blob: 61e66f026563b4b473fc73922d58a984c490c827 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
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
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
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.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
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.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
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>
Hugh Dickins9a840892009-09-21 17:02:01 -070048#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/rmap.h>
50#include <linux/module.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070053#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080054#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070055#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080056#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090059#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Alexey Dobriyan6952b612009-09-18 23:55:55 +040061#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062#include <asm/pgalloc.h>
63#include <asm/uaccess.h>
64#include <asm/tlb.h>
65#include <asm/tlbflush.h>
66#include <asm/pgtable.h>
67
Jan Beulich42b77722008-07-23 21:27:10 -070068#include "internal.h"
69
Andy Whitcroftd41dee32005-06-23 00:07:54 -070070#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070071/* use the per-pgdat data instead for discontigmem - mbligh */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned long num_physpages;
80/*
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.
86 */
87void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89EXPORT_SYMBOL(num_physpages);
90EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Ingo Molnar32a93232008-02-06 22:39:44 +010092/*
93 * Randomize the address space (stacks, mmaps, brk, etc.).
94 *
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
97 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100 1;
101#else
102 2;
103#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100104
105static int __init disable_randmaps(char *s)
106{
107 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800108 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100109}
110__setup("norandmaps", disable_randmaps);
111
Hugh Dickins62eede62009-09-21 17:03:34 -0700112unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700113unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
122}
123core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100124
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800125
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800126#if defined(SPLIT_RSS_COUNTING)
127
KAMEZAWA Hiroyukia3a2e762010-04-06 14:34:42 -0700128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800129{
130 int i;
131
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;
136 }
137 }
138 task->rss_stat.events = 0;
139}
140
141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142{
143 struct task_struct *task = current;
144
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
149}
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)
152
153/* sync counter once per 64 page faults */
154#define TASK_RSS_EVENTS_THRESH (64)
155static void check_sync_rss_stat(struct task_struct *task)
156{
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);
161}
162
163unsigned long get_mm_counter(struct mm_struct *mm, int member)
164{
165 long val = 0;
166
167 /*
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
170 */
171 val = atomic_long_read(&mm->rss_stat.count[member]);
172 /*
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
175 */
176 if (val < 0)
177 return 0;
178 return (unsigned long)val;
179}
180
181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182{
183 __sync_task_rss_stat(task, mm);
184}
185#else
186
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)
189
190static void check_sync_rss_stat(struct task_struct *task)
191{
192}
193
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800194#endif
195
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196/*
197 * If a p?d_bad entry is found while walking page tables, report
198 * the error, before resetting entry to p?d_none. Usually (but
199 * very seldom) called out from the p?d_none_or_clear_bad macros.
200 */
201
202void pgd_clear_bad(pgd_t *pgd)
203{
204 pgd_ERROR(*pgd);
205 pgd_clear(pgd);
206}
207
208void pud_clear_bad(pud_t *pud)
209{
210 pud_ERROR(*pud);
211 pud_clear(pud);
212}
213
214void pmd_clear_bad(pmd_t *pmd)
215{
216 pmd_ERROR(*pmd);
217 pmd_clear(pmd);
218}
219
220/*
221 * Note: this doesn't free the actual pages themselves. That
222 * has been handled earlier when unmapping all the memory regions.
223 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000224static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
225 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800227 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700228 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000229 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700230 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
Hugh Dickinse0da3822005-04-19 13:29:15 -0700233static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
234 unsigned long addr, unsigned long end,
235 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 pmd_t *pmd;
238 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700239 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240
Hugh Dickinse0da3822005-04-19 13:29:15 -0700241 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 do {
244 next = pmd_addr_end(addr, end);
245 if (pmd_none_or_clear_bad(pmd))
246 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000247 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 } while (pmd++, addr = next, addr != end);
249
Hugh Dickinse0da3822005-04-19 13:29:15 -0700250 start &= PUD_MASK;
251 if (start < floor)
252 return;
253 if (ceiling) {
254 ceiling &= PUD_MASK;
255 if (!ceiling)
256 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700258 if (end - 1 > ceiling - 1)
259 return;
260
261 pmd = pmd_offset(pud, start);
262 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000263 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264}
265
Hugh Dickinse0da3822005-04-19 13:29:15 -0700266static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267 unsigned long addr, unsigned long end,
268 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269{
270 pud_t *pud;
271 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700272 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Hugh Dickinse0da3822005-04-19 13:29:15 -0700274 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 do {
277 next = pud_addr_end(addr, end);
278 if (pud_none_or_clear_bad(pud))
279 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700280 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 } while (pud++, addr = next, addr != end);
282
Hugh Dickinse0da3822005-04-19 13:29:15 -0700283 start &= PGDIR_MASK;
284 if (start < floor)
285 return;
286 if (ceiling) {
287 ceiling &= PGDIR_MASK;
288 if (!ceiling)
289 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700291 if (end - 1 > ceiling - 1)
292 return;
293
294 pud = pud_offset(pgd, start);
295 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000296 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297}
298
299/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700300 * This function frees user-level page tables of a process.
301 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 * Must be called with pagetable lock held.
303 */
Jan Beulich42b77722008-07-23 21:27:10 -0700304void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700305 unsigned long addr, unsigned long end,
306 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307{
308 pgd_t *pgd;
309 unsigned long next;
310
Hugh Dickinse0da3822005-04-19 13:29:15 -0700311 /*
312 * The next few lines have given us lots of grief...
313 *
314 * Why are we testing PMD* at this top level? Because often
315 * there will be no work to do at all, and we'd prefer not to
316 * go all the way down to the bottom just to discover that.
317 *
318 * Why all these "- 1"s? Because 0 represents both the bottom
319 * of the address space and the top of it (using -1 for the
320 * top wouldn't help much: the masks would do the wrong thing).
321 * The rule is that addr 0 and floor 0 refer to the bottom of
322 * the address space, but end 0 and ceiling 0 refer to the top
323 * Comparisons need to use "end - 1" and "ceiling - 1" (though
324 * that end 0 case should be mythical).
325 *
326 * Wherever addr is brought up or ceiling brought down, we must
327 * be careful to reject "the opposite 0" before it confuses the
328 * subsequent tests. But what about where end is brought down
329 * by PMD_SIZE below? no, end can't go down to 0 there.
330 *
331 * Whereas we round start (addr) and ceiling down, by different
332 * masks at different levels, in order to test whether a table
333 * now has no other vmas using it, so can be freed, we don't
334 * bother to round floor or end up - the tests don't need that.
335 */
336
337 addr &= PMD_MASK;
338 if (addr < floor) {
339 addr += PMD_SIZE;
340 if (!addr)
341 return;
342 }
343 if (ceiling) {
344 ceiling &= PMD_MASK;
345 if (!ceiling)
346 return;
347 }
348 if (end - 1 > ceiling - 1)
349 end -= PMD_SIZE;
350 if (addr > end - 1)
351 return;
352
Jan Beulich42b77722008-07-23 21:27:10 -0700353 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 do {
355 next = pgd_addr_end(addr, end);
356 if (pgd_none_or_clear_bad(pgd))
357 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700358 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700360}
361
Jan Beulich42b77722008-07-23 21:27:10 -0700362void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700363 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700364{
365 while (vma) {
366 struct vm_area_struct *next = vma->vm_next;
367 unsigned long addr = vma->vm_start;
368
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700369 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000370 * Hide vma from rmap and truncate_pagecache before freeing
371 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700372 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800373 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700374 unlink_file_vma(vma);
375
David Gibson9da61ae2006-03-22 00:08:57 -0800376 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700377 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700378 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700379 } else {
380 /*
381 * Optimization: gather nearby vmas into one call down
382 */
383 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800384 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700385 vma = next;
386 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800387 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700388 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700389 }
390 free_pgd_range(tlb, addr, vma->vm_end,
391 floor, next? next->vm_start: ceiling);
392 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700393 vma = next;
394 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395}
396
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800397int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
398 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800400 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800401 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700402 if (!new)
403 return -ENOMEM;
404
Nick Piggin362a61a2008-05-14 06:37:36 +0200405 /*
406 * Ensure all pte setup (eg. pte page lock and page clearing) are
407 * visible before the pte is made visible to other CPUs by being
408 * put into page tables.
409 *
410 * The other side of the story is the pointer chasing in the page
411 * table walking code (when walking the page table without locking;
412 * ie. most of the time). Fortunately, these data accesses consist
413 * of a chain of data-dependent loads, meaning most CPUs (alpha
414 * being the notable exception) will already guarantee loads are
415 * seen in-order. See the alpha page table accessors for the
416 * smp_read_barrier_depends() barriers in page table walking code.
417 */
418 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
419
Hugh Dickinsc74df322005-10-29 18:16:23 -0700420 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800421 wait_split_huge_page = 0;
422 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800425 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800426 } else if (unlikely(pmd_trans_splitting(*pmd)))
427 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700428 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800429 if (new)
430 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800431 if (wait_split_huge_page)
432 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700433 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
Hugh Dickins1bb36302005-10-29 18:16:22 -0700436int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700438 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
439 if (!new)
440 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Nick Piggin362a61a2008-05-14 06:37:36 +0200442 smp_wmb(); /* See comment in __pte_alloc */
443
Hugh Dickins1bb36302005-10-29 18:16:22 -0700444 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800445 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700446 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800447 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800448 } else
449 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700450 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800451 if (new)
452 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700453 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454}
455
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800456static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700457{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800458 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459}
460
461static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462{
463 int i;
464
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800465 if (current->mm == mm)
466 sync_mm_rss(current, mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800467 for (i = 0; i < NR_MM_COUNTERS; i++)
468 if (rss[i])
469 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700470}
471
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800473 * This function is called to print an error when a bad pte
474 * is found. For example, we might have a PFN-mapped pte in
475 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700476 *
477 * The calling function must still handle the error.
478 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800479static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700481{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800482 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483 pud_t *pud = pud_offset(pgd, addr);
484 pmd_t *pmd = pmd_offset(pud, addr);
485 struct address_space *mapping;
486 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800487 static unsigned long resume;
488 static unsigned long nr_shown;
489 static unsigned long nr_unshown;
490
491 /*
492 * Allow a burst of 60 reports, then keep quiet for that minute;
493 * or allow a steady drip of one report per second.
494 */
495 if (nr_shown == 60) {
496 if (time_before(jiffies, resume)) {
497 nr_unshown++;
498 return;
499 }
500 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800501 printk(KERN_ALERT
502 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800503 nr_unshown);
504 nr_unshown = 0;
505 }
506 nr_shown = 0;
507 }
508 if (nr_shown++ == 0)
509 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800510
511 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512 index = linear_page_index(vma, addr);
513
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800514 printk(KERN_ALERT
515 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800516 current->comm,
517 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800518 if (page)
519 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800520 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800521 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523 /*
524 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
525 */
526 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800527 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800528 (unsigned long)vma->vm_ops->fault);
529 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800530 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800531 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700532 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800533 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700534}
535
Linus Torvalds67121172005-12-11 20:38:17 -0800536static inline int is_cow_mapping(unsigned int flags)
537{
538 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
539}
540
Hugh Dickins62eede62009-09-21 17:03:34 -0700541#ifndef is_zero_pfn
542static inline int is_zero_pfn(unsigned long pfn)
543{
544 return pfn == zero_pfn;
545}
546#endif
547
548#ifndef my_zero_pfn
549static inline unsigned long my_zero_pfn(unsigned long addr)
550{
551 return zero_pfn;
552}
553#endif
554
Nick Pigginb5810032005-10-29 18:16:12 -0700555/*
Nick Piggin7e675132008-04-28 02:13:00 -0700556 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800557 *
Nick Piggin7e675132008-04-28 02:13:00 -0700558 * "Special" mappings do not wish to be associated with a "struct page" (either
559 * it doesn't exist, or it exists but they don't want to touch it). In this
560 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700561 *
Nick Piggin7e675132008-04-28 02:13:00 -0700562 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
563 * pte bit, in which case this function is trivial. Secondly, an architecture
564 * may not have a spare pte bit, which requires a more complicated scheme,
565 * described below.
566 *
567 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
568 * special mapping (even if there are underlying and valid "struct pages").
569 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800570 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700571 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
572 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700573 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
574 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800575 *
576 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
577 *
Nick Piggin7e675132008-04-28 02:13:00 -0700578 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700579 *
Nick Piggin7e675132008-04-28 02:13:00 -0700580 * This restricts such mappings to be a linear translation from virtual address
581 * to pfn. To get around this restriction, we allow arbitrary mappings so long
582 * as the vma is not a COW mapping; in that case, we know that all ptes are
583 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700584 *
585 *
Nick Piggin7e675132008-04-28 02:13:00 -0700586 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
587 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700588 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
589 * page" backing, however the difference is that _all_ pages with a struct
590 * page (that is, those where pfn_valid is true) are refcounted and considered
591 * normal pages by the VM. The disadvantage is that pages are refcounted
592 * (which can be slower and simply not an option for some PFNMAP users). The
593 * advantage is that we don't have to follow the strict linearity rule of
594 * PFNMAP mappings in order to support COWable mappings.
595 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800596 */
Nick Piggin7e675132008-04-28 02:13:00 -0700597#ifdef __HAVE_ARCH_PTE_SPECIAL
598# define HAVE_PTE_SPECIAL 1
599#else
600# define HAVE_PTE_SPECIAL 0
601#endif
602struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
603 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800604{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800605 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700606
607 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800608 if (likely(!pte_special(pte)))
609 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700610 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
611 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700612 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800613 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700614 return NULL;
615 }
616
617 /* !HAVE_PTE_SPECIAL case follows: */
618
Jared Hulbertb379d792008-04-28 02:12:58 -0700619 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620 if (vma->vm_flags & VM_MIXEDMAP) {
621 if (!pfn_valid(pfn))
622 return NULL;
623 goto out;
624 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700625 unsigned long off;
626 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700627 if (pfn == vma->vm_pgoff + off)
628 return NULL;
629 if (!is_cow_mapping(vma->vm_flags))
630 return NULL;
631 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800632 }
633
Hugh Dickins62eede62009-09-21 17:03:34 -0700634 if (is_zero_pfn(pfn))
635 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800636check_pfn:
637 if (unlikely(pfn > highest_memmap_pfn)) {
638 print_bad_pte(vma, addr, pte, NULL);
639 return NULL;
640 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800641
642 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700643 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700644 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800645 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700646out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800647 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800648}
649
650/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 * copy one vm_area from one task to the other. Assumes the page tables
652 * already present in the new task to be cleared in the whole range
653 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 */
655
Hugh Dickins570a3352009-12-14 17:58:46 -0800656static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700658 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700659 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660{
Nick Pigginb5810032005-10-29 18:16:12 -0700661 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 pte_t pte = *src_pte;
663 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665 /* pte contains position in swap or file, so copy. */
666 if (unlikely(!pte_present(pte))) {
667 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700668 swp_entry_t entry = pte_to_swp_entry(pte);
669
Hugh Dickins570a3352009-12-14 17:58:46 -0800670 if (swap_duplicate(entry) < 0)
671 return entry.val;
672
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 /* make sure dst_mm is on swapoff's mmlist. */
674 if (unlikely(list_empty(&dst_mm->mmlist))) {
675 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700676 if (list_empty(&dst_mm->mmlist))
677 list_add(&dst_mm->mmlist,
678 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 spin_unlock(&mmlist_lock);
680 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800681 if (likely(!non_swap_entry(entry)))
682 rss[MM_SWAPENTS]++;
683 else if (is_write_migration_entry(entry) &&
Christoph Lameter06972122006-06-23 02:03:35 -0700684 is_cow_mapping(vm_flags)) {
685 /*
686 * COW mappings require pages in both parent
687 * and child to be set to read.
688 */
689 make_migration_entry_read(&entry);
690 pte = swp_entry_to_pte(entry);
691 set_pte_at(src_mm, addr, src_pte, pte);
692 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700694 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 }
696
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 /*
698 * If it's a COW mapping, write protect it both
699 * in the parent and the child
700 */
Linus Torvalds67121172005-12-11 20:38:17 -0800701 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700703 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 }
705
706 /*
707 * If it's a shared mapping, mark it clean in
708 * the child
709 */
710 if (vm_flags & VM_SHARED)
711 pte = pte_mkclean(pte);
712 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800713
714 page = vm_normal_page(vma, addr, pte);
715 if (page) {
716 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700717 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800718 if (PageAnon(page))
719 rss[MM_ANONPAGES]++;
720 else
721 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800722 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700723
724out_set_pte:
725 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a3352009-12-14 17:58:46 -0800726 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727}
728
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800729int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
730 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
731 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700733 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700735 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700736 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800737 int rss[NR_MM_COUNTERS];
Hugh Dickins570a3352009-12-14 17:58:46 -0800738 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739
740again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800741 init_rss_vec(rss);
742
Hugh Dickinsc74df322005-10-29 18:16:23 -0700743 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 if (!dst_pte)
745 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700746 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700747 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700748 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700749 orig_src_pte = src_pte;
750 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700751 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 do {
754 /*
755 * We are holding two locks at this point - either of them
756 * could generate latencies in another task on another CPU.
757 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700758 if (progress >= 32) {
759 progress = 0;
760 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100761 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700762 break;
763 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 if (pte_none(*src_pte)) {
765 progress++;
766 continue;
767 }
Hugh Dickins570a3352009-12-14 17:58:46 -0800768 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
769 vma, addr, rss);
770 if (entry.val)
771 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 progress += 8;
773 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700775 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700776 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700777 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800778 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700779 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700780 cond_resched();
Hugh Dickins570a3352009-12-14 17:58:46 -0800781
782 if (entry.val) {
783 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
784 return -ENOMEM;
785 progress = 0;
786 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 if (addr != end)
788 goto again;
789 return 0;
790}
791
792static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
793 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
794 unsigned long addr, unsigned long end)
795{
796 pmd_t *src_pmd, *dst_pmd;
797 unsigned long next;
798
799 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
800 if (!dst_pmd)
801 return -ENOMEM;
802 src_pmd = pmd_offset(src_pud, addr);
803 do {
804 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800805 if (pmd_trans_huge(*src_pmd)) {
806 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800807 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800808 err = copy_huge_pmd(dst_mm, src_mm,
809 dst_pmd, src_pmd, addr, vma);
810 if (err == -ENOMEM)
811 return -ENOMEM;
812 if (!err)
813 continue;
814 /* fall through */
815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 if (pmd_none_or_clear_bad(src_pmd))
817 continue;
818 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
819 vma, addr, next))
820 return -ENOMEM;
821 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
822 return 0;
823}
824
825static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
826 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
827 unsigned long addr, unsigned long end)
828{
829 pud_t *src_pud, *dst_pud;
830 unsigned long next;
831
832 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
833 if (!dst_pud)
834 return -ENOMEM;
835 src_pud = pud_offset(src_pgd, addr);
836 do {
837 next = pud_addr_end(addr, end);
838 if (pud_none_or_clear_bad(src_pud))
839 continue;
840 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
841 vma, addr, next))
842 return -ENOMEM;
843 } while (dst_pud++, src_pud++, addr = next, addr != end);
844 return 0;
845}
846
847int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848 struct vm_area_struct *vma)
849{
850 pgd_t *src_pgd, *dst_pgd;
851 unsigned long next;
852 unsigned long addr = vma->vm_start;
853 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700854 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
Nick Piggind9928952005-08-28 16:49:11 +1000856 /*
857 * Don't copy ptes where a page fault will fill them correctly.
858 * Fork becomes much lighter when there are big shared or private
859 * readonly mappings. The tradeoff is that copy_page_range is more
860 * efficient than faulting.
861 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -0800862 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +1000863 if (!vma->anon_vma)
864 return 0;
865 }
866
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 if (is_vm_hugetlb_page(vma))
868 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
869
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -0800870 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -0800871 /*
872 * We do not free on error cases below as remove_vma
873 * gets called on error from higher level routine
874 */
875 ret = track_pfn_vma_copy(vma);
876 if (ret)
877 return ret;
878 }
879
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700880 /*
881 * We need to invalidate the secondary MMU mappings only when
882 * there could be a permission downgrade on the ptes of the
883 * parent mm. And a permission downgrade will only happen if
884 * is_cow_mapping() returns true.
885 */
886 if (is_cow_mapping(vma->vm_flags))
887 mmu_notifier_invalidate_range_start(src_mm, addr, end);
888
889 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 dst_pgd = pgd_offset(dst_mm, addr);
891 src_pgd = pgd_offset(src_mm, addr);
892 do {
893 next = pgd_addr_end(addr, end);
894 if (pgd_none_or_clear_bad(src_pgd))
895 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700896 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
897 vma, addr, next))) {
898 ret = -ENOMEM;
899 break;
900 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700902
903 if (is_cow_mapping(vma->vm_flags))
904 mmu_notifier_invalidate_range_end(src_mm,
905 vma->vm_start, end);
906 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907}
908
Robin Holt51c6f662005-11-13 16:06:42 -0800909static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700910 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800912 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
Nick Pigginb5810032005-10-29 18:16:12 -0700914 struct mm_struct *mm = tlb->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 pte_t *pte;
Hugh Dickins508034a2005-10-29 18:16:30 -0700916 spinlock_t *ptl;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800917 int rss[NR_MM_COUNTERS];
918
919 init_rss_vec(rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
Hugh Dickins508034a2005-10-29 18:16:30 -0700921 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700922 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 do {
924 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -0800925 if (pte_none(ptent)) {
926 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800928 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -0800929
930 (*zap_work) -= PAGE_SIZE;
931
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800933 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -0800934
Linus Torvalds6aab3412005-11-28 14:34:23 -0800935 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 if (unlikely(details) && page) {
937 /*
938 * unmap_shared_mapping_pages() wants to
939 * invalidate cache without truncating:
940 * unmap shared but keep private pages.
941 */
942 if (details->check_mapping &&
943 details->check_mapping != page->mapping)
944 continue;
945 /*
946 * Each page->index must be checked when
947 * invalidating or truncating nonlinear.
948 */
949 if (details->nonlinear_vma &&
950 (page->index < details->first_index ||
951 page->index > details->last_index))
952 continue;
953 }
Nick Pigginb5810032005-10-29 18:16:12 -0700954 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -0700955 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 tlb_remove_tlb_entry(tlb, pte, addr);
957 if (unlikely(!page))
958 continue;
959 if (unlikely(details) && details->nonlinear_vma
960 && linear_page_index(details->nonlinear_vma,
961 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -0700962 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800965 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700966 else {
967 if (pte_dirty(ptent))
968 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800969 if (pte_young(ptent) &&
970 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -0800971 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800972 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700973 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800974 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800975 if (unlikely(page_mapcount(page) < 0))
976 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 tlb_remove_page(tlb, page);
978 continue;
979 }
980 /*
981 * If details->check_mapping, we leave swap entries;
982 * if details->nonlinear_vma, we leave file entries.
983 */
984 if (unlikely(details))
985 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -0800986 if (pte_file(ptent)) {
987 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
988 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800989 } else {
990 swp_entry_t entry = pte_to_swp_entry(ptent);
991
992 if (!non_swap_entry(entry))
993 rss[MM_SWAPENTS]--;
994 if (unlikely(!free_swap_and_cache(entry)))
995 print_bad_pte(vma, addr, ptent, NULL);
996 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -0700997 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Robin Holt51c6f662005-11-13 16:06:42 -0800998 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
Hugh Dickinsae859762005-10-29 18:16:05 -0700999
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001000 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001001 arch_leave_lazy_mmu_mode();
Hugh Dickins508034a2005-10-29 18:16:30 -07001002 pte_unmap_unlock(pte - 1, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001003
1004 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005}
1006
Robin Holt51c6f662005-11-13 16:06:42 -08001007static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001008 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -08001010 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011{
1012 pmd_t *pmd;
1013 unsigned long next;
1014
1015 pmd = pmd_offset(pud, addr);
1016 do {
1017 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001018 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001019 if (next-addr != HPAGE_PMD_SIZE) {
1020 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001021 split_huge_page_pmd(vma->vm_mm, pmd);
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001022 } else if (zap_huge_pmd(tlb, vma, pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001023 (*zap_work)--;
1024 continue;
1025 }
1026 /* fall through */
1027 }
Robin Holt51c6f662005-11-13 16:06:42 -08001028 if (pmd_none_or_clear_bad(pmd)) {
1029 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001031 }
1032 next = zap_pte_range(tlb, vma, pmd, addr, next,
1033 zap_work, details);
1034 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1035
1036 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037}
1038
Robin Holt51c6f662005-11-13 16:06:42 -08001039static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001040 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -08001042 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043{
1044 pud_t *pud;
1045 unsigned long next;
1046
1047 pud = pud_offset(pgd, addr);
1048 do {
1049 next = pud_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -08001050 if (pud_none_or_clear_bad(pud)) {
1051 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001053 }
1054 next = zap_pmd_range(tlb, vma, pud, addr, next,
1055 zap_work, details);
1056 } while (pud++, addr = next, (addr != end && *zap_work > 0));
1057
1058 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059}
1060
Robin Holt51c6f662005-11-13 16:06:42 -08001061static unsigned long unmap_page_range(struct mmu_gather *tlb,
1062 struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -08001064 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065{
1066 pgd_t *pgd;
1067 unsigned long next;
1068
1069 if (details && !details->check_mapping && !details->nonlinear_vma)
1070 details = NULL;
1071
1072 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001073 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 tlb_start_vma(tlb, vma);
1075 pgd = pgd_offset(vma->vm_mm, addr);
1076 do {
1077 next = pgd_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -08001078 if (pgd_none_or_clear_bad(pgd)) {
1079 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001081 }
1082 next = zap_pud_range(tlb, vma, pgd, addr, next,
1083 zap_work, details);
1084 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001086 mem_cgroup_uncharge_end();
Robin Holt51c6f662005-11-13 16:06:42 -08001087
1088 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
1091#ifdef CONFIG_PREEMPT
1092# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1093#else
1094/* No preempt: go for improved straight-line efficiency */
1095# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1096#endif
1097
1098/**
1099 * unmap_vmas - unmap a range of memory covered by a list of vma's
1100 * @tlbp: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 * @vma: the starting vma
1102 * @start_addr: virtual address at which to start unmapping
1103 * @end_addr: virtual address at which to end unmapping
1104 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1105 * @details: details of nonlinear truncation or shared cache invalidation
1106 *
Hugh Dickinsee39b372005-04-19 13:29:15 -07001107 * Returns the end address of the unmapping (restart addr if interrupted).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001109 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001111 * We aim to not hold locks for too long (for scheduling latency reasons).
1112 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 * return the ending mmu_gather to the caller.
1114 *
1115 * Only addresses between `start' and `end' will be unmapped.
1116 *
1117 * The VMA list must be sorted in ascending virtual address order.
1118 *
1119 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1120 * range after unmap_vmas() returns. So the only responsibility here is to
1121 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1122 * drops the lock and schedules.
1123 */
Hugh Dickins508034a2005-10-29 18:16:30 -07001124unsigned long unmap_vmas(struct mmu_gather **tlbp,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 struct vm_area_struct *vma, unsigned long start_addr,
1126 unsigned long end_addr, unsigned long *nr_accounted,
1127 struct zap_details *details)
1128{
Robin Holt51c6f662005-11-13 16:06:42 -08001129 long zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
1131 int tlb_start_valid = 0;
Hugh Dickinsee39b372005-04-19 13:29:15 -07001132 unsigned long start = start_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
Hugh Dickins4d6ddfa2005-10-29 18:16:02 -07001134 int fullmm = (*tlbp)->fullmm;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001135 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001137 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 unsigned long end;
1140
1141 start = max(vma->vm_start, start_addr);
1142 if (start >= vma->vm_end)
1143 continue;
1144 end = min(vma->vm_end, end_addr);
1145 if (end <= vma->vm_start)
1146 continue;
1147
1148 if (vma->vm_flags & VM_ACCOUNT)
1149 *nr_accounted += (end - start) >> PAGE_SHIFT;
1150
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001151 if (unlikely(is_pfn_mapping(vma)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001152 untrack_pfn_vma(vma, 0, 0);
1153
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 while (start != end) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 if (!tlb_start_valid) {
1156 tlb_start = start;
1157 tlb_start_valid = 1;
1158 }
1159
Robin Holt51c6f662005-11-13 16:06:42 -08001160 if (unlikely(is_vm_hugetlb_page(vma))) {
Andi Kleena137e1c2008-07-23 21:27:43 -07001161 /*
1162 * It is undesirable to test vma->vm_file as it
1163 * should be non-null for valid hugetlb area.
1164 * However, vm_file will be NULL in the error
1165 * cleanup path of do_mmap_pgoff. When
1166 * hugetlbfs ->mmap method fails,
1167 * do_mmap_pgoff() nullifies vma->vm_file
1168 * before calling this function to clean up.
1169 * Since no pte has actually been setup, it is
1170 * safe to do nothing in this case.
1171 */
1172 if (vma->vm_file) {
1173 unmap_hugepage_range(vma, start, end, NULL);
1174 zap_work -= (end - start) /
Andi Kleena5516432008-07-23 21:27:41 -07001175 pages_per_huge_page(hstate_vma(vma));
Andi Kleena137e1c2008-07-23 21:27:43 -07001176 }
1177
Robin Holt51c6f662005-11-13 16:06:42 -08001178 start = end;
1179 } else
1180 start = unmap_page_range(*tlbp, vma,
1181 start, end, &zap_work, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182
Robin Holt51c6f662005-11-13 16:06:42 -08001183 if (zap_work > 0) {
1184 BUG_ON(start != end);
1185 break;
1186 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187
1188 tlb_finish_mmu(*tlbp, tlb_start, start);
1189
1190 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +01001191 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 if (i_mmap_lock) {
Hugh Dickins508034a2005-10-29 18:16:30 -07001193 *tlbp = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 goto out;
1195 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 }
1198
Hugh Dickins508034a2005-10-29 18:16:30 -07001199 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 tlb_start_valid = 0;
Robin Holt51c6f662005-11-13 16:06:42 -08001201 zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 }
1203 }
1204out:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001205 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001206 return start; /* which is now the end (or restart) address */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207}
1208
1209/**
1210 * zap_page_range - remove user pages in a given range
1211 * @vma: vm_area_struct holding the applicable pages
1212 * @address: starting address of pages to zap
1213 * @size: number of bytes to zap
1214 * @details: details of nonlinear truncation or shared cache invalidation
1215 */
Hugh Dickinsee39b372005-04-19 13:29:15 -07001216unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 unsigned long size, struct zap_details *details)
1218{
1219 struct mm_struct *mm = vma->vm_mm;
1220 struct mmu_gather *tlb;
1221 unsigned long end = address + size;
1222 unsigned long nr_accounted = 0;
1223
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 lru_add_drain();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 tlb = tlb_gather_mmu(mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001226 update_hiwater_rss(mm);
Hugh Dickins508034a2005-10-29 18:16:30 -07001227 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1228 if (tlb)
1229 tlb_finish_mmu(tlb, address, end);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001230 return end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231}
1232
Jack Steinerc627f9c2008-07-29 22:33:53 -07001233/**
1234 * zap_vma_ptes - remove ptes mapping the vma
1235 * @vma: vm_area_struct holding ptes to be zapped
1236 * @address: starting address of pages to zap
1237 * @size: number of bytes to zap
1238 *
1239 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1240 *
1241 * The entire address range must be fully contained within the vma.
1242 *
1243 * Returns 0 if successful.
1244 */
1245int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1246 unsigned long size)
1247{
1248 if (address < vma->vm_start || address + size > vma->vm_end ||
1249 !(vma->vm_flags & VM_PFNMAP))
1250 return -1;
1251 zap_page_range(vma, address, size, NULL);
1252 return 0;
1253}
1254EXPORT_SYMBOL_GPL(zap_vma_ptes);
1255
Johannes Weiner142762b2010-05-24 14:32:39 -07001256/**
1257 * follow_page - look up a page descriptor from a user-virtual address
1258 * @vma: vm_area_struct mapping @address
1259 * @address: virtual address to look up
1260 * @flags: flags modifying lookup behaviour
1261 *
1262 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1263 *
1264 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1265 * an error pointer if there is a mapping to something not represented
1266 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001268struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001269 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270{
1271 pgd_t *pgd;
1272 pud_t *pud;
1273 pmd_t *pmd;
1274 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001275 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001277 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001279 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1280 if (!IS_ERR(page)) {
1281 BUG_ON(flags & FOLL_GET);
1282 goto out;
1283 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001285 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 pgd = pgd_offset(mm, address);
1287 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001288 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289
1290 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001291 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001292 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001293 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001294 BUG_ON(flags & FOLL_GET);
1295 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1296 goto out;
1297 }
1298 if (unlikely(pud_bad(*pud)))
1299 goto no_page_table;
1300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001302 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001303 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001304 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001305 BUG_ON(flags & FOLL_GET);
1306 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1307 goto out;
1308 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001309 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001310 if (flags & FOLL_SPLIT) {
1311 split_huge_page_pmd(mm, pmd);
1312 goto split_fallthrough;
1313 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001314 spin_lock(&mm->page_table_lock);
1315 if (likely(pmd_trans_huge(*pmd))) {
1316 if (unlikely(pmd_trans_splitting(*pmd))) {
1317 spin_unlock(&mm->page_table_lock);
1318 wait_split_huge_page(vma->anon_vma, pmd);
1319 } else {
1320 page = follow_trans_huge_pmd(mm, address,
1321 pmd, flags);
1322 spin_unlock(&mm->page_table_lock);
1323 goto out;
1324 }
1325 } else
1326 spin_unlock(&mm->page_table_lock);
1327 /* fall through */
1328 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001329split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001330 if (unlikely(pmd_bad(*pmd)))
1331 goto no_page_table;
1332
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001333 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334
1335 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001336 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001337 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001338 if ((flags & FOLL_WRITE) && !pte_write(pte))
1339 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001340
Linus Torvalds6aab3412005-11-28 14:34:23 -08001341 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001342 if (unlikely(!page)) {
1343 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001344 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001345 goto bad_page;
1346 page = pte_page(pte);
1347 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001348
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001349 if (flags & FOLL_GET)
1350 get_page(page);
1351 if (flags & FOLL_TOUCH) {
1352 if ((flags & FOLL_WRITE) &&
1353 !pte_dirty(pte) && !PageDirty(page))
1354 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001355 /*
1356 * pte_mkyoung() would be more correct here, but atomic care
1357 * is needed to avoid losing the dirty bit: it is easier to use
1358 * mark_page_accessed().
1359 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001360 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001362 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001363 /*
1364 * The preliminary mapping check is mainly to avoid the
1365 * pointless overhead of lock_page on the ZERO_PAGE
1366 * which might bounce very badly if there is contention.
1367 *
1368 * If the page is already locked, we don't need to
1369 * handle it now - vmscan will handle it later if and
1370 * when it attempts to reclaim the page.
1371 */
1372 if (page->mapping && trylock_page(page)) {
1373 lru_add_drain(); /* push cached pages to LRU */
1374 /*
1375 * Because we lock page here and migration is
1376 * blocked by the pte's page reference, we need
1377 * only check for file-cache page truncation.
1378 */
1379 if (page->mapping)
1380 mlock_vma_page(page);
1381 unlock_page(page);
1382 }
1383 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001384unlock:
1385 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001387 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001389bad_page:
1390 pte_unmap_unlock(ptep, ptl);
1391 return ERR_PTR(-EFAULT);
1392
1393no_page:
1394 pte_unmap_unlock(ptep, ptl);
1395 if (!pte_none(pte))
1396 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001397
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001398no_page_table:
1399 /*
1400 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001401 * has touched so far, we don't want to allocate unnecessary pages or
1402 * page tables. Return error instead of NULL to skip handle_mm_fault,
1403 * then get_dump_page() will return NULL to leave a hole in the dump.
1404 * But we can only make this optimization where a hole would surely
1405 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001406 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001407 if ((flags & FOLL_DUMP) &&
1408 (!vma->vm_ops || !vma->vm_ops->fault))
1409 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001410 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411}
1412
Linus Torvalds95042f92011-04-12 14:15:51 -07001413static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1414{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001415 return stack_guard_page_start(vma, addr) ||
1416 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001417}
1418
Huang Ying0014bd92011-01-30 11:15:47 +08001419/**
1420 * __get_user_pages() - pin user pages in memory
1421 * @tsk: task_struct of target task
1422 * @mm: mm_struct of target mm
1423 * @start: starting user address
1424 * @nr_pages: number of pages from start to pin
1425 * @gup_flags: flags modifying pin behaviour
1426 * @pages: array that receives pointers to the pages pinned.
1427 * Should be at least nr_pages long. Or NULL, if caller
1428 * only intends to ensure the pages are faulted in.
1429 * @vmas: array of pointers to vmas corresponding to each page.
1430 * Or NULL if the caller does not require them.
1431 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1432 *
1433 * Returns number of pages pinned. This may be fewer than the number
1434 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1435 * were pinned, returns -errno. Each page returned must be released
1436 * with a put_page() call when it is finished with. vmas will only
1437 * remain valid while mmap_sem is held.
1438 *
1439 * Must be called with mmap_sem held for read or write.
1440 *
1441 * __get_user_pages walks a process's page tables and takes a reference to
1442 * each struct page that each user address corresponds to at a given
1443 * instant. That is, it takes the page that would be accessed if a user
1444 * thread accesses the given user virtual address at that instant.
1445 *
1446 * This does not guarantee that the page exists in the user mappings when
1447 * __get_user_pages returns, and there may even be a completely different
1448 * page there in some cases (eg. if mmapped pagecache has been invalidated
1449 * and subsequently re faulted). However it does guarantee that the page
1450 * won't be freed completely. And mostly callers simply care that the page
1451 * contains data that was valid *at some point in time*. Typically, an IO
1452 * or similar operation cannot guarantee anything stronger anyway because
1453 * locks can't be held over the syscall boundary.
1454 *
1455 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1456 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1457 * appropriate) must be called after the page is finished with, and
1458 * before put_page is called.
1459 *
1460 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1461 * or mmap_sem contention, and if waiting is needed to pin all pages,
1462 * *@nonblocking will be set to 0.
1463 *
1464 * In most cases, get_user_pages or get_user_pages_fast should be used
1465 * instead of __get_user_pages. __get_user_pages should be used only if
1466 * you need some special @gup_flags.
1467 */
Nick Pigginb291f002008-10-18 20:26:44 -07001468int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001469 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001470 struct page **pages, struct vm_area_struct **vmas,
1471 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
1473 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001474 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
Peter Zijlstra9d737772009-06-25 11:58:55 +02001476 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001477 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001478
1479 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1480
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 /*
1482 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001483 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001485 vm_flags = (gup_flags & FOLL_WRITE) ?
1486 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1487 vm_flags &= (gup_flags & FOLL_FORCE) ?
1488 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 i = 0;
1490
1491 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001495 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 pgd_t *pgd;
1498 pud_t *pud;
1499 pmd_t *pmd;
1500 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001501
1502 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001503 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 return i ? : -EFAULT;
1505 if (pg > TASK_SIZE)
1506 pgd = pgd_offset_k(pg);
1507 else
1508 pgd = pgd_offset_gate(mm, pg);
1509 BUG_ON(pgd_none(*pgd));
1510 pud = pud_offset(pgd, pg);
1511 BUG_ON(pud_none(*pud));
1512 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001513 if (pmd_none(*pmd))
1514 return i ? : -EFAULT;
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001515 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001517 if (pte_none(*pte)) {
1518 pte_unmap(pte);
1519 return i ? : -EFAULT;
1520 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001521 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 if (pages) {
Hugh Dickinsde512572010-07-30 10:58:26 -07001523 struct page *page;
1524
Linus Torvalds95042f92011-04-12 14:15:51 -07001525 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde512572010-07-30 10:58:26 -07001526 if (!page) {
1527 if (!(gup_flags & FOLL_DUMP) &&
1528 is_zero_pfn(pte_pfn(*pte)))
1529 page = pte_page(*pte);
1530 else {
1531 pte_unmap(pte);
1532 return i ? : -EFAULT;
1533 }
1534 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001535 pages[i] = page;
Hugh Dickinsde512572010-07-30 10:58:26 -07001536 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 }
1538 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001539 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 }
1541
Nick Pigginb291f002008-10-18 20:26:44 -07001542 if (!vma ||
1543 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001544 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 return i ? : -EFAULT;
1546
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001547 if (is_vm_hugetlb_page(vma)) {
1548 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001549 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001550 continue;
1551 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001552
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001554 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001555 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
Ethan Solomita462e00c2007-07-15 23:38:16 -07001557 /*
Ying Han47792802009-01-06 14:40:18 -08001558 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001559 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001560 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001561 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001562 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001563
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001564 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001565 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001566 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001567 unsigned int fault_flags = 0;
1568
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001569 /* For mlock, just skip the stack guard page. */
1570 if (foll_flags & FOLL_MLOCK) {
1571 if (stack_guard_page(vma, start))
1572 goto next_page;
1573 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001574 if (foll_flags & FOLL_WRITE)
1575 fault_flags |= FAULT_FLAG_WRITE;
1576 if (nonblocking)
1577 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001578 if (foll_flags & FOLL_NOWAIT)
1579 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001580
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001581 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001582 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001583
Nick Piggin83c54072007-07-19 01:47:05 -07001584 if (ret & VM_FAULT_ERROR) {
1585 if (ret & VM_FAULT_OOM)
1586 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001587 if (ret & (VM_FAULT_HWPOISON |
1588 VM_FAULT_HWPOISON_LARGE)) {
1589 if (i)
1590 return i;
1591 else if (gup_flags & FOLL_HWPOISON)
1592 return -EHWPOISON;
1593 else
1594 return -EFAULT;
1595 }
1596 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001597 return i ? i : -EFAULT;
1598 BUG();
1599 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001600
1601 if (tsk) {
1602 if (ret & VM_FAULT_MAJOR)
1603 tsk->maj_flt++;
1604 else
1605 tsk->min_flt++;
1606 }
Nick Piggin83c54072007-07-19 01:47:05 -07001607
Michel Lespinasse53a77062011-01-13 15:46:14 -08001608 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001609 if (nonblocking)
1610 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001611 return i;
1612 }
1613
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001614 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001615 * The VM_FAULT_WRITE bit tells us that
1616 * do_wp_page has broken COW when necessary,
1617 * even if maybe_mkwrite decided not to set
1618 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001619 * page lookups as if they were reads. But only
1620 * do so when looping for pte_write is futile:
1621 * in some cases userspace may also be wanting
1622 * to write to the gotten user page, which a
1623 * read fault here might prevent (a readonly
1624 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001625 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001626 if ((ret & VM_FAULT_WRITE) &&
1627 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001628 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001629
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001630 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001632 if (IS_ERR(page))
1633 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001635 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001636
Russell Kinga6f36be2006-12-30 22:24:19 +00001637 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001638 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001640next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 if (vmas)
1642 vmas[i] = vma;
1643 i++;
1644 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001645 nr_pages--;
1646 } while (nr_pages && start < vma->vm_end);
1647 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 return i;
1649}
Huang Ying0014bd92011-01-30 11:15:47 +08001650EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001651
Nick Piggind2bf6be2009-06-16 15:31:39 -07001652/**
1653 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001654 * @tsk: the task_struct to use for page fault accounting, or
1655 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001656 * @mm: mm_struct of target mm
1657 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001658 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001659 * @write: whether pages will be written to by the caller
1660 * @force: whether to force write access even if user mapping is
1661 * readonly. This will result in the page being COWed even
1662 * in MAP_SHARED mappings. You do not want this.
1663 * @pages: array that receives pointers to the pages pinned.
1664 * Should be at least nr_pages long. Or NULL, if caller
1665 * only intends to ensure the pages are faulted in.
1666 * @vmas: array of pointers to vmas corresponding to each page.
1667 * Or NULL if the caller does not require them.
1668 *
1669 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001670 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001671 * were pinned, returns -errno. Each page returned must be released
1672 * with a put_page() call when it is finished with. vmas will only
1673 * remain valid while mmap_sem is held.
1674 *
1675 * Must be called with mmap_sem held for read or write.
1676 *
1677 * get_user_pages walks a process's page tables and takes a reference to
1678 * each struct page that each user address corresponds to at a given
1679 * instant. That is, it takes the page that would be accessed if a user
1680 * thread accesses the given user virtual address at that instant.
1681 *
1682 * This does not guarantee that the page exists in the user mappings when
1683 * get_user_pages returns, and there may even be a completely different
1684 * page there in some cases (eg. if mmapped pagecache has been invalidated
1685 * and subsequently re faulted). However it does guarantee that the page
1686 * won't be freed completely. And mostly callers simply care that the page
1687 * contains data that was valid *at some point in time*. Typically, an IO
1688 * or similar operation cannot guarantee anything stronger anyway because
1689 * locks can't be held over the syscall boundary.
1690 *
1691 * If write=0, the page must not be written to. If the page is written to,
1692 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1693 * after the page is finished with, and before put_page is called.
1694 *
1695 * get_user_pages is typically used for fewer-copy IO operations, to get a
1696 * handle on the memory by some means other than accesses via the user virtual
1697 * addresses. The pages may be submitted for DMA to devices or accessed via
1698 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1699 * use the correct cache flushing APIs.
1700 *
1701 * See also get_user_pages_fast, for performance critical applications.
1702 */
Nick Pigginb291f002008-10-18 20:26:44 -07001703int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001704 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001705 struct page **pages, struct vm_area_struct **vmas)
1706{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001707 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001708
Hugh Dickins58fa8792009-09-21 17:03:31 -07001709 if (pages)
1710 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001711 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001712 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001713 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001714 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001715
Michel Lespinasse53a77062011-01-13 15:46:14 -08001716 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1717 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001718}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719EXPORT_SYMBOL(get_user_pages);
1720
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001721/**
1722 * get_dump_page() - pin user page in memory while writing it to core dump
1723 * @addr: user address
1724 *
1725 * Returns struct page pointer of user page pinned for dump,
1726 * to be freed afterwards by page_cache_release() or put_page().
1727 *
1728 * Returns NULL on any kind of failure - a hole must then be inserted into
1729 * the corefile, to preserve alignment with its headers; and also returns
1730 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1731 * allowing a hole to be left in the corefile to save diskspace.
1732 *
1733 * Called without mmap_sem, but after all other threads have been killed.
1734 */
1735#ifdef CONFIG_ELF_CORE
1736struct page *get_dump_page(unsigned long addr)
1737{
1738 struct vm_area_struct *vma;
1739 struct page *page;
1740
1741 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001742 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1743 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001744 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001745 flush_cache_page(vma, addr, page_to_pfn(page));
1746 return page;
1747}
1748#endif /* CONFIG_ELF_CORE */
1749
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001750pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001751 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001752{
1753 pgd_t * pgd = pgd_offset(mm, addr);
1754 pud_t * pud = pud_alloc(mm, pgd, addr);
1755 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001756 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001757 if (pmd) {
1758 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001759 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001760 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001761 }
1762 return NULL;
1763}
1764
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001766 * This is the old fallback for page remapping.
1767 *
1768 * For historical reasons, it only allows reserved pages. Only
1769 * old drivers should use this, and they needed to mark their
1770 * pages reserved for the old functions anyway.
1771 */
Nick Piggin423bad62008-04-28 02:13:01 -07001772static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1773 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001774{
Nick Piggin423bad62008-04-28 02:13:01 -07001775 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001776 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001777 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001778 spinlock_t *ptl;
1779
Linus Torvalds238f58d2005-11-29 13:01:56 -08001780 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001781 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001782 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001783 retval = -ENOMEM;
1784 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001785 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001786 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001787 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001788 retval = -EBUSY;
1789 if (!pte_none(*pte))
1790 goto out_unlock;
1791
1792 /* Ok, finally just insert the thing.. */
1793 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001794 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001795 page_add_file_rmap(page);
1796 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1797
1798 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001799 pte_unmap_unlock(pte, ptl);
1800 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001801out_unlock:
1802 pte_unmap_unlock(pte, ptl);
1803out:
1804 return retval;
1805}
1806
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001807/**
1808 * vm_insert_page - insert single page into user vma
1809 * @vma: user vma to map to
1810 * @addr: target user address of this page
1811 * @page: source kernel page
1812 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001813 * This allows drivers to insert individual pages they've allocated
1814 * into a user vma.
1815 *
1816 * The page has to be a nice clean _individual_ kernel allocation.
1817 * If you allocate a compound page, you need to have marked it as
1818 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001819 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001820 *
1821 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1822 * took an arbitrary page protection parameter. This doesn't allow
1823 * that. Your vma protection will have to be set up correctly, which
1824 * means that if you want a shared writable mapping, you'd better
1825 * ask for a shared writable mapping!
1826 *
1827 * The page does not need to be reserved.
1828 */
Nick Piggin423bad62008-04-28 02:13:01 -07001829int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1830 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001831{
1832 if (addr < vma->vm_start || addr >= vma->vm_end)
1833 return -EFAULT;
1834 if (!page_count(page))
1835 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001836 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07001837 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001838}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001839EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001840
Nick Piggin423bad62008-04-28 02:13:01 -07001841static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1842 unsigned long pfn, pgprot_t prot)
1843{
1844 struct mm_struct *mm = vma->vm_mm;
1845 int retval;
1846 pte_t *pte, entry;
1847 spinlock_t *ptl;
1848
1849 retval = -ENOMEM;
1850 pte = get_locked_pte(mm, addr, &ptl);
1851 if (!pte)
1852 goto out;
1853 retval = -EBUSY;
1854 if (!pte_none(*pte))
1855 goto out_unlock;
1856
1857 /* Ok, finally just insert the thing.. */
1858 entry = pte_mkspecial(pfn_pte(pfn, prot));
1859 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001860 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001861
1862 retval = 0;
1863out_unlock:
1864 pte_unmap_unlock(pte, ptl);
1865out:
1866 return retval;
1867}
1868
Nick Piggine0dc0d82007-02-12 00:51:36 -08001869/**
1870 * vm_insert_pfn - insert single pfn into user vma
1871 * @vma: user vma to map to
1872 * @addr: target user address of this page
1873 * @pfn: source kernel pfn
1874 *
1875 * Similar to vm_inert_page, this allows drivers to insert individual pages
1876 * they've allocated into a user vma. Same comments apply.
1877 *
1878 * This function should only be called from a vm_ops->fault handler, and
1879 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001880 *
1881 * vma cannot be a COW mapping.
1882 *
1883 * As this is called only for pages that do not currently exist, we
1884 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001885 */
1886int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001887 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001888{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001889 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001890 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001891 /*
1892 * Technically, architectures with pte_special can avoid all these
1893 * restrictions (same for remap_pfn_range). However we would like
1894 * consistency in testing and feature parity among all, so we should
1895 * try to keep these invariants in place for everybody.
