blob: 2466d1250231f3e2405429ea4de4a97c597d0074 [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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040050#include <linux/export.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
David Rientjesea48cf72012-03-21 16:34:13 -0700128void sync_mm_rss(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++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700133 if (current->rss_stat.count[i]) {
134 add_mm_counter(mm, i, current->rss_stat.count[i]);
135 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800136 }
137 }
David Rientjes05af2e12012-03-21 16:34:13 -0700138 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800139}
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))
David Rientjesea48cf72012-03-21 16:34:13 -0700160 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800161}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700162#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800163
164#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
165#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
166
167static void check_sync_rss_stat(struct task_struct *task)
168{
169}
170
Peter Zijlstra9547d012011-05-24 17:12:14 -0700171#endif /* SPLIT_RSS_COUNTING */
172
173#ifdef HAVE_GENERIC_MMU_GATHER
174
175static int tlb_next_batch(struct mmu_gather *tlb)
176{
177 struct mmu_gather_batch *batch;
178
179 batch = tlb->active;
180 if (batch->next) {
181 tlb->active = batch->next;
182 return 1;
183 }
184
185 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
186 if (!batch)
187 return 0;
188
189 batch->next = NULL;
190 batch->nr = 0;
191 batch->max = MAX_GATHER_BATCH;
192
193 tlb->active->next = batch;
194 tlb->active = batch;
195
196 return 1;
197}
198
199/* tlb_gather_mmu
200 * Called to initialize an (on-stack) mmu_gather structure for page-table
201 * tear-down from @mm. The @fullmm argument is used when @mm is without
202 * users and we're going to destroy the full address space (exit/execve).
203 */
204void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
205{
206 tlb->mm = mm;
207
208 tlb->fullmm = fullmm;
209 tlb->need_flush = 0;
210 tlb->fast_mode = (num_possible_cpus() == 1);
211 tlb->local.next = NULL;
212 tlb->local.nr = 0;
213 tlb->local.max = ARRAY_SIZE(tlb->__pages);
214 tlb->active = &tlb->local;
215
216#ifdef CONFIG_HAVE_RCU_TABLE_FREE
217 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800218#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700219}
220
221void tlb_flush_mmu(struct mmu_gather *tlb)
222{
223 struct mmu_gather_batch *batch;
224
225 if (!tlb->need_flush)
226 return;
227 tlb->need_flush = 0;
228 tlb_flush(tlb);
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb_table_flush(tlb);
231#endif
232
233 if (tlb_fast_mode(tlb))
234 return;
235
236 for (batch = &tlb->local; batch; batch = batch->next) {
237 free_pages_and_swap_cache(batch->pages, batch->nr);
238 batch->nr = 0;
239 }
240 tlb->active = &tlb->local;
241}
242
243/* tlb_finish_mmu
244 * Called at the end of the shootdown operation to free up any resources
245 * that were required.
246 */
247void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
248{
249 struct mmu_gather_batch *batch, *next;
250
251 tlb_flush_mmu(tlb);
252
253 /* keep the page table cache within bounds */
254 check_pgt_cache();
255
256 for (batch = tlb->local.next; batch; batch = next) {
257 next = batch->next;
258 free_pages((unsigned long)batch, 0);
259 }
260 tlb->local.next = NULL;
261}
262
263/* __tlb_remove_page
264 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
265 * handling the additional races in SMP caused by other CPUs caching valid
266 * mappings in their TLBs. Returns the number of free page slots left.
267 * When out of page slots we must call tlb_flush_mmu().
268 */
269int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
270{
271 struct mmu_gather_batch *batch;
272
Shaohua Lif21760b2012-01-12 17:19:16 -0800273 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700274
275 if (tlb_fast_mode(tlb)) {
276 free_page_and_swap_cache(page);
277 return 1; /* avoid calling tlb_flush_mmu() */
278 }
279
280 batch = tlb->active;
281 batch->pages[batch->nr++] = page;
282 if (batch->nr == batch->max) {
283 if (!tlb_next_batch(tlb))
284 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700285 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700286 }
287 VM_BUG_ON(batch->nr > batch->max);
288
289 return batch->max - batch->nr;
290}
291
292#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800293
Peter Zijlstra26723912011-05-24 17:12:00 -0700294#ifdef CONFIG_HAVE_RCU_TABLE_FREE
295
296/*
297 * See the comment near struct mmu_table_batch.
298 */
299
300static void tlb_remove_table_smp_sync(void *arg)
301{
302 /* Simply deliver the interrupt */
303}
304
305static void tlb_remove_table_one(void *table)
306{
307 /*
308 * This isn't an RCU grace period and hence the page-tables cannot be
309 * assumed to be actually RCU-freed.
310 *
311 * It is however sufficient for software page-table walkers that rely on
312 * IRQ disabling. See the comment near struct mmu_table_batch.
313 */
314 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
315 __tlb_remove_table(table);
316}
317
318static void tlb_remove_table_rcu(struct rcu_head *head)
319{
320 struct mmu_table_batch *batch;
321 int i;
322
323 batch = container_of(head, struct mmu_table_batch, rcu);
324
325 for (i = 0; i < batch->nr; i++)
326 __tlb_remove_table(batch->tables[i]);
327
328 free_page((unsigned long)batch);
329}
330
331void tlb_table_flush(struct mmu_gather *tlb)
332{
333 struct mmu_table_batch **batch = &tlb->batch;
334
335 if (*batch) {
336 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
337 *batch = NULL;
338 }
339}
340
341void tlb_remove_table(struct mmu_gather *tlb, void *table)
342{
343 struct mmu_table_batch **batch = &tlb->batch;
344
345 tlb->need_flush = 1;
346
347 /*
348 * When there's less then two users of this mm there cannot be a
349 * concurrent page-table walk.
350 */
351 if (atomic_read(&tlb->mm->mm_users) < 2) {
352 __tlb_remove_table(table);
353 return;
354 }
355
356 if (*batch == NULL) {
357 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
358 if (*batch == NULL) {
359 tlb_remove_table_one(table);
360 return;
361 }
362 (*batch)->nr = 0;
363 }
364 (*batch)->tables[(*batch)->nr++] = table;
365 if ((*batch)->nr == MAX_TABLE_BATCH)
366 tlb_table_flush(tlb);
367}
368
Peter Zijlstra9547d012011-05-24 17:12:14 -0700369#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700370
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371/*
372 * If a p?d_bad entry is found while walking page tables, report
373 * the error, before resetting entry to p?d_none. Usually (but
374 * very seldom) called out from the p?d_none_or_clear_bad macros.
375 */
376
377void pgd_clear_bad(pgd_t *pgd)
378{
379 pgd_ERROR(*pgd);
380 pgd_clear(pgd);
381}
382
383void pud_clear_bad(pud_t *pud)
384{
385 pud_ERROR(*pud);
386 pud_clear(pud);
387}
388
389void pmd_clear_bad(pmd_t *pmd)
390{
391 pmd_ERROR(*pmd);
392 pmd_clear(pmd);
393}
394
395/*
396 * Note: this doesn't free the actual pages themselves. That
397 * has been handled earlier when unmapping all the memory regions.
398 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000399static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
400 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800402 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700403 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000404 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700405 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406}
407
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
409 unsigned long addr, unsigned long end,
410 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
412 pmd_t *pmd;
413 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700414 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Hugh Dickinse0da3822005-04-19 13:29:15 -0700416 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 do {
419 next = pmd_addr_end(addr, end);
420 if (pmd_none_or_clear_bad(pmd))
421 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000422 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 } while (pmd++, addr = next, addr != end);
424
Hugh Dickinse0da3822005-04-19 13:29:15 -0700425 start &= PUD_MASK;
426 if (start < floor)
427 return;
428 if (ceiling) {
429 ceiling &= PUD_MASK;
430 if (!ceiling)
431 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700433 if (end - 1 > ceiling - 1)
434 return;
435
436 pmd = pmd_offset(pud, start);
437 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000438 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439}
440
Hugh Dickinse0da3822005-04-19 13:29:15 -0700441static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444{
445 pud_t *pud;
446 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
Hugh Dickinse0da3822005-04-19 13:29:15 -0700449 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 do {
452 next = pud_addr_end(addr, end);
453 if (pud_none_or_clear_bad(pud))
454 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 } while (pud++, addr = next, addr != end);
457
Hugh Dickinse0da3822005-04-19 13:29:15 -0700458 start &= PGDIR_MASK;
459 if (start < floor)
460 return;
461 if (ceiling) {
462 ceiling &= PGDIR_MASK;
463 if (!ceiling)
464 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700466 if (end - 1 > ceiling - 1)
467 return;
468
469 pud = pud_offset(pgd, start);
470 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000471 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472}
473
474/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700475 * This function frees user-level page tables of a process.
476 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 * Must be called with pagetable lock held.
478 */
Jan Beulich42b77722008-07-23 21:27:10 -0700479void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700480 unsigned long addr, unsigned long end,
481 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482{
483 pgd_t *pgd;
484 unsigned long next;
485
Hugh Dickinse0da3822005-04-19 13:29:15 -0700486 /*
487 * The next few lines have given us lots of grief...
488 *
489 * Why are we testing PMD* at this top level? Because often
490 * there will be no work to do at all, and we'd prefer not to
491 * go all the way down to the bottom just to discover that.
492 *
493 * Why all these "- 1"s? Because 0 represents both the bottom
494 * of the address space and the top of it (using -1 for the
495 * top wouldn't help much: the masks would do the wrong thing).
496 * The rule is that addr 0 and floor 0 refer to the bottom of
497 * the address space, but end 0 and ceiling 0 refer to the top
498 * Comparisons need to use "end - 1" and "ceiling - 1" (though
499 * that end 0 case should be mythical).
500 *
501 * Wherever addr is brought up or ceiling brought down, we must
502 * be careful to reject "the opposite 0" before it confuses the
503 * subsequent tests. But what about where end is brought down
504 * by PMD_SIZE below? no, end can't go down to 0 there.
505 *
506 * Whereas we round start (addr) and ceiling down, by different
507 * masks at different levels, in order to test whether a table
508 * now has no other vmas using it, so can be freed, we don't
509 * bother to round floor or end up - the tests don't need that.
510 */
511
512 addr &= PMD_MASK;
513 if (addr < floor) {
514 addr += PMD_SIZE;
515 if (!addr)
516 return;
517 }
518 if (ceiling) {
519 ceiling &= PMD_MASK;
520 if (!ceiling)
521 return;
522 }
523 if (end - 1 > ceiling - 1)
524 end -= PMD_SIZE;
525 if (addr > end - 1)
526 return;
527
Jan Beulich42b77722008-07-23 21:27:10 -0700528 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 do {
530 next = pgd_addr_end(addr, end);
531 if (pgd_none_or_clear_bad(pgd))
532 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700533 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700535}
536
Jan Beulich42b77722008-07-23 21:27:10 -0700537void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700538 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700539{
540 while (vma) {
541 struct vm_area_struct *next = vma->vm_next;
542 unsigned long addr = vma->vm_start;
543
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700544 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000545 * Hide vma from rmap and truncate_pagecache before freeing
546 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700547 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800548 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700549 unlink_file_vma(vma);
550
David Gibson9da61ae2006-03-22 00:08:57 -0800551 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700552 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700553 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700554 } else {
555 /*
556 * Optimization: gather nearby vmas into one call down
557 */
558 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800559 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700560 vma = next;
561 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800562 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700563 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700564 }
565 free_pgd_range(tlb, addr, vma->vm_end,
566 floor, next? next->vm_start: ceiling);
567 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700568 vma = next;
569 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800572int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
573 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800575 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800576 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700577 if (!new)
578 return -ENOMEM;
579
Nick Piggin362a61a2008-05-14 06:37:36 +0200580 /*
581 * Ensure all pte setup (eg. pte page lock and page clearing) are
582 * visible before the pte is made visible to other CPUs by being
583 * put into page tables.
584 *
585 * The other side of the story is the pointer chasing in the page
586 * table walking code (when walking the page table without locking;
587 * ie. most of the time). Fortunately, these data accesses consist
588 * of a chain of data-dependent loads, meaning most CPUs (alpha
589 * being the notable exception) will already guarantee loads are
590 * seen in-order. See the alpha page table accessors for the
591 * smp_read_barrier_depends() barriers in page table walking code.
592 */
593 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
594
Hugh Dickinsc74df322005-10-29 18:16:23 -0700595 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800596 wait_split_huge_page = 0;
597 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800600 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800601 } else if (unlikely(pmd_trans_splitting(*pmd)))
602 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700603 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800604 if (new)
605 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800606 if (wait_split_huge_page)
607 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700608 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609}
610
Hugh Dickins1bb36302005-10-29 18:16:22 -0700611int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
614 if (!new)
615 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
Nick Piggin362a61a2008-05-14 06:37:36 +0200617 smp_wmb(); /* See comment in __pte_alloc */
618
Hugh Dickins1bb36302005-10-29 18:16:22 -0700619 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700621 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800622 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800623 } else
624 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700625 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800626 if (new)
627 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700628 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629}
630
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800631static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700632{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634}
635
636static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
637{
638 int i;
639
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800640 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700641 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800642 for (i = 0; i < NR_MM_COUNTERS; i++)
643 if (rss[i])
644 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700645}
646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700651 *
652 * The calling function must still handle the error.
653 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800654static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700656{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658 pud_t *pud = pud_offset(pgd, addr);
659 pmd_t *pmd = pmd_offset(pud, addr);
660 struct address_space *mapping;
661 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800662 static unsigned long resume;
663 static unsigned long nr_shown;
664 static unsigned long nr_unshown;
665
666 /*
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
669 */
670 if (nr_shown == 60) {
671 if (time_before(jiffies, resume)) {
672 nr_unshown++;
673 return;
674 }
675 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800676 printk(KERN_ALERT
677 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800678 nr_unshown);
679 nr_unshown = 0;
680 }
681 nr_shown = 0;
682 }
683 if (nr_shown++ == 0)
684 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800685
686 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
687 index = linear_page_index(vma, addr);
688
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800689 printk(KERN_ALERT
690 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800691 current->comm,
692 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800693 if (page)
694 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800695 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800696 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
698 /*
699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
700 */
701 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800702 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800703 (unsigned long)vma->vm_ops->fault);
704 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800705 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800706 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700707 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800708 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700709}
710
KOSAKI Motohiroca16d142011-05-26 19:16:19 +0900711static inline int is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800712{
713 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
714}
715
Hugh Dickins62eede62009-09-21 17:03:34 -0700716#ifndef is_zero_pfn
717static inline int is_zero_pfn(unsigned long pfn)
718{
719 return pfn == zero_pfn;
720}
721#endif
722
723#ifndef my_zero_pfn
724static inline unsigned long my_zero_pfn(unsigned long addr)
725{
726 return zero_pfn;
727}
728#endif
729
Nick Pigginb5810032005-10-29 18:16:12 -0700730/*
Nick Piggin7e675132008-04-28 02:13:00 -0700731 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800732 *
Nick Piggin7e675132008-04-28 02:13:00 -0700733 * "Special" mappings do not wish to be associated with a "struct page" (either
734 * it doesn't exist, or it exists but they don't want to touch it). In this
735 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700736 *
Nick Piggin7e675132008-04-28 02:13:00 -0700737 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
738 * pte bit, in which case this function is trivial. Secondly, an architecture
739 * may not have a spare pte bit, which requires a more complicated scheme,
740 * described below.
741 *
742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
743 * special mapping (even if there are underlying and valid "struct pages").
744 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800745 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
749 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800750 *
751 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
752 *
Nick Piggin7e675132008-04-28 02:13:00 -0700753 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700754 *
Nick Piggin7e675132008-04-28 02:13:00 -0700755 * This restricts such mappings to be a linear translation from virtual address
756 * to pfn. To get around this restriction, we allow arbitrary mappings so long
757 * as the vma is not a COW mapping; in that case, we know that all ptes are
758 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700759 *
760 *
Nick Piggin7e675132008-04-28 02:13:00 -0700761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
762 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
764 * page" backing, however the difference is that _all_ pages with a struct
765 * page (that is, those where pfn_valid is true) are refcounted and considered
766 * normal pages by the VM. The disadvantage is that pages are refcounted
767 * (which can be slower and simply not an option for some PFNMAP users). The
768 * advantage is that we don't have to follow the strict linearity rule of
769 * PFNMAP mappings in order to support COWable mappings.
770 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800771 */
Nick Piggin7e675132008-04-28 02:13:00 -0700772#ifdef __HAVE_ARCH_PTE_SPECIAL
773# define HAVE_PTE_SPECIAL 1
774#else
775# define HAVE_PTE_SPECIAL 0
776#endif
777struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
778 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800779{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800780 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700781
782 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800783 if (likely(!pte_special(pte)))
784 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700785 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
786 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700787 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800788 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700789 return NULL;
790 }
791
792 /* !HAVE_PTE_SPECIAL case follows: */
793
Jared Hulbertb379d792008-04-28 02:12:58 -0700794 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
795 if (vma->vm_flags & VM_MIXEDMAP) {
796 if (!pfn_valid(pfn))
797 return NULL;
798 goto out;
799 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700800 unsigned long off;
801 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700802 if (pfn == vma->vm_pgoff + off)
803 return NULL;
804 if (!is_cow_mapping(vma->vm_flags))
805 return NULL;
806 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800807 }
808
Hugh Dickins62eede62009-09-21 17:03:34 -0700809 if (is_zero_pfn(pfn))
810 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800811check_pfn:
812 if (unlikely(pfn > highest_memmap_pfn)) {
813 print_bad_pte(vma, addr, pte, NULL);
814 return NULL;
815 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800816
817 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700818 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700819 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800820 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700821out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800822 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800823}
824
825/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 * copy one vm_area from one task to the other. Assumes the page tables
827 * already present in the new task to be cleared in the whole range
828 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 */
830
Hugh Dickins570a335b2009-12-14 17:58:46 -0800831static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700833 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700834 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835{
Nick Pigginb5810032005-10-29 18:16:12 -0700836 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 pte_t pte = *src_pte;
838 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839
840 /* pte contains position in swap or file, so copy. */
841 if (unlikely(!pte_present(pte))) {
842 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700843 swp_entry_t entry = pte_to_swp_entry(pte);
844
Hugh Dickins570a335b2009-12-14 17:58:46 -0800845 if (swap_duplicate(entry) < 0)
846 return entry.val;
847
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 /* make sure dst_mm is on swapoff's mmlist. */
849 if (unlikely(list_empty(&dst_mm->mmlist))) {
850 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700851 if (list_empty(&dst_mm->mmlist))
852 list_add(&dst_mm->mmlist,
853 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 spin_unlock(&mmlist_lock);
855 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800856 if (likely(!non_swap_entry(entry)))
857 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800858 else if (is_migration_entry(entry)) {
859 page = migration_entry_to_page(entry);
860
861 if (PageAnon(page))
862 rss[MM_ANONPAGES]++;
863 else
864 rss[MM_FILEPAGES]++;
865
866 if (is_write_migration_entry(entry) &&
867 is_cow_mapping(vm_flags)) {
868 /*
869 * COW mappings require pages in both
870 * parent and child to be set to read.
871 */
872 make_migration_entry_read(&entry);
873 pte = swp_entry_to_pte(entry);
874 set_pte_at(src_mm, addr, src_pte, pte);
875 }
Christoph Lameter06972122006-06-23 02:03:35 -0700876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700878 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 }
880
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 /*
882 * If it's a COW mapping, write protect it both
883 * in the parent and the child
884 */
Linus Torvalds67121172005-12-11 20:38:17 -0800885 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700887 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 }
889
890 /*
891 * If it's a shared mapping, mark it clean in
892 * the child
893 */
894 if (vm_flags & VM_SHARED)
895 pte = pte_mkclean(pte);
896 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800897
898 page = vm_normal_page(vma, addr, pte);
899 if (page) {
900 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700901 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800902 if (PageAnon(page))
903 rss[MM_ANONPAGES]++;
904 else
905 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800906 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700907
908out_set_pte:
909 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800910 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911}
912
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800913int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
914 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
915 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700917 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700919 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700920 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800921 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800922 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923
924again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800925 init_rss_vec(rss);
926
Hugh Dickinsc74df322005-10-29 18:16:23 -0700927 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 if (!dst_pte)
929 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700930 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700931 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700933 orig_src_pte = src_pte;
934 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700935 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 do {
938 /*
939 * We are holding two locks at this point - either of them
940 * could generate latencies in another task on another CPU.
941 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700942 if (progress >= 32) {
943 progress = 0;
944 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100945 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700946 break;
947 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 if (pte_none(*src_pte)) {
949 progress++;
950 continue;
951 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800952 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
953 vma, addr, rss);
954 if (entry.val)
955 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 progress += 8;
957 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700959 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700960 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700961 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800962 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700963 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700964 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800965
966 if (entry.val) {
967 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
968 return -ENOMEM;
969 progress = 0;
970 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 if (addr != end)
972 goto again;
973 return 0;
974}
975
976static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
977 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
978 unsigned long addr, unsigned long end)
979{
980 pmd_t *src_pmd, *dst_pmd;
981 unsigned long next;
982
983 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
984 if (!dst_pmd)
985 return -ENOMEM;
986 src_pmd = pmd_offset(src_pud, addr);
987 do {
988 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800989 if (pmd_trans_huge(*src_pmd)) {
990 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800991 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800992 err = copy_huge_pmd(dst_mm, src_mm,
993 dst_pmd, src_pmd, addr, vma);
994 if (err == -ENOMEM)
995 return -ENOMEM;
996 if (!err)
997 continue;
998 /* fall through */
999 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 if (pmd_none_or_clear_bad(src_pmd))
1001 continue;
1002 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1011 unsigned long addr, unsigned long end)
1012{
1013 pud_t *src_pud, *dst_pud;
1014 unsigned long next;
1015
1016 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1017 if (!dst_pud)
1018 return -ENOMEM;
1019 src_pud = pud_offset(src_pgd, addr);
1020 do {
1021 next = pud_addr_end(addr, end);
1022 if (pud_none_or_clear_bad(src_pud))
1023 continue;
1024 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1025 vma, addr, next))
1026 return -ENOMEM;
1027 } while (dst_pud++, src_pud++, addr = next, addr != end);
1028 return 0;
1029}
1030
1031int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032 struct vm_area_struct *vma)
1033{
1034 pgd_t *src_pgd, *dst_pgd;
1035 unsigned long next;
1036 unsigned long addr = vma->vm_start;
1037 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001038 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039
Nick Piggind9928952005-08-28 16:49:11 +10001040 /*
1041 * Don't copy ptes where a page fault will fill them correctly.
1042 * Fork becomes much lighter when there are big shared or private
1043 * readonly mappings. The tradeoff is that copy_page_range is more
1044 * efficient than faulting.
1045 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001046 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +10001047 if (!vma->anon_vma)
1048 return 0;
1049 }
1050
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 if (is_vm_hugetlb_page(vma))
1052 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1053
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001054 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001055 /*
1056 * We do not free on error cases below as remove_vma
1057 * gets called on error from higher level routine
1058 */
1059 ret = track_pfn_vma_copy(vma);
1060 if (ret)
1061 return ret;
1062 }
1063
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001064 /*
1065 * We need to invalidate the secondary MMU mappings only when
1066 * there could be a permission downgrade on the ptes of the
1067 * parent mm. And a permission downgrade will only happen if
1068 * is_cow_mapping() returns true.
1069 */
1070 if (is_cow_mapping(vma->vm_flags))
1071 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1072
1073 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 dst_pgd = pgd_offset(dst_mm, addr);
1075 src_pgd = pgd_offset(src_mm, addr);
1076 do {
1077 next = pgd_addr_end(addr, end);
1078 if (pgd_none_or_clear_bad(src_pgd))
1079 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001080 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1081 vma, addr, next))) {
1082 ret = -ENOMEM;
1083 break;
1084 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001086
1087 if (is_cow_mapping(vma->vm_flags))
1088 mmu_notifier_invalidate_range_end(src_mm,
1089 vma->vm_start, end);
1090 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091}
1092
Robin Holt51c6f662005-11-13 16:06:42 -08001093static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001094 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001096 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097{
Nick Pigginb5810032005-10-29 18:16:12 -07001098 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001099 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001100 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001101 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001102 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001103 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001104
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001105again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001106 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001107 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001109 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 do {
1111 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001112 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001114 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001115
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001117 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001118
Linus Torvalds6aab3412005-11-28 14:34:23 -08001119 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 if (unlikely(details) && page) {
1121 /*
1122 * unmap_shared_mapping_pages() wants to
1123 * invalidate cache without truncating:
1124 * unmap shared but keep private pages.
1125 */
1126 if (details->check_mapping &&
1127 details->check_mapping != page->mapping)
1128 continue;
1129 /*
1130 * Each page->index must be checked when
1131 * invalidating or truncating nonlinear.
1132 */
1133 if (details->nonlinear_vma &&
1134 (page->index < details->first_index ||
1135 page->index > details->last_index))
1136 continue;
1137 }
Nick Pigginb5810032005-10-29 18:16:12 -07001138 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001139 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 tlb_remove_tlb_entry(tlb, pte, addr);
1141 if (unlikely(!page))
1142 continue;
1143 if (unlikely(details) && details->nonlinear_vma
1144 && linear_page_index(details->nonlinear_vma,
1145 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001146 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001149 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001150 else {
1151 if (pte_dirty(ptent))
1152 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001153 if (pte_young(ptent) &&
1154 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001155 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001156 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001157 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001158 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001159 if (unlikely(page_mapcount(page) < 0))
1160 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001161 force_flush = !__tlb_remove_page(tlb, page);
1162 if (force_flush)
1163 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 continue;
1165 }
1166 /*
1167 * If details->check_mapping, we leave swap entries;
1168 * if details->nonlinear_vma, we leave file entries.
1169 */
1170 if (unlikely(details))
1171 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001172 if (pte_file(ptent)) {
1173 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001175 } else {
1176 swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178 if (!non_swap_entry(entry))
1179 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001180 else if (is_migration_entry(entry)) {
1181 struct page *page;
1182
1183 page = migration_entry_to_page(entry);
1184
1185 if (PageAnon(page))
1186 rss[MM_ANONPAGES]--;
1187 else
1188 rss[MM_FILEPAGES]--;
1189 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001190 if (unlikely(!free_swap_and_cache(entry)))
1191 print_bad_pte(vma, addr, ptent, NULL);
1192 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001193 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001194 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001195
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001196 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001197 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001198 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001199
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001200 /*
1201 * mmu_gather ran out of room to batch pages, we break out of
1202 * the PTE lock to avoid doing the potential expensive TLB invalidate
1203 * and page-free while holding it.
1204 */
1205 if (force_flush) {
1206 force_flush = 0;
1207 tlb_flush_mmu(tlb);
1208 if (addr != end)
1209 goto again;
1210 }
1211
Robin Holt51c6f662005-11-13 16:06:42 -08001212 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213}
1214
Robin Holt51c6f662005-11-13 16:06:42 -08001215static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001216 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001218 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219{
1220 pmd_t *pmd;
1221 unsigned long next;
1222
1223 pmd = pmd_offset(pud, addr);
1224 do {
1225 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001226 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001227 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001228#ifdef CONFIG_DEBUG_VM
1229 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1230 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1231 __func__, addr, end,
1232 vma->vm_start,
1233 vma->vm_end);
1234 BUG();
1235 }
1236#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001237 split_huge_page_pmd(vma->vm_mm, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001238 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001239 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001240 /* fall through */
1241 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001242 /*
1243 * Here there can be other concurrent MADV_DONTNEED or
1244 * trans huge page faults running, and if the pmd is
1245 * none or trans huge it can change under us. This is
1246 * because MADV_DONTNEED holds the mmap_sem in read
1247 * mode.
1248 */
1249 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1250 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001251 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001252next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001253 cond_resched();
1254 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001255
1256 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257}
1258
Robin Holt51c6f662005-11-13 16:06:42 -08001259static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001260 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001262 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263{
1264 pud_t *pud;
1265 unsigned long next;
1266
1267 pud = pud_offset(pgd, addr);
1268 do {
1269 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001270 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001272 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1273 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001274
1275 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276}
1277
Al Viro038c7aa2012-03-05 13:25:09 -05001278static void unmap_page_range(struct mmu_gather *tlb,
1279 struct vm_area_struct *vma,
1280 unsigned long addr, unsigned long end,
1281 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282{
1283 pgd_t *pgd;
1284 unsigned long next;
1285
1286 if (details && !details->check_mapping && !details->nonlinear_vma)
1287 details = NULL;
1288
1289 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001290 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 tlb_start_vma(tlb, vma);
1292 pgd = pgd_offset(vma->vm_mm, addr);
1293 do {
1294 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001295 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001297 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1298 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001300 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302
Al Virof5cc4ee2012-03-05 14:14:20 -05001303
1304static void unmap_single_vma(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001306 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001307 struct zap_details *details)
1308{
1309 unsigned long start = max(vma->vm_start, start_addr);
1310 unsigned long end;
1311
1312 if (start >= vma->vm_end)
1313 return;
1314 end = min(vma->vm_end, end_addr);
1315 if (end <= vma->vm_start)
1316 return;
1317
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301318 if (vma->vm_file)
1319 uprobe_munmap(vma, start, end);
1320
Al Virof5cc4ee2012-03-05 14:14:20 -05001321 if (unlikely(is_pfn_mapping(vma)))
1322 untrack_pfn_vma(vma, 0, 0);
1323
1324 if (start != end) {
1325 if (unlikely(is_vm_hugetlb_page(vma))) {
1326 /*
1327 * It is undesirable to test vma->vm_file as it
1328 * should be non-null for valid hugetlb area.
1329 * However, vm_file will be NULL in the error
1330 * cleanup path of do_mmap_pgoff. When
1331 * hugetlbfs ->mmap method fails,
1332 * do_mmap_pgoff() nullifies vma->vm_file
1333 * before calling this function to clean up.
1334 * Since no pte has actually been setup, it is
1335 * safe to do nothing in this case.
1336 */
1337 if (vma->vm_file)
1338 unmap_hugepage_range(vma, start, end, NULL);
1339 } else
1340 unmap_page_range(tlb, vma, start, end, details);
1341 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342}
1343
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344/**
1345 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001346 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 * @vma: the starting vma
1348 * @start_addr: virtual address at which to start unmapping
1349 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001351 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 * Only addresses between `start' and `end' will be unmapped.
1354 *
1355 * The VMA list must be sorted in ascending virtual address order.
1356 *
1357 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358 * range after unmap_vmas() returns. So the only responsibility here is to
1359 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360 * drops the lock and schedules.
1361 */
Al Viro6e8bb012012-03-05 13:41:15 -05001362void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001364 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001366 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001368 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001369 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001370 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001371 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001377 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 * @size: number of bytes to zap
1379 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001380 *
1381 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001383void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 unsigned long size, struct zap_details *details)
1385{
1386 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001387 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001388 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001391 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001392 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001393 mmu_notifier_invalidate_range_start(mm, start, end);
1394 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001395 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001396 mmu_notifier_invalidate_range_end(mm, start, end);
1397 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398}
1399
Jack Steinerc627f9c2008-07-29 22:33:53 -07001400/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001401 * zap_page_range_single - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of nonlinear truncation or shared cache invalidation
1406 *
1407 * The range must fit into one VMA.
1408 */
1409static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410 unsigned long size, struct zap_details *details)
1411{
1412 struct mm_struct *mm = vma->vm_mm;
1413 struct mmu_gather tlb;
1414 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001415
1416 lru_add_drain();
1417 tlb_gather_mmu(&tlb, mm, 0);
1418 update_hiwater_rss(mm);
1419 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001420 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001421 mmu_notifier_invalidate_range_end(mm, address, end);
1422 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423}
1424
Jack Steinerc627f9c2008-07-29 22:33:53 -07001425/**
1426 * zap_vma_ptes - remove ptes mapping the vma
1427 * @vma: vm_area_struct holding ptes to be zapped
1428 * @address: starting address of pages to zap
1429 * @size: number of bytes to zap
1430 *
1431 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432 *
1433 * The entire address range must be fully contained within the vma.
1434 *
1435 * Returns 0 if successful.
1436 */
1437int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438 unsigned long size)
1439{
1440 if (address < vma->vm_start || address + size > vma->vm_end ||
1441 !(vma->vm_flags & VM_PFNMAP))
1442 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001443 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001444 return 0;
1445}
1446EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447
Johannes Weiner142762b2010-05-24 14:32:39 -07001448/**
1449 * follow_page - look up a page descriptor from a user-virtual address
1450 * @vma: vm_area_struct mapping @address
1451 * @address: virtual address to look up
1452 * @flags: flags modifying lookup behaviour
1453 *
1454 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1455 *
1456 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1457 * an error pointer if there is a mapping to something not represented
1458 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001460struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001461 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462{
1463 pgd_t *pgd;
1464 pud_t *pud;
1465 pmd_t *pmd;
1466 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001467 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001469 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001471 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472 if (!IS_ERR(page)) {
1473 BUG_ON(flags & FOLL_GET);
1474 goto out;
1475 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001477 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 pgd = pgd_offset(mm, address);
1479 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001480 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
1482 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001483 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001484 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001485 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001486 BUG_ON(flags & FOLL_GET);
1487 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1488 goto out;
1489 }
1490 if (unlikely(pud_bad(*pud)))
1491 goto no_page_table;
1492
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001494 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001495 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001496 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001497 BUG_ON(flags & FOLL_GET);
1498 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1499 goto out;
1500 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001501 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001502 if (flags & FOLL_SPLIT) {
1503 split_huge_page_pmd(mm, pmd);
1504 goto split_fallthrough;
1505 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001506 spin_lock(&mm->page_table_lock);
1507 if (likely(pmd_trans_huge(*pmd))) {
1508 if (unlikely(pmd_trans_splitting(*pmd))) {
1509 spin_unlock(&mm->page_table_lock);
1510 wait_split_huge_page(vma->anon_vma, pmd);
1511 } else {
1512 page = follow_trans_huge_pmd(mm, address,
1513 pmd, flags);
1514 spin_unlock(&mm->page_table_lock);
1515 goto out;
1516 }
1517 } else
1518 spin_unlock(&mm->page_table_lock);
1519 /* fall through */
1520 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001521split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001522 if (unlikely(pmd_bad(*pmd)))
1523 goto no_page_table;
1524
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001525 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
1527 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001528 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001529 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001530 if ((flags & FOLL_WRITE) && !pte_write(pte))
1531 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001532
Linus Torvalds6aab3412005-11-28 14:34:23 -08001533 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001534 if (unlikely(!page)) {
1535 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001536 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001537 goto bad_page;
1538 page = pte_page(pte);
1539 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001540
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001541 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001542 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001543 if (flags & FOLL_TOUCH) {
1544 if ((flags & FOLL_WRITE) &&
1545 !pte_dirty(pte) && !PageDirty(page))
1546 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001547 /*
1548 * pte_mkyoung() would be more correct here, but atomic care
1549 * is needed to avoid losing the dirty bit: it is easier to use
1550 * mark_page_accessed().
1551 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001552 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001554 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001555 /*
1556 * The preliminary mapping check is mainly to avoid the
1557 * pointless overhead of lock_page on the ZERO_PAGE
1558 * which might bounce very badly if there is contention.
1559 *
1560 * If the page is already locked, we don't need to
1561 * handle it now - vmscan will handle it later if and
1562 * when it attempts to reclaim the page.
1563 */
1564 if (page->mapping && trylock_page(page)) {
1565 lru_add_drain(); /* push cached pages to LRU */
1566 /*
1567 * Because we lock page here and migration is
1568 * blocked by the pte's page reference, we need
1569 * only check for file-cache page truncation.
1570 */
1571 if (page->mapping)
1572 mlock_vma_page(page);
1573 unlock_page(page);
1574 }
1575 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001576unlock:
1577 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001579 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001581bad_page:
1582 pte_unmap_unlock(ptep, ptl);
1583 return ERR_PTR(-EFAULT);
1584
1585no_page:
1586 pte_unmap_unlock(ptep, ptl);
1587 if (!pte_none(pte))
1588 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001589
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001590no_page_table:
1591 /*
1592 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001593 * has touched so far, we don't want to allocate unnecessary pages or
1594 * page tables. Return error instead of NULL to skip handle_mm_fault,
1595 * then get_dump_page() will return NULL to leave a hole in the dump.
1596 * But we can only make this optimization where a hole would surely
1597 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001598 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001599 if ((flags & FOLL_DUMP) &&
1600 (!vma->vm_ops || !vma->vm_ops->fault))
1601 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001602 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603}
1604
Linus Torvalds95042f92011-04-12 14:15:51 -07001605static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1606{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001607 return stack_guard_page_start(vma, addr) ||
1608 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001609}
1610
Huang Ying0014bd92011-01-30 11:15:47 +08001611/**
1612 * __get_user_pages() - pin user pages in memory
1613 * @tsk: task_struct of target task
1614 * @mm: mm_struct of target mm
1615 * @start: starting user address
1616 * @nr_pages: number of pages from start to pin
1617 * @gup_flags: flags modifying pin behaviour
1618 * @pages: array that receives pointers to the pages pinned.
1619 * Should be at least nr_pages long. Or NULL, if caller
1620 * only intends to ensure the pages are faulted in.
1621 * @vmas: array of pointers to vmas corresponding to each page.
1622 * Or NULL if the caller does not require them.
1623 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1624 *
1625 * Returns number of pages pinned. This may be fewer than the number
1626 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1627 * were pinned, returns -errno. Each page returned must be released
1628 * with a put_page() call when it is finished with. vmas will only
1629 * remain valid while mmap_sem is held.
1630 *
1631 * Must be called with mmap_sem held for read or write.
1632 *
1633 * __get_user_pages walks a process's page tables and takes a reference to
1634 * each struct page that each user address corresponds to at a given
1635 * instant. That is, it takes the page that would be accessed if a user
1636 * thread accesses the given user virtual address at that instant.
1637 *
1638 * This does not guarantee that the page exists in the user mappings when
1639 * __get_user_pages returns, and there may even be a completely different
1640 * page there in some cases (eg. if mmapped pagecache has been invalidated
1641 * and subsequently re faulted). However it does guarantee that the page
1642 * won't be freed completely. And mostly callers simply care that the page
1643 * contains data that was valid *at some point in time*. Typically, an IO
1644 * or similar operation cannot guarantee anything stronger anyway because
1645 * locks can't be held over the syscall boundary.
1646 *
1647 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1648 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1649 * appropriate) must be called after the page is finished with, and
1650 * before put_page is called.
1651 *
1652 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1653 * or mmap_sem contention, and if waiting is needed to pin all pages,
1654 * *@nonblocking will be set to 0.
1655 *
1656 * In most cases, get_user_pages or get_user_pages_fast should be used
1657 * instead of __get_user_pages. __get_user_pages should be used only if
1658 * you need some special @gup_flags.
1659 */
Nick Pigginb291f002008-10-18 20:26:44 -07001660int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001661 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001662 struct page **pages, struct vm_area_struct **vmas,
1663 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664{
1665 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001666 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667
Peter Zijlstra9d737772009-06-25 11:58:55 +02001668 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001669 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001670
1671 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1672
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 /*
1674 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001675 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001677 vm_flags = (gup_flags & FOLL_WRITE) ?
1678 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1679 vm_flags &= (gup_flags & FOLL_FORCE) ?
1680 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 i = 0;
1682
1683 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001684 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685
1686 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001687 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 pgd_t *pgd;
1690 pud_t *pud;
1691 pmd_t *pmd;
1692 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001693
1694 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001695 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 return i ? : -EFAULT;
1697 if (pg > TASK_SIZE)
1698 pgd = pgd_offset_k(pg);
1699 else
1700 pgd = pgd_offset_gate(mm, pg);
1701 BUG_ON(pgd_none(*pgd));
1702 pud = pud_offset(pgd, pg);
1703 BUG_ON(pud_none(*pud));
1704 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001705 if (pmd_none(*pmd))
1706 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001707 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001709 if (pte_none(*pte)) {
1710 pte_unmap(pte);
1711 return i ? : -EFAULT;
1712 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001713 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001715 struct page *page;
1716
Linus Torvalds95042f92011-04-12 14:15:51 -07001717 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001718 if (!page) {
1719 if (!(gup_flags & FOLL_DUMP) &&
1720 is_zero_pfn(pte_pfn(*pte)))
1721 page = pte_page(*pte);
1722 else {
1723 pte_unmap(pte);
1724 return i ? : -EFAULT;
1725 }
1726 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001727 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001728 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001731 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 }
1733
Nick Pigginb291f002008-10-18 20:26:44 -07001734 if (!vma ||
1735 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001736 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 return i ? : -EFAULT;
1738
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001739 if (is_vm_hugetlb_page(vma)) {
1740 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001741 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001742 continue;
1743 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001744
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001746 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001747 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748
Ethan Solomita462e00c2007-07-15 23:38:16 -07001749 /*
Ying Han47792802009-01-06 14:40:18 -08001750 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001751 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001752 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001753 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001754 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001755
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001756 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001757 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001758 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001759 unsigned int fault_flags = 0;
1760
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001761 /* For mlock, just skip the stack guard page. */
1762 if (foll_flags & FOLL_MLOCK) {
1763 if (stack_guard_page(vma, start))
1764 goto next_page;
1765 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001766 if (foll_flags & FOLL_WRITE)
1767 fault_flags |= FAULT_FLAG_WRITE;
1768 if (nonblocking)
1769 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001770 if (foll_flags & FOLL_NOWAIT)
1771 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001772
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001773 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001774 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001775
Nick Piggin83c54072007-07-19 01:47:05 -07001776 if (ret & VM_FAULT_ERROR) {
1777 if (ret & VM_FAULT_OOM)
1778 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001779 if (ret & (VM_FAULT_HWPOISON |
1780 VM_FAULT_HWPOISON_LARGE)) {
1781 if (i)
1782 return i;
1783 else if (gup_flags & FOLL_HWPOISON)
1784 return -EHWPOISON;
1785 else
1786 return -EFAULT;
1787 }
1788 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001789 return i ? i : -EFAULT;
1790 BUG();
1791 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001792
1793 if (tsk) {
1794 if (ret & VM_FAULT_MAJOR)
1795 tsk->maj_flt++;
1796 else
1797 tsk->min_flt++;
1798 }
Nick Piggin83c54072007-07-19 01:47:05 -07001799
Michel Lespinasse53a77062011-01-13 15:46:14 -08001800 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001801 if (nonblocking)
1802 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001803 return i;
1804 }
1805
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001806 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001807 * The VM_FAULT_WRITE bit tells us that
1808 * do_wp_page has broken COW when necessary,
1809 * even if maybe_mkwrite decided not to set
1810 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001811 * page lookups as if they were reads. But only
1812 * do so when looping for pte_write is futile:
1813 * in some cases userspace may also be wanting
1814 * to write to the gotten user page, which a
1815 * read fault here might prevent (a readonly
1816 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001817 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001818 if ((ret & VM_FAULT_WRITE) &&
1819 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001820 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001821
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001822 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001824 if (IS_ERR(page))
1825 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001827 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001828
Russell Kinga6f36be2006-12-30 22:24:19 +00001829 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001830 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001832next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 if (vmas)
1834 vmas[i] = vma;
1835 i++;
1836 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001837 nr_pages--;
1838 } while (nr_pages && start < vma->vm_end);
1839 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 return i;
1841}
Huang Ying0014bd92011-01-30 11:15:47 +08001842EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001843
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001844/*
1845 * fixup_user_fault() - manually resolve a user page fault
1846 * @tsk: the task_struct to use for page fault accounting, or
1847 * NULL if faults are not to be recorded.
1848 * @mm: mm_struct of target mm
1849 * @address: user address
1850 * @fault_flags:flags to pass down to handle_mm_fault()
1851 *
1852 * This is meant to be called in the specific scenario where for locking reasons
1853 * we try to access user memory in atomic context (within a pagefault_disable()
1854 * section), this returns -EFAULT, and we want to resolve the user fault before
1855 * trying again.
1856 *
1857 * Typically this is meant to be used by the futex code.
1858 *
1859 * The main difference with get_user_pages() is that this function will
1860 * unconditionally call handle_mm_fault() which will in turn perform all the
1861 * necessary SW fixup of the dirty and young bits in the PTE, while
1862 * handle_mm_fault() only guarantees to update these in the struct page.
1863 *
1864 * This is important for some architectures where those bits also gate the
1865 * access permission to the page because they are maintained in software. On
1866 * such architectures, gup() will not be enough to make a subsequent access
1867 * succeed.
1868 *
1869 * This should be called with the mm_sem held for read.
1870 */
1871int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1872 unsigned long address, unsigned int fault_flags)
1873{
1874 struct vm_area_struct *vma;
1875 int ret;
1876
1877 vma = find_extend_vma(mm, address);
1878 if (!vma || address < vma->vm_start)
1879 return -EFAULT;
1880
1881 ret = handle_mm_fault(mm, vma, address, fault_flags);
1882 if (ret & VM_FAULT_ERROR) {
1883 if (ret & VM_FAULT_OOM)
1884 return -ENOMEM;
1885 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1886 return -EHWPOISON;
1887 if (ret & VM_FAULT_SIGBUS)
1888 return -EFAULT;
1889 BUG();
1890 }
1891 if (tsk) {
1892 if (ret & VM_FAULT_MAJOR)
1893 tsk->maj_flt++;
1894 else
1895 tsk->min_flt++;
1896 }
1897 return 0;
1898}
1899
1900/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001901 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001902 * @tsk: the task_struct to use for page fault accounting, or
1903 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001904 * @mm: mm_struct of target mm
1905 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001906 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001907 * @write: whether pages will be written to by the caller
1908 * @force: whether to force write access even if user mapping is
1909 * readonly. This will result in the page being COWed even
1910 * in MAP_SHARED mappings. You do not want this.
1911 * @pages: array that receives pointers to the pages pinned.
1912 * Should be at least nr_pages long. Or NULL, if caller
1913 * only intends to ensure the pages are faulted in.
1914 * @vmas: array of pointers to vmas corresponding to each page.
1915 * Or NULL if the caller does not require them.
1916 *
1917 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001918 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001919 * were pinned, returns -errno. Each page returned must be released
1920 * with a put_page() call when it is finished with. vmas will only
1921 * remain valid while mmap_sem is held.
1922 *
1923 * Must be called with mmap_sem held for read or write.
1924 *
1925 * get_user_pages walks a process's page tables and takes a reference to
1926 * each struct page that each user address corresponds to at a given
1927 * instant. That is, it takes the page that would be accessed if a user
1928 * thread accesses the given user virtual address at that instant.
1929 *
1930 * This does not guarantee that the page exists in the user mappings when
1931 * get_user_pages returns, and there may even be a completely different
1932 * page there in some cases (eg. if mmapped pagecache has been invalidated
1933 * and subsequently re faulted). However it does guarantee that the page
1934 * won't be freed completely. And mostly callers simply care that the page
1935 * contains data that was valid *at some point in time*. Typically, an IO
1936 * or similar operation cannot guarantee anything stronger anyway because
1937 * locks can't be held over the syscall boundary.
1938 *
1939 * If write=0, the page must not be written to. If the page is written to,
1940 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1941 * after the page is finished with, and before put_page is called.
1942 *
1943 * get_user_pages is typically used for fewer-copy IO operations, to get a
1944 * handle on the memory by some means other than accesses via the user virtual
1945 * addresses. The pages may be submitted for DMA to devices or accessed via
1946 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1947 * use the correct cache flushing APIs.
1948 *
1949 * See also get_user_pages_fast, for performance critical applications.
1950 */
Nick Pigginb291f002008-10-18 20:26:44 -07001951int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001952 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001953 struct page **pages, struct vm_area_struct **vmas)
1954{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001955 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001956
Hugh Dickins58fa8792009-09-21 17:03:31 -07001957 if (pages)
1958 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001959 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001960 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001961 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001962 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001963
Michel Lespinasse53a77062011-01-13 15:46:14 -08001964 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1965 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001966}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967EXPORT_SYMBOL(get_user_pages);
1968
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001969/**
1970 * get_dump_page() - pin user page in memory while writing it to core dump
1971 * @addr: user address
1972 *
1973 * Returns struct page pointer of user page pinned for dump,
1974 * to be freed afterwards by page_cache_release() or put_page().
1975 *
1976 * Returns NULL on any kind of failure - a hole must then be inserted into
1977 * the corefile, to preserve alignment with its headers; and also returns
1978 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1979 * allowing a hole to be left in the corefile to save diskspace.
1980 *
1981 * Called without mmap_sem, but after all other threads have been killed.
1982 */
1983#ifdef CONFIG_ELF_CORE
1984struct page *get_dump_page(unsigned long addr)
1985{
1986 struct vm_area_struct *vma;
1987 struct page *page;
1988
1989 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001990 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1991 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001992 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001993 flush_cache_page(vma, addr, page_to_pfn(page));
1994 return page;
1995}
1996#endif /* CONFIG_ELF_CORE */
1997
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001998pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001999 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002000{
2001 pgd_t * pgd = pgd_offset(mm, addr);
2002 pud_t * pud = pud_alloc(mm, pgd, addr);
2003 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002004 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002005 if (pmd) {
2006 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002007 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002008 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002009 }
2010 return NULL;
2011}
2012
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002014 * This is the old fallback for page remapping.
2015 *
2016 * For historical reasons, it only allows reserved pages. Only
2017 * old drivers should use this, and they needed to mark their
2018 * pages reserved for the old functions anyway.
2019 */
Nick Piggin423bad62008-04-28 02:13:01 -07002020static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2021 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002022{
Nick Piggin423bad62008-04-28 02:13:01 -07002023 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002024 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002025 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002026 spinlock_t *ptl;
2027
Linus Torvalds238f58d2005-11-29 13:01:56 -08002028 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002029 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002030 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002031 retval = -ENOMEM;
2032 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002033 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002034 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002035 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002036 retval = -EBUSY;
2037 if (!pte_none(*pte))
2038 goto out_unlock;
2039
2040 /* Ok, finally just insert the thing.. */
2041 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002042 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002043 page_add_file_rmap(page);
2044 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2045
2046 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002047 pte_unmap_unlock(pte, ptl);
2048 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002049out_unlock:
2050 pte_unmap_unlock(pte, ptl);
2051out:
2052 return retval;
2053}
2054
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002055/**
2056 * vm_insert_page - insert single page into user vma
2057 * @vma: user vma to map to
2058 * @addr: target user address of this page
2059 * @page: source kernel page
2060 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002061 * This allows drivers to insert individual pages they've allocated
2062 * into a user vma.
2063 *
2064 * The page has to be a nice clean _individual_ kernel allocation.
2065 * If you allocate a compound page, you need to have marked it as
2066 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002067 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002068 *
2069 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2070 * took an arbitrary page protection parameter. This doesn't allow
2071 * that. Your vma protection will have to be set up correctly, which
2072 * means that if you want a shared writable mapping, you'd better
2073 * ask for a shared writable mapping!
2074 *
2075 * The page does not need to be reserved.
2076 */
Nick Piggin423bad62008-04-28 02:13:01 -07002077int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2078 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002079{
2080 if (addr < vma->vm_start || addr >= vma->vm_end)
2081 return -EFAULT;
2082 if (!page_count(page))
2083 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08002084 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07002085 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002086}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002087EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002088
Nick Piggin423bad62008-04-28 02:13:01 -07002089static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2090 unsigned long pfn, pgprot_t prot)
2091{
2092 struct mm_struct *mm = vma->vm_mm;
2093 int retval;
2094 pte_t *pte, entry;
2095 spinlock_t *ptl;
2096
2097 retval = -ENOMEM;
2098 pte = get_locked_pte(mm, addr, &ptl);
2099 if (!pte)
2100 goto out;
2101 retval = -EBUSY;
2102 if (!pte_none(*pte))
2103 goto out_unlock;
2104
2105 /* Ok, finally just insert the thing.. */
2106 entry = pte_mkspecial(pfn_pte(pfn, prot));
2107 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002108 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002109
2110 retval = 0;
2111out_unlock:
2112 pte_unmap_unlock(pte, ptl);
2113out:
2114 return retval;
2115}
2116
Nick Piggine0dc0d82007-02-12 00:51:36 -08002117/**
2118 * vm_insert_pfn - insert single pfn into user vma
2119 * @vma: user vma to map to
2120 * @addr: target user address of this page
2121 * @pfn: source kernel pfn
2122 *
2123 * Similar to vm_inert_page, this allows drivers to insert individual pages
2124 * they've allocated into a user vma. Same comments apply.
2125 *
2126 * This function should only be called from a vm_ops->fault handler, and
2127 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002128 *
2129 * vma cannot be a COW mapping.
2130 *
2131 * As this is called only for pages that do not currently exist, we
2132 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002133 */
2134int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002135 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002136{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002137 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002138 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002139 /*
2140 * Technically, architectures with pte_special can avoid all these
2141 * restrictions (same for remap_pfn_range). However we would like
2142 * consistency in testing and feature parity among all, so we should
2143 * try to keep these invariants in place for everybody.
2144 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002145 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2146 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2147 (VM_PFNMAP|VM_MIXEDMAP));
2148 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2149 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002150
Nick Piggin423bad62008-04-28 02:13:01 -07002151 if (addr < vma->vm_start || addr >= vma->vm_end)
2152 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002153 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002154 return -EINVAL;
2155
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002156 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002157
2158 if (ret)
2159 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2160
2161 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002162}
2163EXPORT_SYMBOL(vm_insert_pfn);
2164
Nick Piggin423bad62008-04-28 02:13:01 -07002165int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2166 unsigned long pfn)
2167{
2168 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2169
2170 if (addr < vma->vm_start || addr >= vma->vm_end)
2171 return -EFAULT;
2172
2173 /*
2174 * If we don't have pte special, then we have to use the pfn_valid()
2175 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2176 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002177 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2178 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002179 */
2180 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2181 struct page *page;
2182
2183 page = pfn_to_page(pfn);
2184 return insert_page(vma, addr, page, vma->vm_page_prot);
2185 }
2186 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2187}
2188EXPORT_SYMBOL(vm_insert_mixed);
2189
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002190/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 * maps a range of physical memory into the requested pages. the old
2192 * mappings are removed. any references to nonexistent pages results
2193 * in null mappings (currently treated as "copy-on-access")
2194 */
2195static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2196 unsigned long addr, unsigned long end,
2197 unsigned long pfn, pgprot_t prot)
2198{
2199 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002200 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201
Hugh Dickinsc74df322005-10-29 18:16:23 -07002202 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 if (!pte)
2204 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002205 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 do {
2207 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002208 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 pfn++;
2210 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002211 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002212 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 return 0;
2214}
2215
2216static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2217 unsigned long addr, unsigned long end,
2218 unsigned long pfn, pgprot_t prot)
2219{
2220 pmd_t *pmd;
2221 unsigned long next;
2222
2223 pfn -= addr >> PAGE_SHIFT;
2224 pmd = pmd_alloc(mm, pud, addr);
2225 if (!pmd)
2226 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002227 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 do {
2229 next = pmd_addr_end(addr, end);
2230 if (remap_pte_range(mm, pmd, addr, next,
2231 pfn + (addr >> PAGE_SHIFT), prot))
2232 return -ENOMEM;
2233 } while (pmd++, addr = next, addr != end);
2234 return 0;
2235}
2236
2237static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2238 unsigned long addr, unsigned long end,
2239 unsigned long pfn, pgprot_t prot)
2240{
2241 pud_t *pud;
2242 unsigned long next;
2243
2244 pfn -= addr >> PAGE_SHIFT;
2245 pud = pud_alloc(mm, pgd, addr);
2246 if (!pud)
2247 return -ENOMEM;
2248 do {
2249 next = pud_addr_end(addr, end);
2250 if (remap_pmd_range(mm, pud, addr, next,
2251 pfn + (addr >> PAGE_SHIFT), prot))
2252 return -ENOMEM;
2253 } while (pud++, addr = next, addr != end);
2254 return 0;
2255}
2256
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002257/**
2258 * remap_pfn_range - remap kernel memory to userspace
2259 * @vma: user vma to map to
2260 * @addr: target user address to start at
2261 * @pfn: physical address of kernel memory
2262 * @size: size of map area
2263 * @prot: page protection flags for this mapping
2264 *
2265 * Note: this is only safe if the mm semaphore is held when called.
2266 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2268 unsigned long pfn, unsigned long size, pgprot_t prot)
2269{
2270 pgd_t *pgd;
2271 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002272 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 struct mm_struct *mm = vma->vm_mm;
2274 int err;
2275
2276 /*
2277 * Physically remapped pages are special. Tell the
2278 * rest of the world about it:
2279 * VM_IO tells people not to look at these pages
2280 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08002281 * VM_RESERVED is specified all over the place, because
2282 * in 2.4 it kept swapout's vma scan off this vma; but
2283 * in 2.6 the LRU scan won't even find its pages, so this
2284 * flag means no more than count its pages in reserved_vm,
2285 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08002286 * VM_PFNMAP tells the core MM that the base pages are just
2287 * raw PFN mappings, and do not have a "struct page" associated
2288 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002289 *
2290 * There's a horrible special case to handle copy-on-write
2291 * behaviour that some programs depend on. We mark the "original"
2292 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002294 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002295 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002296 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002297 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002298 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002299
Linus Torvalds6aab3412005-11-28 14:34:23 -08002300 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002302 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002303 if (err) {
2304 /*
2305 * To indicate that track_pfn related cleanup is not
2306 * needed from higher level routine calling unmap_vmas
2307 */
2308 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002309 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002310 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002311 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002312
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 BUG_ON(addr >= end);
2314 pfn -= addr >> PAGE_SHIFT;
2315 pgd = pgd_offset(mm, addr);
2316 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 do {
2318 next = pgd_addr_end(addr, end);
2319 err = remap_pud_range(mm, pgd, addr, next,
2320 pfn + (addr >> PAGE_SHIFT), prot);
2321 if (err)
2322 break;
2323 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002324
2325 if (err)
2326 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2327
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 return err;
2329}
2330EXPORT_SYMBOL(remap_pfn_range);
2331
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002332static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2333 unsigned long addr, unsigned long end,
2334 pte_fn_t fn, void *data)
2335{
2336 pte_t *pte;
2337 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002338 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002339 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002340
2341 pte = (mm == &init_mm) ?
2342 pte_alloc_kernel(pmd, addr) :
2343 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2344 if (!pte)
2345 return -ENOMEM;
2346
2347 BUG_ON(pmd_huge(*pmd));
2348
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002349 arch_enter_lazy_mmu_mode();
2350
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002351 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002352
2353 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002354 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002355 if (err)
2356 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002357 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002358
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002359 arch_leave_lazy_mmu_mode();
2360
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002361 if (mm != &init_mm)
2362 pte_unmap_unlock(pte-1, ptl);
2363 return err;
2364}
2365
2366static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2367 unsigned long addr, unsigned long end,
2368 pte_fn_t fn, void *data)
2369{
2370 pmd_t *pmd;
2371 unsigned long next;
2372 int err;
2373
Andi Kleenceb86872008-07-23 21:27:50 -07002374 BUG_ON(pud_huge(*pud));
2375
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002376 pmd = pmd_alloc(mm, pud, addr);
2377 if (!pmd)
2378 return -ENOMEM;
2379 do {
2380 next = pmd_addr_end(addr, end);
2381 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2382 if (err)
2383 break;
2384 } while (pmd++, addr = next, addr != end);
2385 return err;
2386}
2387
2388static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2389 unsigned long addr, unsigned long end,
2390 pte_fn_t fn, void *data)
2391{
2392 pud_t *pud;
2393 unsigned long next;
2394 int err;
2395
2396 pud = pud_alloc(mm, pgd, addr);
2397 if (!pud)
2398 return -ENOMEM;
2399 do {
2400 next = pud_addr_end(addr, end);
2401 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2402 if (err)
2403 break;
2404 } while (pud++, addr = next, addr != end);
2405 return err;
2406}
2407
2408/*
2409 * Scan a region of virtual memory, filling in page tables as necessary
2410 * and calling a provided function on each leaf page table.
2411 */
2412int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2413 unsigned long size, pte_fn_t fn, void *data)
2414{
2415 pgd_t *pgd;
2416 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002417 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002418 int err;
2419
2420 BUG_ON(addr >= end);
2421 pgd = pgd_offset(mm, addr);
2422 do {
2423 next = pgd_addr_end(addr, end);
2424 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2425 if (err)
2426 break;
2427 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002428
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002429 return err;
2430}
2431EXPORT_SYMBOL_GPL(apply_to_page_range);
2432
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002434 * handle_pte_fault chooses page fault handler according to an entry
2435 * which was read non-atomically. Before making any commitment, on
2436 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002437 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002438 * must check under lock before unmapping the pte and proceeding
2439 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002440 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002441 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002442static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002443 pte_t *page_table, pte_t orig_pte)
2444{
2445 int same = 1;
2446#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2447 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002448 spinlock_t *ptl = pte_lockptr(mm, pmd);
2449 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002450 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002451 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002452 }
2453#endif
2454 pte_unmap(page_table);
2455 return same;
2456}
2457
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002458static 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 -08002459{
2460 /*
2461 * If the source page was a PFN mapping, we don't have
2462 * a "struct page" for it. We do a best-effort copy by
2463 * just copying from the original user address. If that
2464 * fails, we just zero-fill it. Live with it.
2465 */
2466 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002467 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002468 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2469
2470 /*
2471 * This really shouldn't fail, because the page is there
2472 * in the page tables. But it might just be unreadable,
2473 * in which case we just give up and fill the result with
2474 * zeroes.
2475 */
2476 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002477 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002478 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002479 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002480 } else
2481 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002482}
2483
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 * This routine handles present pages, when users try to write
2486 * to a shared page. It is done by copying the page to a new address
2487 * and decrementing the shared-page counter for the old page.
2488 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 * Note that this routine assumes that the protection checks have been
2490 * done by the caller (the low-level page fault routine in most cases).
2491 * Thus we can safely just mark it writable once we've done any necessary
2492 * COW.
2493 *
2494 * We also mark the page dirty at this point even though the page will
2495 * change only once the write actually happens. This avoids a few races,
2496 * and potentially makes it more efficient.
2497 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002498 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2499 * but allow concurrent faults), with pte both mapped and locked.
2500 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002502static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2503 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002504 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002505 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002507 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002509 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002510 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002511 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512
Linus Torvalds6aab3412005-11-28 14:34:23 -08002513 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002514 if (!old_page) {
2515 /*
2516 * VM_MIXEDMAP !pfn_valid() case
2517 *
2518 * We should not cow pages in a shared writeable mapping.
2519 * Just mark the pages writable as we can't do any dirty
2520 * accounting on raw pfn maps.
2521 */
2522 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2523 (VM_WRITE|VM_SHARED))
2524 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002525 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002526 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002528 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002529 * Take out anonymous pages first, anonymous shared vmas are
2530 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002531 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002532 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002533 if (!trylock_page(old_page)) {
2534 page_cache_get(old_page);
2535 pte_unmap_unlock(page_table, ptl);
2536 lock_page(old_page);
2537 page_table = pte_offset_map_lock(mm, pmd, address,
2538 &ptl);
2539 if (!pte_same(*page_table, orig_pte)) {
2540 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002541 goto unlock;
2542 }
2543 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002544 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002545 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002546 /*
2547 * The page is all ours. Move it to our anon_vma so
2548 * the rmap code will not search our parent or siblings.
2549 * Protected against the rmap code by the page lock.
2550 */
2551 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002552 unlock_page(old_page);
2553 goto reuse;
2554 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002555 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002556 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002557 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002558 /*
2559 * Only catch write-faults on shared writable pages,
2560 * read-only shared pages can get COWed by
2561 * get_user_pages(.write=1, .force=1).
2562 */
David Howells9637a5e2006-06-23 02:03:43 -07002563 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002564 struct vm_fault vmf;
2565 int tmp;
2566
2567 vmf.virtual_address = (void __user *)(address &
2568 PAGE_MASK);
2569 vmf.pgoff = old_page->index;
2570 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2571 vmf.page = old_page;
2572
David Howells9637a5e2006-06-23 02:03:43 -07002573 /*
2574 * Notify the address space that the page is about to
2575 * become writable so that it can prohibit this or wait
2576 * for the page to get into an appropriate state.
2577 *
2578 * We do this without the lock held, so that it can
2579 * sleep if it needs to.
2580 */
2581 page_cache_get(old_page);
2582 pte_unmap_unlock(page_table, ptl);
2583
Nick Pigginc2ec1752009-03-31 15:23:21 -07002584 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2585 if (unlikely(tmp &
2586 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2587 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002588 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002589 }
Nick Pigginb827e492009-04-30 15:08:16 -07002590 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2591 lock_page(old_page);
2592 if (!old_page->mapping) {
2593 ret = 0; /* retry the fault */
2594 unlock_page(old_page);
2595 goto unwritable_page;
2596 }
2597 } else
2598 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002599
David Howells9637a5e2006-06-23 02:03:43 -07002600 /*
2601 * Since we dropped the lock we need to revalidate
2602 * the PTE as someone else may have changed it. If
2603 * they did, we just return, as we can count on the
2604 * MMU to tell us if they didn't also make it writable.
2605 */
2606 page_table = pte_offset_map_lock(mm, pmd, address,
2607 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002608 if (!pte_same(*page_table, orig_pte)) {
2609 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002610 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002611 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002612
2613 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002615 dirty_page = old_page;
2616 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002617
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002618reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002619 flush_cache_page(vma, address, pte_pfn(orig_pte));
2620 entry = pte_mkyoung(orig_pte);
2621 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002622 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002623 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002624 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002625 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002626
2627 if (!dirty_page)
2628 return ret;
2629
2630 /*
2631 * Yes, Virginia, this is actually required to prevent a race
2632 * with clear_page_dirty_for_io() from clearing the page dirty
2633 * bit after it clear all dirty ptes, but before a racing
2634 * do_wp_page installs a dirty pte.
2635 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002636 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002637 */
2638 if (!page_mkwrite) {
2639 wait_on_page_locked(dirty_page);
2640 set_page_dirty_balance(dirty_page, page_mkwrite);
2641 }
2642 put_page(dirty_page);
2643 if (page_mkwrite) {
2644 struct address_space *mapping = dirty_page->mapping;
2645
2646 set_page_dirty(dirty_page);
2647 unlock_page(dirty_page);
2648 page_cache_release(dirty_page);
2649 if (mapping) {
2650 /*
2651 * Some device drivers do not set page.mapping
2652 * but still dirty their pages
2653 */
2654 balance_dirty_pages_ratelimited(mapping);
2655 }
2656 }
2657
2658 /* file_update_time outside page_lock */
2659 if (vma->vm_file)
2660 file_update_time(vma->vm_file);
2661
2662 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664
2665 /*
2666 * Ok, we need to copy. Oh, well..
2667 */
Nick Pigginb5810032005-10-29 18:16:12 -07002668 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002669gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002670 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671
2672 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002673 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002674
Hugh Dickins62eede62009-09-21 17:03:34 -07002675 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002676 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2677 if (!new_page)
2678 goto oom;
2679 } else {
2680 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2681 if (!new_page)
2682 goto oom;
2683 cow_user_page(new_page, old_page, address, vma);
2684 }
2685 __SetPageUptodate(new_page);
2686
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002687 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002688 goto oom_free_new;
2689
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 /*
2691 * Re-check the pte - we dropped the lock
2692 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002693 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002694 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002695 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002696 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002697 dec_mm_counter_fast(mm, MM_FILEPAGES);
2698 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002699 }
2700 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002701 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002702 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002703 entry = mk_pte(new_page, vma->vm_page_prot);
2704 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002705 /*
2706 * Clear the pte entry and flush it first, before updating the
2707 * pte with the new entry. This will avoid a race condition
2708 * seen in the presence of one thread doing SMC and another
2709 * thread doing COW.
2710 */
Izik Eidus828502d2009-09-21 17:01:51 -07002711 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002712 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002713 /*
2714 * We call the notify macro here because, when using secondary
2715 * mmu page tables (such as kvm shadow page tables), we want the
2716 * new page to be mapped directly into the secondary page table.
2717 */
2718 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002719 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002720 if (old_page) {
2721 /*
2722 * Only after switching the pte to the new page may
2723 * we remove the mapcount here. Otherwise another
2724 * process may come and find the rmap count decremented
2725 * before the pte is switched to the new page, and
2726 * "reuse" the old page writing into it while our pte
2727 * here still points into it and can be read by other
2728 * threads.
2729 *
2730 * The critical issue is to order this
2731 * page_remove_rmap with the ptp_clear_flush above.
2732 * Those stores are ordered by (if nothing else,)
2733 * the barrier present in the atomic_add_negative
2734 * in page_remove_rmap.
2735 *
2736 * Then the TLB flush in ptep_clear_flush ensures that
2737 * no process can access the old page before the
2738 * decremented mapcount is visible. And the old page
2739 * cannot be reused until after the decremented
2740 * mapcount is visible. So transitively, TLBs to
2741 * old page will be flushed before it can be reused.
2742 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002743 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002744 }
2745
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 /* Free the old page.. */
2747 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002748 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002749 } else
2750 mem_cgroup_uncharge_page(new_page);
2751
Hugh Dickins920fc352005-11-21 21:32:17 -08002752 if (new_page)
2753 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002754unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002755 pte_unmap_unlock(page_table, ptl);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002756 if (old_page) {
2757 /*
2758 * Don't let another task, with possibly unlocked vma,
2759 * keep the mlocked page.
2760 */
2761 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2762 lock_page(old_page); /* LRU manipulation */
2763 munlock_vma_page(old_page);
2764 unlock_page(old_page);
2765 }
2766 page_cache_release(old_page);
2767 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002768 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002769oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002770 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002771oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002772 if (old_page) {
2773 if (page_mkwrite) {
2774 unlock_page(old_page);
2775 page_cache_release(old_page);
2776 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002777 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002778 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002780
2781unwritable_page:
2782 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002783 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784}
2785
Peter Zijlstra97a89412011-05-24 17:12:04 -07002786static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 unsigned long start_addr, unsigned long end_addr,
2788 struct zap_details *details)
2789{
Al Virof5cc4ee2012-03-05 14:14:20 -05002790 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791}
2792
2793static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2794 struct zap_details *details)
2795{
2796 struct vm_area_struct *vma;
2797 struct prio_tree_iter iter;
2798 pgoff_t vba, vea, zba, zea;
2799
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 vma_prio_tree_foreach(vma, &iter, root,
2801 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802
2803 vba = vma->vm_pgoff;
2804 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2805 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2806 zba = details->first_index;
2807 if (zba < vba)
2808 zba = vba;
2809 zea = details->last_index;
2810 if (zea > vea)
2811 zea = vea;
2812
Peter Zijlstra97a89412011-05-24 17:12:04 -07002813 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2815 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002816 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 }
2818}
2819
2820static inline void unmap_mapping_range_list(struct list_head *head,
2821 struct zap_details *details)
2822{
2823 struct vm_area_struct *vma;
2824
2825 /*
2826 * In nonlinear VMAs there is no correspondence between virtual address
2827 * offset and file offset. So we must perform an exhaustive search
2828 * across *all* the pages in each nonlinear VMA, not just the pages
2829 * whose virtual address lies outside the file truncation point.
2830 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 list_for_each_entry(vma, head, shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002833 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 }
2835}
2836
2837/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002838 * 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 -07002839 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 * @holebegin: byte in first page to unmap, relative to the start of
2841 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002842 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 * must keep the partial page. In contrast, we must get rid of
2844 * partial pages.
2845 * @holelen: size of prospective hole in bytes. This will be rounded
2846 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2847 * end of the file.
2848 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2849 * but 0 when invalidating pagecache, don't throw away private data.
2850 */
2851void unmap_mapping_range(struct address_space *mapping,
2852 loff_t const holebegin, loff_t const holelen, int even_cows)
2853{
2854 struct zap_details details;
2855 pgoff_t hba = holebegin >> PAGE_SHIFT;
2856 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857
2858 /* Check for overflow. */
2859 if (sizeof(holelen) > sizeof(hlen)) {
2860 long long holeend =
2861 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862 if (holeend & ~(long long)ULONG_MAX)
2863 hlen = ULONG_MAX - hba + 1;
2864 }
2865
2866 details.check_mapping = even_cows? NULL: mapping;
2867 details.nonlinear_vma = NULL;
2868 details.first_index = hba;
2869 details.last_index = hba + hlen - 1;
2870 if (details.last_index < details.first_index)
2871 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872
Peter Zijlstra97a89412011-05-24 17:12:04 -07002873
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002874 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2876 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2877 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2878 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002879 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880}
2881EXPORT_SYMBOL(unmap_mapping_range);
2882
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002884 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2885 * but allow concurrent faults), and pte mapped but not yet locked.
2886 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002888static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2889 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002890 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002892 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002893 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002894 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002896 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002897 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002898 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002899 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002901 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002902 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002903
2904 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002905 if (unlikely(non_swap_entry(entry))) {
2906 if (is_migration_entry(entry)) {
2907 migration_entry_wait(mm, pmd, address);
2908 } else if (is_hwpoison_entry(entry)) {
2909 ret = VM_FAULT_HWPOISON;
2910 } else {
2911 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002912 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002913 }
Christoph Lameter06972122006-06-23 02:03:35 -07002914 goto out;
2915 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002916 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917 page = lookup_swap_cache(entry);
2918 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002919 page = swapin_readahead(entry,
2920 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 if (!page) {
2922 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002923 * Back out if somebody else faulted in this pte
2924 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002926 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 if (likely(pte_same(*page_table, orig_pte)))
2928 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002929 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002930 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 }
2932
2933 /* Had to read the page from swap area: Major fault */
2934 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002935 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002936 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002937 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002938 /*
2939 * hwpoisoned dirty swapcache pages are kept for killing
2940 * owner processes (which may be unknown at hwpoison time)
2941 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002942 ret = VM_FAULT_HWPOISON;
2943 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002944 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 }
2946
Michel Lespinassed065bd82010-10-26 14:21:57 -07002947 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002948
Balbir Singh20a10222007-11-14 17:00:33 -08002949 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002950 if (!locked) {
2951 ret |= VM_FAULT_RETRY;
2952 goto out_release;
2953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002955 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002956 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2957 * release the swapcache from under us. The page pin, and pte_same
2958 * test below, are not enough to exclude that. Even if it is still
2959 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002960 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002961 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002962 goto out_page;
2963
2964 if (ksm_might_need_to_copy(page, vma, address)) {
2965 swapcache = page;
2966 page = ksm_does_need_to_copy(page, vma, address);
2967
2968 if (unlikely(!page)) {
2969 ret = VM_FAULT_OOM;
2970 page = swapcache;
2971 swapcache = NULL;
2972 goto out_page;
2973 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002974 }
2975
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002976 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002977 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002978 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002979 }
2980
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002982 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002984 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002985 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002986 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002987
2988 if (unlikely(!PageUptodate(page))) {
2989 ret = VM_FAULT_SIGBUS;
2990 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 }
2992
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002993 /*
2994 * The page isn't present yet, go ahead with the fault.
2995 *
2996 * Be careful about the sequence of operations here.
2997 * To get its accounting right, reuse_swap_page() must be called
2998 * while the page is counted on swap but not yet in mapcount i.e.
2999 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3000 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003001 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3002 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3003 * in page->private. In this case, a record in swap_cgroup is silently
3004 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003005 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003007 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003008 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003010 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003012 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003013 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003014 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 flush_icache_page(vma, page);
3017 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003018 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003019 /* It's better to call commit-charge after rmap is established */
3020 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003022 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003023 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003024 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003025 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003026 if (swapcache) {
3027 /*
3028 * Hold the lock to avoid the swap entry to be reused
3029 * until we take the PT lock for the pte_same() check
3030 * (to avoid false positives from pte_same). For
3031 * further safety release the lock after the swap_free
3032 * so that the swap count won't change under a
3033 * parallel locked swapcache.
3034 */
3035 unlock_page(swapcache);
3036 page_cache_release(swapcache);
3037 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003038
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003039 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003040 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3041 if (ret & VM_FAULT_ERROR)
3042 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 goto out;
3044 }
3045
3046 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003047 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003048unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003049 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050out:
3051 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003052out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003053 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003054 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003055out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003056 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003057out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003058 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003059 if (swapcache) {
3060 unlock_page(swapcache);
3061 page_cache_release(swapcache);
3062 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003063 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064}
3065
3066/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003067 * This is like a special single-page "expand_{down|up}wards()",
3068 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003069 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003070 */
3071static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3072{
3073 address &= PAGE_MASK;
3074 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003075 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003076
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003077 /*
3078 * Is there a mapping abutting this one below?
3079 *
3080 * That's only ok if it's the same stack mapping
3081 * that has gotten split..
3082 */
3083 if (prev && prev->vm_end == address)
3084 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3085
Michal Hockod05f3162011-05-24 17:11:44 -07003086 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003087 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003088 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3089 struct vm_area_struct *next = vma->vm_next;
3090
3091 /* As VM_GROWSDOWN but s/below/above/ */
3092 if (next && next->vm_start == address + PAGE_SIZE)
3093 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3094
3095 expand_upwards(vma, address + PAGE_SIZE);
3096 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003097 return 0;
3098}
3099
3100/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003101 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3102 * but allow concurrent faults), and pte mapped but not yet locked.
3103 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003105static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3106 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003107 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003109 struct page *page;
3110 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112
Linus Torvalds11ac5522010-08-14 11:44:56 -07003113 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003114
Linus Torvalds11ac5522010-08-14 11:44:56 -07003115 /* Check if we need to add a guard page to the stack */
3116 if (check_stack_guard_page(vma, address) < 0)
3117 return VM_FAULT_SIGBUS;
3118
3119 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003120 if (!(flags & FAULT_FLAG_WRITE)) {
3121 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3122 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003123 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003124 if (!pte_none(*page_table))
3125 goto unlock;
3126 goto setpte;
3127 }
3128
Nick Piggin557ed1f2007-10-16 01:24:40 -07003129 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003130 if (unlikely(anon_vma_prepare(vma)))
3131 goto oom;
3132 page = alloc_zeroed_user_highpage_movable(vma, address);
3133 if (!page)
3134 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003135 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003137 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003138 goto oom_free_page;
3139
Nick Piggin557ed1f2007-10-16 01:24:40 -07003140 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003141 if (vma->vm_flags & VM_WRITE)
3142 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003143
Nick Piggin557ed1f2007-10-16 01:24:40 -07003144 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003145 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003146 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003147
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003148 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003149 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003150setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003151 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152
3153 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003154 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003155unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003156 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003157 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003158release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003159 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003160 page_cache_release(page);
3161 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003162oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003163 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003164oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 return VM_FAULT_OOM;
3166}
3167
3168/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003169 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003171 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3172 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 *
3174 * As this is called only for pages that do not currently exist, we
3175 * do not need to flush old virtual caches or the TLB.
3176 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003178 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003179 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003181static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003182 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003183 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003185 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003186 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003187 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003188 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003191 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003192 struct vm_fault vmf;
3193 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003194 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003195
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003196 /*
3197 * If we do COW later, allocate page befor taking lock_page()
3198 * on the file cache page. This will reduce lock holding time.
3199 */
3200 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3201
3202 if (unlikely(anon_vma_prepare(vma)))
3203 return VM_FAULT_OOM;
3204
3205 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3206 if (!cow_page)
3207 return VM_FAULT_OOM;
3208
3209 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3210 page_cache_release(cow_page);
3211 return VM_FAULT_OOM;
3212 }
3213 } else
3214 cow_page = NULL;
3215
Nick Piggind0217ac2007-07-19 01:47:03 -07003216 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3217 vmf.pgoff = pgoff;
3218 vmf.flags = flags;
3219 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003221 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003222 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3223 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003224 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225
Andi Kleena3b947e2009-09-16 11:50:08 +02003226 if (unlikely(PageHWPoison(vmf.page))) {
3227 if (ret & VM_FAULT_LOCKED)
3228 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003229 ret = VM_FAULT_HWPOISON;
3230 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003231 }
3232
Nick Piggind00806b2007-07-19 01:46:57 -07003233 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003234 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003235 * locked.
3236 */
Nick Piggin83c54072007-07-19 01:47:05 -07003237 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003238 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003239 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003240 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003241
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 /*
3243 * Should we do an early C-O-W break?
3244 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003245 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003246 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003247 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003248 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003249 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003250 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003251 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003252 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003253 /*
3254 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003255 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003256 * to become writable
3257 */
Mark Fasheh69676142007-07-19 01:47:00 -07003258 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003259 int tmp;
3260
Mark Fasheh69676142007-07-19 01:47:00 -07003261 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003262 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003263 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3264 if (unlikely(tmp &
3265 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3266 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003267 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003268 }
Nick Pigginb827e492009-04-30 15:08:16 -07003269 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3270 lock_page(page);
3271 if (!page->mapping) {
3272 ret = 0; /* retry the fault */
3273 unlock_page(page);
3274 goto unwritable_page;
3275 }
3276 } else
3277 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003278 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003279 }
3280 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003281
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 }
3283
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003284 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285
3286 /*
3287 * This silly early PAGE_DIRTY setting removes a race
3288 * due to the bad i386 page protection. But it's valid
3289 * for other architectures too.
3290 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003291 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 * an exclusive copy of the page, or this is a shared mapping,
3293 * so we can make it writable and dirty to avoid having to
3294 * handle that later.
3295 */
3296 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003297 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003298 flush_icache_page(vma, page);
3299 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003300 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003303 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003304 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003305 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003306 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003307 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003308 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003309 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003310 get_page(dirty_page);
3311 }
Hugh Dickins42946212005-10-29 18:16:05 -07003312 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003313 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003314
3315 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003316 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003318 if (cow_page)
3319 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003320 if (anon)
3321 page_cache_release(page);
3322 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003323 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 }
3325
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003326 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003327
Nick Pigginb827e492009-04-30 15:08:16 -07003328 if (dirty_page) {
3329 struct address_space *mapping = page->mapping;
3330
3331 if (set_page_dirty(dirty_page))
3332 page_mkwrite = 1;
3333 unlock_page(dirty_page);
3334 put_page(dirty_page);
3335 if (page_mkwrite && mapping) {
3336 /*
3337 * Some device drivers do not set page.mapping but still
3338 * dirty their pages
3339 */
3340 balance_dirty_pages_ratelimited(mapping);
3341 }
3342
3343 /* file_update_time outside page_lock */
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003344 if (vma->vm_file)
3345 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003346 } else {
3347 unlock_page(vmf.page);
3348 if (anon)
3349 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003350 }
Nick Piggind00806b2007-07-19 01:46:57 -07003351
Nick Piggin83c54072007-07-19 01:47:05 -07003352 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003353
3354unwritable_page:
3355 page_cache_release(page);
3356 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003357uncharge_out:
3358 /* fs's fault handler get error */
3359 if (cow_page) {
3360 mem_cgroup_uncharge_page(cow_page);
3361 page_cache_release(cow_page);
3362 }
3363 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003364}
Nick Piggind00806b2007-07-19 01:46:57 -07003365
Nick Piggin54cb8822007-07-19 01:46:59 -07003366static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3367 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003368 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003369{
3370 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003371 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003372
Hugh Dickins16abfa02007-10-04 16:56:06 +01003373 pte_unmap(page_table);
3374 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003375}
3376
Jes Sorensenf4b81802006-09-27 01:50:10 -07003377/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 * Fault of a previously existing named mapping. Repopulate the pte
3379 * from the encoded file_pte if possible. This enables swappable
3380 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003381 *
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003386static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003387 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003388 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389{
Hugh Dickins65500d22005-10-29 18:15:59 -07003390 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003392 flags |= FAULT_FLAG_NONLINEAR;
3393
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003394 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003395 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
Hugh Dickins2509ef22009-01-06 14:40:10 -08003397 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003398 /*
3399 * Page table corrupted: show pte and kill process.
3400 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003401 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003402 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003403 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003404
3405 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003406 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407}
3408
3409/*
3410 * These routines also need to handle stuff like marking pages dirty
3411 * and/or accessed for architectures that don't do it in hardware (most
3412 * RISC architectures). The early dirtying is also good on the i386.
3413 *
3414 * There is also a hook called "update_mmu_cache()" that architectures
3415 * with external mmu caches can use to update those (ie the Sparc or
3416 * PowerPC hashed page tables that act as extended TLBs).
3417 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003418 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3419 * but allow concurrent faults), and pte mapped but not yet locked.
3420 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003422int handle_pte_fault(struct mm_struct *mm,
3423 struct vm_area_struct *vma, unsigned long address,
3424 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425{
3426 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003427 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003429 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003431 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003432 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003433 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003434 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003435 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003436 }
3437 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003438 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003439 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003441 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003442 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003443 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003444 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 }
3446
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003447 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003448 spin_lock(ptl);
3449 if (unlikely(!pte_same(*pte, entry)))
3450 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003451 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003453 return do_wp_page(mm, vma, address,
3454 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455 entry = pte_mkdirty(entry);
3456 }
3457 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003458 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003459 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003460 } else {
3461 /*
3462 * This is needed only for protection faults but the arch code
3463 * is not yet telling us if this is a protection fault or not.
3464 * This still avoids useless tlb flushes for .text page faults
3465 * with threads.
3466 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003467 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003468 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003469 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003470unlock:
3471 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003472 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473}
3474
3475/*
3476 * By the time we get here, we already hold the mm semaphore
3477 */
Nick Piggin83c54072007-07-19 01:47:05 -07003478int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003479 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480{
3481 pgd_t *pgd;
3482 pud_t *pud;
3483 pmd_t *pmd;
3484 pte_t *pte;
3485
3486 __set_current_state(TASK_RUNNING);
3487
Christoph Lameterf8891e52006-06-30 01:55:45 -07003488 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003489 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003491 /* do counter updates before entering really critical section. */
3492 check_sync_rss_stat(current);
3493
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003494 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003495 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496
David Rientjes1f1d06c2012-05-29 15:06:23 -07003497retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 pud = pud_alloc(mm, pgd, address);
3500 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003501 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 pmd = pmd_alloc(mm, pud, address);
3503 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003504 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003505 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3506 if (!vma->vm_ops)
3507 return do_huge_pmd_anonymous_page(mm, vma, address,
3508 pmd, flags);
3509 } else {
3510 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003511 int ret;
3512
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003513 barrier();
3514 if (pmd_trans_huge(orig_pmd)) {
3515 if (flags & FAULT_FLAG_WRITE &&
3516 !pmd_write(orig_pmd) &&
David Rientjes1f1d06c2012-05-29 15:06:23 -07003517 !pmd_trans_splitting(orig_pmd)) {
3518 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3519 orig_pmd);
3520 /*
3521 * If COW results in an oom, the huge pmd will
3522 * have been split, so retry the fault on the
3523 * pte for a smaller charge.
3524 */
3525 if (unlikely(ret & VM_FAULT_OOM))
3526 goto retry;
3527 return ret;
3528 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003529 return 0;
3530 }
3531 }
3532
3533 /*
3534 * Use __pte_alloc instead of pte_alloc_map, because we can't
3535 * run pte_offset_map on the pmd, if an huge pmd could
3536 * materialize from under us from a different thread.
3537 */
Mel Gormancc036382011-04-27 15:26:56 -07003538 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003539 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003540 /* if an huge pmd materialized from under us just retry later */
3541 if (unlikely(pmd_trans_huge(*pmd)))
3542 return 0;
3543 /*
3544 * A regular pmd is established and it can't morph into a huge pmd
3545 * from under us anymore at this point because we hold the mmap_sem
3546 * read mode and khugepaged takes it in write mode. So now it's
3547 * safe to run pte_offset_map().
3548 */
3549 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003551 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552}
3553
3554#ifndef __PAGETABLE_PUD_FOLDED
3555/*
3556 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003557 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003559int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003561 pud_t *new = pud_alloc_one(mm, address);
3562 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003563 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564
Nick Piggin362a61a2008-05-14 06:37:36 +02003565 smp_wmb(); /* See comment in __pte_alloc */
3566
Hugh Dickins872fec12005-10-29 18:16:21 -07003567 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003568 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003569 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003570 else
3571 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003572 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003573 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
3575#endif /* __PAGETABLE_PUD_FOLDED */
3576
3577#ifndef __PAGETABLE_PMD_FOLDED
3578/*
3579 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003580 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003582int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003584 pmd_t *new = pmd_alloc_one(mm, address);
3585 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003586 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
Nick Piggin362a61a2008-05-14 06:37:36 +02003588 smp_wmb(); /* See comment in __pte_alloc */
3589
Hugh Dickins872fec12005-10-29 18:16:21 -07003590 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003592 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003593 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003594 else
3595 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003597 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003598 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003599 else
3600 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003602 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003603 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604}
3605#endif /* __PAGETABLE_PMD_FOLDED */
3606
3607int make_pages_present(unsigned long addr, unsigned long end)
3608{
3609 int ret, len, write;
3610 struct vm_area_struct * vma;
3611
3612 vma = find_vma(current->mm, addr);
3613 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003614 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003615 /*
3616 * We want to touch writable mappings with a write fault in order
3617 * to break COW, except for shared mappings because these don't COW
3618 * and we would not want to dirty them for nothing.
3619 */
3620 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003621 BUG_ON(addr >= end);
3622 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003623 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 ret = get_user_pages(current, current->mm, addr,
3625 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003626 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003628 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629}
3630
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631#if !defined(__HAVE_ARCH_GATE_AREA)
3632
3633#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003634static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635
3636static int __init gate_vma_init(void)
3637{
3638 gate_vma.vm_mm = NULL;
3639 gate_vma.vm_start = FIXADDR_USER_START;
3640 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003641 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3642 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003643
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 return 0;
3645}
3646__initcall(gate_vma_init);
3647#endif
3648
Stephen Wilson31db58b2011-03-13 15:49:15 -04003649struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650{
3651#ifdef AT_SYSINFO_EHDR
3652 return &gate_vma;
3653#else
3654 return NULL;
3655#endif
3656}
3657
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003658int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659{
3660#ifdef AT_SYSINFO_EHDR
3661 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3662 return 1;
3663#endif
3664 return 0;
3665}
3666
3667#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003668
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003669static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003670 pte_t **ptepp, spinlock_t **ptlp)
3671{
3672 pgd_t *pgd;
3673 pud_t *pud;
3674 pmd_t *pmd;
3675 pte_t *ptep;
3676
3677 pgd = pgd_offset(mm, address);
3678 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3679 goto out;
3680
3681 pud = pud_offset(pgd, address);
3682 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3683 goto out;
3684
3685 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003686 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003687 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3688 goto out;
3689
3690 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3691 if (pmd_huge(*pmd))
3692 goto out;
3693
3694 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3695 if (!ptep)
3696 goto out;
3697 if (!pte_present(*ptep))
3698 goto unlock;
3699 *ptepp = ptep;
3700 return 0;
3701unlock:
3702 pte_unmap_unlock(ptep, *ptlp);
3703out:
3704 return -EINVAL;
3705}
3706
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003707static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3708 pte_t **ptepp, spinlock_t **ptlp)
3709{
3710 int res;
3711
3712 /* (void) is needed to make gcc happy */
3713 (void) __cond_lock(*ptlp,
3714 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3715 return res;
3716}
3717
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003718/**
3719 * follow_pfn - look up PFN at a user virtual address
3720 * @vma: memory mapping
3721 * @address: user virtual address
3722 * @pfn: location to store found PFN
3723 *
3724 * Only IO mappings and raw PFN mappings are allowed.
3725 *
3726 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3727 */
3728int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3729 unsigned long *pfn)
3730{
3731 int ret = -EINVAL;
3732 spinlock_t *ptl;
3733 pte_t *ptep;
3734
3735 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3736 return ret;
3737
3738 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3739 if (ret)
3740 return ret;
3741 *pfn = pte_pfn(*ptep);
3742 pte_unmap_unlock(ptep, ptl);
3743 return 0;
3744}
3745EXPORT_SYMBOL(follow_pfn);
3746
Rik van Riel28b2ee22008-07-23 21:27:05 -07003747#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003748int follow_phys(struct vm_area_struct *vma,
3749 unsigned long address, unsigned int flags,
3750 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003751{
Johannes Weiner03668a42009-06-16 15:32:34 -07003752 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003753 pte_t *ptep, pte;
3754 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003755
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003756 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3757 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003758
Johannes Weiner03668a42009-06-16 15:32:34 -07003759 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003760 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003761 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003762
Rik van Riel28b2ee22008-07-23 21:27:05 -07003763 if ((flags & FOLL_WRITE) && !pte_write(pte))
3764 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003765
3766 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003767 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003768
Johannes Weiner03668a42009-06-16 15:32:34 -07003769 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003770unlock:
3771 pte_unmap_unlock(ptep, ptl);
3772out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003773 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003774}
3775
3776int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3777 void *buf, int len, int write)
3778{
3779 resource_size_t phys_addr;
3780 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003781 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003782 int offset = addr & (PAGE_SIZE-1);
3783
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003784 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003785 return -EINVAL;
3786
3787 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3788 if (write)
3789 memcpy_toio(maddr + offset, buf, len);
3790 else
3791 memcpy_fromio(buf, maddr + offset, len);
3792 iounmap(maddr);
3793
3794 return len;
3795}
3796#endif
3797
David Howells0ec76a12006-09-27 01:50:15 -07003798/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003799 * Access another process' address space as given in mm. If non-NULL, use the
3800 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003801 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003802static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3803 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003804{
David Howells0ec76a12006-09-27 01:50:15 -07003805 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003806 void *old_buf = buf;
3807
David Howells0ec76a12006-09-27 01:50:15 -07003808 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003809 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003810 while (len) {
3811 int bytes, ret, offset;
3812 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003813 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003814
3815 ret = get_user_pages(tsk, mm, addr, 1,
3816 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003817 if (ret <= 0) {
3818 /*
3819 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3820 * we can access using slightly different code.
3821 */
3822#ifdef CONFIG_HAVE_IOREMAP_PROT
3823 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003824 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003825 break;
3826 if (vma->vm_ops && vma->vm_ops->access)
3827 ret = vma->vm_ops->access(vma, addr, buf,
3828 len, write);
3829 if (ret <= 0)
3830#endif
3831 break;
3832 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003833 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003834 bytes = len;
3835 offset = addr & (PAGE_SIZE-1);
3836 if (bytes > PAGE_SIZE-offset)
3837 bytes = PAGE_SIZE-offset;
3838
3839 maddr = kmap(page);
3840 if (write) {
3841 copy_to_user_page(vma, page, addr,
3842 maddr + offset, buf, bytes);
3843 set_page_dirty_lock(page);
3844 } else {
3845 copy_from_user_page(vma, page, addr,
3846 buf, maddr + offset, bytes);
3847 }
3848 kunmap(page);
3849 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003850 }
David Howells0ec76a12006-09-27 01:50:15 -07003851 len -= bytes;
3852 buf += bytes;
3853 addr += bytes;
3854 }
3855 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003856
3857 return buf - old_buf;
3858}
Andi Kleen03252912008-01-30 13:33:18 +01003859
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003860/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003861 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003862 * @mm: the mm_struct of the target address space
3863 * @addr: start address to access
3864 * @buf: source or destination buffer
3865 * @len: number of bytes to transfer
3866 * @write: whether the access is a write
3867 *
3868 * The caller must hold a reference on @mm.
3869 */
3870int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3871 void *buf, int len, int write)
3872{
3873 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3874}
3875
Andi Kleen03252912008-01-30 13:33:18 +01003876/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003877 * Access another process' address space.
3878 * Source/target buffer must be kernel space,
3879 * Do not walk the page table directly, use get_user_pages
3880 */
3881int access_process_vm(struct task_struct *tsk, unsigned long addr,
3882 void *buf, int len, int write)
3883{
3884 struct mm_struct *mm;
3885 int ret;
3886
3887 mm = get_task_mm(tsk);
3888 if (!mm)
3889 return 0;
3890
3891 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3892 mmput(mm);
3893
3894 return ret;
3895}
3896
Andi Kleen03252912008-01-30 13:33:18 +01003897/*
3898 * Print the name of a VMA.
3899 */
3900void print_vma_addr(char *prefix, unsigned long ip)
3901{
3902 struct mm_struct *mm = current->mm;
3903 struct vm_area_struct *vma;
3904
Ingo Molnare8bff742008-02-13 20:21:06 +01003905 /*
3906 * Do not print if we are in atomic
3907 * contexts (in exception stacks, etc.):
3908 */
3909 if (preempt_count())
3910 return;
3911
Andi Kleen03252912008-01-30 13:33:18 +01003912 down_read(&mm->mmap_sem);
3913 vma = find_vma(mm, ip);
3914 if (vma && vma->vm_file) {
3915 struct file *f = vma->vm_file;
3916 char *buf = (char *)__get_free_page(GFP_KERNEL);
3917 if (buf) {
3918 char *p, *s;
3919
Jan Blunckcf28b482008-02-14 19:38:44 -08003920 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003921 if (IS_ERR(p))
3922 p = "?";
3923 s = strrchr(p, '/');
3924 if (s)
3925 p = s+1;
3926 printk("%s%s[%lx+%lx]", prefix, p,
3927 vma->vm_start,
3928 vma->vm_end - vma->vm_start);
3929 free_page((unsigned long)buf);
3930 }
3931 }
3932 up_read(&current->mm->mmap_sem);
3933}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003934
3935#ifdef CONFIG_PROVE_LOCKING
3936void might_fault(void)
3937{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003938 /*
3939 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3940 * holding the mmap_sem, this is safe because kernel memory doesn't
3941 * get paged out, therefore we'll never actually fault, and the
3942 * below annotations will generate false positives.
3943 */
3944 if (segment_eq(get_fs(), KERNEL_DS))
3945 return;
3946
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003947 might_sleep();
3948 /*
3949 * it would be nicer only to annotate paths which are not under
3950 * pagefault_disable, however that requires a larger audit and
3951 * providing helpers like get_user_atomic.
3952 */
3953 if (!in_atomic() && current->mm)
3954 might_lock_read(&current->mm->mmap_sem);
3955}
3956EXPORT_SYMBOL(might_fault);
3957#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003958
3959#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3960static void clear_gigantic_page(struct page *page,
3961 unsigned long addr,
3962 unsigned int pages_per_huge_page)
3963{
3964 int i;
3965 struct page *p = page;
3966
3967 might_sleep();
3968 for (i = 0; i < pages_per_huge_page;
3969 i++, p = mem_map_next(p, page, i)) {
3970 cond_resched();
3971 clear_user_highpage(p, addr + i * PAGE_SIZE);
3972 }
3973}
3974void clear_huge_page(struct page *page,
3975 unsigned long addr, unsigned int pages_per_huge_page)
3976{
3977 int i;
3978
3979 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3980 clear_gigantic_page(page, addr, pages_per_huge_page);
3981 return;
3982 }
3983
3984 might_sleep();
3985 for (i = 0; i < pages_per_huge_page; i++) {
3986 cond_resched();
3987 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3988 }
3989}
3990
3991static void copy_user_gigantic_page(struct page *dst, struct page *src,
3992 unsigned long addr,
3993 struct vm_area_struct *vma,
3994 unsigned int pages_per_huge_page)
3995{
3996 int i;
3997 struct page *dst_base = dst;
3998 struct page *src_base = src;
3999
4000 for (i = 0; i < pages_per_huge_page; ) {
4001 cond_resched();
4002 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4003
4004 i++;
4005 dst = mem_map_next(dst, dst_base, i);
4006 src = mem_map_next(src, src_base, i);
4007 }
4008}
4009
4010void copy_user_huge_page(struct page *dst, struct page *src,
4011 unsigned long addr, struct vm_area_struct *vma,
4012 unsigned int pages_per_huge_page)
4013{
4014 int i;
4015
4016 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4017 copy_user_gigantic_page(dst, src, addr, vma,
4018 pages_per_huge_page);
4019 return;
4020 }
4021
4022 might_sleep();
4023 for (i = 0; i < pages_per_huge_page; i++) {
4024 cond_resched();
4025 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4026 }
4027}
4028#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */