) {
2576 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2577 write_access = 0;
2579 flush_icache_page(vma, page);
2580 set_pte_at(mm, address, page_table, pte);
2581 page_add_anon_rmap(page, vma, address);
2582 /* It's better to call commit-charge after rmap is established */
2583 mem_cgroup_commit_charge_swapin(page, ptr);
2585 swap_free(entry);
2586 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2587 try_to_free_swap(page);
2588 unlock_page(page);
2590 if (write_access) {
2591 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2592 if (ret & VM_FAULT_ERROR)
2593 ret &= VM_FAULT_ERROR;
2594 goto out;
2597 /* No need to invalidate - it was non-present before */
2598 update_mmu_cache(vma, address, pte);
2599 unlock:
2600 pte_unmap_unlock(page_table, ptl);
2601 out:
2602 return ret;
2603 out_nomap:
2604 mem_cgroup_cancel_charge_swapin(ptr);
2605 pte_unmap_unlock(page_table, ptl);
2606 out_page:
2607 unlock_page(page);
2608 page_cache_release(page);
2609 return ret;
2613 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2614 * but allow concurrent faults), and pte mapped but not yet locked.
2615 * We return with mmap_sem still held, but pte unmapped and unlocked.
2617 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2618 unsigned long address, pte_t *page_table, pmd_t *pmd,
2619 int write_access)
2621 struct page *page;
2622 spinlock_t *ptl;
2623 pte_t entry;
2625 /* Allocate our own private page. */
2626 pte_unmap(page_table);
2628 if (unlikely(anon_vma_prepare(vma)))
2629 goto oom;
2630 page = alloc_zeroed_user_highpage_movable(vma, address);
2631 if (!page)
2632 goto oom;
2633 __SetPageUptodate(page);
2635 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2636 goto oom_free_page;
2638 entry = mk_pte(page, vma->vm_page_prot);
2639 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2641 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2642 if (!pte_none(*page_table))
2643 goto release;
2644 inc_mm_counter(mm, anon_rss);
2645 page_add_new_anon_rmap(page, vma, address);
2646 set_pte_at(mm, address, page_table, entry);
2648 /* No need to invalidate - it was non-present before */
2649 update_mmu_cache(vma, address, entry);
2650 unlock:
2651 pte_unmap_unlock(page_table, ptl);
2652 return 0;
2653 release:
2654 mem_cgroup_uncharge_page(page);
2655 page_cache_release(page);
2656 goto unlock;
2657 oom_free_page:
2658 page_cache_release(page);
2659 oom:
2660 return VM_FAULT_OOM;
2664 * __do_fault() tries to create a new page mapping. It aggressively
2665 * tries to share with existing pages, but makes a separate copy if
2666 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2667 * the next page fault.
2669 * As this is called only for pages that do not currently exist, we
2670 * do not need to flush old virtual caches or the TLB.
2672 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2673 * but allow concurrent faults), and pte neither mapped nor locked.
2674 * We return with mmap_sem still held, but pte unmapped and unlocked.
2676 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2677 unsigned long address, pmd_t *pmd,
2678 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2680 pte_t *page_table;
2681 spinlock_t *ptl;
2682 struct page *page;
2683 pte_t entry;
2684 int anon = 0;
2685 int charged = 0;
2686 struct page *dirty_page = NULL;
2687 struct vm_fault vmf;
2688 int ret;
2689 int page_mkwrite = 0;
2691 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2692 vmf.pgoff = pgoff;
2693 vmf.flags = flags;
2694 vmf.page = NULL;
2696 ret = vma->vm_ops->fault(vma, &vmf);
2697 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2698 return ret;
2701 * For consistency in subsequent calls, make the faulted page always
2702 * locked.
2704 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2705 lock_page(vmf.page);
2706 else
2707 VM_BUG_ON(!PageLocked(vmf.page));
2710 * Should we do an early C-O-W break?
2712 page = vmf.page;
2713 if (flags & FAULT_FLAG_WRITE) {
2714 if (!(vma->vm_flags & VM_SHARED)) {
2715 anon = 1;
2716 if (unlikely(anon_vma_prepare(vma))) {
2717 ret = VM_FAULT_OOM;
2718 goto out;
2720 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2721 vma, address);
2722 if (!page) {
2723 ret = VM_FAULT_OOM;
2724 goto out;
2726 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2727 ret = VM_FAULT_OOM;
2728 page_cache_release(page);
2729 goto out;
2731 charged = 1;
2733 * Don't let another task, with possibly unlocked vma,
2734 * keep the mlocked page.
2736 if (vma->vm_flags & VM_LOCKED)
2737 clear_page_mlock(vmf.page);
2738 copy_user_highpage(page, vmf.page, address, vma);
2739 __SetPageUptodate(page);
2740 } else {
2742 * If the page will be shareable, see if the backing
2743 * address space wants to know that the page is about
2744 * to become writable
2746 if (vma->vm_ops->page_mkwrite) {
2747 int tmp;
2749 unlock_page(page);
2750 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2751 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2752 if (unlikely(tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2754 ret = tmp;
2755 goto unwritable_page;
2757 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2758 lock_page(page);
2759 if (!page->mapping) {
2760 ret = 0; /* retry the fault */
2761 unlock_page(page);
2762 goto unwritable_page;
2764 } else
2765 VM_BUG_ON(!PageLocked(page));
2766 page_mkwrite = 1;
2772 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2775 * This silly early PAGE_DIRTY setting removes a race
2776 * due to the bad i386 page protection. But it's valid
2777 * for other architectures too.
2779 * Note that if write_access is true, we either now have
2780 * an exclusive copy of the page, or this is a shared mapping,
2781 * so we can make it writable and dirty to avoid having to
2782 * handle that later.
2784 /* Only go through if we didn't race with anybody else... */
2785 if (likely(pte_same(*page_table, orig_pte))) {
2786 flush_icache_page(vma, page);
2787 entry = mk_pte(page, vma->vm_page_prot);
2788 if (flags & FAULT_FLAG_WRITE)
2789 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2790 if (anon) {
2791 inc_mm_counter(mm, anon_rss);
2792 page_add_new_anon_rmap(page, vma, address);
2793 } else {
2794 inc_mm_counter(mm, file_rss);
2795 page_add_file_rmap(page);
2796 if (flags & FAULT_FLAG_WRITE) {
2797 dirty_page = page;
2798 get_page(dirty_page);
2801 set_pte_at(mm, address, page_table, entry);
2803 /* no need to invalidate: a not-present page won't be cached */
2804 update_mmu_cache(vma, address, entry);
2805 } else {
2806 if (charged)
2807 mem_cgroup_uncharge_page(page);
2808 if (anon)
2809 page_cache_release(page);
2810 else
2811 anon = 1; /* no anon but release faulted_page */
2814 pte_unmap_unlock(page_table, ptl);
2816 out:
2817 if (dirty_page) {
2818 struct address_space *mapping = page->mapping;
2820 if (set_page_dirty(dirty_page))
2821 page_mkwrite = 1;
2822 unlock_page(dirty_page);
2823 put_page(dirty_page);
2824 if (page_mkwrite && mapping) {
2826 * Some device drivers do not set page.mapping but still
2827 * dirty their pages
2829 balance_dirty_pages_ratelimited(mapping);
2832 /* file_update_time outside page_lock */
2833 if (vma->vm_file)
2834 file_update_time(vma->vm_file);
2835 } else {
2836 unlock_page(vmf.page);
2837 if (anon)
2838 page_cache_release(vmf.page);
2841 return ret;
2843 unwritable_page:
2844 page_cache_release(page);
2845 return ret;
2848 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2849 unsigned long address, pte_t *page_table, pmd_t *pmd,
2850 int write_access, pte_t orig_pte)
2852 pgoff_t pgoff = (((address & PAGE_MASK)
2853 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2854 unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2856 pte_unmap(page_table);
2857 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2861 * Fault of a previously existing named mapping. Repopulate the pte
2862 * from the encoded file_pte if possible. This enables swappable
2863 * nonlinear vmas.
2865 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2866 * but allow concurrent faults), and pte mapped but not yet locked.
2867 * We return with mmap_sem still held, but pte unmapped and unlocked.
2869 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2870 unsigned long address, pte_t *page_table, pmd_t *pmd,
2871 int write_access, pte_t orig_pte)
2873 unsigned int flags = FAULT_FLAG_NONLINEAR |
2874 (write_access ? FAULT_FLAG_WRITE : 0);
2875 pgoff_t pgoff;
2877 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2878 return 0;
2880 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2882 * Page table corrupted: show pte and kill process.
2884 print_bad_pte(vma, address, orig_pte, NULL);
2885 return VM_FAULT_OOM;
2888 pgoff = pte_to_pgoff(orig_pte);
2889 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2893 * These routines also need to handle stuff like marking pages dirty
2894 * and/or accessed for architectures that don't do it in hardware (most
2895 * RISC architectures). The early dirtying is also good on the i386.
2897 * There is also a hook called "update_mmu_cache()" that architectures
2898 * with external mmu caches can use to update those (ie the Sparc or
2899 * PowerPC hashed page tables that act as extended TLBs).
2901 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2902 * but allow concurrent faults), and pte mapped but not yet locked.
2903 * We return with mmap_sem still held, but pte unmapped and unlocked.
2905 static inline int handle_pte_fault(struct mm_struct *mm,
2906 struct vm_area_struct *vma, unsigned long address,
2907 pte_t *pte, pmd_t *pmd, int write_access)
2909 pte_t entry;
2910 spinlock_t *ptl;
2912 entry = *pte;
2913 if (!pte_present(entry)) {
2914 if (pte_none(entry)) {
2915 if (vma->vm_ops) {
2916 if (likely(vma->vm_ops->fault))
2917 return do_linear_fault(mm, vma, address,
2918 pte, pmd, write_access, entry);
2920 return do_anonymous_page(mm, vma, address,
2921 pte, pmd, write_access);
2923 if (pte_file(entry))
2924 return do_nonlinear_fault(mm, vma, address,
2925 pte, pmd, write_access, entry);
2926 return do_swap_page(mm, vma, address,
2927 pte, pmd, write_access, entry);
2930 ptl = pte_lockptr(mm, pmd);
2931 spin_lock(ptl);
2932 if (unlikely(!pte_same(*pte, entry)))
2933 goto unlock;
2934 if (write_access) {
2935 if (!pte_write(entry))
2936 return do_wp_page(mm, vma, address,
2937 pte, pmd, ptl, entry);
2938 entry = pte_mkdirty(entry);
2940 entry = pte_mkyoung(entry);
2941 if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2942 update_mmu_cache(vma, address, entry);
2943 } else {
2945 * This is needed only for protection faults but the arch code
2946 * is not yet telling us if this is a protection fault or not.
2947 * This still avoids useless tlb flushes for .text page faults
2948 * with threads.
2950 if (write_access)
2951 flush_tlb_page(vma, address);
2953 unlock:
2954 pte_unmap_unlock(pte, ptl);
2955 return 0;
2959 * By the time we get here, we already hold the mm semaphore
2961 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2962 unsigned long address, int write_access)
2964 pgd_t *pgd;
2965 pud_t *pud;
2966 pmd_t *pmd;
2967 pte_t *pte;
2969 __set_current_state(TASK_RUNNING);
2971 count_vm_event(PGFAULT);
2973 if (unlikely(is_vm_hugetlb_page(vma)))
2974 return hugetlb_fault(mm, vma, address, write_access);
2976 pgd = pgd_offset(mm, address);
2977 pud = pud_alloc(mm, pgd, address);
2978 if (!pud)
2979 return VM_FAULT_OOM;
2980 pmd = pmd_alloc(mm, pud, address);
2981 if (!pmd)
2982 return VM_FAULT_OOM;
2983 pte = pte_alloc_map(mm, pmd, address);
2984 if (!pte)
2985 return VM_FAULT_OOM;
2987 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2990 #ifndef __PAGETABLE_PUD_FOLDED
2992 * Allocate page upper directory.
2993 * We've already handled the fast-path in-line.
2995 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2997 pud_t *new = pud_alloc_one(mm, address);
2998 if (!new)
2999 return -ENOMEM;
3001 smp_wmb(); /* See comment in __pte_alloc */
3003 spin_lock(&mm->page_table_lock);
3004 if (pgd_present(*pgd)) /* Another has populated it */
3005 pud_free(mm, new);
3006 else
3007 pgd_populate(mm, pgd, new);
3008 spin_unlock(&mm->page_table_lock);
3009 return 0;
3011 #endif /* __PAGETABLE_PUD_FOLDED */
3013 #ifndef __PAGETABLE_PMD_FOLDED
3015 * Allocate page middle directory.
3016 * We've already handled the fast-path in-line.
3018 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3020 pmd_t *new = pmd_alloc_one(mm, address);
3021 if (!new)
3022 return -ENOMEM;
3024 smp_wmb(); /* See comment in __pte_alloc */
3026 spin_lock(&mm->page_table_lock);
3027 #ifndef __ARCH_HAS_4LEVEL_HACK
3028 if (pud_present(*pud)) /* Another has populated it */
3029 pmd_free(mm, new);
3030 else
3031 pud_populate(mm, pud, new);
3032 #else
3033 if (pgd_present(*pud)) /* Another has populated it */
3034 pmd_free(mm, new);
3035 else
3036 pgd_populate(mm, pud, new);
3037 #endif /* __ARCH_HAS_4LEVEL_HACK */
3038 spin_unlock(&mm->page_table_lock);
3039 return 0;
3041 #endif /* __PAGETABLE_PMD_FOLDED */
3043 int make_pages_present(unsigned long addr, unsigned long end)
3045 int ret, len, write;
3046 struct vm_area_struct * vma;
3048 vma = find_vma(current->mm, addr);
3049 if (!vma)
3050 return -ENOMEM;
3051 write = (vma->vm_flags & VM_WRITE) != 0;
3052 BUG_ON(addr >= end);
3053 BUG_ON(end > vma->vm_end);
3054 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3055 ret = get_user_pages(current, current->mm, addr,
3056 len, write, 0, NULL, NULL);
3057 if (ret < 0)
3058 return ret;
3059 return ret == len ? 0 : -EFAULT;
3062 #if !defined(__HAVE_ARCH_GATE_AREA)
3064 #if defined(AT_SYSINFO_EHDR)
3065 static struct vm_area_struct gate_vma;
3067 static int __init gate_vma_init(void)
3069 gate_vma.vm_mm = NULL;
3070 gate_vma.vm_start = FIXADDR_USER_START;
3071 gate_vma.vm_end = FIXADDR_USER_END;
3072 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3073 gate_vma.vm_page_prot = __P101;
3075 * Make sure the vDSO gets into every core dump.
3076 * Dumping its contents makes post-mortem fully interpretable later
3077 * without matching up the same kernel and hardware config to see
3078 * what PC values meant.
3080 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3081 return 0;
3083 __initcall(gate_vma_init);
3084 #endif
3086 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3088 #ifdef AT_SYSINFO_EHDR
3089 return &gate_vma;
3090 #else
3091 return NULL;
3092 #endif
3095 int in_gate_area_no_task(unsigned long addr)
3097 #ifdef AT_SYSINFO_EHDR
3098 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3099 return 1;
3100 #endif
3101 return 0;
3104 #endif /* __HAVE_ARCH_GATE_AREA */
3106 static int follow_pte(struct mm_struct *mm, unsigned long address,
3107 pte_t **ptepp, spinlock_t **ptlp)
3109 pgd_t *pgd;
3110 pud_t *pud;
3111 pmd_t *pmd;
3112 pte_t *ptep;
3114 pgd = pgd_offset(mm, address);
3115 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3116 goto out;
3118 pud = pud_offset(pgd, address);
3119 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3120 goto out;
3122 pmd = pmd_offset(pud, address);
3123 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3124 goto out;
3126 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3127 if (pmd_huge(*pmd))
3128 goto out;
3130 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3131 if (!ptep)
3132 goto out;
3133 if (!pte_present(*ptep))
3134 goto unlock;
3135 *ptepp = ptep;
3136 return 0;
3137 unlock:
3138 pte_unmap_unlock(ptep, *ptlp);
3139 out:
3140 return -EINVAL;
3144 * follow_pfn - look up PFN at a user virtual address
3145 * @vma: memory mapping
3146 * @address: user virtual address
3147 * @pfn: location to store found PFN
3149 * Only IO mappings and raw PFN mappings are allowed.
3151 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3153 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3154 unsigned long *pfn)
3156 int ret = -EINVAL;
3157 spinlock_t *ptl;
3158 pte_t *ptep;
3160 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3161 return ret;
3163 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3164 if (ret)
3165 return ret;
3166 *pfn = pte_pfn(*ptep);
3167 pte_unmap_unlock(ptep, ptl);
3168 return 0;
3170 EXPORT_SYMBOL(follow_pfn);
3172 #ifdef CONFIG_HAVE_IOREMAP_PROT
3173 int follow_phys(struct vm_area_struct *vma,
3174 unsigned long address, unsigned int flags,
3175 unsigned long *prot, resource_size_t *phys)
3177 int ret = -EINVAL;
3178 pte_t *ptep, pte;
3179 spinlock_t *ptl;
3181 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3182 goto out;
3184 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3185 goto out;
3186 pte = *ptep;
3188 if ((flags & FOLL_WRITE) && !pte_write(pte))
3189 goto unlock;
3191 *prot = pgprot_val(pte_pgprot(pte));
3192 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3194 ret = 0;
3195 unlock:
3196 pte_unmap_unlock(ptep, ptl);
3197 out:
3198 return ret;
3201 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3202 void *buf, int len, int write)
3204 resource_size_t phys_addr;
3205 unsigned long prot = 0;
3206 void __iomem *maddr;
3207 int offset = addr & (PAGE_SIZE-1);
3209 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3210 return -EINVAL;
3212 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3213 if (write)
3214 memcpy_toio(maddr + offset, buf, len);
3215 else
3216 memcpy_fromio(buf, maddr + offset, len);
3217 iounmap(maddr);
3219 return len;
3221 #endif
3224 * Access another process' address space.
3225 * Source/target buffer must be kernel space,
3226 * Do not walk the page table directly, use get_user_pages
3228 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3230 struct mm_struct *mm;
3231 struct vm_area_struct *vma;
3232 void *old_buf = buf;
3234 mm = get_task_mm(tsk);
3235 if (!mm)
3236 return 0;
3238 down_read(&mm->mmap_sem);
3239 /* ignore errors, just check how much was successfully transferred */
3240 while (len) {
3241 int bytes, ret, offset;
3242 void *maddr;
3243 struct page *page = NULL;
3245 ret = get_user_pages(tsk, mm, addr, 1,
3246 write, 1, &page, &vma);
3247 if (ret <= 0) {
3249 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3250 * we can access using slightly different code.
3252 #ifdef CONFIG_HAVE_IOREMAP_PROT
3253 vma = find_vma(mm, addr);
3254 if (!vma)
3255 break;
3256 if (vma->vm_ops && vma->vm_ops->access)
3257 ret = vma->vm_ops->access(vma, addr, buf,
3258 len, write);
3259 if (ret <= 0)
3260 #endif
3261 break;
3262 bytes = ret;
3263 } else {
3264 bytes = len;
3265 offset = addr & (PAGE_SIZE-1);
3266 if (bytes > PAGE_SIZE-offset)
3267 bytes = PAGE_SIZE-offset;
3269 maddr = kmap(page);
3270 if (write) {
3271 copy_to_user_page(vma, page, addr,
3272 maddr + offset, buf, bytes);
3273 set_page_dirty_lock(page);
3274 } else {
3275 copy_from_user_page(vma, page, addr,
3276 buf, maddr + offset, bytes);
3278 kunmap(page);
3279 page_cache_release(page);
3281 len -= bytes;
3282 buf += bytes;
3283 addr += bytes;
3285 up_read(&mm->mmap_sem);
3286 mmput(mm);
3288 return buf - old_buf;
3292 * Print the name of a VMA.
3294 void print_vma_addr(char *prefix, unsigned long ip)
3296 struct mm_struct *mm = current->mm;
3297 struct vm_area_struct *vma;
3300 * Do not print if we are in atomic
3301 * contexts (in exception stacks, etc.):
3303 if (preempt_count())
3304 return;
3306 down_read(&mm->mmap_sem);
3307 vma = find_vma(mm, ip);
3308 if (vma && vma->vm_file) {
3309 struct file *f = vma->vm_file;
3310 char *buf = (char *)__get_free_page(GFP_KERNEL);
3311 if (buf) {
3312 char *p, *s;
3314 p = d_path(&f->f_path, buf, PAGE_SIZE);
3315 if (IS_ERR(p))
3316 p = "?";
3317 s = strrchr(p, '/');
3318 if (s)
3319 p = s+1;
3320 printk("%s%s[%lx+%lx]", prefix, p,
3321 vma->vm_start,
3322 vma->vm_end - vma->vm_start);
3323 free_page((unsigned long)buf);
3326 up_read(&current->mm->mmap_sem);
3329 #ifdef CONFIG_PROVE_LOCKING
3330 void might_fault(void)
3333 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3334 * holding the mmap_sem, this is safe because kernel memory doesn't
3335 * get paged out, therefore we'll never actually fault, and the
3336 * below annotations will generate false positives.
3338 if (segment_eq(get_fs(), KERNEL_DS))
3339 return;
3341 might_sleep();
3343 * it would be nicer only to annotate paths which are not under
3344 * pagefault_disable, however that requires a larger audit and
3345 * providing helpers like get_user_atomic.
3347 if (!in_atomic() && current->mm)
3348 might_lock_read(&current->mm->mmap_sem);
3350 EXPORT_SYMBOL(might_fault);
3351 #endif