1896 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001897 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1898 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1899 (VM_PFNMAP|VM_MIXEDMAP));
1900 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1901 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001902
Nick Piggin423bad62008-04-28 02:13:01 -07001903 if (addr < vma->vm_start || addr >= vma->vm_end)
1904 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001905 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001906 return -EINVAL;
1907
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001908 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001909
1910 if (ret)
1911 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1912
1913 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001914}
1915EXPORT_SYMBOL(vm_insert_pfn);
1916
Nick Piggin423bad62008-04-28 02:13:01 -07001917int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918 unsigned long pfn)
1919{
1920 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1921
1922 if (addr < vma->vm_start || addr >= vma->vm_end)
1923 return -EFAULT;
1924
1925 /*
1926 * If we don't have pte special, then we have to use the pfn_valid()
1927 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1928 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001929 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1930 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001931 */
1932 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1933 struct page *page;
1934
1935 page = pfn_to_page(pfn);
1936 return insert_page(vma, addr, page, vma->vm_page_prot);
1937 }
1938 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1939}
1940EXPORT_SYMBOL(vm_insert_mixed);
1941
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001942/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 * maps a range of physical memory into the requested pages. the old
1944 * mappings are removed. any references to nonexistent pages results
1945 * in null mappings (currently treated as "copy-on-access")
1946 */
1947static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1948 unsigned long addr, unsigned long end,
1949 unsigned long pfn, pgprot_t prot)
1950{
1951 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001952 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
Hugh Dickinsc74df322005-10-29 18:16:23 -07001954 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 if (!pte)
1956 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001957 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 do {
1959 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001960 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 pfn++;
1962 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001963 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001964 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 return 0;
1966}
1967
1968static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1969 unsigned long addr, unsigned long end,
1970 unsigned long pfn, pgprot_t prot)
1971{
1972 pmd_t *pmd;
1973 unsigned long next;
1974
1975 pfn -= addr >> PAGE_SHIFT;
1976 pmd = pmd_alloc(mm, pud, addr);
1977 if (!pmd)
1978 return -ENOMEM;
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001979 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 do {
1981 next = pmd_addr_end(addr, end);
1982 if (remap_pte_range(mm, pmd, addr, next,
1983 pfn + (addr >> PAGE_SHIFT), prot))
1984 return -ENOMEM;
1985 } while (pmd++, addr = next, addr != end);
1986 return 0;
1987}
1988
1989static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1990 unsigned long addr, unsigned long end,
1991 unsigned long pfn, pgprot_t prot)
1992{
1993 pud_t *pud;
1994 unsigned long next;
1995
1996 pfn -= addr >> PAGE_SHIFT;
1997 pud = pud_alloc(mm, pgd, addr);
1998 if (!pud)
1999 return -ENOMEM;
2000 do {
2001 next = pud_addr_end(addr, end);
2002 if (remap_pmd_range(mm, pud, addr, next,
2003 pfn + (addr >> PAGE_SHIFT), prot))
2004 return -ENOMEM;
2005 } while (pud++, addr = next, addr != end);
2006 return 0;
2007}
2008
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002009/**
2010 * remap_pfn_range - remap kernel memory to userspace
2011 * @vma: user vma to map to
2012 * @addr: target user address to start at
2013 * @pfn: physical address of kernel memory
2014 * @size: size of map area
2015 * @prot: page protection flags for this mapping
2016 *
2017 * Note: this is only safe if the mm semaphore is held when called.
2018 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2020 unsigned long pfn, unsigned long size, pgprot_t prot)
2021{
2022 pgd_t *pgd;
2023 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002024 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 struct mm_struct *mm = vma->vm_mm;
2026 int err;
2027
2028 /*
2029 * Physically remapped pages are special. Tell the
2030 * rest of the world about it:
2031 * VM_IO tells people not to look at these pages
2032 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08002033 * VM_RESERVED is specified all over the place, because
2034 * in 2.4 it kept swapout's vma scan off this vma; but
2035 * in 2.6 the LRU scan won't even find its pages, so this
2036 * flag means no more than count its pages in reserved_vm,
2037 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08002038 * VM_PFNMAP tells the core MM that the base pages are just
2039 * raw PFN mappings, and do not have a "struct page" associated
2040 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002041 *
2042 * There's a horrible special case to handle copy-on-write
2043 * behaviour that some programs depend on. We mark the "original"
2044 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002046 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002047 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002048 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002049 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002050 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002051
Linus Torvalds6aab3412005-11-28 14:34:23 -08002052 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002054 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002055 if (err) {
2056 /*
2057 * To indicate that track_pfn related cleanup is not
2058 * needed from higher level routine calling unmap_vmas
2059 */
2060 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002061 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002062 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002063 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002064
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 BUG_ON(addr >= end);
2066 pfn -= addr >> PAGE_SHIFT;
2067 pgd = pgd_offset(mm, addr);
2068 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 do {
2070 next = pgd_addr_end(addr, end);
2071 err = remap_pud_range(mm, pgd, addr, next,
2072 pfn + (addr >> PAGE_SHIFT), prot);
2073 if (err)
2074 break;
2075 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002076
2077 if (err)
2078 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2079
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 return err;
2081}
2082EXPORT_SYMBOL(remap_pfn_range);
2083
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002084static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2085 unsigned long addr, unsigned long end,
2086 pte_fn_t fn, void *data)
2087{
2088 pte_t *pte;
2089 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002090 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002091 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002092
2093 pte = (mm == &init_mm) ?
2094 pte_alloc_kernel(pmd, addr) :
2095 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2096 if (!pte)
2097 return -ENOMEM;
2098
2099 BUG_ON(pmd_huge(*pmd));
2100
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002101 arch_enter_lazy_mmu_mode();
2102
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002103 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002104
2105 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002106 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002107 if (err)
2108 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002109 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002110
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002111 arch_leave_lazy_mmu_mode();
2112
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002113 if (mm != &init_mm)
2114 pte_unmap_unlock(pte-1, ptl);
2115 return err;
2116}
2117
2118static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2119 unsigned long addr, unsigned long end,
2120 pte_fn_t fn, void *data)
2121{
2122 pmd_t *pmd;
2123 unsigned long next;
2124 int err;
2125
Andi Kleenceb86872008-07-23 21:27:50 -07002126 BUG_ON(pud_huge(*pud));
2127
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002128 pmd = pmd_alloc(mm, pud, addr);
2129 if (!pmd)
2130 return -ENOMEM;
2131 do {
2132 next = pmd_addr_end(addr, end);
2133 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2134 if (err)
2135 break;
2136 } while (pmd++, addr = next, addr != end);
2137 return err;
2138}
2139
2140static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2141 unsigned long addr, unsigned long end,
2142 pte_fn_t fn, void *data)
2143{
2144 pud_t *pud;
2145 unsigned long next;
2146 int err;
2147
2148 pud = pud_alloc(mm, pgd, addr);
2149 if (!pud)
2150 return -ENOMEM;
2151 do {
2152 next = pud_addr_end(addr, end);
2153 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2154 if (err)
2155 break;
2156 } while (pud++, addr = next, addr != end);
2157 return err;
2158}
2159
2160/*
2161 * Scan a region of virtual memory, filling in page tables as necessary
2162 * and calling a provided function on each leaf page table.
2163 */
2164int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2165 unsigned long size, pte_fn_t fn, void *data)
2166{
2167 pgd_t *pgd;
2168 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002169 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002170 int err;
2171
2172 BUG_ON(addr >= end);
2173 pgd = pgd_offset(mm, addr);
2174 do {
2175 next = pgd_addr_end(addr, end);
2176 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2177 if (err)
2178 break;
2179 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002180
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002181 return err;
2182}
2183EXPORT_SYMBOL_GPL(apply_to_page_range);
2184
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002186 * handle_pte_fault chooses page fault handler according to an entry
2187 * which was read non-atomically. Before making any commitment, on
2188 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002189 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002190 * must check under lock before unmapping the pte and proceeding
2191 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002192 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002193 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002194static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002195 pte_t *page_table, pte_t orig_pte)
2196{
2197 int same = 1;
2198#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2199 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002200 spinlock_t *ptl = pte_lockptr(mm, pmd);
2201 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002202 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002203 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002204 }
2205#endif
2206 pte_unmap(page_table);
2207 return same;
2208}
2209
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002210static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
Linus Torvalds6aab3412005-11-28 14:34:23 -08002211{
2212 /*
2213 * If the source page was a PFN mapping, we don't have
2214 * a "struct page" for it. We do a best-effort copy by
2215 * just copying from the original user address. If that
2216 * fails, we just zero-fill it. Live with it.
2217 */
2218 if (unlikely(!src)) {
2219 void *kaddr = kmap_atomic(dst, KM_USER0);
Linus Torvalds5d2a2db2005-11-29 14:07:55 -08002220 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2221
2222 /*
2223 * This really shouldn't fail, because the page is there
2224 * in the page tables. But it might just be unreadable,
2225 * in which case we just give up and fill the result with
2226 * zeroes.
2227 */
2228 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002229 clear_page(kaddr);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002230 kunmap_atomic(kaddr, KM_USER0);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002231 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002232 } else
2233 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002234}
2235
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 * This routine handles present pages, when users try to write
2238 * to a shared page. It is done by copying the page to a new address
2239 * and decrementing the shared-page counter for the old page.
2240 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 * Note that this routine assumes that the protection checks have been
2242 * done by the caller (the low-level page fault routine in most cases).
2243 * Thus we can safely just mark it writable once we've done any necessary
2244 * COW.
2245 *
2246 * We also mark the page dirty at this point even though the page will
2247 * change only once the write actually happens. This avoids a few races,
2248 * and potentially makes it more efficient.
2249 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002250 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2251 * but allow concurrent faults), with pte both mapped and locked.
2252 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002254static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2255 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002256 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002257 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002259 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002261 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002262 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002263 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264
Linus Torvalds6aab3412005-11-28 14:34:23 -08002265 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002266 if (!old_page) {
2267 /*
2268 * VM_MIXEDMAP !pfn_valid() case
2269 *
2270 * We should not cow pages in a shared writeable mapping.
2271 * Just mark the pages writable as we can't do any dirty
2272 * accounting on raw pfn maps.
2273 */
2274 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2275 (VM_WRITE|VM_SHARED))
2276 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002277 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002278 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002280 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002281 * Take out anonymous pages first, anonymous shared vmas are
2282 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002283 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002284 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002285 if (!trylock_page(old_page)) {
2286 page_cache_get(old_page);
2287 pte_unmap_unlock(page_table, ptl);
2288 lock_page(old_page);
2289 page_table = pte_offset_map_lock(mm, pmd, address,
2290 &ptl);
2291 if (!pte_same(*page_table, orig_pte)) {
2292 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002293 goto unlock;
2294 }
2295 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002296 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002297 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002298 /*
2299 * The page is all ours. Move it to our anon_vma so
2300 * the rmap code will not search our parent or siblings.
2301 * Protected against the rmap code by the page lock.
2302 */
2303 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002304 unlock_page(old_page);
2305 goto reuse;
2306 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002307 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002308 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002309 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002310 /*
2311 * Only catch write-faults on shared writable pages,
2312 * read-only shared pages can get COWed by
2313 * get_user_pages(.write=1, .force=1).
2314 */
David Howells9637a5e2006-06-23 02:03:43 -07002315 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002316 struct vm_fault vmf;
2317 int tmp;
2318
2319 vmf.virtual_address = (void __user *)(address &
2320 PAGE_MASK);
2321 vmf.pgoff = old_page->index;
2322 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2323 vmf.page = old_page;
2324
David Howells9637a5e2006-06-23 02:03:43 -07002325 /*
2326 * Notify the address space that the page is about to
2327 * become writable so that it can prohibit this or wait
2328 * for the page to get into an appropriate state.
2329 *
2330 * We do this without the lock held, so that it can
2331 * sleep if it needs to.
2332 */
2333 page_cache_get(old_page);
2334 pte_unmap_unlock(page_table, ptl);
2335
Nick Pigginc2ec1752009-03-31 15:23:21 -07002336 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2337 if (unlikely(tmp &
2338 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2339 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002340 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002341 }
Nick Pigginb827e492009-04-30 15:08:16 -07002342 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2343 lock_page(old_page);
2344 if (!old_page->mapping) {
2345 ret = 0; /* retry the fault */
2346 unlock_page(old_page);
2347 goto unwritable_page;
2348 }
2349 } else
2350 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002351
David Howells9637a5e2006-06-23 02:03:43 -07002352 /*
2353 * Since we dropped the lock we need to revalidate
2354 * the PTE as someone else may have changed it. If
2355 * they did, we just return, as we can count on the
2356 * MMU to tell us if they didn't also make it writable.
2357 */
2358 page_table = pte_offset_map_lock(mm, pmd, address,
2359 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002360 if (!pte_same(*page_table, orig_pte)) {
2361 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002362 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002363 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002364
2365 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002367 dirty_page = old_page;
2368 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002369
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002370reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002371 flush_cache_page(vma, address, pte_pfn(orig_pte));
2372 entry = pte_mkyoung(orig_pte);
2373 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002374 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002375 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002376 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002377 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002378
2379 if (!dirty_page)
2380 return ret;
2381
2382 /*
2383 * Yes, Virginia, this is actually required to prevent a race
2384 * with clear_page_dirty_for_io() from clearing the page dirty
2385 * bit after it clear all dirty ptes, but before a racing
2386 * do_wp_page installs a dirty pte.
2387 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002388 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002389 */
2390 if (!page_mkwrite) {
2391 wait_on_page_locked(dirty_page);
2392 set_page_dirty_balance(dirty_page, page_mkwrite);
2393 }
2394 put_page(dirty_page);
2395 if (page_mkwrite) {
2396 struct address_space *mapping = dirty_page->mapping;
2397
2398 set_page_dirty(dirty_page);
2399 unlock_page(dirty_page);
2400 page_cache_release(dirty_page);
2401 if (mapping) {
2402 /*
2403 * Some device drivers do not set page.mapping
2404 * but still dirty their pages
2405 */
2406 balance_dirty_pages_ratelimited(mapping);
2407 }
2408 }
2409
2410 /* file_update_time outside page_lock */
2411 if (vma->vm_file)
2412 file_update_time(vma->vm_file);
2413
2414 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
2417 /*
2418 * Ok, we need to copy. Oh, well..
2419 */
Nick Pigginb5810032005-10-29 18:16:12 -07002420 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002421gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002422 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423
2424 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002425 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002426
Hugh Dickins62eede62009-09-21 17:03:34 -07002427 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002428 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2429 if (!new_page)
2430 goto oom;
2431 } else {
2432 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2433 if (!new_page)
2434 goto oom;
2435 cow_user_page(new_page, old_page, address, vma);
2436 }
2437 __SetPageUptodate(new_page);
2438
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002439 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002440 goto oom_free_new;
2441
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 /*
2443 * Re-check the pte - we dropped the lock
2444 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002445 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002446 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002447 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002448 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002449 dec_mm_counter_fast(mm, MM_FILEPAGES);
2450 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002451 }
2452 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002453 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002454 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002455 entry = mk_pte(new_page, vma->vm_page_prot);
2456 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002457 /*
2458 * Clear the pte entry and flush it first, before updating the
2459 * pte with the new entry. This will avoid a race condition
2460 * seen in the presence of one thread doing SMC and another
2461 * thread doing COW.
2462 */
Izik Eidus828502d2009-09-21 17:01:51 -07002463 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002464 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002465 /*
2466 * We call the notify macro here because, when using secondary
2467 * mmu page tables (such as kvm shadow page tables), we want the
2468 * new page to be mapped directly into the secondary page table.
2469 */
2470 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002471 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002472 if (old_page) {
2473 /*
2474 * Only after switching the pte to the new page may
2475 * we remove the mapcount here. Otherwise another
2476 * process may come and find the rmap count decremented
2477 * before the pte is switched to the new page, and
2478 * "reuse" the old page writing into it while our pte
2479 * here still points into it and can be read by other
2480 * threads.
2481 *
2482 * The critical issue is to order this
2483 * page_remove_rmap with the ptp_clear_flush above.
2484 * Those stores are ordered by (if nothing else,)
2485 * the barrier present in the atomic_add_negative
2486 * in page_remove_rmap.
2487 *
2488 * Then the TLB flush in ptep_clear_flush ensures that
2489 * no process can access the old page before the
2490 * decremented mapcount is visible. And the old page
2491 * cannot be reused until after the decremented
2492 * mapcount is visible. So transitively, TLBs to
2493 * old page will be flushed before it can be reused.
2494 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002495 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002496 }
2497
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 /* Free the old page.. */
2499 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002500 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002501 } else
2502 mem_cgroup_uncharge_page(new_page);
2503
Hugh Dickins920fc352005-11-21 21:32:17 -08002504 if (new_page)
2505 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002506unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002507 pte_unmap_unlock(page_table, ptl);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002508 if (old_page) {
2509 /*
2510 * Don't let another task, with possibly unlocked vma,
2511 * keep the mlocked page.
2512 */
2513 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2514 lock_page(old_page); /* LRU manipulation */
2515 munlock_vma_page(old_page);
2516 unlock_page(old_page);
2517 }
2518 page_cache_release(old_page);
2519 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002520 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002521oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002522 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002523oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002524 if (old_page) {
2525 if (page_mkwrite) {
2526 unlock_page(old_page);
2527 page_cache_release(old_page);
2528 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002529 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002530 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002532
2533unwritable_page:
2534 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002535 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536}
2537
2538/*
2539 * Helper functions for unmap_mapping_range().
2540 *
2541 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2542 *
2543 * We have to restart searching the prio_tree whenever we drop the lock,
2544 * since the iterator is only valid while the lock is held, and anyway
2545 * a later vma might be split and reinserted earlier while lock dropped.
2546 *
2547 * The list of nonlinear vmas could be handled more efficiently, using
2548 * a placeholder, but handle it in the same way until a need is shown.
2549 * It is important to search the prio_tree before nonlinear list: a vma
2550 * may become nonlinear and be shifted from prio_tree to nonlinear list
2551 * while the lock is dropped; but never shifted from list to prio_tree.
2552 *
2553 * In order to make forward progress despite restarting the search,
2554 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2555 * quickly skip it next time around. Since the prio_tree search only
2556 * shows us those vmas affected by unmapping the range in question, we
2557 * can't efficiently keep all vmas in step with mapping->truncate_count:
2558 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2559 * mapping->truncate_count and vma->vm_truncate_count are protected by
2560 * i_mmap_lock.
2561 *
2562 * In order to make forward progress despite repeatedly restarting some
Hugh Dickinsee39b372005-04-19 13:29:15 -07002563 * large vma, note the restart_addr from unmap_vmas when it breaks out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 * and restart from that address when we reach that vma again. It might
2565 * have been split or merged, shrunk or extended, but never shifted: so
2566 * restart_addr remains valid so long as it remains in the vma's range.
2567 * unmap_mapping_range forces truncate_count to leap over page-aligned
2568 * values so we can save vma's restart_addr in its truncate_count field.
2569 */
2570#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2571
2572static void reset_vma_truncate_counts(struct address_space *mapping)
2573{
2574 struct vm_area_struct *vma;
2575 struct prio_tree_iter iter;
2576
2577 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2578 vma->vm_truncate_count = 0;
2579 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2580 vma->vm_truncate_count = 0;
2581}
2582
2583static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2584 unsigned long start_addr, unsigned long end_addr,
2585 struct zap_details *details)
2586{
2587 unsigned long restart_addr;
2588 int need_break;
2589
Nick Piggind00806b2007-07-19 01:46:57 -07002590 /*
2591 * files that support invalidating or truncating portions of the
Nick Piggind0217ac2007-07-19 01:47:03 -07002592 * file from under mmaped areas must have their ->fault function
Nick Piggin83c54072007-07-19 01:47:05 -07002593 * return a locked page (and set VM_FAULT_LOCKED in the return).
2594 * This provides synchronisation against concurrent unmapping here.
Nick Piggind00806b2007-07-19 01:46:57 -07002595 */
Nick Piggind00806b2007-07-19 01:46:57 -07002596
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597again:
2598 restart_addr = vma->vm_truncate_count;
2599 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2600 start_addr = restart_addr;
2601 if (start_addr >= end_addr) {
2602 /* Top of vma has been split off since last time */
2603 vma->vm_truncate_count = details->truncate_count;
2604 return 0;
2605 }
2606 }
2607
Hugh Dickinsee39b372005-04-19 13:29:15 -07002608 restart_addr = zap_page_range(vma, start_addr,
2609 end_addr - start_addr, details);
Nick Piggin95c354f2008-01-30 13:31:20 +01002610 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611
Hugh Dickinsee39b372005-04-19 13:29:15 -07002612 if (restart_addr >= end_addr) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 /* We have now completed this vma: mark it so */
2614 vma->vm_truncate_count = details->truncate_count;
2615 if (!need_break)
2616 return 0;
2617 } else {
2618 /* Note restart_addr in vma's truncate_count field */
Hugh Dickinsee39b372005-04-19 13:29:15 -07002619 vma->vm_truncate_count = restart_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 if (!need_break)
2621 goto again;
2622 }
2623
2624 spin_unlock(details->i_mmap_lock);
2625 cond_resched();
2626 spin_lock(details->i_mmap_lock);
2627 return -EINTR;
2628}
2629
2630static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2631 struct zap_details *details)
2632{
2633 struct vm_area_struct *vma;
2634 struct prio_tree_iter iter;
2635 pgoff_t vba, vea, zba, zea;
2636
2637restart:
2638 vma_prio_tree_foreach(vma, &iter, root,
2639 details->first_index, details->last_index) {
2640 /* Skip quickly over those we have already dealt with */
2641 if (vma->vm_truncate_count == details->truncate_count)
2642 continue;
2643
2644 vba = vma->vm_pgoff;
2645 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2646 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2647 zba = details->first_index;
2648 if (zba < vba)
2649 zba = vba;
2650 zea = details->last_index;
2651 if (zea > vea)
2652 zea = vea;
2653
2654 if (unmap_mapping_range_vma(vma,
2655 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2656 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2657 details) < 0)
2658 goto restart;
2659 }
2660}
2661
2662static inline void unmap_mapping_range_list(struct list_head *head,
2663 struct zap_details *details)
2664{
2665 struct vm_area_struct *vma;
2666
2667 /*
2668 * In nonlinear VMAs there is no correspondence between virtual address
2669 * offset and file offset. So we must perform an exhaustive search
2670 * across *all* the pages in each nonlinear VMA, not just the pages
2671 * whose virtual address lies outside the file truncation point.
2672 */
2673restart:
2674 list_for_each_entry(vma, head, shared.vm_set.list) {
2675 /* Skip quickly over those we have already dealt with */
2676 if (vma->vm_truncate_count == details->truncate_count)
2677 continue;
2678 details->nonlinear_vma = vma;
2679 if (unmap_mapping_range_vma(vma, vma->vm_start,
2680 vma->vm_end, details) < 0)
2681 goto restart;
2682 }
2683}
2684
2685/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002686 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
Martin Waitz3d410882005-06-23 22:05:21 -07002687 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 * @holebegin: byte in first page to unmap, relative to the start of
2689 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002690 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 * must keep the partial page. In contrast, we must get rid of
2692 * partial pages.
2693 * @holelen: size of prospective hole in bytes. This will be rounded
2694 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2695 * end of the file.
2696 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2697 * but 0 when invalidating pagecache, don't throw away private data.
2698 */
2699void unmap_mapping_range(struct address_space *mapping,
2700 loff_t const holebegin, loff_t const holelen, int even_cows)
2701{
2702 struct zap_details details;
2703 pgoff_t hba = holebegin >> PAGE_SHIFT;
2704 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2705
2706 /* Check for overflow. */
2707 if (sizeof(holelen) > sizeof(hlen)) {
2708 long long holeend =
2709 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2710 if (holeend & ~(long long)ULONG_MAX)
2711 hlen = ULONG_MAX - hba + 1;
2712 }
2713
2714 details.check_mapping = even_cows? NULL: mapping;
2715 details.nonlinear_vma = NULL;
2716 details.first_index = hba;
2717 details.last_index = hba + hlen - 1;
2718 if (details.last_index < details.first_index)
2719 details.last_index = ULONG_MAX;
2720 details.i_mmap_lock = &mapping->i_mmap_lock;
2721
Miklos Szeredi2aa15892011-02-23 13:49:47 +01002722 mutex_lock(&mapping->unmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 spin_lock(&mapping->i_mmap_lock);
2724
Nick Piggind00806b2007-07-19 01:46:57 -07002725 /* Protect against endless unmapping loops */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 mapping->truncate_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2728 if (mapping->truncate_count == 0)
2729 reset_vma_truncate_counts(mapping);
2730 mapping->truncate_count++;
2731 }
2732 details.truncate_count = mapping->truncate_count;
2733
2734 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2735 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2736 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2737 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2738 spin_unlock(&mapping->i_mmap_lock);
Miklos Szeredi2aa15892011-02-23 13:49:47 +01002739 mutex_unlock(&mapping->unmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740}
2741EXPORT_SYMBOL(unmap_mapping_range);
2742
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002743int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2744{
2745 struct address_space *mapping = inode->i_mapping;
2746
2747 /*
2748 * If the underlying filesystem is not going to provide
2749 * a way to truncate a range of blocks (punch a hole) -
2750 * we should return failure right now.
2751 */
Al Viroacfa4382008-12-04 10:06:33 -05002752 if (!inode->i_op->truncate_range)
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002753 return -ENOSYS;
2754
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002755 mutex_lock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002756 down_write(&inode->i_alloc_sem);
2757 unmap_mapping_range(mapping, offset, (end - offset), 1);
2758 truncate_inode_pages_range(mapping, offset, end);
Nick Piggind00806b2007-07-19 01:46:57 -07002759 unmap_mapping_range(mapping, offset, (end - offset), 1);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002760 inode->i_op->truncate_range(inode, offset, end);
2761 up_write(&inode->i_alloc_sem);
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002762 mutex_unlock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002763
2764 return 0;
2765}
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002766
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002768 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2769 * but allow concurrent faults), and pte mapped but not yet locked.
2770 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002772static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2773 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002774 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002776 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002777 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002778 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002780 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002781 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002782 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002783 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002785 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002786 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002787
2788 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002789 if (unlikely(non_swap_entry(entry))) {
2790 if (is_migration_entry(entry)) {
2791 migration_entry_wait(mm, pmd, address);
2792 } else if (is_hwpoison_entry(entry)) {
2793 ret = VM_FAULT_HWPOISON;
2794 } else {
2795 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002796 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002797 }
Christoph Lameter06972122006-06-23 02:03:35 -07002798 goto out;
2799 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002800 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 page = lookup_swap_cache(entry);
2802 if (!page) {
Hugh Dickinsa5c9b692009-06-23 12:36:58 -07002803 grab_swap_token(mm); /* Contend for token _before_ read-in */
Hugh Dickins02098fe2008-02-04 22:28:42 -08002804 page = swapin_readahead(entry,
2805 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 if (!page) {
2807 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002808 * Back out if somebody else faulted in this pte
2809 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002811 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 if (likely(pte_same(*page_table, orig_pte)))
2813 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002814 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002815 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 }
2817
2818 /* Had to read the page from swap area: Major fault */
2819 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002820 count_vm_event(PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002821 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002822 /*
2823 * hwpoisoned dirty swapcache pages are kept for killing
2824 * owner processes (which may be unknown at hwpoison time)
2825 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002826 ret = VM_FAULT_HWPOISON;
2827 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002828 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 }
2830
Michel Lespinassed065bd82010-10-26 14:21:57 -07002831 locked = lock_page_or_retry(page, mm, flags);
Balbir Singh20a10222007-11-14 17:00:33 -08002832 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002833 if (!locked) {
2834 ret |= VM_FAULT_RETRY;
2835 goto out_release;
2836 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002838 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002839 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2840 * release the swapcache from under us. The page pin, and pte_same
2841 * test below, are not enough to exclude that. Even if it is still
2842 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002843 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002844 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002845 goto out_page;
2846
2847 if (ksm_might_need_to_copy(page, vma, address)) {
2848 swapcache = page;
2849 page = ksm_does_need_to_copy(page, vma, address);
2850
2851 if (unlikely(!page)) {
2852 ret = VM_FAULT_OOM;
2853 page = swapcache;
2854 swapcache = NULL;
2855 goto out_page;
2856 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002857 }
2858
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002859 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002860 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002861 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002862 }
2863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002865 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002867 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002868 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002869 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002870
2871 if (unlikely(!PageUptodate(page))) {
2872 ret = VM_FAULT_SIGBUS;
2873 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 }
2875
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002876 /*
2877 * The page isn't present yet, go ahead with the fault.
2878 *
2879 * Be careful about the sequence of operations here.
2880 * To get its accounting right, reuse_swap_page() must be called
2881 * while the page is counted on swap but not yet in mapcount i.e.
2882 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2883 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002884 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2885 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2886 * in page->private. In this case, a record in swap_cgroup is silently
2887 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002888 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002890 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002891 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002893 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002895 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002896 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002897 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 flush_icache_page(vma, page);
2900 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002901 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002902 /* It's better to call commit-charge after rmap is established */
2903 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002905 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002906 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002907 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002908 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002909 if (swapcache) {
2910 /*
2911 * Hold the lock to avoid the swap entry to be reused
2912 * until we take the PT lock for the pte_same() check
2913 * (to avoid false positives from pte_same). For
2914 * further safety release the lock after the swap_free
2915 * so that the swap count won't change under a
2916 * parallel locked swapcache.
2917 */
2918 unlock_page(swapcache);
2919 page_cache_release(swapcache);
2920 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002921
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002922 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002923 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2924 if (ret & VM_FAULT_ERROR)
2925 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 goto out;
2927 }
2928
2929 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002930 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002931unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002932 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933out:
2934 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002935out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08002936 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002937 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002938out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002939 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002940out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002941 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002942 if (swapcache) {
2943 unlock_page(swapcache);
2944 page_cache_release(swapcache);
2945 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002946 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947}
2948
2949/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002950 * This is like a special single-page "expand_{down|up}wards()",
2951 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002952 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002953 */
2954static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2955{
2956 address &= PAGE_MASK;
2957 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002958 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002959
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002960 /*
2961 * Is there a mapping abutting this one below?
2962 *
2963 * That's only ok if it's the same stack mapping
2964 * that has gotten split..
2965 */
2966 if (prev && prev->vm_end == address)
2967 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2968
2969 expand_stack(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002970 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002971 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2972 struct vm_area_struct *next = vma->vm_next;
2973
2974 /* As VM_GROWSDOWN but s/below/above/ */
2975 if (next && next->vm_start == address + PAGE_SIZE)
2976 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2977
2978 expand_upwards(vma, address + PAGE_SIZE);
2979 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002980 return 0;
2981}
2982
2983/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002984 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2985 * but allow concurrent faults), and pte mapped but not yet locked.
2986 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002988static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2989 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002990 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002992 struct page *page;
2993 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995
Linus Torvalds11ac5522010-08-14 11:44:56 -07002996 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002997
Linus Torvalds11ac5522010-08-14 11:44:56 -07002998 /* Check if we need to add a guard page to the stack */
2999 if (check_stack_guard_page(vma, address) < 0)
3000 return VM_FAULT_SIGBUS;
3001
3002 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003003 if (!(flags & FAULT_FLAG_WRITE)) {
3004 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3005 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003006 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003007 if (!pte_none(*page_table))
3008 goto unlock;
3009 goto setpte;
3010 }
3011
Nick Piggin557ed1f2007-10-16 01:24:40 -07003012 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003013 if (unlikely(anon_vma_prepare(vma)))
3014 goto oom;
3015 page = alloc_zeroed_user_highpage_movable(vma, address);
3016 if (!page)
3017 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003018 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003020 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003021 goto oom_free_page;
3022
Nick Piggin557ed1f2007-10-16 01:24:40 -07003023 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003024 if (vma->vm_flags & VM_WRITE)
3025 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003026
Nick Piggin557ed1f2007-10-16 01:24:40 -07003027 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003028 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003029 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003030
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003031 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003032 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003033setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003034 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
3036 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003037 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003038unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003039 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003040 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003041release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003042 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003043 page_cache_release(page);
3044 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003045oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003046 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003047oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 return VM_FAULT_OOM;
3049}
3050
3051/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003052 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003054 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3055 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 *
3057 * As this is called only for pages that do not currently exist, we
3058 * do not need to flush old virtual caches or the TLB.
3059 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003060 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003061 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003062 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003064static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003065 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003066 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003068 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003069 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003070 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 int anon = 0;
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003073 int charged = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003074 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003075 struct vm_fault vmf;
3076 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003077 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003078
Nick Piggind0217ac2007-07-19 01:47:03 -07003079 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3080 vmf.pgoff = pgoff;
3081 vmf.flags = flags;
3082 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003084 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003085 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3086 VM_FAULT_RETRY)))
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003087 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088
Andi Kleena3b947e2009-09-16 11:50:08 +02003089 if (unlikely(PageHWPoison(vmf.page))) {
3090 if (ret & VM_FAULT_LOCKED)
3091 unlock_page(vmf.page);
3092 return VM_FAULT_HWPOISON;
3093 }
3094
Nick Piggind00806b2007-07-19 01:46:57 -07003095 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003096 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003097 * locked.
3098 */
Nick Piggin83c54072007-07-19 01:47:05 -07003099 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003100 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003101 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003102 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003103
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 /*
3105 * Should we do an early C-O-W break?
3106 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003107 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003108 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003109 if (!(vma->vm_flags & VM_SHARED)) {
Nick Piggin54cb8822007-07-19 01:46:59 -07003110 anon = 1;
Nick Piggind00806b2007-07-19 01:46:57 -07003111 if (unlikely(anon_vma_prepare(vma))) {
Nick Piggind0217ac2007-07-19 01:47:03 -07003112 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07003113 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07003114 }
Nick Piggin83c54072007-07-19 01:47:05 -07003115 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3116 vma, address);
Nick Piggind00806b2007-07-19 01:46:57 -07003117 if (!page) {
Nick Piggind0217ac2007-07-19 01:47:03 -07003118 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07003119 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07003120 }
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003121 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003122 ret = VM_FAULT_OOM;
3123 page_cache_release(page);
3124 goto out;
3125 }
3126 charged = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003127 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003128 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003129 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003130 /*
3131 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003132 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003133 * to become writable
3134 */
Mark Fasheh69676142007-07-19 01:47:00 -07003135 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003136 int tmp;
3137
Mark Fasheh69676142007-07-19 01:47:00 -07003138 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003139 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003140 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3141 if (unlikely(tmp &
3142 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3143 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003144 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003145 }
Nick Pigginb827e492009-04-30 15:08:16 -07003146 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3147 lock_page(page);
3148 if (!page->mapping) {
3149 ret = 0; /* retry the fault */
3150 unlock_page(page);
3151 goto unwritable_page;
3152 }
3153 } else
3154 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003155 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003156 }
3157 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003158
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 }
3160
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003161 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162
3163 /*
3164 * This silly early PAGE_DIRTY setting removes a race
3165 * due to the bad i386 page protection. But it's valid
3166 * for other architectures too.
3167 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003168 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 * an exclusive copy of the page, or this is a shared mapping,
3170 * so we can make it writable and dirty to avoid having to
3171 * handle that later.
3172 */
3173 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003174 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003175 flush_icache_page(vma, page);
3176 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003177 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003180 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003181 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003182 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003183 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003184 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003185 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003186 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003187 get_page(dirty_page);
3188 }
Hugh Dickins42946212005-10-29 18:16:05 -07003189 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003190 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003191
3192 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003193 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 } else {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003195 if (charged)
3196 mem_cgroup_uncharge_page(page);
Nick Piggind00806b2007-07-19 01:46:57 -07003197 if (anon)
3198 page_cache_release(page);
3199 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003200 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 }
3202
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003203 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003204
3205out:
Nick Pigginb827e492009-04-30 15:08:16 -07003206 if (dirty_page) {
3207 struct address_space *mapping = page->mapping;
3208
3209 if (set_page_dirty(dirty_page))
3210 page_mkwrite = 1;
3211 unlock_page(dirty_page);
3212 put_page(dirty_page);
3213 if (page_mkwrite && mapping) {
3214 /*
3215 * Some device drivers do not set page.mapping but still
3216 * dirty their pages
3217 */
3218 balance_dirty_pages_ratelimited(mapping);
3219 }
3220
3221 /* file_update_time outside page_lock */
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003222 if (vma->vm_file)
3223 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003224 } else {
3225 unlock_page(vmf.page);
3226 if (anon)
3227 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003228 }
Nick Piggind00806b2007-07-19 01:46:57 -07003229
Nick Piggin83c54072007-07-19 01:47:05 -07003230 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003231
3232unwritable_page:
3233 page_cache_release(page);
3234 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003235}
Nick Piggind00806b2007-07-19 01:46:57 -07003236
Nick Piggin54cb8822007-07-19 01:46:59 -07003237static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3238 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003239 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003240{
3241 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003242 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003243
Hugh Dickins16abfa02007-10-04 16:56:06 +01003244 pte_unmap(page_table);
3245 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003246}
3247
Jes Sorensenf4b81802006-09-27 01:50:10 -07003248/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249 * Fault of a previously existing named mapping. Repopulate the pte
3250 * from the encoded file_pte if possible. This enables swappable
3251 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003252 *
3253 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3254 * but allow concurrent faults), and pte mapped but not yet locked.
3255 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003257static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003258 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003259 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260{
Hugh Dickins65500d22005-10-29 18:15:59 -07003261 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003263 flags |= FAULT_FLAG_NONLINEAR;
3264
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003265 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003266 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267
Hugh Dickins2509ef22009-01-06 14:40:10 -08003268 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003269 /*
3270 * Page table corrupted: show pte and kill process.
3271 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003272 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003273 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003274 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003275
3276 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003277 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278}
3279
3280/*
3281 * These routines also need to handle stuff like marking pages dirty
3282 * and/or accessed for architectures that don't do it in hardware (most
3283 * RISC architectures). The early dirtying is also good on the i386.
3284 *
3285 * There is also a hook called "update_mmu_cache()" that architectures
3286 * with external mmu caches can use to update those (ie the Sparc or
3287 * PowerPC hashed page tables that act as extended TLBs).
3288 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003289 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3290 * but allow concurrent faults), and pte mapped but not yet locked.
3291 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003293int handle_pte_fault(struct mm_struct *mm,
3294 struct vm_area_struct *vma, unsigned long address,
3295 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296{
3297 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003298 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003300 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003302 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003303 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003304 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003305 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003306 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003307 }
3308 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003309 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003310 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003312 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003313 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003314 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003315 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 }
3317
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003318 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003319 spin_lock(ptl);
3320 if (unlikely(!pte_same(*pte, entry)))
3321 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003322 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003324 return do_wp_page(mm, vma, address,
3325 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 entry = pte_mkdirty(entry);
3327 }
3328 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003329 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003330 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003331 } else {
3332 /*
3333 * This is needed only for protection faults but the arch code
3334 * is not yet telling us if this is a protection fault or not.
3335 * This still avoids useless tlb flushes for .text page faults
3336 * with threads.
3337 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003338 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003339 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003340 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003341unlock:
3342 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003343 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344}
3345
3346/*
3347 * By the time we get here, we already hold the mm semaphore
3348 */
Nick Piggin83c54072007-07-19 01:47:05 -07003349int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003350 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351{
3352 pgd_t *pgd;
3353 pud_t *pud;
3354 pmd_t *pmd;
3355 pte_t *pte;
3356
3357 __set_current_state(TASK_RUNNING);
3358
Christoph Lameterf8891e52006-06-30 01:55:45 -07003359 count_vm_event(PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003361 /* do counter updates before entering really critical section. */
3362 check_sync_rss_stat(current);
3363
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003364 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003365 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 pud = pud_alloc(mm, pgd, address);
3369 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003370 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 pmd = pmd_alloc(mm, pud, address);
3372 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003373 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003374 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3375 if (!vma->vm_ops)
3376 return do_huge_pmd_anonymous_page(mm, vma, address,
3377 pmd, flags);
3378 } else {
3379 pmd_t orig_pmd = *pmd;
3380 barrier();
3381 if (pmd_trans_huge(orig_pmd)) {
3382 if (flags & FAULT_FLAG_WRITE &&
3383 !pmd_write(orig_pmd) &&
3384 !pmd_trans_splitting(orig_pmd))
3385 return do_huge_pmd_wp_page(mm, vma, address,
3386 pmd, orig_pmd);
3387 return 0;
3388 }
3389 }
3390
3391 /*
3392 * Use __pte_alloc instead of pte_alloc_map, because we can't
3393 * run pte_offset_map on the pmd, if an huge pmd could
3394 * materialize from under us from a different thread.
3395 */
Mel Gormancc036382011-04-27 15:26:56 -07003396 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003397 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003398 /* if an huge pmd materialized from under us just retry later */
3399 if (unlikely(pmd_trans_huge(*pmd)))
3400 return 0;
3401 /*
3402 * A regular pmd is established and it can't morph into a huge pmd
3403 * from under us anymore at this point because we hold the mmap_sem
3404 * read mode and khugepaged takes it in write mode. So now it's
3405 * safe to run pte_offset_map().
3406 */
3407 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003409 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410}
3411
3412#ifndef __PAGETABLE_PUD_FOLDED
3413/*
3414 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003415 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003417int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003419 pud_t *new = pud_alloc_one(mm, address);
3420 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003421 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Nick Piggin362a61a2008-05-14 06:37:36 +02003423 smp_wmb(); /* See comment in __pte_alloc */
3424
Hugh Dickins872fec12005-10-29 18:16:21 -07003425 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003426 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003427 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003428 else
3429 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003430 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003431 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432}
3433#endif /* __PAGETABLE_PUD_FOLDED */
3434
3435#ifndef __PAGETABLE_PMD_FOLDED
3436/*
3437 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003438 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003440int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003442 pmd_t *new = pmd_alloc_one(mm, address);
3443 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003444 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
Nick Piggin362a61a2008-05-14 06:37:36 +02003446 smp_wmb(); /* See comment in __pte_alloc */
3447
Hugh Dickins872fec12005-10-29 18:16:21 -07003448 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003450 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003451 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003452 else
3453 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003455 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003456 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003457 else
3458 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003460 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003461 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462}
3463#endif /* __PAGETABLE_PMD_FOLDED */
3464
3465int make_pages_present(unsigned long addr, unsigned long end)
3466{
3467 int ret, len, write;
3468 struct vm_area_struct * vma;
3469
3470 vma = find_vma(current->mm, addr);
3471 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003472 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003473 /*
3474 * We want to touch writable mappings with a write fault in order
3475 * to break COW, except for shared mappings because these don't COW
3476 * and we would not want to dirty them for nothing.
3477 */
3478 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003479 BUG_ON(addr >= end);
3480 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003481 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 ret = get_user_pages(current, current->mm, addr,
3483 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003484 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003486 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487}
3488
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489#if !defined(__HAVE_ARCH_GATE_AREA)
3490
3491#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003492static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
3494static int __init gate_vma_init(void)
3495{
3496 gate_vma.vm_mm = NULL;
3497 gate_vma.vm_start = FIXADDR_USER_START;
3498 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003499 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3500 gate_vma.vm_page_prot = __P101;
Roland McGrathf47aef52007-01-26 00:56:49 -08003501 /*
3502 * Make sure the vDSO gets into every core dump.
3503 * Dumping its contents makes post-mortem fully interpretable later
3504 * without matching up the same kernel and hardware config to see
3505 * what PC values meant.
3506 */
3507 gate_vma.vm_flags |= VM_ALWAYSDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 return 0;
3509}
3510__initcall(gate_vma_init);
3511#endif
3512
Stephen Wilson31db58b2011-03-13 15:49:15 -04003513struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514{
3515#ifdef AT_SYSINFO_EHDR
3516 return &gate_vma;
3517#else
3518 return NULL;
3519#endif
3520}
3521
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003522int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523{
3524#ifdef AT_SYSINFO_EHDR
3525 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3526 return 1;
3527#endif
3528 return 0;
3529}
3530
3531#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003532
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003533static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003534 pte_t **ptepp, spinlock_t **ptlp)
3535{
3536 pgd_t *pgd;
3537 pud_t *pud;
3538 pmd_t *pmd;
3539 pte_t *ptep;
3540
3541 pgd = pgd_offset(mm, address);
3542 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3543 goto out;
3544
3545 pud = pud_offset(pgd, address);
3546 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3547 goto out;
3548
3549 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08003550 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003551 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3552 goto out;
3553
3554 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3555 if (pmd_huge(*pmd))
3556 goto out;
3557
3558 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3559 if (!ptep)
3560 goto out;
3561 if (!pte_present(*ptep))
3562 goto unlock;
3563 *ptepp = ptep;
3564 return 0;
3565unlock:
3566 pte_unmap_unlock(ptep, *ptlp);
3567out:
3568 return -EINVAL;
3569}
3570
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003571static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3572 pte_t **ptepp, spinlock_t **ptlp)
3573{
3574 int res;
3575
3576 /* (void) is needed to make gcc happy */
3577 (void) __cond_lock(*ptlp,
3578 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3579 return res;
3580}
3581
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003582/**
3583 * follow_pfn - look up PFN at a user virtual address
3584 * @vma: memory mapping
3585 * @address: user virtual address
3586 * @pfn: location to store found PFN
3587 *
3588 * Only IO mappings and raw PFN mappings are allowed.
3589 *
3590 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3591 */
3592int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3593 unsigned long *pfn)
3594{
3595 int ret = -EINVAL;
3596 spinlock_t *ptl;
3597 pte_t *ptep;
3598
3599 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3600 return ret;
3601
3602 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3603 if (ret)
3604 return ret;
3605 *pfn = pte_pfn(*ptep);
3606 pte_unmap_unlock(ptep, ptl);
3607 return 0;
3608}
3609EXPORT_SYMBOL(follow_pfn);
3610
Rik van Riel28b2ee22008-07-23 21:27:05 -07003611#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003612int follow_phys(struct vm_area_struct *vma,
3613 unsigned long address, unsigned int flags,
3614 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003615{
Johannes Weiner03668a42009-06-16 15:32:34 -07003616 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003617 pte_t *ptep, pte;
3618 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003619
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003620 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3621 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003622
Johannes Weiner03668a42009-06-16 15:32:34 -07003623 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003624 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003625 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003626
Rik van Riel28b2ee22008-07-23 21:27:05 -07003627 if ((flags & FOLL_WRITE) && !pte_write(pte))
3628 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003629
3630 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003631 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003632
Johannes Weiner03668a42009-06-16 15:32:34 -07003633 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003634unlock:
3635 pte_unmap_unlock(ptep, ptl);
3636out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003637 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003638}
3639
3640int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3641 void *buf, int len, int write)
3642{
3643 resource_size_t phys_addr;
3644 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003645 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003646 int offset = addr & (PAGE_SIZE-1);
3647
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003648 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003649 return -EINVAL;
3650
3651 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3652 if (write)
3653 memcpy_toio(maddr + offset, buf, len);
3654 else
3655 memcpy_fromio(buf, maddr + offset, len);
3656 iounmap(maddr);
3657
3658 return len;
3659}
3660#endif
3661
David Howells0ec76a12006-09-27 01:50:15 -07003662/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003663 * Access another process' address space as given in mm. If non-NULL, use the
3664 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003665 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003666static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3667 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003668{
David Howells0ec76a12006-09-27 01:50:15 -07003669 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003670 void *old_buf = buf;
3671
David Howells0ec76a12006-09-27 01:50:15 -07003672 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003673 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003674 while (len) {
3675 int bytes, ret, offset;
3676 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003677 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003678
3679 ret = get_user_pages(tsk, mm, addr, 1,
3680 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003681 if (ret <= 0) {
3682 /*
3683 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3684 * we can access using slightly different code.
3685 */
3686#ifdef CONFIG_HAVE_IOREMAP_PROT
3687 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003688 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003689 break;
3690 if (vma->vm_ops && vma->vm_ops->access)
3691 ret = vma->vm_ops->access(vma, addr, buf,
3692 len, write);
3693 if (ret <= 0)
3694#endif
3695 break;
3696 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003697 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003698 bytes = len;
3699 offset = addr & (PAGE_SIZE-1);
3700 if (bytes > PAGE_SIZE-offset)
3701 bytes = PAGE_SIZE-offset;
3702
3703 maddr = kmap(page);
3704 if (write) {
3705 copy_to_user_page(vma, page, addr,
3706 maddr + offset, buf, bytes);
3707 set_page_dirty_lock(page);
3708 } else {
3709 copy_from_user_page(vma, page, addr,
3710 buf, maddr + offset, bytes);
3711 }
3712 kunmap(page);
3713 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003714 }
David Howells0ec76a12006-09-27 01:50:15 -07003715 len -= bytes;
3716 buf += bytes;
3717 addr += bytes;
3718 }
3719 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003720
3721 return buf - old_buf;
3722}
Andi Kleen03252912008-01-30 13:33:18 +01003723
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003724/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003725 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003726 * @mm: the mm_struct of the target address space
3727 * @addr: start address to access
3728 * @buf: source or destination buffer
3729 * @len: number of bytes to transfer
3730 * @write: whether the access is a write
3731 *
3732 * The caller must hold a reference on @mm.
3733 */
3734int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3735 void *buf, int len, int write)
3736{
3737 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3738}
3739
Andi Kleen03252912008-01-30 13:33:18 +01003740/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003741 * Access another process' address space.
3742 * Source/target buffer must be kernel space,
3743 * Do not walk the page table directly, use get_user_pages
3744 */
3745int access_process_vm(struct task_struct *tsk, unsigned long addr,
3746 void *buf, int len, int write)
3747{
3748 struct mm_struct *mm;
3749 int ret;
3750
3751 mm = get_task_mm(tsk);
3752 if (!mm)
3753 return 0;
3754
3755 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3756 mmput(mm);
3757
3758 return ret;
3759}
3760
Andi Kleen03252912008-01-30 13:33:18 +01003761/*
3762 * Print the name of a VMA.
3763 */
3764void print_vma_addr(char *prefix, unsigned long ip)
3765{
3766 struct mm_struct *mm = current->mm;
3767 struct vm_area_struct *vma;
3768
Ingo Molnare8bff742008-02-13 20:21:06 +01003769 /*
3770 * Do not print if we are in atomic
3771 * contexts (in exception stacks, etc.):
3772 */
3773 if (preempt_count())
3774 return;
3775
Andi Kleen03252912008-01-30 13:33:18 +01003776 down_read(&mm->mmap_sem);
3777 vma = find_vma(mm, ip);
3778 if (vma && vma->vm_file) {
3779 struct file *f = vma->vm_file;
3780 char *buf = (char *)__get_free_page(GFP_KERNEL);
3781 if (buf) {
3782 char *p, *s;
3783
Jan Blunckcf28b482008-02-14 19:38:44 -08003784 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003785 if (IS_ERR(p))
3786 p = "?";
3787 s = strrchr(p, '/');
3788 if (s)
3789 p = s+1;
3790 printk("%s%s[%lx+%lx]", prefix, p,
3791 vma->vm_start,
3792 vma->vm_end - vma->vm_start);
3793 free_page((unsigned long)buf);
3794 }
3795 }
3796 up_read(&current->mm->mmap_sem);
3797}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003798
3799#ifdef CONFIG_PROVE_LOCKING
3800void might_fault(void)
3801{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003802 /*
3803 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3804 * holding the mmap_sem, this is safe because kernel memory doesn't
3805 * get paged out, therefore we'll never actually fault, and the
3806 * below annotations will generate false positives.
3807 */
3808 if (segment_eq(get_fs(), KERNEL_DS))
3809 return;
3810
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003811 might_sleep();
3812 /*
3813 * it would be nicer only to annotate paths which are not under
3814 * pagefault_disable, however that requires a larger audit and
3815 * providing helpers like get_user_atomic.
3816 */
3817 if (!in_atomic() && current->mm)
3818 might_lock_read(&current->mm->mmap_sem);
3819}
3820EXPORT_SYMBOL(might_fault);
3821#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003822
3823#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3824static void clear_gigantic_page(struct page *page,
3825 unsigned long addr,
3826 unsigned int pages_per_huge_page)
3827{
3828 int i;
3829 struct page *p = page;
3830
3831 might_sleep();
3832 for (i = 0; i < pages_per_huge_page;
3833 i++, p = mem_map_next(p, page, i)) {
3834 cond_resched();
3835 clear_user_highpage(p, addr + i * PAGE_SIZE);
3836 }
3837}
3838void clear_huge_page(struct page *page,
3839 unsigned long addr, unsigned int pages_per_huge_page)
3840{
3841 int i;
3842
3843 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3844 clear_gigantic_page(page, addr, pages_per_huge_page);
3845 return;
3846 }
3847
3848 might_sleep();
3849 for (i = 0; i < pages_per_huge_page; i++) {
3850 cond_resched();
3851 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3852 }
3853}
3854
3855static void copy_user_gigantic_page(struct page *dst, struct page *src,
3856 unsigned long addr,
3857 struct vm_area_struct *vma,
3858 unsigned int pages_per_huge_page)
3859{
3860 int i;
3861 struct page *dst_base = dst;
3862 struct page *src_base = src;
3863
3864 for (i = 0; i < pages_per_huge_page; ) {
3865 cond_resched();
3866 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3867
3868 i++;
3869 dst = mem_map_next(dst, dst_base, i);
3870 src = mem_map_next(src, src_base, i);
3871 }
3872}
3873
3874void copy_user_huge_page(struct page *dst, struct page *src,
3875 unsigned long addr, struct vm_area_struct *vma,
3876 unsigned int pages_per_huge_page)
3877{
3878 int i;
3879
3880 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3881 copy_user_gigantic_page(dst, src, addr, vma,
3882 pages_per_huge_page);
3883 return;
3884 }
3885
3886 might_sleep();
3887 for (i = 0; i < pages_per_huge_page; i++) {
3888 cond_resched();
3889 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3890 }
3891}
3892#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */