blob: af82741caaa496886efa24b6c57817c4ccf80c4c [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
Hugh Dickins1bb36302005-10-29 18:16:22 -0700397int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800399 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700400 if (!new)
401 return -ENOMEM;
402
Nick Piggin362a61a2008-05-14 06:37:36 +0200403 /*
404 * Ensure all pte setup (eg. pte page lock and page clearing) are
405 * visible before the pte is made visible to other CPUs by being
406 * put into page tables.
407 *
408 * The other side of the story is the pointer chasing in the page
409 * table walking code (when walking the page table without locking;
410 * ie. most of the time). Fortunately, these data accesses consist
411 * of a chain of data-dependent loads, meaning most CPUs (alpha
412 * being the notable exception) will already guarantee loads are
413 * seen in-order. See the alpha page table accessors for the
414 * smp_read_barrier_depends() barriers in page table walking code.
415 */
416 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
417
Hugh Dickinsc74df322005-10-29 18:16:23 -0700418 spin_lock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800419 if (!pmd_present(*pmd)) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800422 new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 }
Hugh Dickinsc74df322005-10-29 18:16:23 -0700424 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800425 if (new)
426 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700427 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Hugh Dickins1bb36302005-10-29 18:16:22 -0700430int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700432 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
433 if (!new)
434 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Nick Piggin362a61a2008-05-14 06:37:36 +0200436 smp_wmb(); /* See comment in __pte_alloc */
437
Hugh Dickins1bb36302005-10-29 18:16:22 -0700438 spin_lock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800439 if (!pmd_present(*pmd)) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700440 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800441 new = NULL;
442 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700443 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800444 if (new)
445 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700446 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447}
448
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800449static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700450{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800451 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
452}
453
454static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
455{
456 int i;
457
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800458 if (current->mm == mm)
459 sync_mm_rss(current, mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800460 for (i = 0; i < NR_MM_COUNTERS; i++)
461 if (rss[i])
462 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700463}
464
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800466 * This function is called to print an error when a bad pte
467 * is found. For example, we might have a PFN-mapped pte in
468 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700469 *
470 * The calling function must still handle the error.
471 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800472static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
473 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700474{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800475 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
476 pud_t *pud = pud_offset(pgd, addr);
477 pmd_t *pmd = pmd_offset(pud, addr);
478 struct address_space *mapping;
479 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800480 static unsigned long resume;
481 static unsigned long nr_shown;
482 static unsigned long nr_unshown;
483
484 /*
485 * Allow a burst of 60 reports, then keep quiet for that minute;
486 * or allow a steady drip of one report per second.
487 */
488 if (nr_shown == 60) {
489 if (time_before(jiffies, resume)) {
490 nr_unshown++;
491 return;
492 }
493 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800494 printk(KERN_ALERT
495 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800496 nr_unshown);
497 nr_unshown = 0;
498 }
499 nr_shown = 0;
500 }
501 if (nr_shown++ == 0)
502 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800503
504 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
505 index = linear_page_index(vma, addr);
506
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800507 printk(KERN_ALERT
508 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800509 current->comm,
510 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800511 if (page)
512 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800513 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800514 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
515 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
516 /*
517 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
518 */
519 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800520 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800521 (unsigned long)vma->vm_ops->fault);
522 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800523 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800524 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700525 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800526 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700527}
528
Linus Torvalds67121172005-12-11 20:38:17 -0800529static inline int is_cow_mapping(unsigned int flags)
530{
531 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
532}
533
Hugh Dickins62eede62009-09-21 17:03:34 -0700534#ifndef is_zero_pfn
535static inline int is_zero_pfn(unsigned long pfn)
536{
537 return pfn == zero_pfn;
538}
539#endif
540
541#ifndef my_zero_pfn
542static inline unsigned long my_zero_pfn(unsigned long addr)
543{
544 return zero_pfn;
545}
546#endif
547
Nick Pigginb5810032005-10-29 18:16:12 -0700548/*
Nick Piggin7e675132008-04-28 02:13:00 -0700549 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800550 *
Nick Piggin7e675132008-04-28 02:13:00 -0700551 * "Special" mappings do not wish to be associated with a "struct page" (either
552 * it doesn't exist, or it exists but they don't want to touch it). In this
553 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700554 *
Nick Piggin7e675132008-04-28 02:13:00 -0700555 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
556 * pte bit, in which case this function is trivial. Secondly, an architecture
557 * may not have a spare pte bit, which requires a more complicated scheme,
558 * described below.
559 *
560 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
561 * special mapping (even if there are underlying and valid "struct pages").
562 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800563 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700564 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
565 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700566 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
567 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800568 *
569 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
570 *
Nick Piggin7e675132008-04-28 02:13:00 -0700571 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700572 *
Nick Piggin7e675132008-04-28 02:13:00 -0700573 * This restricts such mappings to be a linear translation from virtual address
574 * to pfn. To get around this restriction, we allow arbitrary mappings so long
575 * as the vma is not a COW mapping; in that case, we know that all ptes are
576 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700577 *
578 *
Nick Piggin7e675132008-04-28 02:13:00 -0700579 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
580 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700581 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
582 * page" backing, however the difference is that _all_ pages with a struct
583 * page (that is, those where pfn_valid is true) are refcounted and considered
584 * normal pages by the VM. The disadvantage is that pages are refcounted
585 * (which can be slower and simply not an option for some PFNMAP users). The
586 * advantage is that we don't have to follow the strict linearity rule of
587 * PFNMAP mappings in order to support COWable mappings.
588 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800589 */
Nick Piggin7e675132008-04-28 02:13:00 -0700590#ifdef __HAVE_ARCH_PTE_SPECIAL
591# define HAVE_PTE_SPECIAL 1
592#else
593# define HAVE_PTE_SPECIAL 0
594#endif
595struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
596 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800597{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800598 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700599
600 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800601 if (likely(!pte_special(pte)))
602 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700603 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
604 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700605 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800606 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700607 return NULL;
608 }
609
610 /* !HAVE_PTE_SPECIAL case follows: */
611
Jared Hulbertb379d792008-04-28 02:12:58 -0700612 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
613 if (vma->vm_flags & VM_MIXEDMAP) {
614 if (!pfn_valid(pfn))
615 return NULL;
616 goto out;
617 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700618 unsigned long off;
619 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700620 if (pfn == vma->vm_pgoff + off)
621 return NULL;
622 if (!is_cow_mapping(vma->vm_flags))
623 return NULL;
624 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800625 }
626
Hugh Dickins62eede62009-09-21 17:03:34 -0700627 if (is_zero_pfn(pfn))
628 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800629check_pfn:
630 if (unlikely(pfn > highest_memmap_pfn)) {
631 print_bad_pte(vma, addr, pte, NULL);
632 return NULL;
633 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800634
635 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700636 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700637 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800638 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700639out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800640 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800641}
642
643/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 * copy one vm_area from one task to the other. Assumes the page tables
645 * already present in the new task to be cleared in the whole range
646 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 */
648
Hugh Dickins570a3352009-12-14 17:58:46 -0800649static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700651 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700652 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
Nick Pigginb5810032005-10-29 18:16:12 -0700654 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 pte_t pte = *src_pte;
656 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
658 /* pte contains position in swap or file, so copy. */
659 if (unlikely(!pte_present(pte))) {
660 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700661 swp_entry_t entry = pte_to_swp_entry(pte);
662
Hugh Dickins570a3352009-12-14 17:58:46 -0800663 if (swap_duplicate(entry) < 0)
664 return entry.val;
665
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 /* make sure dst_mm is on swapoff's mmlist. */
667 if (unlikely(list_empty(&dst_mm->mmlist))) {
668 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700669 if (list_empty(&dst_mm->mmlist))
670 list_add(&dst_mm->mmlist,
671 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 spin_unlock(&mmlist_lock);
673 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800674 if (likely(!non_swap_entry(entry)))
675 rss[MM_SWAPENTS]++;
676 else if (is_write_migration_entry(entry) &&
Christoph Lameter06972122006-06-23 02:03:35 -0700677 is_cow_mapping(vm_flags)) {
678 /*
679 * COW mappings require pages in both parent
680 * and child to be set to read.
681 */
682 make_migration_entry_read(&entry);
683 pte = swp_entry_to_pte(entry);
684 set_pte_at(src_mm, addr, src_pte, pte);
685 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700687 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 }
689
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 /*
691 * If it's a COW mapping, write protect it both
692 * in the parent and the child
693 */
Linus Torvalds67121172005-12-11 20:38:17 -0800694 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700696 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 }
698
699 /*
700 * If it's a shared mapping, mark it clean in
701 * the child
702 */
703 if (vm_flags & VM_SHARED)
704 pte = pte_mkclean(pte);
705 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800706
707 page = vm_normal_page(vma, addr, pte);
708 if (page) {
709 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700710 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800711 if (PageAnon(page))
712 rss[MM_ANONPAGES]++;
713 else
714 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800715 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700716
717out_set_pte:
718 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a3352009-12-14 17:58:46 -0800719 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720}
721
722static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
723 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
724 unsigned long addr, unsigned long end)
725{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700726 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700728 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700729 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800730 int rss[NR_MM_COUNTERS];
Hugh Dickins570a3352009-12-14 17:58:46 -0800731 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732
733again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800734 init_rss_vec(rss);
735
Hugh Dickinsc74df322005-10-29 18:16:23 -0700736 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 if (!dst_pte)
738 return -ENOMEM;
739 src_pte = pte_offset_map_nested(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700740 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700741 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700742 orig_src_pte = src_pte;
743 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700744 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 do {
747 /*
748 * We are holding two locks at this point - either of them
749 * could generate latencies in another task on another CPU.
750 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700751 if (progress >= 32) {
752 progress = 0;
753 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100754 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700755 break;
756 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 if (pte_none(*src_pte)) {
758 progress++;
759 continue;
760 }
Hugh Dickins570a3352009-12-14 17:58:46 -0800761 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
762 vma, addr, rss);
763 if (entry.val)
764 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 progress += 8;
766 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700768 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700769 spin_unlock(src_ptl);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700770 pte_unmap_nested(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800771 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700772 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700773 cond_resched();
Hugh Dickins570a3352009-12-14 17:58:46 -0800774
775 if (entry.val) {
776 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
777 return -ENOMEM;
778 progress = 0;
779 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 if (addr != end)
781 goto again;
782 return 0;
783}
784
785static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
786 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
787 unsigned long addr, unsigned long end)
788{
789 pmd_t *src_pmd, *dst_pmd;
790 unsigned long next;
791
792 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
793 if (!dst_pmd)
794 return -ENOMEM;
795 src_pmd = pmd_offset(src_pud, addr);
796 do {
797 next = pmd_addr_end(addr, end);
798 if (pmd_none_or_clear_bad(src_pmd))
799 continue;
800 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
801 vma, addr, next))
802 return -ENOMEM;
803 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
804 return 0;
805}
806
807static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
808 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
809 unsigned long addr, unsigned long end)
810{
811 pud_t *src_pud, *dst_pud;
812 unsigned long next;
813
814 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
815 if (!dst_pud)
816 return -ENOMEM;
817 src_pud = pud_offset(src_pgd, addr);
818 do {
819 next = pud_addr_end(addr, end);
820 if (pud_none_or_clear_bad(src_pud))
821 continue;
822 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
823 vma, addr, next))
824 return -ENOMEM;
825 } while (dst_pud++, src_pud++, addr = next, addr != end);
826 return 0;
827}
828
829int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
830 struct vm_area_struct *vma)
831{
832 pgd_t *src_pgd, *dst_pgd;
833 unsigned long next;
834 unsigned long addr = vma->vm_start;
835 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700836 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
Nick Piggind9928952005-08-28 16:49:11 +1000838 /*
839 * Don't copy ptes where a page fault will fill them correctly.
840 * Fork becomes much lighter when there are big shared or private
841 * readonly mappings. The tradeoff is that copy_page_range is more
842 * efficient than faulting.
843 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -0800844 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +1000845 if (!vma->anon_vma)
846 return 0;
847 }
848
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 if (is_vm_hugetlb_page(vma))
850 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
851
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -0800852 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -0800853 /*
854 * We do not free on error cases below as remove_vma
855 * gets called on error from higher level routine
856 */
857 ret = track_pfn_vma_copy(vma);
858 if (ret)
859 return ret;
860 }
861
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700862 /*
863 * We need to invalidate the secondary MMU mappings only when
864 * there could be a permission downgrade on the ptes of the
865 * parent mm. And a permission downgrade will only happen if
866 * is_cow_mapping() returns true.
867 */
868 if (is_cow_mapping(vma->vm_flags))
869 mmu_notifier_invalidate_range_start(src_mm, addr, end);
870
871 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 dst_pgd = pgd_offset(dst_mm, addr);
873 src_pgd = pgd_offset(src_mm, addr);
874 do {
875 next = pgd_addr_end(addr, end);
876 if (pgd_none_or_clear_bad(src_pgd))
877 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700878 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
879 vma, addr, next))) {
880 ret = -ENOMEM;
881 break;
882 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700884
885 if (is_cow_mapping(vma->vm_flags))
886 mmu_notifier_invalidate_range_end(src_mm,
887 vma->vm_start, end);
888 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889}
890
Robin Holt51c6f662005-11-13 16:06:42 -0800891static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700892 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800894 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895{
Nick Pigginb5810032005-10-29 18:16:12 -0700896 struct mm_struct *mm = tlb->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 pte_t *pte;
Hugh Dickins508034a2005-10-29 18:16:30 -0700898 spinlock_t *ptl;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800899 int rss[NR_MM_COUNTERS];
900
901 init_rss_vec(rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902
Hugh Dickins508034a2005-10-29 18:16:30 -0700903 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700904 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 do {
906 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -0800907 if (pte_none(ptent)) {
908 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800910 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -0800911
912 (*zap_work) -= PAGE_SIZE;
913
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800915 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -0800916
Linus Torvalds6aab3412005-11-28 14:34:23 -0800917 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 if (unlikely(details) && page) {
919 /*
920 * unmap_shared_mapping_pages() wants to
921 * invalidate cache without truncating:
922 * unmap shared but keep private pages.
923 */
924 if (details->check_mapping &&
925 details->check_mapping != page->mapping)
926 continue;
927 /*
928 * Each page->index must be checked when
929 * invalidating or truncating nonlinear.
930 */
931 if (details->nonlinear_vma &&
932 (page->index < details->first_index ||
933 page->index > details->last_index))
934 continue;
935 }
Nick Pigginb5810032005-10-29 18:16:12 -0700936 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -0700937 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 tlb_remove_tlb_entry(tlb, pte, addr);
939 if (unlikely(!page))
940 continue;
941 if (unlikely(details) && details->nonlinear_vma
942 && linear_page_index(details->nonlinear_vma,
943 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -0700944 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800947 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700948 else {
949 if (pte_dirty(ptent))
950 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800951 if (pte_young(ptent) &&
952 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -0800953 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800954 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700955 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800956 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800957 if (unlikely(page_mapcount(page) < 0))
958 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 tlb_remove_page(tlb, page);
960 continue;
961 }
962 /*
963 * If details->check_mapping, we leave swap entries;
964 * if details->nonlinear_vma, we leave file entries.
965 */
966 if (unlikely(details))
967 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -0800968 if (pte_file(ptent)) {
969 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
970 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800971 } else {
972 swp_entry_t entry = pte_to_swp_entry(ptent);
973
974 if (!non_swap_entry(entry))
975 rss[MM_SWAPENTS]--;
976 if (unlikely(!free_swap_and_cache(entry)))
977 print_bad_pte(vma, addr, ptent, NULL);
978 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -0700979 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Robin Holt51c6f662005-11-13 16:06:42 -0800980 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
Hugh Dickinsae859762005-10-29 18:16:05 -0700981
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800982 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700983 arch_leave_lazy_mmu_mode();
Hugh Dickins508034a2005-10-29 18:16:30 -0700984 pte_unmap_unlock(pte - 1, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -0800985
986 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987}
988
Robin Holt51c6f662005-11-13 16:06:42 -0800989static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700990 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800992 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993{
994 pmd_t *pmd;
995 unsigned long next;
996
997 pmd = pmd_offset(pud, addr);
998 do {
999 next = pmd_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -08001000 if (pmd_none_or_clear_bad(pmd)) {
1001 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001003 }
1004 next = zap_pte_range(tlb, vma, pmd, addr, next,
1005 zap_work, details);
1006 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1007
1008 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009}
1010
Robin Holt51c6f662005-11-13 16:06:42 -08001011static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001012 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -08001014 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015{
1016 pud_t *pud;
1017 unsigned long next;
1018
1019 pud = pud_offset(pgd, addr);
1020 do {
1021 next = pud_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -08001022 if (pud_none_or_clear_bad(pud)) {
1023 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001025 }
1026 next = zap_pmd_range(tlb, vma, pud, addr, next,
1027 zap_work, details);
1028 } while (pud++, addr = next, (addr != end && *zap_work > 0));
1029
1030 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031}
1032
Robin Holt51c6f662005-11-13 16:06:42 -08001033static unsigned long unmap_page_range(struct mmu_gather *tlb,
1034 struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -08001036 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037{
1038 pgd_t *pgd;
1039 unsigned long next;
1040
1041 if (details && !details->check_mapping && !details->nonlinear_vma)
1042 details = NULL;
1043
1044 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001045 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 tlb_start_vma(tlb, vma);
1047 pgd = pgd_offset(vma->vm_mm, addr);
1048 do {
1049 next = pgd_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -08001050 if (pgd_none_or_clear_bad(pgd)) {
1051 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001053 }
1054 next = zap_pud_range(tlb, vma, pgd, addr, next,
1055 zap_work, details);
1056 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001058 mem_cgroup_uncharge_end();
Robin Holt51c6f662005-11-13 16:06:42 -08001059
1060 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061}
1062
1063#ifdef CONFIG_PREEMPT
1064# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1065#else
1066/* No preempt: go for improved straight-line efficiency */
1067# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1068#endif
1069
1070/**
1071 * unmap_vmas - unmap a range of memory covered by a list of vma's
1072 * @tlbp: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 * @vma: the starting vma
1074 * @start_addr: virtual address at which to start unmapping
1075 * @end_addr: virtual address at which to end unmapping
1076 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1077 * @details: details of nonlinear truncation or shared cache invalidation
1078 *
Hugh Dickinsee39b372005-04-19 13:29:15 -07001079 * Returns the end address of the unmapping (restart addr if interrupted).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001081 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001083 * We aim to not hold locks for too long (for scheduling latency reasons).
1084 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 * return the ending mmu_gather to the caller.
1086 *
1087 * Only addresses between `start' and `end' will be unmapped.
1088 *
1089 * The VMA list must be sorted in ascending virtual address order.
1090 *
1091 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1092 * range after unmap_vmas() returns. So the only responsibility here is to
1093 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1094 * drops the lock and schedules.
1095 */
Hugh Dickins508034a2005-10-29 18:16:30 -07001096unsigned long unmap_vmas(struct mmu_gather **tlbp,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 struct vm_area_struct *vma, unsigned long start_addr,
1098 unsigned long end_addr, unsigned long *nr_accounted,
1099 struct zap_details *details)
1100{
Robin Holt51c6f662005-11-13 16:06:42 -08001101 long zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
1103 int tlb_start_valid = 0;
Hugh Dickinsee39b372005-04-19 13:29:15 -07001104 unsigned long start = start_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
Hugh Dickins4d6ddfa2005-10-29 18:16:02 -07001106 int fullmm = (*tlbp)->fullmm;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001107 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001109 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 unsigned long end;
1112
1113 start = max(vma->vm_start, start_addr);
1114 if (start >= vma->vm_end)
1115 continue;
1116 end = min(vma->vm_end, end_addr);
1117 if (end <= vma->vm_start)
1118 continue;
1119
1120 if (vma->vm_flags & VM_ACCOUNT)
1121 *nr_accounted += (end - start) >> PAGE_SHIFT;
1122
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001123 if (unlikely(is_pfn_mapping(vma)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001124 untrack_pfn_vma(vma, 0, 0);
1125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 while (start != end) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 if (!tlb_start_valid) {
1128 tlb_start = start;
1129 tlb_start_valid = 1;
1130 }
1131
Robin Holt51c6f662005-11-13 16:06:42 -08001132 if (unlikely(is_vm_hugetlb_page(vma))) {
Andi Kleena137e1c2008-07-23 21:27:43 -07001133 /*
1134 * It is undesirable to test vma->vm_file as it
1135 * should be non-null for valid hugetlb area.
1136 * However, vm_file will be NULL in the error
1137 * cleanup path of do_mmap_pgoff. When
1138 * hugetlbfs ->mmap method fails,
1139 * do_mmap_pgoff() nullifies vma->vm_file
1140 * before calling this function to clean up.
1141 * Since no pte has actually been setup, it is
1142 * safe to do nothing in this case.
1143 */
1144 if (vma->vm_file) {
1145 unmap_hugepage_range(vma, start, end, NULL);
1146 zap_work -= (end - start) /
Andi Kleena5516432008-07-23 21:27:41 -07001147 pages_per_huge_page(hstate_vma(vma));
Andi Kleena137e1c2008-07-23 21:27:43 -07001148 }
1149
Robin Holt51c6f662005-11-13 16:06:42 -08001150 start = end;
1151 } else
1152 start = unmap_page_range(*tlbp, vma,
1153 start, end, &zap_work, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
Robin Holt51c6f662005-11-13 16:06:42 -08001155 if (zap_work > 0) {
1156 BUG_ON(start != end);
1157 break;
1158 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
1160 tlb_finish_mmu(*tlbp, tlb_start, start);
1161
1162 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +01001163 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 if (i_mmap_lock) {
Hugh Dickins508034a2005-10-29 18:16:30 -07001165 *tlbp = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 goto out;
1167 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 }
1170
Hugh Dickins508034a2005-10-29 18:16:30 -07001171 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 tlb_start_valid = 0;
Robin Holt51c6f662005-11-13 16:06:42 -08001173 zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 }
1175 }
1176out:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001177 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001178 return start; /* which is now the end (or restart) address */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179}
1180
1181/**
1182 * zap_page_range - remove user pages in a given range
1183 * @vma: vm_area_struct holding the applicable pages
1184 * @address: starting address of pages to zap
1185 * @size: number of bytes to zap
1186 * @details: details of nonlinear truncation or shared cache invalidation
1187 */
Hugh Dickinsee39b372005-04-19 13:29:15 -07001188unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 unsigned long size, struct zap_details *details)
1190{
1191 struct mm_struct *mm = vma->vm_mm;
1192 struct mmu_gather *tlb;
1193 unsigned long end = address + size;
1194 unsigned long nr_accounted = 0;
1195
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 lru_add_drain();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 tlb = tlb_gather_mmu(mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001198 update_hiwater_rss(mm);
Hugh Dickins508034a2005-10-29 18:16:30 -07001199 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1200 if (tlb)
1201 tlb_finish_mmu(tlb, address, end);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001202 return end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203}
1204
Jack Steinerc627f9c2008-07-29 22:33:53 -07001205/**
1206 * zap_vma_ptes - remove ptes mapping the vma
1207 * @vma: vm_area_struct holding ptes to be zapped
1208 * @address: starting address of pages to zap
1209 * @size: number of bytes to zap
1210 *
1211 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1212 *
1213 * The entire address range must be fully contained within the vma.
1214 *
1215 * Returns 0 if successful.
1216 */
1217int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1218 unsigned long size)
1219{
1220 if (address < vma->vm_start || address + size > vma->vm_end ||
1221 !(vma->vm_flags & VM_PFNMAP))
1222 return -1;
1223 zap_page_range(vma, address, size, NULL);
1224 return 0;
1225}
1226EXPORT_SYMBOL_GPL(zap_vma_ptes);
1227
Johannes Weiner142762b2010-05-24 14:32:39 -07001228/**
1229 * follow_page - look up a page descriptor from a user-virtual address
1230 * @vma: vm_area_struct mapping @address
1231 * @address: virtual address to look up
1232 * @flags: flags modifying lookup behaviour
1233 *
1234 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1235 *
1236 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1237 * an error pointer if there is a mapping to something not represented
1238 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001240struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001241 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242{
1243 pgd_t *pgd;
1244 pud_t *pud;
1245 pmd_t *pmd;
1246 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001247 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001249 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001251 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1252 if (!IS_ERR(page)) {
1253 BUG_ON(flags & FOLL_GET);
1254 goto out;
1255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001257 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 pgd = pgd_offset(mm, address);
1259 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001260 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261
1262 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001263 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001264 goto no_page_table;
Andi Kleenceb86872008-07-23 21:27:50 -07001265 if (pud_huge(*pud)) {
1266 BUG_ON(flags & FOLL_GET);
1267 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1268 goto out;
1269 }
1270 if (unlikely(pud_bad(*pud)))
1271 goto no_page_table;
1272
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001274 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001275 goto no_page_table;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001276 if (pmd_huge(*pmd)) {
1277 BUG_ON(flags & FOLL_GET);
1278 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1279 goto out;
1280 }
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001281 if (unlikely(pmd_bad(*pmd)))
1282 goto no_page_table;
1283
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001284 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285
1286 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001287 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001288 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001289 if ((flags & FOLL_WRITE) && !pte_write(pte))
1290 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001291
Linus Torvalds6aab3412005-11-28 14:34:23 -08001292 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001293 if (unlikely(!page)) {
1294 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001295 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001296 goto bad_page;
1297 page = pte_page(pte);
1298 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001299
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001300 if (flags & FOLL_GET)
1301 get_page(page);
1302 if (flags & FOLL_TOUCH) {
1303 if ((flags & FOLL_WRITE) &&
1304 !pte_dirty(pte) && !PageDirty(page))
1305 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001306 /*
1307 * pte_mkyoung() would be more correct here, but atomic care
1308 * is needed to avoid losing the dirty bit: it is easier to use
1309 * mark_page_accessed().
1310 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001311 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001313unlock:
1314 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001316 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001318bad_page:
1319 pte_unmap_unlock(ptep, ptl);
1320 return ERR_PTR(-EFAULT);
1321
1322no_page:
1323 pte_unmap_unlock(ptep, ptl);
1324 if (!pte_none(pte))
1325 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001326
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001327no_page_table:
1328 /*
1329 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001330 * has touched so far, we don't want to allocate unnecessary pages or
1331 * page tables. Return error instead of NULL to skip handle_mm_fault,
1332 * then get_dump_page() will return NULL to leave a hole in the dump.
1333 * But we can only make this optimization where a hole would surely
1334 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001335 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001336 if ((flags & FOLL_DUMP) &&
1337 (!vma->vm_ops || !vma->vm_ops->fault))
1338 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001339 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340}
1341
Nick Pigginb291f002008-10-18 20:26:44 -07001342int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001343 unsigned long start, int nr_pages, unsigned int gup_flags,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001344 struct page **pages, struct vm_area_struct **vmas)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345{
1346 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001347 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Peter Zijlstra9d737772009-06-25 11:58:55 +02001349 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001350 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001351
1352 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1353
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 /*
1355 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001356 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001358 vm_flags = (gup_flags & FOLL_WRITE) ?
1359 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1360 vm_flags &= (gup_flags & FOLL_FORCE) ?
1361 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 i = 0;
1363
1364 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001365 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
1367 vma = find_extend_vma(mm, start);
1368 if (!vma && in_gate_area(tsk, start)) {
1369 unsigned long pg = start & PAGE_MASK;
1370 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1371 pgd_t *pgd;
1372 pud_t *pud;
1373 pmd_t *pmd;
1374 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001375
1376 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001377 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 return i ? : -EFAULT;
1379 if (pg > TASK_SIZE)
1380 pgd = pgd_offset_k(pg);
1381 else
1382 pgd = pgd_offset_gate(mm, pg);
1383 BUG_ON(pgd_none(*pgd));
1384 pud = pud_offset(pgd, pg);
1385 BUG_ON(pud_none(*pud));
1386 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001387 if (pmd_none(*pmd))
1388 return i ? : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001390 if (pte_none(*pte)) {
1391 pte_unmap(pte);
1392 return i ? : -EFAULT;
1393 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 if (pages) {
Hugh Dickinsde512572010-07-30 10:58:26 -07001395 struct page *page;
1396
1397 page = vm_normal_page(gate_vma, start, *pte);
1398 if (!page) {
1399 if (!(gup_flags & FOLL_DUMP) &&
1400 is_zero_pfn(pte_pfn(*pte)))
1401 page = pte_page(*pte);
1402 else {
1403 pte_unmap(pte);
1404 return i ? : -EFAULT;
1405 }
1406 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001407 pages[i] = page;
Hugh Dickinsde512572010-07-30 10:58:26 -07001408 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 }
1410 pte_unmap(pte);
1411 if (vmas)
1412 vmas[i] = gate_vma;
1413 i++;
1414 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001415 nr_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 continue;
1417 }
1418
Nick Pigginb291f002008-10-18 20:26:44 -07001419 if (!vma ||
1420 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001421 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 return i ? : -EFAULT;
1423
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001424 if (is_vm_hugetlb_page(vma)) {
1425 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001426 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001427 continue;
1428 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001429
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001431 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001432 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433
Ethan Solomita462e00c2007-07-15 23:38:16 -07001434 /*
Ying Han47792802009-01-06 14:40:18 -08001435 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001436 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001437 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001438 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001439 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001440
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001441 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001442 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001443 int ret;
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001444
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001445 ret = handle_mm_fault(mm, vma, start,
1446 (foll_flags & FOLL_WRITE) ?
1447 FAULT_FLAG_WRITE : 0);
1448
Nick Piggin83c54072007-07-19 01:47:05 -07001449 if (ret & VM_FAULT_ERROR) {
1450 if (ret & VM_FAULT_OOM)
1451 return i ? i : -ENOMEM;
Andi Kleend1737fd2009-09-16 11:50:06 +02001452 if (ret &
Andi Kleenaa50d3a2010-10-06 21:45:00 +02001453 (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE|
1454 VM_FAULT_SIGBUS))
Nick Piggin83c54072007-07-19 01:47:05 -07001455 return i ? i : -EFAULT;
1456 BUG();
1457 }
1458 if (ret & VM_FAULT_MAJOR)
1459 tsk->maj_flt++;
1460 else
1461 tsk->min_flt++;
1462
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001463 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001464 * The VM_FAULT_WRITE bit tells us that
1465 * do_wp_page has broken COW when necessary,
1466 * even if maybe_mkwrite decided not to set
1467 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001468 * page lookups as if they were reads. But only
1469 * do so when looping for pte_write is futile:
1470 * in some cases userspace may also be wanting
1471 * to write to the gotten user page, which a
1472 * read fault here might prevent (a readonly
1473 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001474 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001475 if ((ret & VM_FAULT_WRITE) &&
1476 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001477 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001478
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001479 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001481 if (IS_ERR(page))
1482 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001484 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001485
Russell Kinga6f36be2006-12-30 22:24:19 +00001486 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001487 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 }
1489 if (vmas)
1490 vmas[i] = vma;
1491 i++;
1492 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001493 nr_pages--;
1494 } while (nr_pages && start < vma->vm_end);
1495 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 return i;
1497}
Nick Pigginb291f002008-10-18 20:26:44 -07001498
Nick Piggind2bf6be2009-06-16 15:31:39 -07001499/**
1500 * get_user_pages() - pin user pages in memory
1501 * @tsk: task_struct of target task
1502 * @mm: mm_struct of target mm
1503 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001504 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001505 * @write: whether pages will be written to by the caller
1506 * @force: whether to force write access even if user mapping is
1507 * readonly. This will result in the page being COWed even
1508 * in MAP_SHARED mappings. You do not want this.
1509 * @pages: array that receives pointers to the pages pinned.
1510 * Should be at least nr_pages long. Or NULL, if caller
1511 * only intends to ensure the pages are faulted in.
1512 * @vmas: array of pointers to vmas corresponding to each page.
1513 * Or NULL if the caller does not require them.
1514 *
1515 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001516 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001517 * were pinned, returns -errno. Each page returned must be released
1518 * with a put_page() call when it is finished with. vmas will only
1519 * remain valid while mmap_sem is held.
1520 *
1521 * Must be called with mmap_sem held for read or write.
1522 *
1523 * get_user_pages walks a process's page tables and takes a reference to
1524 * each struct page that each user address corresponds to at a given
1525 * instant. That is, it takes the page that would be accessed if a user
1526 * thread accesses the given user virtual address at that instant.
1527 *
1528 * This does not guarantee that the page exists in the user mappings when
1529 * get_user_pages returns, and there may even be a completely different
1530 * page there in some cases (eg. if mmapped pagecache has been invalidated
1531 * and subsequently re faulted). However it does guarantee that the page
1532 * won't be freed completely. And mostly callers simply care that the page
1533 * contains data that was valid *at some point in time*. Typically, an IO
1534 * or similar operation cannot guarantee anything stronger anyway because
1535 * locks can't be held over the syscall boundary.
1536 *
1537 * If write=0, the page must not be written to. If the page is written to,
1538 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1539 * after the page is finished with, and before put_page is called.
1540 *
1541 * get_user_pages is typically used for fewer-copy IO operations, to get a
1542 * handle on the memory by some means other than accesses via the user virtual
1543 * addresses. The pages may be submitted for DMA to devices or accessed via
1544 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1545 * use the correct cache flushing APIs.
1546 *
1547 * See also get_user_pages_fast, for performance critical applications.
1548 */
Nick Pigginb291f002008-10-18 20:26:44 -07001549int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001550 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001551 struct page **pages, struct vm_area_struct **vmas)
1552{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001553 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001554
Hugh Dickins58fa8792009-09-21 17:03:31 -07001555 if (pages)
1556 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001557 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001558 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001559 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001560 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001561
Peter Zijlstra9d737772009-06-25 11:58:55 +02001562 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
Nick Pigginb291f002008-10-18 20:26:44 -07001563}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564EXPORT_SYMBOL(get_user_pages);
1565
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001566/**
1567 * get_dump_page() - pin user page in memory while writing it to core dump
1568 * @addr: user address
1569 *
1570 * Returns struct page pointer of user page pinned for dump,
1571 * to be freed afterwards by page_cache_release() or put_page().
1572 *
1573 * Returns NULL on any kind of failure - a hole must then be inserted into
1574 * the corefile, to preserve alignment with its headers; and also returns
1575 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1576 * allowing a hole to be left in the corefile to save diskspace.
1577 *
1578 * Called without mmap_sem, but after all other threads have been killed.
1579 */
1580#ifdef CONFIG_ELF_CORE
1581struct page *get_dump_page(unsigned long addr)
1582{
1583 struct vm_area_struct *vma;
1584 struct page *page;
1585
1586 if (__get_user_pages(current, current->mm, addr, 1,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001587 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001588 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001589 flush_cache_page(vma, addr, page_to_pfn(page));
1590 return page;
1591}
1592#endif /* CONFIG_ELF_CORE */
1593
Harvey Harrison920c7a52008-02-04 22:29:26 -08001594pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1595 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001596{
1597 pgd_t * pgd = pgd_offset(mm, addr);
1598 pud_t * pud = pud_alloc(mm, pgd, addr);
1599 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001600 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001601 if (pmd)
1602 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1603 }
1604 return NULL;
1605}
1606
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001608 * This is the old fallback for page remapping.
1609 *
1610 * For historical reasons, it only allows reserved pages. Only
1611 * old drivers should use this, and they needed to mark their
1612 * pages reserved for the old functions anyway.
1613 */
Nick Piggin423bad62008-04-28 02:13:01 -07001614static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1615 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001616{
Nick Piggin423bad62008-04-28 02:13:01 -07001617 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001618 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001619 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001620 spinlock_t *ptl;
1621
Linus Torvalds238f58d2005-11-29 13:01:56 -08001622 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001623 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001624 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001625 retval = -ENOMEM;
1626 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001627 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001628 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001629 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001630 retval = -EBUSY;
1631 if (!pte_none(*pte))
1632 goto out_unlock;
1633
1634 /* Ok, finally just insert the thing.. */
1635 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001636 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001637 page_add_file_rmap(page);
1638 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1639
1640 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001641 pte_unmap_unlock(pte, ptl);
1642 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001643out_unlock:
1644 pte_unmap_unlock(pte, ptl);
1645out:
1646 return retval;
1647}
1648
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001649/**
1650 * vm_insert_page - insert single page into user vma
1651 * @vma: user vma to map to
1652 * @addr: target user address of this page
1653 * @page: source kernel page
1654 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001655 * This allows drivers to insert individual pages they've allocated
1656 * into a user vma.
1657 *
1658 * The page has to be a nice clean _individual_ kernel allocation.
1659 * If you allocate a compound page, you need to have marked it as
1660 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001661 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001662 *
1663 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1664 * took an arbitrary page protection parameter. This doesn't allow
1665 * that. Your vma protection will have to be set up correctly, which
1666 * means that if you want a shared writable mapping, you'd better
1667 * ask for a shared writable mapping!
1668 *
1669 * The page does not need to be reserved.
1670 */
Nick Piggin423bad62008-04-28 02:13:01 -07001671int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1672 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001673{
1674 if (addr < vma->vm_start || addr >= vma->vm_end)
1675 return -EFAULT;
1676 if (!page_count(page))
1677 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001678 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07001679 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001680}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001681EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001682
Nick Piggin423bad62008-04-28 02:13:01 -07001683static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1684 unsigned long pfn, pgprot_t prot)
1685{
1686 struct mm_struct *mm = vma->vm_mm;
1687 int retval;
1688 pte_t *pte, entry;
1689 spinlock_t *ptl;
1690
1691 retval = -ENOMEM;
1692 pte = get_locked_pte(mm, addr, &ptl);
1693 if (!pte)
1694 goto out;
1695 retval = -EBUSY;
1696 if (!pte_none(*pte))
1697 goto out_unlock;
1698
1699 /* Ok, finally just insert the thing.. */
1700 entry = pte_mkspecial(pfn_pte(pfn, prot));
1701 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001702 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001703
1704 retval = 0;
1705out_unlock:
1706 pte_unmap_unlock(pte, ptl);
1707out:
1708 return retval;
1709}
1710
Nick Piggine0dc0d82007-02-12 00:51:36 -08001711/**
1712 * vm_insert_pfn - insert single pfn into user vma
1713 * @vma: user vma to map to
1714 * @addr: target user address of this page
1715 * @pfn: source kernel pfn
1716 *
1717 * Similar to vm_inert_page, this allows drivers to insert individual pages
1718 * they've allocated into a user vma. Same comments apply.
1719 *
1720 * This function should only be called from a vm_ops->fault handler, and
1721 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001722 *
1723 * vma cannot be a COW mapping.
1724 *
1725 * As this is called only for pages that do not currently exist, we
1726 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001727 */
1728int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001729 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001730{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001731 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001732 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001733 /*
1734 * Technically, architectures with pte_special can avoid all these
1735 * restrictions (same for remap_pfn_range). However we would like
1736 * consistency in testing and feature parity among all, so we should
1737 * try to keep these invariants in place for everybody.
1738 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001739 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1740 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1741 (VM_PFNMAP|VM_MIXEDMAP));
1742 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1743 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001744
Nick Piggin423bad62008-04-28 02:13:01 -07001745 if (addr < vma->vm_start || addr >= vma->vm_end)
1746 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001747 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001748 return -EINVAL;
1749
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001750 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001751
1752 if (ret)
1753 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1754
1755 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001756}
1757EXPORT_SYMBOL(vm_insert_pfn);
1758
Nick Piggin423bad62008-04-28 02:13:01 -07001759int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1760 unsigned long pfn)
1761{
1762 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1763
1764 if (addr < vma->vm_start || addr >= vma->vm_end)
1765 return -EFAULT;
1766
1767 /*
1768 * If we don't have pte special, then we have to use the pfn_valid()
1769 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1770 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001771 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1772 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001773 */
1774 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1775 struct page *page;
1776
1777 page = pfn_to_page(pfn);
1778 return insert_page(vma, addr, page, vma->vm_page_prot);
1779 }
1780 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1781}
1782EXPORT_SYMBOL(vm_insert_mixed);
1783
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001784/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 * maps a range of physical memory into the requested pages. the old
1786 * mappings are removed. any references to nonexistent pages results
1787 * in null mappings (currently treated as "copy-on-access")
1788 */
1789static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1790 unsigned long addr, unsigned long end,
1791 unsigned long pfn, pgprot_t prot)
1792{
1793 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001794 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
Hugh Dickinsc74df322005-10-29 18:16:23 -07001796 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 if (!pte)
1798 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001799 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 do {
1801 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001802 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 pfn++;
1804 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001805 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001806 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 return 0;
1808}
1809
1810static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1811 unsigned long addr, unsigned long end,
1812 unsigned long pfn, pgprot_t prot)
1813{
1814 pmd_t *pmd;
1815 unsigned long next;
1816
1817 pfn -= addr >> PAGE_SHIFT;
1818 pmd = pmd_alloc(mm, pud, addr);
1819 if (!pmd)
1820 return -ENOMEM;
1821 do {
1822 next = pmd_addr_end(addr, end);
1823 if (remap_pte_range(mm, pmd, addr, next,
1824 pfn + (addr >> PAGE_SHIFT), prot))
1825 return -ENOMEM;
1826 } while (pmd++, addr = next, addr != end);
1827 return 0;
1828}
1829
1830static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1831 unsigned long addr, unsigned long end,
1832 unsigned long pfn, pgprot_t prot)
1833{
1834 pud_t *pud;
1835 unsigned long next;
1836
1837 pfn -= addr >> PAGE_SHIFT;
1838 pud = pud_alloc(mm, pgd, addr);
1839 if (!pud)
1840 return -ENOMEM;
1841 do {
1842 next = pud_addr_end(addr, end);
1843 if (remap_pmd_range(mm, pud, addr, next,
1844 pfn + (addr >> PAGE_SHIFT), prot))
1845 return -ENOMEM;
1846 } while (pud++, addr = next, addr != end);
1847 return 0;
1848}
1849
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001850/**
1851 * remap_pfn_range - remap kernel memory to userspace
1852 * @vma: user vma to map to
1853 * @addr: target user address to start at
1854 * @pfn: physical address of kernel memory
1855 * @size: size of map area
1856 * @prot: page protection flags for this mapping
1857 *
1858 * Note: this is only safe if the mm semaphore is held when called.
1859 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1861 unsigned long pfn, unsigned long size, pgprot_t prot)
1862{
1863 pgd_t *pgd;
1864 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001865 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 struct mm_struct *mm = vma->vm_mm;
1867 int err;
1868
1869 /*
1870 * Physically remapped pages are special. Tell the
1871 * rest of the world about it:
1872 * VM_IO tells people not to look at these pages
1873 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08001874 * VM_RESERVED is specified all over the place, because
1875 * in 2.4 it kept swapout's vma scan off this vma; but
1876 * in 2.6 the LRU scan won't even find its pages, so this
1877 * flag means no more than count its pages in reserved_vm,
1878 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08001879 * VM_PFNMAP tells the core MM that the base pages are just
1880 * raw PFN mappings, and do not have a "struct page" associated
1881 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001882 *
1883 * There's a horrible special case to handle copy-on-write
1884 * behaviour that some programs depend on. We mark the "original"
1885 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07001887 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001888 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07001889 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07001890 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001891 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001892
Linus Torvalds6aab3412005-11-28 14:34:23 -08001893 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001895 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08001896 if (err) {
1897 /*
1898 * To indicate that track_pfn related cleanup is not
1899 * needed from higher level routine calling unmap_vmas
1900 */
1901 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07001902 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001903 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08001904 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001905
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 BUG_ON(addr >= end);
1907 pfn -= addr >> PAGE_SHIFT;
1908 pgd = pgd_offset(mm, addr);
1909 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 do {
1911 next = pgd_addr_end(addr, end);
1912 err = remap_pud_range(mm, pgd, addr, next,
1913 pfn + (addr >> PAGE_SHIFT), prot);
1914 if (err)
1915 break;
1916 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001917
1918 if (err)
1919 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1920
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 return err;
1922}
1923EXPORT_SYMBOL(remap_pfn_range);
1924
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001925static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1926 unsigned long addr, unsigned long end,
1927 pte_fn_t fn, void *data)
1928{
1929 pte_t *pte;
1930 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001931 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001932 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001933
1934 pte = (mm == &init_mm) ?
1935 pte_alloc_kernel(pmd, addr) :
1936 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1937 if (!pte)
1938 return -ENOMEM;
1939
1940 BUG_ON(pmd_huge(*pmd));
1941
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001942 arch_enter_lazy_mmu_mode();
1943
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001944 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001945
1946 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001947 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001948 if (err)
1949 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001950 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001951
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001952 arch_leave_lazy_mmu_mode();
1953
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001954 if (mm != &init_mm)
1955 pte_unmap_unlock(pte-1, ptl);
1956 return err;
1957}
1958
1959static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1960 unsigned long addr, unsigned long end,
1961 pte_fn_t fn, void *data)
1962{
1963 pmd_t *pmd;
1964 unsigned long next;
1965 int err;
1966
Andi Kleenceb86872008-07-23 21:27:50 -07001967 BUG_ON(pud_huge(*pud));
1968
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001969 pmd = pmd_alloc(mm, pud, addr);
1970 if (!pmd)
1971 return -ENOMEM;
1972 do {
1973 next = pmd_addr_end(addr, end);
1974 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1975 if (err)
1976 break;
1977 } while (pmd++, addr = next, addr != end);
1978 return err;
1979}
1980
1981static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1982 unsigned long addr, unsigned long end,
1983 pte_fn_t fn, void *data)
1984{
1985 pud_t *pud;
1986 unsigned long next;
1987 int err;
1988
1989 pud = pud_alloc(mm, pgd, addr);
1990 if (!pud)
1991 return -ENOMEM;
1992 do {
1993 next = pud_addr_end(addr, end);
1994 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1995 if (err)
1996 break;
1997 } while (pud++, addr = next, addr != end);
1998 return err;
1999}
2000
2001/*
2002 * Scan a region of virtual memory, filling in page tables as necessary
2003 * and calling a provided function on each leaf page table.
2004 */
2005int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2006 unsigned long size, pte_fn_t fn, void *data)
2007{
2008 pgd_t *pgd;
2009 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002010 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002011 int err;
2012
2013 BUG_ON(addr >= end);
2014 pgd = pgd_offset(mm, addr);
2015 do {
2016 next = pgd_addr_end(addr, end);
2017 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2018 if (err)
2019 break;
2020 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002021
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002022 return err;
2023}
2024EXPORT_SYMBOL_GPL(apply_to_page_range);
2025
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002027 * handle_pte_fault chooses page fault handler according to an entry
2028 * which was read non-atomically. Before making any commitment, on
2029 * those architectures or configurations (e.g. i386 with PAE) which
2030 * might give a mix of unmatched parts, do_swap_page and do_file_page
2031 * must check under lock before unmapping the pte and proceeding
2032 * (but do_wp_page is only called after already making such a check;
2033 * and do_anonymous_page and do_no_page can safely check later on).
2034 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002035static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002036 pte_t *page_table, pte_t orig_pte)
2037{
2038 int same = 1;
2039#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2040 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002041 spinlock_t *ptl = pte_lockptr(mm, pmd);
2042 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002043 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002044 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002045 }
2046#endif
2047 pte_unmap(page_table);
2048 return same;
2049}
2050
2051/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
2053 * servicing faults for write access. In the normal case, do always want
2054 * pte_mkwrite. But get_user_pages can cause write faults for mappings
2055 * that do not have writing enabled, when used by access_process_vm.
2056 */
2057static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
2058{
2059 if (likely(vma->vm_flags & VM_WRITE))
2060 pte = pte_mkwrite(pte);
2061 return pte;
2062}
2063
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002064static 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 -08002065{
2066 /*
2067 * If the source page was a PFN mapping, we don't have
2068 * a "struct page" for it. We do a best-effort copy by
2069 * just copying from the original user address. If that
2070 * fails, we just zero-fill it. Live with it.
2071 */
2072 if (unlikely(!src)) {
2073 void *kaddr = kmap_atomic(dst, KM_USER0);
Linus Torvalds5d2a2db2005-11-29 14:07:55 -08002074 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2075
2076 /*
2077 * This really shouldn't fail, because the page is there
2078 * in the page tables. But it might just be unreadable,
2079 * in which case we just give up and fill the result with
2080 * zeroes.
2081 */
2082 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Linus Torvalds6aab3412005-11-28 14:34:23 -08002083 memset(kaddr, 0, PAGE_SIZE);
2084 kunmap_atomic(kaddr, KM_USER0);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002085 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002086 } else
2087 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002088}
2089
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 * This routine handles present pages, when users try to write
2092 * to a shared page. It is done by copying the page to a new address
2093 * and decrementing the shared-page counter for the old page.
2094 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 * Note that this routine assumes that the protection checks have been
2096 * done by the caller (the low-level page fault routine in most cases).
2097 * Thus we can safely just mark it writable once we've done any necessary
2098 * COW.
2099 *
2100 * We also mark the page dirty at this point even though the page will
2101 * change only once the write actually happens. This avoids a few races,
2102 * and potentially makes it more efficient.
2103 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002104 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2105 * but allow concurrent faults), with pte both mapped and locked.
2106 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002108static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2109 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002110 spinlock_t *ptl, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002112 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 pte_t entry;
Nick Piggin83c54072007-07-19 01:47:05 -07002114 int reuse = 0, ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002115 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002116 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Linus Torvalds6aab3412005-11-28 14:34:23 -08002118 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002119 if (!old_page) {
2120 /*
2121 * VM_MIXEDMAP !pfn_valid() case
2122 *
2123 * We should not cow pages in a shared writeable mapping.
2124 * Just mark the pages writable as we can't do any dirty
2125 * accounting on raw pfn maps.
2126 */
2127 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2128 (VM_WRITE|VM_SHARED))
2129 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002130 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002131 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002133 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002134 * Take out anonymous pages first, anonymous shared vmas are
2135 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002136 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002137 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002138 if (!trylock_page(old_page)) {
2139 page_cache_get(old_page);
2140 pte_unmap_unlock(page_table, ptl);
2141 lock_page(old_page);
2142 page_table = pte_offset_map_lock(mm, pmd, address,
2143 &ptl);
2144 if (!pte_same(*page_table, orig_pte)) {
2145 unlock_page(old_page);
2146 page_cache_release(old_page);
2147 goto unlock;
2148 }
2149 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002150 }
Hugh Dickins7b1fe592009-01-06 14:39:34 -08002151 reuse = reuse_swap_page(old_page);
Rik van Rielc44b6742010-03-05 13:42:09 -08002152 if (reuse)
2153 /*
2154 * The page is all ours. Move it to our anon_vma so
2155 * the rmap code will not search our parent or siblings.
2156 * Protected against the rmap code by the page lock.
2157 */
2158 page_move_anon_rmap(old_page, vma, address);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002159 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002160 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002161 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002162 /*
2163 * Only catch write-faults on shared writable pages,
2164 * read-only shared pages can get COWed by
2165 * get_user_pages(.write=1, .force=1).
2166 */
David Howells9637a5e2006-06-23 02:03:43 -07002167 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002168 struct vm_fault vmf;
2169 int tmp;
2170
2171 vmf.virtual_address = (void __user *)(address &
2172 PAGE_MASK);
2173 vmf.pgoff = old_page->index;
2174 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2175 vmf.page = old_page;
2176
David Howells9637a5e2006-06-23 02:03:43 -07002177 /*
2178 * Notify the address space that the page is about to
2179 * become writable so that it can prohibit this or wait
2180 * for the page to get into an appropriate state.
2181 *
2182 * We do this without the lock held, so that it can
2183 * sleep if it needs to.
2184 */
2185 page_cache_get(old_page);
2186 pte_unmap_unlock(page_table, ptl);
2187
Nick Pigginc2ec1752009-03-31 15:23:21 -07002188 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2189 if (unlikely(tmp &
2190 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2191 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002192 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002193 }
Nick Pigginb827e492009-04-30 15:08:16 -07002194 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2195 lock_page(old_page);
2196 if (!old_page->mapping) {
2197 ret = 0; /* retry the fault */
2198 unlock_page(old_page);
2199 goto unwritable_page;
2200 }
2201 } else
2202 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002203
David Howells9637a5e2006-06-23 02:03:43 -07002204 /*
2205 * Since we dropped the lock we need to revalidate
2206 * the PTE as someone else may have changed it. If
2207 * they did, we just return, as we can count on the
2208 * MMU to tell us if they didn't also make it writable.
2209 */
2210 page_table = pte_offset_map_lock(mm, pmd, address,
2211 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002212 if (!pte_same(*page_table, orig_pte)) {
2213 unlock_page(old_page);
2214 page_cache_release(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002215 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002216 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002217
2218 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002220 dirty_page = old_page;
2221 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002222 reuse = 1;
David Howells9637a5e2006-06-23 02:03:43 -07002223 }
2224
2225 if (reuse) {
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002226reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002227 flush_cache_page(vma, address, pte_pfn(orig_pte));
2228 entry = pte_mkyoung(orig_pte);
2229 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002230 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002231 update_mmu_cache(vma, address, page_table);
David Howells9637a5e2006-06-23 02:03:43 -07002232 ret |= VM_FAULT_WRITE;
2233 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
2236 /*
2237 * Ok, we need to copy. Oh, well..
2238 */
Nick Pigginb5810032005-10-29 18:16:12 -07002239 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002240gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002241 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242
2243 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002244 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002245
Hugh Dickins62eede62009-09-21 17:03:34 -07002246 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002247 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2248 if (!new_page)
2249 goto oom;
2250 } else {
2251 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2252 if (!new_page)
2253 goto oom;
2254 cow_user_page(new_page, old_page, address, vma);
2255 }
2256 __SetPageUptodate(new_page);
2257
Nick Pigginb291f002008-10-18 20:26:44 -07002258 /*
2259 * Don't let another task, with possibly unlocked vma,
2260 * keep the mlocked page.
2261 */
Carsten Otteab926612009-02-04 15:12:16 -08002262 if ((vma->vm_flags & VM_LOCKED) && old_page) {
Nick Pigginb291f002008-10-18 20:26:44 -07002263 lock_page(old_page); /* for LRU manipulation */
2264 clear_page_mlock(old_page);
2265 unlock_page(old_page);
2266 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002267
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002268 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002269 goto oom_free_new;
2270
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 /*
2272 * Re-check the pte - we dropped the lock
2273 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002274 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002275 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002276 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002277 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002278 dec_mm_counter_fast(mm, MM_FILEPAGES);
2279 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002280 }
2281 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002282 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002283 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002284 entry = mk_pte(new_page, vma->vm_page_prot);
2285 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002286 /*
2287 * Clear the pte entry and flush it first, before updating the
2288 * pte with the new entry. This will avoid a race condition
2289 * seen in the presence of one thread doing SMC and another
2290 * thread doing COW.
2291 */
Izik Eidus828502d2009-09-21 17:01:51 -07002292 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002293 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002294 /*
2295 * We call the notify macro here because, when using secondary
2296 * mmu page tables (such as kvm shadow page tables), we want the
2297 * new page to be mapped directly into the secondary page table.
2298 */
2299 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002300 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002301 if (old_page) {
2302 /*
2303 * Only after switching the pte to the new page may
2304 * we remove the mapcount here. Otherwise another
2305 * process may come and find the rmap count decremented
2306 * before the pte is switched to the new page, and
2307 * "reuse" the old page writing into it while our pte
2308 * here still points into it and can be read by other
2309 * threads.
2310 *
2311 * The critical issue is to order this
2312 * page_remove_rmap with the ptp_clear_flush above.
2313 * Those stores are ordered by (if nothing else,)
2314 * the barrier present in the atomic_add_negative
2315 * in page_remove_rmap.
2316 *
2317 * Then the TLB flush in ptep_clear_flush ensures that
2318 * no process can access the old page before the
2319 * decremented mapcount is visible. And the old page
2320 * cannot be reused until after the decremented
2321 * mapcount is visible. So transitively, TLBs to
2322 * old page will be flushed before it can be reused.
2323 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002324 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002325 }
2326
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 /* Free the old page.. */
2328 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002329 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002330 } else
2331 mem_cgroup_uncharge_page(new_page);
2332
Hugh Dickins920fc352005-11-21 21:32:17 -08002333 if (new_page)
2334 page_cache_release(new_page);
2335 if (old_page)
2336 page_cache_release(old_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002337unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002338 pte_unmap_unlock(page_table, ptl);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002339 if (dirty_page) {
Nick Piggin79352892007-07-19 01:47:22 -07002340 /*
2341 * Yes, Virginia, this is actually required to prevent a race
2342 * with clear_page_dirty_for_io() from clearing the page dirty
2343 * bit after it clear all dirty ptes, but before a racing
2344 * do_wp_page installs a dirty pte.
2345 *
2346 * do_no_page is protected similarly.
2347 */
Nick Pigginb827e492009-04-30 15:08:16 -07002348 if (!page_mkwrite) {
2349 wait_on_page_locked(dirty_page);
2350 set_page_dirty_balance(dirty_page, page_mkwrite);
2351 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002352 put_page(dirty_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002353 if (page_mkwrite) {
2354 struct address_space *mapping = dirty_page->mapping;
2355
2356 set_page_dirty(dirty_page);
2357 unlock_page(dirty_page);
2358 page_cache_release(dirty_page);
2359 if (mapping) {
2360 /*
2361 * Some device drivers do not set page.mapping
2362 * but still dirty their pages
2363 */
2364 balance_dirty_pages_ratelimited(mapping);
2365 }
2366 }
2367
2368 /* file_update_time outside page_lock */
2369 if (vma->vm_file)
2370 file_update_time(vma->vm_file);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002371 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002372 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002373oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002374 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002375oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002376 if (old_page) {
2377 if (page_mkwrite) {
2378 unlock_page(old_page);
2379 page_cache_release(old_page);
2380 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002381 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002382 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002384
2385unwritable_page:
2386 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002387 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388}
2389
2390/*
2391 * Helper functions for unmap_mapping_range().
2392 *
2393 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2394 *
2395 * We have to restart searching the prio_tree whenever we drop the lock,
2396 * since the iterator is only valid while the lock is held, and anyway
2397 * a later vma might be split and reinserted earlier while lock dropped.
2398 *
2399 * The list of nonlinear vmas could be handled more efficiently, using
2400 * a placeholder, but handle it in the same way until a need is shown.
2401 * It is important to search the prio_tree before nonlinear list: a vma
2402 * may become nonlinear and be shifted from prio_tree to nonlinear list
2403 * while the lock is dropped; but never shifted from list to prio_tree.
2404 *
2405 * In order to make forward progress despite restarting the search,
2406 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2407 * quickly skip it next time around. Since the prio_tree search only
2408 * shows us those vmas affected by unmapping the range in question, we
2409 * can't efficiently keep all vmas in step with mapping->truncate_count:
2410 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2411 * mapping->truncate_count and vma->vm_truncate_count are protected by
2412 * i_mmap_lock.
2413 *
2414 * In order to make forward progress despite repeatedly restarting some
Hugh Dickinsee39b372005-04-19 13:29:15 -07002415 * large vma, note the restart_addr from unmap_vmas when it breaks out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 * and restart from that address when we reach that vma again. It might
2417 * have been split or merged, shrunk or extended, but never shifted: so
2418 * restart_addr remains valid so long as it remains in the vma's range.
2419 * unmap_mapping_range forces truncate_count to leap over page-aligned
2420 * values so we can save vma's restart_addr in its truncate_count field.
2421 */
2422#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2423
2424static void reset_vma_truncate_counts(struct address_space *mapping)
2425{
2426 struct vm_area_struct *vma;
2427 struct prio_tree_iter iter;
2428
2429 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2430 vma->vm_truncate_count = 0;
2431 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2432 vma->vm_truncate_count = 0;
2433}
2434
2435static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2436 unsigned long start_addr, unsigned long end_addr,
2437 struct zap_details *details)
2438{
2439 unsigned long restart_addr;
2440 int need_break;
2441
Nick Piggind00806b2007-07-19 01:46:57 -07002442 /*
2443 * files that support invalidating or truncating portions of the
Nick Piggind0217ac2007-07-19 01:47:03 -07002444 * file from under mmaped areas must have their ->fault function
Nick Piggin83c54072007-07-19 01:47:05 -07002445 * return a locked page (and set VM_FAULT_LOCKED in the return).
2446 * This provides synchronisation against concurrent unmapping here.
Nick Piggind00806b2007-07-19 01:46:57 -07002447 */
Nick Piggind00806b2007-07-19 01:46:57 -07002448
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449again:
2450 restart_addr = vma->vm_truncate_count;
2451 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2452 start_addr = restart_addr;
2453 if (start_addr >= end_addr) {
2454 /* Top of vma has been split off since last time */
2455 vma->vm_truncate_count = details->truncate_count;
2456 return 0;
2457 }
2458 }
2459
Hugh Dickinsee39b372005-04-19 13:29:15 -07002460 restart_addr = zap_page_range(vma, start_addr,
2461 end_addr - start_addr, details);
Nick Piggin95c354f2008-01-30 13:31:20 +01002462 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
Hugh Dickinsee39b372005-04-19 13:29:15 -07002464 if (restart_addr >= end_addr) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 /* We have now completed this vma: mark it so */
2466 vma->vm_truncate_count = details->truncate_count;
2467 if (!need_break)
2468 return 0;
2469 } else {
2470 /* Note restart_addr in vma's truncate_count field */
Hugh Dickinsee39b372005-04-19 13:29:15 -07002471 vma->vm_truncate_count = restart_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 if (!need_break)
2473 goto again;
2474 }
2475
2476 spin_unlock(details->i_mmap_lock);
2477 cond_resched();
2478 spin_lock(details->i_mmap_lock);
2479 return -EINTR;
2480}
2481
2482static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2483 struct zap_details *details)
2484{
2485 struct vm_area_struct *vma;
2486 struct prio_tree_iter iter;
2487 pgoff_t vba, vea, zba, zea;
2488
2489restart:
2490 vma_prio_tree_foreach(vma, &iter, root,
2491 details->first_index, details->last_index) {
2492 /* Skip quickly over those we have already dealt with */
2493 if (vma->vm_truncate_count == details->truncate_count)
2494 continue;
2495
2496 vba = vma->vm_pgoff;
2497 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2498 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2499 zba = details->first_index;
2500 if (zba < vba)
2501 zba = vba;
2502 zea = details->last_index;
2503 if (zea > vea)
2504 zea = vea;
2505
2506 if (unmap_mapping_range_vma(vma,
2507 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2508 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2509 details) < 0)
2510 goto restart;
2511 }
2512}
2513
2514static inline void unmap_mapping_range_list(struct list_head *head,
2515 struct zap_details *details)
2516{
2517 struct vm_area_struct *vma;
2518
2519 /*
2520 * In nonlinear VMAs there is no correspondence between virtual address
2521 * offset and file offset. So we must perform an exhaustive search
2522 * across *all* the pages in each nonlinear VMA, not just the pages
2523 * whose virtual address lies outside the file truncation point.
2524 */
2525restart:
2526 list_for_each_entry(vma, head, shared.vm_set.list) {
2527 /* Skip quickly over those we have already dealt with */
2528 if (vma->vm_truncate_count == details->truncate_count)
2529 continue;
2530 details->nonlinear_vma = vma;
2531 if (unmap_mapping_range_vma(vma, vma->vm_start,
2532 vma->vm_end, details) < 0)
2533 goto restart;
2534 }
2535}
2536
2537/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002538 * 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 -07002539 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 * @holebegin: byte in first page to unmap, relative to the start of
2541 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002542 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 * must keep the partial page. In contrast, we must get rid of
2544 * partial pages.
2545 * @holelen: size of prospective hole in bytes. This will be rounded
2546 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2547 * end of the file.
2548 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2549 * but 0 when invalidating pagecache, don't throw away private data.
2550 */
2551void unmap_mapping_range(struct address_space *mapping,
2552 loff_t const holebegin, loff_t const holelen, int even_cows)
2553{
2554 struct zap_details details;
2555 pgoff_t hba = holebegin >> PAGE_SHIFT;
2556 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2557
2558 /* Check for overflow. */
2559 if (sizeof(holelen) > sizeof(hlen)) {
2560 long long holeend =
2561 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2562 if (holeend & ~(long long)ULONG_MAX)
2563 hlen = ULONG_MAX - hba + 1;
2564 }
2565
2566 details.check_mapping = even_cows? NULL: mapping;
2567 details.nonlinear_vma = NULL;
2568 details.first_index = hba;
2569 details.last_index = hba + hlen - 1;
2570 if (details.last_index < details.first_index)
2571 details.last_index = ULONG_MAX;
2572 details.i_mmap_lock = &mapping->i_mmap_lock;
2573
2574 spin_lock(&mapping->i_mmap_lock);
2575
Nick Piggind00806b2007-07-19 01:46:57 -07002576 /* Protect against endless unmapping loops */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 mapping->truncate_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2579 if (mapping->truncate_count == 0)
2580 reset_vma_truncate_counts(mapping);
2581 mapping->truncate_count++;
2582 }
2583 details.truncate_count = mapping->truncate_count;
2584
2585 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2586 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2587 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2588 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2589 spin_unlock(&mapping->i_mmap_lock);
2590}
2591EXPORT_SYMBOL(unmap_mapping_range);
2592
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002593int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2594{
2595 struct address_space *mapping = inode->i_mapping;
2596
2597 /*
2598 * If the underlying filesystem is not going to provide
2599 * a way to truncate a range of blocks (punch a hole) -
2600 * we should return failure right now.
2601 */
Al Viroacfa4382008-12-04 10:06:33 -05002602 if (!inode->i_op->truncate_range)
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002603 return -ENOSYS;
2604
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002605 mutex_lock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002606 down_write(&inode->i_alloc_sem);
2607 unmap_mapping_range(mapping, offset, (end - offset), 1);
2608 truncate_inode_pages_range(mapping, offset, end);
Nick Piggind00806b2007-07-19 01:46:57 -07002609 unmap_mapping_range(mapping, offset, (end - offset), 1);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002610 inode->i_op->truncate_range(inode, offset, end);
2611 up_write(&inode->i_alloc_sem);
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002612 mutex_unlock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002613
2614 return 0;
2615}
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002618 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2619 * but allow concurrent faults), and pte mapped but not yet locked.
2620 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002622static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2623 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002624 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002626 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002627 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002628 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 pte_t pte;
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08002630 struct mem_cgroup *ptr = NULL;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002631 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002632 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002634 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002635 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002636
2637 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002638 if (unlikely(non_swap_entry(entry))) {
2639 if (is_migration_entry(entry)) {
2640 migration_entry_wait(mm, pmd, address);
2641 } else if (is_hwpoison_entry(entry)) {
2642 ret = VM_FAULT_HWPOISON;
2643 } else {
2644 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002645 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002646 }
Christoph Lameter06972122006-06-23 02:03:35 -07002647 goto out;
2648 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002649 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 page = lookup_swap_cache(entry);
2651 if (!page) {
Hugh Dickinsa5c9b692009-06-23 12:36:58 -07002652 grab_swap_token(mm); /* Contend for token _before_ read-in */
Hugh Dickins02098fe2008-02-04 22:28:42 -08002653 page = swapin_readahead(entry,
2654 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 if (!page) {
2656 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002657 * Back out if somebody else faulted in this pte
2658 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002660 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 if (likely(pte_same(*page_table, orig_pte)))
2662 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002663 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002664 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 }
2666
2667 /* Had to read the page from swap area: Major fault */
2668 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002669 count_vm_event(PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002670 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002671 /*
2672 * hwpoisoned dirty swapcache pages are kept for killing
2673 * owner processes (which may be unknown at hwpoison time)
2674 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002675 ret = VM_FAULT_HWPOISON;
2676 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002677 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 }
2679
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 lock_page(page);
Balbir Singh20a10222007-11-14 17:00:33 -08002681 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002683 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002684 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2685 * release the swapcache from under us. The page pin, and pte_same
2686 * test below, are not enough to exclude that. Even if it is still
2687 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002688 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002689 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002690 goto out_page;
2691
2692 if (ksm_might_need_to_copy(page, vma, address)) {
2693 swapcache = page;
2694 page = ksm_does_need_to_copy(page, vma, address);
2695
2696 if (unlikely(!page)) {
2697 ret = VM_FAULT_OOM;
2698 page = swapcache;
2699 swapcache = NULL;
2700 goto out_page;
2701 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002702 }
2703
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002704 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002705 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002706 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002707 }
2708
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002710 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002712 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002713 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002714 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002715
2716 if (unlikely(!PageUptodate(page))) {
2717 ret = VM_FAULT_SIGBUS;
2718 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 }
2720
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002721 /*
2722 * The page isn't present yet, go ahead with the fault.
2723 *
2724 * Be careful about the sequence of operations here.
2725 * To get its accounting right, reuse_swap_page() must be called
2726 * while the page is counted on swap but not yet in mapcount i.e.
2727 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2728 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002729 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2730 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2731 * in page->private. In this case, a record in swap_cgroup is silently
2732 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002733 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002735 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002736 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002738 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002740 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002741 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002742 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 flush_icache_page(vma, page);
2745 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002746 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002747 /* It's better to call commit-charge after rmap is established */
2748 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002750 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002751 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002752 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002753 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002754 if (swapcache) {
2755 /*
2756 * Hold the lock to avoid the swap entry to be reused
2757 * until we take the PT lock for the pte_same() check
2758 * (to avoid false positives from pte_same). For
2759 * further safety release the lock after the swap_free
2760 * so that the swap count won't change under a
2761 * parallel locked swapcache.
2762 */
2763 unlock_page(swapcache);
2764 page_cache_release(swapcache);
2765 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002766
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002767 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002768 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2769 if (ret & VM_FAULT_ERROR)
2770 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 goto out;
2772 }
2773
2774 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002775 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002776unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002777 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778out:
2779 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002780out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08002781 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002782 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002783out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002784 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002785out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002786 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002787 if (swapcache) {
2788 unlock_page(swapcache);
2789 page_cache_release(swapcache);
2790 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002791 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792}
2793
2794/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002795 * This is like a special single-page "expand_{down|up}wards()",
2796 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002797 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002798 */
2799static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2800{
2801 address &= PAGE_MASK;
2802 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002803 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002804
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002805 /*
2806 * Is there a mapping abutting this one below?
2807 *
2808 * That's only ok if it's the same stack mapping
2809 * that has gotten split..
2810 */
2811 if (prev && prev->vm_end == address)
2812 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2813
2814 expand_stack(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002815 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002816 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2817 struct vm_area_struct *next = vma->vm_next;
2818
2819 /* As VM_GROWSDOWN but s/below/above/ */
2820 if (next && next->vm_start == address + PAGE_SIZE)
2821 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2822
2823 expand_upwards(vma, address + PAGE_SIZE);
2824 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002825 return 0;
2826}
2827
2828/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002829 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2830 * but allow concurrent faults), and pte mapped but not yet locked.
2831 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002833static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2834 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002835 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002837 struct page *page;
2838 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840
Linus Torvalds11ac5522010-08-14 11:44:56 -07002841 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002842
Linus Torvalds11ac5522010-08-14 11:44:56 -07002843 /* Check if we need to add a guard page to the stack */
2844 if (check_stack_guard_page(vma, address) < 0)
2845 return VM_FAULT_SIGBUS;
2846
2847 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07002848 if (!(flags & FAULT_FLAG_WRITE)) {
2849 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2850 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002851 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002852 if (!pte_none(*page_table))
2853 goto unlock;
2854 goto setpte;
2855 }
2856
Nick Piggin557ed1f2007-10-16 01:24:40 -07002857 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002858 if (unlikely(anon_vma_prepare(vma)))
2859 goto oom;
2860 page = alloc_zeroed_user_highpage_movable(vma, address);
2861 if (!page)
2862 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08002863 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002865 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002866 goto oom_free_page;
2867
Nick Piggin557ed1f2007-10-16 01:24:40 -07002868 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002869 if (vma->vm_flags & VM_WRITE)
2870 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002871
Nick Piggin557ed1f2007-10-16 01:24:40 -07002872 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002873 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002874 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002875
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002876 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002877 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002878setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002879 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880
2881 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002882 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002883unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002884 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002885 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002886release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002887 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002888 page_cache_release(page);
2889 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002890oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002891 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002892oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 return VM_FAULT_OOM;
2894}
2895
2896/*
Nick Piggin54cb8822007-07-19 01:46:59 -07002897 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07002899 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2900 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 *
2902 * As this is called only for pages that do not currently exist, we
2903 * do not need to flush old virtual caches or the TLB.
2904 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002905 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01002906 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002907 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 */
Nick Piggin54cb8822007-07-19 01:46:59 -07002909static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01002910 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07002911 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912{
Hugh Dickins16abfa02007-10-04 16:56:06 +01002913 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002914 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07002915 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917 int anon = 0;
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002918 int charged = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002919 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07002920 struct vm_fault vmf;
2921 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002922 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07002923
Nick Piggind0217ac2007-07-19 01:47:03 -07002924 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2925 vmf.pgoff = pgoff;
2926 vmf.flags = flags;
2927 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928
Nick Piggin3c18ddd2008-04-28 02:12:10 -07002929 ret = vma->vm_ops->fault(vma, &vmf);
2930 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2931 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932
Andi Kleena3b947e2009-09-16 11:50:08 +02002933 if (unlikely(PageHWPoison(vmf.page))) {
2934 if (ret & VM_FAULT_LOCKED)
2935 unlock_page(vmf.page);
2936 return VM_FAULT_HWPOISON;
2937 }
2938
Nick Piggind00806b2007-07-19 01:46:57 -07002939 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07002940 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07002941 * locked.
2942 */
Nick Piggin83c54072007-07-19 01:47:05 -07002943 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07002944 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07002945 else
Nick Piggind0217ac2007-07-19 01:47:03 -07002946 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07002947
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 /*
2949 * Should we do an early C-O-W break?
2950 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002951 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07002952 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07002953 if (!(vma->vm_flags & VM_SHARED)) {
Nick Piggin54cb8822007-07-19 01:46:59 -07002954 anon = 1;
Nick Piggind00806b2007-07-19 01:46:57 -07002955 if (unlikely(anon_vma_prepare(vma))) {
Nick Piggind0217ac2007-07-19 01:47:03 -07002956 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07002957 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07002958 }
Nick Piggin83c54072007-07-19 01:47:05 -07002959 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2960 vma, address);
Nick Piggind00806b2007-07-19 01:46:57 -07002961 if (!page) {
Nick Piggind0217ac2007-07-19 01:47:03 -07002962 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07002963 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07002964 }
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002965 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002966 ret = VM_FAULT_OOM;
2967 page_cache_release(page);
2968 goto out;
2969 }
2970 charged = 1;
Nick Pigginb291f002008-10-18 20:26:44 -07002971 /*
2972 * Don't let another task, with possibly unlocked vma,
2973 * keep the mlocked page.
2974 */
2975 if (vma->vm_flags & VM_LOCKED)
2976 clear_page_mlock(vmf.page);
Nick Piggind0217ac2007-07-19 01:47:03 -07002977 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002978 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07002979 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07002980 /*
2981 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07002982 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07002983 * to become writable
2984 */
Mark Fasheh69676142007-07-19 01:47:00 -07002985 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002986 int tmp;
2987
Mark Fasheh69676142007-07-19 01:47:00 -07002988 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07002989 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002990 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2991 if (unlikely(tmp &
2992 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2993 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07002994 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07002995 }
Nick Pigginb827e492009-04-30 15:08:16 -07002996 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2997 lock_page(page);
2998 if (!page->mapping) {
2999 ret = 0; /* retry the fault */
3000 unlock_page(page);
3001 goto unwritable_page;
3002 }
3003 } else
3004 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003005 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003006 }
3007 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003008
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 }
3010
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003011 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
3013 /*
3014 * This silly early PAGE_DIRTY setting removes a race
3015 * due to the bad i386 page protection. But it's valid
3016 * for other architectures too.
3017 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003018 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 * an exclusive copy of the page, or this is a shared mapping,
3020 * so we can make it writable and dirty to avoid having to
3021 * handle that later.
3022 */
3023 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003024 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003025 flush_icache_page(vma, page);
3026 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003027 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003030 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003031 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003032 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003033 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003034 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003035 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003036 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003037 get_page(dirty_page);
3038 }
Hugh Dickins42946212005-10-29 18:16:05 -07003039 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003040 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003041
3042 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003043 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 } else {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003045 if (charged)
3046 mem_cgroup_uncharge_page(page);
Nick Piggind00806b2007-07-19 01:46:57 -07003047 if (anon)
3048 page_cache_release(page);
3049 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003050 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 }
3052
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003053 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003054
3055out:
Nick Pigginb827e492009-04-30 15:08:16 -07003056 if (dirty_page) {
3057 struct address_space *mapping = page->mapping;
3058
3059 if (set_page_dirty(dirty_page))
3060 page_mkwrite = 1;
3061 unlock_page(dirty_page);
3062 put_page(dirty_page);
3063 if (page_mkwrite && mapping) {
3064 /*
3065 * Some device drivers do not set page.mapping but still
3066 * dirty their pages
3067 */
3068 balance_dirty_pages_ratelimited(mapping);
3069 }
3070
3071 /* file_update_time outside page_lock */
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003072 if (vma->vm_file)
3073 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003074 } else {
3075 unlock_page(vmf.page);
3076 if (anon)
3077 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003078 }
Nick Piggind00806b2007-07-19 01:46:57 -07003079
Nick Piggin83c54072007-07-19 01:47:05 -07003080 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003081
3082unwritable_page:
3083 page_cache_release(page);
3084 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003085}
Nick Piggind00806b2007-07-19 01:46:57 -07003086
Nick Piggin54cb8822007-07-19 01:46:59 -07003087static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3088 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003089 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003090{
3091 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003092 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003093
Hugh Dickins16abfa02007-10-04 16:56:06 +01003094 pte_unmap(page_table);
3095 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003096}
3097
Jes Sorensenf4b81802006-09-27 01:50:10 -07003098/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 * Fault of a previously existing named mapping. Repopulate the pte
3100 * from the encoded file_pte if possible. This enables swappable
3101 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003102 *
3103 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3104 * but allow concurrent faults), and pte mapped but not yet locked.
3105 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003107static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003108 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003109 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110{
Hugh Dickins65500d22005-10-29 18:15:59 -07003111 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003113 flags |= FAULT_FLAG_NONLINEAR;
3114
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003115 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003116 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117
Hugh Dickins2509ef22009-01-06 14:40:10 -08003118 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003119 /*
3120 * Page table corrupted: show pte and kill process.
3121 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003122 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003123 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003124 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003125
3126 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003127 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128}
3129
3130/*
3131 * These routines also need to handle stuff like marking pages dirty
3132 * and/or accessed for architectures that don't do it in hardware (most
3133 * RISC architectures). The early dirtying is also good on the i386.
3134 *
3135 * There is also a hook called "update_mmu_cache()" that architectures
3136 * with external mmu caches can use to update those (ie the Sparc or
3137 * PowerPC hashed page tables that act as extended TLBs).
3138 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003139 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3140 * but allow concurrent faults), and pte mapped but not yet locked.
3141 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142 */
3143static inline int handle_pte_fault(struct mm_struct *mm,
Hugh Dickins65500d22005-10-29 18:15:59 -07003144 struct vm_area_struct *vma, unsigned long address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003145 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146{
3147 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003148 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003150 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003152 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003153 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003154 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003155 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003156 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003157 }
3158 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003159 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003160 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003162 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003163 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003164 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003165 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 }
3167
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003168 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003169 spin_lock(ptl);
3170 if (unlikely(!pte_same(*pte, entry)))
3171 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003172 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003174 return do_wp_page(mm, vma, address,
3175 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 entry = pte_mkdirty(entry);
3177 }
3178 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003179 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003180 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003181 } else {
3182 /*
3183 * This is needed only for protection faults but the arch code
3184 * is not yet telling us if this is a protection fault or not.
3185 * This still avoids useless tlb flushes for .text page faults
3186 * with threads.
3187 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003188 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003189 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003190 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003191unlock:
3192 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003193 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194}
3195
3196/*
3197 * By the time we get here, we already hold the mm semaphore
3198 */
Nick Piggin83c54072007-07-19 01:47:05 -07003199int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003200 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201{
3202 pgd_t *pgd;
3203 pud_t *pud;
3204 pmd_t *pmd;
3205 pte_t *pte;
3206
3207 __set_current_state(TASK_RUNNING);
3208
Christoph Lameterf8891e52006-06-30 01:55:45 -07003209 count_vm_event(PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003211 /* do counter updates before entering really critical section. */
3212 check_sync_rss_stat(current);
3213
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003214 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003215 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 pud = pud_alloc(mm, pgd, address);
3219 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003220 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 pmd = pmd_alloc(mm, pud, address);
3222 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003223 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224 pte = pte_alloc_map(mm, pmd, address);
3225 if (!pte)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003226 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003228 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229}
3230
3231#ifndef __PAGETABLE_PUD_FOLDED
3232/*
3233 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003234 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003236int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003238 pud_t *new = pud_alloc_one(mm, address);
3239 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003240 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
Nick Piggin362a61a2008-05-14 06:37:36 +02003242 smp_wmb(); /* See comment in __pte_alloc */
3243
Hugh Dickins872fec12005-10-29 18:16:21 -07003244 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003245 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003246 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003247 else
3248 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003249 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003250 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251}
3252#endif /* __PAGETABLE_PUD_FOLDED */
3253
3254#ifndef __PAGETABLE_PMD_FOLDED
3255/*
3256 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003257 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003259int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003261 pmd_t *new = pmd_alloc_one(mm, address);
3262 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003263 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264
Nick Piggin362a61a2008-05-14 06:37:36 +02003265 smp_wmb(); /* See comment in __pte_alloc */
3266
Hugh Dickins872fec12005-10-29 18:16:21 -07003267 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003269 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003270 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003271 else
3272 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003274 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003275 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003276 else
3277 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003279 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003280 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281}
3282#endif /* __PAGETABLE_PMD_FOLDED */
3283
3284int make_pages_present(unsigned long addr, unsigned long end)
3285{
3286 int ret, len, write;
3287 struct vm_area_struct * vma;
3288
3289 vma = find_vma(current->mm, addr);
3290 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003291 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 write = (vma->vm_flags & VM_WRITE) != 0;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003293 BUG_ON(addr >= end);
3294 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003295 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 ret = get_user_pages(current, current->mm, addr,
3297 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003298 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003300 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301}
3302
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303#if !defined(__HAVE_ARCH_GATE_AREA)
3304
3305#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003306static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307
3308static int __init gate_vma_init(void)
3309{
3310 gate_vma.vm_mm = NULL;
3311 gate_vma.vm_start = FIXADDR_USER_START;
3312 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003313 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3314 gate_vma.vm_page_prot = __P101;
Roland McGrathf47aef52007-01-26 00:56:49 -08003315 /*
3316 * Make sure the vDSO gets into every core dump.
3317 * Dumping its contents makes post-mortem fully interpretable later
3318 * without matching up the same kernel and hardware config to see
3319 * what PC values meant.
3320 */
3321 gate_vma.vm_flags |= VM_ALWAYSDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 return 0;
3323}
3324__initcall(gate_vma_init);
3325#endif
3326
3327struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3328{
3329#ifdef AT_SYSINFO_EHDR
3330 return &gate_vma;
3331#else
3332 return NULL;
3333#endif
3334}
3335
3336int in_gate_area_no_task(unsigned long addr)
3337{
3338#ifdef AT_SYSINFO_EHDR
3339 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3340 return 1;
3341#endif
3342 return 0;
3343}
3344
3345#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003346
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003347static int follow_pte(struct mm_struct *mm, unsigned long address,
3348 pte_t **ptepp, spinlock_t **ptlp)
3349{
3350 pgd_t *pgd;
3351 pud_t *pud;
3352 pmd_t *pmd;
3353 pte_t *ptep;
3354
3355 pgd = pgd_offset(mm, address);
3356 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3357 goto out;
3358
3359 pud = pud_offset(pgd, address);
3360 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3361 goto out;
3362
3363 pmd = pmd_offset(pud, address);
3364 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3365 goto out;
3366
3367 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3368 if (pmd_huge(*pmd))
3369 goto out;
3370
3371 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3372 if (!ptep)
3373 goto out;
3374 if (!pte_present(*ptep))
3375 goto unlock;
3376 *ptepp = ptep;
3377 return 0;
3378unlock:
3379 pte_unmap_unlock(ptep, *ptlp);
3380out:
3381 return -EINVAL;
3382}
3383
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003384/**
3385 * follow_pfn - look up PFN at a user virtual address
3386 * @vma: memory mapping
3387 * @address: user virtual address
3388 * @pfn: location to store found PFN
3389 *
3390 * Only IO mappings and raw PFN mappings are allowed.
3391 *
3392 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3393 */
3394int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3395 unsigned long *pfn)
3396{
3397 int ret = -EINVAL;
3398 spinlock_t *ptl;
3399 pte_t *ptep;
3400
3401 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3402 return ret;
3403
3404 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3405 if (ret)
3406 return ret;
3407 *pfn = pte_pfn(*ptep);
3408 pte_unmap_unlock(ptep, ptl);
3409 return 0;
3410}
3411EXPORT_SYMBOL(follow_pfn);
3412
Rik van Riel28b2ee22008-07-23 21:27:05 -07003413#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003414int follow_phys(struct vm_area_struct *vma,
3415 unsigned long address, unsigned int flags,
3416 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003417{
Johannes Weiner03668a42009-06-16 15:32:34 -07003418 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003419 pte_t *ptep, pte;
3420 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003421
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003422 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3423 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003424
Johannes Weiner03668a42009-06-16 15:32:34 -07003425 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003426 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003427 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003428
Rik van Riel28b2ee22008-07-23 21:27:05 -07003429 if ((flags & FOLL_WRITE) && !pte_write(pte))
3430 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003431
3432 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003433 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003434
Johannes Weiner03668a42009-06-16 15:32:34 -07003435 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003436unlock:
3437 pte_unmap_unlock(ptep, ptl);
3438out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003439 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003440}
3441
3442int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3443 void *buf, int len, int write)
3444{
3445 resource_size_t phys_addr;
3446 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003447 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003448 int offset = addr & (PAGE_SIZE-1);
3449
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003450 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003451 return -EINVAL;
3452
3453 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3454 if (write)
3455 memcpy_toio(maddr + offset, buf, len);
3456 else
3457 memcpy_fromio(buf, maddr + offset, len);
3458 iounmap(maddr);
3459
3460 return len;
3461}
3462#endif
3463
David Howells0ec76a12006-09-27 01:50:15 -07003464/*
3465 * Access another process' address space.
3466 * Source/target buffer must be kernel space,
3467 * Do not walk the page table directly, use get_user_pages
3468 */
3469int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3470{
3471 struct mm_struct *mm;
3472 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003473 void *old_buf = buf;
3474
3475 mm = get_task_mm(tsk);
3476 if (!mm)
3477 return 0;
3478
3479 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003480 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003481 while (len) {
3482 int bytes, ret, offset;
3483 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003484 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003485
3486 ret = get_user_pages(tsk, mm, addr, 1,
3487 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003488 if (ret <= 0) {
3489 /*
3490 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3491 * we can access using slightly different code.
3492 */
3493#ifdef CONFIG_HAVE_IOREMAP_PROT
3494 vma = find_vma(mm, addr);
3495 if (!vma)
3496 break;
3497 if (vma->vm_ops && vma->vm_ops->access)
3498 ret = vma->vm_ops->access(vma, addr, buf,
3499 len, write);
3500 if (ret <= 0)
3501#endif
3502 break;
3503 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003504 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003505 bytes = len;
3506 offset = addr & (PAGE_SIZE-1);
3507 if (bytes > PAGE_SIZE-offset)
3508 bytes = PAGE_SIZE-offset;
3509
3510 maddr = kmap(page);
3511 if (write) {
3512 copy_to_user_page(vma, page, addr,
3513 maddr + offset, buf, bytes);
3514 set_page_dirty_lock(page);
3515 } else {
3516 copy_from_user_page(vma, page, addr,
3517 buf, maddr + offset, bytes);
3518 }
3519 kunmap(page);
3520 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003521 }
David Howells0ec76a12006-09-27 01:50:15 -07003522 len -= bytes;
3523 buf += bytes;
3524 addr += bytes;
3525 }
3526 up_read(&mm->mmap_sem);
3527 mmput(mm);
3528
3529 return buf - old_buf;
3530}
Andi Kleen03252912008-01-30 13:33:18 +01003531
3532/*
3533 * Print the name of a VMA.
3534 */
3535void print_vma_addr(char *prefix, unsigned long ip)
3536{
3537 struct mm_struct *mm = current->mm;
3538 struct vm_area_struct *vma;
3539
Ingo Molnare8bff742008-02-13 20:21:06 +01003540 /*
3541 * Do not print if we are in atomic
3542 * contexts (in exception stacks, etc.):
3543 */
3544 if (preempt_count())
3545 return;
3546
Andi Kleen03252912008-01-30 13:33:18 +01003547 down_read(&mm->mmap_sem);
3548 vma = find_vma(mm, ip);
3549 if (vma && vma->vm_file) {
3550 struct file *f = vma->vm_file;
3551 char *buf = (char *)__get_free_page(GFP_KERNEL);
3552 if (buf) {
3553 char *p, *s;
3554
Jan Blunckcf28b482008-02-14 19:38:44 -08003555 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003556 if (IS_ERR(p))
3557 p = "?";
3558 s = strrchr(p, '/');
3559 if (s)
3560 p = s+1;
3561 printk("%s%s[%lx+%lx]", prefix, p,
3562 vma->vm_start,
3563 vma->vm_end - vma->vm_start);
3564 free_page((unsigned long)buf);
3565 }
3566 }
3567 up_read(&current->mm->mmap_sem);
3568}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003569
3570#ifdef CONFIG_PROVE_LOCKING
3571void might_fault(void)
3572{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003573 /*
3574 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3575 * holding the mmap_sem, this is safe because kernel memory doesn't
3576 * get paged out, therefore we'll never actually fault, and the
3577 * below annotations will generate false positives.
3578 */
3579 if (segment_eq(get_fs(), KERNEL_DS))
3580 return;
3581
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003582 might_sleep();
3583 /*
3584 * it would be nicer only to annotate paths which are not under
3585 * pagefault_disable, however that requires a larger audit and
3586 * providing helpers like get_user_atomic.
3587 */
3588 if (!in_atomic() && current->mm)
3589 might_lock_read(&current->mm->mmap_sem);
3590}
3591EXPORT_SYMBOL(might_fault);
3592#endif