tpm_tis: Introduce durations sysfs entry
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / memory.c
blob40b7531ee8badb288823b64715265d0f16be397d
1 /*
2 * linux/mm/memory.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
7 /*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
68 #include "internal.h"
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
79 unsigned long num_physpages;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 * and ZONE_HIGHMEM.
87 void * high_memory;
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
101 #else
103 #endif
105 static int __init disable_randmaps(char *s)
107 randomize_va_space = 0;
108 return 1;
110 __setup("norandmaps", disable_randmaps);
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init init_zero_pfn(void)
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
123 core_initcall(init_zero_pfn);
126 #if defined(SPLIT_RSS_COUNTING)
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
130 int i;
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
138 task->rss_stat.events = 0;
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 struct task_struct *task = current;
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct *task)
157 if (unlikely(task != current))
158 return;
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
165 long val = 0;
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
171 val = atomic_long_read(&mm->rss_stat.count[member]);
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
176 if (val < 0)
177 return 0;
178 return (unsigned long)val;
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
183 __sync_task_rss_stat(task, mm);
185 #else /* SPLIT_RSS_COUNTING */
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
190 static void check_sync_rss_stat(struct task_struct *task)
194 #endif /* SPLIT_RSS_COUNTING */
196 #ifdef HAVE_GENERIC_MMU_GATHER
198 static int tlb_next_batch(struct mmu_gather *tlb)
200 struct mmu_gather_batch *batch;
202 batch = tlb->active;
203 if (batch->next) {
204 tlb->active = batch->next;
205 return 1;
208 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209 if (!batch)
210 return 0;
212 batch->next = NULL;
213 batch->nr = 0;
214 batch->max = MAX_GATHER_BATCH;
216 tlb->active->next = batch;
217 tlb->active = batch;
219 return 1;
222 /* tlb_gather_mmu
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
227 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
229 tlb->mm = mm;
231 tlb->fullmm = fullmm;
232 tlb->need_flush = 0;
233 tlb->fast_mode = (num_possible_cpus() == 1);
234 tlb->local.next = NULL;
235 tlb->local.nr = 0;
236 tlb->local.max = ARRAY_SIZE(tlb->__pages);
237 tlb->active = &tlb->local;
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb->batch = NULL;
241 #endif
244 void tlb_flush_mmu(struct mmu_gather *tlb)
246 struct mmu_gather_batch *batch;
248 if (!tlb->need_flush)
249 return;
250 tlb->need_flush = 0;
251 tlb_flush(tlb);
252 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
253 tlb_table_flush(tlb);
254 #endif
256 if (tlb_fast_mode(tlb))
257 return;
259 for (batch = &tlb->local; batch; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
261 batch->nr = 0;
263 tlb->active = &tlb->local;
266 /* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
272 struct mmu_gather_batch *batch, *next;
274 tlb_flush_mmu(tlb);
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
283 tlb->local.next = NULL;
286 /* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
294 struct mmu_gather_batch *batch;
296 tlb->need_flush = 1;
298 if (tlb_fast_mode(tlb)) {
299 free_page_and_swap_cache(page);
300 return 1; /* avoid calling tlb_flush_mmu() */
303 batch = tlb->active;
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
307 return 0;
309 VM_BUG_ON(batch->nr > batch->max);
311 return batch->max - batch->nr;
314 #endif /* HAVE_GENERIC_MMU_GATHER */
316 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
319 * See the comment near struct mmu_table_batch.
322 static void tlb_remove_table_smp_sync(void *arg)
324 /* Simply deliver the interrupt */
327 static void tlb_remove_table_one(void *table)
330 * This isn't an RCU grace period and hence the page-tables cannot be
331 * assumed to be actually RCU-freed.
333 * It is however sufficient for software page-table walkers that rely on
334 * IRQ disabling. See the comment near struct mmu_table_batch.
336 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
337 __tlb_remove_table(table);
340 static void tlb_remove_table_rcu(struct rcu_head *head)
342 struct mmu_table_batch *batch;
343 int i;
345 batch = container_of(head, struct mmu_table_batch, rcu);
347 for (i = 0; i < batch->nr; i++)
348 __tlb_remove_table(batch->tables[i]);
350 free_page((unsigned long)batch);
353 void tlb_table_flush(struct mmu_gather *tlb)
355 struct mmu_table_batch **batch = &tlb->batch;
357 if (*batch) {
358 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
359 *batch = NULL;
363 void tlb_remove_table(struct mmu_gather *tlb, void *table)
365 struct mmu_table_batch **batch = &tlb->batch;
367 tlb->need_flush = 1;
370 * When there's less then two users of this mm there cannot be a
371 * concurrent page-table walk.
373 if (atomic_read(&tlb->mm->mm_users) < 2) {
374 __tlb_remove_table(table);
375 return;
378 if (*batch == NULL) {
379 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
380 if (*batch == NULL) {
381 tlb_remove_table_one(table);
382 return;
384 (*batch)->nr = 0;
386 (*batch)->tables[(*batch)->nr++] = table;
387 if ((*batch)->nr == MAX_TABLE_BATCH)
388 tlb_table_flush(tlb);
391 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
394 * If a p?d_bad entry is found while walking page tables, report
395 * the error, before resetting entry to p?d_none. Usually (but
396 * very seldom) called out from the p?d_none_or_clear_bad macros.
399 void pgd_clear_bad(pgd_t *pgd)
401 pgd_ERROR(*pgd);
402 pgd_clear(pgd);
405 void pud_clear_bad(pud_t *pud)
407 pud_ERROR(*pud);
408 pud_clear(pud);
411 void pmd_clear_bad(pmd_t *pmd)
413 pmd_ERROR(*pmd);
414 pmd_clear(pmd);
418 * Note: this doesn't free the actual pages themselves. That
419 * has been handled earlier when unmapping all the memory regions.
421 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
422 unsigned long addr)
424 pgtable_t token = pmd_pgtable(*pmd);
425 pmd_clear(pmd);
426 pte_free_tlb(tlb, token, addr);
427 tlb->mm->nr_ptes--;
430 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
431 unsigned long addr, unsigned long end,
432 unsigned long floor, unsigned long ceiling)
434 pmd_t *pmd;
435 unsigned long next;
436 unsigned long start;
438 start = addr;
439 pmd = pmd_offset(pud, addr);
440 do {
441 next = pmd_addr_end(addr, end);
442 if (pmd_none_or_clear_bad(pmd))
443 continue;
444 free_pte_range(tlb, pmd, addr);
445 } while (pmd++, addr = next, addr != end);
447 start &= PUD_MASK;
448 if (start < floor)
449 return;
450 if (ceiling) {
451 ceiling &= PUD_MASK;
452 if (!ceiling)
453 return;
455 if (end - 1 > ceiling - 1)
456 return;
458 pmd = pmd_offset(pud, start);
459 pud_clear(pud);
460 pmd_free_tlb(tlb, pmd, start);
463 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
464 unsigned long addr, unsigned long end,
465 unsigned long floor, unsigned long ceiling)
467 pud_t *pud;
468 unsigned long next;
469 unsigned long start;
471 start = addr;
472 pud = pud_offset(pgd, addr);
473 do {
474 next = pud_addr_end(addr, end);
475 if (pud_none_or_clear_bad(pud))
476 continue;
477 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
478 } while (pud++, addr = next, addr != end);
480 start &= PGDIR_MASK;
481 if (start < floor)
482 return;
483 if (ceiling) {
484 ceiling &= PGDIR_MASK;
485 if (!ceiling)
486 return;
488 if (end - 1 > ceiling - 1)
489 return;
491 pud = pud_offset(pgd, start);
492 pgd_clear(pgd);
493 pud_free_tlb(tlb, pud, start);
497 * This function frees user-level page tables of a process.
499 * Must be called with pagetable lock held.
501 void free_pgd_range(struct mmu_gather *tlb,
502 unsigned long addr, unsigned long end,
503 unsigned long floor, unsigned long ceiling)
505 pgd_t *pgd;
506 unsigned long next;
509 * The next few lines have given us lots of grief...
511 * Why are we testing PMD* at this top level? Because often
512 * there will be no work to do at all, and we'd prefer not to
513 * go all the way down to the bottom just to discover that.
515 * Why all these "- 1"s? Because 0 represents both the bottom
516 * of the address space and the top of it (using -1 for the
517 * top wouldn't help much: the masks would do the wrong thing).
518 * The rule is that addr 0 and floor 0 refer to the bottom of
519 * the address space, but end 0 and ceiling 0 refer to the top
520 * Comparisons need to use "end - 1" and "ceiling - 1" (though
521 * that end 0 case should be mythical).
523 * Wherever addr is brought up or ceiling brought down, we must
524 * be careful to reject "the opposite 0" before it confuses the
525 * subsequent tests. But what about where end is brought down
526 * by PMD_SIZE below? no, end can't go down to 0 there.
528 * Whereas we round start (addr) and ceiling down, by different
529 * masks at different levels, in order to test whether a table
530 * now has no other vmas using it, so can be freed, we don't
531 * bother to round floor or end up - the tests don't need that.
534 addr &= PMD_MASK;
535 if (addr < floor) {
536 addr += PMD_SIZE;
537 if (!addr)
538 return;
540 if (ceiling) {
541 ceiling &= PMD_MASK;
542 if (!ceiling)
543 return;
545 if (end - 1 > ceiling - 1)
546 end -= PMD_SIZE;
547 if (addr > end - 1)
548 return;
550 pgd = pgd_offset(tlb->mm, addr);
551 do {
552 next = pgd_addr_end(addr, end);
553 if (pgd_none_or_clear_bad(pgd))
554 continue;
555 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
556 } while (pgd++, addr = next, addr != end);
559 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
560 unsigned long floor, unsigned long ceiling)
562 while (vma) {
563 struct vm_area_struct *next = vma->vm_next;
564 unsigned long addr = vma->vm_start;
567 * Hide vma from rmap and truncate_pagecache before freeing
568 * pgtables
570 unlink_anon_vmas(vma);
571 unlink_file_vma(vma);
573 if (is_vm_hugetlb_page(vma)) {
574 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
575 floor, next? next->vm_start: ceiling);
576 } else {
578 * Optimization: gather nearby vmas into one call down
580 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
581 && !is_vm_hugetlb_page(next)) {
582 vma = next;
583 next = vma->vm_next;
584 unlink_anon_vmas(vma);
585 unlink_file_vma(vma);
587 free_pgd_range(tlb, addr, vma->vm_end,
588 floor, next? next->vm_start: ceiling);
590 vma = next;
594 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
595 pmd_t *pmd, unsigned long address)
597 pgtable_t new = pte_alloc_one(mm, address);
598 int wait_split_huge_page;
599 if (!new)
600 return -ENOMEM;
603 * Ensure all pte setup (eg. pte page lock and page clearing) are
604 * visible before the pte is made visible to other CPUs by being
605 * put into page tables.
607 * The other side of the story is the pointer chasing in the page
608 * table walking code (when walking the page table without locking;
609 * ie. most of the time). Fortunately, these data accesses consist
610 * of a chain of data-dependent loads, meaning most CPUs (alpha
611 * being the notable exception) will already guarantee loads are
612 * seen in-order. See the alpha page table accessors for the
613 * smp_read_barrier_depends() barriers in page table walking code.
615 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617 spin_lock(&mm->page_table_lock);
618 wait_split_huge_page = 0;
619 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
620 mm->nr_ptes++;
621 pmd_populate(mm, pmd, new);
622 new = NULL;
623 } else if (unlikely(pmd_trans_splitting(*pmd)))
624 wait_split_huge_page = 1;
625 spin_unlock(&mm->page_table_lock);
626 if (new)
627 pte_free(mm, new);
628 if (wait_split_huge_page)
629 wait_split_huge_page(vma->anon_vma, pmd);
630 return 0;
633 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
635 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
636 if (!new)
637 return -ENOMEM;
639 smp_wmb(); /* See comment in __pte_alloc */
641 spin_lock(&init_mm.page_table_lock);
642 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
643 pmd_populate_kernel(&init_mm, pmd, new);
644 new = NULL;
645 } else
646 VM_BUG_ON(pmd_trans_splitting(*pmd));
647 spin_unlock(&init_mm.page_table_lock);
648 if (new)
649 pte_free_kernel(&init_mm, new);
650 return 0;
653 static inline void init_rss_vec(int *rss)
655 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
658 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660 int i;
662 if (current->mm == mm)
663 sync_mm_rss(current, mm);
664 for (i = 0; i < NR_MM_COUNTERS; i++)
665 if (rss[i])
666 add_mm_counter(mm, i, rss[i]);
670 * This function is called to print an error when a bad pte
671 * is found. For example, we might have a PFN-mapped pte in
672 * a region that doesn't allow it.
674 * The calling function must still handle the error.
676 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
677 pte_t pte, struct page *page)
679 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
680 pud_t *pud = pud_offset(pgd, addr);
681 pmd_t *pmd = pmd_offset(pud, addr);
682 struct address_space *mapping;
683 pgoff_t index;
684 static unsigned long resume;
685 static unsigned long nr_shown;
686 static unsigned long nr_unshown;
689 * Allow a burst of 60 reports, then keep quiet for that minute;
690 * or allow a steady drip of one report per second.
692 if (nr_shown == 60) {
693 if (time_before(jiffies, resume)) {
694 nr_unshown++;
695 return;
697 if (nr_unshown) {
698 printk(KERN_ALERT
699 "BUG: Bad page map: %lu messages suppressed\n",
700 nr_unshown);
701 nr_unshown = 0;
703 nr_shown = 0;
705 if (nr_shown++ == 0)
706 resume = jiffies + 60 * HZ;
708 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
709 index = linear_page_index(vma, addr);
711 printk(KERN_ALERT
712 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
713 current->comm,
714 (long long)pte_val(pte), (long long)pmd_val(*pmd));
715 if (page)
716 dump_page(page);
717 printk(KERN_ALERT
718 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
719 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723 if (vma->vm_ops)
724 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
725 (unsigned long)vma->vm_ops->fault);
726 if (vma->vm_file && vma->vm_file->f_op)
727 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
728 (unsigned long)vma->vm_file->f_op->mmap);
729 dump_stack();
730 add_taint(TAINT_BAD_PAGE);
733 static inline int is_cow_mapping(vm_flags_t flags)
735 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
738 #ifndef is_zero_pfn
739 static inline int is_zero_pfn(unsigned long pfn)
741 return pfn == zero_pfn;
743 #endif
745 #ifndef my_zero_pfn
746 static inline unsigned long my_zero_pfn(unsigned long addr)
748 return zero_pfn;
750 #endif
753 * vm_normal_page -- This function gets the "struct page" associated with a pte.
755 * "Special" mappings do not wish to be associated with a "struct page" (either
756 * it doesn't exist, or it exists but they don't want to touch it). In this
757 * case, NULL is returned here. "Normal" mappings do have a struct page.
759 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
760 * pte bit, in which case this function is trivial. Secondly, an architecture
761 * may not have a spare pte bit, which requires a more complicated scheme,
762 * described below.
764 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
765 * special mapping (even if there are underlying and valid "struct pages").
766 * COWed pages of a VM_PFNMAP are always normal.
768 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
769 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
770 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
771 * mapping will always honor the rule
773 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775 * And for normal mappings this is false.
777 * This restricts such mappings to be a linear translation from virtual address
778 * to pfn. To get around this restriction, we allow arbitrary mappings so long
779 * as the vma is not a COW mapping; in that case, we know that all ptes are
780 * special (because none can have been COWed).
783 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
786 * page" backing, however the difference is that _all_ pages with a struct
787 * page (that is, those where pfn_valid is true) are refcounted and considered
788 * normal pages by the VM. The disadvantage is that pages are refcounted
789 * (which can be slower and simply not an option for some PFNMAP users). The
790 * advantage is that we don't have to follow the strict linearity rule of
791 * PFNMAP mappings in order to support COWable mappings.
794 #ifdef __HAVE_ARCH_PTE_SPECIAL
795 # define HAVE_PTE_SPECIAL 1
796 #else
797 # define HAVE_PTE_SPECIAL 0
798 #endif
799 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
800 pte_t pte)
802 unsigned long pfn = pte_pfn(pte);
804 if (HAVE_PTE_SPECIAL) {
805 if (likely(!pte_special(pte)))
806 goto check_pfn;
807 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
808 return NULL;
809 if (!is_zero_pfn(pfn))
810 print_bad_pte(vma, addr, pte, NULL);
811 return NULL;
814 /* !HAVE_PTE_SPECIAL case follows: */
816 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
817 if (vma->vm_flags & VM_MIXEDMAP) {
818 if (!pfn_valid(pfn))
819 return NULL;
820 goto out;
821 } else {
822 unsigned long off;
823 off = (addr - vma->vm_start) >> PAGE_SHIFT;
824 if (pfn == vma->vm_pgoff + off)
825 return NULL;
826 if (!is_cow_mapping(vma->vm_flags))
827 return NULL;
831 if (is_zero_pfn(pfn))
832 return NULL;
833 check_pfn:
834 if (unlikely(pfn > highest_memmap_pfn)) {
835 print_bad_pte(vma, addr, pte, NULL);
836 return NULL;
840 * NOTE! We still have PageReserved() pages in the page tables.
841 * eg. VDSO mappings can cause them to exist.
843 out:
844 return pfn_to_page(pfn);
848 * copy one vm_area from one task to the other. Assumes the page tables
849 * already present in the new task to be cleared in the whole range
850 * covered by this vma.
853 static inline unsigned long
854 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
855 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
856 unsigned long addr, int *rss)
858 unsigned long vm_flags = vma->vm_flags;
859 pte_t pte = *src_pte;
860 struct page *page;
862 /* pte contains position in swap or file, so copy. */
863 if (unlikely(!pte_present(pte))) {
864 if (!pte_file(pte)) {
865 swp_entry_t entry = pte_to_swp_entry(pte);
867 if (swap_duplicate(entry) < 0)
868 return entry.val;
870 /* make sure dst_mm is on swapoff's mmlist. */
871 if (unlikely(list_empty(&dst_mm->mmlist))) {
872 spin_lock(&mmlist_lock);
873 if (list_empty(&dst_mm->mmlist))
874 list_add(&dst_mm->mmlist,
875 &src_mm->mmlist);
876 spin_unlock(&mmlist_lock);
878 if (likely(!non_swap_entry(entry)))
879 rss[MM_SWAPENTS]++;
880 else if (is_write_migration_entry(entry) &&
881 is_cow_mapping(vm_flags)) {
883 * COW mappings require pages in both parent
884 * and child to be set to read.
886 make_migration_entry_read(&entry);
887 pte = swp_entry_to_pte(entry);
888 set_pte_at(src_mm, addr, src_pte, pte);
891 goto out_set_pte;
895 * If it's a COW mapping, write protect it both
896 * in the parent and the child
898 if (is_cow_mapping(vm_flags)) {
899 ptep_set_wrprotect(src_mm, addr, src_pte);
900 pte = pte_wrprotect(pte);
904 * If it's a shared mapping, mark it clean in
905 * the child
907 if (vm_flags & VM_SHARED)
908 pte = pte_mkclean(pte);
909 pte = pte_mkold(pte);
911 page = vm_normal_page(vma, addr, pte);
912 if (page) {
913 get_page(page);
914 page_dup_rmap(page);
915 if (PageAnon(page))
916 rss[MM_ANONPAGES]++;
917 else
918 rss[MM_FILEPAGES]++;
921 out_set_pte:
922 set_pte_at(dst_mm, addr, dst_pte, pte);
923 return 0;
926 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
927 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
928 unsigned long addr, unsigned long end)
930 pte_t *orig_src_pte, *orig_dst_pte;
931 pte_t *src_pte, *dst_pte;
932 spinlock_t *src_ptl, *dst_ptl;
933 int progress = 0;
934 int rss[NR_MM_COUNTERS];
935 swp_entry_t entry = (swp_entry_t){0};
937 again:
938 init_rss_vec(rss);
940 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
941 if (!dst_pte)
942 return -ENOMEM;
943 src_pte = pte_offset_map(src_pmd, addr);
944 src_ptl = pte_lockptr(src_mm, src_pmd);
945 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
946 orig_src_pte = src_pte;
947 orig_dst_pte = dst_pte;
948 arch_enter_lazy_mmu_mode();
950 do {
952 * We are holding two locks at this point - either of them
953 * could generate latencies in another task on another CPU.
955 if (progress >= 32) {
956 progress = 0;
957 if (need_resched() ||
958 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
959 break;
961 if (pte_none(*src_pte)) {
962 progress++;
963 continue;
965 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
966 vma, addr, rss);
967 if (entry.val)
968 break;
969 progress += 8;
970 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
972 arch_leave_lazy_mmu_mode();
973 spin_unlock(src_ptl);
974 pte_unmap(orig_src_pte);
975 add_mm_rss_vec(dst_mm, rss);
976 pte_unmap_unlock(orig_dst_pte, dst_ptl);
977 cond_resched();
979 if (entry.val) {
980 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
981 return -ENOMEM;
982 progress = 0;
984 if (addr != end)
985 goto again;
986 return 0;
989 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
990 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
991 unsigned long addr, unsigned long end)
993 pmd_t *src_pmd, *dst_pmd;
994 unsigned long next;
996 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
997 if (!dst_pmd)
998 return -ENOMEM;
999 src_pmd = pmd_offset(src_pud, addr);
1000 do {
1001 next = pmd_addr_end(addr, end);
1002 if (pmd_trans_huge(*src_pmd)) {
1003 int err;
1004 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1005 err = copy_huge_pmd(dst_mm, src_mm,
1006 dst_pmd, src_pmd, addr, vma);
1007 if (err == -ENOMEM)
1008 return -ENOMEM;
1009 if (!err)
1010 continue;
1011 /* fall through */
1013 if (pmd_none_or_clear_bad(src_pmd))
1014 continue;
1015 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1019 return 0;
1022 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1024 unsigned long addr, unsigned long end)
1026 pud_t *src_pud, *dst_pud;
1027 unsigned long next;
1029 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030 if (!dst_pud)
1031 return -ENOMEM;
1032 src_pud = pud_offset(src_pgd, addr);
1033 do {
1034 next = pud_addr_end(addr, end);
1035 if (pud_none_or_clear_bad(src_pud))
1036 continue;
1037 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038 vma, addr, next))
1039 return -ENOMEM;
1040 } while (dst_pud++, src_pud++, addr = next, addr != end);
1041 return 0;
1044 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 struct vm_area_struct *vma)
1047 pgd_t *src_pgd, *dst_pgd;
1048 unsigned long next;
1049 unsigned long addr = vma->vm_start;
1050 unsigned long end = vma->vm_end;
1051 int ret;
1054 * Don't copy ptes where a page fault will fill them correctly.
1055 * Fork becomes much lighter when there are big shared or private
1056 * readonly mappings. The tradeoff is that copy_page_range is more
1057 * efficient than faulting.
1059 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1060 if (!vma->anon_vma)
1061 return 0;
1064 if (is_vm_hugetlb_page(vma))
1065 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067 if (unlikely(is_pfn_mapping(vma))) {
1069 * We do not free on error cases below as remove_vma
1070 * gets called on error from higher level routine
1072 ret = track_pfn_vma_copy(vma);
1073 if (ret)
1074 return ret;
1078 * We need to invalidate the secondary MMU mappings only when
1079 * there could be a permission downgrade on the ptes of the
1080 * parent mm. And a permission downgrade will only happen if
1081 * is_cow_mapping() returns true.
1083 if (is_cow_mapping(vma->vm_flags))
1084 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1086 ret = 0;
1087 dst_pgd = pgd_offset(dst_mm, addr);
1088 src_pgd = pgd_offset(src_mm, addr);
1089 do {
1090 next = pgd_addr_end(addr, end);
1091 if (pgd_none_or_clear_bad(src_pgd))
1092 continue;
1093 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1094 vma, addr, next))) {
1095 ret = -ENOMEM;
1096 break;
1098 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1100 if (is_cow_mapping(vma->vm_flags))
1101 mmu_notifier_invalidate_range_end(src_mm,
1102 vma->vm_start, end);
1103 return ret;
1106 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1107 struct vm_area_struct *vma, pmd_t *pmd,
1108 unsigned long addr, unsigned long end,
1109 struct zap_details *details)
1111 struct mm_struct *mm = tlb->mm;
1112 int force_flush = 0;
1113 int rss[NR_MM_COUNTERS];
1114 spinlock_t *ptl;
1115 pte_t *start_pte;
1116 pte_t *pte;
1118 again:
1119 init_rss_vec(rss);
1120 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1121 pte = start_pte;
1122 arch_enter_lazy_mmu_mode();
1123 do {
1124 pte_t ptent = *pte;
1125 if (pte_none(ptent)) {
1126 continue;
1129 if (pte_present(ptent)) {
1130 struct page *page;
1132 page = vm_normal_page(vma, addr, ptent);
1133 if (unlikely(details) && page) {
1135 * unmap_shared_mapping_pages() wants to
1136 * invalidate cache without truncating:
1137 * unmap shared but keep private pages.
1139 if (details->check_mapping &&
1140 details->check_mapping != page->mapping)
1141 continue;
1143 * Each page->index must be checked when
1144 * invalidating or truncating nonlinear.
1146 if (details->nonlinear_vma &&
1147 (page->index < details->first_index ||
1148 page->index > details->last_index))
1149 continue;
1151 ptent = ptep_get_and_clear_full(mm, addr, pte,
1152 tlb->fullmm);
1153 tlb_remove_tlb_entry(tlb, pte, addr);
1154 if (unlikely(!page))
1155 continue;
1156 if (unlikely(details) && details->nonlinear_vma
1157 && linear_page_index(details->nonlinear_vma,
1158 addr) != page->index)
1159 set_pte_at(mm, addr, pte,
1160 pgoff_to_pte(page->index));
1161 if (PageAnon(page))
1162 rss[MM_ANONPAGES]--;
1163 else {
1164 if (pte_dirty(ptent))
1165 set_page_dirty(page);
1166 if (pte_young(ptent) &&
1167 likely(!VM_SequentialReadHint(vma)))
1168 mark_page_accessed(page);
1169 rss[MM_FILEPAGES]--;
1171 page_remove_rmap(page);
1172 if (unlikely(page_mapcount(page) < 0))
1173 print_bad_pte(vma, addr, ptent, page);
1174 force_flush = !__tlb_remove_page(tlb, page);
1175 if (force_flush)
1176 break;
1177 continue;
1180 * If details->check_mapping, we leave swap entries;
1181 * if details->nonlinear_vma, we leave file entries.
1183 if (unlikely(details))
1184 continue;
1185 if (pte_file(ptent)) {
1186 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1187 print_bad_pte(vma, addr, ptent, NULL);
1188 } else {
1189 swp_entry_t entry = pte_to_swp_entry(ptent);
1191 if (!non_swap_entry(entry))
1192 rss[MM_SWAPENTS]--;
1193 if (unlikely(!free_swap_and_cache(entry)))
1194 print_bad_pte(vma, addr, ptent, NULL);
1196 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1197 } while (pte++, addr += PAGE_SIZE, addr != end);
1199 add_mm_rss_vec(mm, rss);
1200 arch_leave_lazy_mmu_mode();
1201 pte_unmap_unlock(start_pte, ptl);
1204 * mmu_gather ran out of room to batch pages, we break out of
1205 * the PTE lock to avoid doing the potential expensive TLB invalidate
1206 * and page-free while holding it.
1208 if (force_flush) {
1209 force_flush = 0;
1210 tlb_flush_mmu(tlb);
1211 if (addr != end)
1212 goto again;
1215 return addr;
1218 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1219 struct vm_area_struct *vma, pud_t *pud,
1220 unsigned long addr, unsigned long end,
1221 struct zap_details *details)
1223 pmd_t *pmd;
1224 unsigned long next;
1226 pmd = pmd_offset(pud, addr);
1227 do {
1228 next = pmd_addr_end(addr, end);
1229 if (pmd_trans_huge(*pmd)) {
1230 if (next-addr != HPAGE_PMD_SIZE) {
1231 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1232 split_huge_page_pmd(vma->vm_mm, pmd);
1233 } else if (zap_huge_pmd(tlb, vma, pmd))
1234 continue;
1235 /* fall through */
1237 if (pmd_none_or_clear_bad(pmd))
1238 continue;
1239 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1240 cond_resched();
1241 } while (pmd++, addr = next, addr != end);
1243 return addr;
1246 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1247 struct vm_area_struct *vma, pgd_t *pgd,
1248 unsigned long addr, unsigned long end,
1249 struct zap_details *details)
1251 pud_t *pud;
1252 unsigned long next;
1254 pud = pud_offset(pgd, addr);
1255 do {
1256 next = pud_addr_end(addr, end);
1257 if (pud_none_or_clear_bad(pud))
1258 continue;
1259 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1260 } while (pud++, addr = next, addr != end);
1262 return addr;
1265 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1266 struct vm_area_struct *vma,
1267 unsigned long addr, unsigned long end,
1268 struct zap_details *details)
1270 pgd_t *pgd;
1271 unsigned long next;
1273 if (details && !details->check_mapping && !details->nonlinear_vma)
1274 details = NULL;
1276 BUG_ON(addr >= end);
1277 mem_cgroup_uncharge_start();
1278 tlb_start_vma(tlb, vma);
1279 pgd = pgd_offset(vma->vm_mm, addr);
1280 do {
1281 next = pgd_addr_end(addr, end);
1282 if (pgd_none_or_clear_bad(pgd))
1283 continue;
1284 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1285 } while (pgd++, addr = next, addr != end);
1286 tlb_end_vma(tlb, vma);
1287 mem_cgroup_uncharge_end();
1289 return addr;
1292 #ifdef CONFIG_PREEMPT
1293 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1294 #else
1295 /* No preempt: go for improved straight-line efficiency */
1296 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1297 #endif
1300 * unmap_vmas - unmap a range of memory covered by a list of vma's
1301 * @tlb: address of the caller's struct mmu_gather
1302 * @vma: the starting vma
1303 * @start_addr: virtual address at which to start unmapping
1304 * @end_addr: virtual address at which to end unmapping
1305 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1306 * @details: details of nonlinear truncation or shared cache invalidation
1308 * Returns the end address of the unmapping (restart addr if interrupted).
1310 * Unmap all pages in the vma list.
1312 * We aim to not hold locks for too long (for scheduling latency reasons).
1313 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1314 * return the ending mmu_gather to the caller.
1316 * Only addresses between `start' and `end' will be unmapped.
1318 * The VMA list must be sorted in ascending virtual address order.
1320 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1321 * range after unmap_vmas() returns. So the only responsibility here is to
1322 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1323 * drops the lock and schedules.
1325 unsigned long unmap_vmas(struct mmu_gather *tlb,
1326 struct vm_area_struct *vma, unsigned long start_addr,
1327 unsigned long end_addr, unsigned long *nr_accounted,
1328 struct zap_details *details)
1330 unsigned long start = start_addr;
1331 struct mm_struct *mm = vma->vm_mm;
1333 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1334 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1335 unsigned long end;
1337 start = max(vma->vm_start, start_addr);
1338 if (start >= vma->vm_end)
1339 continue;
1340 end = min(vma->vm_end, end_addr);
1341 if (end <= vma->vm_start)
1342 continue;
1344 if (vma->vm_flags & VM_ACCOUNT)
1345 *nr_accounted += (end - start) >> PAGE_SHIFT;
1347 if (unlikely(is_pfn_mapping(vma)))
1348 untrack_pfn_vma(vma, 0, 0);
1350 while (start != end) {
1351 if (unlikely(is_vm_hugetlb_page(vma))) {
1353 * It is undesirable to test vma->vm_file as it
1354 * should be non-null for valid hugetlb area.
1355 * However, vm_file will be NULL in the error
1356 * cleanup path of do_mmap_pgoff. When
1357 * hugetlbfs ->mmap method fails,
1358 * do_mmap_pgoff() nullifies vma->vm_file
1359 * before calling this function to clean up.
1360 * Since no pte has actually been setup, it is
1361 * safe to do nothing in this case.
1363 if (vma->vm_file)
1364 unmap_hugepage_range(vma, start, end, NULL);
1366 start = end;
1367 } else
1368 start = unmap_page_range(tlb, vma, start, end, details);
1372 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1373 return start; /* which is now the end (or restart) address */
1377 * zap_page_range - remove user pages in a given range
1378 * @vma: vm_area_struct holding the applicable pages
1379 * @address: starting address of pages to zap
1380 * @size: number of bytes to zap
1381 * @details: details of nonlinear truncation or shared cache invalidation
1383 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1384 unsigned long size, struct zap_details *details)
1386 struct mm_struct *mm = vma->vm_mm;
1387 struct mmu_gather tlb;
1388 unsigned long end = address + size;
1389 unsigned long nr_accounted = 0;
1391 lru_add_drain();
1392 tlb_gather_mmu(&tlb, mm, 0);
1393 update_hiwater_rss(mm);
1394 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1395 tlb_finish_mmu(&tlb, address, end);
1396 return end;
1400 * zap_vma_ptes - remove ptes mapping the vma
1401 * @vma: vm_area_struct holding ptes to be zapped
1402 * @address: starting address of pages to zap
1403 * @size: number of bytes to zap
1405 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 * The entire address range must be fully contained within the vma.
1409 * Returns 0 if successful.
1411 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1412 unsigned long size)
1414 if (address < vma->vm_start || address + size > vma->vm_end ||
1415 !(vma->vm_flags & VM_PFNMAP))
1416 return -1;
1417 zap_page_range(vma, address, size, NULL);
1418 return 0;
1420 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1423 * follow_page - look up a page descriptor from a user-virtual address
1424 * @vma: vm_area_struct mapping @address
1425 * @address: virtual address to look up
1426 * @flags: flags modifying lookup behaviour
1428 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1430 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1431 * an error pointer if there is a mapping to something not represented
1432 * by a page descriptor (see also vm_normal_page()).
1434 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1435 unsigned int flags)
1437 pgd_t *pgd;
1438 pud_t *pud;
1439 pmd_t *pmd;
1440 pte_t *ptep, pte;
1441 spinlock_t *ptl;
1442 struct page *page;
1443 struct mm_struct *mm = vma->vm_mm;
1445 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1446 if (!IS_ERR(page)) {
1447 BUG_ON(flags & FOLL_GET);
1448 goto out;
1451 page = NULL;
1452 pgd = pgd_offset(mm, address);
1453 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1454 goto no_page_table;
1456 pud = pud_offset(pgd, address);
1457 if (pud_none(*pud))
1458 goto no_page_table;
1459 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1460 BUG_ON(flags & FOLL_GET);
1461 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1462 goto out;
1464 if (unlikely(pud_bad(*pud)))
1465 goto no_page_table;
1467 pmd = pmd_offset(pud, address);
1468 if (pmd_none(*pmd))
1469 goto no_page_table;
1470 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1471 BUG_ON(flags & FOLL_GET);
1472 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1473 goto out;
1475 if (pmd_trans_huge(*pmd)) {
1476 if (flags & FOLL_SPLIT) {
1477 split_huge_page_pmd(mm, pmd);
1478 goto split_fallthrough;
1480 spin_lock(&mm->page_table_lock);
1481 if (likely(pmd_trans_huge(*pmd))) {
1482 if (unlikely(pmd_trans_splitting(*pmd))) {
1483 spin_unlock(&mm->page_table_lock);
1484 wait_split_huge_page(vma->anon_vma, pmd);
1485 } else {
1486 page = follow_trans_huge_pmd(mm, address,
1487 pmd, flags);
1488 spin_unlock(&mm->page_table_lock);
1489 goto out;
1491 } else
1492 spin_unlock(&mm->page_table_lock);
1493 /* fall through */
1495 split_fallthrough:
1496 if (unlikely(pmd_bad(*pmd)))
1497 goto no_page_table;
1499 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1501 pte = *ptep;
1502 if (!pte_present(pte))
1503 goto no_page;
1504 if ((flags & FOLL_WRITE) && !pte_write(pte))
1505 goto unlock;
1507 page = vm_normal_page(vma, address, pte);
1508 if (unlikely(!page)) {
1509 if ((flags & FOLL_DUMP) ||
1510 !is_zero_pfn(pte_pfn(pte)))
1511 goto bad_page;
1512 page = pte_page(pte);
1515 if (flags & FOLL_GET)
1516 get_page(page);
1517 if (flags & FOLL_TOUCH) {
1518 if ((flags & FOLL_WRITE) &&
1519 !pte_dirty(pte) && !PageDirty(page))
1520 set_page_dirty(page);
1522 * pte_mkyoung() would be more correct here, but atomic care
1523 * is needed to avoid losing the dirty bit: it is easier to use
1524 * mark_page_accessed().
1526 mark_page_accessed(page);
1528 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1530 * The preliminary mapping check is mainly to avoid the
1531 * pointless overhead of lock_page on the ZERO_PAGE
1532 * which might bounce very badly if there is contention.
1534 * If the page is already locked, we don't need to
1535 * handle it now - vmscan will handle it later if and
1536 * when it attempts to reclaim the page.
1538 if (page->mapping && trylock_page(page)) {
1539 lru_add_drain(); /* push cached pages to LRU */
1541 * Because we lock page here and migration is
1542 * blocked by the pte's page reference, we need
1543 * only check for file-cache page truncation.
1545 if (page->mapping)
1546 mlock_vma_page(page);
1547 unlock_page(page);
1550 unlock:
1551 pte_unmap_unlock(ptep, ptl);
1552 out:
1553 return page;
1555 bad_page:
1556 pte_unmap_unlock(ptep, ptl);
1557 return ERR_PTR(-EFAULT);
1559 no_page:
1560 pte_unmap_unlock(ptep, ptl);
1561 if (!pte_none(pte))
1562 return page;
1564 no_page_table:
1566 * When core dumping an enormous anonymous area that nobody
1567 * has touched so far, we don't want to allocate unnecessary pages or
1568 * page tables. Return error instead of NULL to skip handle_mm_fault,
1569 * then get_dump_page() will return NULL to leave a hole in the dump.
1570 * But we can only make this optimization where a hole would surely
1571 * be zero-filled if handle_mm_fault() actually did handle it.
1573 if ((flags & FOLL_DUMP) &&
1574 (!vma->vm_ops || !vma->vm_ops->fault))
1575 return ERR_PTR(-EFAULT);
1576 return page;
1579 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1581 return stack_guard_page_start(vma, addr) ||
1582 stack_guard_page_end(vma, addr+PAGE_SIZE);
1586 * __get_user_pages() - pin user pages in memory
1587 * @tsk: task_struct of target task
1588 * @mm: mm_struct of target mm
1589 * @start: starting user address
1590 * @nr_pages: number of pages from start to pin
1591 * @gup_flags: flags modifying pin behaviour
1592 * @pages: array that receives pointers to the pages pinned.
1593 * Should be at least nr_pages long. Or NULL, if caller
1594 * only intends to ensure the pages are faulted in.
1595 * @vmas: array of pointers to vmas corresponding to each page.
1596 * Or NULL if the caller does not require them.
1597 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1599 * Returns number of pages pinned. This may be fewer than the number
1600 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1601 * were pinned, returns -errno. Each page returned must be released
1602 * with a put_page() call when it is finished with. vmas will only
1603 * remain valid while mmap_sem is held.
1605 * Must be called with mmap_sem held for read or write.
1607 * __get_user_pages walks a process's page tables and takes a reference to
1608 * each struct page that each user address corresponds to at a given
1609 * instant. That is, it takes the page that would be accessed if a user
1610 * thread accesses the given user virtual address at that instant.
1612 * This does not guarantee that the page exists in the user mappings when
1613 * __get_user_pages returns, and there may even be a completely different
1614 * page there in some cases (eg. if mmapped pagecache has been invalidated
1615 * and subsequently re faulted). However it does guarantee that the page
1616 * won't be freed completely. And mostly callers simply care that the page
1617 * contains data that was valid *at some point in time*. Typically, an IO
1618 * or similar operation cannot guarantee anything stronger anyway because
1619 * locks can't be held over the syscall boundary.
1621 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1622 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1623 * appropriate) must be called after the page is finished with, and
1624 * before put_page is called.
1626 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1627 * or mmap_sem contention, and if waiting is needed to pin all pages,
1628 * *@nonblocking will be set to 0.
1630 * In most cases, get_user_pages or get_user_pages_fast should be used
1631 * instead of __get_user_pages. __get_user_pages should be used only if
1632 * you need some special @gup_flags.
1634 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1635 unsigned long start, int nr_pages, unsigned int gup_flags,
1636 struct page **pages, struct vm_area_struct **vmas,
1637 int *nonblocking)
1639 int i;
1640 unsigned long vm_flags;
1642 if (nr_pages <= 0)
1643 return 0;
1645 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1648 * Require read or write permissions.
1649 * If FOLL_FORCE is set, we only require the "MAY" flags.
1651 vm_flags = (gup_flags & FOLL_WRITE) ?
1652 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1653 vm_flags &= (gup_flags & FOLL_FORCE) ?
1654 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1655 i = 0;
1657 do {
1658 struct vm_area_struct *vma;
1660 vma = find_extend_vma(mm, start);
1661 if (!vma && in_gate_area(mm, start)) {
1662 unsigned long pg = start & PAGE_MASK;
1663 pgd_t *pgd;
1664 pud_t *pud;
1665 pmd_t *pmd;
1666 pte_t *pte;
1668 /* user gate pages are read-only */
1669 if (gup_flags & FOLL_WRITE)
1670 return i ? : -EFAULT;
1671 if (pg > TASK_SIZE)
1672 pgd = pgd_offset_k(pg);
1673 else
1674 pgd = pgd_offset_gate(mm, pg);
1675 BUG_ON(pgd_none(*pgd));
1676 pud = pud_offset(pgd, pg);
1677 BUG_ON(pud_none(*pud));
1678 pmd = pmd_offset(pud, pg);
1679 if (pmd_none(*pmd))
1680 return i ? : -EFAULT;
1681 VM_BUG_ON(pmd_trans_huge(*pmd));
1682 pte = pte_offset_map(pmd, pg);
1683 if (pte_none(*pte)) {
1684 pte_unmap(pte);
1685 return i ? : -EFAULT;
1687 vma = get_gate_vma(mm);
1688 if (pages) {
1689 struct page *page;
1691 page = vm_normal_page(vma, start, *pte);
1692 if (!page) {
1693 if (!(gup_flags & FOLL_DUMP) &&
1694 is_zero_pfn(pte_pfn(*pte)))
1695 page = pte_page(*pte);
1696 else {
1697 pte_unmap(pte);
1698 return i ? : -EFAULT;
1701 pages[i] = page;
1702 get_page(page);
1704 pte_unmap(pte);
1705 goto next_page;
1708 if (!vma ||
1709 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1710 !(vm_flags & vma->vm_flags))
1711 return i ? : -EFAULT;
1713 if (is_vm_hugetlb_page(vma)) {
1714 i = follow_hugetlb_page(mm, vma, pages, vmas,
1715 &start, &nr_pages, i, gup_flags);
1716 continue;
1719 do {
1720 struct page *page;
1721 unsigned int foll_flags = gup_flags;
1724 * If we have a pending SIGKILL, don't keep faulting
1725 * pages and potentially allocating memory.
1727 if (unlikely(fatal_signal_pending(current)))
1728 return i ? i : -ERESTARTSYS;
1730 cond_resched();
1731 while (!(page = follow_page(vma, start, foll_flags))) {
1732 int ret;
1733 unsigned int fault_flags = 0;
1735 /* For mlock, just skip the stack guard page. */
1736 if (foll_flags & FOLL_MLOCK) {
1737 if (stack_guard_page(vma, start))
1738 goto next_page;
1740 if (foll_flags & FOLL_WRITE)
1741 fault_flags |= FAULT_FLAG_WRITE;
1742 if (nonblocking)
1743 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1744 if (foll_flags & FOLL_NOWAIT)
1745 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1747 ret = handle_mm_fault(mm, vma, start,
1748 fault_flags);
1750 if (ret & VM_FAULT_ERROR) {
1751 if (ret & VM_FAULT_OOM)
1752 return i ? i : -ENOMEM;
1753 if (ret & (VM_FAULT_HWPOISON |
1754 VM_FAULT_HWPOISON_LARGE)) {
1755 if (i)
1756 return i;
1757 else if (gup_flags & FOLL_HWPOISON)
1758 return -EHWPOISON;
1759 else
1760 return -EFAULT;
1762 if (ret & VM_FAULT_SIGBUS)
1763 return i ? i : -EFAULT;
1764 BUG();
1767 if (tsk) {
1768 if (ret & VM_FAULT_MAJOR)
1769 tsk->maj_flt++;
1770 else
1771 tsk->min_flt++;
1774 if (ret & VM_FAULT_RETRY) {
1775 if (nonblocking)
1776 *nonblocking = 0;
1777 return i;
1781 * The VM_FAULT_WRITE bit tells us that
1782 * do_wp_page has broken COW when necessary,
1783 * even if maybe_mkwrite decided not to set
1784 * pte_write. We can thus safely do subsequent
1785 * page lookups as if they were reads. But only
1786 * do so when looping for pte_write is futile:
1787 * in some cases userspace may also be wanting
1788 * to write to the gotten user page, which a
1789 * read fault here might prevent (a readonly
1790 * page might get reCOWed by userspace write).
1792 if ((ret & VM_FAULT_WRITE) &&
1793 !(vma->vm_flags & VM_WRITE))
1794 foll_flags &= ~FOLL_WRITE;
1796 cond_resched();
1798 if (IS_ERR(page))
1799 return i ? i : PTR_ERR(page);
1800 if (pages) {
1801 pages[i] = page;
1803 flush_anon_page(vma, page, start);
1804 flush_dcache_page(page);
1806 next_page:
1807 if (vmas)
1808 vmas[i] = vma;
1809 i++;
1810 start += PAGE_SIZE;
1811 nr_pages--;
1812 } while (nr_pages && start < vma->vm_end);
1813 } while (nr_pages);
1814 return i;
1816 EXPORT_SYMBOL(__get_user_pages);
1819 * get_user_pages() - pin user pages in memory
1820 * @tsk: the task_struct to use for page fault accounting, or
1821 * NULL if faults are not to be recorded.
1822 * @mm: mm_struct of target mm
1823 * @start: starting user address
1824 * @nr_pages: number of pages from start to pin
1825 * @write: whether pages will be written to by the caller
1826 * @force: whether to force write access even if user mapping is
1827 * readonly. This will result in the page being COWed even
1828 * in MAP_SHARED mappings. You do not want this.
1829 * @pages: array that receives pointers to the pages pinned.
1830 * Should be at least nr_pages long. Or NULL, if caller
1831 * only intends to ensure the pages are faulted in.
1832 * @vmas: array of pointers to vmas corresponding to each page.
1833 * Or NULL if the caller does not require them.
1835 * Returns number of pages pinned. This may be fewer than the number
1836 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1837 * were pinned, returns -errno. Each page returned must be released
1838 * with a put_page() call when it is finished with. vmas will only
1839 * remain valid while mmap_sem is held.
1841 * Must be called with mmap_sem held for read or write.
1843 * get_user_pages walks a process's page tables and takes a reference to
1844 * each struct page that each user address corresponds to at a given
1845 * instant. That is, it takes the page that would be accessed if a user
1846 * thread accesses the given user virtual address at that instant.
1848 * This does not guarantee that the page exists in the user mappings when
1849 * get_user_pages returns, and there may even be a completely different
1850 * page there in some cases (eg. if mmapped pagecache has been invalidated
1851 * and subsequently re faulted). However it does guarantee that the page
1852 * won't be freed completely. And mostly callers simply care that the page
1853 * contains data that was valid *at some point in time*. Typically, an IO
1854 * or similar operation cannot guarantee anything stronger anyway because
1855 * locks can't be held over the syscall boundary.
1857 * If write=0, the page must not be written to. If the page is written to,
1858 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1859 * after the page is finished with, and before put_page is called.
1861 * get_user_pages is typically used for fewer-copy IO operations, to get a
1862 * handle on the memory by some means other than accesses via the user virtual
1863 * addresses. The pages may be submitted for DMA to devices or accessed via
1864 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1865 * use the correct cache flushing APIs.
1867 * See also get_user_pages_fast, for performance critical applications.
1869 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1870 unsigned long start, int nr_pages, int write, int force,
1871 struct page **pages, struct vm_area_struct **vmas)
1873 int flags = FOLL_TOUCH;
1875 if (pages)
1876 flags |= FOLL_GET;
1877 if (write)
1878 flags |= FOLL_WRITE;
1879 if (force)
1880 flags |= FOLL_FORCE;
1882 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1883 NULL);
1885 EXPORT_SYMBOL(get_user_pages);
1888 * get_dump_page() - pin user page in memory while writing it to core dump
1889 * @addr: user address
1891 * Returns struct page pointer of user page pinned for dump,
1892 * to be freed afterwards by page_cache_release() or put_page().
1894 * Returns NULL on any kind of failure - a hole must then be inserted into
1895 * the corefile, to preserve alignment with its headers; and also returns
1896 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1897 * allowing a hole to be left in the corefile to save diskspace.
1899 * Called without mmap_sem, but after all other threads have been killed.
1901 #ifdef CONFIG_ELF_CORE
1902 struct page *get_dump_page(unsigned long addr)
1904 struct vm_area_struct *vma;
1905 struct page *page;
1907 if (__get_user_pages(current, current->mm, addr, 1,
1908 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1909 NULL) < 1)
1910 return NULL;
1911 flush_cache_page(vma, addr, page_to_pfn(page));
1912 return page;
1914 #endif /* CONFIG_ELF_CORE */
1916 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1917 spinlock_t **ptl)
1919 pgd_t * pgd = pgd_offset(mm, addr);
1920 pud_t * pud = pud_alloc(mm, pgd, addr);
1921 if (pud) {
1922 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1923 if (pmd) {
1924 VM_BUG_ON(pmd_trans_huge(*pmd));
1925 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1928 return NULL;
1932 * This is the old fallback for page remapping.
1934 * For historical reasons, it only allows reserved pages. Only
1935 * old drivers should use this, and they needed to mark their
1936 * pages reserved for the old functions anyway.
1938 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1939 struct page *page, pgprot_t prot)
1941 struct mm_struct *mm = vma->vm_mm;
1942 int retval;
1943 pte_t *pte;
1944 spinlock_t *ptl;
1946 retval = -EINVAL;
1947 if (PageAnon(page))
1948 goto out;
1949 retval = -ENOMEM;
1950 flush_dcache_page(page);
1951 pte = get_locked_pte(mm, addr, &ptl);
1952 if (!pte)
1953 goto out;
1954 retval = -EBUSY;
1955 if (!pte_none(*pte))
1956 goto out_unlock;
1958 /* Ok, finally just insert the thing.. */
1959 get_page(page);
1960 inc_mm_counter_fast(mm, MM_FILEPAGES);
1961 page_add_file_rmap(page);
1962 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1964 retval = 0;
1965 pte_unmap_unlock(pte, ptl);
1966 return retval;
1967 out_unlock:
1968 pte_unmap_unlock(pte, ptl);
1969 out:
1970 return retval;
1974 * vm_insert_page - insert single page into user vma
1975 * @vma: user vma to map to
1976 * @addr: target user address of this page
1977 * @page: source kernel page
1979 * This allows drivers to insert individual pages they've allocated
1980 * into a user vma.
1982 * The page has to be a nice clean _individual_ kernel allocation.
1983 * If you allocate a compound page, you need to have marked it as
1984 * such (__GFP_COMP), or manually just split the page up yourself
1985 * (see split_page()).
1987 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1988 * took an arbitrary page protection parameter. This doesn't allow
1989 * that. Your vma protection will have to be set up correctly, which
1990 * means that if you want a shared writable mapping, you'd better
1991 * ask for a shared writable mapping!
1993 * The page does not need to be reserved.
1995 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1996 struct page *page)
1998 if (addr < vma->vm_start || addr >= vma->vm_end)
1999 return -EFAULT;
2000 if (!page_count(page))
2001 return -EINVAL;
2002 vma->vm_flags |= VM_INSERTPAGE;
2003 return insert_page(vma, addr, page, vma->vm_page_prot);
2005 EXPORT_SYMBOL(vm_insert_page);
2007 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2008 unsigned long pfn, pgprot_t prot)
2010 struct mm_struct *mm = vma->vm_mm;
2011 int retval;
2012 pte_t *pte, entry;
2013 spinlock_t *ptl;
2015 retval = -ENOMEM;
2016 pte = get_locked_pte(mm, addr, &ptl);
2017 if (!pte)
2018 goto out;
2019 retval = -EBUSY;
2020 if (!pte_none(*pte))
2021 goto out_unlock;
2023 /* Ok, finally just insert the thing.. */
2024 entry = pte_mkspecial(pfn_pte(pfn, prot));
2025 set_pte_at(mm, addr, pte, entry);
2026 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2028 retval = 0;
2029 out_unlock:
2030 pte_unmap_unlock(pte, ptl);
2031 out:
2032 return retval;
2036 * vm_insert_pfn - insert single pfn into user vma
2037 * @vma: user vma to map to
2038 * @addr: target user address of this page
2039 * @pfn: source kernel pfn
2041 * Similar to vm_inert_page, this allows drivers to insert individual pages
2042 * they've allocated into a user vma. Same comments apply.
2044 * This function should only be called from a vm_ops->fault handler, and
2045 * in that case the handler should return NULL.
2047 * vma cannot be a COW mapping.
2049 * As this is called only for pages that do not currently exist, we
2050 * do not need to flush old virtual caches or the TLB.
2052 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2053 unsigned long pfn)
2055 int ret;
2056 pgprot_t pgprot = vma->vm_page_prot;
2058 * Technically, architectures with pte_special can avoid all these
2059 * restrictions (same for remap_pfn_range). However we would like
2060 * consistency in testing and feature parity among all, so we should
2061 * try to keep these invariants in place for everybody.
2063 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2064 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2065 (VM_PFNMAP|VM_MIXEDMAP));
2066 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2067 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2069 if (addr < vma->vm_start || addr >= vma->vm_end)
2070 return -EFAULT;
2071 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2072 return -EINVAL;
2074 ret = insert_pfn(vma, addr, pfn, pgprot);
2076 if (ret)
2077 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2079 return ret;
2081 EXPORT_SYMBOL(vm_insert_pfn);
2083 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2084 unsigned long pfn)
2086 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2088 if (addr < vma->vm_start || addr >= vma->vm_end)
2089 return -EFAULT;
2092 * If we don't have pte special, then we have to use the pfn_valid()
2093 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2094 * refcount the page if pfn_valid is true (hence insert_page rather
2095 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2096 * without pte special, it would there be refcounted as a normal page.
2098 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2099 struct page *page;
2101 page = pfn_to_page(pfn);
2102 return insert_page(vma, addr, page, vma->vm_page_prot);
2104 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2106 EXPORT_SYMBOL(vm_insert_mixed);
2109 * maps a range of physical memory into the requested pages. the old
2110 * mappings are removed. any references to nonexistent pages results
2111 * in null mappings (currently treated as "copy-on-access")
2113 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2114 unsigned long addr, unsigned long end,
2115 unsigned long pfn, pgprot_t prot)
2117 pte_t *pte;
2118 spinlock_t *ptl;
2120 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2121 if (!pte)
2122 return -ENOMEM;
2123 arch_enter_lazy_mmu_mode();
2124 do {
2125 BUG_ON(!pte_none(*pte));
2126 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2127 pfn++;
2128 } while (pte++, addr += PAGE_SIZE, addr != end);
2129 arch_leave_lazy_mmu_mode();
2130 pte_unmap_unlock(pte - 1, ptl);
2131 return 0;
2134 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2135 unsigned long addr, unsigned long end,
2136 unsigned long pfn, pgprot_t prot)
2138 pmd_t *pmd;
2139 unsigned long next;
2141 pfn -= addr >> PAGE_SHIFT;
2142 pmd = pmd_alloc(mm, pud, addr);
2143 if (!pmd)
2144 return -ENOMEM;
2145 VM_BUG_ON(pmd_trans_huge(*pmd));
2146 do {
2147 next = pmd_addr_end(addr, end);
2148 if (remap_pte_range(mm, pmd, addr, next,
2149 pfn + (addr >> PAGE_SHIFT), prot))
2150 return -ENOMEM;
2151 } while (pmd++, addr = next, addr != end);
2152 return 0;
2155 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2156 unsigned long addr, unsigned long end,
2157 unsigned long pfn, pgprot_t prot)
2159 pud_t *pud;
2160 unsigned long next;
2162 pfn -= addr >> PAGE_SHIFT;
2163 pud = pud_alloc(mm, pgd, addr);
2164 if (!pud)
2165 return -ENOMEM;
2166 do {
2167 next = pud_addr_end(addr, end);
2168 if (remap_pmd_range(mm, pud, addr, next,
2169 pfn + (addr >> PAGE_SHIFT), prot))
2170 return -ENOMEM;
2171 } while (pud++, addr = next, addr != end);
2172 return 0;
2176 * remap_pfn_range - remap kernel memory to userspace
2177 * @vma: user vma to map to
2178 * @addr: target user address to start at
2179 * @pfn: physical address of kernel memory
2180 * @size: size of map area
2181 * @prot: page protection flags for this mapping
2183 * Note: this is only safe if the mm semaphore is held when called.
2185 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2186 unsigned long pfn, unsigned long size, pgprot_t prot)
2188 pgd_t *pgd;
2189 unsigned long next;
2190 unsigned long end = addr + PAGE_ALIGN(size);
2191 struct mm_struct *mm = vma->vm_mm;
2192 int err;
2195 * Physically remapped pages are special. Tell the
2196 * rest of the world about it:
2197 * VM_IO tells people not to look at these pages
2198 * (accesses can have side effects).
2199 * VM_RESERVED is specified all over the place, because
2200 * in 2.4 it kept swapout's vma scan off this vma; but
2201 * in 2.6 the LRU scan won't even find its pages, so this
2202 * flag means no more than count its pages in reserved_vm,
2203 * and omit it from core dump, even when VM_IO turned off.
2204 * VM_PFNMAP tells the core MM that the base pages are just
2205 * raw PFN mappings, and do not have a "struct page" associated
2206 * with them.
2208 * There's a horrible special case to handle copy-on-write
2209 * behaviour that some programs depend on. We mark the "original"
2210 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2212 if (addr == vma->vm_start && end == vma->vm_end) {
2213 vma->vm_pgoff = pfn;
2214 vma->vm_flags |= VM_PFN_AT_MMAP;
2215 } else if (is_cow_mapping(vma->vm_flags))
2216 return -EINVAL;
2218 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2220 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2221 if (err) {
2223 * To indicate that track_pfn related cleanup is not
2224 * needed from higher level routine calling unmap_vmas
2226 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2227 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2228 return -EINVAL;
2231 BUG_ON(addr >= end);
2232 pfn -= addr >> PAGE_SHIFT;
2233 pgd = pgd_offset(mm, addr);
2234 flush_cache_range(vma, addr, end);
2235 do {
2236 next = pgd_addr_end(addr, end);
2237 err = remap_pud_range(mm, pgd, addr, next,
2238 pfn + (addr >> PAGE_SHIFT), prot);
2239 if (err)
2240 break;
2241 } while (pgd++, addr = next, addr != end);
2243 if (err)
2244 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2246 return err;
2248 EXPORT_SYMBOL(remap_pfn_range);
2250 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2251 unsigned long addr, unsigned long end,
2252 pte_fn_t fn, void *data)
2254 pte_t *pte;
2255 int err;
2256 pgtable_t token;
2257 spinlock_t *uninitialized_var(ptl);
2259 pte = (mm == &init_mm) ?
2260 pte_alloc_kernel(pmd, addr) :
2261 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2262 if (!pte)
2263 return -ENOMEM;
2265 BUG_ON(pmd_huge(*pmd));
2267 arch_enter_lazy_mmu_mode();
2269 token = pmd_pgtable(*pmd);
2271 do {
2272 err = fn(pte++, token, addr, data);
2273 if (err)
2274 break;
2275 } while (addr += PAGE_SIZE, addr != end);
2277 arch_leave_lazy_mmu_mode();
2279 if (mm != &init_mm)
2280 pte_unmap_unlock(pte-1, ptl);
2281 return err;
2284 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2285 unsigned long addr, unsigned long end,
2286 pte_fn_t fn, void *data)
2288 pmd_t *pmd;
2289 unsigned long next;
2290 int err;
2292 BUG_ON(pud_huge(*pud));
2294 pmd = pmd_alloc(mm, pud, addr);
2295 if (!pmd)
2296 return -ENOMEM;
2297 do {
2298 next = pmd_addr_end(addr, end);
2299 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2300 if (err)
2301 break;
2302 } while (pmd++, addr = next, addr != end);
2303 return err;
2306 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2307 unsigned long addr, unsigned long end,
2308 pte_fn_t fn, void *data)
2310 pud_t *pud;
2311 unsigned long next;
2312 int err;
2314 pud = pud_alloc(mm, pgd, addr);
2315 if (!pud)
2316 return -ENOMEM;
2317 do {
2318 next = pud_addr_end(addr, end);
2319 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2320 if (err)
2321 break;
2322 } while (pud++, addr = next, addr != end);
2323 return err;
2327 * Scan a region of virtual memory, filling in page tables as necessary
2328 * and calling a provided function on each leaf page table.
2330 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2331 unsigned long size, pte_fn_t fn, void *data)
2333 pgd_t *pgd;
2334 unsigned long next;
2335 unsigned long end = addr + size;
2336 int err;
2338 BUG_ON(addr >= end);
2339 pgd = pgd_offset(mm, addr);
2340 do {
2341 next = pgd_addr_end(addr, end);
2342 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2343 if (err)
2344 break;
2345 } while (pgd++, addr = next, addr != end);
2347 return err;
2349 EXPORT_SYMBOL_GPL(apply_to_page_range);
2352 * handle_pte_fault chooses page fault handler according to an entry
2353 * which was read non-atomically. Before making any commitment, on
2354 * those architectures or configurations (e.g. i386 with PAE) which
2355 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2356 * must check under lock before unmapping the pte and proceeding
2357 * (but do_wp_page is only called after already making such a check;
2358 * and do_anonymous_page can safely check later on).
2360 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2361 pte_t *page_table, pte_t orig_pte)
2363 int same = 1;
2364 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2365 if (sizeof(pte_t) > sizeof(unsigned long)) {
2366 spinlock_t *ptl = pte_lockptr(mm, pmd);
2367 spin_lock(ptl);
2368 same = pte_same(*page_table, orig_pte);
2369 spin_unlock(ptl);
2371 #endif
2372 pte_unmap(page_table);
2373 return same;
2376 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2379 * If the source page was a PFN mapping, we don't have
2380 * a "struct page" for it. We do a best-effort copy by
2381 * just copying from the original user address. If that
2382 * fails, we just zero-fill it. Live with it.
2384 if (unlikely(!src)) {
2385 void *kaddr = kmap_atomic(dst, KM_USER0);
2386 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2389 * This really shouldn't fail, because the page is there
2390 * in the page tables. But it might just be unreadable,
2391 * in which case we just give up and fill the result with
2392 * zeroes.
2394 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2395 clear_page(kaddr);
2396 kunmap_atomic(kaddr, KM_USER0);
2397 flush_dcache_page(dst);
2398 } else
2399 copy_user_highpage(dst, src, va, vma);
2403 * This routine handles present pages, when users try to write
2404 * to a shared page. It is done by copying the page to a new address
2405 * and decrementing the shared-page counter for the old page.
2407 * Note that this routine assumes that the protection checks have been
2408 * done by the caller (the low-level page fault routine in most cases).
2409 * Thus we can safely just mark it writable once we've done any necessary
2410 * COW.
2412 * We also mark the page dirty at this point even though the page will
2413 * change only once the write actually happens. This avoids a few races,
2414 * and potentially makes it more efficient.
2416 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2417 * but allow concurrent faults), with pte both mapped and locked.
2418 * We return with mmap_sem still held, but pte unmapped and unlocked.
2420 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2421 unsigned long address, pte_t *page_table, pmd_t *pmd,
2422 spinlock_t *ptl, pte_t orig_pte)
2423 __releases(ptl)
2425 struct page *old_page, *new_page;
2426 pte_t entry;
2427 int ret = 0;
2428 int page_mkwrite = 0;
2429 struct page *dirty_page = NULL;
2431 old_page = vm_normal_page(vma, address, orig_pte);
2432 if (!old_page) {
2434 * VM_MIXEDMAP !pfn_valid() case
2436 * We should not cow pages in a shared writeable mapping.
2437 * Just mark the pages writable as we can't do any dirty
2438 * accounting on raw pfn maps.
2440 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2441 (VM_WRITE|VM_SHARED))
2442 goto reuse;
2443 goto gotten;
2447 * Take out anonymous pages first, anonymous shared vmas are
2448 * not dirty accountable.
2450 if (PageAnon(old_page) && !PageKsm(old_page)) {
2451 if (!trylock_page(old_page)) {
2452 page_cache_get(old_page);
2453 pte_unmap_unlock(page_table, ptl);
2454 lock_page(old_page);
2455 page_table = pte_offset_map_lock(mm, pmd, address,
2456 &ptl);
2457 if (!pte_same(*page_table, orig_pte)) {
2458 unlock_page(old_page);
2459 goto unlock;
2461 page_cache_release(old_page);
2463 if (reuse_swap_page(old_page)) {
2465 * The page is all ours. Move it to our anon_vma so
2466 * the rmap code will not search our parent or siblings.
2467 * Protected against the rmap code by the page lock.
2469 page_move_anon_rmap(old_page, vma, address);
2470 unlock_page(old_page);
2471 goto reuse;
2473 unlock_page(old_page);
2474 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2475 (VM_WRITE|VM_SHARED))) {
2477 * Only catch write-faults on shared writable pages,
2478 * read-only shared pages can get COWed by
2479 * get_user_pages(.write=1, .force=1).
2481 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2482 struct vm_fault vmf;
2483 int tmp;
2485 vmf.virtual_address = (void __user *)(address &
2486 PAGE_MASK);
2487 vmf.pgoff = old_page->index;
2488 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2489 vmf.page = old_page;
2492 * Notify the address space that the page is about to
2493 * become writable so that it can prohibit this or wait
2494 * for the page to get into an appropriate state.
2496 * We do this without the lock held, so that it can
2497 * sleep if it needs to.
2499 page_cache_get(old_page);
2500 pte_unmap_unlock(page_table, ptl);
2502 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2503 if (unlikely(tmp &
2504 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2505 ret = tmp;
2506 goto unwritable_page;
2508 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2509 lock_page(old_page);
2510 if (!old_page->mapping) {
2511 ret = 0; /* retry the fault */
2512 unlock_page(old_page);
2513 goto unwritable_page;
2515 } else
2516 VM_BUG_ON(!PageLocked(old_page));
2519 * Since we dropped the lock we need to revalidate
2520 * the PTE as someone else may have changed it. If
2521 * they did, we just return, as we can count on the
2522 * MMU to tell us if they didn't also make it writable.
2524 page_table = pte_offset_map_lock(mm, pmd, address,
2525 &ptl);
2526 if (!pte_same(*page_table, orig_pte)) {
2527 unlock_page(old_page);
2528 goto unlock;
2531 page_mkwrite = 1;
2533 dirty_page = old_page;
2534 get_page(dirty_page);
2536 reuse:
2537 flush_cache_page(vma, address, pte_pfn(orig_pte));
2538 entry = pte_mkyoung(orig_pte);
2539 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2540 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2541 update_mmu_cache(vma, address, page_table);
2542 pte_unmap_unlock(page_table, ptl);
2543 ret |= VM_FAULT_WRITE;
2545 if (!dirty_page)
2546 return ret;
2549 * Yes, Virginia, this is actually required to prevent a race
2550 * with clear_page_dirty_for_io() from clearing the page dirty
2551 * bit after it clear all dirty ptes, but before a racing
2552 * do_wp_page installs a dirty pte.
2554 * __do_fault is protected similarly.
2556 if (!page_mkwrite) {
2557 wait_on_page_locked(dirty_page);
2558 set_page_dirty_balance(dirty_page, page_mkwrite);
2560 put_page(dirty_page);
2561 if (page_mkwrite) {
2562 struct address_space *mapping = dirty_page->mapping;
2564 set_page_dirty(dirty_page);
2565 unlock_page(dirty_page);
2566 page_cache_release(dirty_page);
2567 if (mapping) {
2569 * Some device drivers do not set page.mapping
2570 * but still dirty their pages
2572 balance_dirty_pages_ratelimited(mapping);
2576 /* file_update_time outside page_lock */
2577 if (vma->vm_file)
2578 file_update_time(vma->vm_file);
2580 return ret;
2584 * Ok, we need to copy. Oh, well..
2586 page_cache_get(old_page);
2587 gotten:
2588 pte_unmap_unlock(page_table, ptl);
2590 if (unlikely(anon_vma_prepare(vma)))
2591 goto oom;
2593 if (is_zero_pfn(pte_pfn(orig_pte))) {
2594 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2595 if (!new_page)
2596 goto oom;
2597 } else {
2598 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2599 if (!new_page)
2600 goto oom;
2601 cow_user_page(new_page, old_page, address, vma);
2603 __SetPageUptodate(new_page);
2605 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2606 goto oom_free_new;
2609 * Re-check the pte - we dropped the lock
2611 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2612 if (likely(pte_same(*page_table, orig_pte))) {
2613 if (old_page) {
2614 if (!PageAnon(old_page)) {
2615 dec_mm_counter_fast(mm, MM_FILEPAGES);
2616 inc_mm_counter_fast(mm, MM_ANONPAGES);
2618 } else
2619 inc_mm_counter_fast(mm, MM_ANONPAGES);
2620 flush_cache_page(vma, address, pte_pfn(orig_pte));
2621 entry = mk_pte(new_page, vma->vm_page_prot);
2622 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2624 * Clear the pte entry and flush it first, before updating the
2625 * pte with the new entry. This will avoid a race condition
2626 * seen in the presence of one thread doing SMC and another
2627 * thread doing COW.
2629 ptep_clear_flush(vma, address, page_table);
2630 page_add_new_anon_rmap(new_page, vma, address);
2632 * We call the notify macro here because, when using secondary
2633 * mmu page tables (such as kvm shadow page tables), we want the
2634 * new page to be mapped directly into the secondary page table.
2636 set_pte_at_notify(mm, address, page_table, entry);
2637 update_mmu_cache(vma, address, page_table);
2638 if (old_page) {
2640 * Only after switching the pte to the new page may
2641 * we remove the mapcount here. Otherwise another
2642 * process may come and find the rmap count decremented
2643 * before the pte is switched to the new page, and
2644 * "reuse" the old page writing into it while our pte
2645 * here still points into it and can be read by other
2646 * threads.
2648 * The critical issue is to order this
2649 * page_remove_rmap with the ptp_clear_flush above.
2650 * Those stores are ordered by (if nothing else,)
2651 * the barrier present in the atomic_add_negative
2652 * in page_remove_rmap.
2654 * Then the TLB flush in ptep_clear_flush ensures that
2655 * no process can access the old page before the
2656 * decremented mapcount is visible. And the old page
2657 * cannot be reused until after the decremented
2658 * mapcount is visible. So transitively, TLBs to
2659 * old page will be flushed before it can be reused.
2661 page_remove_rmap(old_page);
2664 /* Free the old page.. */
2665 new_page = old_page;
2666 ret |= VM_FAULT_WRITE;
2667 } else
2668 mem_cgroup_uncharge_page(new_page);
2670 if (new_page)
2671 page_cache_release(new_page);
2672 unlock:
2673 pte_unmap_unlock(page_table, ptl);
2674 if (old_page) {
2676 * Don't let another task, with possibly unlocked vma,
2677 * keep the mlocked page.
2679 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2680 lock_page(old_page); /* LRU manipulation */
2681 munlock_vma_page(old_page);
2682 unlock_page(old_page);
2684 page_cache_release(old_page);
2686 return ret;
2687 oom_free_new:
2688 page_cache_release(new_page);
2689 oom:
2690 if (old_page) {
2691 if (page_mkwrite) {
2692 unlock_page(old_page);
2693 page_cache_release(old_page);
2695 page_cache_release(old_page);
2697 return VM_FAULT_OOM;
2699 unwritable_page:
2700 page_cache_release(old_page);
2701 return ret;
2704 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2705 unsigned long start_addr, unsigned long end_addr,
2706 struct zap_details *details)
2708 zap_page_range(vma, start_addr, end_addr - start_addr, details);
2711 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2712 struct zap_details *details)
2714 struct vm_area_struct *vma;
2715 struct prio_tree_iter iter;
2716 pgoff_t vba, vea, zba, zea;
2718 vma_prio_tree_foreach(vma, &iter, root,
2719 details->first_index, details->last_index) {
2721 vba = vma->vm_pgoff;
2722 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2723 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2724 zba = details->first_index;
2725 if (zba < vba)
2726 zba = vba;
2727 zea = details->last_index;
2728 if (zea > vea)
2729 zea = vea;
2731 unmap_mapping_range_vma(vma,
2732 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2733 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2734 details);
2738 static inline void unmap_mapping_range_list(struct list_head *head,
2739 struct zap_details *details)
2741 struct vm_area_struct *vma;
2744 * In nonlinear VMAs there is no correspondence between virtual address
2745 * offset and file offset. So we must perform an exhaustive search
2746 * across *all* the pages in each nonlinear VMA, not just the pages
2747 * whose virtual address lies outside the file truncation point.
2749 list_for_each_entry(vma, head, shared.vm_set.list) {
2750 details->nonlinear_vma = vma;
2751 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2756 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2757 * @mapping: the address space containing mmaps to be unmapped.
2758 * @holebegin: byte in first page to unmap, relative to the start of
2759 * the underlying file. This will be rounded down to a PAGE_SIZE
2760 * boundary. Note that this is different from truncate_pagecache(), which
2761 * must keep the partial page. In contrast, we must get rid of
2762 * partial pages.
2763 * @holelen: size of prospective hole in bytes. This will be rounded
2764 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2765 * end of the file.
2766 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2767 * but 0 when invalidating pagecache, don't throw away private data.
2769 void unmap_mapping_range(struct address_space *mapping,
2770 loff_t const holebegin, loff_t const holelen, int even_cows)
2772 struct zap_details details;
2773 pgoff_t hba = holebegin >> PAGE_SHIFT;
2774 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2776 /* Check for overflow. */
2777 if (sizeof(holelen) > sizeof(hlen)) {
2778 long long holeend =
2779 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2780 if (holeend & ~(long long)ULONG_MAX)
2781 hlen = ULONG_MAX - hba + 1;
2784 details.check_mapping = even_cows? NULL: mapping;
2785 details.nonlinear_vma = NULL;
2786 details.first_index = hba;
2787 details.last_index = hba + hlen - 1;
2788 if (details.last_index < details.first_index)
2789 details.last_index = ULONG_MAX;
2792 mutex_lock(&mapping->i_mmap_mutex);
2793 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2794 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2795 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2796 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2797 mutex_unlock(&mapping->i_mmap_mutex);
2799 EXPORT_SYMBOL(unmap_mapping_range);
2802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2803 * but allow concurrent faults), and pte mapped but not yet locked.
2804 * We return with mmap_sem still held, but pte unmapped and unlocked.
2806 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2807 unsigned long address, pte_t *page_table, pmd_t *pmd,
2808 unsigned int flags, pte_t orig_pte)
2810 spinlock_t *ptl;
2811 struct page *page, *swapcache = NULL;
2812 swp_entry_t entry;
2813 pte_t pte;
2814 int locked;
2815 struct mem_cgroup *ptr;
2816 int exclusive = 0;
2817 int ret = 0;
2819 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2820 goto out;
2822 entry = pte_to_swp_entry(orig_pte);
2823 if (unlikely(non_swap_entry(entry))) {
2824 if (is_migration_entry(entry)) {
2825 migration_entry_wait(mm, pmd, address);
2826 } else if (is_hwpoison_entry(entry)) {
2827 ret = VM_FAULT_HWPOISON;
2828 } else {
2829 print_bad_pte(vma, address, orig_pte, NULL);
2830 ret = VM_FAULT_SIGBUS;
2832 goto out;
2834 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2835 page = lookup_swap_cache(entry);
2836 if (!page) {
2837 grab_swap_token(mm); /* Contend for token _before_ read-in */
2838 page = swapin_readahead(entry,
2839 GFP_HIGHUSER_MOVABLE, vma, address);
2840 if (!page) {
2842 * Back out if somebody else faulted in this pte
2843 * while we released the pte lock.
2845 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2846 if (likely(pte_same(*page_table, orig_pte)))
2847 ret = VM_FAULT_OOM;
2848 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2849 goto unlock;
2852 /* Had to read the page from swap area: Major fault */
2853 ret = VM_FAULT_MAJOR;
2854 count_vm_event(PGMAJFAULT);
2855 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2856 } else if (PageHWPoison(page)) {
2858 * hwpoisoned dirty swapcache pages are kept for killing
2859 * owner processes (which may be unknown at hwpoison time)
2861 ret = VM_FAULT_HWPOISON;
2862 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2863 goto out_release;
2866 locked = lock_page_or_retry(page, mm, flags);
2867 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2868 if (!locked) {
2869 ret |= VM_FAULT_RETRY;
2870 goto out_release;
2874 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2875 * release the swapcache from under us. The page pin, and pte_same
2876 * test below, are not enough to exclude that. Even if it is still
2877 * swapcache, we need to check that the page's swap has not changed.
2879 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2880 goto out_page;
2882 if (ksm_might_need_to_copy(page, vma, address)) {
2883 swapcache = page;
2884 page = ksm_does_need_to_copy(page, vma, address);
2886 if (unlikely(!page)) {
2887 ret = VM_FAULT_OOM;
2888 page = swapcache;
2889 swapcache = NULL;
2890 goto out_page;
2894 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2895 ret = VM_FAULT_OOM;
2896 goto out_page;
2900 * Back out if somebody else already faulted in this pte.
2902 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2903 if (unlikely(!pte_same(*page_table, orig_pte)))
2904 goto out_nomap;
2906 if (unlikely(!PageUptodate(page))) {
2907 ret = VM_FAULT_SIGBUS;
2908 goto out_nomap;
2912 * The page isn't present yet, go ahead with the fault.
2914 * Be careful about the sequence of operations here.
2915 * To get its accounting right, reuse_swap_page() must be called
2916 * while the page is counted on swap but not yet in mapcount i.e.
2917 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2918 * must be called after the swap_free(), or it will never succeed.
2919 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2920 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2921 * in page->private. In this case, a record in swap_cgroup is silently
2922 * discarded at swap_free().
2925 inc_mm_counter_fast(mm, MM_ANONPAGES);
2926 dec_mm_counter_fast(mm, MM_SWAPENTS);
2927 pte = mk_pte(page, vma->vm_page_prot);
2928 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2929 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2930 flags &= ~FAULT_FLAG_WRITE;
2931 ret |= VM_FAULT_WRITE;
2932 exclusive = 1;
2934 flush_icache_page(vma, page);
2935 set_pte_at(mm, address, page_table, pte);
2936 do_page_add_anon_rmap(page, vma, address, exclusive);
2937 /* It's better to call commit-charge after rmap is established */
2938 mem_cgroup_commit_charge_swapin(page, ptr);
2940 swap_free(entry);
2941 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2942 try_to_free_swap(page);
2943 unlock_page(page);
2944 if (swapcache) {
2946 * Hold the lock to avoid the swap entry to be reused
2947 * until we take the PT lock for the pte_same() check
2948 * (to avoid false positives from pte_same). For
2949 * further safety release the lock after the swap_free
2950 * so that the swap count won't change under a
2951 * parallel locked swapcache.
2953 unlock_page(swapcache);
2954 page_cache_release(swapcache);
2957 if (flags & FAULT_FLAG_WRITE) {
2958 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2959 if (ret & VM_FAULT_ERROR)
2960 ret &= VM_FAULT_ERROR;
2961 goto out;
2964 /* No need to invalidate - it was non-present before */
2965 update_mmu_cache(vma, address, page_table);
2966 unlock:
2967 pte_unmap_unlock(page_table, ptl);
2968 out:
2969 return ret;
2970 out_nomap:
2971 mem_cgroup_cancel_charge_swapin(ptr);
2972 pte_unmap_unlock(page_table, ptl);
2973 out_page:
2974 unlock_page(page);
2975 out_release:
2976 page_cache_release(page);
2977 if (swapcache) {
2978 unlock_page(swapcache);
2979 page_cache_release(swapcache);
2981 return ret;
2985 * This is like a special single-page "expand_{down|up}wards()",
2986 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2987 * doesn't hit another vma.
2989 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2991 address &= PAGE_MASK;
2992 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2993 struct vm_area_struct *prev = vma->vm_prev;
2996 * Is there a mapping abutting this one below?
2998 * That's only ok if it's the same stack mapping
2999 * that has gotten split..
3001 if (prev && prev->vm_end == address)
3002 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3004 expand_downwards(vma, address - PAGE_SIZE);
3006 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3007 struct vm_area_struct *next = vma->vm_next;
3009 /* As VM_GROWSDOWN but s/below/above/ */
3010 if (next && next->vm_start == address + PAGE_SIZE)
3011 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3013 expand_upwards(vma, address + PAGE_SIZE);
3015 return 0;
3019 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3020 * but allow concurrent faults), and pte mapped but not yet locked.
3021 * We return with mmap_sem still held, but pte unmapped and unlocked.
3023 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3024 unsigned long address, pte_t *page_table, pmd_t *pmd,
3025 unsigned int flags)
3027 struct page *page;
3028 spinlock_t *ptl;
3029 pte_t entry;
3031 pte_unmap(page_table);
3033 /* Check if we need to add a guard page to the stack */
3034 if (check_stack_guard_page(vma, address) < 0)
3035 return VM_FAULT_SIGBUS;
3037 /* Use the zero-page for reads */
3038 if (!(flags & FAULT_FLAG_WRITE)) {
3039 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3040 vma->vm_page_prot));
3041 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3042 if (!pte_none(*page_table))
3043 goto unlock;
3044 goto setpte;
3047 /* Allocate our own private page. */
3048 if (unlikely(anon_vma_prepare(vma)))
3049 goto oom;
3050 page = alloc_zeroed_user_highpage_movable(vma, address);
3051 if (!page)
3052 goto oom;
3053 __SetPageUptodate(page);
3055 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3056 goto oom_free_page;
3058 entry = mk_pte(page, vma->vm_page_prot);
3059 if (vma->vm_flags & VM_WRITE)
3060 entry = pte_mkwrite(pte_mkdirty(entry));
3062 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3063 if (!pte_none(*page_table))
3064 goto release;
3066 inc_mm_counter_fast(mm, MM_ANONPAGES);
3067 page_add_new_anon_rmap(page, vma, address);
3068 setpte:
3069 set_pte_at(mm, address, page_table, entry);
3071 /* No need to invalidate - it was non-present before */
3072 update_mmu_cache(vma, address, page_table);
3073 unlock:
3074 pte_unmap_unlock(page_table, ptl);
3075 return 0;
3076 release:
3077 mem_cgroup_uncharge_page(page);
3078 page_cache_release(page);
3079 goto unlock;
3080 oom_free_page:
3081 page_cache_release(page);
3082 oom:
3083 return VM_FAULT_OOM;
3087 * __do_fault() tries to create a new page mapping. It aggressively
3088 * tries to share with existing pages, but makes a separate copy if
3089 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3090 * the next page fault.
3092 * As this is called only for pages that do not currently exist, we
3093 * do not need to flush old virtual caches or the TLB.
3095 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3096 * but allow concurrent faults), and pte neither mapped nor locked.
3097 * We return with mmap_sem still held, but pte unmapped and unlocked.
3099 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3100 unsigned long address, pmd_t *pmd,
3101 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3103 pte_t *page_table;
3104 spinlock_t *ptl;
3105 struct page *page;
3106 pte_t entry;
3107 int anon = 0;
3108 int charged = 0;
3109 struct page *dirty_page = NULL;
3110 struct vm_fault vmf;
3111 int ret;
3112 int page_mkwrite = 0;
3114 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3115 vmf.pgoff = pgoff;
3116 vmf.flags = flags;
3117 vmf.page = NULL;
3119 ret = vma->vm_ops->fault(vma, &vmf);
3120 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3121 VM_FAULT_RETRY)))
3122 return ret;
3124 if (unlikely(PageHWPoison(vmf.page))) {
3125 if (ret & VM_FAULT_LOCKED)
3126 unlock_page(vmf.page);
3127 return VM_FAULT_HWPOISON;
3131 * For consistency in subsequent calls, make the faulted page always
3132 * locked.
3134 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3135 lock_page(vmf.page);
3136 else
3137 VM_BUG_ON(!PageLocked(vmf.page));
3140 * Should we do an early C-O-W break?
3142 page = vmf.page;
3143 if (flags & FAULT_FLAG_WRITE) {
3144 if (!(vma->vm_flags & VM_SHARED)) {
3145 anon = 1;
3146 if (unlikely(anon_vma_prepare(vma))) {
3147 ret = VM_FAULT_OOM;
3148 goto out;
3150 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3151 vma, address);
3152 if (!page) {
3153 ret = VM_FAULT_OOM;
3154 goto out;
3156 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3157 ret = VM_FAULT_OOM;
3158 page_cache_release(page);
3159 goto out;
3161 charged = 1;
3162 copy_user_highpage(page, vmf.page, address, vma);
3163 __SetPageUptodate(page);
3164 } else {
3166 * If the page will be shareable, see if the backing
3167 * address space wants to know that the page is about
3168 * to become writable
3170 if (vma->vm_ops->page_mkwrite) {
3171 int tmp;
3173 unlock_page(page);
3174 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3175 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3176 if (unlikely(tmp &
3177 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3178 ret = tmp;
3179 goto unwritable_page;
3181 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3182 lock_page(page);
3183 if (!page->mapping) {
3184 ret = 0; /* retry the fault */
3185 unlock_page(page);
3186 goto unwritable_page;
3188 } else
3189 VM_BUG_ON(!PageLocked(page));
3190 page_mkwrite = 1;
3196 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3199 * This silly early PAGE_DIRTY setting removes a race
3200 * due to the bad i386 page protection. But it's valid
3201 * for other architectures too.
3203 * Note that if FAULT_FLAG_WRITE is set, we either now have
3204 * an exclusive copy of the page, or this is a shared mapping,
3205 * so we can make it writable and dirty to avoid having to
3206 * handle that later.
3208 /* Only go through if we didn't race with anybody else... */
3209 if (likely(pte_same(*page_table, orig_pte))) {
3210 flush_icache_page(vma, page);
3211 entry = mk_pte(page, vma->vm_page_prot);
3212 if (flags & FAULT_FLAG_WRITE)
3213 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3214 if (anon) {
3215 inc_mm_counter_fast(mm, MM_ANONPAGES);
3216 page_add_new_anon_rmap(page, vma, address);
3217 } else {
3218 inc_mm_counter_fast(mm, MM_FILEPAGES);
3219 page_add_file_rmap(page);
3220 if (flags & FAULT_FLAG_WRITE) {
3221 dirty_page = page;
3222 get_page(dirty_page);
3225 set_pte_at(mm, address, page_table, entry);
3227 /* no need to invalidate: a not-present page won't be cached */
3228 update_mmu_cache(vma, address, page_table);
3229 } else {
3230 if (charged)
3231 mem_cgroup_uncharge_page(page);
3232 if (anon)
3233 page_cache_release(page);
3234 else
3235 anon = 1; /* no anon but release faulted_page */
3238 pte_unmap_unlock(page_table, ptl);
3240 out:
3241 if (dirty_page) {
3242 struct address_space *mapping = page->mapping;
3244 if (set_page_dirty(dirty_page))
3245 page_mkwrite = 1;
3246 unlock_page(dirty_page);
3247 put_page(dirty_page);
3248 if (page_mkwrite && mapping) {
3250 * Some device drivers do not set page.mapping but still
3251 * dirty their pages
3253 balance_dirty_pages_ratelimited(mapping);
3256 /* file_update_time outside page_lock */
3257 if (vma->vm_file)
3258 file_update_time(vma->vm_file);
3259 } else {
3260 unlock_page(vmf.page);
3261 if (anon)
3262 page_cache_release(vmf.page);
3265 return ret;
3267 unwritable_page:
3268 page_cache_release(page);
3269 return ret;
3272 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3273 unsigned long address, pte_t *page_table, pmd_t *pmd,
3274 unsigned int flags, pte_t orig_pte)
3276 pgoff_t pgoff = (((address & PAGE_MASK)
3277 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3279 pte_unmap(page_table);
3280 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3284 * Fault of a previously existing named mapping. Repopulate the pte
3285 * from the encoded file_pte if possible. This enables swappable
3286 * nonlinear vmas.
3288 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3289 * but allow concurrent faults), and pte mapped but not yet locked.
3290 * We return with mmap_sem still held, but pte unmapped and unlocked.
3292 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3293 unsigned long address, pte_t *page_table, pmd_t *pmd,
3294 unsigned int flags, pte_t orig_pte)
3296 pgoff_t pgoff;
3298 flags |= FAULT_FLAG_NONLINEAR;
3300 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3301 return 0;
3303 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3305 * Page table corrupted: show pte and kill process.
3307 print_bad_pte(vma, address, orig_pte, NULL);
3308 return VM_FAULT_SIGBUS;
3311 pgoff = pte_to_pgoff(orig_pte);
3312 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3316 * These routines also need to handle stuff like marking pages dirty
3317 * and/or accessed for architectures that don't do it in hardware (most
3318 * RISC architectures). The early dirtying is also good on the i386.
3320 * There is also a hook called "update_mmu_cache()" that architectures
3321 * with external mmu caches can use to update those (ie the Sparc or
3322 * PowerPC hashed page tables that act as extended TLBs).
3324 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3325 * but allow concurrent faults), and pte mapped but not yet locked.
3326 * We return with mmap_sem still held, but pte unmapped and unlocked.
3328 int handle_pte_fault(struct mm_struct *mm,
3329 struct vm_area_struct *vma, unsigned long address,
3330 pte_t *pte, pmd_t *pmd, unsigned int flags)
3332 pte_t entry;
3333 spinlock_t *ptl;
3335 entry = *pte;
3336 if (!pte_present(entry)) {
3337 if (pte_none(entry)) {
3338 if (vma->vm_ops) {
3339 if (likely(vma->vm_ops->fault))
3340 return do_linear_fault(mm, vma, address,
3341 pte, pmd, flags, entry);
3343 return do_anonymous_page(mm, vma, address,
3344 pte, pmd, flags);
3346 if (pte_file(entry))
3347 return do_nonlinear_fault(mm, vma, address,
3348 pte, pmd, flags, entry);
3349 return do_swap_page(mm, vma, address,
3350 pte, pmd, flags, entry);
3353 ptl = pte_lockptr(mm, pmd);
3354 spin_lock(ptl);
3355 if (unlikely(!pte_same(*pte, entry)))
3356 goto unlock;
3357 if (flags & FAULT_FLAG_WRITE) {
3358 if (!pte_write(entry))
3359 return do_wp_page(mm, vma, address,
3360 pte, pmd, ptl, entry);
3361 entry = pte_mkdirty(entry);
3363 entry = pte_mkyoung(entry);
3364 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3365 update_mmu_cache(vma, address, pte);
3366 } else {
3368 * This is needed only for protection faults but the arch code
3369 * is not yet telling us if this is a protection fault or not.
3370 * This still avoids useless tlb flushes for .text page faults
3371 * with threads.
3373 if (flags & FAULT_FLAG_WRITE)
3374 flush_tlb_fix_spurious_fault(vma, address);
3376 unlock:
3377 pte_unmap_unlock(pte, ptl);
3378 return 0;
3382 * By the time we get here, we already hold the mm semaphore
3384 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3385 unsigned long address, unsigned int flags)
3387 pgd_t *pgd;
3388 pud_t *pud;
3389 pmd_t *pmd;
3390 pte_t *pte;
3392 __set_current_state(TASK_RUNNING);
3394 count_vm_event(PGFAULT);
3395 mem_cgroup_count_vm_event(mm, PGFAULT);
3397 /* do counter updates before entering really critical section. */
3398 check_sync_rss_stat(current);
3400 if (unlikely(is_vm_hugetlb_page(vma)))
3401 return hugetlb_fault(mm, vma, address, flags);
3403 pgd = pgd_offset(mm, address);
3404 pud = pud_alloc(mm, pgd, address);
3405 if (!pud)
3406 return VM_FAULT_OOM;
3407 pmd = pmd_alloc(mm, pud, address);
3408 if (!pmd)
3409 return VM_FAULT_OOM;
3410 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3411 if (!vma->vm_ops)
3412 return do_huge_pmd_anonymous_page(mm, vma, address,
3413 pmd, flags);
3414 } else {
3415 pmd_t orig_pmd = *pmd;
3416 barrier();
3417 if (pmd_trans_huge(orig_pmd)) {
3418 if (flags & FAULT_FLAG_WRITE &&
3419 !pmd_write(orig_pmd) &&
3420 !pmd_trans_splitting(orig_pmd))
3421 return do_huge_pmd_wp_page(mm, vma, address,
3422 pmd, orig_pmd);
3423 return 0;
3428 * Use __pte_alloc instead of pte_alloc_map, because we can't
3429 * run pte_offset_map on the pmd, if an huge pmd could
3430 * materialize from under us from a different thread.
3432 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3433 return VM_FAULT_OOM;
3434 /* if an huge pmd materialized from under us just retry later */
3435 if (unlikely(pmd_trans_huge(*pmd)))
3436 return 0;
3438 * A regular pmd is established and it can't morph into a huge pmd
3439 * from under us anymore at this point because we hold the mmap_sem
3440 * read mode and khugepaged takes it in write mode. So now it's
3441 * safe to run pte_offset_map().
3443 pte = pte_offset_map(pmd, address);
3445 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3448 #ifndef __PAGETABLE_PUD_FOLDED
3450 * Allocate page upper directory.
3451 * We've already handled the fast-path in-line.
3453 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3455 pud_t *new = pud_alloc_one(mm, address);
3456 if (!new)
3457 return -ENOMEM;
3459 smp_wmb(); /* See comment in __pte_alloc */
3461 spin_lock(&mm->page_table_lock);
3462 if (pgd_present(*pgd)) /* Another has populated it */
3463 pud_free(mm, new);
3464 else
3465 pgd_populate(mm, pgd, new);
3466 spin_unlock(&mm->page_table_lock);
3467 return 0;
3469 #endif /* __PAGETABLE_PUD_FOLDED */
3471 #ifndef __PAGETABLE_PMD_FOLDED
3473 * Allocate page middle directory.
3474 * We've already handled the fast-path in-line.
3476 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3478 pmd_t *new = pmd_alloc_one(mm, address);
3479 if (!new)
3480 return -ENOMEM;
3482 smp_wmb(); /* See comment in __pte_alloc */
3484 spin_lock(&mm->page_table_lock);
3485 #ifndef __ARCH_HAS_4LEVEL_HACK
3486 if (pud_present(*pud)) /* Another has populated it */
3487 pmd_free(mm, new);
3488 else
3489 pud_populate(mm, pud, new);
3490 #else
3491 if (pgd_present(*pud)) /* Another has populated it */
3492 pmd_free(mm, new);
3493 else
3494 pgd_populate(mm, pud, new);
3495 #endif /* __ARCH_HAS_4LEVEL_HACK */
3496 spin_unlock(&mm->page_table_lock);
3497 return 0;
3499 #endif /* __PAGETABLE_PMD_FOLDED */
3501 int make_pages_present(unsigned long addr, unsigned long end)
3503 int ret, len, write;
3504 struct vm_area_struct * vma;
3506 vma = find_vma(current->mm, addr);
3507 if (!vma)
3508 return -ENOMEM;
3510 * We want to touch writable mappings with a write fault in order
3511 * to break COW, except for shared mappings because these don't COW
3512 * and we would not want to dirty them for nothing.
3514 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3515 BUG_ON(addr >= end);
3516 BUG_ON(end > vma->vm_end);
3517 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3518 ret = get_user_pages(current, current->mm, addr,
3519 len, write, 0, NULL, NULL);
3520 if (ret < 0)
3521 return ret;
3522 return ret == len ? 0 : -EFAULT;
3525 #if !defined(__HAVE_ARCH_GATE_AREA)
3527 #if defined(AT_SYSINFO_EHDR)
3528 static struct vm_area_struct gate_vma;
3530 static int __init gate_vma_init(void)
3532 gate_vma.vm_mm = NULL;
3533 gate_vma.vm_start = FIXADDR_USER_START;
3534 gate_vma.vm_end = FIXADDR_USER_END;
3535 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3536 gate_vma.vm_page_prot = __P101;
3538 * Make sure the vDSO gets into every core dump.
3539 * Dumping its contents makes post-mortem fully interpretable later
3540 * without matching up the same kernel and hardware config to see
3541 * what PC values meant.
3543 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3544 return 0;
3546 __initcall(gate_vma_init);
3547 #endif
3549 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3551 #ifdef AT_SYSINFO_EHDR
3552 return &gate_vma;
3553 #else
3554 return NULL;
3555 #endif
3558 int in_gate_area_no_mm(unsigned long addr)
3560 #ifdef AT_SYSINFO_EHDR
3561 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3562 return 1;
3563 #endif
3564 return 0;
3567 #endif /* __HAVE_ARCH_GATE_AREA */
3569 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3570 pte_t **ptepp, spinlock_t **ptlp)
3572 pgd_t *pgd;
3573 pud_t *pud;
3574 pmd_t *pmd;
3575 pte_t *ptep;
3577 pgd = pgd_offset(mm, address);
3578 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3579 goto out;
3581 pud = pud_offset(pgd, address);
3582 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3583 goto out;
3585 pmd = pmd_offset(pud, address);
3586 VM_BUG_ON(pmd_trans_huge(*pmd));
3587 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3588 goto out;
3590 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3591 if (pmd_huge(*pmd))
3592 goto out;
3594 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3595 if (!ptep)
3596 goto out;
3597 if (!pte_present(*ptep))
3598 goto unlock;
3599 *ptepp = ptep;
3600 return 0;
3601 unlock:
3602 pte_unmap_unlock(ptep, *ptlp);
3603 out:
3604 return -EINVAL;
3607 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3608 pte_t **ptepp, spinlock_t **ptlp)
3610 int res;
3612 /* (void) is needed to make gcc happy */
3613 (void) __cond_lock(*ptlp,
3614 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3615 return res;
3619 * follow_pfn - look up PFN at a user virtual address
3620 * @vma: memory mapping
3621 * @address: user virtual address
3622 * @pfn: location to store found PFN
3624 * Only IO mappings and raw PFN mappings are allowed.
3626 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3628 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3629 unsigned long *pfn)
3631 int ret = -EINVAL;
3632 spinlock_t *ptl;
3633 pte_t *ptep;
3635 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3636 return ret;
3638 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3639 if (ret)
3640 return ret;
3641 *pfn = pte_pfn(*ptep);
3642 pte_unmap_unlock(ptep, ptl);
3643 return 0;
3645 EXPORT_SYMBOL(follow_pfn);
3647 #ifdef CONFIG_HAVE_IOREMAP_PROT
3648 int follow_phys(struct vm_area_struct *vma,
3649 unsigned long address, unsigned int flags,
3650 unsigned long *prot, resource_size_t *phys)
3652 int ret = -EINVAL;
3653 pte_t *ptep, pte;
3654 spinlock_t *ptl;
3656 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3657 goto out;
3659 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3660 goto out;
3661 pte = *ptep;
3663 if ((flags & FOLL_WRITE) && !pte_write(pte))
3664 goto unlock;
3666 *prot = pgprot_val(pte_pgprot(pte));
3667 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3669 ret = 0;
3670 unlock:
3671 pte_unmap_unlock(ptep, ptl);
3672 out:
3673 return ret;
3676 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3677 void *buf, int len, int write)
3679 resource_size_t phys_addr;
3680 unsigned long prot = 0;
3681 void __iomem *maddr;
3682 int offset = addr & (PAGE_SIZE-1);
3684 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3685 return -EINVAL;
3687 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3688 if (write)
3689 memcpy_toio(maddr + offset, buf, len);
3690 else
3691 memcpy_fromio(buf, maddr + offset, len);
3692 iounmap(maddr);
3694 return len;
3696 #endif
3699 * Access another process' address space as given in mm. If non-NULL, use the
3700 * given task for page fault accounting.
3702 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3703 unsigned long addr, void *buf, int len, int write)
3705 struct vm_area_struct *vma;
3706 void *old_buf = buf;
3708 down_read(&mm->mmap_sem);
3709 /* ignore errors, just check how much was successfully transferred */
3710 while (len) {
3711 int bytes, ret, offset;
3712 void *maddr;
3713 struct page *page = NULL;
3715 ret = get_user_pages(tsk, mm, addr, 1,
3716 write, 1, &page, &vma);
3717 if (ret <= 0) {
3719 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3720 * we can access using slightly different code.
3722 #ifdef CONFIG_HAVE_IOREMAP_PROT
3723 vma = find_vma(mm, addr);
3724 if (!vma || vma->vm_start > addr)
3725 break;
3726 if (vma->vm_ops && vma->vm_ops->access)
3727 ret = vma->vm_ops->access(vma, addr, buf,
3728 len, write);
3729 if (ret <= 0)
3730 #endif
3731 break;
3732 bytes = ret;
3733 } else {
3734 bytes = len;
3735 offset = addr & (PAGE_SIZE-1);
3736 if (bytes > PAGE_SIZE-offset)
3737 bytes = PAGE_SIZE-offset;
3739 maddr = kmap(page);
3740 if (write) {
3741 copy_to_user_page(vma, page, addr,
3742 maddr + offset, buf, bytes);
3743 set_page_dirty_lock(page);
3744 } else {
3745 copy_from_user_page(vma, page, addr,
3746 buf, maddr + offset, bytes);
3748 kunmap(page);
3749 page_cache_release(page);
3751 len -= bytes;
3752 buf += bytes;
3753 addr += bytes;
3755 up_read(&mm->mmap_sem);
3757 return buf - old_buf;
3761 * access_remote_vm - access another process' address space
3762 * @mm: the mm_struct of the target address space
3763 * @addr: start address to access
3764 * @buf: source or destination buffer
3765 * @len: number of bytes to transfer
3766 * @write: whether the access is a write
3768 * The caller must hold a reference on @mm.
3770 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3771 void *buf, int len, int write)
3773 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3777 * Access another process' address space.
3778 * Source/target buffer must be kernel space,
3779 * Do not walk the page table directly, use get_user_pages
3781 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3782 void *buf, int len, int write)
3784 struct mm_struct *mm;
3785 int ret;
3787 mm = get_task_mm(tsk);
3788 if (!mm)
3789 return 0;
3791 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3792 mmput(mm);
3794 return ret;
3798 * Print the name of a VMA.
3800 void print_vma_addr(char *prefix, unsigned long ip)
3802 struct mm_struct *mm = current->mm;
3803 struct vm_area_struct *vma;
3806 * Do not print if we are in atomic
3807 * contexts (in exception stacks, etc.):
3809 if (preempt_count())
3810 return;
3812 down_read(&mm->mmap_sem);
3813 vma = find_vma(mm, ip);
3814 if (vma && vma->vm_file) {
3815 struct file *f = vma->vm_file;
3816 char *buf = (char *)__get_free_page(GFP_KERNEL);
3817 if (buf) {
3818 char *p, *s;
3820 p = d_path(&f->f_path, buf, PAGE_SIZE);
3821 if (IS_ERR(p))
3822 p = "?";
3823 s = strrchr(p, '/');
3824 if (s)
3825 p = s+1;
3826 printk("%s%s[%lx+%lx]", prefix, p,
3827 vma->vm_start,
3828 vma->vm_end - vma->vm_start);
3829 free_page((unsigned long)buf);
3832 up_read(&current->mm->mmap_sem);
3835 #ifdef CONFIG_PROVE_LOCKING
3836 void might_fault(void)
3839 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3840 * holding the mmap_sem, this is safe because kernel memory doesn't
3841 * get paged out, therefore we'll never actually fault, and the
3842 * below annotations will generate false positives.
3844 if (segment_eq(get_fs(), KERNEL_DS))
3845 return;
3847 might_sleep();
3849 * it would be nicer only to annotate paths which are not under
3850 * pagefault_disable, however that requires a larger audit and
3851 * providing helpers like get_user_atomic.
3853 if (!in_atomic() && current->mm)
3854 might_lock_read(&current->mm->mmap_sem);
3856 EXPORT_SYMBOL(might_fault);
3857 #endif
3859 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3860 static void clear_gigantic_page(struct page *page,
3861 unsigned long addr,
3862 unsigned int pages_per_huge_page)
3864 int i;
3865 struct page *p = page;
3867 might_sleep();
3868 for (i = 0; i < pages_per_huge_page;
3869 i++, p = mem_map_next(p, page, i)) {
3870 cond_resched();
3871 clear_user_highpage(p, addr + i * PAGE_SIZE);
3874 void clear_huge_page(struct page *page,
3875 unsigned long addr, unsigned int pages_per_huge_page)
3877 int i;
3879 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3880 clear_gigantic_page(page, addr, pages_per_huge_page);
3881 return;
3884 might_sleep();
3885 for (i = 0; i < pages_per_huge_page; i++) {
3886 cond_resched();
3887 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3891 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3892 unsigned long addr,
3893 struct vm_area_struct *vma,
3894 unsigned int pages_per_huge_page)
3896 int i;
3897 struct page *dst_base = dst;
3898 struct page *src_base = src;
3900 for (i = 0; i < pages_per_huge_page; ) {
3901 cond_resched();
3902 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3904 i++;
3905 dst = mem_map_next(dst, dst_base, i);
3906 src = mem_map_next(src, src_base, i);
3910 void copy_user_huge_page(struct page *dst, struct page *src,
3911 unsigned long addr, struct vm_area_struct *vma,
3912 unsigned int pages_per_huge_page)
3914 int i;
3916 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3917 copy_user_gigantic_page(dst, src, addr, vma,
3918 pages_per_huge_page);
3919 return;
3922 might_sleep();
3923 for (i = 0; i < pages_per_huge_page; i++) {
3924 cond_resched();
3925 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3928 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */