blob: fb135ba4aba90349e58c79d49838860cb7db8105 [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
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700715static inline bool 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;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001042 unsigned long mmun_start; /* For mmu_notifiers */
1043 unsigned long mmun_end; /* For mmu_notifiers */
1044 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001045 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046
Nick Piggind9928952005-08-28 16:49:11 +10001047 /*
1048 * Don't copy ptes where a page fault will fill them correctly.
1049 * Fork becomes much lighter when there are big shared or private
1050 * readonly mappings. The tradeoff is that copy_page_range is more
1051 * efficient than faulting.
1052 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001053 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1054 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001055 if (!vma->anon_vma)
1056 return 0;
1057 }
1058
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 if (is_vm_hugetlb_page(vma))
1060 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1061
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001062 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001063 /*
1064 * We do not free on error cases below as remove_vma
1065 * gets called on error from higher level routine
1066 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001067 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001068 if (ret)
1069 return ret;
1070 }
1071
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001072 /*
1073 * We need to invalidate the secondary MMU mappings only when
1074 * there could be a permission downgrade on the ptes of the
1075 * parent mm. And a permission downgrade will only happen if
1076 * is_cow_mapping() returns true.
1077 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001078 is_cow = is_cow_mapping(vma->vm_flags);
1079 mmun_start = addr;
1080 mmun_end = end;
1081 if (is_cow)
1082 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1083 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001084
1085 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 dst_pgd = pgd_offset(dst_mm, addr);
1087 src_pgd = pgd_offset(src_mm, addr);
1088 do {
1089 next = pgd_addr_end(addr, end);
1090 if (pgd_none_or_clear_bad(src_pgd))
1091 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001092 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1093 vma, addr, next))) {
1094 ret = -ENOMEM;
1095 break;
1096 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001098
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001099 if (is_cow)
1100 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001101 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102}
1103
Robin Holt51c6f662005-11-13 16:06:42 -08001104static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001105 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001107 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108{
Nick Pigginb5810032005-10-29 18:16:12 -07001109 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001110 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001111 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001112 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001113 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001114 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001115
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001116again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001117 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001118 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1119 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001120 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 do {
1122 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001123 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001125 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001126
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001128 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001129
Linus Torvalds6aab3412005-11-28 14:34:23 -08001130 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 if (unlikely(details) && page) {
1132 /*
1133 * unmap_shared_mapping_pages() wants to
1134 * invalidate cache without truncating:
1135 * unmap shared but keep private pages.
1136 */
1137 if (details->check_mapping &&
1138 details->check_mapping != page->mapping)
1139 continue;
1140 /*
1141 * Each page->index must be checked when
1142 * invalidating or truncating nonlinear.
1143 */
1144 if (details->nonlinear_vma &&
1145 (page->index < details->first_index ||
1146 page->index > details->last_index))
1147 continue;
1148 }
Nick Pigginb5810032005-10-29 18:16:12 -07001149 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001150 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 tlb_remove_tlb_entry(tlb, pte, addr);
1152 if (unlikely(!page))
1153 continue;
1154 if (unlikely(details) && details->nonlinear_vma
1155 && linear_page_index(details->nonlinear_vma,
1156 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001157 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001160 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001161 else {
1162 if (pte_dirty(ptent))
1163 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001164 if (pte_young(ptent) &&
1165 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001166 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001167 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001168 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001169 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001170 if (unlikely(page_mapcount(page) < 0))
1171 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001172 force_flush = !__tlb_remove_page(tlb, page);
1173 if (force_flush)
1174 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 continue;
1176 }
1177 /*
1178 * If details->check_mapping, we leave swap entries;
1179 * if details->nonlinear_vma, we leave file entries.
1180 */
1181 if (unlikely(details))
1182 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001183 if (pte_file(ptent)) {
1184 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1185 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001186 } else {
1187 swp_entry_t entry = pte_to_swp_entry(ptent);
1188
1189 if (!non_swap_entry(entry))
1190 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001191 else if (is_migration_entry(entry)) {
1192 struct page *page;
1193
1194 page = migration_entry_to_page(entry);
1195
1196 if (PageAnon(page))
1197 rss[MM_ANONPAGES]--;
1198 else
1199 rss[MM_FILEPAGES]--;
1200 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001201 if (unlikely(!free_swap_and_cache(entry)))
1202 print_bad_pte(vma, addr, ptent, NULL);
1203 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001204 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001205 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001206
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001207 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001208 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001209 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001210
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001211 /*
1212 * mmu_gather ran out of room to batch pages, we break out of
1213 * the PTE lock to avoid doing the potential expensive TLB invalidate
1214 * and page-free while holding it.
1215 */
1216 if (force_flush) {
1217 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001218
1219#ifdef HAVE_GENERIC_MMU_GATHER
1220 tlb->start = addr;
1221 tlb->end = end;
1222#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001223 tlb_flush_mmu(tlb);
1224 if (addr != end)
1225 goto again;
1226 }
1227
Robin Holt51c6f662005-11-13 16:06:42 -08001228 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229}
1230
Robin Holt51c6f662005-11-13 16:06:42 -08001231static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001232 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001234 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235{
1236 pmd_t *pmd;
1237 unsigned long next;
1238
1239 pmd = pmd_offset(pud, addr);
1240 do {
1241 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001242 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001243 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001244#ifdef CONFIG_DEBUG_VM
1245 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1246 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1247 __func__, addr, end,
1248 vma->vm_start,
1249 vma->vm_end);
1250 BUG();
1251 }
1252#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001253 split_huge_page_pmd(vma->vm_mm, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001254 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001255 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001256 /* fall through */
1257 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001258 /*
1259 * Here there can be other concurrent MADV_DONTNEED or
1260 * trans huge page faults running, and if the pmd is
1261 * none or trans huge it can change under us. This is
1262 * because MADV_DONTNEED holds the mmap_sem in read
1263 * mode.
1264 */
1265 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1266 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001267 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001268next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001269 cond_resched();
1270 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001271
1272 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273}
1274
Robin Holt51c6f662005-11-13 16:06:42 -08001275static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001276 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001278 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279{
1280 pud_t *pud;
1281 unsigned long next;
1282
1283 pud = pud_offset(pgd, addr);
1284 do {
1285 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001286 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001288 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1289 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001290
1291 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292}
1293
Al Viro038c7aa2012-03-05 13:25:09 -05001294static void unmap_page_range(struct mmu_gather *tlb,
1295 struct vm_area_struct *vma,
1296 unsigned long addr, unsigned long end,
1297 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298{
1299 pgd_t *pgd;
1300 unsigned long next;
1301
1302 if (details && !details->check_mapping && !details->nonlinear_vma)
1303 details = NULL;
1304
1305 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001306 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 tlb_start_vma(tlb, vma);
1308 pgd = pgd_offset(vma->vm_mm, addr);
1309 do {
1310 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001311 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001313 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1314 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001316 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318
Al Virof5cc4ee2012-03-05 14:14:20 -05001319
1320static void unmap_single_vma(struct mmu_gather *tlb,
1321 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001322 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001323 struct zap_details *details)
1324{
1325 unsigned long start = max(vma->vm_start, start_addr);
1326 unsigned long end;
1327
1328 if (start >= vma->vm_end)
1329 return;
1330 end = min(vma->vm_end, end_addr);
1331 if (end <= vma->vm_start)
1332 return;
1333
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301334 if (vma->vm_file)
1335 uprobe_munmap(vma, start, end);
1336
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001337 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001338 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001339
1340 if (start != end) {
1341 if (unlikely(is_vm_hugetlb_page(vma))) {
1342 /*
1343 * It is undesirable to test vma->vm_file as it
1344 * should be non-null for valid hugetlb area.
1345 * However, vm_file will be NULL in the error
1346 * cleanup path of do_mmap_pgoff. When
1347 * hugetlbfs ->mmap method fails,
1348 * do_mmap_pgoff() nullifies vma->vm_file
1349 * before calling this function to clean up.
1350 * Since no pte has actually been setup, it is
1351 * safe to do nothing in this case.
1352 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001353 if (vma->vm_file) {
1354 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001355 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001356 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1357 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001358 } else
1359 unmap_page_range(tlb, vma, start, end, details);
1360 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361}
1362
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363/**
1364 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001365 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 * @vma: the starting vma
1367 * @start_addr: virtual address at which to start unmapping
1368 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001370 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 * Only addresses between `start' and `end' will be unmapped.
1373 *
1374 * The VMA list must be sorted in ascending virtual address order.
1375 *
1376 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1377 * range after unmap_vmas() returns. So the only responsibility here is to
1378 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1379 * drops the lock and schedules.
1380 */
Al Viro6e8bb012012-03-05 13:41:15 -05001381void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001383 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001385 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001387 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001388 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001389 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001390 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391}
1392
1393/**
1394 * zap_page_range - remove user pages in a given range
1395 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001396 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 * @size: number of bytes to zap
1398 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001399 *
1400 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001402void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 unsigned long size, struct zap_details *details)
1404{
1405 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001406 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001407 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001410 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001411 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001412 mmu_notifier_invalidate_range_start(mm, start, end);
1413 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001414 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001415 mmu_notifier_invalidate_range_end(mm, start, end);
1416 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417}
1418
Jack Steinerc627f9c2008-07-29 22:33:53 -07001419/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001420 * zap_page_range_single - remove user pages in a given range
1421 * @vma: vm_area_struct holding the applicable pages
1422 * @address: starting address of pages to zap
1423 * @size: number of bytes to zap
1424 * @details: details of nonlinear truncation or shared cache invalidation
1425 *
1426 * The range must fit into one VMA.
1427 */
1428static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1429 unsigned long size, struct zap_details *details)
1430{
1431 struct mm_struct *mm = vma->vm_mm;
1432 struct mmu_gather tlb;
1433 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001434
1435 lru_add_drain();
1436 tlb_gather_mmu(&tlb, mm, 0);
1437 update_hiwater_rss(mm);
1438 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001439 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001440 mmu_notifier_invalidate_range_end(mm, address, end);
1441 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442}
1443
Jack Steinerc627f9c2008-07-29 22:33:53 -07001444/**
1445 * zap_vma_ptes - remove ptes mapping the vma
1446 * @vma: vm_area_struct holding ptes to be zapped
1447 * @address: starting address of pages to zap
1448 * @size: number of bytes to zap
1449 *
1450 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1451 *
1452 * The entire address range must be fully contained within the vma.
1453 *
1454 * Returns 0 if successful.
1455 */
1456int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1457 unsigned long size)
1458{
1459 if (address < vma->vm_start || address + size > vma->vm_end ||
1460 !(vma->vm_flags & VM_PFNMAP))
1461 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001462 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001463 return 0;
1464}
1465EXPORT_SYMBOL_GPL(zap_vma_ptes);
1466
Johannes Weiner142762b2010-05-24 14:32:39 -07001467/**
1468 * follow_page - look up a page descriptor from a user-virtual address
1469 * @vma: vm_area_struct mapping @address
1470 * @address: virtual address to look up
1471 * @flags: flags modifying lookup behaviour
1472 *
1473 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1474 *
1475 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1476 * an error pointer if there is a mapping to something not represented
1477 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001479struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001480 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481{
1482 pgd_t *pgd;
1483 pud_t *pud;
1484 pmd_t *pmd;
1485 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001486 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001488 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001490 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1491 if (!IS_ERR(page)) {
1492 BUG_ON(flags & FOLL_GET);
1493 goto out;
1494 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 pgd = pgd_offset(mm, address);
1498 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001499 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
1501 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001502 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001503 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001504 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001505 BUG_ON(flags & FOLL_GET);
1506 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1507 goto out;
1508 }
1509 if (unlikely(pud_bad(*pud)))
1510 goto no_page_table;
1511
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001513 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001514 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001515 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001516 BUG_ON(flags & FOLL_GET);
1517 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1518 goto out;
1519 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001520 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001521 if (flags & FOLL_SPLIT) {
1522 split_huge_page_pmd(mm, pmd);
1523 goto split_fallthrough;
1524 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001525 spin_lock(&mm->page_table_lock);
1526 if (likely(pmd_trans_huge(*pmd))) {
1527 if (unlikely(pmd_trans_splitting(*pmd))) {
1528 spin_unlock(&mm->page_table_lock);
1529 wait_split_huge_page(vma->anon_vma, pmd);
1530 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001531 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001532 pmd, flags);
1533 spin_unlock(&mm->page_table_lock);
1534 goto out;
1535 }
1536 } else
1537 spin_unlock(&mm->page_table_lock);
1538 /* fall through */
1539 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001540split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001541 if (unlikely(pmd_bad(*pmd)))
1542 goto no_page_table;
1543
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001544 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545
1546 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001547 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001548 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001549 if ((flags & FOLL_WRITE) && !pte_write(pte))
1550 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001551
Linus Torvalds6aab3412005-11-28 14:34:23 -08001552 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001553 if (unlikely(!page)) {
1554 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001555 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001556 goto bad_page;
1557 page = pte_page(pte);
1558 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001559
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001560 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001561 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001562 if (flags & FOLL_TOUCH) {
1563 if ((flags & FOLL_WRITE) &&
1564 !pte_dirty(pte) && !PageDirty(page))
1565 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001566 /*
1567 * pte_mkyoung() would be more correct here, but atomic care
1568 * is needed to avoid losing the dirty bit: it is easier to use
1569 * mark_page_accessed().
1570 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001571 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001573 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001574 /*
1575 * The preliminary mapping check is mainly to avoid the
1576 * pointless overhead of lock_page on the ZERO_PAGE
1577 * which might bounce very badly if there is contention.
1578 *
1579 * If the page is already locked, we don't need to
1580 * handle it now - vmscan will handle it later if and
1581 * when it attempts to reclaim the page.
1582 */
1583 if (page->mapping && trylock_page(page)) {
1584 lru_add_drain(); /* push cached pages to LRU */
1585 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001586 * Because we lock page here, and migration is
1587 * blocked by the pte's page reference, and we
1588 * know the page is still mapped, we don't even
1589 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001590 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001591 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001592 unlock_page(page);
1593 }
1594 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001595unlock:
1596 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001598 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001600bad_page:
1601 pte_unmap_unlock(ptep, ptl);
1602 return ERR_PTR(-EFAULT);
1603
1604no_page:
1605 pte_unmap_unlock(ptep, ptl);
1606 if (!pte_none(pte))
1607 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001608
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001609no_page_table:
1610 /*
1611 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001612 * has touched so far, we don't want to allocate unnecessary pages or
1613 * page tables. Return error instead of NULL to skip handle_mm_fault,
1614 * then get_dump_page() will return NULL to leave a hole in the dump.
1615 * But we can only make this optimization where a hole would surely
1616 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001617 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001618 if ((flags & FOLL_DUMP) &&
1619 (!vma->vm_ops || !vma->vm_ops->fault))
1620 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001621 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622}
1623
Linus Torvalds95042f92011-04-12 14:15:51 -07001624static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1625{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001626 return stack_guard_page_start(vma, addr) ||
1627 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001628}
1629
Huang Ying0014bd92011-01-30 11:15:47 +08001630/**
1631 * __get_user_pages() - pin user pages in memory
1632 * @tsk: task_struct of target task
1633 * @mm: mm_struct of target mm
1634 * @start: starting user address
1635 * @nr_pages: number of pages from start to pin
1636 * @gup_flags: flags modifying pin behaviour
1637 * @pages: array that receives pointers to the pages pinned.
1638 * Should be at least nr_pages long. Or NULL, if caller
1639 * only intends to ensure the pages are faulted in.
1640 * @vmas: array of pointers to vmas corresponding to each page.
1641 * Or NULL if the caller does not require them.
1642 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1643 *
1644 * Returns number of pages pinned. This may be fewer than the number
1645 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1646 * were pinned, returns -errno. Each page returned must be released
1647 * with a put_page() call when it is finished with. vmas will only
1648 * remain valid while mmap_sem is held.
1649 *
1650 * Must be called with mmap_sem held for read or write.
1651 *
1652 * __get_user_pages walks a process's page tables and takes a reference to
1653 * each struct page that each user address corresponds to at a given
1654 * instant. That is, it takes the page that would be accessed if a user
1655 * thread accesses the given user virtual address at that instant.
1656 *
1657 * This does not guarantee that the page exists in the user mappings when
1658 * __get_user_pages returns, and there may even be a completely different
1659 * page there in some cases (eg. if mmapped pagecache has been invalidated
1660 * and subsequently re faulted). However it does guarantee that the page
1661 * won't be freed completely. And mostly callers simply care that the page
1662 * contains data that was valid *at some point in time*. Typically, an IO
1663 * or similar operation cannot guarantee anything stronger anyway because
1664 * locks can't be held over the syscall boundary.
1665 *
1666 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1667 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1668 * appropriate) must be called after the page is finished with, and
1669 * before put_page is called.
1670 *
1671 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1672 * or mmap_sem contention, and if waiting is needed to pin all pages,
1673 * *@nonblocking will be set to 0.
1674 *
1675 * In most cases, get_user_pages or get_user_pages_fast should be used
1676 * instead of __get_user_pages. __get_user_pages should be used only if
1677 * you need some special @gup_flags.
1678 */
Nick Pigginb291f002008-10-18 20:26:44 -07001679int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001680 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001681 struct page **pages, struct vm_area_struct **vmas,
1682 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
1684 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001685 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Peter Zijlstra9d737772009-06-25 11:58:55 +02001687 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001688 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001689
1690 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1691
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 /*
1693 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001694 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001696 vm_flags = (gup_flags & FOLL_WRITE) ?
1697 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1698 vm_flags &= (gup_flags & FOLL_FORCE) ?
1699 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 i = 0;
1701
1702 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001703 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704
1705 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001706 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 pgd_t *pgd;
1709 pud_t *pud;
1710 pmd_t *pmd;
1711 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001712
1713 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001714 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 return i ? : -EFAULT;
1716 if (pg > TASK_SIZE)
1717 pgd = pgd_offset_k(pg);
1718 else
1719 pgd = pgd_offset_gate(mm, pg);
1720 BUG_ON(pgd_none(*pgd));
1721 pud = pud_offset(pgd, pg);
1722 BUG_ON(pud_none(*pud));
1723 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001724 if (pmd_none(*pmd))
1725 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001726 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001728 if (pte_none(*pte)) {
1729 pte_unmap(pte);
1730 return i ? : -EFAULT;
1731 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001732 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001734 struct page *page;
1735
Linus Torvalds95042f92011-04-12 14:15:51 -07001736 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001737 if (!page) {
1738 if (!(gup_flags & FOLL_DUMP) &&
1739 is_zero_pfn(pte_pfn(*pte)))
1740 page = pte_page(*pte);
1741 else {
1742 pte_unmap(pte);
1743 return i ? : -EFAULT;
1744 }
1745 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001746 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001747 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 }
1749 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001750 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 }
1752
Nick Pigginb291f002008-10-18 20:26:44 -07001753 if (!vma ||
1754 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001755 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 return i ? : -EFAULT;
1757
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001758 if (is_vm_hugetlb_page(vma)) {
1759 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001760 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001761 continue;
1762 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001763
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001765 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001766 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767
Ethan Solomita462e00c2007-07-15 23:38:16 -07001768 /*
Ying Han47792802009-01-06 14:40:18 -08001769 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001770 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001771 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001772 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001773 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001774
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001775 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001776 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001777 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001778 unsigned int fault_flags = 0;
1779
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001780 /* For mlock, just skip the stack guard page. */
1781 if (foll_flags & FOLL_MLOCK) {
1782 if (stack_guard_page(vma, start))
1783 goto next_page;
1784 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001785 if (foll_flags & FOLL_WRITE)
1786 fault_flags |= FAULT_FLAG_WRITE;
1787 if (nonblocking)
1788 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001789 if (foll_flags & FOLL_NOWAIT)
1790 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001791
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001792 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001793 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001794
Nick Piggin83c54072007-07-19 01:47:05 -07001795 if (ret & VM_FAULT_ERROR) {
1796 if (ret & VM_FAULT_OOM)
1797 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001798 if (ret & (VM_FAULT_HWPOISON |
1799 VM_FAULT_HWPOISON_LARGE)) {
1800 if (i)
1801 return i;
1802 else if (gup_flags & FOLL_HWPOISON)
1803 return -EHWPOISON;
1804 else
1805 return -EFAULT;
1806 }
1807 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001808 return i ? i : -EFAULT;
1809 BUG();
1810 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001811
1812 if (tsk) {
1813 if (ret & VM_FAULT_MAJOR)
1814 tsk->maj_flt++;
1815 else
1816 tsk->min_flt++;
1817 }
Nick Piggin83c54072007-07-19 01:47:05 -07001818
Michel Lespinasse53a77062011-01-13 15:46:14 -08001819 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001820 if (nonblocking)
1821 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001822 return i;
1823 }
1824
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001825 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001826 * The VM_FAULT_WRITE bit tells us that
1827 * do_wp_page has broken COW when necessary,
1828 * even if maybe_mkwrite decided not to set
1829 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001830 * page lookups as if they were reads. But only
1831 * do so when looping for pte_write is futile:
1832 * in some cases userspace may also be wanting
1833 * to write to the gotten user page, which a
1834 * read fault here might prevent (a readonly
1835 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001836 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001837 if ((ret & VM_FAULT_WRITE) &&
1838 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001839 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001840
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001841 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001843 if (IS_ERR(page))
1844 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001846 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001847
Russell Kinga6f36be2006-12-30 22:24:19 +00001848 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001849 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001851next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 if (vmas)
1853 vmas[i] = vma;
1854 i++;
1855 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001856 nr_pages--;
1857 } while (nr_pages && start < vma->vm_end);
1858 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 return i;
1860}
Huang Ying0014bd92011-01-30 11:15:47 +08001861EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001862
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001863/*
1864 * fixup_user_fault() - manually resolve a user page fault
1865 * @tsk: the task_struct to use for page fault accounting, or
1866 * NULL if faults are not to be recorded.
1867 * @mm: mm_struct of target mm
1868 * @address: user address
1869 * @fault_flags:flags to pass down to handle_mm_fault()
1870 *
1871 * This is meant to be called in the specific scenario where for locking reasons
1872 * we try to access user memory in atomic context (within a pagefault_disable()
1873 * section), this returns -EFAULT, and we want to resolve the user fault before
1874 * trying again.
1875 *
1876 * Typically this is meant to be used by the futex code.
1877 *
1878 * The main difference with get_user_pages() is that this function will
1879 * unconditionally call handle_mm_fault() which will in turn perform all the
1880 * necessary SW fixup of the dirty and young bits in the PTE, while
1881 * handle_mm_fault() only guarantees to update these in the struct page.
1882 *
1883 * This is important for some architectures where those bits also gate the
1884 * access permission to the page because they are maintained in software. On
1885 * such architectures, gup() will not be enough to make a subsequent access
1886 * succeed.
1887 *
1888 * This should be called with the mm_sem held for read.
1889 */
1890int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1891 unsigned long address, unsigned int fault_flags)
1892{
1893 struct vm_area_struct *vma;
1894 int ret;
1895
1896 vma = find_extend_vma(mm, address);
1897 if (!vma || address < vma->vm_start)
1898 return -EFAULT;
1899
1900 ret = handle_mm_fault(mm, vma, address, fault_flags);
1901 if (ret & VM_FAULT_ERROR) {
1902 if (ret & VM_FAULT_OOM)
1903 return -ENOMEM;
1904 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1905 return -EHWPOISON;
1906 if (ret & VM_FAULT_SIGBUS)
1907 return -EFAULT;
1908 BUG();
1909 }
1910 if (tsk) {
1911 if (ret & VM_FAULT_MAJOR)
1912 tsk->maj_flt++;
1913 else
1914 tsk->min_flt++;
1915 }
1916 return 0;
1917}
1918
1919/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001920 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001921 * @tsk: the task_struct to use for page fault accounting, or
1922 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001923 * @mm: mm_struct of target mm
1924 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001925 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001926 * @write: whether pages will be written to by the caller
1927 * @force: whether to force write access even if user mapping is
1928 * readonly. This will result in the page being COWed even
1929 * in MAP_SHARED mappings. You do not want this.
1930 * @pages: array that receives pointers to the pages pinned.
1931 * Should be at least nr_pages long. Or NULL, if caller
1932 * only intends to ensure the pages are faulted in.
1933 * @vmas: array of pointers to vmas corresponding to each page.
1934 * Or NULL if the caller does not require them.
1935 *
1936 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001937 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001938 * were pinned, returns -errno. Each page returned must be released
1939 * with a put_page() call when it is finished with. vmas will only
1940 * remain valid while mmap_sem is held.
1941 *
1942 * Must be called with mmap_sem held for read or write.
1943 *
1944 * get_user_pages walks a process's page tables and takes a reference to
1945 * each struct page that each user address corresponds to at a given
1946 * instant. That is, it takes the page that would be accessed if a user
1947 * thread accesses the given user virtual address at that instant.
1948 *
1949 * This does not guarantee that the page exists in the user mappings when
1950 * get_user_pages returns, and there may even be a completely different
1951 * page there in some cases (eg. if mmapped pagecache has been invalidated
1952 * and subsequently re faulted). However it does guarantee that the page
1953 * won't be freed completely. And mostly callers simply care that the page
1954 * contains data that was valid *at some point in time*. Typically, an IO
1955 * or similar operation cannot guarantee anything stronger anyway because
1956 * locks can't be held over the syscall boundary.
1957 *
1958 * If write=0, the page must not be written to. If the page is written to,
1959 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1960 * after the page is finished with, and before put_page is called.
1961 *
1962 * get_user_pages is typically used for fewer-copy IO operations, to get a
1963 * handle on the memory by some means other than accesses via the user virtual
1964 * addresses. The pages may be submitted for DMA to devices or accessed via
1965 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1966 * use the correct cache flushing APIs.
1967 *
1968 * See also get_user_pages_fast, for performance critical applications.
1969 */
Nick Pigginb291f002008-10-18 20:26:44 -07001970int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001971 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001972 struct page **pages, struct vm_area_struct **vmas)
1973{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001974 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001975
Hugh Dickins58fa8792009-09-21 17:03:31 -07001976 if (pages)
1977 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001978 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001979 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001980 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001981 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001982
Michel Lespinasse53a77062011-01-13 15:46:14 -08001983 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1984 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001985}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986EXPORT_SYMBOL(get_user_pages);
1987
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001988/**
1989 * get_dump_page() - pin user page in memory while writing it to core dump
1990 * @addr: user address
1991 *
1992 * Returns struct page pointer of user page pinned for dump,
1993 * to be freed afterwards by page_cache_release() or put_page().
1994 *
1995 * Returns NULL on any kind of failure - a hole must then be inserted into
1996 * the corefile, to preserve alignment with its headers; and also returns
1997 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1998 * allowing a hole to be left in the corefile to save diskspace.
1999 *
2000 * Called without mmap_sem, but after all other threads have been killed.
2001 */
2002#ifdef CONFIG_ELF_CORE
2003struct page *get_dump_page(unsigned long addr)
2004{
2005 struct vm_area_struct *vma;
2006 struct page *page;
2007
2008 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002009 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2010 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002011 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002012 flush_cache_page(vma, addr, page_to_pfn(page));
2013 return page;
2014}
2015#endif /* CONFIG_ELF_CORE */
2016
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002017pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002018 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002019{
2020 pgd_t * pgd = pgd_offset(mm, addr);
2021 pud_t * pud = pud_alloc(mm, pgd, addr);
2022 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002023 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002024 if (pmd) {
2025 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002026 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002027 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002028 }
2029 return NULL;
2030}
2031
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002033 * This is the old fallback for page remapping.
2034 *
2035 * For historical reasons, it only allows reserved pages. Only
2036 * old drivers should use this, and they needed to mark their
2037 * pages reserved for the old functions anyway.
2038 */
Nick Piggin423bad602008-04-28 02:13:01 -07002039static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2040 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002041{
Nick Piggin423bad602008-04-28 02:13:01 -07002042 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002043 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002044 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002045 spinlock_t *ptl;
2046
Linus Torvalds238f58d2005-11-29 13:01:56 -08002047 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002048 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002049 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002050 retval = -ENOMEM;
2051 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002052 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002053 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002054 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002055 retval = -EBUSY;
2056 if (!pte_none(*pte))
2057 goto out_unlock;
2058
2059 /* Ok, finally just insert the thing.. */
2060 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002061 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002062 page_add_file_rmap(page);
2063 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2064
2065 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002066 pte_unmap_unlock(pte, ptl);
2067 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002068out_unlock:
2069 pte_unmap_unlock(pte, ptl);
2070out:
2071 return retval;
2072}
2073
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002074/**
2075 * vm_insert_page - insert single page into user vma
2076 * @vma: user vma to map to
2077 * @addr: target user address of this page
2078 * @page: source kernel page
2079 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002080 * This allows drivers to insert individual pages they've allocated
2081 * into a user vma.
2082 *
2083 * The page has to be a nice clean _individual_ kernel allocation.
2084 * If you allocate a compound page, you need to have marked it as
2085 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002086 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002087 *
2088 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2089 * took an arbitrary page protection parameter. This doesn't allow
2090 * that. Your vma protection will have to be set up correctly, which
2091 * means that if you want a shared writable mapping, you'd better
2092 * ask for a shared writable mapping!
2093 *
2094 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002095 *
2096 * Usually this function is called from f_op->mmap() handler
2097 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2098 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2099 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002100 */
Nick Piggin423bad602008-04-28 02:13:01 -07002101int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2102 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002103{
2104 if (addr < vma->vm_start || addr >= vma->vm_end)
2105 return -EFAULT;
2106 if (!page_count(page))
2107 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002108 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2109 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2110 BUG_ON(vma->vm_flags & VM_PFNMAP);
2111 vma->vm_flags |= VM_MIXEDMAP;
2112 }
Nick Piggin423bad602008-04-28 02:13:01 -07002113 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002114}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002115EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002116
Nick Piggin423bad602008-04-28 02:13:01 -07002117static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2118 unsigned long pfn, pgprot_t prot)
2119{
2120 struct mm_struct *mm = vma->vm_mm;
2121 int retval;
2122 pte_t *pte, entry;
2123 spinlock_t *ptl;
2124
2125 retval = -ENOMEM;
2126 pte = get_locked_pte(mm, addr, &ptl);
2127 if (!pte)
2128 goto out;
2129 retval = -EBUSY;
2130 if (!pte_none(*pte))
2131 goto out_unlock;
2132
2133 /* Ok, finally just insert the thing.. */
2134 entry = pte_mkspecial(pfn_pte(pfn, prot));
2135 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002136 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad602008-04-28 02:13:01 -07002137
2138 retval = 0;
2139out_unlock:
2140 pte_unmap_unlock(pte, ptl);
2141out:
2142 return retval;
2143}
2144
Nick Piggine0dc0d82007-02-12 00:51:36 -08002145/**
2146 * vm_insert_pfn - insert single pfn into user vma
2147 * @vma: user vma to map to
2148 * @addr: target user address of this page
2149 * @pfn: source kernel pfn
2150 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002151 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002152 * they've allocated into a user vma. Same comments apply.
2153 *
2154 * This function should only be called from a vm_ops->fault handler, and
2155 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002156 *
2157 * vma cannot be a COW mapping.
2158 *
2159 * As this is called only for pages that do not currently exist, we
2160 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002161 */
2162int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad602008-04-28 02:13:01 -07002163 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002164{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002165 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002166 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002167 /*
2168 * Technically, architectures with pte_special can avoid all these
2169 * restrictions (same for remap_pfn_range). However we would like
2170 * consistency in testing and feature parity among all, so we should
2171 * try to keep these invariants in place for everybody.
2172 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002173 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2174 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2175 (VM_PFNMAP|VM_MIXEDMAP));
2176 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2177 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002178
Nick Piggin423bad602008-04-28 02:13:01 -07002179 if (addr < vma->vm_start || addr >= vma->vm_end)
2180 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002181 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002182 return -EINVAL;
2183
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002184 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002185
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002186 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002187}
2188EXPORT_SYMBOL(vm_insert_pfn);
2189
Nick Piggin423bad602008-04-28 02:13:01 -07002190int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2191 unsigned long pfn)
2192{
2193 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2194
2195 if (addr < vma->vm_start || addr >= vma->vm_end)
2196 return -EFAULT;
2197
2198 /*
2199 * If we don't have pte special, then we have to use the pfn_valid()
2200 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2201 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002202 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2203 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad602008-04-28 02:13:01 -07002204 */
2205 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2206 struct page *page;
2207
2208 page = pfn_to_page(pfn);
2209 return insert_page(vma, addr, page, vma->vm_page_prot);
2210 }
2211 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2212}
2213EXPORT_SYMBOL(vm_insert_mixed);
2214
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002215/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * maps a range of physical memory into the requested pages. the old
2217 * mappings are removed. any references to nonexistent pages results
2218 * in null mappings (currently treated as "copy-on-access")
2219 */
2220static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2221 unsigned long addr, unsigned long end,
2222 unsigned long pfn, pgprot_t prot)
2223{
2224 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002225 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226
Hugh Dickinsc74df322005-10-29 18:16:23 -07002227 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 if (!pte)
2229 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002230 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 do {
2232 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002233 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 pfn++;
2235 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002236 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002237 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 return 0;
2239}
2240
2241static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2242 unsigned long addr, unsigned long end,
2243 unsigned long pfn, pgprot_t prot)
2244{
2245 pmd_t *pmd;
2246 unsigned long next;
2247
2248 pfn -= addr >> PAGE_SHIFT;
2249 pmd = pmd_alloc(mm, pud, addr);
2250 if (!pmd)
2251 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002252 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 do {
2254 next = pmd_addr_end(addr, end);
2255 if (remap_pte_range(mm, pmd, addr, next,
2256 pfn + (addr >> PAGE_SHIFT), prot))
2257 return -ENOMEM;
2258 } while (pmd++, addr = next, addr != end);
2259 return 0;
2260}
2261
2262static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2263 unsigned long addr, unsigned long end,
2264 unsigned long pfn, pgprot_t prot)
2265{
2266 pud_t *pud;
2267 unsigned long next;
2268
2269 pfn -= addr >> PAGE_SHIFT;
2270 pud = pud_alloc(mm, pgd, addr);
2271 if (!pud)
2272 return -ENOMEM;
2273 do {
2274 next = pud_addr_end(addr, end);
2275 if (remap_pmd_range(mm, pud, addr, next,
2276 pfn + (addr >> PAGE_SHIFT), prot))
2277 return -ENOMEM;
2278 } while (pud++, addr = next, addr != end);
2279 return 0;
2280}
2281
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002282/**
2283 * remap_pfn_range - remap kernel memory to userspace
2284 * @vma: user vma to map to
2285 * @addr: target user address to start at
2286 * @pfn: physical address of kernel memory
2287 * @size: size of map area
2288 * @prot: page protection flags for this mapping
2289 *
2290 * Note: this is only safe if the mm semaphore is held when called.
2291 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2293 unsigned long pfn, unsigned long size, pgprot_t prot)
2294{
2295 pgd_t *pgd;
2296 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002297 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 struct mm_struct *mm = vma->vm_mm;
2299 int err;
2300
2301 /*
2302 * Physically remapped pages are special. Tell the
2303 * rest of the world about it:
2304 * VM_IO tells people not to look at these pages
2305 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002306 * VM_PFNMAP tells the core MM that the base pages are just
2307 * raw PFN mappings, and do not have a "struct page" associated
2308 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002309 * VM_DONTEXPAND
2310 * Disable vma merging and expanding with mremap().
2311 * VM_DONTDUMP
2312 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002313 *
2314 * There's a horrible special case to handle copy-on-write
2315 * behaviour that some programs depend on. We mark the "original"
2316 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002317 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002319 if (is_cow_mapping(vma->vm_flags)) {
2320 if (addr != vma->vm_start || end != vma->vm_end)
2321 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002322 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002323 }
2324
2325 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2326 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002327 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002328
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002329 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
2331 BUG_ON(addr >= end);
2332 pfn -= addr >> PAGE_SHIFT;
2333 pgd = pgd_offset(mm, addr);
2334 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 do {
2336 next = pgd_addr_end(addr, end);
2337 err = remap_pud_range(mm, pgd, addr, next,
2338 pfn + (addr >> PAGE_SHIFT), prot);
2339 if (err)
2340 break;
2341 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002342
2343 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002344 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002345
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 return err;
2347}
2348EXPORT_SYMBOL(remap_pfn_range);
2349
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002350static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2351 unsigned long addr, unsigned long end,
2352 pte_fn_t fn, void *data)
2353{
2354 pte_t *pte;
2355 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002356 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002357 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002358
2359 pte = (mm == &init_mm) ?
2360 pte_alloc_kernel(pmd, addr) :
2361 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2362 if (!pte)
2363 return -ENOMEM;
2364
2365 BUG_ON(pmd_huge(*pmd));
2366
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002367 arch_enter_lazy_mmu_mode();
2368
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002369 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002370
2371 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002372 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002373 if (err)
2374 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002375 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002376
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002377 arch_leave_lazy_mmu_mode();
2378
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002379 if (mm != &init_mm)
2380 pte_unmap_unlock(pte-1, ptl);
2381 return err;
2382}
2383
2384static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2385 unsigned long addr, unsigned long end,
2386 pte_fn_t fn, void *data)
2387{
2388 pmd_t *pmd;
2389 unsigned long next;
2390 int err;
2391
Andi Kleenceb86872008-07-23 21:27:50 -07002392 BUG_ON(pud_huge(*pud));
2393
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002394 pmd = pmd_alloc(mm, pud, addr);
2395 if (!pmd)
2396 return -ENOMEM;
2397 do {
2398 next = pmd_addr_end(addr, end);
2399 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2400 if (err)
2401 break;
2402 } while (pmd++, addr = next, addr != end);
2403 return err;
2404}
2405
2406static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2407 unsigned long addr, unsigned long end,
2408 pte_fn_t fn, void *data)
2409{
2410 pud_t *pud;
2411 unsigned long next;
2412 int err;
2413
2414 pud = pud_alloc(mm, pgd, addr);
2415 if (!pud)
2416 return -ENOMEM;
2417 do {
2418 next = pud_addr_end(addr, end);
2419 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2420 if (err)
2421 break;
2422 } while (pud++, addr = next, addr != end);
2423 return err;
2424}
2425
2426/*
2427 * Scan a region of virtual memory, filling in page tables as necessary
2428 * and calling a provided function on each leaf page table.
2429 */
2430int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2431 unsigned long size, pte_fn_t fn, void *data)
2432{
2433 pgd_t *pgd;
2434 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002435 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002436 int err;
2437
2438 BUG_ON(addr >= end);
2439 pgd = pgd_offset(mm, addr);
2440 do {
2441 next = pgd_addr_end(addr, end);
2442 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2443 if (err)
2444 break;
2445 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002446
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002447 return err;
2448}
2449EXPORT_SYMBOL_GPL(apply_to_page_range);
2450
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002452 * handle_pte_fault chooses page fault handler according to an entry
2453 * which was read non-atomically. Before making any commitment, on
2454 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002455 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002456 * must check under lock before unmapping the pte and proceeding
2457 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002458 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002459 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002460static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002461 pte_t *page_table, pte_t orig_pte)
2462{
2463 int same = 1;
2464#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2465 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002466 spinlock_t *ptl = pte_lockptr(mm, pmd);
2467 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002468 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002469 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002470 }
2471#endif
2472 pte_unmap(page_table);
2473 return same;
2474}
2475
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002476static 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 -08002477{
2478 /*
2479 * If the source page was a PFN mapping, we don't have
2480 * a "struct page" for it. We do a best-effort copy by
2481 * just copying from the original user address. If that
2482 * fails, we just zero-fill it. Live with it.
2483 */
2484 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002485 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002486 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2487
2488 /*
2489 * This really shouldn't fail, because the page is there
2490 * in the page tables. But it might just be unreadable,
2491 * in which case we just give up and fill the result with
2492 * zeroes.
2493 */
2494 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002495 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002496 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002497 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002498 } else
2499 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002500}
2501
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 * This routine handles present pages, when users try to write
2504 * to a shared page. It is done by copying the page to a new address
2505 * and decrementing the shared-page counter for the old page.
2506 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 * Note that this routine assumes that the protection checks have been
2508 * done by the caller (the low-level page fault routine in most cases).
2509 * Thus we can safely just mark it writable once we've done any necessary
2510 * COW.
2511 *
2512 * We also mark the page dirty at this point even though the page will
2513 * change only once the write actually happens. This avoids a few races,
2514 * and potentially makes it more efficient.
2515 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002516 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2517 * but allow concurrent faults), with pte both mapped and locked.
2518 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002520static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2521 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002522 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002523 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002525 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002527 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002528 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002529 struct page *dirty_page = NULL;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002530 unsigned long mmun_start; /* For mmu_notifiers */
2531 unsigned long mmun_end; /* For mmu_notifiers */
2532 bool mmun_called = false; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
Linus Torvalds6aab3412005-11-28 14:34:23 -08002534 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002535 if (!old_page) {
2536 /*
2537 * VM_MIXEDMAP !pfn_valid() case
2538 *
2539 * We should not cow pages in a shared writeable mapping.
2540 * Just mark the pages writable as we can't do any dirty
2541 * accounting on raw pfn maps.
2542 */
2543 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2544 (VM_WRITE|VM_SHARED))
2545 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002546 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002547 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002549 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002550 * Take out anonymous pages first, anonymous shared vmas are
2551 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002552 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002553 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002554 if (!trylock_page(old_page)) {
2555 page_cache_get(old_page);
2556 pte_unmap_unlock(page_table, ptl);
2557 lock_page(old_page);
2558 page_table = pte_offset_map_lock(mm, pmd, address,
2559 &ptl);
2560 if (!pte_same(*page_table, orig_pte)) {
2561 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002562 goto unlock;
2563 }
2564 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002565 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002566 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002567 /*
2568 * The page is all ours. Move it to our anon_vma so
2569 * the rmap code will not search our parent or siblings.
2570 * Protected against the rmap code by the page lock.
2571 */
2572 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002573 unlock_page(old_page);
2574 goto reuse;
2575 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002576 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002577 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002578 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002579 /*
2580 * Only catch write-faults on shared writable pages,
2581 * read-only shared pages can get COWed by
2582 * get_user_pages(.write=1, .force=1).
2583 */
David Howells9637a5e2006-06-23 02:03:43 -07002584 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002585 struct vm_fault vmf;
2586 int tmp;
2587
2588 vmf.virtual_address = (void __user *)(address &
2589 PAGE_MASK);
2590 vmf.pgoff = old_page->index;
2591 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2592 vmf.page = old_page;
2593
David Howells9637a5e2006-06-23 02:03:43 -07002594 /*
2595 * Notify the address space that the page is about to
2596 * become writable so that it can prohibit this or wait
2597 * for the page to get into an appropriate state.
2598 *
2599 * We do this without the lock held, so that it can
2600 * sleep if it needs to.
2601 */
2602 page_cache_get(old_page);
2603 pte_unmap_unlock(page_table, ptl);
2604
Nick Pigginc2ec1752009-03-31 15:23:21 -07002605 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2606 if (unlikely(tmp &
2607 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2608 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002609 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002610 }
Nick Pigginb827e492009-04-30 15:08:16 -07002611 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2612 lock_page(old_page);
2613 if (!old_page->mapping) {
2614 ret = 0; /* retry the fault */
2615 unlock_page(old_page);
2616 goto unwritable_page;
2617 }
2618 } else
2619 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002620
David Howells9637a5e2006-06-23 02:03:43 -07002621 /*
2622 * Since we dropped the lock we need to revalidate
2623 * the PTE as someone else may have changed it. If
2624 * they did, we just return, as we can count on the
2625 * MMU to tell us if they didn't also make it writable.
2626 */
2627 page_table = pte_offset_map_lock(mm, pmd, address,
2628 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002629 if (!pte_same(*page_table, orig_pte)) {
2630 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002631 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002632 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002633
2634 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002636 dirty_page = old_page;
2637 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002638
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002639reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002640 flush_cache_page(vma, address, pte_pfn(orig_pte));
2641 entry = pte_mkyoung(orig_pte);
2642 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002643 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002644 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002645 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002646 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002647
2648 if (!dirty_page)
2649 return ret;
2650
2651 /*
2652 * Yes, Virginia, this is actually required to prevent a race
2653 * with clear_page_dirty_for_io() from clearing the page dirty
2654 * bit after it clear all dirty ptes, but before a racing
2655 * do_wp_page installs a dirty pte.
2656 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002657 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002658 */
2659 if (!page_mkwrite) {
2660 wait_on_page_locked(dirty_page);
2661 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002662 /* file_update_time outside page_lock */
2663 if (vma->vm_file)
2664 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002665 }
2666 put_page(dirty_page);
2667 if (page_mkwrite) {
2668 struct address_space *mapping = dirty_page->mapping;
2669
2670 set_page_dirty(dirty_page);
2671 unlock_page(dirty_page);
2672 page_cache_release(dirty_page);
2673 if (mapping) {
2674 /*
2675 * Some device drivers do not set page.mapping
2676 * but still dirty their pages
2677 */
2678 balance_dirty_pages_ratelimited(mapping);
2679 }
2680 }
2681
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002682 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684
2685 /*
2686 * Ok, we need to copy. Oh, well..
2687 */
Nick Pigginb5810032005-10-29 18:16:12 -07002688 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002689gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002690 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691
2692 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002693 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002694
Hugh Dickins62eede62009-09-21 17:03:34 -07002695 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002696 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2697 if (!new_page)
2698 goto oom;
2699 } else {
2700 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2701 if (!new_page)
2702 goto oom;
2703 cow_user_page(new_page, old_page, address, vma);
2704 }
2705 __SetPageUptodate(new_page);
2706
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002707 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002708 goto oom_free_new;
2709
Haggai Eran6bdb9132012-10-08 16:33:35 -07002710 mmun_start = address & PAGE_MASK;
2711 mmun_end = (address & PAGE_MASK) + PAGE_SIZE;
2712 mmun_called = true;
2713 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2714
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 /*
2716 * Re-check the pte - we dropped the lock
2717 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002718 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002719 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002720 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002721 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002722 dec_mm_counter_fast(mm, MM_FILEPAGES);
2723 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002724 }
2725 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002726 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002727 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002728 entry = mk_pte(new_page, vma->vm_page_prot);
2729 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002730 /*
2731 * Clear the pte entry and flush it first, before updating the
2732 * pte with the new entry. This will avoid a race condition
2733 * seen in the presence of one thread doing SMC and another
2734 * thread doing COW.
2735 */
Izik Eidus828502d2009-09-21 17:01:51 -07002736 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002737 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002738 /*
2739 * We call the notify macro here because, when using secondary
2740 * mmu page tables (such as kvm shadow page tables), we want the
2741 * new page to be mapped directly into the secondary page table.
2742 */
2743 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002744 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002745 if (old_page) {
2746 /*
2747 * Only after switching the pte to the new page may
2748 * we remove the mapcount here. Otherwise another
2749 * process may come and find the rmap count decremented
2750 * before the pte is switched to the new page, and
2751 * "reuse" the old page writing into it while our pte
2752 * here still points into it and can be read by other
2753 * threads.
2754 *
2755 * The critical issue is to order this
2756 * page_remove_rmap with the ptp_clear_flush above.
2757 * Those stores are ordered by (if nothing else,)
2758 * the barrier present in the atomic_add_negative
2759 * in page_remove_rmap.
2760 *
2761 * Then the TLB flush in ptep_clear_flush ensures that
2762 * no process can access the old page before the
2763 * decremented mapcount is visible. And the old page
2764 * cannot be reused until after the decremented
2765 * mapcount is visible. So transitively, TLBs to
2766 * old page will be flushed before it can be reused.
2767 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002768 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002769 }
2770
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 /* Free the old page.. */
2772 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002773 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002774 } else
2775 mem_cgroup_uncharge_page(new_page);
2776
Haggai Eran6bdb9132012-10-08 16:33:35 -07002777 if (new_page)
2778 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002779unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002780 pte_unmap_unlock(page_table, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -07002781 if (mmun_called)
2782 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002783 if (old_page) {
2784 /*
2785 * Don't let another task, with possibly unlocked vma,
2786 * keep the mlocked page.
2787 */
2788 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2789 lock_page(old_page); /* LRU manipulation */
2790 munlock_vma_page(old_page);
2791 unlock_page(old_page);
2792 }
2793 page_cache_release(old_page);
2794 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002795 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002796oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002797 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002798oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002799 if (old_page) {
2800 if (page_mkwrite) {
2801 unlock_page(old_page);
2802 page_cache_release(old_page);
2803 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002804 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002805 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002807
2808unwritable_page:
2809 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002810 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811}
2812
Peter Zijlstra97a89412011-05-24 17:12:04 -07002813static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 unsigned long start_addr, unsigned long end_addr,
2815 struct zap_details *details)
2816{
Al Virof5cc4ee2012-03-05 14:14:20 -05002817 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818}
2819
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002820static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 struct zap_details *details)
2822{
2823 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 pgoff_t vba, vea, zba, zea;
2825
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002826 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
2829 vba = vma->vm_pgoff;
2830 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2831 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2832 zba = details->first_index;
2833 if (zba < vba)
2834 zba = vba;
2835 zea = details->last_index;
2836 if (zea > vea)
2837 zea = vea;
2838
Peter Zijlstra97a89412011-05-24 17:12:04 -07002839 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2841 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002842 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 }
2844}
2845
2846static inline void unmap_mapping_range_list(struct list_head *head,
2847 struct zap_details *details)
2848{
2849 struct vm_area_struct *vma;
2850
2851 /*
2852 * In nonlinear VMAs there is no correspondence between virtual address
2853 * offset and file offset. So we must perform an exhaustive search
2854 * across *all* the pages in each nonlinear VMA, not just the pages
2855 * whose virtual address lies outside the file truncation point.
2856 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002857 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002859 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 }
2861}
2862
2863/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002864 * 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 -07002865 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 * @holebegin: byte in first page to unmap, relative to the start of
2867 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002868 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 * must keep the partial page. In contrast, we must get rid of
2870 * partial pages.
2871 * @holelen: size of prospective hole in bytes. This will be rounded
2872 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2873 * end of the file.
2874 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2875 * but 0 when invalidating pagecache, don't throw away private data.
2876 */
2877void unmap_mapping_range(struct address_space *mapping,
2878 loff_t const holebegin, loff_t const holelen, int even_cows)
2879{
2880 struct zap_details details;
2881 pgoff_t hba = holebegin >> PAGE_SHIFT;
2882 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2883
2884 /* Check for overflow. */
2885 if (sizeof(holelen) > sizeof(hlen)) {
2886 long long holeend =
2887 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2888 if (holeend & ~(long long)ULONG_MAX)
2889 hlen = ULONG_MAX - hba + 1;
2890 }
2891
2892 details.check_mapping = even_cows? NULL: mapping;
2893 details.nonlinear_vma = NULL;
2894 details.first_index = hba;
2895 details.last_index = hba + hlen - 1;
2896 if (details.last_index < details.first_index)
2897 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898
Peter Zijlstra97a89412011-05-24 17:12:04 -07002899
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002900 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002901 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2903 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2904 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002905 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906}
2907EXPORT_SYMBOL(unmap_mapping_range);
2908
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002910 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2911 * but allow concurrent faults), and pte mapped but not yet locked.
2912 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002914static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2915 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002916 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002918 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002919 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002920 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002922 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002923 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002924 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002925 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002927 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002928 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002929
2930 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002931 if (unlikely(non_swap_entry(entry))) {
2932 if (is_migration_entry(entry)) {
2933 migration_entry_wait(mm, pmd, address);
2934 } else if (is_hwpoison_entry(entry)) {
2935 ret = VM_FAULT_HWPOISON;
2936 } else {
2937 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002938 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002939 }
Christoph Lameter06972122006-06-23 02:03:35 -07002940 goto out;
2941 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002942 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 page = lookup_swap_cache(entry);
2944 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002945 page = swapin_readahead(entry,
2946 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 if (!page) {
2948 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002949 * Back out if somebody else faulted in this pte
2950 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002952 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 if (likely(pte_same(*page_table, orig_pte)))
2954 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002955 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002956 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 }
2958
2959 /* Had to read the page from swap area: Major fault */
2960 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002961 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002962 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002963 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002964 /*
2965 * hwpoisoned dirty swapcache pages are kept for killing
2966 * owner processes (which may be unknown at hwpoison time)
2967 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002968 ret = VM_FAULT_HWPOISON;
2969 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002970 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 }
2972
Michel Lespinassed065bd82010-10-26 14:21:57 -07002973 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002974
Balbir Singh20a10222007-11-14 17:00:33 -08002975 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002976 if (!locked) {
2977 ret |= VM_FAULT_RETRY;
2978 goto out_release;
2979 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002981 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002982 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2983 * release the swapcache from under us. The page pin, and pte_same
2984 * test below, are not enough to exclude that. Even if it is still
2985 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002986 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002987 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002988 goto out_page;
2989
2990 if (ksm_might_need_to_copy(page, vma, address)) {
2991 swapcache = page;
2992 page = ksm_does_need_to_copy(page, vma, address);
2993
2994 if (unlikely(!page)) {
2995 ret = VM_FAULT_OOM;
2996 page = swapcache;
2997 swapcache = NULL;
2998 goto out_page;
2999 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08003000 }
3001
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003002 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003003 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003004 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003005 }
3006
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003008 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003010 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003011 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003012 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003013
3014 if (unlikely(!PageUptodate(page))) {
3015 ret = VM_FAULT_SIGBUS;
3016 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 }
3018
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003019 /*
3020 * The page isn't present yet, go ahead with the fault.
3021 *
3022 * Be careful about the sequence of operations here.
3023 * To get its accounting right, reuse_swap_page() must be called
3024 * while the page is counted on swap but not yet in mapcount i.e.
3025 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3026 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003027 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3028 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3029 * in page->private. In this case, a record in swap_cgroup is silently
3030 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003031 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003033 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003034 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003036 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003038 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003039 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003040 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042 flush_icache_page(vma, page);
3043 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003044 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003045 /* It's better to call commit-charge after rmap is established */
3046 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003048 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003049 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003050 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003051 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003052 if (swapcache) {
3053 /*
3054 * Hold the lock to avoid the swap entry to be reused
3055 * until we take the PT lock for the pte_same() check
3056 * (to avoid false positives from pte_same). For
3057 * further safety release the lock after the swap_free
3058 * so that the swap count won't change under a
3059 * parallel locked swapcache.
3060 */
3061 unlock_page(swapcache);
3062 page_cache_release(swapcache);
3063 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003064
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003065 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003066 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3067 if (ret & VM_FAULT_ERROR)
3068 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069 goto out;
3070 }
3071
3072 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003073 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003074unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003075 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076out:
3077 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003078out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003079 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003080 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003081out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003082 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003083out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003084 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003085 if (swapcache) {
3086 unlock_page(swapcache);
3087 page_cache_release(swapcache);
3088 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003089 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090}
3091
3092/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003093 * This is like a special single-page "expand_{down|up}wards()",
3094 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003095 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003096 */
3097static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3098{
3099 address &= PAGE_MASK;
3100 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003101 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003102
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003103 /*
3104 * Is there a mapping abutting this one below?
3105 *
3106 * That's only ok if it's the same stack mapping
3107 * that has gotten split..
3108 */
3109 if (prev && prev->vm_end == address)
3110 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3111
Michal Hockod05f3162011-05-24 17:11:44 -07003112 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003113 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003114 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3115 struct vm_area_struct *next = vma->vm_next;
3116
3117 /* As VM_GROWSDOWN but s/below/above/ */
3118 if (next && next->vm_start == address + PAGE_SIZE)
3119 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3120
3121 expand_upwards(vma, address + PAGE_SIZE);
3122 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003123 return 0;
3124}
3125
3126/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003127 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3128 * but allow concurrent faults), and pte mapped but not yet locked.
3129 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003131static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3132 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003133 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003135 struct page *page;
3136 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
Linus Torvalds11ac5522010-08-14 11:44:56 -07003139 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003140
Linus Torvalds11ac5522010-08-14 11:44:56 -07003141 /* Check if we need to add a guard page to the stack */
3142 if (check_stack_guard_page(vma, address) < 0)
3143 return VM_FAULT_SIGBUS;
3144
3145 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003146 if (!(flags & FAULT_FLAG_WRITE)) {
3147 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3148 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003149 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003150 if (!pte_none(*page_table))
3151 goto unlock;
3152 goto setpte;
3153 }
3154
Nick Piggin557ed1f2007-10-16 01:24:40 -07003155 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003156 if (unlikely(anon_vma_prepare(vma)))
3157 goto oom;
3158 page = alloc_zeroed_user_highpage_movable(vma, address);
3159 if (!page)
3160 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003161 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003163 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003164 goto oom_free_page;
3165
Nick Piggin557ed1f2007-10-16 01:24:40 -07003166 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003167 if (vma->vm_flags & VM_WRITE)
3168 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003169
Nick Piggin557ed1f2007-10-16 01:24:40 -07003170 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003171 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003172 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003173
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003174 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003175 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003176setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003177 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178
3179 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003180 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003181unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003182 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003183 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003184release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003185 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003186 page_cache_release(page);
3187 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003188oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003189 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003190oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191 return VM_FAULT_OOM;
3192}
3193
3194/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003195 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003197 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3198 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 *
3200 * As this is called only for pages that do not currently exist, we
3201 * do not need to flush old virtual caches or the TLB.
3202 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003203 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003204 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003205 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003207static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003208 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003209 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003211 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003212 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003213 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003214 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003217 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003218 struct vm_fault vmf;
3219 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003220 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003221
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003222 /*
3223 * If we do COW later, allocate page befor taking lock_page()
3224 * on the file cache page. This will reduce lock holding time.
3225 */
3226 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3227
3228 if (unlikely(anon_vma_prepare(vma)))
3229 return VM_FAULT_OOM;
3230
3231 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3232 if (!cow_page)
3233 return VM_FAULT_OOM;
3234
3235 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3236 page_cache_release(cow_page);
3237 return VM_FAULT_OOM;
3238 }
3239 } else
3240 cow_page = NULL;
3241
Nick Piggind0217ac2007-07-19 01:47:03 -07003242 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3243 vmf.pgoff = pgoff;
3244 vmf.flags = flags;
3245 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003247 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003248 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3249 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003250 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251
Andi Kleena3b947e2009-09-16 11:50:08 +02003252 if (unlikely(PageHWPoison(vmf.page))) {
3253 if (ret & VM_FAULT_LOCKED)
3254 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003255 ret = VM_FAULT_HWPOISON;
3256 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003257 }
3258
Nick Piggind00806b2007-07-19 01:46:57 -07003259 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003260 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003261 * locked.
3262 */
Nick Piggin83c54072007-07-19 01:47:05 -07003263 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003264 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003265 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003266 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003267
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 /*
3269 * Should we do an early C-O-W break?
3270 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003271 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003272 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003273 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003274 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003275 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003276 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003277 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003278 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003279 /*
3280 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003281 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003282 * to become writable
3283 */
Mark Fasheh69676142007-07-19 01:47:00 -07003284 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003285 int tmp;
3286
Mark Fasheh69676142007-07-19 01:47:00 -07003287 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003288 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003289 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3290 if (unlikely(tmp &
3291 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3292 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003293 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003294 }
Nick Pigginb827e492009-04-30 15:08:16 -07003295 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3296 lock_page(page);
3297 if (!page->mapping) {
3298 ret = 0; /* retry the fault */
3299 unlock_page(page);
3300 goto unwritable_page;
3301 }
3302 } else
3303 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003304 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003305 }
3306 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003307
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 }
3309
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003310 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311
3312 /*
3313 * This silly early PAGE_DIRTY setting removes a race
3314 * due to the bad i386 page protection. But it's valid
3315 * for other architectures too.
3316 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003317 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 * an exclusive copy of the page, or this is a shared mapping,
3319 * so we can make it writable and dirty to avoid having to
3320 * handle that later.
3321 */
3322 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003323 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003324 flush_icache_page(vma, page);
3325 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003326 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003329 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003330 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003331 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003332 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003333 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003334 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003335 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003336 get_page(dirty_page);
3337 }
Hugh Dickins42946212005-10-29 18:16:05 -07003338 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003339 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003340
3341 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003342 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003344 if (cow_page)
3345 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003346 if (anon)
3347 page_cache_release(page);
3348 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003349 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350 }
3351
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003352 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003353
Nick Pigginb827e492009-04-30 15:08:16 -07003354 if (dirty_page) {
3355 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003356 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003357
3358 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003359 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003360 unlock_page(dirty_page);
3361 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003362 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003363 /*
3364 * Some device drivers do not set page.mapping but still
3365 * dirty their pages
3366 */
3367 balance_dirty_pages_ratelimited(mapping);
3368 }
3369
3370 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003371 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003372 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003373 } else {
3374 unlock_page(vmf.page);
3375 if (anon)
3376 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003377 }
Nick Piggind00806b2007-07-19 01:46:57 -07003378
Nick Piggin83c54072007-07-19 01:47:05 -07003379 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003380
3381unwritable_page:
3382 page_cache_release(page);
3383 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003384uncharge_out:
3385 /* fs's fault handler get error */
3386 if (cow_page) {
3387 mem_cgroup_uncharge_page(cow_page);
3388 page_cache_release(cow_page);
3389 }
3390 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003391}
Nick Piggind00806b2007-07-19 01:46:57 -07003392
Nick Piggin54cb8822007-07-19 01:46:59 -07003393static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3394 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003395 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003396{
3397 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003398 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003399
Hugh Dickins16abfa02007-10-04 16:56:06 +01003400 pte_unmap(page_table);
3401 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003402}
3403
Jes Sorensenf4b81802006-09-27 01:50:10 -07003404/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 * Fault of a previously existing named mapping. Repopulate the pte
3406 * from the encoded file_pte if possible. This enables swappable
3407 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003408 *
3409 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3410 * but allow concurrent faults), and pte mapped but not yet locked.
3411 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003413static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003414 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003415 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416{
Hugh Dickins65500d22005-10-29 18:15:59 -07003417 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003419 flags |= FAULT_FLAG_NONLINEAR;
3420
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003421 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003422 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
Hugh Dickins2509ef22009-01-06 14:40:10 -08003424 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003425 /*
3426 * Page table corrupted: show pte and kill process.
3427 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003428 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003429 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003430 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003431
3432 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003433 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434}
3435
3436/*
3437 * These routines also need to handle stuff like marking pages dirty
3438 * and/or accessed for architectures that don't do it in hardware (most
3439 * RISC architectures). The early dirtying is also good on the i386.
3440 *
3441 * There is also a hook called "update_mmu_cache()" that architectures
3442 * with external mmu caches can use to update those (ie the Sparc or
3443 * PowerPC hashed page tables that act as extended TLBs).
3444 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003445 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3446 * but allow concurrent faults), and pte mapped but not yet locked.
3447 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003449int handle_pte_fault(struct mm_struct *mm,
3450 struct vm_area_struct *vma, unsigned long address,
3451 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452{
3453 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003454 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003456 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003458 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003459 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003460 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003461 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003462 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003463 }
3464 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003465 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003466 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003468 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003469 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003470 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003471 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 }
3473
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003474 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003475 spin_lock(ptl);
3476 if (unlikely(!pte_same(*pte, entry)))
3477 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003478 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003480 return do_wp_page(mm, vma, address,
3481 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 entry = pte_mkdirty(entry);
3483 }
3484 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003485 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003486 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003487 } else {
3488 /*
3489 * This is needed only for protection faults but the arch code
3490 * is not yet telling us if this is a protection fault or not.
3491 * This still avoids useless tlb flushes for .text page faults
3492 * with threads.
3493 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003494 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003495 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003496 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003497unlock:
3498 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003499 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500}
3501
3502/*
3503 * By the time we get here, we already hold the mm semaphore
3504 */
Nick Piggin83c54072007-07-19 01:47:05 -07003505int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003506 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507{
3508 pgd_t *pgd;
3509 pud_t *pud;
3510 pmd_t *pmd;
3511 pte_t *pte;
3512
3513 __set_current_state(TASK_RUNNING);
3514
Christoph Lameterf8891e52006-06-30 01:55:45 -07003515 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003516 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003518 /* do counter updates before entering really critical section. */
3519 check_sync_rss_stat(current);
3520
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003521 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003522 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523
David Rientjes1f1d06c2012-05-29 15:06:23 -07003524retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 pud = pud_alloc(mm, pgd, address);
3527 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003528 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 pmd = pmd_alloc(mm, pud, address);
3530 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003531 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003532 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3533 if (!vma->vm_ops)
3534 return do_huge_pmd_anonymous_page(mm, vma, address,
3535 pmd, flags);
3536 } else {
3537 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003538 int ret;
3539
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003540 barrier();
3541 if (pmd_trans_huge(orig_pmd)) {
3542 if (flags & FAULT_FLAG_WRITE &&
3543 !pmd_write(orig_pmd) &&
David Rientjes1f1d06c2012-05-29 15:06:23 -07003544 !pmd_trans_splitting(orig_pmd)) {
3545 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3546 orig_pmd);
3547 /*
3548 * If COW results in an oom, the huge pmd will
3549 * have been split, so retry the fault on the
3550 * pte for a smaller charge.
3551 */
3552 if (unlikely(ret & VM_FAULT_OOM))
3553 goto retry;
3554 return ret;
3555 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003556 return 0;
3557 }
3558 }
3559
3560 /*
3561 * Use __pte_alloc instead of pte_alloc_map, because we can't
3562 * run pte_offset_map on the pmd, if an huge pmd could
3563 * materialize from under us from a different thread.
3564 */
Mel Gormancc036382011-04-27 15:26:56 -07003565 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003566 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003567 /* if an huge pmd materialized from under us just retry later */
3568 if (unlikely(pmd_trans_huge(*pmd)))
3569 return 0;
3570 /*
3571 * A regular pmd is established and it can't morph into a huge pmd
3572 * from under us anymore at this point because we hold the mmap_sem
3573 * read mode and khugepaged takes it in write mode. So now it's
3574 * safe to run pte_offset_map().
3575 */
3576 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003578 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579}
3580
3581#ifndef __PAGETABLE_PUD_FOLDED
3582/*
3583 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003584 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003586int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003588 pud_t *new = pud_alloc_one(mm, address);
3589 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003590 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591
Nick Piggin362a61a2008-05-14 06:37:36 +02003592 smp_wmb(); /* See comment in __pte_alloc */
3593
Hugh Dickins872fec12005-10-29 18:16:21 -07003594 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003595 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003596 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003597 else
3598 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003599 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003600 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601}
3602#endif /* __PAGETABLE_PUD_FOLDED */
3603
3604#ifndef __PAGETABLE_PMD_FOLDED
3605/*
3606 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003607 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003609int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003611 pmd_t *new = pmd_alloc_one(mm, address);
3612 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003613 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614
Nick Piggin362a61a2008-05-14 06:37:36 +02003615 smp_wmb(); /* See comment in __pte_alloc */
3616
Hugh Dickins872fec12005-10-29 18:16:21 -07003617 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003619 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003620 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003621 else
3622 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003624 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003625 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003626 else
3627 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003629 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003630 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631}
3632#endif /* __PAGETABLE_PMD_FOLDED */
3633
3634int make_pages_present(unsigned long addr, unsigned long end)
3635{
3636 int ret, len, write;
3637 struct vm_area_struct * vma;
3638
3639 vma = find_vma(current->mm, addr);
3640 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003641 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003642 /*
3643 * We want to touch writable mappings with a write fault in order
3644 * to break COW, except for shared mappings because these don't COW
3645 * and we would not want to dirty them for nothing.
3646 */
3647 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003648 BUG_ON(addr >= end);
3649 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003650 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 ret = get_user_pages(current, current->mm, addr,
3652 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003653 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003655 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656}
3657
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658#if !defined(__HAVE_ARCH_GATE_AREA)
3659
3660#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003661static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662
3663static int __init gate_vma_init(void)
3664{
3665 gate_vma.vm_mm = NULL;
3666 gate_vma.vm_start = FIXADDR_USER_START;
3667 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003668 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3669 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003670
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 return 0;
3672}
3673__initcall(gate_vma_init);
3674#endif
3675
Stephen Wilson31db58b2011-03-13 15:49:15 -04003676struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677{
3678#ifdef AT_SYSINFO_EHDR
3679 return &gate_vma;
3680#else
3681 return NULL;
3682#endif
3683}
3684
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003685int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686{
3687#ifdef AT_SYSINFO_EHDR
3688 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3689 return 1;
3690#endif
3691 return 0;
3692}
3693
3694#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003695
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003696static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003697 pte_t **ptepp, spinlock_t **ptlp)
3698{
3699 pgd_t *pgd;
3700 pud_t *pud;
3701 pmd_t *pmd;
3702 pte_t *ptep;
3703
3704 pgd = pgd_offset(mm, address);
3705 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3706 goto out;
3707
3708 pud = pud_offset(pgd, address);
3709 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3710 goto out;
3711
3712 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003713 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003714 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3715 goto out;
3716
3717 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3718 if (pmd_huge(*pmd))
3719 goto out;
3720
3721 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3722 if (!ptep)
3723 goto out;
3724 if (!pte_present(*ptep))
3725 goto unlock;
3726 *ptepp = ptep;
3727 return 0;
3728unlock:
3729 pte_unmap_unlock(ptep, *ptlp);
3730out:
3731 return -EINVAL;
3732}
3733
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003734static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3735 pte_t **ptepp, spinlock_t **ptlp)
3736{
3737 int res;
3738
3739 /* (void) is needed to make gcc happy */
3740 (void) __cond_lock(*ptlp,
3741 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3742 return res;
3743}
3744
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003745/**
3746 * follow_pfn - look up PFN at a user virtual address
3747 * @vma: memory mapping
3748 * @address: user virtual address
3749 * @pfn: location to store found PFN
3750 *
3751 * Only IO mappings and raw PFN mappings are allowed.
3752 *
3753 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3754 */
3755int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3756 unsigned long *pfn)
3757{
3758 int ret = -EINVAL;
3759 spinlock_t *ptl;
3760 pte_t *ptep;
3761
3762 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3763 return ret;
3764
3765 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3766 if (ret)
3767 return ret;
3768 *pfn = pte_pfn(*ptep);
3769 pte_unmap_unlock(ptep, ptl);
3770 return 0;
3771}
3772EXPORT_SYMBOL(follow_pfn);
3773
Rik van Riel28b2ee22008-07-23 21:27:05 -07003774#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003775int follow_phys(struct vm_area_struct *vma,
3776 unsigned long address, unsigned int flags,
3777 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003778{
Johannes Weiner03668a42009-06-16 15:32:34 -07003779 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003780 pte_t *ptep, pte;
3781 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003782
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003783 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3784 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003785
Johannes Weiner03668a42009-06-16 15:32:34 -07003786 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003787 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003788 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003789
Rik van Riel28b2ee22008-07-23 21:27:05 -07003790 if ((flags & FOLL_WRITE) && !pte_write(pte))
3791 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003792
3793 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003794 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003795
Johannes Weiner03668a42009-06-16 15:32:34 -07003796 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003797unlock:
3798 pte_unmap_unlock(ptep, ptl);
3799out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003800 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003801}
3802
3803int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3804 void *buf, int len, int write)
3805{
3806 resource_size_t phys_addr;
3807 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003808 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003809 int offset = addr & (PAGE_SIZE-1);
3810
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003811 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003812 return -EINVAL;
3813
3814 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3815 if (write)
3816 memcpy_toio(maddr + offset, buf, len);
3817 else
3818 memcpy_fromio(buf, maddr + offset, len);
3819 iounmap(maddr);
3820
3821 return len;
3822}
3823#endif
3824
David Howells0ec76a12006-09-27 01:50:15 -07003825/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003826 * Access another process' address space as given in mm. If non-NULL, use the
3827 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003828 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003829static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3830 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003831{
David Howells0ec76a12006-09-27 01:50:15 -07003832 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003833 void *old_buf = buf;
3834
David Howells0ec76a12006-09-27 01:50:15 -07003835 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003836 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003837 while (len) {
3838 int bytes, ret, offset;
3839 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003840 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003841
3842 ret = get_user_pages(tsk, mm, addr, 1,
3843 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003844 if (ret <= 0) {
3845 /*
3846 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3847 * we can access using slightly different code.
3848 */
3849#ifdef CONFIG_HAVE_IOREMAP_PROT
3850 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003851 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003852 break;
3853 if (vma->vm_ops && vma->vm_ops->access)
3854 ret = vma->vm_ops->access(vma, addr, buf,
3855 len, write);
3856 if (ret <= 0)
3857#endif
3858 break;
3859 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003860 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003861 bytes = len;
3862 offset = addr & (PAGE_SIZE-1);
3863 if (bytes > PAGE_SIZE-offset)
3864 bytes = PAGE_SIZE-offset;
3865
3866 maddr = kmap(page);
3867 if (write) {
3868 copy_to_user_page(vma, page, addr,
3869 maddr + offset, buf, bytes);
3870 set_page_dirty_lock(page);
3871 } else {
3872 copy_from_user_page(vma, page, addr,
3873 buf, maddr + offset, bytes);
3874 }
3875 kunmap(page);
3876 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003877 }
David Howells0ec76a12006-09-27 01:50:15 -07003878 len -= bytes;
3879 buf += bytes;
3880 addr += bytes;
3881 }
3882 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003883
3884 return buf - old_buf;
3885}
Andi Kleen03252912008-01-30 13:33:18 +01003886
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003887/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003888 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003889 * @mm: the mm_struct of the target address space
3890 * @addr: start address to access
3891 * @buf: source or destination buffer
3892 * @len: number of bytes to transfer
3893 * @write: whether the access is a write
3894 *
3895 * The caller must hold a reference on @mm.
3896 */
3897int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3898 void *buf, int len, int write)
3899{
3900 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3901}
3902
Andi Kleen03252912008-01-30 13:33:18 +01003903/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003904 * Access another process' address space.
3905 * Source/target buffer must be kernel space,
3906 * Do not walk the page table directly, use get_user_pages
3907 */
3908int access_process_vm(struct task_struct *tsk, unsigned long addr,
3909 void *buf, int len, int write)
3910{
3911 struct mm_struct *mm;
3912 int ret;
3913
3914 mm = get_task_mm(tsk);
3915 if (!mm)
3916 return 0;
3917
3918 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3919 mmput(mm);
3920
3921 return ret;
3922}
3923
Andi Kleen03252912008-01-30 13:33:18 +01003924/*
3925 * Print the name of a VMA.
3926 */
3927void print_vma_addr(char *prefix, unsigned long ip)
3928{
3929 struct mm_struct *mm = current->mm;
3930 struct vm_area_struct *vma;
3931
Ingo Molnare8bff742008-02-13 20:21:06 +01003932 /*
3933 * Do not print if we are in atomic
3934 * contexts (in exception stacks, etc.):
3935 */
3936 if (preempt_count())
3937 return;
3938
Andi Kleen03252912008-01-30 13:33:18 +01003939 down_read(&mm->mmap_sem);
3940 vma = find_vma(mm, ip);
3941 if (vma && vma->vm_file) {
3942 struct file *f = vma->vm_file;
3943 char *buf = (char *)__get_free_page(GFP_KERNEL);
3944 if (buf) {
3945 char *p, *s;
3946
Jan Blunckcf28b482008-02-14 19:38:44 -08003947 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003948 if (IS_ERR(p))
3949 p = "?";
3950 s = strrchr(p, '/');
3951 if (s)
3952 p = s+1;
3953 printk("%s%s[%lx+%lx]", prefix, p,
3954 vma->vm_start,
3955 vma->vm_end - vma->vm_start);
3956 free_page((unsigned long)buf);
3957 }
3958 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003959 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003960}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003961
3962#ifdef CONFIG_PROVE_LOCKING
3963void might_fault(void)
3964{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003965 /*
3966 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3967 * holding the mmap_sem, this is safe because kernel memory doesn't
3968 * get paged out, therefore we'll never actually fault, and the
3969 * below annotations will generate false positives.
3970 */
3971 if (segment_eq(get_fs(), KERNEL_DS))
3972 return;
3973
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003974 might_sleep();
3975 /*
3976 * it would be nicer only to annotate paths which are not under
3977 * pagefault_disable, however that requires a larger audit and
3978 * providing helpers like get_user_atomic.
3979 */
3980 if (!in_atomic() && current->mm)
3981 might_lock_read(&current->mm->mmap_sem);
3982}
3983EXPORT_SYMBOL(might_fault);
3984#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003985
3986#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3987static void clear_gigantic_page(struct page *page,
3988 unsigned long addr,
3989 unsigned int pages_per_huge_page)
3990{
3991 int i;
3992 struct page *p = page;
3993
3994 might_sleep();
3995 for (i = 0; i < pages_per_huge_page;
3996 i++, p = mem_map_next(p, page, i)) {
3997 cond_resched();
3998 clear_user_highpage(p, addr + i * PAGE_SIZE);
3999 }
4000}
4001void clear_huge_page(struct page *page,
4002 unsigned long addr, unsigned int pages_per_huge_page)
4003{
4004 int i;
4005
4006 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4007 clear_gigantic_page(page, addr, pages_per_huge_page);
4008 return;
4009 }
4010
4011 might_sleep();
4012 for (i = 0; i < pages_per_huge_page; i++) {
4013 cond_resched();
4014 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4015 }
4016}
4017
4018static void copy_user_gigantic_page(struct page *dst, struct page *src,
4019 unsigned long addr,
4020 struct vm_area_struct *vma,
4021 unsigned int pages_per_huge_page)
4022{
4023 int i;
4024 struct page *dst_base = dst;
4025 struct page *src_base = src;
4026
4027 for (i = 0; i < pages_per_huge_page; ) {
4028 cond_resched();
4029 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4030
4031 i++;
4032 dst = mem_map_next(dst, dst_base, i);
4033 src = mem_map_next(src, src_base, i);
4034 }
4035}
4036
4037void copy_user_huge_page(struct page *dst, struct page *src,
4038 unsigned long addr, struct vm_area_struct *vma,
4039 unsigned int pages_per_huge_page)
4040{
4041 int i;
4042
4043 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4044 copy_user_gigantic_page(dst, src, addr, vma,
4045 pages_per_huge_page);
4046 return;
4047 }
4048
4049 might_sleep();
4050 for (i = 0; i < pages_per_huge_page; i++) {
4051 cond_resched();
4052 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4053 }
4054}
4055#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */