blob: 57361708d1a57d7bc11c8f34d269a35c50317dbb [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;
Alex Shi597e1c32012-06-28 09:02:21 +0800209 tlb->start = -1UL;
210 tlb->end = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700211 tlb->need_flush = 0;
212 tlb->fast_mode = (num_possible_cpus() == 1);
213 tlb->local.next = NULL;
214 tlb->local.nr = 0;
215 tlb->local.max = ARRAY_SIZE(tlb->__pages);
216 tlb->active = &tlb->local;
217
218#ifdef CONFIG_HAVE_RCU_TABLE_FREE
219 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800220#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700221}
222
223void tlb_flush_mmu(struct mmu_gather *tlb)
224{
225 struct mmu_gather_batch *batch;
226
227 if (!tlb->need_flush)
228 return;
229 tlb->need_flush = 0;
230 tlb_flush(tlb);
231#ifdef CONFIG_HAVE_RCU_TABLE_FREE
232 tlb_table_flush(tlb);
233#endif
234
235 if (tlb_fast_mode(tlb))
236 return;
237
238 for (batch = &tlb->local; batch; batch = batch->next) {
239 free_pages_and_swap_cache(batch->pages, batch->nr);
240 batch->nr = 0;
241 }
242 tlb->active = &tlb->local;
243}
244
245/* tlb_finish_mmu
246 * Called at the end of the shootdown operation to free up any resources
247 * that were required.
248 */
249void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
250{
251 struct mmu_gather_batch *batch, *next;
252
Alex Shi597e1c32012-06-28 09:02:21 +0800253 tlb->start = start;
254 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700255 tlb_flush_mmu(tlb);
256
257 /* keep the page table cache within bounds */
258 check_pgt_cache();
259
260 for (batch = tlb->local.next; batch; batch = next) {
261 next = batch->next;
262 free_pages((unsigned long)batch, 0);
263 }
264 tlb->local.next = NULL;
265}
266
267/* __tlb_remove_page
268 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
269 * handling the additional races in SMP caused by other CPUs caching valid
270 * mappings in their TLBs. Returns the number of free page slots left.
271 * When out of page slots we must call tlb_flush_mmu().
272 */
273int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
274{
275 struct mmu_gather_batch *batch;
276
Shaohua Lif21760b2012-01-12 17:19:16 -0800277 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700278
279 if (tlb_fast_mode(tlb)) {
280 free_page_and_swap_cache(page);
281 return 1; /* avoid calling tlb_flush_mmu() */
282 }
283
284 batch = tlb->active;
285 batch->pages[batch->nr++] = page;
286 if (batch->nr == batch->max) {
287 if (!tlb_next_batch(tlb))
288 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700289 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700290 }
291 VM_BUG_ON(batch->nr > batch->max);
292
293 return batch->max - batch->nr;
294}
295
296#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800297
Peter Zijlstra26723912011-05-24 17:12:00 -0700298#ifdef CONFIG_HAVE_RCU_TABLE_FREE
299
300/*
301 * See the comment near struct mmu_table_batch.
302 */
303
304static void tlb_remove_table_smp_sync(void *arg)
305{
306 /* Simply deliver the interrupt */
307}
308
309static void tlb_remove_table_one(void *table)
310{
311 /*
312 * This isn't an RCU grace period and hence the page-tables cannot be
313 * assumed to be actually RCU-freed.
314 *
315 * It is however sufficient for software page-table walkers that rely on
316 * IRQ disabling. See the comment near struct mmu_table_batch.
317 */
318 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
319 __tlb_remove_table(table);
320}
321
322static void tlb_remove_table_rcu(struct rcu_head *head)
323{
324 struct mmu_table_batch *batch;
325 int i;
326
327 batch = container_of(head, struct mmu_table_batch, rcu);
328
329 for (i = 0; i < batch->nr; i++)
330 __tlb_remove_table(batch->tables[i]);
331
332 free_page((unsigned long)batch);
333}
334
335void tlb_table_flush(struct mmu_gather *tlb)
336{
337 struct mmu_table_batch **batch = &tlb->batch;
338
339 if (*batch) {
340 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341 *batch = NULL;
342 }
343}
344
345void tlb_remove_table(struct mmu_gather *tlb, void *table)
346{
347 struct mmu_table_batch **batch = &tlb->batch;
348
349 tlb->need_flush = 1;
350
351 /*
352 * When there's less then two users of this mm there cannot be a
353 * concurrent page-table walk.
354 */
355 if (atomic_read(&tlb->mm->mm_users) < 2) {
356 __tlb_remove_table(table);
357 return;
358 }
359
360 if (*batch == NULL) {
361 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
362 if (*batch == NULL) {
363 tlb_remove_table_one(table);
364 return;
365 }
366 (*batch)->nr = 0;
367 }
368 (*batch)->tables[(*batch)->nr++] = table;
369 if ((*batch)->nr == MAX_TABLE_BATCH)
370 tlb_table_flush(tlb);
371}
372
Peter Zijlstra9547d012011-05-24 17:12:14 -0700373#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700374
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375/*
376 * If a p?d_bad entry is found while walking page tables, report
377 * the error, before resetting entry to p?d_none. Usually (but
378 * very seldom) called out from the p?d_none_or_clear_bad macros.
379 */
380
381void pgd_clear_bad(pgd_t *pgd)
382{
383 pgd_ERROR(*pgd);
384 pgd_clear(pgd);
385}
386
387void pud_clear_bad(pud_t *pud)
388{
389 pud_ERROR(*pud);
390 pud_clear(pud);
391}
392
393void pmd_clear_bad(pmd_t *pmd)
394{
395 pmd_ERROR(*pmd);
396 pmd_clear(pmd);
397}
398
399/*
400 * Note: this doesn't free the actual pages themselves. That
401 * has been handled earlier when unmapping all the memory regions.
402 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000403static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
404 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800406 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700407 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000408 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700409 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410}
411
Hugh Dickinse0da3822005-04-19 13:29:15 -0700412static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
413 unsigned long addr, unsigned long end,
414 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415{
416 pmd_t *pmd;
417 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700418 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Hugh Dickinse0da3822005-04-19 13:29:15 -0700420 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 do {
423 next = pmd_addr_end(addr, end);
424 if (pmd_none_or_clear_bad(pmd))
425 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000426 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 } while (pmd++, addr = next, addr != end);
428
Hugh Dickinse0da3822005-04-19 13:29:15 -0700429 start &= PUD_MASK;
430 if (start < floor)
431 return;
432 if (ceiling) {
433 ceiling &= PUD_MASK;
434 if (!ceiling)
435 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700437 if (end - 1 > ceiling - 1)
438 return;
439
440 pmd = pmd_offset(pud, start);
441 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000442 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
Hugh Dickinse0da3822005-04-19 13:29:15 -0700445static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
446 unsigned long addr, unsigned long end,
447 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448{
449 pud_t *pud;
450 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700451 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Hugh Dickinse0da3822005-04-19 13:29:15 -0700453 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 do {
456 next = pud_addr_end(addr, end);
457 if (pud_none_or_clear_bad(pud))
458 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700459 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 } while (pud++, addr = next, addr != end);
461
Hugh Dickinse0da3822005-04-19 13:29:15 -0700462 start &= PGDIR_MASK;
463 if (start < floor)
464 return;
465 if (ceiling) {
466 ceiling &= PGDIR_MASK;
467 if (!ceiling)
468 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700470 if (end - 1 > ceiling - 1)
471 return;
472
473 pud = pud_offset(pgd, start);
474 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000475 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476}
477
478/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700479 * This function frees user-level page tables of a process.
480 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 * Must be called with pagetable lock held.
482 */
Jan Beulich42b77722008-07-23 21:27:10 -0700483void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700484 unsigned long addr, unsigned long end,
485 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486{
487 pgd_t *pgd;
488 unsigned long next;
489
Hugh Dickinse0da3822005-04-19 13:29:15 -0700490 /*
491 * The next few lines have given us lots of grief...
492 *
493 * Why are we testing PMD* at this top level? Because often
494 * there will be no work to do at all, and we'd prefer not to
495 * go all the way down to the bottom just to discover that.
496 *
497 * Why all these "- 1"s? Because 0 represents both the bottom
498 * of the address space and the top of it (using -1 for the
499 * top wouldn't help much: the masks would do the wrong thing).
500 * The rule is that addr 0 and floor 0 refer to the bottom of
501 * the address space, but end 0 and ceiling 0 refer to the top
502 * Comparisons need to use "end - 1" and "ceiling - 1" (though
503 * that end 0 case should be mythical).
504 *
505 * Wherever addr is brought up or ceiling brought down, we must
506 * be careful to reject "the opposite 0" before it confuses the
507 * subsequent tests. But what about where end is brought down
508 * by PMD_SIZE below? no, end can't go down to 0 there.
509 *
510 * Whereas we round start (addr) and ceiling down, by different
511 * masks at different levels, in order to test whether a table
512 * now has no other vmas using it, so can be freed, we don't
513 * bother to round floor or end up - the tests don't need that.
514 */
515
516 addr &= PMD_MASK;
517 if (addr < floor) {
518 addr += PMD_SIZE;
519 if (!addr)
520 return;
521 }
522 if (ceiling) {
523 ceiling &= PMD_MASK;
524 if (!ceiling)
525 return;
526 }
527 if (end - 1 > ceiling - 1)
528 end -= PMD_SIZE;
529 if (addr > end - 1)
530 return;
531
Jan Beulich42b77722008-07-23 21:27:10 -0700532 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 do {
534 next = pgd_addr_end(addr, end);
535 if (pgd_none_or_clear_bad(pgd))
536 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700537 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700539}
540
Jan Beulich42b77722008-07-23 21:27:10 -0700541void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700542 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700543{
544 while (vma) {
545 struct vm_area_struct *next = vma->vm_next;
546 unsigned long addr = vma->vm_start;
547
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700548 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000549 * Hide vma from rmap and truncate_pagecache before freeing
550 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700551 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800552 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700553 unlink_file_vma(vma);
554
David Gibson9da61ae2006-03-22 00:08:57 -0800555 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700556 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700557 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700558 } else {
559 /*
560 * Optimization: gather nearby vmas into one call down
561 */
562 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800563 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700564 vma = next;
565 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800566 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700567 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700568 }
569 free_pgd_range(tlb, addr, vma->vm_end,
570 floor, next? next->vm_start: ceiling);
571 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700572 vma = next;
573 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800576int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
577 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800579 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800580 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700581 if (!new)
582 return -ENOMEM;
583
Nick Piggin362a61a2008-05-14 06:37:36 +0200584 /*
585 * Ensure all pte setup (eg. pte page lock and page clearing) are
586 * visible before the pte is made visible to other CPUs by being
587 * put into page tables.
588 *
589 * The other side of the story is the pointer chasing in the page
590 * table walking code (when walking the page table without locking;
591 * ie. most of the time). Fortunately, these data accesses consist
592 * of a chain of data-dependent loads, meaning most CPUs (alpha
593 * being the notable exception) will already guarantee loads are
594 * seen in-order. See the alpha page table accessors for the
595 * smp_read_barrier_depends() barriers in page table walking code.
596 */
597 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
598
Hugh Dickinsc74df322005-10-29 18:16:23 -0700599 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800600 wait_split_huge_page = 0;
601 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800604 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800605 } else if (unlikely(pmd_trans_splitting(*pmd)))
606 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700607 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800608 if (new)
609 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800610 if (wait_split_huge_page)
611 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700612 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613}
614
Hugh Dickins1bb36302005-10-29 18:16:22 -0700615int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700617 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
618 if (!new)
619 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
Nick Piggin362a61a2008-05-14 06:37:36 +0200621 smp_wmb(); /* See comment in __pte_alloc */
622
Hugh Dickins1bb36302005-10-29 18:16:22 -0700623 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800624 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700625 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800626 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800627 } else
628 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700629 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800630 if (new)
631 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700632 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633}
634
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800635static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700636{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800637 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
638}
639
640static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
641{
642 int i;
643
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800644 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700645 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800646 for (i = 0; i < NR_MM_COUNTERS; i++)
647 if (rss[i])
648 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700649}
650
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800652 * This function is called to print an error when a bad pte
653 * is found. For example, we might have a PFN-mapped pte in
654 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700655 *
656 * The calling function must still handle the error.
657 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800658static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
659 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700660{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800661 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
662 pud_t *pud = pud_offset(pgd, addr);
663 pmd_t *pmd = pmd_offset(pud, addr);
664 struct address_space *mapping;
665 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800666 static unsigned long resume;
667 static unsigned long nr_shown;
668 static unsigned long nr_unshown;
669
670 /*
671 * Allow a burst of 60 reports, then keep quiet for that minute;
672 * or allow a steady drip of one report per second.
673 */
674 if (nr_shown == 60) {
675 if (time_before(jiffies, resume)) {
676 nr_unshown++;
677 return;
678 }
679 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800680 printk(KERN_ALERT
681 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800682 nr_unshown);
683 nr_unshown = 0;
684 }
685 nr_shown = 0;
686 }
687 if (nr_shown++ == 0)
688 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800689
690 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
691 index = linear_page_index(vma, addr);
692
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800693 printk(KERN_ALERT
694 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800695 current->comm,
696 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800697 if (page)
698 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800699 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800700 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
701 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
702 /*
703 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
704 */
705 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800706 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800707 (unsigned long)vma->vm_ops->fault);
708 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800709 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800710 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700711 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800712 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700713}
714
KOSAKI Motohiroca16d142011-05-26 19:16:19 +0900715static inline int is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800716{
717 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
718}
719
Hugh Dickins62eede62009-09-21 17:03:34 -0700720#ifndef is_zero_pfn
721static inline int is_zero_pfn(unsigned long pfn)
722{
723 return pfn == zero_pfn;
724}
725#endif
726
727#ifndef my_zero_pfn
728static inline unsigned long my_zero_pfn(unsigned long addr)
729{
730 return zero_pfn;
731}
732#endif
733
Nick Pigginb5810032005-10-29 18:16:12 -0700734/*
Nick Piggin7e675132008-04-28 02:13:00 -0700735 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800736 *
Nick Piggin7e675132008-04-28 02:13:00 -0700737 * "Special" mappings do not wish to be associated with a "struct page" (either
738 * it doesn't exist, or it exists but they don't want to touch it). In this
739 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700740 *
Nick Piggin7e675132008-04-28 02:13:00 -0700741 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
742 * pte bit, in which case this function is trivial. Secondly, an architecture
743 * may not have a spare pte bit, which requires a more complicated scheme,
744 * described below.
745 *
746 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
747 * special mapping (even if there are underlying and valid "struct pages").
748 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800749 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700750 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
751 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700752 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
753 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800754 *
755 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
756 *
Nick Piggin7e675132008-04-28 02:13:00 -0700757 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700758 *
Nick Piggin7e675132008-04-28 02:13:00 -0700759 * This restricts such mappings to be a linear translation from virtual address
760 * to pfn. To get around this restriction, we allow arbitrary mappings so long
761 * as the vma is not a COW mapping; in that case, we know that all ptes are
762 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700763 *
764 *
Nick Piggin7e675132008-04-28 02:13:00 -0700765 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
766 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700767 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
768 * page" backing, however the difference is that _all_ pages with a struct
769 * page (that is, those where pfn_valid is true) are refcounted and considered
770 * normal pages by the VM. The disadvantage is that pages are refcounted
771 * (which can be slower and simply not an option for some PFNMAP users). The
772 * advantage is that we don't have to follow the strict linearity rule of
773 * PFNMAP mappings in order to support COWable mappings.
774 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800775 */
Nick Piggin7e675132008-04-28 02:13:00 -0700776#ifdef __HAVE_ARCH_PTE_SPECIAL
777# define HAVE_PTE_SPECIAL 1
778#else
779# define HAVE_PTE_SPECIAL 0
780#endif
781struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
782 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800783{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800784 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700785
786 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800787 if (likely(!pte_special(pte)))
788 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700789 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
790 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700791 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800792 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700793 return NULL;
794 }
795
796 /* !HAVE_PTE_SPECIAL case follows: */
797
Jared Hulbertb379d792008-04-28 02:12:58 -0700798 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
799 if (vma->vm_flags & VM_MIXEDMAP) {
800 if (!pfn_valid(pfn))
801 return NULL;
802 goto out;
803 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700804 unsigned long off;
805 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700806 if (pfn == vma->vm_pgoff + off)
807 return NULL;
808 if (!is_cow_mapping(vma->vm_flags))
809 return NULL;
810 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800811 }
812
Hugh Dickins62eede62009-09-21 17:03:34 -0700813 if (is_zero_pfn(pfn))
814 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800815check_pfn:
816 if (unlikely(pfn > highest_memmap_pfn)) {
817 print_bad_pte(vma, addr, pte, NULL);
818 return NULL;
819 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800820
821 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700822 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700823 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800824 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700825out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800826 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800827}
828
829/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 * copy one vm_area from one task to the other. Assumes the page tables
831 * already present in the new task to be cleared in the whole range
832 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 */
834
Hugh Dickins570a335b2009-12-14 17:58:46 -0800835static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700837 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700838 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
Nick Pigginb5810032005-10-29 18:16:12 -0700840 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 pte_t pte = *src_pte;
842 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
844 /* pte contains position in swap or file, so copy. */
845 if (unlikely(!pte_present(pte))) {
846 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700847 swp_entry_t entry = pte_to_swp_entry(pte);
848
Hugh Dickins570a335b2009-12-14 17:58:46 -0800849 if (swap_duplicate(entry) < 0)
850 return entry.val;
851
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 /* make sure dst_mm is on swapoff's mmlist. */
853 if (unlikely(list_empty(&dst_mm->mmlist))) {
854 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700855 if (list_empty(&dst_mm->mmlist))
856 list_add(&dst_mm->mmlist,
857 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 spin_unlock(&mmlist_lock);
859 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800860 if (likely(!non_swap_entry(entry)))
861 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800862 else if (is_migration_entry(entry)) {
863 page = migration_entry_to_page(entry);
864
865 if (PageAnon(page))
866 rss[MM_ANONPAGES]++;
867 else
868 rss[MM_FILEPAGES]++;
869
870 if (is_write_migration_entry(entry) &&
871 is_cow_mapping(vm_flags)) {
872 /*
873 * COW mappings require pages in both
874 * parent and child to be set to read.
875 */
876 make_migration_entry_read(&entry);
877 pte = swp_entry_to_pte(entry);
878 set_pte_at(src_mm, addr, src_pte, pte);
879 }
Christoph Lameter06972122006-06-23 02:03:35 -0700880 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700882 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 }
884
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 /*
886 * If it's a COW mapping, write protect it both
887 * in the parent and the child
888 */
Linus Torvalds67121172005-12-11 20:38:17 -0800889 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700891 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 }
893
894 /*
895 * If it's a shared mapping, mark it clean in
896 * the child
897 */
898 if (vm_flags & VM_SHARED)
899 pte = pte_mkclean(pte);
900 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800901
902 page = vm_normal_page(vma, addr, pte);
903 if (page) {
904 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700905 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800906 if (PageAnon(page))
907 rss[MM_ANONPAGES]++;
908 else
909 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800910 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700911
912out_set_pte:
913 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800914 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915}
916
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800917int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
918 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
919 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700921 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700923 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700924 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800925 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800926 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927
928again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800929 init_rss_vec(rss);
930
Hugh Dickinsc74df322005-10-29 18:16:23 -0700931 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 if (!dst_pte)
933 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700934 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700935 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700936 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700937 orig_src_pte = src_pte;
938 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700939 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 do {
942 /*
943 * We are holding two locks at this point - either of them
944 * could generate latencies in another task on another CPU.
945 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700946 if (progress >= 32) {
947 progress = 0;
948 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100949 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700950 break;
951 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 if (pte_none(*src_pte)) {
953 progress++;
954 continue;
955 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800956 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
957 vma, addr, rss);
958 if (entry.val)
959 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 progress += 8;
961 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700963 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700964 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700965 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800966 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700967 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700968 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800969
970 if (entry.val) {
971 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
972 return -ENOMEM;
973 progress = 0;
974 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 if (addr != end)
976 goto again;
977 return 0;
978}
979
980static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
981 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
982 unsigned long addr, unsigned long end)
983{
984 pmd_t *src_pmd, *dst_pmd;
985 unsigned long next;
986
987 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
988 if (!dst_pmd)
989 return -ENOMEM;
990 src_pmd = pmd_offset(src_pud, addr);
991 do {
992 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800993 if (pmd_trans_huge(*src_pmd)) {
994 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800995 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800996 err = copy_huge_pmd(dst_mm, src_mm,
997 dst_pmd, src_pmd, addr, vma);
998 if (err == -ENOMEM)
999 return -ENOMEM;
1000 if (!err)
1001 continue;
1002 /* fall through */
1003 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 if (pmd_none_or_clear_bad(src_pmd))
1005 continue;
1006 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1007 vma, addr, next))
1008 return -ENOMEM;
1009 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1010 return 0;
1011}
1012
1013static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1015 unsigned long addr, unsigned long end)
1016{
1017 pud_t *src_pud, *dst_pud;
1018 unsigned long next;
1019
1020 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1021 if (!dst_pud)
1022 return -ENOMEM;
1023 src_pud = pud_offset(src_pgd, addr);
1024 do {
1025 next = pud_addr_end(addr, end);
1026 if (pud_none_or_clear_bad(src_pud))
1027 continue;
1028 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1029 vma, addr, next))
1030 return -ENOMEM;
1031 } while (dst_pud++, src_pud++, addr = next, addr != end);
1032 return 0;
1033}
1034
1035int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1036 struct vm_area_struct *vma)
1037{
1038 pgd_t *src_pgd, *dst_pgd;
1039 unsigned long next;
1040 unsigned long addr = vma->vm_start;
1041 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001042 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
Nick Piggind9928952005-08-28 16:49:11 +10001044 /*
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1049 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001050 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +10001051 if (!vma->anon_vma)
1052 return 0;
1053 }
1054
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 if (is_vm_hugetlb_page(vma))
1056 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1057
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001058 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001059 /*
1060 * We do not free on error cases below as remove_vma
1061 * gets called on error from higher level routine
1062 */
1063 ret = track_pfn_vma_copy(vma);
1064 if (ret)
1065 return ret;
1066 }
1067
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001068 /*
1069 * We need to invalidate the secondary MMU mappings only when
1070 * there could be a permission downgrade on the ptes of the
1071 * parent mm. And a permission downgrade will only happen if
1072 * is_cow_mapping() returns true.
1073 */
1074 if (is_cow_mapping(vma->vm_flags))
1075 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1076
1077 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 dst_pgd = pgd_offset(dst_mm, addr);
1079 src_pgd = pgd_offset(src_mm, addr);
1080 do {
1081 next = pgd_addr_end(addr, end);
1082 if (pgd_none_or_clear_bad(src_pgd))
1083 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001084 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1085 vma, addr, next))) {
1086 ret = -ENOMEM;
1087 break;
1088 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001090
1091 if (is_cow_mapping(vma->vm_flags))
1092 mmu_notifier_invalidate_range_end(src_mm,
1093 vma->vm_start, end);
1094 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095}
1096
Robin Holt51c6f662005-11-13 16:06:42 -08001097static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001098 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001100 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101{
Nick Pigginb5810032005-10-29 18:16:12 -07001102 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001103 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001104 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001105 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001106 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001107 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001108
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001109again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001110 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001111 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1112 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001113 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 do {
1115 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001116 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001118 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001119
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001121 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001122
Linus Torvalds6aab3412005-11-28 14:34:23 -08001123 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 if (unlikely(details) && page) {
1125 /*
1126 * unmap_shared_mapping_pages() wants to
1127 * invalidate cache without truncating:
1128 * unmap shared but keep private pages.
1129 */
1130 if (details->check_mapping &&
1131 details->check_mapping != page->mapping)
1132 continue;
1133 /*
1134 * Each page->index must be checked when
1135 * invalidating or truncating nonlinear.
1136 */
1137 if (details->nonlinear_vma &&
1138 (page->index < details->first_index ||
1139 page->index > details->last_index))
1140 continue;
1141 }
Nick Pigginb5810032005-10-29 18:16:12 -07001142 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001143 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 tlb_remove_tlb_entry(tlb, pte, addr);
1145 if (unlikely(!page))
1146 continue;
1147 if (unlikely(details) && details->nonlinear_vma
1148 && linear_page_index(details->nonlinear_vma,
1149 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001150 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001153 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001154 else {
1155 if (pte_dirty(ptent))
1156 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001157 if (pte_young(ptent) &&
1158 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001159 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001160 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001161 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001162 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001163 if (unlikely(page_mapcount(page) < 0))
1164 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001165 force_flush = !__tlb_remove_page(tlb, page);
1166 if (force_flush)
1167 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 continue;
1169 }
1170 /*
1171 * If details->check_mapping, we leave swap entries;
1172 * if details->nonlinear_vma, we leave file entries.
1173 */
1174 if (unlikely(details))
1175 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001176 if (pte_file(ptent)) {
1177 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1178 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001179 } else {
1180 swp_entry_t entry = pte_to_swp_entry(ptent);
1181
1182 if (!non_swap_entry(entry))
1183 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001184 else if (is_migration_entry(entry)) {
1185 struct page *page;
1186
1187 page = migration_entry_to_page(entry);
1188
1189 if (PageAnon(page))
1190 rss[MM_ANONPAGES]--;
1191 else
1192 rss[MM_FILEPAGES]--;
1193 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
1196 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001198 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001199
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001200 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001201 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001202 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001203
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001204 /*
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1208 */
1209 if (force_flush) {
1210 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001211
1212#ifdef HAVE_GENERIC_MMU_GATHER
1213 tlb->start = addr;
1214 tlb->end = end;
1215#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001216 tlb_flush_mmu(tlb);
1217 if (addr != end)
1218 goto again;
1219 }
1220
Robin Holt51c6f662005-11-13 16:06:42 -08001221 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222}
1223
Robin Holt51c6f662005-11-13 16:06:42 -08001224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001225 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001227 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001236 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246 split_huge_page_pmd(vma->vm_mm, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001248 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001249 /* fall through */
1250 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001261next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001264
1265 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266}
1267
Robin Holt51c6f662005-11-13 16:06:42 -08001268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001269 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001271 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001279 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001283
1284 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Al Viro038c7aa2012-03-05 13:25:09 -05001287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001299 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001304 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001309 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
Al Virof5cc4ee2012-03-05 14:14:20 -05001312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001315 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
Al Virof5cc4ee2012-03-05 14:14:20 -05001330 if (unlikely(is_pfn_mapping(vma)))
1331 untrack_pfn_vma(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of do_mmap_pgoff. When
1340 * hugetlbfs ->mmap method fails,
1341 * do_mmap_pgoff() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354}
1355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001358 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001363 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
Al Viro6e8bb012012-03-05 13:41:15 -05001374void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001376 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001378 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001389 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001392 *
1393 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001399 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001400 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001403 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001404 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001407 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410}
1411
Jack Steinerc627f9c2008-07-29 22:33:53 -07001412/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, 0);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001432 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435}
1436
Jack Steinerc627f9c2008-07-29 22:33:53 -07001437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001455 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
Johannes Weiner142762b2010-05-24 14:32:39 -07001460/**
1461 * follow_page - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 *
1466 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1467 *
1468 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469 * an error pointer if there is a mapping to something not represented
1470 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001472struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001473 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474{
1475 pgd_t *pgd;
1476 pud_t *pud;
1477 pmd_t *pmd;
1478 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001479 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001481 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001483 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484 if (!IS_ERR(page)) {
1485 BUG_ON(flags & FOLL_GET);
1486 goto out;
1487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001489 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 pgd = pgd_offset(mm, address);
1491 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001495 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001497 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001498 BUG_ON(flags & FOLL_GET);
1499 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500 goto out;
1501 }
1502 if (unlikely(pud_bad(*pud)))
1503 goto no_page_table;
1504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001506 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001507 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001508 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001509 BUG_ON(flags & FOLL_GET);
1510 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511 goto out;
1512 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001513 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001514 if (flags & FOLL_SPLIT) {
1515 split_huge_page_pmd(mm, pmd);
1516 goto split_fallthrough;
1517 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001518 spin_lock(&mm->page_table_lock);
1519 if (likely(pmd_trans_huge(*pmd))) {
1520 if (unlikely(pmd_trans_splitting(*pmd))) {
1521 spin_unlock(&mm->page_table_lock);
1522 wait_split_huge_page(vma->anon_vma, pmd);
1523 } else {
1524 page = follow_trans_huge_pmd(mm, address,
1525 pmd, flags);
1526 spin_unlock(&mm->page_table_lock);
1527 goto out;
1528 }
1529 } else
1530 spin_unlock(&mm->page_table_lock);
1531 /* fall through */
1532 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001533split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001534 if (unlikely(pmd_bad(*pmd)))
1535 goto no_page_table;
1536
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001537 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538
1539 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001540 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001541 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001542 if ((flags & FOLL_WRITE) && !pte_write(pte))
1543 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001544
Linus Torvalds6aab3412005-11-28 14:34:23 -08001545 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001546 if (unlikely(!page)) {
1547 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001548 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001549 goto bad_page;
1550 page = pte_page(pte);
1551 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001552
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001553 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001554 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001555 if (flags & FOLL_TOUCH) {
1556 if ((flags & FOLL_WRITE) &&
1557 !pte_dirty(pte) && !PageDirty(page))
1558 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001559 /*
1560 * pte_mkyoung() would be more correct here, but atomic care
1561 * is needed to avoid losing the dirty bit: it is easier to use
1562 * mark_page_accessed().
1563 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001564 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001566 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001567 /*
1568 * The preliminary mapping check is mainly to avoid the
1569 * pointless overhead of lock_page on the ZERO_PAGE
1570 * which might bounce very badly if there is contention.
1571 *
1572 * If the page is already locked, we don't need to
1573 * handle it now - vmscan will handle it later if and
1574 * when it attempts to reclaim the page.
1575 */
1576 if (page->mapping && trylock_page(page)) {
1577 lru_add_drain(); /* push cached pages to LRU */
1578 /*
1579 * Because we lock page here and migration is
1580 * blocked by the pte's page reference, we need
1581 * only check for file-cache page truncation.
1582 */
1583 if (page->mapping)
1584 mlock_vma_page(page);
1585 unlock_page(page);
1586 }
1587 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001588unlock:
1589 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001591 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001593bad_page:
1594 pte_unmap_unlock(ptep, ptl);
1595 return ERR_PTR(-EFAULT);
1596
1597no_page:
1598 pte_unmap_unlock(ptep, ptl);
1599 if (!pte_none(pte))
1600 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001601
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001602no_page_table:
1603 /*
1604 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001605 * has touched so far, we don't want to allocate unnecessary pages or
1606 * page tables. Return error instead of NULL to skip handle_mm_fault,
1607 * then get_dump_page() will return NULL to leave a hole in the dump.
1608 * But we can only make this optimization where a hole would surely
1609 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001610 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001611 if ((flags & FOLL_DUMP) &&
1612 (!vma->vm_ops || !vma->vm_ops->fault))
1613 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001614 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615}
1616
Linus Torvalds95042f92011-04-12 14:15:51 -07001617static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1618{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001619 return stack_guard_page_start(vma, addr) ||
1620 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001621}
1622
Huang Ying0014bd92011-01-30 11:15:47 +08001623/**
1624 * __get_user_pages() - pin user pages in memory
1625 * @tsk: task_struct of target task
1626 * @mm: mm_struct of target mm
1627 * @start: starting user address
1628 * @nr_pages: number of pages from start to pin
1629 * @gup_flags: flags modifying pin behaviour
1630 * @pages: array that receives pointers to the pages pinned.
1631 * Should be at least nr_pages long. Or NULL, if caller
1632 * only intends to ensure the pages are faulted in.
1633 * @vmas: array of pointers to vmas corresponding to each page.
1634 * Or NULL if the caller does not require them.
1635 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1636 *
1637 * Returns number of pages pinned. This may be fewer than the number
1638 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1639 * were pinned, returns -errno. Each page returned must be released
1640 * with a put_page() call when it is finished with. vmas will only
1641 * remain valid while mmap_sem is held.
1642 *
1643 * Must be called with mmap_sem held for read or write.
1644 *
1645 * __get_user_pages walks a process's page tables and takes a reference to
1646 * each struct page that each user address corresponds to at a given
1647 * instant. That is, it takes the page that would be accessed if a user
1648 * thread accesses the given user virtual address at that instant.
1649 *
1650 * This does not guarantee that the page exists in the user mappings when
1651 * __get_user_pages returns, and there may even be a completely different
1652 * page there in some cases (eg. if mmapped pagecache has been invalidated
1653 * and subsequently re faulted). However it does guarantee that the page
1654 * won't be freed completely. And mostly callers simply care that the page
1655 * contains data that was valid *at some point in time*. Typically, an IO
1656 * or similar operation cannot guarantee anything stronger anyway because
1657 * locks can't be held over the syscall boundary.
1658 *
1659 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1660 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1661 * appropriate) must be called after the page is finished with, and
1662 * before put_page is called.
1663 *
1664 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1665 * or mmap_sem contention, and if waiting is needed to pin all pages,
1666 * *@nonblocking will be set to 0.
1667 *
1668 * In most cases, get_user_pages or get_user_pages_fast should be used
1669 * instead of __get_user_pages. __get_user_pages should be used only if
1670 * you need some special @gup_flags.
1671 */
Nick Pigginb291f002008-10-18 20:26:44 -07001672int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001673 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001674 struct page **pages, struct vm_area_struct **vmas,
1675 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676{
1677 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001678 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
Peter Zijlstra9d737772009-06-25 11:58:55 +02001680 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001681 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001682
1683 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1684
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 /*
1686 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001687 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001689 vm_flags = (gup_flags & FOLL_WRITE) ?
1690 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1691 vm_flags &= (gup_flags & FOLL_FORCE) ?
1692 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 i = 0;
1694
1695 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001696 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697
1698 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001699 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 pgd_t *pgd;
1702 pud_t *pud;
1703 pmd_t *pmd;
1704 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001705
1706 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001707 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 return i ? : -EFAULT;
1709 if (pg > TASK_SIZE)
1710 pgd = pgd_offset_k(pg);
1711 else
1712 pgd = pgd_offset_gate(mm, pg);
1713 BUG_ON(pgd_none(*pgd));
1714 pud = pud_offset(pgd, pg);
1715 BUG_ON(pud_none(*pud));
1716 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001717 if (pmd_none(*pmd))
1718 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001719 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001721 if (pte_none(*pte)) {
1722 pte_unmap(pte);
1723 return i ? : -EFAULT;
1724 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001725 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001727 struct page *page;
1728
Linus Torvalds95042f92011-04-12 14:15:51 -07001729 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001730 if (!page) {
1731 if (!(gup_flags & FOLL_DUMP) &&
1732 is_zero_pfn(pte_pfn(*pte)))
1733 page = pte_page(*pte);
1734 else {
1735 pte_unmap(pte);
1736 return i ? : -EFAULT;
1737 }
1738 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001739 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001740 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 }
1742 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001743 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 }
1745
Nick Pigginb291f002008-10-18 20:26:44 -07001746 if (!vma ||
1747 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001748 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 return i ? : -EFAULT;
1750
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001751 if (is_vm_hugetlb_page(vma)) {
1752 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001753 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001754 continue;
1755 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001756
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001758 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001759 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Ethan Solomita462e00c2007-07-15 23:38:16 -07001761 /*
Ying Han47792802009-01-06 14:40:18 -08001762 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001763 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001764 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001765 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001766 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001767
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001768 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001769 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001770 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001771 unsigned int fault_flags = 0;
1772
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001773 /* For mlock, just skip the stack guard page. */
1774 if (foll_flags & FOLL_MLOCK) {
1775 if (stack_guard_page(vma, start))
1776 goto next_page;
1777 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001778 if (foll_flags & FOLL_WRITE)
1779 fault_flags |= FAULT_FLAG_WRITE;
1780 if (nonblocking)
1781 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001782 if (foll_flags & FOLL_NOWAIT)
1783 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001784
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001785 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001786 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001787
Nick Piggin83c54072007-07-19 01:47:05 -07001788 if (ret & VM_FAULT_ERROR) {
1789 if (ret & VM_FAULT_OOM)
1790 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001791 if (ret & (VM_FAULT_HWPOISON |
1792 VM_FAULT_HWPOISON_LARGE)) {
1793 if (i)
1794 return i;
1795 else if (gup_flags & FOLL_HWPOISON)
1796 return -EHWPOISON;
1797 else
1798 return -EFAULT;
1799 }
1800 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001801 return i ? i : -EFAULT;
1802 BUG();
1803 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001804
1805 if (tsk) {
1806 if (ret & VM_FAULT_MAJOR)
1807 tsk->maj_flt++;
1808 else
1809 tsk->min_flt++;
1810 }
Nick Piggin83c54072007-07-19 01:47:05 -07001811
Michel Lespinasse53a77062011-01-13 15:46:14 -08001812 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001813 if (nonblocking)
1814 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001815 return i;
1816 }
1817
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001818 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001819 * The VM_FAULT_WRITE bit tells us that
1820 * do_wp_page has broken COW when necessary,
1821 * even if maybe_mkwrite decided not to set
1822 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001823 * page lookups as if they were reads. But only
1824 * do so when looping for pte_write is futile:
1825 * in some cases userspace may also be wanting
1826 * to write to the gotten user page, which a
1827 * read fault here might prevent (a readonly
1828 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001829 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001830 if ((ret & VM_FAULT_WRITE) &&
1831 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001832 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001833
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001834 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001836 if (IS_ERR(page))
1837 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001839 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001840
Russell Kinga6f36be2006-12-30 22:24:19 +00001841 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001842 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001844next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 if (vmas)
1846 vmas[i] = vma;
1847 i++;
1848 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001849 nr_pages--;
1850 } while (nr_pages && start < vma->vm_end);
1851 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 return i;
1853}
Huang Ying0014bd92011-01-30 11:15:47 +08001854EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001855
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001856/*
1857 * fixup_user_fault() - manually resolve a user page fault
1858 * @tsk: the task_struct to use for page fault accounting, or
1859 * NULL if faults are not to be recorded.
1860 * @mm: mm_struct of target mm
1861 * @address: user address
1862 * @fault_flags:flags to pass down to handle_mm_fault()
1863 *
1864 * This is meant to be called in the specific scenario where for locking reasons
1865 * we try to access user memory in atomic context (within a pagefault_disable()
1866 * section), this returns -EFAULT, and we want to resolve the user fault before
1867 * trying again.
1868 *
1869 * Typically this is meant to be used by the futex code.
1870 *
1871 * The main difference with get_user_pages() is that this function will
1872 * unconditionally call handle_mm_fault() which will in turn perform all the
1873 * necessary SW fixup of the dirty and young bits in the PTE, while
1874 * handle_mm_fault() only guarantees to update these in the struct page.
1875 *
1876 * This is important for some architectures where those bits also gate the
1877 * access permission to the page because they are maintained in software. On
1878 * such architectures, gup() will not be enough to make a subsequent access
1879 * succeed.
1880 *
1881 * This should be called with the mm_sem held for read.
1882 */
1883int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1884 unsigned long address, unsigned int fault_flags)
1885{
1886 struct vm_area_struct *vma;
1887 int ret;
1888
1889 vma = find_extend_vma(mm, address);
1890 if (!vma || address < vma->vm_start)
1891 return -EFAULT;
1892
1893 ret = handle_mm_fault(mm, vma, address, fault_flags);
1894 if (ret & VM_FAULT_ERROR) {
1895 if (ret & VM_FAULT_OOM)
1896 return -ENOMEM;
1897 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1898 return -EHWPOISON;
1899 if (ret & VM_FAULT_SIGBUS)
1900 return -EFAULT;
1901 BUG();
1902 }
1903 if (tsk) {
1904 if (ret & VM_FAULT_MAJOR)
1905 tsk->maj_flt++;
1906 else
1907 tsk->min_flt++;
1908 }
1909 return 0;
1910}
1911
1912/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001913 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001914 * @tsk: the task_struct to use for page fault accounting, or
1915 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001916 * @mm: mm_struct of target mm
1917 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001918 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001919 * @write: whether pages will be written to by the caller
1920 * @force: whether to force write access even if user mapping is
1921 * readonly. This will result in the page being COWed even
1922 * in MAP_SHARED mappings. You do not want this.
1923 * @pages: array that receives pointers to the pages pinned.
1924 * Should be at least nr_pages long. Or NULL, if caller
1925 * only intends to ensure the pages are faulted in.
1926 * @vmas: array of pointers to vmas corresponding to each page.
1927 * Or NULL if the caller does not require them.
1928 *
1929 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001930 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001931 * were pinned, returns -errno. Each page returned must be released
1932 * with a put_page() call when it is finished with. vmas will only
1933 * remain valid while mmap_sem is held.
1934 *
1935 * Must be called with mmap_sem held for read or write.
1936 *
1937 * get_user_pages walks a process's page tables and takes a reference to
1938 * each struct page that each user address corresponds to at a given
1939 * instant. That is, it takes the page that would be accessed if a user
1940 * thread accesses the given user virtual address at that instant.
1941 *
1942 * This does not guarantee that the page exists in the user mappings when
1943 * get_user_pages returns, and there may even be a completely different
1944 * page there in some cases (eg. if mmapped pagecache has been invalidated
1945 * and subsequently re faulted). However it does guarantee that the page
1946 * won't be freed completely. And mostly callers simply care that the page
1947 * contains data that was valid *at some point in time*. Typically, an IO
1948 * or similar operation cannot guarantee anything stronger anyway because
1949 * locks can't be held over the syscall boundary.
1950 *
1951 * If write=0, the page must not be written to. If the page is written to,
1952 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1953 * after the page is finished with, and before put_page is called.
1954 *
1955 * get_user_pages is typically used for fewer-copy IO operations, to get a
1956 * handle on the memory by some means other than accesses via the user virtual
1957 * addresses. The pages may be submitted for DMA to devices or accessed via
1958 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1959 * use the correct cache flushing APIs.
1960 *
1961 * See also get_user_pages_fast, for performance critical applications.
1962 */
Nick Pigginb291f002008-10-18 20:26:44 -07001963int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001964 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001965 struct page **pages, struct vm_area_struct **vmas)
1966{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001967 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001968
Hugh Dickins58fa8792009-09-21 17:03:31 -07001969 if (pages)
1970 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001971 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001972 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001973 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001974 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001975
Michel Lespinasse53a77062011-01-13 15:46:14 -08001976 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1977 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001978}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979EXPORT_SYMBOL(get_user_pages);
1980
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001981/**
1982 * get_dump_page() - pin user page in memory while writing it to core dump
1983 * @addr: user address
1984 *
1985 * Returns struct page pointer of user page pinned for dump,
1986 * to be freed afterwards by page_cache_release() or put_page().
1987 *
1988 * Returns NULL on any kind of failure - a hole must then be inserted into
1989 * the corefile, to preserve alignment with its headers; and also returns
1990 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1991 * allowing a hole to be left in the corefile to save diskspace.
1992 *
1993 * Called without mmap_sem, but after all other threads have been killed.
1994 */
1995#ifdef CONFIG_ELF_CORE
1996struct page *get_dump_page(unsigned long addr)
1997{
1998 struct vm_area_struct *vma;
1999 struct page *page;
2000
2001 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002002 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2003 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002004 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002005 flush_cache_page(vma, addr, page_to_pfn(page));
2006 return page;
2007}
2008#endif /* CONFIG_ELF_CORE */
2009
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002010pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002011 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002012{
2013 pgd_t * pgd = pgd_offset(mm, addr);
2014 pud_t * pud = pud_alloc(mm, pgd, addr);
2015 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002016 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002017 if (pmd) {
2018 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002019 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002020 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002021 }
2022 return NULL;
2023}
2024
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002026 * This is the old fallback for page remapping.
2027 *
2028 * For historical reasons, it only allows reserved pages. Only
2029 * old drivers should use this, and they needed to mark their
2030 * pages reserved for the old functions anyway.
2031 */
Nick Piggin423bad62008-04-28 02:13:01 -07002032static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2033 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002034{
Nick Piggin423bad62008-04-28 02:13:01 -07002035 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002036 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002037 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002038 spinlock_t *ptl;
2039
Linus Torvalds238f58d2005-11-29 13:01:56 -08002040 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002041 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002042 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002043 retval = -ENOMEM;
2044 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002045 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002046 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002047 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002048 retval = -EBUSY;
2049 if (!pte_none(*pte))
2050 goto out_unlock;
2051
2052 /* Ok, finally just insert the thing.. */
2053 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002054 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002055 page_add_file_rmap(page);
2056 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2057
2058 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002059 pte_unmap_unlock(pte, ptl);
2060 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002061out_unlock:
2062 pte_unmap_unlock(pte, ptl);
2063out:
2064 return retval;
2065}
2066
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002067/**
2068 * vm_insert_page - insert single page into user vma
2069 * @vma: user vma to map to
2070 * @addr: target user address of this page
2071 * @page: source kernel page
2072 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002073 * This allows drivers to insert individual pages they've allocated
2074 * into a user vma.
2075 *
2076 * The page has to be a nice clean _individual_ kernel allocation.
2077 * If you allocate a compound page, you need to have marked it as
2078 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002079 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002080 *
2081 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2082 * took an arbitrary page protection parameter. This doesn't allow
2083 * that. Your vma protection will have to be set up correctly, which
2084 * means that if you want a shared writable mapping, you'd better
2085 * ask for a shared writable mapping!
2086 *
2087 * The page does not need to be reserved.
2088 */
Nick Piggin423bad62008-04-28 02:13:01 -07002089int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2090 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002091{
2092 if (addr < vma->vm_start || addr >= vma->vm_end)
2093 return -EFAULT;
2094 if (!page_count(page))
2095 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08002096 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07002097 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002098}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002099EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002100
Nick Piggin423bad62008-04-28 02:13:01 -07002101static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2102 unsigned long pfn, pgprot_t prot)
2103{
2104 struct mm_struct *mm = vma->vm_mm;
2105 int retval;
2106 pte_t *pte, entry;
2107 spinlock_t *ptl;
2108
2109 retval = -ENOMEM;
2110 pte = get_locked_pte(mm, addr, &ptl);
2111 if (!pte)
2112 goto out;
2113 retval = -EBUSY;
2114 if (!pte_none(*pte))
2115 goto out_unlock;
2116
2117 /* Ok, finally just insert the thing.. */
2118 entry = pte_mkspecial(pfn_pte(pfn, prot));
2119 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002120 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002121
2122 retval = 0;
2123out_unlock:
2124 pte_unmap_unlock(pte, ptl);
2125out:
2126 return retval;
2127}
2128
Nick Piggine0dc0d82007-02-12 00:51:36 -08002129/**
2130 * vm_insert_pfn - insert single pfn into user vma
2131 * @vma: user vma to map to
2132 * @addr: target user address of this page
2133 * @pfn: source kernel pfn
2134 *
2135 * Similar to vm_inert_page, this allows drivers to insert individual pages
2136 * they've allocated into a user vma. Same comments apply.
2137 *
2138 * This function should only be called from a vm_ops->fault handler, and
2139 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002140 *
2141 * vma cannot be a COW mapping.
2142 *
2143 * As this is called only for pages that do not currently exist, we
2144 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002145 */
2146int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002147 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002148{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002149 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002150 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002151 /*
2152 * Technically, architectures with pte_special can avoid all these
2153 * restrictions (same for remap_pfn_range). However we would like
2154 * consistency in testing and feature parity among all, so we should
2155 * try to keep these invariants in place for everybody.
2156 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002157 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2158 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2159 (VM_PFNMAP|VM_MIXEDMAP));
2160 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2161 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002162
Nick Piggin423bad62008-04-28 02:13:01 -07002163 if (addr < vma->vm_start || addr >= vma->vm_end)
2164 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002165 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002166 return -EINVAL;
2167
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002168 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002169
2170 if (ret)
2171 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2172
2173 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002174}
2175EXPORT_SYMBOL(vm_insert_pfn);
2176
Nick Piggin423bad62008-04-28 02:13:01 -07002177int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2178 unsigned long pfn)
2179{
2180 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2181
2182 if (addr < vma->vm_start || addr >= vma->vm_end)
2183 return -EFAULT;
2184
2185 /*
2186 * If we don't have pte special, then we have to use the pfn_valid()
2187 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2188 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002189 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2190 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002191 */
2192 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2193 struct page *page;
2194
2195 page = pfn_to_page(pfn);
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197 }
2198 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2199}
2200EXPORT_SYMBOL(vm_insert_mixed);
2201
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002202/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 * maps a range of physical memory into the requested pages. the old
2204 * mappings are removed. any references to nonexistent pages results
2205 * in null mappings (currently treated as "copy-on-access")
2206 */
2207static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2208 unsigned long addr, unsigned long end,
2209 unsigned long pfn, pgprot_t prot)
2210{
2211 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002212 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213
Hugh Dickinsc74df322005-10-29 18:16:23 -07002214 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 if (!pte)
2216 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002217 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 do {
2219 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002220 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 pfn++;
2222 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002223 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002224 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 return 0;
2226}
2227
2228static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2229 unsigned long addr, unsigned long end,
2230 unsigned long pfn, pgprot_t prot)
2231{
2232 pmd_t *pmd;
2233 unsigned long next;
2234
2235 pfn -= addr >> PAGE_SHIFT;
2236 pmd = pmd_alloc(mm, pud, addr);
2237 if (!pmd)
2238 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002239 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 do {
2241 next = pmd_addr_end(addr, end);
2242 if (remap_pte_range(mm, pmd, addr, next,
2243 pfn + (addr >> PAGE_SHIFT), prot))
2244 return -ENOMEM;
2245 } while (pmd++, addr = next, addr != end);
2246 return 0;
2247}
2248
2249static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2250 unsigned long addr, unsigned long end,
2251 unsigned long pfn, pgprot_t prot)
2252{
2253 pud_t *pud;
2254 unsigned long next;
2255
2256 pfn -= addr >> PAGE_SHIFT;
2257 pud = pud_alloc(mm, pgd, addr);
2258 if (!pud)
2259 return -ENOMEM;
2260 do {
2261 next = pud_addr_end(addr, end);
2262 if (remap_pmd_range(mm, pud, addr, next,
2263 pfn + (addr >> PAGE_SHIFT), prot))
2264 return -ENOMEM;
2265 } while (pud++, addr = next, addr != end);
2266 return 0;
2267}
2268
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002269/**
2270 * remap_pfn_range - remap kernel memory to userspace
2271 * @vma: user vma to map to
2272 * @addr: target user address to start at
2273 * @pfn: physical address of kernel memory
2274 * @size: size of map area
2275 * @prot: page protection flags for this mapping
2276 *
2277 * Note: this is only safe if the mm semaphore is held when called.
2278 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2280 unsigned long pfn, unsigned long size, pgprot_t prot)
2281{
2282 pgd_t *pgd;
2283 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002284 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 struct mm_struct *mm = vma->vm_mm;
2286 int err;
2287
2288 /*
2289 * Physically remapped pages are special. Tell the
2290 * rest of the world about it:
2291 * VM_IO tells people not to look at these pages
2292 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08002293 * VM_RESERVED is specified all over the place, because
2294 * in 2.4 it kept swapout's vma scan off this vma; but
2295 * in 2.6 the LRU scan won't even find its pages, so this
2296 * flag means no more than count its pages in reserved_vm,
2297 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08002298 * VM_PFNMAP tells the core MM that the base pages are just
2299 * raw PFN mappings, and do not have a "struct page" associated
2300 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002301 *
2302 * There's a horrible special case to handle copy-on-write
2303 * behaviour that some programs depend on. We mark the "original"
2304 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002306 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002307 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002308 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002309 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002310 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002311
Linus Torvalds6aab3412005-11-28 14:34:23 -08002312 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002314 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002315 if (err) {
2316 /*
2317 * To indicate that track_pfn related cleanup is not
2318 * needed from higher level routine calling unmap_vmas
2319 */
2320 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002321 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002322 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002323 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 BUG_ON(addr >= end);
2326 pfn -= addr >> PAGE_SHIFT;
2327 pgd = pgd_offset(mm, addr);
2328 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 do {
2330 next = pgd_addr_end(addr, end);
2331 err = remap_pud_range(mm, pgd, addr, next,
2332 pfn + (addr >> PAGE_SHIFT), prot);
2333 if (err)
2334 break;
2335 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002336
2337 if (err)
2338 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2339
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 return err;
2341}
2342EXPORT_SYMBOL(remap_pfn_range);
2343
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002344static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2345 unsigned long addr, unsigned long end,
2346 pte_fn_t fn, void *data)
2347{
2348 pte_t *pte;
2349 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002350 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002351 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002352
2353 pte = (mm == &init_mm) ?
2354 pte_alloc_kernel(pmd, addr) :
2355 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2356 if (!pte)
2357 return -ENOMEM;
2358
2359 BUG_ON(pmd_huge(*pmd));
2360
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002361 arch_enter_lazy_mmu_mode();
2362
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002363 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002364
2365 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002366 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002367 if (err)
2368 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002369 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002370
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002371 arch_leave_lazy_mmu_mode();
2372
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002373 if (mm != &init_mm)
2374 pte_unmap_unlock(pte-1, ptl);
2375 return err;
2376}
2377
2378static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2379 unsigned long addr, unsigned long end,
2380 pte_fn_t fn, void *data)
2381{
2382 pmd_t *pmd;
2383 unsigned long next;
2384 int err;
2385
Andi Kleenceb86872008-07-23 21:27:50 -07002386 BUG_ON(pud_huge(*pud));
2387
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002388 pmd = pmd_alloc(mm, pud, addr);
2389 if (!pmd)
2390 return -ENOMEM;
2391 do {
2392 next = pmd_addr_end(addr, end);
2393 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2394 if (err)
2395 break;
2396 } while (pmd++, addr = next, addr != end);
2397 return err;
2398}
2399
2400static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2401 unsigned long addr, unsigned long end,
2402 pte_fn_t fn, void *data)
2403{
2404 pud_t *pud;
2405 unsigned long next;
2406 int err;
2407
2408 pud = pud_alloc(mm, pgd, addr);
2409 if (!pud)
2410 return -ENOMEM;
2411 do {
2412 next = pud_addr_end(addr, end);
2413 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2414 if (err)
2415 break;
2416 } while (pud++, addr = next, addr != end);
2417 return err;
2418}
2419
2420/*
2421 * Scan a region of virtual memory, filling in page tables as necessary
2422 * and calling a provided function on each leaf page table.
2423 */
2424int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2425 unsigned long size, pte_fn_t fn, void *data)
2426{
2427 pgd_t *pgd;
2428 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002429 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002430 int err;
2431
2432 BUG_ON(addr >= end);
2433 pgd = pgd_offset(mm, addr);
2434 do {
2435 next = pgd_addr_end(addr, end);
2436 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2437 if (err)
2438 break;
2439 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002440
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002441 return err;
2442}
2443EXPORT_SYMBOL_GPL(apply_to_page_range);
2444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002446 * handle_pte_fault chooses page fault handler according to an entry
2447 * which was read non-atomically. Before making any commitment, on
2448 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002449 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002450 * must check under lock before unmapping the pte and proceeding
2451 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002452 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002453 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002454static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002455 pte_t *page_table, pte_t orig_pte)
2456{
2457 int same = 1;
2458#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2459 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002460 spinlock_t *ptl = pte_lockptr(mm, pmd);
2461 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002462 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002463 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002464 }
2465#endif
2466 pte_unmap(page_table);
2467 return same;
2468}
2469
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002470static 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 -08002471{
2472 /*
2473 * If the source page was a PFN mapping, we don't have
2474 * a "struct page" for it. We do a best-effort copy by
2475 * just copying from the original user address. If that
2476 * fails, we just zero-fill it. Live with it.
2477 */
2478 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002479 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002480 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2481
2482 /*
2483 * This really shouldn't fail, because the page is there
2484 * in the page tables. But it might just be unreadable,
2485 * in which case we just give up and fill the result with
2486 * zeroes.
2487 */
2488 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002489 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002490 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002491 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002492 } else
2493 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002494}
2495
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 * This routine handles present pages, when users try to write
2498 * to a shared page. It is done by copying the page to a new address
2499 * and decrementing the shared-page counter for the old page.
2500 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 * Note that this routine assumes that the protection checks have been
2502 * done by the caller (the low-level page fault routine in most cases).
2503 * Thus we can safely just mark it writable once we've done any necessary
2504 * COW.
2505 *
2506 * We also mark the page dirty at this point even though the page will
2507 * change only once the write actually happens. This avoids a few races,
2508 * and potentially makes it more efficient.
2509 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002510 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2511 * but allow concurrent faults), with pte both mapped and locked.
2512 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002514static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2515 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002516 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002517 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002519 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002521 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002522 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002523 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524
Linus Torvalds6aab3412005-11-28 14:34:23 -08002525 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002526 if (!old_page) {
2527 /*
2528 * VM_MIXEDMAP !pfn_valid() case
2529 *
2530 * We should not cow pages in a shared writeable mapping.
2531 * Just mark the pages writable as we can't do any dirty
2532 * accounting on raw pfn maps.
2533 */
2534 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2535 (VM_WRITE|VM_SHARED))
2536 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002537 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002538 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002540 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002541 * Take out anonymous pages first, anonymous shared vmas are
2542 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002543 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002544 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002545 if (!trylock_page(old_page)) {
2546 page_cache_get(old_page);
2547 pte_unmap_unlock(page_table, ptl);
2548 lock_page(old_page);
2549 page_table = pte_offset_map_lock(mm, pmd, address,
2550 &ptl);
2551 if (!pte_same(*page_table, orig_pte)) {
2552 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002553 goto unlock;
2554 }
2555 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002556 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002557 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002558 /*
2559 * The page is all ours. Move it to our anon_vma so
2560 * the rmap code will not search our parent or siblings.
2561 * Protected against the rmap code by the page lock.
2562 */
2563 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002564 unlock_page(old_page);
2565 goto reuse;
2566 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002567 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002568 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002569 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002570 /*
2571 * Only catch write-faults on shared writable pages,
2572 * read-only shared pages can get COWed by
2573 * get_user_pages(.write=1, .force=1).
2574 */
David Howells9637a5e2006-06-23 02:03:43 -07002575 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002576 struct vm_fault vmf;
2577 int tmp;
2578
2579 vmf.virtual_address = (void __user *)(address &
2580 PAGE_MASK);
2581 vmf.pgoff = old_page->index;
2582 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2583 vmf.page = old_page;
2584
David Howells9637a5e2006-06-23 02:03:43 -07002585 /*
2586 * Notify the address space that the page is about to
2587 * become writable so that it can prohibit this or wait
2588 * for the page to get into an appropriate state.
2589 *
2590 * We do this without the lock held, so that it can
2591 * sleep if it needs to.
2592 */
2593 page_cache_get(old_page);
2594 pte_unmap_unlock(page_table, ptl);
2595
Nick Pigginc2ec1752009-03-31 15:23:21 -07002596 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2597 if (unlikely(tmp &
2598 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2599 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002600 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002601 }
Nick Pigginb827e492009-04-30 15:08:16 -07002602 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2603 lock_page(old_page);
2604 if (!old_page->mapping) {
2605 ret = 0; /* retry the fault */
2606 unlock_page(old_page);
2607 goto unwritable_page;
2608 }
2609 } else
2610 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002611
David Howells9637a5e2006-06-23 02:03:43 -07002612 /*
2613 * Since we dropped the lock we need to revalidate
2614 * the PTE as someone else may have changed it. If
2615 * they did, we just return, as we can count on the
2616 * MMU to tell us if they didn't also make it writable.
2617 */
2618 page_table = pte_offset_map_lock(mm, pmd, address,
2619 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002620 if (!pte_same(*page_table, orig_pte)) {
2621 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002622 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002623 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002624
2625 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002627 dirty_page = old_page;
2628 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002629
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002630reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002631 flush_cache_page(vma, address, pte_pfn(orig_pte));
2632 entry = pte_mkyoung(orig_pte);
2633 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002634 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002635 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002636 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002637 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002638
2639 if (!dirty_page)
2640 return ret;
2641
2642 /*
2643 * Yes, Virginia, this is actually required to prevent a race
2644 * with clear_page_dirty_for_io() from clearing the page dirty
2645 * bit after it clear all dirty ptes, but before a racing
2646 * do_wp_page installs a dirty pte.
2647 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002648 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002649 */
2650 if (!page_mkwrite) {
2651 wait_on_page_locked(dirty_page);
2652 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002653 /* file_update_time outside page_lock */
2654 if (vma->vm_file)
2655 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002656 }
2657 put_page(dirty_page);
2658 if (page_mkwrite) {
2659 struct address_space *mapping = dirty_page->mapping;
2660
2661 set_page_dirty(dirty_page);
2662 unlock_page(dirty_page);
2663 page_cache_release(dirty_page);
2664 if (mapping) {
2665 /*
2666 * Some device drivers do not set page.mapping
2667 * but still dirty their pages
2668 */
2669 balance_dirty_pages_ratelimited(mapping);
2670 }
2671 }
2672
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002673 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675
2676 /*
2677 * Ok, we need to copy. Oh, well..
2678 */
Nick Pigginb5810032005-10-29 18:16:12 -07002679 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002680gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002681 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682
2683 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002684 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002685
Hugh Dickins62eede62009-09-21 17:03:34 -07002686 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002687 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2688 if (!new_page)
2689 goto oom;
2690 } else {
2691 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2692 if (!new_page)
2693 goto oom;
2694 cow_user_page(new_page, old_page, address, vma);
2695 }
2696 __SetPageUptodate(new_page);
2697
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002698 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002699 goto oom_free_new;
2700
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 /*
2702 * Re-check the pte - we dropped the lock
2703 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002704 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002705 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002706 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002707 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002708 dec_mm_counter_fast(mm, MM_FILEPAGES);
2709 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002710 }
2711 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002712 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002713 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002714 entry = mk_pte(new_page, vma->vm_page_prot);
2715 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002716 /*
2717 * Clear the pte entry and flush it first, before updating the
2718 * pte with the new entry. This will avoid a race condition
2719 * seen in the presence of one thread doing SMC and another
2720 * thread doing COW.
2721 */
Izik Eidus828502d2009-09-21 17:01:51 -07002722 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002723 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002724 /*
2725 * We call the notify macro here because, when using secondary
2726 * mmu page tables (such as kvm shadow page tables), we want the
2727 * new page to be mapped directly into the secondary page table.
2728 */
2729 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002730 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002731 if (old_page) {
2732 /*
2733 * Only after switching the pte to the new page may
2734 * we remove the mapcount here. Otherwise another
2735 * process may come and find the rmap count decremented
2736 * before the pte is switched to the new page, and
2737 * "reuse" the old page writing into it while our pte
2738 * here still points into it and can be read by other
2739 * threads.
2740 *
2741 * The critical issue is to order this
2742 * page_remove_rmap with the ptp_clear_flush above.
2743 * Those stores are ordered by (if nothing else,)
2744 * the barrier present in the atomic_add_negative
2745 * in page_remove_rmap.
2746 *
2747 * Then the TLB flush in ptep_clear_flush ensures that
2748 * no process can access the old page before the
2749 * decremented mapcount is visible. And the old page
2750 * cannot be reused until after the decremented
2751 * mapcount is visible. So transitively, TLBs to
2752 * old page will be flushed before it can be reused.
2753 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002754 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002755 }
2756
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 /* Free the old page.. */
2758 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002759 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002760 } else
2761 mem_cgroup_uncharge_page(new_page);
2762
Hugh Dickins920fc352005-11-21 21:32:17 -08002763 if (new_page)
2764 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002765unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002766 pte_unmap_unlock(page_table, ptl);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002767 if (old_page) {
2768 /*
2769 * Don't let another task, with possibly unlocked vma,
2770 * keep the mlocked page.
2771 */
2772 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2773 lock_page(old_page); /* LRU manipulation */
2774 munlock_vma_page(old_page);
2775 unlock_page(old_page);
2776 }
2777 page_cache_release(old_page);
2778 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002779 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002780oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002781 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002782oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002783 if (old_page) {
2784 if (page_mkwrite) {
2785 unlock_page(old_page);
2786 page_cache_release(old_page);
2787 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002788 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002789 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002791
2792unwritable_page:
2793 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002794 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795}
2796
Peter Zijlstra97a89412011-05-24 17:12:04 -07002797static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 unsigned long start_addr, unsigned long end_addr,
2799 struct zap_details *details)
2800{
Al Virof5cc4ee2012-03-05 14:14:20 -05002801 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802}
2803
2804static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2805 struct zap_details *details)
2806{
2807 struct vm_area_struct *vma;
2808 struct prio_tree_iter iter;
2809 pgoff_t vba, vea, zba, zea;
2810
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 vma_prio_tree_foreach(vma, &iter, root,
2812 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813
2814 vba = vma->vm_pgoff;
2815 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2816 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2817 zba = details->first_index;
2818 if (zba < vba)
2819 zba = vba;
2820 zea = details->last_index;
2821 if (zea > vea)
2822 zea = vea;
2823
Peter Zijlstra97a89412011-05-24 17:12:04 -07002824 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2826 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002827 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 }
2829}
2830
2831static inline void unmap_mapping_range_list(struct list_head *head,
2832 struct zap_details *details)
2833{
2834 struct vm_area_struct *vma;
2835
2836 /*
2837 * In nonlinear VMAs there is no correspondence between virtual address
2838 * offset and file offset. So we must perform an exhaustive search
2839 * across *all* the pages in each nonlinear VMA, not just the pages
2840 * whose virtual address lies outside the file truncation point.
2841 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 list_for_each_entry(vma, head, shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002844 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 }
2846}
2847
2848/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002849 * 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 -07002850 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 * @holebegin: byte in first page to unmap, relative to the start of
2852 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002853 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 * must keep the partial page. In contrast, we must get rid of
2855 * partial pages.
2856 * @holelen: size of prospective hole in bytes. This will be rounded
2857 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2858 * end of the file.
2859 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2860 * but 0 when invalidating pagecache, don't throw away private data.
2861 */
2862void unmap_mapping_range(struct address_space *mapping,
2863 loff_t const holebegin, loff_t const holelen, int even_cows)
2864{
2865 struct zap_details details;
2866 pgoff_t hba = holebegin >> PAGE_SHIFT;
2867 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2868
2869 /* Check for overflow. */
2870 if (sizeof(holelen) > sizeof(hlen)) {
2871 long long holeend =
2872 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2873 if (holeend & ~(long long)ULONG_MAX)
2874 hlen = ULONG_MAX - hba + 1;
2875 }
2876
2877 details.check_mapping = even_cows? NULL: mapping;
2878 details.nonlinear_vma = NULL;
2879 details.first_index = hba;
2880 details.last_index = hba + hlen - 1;
2881 if (details.last_index < details.first_index)
2882 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883
Peter Zijlstra97a89412011-05-24 17:12:04 -07002884
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002885 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2887 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2888 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2889 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002890 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891}
2892EXPORT_SYMBOL(unmap_mapping_range);
2893
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002895 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2896 * but allow concurrent faults), and pte mapped but not yet locked.
2897 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002899static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2900 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002901 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002903 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002904 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002905 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002907 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002908 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002909 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002910 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002912 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002913 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002914
2915 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002916 if (unlikely(non_swap_entry(entry))) {
2917 if (is_migration_entry(entry)) {
2918 migration_entry_wait(mm, pmd, address);
2919 } else if (is_hwpoison_entry(entry)) {
2920 ret = VM_FAULT_HWPOISON;
2921 } else {
2922 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002923 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002924 }
Christoph Lameter06972122006-06-23 02:03:35 -07002925 goto out;
2926 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002927 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 page = lookup_swap_cache(entry);
2929 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002930 page = swapin_readahead(entry,
2931 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 if (!page) {
2933 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002934 * Back out if somebody else faulted in this pte
2935 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002937 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 if (likely(pte_same(*page_table, orig_pte)))
2939 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002940 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002941 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 }
2943
2944 /* Had to read the page from swap area: Major fault */
2945 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002946 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002947 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002948 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002949 /*
2950 * hwpoisoned dirty swapcache pages are kept for killing
2951 * owner processes (which may be unknown at hwpoison time)
2952 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002953 ret = VM_FAULT_HWPOISON;
2954 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002955 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 }
2957
Michel Lespinassed065bd82010-10-26 14:21:57 -07002958 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002959
Balbir Singh20a10222007-11-14 17:00:33 -08002960 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002961 if (!locked) {
2962 ret |= VM_FAULT_RETRY;
2963 goto out_release;
2964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002966 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002967 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2968 * release the swapcache from under us. The page pin, and pte_same
2969 * test below, are not enough to exclude that. Even if it is still
2970 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002971 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002972 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002973 goto out_page;
2974
2975 if (ksm_might_need_to_copy(page, vma, address)) {
2976 swapcache = page;
2977 page = ksm_does_need_to_copy(page, vma, address);
2978
2979 if (unlikely(!page)) {
2980 ret = VM_FAULT_OOM;
2981 page = swapcache;
2982 swapcache = NULL;
2983 goto out_page;
2984 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002985 }
2986
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002987 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002988 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002989 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002990 }
2991
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002993 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002995 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002996 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002997 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002998
2999 if (unlikely(!PageUptodate(page))) {
3000 ret = VM_FAULT_SIGBUS;
3001 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 }
3003
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003004 /*
3005 * The page isn't present yet, go ahead with the fault.
3006 *
3007 * Be careful about the sequence of operations here.
3008 * To get its accounting right, reuse_swap_page() must be called
3009 * while the page is counted on swap but not yet in mapcount i.e.
3010 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3011 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003012 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3013 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3014 * in page->private. In this case, a record in swap_cgroup is silently
3015 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003016 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003018 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003019 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003021 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003023 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003024 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003025 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 flush_icache_page(vma, page);
3028 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003029 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003030 /* It's better to call commit-charge after rmap is established */
3031 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003033 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003034 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003035 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003036 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003037 if (swapcache) {
3038 /*
3039 * Hold the lock to avoid the swap entry to be reused
3040 * until we take the PT lock for the pte_same() check
3041 * (to avoid false positives from pte_same). For
3042 * further safety release the lock after the swap_free
3043 * so that the swap count won't change under a
3044 * parallel locked swapcache.
3045 */
3046 unlock_page(swapcache);
3047 page_cache_release(swapcache);
3048 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003049
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003050 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003051 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3052 if (ret & VM_FAULT_ERROR)
3053 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 goto out;
3055 }
3056
3057 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003058 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003059unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003060 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061out:
3062 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003063out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003064 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003065 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003066out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003067 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003068out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003069 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003070 if (swapcache) {
3071 unlock_page(swapcache);
3072 page_cache_release(swapcache);
3073 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003074 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075}
3076
3077/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003078 * This is like a special single-page "expand_{down|up}wards()",
3079 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003080 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003081 */
3082static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3083{
3084 address &= PAGE_MASK;
3085 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003086 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003087
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003088 /*
3089 * Is there a mapping abutting this one below?
3090 *
3091 * That's only ok if it's the same stack mapping
3092 * that has gotten split..
3093 */
3094 if (prev && prev->vm_end == address)
3095 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3096
Michal Hockod05f3162011-05-24 17:11:44 -07003097 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003098 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003099 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3100 struct vm_area_struct *next = vma->vm_next;
3101
3102 /* As VM_GROWSDOWN but s/below/above/ */
3103 if (next && next->vm_start == address + PAGE_SIZE)
3104 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3105
3106 expand_upwards(vma, address + PAGE_SIZE);
3107 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003108 return 0;
3109}
3110
3111/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003112 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3113 * but allow concurrent faults), and pte mapped but not yet locked.
3114 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003116static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3117 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003118 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003120 struct page *page;
3121 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123
Linus Torvalds11ac5522010-08-14 11:44:56 -07003124 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003125
Linus Torvalds11ac5522010-08-14 11:44:56 -07003126 /* Check if we need to add a guard page to the stack */
3127 if (check_stack_guard_page(vma, address) < 0)
3128 return VM_FAULT_SIGBUS;
3129
3130 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003131 if (!(flags & FAULT_FLAG_WRITE)) {
3132 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3133 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003134 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003135 if (!pte_none(*page_table))
3136 goto unlock;
3137 goto setpte;
3138 }
3139
Nick Piggin557ed1f2007-10-16 01:24:40 -07003140 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003141 if (unlikely(anon_vma_prepare(vma)))
3142 goto oom;
3143 page = alloc_zeroed_user_highpage_movable(vma, address);
3144 if (!page)
3145 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003146 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003148 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003149 goto oom_free_page;
3150
Nick Piggin557ed1f2007-10-16 01:24:40 -07003151 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003152 if (vma->vm_flags & VM_WRITE)
3153 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003154
Nick Piggin557ed1f2007-10-16 01:24:40 -07003155 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003156 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003157 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003158
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003159 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003160 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003161setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003162 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163
3164 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003165 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003166unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003167 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003168 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003169release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003170 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003171 page_cache_release(page);
3172 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003173oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003174 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003175oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 return VM_FAULT_OOM;
3177}
3178
3179/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003180 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003182 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3183 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 *
3185 * As this is called only for pages that do not currently exist, we
3186 * do not need to flush old virtual caches or the TLB.
3187 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003188 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003189 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003190 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003192static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003193 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003194 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003196 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003197 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003198 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003199 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003202 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003203 struct vm_fault vmf;
3204 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003205 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003206
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003207 /*
3208 * If we do COW later, allocate page befor taking lock_page()
3209 * on the file cache page. This will reduce lock holding time.
3210 */
3211 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3212
3213 if (unlikely(anon_vma_prepare(vma)))
3214 return VM_FAULT_OOM;
3215
3216 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3217 if (!cow_page)
3218 return VM_FAULT_OOM;
3219
3220 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3221 page_cache_release(cow_page);
3222 return VM_FAULT_OOM;
3223 }
3224 } else
3225 cow_page = NULL;
3226
Nick Piggind0217ac2007-07-19 01:47:03 -07003227 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3228 vmf.pgoff = pgoff;
3229 vmf.flags = flags;
3230 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003232 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003233 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3234 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003235 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236
Andi Kleena3b947e2009-09-16 11:50:08 +02003237 if (unlikely(PageHWPoison(vmf.page))) {
3238 if (ret & VM_FAULT_LOCKED)
3239 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003240 ret = VM_FAULT_HWPOISON;
3241 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003242 }
3243
Nick Piggind00806b2007-07-19 01:46:57 -07003244 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003245 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003246 * locked.
3247 */
Nick Piggin83c54072007-07-19 01:47:05 -07003248 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003249 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003250 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003251 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003252
Linus Torvalds1da177e2005-04-16 15:20:36 -07003253 /*
3254 * Should we do an early C-O-W break?
3255 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003256 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003257 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003258 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003259 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003260 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003261 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003262 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003263 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003264 /*
3265 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003266 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003267 * to become writable
3268 */
Mark Fasheh69676142007-07-19 01:47:00 -07003269 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003270 int tmp;
3271
Mark Fasheh69676142007-07-19 01:47:00 -07003272 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003273 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003274 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3275 if (unlikely(tmp &
3276 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3277 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003278 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003279 }
Nick Pigginb827e492009-04-30 15:08:16 -07003280 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3281 lock_page(page);
3282 if (!page->mapping) {
3283 ret = 0; /* retry the fault */
3284 unlock_page(page);
3285 goto unwritable_page;
3286 }
3287 } else
3288 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003289 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003290 }
3291 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003292
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 }
3294
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003295 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296
3297 /*
3298 * This silly early PAGE_DIRTY setting removes a race
3299 * due to the bad i386 page protection. But it's valid
3300 * for other architectures too.
3301 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003302 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 * an exclusive copy of the page, or this is a shared mapping,
3304 * so we can make it writable and dirty to avoid having to
3305 * handle that later.
3306 */
3307 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003308 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003309 flush_icache_page(vma, page);
3310 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003311 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003314 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003315 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003316 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003317 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003318 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003319 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003320 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003321 get_page(dirty_page);
3322 }
Hugh Dickins42946212005-10-29 18:16:05 -07003323 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003324 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003325
3326 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003327 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003329 if (cow_page)
3330 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003331 if (anon)
3332 page_cache_release(page);
3333 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003334 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335 }
3336
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003337 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003338
Nick Pigginb827e492009-04-30 15:08:16 -07003339 if (dirty_page) {
3340 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003341 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003342
3343 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003344 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003345 unlock_page(dirty_page);
3346 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003347 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003348 /*
3349 * Some device drivers do not set page.mapping but still
3350 * dirty their pages
3351 */
3352 balance_dirty_pages_ratelimited(mapping);
3353 }
3354
3355 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003356 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003357 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003358 } else {
3359 unlock_page(vmf.page);
3360 if (anon)
3361 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003362 }
Nick Piggind00806b2007-07-19 01:46:57 -07003363
Nick Piggin83c54072007-07-19 01:47:05 -07003364 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003365
3366unwritable_page:
3367 page_cache_release(page);
3368 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003369uncharge_out:
3370 /* fs's fault handler get error */
3371 if (cow_page) {
3372 mem_cgroup_uncharge_page(cow_page);
3373 page_cache_release(cow_page);
3374 }
3375 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003376}
Nick Piggind00806b2007-07-19 01:46:57 -07003377
Nick Piggin54cb8822007-07-19 01:46:59 -07003378static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3379 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003380 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003381{
3382 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003383 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003384
Hugh Dickins16abfa02007-10-04 16:56:06 +01003385 pte_unmap(page_table);
3386 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003387}
3388
Jes Sorensenf4b81802006-09-27 01:50:10 -07003389/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 * Fault of a previously existing named mapping. Repopulate the pte
3391 * from the encoded file_pte if possible. This enables swappable
3392 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003393 *
3394 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3395 * but allow concurrent faults), and pte mapped but not yet locked.
3396 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003398static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003399 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003400 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401{
Hugh Dickins65500d22005-10-29 18:15:59 -07003402 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003404 flags |= FAULT_FLAG_NONLINEAR;
3405
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003406 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003407 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408
Hugh Dickins2509ef22009-01-06 14:40:10 -08003409 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003410 /*
3411 * Page table corrupted: show pte and kill process.
3412 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003413 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003414 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003415 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003416
3417 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003418 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419}
3420
3421/*
3422 * These routines also need to handle stuff like marking pages dirty
3423 * and/or accessed for architectures that don't do it in hardware (most
3424 * RISC architectures). The early dirtying is also good on the i386.
3425 *
3426 * There is also a hook called "update_mmu_cache()" that architectures
3427 * with external mmu caches can use to update those (ie the Sparc or
3428 * PowerPC hashed page tables that act as extended TLBs).
3429 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003430 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3431 * but allow concurrent faults), and pte mapped but not yet locked.
3432 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003434int handle_pte_fault(struct mm_struct *mm,
3435 struct vm_area_struct *vma, unsigned long address,
3436 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437{
3438 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003439 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003441 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003443 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003444 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003445 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003446 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003447 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003448 }
3449 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003450 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003451 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003453 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003454 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003455 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003456 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 }
3458
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003459 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003460 spin_lock(ptl);
3461 if (unlikely(!pte_same(*pte, entry)))
3462 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003463 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003465 return do_wp_page(mm, vma, address,
3466 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 entry = pte_mkdirty(entry);
3468 }
3469 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003470 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003471 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003472 } else {
3473 /*
3474 * This is needed only for protection faults but the arch code
3475 * is not yet telling us if this is a protection fault or not.
3476 * This still avoids useless tlb flushes for .text page faults
3477 * with threads.
3478 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003479 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003480 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003481 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003482unlock:
3483 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003484 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485}
3486
3487/*
3488 * By the time we get here, we already hold the mm semaphore
3489 */
Nick Piggin83c54072007-07-19 01:47:05 -07003490int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003491 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492{
3493 pgd_t *pgd;
3494 pud_t *pud;
3495 pmd_t *pmd;
3496 pte_t *pte;
3497
3498 __set_current_state(TASK_RUNNING);
3499
Christoph Lameterf8891e52006-06-30 01:55:45 -07003500 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003501 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003503 /* do counter updates before entering really critical section. */
3504 check_sync_rss_stat(current);
3505
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003506 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003507 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508
David Rientjes1f1d06c2012-05-29 15:06:23 -07003509retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511 pud = pud_alloc(mm, pgd, address);
3512 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003513 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 pmd = pmd_alloc(mm, pud, address);
3515 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003516 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003517 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3518 if (!vma->vm_ops)
3519 return do_huge_pmd_anonymous_page(mm, vma, address,
3520 pmd, flags);
3521 } else {
3522 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003523 int ret;
3524
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003525 barrier();
3526 if (pmd_trans_huge(orig_pmd)) {
3527 if (flags & FAULT_FLAG_WRITE &&
3528 !pmd_write(orig_pmd) &&
David Rientjes1f1d06c2012-05-29 15:06:23 -07003529 !pmd_trans_splitting(orig_pmd)) {
3530 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3531 orig_pmd);
3532 /*
3533 * If COW results in an oom, the huge pmd will
3534 * have been split, so retry the fault on the
3535 * pte for a smaller charge.
3536 */
3537 if (unlikely(ret & VM_FAULT_OOM))
3538 goto retry;
3539 return ret;
3540 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003541 return 0;
3542 }
3543 }
3544
3545 /*
3546 * Use __pte_alloc instead of pte_alloc_map, because we can't
3547 * run pte_offset_map on the pmd, if an huge pmd could
3548 * materialize from under us from a different thread.
3549 */
Mel Gormancc036382011-04-27 15:26:56 -07003550 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003551 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003552 /* if an huge pmd materialized from under us just retry later */
3553 if (unlikely(pmd_trans_huge(*pmd)))
3554 return 0;
3555 /*
3556 * A regular pmd is established and it can't morph into a huge pmd
3557 * from under us anymore at this point because we hold the mmap_sem
3558 * read mode and khugepaged takes it in write mode. So now it's
3559 * safe to run pte_offset_map().
3560 */
3561 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003563 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564}
3565
3566#ifndef __PAGETABLE_PUD_FOLDED
3567/*
3568 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003569 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003571int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003573 pud_t *new = pud_alloc_one(mm, address);
3574 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003575 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Nick Piggin362a61a2008-05-14 06:37:36 +02003577 smp_wmb(); /* See comment in __pte_alloc */
3578
Hugh Dickins872fec12005-10-29 18:16:21 -07003579 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003580 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003581 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003582 else
3583 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003584 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003585 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586}
3587#endif /* __PAGETABLE_PUD_FOLDED */
3588
3589#ifndef __PAGETABLE_PMD_FOLDED
3590/*
3591 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003592 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003594int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003596 pmd_t *new = pmd_alloc_one(mm, address);
3597 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003598 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599
Nick Piggin362a61a2008-05-14 06:37:36 +02003600 smp_wmb(); /* See comment in __pte_alloc */
3601
Hugh Dickins872fec12005-10-29 18:16:21 -07003602 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003604 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003605 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003606 else
3607 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003609 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003610 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003611 else
3612 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003614 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003615 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616}
3617#endif /* __PAGETABLE_PMD_FOLDED */
3618
3619int make_pages_present(unsigned long addr, unsigned long end)
3620{
3621 int ret, len, write;
3622 struct vm_area_struct * vma;
3623
3624 vma = find_vma(current->mm, addr);
3625 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003626 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003627 /*
3628 * We want to touch writable mappings with a write fault in order
3629 * to break COW, except for shared mappings because these don't COW
3630 * and we would not want to dirty them for nothing.
3631 */
3632 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003633 BUG_ON(addr >= end);
3634 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003635 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 ret = get_user_pages(current, current->mm, addr,
3637 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003638 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003640 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641}
3642
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643#if !defined(__HAVE_ARCH_GATE_AREA)
3644
3645#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003646static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647
3648static int __init gate_vma_init(void)
3649{
3650 gate_vma.vm_mm = NULL;
3651 gate_vma.vm_start = FIXADDR_USER_START;
3652 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003653 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3654 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003655
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656 return 0;
3657}
3658__initcall(gate_vma_init);
3659#endif
3660
Stephen Wilson31db58b2011-03-13 15:49:15 -04003661struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662{
3663#ifdef AT_SYSINFO_EHDR
3664 return &gate_vma;
3665#else
3666 return NULL;
3667#endif
3668}
3669
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003670int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671{
3672#ifdef AT_SYSINFO_EHDR
3673 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3674 return 1;
3675#endif
3676 return 0;
3677}
3678
3679#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003680
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003681static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003682 pte_t **ptepp, spinlock_t **ptlp)
3683{
3684 pgd_t *pgd;
3685 pud_t *pud;
3686 pmd_t *pmd;
3687 pte_t *ptep;
3688
3689 pgd = pgd_offset(mm, address);
3690 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3691 goto out;
3692
3693 pud = pud_offset(pgd, address);
3694 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3695 goto out;
3696
3697 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003698 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003699 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3700 goto out;
3701
3702 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3703 if (pmd_huge(*pmd))
3704 goto out;
3705
3706 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3707 if (!ptep)
3708 goto out;
3709 if (!pte_present(*ptep))
3710 goto unlock;
3711 *ptepp = ptep;
3712 return 0;
3713unlock:
3714 pte_unmap_unlock(ptep, *ptlp);
3715out:
3716 return -EINVAL;
3717}
3718
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003719static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3720 pte_t **ptepp, spinlock_t **ptlp)
3721{
3722 int res;
3723
3724 /* (void) is needed to make gcc happy */
3725 (void) __cond_lock(*ptlp,
3726 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3727 return res;
3728}
3729
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003730/**
3731 * follow_pfn - look up PFN at a user virtual address
3732 * @vma: memory mapping
3733 * @address: user virtual address
3734 * @pfn: location to store found PFN
3735 *
3736 * Only IO mappings and raw PFN mappings are allowed.
3737 *
3738 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3739 */
3740int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3741 unsigned long *pfn)
3742{
3743 int ret = -EINVAL;
3744 spinlock_t *ptl;
3745 pte_t *ptep;
3746
3747 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3748 return ret;
3749
3750 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3751 if (ret)
3752 return ret;
3753 *pfn = pte_pfn(*ptep);
3754 pte_unmap_unlock(ptep, ptl);
3755 return 0;
3756}
3757EXPORT_SYMBOL(follow_pfn);
3758
Rik van Riel28b2ee22008-07-23 21:27:05 -07003759#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003760int follow_phys(struct vm_area_struct *vma,
3761 unsigned long address, unsigned int flags,
3762 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003763{
Johannes Weiner03668a42009-06-16 15:32:34 -07003764 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003765 pte_t *ptep, pte;
3766 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003767
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003768 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3769 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003770
Johannes Weiner03668a42009-06-16 15:32:34 -07003771 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003772 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003773 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003774
Rik van Riel28b2ee22008-07-23 21:27:05 -07003775 if ((flags & FOLL_WRITE) && !pte_write(pte))
3776 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003777
3778 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003779 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003780
Johannes Weiner03668a42009-06-16 15:32:34 -07003781 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003782unlock:
3783 pte_unmap_unlock(ptep, ptl);
3784out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003785 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003786}
3787
3788int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3789 void *buf, int len, int write)
3790{
3791 resource_size_t phys_addr;
3792 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003793 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003794 int offset = addr & (PAGE_SIZE-1);
3795
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003796 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003797 return -EINVAL;
3798
3799 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3800 if (write)
3801 memcpy_toio(maddr + offset, buf, len);
3802 else
3803 memcpy_fromio(buf, maddr + offset, len);
3804 iounmap(maddr);
3805
3806 return len;
3807}
3808#endif
3809
David Howells0ec76a12006-09-27 01:50:15 -07003810/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003811 * Access another process' address space as given in mm. If non-NULL, use the
3812 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003813 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003814static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3815 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003816{
David Howells0ec76a12006-09-27 01:50:15 -07003817 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003818 void *old_buf = buf;
3819
David Howells0ec76a12006-09-27 01:50:15 -07003820 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003821 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003822 while (len) {
3823 int bytes, ret, offset;
3824 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003825 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003826
3827 ret = get_user_pages(tsk, mm, addr, 1,
3828 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003829 if (ret <= 0) {
3830 /*
3831 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3832 * we can access using slightly different code.
3833 */
3834#ifdef CONFIG_HAVE_IOREMAP_PROT
3835 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003836 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003837 break;
3838 if (vma->vm_ops && vma->vm_ops->access)
3839 ret = vma->vm_ops->access(vma, addr, buf,
3840 len, write);
3841 if (ret <= 0)
3842#endif
3843 break;
3844 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003845 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003846 bytes = len;
3847 offset = addr & (PAGE_SIZE-1);
3848 if (bytes > PAGE_SIZE-offset)
3849 bytes = PAGE_SIZE-offset;
3850
3851 maddr = kmap(page);
3852 if (write) {
3853 copy_to_user_page(vma, page, addr,
3854 maddr + offset, buf, bytes);
3855 set_page_dirty_lock(page);
3856 } else {
3857 copy_from_user_page(vma, page, addr,
3858 buf, maddr + offset, bytes);
3859 }
3860 kunmap(page);
3861 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003862 }
David Howells0ec76a12006-09-27 01:50:15 -07003863 len -= bytes;
3864 buf += bytes;
3865 addr += bytes;
3866 }
3867 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003868
3869 return buf - old_buf;
3870}
Andi Kleen03252912008-01-30 13:33:18 +01003871
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003872/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003873 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003874 * @mm: the mm_struct of the target address space
3875 * @addr: start address to access
3876 * @buf: source or destination buffer
3877 * @len: number of bytes to transfer
3878 * @write: whether the access is a write
3879 *
3880 * The caller must hold a reference on @mm.
3881 */
3882int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3883 void *buf, int len, int write)
3884{
3885 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3886}
3887
Andi Kleen03252912008-01-30 13:33:18 +01003888/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003889 * Access another process' address space.
3890 * Source/target buffer must be kernel space,
3891 * Do not walk the page table directly, use get_user_pages
3892 */
3893int access_process_vm(struct task_struct *tsk, unsigned long addr,
3894 void *buf, int len, int write)
3895{
3896 struct mm_struct *mm;
3897 int ret;
3898
3899 mm = get_task_mm(tsk);
3900 if (!mm)
3901 return 0;
3902
3903 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3904 mmput(mm);
3905
3906 return ret;
3907}
3908
Andi Kleen03252912008-01-30 13:33:18 +01003909/*
3910 * Print the name of a VMA.
3911 */
3912void print_vma_addr(char *prefix, unsigned long ip)
3913{
3914 struct mm_struct *mm = current->mm;
3915 struct vm_area_struct *vma;
3916
Ingo Molnare8bff742008-02-13 20:21:06 +01003917 /*
3918 * Do not print if we are in atomic
3919 * contexts (in exception stacks, etc.):
3920 */
3921 if (preempt_count())
3922 return;
3923
Andi Kleen03252912008-01-30 13:33:18 +01003924 down_read(&mm->mmap_sem);
3925 vma = find_vma(mm, ip);
3926 if (vma && vma->vm_file) {
3927 struct file *f = vma->vm_file;
3928 char *buf = (char *)__get_free_page(GFP_KERNEL);
3929 if (buf) {
3930 char *p, *s;
3931
Jan Blunckcf28b482008-02-14 19:38:44 -08003932 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003933 if (IS_ERR(p))
3934 p = "?";
3935 s = strrchr(p, '/');
3936 if (s)
3937 p = s+1;
3938 printk("%s%s[%lx+%lx]", prefix, p,
3939 vma->vm_start,
3940 vma->vm_end - vma->vm_start);
3941 free_page((unsigned long)buf);
3942 }
3943 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003944 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003945}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003946
3947#ifdef CONFIG_PROVE_LOCKING
3948void might_fault(void)
3949{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003950 /*
3951 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3952 * holding the mmap_sem, this is safe because kernel memory doesn't
3953 * get paged out, therefore we'll never actually fault, and the
3954 * below annotations will generate false positives.
3955 */
3956 if (segment_eq(get_fs(), KERNEL_DS))
3957 return;
3958
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003959 might_sleep();
3960 /*
3961 * it would be nicer only to annotate paths which are not under
3962 * pagefault_disable, however that requires a larger audit and
3963 * providing helpers like get_user_atomic.
3964 */
3965 if (!in_atomic() && current->mm)
3966 might_lock_read(&current->mm->mmap_sem);
3967}
3968EXPORT_SYMBOL(might_fault);
3969#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003970
3971#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3972static void clear_gigantic_page(struct page *page,
3973 unsigned long addr,
3974 unsigned int pages_per_huge_page)
3975{
3976 int i;
3977 struct page *p = page;
3978
3979 might_sleep();
3980 for (i = 0; i < pages_per_huge_page;
3981 i++, p = mem_map_next(p, page, i)) {
3982 cond_resched();
3983 clear_user_highpage(p, addr + i * PAGE_SIZE);
3984 }
3985}
3986void clear_huge_page(struct page *page,
3987 unsigned long addr, unsigned int pages_per_huge_page)
3988{
3989 int i;
3990
3991 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3992 clear_gigantic_page(page, addr, pages_per_huge_page);
3993 return;
3994 }
3995
3996 might_sleep();
3997 for (i = 0; i < pages_per_huge_page; i++) {
3998 cond_resched();
3999 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4000 }
4001}
4002
4003static void copy_user_gigantic_page(struct page *dst, struct page *src,
4004 unsigned long addr,
4005 struct vm_area_struct *vma,
4006 unsigned int pages_per_huge_page)
4007{
4008 int i;
4009 struct page *dst_base = dst;
4010 struct page *src_base = src;
4011
4012 for (i = 0; i < pages_per_huge_page; ) {
4013 cond_resched();
4014 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4015
4016 i++;
4017 dst = mem_map_next(dst, dst_base, i);
4018 src = mem_map_next(src, src_base, i);
4019 }
4020}
4021
4022void copy_user_huge_page(struct page *dst, struct page *src,
4023 unsigned long addr, struct vm_area_struct *vma,
4024 unsigned int pages_per_huge_page)
4025{
4026 int i;
4027
4028 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4029 copy_user_gigantic_page(dst, src, addr, vma,
4030 pages_per_huge_page);
4031 return;
4032 }
4033
4034 might_sleep();
4035 for (i = 0; i < pages_per_huge_page; i++) {
4036 cond_resched();
4037 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4038 }
4039}
4040#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */