blob: 61a262b08e53efa2f69c1387e488bea6f87bb0f9 [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>
Mel Gorman4daae3b2012-11-02 11:33:45 +000060#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080061#include <linux/string.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Alexey Dobriyan6952b612009-09-18 23:55:55 +040063#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064#include <asm/pgalloc.h>
65#include <asm/uaccess.h>
66#include <asm/tlb.h>
67#include <asm/tlbflush.h>
68#include <asm/pgtable.h>
69
Jan Beulich42b77722008-07-23 21:27:10 -070070#include "internal.h"
71
Peter Zijlstra75980e92013-02-22 16:34:32 -080072#ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
74#endif
75
Andy Whitcroftd41dee32005-06-23 00:07:54 -070076#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070077/* use the per-pgdat data instead for discontigmem - mbligh */
78unsigned long max_mapnr;
79struct page *mem_map;
80
81EXPORT_SYMBOL(max_mapnr);
82EXPORT_SYMBOL(mem_map);
83#endif
84
85unsigned long num_physpages;
86/*
87 * A number of key systems in x86 including ioremap() rely on the assumption
88 * that high_memory defines the upper bound on direct map memory, then end
89 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
90 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
91 * and ZONE_HIGHMEM.
92 */
93void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
95EXPORT_SYMBOL(num_physpages);
96EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Ingo Molnar32a93232008-02-06 22:39:44 +010098/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800114 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100115}
116__setup("norandmaps", disable_randmaps);
117
Hugh Dickins62eede62009-09-21 17:03:34 -0700118unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700119unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100130
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800131
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800132#if defined(SPLIT_RSS_COUNTING)
133
David Rientjesea48cf72012-03-21 16:34:13 -0700134void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800142 }
143 }
David Rientjes05af2e12012-03-21 16:34:13 -0700144 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700166 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800167}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700168#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
Peter Zijlstra9547d012011-05-24 17:12:14 -0700177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
Michal Hocko53a59fc2013-01-04 15:35:12 -0800191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
Peter Zijlstra9547d012011-05-24 17:12:14 -0700194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
Michal Hocko53a59fc2013-01-04 15:35:12 -0800198 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
215{
216 tlb->mm = mm;
217
218 tlb->fullmm = fullmm;
Dave Hansen1de14c32013-04-12 16:23:54 -0700219 tlb->need_flush_all = 0;
Alex Shi597e1c32012-06-28 09:02:21 +0800220 tlb->start = -1UL;
221 tlb->end = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700222 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700223 tlb->local.next = NULL;
224 tlb->local.nr = 0;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800227 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700228
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800231#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700232}
233
234void tlb_flush_mmu(struct mmu_gather *tlb)
235{
236 struct mmu_gather_batch *batch;
237
238 if (!tlb->need_flush)
239 return;
240 tlb->need_flush = 0;
241 tlb_flush(tlb);
242#ifdef CONFIG_HAVE_RCU_TABLE_FREE
243 tlb_table_flush(tlb);
244#endif
245
Peter Zijlstra9547d012011-05-24 17:12:14 -0700246 for (batch = &tlb->local; batch; batch = batch->next) {
247 free_pages_and_swap_cache(batch->pages, batch->nr);
248 batch->nr = 0;
249 }
250 tlb->active = &tlb->local;
251}
252
253/* tlb_finish_mmu
254 * Called at the end of the shootdown operation to free up any resources
255 * that were required.
256 */
257void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
258{
259 struct mmu_gather_batch *batch, *next;
260
Alex Shi597e1c32012-06-28 09:02:21 +0800261 tlb->start = start;
262 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700263 tlb_flush_mmu(tlb);
264
265 /* keep the page table cache within bounds */
266 check_pgt_cache();
267
268 for (batch = tlb->local.next; batch; batch = next) {
269 next = batch->next;
270 free_pages((unsigned long)batch, 0);
271 }
272 tlb->local.next = NULL;
273}
274
275/* __tlb_remove_page
276 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
277 * handling the additional races in SMP caused by other CPUs caching valid
278 * mappings in their TLBs. Returns the number of free page slots left.
279 * When out of page slots we must call tlb_flush_mmu().
280 */
281int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
282{
283 struct mmu_gather_batch *batch;
284
Shaohua Lif21760b2012-01-12 17:19:16 -0800285 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700286
Peter Zijlstra9547d012011-05-24 17:12:14 -0700287 batch = tlb->active;
288 batch->pages[batch->nr++] = page;
289 if (batch->nr == batch->max) {
290 if (!tlb_next_batch(tlb))
291 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700292 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700293 }
294 VM_BUG_ON(batch->nr > batch->max);
295
296 return batch->max - batch->nr;
297}
298
299#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800300
Peter Zijlstra26723912011-05-24 17:12:00 -0700301#ifdef CONFIG_HAVE_RCU_TABLE_FREE
302
303/*
304 * See the comment near struct mmu_table_batch.
305 */
306
307static void tlb_remove_table_smp_sync(void *arg)
308{
309 /* Simply deliver the interrupt */
310}
311
312static void tlb_remove_table_one(void *table)
313{
314 /*
315 * This isn't an RCU grace period and hence the page-tables cannot be
316 * assumed to be actually RCU-freed.
317 *
318 * It is however sufficient for software page-table walkers that rely on
319 * IRQ disabling. See the comment near struct mmu_table_batch.
320 */
321 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
322 __tlb_remove_table(table);
323}
324
325static void tlb_remove_table_rcu(struct rcu_head *head)
326{
327 struct mmu_table_batch *batch;
328 int i;
329
330 batch = container_of(head, struct mmu_table_batch, rcu);
331
332 for (i = 0; i < batch->nr; i++)
333 __tlb_remove_table(batch->tables[i]);
334
335 free_page((unsigned long)batch);
336}
337
338void tlb_table_flush(struct mmu_gather *tlb)
339{
340 struct mmu_table_batch **batch = &tlb->batch;
341
342 if (*batch) {
343 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
344 *batch = NULL;
345 }
346}
347
348void tlb_remove_table(struct mmu_gather *tlb, void *table)
349{
350 struct mmu_table_batch **batch = &tlb->batch;
351
352 tlb->need_flush = 1;
353
354 /*
355 * When there's less then two users of this mm there cannot be a
356 * concurrent page-table walk.
357 */
358 if (atomic_read(&tlb->mm->mm_users) < 2) {
359 __tlb_remove_table(table);
360 return;
361 }
362
363 if (*batch == NULL) {
364 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
365 if (*batch == NULL) {
366 tlb_remove_table_one(table);
367 return;
368 }
369 (*batch)->nr = 0;
370 }
371 (*batch)->tables[(*batch)->nr++] = table;
372 if ((*batch)->nr == MAX_TABLE_BATCH)
373 tlb_table_flush(tlb);
374}
375
Peter Zijlstra9547d012011-05-24 17:12:14 -0700376#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700377
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378/*
379 * If a p?d_bad entry is found while walking page tables, report
380 * the error, before resetting entry to p?d_none. Usually (but
381 * very seldom) called out from the p?d_none_or_clear_bad macros.
382 */
383
384void pgd_clear_bad(pgd_t *pgd)
385{
386 pgd_ERROR(*pgd);
387 pgd_clear(pgd);
388}
389
390void pud_clear_bad(pud_t *pud)
391{
392 pud_ERROR(*pud);
393 pud_clear(pud);
394}
395
396void pmd_clear_bad(pmd_t *pmd)
397{
398 pmd_ERROR(*pmd);
399 pmd_clear(pmd);
400}
401
402/*
403 * Note: this doesn't free the actual pages themselves. That
404 * has been handled earlier when unmapping all the memory regions.
405 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000406static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
407 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800409 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700410 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000411 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700412 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413}
414
Hugh Dickinse0da3822005-04-19 13:29:15 -0700415static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
416 unsigned long addr, unsigned long end,
417 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418{
419 pmd_t *pmd;
420 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
Hugh Dickinse0da3822005-04-19 13:29:15 -0700423 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 do {
426 next = pmd_addr_end(addr, end);
427 if (pmd_none_or_clear_bad(pmd))
428 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000429 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 } while (pmd++, addr = next, addr != end);
431
Hugh Dickinse0da3822005-04-19 13:29:15 -0700432 start &= PUD_MASK;
433 if (start < floor)
434 return;
435 if (ceiling) {
436 ceiling &= PUD_MASK;
437 if (!ceiling)
438 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700440 if (end - 1 > ceiling - 1)
441 return;
442
443 pmd = pmd_offset(pud, start);
444 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000445 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446}
447
Hugh Dickinse0da3822005-04-19 13:29:15 -0700448static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
449 unsigned long addr, unsigned long end,
450 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451{
452 pud_t *pud;
453 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700454 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Hugh Dickinse0da3822005-04-19 13:29:15 -0700456 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 do {
459 next = pud_addr_end(addr, end);
460 if (pud_none_or_clear_bad(pud))
461 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700462 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 } while (pud++, addr = next, addr != end);
464
Hugh Dickinse0da3822005-04-19 13:29:15 -0700465 start &= PGDIR_MASK;
466 if (start < floor)
467 return;
468 if (ceiling) {
469 ceiling &= PGDIR_MASK;
470 if (!ceiling)
471 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700473 if (end - 1 > ceiling - 1)
474 return;
475
476 pud = pud_offset(pgd, start);
477 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000478 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479}
480
481/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700482 * This function frees user-level page tables of a process.
483 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 * Must be called with pagetable lock held.
485 */
Jan Beulich42b77722008-07-23 21:27:10 -0700486void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700487 unsigned long addr, unsigned long end,
488 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489{
490 pgd_t *pgd;
491 unsigned long next;
492
Hugh Dickinse0da3822005-04-19 13:29:15 -0700493 /*
494 * The next few lines have given us lots of grief...
495 *
496 * Why are we testing PMD* at this top level? Because often
497 * there will be no work to do at all, and we'd prefer not to
498 * go all the way down to the bottom just to discover that.
499 *
500 * Why all these "- 1"s? Because 0 represents both the bottom
501 * of the address space and the top of it (using -1 for the
502 * top wouldn't help much: the masks would do the wrong thing).
503 * The rule is that addr 0 and floor 0 refer to the bottom of
504 * the address space, but end 0 and ceiling 0 refer to the top
505 * Comparisons need to use "end - 1" and "ceiling - 1" (though
506 * that end 0 case should be mythical).
507 *
508 * Wherever addr is brought up or ceiling brought down, we must
509 * be careful to reject "the opposite 0" before it confuses the
510 * subsequent tests. But what about where end is brought down
511 * by PMD_SIZE below? no, end can't go down to 0 there.
512 *
513 * Whereas we round start (addr) and ceiling down, by different
514 * masks at different levels, in order to test whether a table
515 * now has no other vmas using it, so can be freed, we don't
516 * bother to round floor or end up - the tests don't need that.
517 */
518
519 addr &= PMD_MASK;
520 if (addr < floor) {
521 addr += PMD_SIZE;
522 if (!addr)
523 return;
524 }
525 if (ceiling) {
526 ceiling &= PMD_MASK;
527 if (!ceiling)
528 return;
529 }
530 if (end - 1 > ceiling - 1)
531 end -= PMD_SIZE;
532 if (addr > end - 1)
533 return;
534
Jan Beulich42b77722008-07-23 21:27:10 -0700535 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 do {
537 next = pgd_addr_end(addr, end);
538 if (pgd_none_or_clear_bad(pgd))
539 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700540 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700542}
543
Jan Beulich42b77722008-07-23 21:27:10 -0700544void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700545 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700546{
547 while (vma) {
548 struct vm_area_struct *next = vma->vm_next;
549 unsigned long addr = vma->vm_start;
550
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700551 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000552 * Hide vma from rmap and truncate_pagecache before freeing
553 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700554 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800555 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700556 unlink_file_vma(vma);
557
David Gibson9da61ae2006-03-22 00:08:57 -0800558 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700559 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700560 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700561 } else {
562 /*
563 * Optimization: gather nearby vmas into one call down
564 */
565 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800566 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700567 vma = next;
568 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800569 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700570 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700571 }
572 free_pgd_range(tlb, addr, vma->vm_end,
573 floor, next? next->vm_start: ceiling);
574 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700575 vma = next;
576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577}
578
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800579int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
580 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800582 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800583 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700584 if (!new)
585 return -ENOMEM;
586
Nick Piggin362a61a2008-05-14 06:37:36 +0200587 /*
588 * Ensure all pte setup (eg. pte page lock and page clearing) are
589 * visible before the pte is made visible to other CPUs by being
590 * put into page tables.
591 *
592 * The other side of the story is the pointer chasing in the page
593 * table walking code (when walking the page table without locking;
594 * ie. most of the time). Fortunately, these data accesses consist
595 * of a chain of data-dependent loads, meaning most CPUs (alpha
596 * being the notable exception) will already guarantee loads are
597 * seen in-order. See the alpha page table accessors for the
598 * smp_read_barrier_depends() barriers in page table walking code.
599 */
600 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
601
Hugh Dickinsc74df322005-10-29 18:16:23 -0700602 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800603 wait_split_huge_page = 0;
604 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800607 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800608 } else if (unlikely(pmd_trans_splitting(*pmd)))
609 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700610 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800611 if (new)
612 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800613 if (wait_split_huge_page)
614 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700615 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616}
617
Hugh Dickins1bb36302005-10-29 18:16:22 -0700618int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700620 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
621 if (!new)
622 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623
Nick Piggin362a61a2008-05-14 06:37:36 +0200624 smp_wmb(); /* See comment in __pte_alloc */
625
Hugh Dickins1bb36302005-10-29 18:16:22 -0700626 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800627 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700628 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800629 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800630 } else
631 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700632 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800633 if (new)
634 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700635 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636}
637
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800638static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700639{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800640 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
641}
642
643static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
644{
645 int i;
646
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800647 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700648 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800649 for (i = 0; i < NR_MM_COUNTERS; i++)
650 if (rss[i])
651 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700652}
653
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800655 * This function is called to print an error when a bad pte
656 * is found. For example, we might have a PFN-mapped pte in
657 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700658 *
659 * The calling function must still handle the error.
660 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800661static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
662 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700663{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800664 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
665 pud_t *pud = pud_offset(pgd, addr);
666 pmd_t *pmd = pmd_offset(pud, addr);
667 struct address_space *mapping;
668 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800669 static unsigned long resume;
670 static unsigned long nr_shown;
671 static unsigned long nr_unshown;
672
673 /*
674 * Allow a burst of 60 reports, then keep quiet for that minute;
675 * or allow a steady drip of one report per second.
676 */
677 if (nr_shown == 60) {
678 if (time_before(jiffies, resume)) {
679 nr_unshown++;
680 return;
681 }
682 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800683 printk(KERN_ALERT
684 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800685 nr_unshown);
686 nr_unshown = 0;
687 }
688 nr_shown = 0;
689 }
690 if (nr_shown++ == 0)
691 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800692
693 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
694 index = linear_page_index(vma, addr);
695
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800696 printk(KERN_ALERT
697 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800698 current->comm,
699 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800700 if (page)
701 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800702 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800703 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
704 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
705 /*
706 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
707 */
708 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800709 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
710 vma->vm_ops->fault);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800711 if (vma->vm_file && vma->vm_file->f_op)
Joe Perches071361d2012-12-12 10:19:12 -0800712 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
713 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700714 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030715 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700716}
717
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700718static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800719{
720 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
721}
722
Nick Pigginb5810032005-10-29 18:16:12 -0700723/*
Nick Piggin7e675132008-04-28 02:13:00 -0700724 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800725 *
Nick Piggin7e675132008-04-28 02:13:00 -0700726 * "Special" mappings do not wish to be associated with a "struct page" (either
727 * it doesn't exist, or it exists but they don't want to touch it). In this
728 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 *
Nick Piggin7e675132008-04-28 02:13:00 -0700730 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
731 * pte bit, in which case this function is trivial. Secondly, an architecture
732 * may not have a spare pte bit, which requires a more complicated scheme,
733 * described below.
734 *
735 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
736 * special mapping (even if there are underlying and valid "struct pages").
737 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800738 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700739 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
740 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700741 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
742 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800743 *
744 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
745 *
Nick Piggin7e675132008-04-28 02:13:00 -0700746 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700747 *
Nick Piggin7e675132008-04-28 02:13:00 -0700748 * This restricts such mappings to be a linear translation from virtual address
749 * to pfn. To get around this restriction, we allow arbitrary mappings so long
750 * as the vma is not a COW mapping; in that case, we know that all ptes are
751 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700752 *
753 *
Nick Piggin7e675132008-04-28 02:13:00 -0700754 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
755 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700756 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
757 * page" backing, however the difference is that _all_ pages with a struct
758 * page (that is, those where pfn_valid is true) are refcounted and considered
759 * normal pages by the VM. The disadvantage is that pages are refcounted
760 * (which can be slower and simply not an option for some PFNMAP users). The
761 * advantage is that we don't have to follow the strict linearity rule of
762 * PFNMAP mappings in order to support COWable mappings.
763 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800764 */
Nick Piggin7e675132008-04-28 02:13:00 -0700765#ifdef __HAVE_ARCH_PTE_SPECIAL
766# define HAVE_PTE_SPECIAL 1
767#else
768# define HAVE_PTE_SPECIAL 0
769#endif
770struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
771 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800772{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800773 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700774
775 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800776 if (likely(!pte_special(pte)))
777 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700778 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
779 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700780 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800781 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700782 return NULL;
783 }
784
785 /* !HAVE_PTE_SPECIAL case follows: */
786
Jared Hulbertb379d792008-04-28 02:12:58 -0700787 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
788 if (vma->vm_flags & VM_MIXEDMAP) {
789 if (!pfn_valid(pfn))
790 return NULL;
791 goto out;
792 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700793 unsigned long off;
794 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700795 if (pfn == vma->vm_pgoff + off)
796 return NULL;
797 if (!is_cow_mapping(vma->vm_flags))
798 return NULL;
799 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800800 }
801
Hugh Dickins62eede62009-09-21 17:03:34 -0700802 if (is_zero_pfn(pfn))
803 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800804check_pfn:
805 if (unlikely(pfn > highest_memmap_pfn)) {
806 print_bad_pte(vma, addr, pte, NULL);
807 return NULL;
808 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800809
810 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700811 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700812 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800813 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700814out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800815 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800816}
817
818/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 * copy one vm_area from one task to the other. Assumes the page tables
820 * already present in the new task to be cleared in the whole range
821 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 */
823
Hugh Dickins570a335b2009-12-14 17:58:46 -0800824static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700826 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700827 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828{
Nick Pigginb5810032005-10-29 18:16:12 -0700829 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 pte_t pte = *src_pte;
831 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
833 /* pte contains position in swap or file, so copy. */
834 if (unlikely(!pte_present(pte))) {
835 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700836 swp_entry_t entry = pte_to_swp_entry(pte);
837
Hugh Dickins570a335b2009-12-14 17:58:46 -0800838 if (swap_duplicate(entry) < 0)
839 return entry.val;
840
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 /* make sure dst_mm is on swapoff's mmlist. */
842 if (unlikely(list_empty(&dst_mm->mmlist))) {
843 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700844 if (list_empty(&dst_mm->mmlist))
845 list_add(&dst_mm->mmlist,
846 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 spin_unlock(&mmlist_lock);
848 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800849 if (likely(!non_swap_entry(entry)))
850 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800851 else if (is_migration_entry(entry)) {
852 page = migration_entry_to_page(entry);
853
854 if (PageAnon(page))
855 rss[MM_ANONPAGES]++;
856 else
857 rss[MM_FILEPAGES]++;
858
859 if (is_write_migration_entry(entry) &&
860 is_cow_mapping(vm_flags)) {
861 /*
862 * COW mappings require pages in both
863 * parent and child to be set to read.
864 */
865 make_migration_entry_read(&entry);
866 pte = swp_entry_to_pte(entry);
867 set_pte_at(src_mm, addr, src_pte, pte);
868 }
Christoph Lameter06972122006-06-23 02:03:35 -0700869 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700871 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 }
873
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 /*
875 * If it's a COW mapping, write protect it both
876 * in the parent and the child
877 */
Linus Torvalds67121172005-12-11 20:38:17 -0800878 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700880 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 }
882
883 /*
884 * If it's a shared mapping, mark it clean in
885 * the child
886 */
887 if (vm_flags & VM_SHARED)
888 pte = pte_mkclean(pte);
889 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800890
891 page = vm_normal_page(vma, addr, pte);
892 if (page) {
893 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700894 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800895 if (PageAnon(page))
896 rss[MM_ANONPAGES]++;
897 else
898 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800899 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700900
901out_set_pte:
902 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800903 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904}
905
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800906int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
907 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
908 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700910 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700912 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700913 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800914 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800915 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916
917again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800918 init_rss_vec(rss);
919
Hugh Dickinsc74df322005-10-29 18:16:23 -0700920 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 if (!dst_pte)
922 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700923 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700924 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700925 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700926 orig_src_pte = src_pte;
927 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700928 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 do {
931 /*
932 * We are holding two locks at this point - either of them
933 * could generate latencies in another task on another CPU.
934 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700935 if (progress >= 32) {
936 progress = 0;
937 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100938 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700939 break;
940 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 if (pte_none(*src_pte)) {
942 progress++;
943 continue;
944 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800945 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
946 vma, addr, rss);
947 if (entry.val)
948 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 progress += 8;
950 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700952 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700953 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700954 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800955 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700956 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700957 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800958
959 if (entry.val) {
960 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
961 return -ENOMEM;
962 progress = 0;
963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 if (addr != end)
965 goto again;
966 return 0;
967}
968
969static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
970 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
971 unsigned long addr, unsigned long end)
972{
973 pmd_t *src_pmd, *dst_pmd;
974 unsigned long next;
975
976 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
977 if (!dst_pmd)
978 return -ENOMEM;
979 src_pmd = pmd_offset(src_pud, addr);
980 do {
981 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800982 if (pmd_trans_huge(*src_pmd)) {
983 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800984 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800985 err = copy_huge_pmd(dst_mm, src_mm,
986 dst_pmd, src_pmd, addr, vma);
987 if (err == -ENOMEM)
988 return -ENOMEM;
989 if (!err)
990 continue;
991 /* fall through */
992 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 if (pmd_none_or_clear_bad(src_pmd))
994 continue;
995 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
996 vma, addr, next))
997 return -ENOMEM;
998 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
999 return 0;
1000}
1001
1002static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1003 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1004 unsigned long addr, unsigned long end)
1005{
1006 pud_t *src_pud, *dst_pud;
1007 unsigned long next;
1008
1009 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1010 if (!dst_pud)
1011 return -ENOMEM;
1012 src_pud = pud_offset(src_pgd, addr);
1013 do {
1014 next = pud_addr_end(addr, end);
1015 if (pud_none_or_clear_bad(src_pud))
1016 continue;
1017 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1018 vma, addr, next))
1019 return -ENOMEM;
1020 } while (dst_pud++, src_pud++, addr = next, addr != end);
1021 return 0;
1022}
1023
1024int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1025 struct vm_area_struct *vma)
1026{
1027 pgd_t *src_pgd, *dst_pgd;
1028 unsigned long next;
1029 unsigned long addr = vma->vm_start;
1030 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001031 unsigned long mmun_start; /* For mmu_notifiers */
1032 unsigned long mmun_end; /* For mmu_notifiers */
1033 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001034 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035
Nick Piggind9928952005-08-28 16:49:11 +10001036 /*
1037 * Don't copy ptes where a page fault will fill them correctly.
1038 * Fork becomes much lighter when there are big shared or private
1039 * readonly mappings. The tradeoff is that copy_page_range is more
1040 * efficient than faulting.
1041 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001042 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1043 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001044 if (!vma->anon_vma)
1045 return 0;
1046 }
1047
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 if (is_vm_hugetlb_page(vma))
1049 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1050
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001051 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001052 /*
1053 * We do not free on error cases below as remove_vma
1054 * gets called on error from higher level routine
1055 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001056 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001057 if (ret)
1058 return ret;
1059 }
1060
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001061 /*
1062 * We need to invalidate the secondary MMU mappings only when
1063 * there could be a permission downgrade on the ptes of the
1064 * parent mm. And a permission downgrade will only happen if
1065 * is_cow_mapping() returns true.
1066 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001067 is_cow = is_cow_mapping(vma->vm_flags);
1068 mmun_start = addr;
1069 mmun_end = end;
1070 if (is_cow)
1071 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1072 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001073
1074 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 dst_pgd = pgd_offset(dst_mm, addr);
1076 src_pgd = pgd_offset(src_mm, addr);
1077 do {
1078 next = pgd_addr_end(addr, end);
1079 if (pgd_none_or_clear_bad(src_pgd))
1080 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001081 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1082 vma, addr, next))) {
1083 ret = -ENOMEM;
1084 break;
1085 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001087
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001088 if (is_cow)
1089 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001090 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091}
1092
Robin Holt51c6f662005-11-13 16:06:42 -08001093static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001094 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001096 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097{
Nick Pigginb5810032005-10-29 18:16:12 -07001098 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001099 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001100 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001101 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001102 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001103 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001104
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001105again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001106 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001107 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001109 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 do {
1111 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001112 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001114 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001115
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001117 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001118
Linus Torvalds6aab3412005-11-28 14:34:23 -08001119 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 if (unlikely(details) && page) {
1121 /*
1122 * unmap_shared_mapping_pages() wants to
1123 * invalidate cache without truncating:
1124 * unmap shared but keep private pages.
1125 */
1126 if (details->check_mapping &&
1127 details->check_mapping != page->mapping)
1128 continue;
1129 /*
1130 * Each page->index must be checked when
1131 * invalidating or truncating nonlinear.
1132 */
1133 if (details->nonlinear_vma &&
1134 (page->index < details->first_index ||
1135 page->index > details->last_index))
1136 continue;
1137 }
Nick Pigginb5810032005-10-29 18:16:12 -07001138 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001139 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 tlb_remove_tlb_entry(tlb, pte, addr);
1141 if (unlikely(!page))
1142 continue;
1143 if (unlikely(details) && details->nonlinear_vma
1144 && linear_page_index(details->nonlinear_vma,
1145 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001146 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001149 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001150 else {
1151 if (pte_dirty(ptent))
1152 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001153 if (pte_young(ptent) &&
1154 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001155 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001156 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001157 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001158 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001159 if (unlikely(page_mapcount(page) < 0))
1160 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001161 force_flush = !__tlb_remove_page(tlb, page);
1162 if (force_flush)
1163 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 continue;
1165 }
1166 /*
1167 * If details->check_mapping, we leave swap entries;
1168 * if details->nonlinear_vma, we leave file entries.
1169 */
1170 if (unlikely(details))
1171 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001172 if (pte_file(ptent)) {
1173 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001175 } else {
1176 swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178 if (!non_swap_entry(entry))
1179 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001180 else if (is_migration_entry(entry)) {
1181 struct page *page;
1182
1183 page = migration_entry_to_page(entry);
1184
1185 if (PageAnon(page))
1186 rss[MM_ANONPAGES]--;
1187 else
1188 rss[MM_FILEPAGES]--;
1189 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001190 if (unlikely(!free_swap_and_cache(entry)))
1191 print_bad_pte(vma, addr, ptent, NULL);
1192 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001193 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001194 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001195
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001196 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001197 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001198 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001199
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001200 /*
1201 * mmu_gather ran out of room to batch pages, we break out of
1202 * the PTE lock to avoid doing the potential expensive TLB invalidate
1203 * and page-free while holding it.
1204 */
1205 if (force_flush) {
1206 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001207
1208#ifdef HAVE_GENERIC_MMU_GATHER
1209 tlb->start = addr;
1210 tlb->end = end;
1211#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001212 tlb_flush_mmu(tlb);
1213 if (addr != end)
1214 goto again;
1215 }
1216
Robin Holt51c6f662005-11-13 16:06:42 -08001217 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218}
1219
Robin Holt51c6f662005-11-13 16:06:42 -08001220static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001221 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001223 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224{
1225 pmd_t *pmd;
1226 unsigned long next;
1227
1228 pmd = pmd_offset(pud, addr);
1229 do {
1230 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001231 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001232 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001233#ifdef CONFIG_DEBUG_VM
1234 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1235 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1236 __func__, addr, end,
1237 vma->vm_start,
1238 vma->vm_end);
1239 BUG();
1240 }
1241#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001242 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001243 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001244 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001245 /* fall through */
1246 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001247 /*
1248 * Here there can be other concurrent MADV_DONTNEED or
1249 * trans huge page faults running, and if the pmd is
1250 * none or trans huge it can change under us. This is
1251 * because MADV_DONTNEED holds the mmap_sem in read
1252 * mode.
1253 */
1254 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1255 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001256 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001257next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001258 cond_resched();
1259 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001260
1261 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262}
1263
Robin Holt51c6f662005-11-13 16:06:42 -08001264static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001265 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001267 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268{
1269 pud_t *pud;
1270 unsigned long next;
1271
1272 pud = pud_offset(pgd, addr);
1273 do {
1274 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001275 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001277 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1278 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001279
1280 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281}
1282
Al Viro038c7aa2012-03-05 13:25:09 -05001283static void unmap_page_range(struct mmu_gather *tlb,
1284 struct vm_area_struct *vma,
1285 unsigned long addr, unsigned long end,
1286 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287{
1288 pgd_t *pgd;
1289 unsigned long next;
1290
1291 if (details && !details->check_mapping && !details->nonlinear_vma)
1292 details = NULL;
1293
1294 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001295 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 tlb_start_vma(tlb, vma);
1297 pgd = pgd_offset(vma->vm_mm, addr);
1298 do {
1299 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001300 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001302 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1303 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001305 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307
Al Virof5cc4ee2012-03-05 14:14:20 -05001308
1309static void unmap_single_vma(struct mmu_gather *tlb,
1310 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001311 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001312 struct zap_details *details)
1313{
1314 unsigned long start = max(vma->vm_start, start_addr);
1315 unsigned long end;
1316
1317 if (start >= vma->vm_end)
1318 return;
1319 end = min(vma->vm_end, end_addr);
1320 if (end <= vma->vm_start)
1321 return;
1322
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301323 if (vma->vm_file)
1324 uprobe_munmap(vma, start, end);
1325
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001326 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001327 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001328
1329 if (start != end) {
1330 if (unlikely(is_vm_hugetlb_page(vma))) {
1331 /*
1332 * It is undesirable to test vma->vm_file as it
1333 * should be non-null for valid hugetlb area.
1334 * However, vm_file will be NULL in the error
1335 * cleanup path of do_mmap_pgoff. When
1336 * hugetlbfs ->mmap method fails,
1337 * do_mmap_pgoff() nullifies vma->vm_file
1338 * before calling this function to clean up.
1339 * Since no pte has actually been setup, it is
1340 * safe to do nothing in this case.
1341 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001342 if (vma->vm_file) {
1343 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001344 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001345 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1346 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001347 } else
1348 unmap_page_range(tlb, vma, start, end, details);
1349 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350}
1351
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352/**
1353 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001354 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 * @vma: the starting vma
1356 * @start_addr: virtual address at which to start unmapping
1357 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001359 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 * Only addresses between `start' and `end' will be unmapped.
1362 *
1363 * The VMA list must be sorted in ascending virtual address order.
1364 *
1365 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1366 * range after unmap_vmas() returns. So the only responsibility here is to
1367 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1368 * drops the lock and schedules.
1369 */
Al Viro6e8bb012012-03-05 13:41:15 -05001370void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001372 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001374 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001376 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001377 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001378 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001379 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
1382/**
1383 * zap_page_range - remove user pages in a given range
1384 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001385 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 * @size: number of bytes to zap
1387 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001388 *
1389 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001391void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 unsigned long size, struct zap_details *details)
1393{
1394 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001395 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001396 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001399 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001400 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001401 mmu_notifier_invalidate_range_start(mm, start, end);
1402 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001403 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001404 mmu_notifier_invalidate_range_end(mm, start, end);
1405 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406}
1407
Jack Steinerc627f9c2008-07-29 22:33:53 -07001408/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001409 * zap_page_range_single - remove user pages in a given range
1410 * @vma: vm_area_struct holding the applicable pages
1411 * @address: starting address of pages to zap
1412 * @size: number of bytes to zap
1413 * @details: details of nonlinear truncation or shared cache invalidation
1414 *
1415 * The range must fit into one VMA.
1416 */
1417static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1418 unsigned long size, struct zap_details *details)
1419{
1420 struct mm_struct *mm = vma->vm_mm;
1421 struct mmu_gather tlb;
1422 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001423
1424 lru_add_drain();
1425 tlb_gather_mmu(&tlb, mm, 0);
1426 update_hiwater_rss(mm);
1427 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001428 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001429 mmu_notifier_invalidate_range_end(mm, address, end);
1430 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
Jack Steinerc627f9c2008-07-29 22:33:53 -07001433/**
1434 * zap_vma_ptes - remove ptes mapping the vma
1435 * @vma: vm_area_struct holding ptes to be zapped
1436 * @address: starting address of pages to zap
1437 * @size: number of bytes to zap
1438 *
1439 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1440 *
1441 * The entire address range must be fully contained within the vma.
1442 *
1443 * Returns 0 if successful.
1444 */
1445int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1446 unsigned long size)
1447{
1448 if (address < vma->vm_start || address + size > vma->vm_end ||
1449 !(vma->vm_flags & VM_PFNMAP))
1450 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001451 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001452 return 0;
1453}
1454EXPORT_SYMBOL_GPL(zap_vma_ptes);
1455
Johannes Weiner142762b2010-05-24 14:32:39 -07001456/**
Michel Lespinasse240aade2013-02-22 16:35:56 -08001457 * follow_page_mask - look up a page descriptor from a user-virtual address
Johannes Weiner142762b2010-05-24 14:32:39 -07001458 * @vma: vm_area_struct mapping @address
1459 * @address: virtual address to look up
1460 * @flags: flags modifying lookup behaviour
Michel Lespinasse240aade2013-02-22 16:35:56 -08001461 * @page_mask: on output, *page_mask is set according to the size of the page
Johannes Weiner142762b2010-05-24 14:32:39 -07001462 *
1463 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1464 *
1465 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1466 * an error pointer if there is a mapping to something not represented
1467 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 */
Michel Lespinasse240aade2013-02-22 16:35:56 -08001469struct page *follow_page_mask(struct vm_area_struct *vma,
1470 unsigned long address, unsigned int flags,
1471 unsigned int *page_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
1473 pgd_t *pgd;
1474 pud_t *pud;
1475 pmd_t *pmd;
1476 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001477 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001479 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Michel Lespinasse240aade2013-02-22 16:35:56 -08001481 *page_mask = 0;
1482
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001483 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484 if (!IS_ERR(page)) {
1485 BUG_ON(flags & FOLL_GET);
1486 goto out;
1487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001489 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 pgd = pgd_offset(mm, address);
1491 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001495 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001497 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001498 BUG_ON(flags & FOLL_GET);
1499 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500 goto out;
1501 }
1502 if (unlikely(pud_bad(*pud)))
1503 goto no_page_table;
1504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001506 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001507 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001508 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001509 BUG_ON(flags & FOLL_GET);
1510 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511 goto out;
1512 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001513 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1514 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001515 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001516 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001517 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001518 goto split_fallthrough;
1519 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001520 spin_lock(&mm->page_table_lock);
1521 if (likely(pmd_trans_huge(*pmd))) {
1522 if (unlikely(pmd_trans_splitting(*pmd))) {
1523 spin_unlock(&mm->page_table_lock);
1524 wait_split_huge_page(vma->anon_vma, pmd);
1525 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001526 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001527 pmd, flags);
1528 spin_unlock(&mm->page_table_lock);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001529 *page_mask = HPAGE_PMD_NR - 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001530 goto out;
1531 }
1532 } else
1533 spin_unlock(&mm->page_table_lock);
1534 /* fall through */
1535 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001536split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001537 if (unlikely(pmd_bad(*pmd)))
1538 goto no_page_table;
1539
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001540 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541
1542 pte = *ptep;
Hugh Dickins5117b3b2013-02-22 16:36:07 -08001543 if (!pte_present(pte)) {
1544 swp_entry_t entry;
1545 /*
1546 * KSM's break_ksm() relies upon recognizing a ksm page
1547 * even while it is being migrated, so for that case we
1548 * need migration_entry_wait().
1549 */
1550 if (likely(!(flags & FOLL_MIGRATION)))
1551 goto no_page;
1552 if (pte_none(pte) || pte_file(pte))
1553 goto no_page;
1554 entry = pte_to_swp_entry(pte);
1555 if (!is_migration_entry(entry))
1556 goto no_page;
1557 pte_unmap_unlock(ptep, ptl);
1558 migration_entry_wait(mm, pmd, address);
1559 goto split_fallthrough;
1560 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001561 if ((flags & FOLL_NUMA) && pte_numa(pte))
1562 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001563 if ((flags & FOLL_WRITE) && !pte_write(pte))
1564 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001565
Linus Torvalds6aab3412005-11-28 14:34:23 -08001566 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001567 if (unlikely(!page)) {
1568 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001569 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001570 goto bad_page;
1571 page = pte_page(pte);
1572 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001573
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001574 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001575 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001576 if (flags & FOLL_TOUCH) {
1577 if ((flags & FOLL_WRITE) &&
1578 !pte_dirty(pte) && !PageDirty(page))
1579 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001580 /*
1581 * pte_mkyoung() would be more correct here, but atomic care
1582 * is needed to avoid losing the dirty bit: it is easier to use
1583 * mark_page_accessed().
1584 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001585 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001587 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001588 /*
1589 * The preliminary mapping check is mainly to avoid the
1590 * pointless overhead of lock_page on the ZERO_PAGE
1591 * which might bounce very badly if there is contention.
1592 *
1593 * If the page is already locked, we don't need to
1594 * handle it now - vmscan will handle it later if and
1595 * when it attempts to reclaim the page.
1596 */
1597 if (page->mapping && trylock_page(page)) {
1598 lru_add_drain(); /* push cached pages to LRU */
1599 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001600 * Because we lock page here, and migration is
1601 * blocked by the pte's page reference, and we
1602 * know the page is still mapped, we don't even
1603 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001604 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001605 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001606 unlock_page(page);
1607 }
1608 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001609unlock:
1610 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001612 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001614bad_page:
1615 pte_unmap_unlock(ptep, ptl);
1616 return ERR_PTR(-EFAULT);
1617
1618no_page:
1619 pte_unmap_unlock(ptep, ptl);
1620 if (!pte_none(pte))
1621 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001622
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001623no_page_table:
1624 /*
1625 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001626 * has touched so far, we don't want to allocate unnecessary pages or
1627 * page tables. Return error instead of NULL to skip handle_mm_fault,
1628 * then get_dump_page() will return NULL to leave a hole in the dump.
1629 * But we can only make this optimization where a hole would surely
1630 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001631 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001632 if ((flags & FOLL_DUMP) &&
1633 (!vma->vm_ops || !vma->vm_ops->fault))
1634 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001635 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636}
1637
Linus Torvalds95042f92011-04-12 14:15:51 -07001638static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1639{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001640 return stack_guard_page_start(vma, addr) ||
1641 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001642}
1643
Huang Ying0014bd92011-01-30 11:15:47 +08001644/**
1645 * __get_user_pages() - pin user pages in memory
1646 * @tsk: task_struct of target task
1647 * @mm: mm_struct of target mm
1648 * @start: starting user address
1649 * @nr_pages: number of pages from start to pin
1650 * @gup_flags: flags modifying pin behaviour
1651 * @pages: array that receives pointers to the pages pinned.
1652 * Should be at least nr_pages long. Or NULL, if caller
1653 * only intends to ensure the pages are faulted in.
1654 * @vmas: array of pointers to vmas corresponding to each page.
1655 * Or NULL if the caller does not require them.
1656 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1657 *
1658 * Returns number of pages pinned. This may be fewer than the number
1659 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1660 * were pinned, returns -errno. Each page returned must be released
1661 * with a put_page() call when it is finished with. vmas will only
1662 * remain valid while mmap_sem is held.
1663 *
1664 * Must be called with mmap_sem held for read or write.
1665 *
1666 * __get_user_pages walks a process's page tables and takes a reference to
1667 * each struct page that each user address corresponds to at a given
1668 * instant. That is, it takes the page that would be accessed if a user
1669 * thread accesses the given user virtual address at that instant.
1670 *
1671 * This does not guarantee that the page exists in the user mappings when
1672 * __get_user_pages returns, and there may even be a completely different
1673 * page there in some cases (eg. if mmapped pagecache has been invalidated
1674 * and subsequently re faulted). However it does guarantee that the page
1675 * won't be freed completely. And mostly callers simply care that the page
1676 * contains data that was valid *at some point in time*. Typically, an IO
1677 * or similar operation cannot guarantee anything stronger anyway because
1678 * locks can't be held over the syscall boundary.
1679 *
1680 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1681 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1682 * appropriate) must be called after the page is finished with, and
1683 * before put_page is called.
1684 *
1685 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1686 * or mmap_sem contention, and if waiting is needed to pin all pages,
1687 * *@nonblocking will be set to 0.
1688 *
1689 * In most cases, get_user_pages or get_user_pages_fast should be used
1690 * instead of __get_user_pages. __get_user_pages should be used only if
1691 * you need some special @gup_flags.
1692 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08001693long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1694 unsigned long start, unsigned long nr_pages,
1695 unsigned int gup_flags, struct page **pages,
1696 struct vm_area_struct **vmas, int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697{
Michel Lespinasse28a35712013-02-22 16:35:55 -08001698 long i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001699 unsigned long vm_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001700 unsigned int page_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
Michel Lespinasse28a35712013-02-22 16:35:55 -08001702 if (!nr_pages)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001703 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001704
1705 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1706
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 /*
1708 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001709 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001711 vm_flags = (gup_flags & FOLL_WRITE) ?
1712 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1713 vm_flags &= (gup_flags & FOLL_FORCE) ?
1714 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001715
1716 /*
1717 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1718 * would be called on PROT_NONE ranges. We must never invoke
1719 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1720 * page faults would unprotect the PROT_NONE ranges if
1721 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1722 * bitflag. So to avoid that, don't set FOLL_NUMA if
1723 * FOLL_FORCE is set.
1724 */
1725 if (!(gup_flags & FOLL_FORCE))
1726 gup_flags |= FOLL_NUMA;
1727
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 i = 0;
1729
1730 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001731 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
1733 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001734 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 pgd_t *pgd;
1737 pud_t *pud;
1738 pmd_t *pmd;
1739 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001740
1741 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001742 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 return i ? : -EFAULT;
1744 if (pg > TASK_SIZE)
1745 pgd = pgd_offset_k(pg);
1746 else
1747 pgd = pgd_offset_gate(mm, pg);
1748 BUG_ON(pgd_none(*pgd));
1749 pud = pud_offset(pgd, pg);
1750 BUG_ON(pud_none(*pud));
1751 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001752 if (pmd_none(*pmd))
1753 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001754 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001756 if (pte_none(*pte)) {
1757 pte_unmap(pte);
1758 return i ? : -EFAULT;
1759 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001760 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001762 struct page *page;
1763
Linus Torvalds95042f92011-04-12 14:15:51 -07001764 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001765 if (!page) {
1766 if (!(gup_flags & FOLL_DUMP) &&
1767 is_zero_pfn(pte_pfn(*pte)))
1768 page = pte_page(*pte);
1769 else {
1770 pte_unmap(pte);
1771 return i ? : -EFAULT;
1772 }
1773 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001774 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001775 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 }
1777 pte_unmap(pte);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001778 page_mask = 0;
Linus Torvalds95042f92011-04-12 14:15:51 -07001779 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 }
1781
Nick Pigginb291f002008-10-18 20:26:44 -07001782 if (!vma ||
1783 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001784 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 return i ? : -EFAULT;
1786
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001787 if (is_vm_hugetlb_page(vma)) {
1788 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001789 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001790 continue;
1791 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001792
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001794 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001795 unsigned int foll_flags = gup_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001796 unsigned int page_increm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797
Ethan Solomita462e00c2007-07-15 23:38:16 -07001798 /*
Ying Han47792802009-01-06 14:40:18 -08001799 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001800 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001801 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001802 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001803 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001804
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001805 cond_resched();
Michel Lespinasse240aade2013-02-22 16:35:56 -08001806 while (!(page = follow_page_mask(vma, start,
1807 foll_flags, &page_mask))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001808 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001809 unsigned int fault_flags = 0;
1810
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001811 /* For mlock, just skip the stack guard page. */
1812 if (foll_flags & FOLL_MLOCK) {
1813 if (stack_guard_page(vma, start))
1814 goto next_page;
1815 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001816 if (foll_flags & FOLL_WRITE)
1817 fault_flags |= FAULT_FLAG_WRITE;
1818 if (nonblocking)
1819 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001820 if (foll_flags & FOLL_NOWAIT)
1821 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001822
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001823 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001824 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001825
Nick Piggin83c54072007-07-19 01:47:05 -07001826 if (ret & VM_FAULT_ERROR) {
1827 if (ret & VM_FAULT_OOM)
1828 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001829 if (ret & (VM_FAULT_HWPOISON |
1830 VM_FAULT_HWPOISON_LARGE)) {
1831 if (i)
1832 return i;
1833 else if (gup_flags & FOLL_HWPOISON)
1834 return -EHWPOISON;
1835 else
1836 return -EFAULT;
1837 }
1838 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001839 return i ? i : -EFAULT;
1840 BUG();
1841 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001842
1843 if (tsk) {
1844 if (ret & VM_FAULT_MAJOR)
1845 tsk->maj_flt++;
1846 else
1847 tsk->min_flt++;
1848 }
Nick Piggin83c54072007-07-19 01:47:05 -07001849
Michel Lespinasse53a77062011-01-13 15:46:14 -08001850 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001851 if (nonblocking)
1852 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001853 return i;
1854 }
1855
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001856 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001857 * The VM_FAULT_WRITE bit tells us that
1858 * do_wp_page has broken COW when necessary,
1859 * even if maybe_mkwrite decided not to set
1860 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001861 * page lookups as if they were reads. But only
1862 * do so when looping for pte_write is futile:
1863 * in some cases userspace may also be wanting
1864 * to write to the gotten user page, which a
1865 * read fault here might prevent (a readonly
1866 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001867 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001868 if ((ret & VM_FAULT_WRITE) &&
1869 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001870 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001871
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001872 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001874 if (IS_ERR(page))
1875 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001877 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001878
Russell Kinga6f36be2006-12-30 22:24:19 +00001879 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001880 flush_dcache_page(page);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001881 page_mask = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001883next_page:
Michel Lespinasse240aade2013-02-22 16:35:56 -08001884 if (vmas) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 vmas[i] = vma;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001886 page_mask = 0;
1887 }
1888 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1889 if (page_increm > nr_pages)
1890 page_increm = nr_pages;
1891 i += page_increm;
1892 start += page_increm * PAGE_SIZE;
1893 nr_pages -= page_increm;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001894 } while (nr_pages && start < vma->vm_end);
1895 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 return i;
1897}
Huang Ying0014bd92011-01-30 11:15:47 +08001898EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001899
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001900/*
1901 * fixup_user_fault() - manually resolve a user page fault
1902 * @tsk: the task_struct to use for page fault accounting, or
1903 * NULL if faults are not to be recorded.
1904 * @mm: mm_struct of target mm
1905 * @address: user address
1906 * @fault_flags:flags to pass down to handle_mm_fault()
1907 *
1908 * This is meant to be called in the specific scenario where for locking reasons
1909 * we try to access user memory in atomic context (within a pagefault_disable()
1910 * section), this returns -EFAULT, and we want to resolve the user fault before
1911 * trying again.
1912 *
1913 * Typically this is meant to be used by the futex code.
1914 *
1915 * The main difference with get_user_pages() is that this function will
1916 * unconditionally call handle_mm_fault() which will in turn perform all the
1917 * necessary SW fixup of the dirty and young bits in the PTE, while
1918 * handle_mm_fault() only guarantees to update these in the struct page.
1919 *
1920 * This is important for some architectures where those bits also gate the
1921 * access permission to the page because they are maintained in software. On
1922 * such architectures, gup() will not be enough to make a subsequent access
1923 * succeed.
1924 *
1925 * This should be called with the mm_sem held for read.
1926 */
1927int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1928 unsigned long address, unsigned int fault_flags)
1929{
1930 struct vm_area_struct *vma;
1931 int ret;
1932
1933 vma = find_extend_vma(mm, address);
1934 if (!vma || address < vma->vm_start)
1935 return -EFAULT;
1936
1937 ret = handle_mm_fault(mm, vma, address, fault_flags);
1938 if (ret & VM_FAULT_ERROR) {
1939 if (ret & VM_FAULT_OOM)
1940 return -ENOMEM;
1941 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1942 return -EHWPOISON;
1943 if (ret & VM_FAULT_SIGBUS)
1944 return -EFAULT;
1945 BUG();
1946 }
1947 if (tsk) {
1948 if (ret & VM_FAULT_MAJOR)
1949 tsk->maj_flt++;
1950 else
1951 tsk->min_flt++;
1952 }
1953 return 0;
1954}
1955
1956/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001957 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001958 * @tsk: the task_struct to use for page fault accounting, or
1959 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001960 * @mm: mm_struct of target mm
1961 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001962 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001963 * @write: whether pages will be written to by the caller
1964 * @force: whether to force write access even if user mapping is
1965 * readonly. This will result in the page being COWed even
1966 * in MAP_SHARED mappings. You do not want this.
1967 * @pages: array that receives pointers to the pages pinned.
1968 * Should be at least nr_pages long. Or NULL, if caller
1969 * only intends to ensure the pages are faulted in.
1970 * @vmas: array of pointers to vmas corresponding to each page.
1971 * Or NULL if the caller does not require them.
1972 *
1973 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001974 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001975 * were pinned, returns -errno. Each page returned must be released
1976 * with a put_page() call when it is finished with. vmas will only
1977 * remain valid while mmap_sem is held.
1978 *
1979 * Must be called with mmap_sem held for read or write.
1980 *
1981 * get_user_pages walks a process's page tables and takes a reference to
1982 * each struct page that each user address corresponds to at a given
1983 * instant. That is, it takes the page that would be accessed if a user
1984 * thread accesses the given user virtual address at that instant.
1985 *
1986 * This does not guarantee that the page exists in the user mappings when
1987 * get_user_pages returns, and there may even be a completely different
1988 * page there in some cases (eg. if mmapped pagecache has been invalidated
1989 * and subsequently re faulted). However it does guarantee that the page
1990 * won't be freed completely. And mostly callers simply care that the page
1991 * contains data that was valid *at some point in time*. Typically, an IO
1992 * or similar operation cannot guarantee anything stronger anyway because
1993 * locks can't be held over the syscall boundary.
1994 *
1995 * If write=0, the page must not be written to. If the page is written to,
1996 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1997 * after the page is finished with, and before put_page is called.
1998 *
1999 * get_user_pages is typically used for fewer-copy IO operations, to get a
2000 * handle on the memory by some means other than accesses via the user virtual
2001 * addresses. The pages may be submitted for DMA to devices or accessed via
2002 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2003 * use the correct cache flushing APIs.
2004 *
2005 * See also get_user_pages_fast, for performance critical applications.
2006 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08002007long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2008 unsigned long start, unsigned long nr_pages, int write,
2009 int force, struct page **pages, struct vm_area_struct **vmas)
Nick Pigginb291f002008-10-18 20:26:44 -07002010{
Hugh Dickins58fa8792009-09-21 17:03:31 -07002011 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07002012
Hugh Dickins58fa8792009-09-21 17:03:31 -07002013 if (pages)
2014 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07002015 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002016 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07002017 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002018 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07002019
Michel Lespinasse53a77062011-01-13 15:46:14 -08002020 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2021 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07002022}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023EXPORT_SYMBOL(get_user_pages);
2024
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002025/**
2026 * get_dump_page() - pin user page in memory while writing it to core dump
2027 * @addr: user address
2028 *
2029 * Returns struct page pointer of user page pinned for dump,
2030 * to be freed afterwards by page_cache_release() or put_page().
2031 *
2032 * Returns NULL on any kind of failure - a hole must then be inserted into
2033 * the corefile, to preserve alignment with its headers; and also returns
2034 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2035 * allowing a hole to be left in the corefile to save diskspace.
2036 *
2037 * Called without mmap_sem, but after all other threads have been killed.
2038 */
2039#ifdef CONFIG_ELF_CORE
2040struct page *get_dump_page(unsigned long addr)
2041{
2042 struct vm_area_struct *vma;
2043 struct page *page;
2044
2045 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002046 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2047 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002048 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002049 flush_cache_page(vma, addr, page_to_pfn(page));
2050 return page;
2051}
2052#endif /* CONFIG_ELF_CORE */
2053
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002054pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002055 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002056{
2057 pgd_t * pgd = pgd_offset(mm, addr);
2058 pud_t * pud = pud_alloc(mm, pgd, addr);
2059 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002060 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002061 if (pmd) {
2062 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002063 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002064 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002065 }
2066 return NULL;
2067}
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002070 * This is the old fallback for page remapping.
2071 *
2072 * For historical reasons, it only allows reserved pages. Only
2073 * old drivers should use this, and they needed to mark their
2074 * pages reserved for the old functions anyway.
2075 */
Nick Piggin423bad62008-04-28 02:13:01 -07002076static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2077 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002078{
Nick Piggin423bad62008-04-28 02:13:01 -07002079 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002080 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002081 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002082 spinlock_t *ptl;
2083
Linus Torvalds238f58d2005-11-29 13:01:56 -08002084 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002085 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002086 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002087 retval = -ENOMEM;
2088 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002089 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002090 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002091 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002092 retval = -EBUSY;
2093 if (!pte_none(*pte))
2094 goto out_unlock;
2095
2096 /* Ok, finally just insert the thing.. */
2097 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002098 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002099 page_add_file_rmap(page);
2100 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2101
2102 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002103 pte_unmap_unlock(pte, ptl);
2104 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002105out_unlock:
2106 pte_unmap_unlock(pte, ptl);
2107out:
2108 return retval;
2109}
2110
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002111/**
2112 * vm_insert_page - insert single page into user vma
2113 * @vma: user vma to map to
2114 * @addr: target user address of this page
2115 * @page: source kernel page
2116 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002117 * This allows drivers to insert individual pages they've allocated
2118 * into a user vma.
2119 *
2120 * The page has to be a nice clean _individual_ kernel allocation.
2121 * If you allocate a compound page, you need to have marked it as
2122 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002123 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002124 *
2125 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2126 * took an arbitrary page protection parameter. This doesn't allow
2127 * that. Your vma protection will have to be set up correctly, which
2128 * means that if you want a shared writable mapping, you'd better
2129 * ask for a shared writable mapping!
2130 *
2131 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002132 *
2133 * Usually this function is called from f_op->mmap() handler
2134 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2135 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2136 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002137 */
Nick Piggin423bad62008-04-28 02:13:01 -07002138int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2139 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002140{
2141 if (addr < vma->vm_start || addr >= vma->vm_end)
2142 return -EFAULT;
2143 if (!page_count(page))
2144 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002145 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2146 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2147 BUG_ON(vma->vm_flags & VM_PFNMAP);
2148 vma->vm_flags |= VM_MIXEDMAP;
2149 }
Nick Piggin423bad62008-04-28 02:13:01 -07002150 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002151}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002152EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002153
Nick Piggin423bad62008-04-28 02:13:01 -07002154static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2155 unsigned long pfn, pgprot_t prot)
2156{
2157 struct mm_struct *mm = vma->vm_mm;
2158 int retval;
2159 pte_t *pte, entry;
2160 spinlock_t *ptl;
2161
2162 retval = -ENOMEM;
2163 pte = get_locked_pte(mm, addr, &ptl);
2164 if (!pte)
2165 goto out;
2166 retval = -EBUSY;
2167 if (!pte_none(*pte))
2168 goto out_unlock;
2169
2170 /* Ok, finally just insert the thing.. */
2171 entry = pte_mkspecial(pfn_pte(pfn, prot));
2172 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002173 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002174
2175 retval = 0;
2176out_unlock:
2177 pte_unmap_unlock(pte, ptl);
2178out:
2179 return retval;
2180}
2181
Nick Piggine0dc0d82007-02-12 00:51:36 -08002182/**
2183 * vm_insert_pfn - insert single pfn into user vma
2184 * @vma: user vma to map to
2185 * @addr: target user address of this page
2186 * @pfn: source kernel pfn
2187 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002188 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002189 * they've allocated into a user vma. Same comments apply.
2190 *
2191 * This function should only be called from a vm_ops->fault handler, and
2192 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002193 *
2194 * vma cannot be a COW mapping.
2195 *
2196 * As this is called only for pages that do not currently exist, we
2197 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002198 */
2199int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002200 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002201{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002202 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002203 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002204 /*
2205 * Technically, architectures with pte_special can avoid all these
2206 * restrictions (same for remap_pfn_range). However we would like
2207 * consistency in testing and feature parity among all, so we should
2208 * try to keep these invariants in place for everybody.
2209 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002210 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2211 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2212 (VM_PFNMAP|VM_MIXEDMAP));
2213 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2214 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002215
Nick Piggin423bad62008-04-28 02:13:01 -07002216 if (addr < vma->vm_start || addr >= vma->vm_end)
2217 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002218 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002219 return -EINVAL;
2220
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002221 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002222
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002223 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002224}
2225EXPORT_SYMBOL(vm_insert_pfn);
2226
Nick Piggin423bad62008-04-28 02:13:01 -07002227int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2228 unsigned long pfn)
2229{
2230 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2231
2232 if (addr < vma->vm_start || addr >= vma->vm_end)
2233 return -EFAULT;
2234
2235 /*
2236 * If we don't have pte special, then we have to use the pfn_valid()
2237 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2238 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002239 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2240 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002241 */
2242 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2243 struct page *page;
2244
2245 page = pfn_to_page(pfn);
2246 return insert_page(vma, addr, page, vma->vm_page_prot);
2247 }
2248 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2249}
2250EXPORT_SYMBOL(vm_insert_mixed);
2251
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002252/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 * maps a range of physical memory into the requested pages. the old
2254 * mappings are removed. any references to nonexistent pages results
2255 * in null mappings (currently treated as "copy-on-access")
2256 */
2257static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2258 unsigned long addr, unsigned long end,
2259 unsigned long pfn, pgprot_t prot)
2260{
2261 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002262 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
Hugh Dickinsc74df322005-10-29 18:16:23 -07002264 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 if (!pte)
2266 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002267 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 do {
2269 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002270 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 pfn++;
2272 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002273 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002274 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 return 0;
2276}
2277
2278static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2279 unsigned long addr, unsigned long end,
2280 unsigned long pfn, pgprot_t prot)
2281{
2282 pmd_t *pmd;
2283 unsigned long next;
2284
2285 pfn -= addr >> PAGE_SHIFT;
2286 pmd = pmd_alloc(mm, pud, addr);
2287 if (!pmd)
2288 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002289 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 do {
2291 next = pmd_addr_end(addr, end);
2292 if (remap_pte_range(mm, pmd, addr, next,
2293 pfn + (addr >> PAGE_SHIFT), prot))
2294 return -ENOMEM;
2295 } while (pmd++, addr = next, addr != end);
2296 return 0;
2297}
2298
2299static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2300 unsigned long addr, unsigned long end,
2301 unsigned long pfn, pgprot_t prot)
2302{
2303 pud_t *pud;
2304 unsigned long next;
2305
2306 pfn -= addr >> PAGE_SHIFT;
2307 pud = pud_alloc(mm, pgd, addr);
2308 if (!pud)
2309 return -ENOMEM;
2310 do {
2311 next = pud_addr_end(addr, end);
2312 if (remap_pmd_range(mm, pud, addr, next,
2313 pfn + (addr >> PAGE_SHIFT), prot))
2314 return -ENOMEM;
2315 } while (pud++, addr = next, addr != end);
2316 return 0;
2317}
2318
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002319/**
2320 * remap_pfn_range - remap kernel memory to userspace
2321 * @vma: user vma to map to
2322 * @addr: target user address to start at
2323 * @pfn: physical address of kernel memory
2324 * @size: size of map area
2325 * @prot: page protection flags for this mapping
2326 *
2327 * Note: this is only safe if the mm semaphore is held when called.
2328 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2330 unsigned long pfn, unsigned long size, pgprot_t prot)
2331{
2332 pgd_t *pgd;
2333 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002334 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 struct mm_struct *mm = vma->vm_mm;
2336 int err;
2337
2338 /*
2339 * Physically remapped pages are special. Tell the
2340 * rest of the world about it:
2341 * VM_IO tells people not to look at these pages
2342 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002343 * VM_PFNMAP tells the core MM that the base pages are just
2344 * raw PFN mappings, and do not have a "struct page" associated
2345 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002346 * VM_DONTEXPAND
2347 * Disable vma merging and expanding with mremap().
2348 * VM_DONTDUMP
2349 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002350 *
2351 * There's a horrible special case to handle copy-on-write
2352 * behaviour that some programs depend on. We mark the "original"
2353 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002354 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002356 if (is_cow_mapping(vma->vm_flags)) {
2357 if (addr != vma->vm_start || end != vma->vm_end)
2358 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002359 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002360 }
2361
2362 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2363 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002364 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002365
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002366 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
2368 BUG_ON(addr >= end);
2369 pfn -= addr >> PAGE_SHIFT;
2370 pgd = pgd_offset(mm, addr);
2371 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 do {
2373 next = pgd_addr_end(addr, end);
2374 err = remap_pud_range(mm, pgd, addr, next,
2375 pfn + (addr >> PAGE_SHIFT), prot);
2376 if (err)
2377 break;
2378 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002379
2380 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002381 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002382
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 return err;
2384}
2385EXPORT_SYMBOL(remap_pfn_range);
2386
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07002387/**
2388 * vm_iomap_memory - remap memory to userspace
2389 * @vma: user vma to map to
2390 * @start: start of area
2391 * @len: size of area
2392 *
2393 * This is a simplified io_remap_pfn_range() for common driver use. The
2394 * driver just needs to give us the physical memory range to be mapped,
2395 * we'll figure out the rest from the vma information.
2396 *
2397 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2398 * whatever write-combining details or similar.
2399 */
2400int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2401{
2402 unsigned long vm_len, pfn, pages;
2403
2404 /* Check that the physical memory area passed in looks valid */
2405 if (start + len < start)
2406 return -EINVAL;
2407 /*
2408 * You *really* shouldn't map things that aren't page-aligned,
2409 * but we've historically allowed it because IO memory might
2410 * just have smaller alignment.
2411 */
2412 len += start & ~PAGE_MASK;
2413 pfn = start >> PAGE_SHIFT;
2414 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2415 if (pfn + pages < pfn)
2416 return -EINVAL;
2417
2418 /* We start the mapping 'vm_pgoff' pages into the area */
2419 if (vma->vm_pgoff > pages)
2420 return -EINVAL;
2421 pfn += vma->vm_pgoff;
2422 pages -= vma->vm_pgoff;
2423
2424 /* Can we fit all of the mapping? */
2425 vm_len = vma->vm_end - vma->vm_start;
2426 if (vm_len >> PAGE_SHIFT > pages)
2427 return -EINVAL;
2428
2429 /* Ok, let it rip */
2430 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2431}
2432EXPORT_SYMBOL(vm_iomap_memory);
2433
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002434static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2435 unsigned long addr, unsigned long end,
2436 pte_fn_t fn, void *data)
2437{
2438 pte_t *pte;
2439 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002440 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002441 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002442
2443 pte = (mm == &init_mm) ?
2444 pte_alloc_kernel(pmd, addr) :
2445 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2446 if (!pte)
2447 return -ENOMEM;
2448
2449 BUG_ON(pmd_huge(*pmd));
2450
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002451 arch_enter_lazy_mmu_mode();
2452
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002453 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002454
2455 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002456 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002457 if (err)
2458 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002459 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002460
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002461 arch_leave_lazy_mmu_mode();
2462
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002463 if (mm != &init_mm)
2464 pte_unmap_unlock(pte-1, ptl);
2465 return err;
2466}
2467
2468static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2469 unsigned long addr, unsigned long end,
2470 pte_fn_t fn, void *data)
2471{
2472 pmd_t *pmd;
2473 unsigned long next;
2474 int err;
2475
Andi Kleenceb86872008-07-23 21:27:50 -07002476 BUG_ON(pud_huge(*pud));
2477
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002478 pmd = pmd_alloc(mm, pud, addr);
2479 if (!pmd)
2480 return -ENOMEM;
2481 do {
2482 next = pmd_addr_end(addr, end);
2483 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2484 if (err)
2485 break;
2486 } while (pmd++, addr = next, addr != end);
2487 return err;
2488}
2489
2490static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2491 unsigned long addr, unsigned long end,
2492 pte_fn_t fn, void *data)
2493{
2494 pud_t *pud;
2495 unsigned long next;
2496 int err;
2497
2498 pud = pud_alloc(mm, pgd, addr);
2499 if (!pud)
2500 return -ENOMEM;
2501 do {
2502 next = pud_addr_end(addr, end);
2503 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2504 if (err)
2505 break;
2506 } while (pud++, addr = next, addr != end);
2507 return err;
2508}
2509
2510/*
2511 * Scan a region of virtual memory, filling in page tables as necessary
2512 * and calling a provided function on each leaf page table.
2513 */
2514int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2515 unsigned long size, pte_fn_t fn, void *data)
2516{
2517 pgd_t *pgd;
2518 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002519 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002520 int err;
2521
2522 BUG_ON(addr >= end);
2523 pgd = pgd_offset(mm, addr);
2524 do {
2525 next = pgd_addr_end(addr, end);
2526 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2527 if (err)
2528 break;
2529 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002530
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002531 return err;
2532}
2533EXPORT_SYMBOL_GPL(apply_to_page_range);
2534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002536 * handle_pte_fault chooses page fault handler according to an entry
2537 * which was read non-atomically. Before making any commitment, on
2538 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002539 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002540 * must check under lock before unmapping the pte and proceeding
2541 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002542 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002543 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002544static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002545 pte_t *page_table, pte_t orig_pte)
2546{
2547 int same = 1;
2548#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2549 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002550 spinlock_t *ptl = pte_lockptr(mm, pmd);
2551 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002552 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002553 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002554 }
2555#endif
2556 pte_unmap(page_table);
2557 return same;
2558}
2559
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002560static 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 -08002561{
2562 /*
2563 * If the source page was a PFN mapping, we don't have
2564 * a "struct page" for it. We do a best-effort copy by
2565 * just copying from the original user address. If that
2566 * fails, we just zero-fill it. Live with it.
2567 */
2568 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002569 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002570 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2571
2572 /*
2573 * This really shouldn't fail, because the page is there
2574 * in the page tables. But it might just be unreadable,
2575 * in which case we just give up and fill the result with
2576 * zeroes.
2577 */
2578 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002579 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002580 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002581 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002582 } else
2583 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002584}
2585
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 * This routine handles present pages, when users try to write
2588 * to a shared page. It is done by copying the page to a new address
2589 * and decrementing the shared-page counter for the old page.
2590 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 * Note that this routine assumes that the protection checks have been
2592 * done by the caller (the low-level page fault routine in most cases).
2593 * Thus we can safely just mark it writable once we've done any necessary
2594 * COW.
2595 *
2596 * We also mark the page dirty at this point even though the page will
2597 * change only once the write actually happens. This avoids a few races,
2598 * and potentially makes it more efficient.
2599 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002600 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2601 * but allow concurrent faults), with pte both mapped and locked.
2602 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002604static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2605 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002606 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002607 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002609 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002611 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002612 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002613 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002614 unsigned long mmun_start = 0; /* For mmu_notifiers */
2615 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616
Linus Torvalds6aab3412005-11-28 14:34:23 -08002617 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002618 if (!old_page) {
2619 /*
2620 * VM_MIXEDMAP !pfn_valid() case
2621 *
2622 * We should not cow pages in a shared writeable mapping.
2623 * Just mark the pages writable as we can't do any dirty
2624 * accounting on raw pfn maps.
2625 */
2626 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2627 (VM_WRITE|VM_SHARED))
2628 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002629 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002630 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002632 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002633 * Take out anonymous pages first, anonymous shared vmas are
2634 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002635 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002636 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002637 if (!trylock_page(old_page)) {
2638 page_cache_get(old_page);
2639 pte_unmap_unlock(page_table, ptl);
2640 lock_page(old_page);
2641 page_table = pte_offset_map_lock(mm, pmd, address,
2642 &ptl);
2643 if (!pte_same(*page_table, orig_pte)) {
2644 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002645 goto unlock;
2646 }
2647 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002648 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002649 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002650 /*
2651 * The page is all ours. Move it to our anon_vma so
2652 * the rmap code will not search our parent or siblings.
2653 * Protected against the rmap code by the page lock.
2654 */
2655 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002656 unlock_page(old_page);
2657 goto reuse;
2658 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002659 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002660 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002661 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002662 /*
2663 * Only catch write-faults on shared writable pages,
2664 * read-only shared pages can get COWed by
2665 * get_user_pages(.write=1, .force=1).
2666 */
David Howells9637a5e2006-06-23 02:03:43 -07002667 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002668 struct vm_fault vmf;
2669 int tmp;
2670
2671 vmf.virtual_address = (void __user *)(address &
2672 PAGE_MASK);
2673 vmf.pgoff = old_page->index;
2674 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2675 vmf.page = old_page;
2676
David Howells9637a5e2006-06-23 02:03:43 -07002677 /*
2678 * Notify the address space that the page is about to
2679 * become writable so that it can prohibit this or wait
2680 * for the page to get into an appropriate state.
2681 *
2682 * We do this without the lock held, so that it can
2683 * sleep if it needs to.
2684 */
2685 page_cache_get(old_page);
2686 pte_unmap_unlock(page_table, ptl);
2687
Nick Pigginc2ec1752009-03-31 15:23:21 -07002688 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2689 if (unlikely(tmp &
2690 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2691 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002692 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002693 }
Nick Pigginb827e492009-04-30 15:08:16 -07002694 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2695 lock_page(old_page);
2696 if (!old_page->mapping) {
2697 ret = 0; /* retry the fault */
2698 unlock_page(old_page);
2699 goto unwritable_page;
2700 }
2701 } else
2702 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002703
David Howells9637a5e2006-06-23 02:03:43 -07002704 /*
2705 * Since we dropped the lock we need to revalidate
2706 * the PTE as someone else may have changed it. If
2707 * they did, we just return, as we can count on the
2708 * MMU to tell us if they didn't also make it writable.
2709 */
2710 page_table = pte_offset_map_lock(mm, pmd, address,
2711 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002712 if (!pte_same(*page_table, orig_pte)) {
2713 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002714 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002715 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002716
2717 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002719 dirty_page = old_page;
2720 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002721
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002722reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002723 flush_cache_page(vma, address, pte_pfn(orig_pte));
2724 entry = pte_mkyoung(orig_pte);
2725 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002726 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002727 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002728 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002729 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002730
2731 if (!dirty_page)
2732 return ret;
2733
2734 /*
2735 * Yes, Virginia, this is actually required to prevent a race
2736 * with clear_page_dirty_for_io() from clearing the page dirty
2737 * bit after it clear all dirty ptes, but before a racing
2738 * do_wp_page installs a dirty pte.
2739 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002740 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002741 */
2742 if (!page_mkwrite) {
2743 wait_on_page_locked(dirty_page);
2744 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002745 /* file_update_time outside page_lock */
2746 if (vma->vm_file)
2747 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002748 }
2749 put_page(dirty_page);
2750 if (page_mkwrite) {
2751 struct address_space *mapping = dirty_page->mapping;
2752
2753 set_page_dirty(dirty_page);
2754 unlock_page(dirty_page);
2755 page_cache_release(dirty_page);
2756 if (mapping) {
2757 /*
2758 * Some device drivers do not set page.mapping
2759 * but still dirty their pages
2760 */
2761 balance_dirty_pages_ratelimited(mapping);
2762 }
2763 }
2764
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002765 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
2768 /*
2769 * Ok, we need to copy. Oh, well..
2770 */
Nick Pigginb5810032005-10-29 18:16:12 -07002771 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002772gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002773 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774
2775 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002776 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002777
Hugh Dickins62eede62009-09-21 17:03:34 -07002778 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002779 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2780 if (!new_page)
2781 goto oom;
2782 } else {
2783 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2784 if (!new_page)
2785 goto oom;
2786 cow_user_page(new_page, old_page, address, vma);
2787 }
2788 __SetPageUptodate(new_page);
2789
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002790 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002791 goto oom_free_new;
2792
Haggai Eran6bdb9132012-10-08 16:33:35 -07002793 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002794 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002795 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2796
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 /*
2798 * Re-check the pte - we dropped the lock
2799 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002800 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002801 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002802 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002803 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002804 dec_mm_counter_fast(mm, MM_FILEPAGES);
2805 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002806 }
2807 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002808 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002809 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002810 entry = mk_pte(new_page, vma->vm_page_prot);
2811 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002812 /*
2813 * Clear the pte entry and flush it first, before updating the
2814 * pte with the new entry. This will avoid a race condition
2815 * seen in the presence of one thread doing SMC and another
2816 * thread doing COW.
2817 */
Izik Eidus828502d2009-09-21 17:01:51 -07002818 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002819 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002820 /*
2821 * We call the notify macro here because, when using secondary
2822 * mmu page tables (such as kvm shadow page tables), we want the
2823 * new page to be mapped directly into the secondary page table.
2824 */
2825 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002826 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002827 if (old_page) {
2828 /*
2829 * Only after switching the pte to the new page may
2830 * we remove the mapcount here. Otherwise another
2831 * process may come and find the rmap count decremented
2832 * before the pte is switched to the new page, and
2833 * "reuse" the old page writing into it while our pte
2834 * here still points into it and can be read by other
2835 * threads.
2836 *
2837 * The critical issue is to order this
2838 * page_remove_rmap with the ptp_clear_flush above.
2839 * Those stores are ordered by (if nothing else,)
2840 * the barrier present in the atomic_add_negative
2841 * in page_remove_rmap.
2842 *
2843 * Then the TLB flush in ptep_clear_flush ensures that
2844 * no process can access the old page before the
2845 * decremented mapcount is visible. And the old page
2846 * cannot be reused until after the decremented
2847 * mapcount is visible. So transitively, TLBs to
2848 * old page will be flushed before it can be reused.
2849 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002850 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002851 }
2852
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 /* Free the old page.. */
2854 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002855 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002856 } else
2857 mem_cgroup_uncharge_page(new_page);
2858
Haggai Eran6bdb9132012-10-08 16:33:35 -07002859 if (new_page)
2860 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002861unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002862 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002863 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002864 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002865 if (old_page) {
2866 /*
2867 * Don't let another task, with possibly unlocked vma,
2868 * keep the mlocked page.
2869 */
2870 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2871 lock_page(old_page); /* LRU manipulation */
2872 munlock_vma_page(old_page);
2873 unlock_page(old_page);
2874 }
2875 page_cache_release(old_page);
2876 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002877 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002878oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002879 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002880oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002881 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002882 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002884
2885unwritable_page:
2886 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002887 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888}
2889
Peter Zijlstra97a89412011-05-24 17:12:04 -07002890static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 unsigned long start_addr, unsigned long end_addr,
2892 struct zap_details *details)
2893{
Al Virof5cc4ee2012-03-05 14:14:20 -05002894 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895}
2896
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002897static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 struct zap_details *details)
2899{
2900 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 pgoff_t vba, vea, zba, zea;
2902
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002903 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
2906 vba = vma->vm_pgoff;
2907 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2908 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2909 zba = details->first_index;
2910 if (zba < vba)
2911 zba = vba;
2912 zea = details->last_index;
2913 if (zea > vea)
2914 zea = vea;
2915
Peter Zijlstra97a89412011-05-24 17:12:04 -07002916 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2918 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002919 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 }
2921}
2922
2923static inline void unmap_mapping_range_list(struct list_head *head,
2924 struct zap_details *details)
2925{
2926 struct vm_area_struct *vma;
2927
2928 /*
2929 * In nonlinear VMAs there is no correspondence between virtual address
2930 * offset and file offset. So we must perform an exhaustive search
2931 * across *all* the pages in each nonlinear VMA, not just the pages
2932 * whose virtual address lies outside the file truncation point.
2933 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002934 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002936 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 }
2938}
2939
2940/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002941 * 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 -07002942 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 * @holebegin: byte in first page to unmap, relative to the start of
2944 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002945 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 * must keep the partial page. In contrast, we must get rid of
2947 * partial pages.
2948 * @holelen: size of prospective hole in bytes. This will be rounded
2949 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2950 * end of the file.
2951 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2952 * but 0 when invalidating pagecache, don't throw away private data.
2953 */
2954void unmap_mapping_range(struct address_space *mapping,
2955 loff_t const holebegin, loff_t const holelen, int even_cows)
2956{
2957 struct zap_details details;
2958 pgoff_t hba = holebegin >> PAGE_SHIFT;
2959 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2960
2961 /* Check for overflow. */
2962 if (sizeof(holelen) > sizeof(hlen)) {
2963 long long holeend =
2964 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2965 if (holeend & ~(long long)ULONG_MAX)
2966 hlen = ULONG_MAX - hba + 1;
2967 }
2968
2969 details.check_mapping = even_cows? NULL: mapping;
2970 details.nonlinear_vma = NULL;
2971 details.first_index = hba;
2972 details.last_index = hba + hlen - 1;
2973 if (details.last_index < details.first_index)
2974 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975
Peter Zijlstra97a89412011-05-24 17:12:04 -07002976
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002977 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002978 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2980 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2981 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002982 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983}
2984EXPORT_SYMBOL(unmap_mapping_range);
2985
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002987 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2988 * but allow concurrent faults), and pte mapped but not yet locked.
2989 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002991static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2992 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002993 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002995 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002996 struct page *page, *swapcache;
Hugh Dickins65500d22005-10-29 18:15:59 -07002997 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002999 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07003000 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003001 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07003002 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003004 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003005 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07003006
3007 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02003008 if (unlikely(non_swap_entry(entry))) {
3009 if (is_migration_entry(entry)) {
3010 migration_entry_wait(mm, pmd, address);
3011 } else if (is_hwpoison_entry(entry)) {
3012 ret = VM_FAULT_HWPOISON;
3013 } else {
3014 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003015 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02003016 }
Christoph Lameter06972122006-06-23 02:03:35 -07003017 goto out;
3018 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003019 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 page = lookup_swap_cache(entry);
3021 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08003022 page = swapin_readahead(entry,
3023 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 if (!page) {
3025 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003026 * Back out if somebody else faulted in this pte
3027 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003029 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 if (likely(pte_same(*page_table, orig_pte)))
3031 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003032 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07003033 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 }
3035
3036 /* Had to read the page from swap area: Major fault */
3037 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07003038 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003039 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02003040 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01003041 /*
3042 * hwpoisoned dirty swapcache pages are kept for killing
3043 * owner processes (which may be unknown at hwpoison time)
3044 */
Andi Kleend1737fd2009-09-16 11:50:06 +02003045 ret = VM_FAULT_HWPOISON;
3046 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08003047 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02003048 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 }
3050
Hugh Dickins56f31802013-02-22 16:36:10 -08003051 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003052 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07003053
Balbir Singh20a10222007-11-14 17:00:33 -08003054 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003055 if (!locked) {
3056 ret |= VM_FAULT_RETRY;
3057 goto out_release;
3058 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003060 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003061 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3062 * release the swapcache from under us. The page pin, and pte_same
3063 * test below, are not enough to exclude that. Even if it is still
3064 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003065 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003066 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003067 goto out_page;
3068
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003069 page = ksm_might_need_to_copy(page, vma, address);
3070 if (unlikely(!page)) {
3071 ret = VM_FAULT_OOM;
3072 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003073 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08003074 }
3075
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003076 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003077 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003078 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003079 }
3080
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003082 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003084 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003085 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003086 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003087
3088 if (unlikely(!PageUptodate(page))) {
3089 ret = VM_FAULT_SIGBUS;
3090 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 }
3092
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003093 /*
3094 * The page isn't present yet, go ahead with the fault.
3095 *
3096 * Be careful about the sequence of operations here.
3097 * To get its accounting right, reuse_swap_page() must be called
3098 * while the page is counted on swap but not yet in mapcount i.e.
3099 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3100 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003101 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3102 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3103 * in page->private. In this case, a record in swap_cgroup is silently
3104 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003105 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003107 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003108 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003110 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003112 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003113 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003114 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 flush_icache_page(vma, page);
3117 set_pte_at(mm, address, page_table, pte);
Hugh Dickins56f31802013-02-22 16:36:10 -08003118 if (page == swapcache)
Johannes Weineraf347702013-02-22 16:32:20 -08003119 do_page_add_anon_rmap(page, vma, address, exclusive);
Hugh Dickins56f31802013-02-22 16:36:10 -08003120 else /* ksm created a completely new copy */
3121 page_add_new_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003122 /* It's better to call commit-charge after rmap is established */
3123 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003125 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003126 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003127 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003128 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003129 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003130 /*
3131 * Hold the lock to avoid the swap entry to be reused
3132 * until we take the PT lock for the pte_same() check
3133 * (to avoid false positives from pte_same). For
3134 * further safety release the lock after the swap_free
3135 * so that the swap count won't change under a
3136 * parallel locked swapcache.
3137 */
3138 unlock_page(swapcache);
3139 page_cache_release(swapcache);
3140 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003141
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003142 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003143 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3144 if (ret & VM_FAULT_ERROR)
3145 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 goto out;
3147 }
3148
3149 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003150 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003151unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003152 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153out:
3154 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003155out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003156 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003157 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003158out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003159 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003160out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003161 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003162 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003163 unlock_page(swapcache);
3164 page_cache_release(swapcache);
3165 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003166 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167}
3168
3169/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003170 * This is like a special single-page "expand_{down|up}wards()",
3171 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003172 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003173 */
3174static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3175{
3176 address &= PAGE_MASK;
3177 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003178 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003179
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003180 /*
3181 * Is there a mapping abutting this one below?
3182 *
3183 * That's only ok if it's the same stack mapping
3184 * that has gotten split..
3185 */
3186 if (prev && prev->vm_end == address)
3187 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3188
Michal Hockod05f3162011-05-24 17:11:44 -07003189 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003190 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003191 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3192 struct vm_area_struct *next = vma->vm_next;
3193
3194 /* As VM_GROWSDOWN but s/below/above/ */
3195 if (next && next->vm_start == address + PAGE_SIZE)
3196 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3197
3198 expand_upwards(vma, address + PAGE_SIZE);
3199 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003200 return 0;
3201}
3202
3203/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003204 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3205 * but allow concurrent faults), and pte mapped but not yet locked.
3206 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003208static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3209 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003210 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003212 struct page *page;
3213 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
Linus Torvalds11ac5522010-08-14 11:44:56 -07003216 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003217
Linus Torvalds11ac5522010-08-14 11:44:56 -07003218 /* Check if we need to add a guard page to the stack */
3219 if (check_stack_guard_page(vma, address) < 0)
3220 return VM_FAULT_SIGBUS;
3221
3222 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003223 if (!(flags & FAULT_FLAG_WRITE)) {
3224 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3225 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003226 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003227 if (!pte_none(*page_table))
3228 goto unlock;
3229 goto setpte;
3230 }
3231
Nick Piggin557ed1f2007-10-16 01:24:40 -07003232 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003233 if (unlikely(anon_vma_prepare(vma)))
3234 goto oom;
3235 page = alloc_zeroed_user_highpage_movable(vma, address);
3236 if (!page)
3237 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07003238 /*
3239 * The memory barrier inside __SetPageUptodate makes sure that
3240 * preceeding stores to the page contents become visible before
3241 * the set_pte_at() write.
3242 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08003243 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003245 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003246 goto oom_free_page;
3247
Nick Piggin557ed1f2007-10-16 01:24:40 -07003248 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003249 if (vma->vm_flags & VM_WRITE)
3250 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003251
Nick Piggin557ed1f2007-10-16 01:24:40 -07003252 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003253 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003254 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003255
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003256 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003257 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003258setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003259 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260
3261 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003262 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003263unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003264 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003265 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003266release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003267 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003268 page_cache_release(page);
3269 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003270oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003271 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003272oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 return VM_FAULT_OOM;
3274}
3275
3276/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003277 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003279 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3280 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 *
3282 * As this is called only for pages that do not currently exist, we
3283 * do not need to flush old virtual caches or the TLB.
3284 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003285 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003286 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003287 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003289static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003290 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003291 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003293 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003294 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003295 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003296 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003299 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003300 struct vm_fault vmf;
3301 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003302 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003303
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003304 /*
3305 * If we do COW later, allocate page befor taking lock_page()
3306 * on the file cache page. This will reduce lock holding time.
3307 */
3308 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3309
3310 if (unlikely(anon_vma_prepare(vma)))
3311 return VM_FAULT_OOM;
3312
3313 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3314 if (!cow_page)
3315 return VM_FAULT_OOM;
3316
3317 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3318 page_cache_release(cow_page);
3319 return VM_FAULT_OOM;
3320 }
3321 } else
3322 cow_page = NULL;
3323
Nick Piggind0217ac2007-07-19 01:47:03 -07003324 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3325 vmf.pgoff = pgoff;
3326 vmf.flags = flags;
3327 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003329 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003330 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3331 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003332 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333
Andi Kleena3b947e2009-09-16 11:50:08 +02003334 if (unlikely(PageHWPoison(vmf.page))) {
3335 if (ret & VM_FAULT_LOCKED)
3336 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003337 ret = VM_FAULT_HWPOISON;
3338 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003339 }
3340
Nick Piggind00806b2007-07-19 01:46:57 -07003341 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003342 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003343 * locked.
3344 */
Nick Piggin83c54072007-07-19 01:47:05 -07003345 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003346 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003347 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003348 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003349
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350 /*
3351 * Should we do an early C-O-W break?
3352 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003353 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003354 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003355 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003356 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003357 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003358 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003359 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003360 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003361 /*
3362 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003363 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003364 * to become writable
3365 */
Mark Fasheh69676142007-07-19 01:47:00 -07003366 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003367 int tmp;
3368
Mark Fasheh69676142007-07-19 01:47:00 -07003369 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003370 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003371 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3372 if (unlikely(tmp &
3373 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3374 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003375 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003376 }
Nick Pigginb827e492009-04-30 15:08:16 -07003377 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3378 lock_page(page);
3379 if (!page->mapping) {
3380 ret = 0; /* retry the fault */
3381 unlock_page(page);
3382 goto unwritable_page;
3383 }
3384 } else
3385 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003386 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003387 }
3388 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003389
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 }
3391
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003392 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393
3394 /*
3395 * This silly early PAGE_DIRTY setting removes a race
3396 * due to the bad i386 page protection. But it's valid
3397 * for other architectures too.
3398 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003399 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 * an exclusive copy of the page, or this is a shared mapping,
3401 * so we can make it writable and dirty to avoid having to
3402 * handle that later.
3403 */
3404 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003405 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003406 flush_icache_page(vma, page);
3407 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003408 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003411 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003412 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003413 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003414 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003415 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003416 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003417 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003418 get_page(dirty_page);
3419 }
Hugh Dickins42946212005-10-29 18:16:05 -07003420 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003421 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003422
3423 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003424 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003426 if (cow_page)
3427 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003428 if (anon)
3429 page_cache_release(page);
3430 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003431 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 }
3433
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003434 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003435
Nick Pigginb827e492009-04-30 15:08:16 -07003436 if (dirty_page) {
3437 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003438 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003439
3440 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003441 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003442 unlock_page(dirty_page);
3443 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003444 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003445 /*
3446 * Some device drivers do not set page.mapping but still
3447 * dirty their pages
3448 */
3449 balance_dirty_pages_ratelimited(mapping);
3450 }
3451
3452 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003453 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003454 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003455 } else {
3456 unlock_page(vmf.page);
3457 if (anon)
3458 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003459 }
Nick Piggind00806b2007-07-19 01:46:57 -07003460
Nick Piggin83c54072007-07-19 01:47:05 -07003461 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003462
3463unwritable_page:
3464 page_cache_release(page);
3465 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003466uncharge_out:
3467 /* fs's fault handler get error */
3468 if (cow_page) {
3469 mem_cgroup_uncharge_page(cow_page);
3470 page_cache_release(cow_page);
3471 }
3472 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003473}
Nick Piggind00806b2007-07-19 01:46:57 -07003474
Nick Piggin54cb8822007-07-19 01:46:59 -07003475static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3476 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003477 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003478{
3479 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003480 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003481
Hugh Dickins16abfa02007-10-04 16:56:06 +01003482 pte_unmap(page_table);
3483 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003484}
3485
Jes Sorensenf4b81802006-09-27 01:50:10 -07003486/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 * Fault of a previously existing named mapping. Repopulate the pte
3488 * from the encoded file_pte if possible. This enables swappable
3489 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003490 *
3491 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3492 * but allow concurrent faults), and pte mapped but not yet locked.
3493 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003495static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003496 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003497 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498{
Hugh Dickins65500d22005-10-29 18:15:59 -07003499 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003501 flags |= FAULT_FLAG_NONLINEAR;
3502
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003503 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003504 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505
Hugh Dickins2509ef22009-01-06 14:40:10 -08003506 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003507 /*
3508 * Page table corrupted: show pte and kill process.
3509 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003510 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003511 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003512 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003513
3514 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003515 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516}
3517
Mel Gorman9532fec2012-11-15 01:24:32 +00003518int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3519 unsigned long addr, int current_nid)
3520{
3521 get_page(page);
3522
3523 count_vm_numa_event(NUMA_HINT_FAULTS);
3524 if (current_nid == numa_node_id())
3525 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3526
3527 return mpol_misplaced(page, vma, addr);
3528}
3529
Mel Gormand10e63f2012-10-25 14:16:31 +02003530int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3531 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3532{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003533 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003534 spinlock_t *ptl;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003535 int current_nid = -1;
3536 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003537 bool migrated = false;
Mel Gormand10e63f2012-10-25 14:16:31 +02003538
3539 /*
3540 * The "pte" at this point cannot be used safely without
3541 * validation through pte_unmap_same(). It's of NUMA type but
3542 * the pfn may be screwed if the read is non atomic.
3543 *
3544 * ptep_modify_prot_start is not called as this is clearing
3545 * the _PAGE_NUMA bit and it is not really expected that there
3546 * would be concurrent hardware modifications to the PTE.
3547 */
3548 ptl = pte_lockptr(mm, pmd);
3549 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003550 if (unlikely(!pte_same(*ptep, pte))) {
3551 pte_unmap_unlock(ptep, ptl);
3552 goto out;
3553 }
3554
Mel Gormand10e63f2012-10-25 14:16:31 +02003555 pte = pte_mknonnuma(pte);
3556 set_pte_at(mm, addr, ptep, pte);
3557 update_mmu_cache(vma, addr, ptep);
3558
3559 page = vm_normal_page(vma, addr, pte);
3560 if (!page) {
3561 pte_unmap_unlock(ptep, ptl);
3562 return 0;
3563 }
3564
Mel Gorman4daae3b2012-11-02 11:33:45 +00003565 current_nid = page_to_nid(page);
Mel Gorman9532fec2012-11-15 01:24:32 +00003566 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
Mel Gormand10e63f2012-10-25 14:16:31 +02003567 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003568 if (target_nid == -1) {
3569 /*
3570 * Account for the fault against the current node if it not
3571 * being replaced regardless of where the page is located.
3572 */
3573 current_nid = numa_node_id();
3574 put_page(page);
3575 goto out;
3576 }
3577
3578 /* Migrate to the requested node */
Mel Gormanb8593bf2012-11-21 01:18:23 +00003579 migrated = migrate_misplaced_page(page, target_nid);
3580 if (migrated)
Mel Gorman4daae3b2012-11-02 11:33:45 +00003581 current_nid = target_nid;
3582
3583out:
Mel Gorman9532fec2012-11-15 01:24:32 +00003584 if (current_nid != -1)
Mel Gormanb8593bf2012-11-21 01:18:23 +00003585 task_numa_fault(current_nid, 1, migrated);
Mel Gormand10e63f2012-10-25 14:16:31 +02003586 return 0;
3587}
3588
3589/* NUMA hinting page fault entry point for regular pmds */
3590#ifdef CONFIG_NUMA_BALANCING
3591static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3592 unsigned long addr, pmd_t *pmdp)
3593{
3594 pmd_t pmd;
3595 pte_t *pte, *orig_pte;
3596 unsigned long _addr = addr & PMD_MASK;
3597 unsigned long offset;
3598 spinlock_t *ptl;
3599 bool numa = false;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00003600 int local_nid = numa_node_id();
Mel Gormand10e63f2012-10-25 14:16:31 +02003601
3602 spin_lock(&mm->page_table_lock);
3603 pmd = *pmdp;
3604 if (pmd_numa(pmd)) {
3605 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3606 numa = true;
3607 }
3608 spin_unlock(&mm->page_table_lock);
3609
3610 if (!numa)
3611 return 0;
3612
3613 /* we're in a page fault so some vma must be in the range */
3614 BUG_ON(!vma);
3615 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3616 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3617 VM_BUG_ON(offset >= PMD_SIZE);
3618 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3619 pte += offset >> PAGE_SHIFT;
3620 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3621 pte_t pteval = *pte;
3622 struct page *page;
Mel Gorman9532fec2012-11-15 01:24:32 +00003623 int curr_nid = local_nid;
3624 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003625 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02003626 if (!pte_present(pteval))
3627 continue;
3628 if (!pte_numa(pteval))
3629 continue;
3630 if (addr >= vma->vm_end) {
3631 vma = find_vma(mm, addr);
3632 /* there's a pte present so there must be a vma */
3633 BUG_ON(!vma);
3634 BUG_ON(addr < vma->vm_start);
3635 }
3636 if (pte_numa(pteval)) {
3637 pteval = pte_mknonnuma(pteval);
3638 set_pte_at(mm, addr, pte, pteval);
3639 }
3640 page = vm_normal_page(vma, addr, pteval);
3641 if (unlikely(!page))
3642 continue;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003643 /* only check non-shared pages */
3644 if (unlikely(page_mapcount(page) != 1))
3645 continue;
Mel Gorman9532fec2012-11-15 01:24:32 +00003646
3647 /*
3648 * Note that the NUMA fault is later accounted to either
3649 * the node that is currently running or where the page is
3650 * migrated to.
3651 */
3652 curr_nid = local_nid;
3653 target_nid = numa_migrate_prep(page, vma, addr,
3654 page_to_nid(page));
3655 if (target_nid == -1) {
3656 put_page(page);
3657 continue;
3658 }
3659
3660 /* Migrate to the requested node */
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003661 pte_unmap_unlock(pte, ptl);
Mel Gormanb8593bf2012-11-21 01:18:23 +00003662 migrated = migrate_misplaced_page(page, target_nid);
3663 if (migrated)
Mel Gorman9532fec2012-11-15 01:24:32 +00003664 curr_nid = target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003665 task_numa_fault(curr_nid, 1, migrated);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003666
3667 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003668 }
3669 pte_unmap_unlock(orig_pte, ptl);
3670
3671 return 0;
3672}
3673#else
3674static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long addr, pmd_t *pmdp)
3676{
3677 BUG();
Andrew Mortonb3dd2072012-12-17 15:59:24 -08003678 return 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003679}
3680#endif /* CONFIG_NUMA_BALANCING */
3681
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682/*
3683 * These routines also need to handle stuff like marking pages dirty
3684 * and/or accessed for architectures that don't do it in hardware (most
3685 * RISC architectures). The early dirtying is also good on the i386.
3686 *
3687 * There is also a hook called "update_mmu_cache()" that architectures
3688 * with external mmu caches can use to update those (ie the Sparc or
3689 * PowerPC hashed page tables that act as extended TLBs).
3690 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003691 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3692 * but allow concurrent faults), and pte mapped but not yet locked.
3693 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003695int handle_pte_fault(struct mm_struct *mm,
3696 struct vm_area_struct *vma, unsigned long address,
3697 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698{
3699 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003700 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003702 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003704 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003705 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003706 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003707 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003708 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003709 }
3710 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003711 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003712 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003714 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003715 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003716 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003717 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 }
3719
Mel Gormand10e63f2012-10-25 14:16:31 +02003720 if (pte_numa(entry))
3721 return do_numa_page(mm, vma, address, entry, pte, pmd);
3722
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003723 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003724 spin_lock(ptl);
3725 if (unlikely(!pte_same(*pte, entry)))
3726 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003727 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003729 return do_wp_page(mm, vma, address,
3730 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 entry = pte_mkdirty(entry);
3732 }
3733 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003734 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003735 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003736 } else {
3737 /*
3738 * This is needed only for protection faults but the arch code
3739 * is not yet telling us if this is a protection fault or not.
3740 * This still avoids useless tlb flushes for .text page faults
3741 * with threads.
3742 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003743 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003744 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003745 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003746unlock:
3747 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003748 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749}
3750
3751/*
3752 * By the time we get here, we already hold the mm semaphore
3753 */
Nick Piggin83c54072007-07-19 01:47:05 -07003754int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003755 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756{
3757 pgd_t *pgd;
3758 pud_t *pud;
3759 pmd_t *pmd;
3760 pte_t *pte;
3761
3762 __set_current_state(TASK_RUNNING);
3763
Christoph Lameterf8891e52006-06-30 01:55:45 -07003764 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003765 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003767 /* do counter updates before entering really critical section. */
3768 check_sync_rss_stat(current);
3769
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003770 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003771 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772
David Rientjes1f1d06c2012-05-29 15:06:23 -07003773retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003774 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 pud = pud_alloc(mm, pgd, address);
3776 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003777 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 pmd = pmd_alloc(mm, pud, address);
3779 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003780 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003781 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3782 if (!vma->vm_ops)
3783 return do_huge_pmd_anonymous_page(mm, vma, address,
3784 pmd, flags);
3785 } else {
3786 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003787 int ret;
3788
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003789 barrier();
3790 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003791 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3792
Linus Torvaldse53289c2013-01-09 08:36:54 -08003793 /*
3794 * If the pmd is splitting, return and retry the
3795 * the fault. Alternative: wait until the split
3796 * is done, and goto retry.
3797 */
3798 if (pmd_trans_splitting(orig_pmd))
3799 return 0;
3800
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003801 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003802 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003803 orig_pmd, pmd);
3804
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003805 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003806 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3807 orig_pmd);
3808 /*
3809 * If COW results in an oom, the huge pmd will
3810 * have been split, so retry the fault on the
3811 * pte for a smaller charge.
3812 */
3813 if (unlikely(ret & VM_FAULT_OOM))
3814 goto retry;
3815 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003816 } else {
3817 huge_pmd_set_accessed(mm, vma, address, pmd,
3818 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003819 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003820
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003821 return 0;
3822 }
3823 }
3824
Mel Gormand10e63f2012-10-25 14:16:31 +02003825 if (pmd_numa(*pmd))
3826 return do_pmd_numa_page(mm, vma, address, pmd);
3827
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003828 /*
3829 * Use __pte_alloc instead of pte_alloc_map, because we can't
3830 * run pte_offset_map on the pmd, if an huge pmd could
3831 * materialize from under us from a different thread.
3832 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003833 if (unlikely(pmd_none(*pmd)) &&
3834 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003835 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003836 /* if an huge pmd materialized from under us just retry later */
3837 if (unlikely(pmd_trans_huge(*pmd)))
3838 return 0;
3839 /*
3840 * A regular pmd is established and it can't morph into a huge pmd
3841 * from under us anymore at this point because we hold the mmap_sem
3842 * read mode and khugepaged takes it in write mode. So now it's
3843 * safe to run pte_offset_map().
3844 */
3845 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003847 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848}
3849
3850#ifndef __PAGETABLE_PUD_FOLDED
3851/*
3852 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003853 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003855int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003857 pud_t *new = pud_alloc_one(mm, address);
3858 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003859 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860
Nick Piggin362a61a2008-05-14 06:37:36 +02003861 smp_wmb(); /* See comment in __pte_alloc */
3862
Hugh Dickins872fec12005-10-29 18:16:21 -07003863 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003864 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003865 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003866 else
3867 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003868 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003869 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870}
3871#endif /* __PAGETABLE_PUD_FOLDED */
3872
3873#ifndef __PAGETABLE_PMD_FOLDED
3874/*
3875 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003876 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003878int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003880 pmd_t *new = pmd_alloc_one(mm, address);
3881 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003882 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883
Nick Piggin362a61a2008-05-14 06:37:36 +02003884 smp_wmb(); /* See comment in __pte_alloc */
3885
Hugh Dickins872fec12005-10-29 18:16:21 -07003886 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003888 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003889 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003890 else
3891 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003893 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003894 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003895 else
3896 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003898 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003899 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900}
3901#endif /* __PAGETABLE_PMD_FOLDED */
3902
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903#if !defined(__HAVE_ARCH_GATE_AREA)
3904
3905#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003906static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907
3908static int __init gate_vma_init(void)
3909{
3910 gate_vma.vm_mm = NULL;
3911 gate_vma.vm_start = FIXADDR_USER_START;
3912 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003913 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3914 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003915
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 return 0;
3917}
3918__initcall(gate_vma_init);
3919#endif
3920
Stephen Wilson31db58b2011-03-13 15:49:15 -04003921struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922{
3923#ifdef AT_SYSINFO_EHDR
3924 return &gate_vma;
3925#else
3926 return NULL;
3927#endif
3928}
3929
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003930int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931{
3932#ifdef AT_SYSINFO_EHDR
3933 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3934 return 1;
3935#endif
3936 return 0;
3937}
3938
3939#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003940
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003941static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003942 pte_t **ptepp, spinlock_t **ptlp)
3943{
3944 pgd_t *pgd;
3945 pud_t *pud;
3946 pmd_t *pmd;
3947 pte_t *ptep;
3948
3949 pgd = pgd_offset(mm, address);
3950 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3951 goto out;
3952
3953 pud = pud_offset(pgd, address);
3954 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3955 goto out;
3956
3957 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003958 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003959 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3960 goto out;
3961
3962 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3963 if (pmd_huge(*pmd))
3964 goto out;
3965
3966 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3967 if (!ptep)
3968 goto out;
3969 if (!pte_present(*ptep))
3970 goto unlock;
3971 *ptepp = ptep;
3972 return 0;
3973unlock:
3974 pte_unmap_unlock(ptep, *ptlp);
3975out:
3976 return -EINVAL;
3977}
3978
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003979static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3980 pte_t **ptepp, spinlock_t **ptlp)
3981{
3982 int res;
3983
3984 /* (void) is needed to make gcc happy */
3985 (void) __cond_lock(*ptlp,
3986 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3987 return res;
3988}
3989
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003990/**
3991 * follow_pfn - look up PFN at a user virtual address
3992 * @vma: memory mapping
3993 * @address: user virtual address
3994 * @pfn: location to store found PFN
3995 *
3996 * Only IO mappings and raw PFN mappings are allowed.
3997 *
3998 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3999 */
4000int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4001 unsigned long *pfn)
4002{
4003 int ret = -EINVAL;
4004 spinlock_t *ptl;
4005 pte_t *ptep;
4006
4007 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4008 return ret;
4009
4010 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4011 if (ret)
4012 return ret;
4013 *pfn = pte_pfn(*ptep);
4014 pte_unmap_unlock(ptep, ptl);
4015 return 0;
4016}
4017EXPORT_SYMBOL(follow_pfn);
4018
Rik van Riel28b2ee22008-07-23 21:27:05 -07004019#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004020int follow_phys(struct vm_area_struct *vma,
4021 unsigned long address, unsigned int flags,
4022 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004023{
Johannes Weiner03668a42009-06-16 15:32:34 -07004024 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004025 pte_t *ptep, pte;
4026 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004027
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004028 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4029 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004030
Johannes Weiner03668a42009-06-16 15:32:34 -07004031 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004032 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004033 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07004034
Rik van Riel28b2ee22008-07-23 21:27:05 -07004035 if ((flags & FOLL_WRITE) && !pte_write(pte))
4036 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004037
4038 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07004039 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004040
Johannes Weiner03668a42009-06-16 15:32:34 -07004041 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004042unlock:
4043 pte_unmap_unlock(ptep, ptl);
4044out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004045 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004046}
4047
4048int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4049 void *buf, int len, int write)
4050{
4051 resource_size_t phys_addr;
4052 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08004053 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004054 int offset = addr & (PAGE_SIZE-1);
4055
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004056 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004057 return -EINVAL;
4058
4059 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4060 if (write)
4061 memcpy_toio(maddr + offset, buf, len);
4062 else
4063 memcpy_fromio(buf, maddr + offset, len);
4064 iounmap(maddr);
4065
4066 return len;
4067}
4068#endif
4069
David Howells0ec76a12006-09-27 01:50:15 -07004070/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004071 * Access another process' address space as given in mm. If non-NULL, use the
4072 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004073 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004074static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4075 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004076{
David Howells0ec76a12006-09-27 01:50:15 -07004077 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004078 void *old_buf = buf;
4079
David Howells0ec76a12006-09-27 01:50:15 -07004080 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004081 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004082 while (len) {
4083 int bytes, ret, offset;
4084 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004085 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004086
4087 ret = get_user_pages(tsk, mm, addr, 1,
4088 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004089 if (ret <= 0) {
4090 /*
4091 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4092 * we can access using slightly different code.
4093 */
4094#ifdef CONFIG_HAVE_IOREMAP_PROT
4095 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004096 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004097 break;
4098 if (vma->vm_ops && vma->vm_ops->access)
4099 ret = vma->vm_ops->access(vma, addr, buf,
4100 len, write);
4101 if (ret <= 0)
4102#endif
4103 break;
4104 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004105 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004106 bytes = len;
4107 offset = addr & (PAGE_SIZE-1);
4108 if (bytes > PAGE_SIZE-offset)
4109 bytes = PAGE_SIZE-offset;
4110
4111 maddr = kmap(page);
4112 if (write) {
4113 copy_to_user_page(vma, page, addr,
4114 maddr + offset, buf, bytes);
4115 set_page_dirty_lock(page);
4116 } else {
4117 copy_from_user_page(vma, page, addr,
4118 buf, maddr + offset, bytes);
4119 }
4120 kunmap(page);
4121 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004122 }
David Howells0ec76a12006-09-27 01:50:15 -07004123 len -= bytes;
4124 buf += bytes;
4125 addr += bytes;
4126 }
4127 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004128
4129 return buf - old_buf;
4130}
Andi Kleen03252912008-01-30 13:33:18 +01004131
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004132/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004133 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004134 * @mm: the mm_struct of the target address space
4135 * @addr: start address to access
4136 * @buf: source or destination buffer
4137 * @len: number of bytes to transfer
4138 * @write: whether the access is a write
4139 *
4140 * The caller must hold a reference on @mm.
4141 */
4142int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4143 void *buf, int len, int write)
4144{
4145 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4146}
4147
Andi Kleen03252912008-01-30 13:33:18 +01004148/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004149 * Access another process' address space.
4150 * Source/target buffer must be kernel space,
4151 * Do not walk the page table directly, use get_user_pages
4152 */
4153int access_process_vm(struct task_struct *tsk, unsigned long addr,
4154 void *buf, int len, int write)
4155{
4156 struct mm_struct *mm;
4157 int ret;
4158
4159 mm = get_task_mm(tsk);
4160 if (!mm)
4161 return 0;
4162
4163 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4164 mmput(mm);
4165
4166 return ret;
4167}
4168
Andi Kleen03252912008-01-30 13:33:18 +01004169/*
4170 * Print the name of a VMA.
4171 */
4172void print_vma_addr(char *prefix, unsigned long ip)
4173{
4174 struct mm_struct *mm = current->mm;
4175 struct vm_area_struct *vma;
4176
Ingo Molnare8bff742008-02-13 20:21:06 +01004177 /*
4178 * Do not print if we are in atomic
4179 * contexts (in exception stacks, etc.):
4180 */
4181 if (preempt_count())
4182 return;
4183
Andi Kleen03252912008-01-30 13:33:18 +01004184 down_read(&mm->mmap_sem);
4185 vma = find_vma(mm, ip);
4186 if (vma && vma->vm_file) {
4187 struct file *f = vma->vm_file;
4188 char *buf = (char *)__get_free_page(GFP_KERNEL);
4189 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004190 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004191
Jan Blunckcf28b482008-02-14 19:38:44 -08004192 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004193 if (IS_ERR(p))
4194 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004195 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004196 vma->vm_start,
4197 vma->vm_end - vma->vm_start);
4198 free_page((unsigned long)buf);
4199 }
4200 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004201 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004202}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004203
4204#ifdef CONFIG_PROVE_LOCKING
4205void might_fault(void)
4206{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004207 /*
4208 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4209 * holding the mmap_sem, this is safe because kernel memory doesn't
4210 * get paged out, therefore we'll never actually fault, and the
4211 * below annotations will generate false positives.
4212 */
4213 if (segment_eq(get_fs(), KERNEL_DS))
4214 return;
4215
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004216 might_sleep();
4217 /*
4218 * it would be nicer only to annotate paths which are not under
4219 * pagefault_disable, however that requires a larger audit and
4220 * providing helpers like get_user_atomic.
4221 */
4222 if (!in_atomic() && current->mm)
4223 might_lock_read(&current->mm->mmap_sem);
4224}
4225EXPORT_SYMBOL(might_fault);
4226#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004227
4228#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4229static void clear_gigantic_page(struct page *page,
4230 unsigned long addr,
4231 unsigned int pages_per_huge_page)
4232{
4233 int i;
4234 struct page *p = page;
4235
4236 might_sleep();
4237 for (i = 0; i < pages_per_huge_page;
4238 i++, p = mem_map_next(p, page, i)) {
4239 cond_resched();
4240 clear_user_highpage(p, addr + i * PAGE_SIZE);
4241 }
4242}
4243void clear_huge_page(struct page *page,
4244 unsigned long addr, unsigned int pages_per_huge_page)
4245{
4246 int i;
4247
4248 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4249 clear_gigantic_page(page, addr, pages_per_huge_page);
4250 return;
4251 }
4252
4253 might_sleep();
4254 for (i = 0; i < pages_per_huge_page; i++) {
4255 cond_resched();
4256 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4257 }
4258}
4259
4260static void copy_user_gigantic_page(struct page *dst, struct page *src,
4261 unsigned long addr,
4262 struct vm_area_struct *vma,
4263 unsigned int pages_per_huge_page)
4264{
4265 int i;
4266 struct page *dst_base = dst;
4267 struct page *src_base = src;
4268
4269 for (i = 0; i < pages_per_huge_page; ) {
4270 cond_resched();
4271 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4272
4273 i++;
4274 dst = mem_map_next(dst, dst_base, i);
4275 src = mem_map_next(src, src_base, i);
4276 }
4277}
4278
4279void copy_user_huge_page(struct page *dst, struct page *src,
4280 unsigned long addr, struct vm_area_struct *vma,
4281 unsigned int pages_per_huge_page)
4282{
4283 int i;
4284
4285 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4286 copy_user_gigantic_page(dst, src, addr, vma,
4287 pages_per_huge_page);
4288 return;
4289 }
4290
4291 might_sleep();
4292 for (i = 0; i < pages_per_huge_page; i++) {
4293 cond_resched();
4294 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4295 }
4296}
4297#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */