blob: 13cbc420fead06d96f82b437a67d03b6ff9d5fca [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;
223 tlb->fast_mode = (num_possible_cpus() == 1);
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800228 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800232#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700233}
234
235void tlb_flush_mmu(struct mmu_gather *tlb)
236{
237 struct mmu_gather_batch *batch;
238
239 if (!tlb->need_flush)
240 return;
241 tlb->need_flush = 0;
242 tlb_flush(tlb);
243#ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb);
245#endif
246
247 if (tlb_fast_mode(tlb))
248 return;
249
250 for (batch = &tlb->local; batch; batch = batch->next) {
251 free_pages_and_swap_cache(batch->pages, batch->nr);
252 batch->nr = 0;
253 }
254 tlb->active = &tlb->local;
255}
256
257/* tlb_finish_mmu
258 * Called at the end of the shootdown operation to free up any resources
259 * that were required.
260 */
261void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
262{
263 struct mmu_gather_batch *batch, *next;
264
Alex Shi597e1c32012-06-28 09:02:21 +0800265 tlb->start = start;
266 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700267 tlb_flush_mmu(tlb);
268
269 /* keep the page table cache within bounds */
270 check_pgt_cache();
271
272 for (batch = tlb->local.next; batch; batch = next) {
273 next = batch->next;
274 free_pages((unsigned long)batch, 0);
275 }
276 tlb->local.next = NULL;
277}
278
279/* __tlb_remove_page
280 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
281 * handling the additional races in SMP caused by other CPUs caching valid
282 * mappings in their TLBs. Returns the number of free page slots left.
283 * When out of page slots we must call tlb_flush_mmu().
284 */
285int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
286{
287 struct mmu_gather_batch *batch;
288
Shaohua Lif21760b2012-01-12 17:19:16 -0800289 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700290
291 if (tlb_fast_mode(tlb)) {
292 free_page_and_swap_cache(page);
293 return 1; /* avoid calling tlb_flush_mmu() */
294 }
295
296 batch = tlb->active;
297 batch->pages[batch->nr++] = page;
298 if (batch->nr == batch->max) {
299 if (!tlb_next_batch(tlb))
300 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700301 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700302 }
303 VM_BUG_ON(batch->nr > batch->max);
304
305 return batch->max - batch->nr;
306}
307
308#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800309
Peter Zijlstra26723912011-05-24 17:12:00 -0700310#ifdef CONFIG_HAVE_RCU_TABLE_FREE
311
312/*
313 * See the comment near struct mmu_table_batch.
314 */
315
316static void tlb_remove_table_smp_sync(void *arg)
317{
318 /* Simply deliver the interrupt */
319}
320
321static void tlb_remove_table_one(void *table)
322{
323 /*
324 * This isn't an RCU grace period and hence the page-tables cannot be
325 * assumed to be actually RCU-freed.
326 *
327 * It is however sufficient for software page-table walkers that rely on
328 * IRQ disabling. See the comment near struct mmu_table_batch.
329 */
330 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
331 __tlb_remove_table(table);
332}
333
334static void tlb_remove_table_rcu(struct rcu_head *head)
335{
336 struct mmu_table_batch *batch;
337 int i;
338
339 batch = container_of(head, struct mmu_table_batch, rcu);
340
341 for (i = 0; i < batch->nr; i++)
342 __tlb_remove_table(batch->tables[i]);
343
344 free_page((unsigned long)batch);
345}
346
347void tlb_table_flush(struct mmu_gather *tlb)
348{
349 struct mmu_table_batch **batch = &tlb->batch;
350
351 if (*batch) {
352 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
353 *batch = NULL;
354 }
355}
356
357void tlb_remove_table(struct mmu_gather *tlb, void *table)
358{
359 struct mmu_table_batch **batch = &tlb->batch;
360
361 tlb->need_flush = 1;
362
363 /*
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
366 */
367 if (atomic_read(&tlb->mm->mm_users) < 2) {
368 __tlb_remove_table(table);
369 return;
370 }
371
372 if (*batch == NULL) {
373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
374 if (*batch == NULL) {
375 tlb_remove_table_one(table);
376 return;
377 }
378 (*batch)->nr = 0;
379 }
380 (*batch)->tables[(*batch)->nr++] = table;
381 if ((*batch)->nr == MAX_TABLE_BATCH)
382 tlb_table_flush(tlb);
383}
384
Peter Zijlstra9547d012011-05-24 17:12:14 -0700385#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700386
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387/*
388 * If a p?d_bad entry is found while walking page tables, report
389 * the error, before resetting entry to p?d_none. Usually (but
390 * very seldom) called out from the p?d_none_or_clear_bad macros.
391 */
392
393void pgd_clear_bad(pgd_t *pgd)
394{
395 pgd_ERROR(*pgd);
396 pgd_clear(pgd);
397}
398
399void pud_clear_bad(pud_t *pud)
400{
401 pud_ERROR(*pud);
402 pud_clear(pud);
403}
404
405void pmd_clear_bad(pmd_t *pmd)
406{
407 pmd_ERROR(*pmd);
408 pmd_clear(pmd);
409}
410
411/*
412 * Note: this doesn't free the actual pages themselves. That
413 * has been handled earlier when unmapping all the memory regions.
414 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000415static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
416 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800418 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000420 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422}
423
Hugh Dickinse0da3822005-04-19 13:29:15 -0700424static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
425 unsigned long addr, unsigned long end,
426 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427{
428 pmd_t *pmd;
429 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Hugh Dickinse0da3822005-04-19 13:29:15 -0700432 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 do {
435 next = pmd_addr_end(addr, end);
436 if (pmd_none_or_clear_bad(pmd))
437 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000438 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 } while (pmd++, addr = next, addr != end);
440
Hugh Dickinse0da3822005-04-19 13:29:15 -0700441 start &= PUD_MASK;
442 if (start < floor)
443 return;
444 if (ceiling) {
445 ceiling &= PUD_MASK;
446 if (!ceiling)
447 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700449 if (end - 1 > ceiling - 1)
450 return;
451
452 pmd = pmd_offset(pud, start);
453 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000454 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455}
456
Hugh Dickinse0da3822005-04-19 13:29:15 -0700457static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
458 unsigned long addr, unsigned long end,
459 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 pud_t *pud;
462 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700463 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
Hugh Dickinse0da3822005-04-19 13:29:15 -0700465 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 do {
468 next = pud_addr_end(addr, end);
469 if (pud_none_or_clear_bad(pud))
470 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 } while (pud++, addr = next, addr != end);
473
Hugh Dickinse0da3822005-04-19 13:29:15 -0700474 start &= PGDIR_MASK;
475 if (start < floor)
476 return;
477 if (ceiling) {
478 ceiling &= PGDIR_MASK;
479 if (!ceiling)
480 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700482 if (end - 1 > ceiling - 1)
483 return;
484
485 pud = pud_offset(pgd, start);
486 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000487 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488}
489
490/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700491 * This function frees user-level page tables of a process.
492 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 * Must be called with pagetable lock held.
494 */
Jan Beulich42b77722008-07-23 21:27:10 -0700495void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700496 unsigned long addr, unsigned long end,
497 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498{
499 pgd_t *pgd;
500 unsigned long next;
501
Hugh Dickinse0da3822005-04-19 13:29:15 -0700502 /*
503 * The next few lines have given us lots of grief...
504 *
505 * Why are we testing PMD* at this top level? Because often
506 * there will be no work to do at all, and we'd prefer not to
507 * go all the way down to the bottom just to discover that.
508 *
509 * Why all these "- 1"s? Because 0 represents both the bottom
510 * of the address space and the top of it (using -1 for the
511 * top wouldn't help much: the masks would do the wrong thing).
512 * The rule is that addr 0 and floor 0 refer to the bottom of
513 * the address space, but end 0 and ceiling 0 refer to the top
514 * Comparisons need to use "end - 1" and "ceiling - 1" (though
515 * that end 0 case should be mythical).
516 *
517 * Wherever addr is brought up or ceiling brought down, we must
518 * be careful to reject "the opposite 0" before it confuses the
519 * subsequent tests. But what about where end is brought down
520 * by PMD_SIZE below? no, end can't go down to 0 there.
521 *
522 * Whereas we round start (addr) and ceiling down, by different
523 * masks at different levels, in order to test whether a table
524 * now has no other vmas using it, so can be freed, we don't
525 * bother to round floor or end up - the tests don't need that.
526 */
527
528 addr &= PMD_MASK;
529 if (addr < floor) {
530 addr += PMD_SIZE;
531 if (!addr)
532 return;
533 }
534 if (ceiling) {
535 ceiling &= PMD_MASK;
536 if (!ceiling)
537 return;
538 }
539 if (end - 1 > ceiling - 1)
540 end -= PMD_SIZE;
541 if (addr > end - 1)
542 return;
543
Jan Beulich42b77722008-07-23 21:27:10 -0700544 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 do {
546 next = pgd_addr_end(addr, end);
547 if (pgd_none_or_clear_bad(pgd))
548 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700549 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700551}
552
Jan Beulich42b77722008-07-23 21:27:10 -0700553void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700554 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700555{
556 while (vma) {
557 struct vm_area_struct *next = vma->vm_next;
558 unsigned long addr = vma->vm_start;
559
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700560 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000561 * Hide vma from rmap and truncate_pagecache before freeing
562 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700563 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800564 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700565 unlink_file_vma(vma);
566
David Gibson9da61ae2006-03-22 00:08:57 -0800567 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700568 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700569 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700570 } else {
571 /*
572 * Optimization: gather nearby vmas into one call down
573 */
574 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800575 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700576 vma = next;
577 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800578 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700579 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700580 }
581 free_pgd_range(tlb, addr, vma->vm_end,
582 floor, next? next->vm_start: ceiling);
583 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700584 vma = next;
585 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586}
587
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800588int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
589 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800591 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800592 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700593 if (!new)
594 return -ENOMEM;
595
Nick Piggin362a61a2008-05-14 06:37:36 +0200596 /*
597 * Ensure all pte setup (eg. pte page lock and page clearing) are
598 * visible before the pte is made visible to other CPUs by being
599 * put into page tables.
600 *
601 * The other side of the story is the pointer chasing in the page
602 * table walking code (when walking the page table without locking;
603 * ie. most of the time). Fortunately, these data accesses consist
604 * of a chain of data-dependent loads, meaning most CPUs (alpha
605 * being the notable exception) will already guarantee loads are
606 * seen in-order. See the alpha page table accessors for the
607 * smp_read_barrier_depends() barriers in page table walking code.
608 */
609 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
610
Hugh Dickinsc74df322005-10-29 18:16:23 -0700611 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800612 wait_split_huge_page = 0;
613 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800616 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800617 } else if (unlikely(pmd_trans_splitting(*pmd)))
618 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700619 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800620 if (new)
621 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800622 if (wait_split_huge_page)
623 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700624 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
Hugh Dickins1bb36302005-10-29 18:16:22 -0700627int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700629 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
630 if (!new)
631 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Nick Piggin362a61a2008-05-14 06:37:36 +0200633 smp_wmb(); /* See comment in __pte_alloc */
634
Hugh Dickins1bb36302005-10-29 18:16:22 -0700635 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800636 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700637 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800638 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800639 } else
640 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700641 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800642 if (new)
643 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700644 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645}
646
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800647static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700648{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800649 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
650}
651
652static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
653{
654 int i;
655
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800656 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700657 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800658 for (i = 0; i < NR_MM_COUNTERS; i++)
659 if (rss[i])
660 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700661}
662
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800664 * This function is called to print an error when a bad pte
665 * is found. For example, we might have a PFN-mapped pte in
666 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700667 *
668 * The calling function must still handle the error.
669 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800670static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
671 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700672{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800673 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
674 pud_t *pud = pud_offset(pgd, addr);
675 pmd_t *pmd = pmd_offset(pud, addr);
676 struct address_space *mapping;
677 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800678 static unsigned long resume;
679 static unsigned long nr_shown;
680 static unsigned long nr_unshown;
681
682 /*
683 * Allow a burst of 60 reports, then keep quiet for that minute;
684 * or allow a steady drip of one report per second.
685 */
686 if (nr_shown == 60) {
687 if (time_before(jiffies, resume)) {
688 nr_unshown++;
689 return;
690 }
691 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800692 printk(KERN_ALERT
693 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800694 nr_unshown);
695 nr_unshown = 0;
696 }
697 nr_shown = 0;
698 }
699 if (nr_shown++ == 0)
700 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800701
702 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
703 index = linear_page_index(vma, addr);
704
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800705 printk(KERN_ALERT
706 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800707 current->comm,
708 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800709 if (page)
710 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800711 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800712 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
713 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
714 /*
715 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
716 */
717 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800718 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800719 (unsigned long)vma->vm_ops->fault);
720 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800721 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800722 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700723 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030724 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700725}
726
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700727static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800728{
729 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
730}
731
Nick Pigginb5810032005-10-29 18:16:12 -0700732/*
Nick Piggin7e675132008-04-28 02:13:00 -0700733 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800734 *
Nick Piggin7e675132008-04-28 02:13:00 -0700735 * "Special" mappings do not wish to be associated with a "struct page" (either
736 * it doesn't exist, or it exists but they don't want to touch it). In this
737 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700738 *
Nick Piggin7e675132008-04-28 02:13:00 -0700739 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
740 * pte bit, in which case this function is trivial. Secondly, an architecture
741 * may not have a spare pte bit, which requires a more complicated scheme,
742 * described below.
743 *
744 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
745 * special mapping (even if there are underlying and valid "struct pages").
746 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800747 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700748 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
749 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700750 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
751 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800752 *
753 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
754 *
Nick Piggin7e675132008-04-28 02:13:00 -0700755 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700756 *
Nick Piggin7e675132008-04-28 02:13:00 -0700757 * This restricts such mappings to be a linear translation from virtual address
758 * to pfn. To get around this restriction, we allow arbitrary mappings so long
759 * as the vma is not a COW mapping; in that case, we know that all ptes are
760 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700761 *
762 *
Nick Piggin7e675132008-04-28 02:13:00 -0700763 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
764 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700765 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
766 * page" backing, however the difference is that _all_ pages with a struct
767 * page (that is, those where pfn_valid is true) are refcounted and considered
768 * normal pages by the VM. The disadvantage is that pages are refcounted
769 * (which can be slower and simply not an option for some PFNMAP users). The
770 * advantage is that we don't have to follow the strict linearity rule of
771 * PFNMAP mappings in order to support COWable mappings.
772 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800773 */
Nick Piggin7e675132008-04-28 02:13:00 -0700774#ifdef __HAVE_ARCH_PTE_SPECIAL
775# define HAVE_PTE_SPECIAL 1
776#else
777# define HAVE_PTE_SPECIAL 0
778#endif
779struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
780 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800781{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800782 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700783
784 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800785 if (likely(!pte_special(pte)))
786 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700787 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
788 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700789 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800790 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700791 return NULL;
792 }
793
794 /* !HAVE_PTE_SPECIAL case follows: */
795
Jared Hulbertb379d792008-04-28 02:12:58 -0700796 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
797 if (vma->vm_flags & VM_MIXEDMAP) {
798 if (!pfn_valid(pfn))
799 return NULL;
800 goto out;
801 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700802 unsigned long off;
803 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700804 if (pfn == vma->vm_pgoff + off)
805 return NULL;
806 if (!is_cow_mapping(vma->vm_flags))
807 return NULL;
808 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800809 }
810
Hugh Dickins62eede62009-09-21 17:03:34 -0700811 if (is_zero_pfn(pfn))
812 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800813check_pfn:
814 if (unlikely(pfn > highest_memmap_pfn)) {
815 print_bad_pte(vma, addr, pte, NULL);
816 return NULL;
817 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800818
819 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700820 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700821 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800822 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700823out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800824 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800825}
826
827/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 * copy one vm_area from one task to the other. Assumes the page tables
829 * already present in the new task to be cleared in the whole range
830 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 */
832
Hugh Dickins570a335b2009-12-14 17:58:46 -0800833static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700835 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700836 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837{
Nick Pigginb5810032005-10-29 18:16:12 -0700838 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 pte_t pte = *src_pte;
840 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841
842 /* pte contains position in swap or file, so copy. */
843 if (unlikely(!pte_present(pte))) {
844 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700845 swp_entry_t entry = pte_to_swp_entry(pte);
846
Hugh Dickins570a335b2009-12-14 17:58:46 -0800847 if (swap_duplicate(entry) < 0)
848 return entry.val;
849
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 /* make sure dst_mm is on swapoff's mmlist. */
851 if (unlikely(list_empty(&dst_mm->mmlist))) {
852 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700853 if (list_empty(&dst_mm->mmlist))
854 list_add(&dst_mm->mmlist,
855 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 spin_unlock(&mmlist_lock);
857 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800858 if (likely(!non_swap_entry(entry)))
859 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800860 else if (is_migration_entry(entry)) {
861 page = migration_entry_to_page(entry);
862
863 if (PageAnon(page))
864 rss[MM_ANONPAGES]++;
865 else
866 rss[MM_FILEPAGES]++;
867
868 if (is_write_migration_entry(entry) &&
869 is_cow_mapping(vm_flags)) {
870 /*
871 * COW mappings require pages in both
872 * parent and child to be set to read.
873 */
874 make_migration_entry_read(&entry);
875 pte = swp_entry_to_pte(entry);
876 set_pte_at(src_mm, addr, src_pte, pte);
877 }
Christoph Lameter06972122006-06-23 02:03:35 -0700878 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700880 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 }
882
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 /*
884 * If it's a COW mapping, write protect it both
885 * in the parent and the child
886 */
Linus Torvalds67121172005-12-11 20:38:17 -0800887 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700889 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 }
891
892 /*
893 * If it's a shared mapping, mark it clean in
894 * the child
895 */
896 if (vm_flags & VM_SHARED)
897 pte = pte_mkclean(pte);
898 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800899
900 page = vm_normal_page(vma, addr, pte);
901 if (page) {
902 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700903 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800904 if (PageAnon(page))
905 rss[MM_ANONPAGES]++;
906 else
907 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800908 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700909
910out_set_pte:
911 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800912 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913}
914
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800915int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700919 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700921 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700922 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800923 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800924 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925
926again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800927 init_rss_vec(rss);
928
Hugh Dickinsc74df322005-10-29 18:16:23 -0700929 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 if (!dst_pte)
931 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700932 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700933 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700934 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700935 orig_src_pte = src_pte;
936 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700937 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 do {
940 /*
941 * We are holding two locks at this point - either of them
942 * could generate latencies in another task on another CPU.
943 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700944 if (progress >= 32) {
945 progress = 0;
946 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100947 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700948 break;
949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 if (pte_none(*src_pte)) {
951 progress++;
952 continue;
953 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800954 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
955 vma, addr, rss);
956 if (entry.val)
957 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 progress += 8;
959 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700961 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700962 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700963 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800964 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700965 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700966 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800967
968 if (entry.val) {
969 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
970 return -ENOMEM;
971 progress = 0;
972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 if (addr != end)
974 goto again;
975 return 0;
976}
977
978static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
979 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
980 unsigned long addr, unsigned long end)
981{
982 pmd_t *src_pmd, *dst_pmd;
983 unsigned long next;
984
985 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
986 if (!dst_pmd)
987 return -ENOMEM;
988 src_pmd = pmd_offset(src_pud, addr);
989 do {
990 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800991 if (pmd_trans_huge(*src_pmd)) {
992 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800993 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800994 err = copy_huge_pmd(dst_mm, src_mm,
995 dst_pmd, src_pmd, addr, vma);
996 if (err == -ENOMEM)
997 return -ENOMEM;
998 if (!err)
999 continue;
1000 /* fall through */
1001 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 if (pmd_none_or_clear_bad(src_pmd))
1003 continue;
1004 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1005 vma, addr, next))
1006 return -ENOMEM;
1007 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1008 return 0;
1009}
1010
1011static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1012 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1013 unsigned long addr, unsigned long end)
1014{
1015 pud_t *src_pud, *dst_pud;
1016 unsigned long next;
1017
1018 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1019 if (!dst_pud)
1020 return -ENOMEM;
1021 src_pud = pud_offset(src_pgd, addr);
1022 do {
1023 next = pud_addr_end(addr, end);
1024 if (pud_none_or_clear_bad(src_pud))
1025 continue;
1026 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1027 vma, addr, next))
1028 return -ENOMEM;
1029 } while (dst_pud++, src_pud++, addr = next, addr != end);
1030 return 0;
1031}
1032
1033int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1034 struct vm_area_struct *vma)
1035{
1036 pgd_t *src_pgd, *dst_pgd;
1037 unsigned long next;
1038 unsigned long addr = vma->vm_start;
1039 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001040 unsigned long mmun_start; /* For mmu_notifiers */
1041 unsigned long mmun_end; /* For mmu_notifiers */
1042 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001043 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
Nick Piggind9928952005-08-28 16:49:11 +10001045 /*
1046 * Don't copy ptes where a page fault will fill them correctly.
1047 * Fork becomes much lighter when there are big shared or private
1048 * readonly mappings. The tradeoff is that copy_page_range is more
1049 * efficient than faulting.
1050 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001051 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1052 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001053 if (!vma->anon_vma)
1054 return 0;
1055 }
1056
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 if (is_vm_hugetlb_page(vma))
1058 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1059
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001060 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001061 /*
1062 * We do not free on error cases below as remove_vma
1063 * gets called on error from higher level routine
1064 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001065 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001066 if (ret)
1067 return ret;
1068 }
1069
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001070 /*
1071 * We need to invalidate the secondary MMU mappings only when
1072 * there could be a permission downgrade on the ptes of the
1073 * parent mm. And a permission downgrade will only happen if
1074 * is_cow_mapping() returns true.
1075 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001076 is_cow = is_cow_mapping(vma->vm_flags);
1077 mmun_start = addr;
1078 mmun_end = end;
1079 if (is_cow)
1080 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1081 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001082
1083 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 dst_pgd = pgd_offset(dst_mm, addr);
1085 src_pgd = pgd_offset(src_mm, addr);
1086 do {
1087 next = pgd_addr_end(addr, end);
1088 if (pgd_none_or_clear_bad(src_pgd))
1089 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001090 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1091 vma, addr, next))) {
1092 ret = -ENOMEM;
1093 break;
1094 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001096
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001097 if (is_cow)
1098 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001099 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100}
1101
Robin Holt51c6f662005-11-13 16:06:42 -08001102static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001103 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001105 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106{
Nick Pigginb5810032005-10-29 18:16:12 -07001107 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001108 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001109 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001110 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001111 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001112 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001113
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001114again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001115 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001116 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1117 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001118 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 do {
1120 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001121 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001123 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001124
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001126 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001127
Linus Torvalds6aab3412005-11-28 14:34:23 -08001128 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 if (unlikely(details) && page) {
1130 /*
1131 * unmap_shared_mapping_pages() wants to
1132 * invalidate cache without truncating:
1133 * unmap shared but keep private pages.
1134 */
1135 if (details->check_mapping &&
1136 details->check_mapping != page->mapping)
1137 continue;
1138 /*
1139 * Each page->index must be checked when
1140 * invalidating or truncating nonlinear.
1141 */
1142 if (details->nonlinear_vma &&
1143 (page->index < details->first_index ||
1144 page->index > details->last_index))
1145 continue;
1146 }
Nick Pigginb5810032005-10-29 18:16:12 -07001147 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001148 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 tlb_remove_tlb_entry(tlb, pte, addr);
1150 if (unlikely(!page))
1151 continue;
1152 if (unlikely(details) && details->nonlinear_vma
1153 && linear_page_index(details->nonlinear_vma,
1154 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001155 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001158 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001159 else {
1160 if (pte_dirty(ptent))
1161 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001162 if (pte_young(ptent) &&
1163 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001164 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001165 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001166 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001167 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001168 if (unlikely(page_mapcount(page) < 0))
1169 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001170 force_flush = !__tlb_remove_page(tlb, page);
1171 if (force_flush)
1172 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 continue;
1174 }
1175 /*
1176 * If details->check_mapping, we leave swap entries;
1177 * if details->nonlinear_vma, we leave file entries.
1178 */
1179 if (unlikely(details))
1180 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001181 if (pte_file(ptent)) {
1182 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1183 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001184 } else {
1185 swp_entry_t entry = pte_to_swp_entry(ptent);
1186
1187 if (!non_swap_entry(entry))
1188 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001189 else if (is_migration_entry(entry)) {
1190 struct page *page;
1191
1192 page = migration_entry_to_page(entry);
1193
1194 if (PageAnon(page))
1195 rss[MM_ANONPAGES]--;
1196 else
1197 rss[MM_FILEPAGES]--;
1198 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001199 if (unlikely(!free_swap_and_cache(entry)))
1200 print_bad_pte(vma, addr, ptent, NULL);
1201 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001202 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001203 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001204
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001205 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001206 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001207 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001208
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001209 /*
1210 * mmu_gather ran out of room to batch pages, we break out of
1211 * the PTE lock to avoid doing the potential expensive TLB invalidate
1212 * and page-free while holding it.
1213 */
1214 if (force_flush) {
1215 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001216
1217#ifdef HAVE_GENERIC_MMU_GATHER
1218 tlb->start = addr;
1219 tlb->end = end;
1220#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001221 tlb_flush_mmu(tlb);
1222 if (addr != end)
1223 goto again;
1224 }
1225
Robin Holt51c6f662005-11-13 16:06:42 -08001226 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227}
1228
Robin Holt51c6f662005-11-13 16:06:42 -08001229static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001230 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001232 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233{
1234 pmd_t *pmd;
1235 unsigned long next;
1236
1237 pmd = pmd_offset(pud, addr);
1238 do {
1239 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001240 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001241 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001242#ifdef CONFIG_DEBUG_VM
1243 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1244 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1245 __func__, addr, end,
1246 vma->vm_start,
1247 vma->vm_end);
1248 BUG();
1249 }
1250#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001251 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001252 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001253 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001254 /* fall through */
1255 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001256 /*
1257 * Here there can be other concurrent MADV_DONTNEED or
1258 * trans huge page faults running, and if the pmd is
1259 * none or trans huge it can change under us. This is
1260 * because MADV_DONTNEED holds the mmap_sem in read
1261 * mode.
1262 */
1263 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1264 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001265 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001266next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001267 cond_resched();
1268 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001269
1270 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271}
1272
Robin Holt51c6f662005-11-13 16:06:42 -08001273static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001274 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001276 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
1278 pud_t *pud;
1279 unsigned long next;
1280
1281 pud = pud_offset(pgd, addr);
1282 do {
1283 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001284 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001286 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1287 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001288
1289 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290}
1291
Al Viro038c7aa2012-03-05 13:25:09 -05001292static void unmap_page_range(struct mmu_gather *tlb,
1293 struct vm_area_struct *vma,
1294 unsigned long addr, unsigned long end,
1295 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296{
1297 pgd_t *pgd;
1298 unsigned long next;
1299
1300 if (details && !details->check_mapping && !details->nonlinear_vma)
1301 details = NULL;
1302
1303 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001304 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 tlb_start_vma(tlb, vma);
1306 pgd = pgd_offset(vma->vm_mm, addr);
1307 do {
1308 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001309 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001311 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1312 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001314 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316
Al Virof5cc4ee2012-03-05 14:14:20 -05001317
1318static void unmap_single_vma(struct mmu_gather *tlb,
1319 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001320 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001321 struct zap_details *details)
1322{
1323 unsigned long start = max(vma->vm_start, start_addr);
1324 unsigned long end;
1325
1326 if (start >= vma->vm_end)
1327 return;
1328 end = min(vma->vm_end, end_addr);
1329 if (end <= vma->vm_start)
1330 return;
1331
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301332 if (vma->vm_file)
1333 uprobe_munmap(vma, start, end);
1334
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001335 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001336 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001337
1338 if (start != end) {
1339 if (unlikely(is_vm_hugetlb_page(vma))) {
1340 /*
1341 * It is undesirable to test vma->vm_file as it
1342 * should be non-null for valid hugetlb area.
1343 * However, vm_file will be NULL in the error
1344 * cleanup path of do_mmap_pgoff. When
1345 * hugetlbfs ->mmap method fails,
1346 * do_mmap_pgoff() nullifies vma->vm_file
1347 * before calling this function to clean up.
1348 * Since no pte has actually been setup, it is
1349 * safe to do nothing in this case.
1350 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001351 if (vma->vm_file) {
1352 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001353 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001354 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1355 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001356 } else
1357 unmap_page_range(tlb, vma, start, end, details);
1358 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359}
1360
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361/**
1362 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001363 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 * @vma: the starting vma
1365 * @start_addr: virtual address at which to start unmapping
1366 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001368 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 * Only addresses between `start' and `end' will be unmapped.
1371 *
1372 * The VMA list must be sorted in ascending virtual address order.
1373 *
1374 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1375 * range after unmap_vmas() returns. So the only responsibility here is to
1376 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1377 * drops the lock and schedules.
1378 */
Al Viro6e8bb012012-03-05 13:41:15 -05001379void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001381 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001383 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001385 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001386 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001387 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001388 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389}
1390
1391/**
1392 * zap_page_range - remove user pages in a given range
1393 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001394 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 * @size: number of bytes to zap
1396 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001397 *
1398 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001400void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 unsigned long size, struct zap_details *details)
1402{
1403 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001404 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001405 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001408 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001409 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001410 mmu_notifier_invalidate_range_start(mm, start, end);
1411 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001412 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001413 mmu_notifier_invalidate_range_end(mm, start, end);
1414 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415}
1416
Jack Steinerc627f9c2008-07-29 22:33:53 -07001417/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001418 * zap_page_range_single - remove user pages in a given range
1419 * @vma: vm_area_struct holding the applicable pages
1420 * @address: starting address of pages to zap
1421 * @size: number of bytes to zap
1422 * @details: details of nonlinear truncation or shared cache invalidation
1423 *
1424 * The range must fit into one VMA.
1425 */
1426static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1427 unsigned long size, struct zap_details *details)
1428{
1429 struct mm_struct *mm = vma->vm_mm;
1430 struct mmu_gather tlb;
1431 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001432
1433 lru_add_drain();
1434 tlb_gather_mmu(&tlb, mm, 0);
1435 update_hiwater_rss(mm);
1436 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001437 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001438 mmu_notifier_invalidate_range_end(mm, address, end);
1439 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440}
1441
Jack Steinerc627f9c2008-07-29 22:33:53 -07001442/**
1443 * zap_vma_ptes - remove ptes mapping the vma
1444 * @vma: vm_area_struct holding ptes to be zapped
1445 * @address: starting address of pages to zap
1446 * @size: number of bytes to zap
1447 *
1448 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1449 *
1450 * The entire address range must be fully contained within the vma.
1451 *
1452 * Returns 0 if successful.
1453 */
1454int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1455 unsigned long size)
1456{
1457 if (address < vma->vm_start || address + size > vma->vm_end ||
1458 !(vma->vm_flags & VM_PFNMAP))
1459 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001460 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001461 return 0;
1462}
1463EXPORT_SYMBOL_GPL(zap_vma_ptes);
1464
Johannes Weiner142762b2010-05-24 14:32:39 -07001465/**
Michel Lespinasse240aade2013-02-22 16:35:56 -08001466 * follow_page_mask - look up a page descriptor from a user-virtual address
Johannes Weiner142762b2010-05-24 14:32:39 -07001467 * @vma: vm_area_struct mapping @address
1468 * @address: virtual address to look up
1469 * @flags: flags modifying lookup behaviour
Michel Lespinasse240aade2013-02-22 16:35:56 -08001470 * @page_mask: on output, *page_mask is set according to the size of the page
Johannes Weiner142762b2010-05-24 14:32:39 -07001471 *
1472 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1473 *
1474 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1475 * an error pointer if there is a mapping to something not represented
1476 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 */
Michel Lespinasse240aade2013-02-22 16:35:56 -08001478struct page *follow_page_mask(struct vm_area_struct *vma,
1479 unsigned long address, unsigned int flags,
1480 unsigned int *page_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481{
1482 pgd_t *pgd;
1483 pud_t *pud;
1484 pmd_t *pmd;
1485 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001486 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001488 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
Michel Lespinasse240aade2013-02-22 16:35:56 -08001490 *page_mask = 0;
1491
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1493 if (!IS_ERR(page)) {
1494 BUG_ON(flags & FOLL_GET);
1495 goto out;
1496 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001498 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 pgd = pgd_offset(mm, address);
1500 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001501 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
1503 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001504 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001505 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001506 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001507 BUG_ON(flags & FOLL_GET);
1508 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1509 goto out;
1510 }
1511 if (unlikely(pud_bad(*pud)))
1512 goto no_page_table;
1513
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001515 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001516 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001517 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001518 BUG_ON(flags & FOLL_GET);
1519 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1520 goto out;
1521 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001522 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1523 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001524 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001525 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001526 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001527 goto split_fallthrough;
1528 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001529 spin_lock(&mm->page_table_lock);
1530 if (likely(pmd_trans_huge(*pmd))) {
1531 if (unlikely(pmd_trans_splitting(*pmd))) {
1532 spin_unlock(&mm->page_table_lock);
1533 wait_split_huge_page(vma->anon_vma, pmd);
1534 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001535 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001536 pmd, flags);
1537 spin_unlock(&mm->page_table_lock);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001538 *page_mask = HPAGE_PMD_NR - 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001539 goto out;
1540 }
1541 } else
1542 spin_unlock(&mm->page_table_lock);
1543 /* fall through */
1544 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001545split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001546 if (unlikely(pmd_bad(*pmd)))
1547 goto no_page_table;
1548
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001549 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550
1551 pte = *ptep;
Hugh Dickins5117b3b2013-02-22 16:36:07 -08001552 if (!pte_present(pte)) {
1553 swp_entry_t entry;
1554 /*
1555 * KSM's break_ksm() relies upon recognizing a ksm page
1556 * even while it is being migrated, so for that case we
1557 * need migration_entry_wait().
1558 */
1559 if (likely(!(flags & FOLL_MIGRATION)))
1560 goto no_page;
1561 if (pte_none(pte) || pte_file(pte))
1562 goto no_page;
1563 entry = pte_to_swp_entry(pte);
1564 if (!is_migration_entry(entry))
1565 goto no_page;
1566 pte_unmap_unlock(ptep, ptl);
1567 migration_entry_wait(mm, pmd, address);
1568 goto split_fallthrough;
1569 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001570 if ((flags & FOLL_NUMA) && pte_numa(pte))
1571 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001572 if ((flags & FOLL_WRITE) && !pte_write(pte))
1573 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001574
Linus Torvalds6aab3412005-11-28 14:34:23 -08001575 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001576 if (unlikely(!page)) {
1577 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001578 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001579 goto bad_page;
1580 page = pte_page(pte);
1581 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001582
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001583 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001584 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001585 if (flags & FOLL_TOUCH) {
1586 if ((flags & FOLL_WRITE) &&
1587 !pte_dirty(pte) && !PageDirty(page))
1588 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001589 /*
1590 * pte_mkyoung() would be more correct here, but atomic care
1591 * is needed to avoid losing the dirty bit: it is easier to use
1592 * mark_page_accessed().
1593 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001594 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001596 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001597 /*
1598 * The preliminary mapping check is mainly to avoid the
1599 * pointless overhead of lock_page on the ZERO_PAGE
1600 * which might bounce very badly if there is contention.
1601 *
1602 * If the page is already locked, we don't need to
1603 * handle it now - vmscan will handle it later if and
1604 * when it attempts to reclaim the page.
1605 */
1606 if (page->mapping && trylock_page(page)) {
1607 lru_add_drain(); /* push cached pages to LRU */
1608 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001609 * Because we lock page here, and migration is
1610 * blocked by the pte's page reference, and we
1611 * know the page is still mapped, we don't even
1612 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001613 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001614 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001615 unlock_page(page);
1616 }
1617 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001618unlock:
1619 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001621 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001623bad_page:
1624 pte_unmap_unlock(ptep, ptl);
1625 return ERR_PTR(-EFAULT);
1626
1627no_page:
1628 pte_unmap_unlock(ptep, ptl);
1629 if (!pte_none(pte))
1630 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001631
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001632no_page_table:
1633 /*
1634 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001635 * has touched so far, we don't want to allocate unnecessary pages or
1636 * page tables. Return error instead of NULL to skip handle_mm_fault,
1637 * then get_dump_page() will return NULL to leave a hole in the dump.
1638 * But we can only make this optimization where a hole would surely
1639 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001640 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001641 if ((flags & FOLL_DUMP) &&
1642 (!vma->vm_ops || !vma->vm_ops->fault))
1643 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001644 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645}
1646
Linus Torvalds95042f92011-04-12 14:15:51 -07001647static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1648{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001649 return stack_guard_page_start(vma, addr) ||
1650 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001651}
1652
Huang Ying0014bd92011-01-30 11:15:47 +08001653/**
1654 * __get_user_pages() - pin user pages in memory
1655 * @tsk: task_struct of target task
1656 * @mm: mm_struct of target mm
1657 * @start: starting user address
1658 * @nr_pages: number of pages from start to pin
1659 * @gup_flags: flags modifying pin behaviour
1660 * @pages: array that receives pointers to the pages pinned.
1661 * Should be at least nr_pages long. Or NULL, if caller
1662 * only intends to ensure the pages are faulted in.
1663 * @vmas: array of pointers to vmas corresponding to each page.
1664 * Or NULL if the caller does not require them.
1665 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1666 *
1667 * Returns number of pages pinned. This may be fewer than the number
1668 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1669 * were pinned, returns -errno. Each page returned must be released
1670 * with a put_page() call when it is finished with. vmas will only
1671 * remain valid while mmap_sem is held.
1672 *
1673 * Must be called with mmap_sem held for read or write.
1674 *
1675 * __get_user_pages walks a process's page tables and takes a reference to
1676 * each struct page that each user address corresponds to at a given
1677 * instant. That is, it takes the page that would be accessed if a user
1678 * thread accesses the given user virtual address at that instant.
1679 *
1680 * This does not guarantee that the page exists in the user mappings when
1681 * __get_user_pages returns, and there may even be a completely different
1682 * page there in some cases (eg. if mmapped pagecache has been invalidated
1683 * and subsequently re faulted). However it does guarantee that the page
1684 * won't be freed completely. And mostly callers simply care that the page
1685 * contains data that was valid *at some point in time*. Typically, an IO
1686 * or similar operation cannot guarantee anything stronger anyway because
1687 * locks can't be held over the syscall boundary.
1688 *
1689 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1690 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1691 * appropriate) must be called after the page is finished with, and
1692 * before put_page is called.
1693 *
1694 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1695 * or mmap_sem contention, and if waiting is needed to pin all pages,
1696 * *@nonblocking will be set to 0.
1697 *
1698 * In most cases, get_user_pages or get_user_pages_fast should be used
1699 * instead of __get_user_pages. __get_user_pages should be used only if
1700 * you need some special @gup_flags.
1701 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08001702long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1703 unsigned long start, unsigned long nr_pages,
1704 unsigned int gup_flags, struct page **pages,
1705 struct vm_area_struct **vmas, int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706{
Michel Lespinasse28a35712013-02-22 16:35:55 -08001707 long i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001708 unsigned long vm_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001709 unsigned int page_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710
Michel Lespinasse28a35712013-02-22 16:35:55 -08001711 if (!nr_pages)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001712 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001713
1714 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1715
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 /*
1717 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001718 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001720 vm_flags = (gup_flags & FOLL_WRITE) ?
1721 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1722 vm_flags &= (gup_flags & FOLL_FORCE) ?
1723 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001724
1725 /*
1726 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1727 * would be called on PROT_NONE ranges. We must never invoke
1728 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1729 * page faults would unprotect the PROT_NONE ranges if
1730 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1731 * bitflag. So to avoid that, don't set FOLL_NUMA if
1732 * FOLL_FORCE is set.
1733 */
1734 if (!(gup_flags & FOLL_FORCE))
1735 gup_flags |= FOLL_NUMA;
1736
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 i = 0;
1738
1739 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001740 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
1742 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001743 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 pgd_t *pgd;
1746 pud_t *pud;
1747 pmd_t *pmd;
1748 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001749
1750 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001751 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 return i ? : -EFAULT;
1753 if (pg > TASK_SIZE)
1754 pgd = pgd_offset_k(pg);
1755 else
1756 pgd = pgd_offset_gate(mm, pg);
1757 BUG_ON(pgd_none(*pgd));
1758 pud = pud_offset(pgd, pg);
1759 BUG_ON(pud_none(*pud));
1760 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001761 if (pmd_none(*pmd))
1762 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001763 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001765 if (pte_none(*pte)) {
1766 pte_unmap(pte);
1767 return i ? : -EFAULT;
1768 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001769 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001771 struct page *page;
1772
Linus Torvalds95042f92011-04-12 14:15:51 -07001773 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001774 if (!page) {
1775 if (!(gup_flags & FOLL_DUMP) &&
1776 is_zero_pfn(pte_pfn(*pte)))
1777 page = pte_page(*pte);
1778 else {
1779 pte_unmap(pte);
1780 return i ? : -EFAULT;
1781 }
1782 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001783 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001784 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 }
1786 pte_unmap(pte);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001787 page_mask = 0;
Linus Torvalds95042f92011-04-12 14:15:51 -07001788 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 }
1790
Nick Pigginb291f002008-10-18 20:26:44 -07001791 if (!vma ||
1792 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001793 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 return i ? : -EFAULT;
1795
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001796 if (is_vm_hugetlb_page(vma)) {
1797 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001798 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001799 continue;
1800 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001801
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001803 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001804 unsigned int foll_flags = gup_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001805 unsigned int page_increm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
Ethan Solomita462e00c2007-07-15 23:38:16 -07001807 /*
Ying Han47792802009-01-06 14:40:18 -08001808 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001809 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001810 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001811 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001812 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001813
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001814 cond_resched();
Michel Lespinasse240aade2013-02-22 16:35:56 -08001815 while (!(page = follow_page_mask(vma, start,
1816 foll_flags, &page_mask))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001817 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001818 unsigned int fault_flags = 0;
1819
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001820 /* For mlock, just skip the stack guard page. */
1821 if (foll_flags & FOLL_MLOCK) {
1822 if (stack_guard_page(vma, start))
1823 goto next_page;
1824 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001825 if (foll_flags & FOLL_WRITE)
1826 fault_flags |= FAULT_FLAG_WRITE;
1827 if (nonblocking)
1828 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001829 if (foll_flags & FOLL_NOWAIT)
1830 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001831
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001832 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001833 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001834
Nick Piggin83c54072007-07-19 01:47:05 -07001835 if (ret & VM_FAULT_ERROR) {
1836 if (ret & VM_FAULT_OOM)
1837 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001838 if (ret & (VM_FAULT_HWPOISON |
1839 VM_FAULT_HWPOISON_LARGE)) {
1840 if (i)
1841 return i;
1842 else if (gup_flags & FOLL_HWPOISON)
1843 return -EHWPOISON;
1844 else
1845 return -EFAULT;
1846 }
1847 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001848 return i ? i : -EFAULT;
1849 BUG();
1850 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001851
1852 if (tsk) {
1853 if (ret & VM_FAULT_MAJOR)
1854 tsk->maj_flt++;
1855 else
1856 tsk->min_flt++;
1857 }
Nick Piggin83c54072007-07-19 01:47:05 -07001858
Michel Lespinasse53a77062011-01-13 15:46:14 -08001859 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001860 if (nonblocking)
1861 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001862 return i;
1863 }
1864
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001865 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001866 * The VM_FAULT_WRITE bit tells us that
1867 * do_wp_page has broken COW when necessary,
1868 * even if maybe_mkwrite decided not to set
1869 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001870 * page lookups as if they were reads. But only
1871 * do so when looping for pte_write is futile:
1872 * in some cases userspace may also be wanting
1873 * to write to the gotten user page, which a
1874 * read fault here might prevent (a readonly
1875 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001876 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001877 if ((ret & VM_FAULT_WRITE) &&
1878 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001879 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001880
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001881 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001883 if (IS_ERR(page))
1884 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001886 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001887
Russell Kinga6f36be2006-12-30 22:24:19 +00001888 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001889 flush_dcache_page(page);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001890 page_mask = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001892next_page:
Michel Lespinasse240aade2013-02-22 16:35:56 -08001893 if (vmas) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 vmas[i] = vma;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001895 page_mask = 0;
1896 }
1897 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1898 if (page_increm > nr_pages)
1899 page_increm = nr_pages;
1900 i += page_increm;
1901 start += page_increm * PAGE_SIZE;
1902 nr_pages -= page_increm;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001903 } while (nr_pages && start < vma->vm_end);
1904 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 return i;
1906}
Huang Ying0014bd92011-01-30 11:15:47 +08001907EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001908
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001909/*
1910 * fixup_user_fault() - manually resolve a user page fault
1911 * @tsk: the task_struct to use for page fault accounting, or
1912 * NULL if faults are not to be recorded.
1913 * @mm: mm_struct of target mm
1914 * @address: user address
1915 * @fault_flags:flags to pass down to handle_mm_fault()
1916 *
1917 * This is meant to be called in the specific scenario where for locking reasons
1918 * we try to access user memory in atomic context (within a pagefault_disable()
1919 * section), this returns -EFAULT, and we want to resolve the user fault before
1920 * trying again.
1921 *
1922 * Typically this is meant to be used by the futex code.
1923 *
1924 * The main difference with get_user_pages() is that this function will
1925 * unconditionally call handle_mm_fault() which will in turn perform all the
1926 * necessary SW fixup of the dirty and young bits in the PTE, while
1927 * handle_mm_fault() only guarantees to update these in the struct page.
1928 *
1929 * This is important for some architectures where those bits also gate the
1930 * access permission to the page because they are maintained in software. On
1931 * such architectures, gup() will not be enough to make a subsequent access
1932 * succeed.
1933 *
1934 * This should be called with the mm_sem held for read.
1935 */
1936int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1937 unsigned long address, unsigned int fault_flags)
1938{
1939 struct vm_area_struct *vma;
1940 int ret;
1941
1942 vma = find_extend_vma(mm, address);
1943 if (!vma || address < vma->vm_start)
1944 return -EFAULT;
1945
1946 ret = handle_mm_fault(mm, vma, address, fault_flags);
1947 if (ret & VM_FAULT_ERROR) {
1948 if (ret & VM_FAULT_OOM)
1949 return -ENOMEM;
1950 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1951 return -EHWPOISON;
1952 if (ret & VM_FAULT_SIGBUS)
1953 return -EFAULT;
1954 BUG();
1955 }
1956 if (tsk) {
1957 if (ret & VM_FAULT_MAJOR)
1958 tsk->maj_flt++;
1959 else
1960 tsk->min_flt++;
1961 }
1962 return 0;
1963}
1964
1965/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001966 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001967 * @tsk: the task_struct to use for page fault accounting, or
1968 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001969 * @mm: mm_struct of target mm
1970 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001971 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001972 * @write: whether pages will be written to by the caller
1973 * @force: whether to force write access even if user mapping is
1974 * readonly. This will result in the page being COWed even
1975 * in MAP_SHARED mappings. You do not want this.
1976 * @pages: array that receives pointers to the pages pinned.
1977 * Should be at least nr_pages long. Or NULL, if caller
1978 * only intends to ensure the pages are faulted in.
1979 * @vmas: array of pointers to vmas corresponding to each page.
1980 * Or NULL if the caller does not require them.
1981 *
1982 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001983 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001984 * were pinned, returns -errno. Each page returned must be released
1985 * with a put_page() call when it is finished with. vmas will only
1986 * remain valid while mmap_sem is held.
1987 *
1988 * Must be called with mmap_sem held for read or write.
1989 *
1990 * get_user_pages walks a process's page tables and takes a reference to
1991 * each struct page that each user address corresponds to at a given
1992 * instant. That is, it takes the page that would be accessed if a user
1993 * thread accesses the given user virtual address at that instant.
1994 *
1995 * This does not guarantee that the page exists in the user mappings when
1996 * get_user_pages returns, and there may even be a completely different
1997 * page there in some cases (eg. if mmapped pagecache has been invalidated
1998 * and subsequently re faulted). However it does guarantee that the page
1999 * won't be freed completely. And mostly callers simply care that the page
2000 * contains data that was valid *at some point in time*. Typically, an IO
2001 * or similar operation cannot guarantee anything stronger anyway because
2002 * locks can't be held over the syscall boundary.
2003 *
2004 * If write=0, the page must not be written to. If the page is written to,
2005 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2006 * after the page is finished with, and before put_page is called.
2007 *
2008 * get_user_pages is typically used for fewer-copy IO operations, to get a
2009 * handle on the memory by some means other than accesses via the user virtual
2010 * addresses. The pages may be submitted for DMA to devices or accessed via
2011 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2012 * use the correct cache flushing APIs.
2013 *
2014 * See also get_user_pages_fast, for performance critical applications.
2015 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08002016long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2017 unsigned long start, unsigned long nr_pages, int write,
2018 int force, struct page **pages, struct vm_area_struct **vmas)
Nick Pigginb291f002008-10-18 20:26:44 -07002019{
Hugh Dickins58fa8792009-09-21 17:03:31 -07002020 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07002021
Hugh Dickins58fa8792009-09-21 17:03:31 -07002022 if (pages)
2023 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07002024 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002025 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07002026 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002027 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07002028
Michel Lespinasse53a77062011-01-13 15:46:14 -08002029 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2030 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07002031}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032EXPORT_SYMBOL(get_user_pages);
2033
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002034/**
2035 * get_dump_page() - pin user page in memory while writing it to core dump
2036 * @addr: user address
2037 *
2038 * Returns struct page pointer of user page pinned for dump,
2039 * to be freed afterwards by page_cache_release() or put_page().
2040 *
2041 * Returns NULL on any kind of failure - a hole must then be inserted into
2042 * the corefile, to preserve alignment with its headers; and also returns
2043 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2044 * allowing a hole to be left in the corefile to save diskspace.
2045 *
2046 * Called without mmap_sem, but after all other threads have been killed.
2047 */
2048#ifdef CONFIG_ELF_CORE
2049struct page *get_dump_page(unsigned long addr)
2050{
2051 struct vm_area_struct *vma;
2052 struct page *page;
2053
2054 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002055 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2056 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002057 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002058 flush_cache_page(vma, addr, page_to_pfn(page));
2059 return page;
2060}
2061#endif /* CONFIG_ELF_CORE */
2062
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002063pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002064 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002065{
2066 pgd_t * pgd = pgd_offset(mm, addr);
2067 pud_t * pud = pud_alloc(mm, pgd, addr);
2068 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002069 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002070 if (pmd) {
2071 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002072 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002073 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002074 }
2075 return NULL;
2076}
2077
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002079 * This is the old fallback for page remapping.
2080 *
2081 * For historical reasons, it only allows reserved pages. Only
2082 * old drivers should use this, and they needed to mark their
2083 * pages reserved for the old functions anyway.
2084 */
Nick Piggin423bad62008-04-28 02:13:01 -07002085static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2086 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002087{
Nick Piggin423bad62008-04-28 02:13:01 -07002088 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002089 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002090 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002091 spinlock_t *ptl;
2092
Linus Torvalds238f58d2005-11-29 13:01:56 -08002093 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002094 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002095 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002096 retval = -ENOMEM;
2097 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002098 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002099 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002100 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002101 retval = -EBUSY;
2102 if (!pte_none(*pte))
2103 goto out_unlock;
2104
2105 /* Ok, finally just insert the thing.. */
2106 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002107 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002108 page_add_file_rmap(page);
2109 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2110
2111 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002112 pte_unmap_unlock(pte, ptl);
2113 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002114out_unlock:
2115 pte_unmap_unlock(pte, ptl);
2116out:
2117 return retval;
2118}
2119
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002120/**
2121 * vm_insert_page - insert single page into user vma
2122 * @vma: user vma to map to
2123 * @addr: target user address of this page
2124 * @page: source kernel page
2125 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002126 * This allows drivers to insert individual pages they've allocated
2127 * into a user vma.
2128 *
2129 * The page has to be a nice clean _individual_ kernel allocation.
2130 * If you allocate a compound page, you need to have marked it as
2131 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002132 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002133 *
2134 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2135 * took an arbitrary page protection parameter. This doesn't allow
2136 * that. Your vma protection will have to be set up correctly, which
2137 * means that if you want a shared writable mapping, you'd better
2138 * ask for a shared writable mapping!
2139 *
2140 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002141 *
2142 * Usually this function is called from f_op->mmap() handler
2143 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2144 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2145 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002146 */
Nick Piggin423bad62008-04-28 02:13:01 -07002147int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2148 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002149{
2150 if (addr < vma->vm_start || addr >= vma->vm_end)
2151 return -EFAULT;
2152 if (!page_count(page))
2153 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002154 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2155 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2156 BUG_ON(vma->vm_flags & VM_PFNMAP);
2157 vma->vm_flags |= VM_MIXEDMAP;
2158 }
Nick Piggin423bad62008-04-28 02:13:01 -07002159 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002160}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002161EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002162
Nick Piggin423bad62008-04-28 02:13:01 -07002163static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2164 unsigned long pfn, pgprot_t prot)
2165{
2166 struct mm_struct *mm = vma->vm_mm;
2167 int retval;
2168 pte_t *pte, entry;
2169 spinlock_t *ptl;
2170
2171 retval = -ENOMEM;
2172 pte = get_locked_pte(mm, addr, &ptl);
2173 if (!pte)
2174 goto out;
2175 retval = -EBUSY;
2176 if (!pte_none(*pte))
2177 goto out_unlock;
2178
2179 /* Ok, finally just insert the thing.. */
2180 entry = pte_mkspecial(pfn_pte(pfn, prot));
2181 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002182 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002183
2184 retval = 0;
2185out_unlock:
2186 pte_unmap_unlock(pte, ptl);
2187out:
2188 return retval;
2189}
2190
Nick Piggine0dc0d82007-02-12 00:51:36 -08002191/**
2192 * vm_insert_pfn - insert single pfn into user vma
2193 * @vma: user vma to map to
2194 * @addr: target user address of this page
2195 * @pfn: source kernel pfn
2196 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002197 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002198 * they've allocated into a user vma. Same comments apply.
2199 *
2200 * This function should only be called from a vm_ops->fault handler, and
2201 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002202 *
2203 * vma cannot be a COW mapping.
2204 *
2205 * As this is called only for pages that do not currently exist, we
2206 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002207 */
2208int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002209 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002210{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002211 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002212 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002213 /*
2214 * Technically, architectures with pte_special can avoid all these
2215 * restrictions (same for remap_pfn_range). However we would like
2216 * consistency in testing and feature parity among all, so we should
2217 * try to keep these invariants in place for everybody.
2218 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002219 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2220 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2221 (VM_PFNMAP|VM_MIXEDMAP));
2222 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2223 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002224
Nick Piggin423bad62008-04-28 02:13:01 -07002225 if (addr < vma->vm_start || addr >= vma->vm_end)
2226 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002227 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002228 return -EINVAL;
2229
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002230 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002231
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002232 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002233}
2234EXPORT_SYMBOL(vm_insert_pfn);
2235
Nick Piggin423bad62008-04-28 02:13:01 -07002236int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2237 unsigned long pfn)
2238{
2239 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2240
2241 if (addr < vma->vm_start || addr >= vma->vm_end)
2242 return -EFAULT;
2243
2244 /*
2245 * If we don't have pte special, then we have to use the pfn_valid()
2246 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2247 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002248 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2249 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002250 */
2251 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2252 struct page *page;
2253
2254 page = pfn_to_page(pfn);
2255 return insert_page(vma, addr, page, vma->vm_page_prot);
2256 }
2257 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2258}
2259EXPORT_SYMBOL(vm_insert_mixed);
2260
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002261/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 * maps a range of physical memory into the requested pages. the old
2263 * mappings are removed. any references to nonexistent pages results
2264 * in null mappings (currently treated as "copy-on-access")
2265 */
2266static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2267 unsigned long addr, unsigned long end,
2268 unsigned long pfn, pgprot_t prot)
2269{
2270 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002271 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
Hugh Dickinsc74df322005-10-29 18:16:23 -07002273 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 if (!pte)
2275 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002276 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 do {
2278 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002279 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 pfn++;
2281 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002282 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002283 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 return 0;
2285}
2286
2287static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2288 unsigned long addr, unsigned long end,
2289 unsigned long pfn, pgprot_t prot)
2290{
2291 pmd_t *pmd;
2292 unsigned long next;
2293
2294 pfn -= addr >> PAGE_SHIFT;
2295 pmd = pmd_alloc(mm, pud, addr);
2296 if (!pmd)
2297 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002298 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 do {
2300 next = pmd_addr_end(addr, end);
2301 if (remap_pte_range(mm, pmd, addr, next,
2302 pfn + (addr >> PAGE_SHIFT), prot))
2303 return -ENOMEM;
2304 } while (pmd++, addr = next, addr != end);
2305 return 0;
2306}
2307
2308static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2309 unsigned long addr, unsigned long end,
2310 unsigned long pfn, pgprot_t prot)
2311{
2312 pud_t *pud;
2313 unsigned long next;
2314
2315 pfn -= addr >> PAGE_SHIFT;
2316 pud = pud_alloc(mm, pgd, addr);
2317 if (!pud)
2318 return -ENOMEM;
2319 do {
2320 next = pud_addr_end(addr, end);
2321 if (remap_pmd_range(mm, pud, addr, next,
2322 pfn + (addr >> PAGE_SHIFT), prot))
2323 return -ENOMEM;
2324 } while (pud++, addr = next, addr != end);
2325 return 0;
2326}
2327
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002328/**
2329 * remap_pfn_range - remap kernel memory to userspace
2330 * @vma: user vma to map to
2331 * @addr: target user address to start at
2332 * @pfn: physical address of kernel memory
2333 * @size: size of map area
2334 * @prot: page protection flags for this mapping
2335 *
2336 * Note: this is only safe if the mm semaphore is held when called.
2337 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2339 unsigned long pfn, unsigned long size, pgprot_t prot)
2340{
2341 pgd_t *pgd;
2342 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002343 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 struct mm_struct *mm = vma->vm_mm;
2345 int err;
2346
2347 /*
2348 * Physically remapped pages are special. Tell the
2349 * rest of the world about it:
2350 * VM_IO tells people not to look at these pages
2351 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002352 * VM_PFNMAP tells the core MM that the base pages are just
2353 * raw PFN mappings, and do not have a "struct page" associated
2354 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002355 * VM_DONTEXPAND
2356 * Disable vma merging and expanding with mremap().
2357 * VM_DONTDUMP
2358 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002359 *
2360 * There's a horrible special case to handle copy-on-write
2361 * behaviour that some programs depend on. We mark the "original"
2362 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002363 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002365 if (is_cow_mapping(vma->vm_flags)) {
2366 if (addr != vma->vm_start || end != vma->vm_end)
2367 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002368 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002369 }
2370
2371 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2372 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002373 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002374
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002375 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
2377 BUG_ON(addr >= end);
2378 pfn -= addr >> PAGE_SHIFT;
2379 pgd = pgd_offset(mm, addr);
2380 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 do {
2382 next = pgd_addr_end(addr, end);
2383 err = remap_pud_range(mm, pgd, addr, next,
2384 pfn + (addr >> PAGE_SHIFT), prot);
2385 if (err)
2386 break;
2387 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002388
2389 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002390 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002391
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 return err;
2393}
2394EXPORT_SYMBOL(remap_pfn_range);
2395
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002396static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2397 unsigned long addr, unsigned long end,
2398 pte_fn_t fn, void *data)
2399{
2400 pte_t *pte;
2401 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002402 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002403 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002404
2405 pte = (mm == &init_mm) ?
2406 pte_alloc_kernel(pmd, addr) :
2407 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2408 if (!pte)
2409 return -ENOMEM;
2410
2411 BUG_ON(pmd_huge(*pmd));
2412
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002413 arch_enter_lazy_mmu_mode();
2414
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002415 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002416
2417 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002418 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002419 if (err)
2420 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002421 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002422
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002423 arch_leave_lazy_mmu_mode();
2424
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002425 if (mm != &init_mm)
2426 pte_unmap_unlock(pte-1, ptl);
2427 return err;
2428}
2429
2430static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2431 unsigned long addr, unsigned long end,
2432 pte_fn_t fn, void *data)
2433{
2434 pmd_t *pmd;
2435 unsigned long next;
2436 int err;
2437
Andi Kleenceb86872008-07-23 21:27:50 -07002438 BUG_ON(pud_huge(*pud));
2439
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002440 pmd = pmd_alloc(mm, pud, addr);
2441 if (!pmd)
2442 return -ENOMEM;
2443 do {
2444 next = pmd_addr_end(addr, end);
2445 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2446 if (err)
2447 break;
2448 } while (pmd++, addr = next, addr != end);
2449 return err;
2450}
2451
2452static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2453 unsigned long addr, unsigned long end,
2454 pte_fn_t fn, void *data)
2455{
2456 pud_t *pud;
2457 unsigned long next;
2458 int err;
2459
2460 pud = pud_alloc(mm, pgd, addr);
2461 if (!pud)
2462 return -ENOMEM;
2463 do {
2464 next = pud_addr_end(addr, end);
2465 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2466 if (err)
2467 break;
2468 } while (pud++, addr = next, addr != end);
2469 return err;
2470}
2471
2472/*
2473 * Scan a region of virtual memory, filling in page tables as necessary
2474 * and calling a provided function on each leaf page table.
2475 */
2476int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2477 unsigned long size, pte_fn_t fn, void *data)
2478{
2479 pgd_t *pgd;
2480 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002481 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002482 int err;
2483
2484 BUG_ON(addr >= end);
2485 pgd = pgd_offset(mm, addr);
2486 do {
2487 next = pgd_addr_end(addr, end);
2488 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2489 if (err)
2490 break;
2491 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002492
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002493 return err;
2494}
2495EXPORT_SYMBOL_GPL(apply_to_page_range);
2496
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002498 * handle_pte_fault chooses page fault handler according to an entry
2499 * which was read non-atomically. Before making any commitment, on
2500 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002501 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002502 * must check under lock before unmapping the pte and proceeding
2503 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002504 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002505 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002506static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002507 pte_t *page_table, pte_t orig_pte)
2508{
2509 int same = 1;
2510#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2511 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002512 spinlock_t *ptl = pte_lockptr(mm, pmd);
2513 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002514 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002515 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002516 }
2517#endif
2518 pte_unmap(page_table);
2519 return same;
2520}
2521
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002522static 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 -08002523{
2524 /*
2525 * If the source page was a PFN mapping, we don't have
2526 * a "struct page" for it. We do a best-effort copy by
2527 * just copying from the original user address. If that
2528 * fails, we just zero-fill it. Live with it.
2529 */
2530 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002531 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002532 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2533
2534 /*
2535 * This really shouldn't fail, because the page is there
2536 * in the page tables. But it might just be unreadable,
2537 * in which case we just give up and fill the result with
2538 * zeroes.
2539 */
2540 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002541 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002542 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002543 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002544 } else
2545 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002546}
2547
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 * This routine handles present pages, when users try to write
2550 * to a shared page. It is done by copying the page to a new address
2551 * and decrementing the shared-page counter for the old page.
2552 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 * Note that this routine assumes that the protection checks have been
2554 * done by the caller (the low-level page fault routine in most cases).
2555 * Thus we can safely just mark it writable once we've done any necessary
2556 * COW.
2557 *
2558 * We also mark the page dirty at this point even though the page will
2559 * change only once the write actually happens. This avoids a few races,
2560 * and potentially makes it more efficient.
2561 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002562 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2563 * but allow concurrent faults), with pte both mapped and locked.
2564 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002566static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2567 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002568 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002569 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002571 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002573 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002574 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002575 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002576 unsigned long mmun_start = 0; /* For mmu_notifiers */
2577 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578
Linus Torvalds6aab3412005-11-28 14:34:23 -08002579 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002580 if (!old_page) {
2581 /*
2582 * VM_MIXEDMAP !pfn_valid() case
2583 *
2584 * We should not cow pages in a shared writeable mapping.
2585 * Just mark the pages writable as we can't do any dirty
2586 * accounting on raw pfn maps.
2587 */
2588 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2589 (VM_WRITE|VM_SHARED))
2590 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002591 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002592 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002594 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002595 * Take out anonymous pages first, anonymous shared vmas are
2596 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002597 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002598 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002599 if (!trylock_page(old_page)) {
2600 page_cache_get(old_page);
2601 pte_unmap_unlock(page_table, ptl);
2602 lock_page(old_page);
2603 page_table = pte_offset_map_lock(mm, pmd, address,
2604 &ptl);
2605 if (!pte_same(*page_table, orig_pte)) {
2606 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002607 goto unlock;
2608 }
2609 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002610 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002611 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002612 /*
2613 * The page is all ours. Move it to our anon_vma so
2614 * the rmap code will not search our parent or siblings.
2615 * Protected against the rmap code by the page lock.
2616 */
2617 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002618 unlock_page(old_page);
2619 goto reuse;
2620 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002621 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002622 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002623 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002624 /*
2625 * Only catch write-faults on shared writable pages,
2626 * read-only shared pages can get COWed by
2627 * get_user_pages(.write=1, .force=1).
2628 */
David Howells9637a5e2006-06-23 02:03:43 -07002629 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002630 struct vm_fault vmf;
2631 int tmp;
2632
2633 vmf.virtual_address = (void __user *)(address &
2634 PAGE_MASK);
2635 vmf.pgoff = old_page->index;
2636 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2637 vmf.page = old_page;
2638
David Howells9637a5e2006-06-23 02:03:43 -07002639 /*
2640 * Notify the address space that the page is about to
2641 * become writable so that it can prohibit this or wait
2642 * for the page to get into an appropriate state.
2643 *
2644 * We do this without the lock held, so that it can
2645 * sleep if it needs to.
2646 */
2647 page_cache_get(old_page);
2648 pte_unmap_unlock(page_table, ptl);
2649
Nick Pigginc2ec1752009-03-31 15:23:21 -07002650 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2651 if (unlikely(tmp &
2652 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2653 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002654 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002655 }
Nick Pigginb827e492009-04-30 15:08:16 -07002656 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2657 lock_page(old_page);
2658 if (!old_page->mapping) {
2659 ret = 0; /* retry the fault */
2660 unlock_page(old_page);
2661 goto unwritable_page;
2662 }
2663 } else
2664 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002665
David Howells9637a5e2006-06-23 02:03:43 -07002666 /*
2667 * Since we dropped the lock we need to revalidate
2668 * the PTE as someone else may have changed it. If
2669 * they did, we just return, as we can count on the
2670 * MMU to tell us if they didn't also make it writable.
2671 */
2672 page_table = pte_offset_map_lock(mm, pmd, address,
2673 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002674 if (!pte_same(*page_table, orig_pte)) {
2675 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002676 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002677 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002678
2679 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002681 dirty_page = old_page;
2682 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002683
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002684reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002685 flush_cache_page(vma, address, pte_pfn(orig_pte));
2686 entry = pte_mkyoung(orig_pte);
2687 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002688 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002689 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002690 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002691 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002692
2693 if (!dirty_page)
2694 return ret;
2695
2696 /*
2697 * Yes, Virginia, this is actually required to prevent a race
2698 * with clear_page_dirty_for_io() from clearing the page dirty
2699 * bit after it clear all dirty ptes, but before a racing
2700 * do_wp_page installs a dirty pte.
2701 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002702 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002703 */
2704 if (!page_mkwrite) {
2705 wait_on_page_locked(dirty_page);
2706 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002707 /* file_update_time outside page_lock */
2708 if (vma->vm_file)
2709 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002710 }
2711 put_page(dirty_page);
2712 if (page_mkwrite) {
2713 struct address_space *mapping = dirty_page->mapping;
2714
2715 set_page_dirty(dirty_page);
2716 unlock_page(dirty_page);
2717 page_cache_release(dirty_page);
2718 if (mapping) {
2719 /*
2720 * Some device drivers do not set page.mapping
2721 * but still dirty their pages
2722 */
2723 balance_dirty_pages_ratelimited(mapping);
2724 }
2725 }
2726
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002727 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729
2730 /*
2731 * Ok, we need to copy. Oh, well..
2732 */
Nick Pigginb5810032005-10-29 18:16:12 -07002733 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002734gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002735 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736
2737 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002738 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002739
Hugh Dickins62eede62009-09-21 17:03:34 -07002740 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002741 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2742 if (!new_page)
2743 goto oom;
2744 } else {
2745 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2746 if (!new_page)
2747 goto oom;
2748 cow_user_page(new_page, old_page, address, vma);
2749 }
2750 __SetPageUptodate(new_page);
2751
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002752 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002753 goto oom_free_new;
2754
Haggai Eran6bdb9132012-10-08 16:33:35 -07002755 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002756 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002757 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2758
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 /*
2760 * Re-check the pte - we dropped the lock
2761 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002762 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002763 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002764 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002765 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002766 dec_mm_counter_fast(mm, MM_FILEPAGES);
2767 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002768 }
2769 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002770 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002771 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002772 entry = mk_pte(new_page, vma->vm_page_prot);
2773 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002774 /*
2775 * Clear the pte entry and flush it first, before updating the
2776 * pte with the new entry. This will avoid a race condition
2777 * seen in the presence of one thread doing SMC and another
2778 * thread doing COW.
2779 */
Izik Eidus828502d2009-09-21 17:01:51 -07002780 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002781 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002782 /*
2783 * We call the notify macro here because, when using secondary
2784 * mmu page tables (such as kvm shadow page tables), we want the
2785 * new page to be mapped directly into the secondary page table.
2786 */
2787 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002788 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002789 if (old_page) {
2790 /*
2791 * Only after switching the pte to the new page may
2792 * we remove the mapcount here. Otherwise another
2793 * process may come and find the rmap count decremented
2794 * before the pte is switched to the new page, and
2795 * "reuse" the old page writing into it while our pte
2796 * here still points into it and can be read by other
2797 * threads.
2798 *
2799 * The critical issue is to order this
2800 * page_remove_rmap with the ptp_clear_flush above.
2801 * Those stores are ordered by (if nothing else,)
2802 * the barrier present in the atomic_add_negative
2803 * in page_remove_rmap.
2804 *
2805 * Then the TLB flush in ptep_clear_flush ensures that
2806 * no process can access the old page before the
2807 * decremented mapcount is visible. And the old page
2808 * cannot be reused until after the decremented
2809 * mapcount is visible. So transitively, TLBs to
2810 * old page will be flushed before it can be reused.
2811 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002812 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002813 }
2814
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 /* Free the old page.. */
2816 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002817 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002818 } else
2819 mem_cgroup_uncharge_page(new_page);
2820
Haggai Eran6bdb9132012-10-08 16:33:35 -07002821 if (new_page)
2822 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002823unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002824 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002825 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002826 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002827 if (old_page) {
2828 /*
2829 * Don't let another task, with possibly unlocked vma,
2830 * keep the mlocked page.
2831 */
2832 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2833 lock_page(old_page); /* LRU manipulation */
2834 munlock_vma_page(old_page);
2835 unlock_page(old_page);
2836 }
2837 page_cache_release(old_page);
2838 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002839 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002840oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002841 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002842oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002843 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002844 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002846
2847unwritable_page:
2848 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002849 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850}
2851
Peter Zijlstra97a89412011-05-24 17:12:04 -07002852static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 unsigned long start_addr, unsigned long end_addr,
2854 struct zap_details *details)
2855{
Al Virof5cc4ee2012-03-05 14:14:20 -05002856 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857}
2858
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002859static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 struct zap_details *details)
2861{
2862 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 pgoff_t vba, vea, zba, zea;
2864
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002865 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
2868 vba = vma->vm_pgoff;
2869 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2870 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2871 zba = details->first_index;
2872 if (zba < vba)
2873 zba = vba;
2874 zea = details->last_index;
2875 if (zea > vea)
2876 zea = vea;
2877
Peter Zijlstra97a89412011-05-24 17:12:04 -07002878 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2880 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002881 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 }
2883}
2884
2885static inline void unmap_mapping_range_list(struct list_head *head,
2886 struct zap_details *details)
2887{
2888 struct vm_area_struct *vma;
2889
2890 /*
2891 * In nonlinear VMAs there is no correspondence between virtual address
2892 * offset and file offset. So we must perform an exhaustive search
2893 * across *all* the pages in each nonlinear VMA, not just the pages
2894 * whose virtual address lies outside the file truncation point.
2895 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002896 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002898 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 }
2900}
2901
2902/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002903 * 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 -07002904 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 * @holebegin: byte in first page to unmap, relative to the start of
2906 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002907 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 * must keep the partial page. In contrast, we must get rid of
2909 * partial pages.
2910 * @holelen: size of prospective hole in bytes. This will be rounded
2911 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2912 * end of the file.
2913 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2914 * but 0 when invalidating pagecache, don't throw away private data.
2915 */
2916void unmap_mapping_range(struct address_space *mapping,
2917 loff_t const holebegin, loff_t const holelen, int even_cows)
2918{
2919 struct zap_details details;
2920 pgoff_t hba = holebegin >> PAGE_SHIFT;
2921 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2922
2923 /* Check for overflow. */
2924 if (sizeof(holelen) > sizeof(hlen)) {
2925 long long holeend =
2926 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2927 if (holeend & ~(long long)ULONG_MAX)
2928 hlen = ULONG_MAX - hba + 1;
2929 }
2930
2931 details.check_mapping = even_cows? NULL: mapping;
2932 details.nonlinear_vma = NULL;
2933 details.first_index = hba;
2934 details.last_index = hba + hlen - 1;
2935 if (details.last_index < details.first_index)
2936 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
Peter Zijlstra97a89412011-05-24 17:12:04 -07002938
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002939 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002940 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2942 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2943 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002944 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945}
2946EXPORT_SYMBOL(unmap_mapping_range);
2947
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002949 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2950 * but allow concurrent faults), and pte mapped but not yet locked.
2951 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002953static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2954 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002955 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002957 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002958 struct page *page, *swapcache;
Hugh Dickins65500d22005-10-29 18:15:59 -07002959 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002961 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002962 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002963 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002964 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002966 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002967 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002968
2969 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002970 if (unlikely(non_swap_entry(entry))) {
2971 if (is_migration_entry(entry)) {
2972 migration_entry_wait(mm, pmd, address);
2973 } else if (is_hwpoison_entry(entry)) {
2974 ret = VM_FAULT_HWPOISON;
2975 } else {
2976 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002977 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002978 }
Christoph Lameter06972122006-06-23 02:03:35 -07002979 goto out;
2980 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002981 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 page = lookup_swap_cache(entry);
2983 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002984 page = swapin_readahead(entry,
2985 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 if (!page) {
2987 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002988 * Back out if somebody else faulted in this pte
2989 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002991 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 if (likely(pte_same(*page_table, orig_pte)))
2993 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002994 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002995 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 }
2997
2998 /* Had to read the page from swap area: Major fault */
2999 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07003000 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003001 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02003002 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01003003 /*
3004 * hwpoisoned dirty swapcache pages are kept for killing
3005 * owner processes (which may be unknown at hwpoison time)
3006 */
Andi Kleend1737fd2009-09-16 11:50:06 +02003007 ret = VM_FAULT_HWPOISON;
3008 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08003009 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02003010 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 }
3012
Hugh Dickins56f31802013-02-22 16:36:10 -08003013 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003014 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07003015
Balbir Singh20a10222007-11-14 17:00:33 -08003016 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003017 if (!locked) {
3018 ret |= VM_FAULT_RETRY;
3019 goto out_release;
3020 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003022 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003023 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3024 * release the swapcache from under us. The page pin, and pte_same
3025 * test below, are not enough to exclude that. Even if it is still
3026 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003027 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003028 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003029 goto out_page;
3030
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003031 page = ksm_might_need_to_copy(page, vma, address);
3032 if (unlikely(!page)) {
3033 ret = VM_FAULT_OOM;
3034 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003035 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08003036 }
3037
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003038 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003039 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003040 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003041 }
3042
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003044 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003046 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003047 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003048 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003049
3050 if (unlikely(!PageUptodate(page))) {
3051 ret = VM_FAULT_SIGBUS;
3052 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 }
3054
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003055 /*
3056 * The page isn't present yet, go ahead with the fault.
3057 *
3058 * Be careful about the sequence of operations here.
3059 * To get its accounting right, reuse_swap_page() must be called
3060 * while the page is counted on swap but not yet in mapcount i.e.
3061 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3062 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003063 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3064 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3065 * in page->private. In this case, a record in swap_cgroup is silently
3066 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003067 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003069 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003070 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003072 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003074 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003075 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003076 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 flush_icache_page(vma, page);
3079 set_pte_at(mm, address, page_table, pte);
Hugh Dickins56f31802013-02-22 16:36:10 -08003080 if (page == swapcache)
Johannes Weineraf347702013-02-22 16:32:20 -08003081 do_page_add_anon_rmap(page, vma, address, exclusive);
Hugh Dickins56f31802013-02-22 16:36:10 -08003082 else /* ksm created a completely new copy */
3083 page_add_new_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003084 /* It's better to call commit-charge after rmap is established */
3085 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003087 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003088 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003089 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003090 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003091 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003092 /*
3093 * Hold the lock to avoid the swap entry to be reused
3094 * until we take the PT lock for the pte_same() check
3095 * (to avoid false positives from pte_same). For
3096 * further safety release the lock after the swap_free
3097 * so that the swap count won't change under a
3098 * parallel locked swapcache.
3099 */
3100 unlock_page(swapcache);
3101 page_cache_release(swapcache);
3102 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003103
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003104 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003105 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3106 if (ret & VM_FAULT_ERROR)
3107 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 goto out;
3109 }
3110
3111 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003112 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003113unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003114 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115out:
3116 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003117out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003118 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003119 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003120out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003121 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003122out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003123 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003124 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003125 unlock_page(swapcache);
3126 page_cache_release(swapcache);
3127 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003128 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129}
3130
3131/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003132 * This is like a special single-page "expand_{down|up}wards()",
3133 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003134 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003135 */
3136static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3137{
3138 address &= PAGE_MASK;
3139 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003140 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003141
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003142 /*
3143 * Is there a mapping abutting this one below?
3144 *
3145 * That's only ok if it's the same stack mapping
3146 * that has gotten split..
3147 */
3148 if (prev && prev->vm_end == address)
3149 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3150
Michal Hockod05f3162011-05-24 17:11:44 -07003151 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003152 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003153 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3154 struct vm_area_struct *next = vma->vm_next;
3155
3156 /* As VM_GROWSDOWN but s/below/above/ */
3157 if (next && next->vm_start == address + PAGE_SIZE)
3158 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3159
3160 expand_upwards(vma, address + PAGE_SIZE);
3161 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003162 return 0;
3163}
3164
3165/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003166 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3167 * but allow concurrent faults), and pte mapped but not yet locked.
3168 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003170static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3171 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003172 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003174 struct page *page;
3175 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177
Linus Torvalds11ac5522010-08-14 11:44:56 -07003178 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003179
Linus Torvalds11ac5522010-08-14 11:44:56 -07003180 /* Check if we need to add a guard page to the stack */
3181 if (check_stack_guard_page(vma, address) < 0)
3182 return VM_FAULT_SIGBUS;
3183
3184 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003185 if (!(flags & FAULT_FLAG_WRITE)) {
3186 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3187 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003188 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003189 if (!pte_none(*page_table))
3190 goto unlock;
3191 goto setpte;
3192 }
3193
Nick Piggin557ed1f2007-10-16 01:24:40 -07003194 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003195 if (unlikely(anon_vma_prepare(vma)))
3196 goto oom;
3197 page = alloc_zeroed_user_highpage_movable(vma, address);
3198 if (!page)
3199 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003200 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003202 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003203 goto oom_free_page;
3204
Nick Piggin557ed1f2007-10-16 01:24:40 -07003205 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003206 if (vma->vm_flags & VM_WRITE)
3207 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003208
Nick Piggin557ed1f2007-10-16 01:24:40 -07003209 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003210 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003211 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003212
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003213 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003214 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003215setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003216 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217
3218 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003219 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003220unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003221 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003222 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003223release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003224 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003225 page_cache_release(page);
3226 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003227oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003228 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003229oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 return VM_FAULT_OOM;
3231}
3232
3233/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003234 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003236 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3237 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238 *
3239 * As this is called only for pages that do not currently exist, we
3240 * do not need to flush old virtual caches or the TLB.
3241 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003242 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003243 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003244 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003246static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003247 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003248 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003250 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003251 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003252 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003253 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003256 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003257 struct vm_fault vmf;
3258 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003259 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003260
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003261 /*
3262 * If we do COW later, allocate page befor taking lock_page()
3263 * on the file cache page. This will reduce lock holding time.
3264 */
3265 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3266
3267 if (unlikely(anon_vma_prepare(vma)))
3268 return VM_FAULT_OOM;
3269
3270 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3271 if (!cow_page)
3272 return VM_FAULT_OOM;
3273
3274 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3275 page_cache_release(cow_page);
3276 return VM_FAULT_OOM;
3277 }
3278 } else
3279 cow_page = NULL;
3280
Nick Piggind0217ac2007-07-19 01:47:03 -07003281 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3282 vmf.pgoff = pgoff;
3283 vmf.flags = flags;
3284 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003286 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003287 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3288 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003289 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Andi Kleena3b947e2009-09-16 11:50:08 +02003291 if (unlikely(PageHWPoison(vmf.page))) {
3292 if (ret & VM_FAULT_LOCKED)
3293 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003294 ret = VM_FAULT_HWPOISON;
3295 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003296 }
3297
Nick Piggind00806b2007-07-19 01:46:57 -07003298 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003299 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003300 * locked.
3301 */
Nick Piggin83c54072007-07-19 01:47:05 -07003302 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003303 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003304 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003305 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003306
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307 /*
3308 * Should we do an early C-O-W break?
3309 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003310 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003311 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003312 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003313 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003314 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003315 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003316 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003317 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003318 /*
3319 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003320 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003321 * to become writable
3322 */
Mark Fasheh69676142007-07-19 01:47:00 -07003323 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003324 int tmp;
3325
Mark Fasheh69676142007-07-19 01:47:00 -07003326 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003327 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003328 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3329 if (unlikely(tmp &
3330 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3331 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003332 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003333 }
Nick Pigginb827e492009-04-30 15:08:16 -07003334 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3335 lock_page(page);
3336 if (!page->mapping) {
3337 ret = 0; /* retry the fault */
3338 unlock_page(page);
3339 goto unwritable_page;
3340 }
3341 } else
3342 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003343 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003344 }
3345 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003346
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 }
3348
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003349 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350
3351 /*
3352 * This silly early PAGE_DIRTY setting removes a race
3353 * due to the bad i386 page protection. But it's valid
3354 * for other architectures too.
3355 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003356 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 * an exclusive copy of the page, or this is a shared mapping,
3358 * so we can make it writable and dirty to avoid having to
3359 * handle that later.
3360 */
3361 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003362 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003363 flush_icache_page(vma, page);
3364 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003365 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003368 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003369 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003370 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003371 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003372 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003373 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003374 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003375 get_page(dirty_page);
3376 }
Hugh Dickins42946212005-10-29 18:16:05 -07003377 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003378 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003379
3380 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003381 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003383 if (cow_page)
3384 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003385 if (anon)
3386 page_cache_release(page);
3387 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003388 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389 }
3390
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003391 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003392
Nick Pigginb827e492009-04-30 15:08:16 -07003393 if (dirty_page) {
3394 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003395 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003396
3397 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003398 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003399 unlock_page(dirty_page);
3400 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003401 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003402 /*
3403 * Some device drivers do not set page.mapping but still
3404 * dirty their pages
3405 */
3406 balance_dirty_pages_ratelimited(mapping);
3407 }
3408
3409 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003410 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003411 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003412 } else {
3413 unlock_page(vmf.page);
3414 if (anon)
3415 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003416 }
Nick Piggind00806b2007-07-19 01:46:57 -07003417
Nick Piggin83c54072007-07-19 01:47:05 -07003418 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003419
3420unwritable_page:
3421 page_cache_release(page);
3422 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003423uncharge_out:
3424 /* fs's fault handler get error */
3425 if (cow_page) {
3426 mem_cgroup_uncharge_page(cow_page);
3427 page_cache_release(cow_page);
3428 }
3429 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003430}
Nick Piggind00806b2007-07-19 01:46:57 -07003431
Nick Piggin54cb8822007-07-19 01:46:59 -07003432static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3433 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003434 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003435{
3436 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003437 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003438
Hugh Dickins16abfa02007-10-04 16:56:06 +01003439 pte_unmap(page_table);
3440 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003441}
3442
Jes Sorensenf4b81802006-09-27 01:50:10 -07003443/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 * Fault of a previously existing named mapping. Repopulate the pte
3445 * from the encoded file_pte if possible. This enables swappable
3446 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003447 *
3448 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3449 * but allow concurrent faults), and pte mapped but not yet locked.
3450 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003452static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003453 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003454 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
Hugh Dickins65500d22005-10-29 18:15:59 -07003456 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003458 flags |= FAULT_FLAG_NONLINEAR;
3459
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003460 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003461 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462
Hugh Dickins2509ef22009-01-06 14:40:10 -08003463 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003464 /*
3465 * Page table corrupted: show pte and kill process.
3466 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003467 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003468 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003469 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003470
3471 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003472 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473}
3474
Mel Gorman9532fec2012-11-15 01:24:32 +00003475int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3476 unsigned long addr, int current_nid)
3477{
3478 get_page(page);
3479
3480 count_vm_numa_event(NUMA_HINT_FAULTS);
3481 if (current_nid == numa_node_id())
3482 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3483
3484 return mpol_misplaced(page, vma, addr);
3485}
3486
Mel Gormand10e63f2012-10-25 14:16:31 +02003487int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3488 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3489{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003490 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003491 spinlock_t *ptl;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003492 int current_nid = -1;
3493 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003494 bool migrated = false;
Mel Gormand10e63f2012-10-25 14:16:31 +02003495
3496 /*
3497 * The "pte" at this point cannot be used safely without
3498 * validation through pte_unmap_same(). It's of NUMA type but
3499 * the pfn may be screwed if the read is non atomic.
3500 *
3501 * ptep_modify_prot_start is not called as this is clearing
3502 * the _PAGE_NUMA bit and it is not really expected that there
3503 * would be concurrent hardware modifications to the PTE.
3504 */
3505 ptl = pte_lockptr(mm, pmd);
3506 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003507 if (unlikely(!pte_same(*ptep, pte))) {
3508 pte_unmap_unlock(ptep, ptl);
3509 goto out;
3510 }
3511
Mel Gormand10e63f2012-10-25 14:16:31 +02003512 pte = pte_mknonnuma(pte);
3513 set_pte_at(mm, addr, ptep, pte);
3514 update_mmu_cache(vma, addr, ptep);
3515
3516 page = vm_normal_page(vma, addr, pte);
3517 if (!page) {
3518 pte_unmap_unlock(ptep, ptl);
3519 return 0;
3520 }
3521
Mel Gorman4daae3b2012-11-02 11:33:45 +00003522 current_nid = page_to_nid(page);
Mel Gorman9532fec2012-11-15 01:24:32 +00003523 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
Mel Gormand10e63f2012-10-25 14:16:31 +02003524 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003525 if (target_nid == -1) {
3526 /*
3527 * Account for the fault against the current node if it not
3528 * being replaced regardless of where the page is located.
3529 */
3530 current_nid = numa_node_id();
3531 put_page(page);
3532 goto out;
3533 }
3534
3535 /* Migrate to the requested node */
Mel Gormanb8593bf2012-11-21 01:18:23 +00003536 migrated = migrate_misplaced_page(page, target_nid);
3537 if (migrated)
Mel Gorman4daae3b2012-11-02 11:33:45 +00003538 current_nid = target_nid;
3539
3540out:
Mel Gorman9532fec2012-11-15 01:24:32 +00003541 if (current_nid != -1)
Mel Gormanb8593bf2012-11-21 01:18:23 +00003542 task_numa_fault(current_nid, 1, migrated);
Mel Gormand10e63f2012-10-25 14:16:31 +02003543 return 0;
3544}
3545
3546/* NUMA hinting page fault entry point for regular pmds */
3547#ifdef CONFIG_NUMA_BALANCING
3548static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3549 unsigned long addr, pmd_t *pmdp)
3550{
3551 pmd_t pmd;
3552 pte_t *pte, *orig_pte;
3553 unsigned long _addr = addr & PMD_MASK;
3554 unsigned long offset;
3555 spinlock_t *ptl;
3556 bool numa = false;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00003557 int local_nid = numa_node_id();
Mel Gormand10e63f2012-10-25 14:16:31 +02003558
3559 spin_lock(&mm->page_table_lock);
3560 pmd = *pmdp;
3561 if (pmd_numa(pmd)) {
3562 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3563 numa = true;
3564 }
3565 spin_unlock(&mm->page_table_lock);
3566
3567 if (!numa)
3568 return 0;
3569
3570 /* we're in a page fault so some vma must be in the range */
3571 BUG_ON(!vma);
3572 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3573 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3574 VM_BUG_ON(offset >= PMD_SIZE);
3575 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3576 pte += offset >> PAGE_SHIFT;
3577 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3578 pte_t pteval = *pte;
3579 struct page *page;
Mel Gorman9532fec2012-11-15 01:24:32 +00003580 int curr_nid = local_nid;
3581 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003582 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02003583 if (!pte_present(pteval))
3584 continue;
3585 if (!pte_numa(pteval))
3586 continue;
3587 if (addr >= vma->vm_end) {
3588 vma = find_vma(mm, addr);
3589 /* there's a pte present so there must be a vma */
3590 BUG_ON(!vma);
3591 BUG_ON(addr < vma->vm_start);
3592 }
3593 if (pte_numa(pteval)) {
3594 pteval = pte_mknonnuma(pteval);
3595 set_pte_at(mm, addr, pte, pteval);
3596 }
3597 page = vm_normal_page(vma, addr, pteval);
3598 if (unlikely(!page))
3599 continue;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003600 /* only check non-shared pages */
3601 if (unlikely(page_mapcount(page) != 1))
3602 continue;
Mel Gorman9532fec2012-11-15 01:24:32 +00003603
3604 /*
3605 * Note that the NUMA fault is later accounted to either
3606 * the node that is currently running or where the page is
3607 * migrated to.
3608 */
3609 curr_nid = local_nid;
3610 target_nid = numa_migrate_prep(page, vma, addr,
3611 page_to_nid(page));
3612 if (target_nid == -1) {
3613 put_page(page);
3614 continue;
3615 }
3616
3617 /* Migrate to the requested node */
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003618 pte_unmap_unlock(pte, ptl);
Mel Gormanb8593bf2012-11-21 01:18:23 +00003619 migrated = migrate_misplaced_page(page, target_nid);
3620 if (migrated)
Mel Gorman9532fec2012-11-15 01:24:32 +00003621 curr_nid = target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003622 task_numa_fault(curr_nid, 1, migrated);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003623
3624 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003625 }
3626 pte_unmap_unlock(orig_pte, ptl);
3627
3628 return 0;
3629}
3630#else
3631static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3632 unsigned long addr, pmd_t *pmdp)
3633{
3634 BUG();
Andrew Mortonb3dd2072012-12-17 15:59:24 -08003635 return 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003636}
3637#endif /* CONFIG_NUMA_BALANCING */
3638
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639/*
3640 * These routines also need to handle stuff like marking pages dirty
3641 * and/or accessed for architectures that don't do it in hardware (most
3642 * RISC architectures). The early dirtying is also good on the i386.
3643 *
3644 * There is also a hook called "update_mmu_cache()" that architectures
3645 * with external mmu caches can use to update those (ie the Sparc or
3646 * PowerPC hashed page tables that act as extended TLBs).
3647 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003648 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3649 * but allow concurrent faults), and pte mapped but not yet locked.
3650 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003652int handle_pte_fault(struct mm_struct *mm,
3653 struct vm_area_struct *vma, unsigned long address,
3654 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655{
3656 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003657 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003659 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003661 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003662 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003663 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003664 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003665 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003666 }
3667 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003668 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003669 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003671 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003672 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003673 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003674 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 }
3676
Mel Gormand10e63f2012-10-25 14:16:31 +02003677 if (pte_numa(entry))
3678 return do_numa_page(mm, vma, address, entry, pte, pmd);
3679
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003680 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003681 spin_lock(ptl);
3682 if (unlikely(!pte_same(*pte, entry)))
3683 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003684 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003686 return do_wp_page(mm, vma, address,
3687 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688 entry = pte_mkdirty(entry);
3689 }
3690 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003691 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003692 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003693 } else {
3694 /*
3695 * This is needed only for protection faults but the arch code
3696 * is not yet telling us if this is a protection fault or not.
3697 * This still avoids useless tlb flushes for .text page faults
3698 * with threads.
3699 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003700 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003701 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003702 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003703unlock:
3704 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003705 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706}
3707
3708/*
3709 * By the time we get here, we already hold the mm semaphore
3710 */
Nick Piggin83c54072007-07-19 01:47:05 -07003711int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003712 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713{
3714 pgd_t *pgd;
3715 pud_t *pud;
3716 pmd_t *pmd;
3717 pte_t *pte;
3718
3719 __set_current_state(TASK_RUNNING);
3720
Christoph Lameterf8891e52006-06-30 01:55:45 -07003721 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003722 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003724 /* do counter updates before entering really critical section. */
3725 check_sync_rss_stat(current);
3726
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003727 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003728 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729
David Rientjes1f1d06c2012-05-29 15:06:23 -07003730retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732 pud = pud_alloc(mm, pgd, address);
3733 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003734 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 pmd = pmd_alloc(mm, pud, address);
3736 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003737 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003738 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3739 if (!vma->vm_ops)
3740 return do_huge_pmd_anonymous_page(mm, vma, address,
3741 pmd, flags);
3742 } else {
3743 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003744 int ret;
3745
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003746 barrier();
3747 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003748 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3749
Linus Torvaldse53289c2013-01-09 08:36:54 -08003750 /*
3751 * If the pmd is splitting, return and retry the
3752 * the fault. Alternative: wait until the split
3753 * is done, and goto retry.
3754 */
3755 if (pmd_trans_splitting(orig_pmd))
3756 return 0;
3757
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003758 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003759 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003760 orig_pmd, pmd);
3761
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003762 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003763 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3764 orig_pmd);
3765 /*
3766 * If COW results in an oom, the huge pmd will
3767 * have been split, so retry the fault on the
3768 * pte for a smaller charge.
3769 */
3770 if (unlikely(ret & VM_FAULT_OOM))
3771 goto retry;
3772 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003773 } else {
3774 huge_pmd_set_accessed(mm, vma, address, pmd,
3775 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003776 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003777
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003778 return 0;
3779 }
3780 }
3781
Mel Gormand10e63f2012-10-25 14:16:31 +02003782 if (pmd_numa(*pmd))
3783 return do_pmd_numa_page(mm, vma, address, pmd);
3784
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003785 /*
3786 * Use __pte_alloc instead of pte_alloc_map, because we can't
3787 * run pte_offset_map on the pmd, if an huge pmd could
3788 * materialize from under us from a different thread.
3789 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003790 if (unlikely(pmd_none(*pmd)) &&
3791 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003792 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003793 /* if an huge pmd materialized from under us just retry later */
3794 if (unlikely(pmd_trans_huge(*pmd)))
3795 return 0;
3796 /*
3797 * A regular pmd is established and it can't morph into a huge pmd
3798 * from under us anymore at this point because we hold the mmap_sem
3799 * read mode and khugepaged takes it in write mode. So now it's
3800 * safe to run pte_offset_map().
3801 */
3802 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003804 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805}
3806
3807#ifndef __PAGETABLE_PUD_FOLDED
3808/*
3809 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003810 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003812int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003814 pud_t *new = pud_alloc_one(mm, address);
3815 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003816 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817
Nick Piggin362a61a2008-05-14 06:37:36 +02003818 smp_wmb(); /* See comment in __pte_alloc */
3819
Hugh Dickins872fec12005-10-29 18:16:21 -07003820 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003821 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003822 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003823 else
3824 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003825 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003826 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827}
3828#endif /* __PAGETABLE_PUD_FOLDED */
3829
3830#ifndef __PAGETABLE_PMD_FOLDED
3831/*
3832 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003833 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003835int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003837 pmd_t *new = pmd_alloc_one(mm, address);
3838 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003839 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840
Nick Piggin362a61a2008-05-14 06:37:36 +02003841 smp_wmb(); /* See comment in __pte_alloc */
3842
Hugh Dickins872fec12005-10-29 18:16:21 -07003843 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003845 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003846 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003847 else
3848 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003850 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003851 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003852 else
3853 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003855 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003856 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857}
3858#endif /* __PAGETABLE_PMD_FOLDED */
3859
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860#if !defined(__HAVE_ARCH_GATE_AREA)
3861
3862#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003863static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864
3865static int __init gate_vma_init(void)
3866{
3867 gate_vma.vm_mm = NULL;
3868 gate_vma.vm_start = FIXADDR_USER_START;
3869 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003870 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3871 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003872
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 return 0;
3874}
3875__initcall(gate_vma_init);
3876#endif
3877
Stephen Wilson31db58b2011-03-13 15:49:15 -04003878struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879{
3880#ifdef AT_SYSINFO_EHDR
3881 return &gate_vma;
3882#else
3883 return NULL;
3884#endif
3885}
3886
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003887int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888{
3889#ifdef AT_SYSINFO_EHDR
3890 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3891 return 1;
3892#endif
3893 return 0;
3894}
3895
3896#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003897
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003898static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003899 pte_t **ptepp, spinlock_t **ptlp)
3900{
3901 pgd_t *pgd;
3902 pud_t *pud;
3903 pmd_t *pmd;
3904 pte_t *ptep;
3905
3906 pgd = pgd_offset(mm, address);
3907 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3908 goto out;
3909
3910 pud = pud_offset(pgd, address);
3911 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3912 goto out;
3913
3914 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003915 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003916 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3917 goto out;
3918
3919 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3920 if (pmd_huge(*pmd))
3921 goto out;
3922
3923 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3924 if (!ptep)
3925 goto out;
3926 if (!pte_present(*ptep))
3927 goto unlock;
3928 *ptepp = ptep;
3929 return 0;
3930unlock:
3931 pte_unmap_unlock(ptep, *ptlp);
3932out:
3933 return -EINVAL;
3934}
3935
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003936static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3937 pte_t **ptepp, spinlock_t **ptlp)
3938{
3939 int res;
3940
3941 /* (void) is needed to make gcc happy */
3942 (void) __cond_lock(*ptlp,
3943 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3944 return res;
3945}
3946
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003947/**
3948 * follow_pfn - look up PFN at a user virtual address
3949 * @vma: memory mapping
3950 * @address: user virtual address
3951 * @pfn: location to store found PFN
3952 *
3953 * Only IO mappings and raw PFN mappings are allowed.
3954 *
3955 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3956 */
3957int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3958 unsigned long *pfn)
3959{
3960 int ret = -EINVAL;
3961 spinlock_t *ptl;
3962 pte_t *ptep;
3963
3964 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3965 return ret;
3966
3967 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3968 if (ret)
3969 return ret;
3970 *pfn = pte_pfn(*ptep);
3971 pte_unmap_unlock(ptep, ptl);
3972 return 0;
3973}
3974EXPORT_SYMBOL(follow_pfn);
3975
Rik van Riel28b2ee22008-07-23 21:27:05 -07003976#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003977int follow_phys(struct vm_area_struct *vma,
3978 unsigned long address, unsigned int flags,
3979 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003980{
Johannes Weiner03668a42009-06-16 15:32:34 -07003981 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003982 pte_t *ptep, pte;
3983 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003984
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003985 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3986 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003987
Johannes Weiner03668a42009-06-16 15:32:34 -07003988 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003989 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003990 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003991
Rik van Riel28b2ee22008-07-23 21:27:05 -07003992 if ((flags & FOLL_WRITE) && !pte_write(pte))
3993 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003994
3995 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003996 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003997
Johannes Weiner03668a42009-06-16 15:32:34 -07003998 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003999unlock:
4000 pte_unmap_unlock(ptep, ptl);
4001out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004002 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004003}
4004
4005int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4006 void *buf, int len, int write)
4007{
4008 resource_size_t phys_addr;
4009 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08004010 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004011 int offset = addr & (PAGE_SIZE-1);
4012
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004013 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004014 return -EINVAL;
4015
4016 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4017 if (write)
4018 memcpy_toio(maddr + offset, buf, len);
4019 else
4020 memcpy_fromio(buf, maddr + offset, len);
4021 iounmap(maddr);
4022
4023 return len;
4024}
4025#endif
4026
David Howells0ec76a12006-09-27 01:50:15 -07004027/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004028 * Access another process' address space as given in mm. If non-NULL, use the
4029 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004030 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004031static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4032 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004033{
David Howells0ec76a12006-09-27 01:50:15 -07004034 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004035 void *old_buf = buf;
4036
David Howells0ec76a12006-09-27 01:50:15 -07004037 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004038 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004039 while (len) {
4040 int bytes, ret, offset;
4041 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004042 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004043
4044 ret = get_user_pages(tsk, mm, addr, 1,
4045 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004046 if (ret <= 0) {
4047 /*
4048 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4049 * we can access using slightly different code.
4050 */
4051#ifdef CONFIG_HAVE_IOREMAP_PROT
4052 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004053 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004054 break;
4055 if (vma->vm_ops && vma->vm_ops->access)
4056 ret = vma->vm_ops->access(vma, addr, buf,
4057 len, write);
4058 if (ret <= 0)
4059#endif
4060 break;
4061 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004062 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004063 bytes = len;
4064 offset = addr & (PAGE_SIZE-1);
4065 if (bytes > PAGE_SIZE-offset)
4066 bytes = PAGE_SIZE-offset;
4067
4068 maddr = kmap(page);
4069 if (write) {
4070 copy_to_user_page(vma, page, addr,
4071 maddr + offset, buf, bytes);
4072 set_page_dirty_lock(page);
4073 } else {
4074 copy_from_user_page(vma, page, addr,
4075 buf, maddr + offset, bytes);
4076 }
4077 kunmap(page);
4078 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004079 }
David Howells0ec76a12006-09-27 01:50:15 -07004080 len -= bytes;
4081 buf += bytes;
4082 addr += bytes;
4083 }
4084 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004085
4086 return buf - old_buf;
4087}
Andi Kleen03252912008-01-30 13:33:18 +01004088
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004089/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004090 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004091 * @mm: the mm_struct of the target address space
4092 * @addr: start address to access
4093 * @buf: source or destination buffer
4094 * @len: number of bytes to transfer
4095 * @write: whether the access is a write
4096 *
4097 * The caller must hold a reference on @mm.
4098 */
4099int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4100 void *buf, int len, int write)
4101{
4102 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4103}
4104
Andi Kleen03252912008-01-30 13:33:18 +01004105/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004106 * Access another process' address space.
4107 * Source/target buffer must be kernel space,
4108 * Do not walk the page table directly, use get_user_pages
4109 */
4110int access_process_vm(struct task_struct *tsk, unsigned long addr,
4111 void *buf, int len, int write)
4112{
4113 struct mm_struct *mm;
4114 int ret;
4115
4116 mm = get_task_mm(tsk);
4117 if (!mm)
4118 return 0;
4119
4120 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4121 mmput(mm);
4122
4123 return ret;
4124}
4125
Andi Kleen03252912008-01-30 13:33:18 +01004126/*
4127 * Print the name of a VMA.
4128 */
4129void print_vma_addr(char *prefix, unsigned long ip)
4130{
4131 struct mm_struct *mm = current->mm;
4132 struct vm_area_struct *vma;
4133
Ingo Molnare8bff742008-02-13 20:21:06 +01004134 /*
4135 * Do not print if we are in atomic
4136 * contexts (in exception stacks, etc.):
4137 */
4138 if (preempt_count())
4139 return;
4140
Andi Kleen03252912008-01-30 13:33:18 +01004141 down_read(&mm->mmap_sem);
4142 vma = find_vma(mm, ip);
4143 if (vma && vma->vm_file) {
4144 struct file *f = vma->vm_file;
4145 char *buf = (char *)__get_free_page(GFP_KERNEL);
4146 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004147 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004148
Jan Blunckcf28b482008-02-14 19:38:44 -08004149 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004150 if (IS_ERR(p))
4151 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004152 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004153 vma->vm_start,
4154 vma->vm_end - vma->vm_start);
4155 free_page((unsigned long)buf);
4156 }
4157 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004158 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004159}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004160
4161#ifdef CONFIG_PROVE_LOCKING
4162void might_fault(void)
4163{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004164 /*
4165 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4166 * holding the mmap_sem, this is safe because kernel memory doesn't
4167 * get paged out, therefore we'll never actually fault, and the
4168 * below annotations will generate false positives.
4169 */
4170 if (segment_eq(get_fs(), KERNEL_DS))
4171 return;
4172
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004173 might_sleep();
4174 /*
4175 * it would be nicer only to annotate paths which are not under
4176 * pagefault_disable, however that requires a larger audit and
4177 * providing helpers like get_user_atomic.
4178 */
4179 if (!in_atomic() && current->mm)
4180 might_lock_read(&current->mm->mmap_sem);
4181}
4182EXPORT_SYMBOL(might_fault);
4183#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004184
4185#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4186static void clear_gigantic_page(struct page *page,
4187 unsigned long addr,
4188 unsigned int pages_per_huge_page)
4189{
4190 int i;
4191 struct page *p = page;
4192
4193 might_sleep();
4194 for (i = 0; i < pages_per_huge_page;
4195 i++, p = mem_map_next(p, page, i)) {
4196 cond_resched();
4197 clear_user_highpage(p, addr + i * PAGE_SIZE);
4198 }
4199}
4200void clear_huge_page(struct page *page,
4201 unsigned long addr, unsigned int pages_per_huge_page)
4202{
4203 int i;
4204
4205 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4206 clear_gigantic_page(page, addr, pages_per_huge_page);
4207 return;
4208 }
4209
4210 might_sleep();
4211 for (i = 0; i < pages_per_huge_page; i++) {
4212 cond_resched();
4213 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4214 }
4215}
4216
4217static void copy_user_gigantic_page(struct page *dst, struct page *src,
4218 unsigned long addr,
4219 struct vm_area_struct *vma,
4220 unsigned int pages_per_huge_page)
4221{
4222 int i;
4223 struct page *dst_base = dst;
4224 struct page *src_base = src;
4225
4226 for (i = 0; i < pages_per_huge_page; ) {
4227 cond_resched();
4228 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4229
4230 i++;
4231 dst = mem_map_next(dst, dst_base, i);
4232 src = mem_map_next(src, src_base, i);
4233 }
4234}
4235
4236void copy_user_huge_page(struct page *dst, struct page *src,
4237 unsigned long addr, struct vm_area_struct *vma,
4238 unsigned int pages_per_huge_page)
4239{
4240 int i;
4241
4242 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4243 copy_user_gigantic_page(dst, src, addr, vma,
4244 pages_per_huge_page);
4245 return;
4246 }
4247
4248 might_sleep();
4249 for (i = 0; i < pages_per_huge_page; i++) {
4250 cond_resched();
4251 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4252 }
4253}
4254#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */