blob: bb1369f7b9b4ba8af51d90a40de8f522aa2470dc [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
Andy Whitcroftd41dee32005-06-23 00:07:54 -070072#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070073/* use the per-pgdat data instead for discontigmem - mbligh */
74unsigned long max_mapnr;
75struct page *mem_map;
76
77EXPORT_SYMBOL(max_mapnr);
78EXPORT_SYMBOL(mem_map);
79#endif
80
81unsigned long num_physpages;
82/*
83 * A number of key systems in x86 including ioremap() rely on the assumption
84 * that high_memory defines the upper bound on direct map memory, then end
85 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
86 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
87 * and ZONE_HIGHMEM.
88 */
89void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
91EXPORT_SYMBOL(num_physpages);
92EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
Ingo Molnar32a93232008-02-06 22:39:44 +010094/*
95 * Randomize the address space (stacks, mmaps, brk, etc.).
96 *
97 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
98 * as ancient (libc5 based) binaries can segfault. )
99 */
100int randomize_va_space __read_mostly =
101#ifdef CONFIG_COMPAT_BRK
102 1;
103#else
104 2;
105#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100106
107static int __init disable_randmaps(char *s)
108{
109 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800110 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100111}
112__setup("norandmaps", disable_randmaps);
113
Hugh Dickins62eede62009-09-21 17:03:34 -0700114unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700115unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700116
117/*
118 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
119 */
120static int __init init_zero_pfn(void)
121{
122 zero_pfn = page_to_pfn(ZERO_PAGE(0));
123 return 0;
124}
125core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100126
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800127
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800128#if defined(SPLIT_RSS_COUNTING)
129
David Rientjesea48cf72012-03-21 16:34:13 -0700130void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800131{
132 int i;
133
134 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700135 if (current->rss_stat.count[i]) {
136 add_mm_counter(mm, i, current->rss_stat.count[i]);
137 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800138 }
139 }
David Rientjes05af2e12012-03-21 16:34:13 -0700140 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800141}
142
143static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
144{
145 struct task_struct *task = current;
146
147 if (likely(task->mm == mm))
148 task->rss_stat.count[member] += val;
149 else
150 add_mm_counter(mm, member, val);
151}
152#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
153#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
154
155/* sync counter once per 64 page faults */
156#define TASK_RSS_EVENTS_THRESH (64)
157static void check_sync_rss_stat(struct task_struct *task)
158{
159 if (unlikely(task != current))
160 return;
161 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700162 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800163}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700164#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800165
166#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
167#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
168
169static void check_sync_rss_stat(struct task_struct *task)
170{
171}
172
Peter Zijlstra9547d012011-05-24 17:12:14 -0700173#endif /* SPLIT_RSS_COUNTING */
174
175#ifdef HAVE_GENERIC_MMU_GATHER
176
177static int tlb_next_batch(struct mmu_gather *tlb)
178{
179 struct mmu_gather_batch *batch;
180
181 batch = tlb->active;
182 if (batch->next) {
183 tlb->active = batch->next;
184 return 1;
185 }
186
Michal Hocko53a59fc2013-01-04 15:35:12 -0800187 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
188 return 0;
189
Peter Zijlstra9547d012011-05-24 17:12:14 -0700190 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
191 if (!batch)
192 return 0;
193
Michal Hocko53a59fc2013-01-04 15:35:12 -0800194 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700195 batch->next = NULL;
196 batch->nr = 0;
197 batch->max = MAX_GATHER_BATCH;
198
199 tlb->active->next = batch;
200 tlb->active = batch;
201
202 return 1;
203}
204
205/* tlb_gather_mmu
206 * Called to initialize an (on-stack) mmu_gather structure for page-table
207 * tear-down from @mm. The @fullmm argument is used when @mm is without
208 * users and we're going to destroy the full address space (exit/execve).
209 */
210void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
211{
212 tlb->mm = mm;
213
214 tlb->fullmm = fullmm;
Alex Shi597e1c32012-06-28 09:02:21 +0800215 tlb->start = -1UL;
216 tlb->end = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700217 tlb->need_flush = 0;
218 tlb->fast_mode = (num_possible_cpus() == 1);
219 tlb->local.next = NULL;
220 tlb->local.nr = 0;
221 tlb->local.max = ARRAY_SIZE(tlb->__pages);
222 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800223 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700224
225#ifdef CONFIG_HAVE_RCU_TABLE_FREE
226 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800227#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700228}
229
230void tlb_flush_mmu(struct mmu_gather *tlb)
231{
232 struct mmu_gather_batch *batch;
233
234 if (!tlb->need_flush)
235 return;
236 tlb->need_flush = 0;
237 tlb_flush(tlb);
238#ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 tlb_table_flush(tlb);
240#endif
241
242 if (tlb_fast_mode(tlb))
243 return;
244
245 for (batch = &tlb->local; batch; batch = batch->next) {
246 free_pages_and_swap_cache(batch->pages, batch->nr);
247 batch->nr = 0;
248 }
249 tlb->active = &tlb->local;
250}
251
252/* tlb_finish_mmu
253 * Called at the end of the shootdown operation to free up any resources
254 * that were required.
255 */
256void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
257{
258 struct mmu_gather_batch *batch, *next;
259
Alex Shi597e1c32012-06-28 09:02:21 +0800260 tlb->start = start;
261 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700262 tlb_flush_mmu(tlb);
263
264 /* keep the page table cache within bounds */
265 check_pgt_cache();
266
267 for (batch = tlb->local.next; batch; batch = next) {
268 next = batch->next;
269 free_pages((unsigned long)batch, 0);
270 }
271 tlb->local.next = NULL;
272}
273
274/* __tlb_remove_page
275 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
276 * handling the additional races in SMP caused by other CPUs caching valid
277 * mappings in their TLBs. Returns the number of free page slots left.
278 * When out of page slots we must call tlb_flush_mmu().
279 */
280int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
281{
282 struct mmu_gather_batch *batch;
283
Shaohua Lif21760b2012-01-12 17:19:16 -0800284 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700285
286 if (tlb_fast_mode(tlb)) {
287 free_page_and_swap_cache(page);
288 return 1; /* avoid calling tlb_flush_mmu() */
289 }
290
291 batch = tlb->active;
292 batch->pages[batch->nr++] = page;
293 if (batch->nr == batch->max) {
294 if (!tlb_next_batch(tlb))
295 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700296 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297 }
298 VM_BUG_ON(batch->nr > batch->max);
299
300 return batch->max - batch->nr;
301}
302
303#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800304
Peter Zijlstra26723912011-05-24 17:12:00 -0700305#ifdef CONFIG_HAVE_RCU_TABLE_FREE
306
307/*
308 * See the comment near struct mmu_table_batch.
309 */
310
311static void tlb_remove_table_smp_sync(void *arg)
312{
313 /* Simply deliver the interrupt */
314}
315
316static void tlb_remove_table_one(void *table)
317{
318 /*
319 * This isn't an RCU grace period and hence the page-tables cannot be
320 * assumed to be actually RCU-freed.
321 *
322 * It is however sufficient for software page-table walkers that rely on
323 * IRQ disabling. See the comment near struct mmu_table_batch.
324 */
325 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
326 __tlb_remove_table(table);
327}
328
329static void tlb_remove_table_rcu(struct rcu_head *head)
330{
331 struct mmu_table_batch *batch;
332 int i;
333
334 batch = container_of(head, struct mmu_table_batch, rcu);
335
336 for (i = 0; i < batch->nr; i++)
337 __tlb_remove_table(batch->tables[i]);
338
339 free_page((unsigned long)batch);
340}
341
342void tlb_table_flush(struct mmu_gather *tlb)
343{
344 struct mmu_table_batch **batch = &tlb->batch;
345
346 if (*batch) {
347 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
348 *batch = NULL;
349 }
350}
351
352void tlb_remove_table(struct mmu_gather *tlb, void *table)
353{
354 struct mmu_table_batch **batch = &tlb->batch;
355
356 tlb->need_flush = 1;
357
358 /*
359 * When there's less then two users of this mm there cannot be a
360 * concurrent page-table walk.
361 */
362 if (atomic_read(&tlb->mm->mm_users) < 2) {
363 __tlb_remove_table(table);
364 return;
365 }
366
367 if (*batch == NULL) {
368 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
369 if (*batch == NULL) {
370 tlb_remove_table_one(table);
371 return;
372 }
373 (*batch)->nr = 0;
374 }
375 (*batch)->tables[(*batch)->nr++] = table;
376 if ((*batch)->nr == MAX_TABLE_BATCH)
377 tlb_table_flush(tlb);
378}
379
Peter Zijlstra9547d012011-05-24 17:12:14 -0700380#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/*
383 * If a p?d_bad entry is found while walking page tables, report
384 * the error, before resetting entry to p?d_none. Usually (but
385 * very seldom) called out from the p?d_none_or_clear_bad macros.
386 */
387
388void pgd_clear_bad(pgd_t *pgd)
389{
390 pgd_ERROR(*pgd);
391 pgd_clear(pgd);
392}
393
394void pud_clear_bad(pud_t *pud)
395{
396 pud_ERROR(*pud);
397 pud_clear(pud);
398}
399
400void pmd_clear_bad(pmd_t *pmd)
401{
402 pmd_ERROR(*pmd);
403 pmd_clear(pmd);
404}
405
406/*
407 * Note: this doesn't free the actual pages themselves. That
408 * has been handled earlier when unmapping all the memory regions.
409 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000410static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
411 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800413 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700414 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000415 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700416 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417}
418
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
420 unsigned long addr, unsigned long end,
421 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422{
423 pmd_t *pmd;
424 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700425 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
Hugh Dickinse0da3822005-04-19 13:29:15 -0700427 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 do {
430 next = pmd_addr_end(addr, end);
431 if (pmd_none_or_clear_bad(pmd))
432 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000433 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 } while (pmd++, addr = next, addr != end);
435
Hugh Dickinse0da3822005-04-19 13:29:15 -0700436 start &= PUD_MASK;
437 if (start < floor)
438 return;
439 if (ceiling) {
440 ceiling &= PUD_MASK;
441 if (!ceiling)
442 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700444 if (end - 1 > ceiling - 1)
445 return;
446
447 pmd = pmd_offset(pud, start);
448 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000449 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450}
451
Hugh Dickinse0da3822005-04-19 13:29:15 -0700452static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
453 unsigned long addr, unsigned long end,
454 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455{
456 pud_t *pud;
457 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700458 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Hugh Dickinse0da3822005-04-19 13:29:15 -0700460 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 do {
463 next = pud_addr_end(addr, end);
464 if (pud_none_or_clear_bad(pud))
465 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700466 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 } while (pud++, addr = next, addr != end);
468
Hugh Dickinse0da3822005-04-19 13:29:15 -0700469 start &= PGDIR_MASK;
470 if (start < floor)
471 return;
472 if (ceiling) {
473 ceiling &= PGDIR_MASK;
474 if (!ceiling)
475 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700477 if (end - 1 > ceiling - 1)
478 return;
479
480 pud = pud_offset(pgd, start);
481 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000482 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483}
484
485/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700486 * This function frees user-level page tables of a process.
487 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 * Must be called with pagetable lock held.
489 */
Jan Beulich42b77722008-07-23 21:27:10 -0700490void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700491 unsigned long addr, unsigned long end,
492 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493{
494 pgd_t *pgd;
495 unsigned long next;
496
Hugh Dickinse0da3822005-04-19 13:29:15 -0700497 /*
498 * The next few lines have given us lots of grief...
499 *
500 * Why are we testing PMD* at this top level? Because often
501 * there will be no work to do at all, and we'd prefer not to
502 * go all the way down to the bottom just to discover that.
503 *
504 * Why all these "- 1"s? Because 0 represents both the bottom
505 * of the address space and the top of it (using -1 for the
506 * top wouldn't help much: the masks would do the wrong thing).
507 * The rule is that addr 0 and floor 0 refer to the bottom of
508 * the address space, but end 0 and ceiling 0 refer to the top
509 * Comparisons need to use "end - 1" and "ceiling - 1" (though
510 * that end 0 case should be mythical).
511 *
512 * Wherever addr is brought up or ceiling brought down, we must
513 * be careful to reject "the opposite 0" before it confuses the
514 * subsequent tests. But what about where end is brought down
515 * by PMD_SIZE below? no, end can't go down to 0 there.
516 *
517 * Whereas we round start (addr) and ceiling down, by different
518 * masks at different levels, in order to test whether a table
519 * now has no other vmas using it, so can be freed, we don't
520 * bother to round floor or end up - the tests don't need that.
521 */
522
523 addr &= PMD_MASK;
524 if (addr < floor) {
525 addr += PMD_SIZE;
526 if (!addr)
527 return;
528 }
529 if (ceiling) {
530 ceiling &= PMD_MASK;
531 if (!ceiling)
532 return;
533 }
534 if (end - 1 > ceiling - 1)
535 end -= PMD_SIZE;
536 if (addr > end - 1)
537 return;
538
Jan Beulich42b77722008-07-23 21:27:10 -0700539 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 do {
541 next = pgd_addr_end(addr, end);
542 if (pgd_none_or_clear_bad(pgd))
543 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700544 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700546}
547
Jan Beulich42b77722008-07-23 21:27:10 -0700548void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700549 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700550{
551 while (vma) {
552 struct vm_area_struct *next = vma->vm_next;
553 unsigned long addr = vma->vm_start;
554
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700555 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000556 * Hide vma from rmap and truncate_pagecache before freeing
557 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700558 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800559 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700560 unlink_file_vma(vma);
561
David Gibson9da61ae2006-03-22 00:08:57 -0800562 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700563 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700564 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700565 } else {
566 /*
567 * Optimization: gather nearby vmas into one call down
568 */
569 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800570 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700571 vma = next;
572 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800573 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700574 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700575 }
576 free_pgd_range(tlb, addr, vma->vm_end,
577 floor, next? next->vm_start: ceiling);
578 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700579 vma = next;
580 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581}
582
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800583int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
584 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800586 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800587 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700588 if (!new)
589 return -ENOMEM;
590
Nick Piggin362a61a2008-05-14 06:37:36 +0200591 /*
592 * Ensure all pte setup (eg. pte page lock and page clearing) are
593 * visible before the pte is made visible to other CPUs by being
594 * put into page tables.
595 *
596 * The other side of the story is the pointer chasing in the page
597 * table walking code (when walking the page table without locking;
598 * ie. most of the time). Fortunately, these data accesses consist
599 * of a chain of data-dependent loads, meaning most CPUs (alpha
600 * being the notable exception) will already guarantee loads are
601 * seen in-order. See the alpha page table accessors for the
602 * smp_read_barrier_depends() barriers in page table walking code.
603 */
604 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
605
Hugh Dickinsc74df322005-10-29 18:16:23 -0700606 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800607 wait_split_huge_page = 0;
608 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800611 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800612 } else if (unlikely(pmd_trans_splitting(*pmd)))
613 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700614 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800615 if (new)
616 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800617 if (wait_split_huge_page)
618 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700619 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620}
621
Hugh Dickins1bb36302005-10-29 18:16:22 -0700622int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700624 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
625 if (!new)
626 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627
Nick Piggin362a61a2008-05-14 06:37:36 +0200628 smp_wmb(); /* See comment in __pte_alloc */
629
Hugh Dickins1bb36302005-10-29 18:16:22 -0700630 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800631 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700632 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800633 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800634 } else
635 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700636 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800637 if (new)
638 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700639 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640}
641
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800642static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700643{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800644 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
645}
646
647static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
648{
649 int i;
650
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800651 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700652 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800653 for (i = 0; i < NR_MM_COUNTERS; i++)
654 if (rss[i])
655 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700656}
657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800659 * This function is called to print an error when a bad pte
660 * is found. For example, we might have a PFN-mapped pte in
661 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700662 *
663 * The calling function must still handle the error.
664 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800665static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
666 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700667{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800668 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
669 pud_t *pud = pud_offset(pgd, addr);
670 pmd_t *pmd = pmd_offset(pud, addr);
671 struct address_space *mapping;
672 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800673 static unsigned long resume;
674 static unsigned long nr_shown;
675 static unsigned long nr_unshown;
676
677 /*
678 * Allow a burst of 60 reports, then keep quiet for that minute;
679 * or allow a steady drip of one report per second.
680 */
681 if (nr_shown == 60) {
682 if (time_before(jiffies, resume)) {
683 nr_unshown++;
684 return;
685 }
686 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800687 printk(KERN_ALERT
688 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800689 nr_unshown);
690 nr_unshown = 0;
691 }
692 nr_shown = 0;
693 }
694 if (nr_shown++ == 0)
695 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800696
697 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
698 index = linear_page_index(vma, addr);
699
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800700 printk(KERN_ALERT
701 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800702 current->comm,
703 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800704 if (page)
705 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800706 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800707 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
708 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
709 /*
710 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
711 */
712 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800713 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800714 (unsigned long)vma->vm_ops->fault);
715 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800716 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800717 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700718 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800719 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700720}
721
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700722static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800723{
724 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
725}
726
Nick Pigginb5810032005-10-29 18:16:12 -0700727/*
Nick Piggin7e675132008-04-28 02:13:00 -0700728 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800729 *
Nick Piggin7e675132008-04-28 02:13:00 -0700730 * "Special" mappings do not wish to be associated with a "struct page" (either
731 * it doesn't exist, or it exists but they don't want to touch it). In this
732 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700733 *
Nick Piggin7e675132008-04-28 02:13:00 -0700734 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
735 * pte bit, in which case this function is trivial. Secondly, an architecture
736 * may not have a spare pte bit, which requires a more complicated scheme,
737 * described below.
738 *
739 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
740 * special mapping (even if there are underlying and valid "struct pages").
741 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800742 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700743 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
744 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700745 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
746 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800747 *
748 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
749 *
Nick Piggin7e675132008-04-28 02:13:00 -0700750 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700751 *
Nick Piggin7e675132008-04-28 02:13:00 -0700752 * This restricts such mappings to be a linear translation from virtual address
753 * to pfn. To get around this restriction, we allow arbitrary mappings so long
754 * as the vma is not a COW mapping; in that case, we know that all ptes are
755 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700756 *
757 *
Nick Piggin7e675132008-04-28 02:13:00 -0700758 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
759 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700760 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
761 * page" backing, however the difference is that _all_ pages with a struct
762 * page (that is, those where pfn_valid is true) are refcounted and considered
763 * normal pages by the VM. The disadvantage is that pages are refcounted
764 * (which can be slower and simply not an option for some PFNMAP users). The
765 * advantage is that we don't have to follow the strict linearity rule of
766 * PFNMAP mappings in order to support COWable mappings.
767 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800768 */
Nick Piggin7e675132008-04-28 02:13:00 -0700769#ifdef __HAVE_ARCH_PTE_SPECIAL
770# define HAVE_PTE_SPECIAL 1
771#else
772# define HAVE_PTE_SPECIAL 0
773#endif
774struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
775 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800776{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800777 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700778
779 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800780 if (likely(!pte_special(pte)))
781 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700782 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
783 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700784 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800785 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700786 return NULL;
787 }
788
789 /* !HAVE_PTE_SPECIAL case follows: */
790
Jared Hulbertb379d792008-04-28 02:12:58 -0700791 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
792 if (vma->vm_flags & VM_MIXEDMAP) {
793 if (!pfn_valid(pfn))
794 return NULL;
795 goto out;
796 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700797 unsigned long off;
798 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700799 if (pfn == vma->vm_pgoff + off)
800 return NULL;
801 if (!is_cow_mapping(vma->vm_flags))
802 return NULL;
803 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800804 }
805
Hugh Dickins62eede62009-09-21 17:03:34 -0700806 if (is_zero_pfn(pfn))
807 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800808check_pfn:
809 if (unlikely(pfn > highest_memmap_pfn)) {
810 print_bad_pte(vma, addr, pte, NULL);
811 return NULL;
812 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800813
814 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700815 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700816 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800817 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700818out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800819 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800820}
821
822/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 * copy one vm_area from one task to the other. Assumes the page tables
824 * already present in the new task to be cleared in the whole range
825 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 */
827
Hugh Dickins570a335b2009-12-14 17:58:46 -0800828static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700830 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700831 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832{
Nick Pigginb5810032005-10-29 18:16:12 -0700833 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 pte_t pte = *src_pte;
835 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
837 /* pte contains position in swap or file, so copy. */
838 if (unlikely(!pte_present(pte))) {
839 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700840 swp_entry_t entry = pte_to_swp_entry(pte);
841
Hugh Dickins570a335b2009-12-14 17:58:46 -0800842 if (swap_duplicate(entry) < 0)
843 return entry.val;
844
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 /* make sure dst_mm is on swapoff's mmlist. */
846 if (unlikely(list_empty(&dst_mm->mmlist))) {
847 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700848 if (list_empty(&dst_mm->mmlist))
849 list_add(&dst_mm->mmlist,
850 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 spin_unlock(&mmlist_lock);
852 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800853 if (likely(!non_swap_entry(entry)))
854 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800855 else if (is_migration_entry(entry)) {
856 page = migration_entry_to_page(entry);
857
858 if (PageAnon(page))
859 rss[MM_ANONPAGES]++;
860 else
861 rss[MM_FILEPAGES]++;
862
863 if (is_write_migration_entry(entry) &&
864 is_cow_mapping(vm_flags)) {
865 /*
866 * COW mappings require pages in both
867 * parent and child to be set to read.
868 */
869 make_migration_entry_read(&entry);
870 pte = swp_entry_to_pte(entry);
871 set_pte_at(src_mm, addr, src_pte, pte);
872 }
Christoph Lameter06972122006-06-23 02:03:35 -0700873 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700875 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 }
877
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 /*
879 * If it's a COW mapping, write protect it both
880 * in the parent and the child
881 */
Linus Torvalds67121172005-12-11 20:38:17 -0800882 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700884 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 }
886
887 /*
888 * If it's a shared mapping, mark it clean in
889 * the child
890 */
891 if (vm_flags & VM_SHARED)
892 pte = pte_mkclean(pte);
893 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800894
895 page = vm_normal_page(vma, addr, pte);
896 if (page) {
897 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700898 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800899 if (PageAnon(page))
900 rss[MM_ANONPAGES]++;
901 else
902 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800903 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700904
905out_set_pte:
906 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800907 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908}
909
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
911 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
912 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700914 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700916 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700917 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800918 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800919 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
921again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800922 init_rss_vec(rss);
923
Hugh Dickinsc74df322005-10-29 18:16:23 -0700924 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 if (!dst_pte)
926 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700927 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700928 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700929 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700930 orig_src_pte = src_pte;
931 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700932 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 do {
935 /*
936 * We are holding two locks at this point - either of them
937 * could generate latencies in another task on another CPU.
938 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700939 if (progress >= 32) {
940 progress = 0;
941 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100942 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700943 break;
944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 if (pte_none(*src_pte)) {
946 progress++;
947 continue;
948 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800949 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
950 vma, addr, rss);
951 if (entry.val)
952 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 progress += 8;
954 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700956 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700957 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700958 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800959 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700960 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700961 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800962
963 if (entry.val) {
964 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
965 return -ENOMEM;
966 progress = 0;
967 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 if (addr != end)
969 goto again;
970 return 0;
971}
972
973static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
974 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
975 unsigned long addr, unsigned long end)
976{
977 pmd_t *src_pmd, *dst_pmd;
978 unsigned long next;
979
980 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
981 if (!dst_pmd)
982 return -ENOMEM;
983 src_pmd = pmd_offset(src_pud, addr);
984 do {
985 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800986 if (pmd_trans_huge(*src_pmd)) {
987 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800988 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800989 err = copy_huge_pmd(dst_mm, src_mm,
990 dst_pmd, src_pmd, addr, vma);
991 if (err == -ENOMEM)
992 return -ENOMEM;
993 if (!err)
994 continue;
995 /* fall through */
996 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 if (pmd_none_or_clear_bad(src_pmd))
998 continue;
999 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1000 vma, addr, next))
1001 return -ENOMEM;
1002 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1003 return 0;
1004}
1005
1006static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1007 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1008 unsigned long addr, unsigned long end)
1009{
1010 pud_t *src_pud, *dst_pud;
1011 unsigned long next;
1012
1013 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1014 if (!dst_pud)
1015 return -ENOMEM;
1016 src_pud = pud_offset(src_pgd, addr);
1017 do {
1018 next = pud_addr_end(addr, end);
1019 if (pud_none_or_clear_bad(src_pud))
1020 continue;
1021 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1022 vma, addr, next))
1023 return -ENOMEM;
1024 } while (dst_pud++, src_pud++, addr = next, addr != end);
1025 return 0;
1026}
1027
1028int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1029 struct vm_area_struct *vma)
1030{
1031 pgd_t *src_pgd, *dst_pgd;
1032 unsigned long next;
1033 unsigned long addr = vma->vm_start;
1034 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001035 unsigned long mmun_start; /* For mmu_notifiers */
1036 unsigned long mmun_end; /* For mmu_notifiers */
1037 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001038 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039
Nick Piggind9928952005-08-28 16:49:11 +10001040 /*
1041 * Don't copy ptes where a page fault will fill them correctly.
1042 * Fork becomes much lighter when there are big shared or private
1043 * readonly mappings. The tradeoff is that copy_page_range is more
1044 * efficient than faulting.
1045 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001046 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1047 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001048 if (!vma->anon_vma)
1049 return 0;
1050 }
1051
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 if (is_vm_hugetlb_page(vma))
1053 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1054
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001055 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001056 /*
1057 * We do not free on error cases below as remove_vma
1058 * gets called on error from higher level routine
1059 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001060 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001061 if (ret)
1062 return ret;
1063 }
1064
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001065 /*
1066 * We need to invalidate the secondary MMU mappings only when
1067 * there could be a permission downgrade on the ptes of the
1068 * parent mm. And a permission downgrade will only happen if
1069 * is_cow_mapping() returns true.
1070 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001071 is_cow = is_cow_mapping(vma->vm_flags);
1072 mmun_start = addr;
1073 mmun_end = end;
1074 if (is_cow)
1075 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1076 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001077
1078 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 dst_pgd = pgd_offset(dst_mm, addr);
1080 src_pgd = pgd_offset(src_mm, addr);
1081 do {
1082 next = pgd_addr_end(addr, end);
1083 if (pgd_none_or_clear_bad(src_pgd))
1084 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001085 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1086 vma, addr, next))) {
1087 ret = -ENOMEM;
1088 break;
1089 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001091
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001092 if (is_cow)
1093 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001094 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095}
1096
Robin Holt51c6f662005-11-13 16:06:42 -08001097static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001098 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001100 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101{
Nick Pigginb5810032005-10-29 18:16:12 -07001102 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001103 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001104 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001105 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001106 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001107 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001108
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001109again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001110 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001111 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1112 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001113 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 do {
1115 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001116 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001118 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001119
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001121 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001122
Linus Torvalds6aab3412005-11-28 14:34:23 -08001123 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 if (unlikely(details) && page) {
1125 /*
1126 * unmap_shared_mapping_pages() wants to
1127 * invalidate cache without truncating:
1128 * unmap shared but keep private pages.
1129 */
1130 if (details->check_mapping &&
1131 details->check_mapping != page->mapping)
1132 continue;
1133 /*
1134 * Each page->index must be checked when
1135 * invalidating or truncating nonlinear.
1136 */
1137 if (details->nonlinear_vma &&
1138 (page->index < details->first_index ||
1139 page->index > details->last_index))
1140 continue;
1141 }
Nick Pigginb5810032005-10-29 18:16:12 -07001142 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001143 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 tlb_remove_tlb_entry(tlb, pte, addr);
1145 if (unlikely(!page))
1146 continue;
1147 if (unlikely(details) && details->nonlinear_vma
1148 && linear_page_index(details->nonlinear_vma,
1149 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001150 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001153 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001154 else {
1155 if (pte_dirty(ptent))
1156 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001157 if (pte_young(ptent) &&
1158 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001159 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001160 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001161 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001162 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001163 if (unlikely(page_mapcount(page) < 0))
1164 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001165 force_flush = !__tlb_remove_page(tlb, page);
1166 if (force_flush)
1167 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 continue;
1169 }
1170 /*
1171 * If details->check_mapping, we leave swap entries;
1172 * if details->nonlinear_vma, we leave file entries.
1173 */
1174 if (unlikely(details))
1175 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001176 if (pte_file(ptent)) {
1177 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1178 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001179 } else {
1180 swp_entry_t entry = pte_to_swp_entry(ptent);
1181
1182 if (!non_swap_entry(entry))
1183 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001184 else if (is_migration_entry(entry)) {
1185 struct page *page;
1186
1187 page = migration_entry_to_page(entry);
1188
1189 if (PageAnon(page))
1190 rss[MM_ANONPAGES]--;
1191 else
1192 rss[MM_FILEPAGES]--;
1193 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
1196 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001198 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001199
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001200 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001201 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001202 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001203
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001204 /*
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1208 */
1209 if (force_flush) {
1210 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001211
1212#ifdef HAVE_GENERIC_MMU_GATHER
1213 tlb->start = addr;
1214 tlb->end = end;
1215#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001216 tlb_flush_mmu(tlb);
1217 if (addr != end)
1218 goto again;
1219 }
1220
Robin Holt51c6f662005-11-13 16:06:42 -08001221 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222}
1223
Robin Holt51c6f662005-11-13 16:06:42 -08001224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001225 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001227 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001236 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001246 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001248 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001249 /* fall through */
1250 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001261next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001264
1265 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266}
1267
Robin Holt51c6f662005-11-13 16:06:42 -08001268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001269 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001271 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001279 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001283
1284 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Al Viro038c7aa2012-03-05 13:25:09 -05001287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001299 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001304 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001309 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
Al Virof5cc4ee2012-03-05 14:14:20 -05001312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001315 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001330 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001331 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of do_mmap_pgoff. When
1340 * hugetlbfs ->mmap method fails,
1341 * do_mmap_pgoff() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354}
1355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001358 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001363 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
Al Viro6e8bb012012-03-05 13:41:15 -05001374void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001376 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001378 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001389 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001392 *
1393 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001399 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001400 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001403 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001404 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001407 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410}
1411
Jack Steinerc627f9c2008-07-29 22:33:53 -07001412/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, 0);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001432 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435}
1436
Jack Steinerc627f9c2008-07-29 22:33:53 -07001437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001455 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
Johannes Weiner142762b2010-05-24 14:32:39 -07001460/**
1461 * follow_page - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 *
1466 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1467 *
1468 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469 * an error pointer if there is a mapping to something not represented
1470 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001472struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001473 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474{
1475 pgd_t *pgd;
1476 pud_t *pud;
1477 pmd_t *pmd;
1478 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001479 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001481 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001483 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484 if (!IS_ERR(page)) {
1485 BUG_ON(flags & FOLL_GET);
1486 goto out;
1487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001489 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 pgd = pgd_offset(mm, address);
1491 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001495 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001497 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001498 BUG_ON(flags & FOLL_GET);
1499 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500 goto out;
1501 }
1502 if (unlikely(pud_bad(*pud)))
1503 goto no_page_table;
1504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001506 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001507 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001508 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001509 BUG_ON(flags & FOLL_GET);
1510 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511 goto out;
1512 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001513 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1514 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001515 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001516 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001517 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001518 goto split_fallthrough;
1519 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001520 spin_lock(&mm->page_table_lock);
1521 if (likely(pmd_trans_huge(*pmd))) {
1522 if (unlikely(pmd_trans_splitting(*pmd))) {
1523 spin_unlock(&mm->page_table_lock);
1524 wait_split_huge_page(vma->anon_vma, pmd);
1525 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001526 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001527 pmd, flags);
1528 spin_unlock(&mm->page_table_lock);
1529 goto out;
1530 }
1531 } else
1532 spin_unlock(&mm->page_table_lock);
1533 /* fall through */
1534 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001535split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001536 if (unlikely(pmd_bad(*pmd)))
1537 goto no_page_table;
1538
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001539 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540
1541 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001542 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001543 goto no_page;
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001544 if ((flags & FOLL_NUMA) && pte_numa(pte))
1545 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001546 if ((flags & FOLL_WRITE) && !pte_write(pte))
1547 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001548
Linus Torvalds6aab3412005-11-28 14:34:23 -08001549 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001550 if (unlikely(!page)) {
1551 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001552 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001553 goto bad_page;
1554 page = pte_page(pte);
1555 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001556
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001557 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001558 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001559 if (flags & FOLL_TOUCH) {
1560 if ((flags & FOLL_WRITE) &&
1561 !pte_dirty(pte) && !PageDirty(page))
1562 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001563 /*
1564 * pte_mkyoung() would be more correct here, but atomic care
1565 * is needed to avoid losing the dirty bit: it is easier to use
1566 * mark_page_accessed().
1567 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001568 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001570 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001571 /*
1572 * The preliminary mapping check is mainly to avoid the
1573 * pointless overhead of lock_page on the ZERO_PAGE
1574 * which might bounce very badly if there is contention.
1575 *
1576 * If the page is already locked, we don't need to
1577 * handle it now - vmscan will handle it later if and
1578 * when it attempts to reclaim the page.
1579 */
1580 if (page->mapping && trylock_page(page)) {
1581 lru_add_drain(); /* push cached pages to LRU */
1582 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001583 * Because we lock page here, and migration is
1584 * blocked by the pte's page reference, and we
1585 * know the page is still mapped, we don't even
1586 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001587 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001588 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001589 unlock_page(page);
1590 }
1591 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001592unlock:
1593 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001595 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001597bad_page:
1598 pte_unmap_unlock(ptep, ptl);
1599 return ERR_PTR(-EFAULT);
1600
1601no_page:
1602 pte_unmap_unlock(ptep, ptl);
1603 if (!pte_none(pte))
1604 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001605
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001606no_page_table:
1607 /*
1608 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001609 * has touched so far, we don't want to allocate unnecessary pages or
1610 * page tables. Return error instead of NULL to skip handle_mm_fault,
1611 * then get_dump_page() will return NULL to leave a hole in the dump.
1612 * But we can only make this optimization where a hole would surely
1613 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001614 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001615 if ((flags & FOLL_DUMP) &&
1616 (!vma->vm_ops || !vma->vm_ops->fault))
1617 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001618 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619}
1620
Linus Torvalds95042f92011-04-12 14:15:51 -07001621static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1622{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001623 return stack_guard_page_start(vma, addr) ||
1624 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001625}
1626
Huang Ying0014bd92011-01-30 11:15:47 +08001627/**
1628 * __get_user_pages() - pin user pages in memory
1629 * @tsk: task_struct of target task
1630 * @mm: mm_struct of target mm
1631 * @start: starting user address
1632 * @nr_pages: number of pages from start to pin
1633 * @gup_flags: flags modifying pin behaviour
1634 * @pages: array that receives pointers to the pages pinned.
1635 * Should be at least nr_pages long. Or NULL, if caller
1636 * only intends to ensure the pages are faulted in.
1637 * @vmas: array of pointers to vmas corresponding to each page.
1638 * Or NULL if the caller does not require them.
1639 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1640 *
1641 * Returns number of pages pinned. This may be fewer than the number
1642 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1643 * were pinned, returns -errno. Each page returned must be released
1644 * with a put_page() call when it is finished with. vmas will only
1645 * remain valid while mmap_sem is held.
1646 *
1647 * Must be called with mmap_sem held for read or write.
1648 *
1649 * __get_user_pages walks a process's page tables and takes a reference to
1650 * each struct page that each user address corresponds to at a given
1651 * instant. That is, it takes the page that would be accessed if a user
1652 * thread accesses the given user virtual address at that instant.
1653 *
1654 * This does not guarantee that the page exists in the user mappings when
1655 * __get_user_pages returns, and there may even be a completely different
1656 * page there in some cases (eg. if mmapped pagecache has been invalidated
1657 * and subsequently re faulted). However it does guarantee that the page
1658 * won't be freed completely. And mostly callers simply care that the page
1659 * contains data that was valid *at some point in time*. Typically, an IO
1660 * or similar operation cannot guarantee anything stronger anyway because
1661 * locks can't be held over the syscall boundary.
1662 *
1663 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1664 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1665 * appropriate) must be called after the page is finished with, and
1666 * before put_page is called.
1667 *
1668 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1669 * or mmap_sem contention, and if waiting is needed to pin all pages,
1670 * *@nonblocking will be set to 0.
1671 *
1672 * In most cases, get_user_pages or get_user_pages_fast should be used
1673 * instead of __get_user_pages. __get_user_pages should be used only if
1674 * you need some special @gup_flags.
1675 */
Nick Pigginb291f002008-10-18 20:26:44 -07001676int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001677 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001678 struct page **pages, struct vm_area_struct **vmas,
1679 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680{
1681 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001682 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683
Peter Zijlstra9d737772009-06-25 11:58:55 +02001684 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001685 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001686
1687 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1688
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 /*
1690 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001691 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001693 vm_flags = (gup_flags & FOLL_WRITE) ?
1694 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1695 vm_flags &= (gup_flags & FOLL_FORCE) ?
1696 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001697
1698 /*
1699 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1700 * would be called on PROT_NONE ranges. We must never invoke
1701 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1702 * page faults would unprotect the PROT_NONE ranges if
1703 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1704 * bitflag. So to avoid that, don't set FOLL_NUMA if
1705 * FOLL_FORCE is set.
1706 */
1707 if (!(gup_flags & FOLL_FORCE))
1708 gup_flags |= FOLL_NUMA;
1709
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 i = 0;
1711
1712 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001713 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714
1715 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001716 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 pgd_t *pgd;
1719 pud_t *pud;
1720 pmd_t *pmd;
1721 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001722
1723 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001724 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 return i ? : -EFAULT;
1726 if (pg > TASK_SIZE)
1727 pgd = pgd_offset_k(pg);
1728 else
1729 pgd = pgd_offset_gate(mm, pg);
1730 BUG_ON(pgd_none(*pgd));
1731 pud = pud_offset(pgd, pg);
1732 BUG_ON(pud_none(*pud));
1733 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001734 if (pmd_none(*pmd))
1735 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001736 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001738 if (pte_none(*pte)) {
1739 pte_unmap(pte);
1740 return i ? : -EFAULT;
1741 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001742 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001744 struct page *page;
1745
Linus Torvalds95042f92011-04-12 14:15:51 -07001746 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001747 if (!page) {
1748 if (!(gup_flags & FOLL_DUMP) &&
1749 is_zero_pfn(pte_pfn(*pte)))
1750 page = pte_page(*pte);
1751 else {
1752 pte_unmap(pte);
1753 return i ? : -EFAULT;
1754 }
1755 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001756 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001757 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 }
1759 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001760 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 }
1762
Nick Pigginb291f002008-10-18 20:26:44 -07001763 if (!vma ||
1764 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001765 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 return i ? : -EFAULT;
1767
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001768 if (is_vm_hugetlb_page(vma)) {
1769 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001770 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001771 continue;
1772 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001773
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001775 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001776 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777
Ethan Solomita462e00c2007-07-15 23:38:16 -07001778 /*
Ying Han47792802009-01-06 14:40:18 -08001779 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001780 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001781 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001782 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001783 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001784
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001785 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001786 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001787 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001788 unsigned int fault_flags = 0;
1789
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001790 /* For mlock, just skip the stack guard page. */
1791 if (foll_flags & FOLL_MLOCK) {
1792 if (stack_guard_page(vma, start))
1793 goto next_page;
1794 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001795 if (foll_flags & FOLL_WRITE)
1796 fault_flags |= FAULT_FLAG_WRITE;
1797 if (nonblocking)
1798 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001799 if (foll_flags & FOLL_NOWAIT)
1800 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001801
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001802 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001803 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001804
Nick Piggin83c54072007-07-19 01:47:05 -07001805 if (ret & VM_FAULT_ERROR) {
1806 if (ret & VM_FAULT_OOM)
1807 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001808 if (ret & (VM_FAULT_HWPOISON |
1809 VM_FAULT_HWPOISON_LARGE)) {
1810 if (i)
1811 return i;
1812 else if (gup_flags & FOLL_HWPOISON)
1813 return -EHWPOISON;
1814 else
1815 return -EFAULT;
1816 }
1817 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001818 return i ? i : -EFAULT;
1819 BUG();
1820 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001821
1822 if (tsk) {
1823 if (ret & VM_FAULT_MAJOR)
1824 tsk->maj_flt++;
1825 else
1826 tsk->min_flt++;
1827 }
Nick Piggin83c54072007-07-19 01:47:05 -07001828
Michel Lespinasse53a77062011-01-13 15:46:14 -08001829 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001830 if (nonblocking)
1831 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001832 return i;
1833 }
1834
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001835 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001836 * The VM_FAULT_WRITE bit tells us that
1837 * do_wp_page has broken COW when necessary,
1838 * even if maybe_mkwrite decided not to set
1839 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001840 * page lookups as if they were reads. But only
1841 * do so when looping for pte_write is futile:
1842 * in some cases userspace may also be wanting
1843 * to write to the gotten user page, which a
1844 * read fault here might prevent (a readonly
1845 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001846 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001847 if ((ret & VM_FAULT_WRITE) &&
1848 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001849 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001850
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001851 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001853 if (IS_ERR(page))
1854 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001856 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001857
Russell Kinga6f36be2006-12-30 22:24:19 +00001858 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001859 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001861next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 if (vmas)
1863 vmas[i] = vma;
1864 i++;
1865 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001866 nr_pages--;
1867 } while (nr_pages && start < vma->vm_end);
1868 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 return i;
1870}
Huang Ying0014bd92011-01-30 11:15:47 +08001871EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001872
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001873/*
1874 * fixup_user_fault() - manually resolve a user page fault
1875 * @tsk: the task_struct to use for page fault accounting, or
1876 * NULL if faults are not to be recorded.
1877 * @mm: mm_struct of target mm
1878 * @address: user address
1879 * @fault_flags:flags to pass down to handle_mm_fault()
1880 *
1881 * This is meant to be called in the specific scenario where for locking reasons
1882 * we try to access user memory in atomic context (within a pagefault_disable()
1883 * section), this returns -EFAULT, and we want to resolve the user fault before
1884 * trying again.
1885 *
1886 * Typically this is meant to be used by the futex code.
1887 *
1888 * The main difference with get_user_pages() is that this function will
1889 * unconditionally call handle_mm_fault() which will in turn perform all the
1890 * necessary SW fixup of the dirty and young bits in the PTE, while
1891 * handle_mm_fault() only guarantees to update these in the struct page.
1892 *
1893 * This is important for some architectures where those bits also gate the
1894 * access permission to the page because they are maintained in software. On
1895 * such architectures, gup() will not be enough to make a subsequent access
1896 * succeed.
1897 *
1898 * This should be called with the mm_sem held for read.
1899 */
1900int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1901 unsigned long address, unsigned int fault_flags)
1902{
1903 struct vm_area_struct *vma;
1904 int ret;
1905
1906 vma = find_extend_vma(mm, address);
1907 if (!vma || address < vma->vm_start)
1908 return -EFAULT;
1909
1910 ret = handle_mm_fault(mm, vma, address, fault_flags);
1911 if (ret & VM_FAULT_ERROR) {
1912 if (ret & VM_FAULT_OOM)
1913 return -ENOMEM;
1914 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1915 return -EHWPOISON;
1916 if (ret & VM_FAULT_SIGBUS)
1917 return -EFAULT;
1918 BUG();
1919 }
1920 if (tsk) {
1921 if (ret & VM_FAULT_MAJOR)
1922 tsk->maj_flt++;
1923 else
1924 tsk->min_flt++;
1925 }
1926 return 0;
1927}
1928
1929/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001930 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001931 * @tsk: the task_struct to use for page fault accounting, or
1932 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001933 * @mm: mm_struct of target mm
1934 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001935 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001936 * @write: whether pages will be written to by the caller
1937 * @force: whether to force write access even if user mapping is
1938 * readonly. This will result in the page being COWed even
1939 * in MAP_SHARED mappings. You do not want this.
1940 * @pages: array that receives pointers to the pages pinned.
1941 * Should be at least nr_pages long. Or NULL, if caller
1942 * only intends to ensure the pages are faulted in.
1943 * @vmas: array of pointers to vmas corresponding to each page.
1944 * Or NULL if the caller does not require them.
1945 *
1946 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001947 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001948 * were pinned, returns -errno. Each page returned must be released
1949 * with a put_page() call when it is finished with. vmas will only
1950 * remain valid while mmap_sem is held.
1951 *
1952 * Must be called with mmap_sem held for read or write.
1953 *
1954 * get_user_pages walks a process's page tables and takes a reference to
1955 * each struct page that each user address corresponds to at a given
1956 * instant. That is, it takes the page that would be accessed if a user
1957 * thread accesses the given user virtual address at that instant.
1958 *
1959 * This does not guarantee that the page exists in the user mappings when
1960 * get_user_pages returns, and there may even be a completely different
1961 * page there in some cases (eg. if mmapped pagecache has been invalidated
1962 * and subsequently re faulted). However it does guarantee that the page
1963 * won't be freed completely. And mostly callers simply care that the page
1964 * contains data that was valid *at some point in time*. Typically, an IO
1965 * or similar operation cannot guarantee anything stronger anyway because
1966 * locks can't be held over the syscall boundary.
1967 *
1968 * If write=0, the page must not be written to. If the page is written to,
1969 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1970 * after the page is finished with, and before put_page is called.
1971 *
1972 * get_user_pages is typically used for fewer-copy IO operations, to get a
1973 * handle on the memory by some means other than accesses via the user virtual
1974 * addresses. The pages may be submitted for DMA to devices or accessed via
1975 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1976 * use the correct cache flushing APIs.
1977 *
1978 * See also get_user_pages_fast, for performance critical applications.
1979 */
Nick Pigginb291f002008-10-18 20:26:44 -07001980int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001981 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001982 struct page **pages, struct vm_area_struct **vmas)
1983{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001984 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001985
Hugh Dickins58fa8792009-09-21 17:03:31 -07001986 if (pages)
1987 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001988 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001989 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001990 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001991 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001992
Michel Lespinasse53a77062011-01-13 15:46:14 -08001993 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1994 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001995}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996EXPORT_SYMBOL(get_user_pages);
1997
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001998/**
1999 * get_dump_page() - pin user page in memory while writing it to core dump
2000 * @addr: user address
2001 *
2002 * Returns struct page pointer of user page pinned for dump,
2003 * to be freed afterwards by page_cache_release() or put_page().
2004 *
2005 * Returns NULL on any kind of failure - a hole must then be inserted into
2006 * the corefile, to preserve alignment with its headers; and also returns
2007 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2008 * allowing a hole to be left in the corefile to save diskspace.
2009 *
2010 * Called without mmap_sem, but after all other threads have been killed.
2011 */
2012#ifdef CONFIG_ELF_CORE
2013struct page *get_dump_page(unsigned long addr)
2014{
2015 struct vm_area_struct *vma;
2016 struct page *page;
2017
2018 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002019 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2020 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002021 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002022 flush_cache_page(vma, addr, page_to_pfn(page));
2023 return page;
2024}
2025#endif /* CONFIG_ELF_CORE */
2026
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002027pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002028 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002029{
2030 pgd_t * pgd = pgd_offset(mm, addr);
2031 pud_t * pud = pud_alloc(mm, pgd, addr);
2032 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002033 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002034 if (pmd) {
2035 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002036 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002037 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002038 }
2039 return NULL;
2040}
2041
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002043 * This is the old fallback for page remapping.
2044 *
2045 * For historical reasons, it only allows reserved pages. Only
2046 * old drivers should use this, and they needed to mark their
2047 * pages reserved for the old functions anyway.
2048 */
Nick Piggin423bad62008-04-28 02:13:01 -07002049static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2050 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002051{
Nick Piggin423bad62008-04-28 02:13:01 -07002052 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002053 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002054 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002055 spinlock_t *ptl;
2056
Linus Torvalds238f58d2005-11-29 13:01:56 -08002057 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002058 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002059 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002060 retval = -ENOMEM;
2061 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002062 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002063 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002064 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002065 retval = -EBUSY;
2066 if (!pte_none(*pte))
2067 goto out_unlock;
2068
2069 /* Ok, finally just insert the thing.. */
2070 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002071 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002072 page_add_file_rmap(page);
2073 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2074
2075 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002076 pte_unmap_unlock(pte, ptl);
2077 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002078out_unlock:
2079 pte_unmap_unlock(pte, ptl);
2080out:
2081 return retval;
2082}
2083
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002084/**
2085 * vm_insert_page - insert single page into user vma
2086 * @vma: user vma to map to
2087 * @addr: target user address of this page
2088 * @page: source kernel page
2089 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002090 * This allows drivers to insert individual pages they've allocated
2091 * into a user vma.
2092 *
2093 * The page has to be a nice clean _individual_ kernel allocation.
2094 * If you allocate a compound page, you need to have marked it as
2095 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002096 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002097 *
2098 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2099 * took an arbitrary page protection parameter. This doesn't allow
2100 * that. Your vma protection will have to be set up correctly, which
2101 * means that if you want a shared writable mapping, you'd better
2102 * ask for a shared writable mapping!
2103 *
2104 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002105 *
2106 * Usually this function is called from f_op->mmap() handler
2107 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2108 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2109 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002110 */
Nick Piggin423bad62008-04-28 02:13:01 -07002111int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2112 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002113{
2114 if (addr < vma->vm_start || addr >= vma->vm_end)
2115 return -EFAULT;
2116 if (!page_count(page))
2117 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002118 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2119 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2120 BUG_ON(vma->vm_flags & VM_PFNMAP);
2121 vma->vm_flags |= VM_MIXEDMAP;
2122 }
Nick Piggin423bad62008-04-28 02:13:01 -07002123 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002124}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002125EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002126
Nick Piggin423bad62008-04-28 02:13:01 -07002127static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2128 unsigned long pfn, pgprot_t prot)
2129{
2130 struct mm_struct *mm = vma->vm_mm;
2131 int retval;
2132 pte_t *pte, entry;
2133 spinlock_t *ptl;
2134
2135 retval = -ENOMEM;
2136 pte = get_locked_pte(mm, addr, &ptl);
2137 if (!pte)
2138 goto out;
2139 retval = -EBUSY;
2140 if (!pte_none(*pte))
2141 goto out_unlock;
2142
2143 /* Ok, finally just insert the thing.. */
2144 entry = pte_mkspecial(pfn_pte(pfn, prot));
2145 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002146 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002147
2148 retval = 0;
2149out_unlock:
2150 pte_unmap_unlock(pte, ptl);
2151out:
2152 return retval;
2153}
2154
Nick Piggine0dc0d82007-02-12 00:51:36 -08002155/**
2156 * vm_insert_pfn - insert single pfn into user vma
2157 * @vma: user vma to map to
2158 * @addr: target user address of this page
2159 * @pfn: source kernel pfn
2160 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002161 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002162 * they've allocated into a user vma. Same comments apply.
2163 *
2164 * This function should only be called from a vm_ops->fault handler, and
2165 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002166 *
2167 * vma cannot be a COW mapping.
2168 *
2169 * As this is called only for pages that do not currently exist, we
2170 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002171 */
2172int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002173 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002174{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002175 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002176 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002177 /*
2178 * Technically, architectures with pte_special can avoid all these
2179 * restrictions (same for remap_pfn_range). However we would like
2180 * consistency in testing and feature parity among all, so we should
2181 * try to keep these invariants in place for everybody.
2182 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002183 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2184 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2185 (VM_PFNMAP|VM_MIXEDMAP));
2186 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2187 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002188
Nick Piggin423bad62008-04-28 02:13:01 -07002189 if (addr < vma->vm_start || addr >= vma->vm_end)
2190 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002191 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002192 return -EINVAL;
2193
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002194 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002195
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002196 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002197}
2198EXPORT_SYMBOL(vm_insert_pfn);
2199
Nick Piggin423bad62008-04-28 02:13:01 -07002200int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn)
2202{
2203 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2204
2205 if (addr < vma->vm_start || addr >= vma->vm_end)
2206 return -EFAULT;
2207
2208 /*
2209 * If we don't have pte special, then we have to use the pfn_valid()
2210 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2211 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002212 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2213 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002214 */
2215 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2216 struct page *page;
2217
2218 page = pfn_to_page(pfn);
2219 return insert_page(vma, addr, page, vma->vm_page_prot);
2220 }
2221 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2222}
2223EXPORT_SYMBOL(vm_insert_mixed);
2224
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002225/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 * maps a range of physical memory into the requested pages. the old
2227 * mappings are removed. any references to nonexistent pages results
2228 * in null mappings (currently treated as "copy-on-access")
2229 */
2230static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2231 unsigned long addr, unsigned long end,
2232 unsigned long pfn, pgprot_t prot)
2233{
2234 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002235 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236
Hugh Dickinsc74df322005-10-29 18:16:23 -07002237 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 if (!pte)
2239 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002240 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 do {
2242 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002243 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 pfn++;
2245 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002246 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002247 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 return 0;
2249}
2250
2251static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2252 unsigned long addr, unsigned long end,
2253 unsigned long pfn, pgprot_t prot)
2254{
2255 pmd_t *pmd;
2256 unsigned long next;
2257
2258 pfn -= addr >> PAGE_SHIFT;
2259 pmd = pmd_alloc(mm, pud, addr);
2260 if (!pmd)
2261 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002262 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 do {
2264 next = pmd_addr_end(addr, end);
2265 if (remap_pte_range(mm, pmd, addr, next,
2266 pfn + (addr >> PAGE_SHIFT), prot))
2267 return -ENOMEM;
2268 } while (pmd++, addr = next, addr != end);
2269 return 0;
2270}
2271
2272static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2273 unsigned long addr, unsigned long end,
2274 unsigned long pfn, pgprot_t prot)
2275{
2276 pud_t *pud;
2277 unsigned long next;
2278
2279 pfn -= addr >> PAGE_SHIFT;
2280 pud = pud_alloc(mm, pgd, addr);
2281 if (!pud)
2282 return -ENOMEM;
2283 do {
2284 next = pud_addr_end(addr, end);
2285 if (remap_pmd_range(mm, pud, addr, next,
2286 pfn + (addr >> PAGE_SHIFT), prot))
2287 return -ENOMEM;
2288 } while (pud++, addr = next, addr != end);
2289 return 0;
2290}
2291
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002292/**
2293 * remap_pfn_range - remap kernel memory to userspace
2294 * @vma: user vma to map to
2295 * @addr: target user address to start at
2296 * @pfn: physical address of kernel memory
2297 * @size: size of map area
2298 * @prot: page protection flags for this mapping
2299 *
2300 * Note: this is only safe if the mm semaphore is held when called.
2301 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2303 unsigned long pfn, unsigned long size, pgprot_t prot)
2304{
2305 pgd_t *pgd;
2306 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002307 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 struct mm_struct *mm = vma->vm_mm;
2309 int err;
2310
2311 /*
2312 * Physically remapped pages are special. Tell the
2313 * rest of the world about it:
2314 * VM_IO tells people not to look at these pages
2315 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002316 * VM_PFNMAP tells the core MM that the base pages are just
2317 * raw PFN mappings, and do not have a "struct page" associated
2318 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002319 * VM_DONTEXPAND
2320 * Disable vma merging and expanding with mremap().
2321 * VM_DONTDUMP
2322 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002323 *
2324 * There's a horrible special case to handle copy-on-write
2325 * behaviour that some programs depend on. We mark the "original"
2326 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002327 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002329 if (is_cow_mapping(vma->vm_flags)) {
2330 if (addr != vma->vm_start || end != vma->vm_end)
2331 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002332 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002333 }
2334
2335 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2336 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002337 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002338
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002339 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340
2341 BUG_ON(addr >= end);
2342 pfn -= addr >> PAGE_SHIFT;
2343 pgd = pgd_offset(mm, addr);
2344 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 do {
2346 next = pgd_addr_end(addr, end);
2347 err = remap_pud_range(mm, pgd, addr, next,
2348 pfn + (addr >> PAGE_SHIFT), prot);
2349 if (err)
2350 break;
2351 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002352
2353 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002354 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002355
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 return err;
2357}
2358EXPORT_SYMBOL(remap_pfn_range);
2359
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002360static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2361 unsigned long addr, unsigned long end,
2362 pte_fn_t fn, void *data)
2363{
2364 pte_t *pte;
2365 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002366 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002367 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002368
2369 pte = (mm == &init_mm) ?
2370 pte_alloc_kernel(pmd, addr) :
2371 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2372 if (!pte)
2373 return -ENOMEM;
2374
2375 BUG_ON(pmd_huge(*pmd));
2376
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002377 arch_enter_lazy_mmu_mode();
2378
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002379 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002380
2381 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002382 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002383 if (err)
2384 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002385 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002386
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002387 arch_leave_lazy_mmu_mode();
2388
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002389 if (mm != &init_mm)
2390 pte_unmap_unlock(pte-1, ptl);
2391 return err;
2392}
2393
2394static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2395 unsigned long addr, unsigned long end,
2396 pte_fn_t fn, void *data)
2397{
2398 pmd_t *pmd;
2399 unsigned long next;
2400 int err;
2401
Andi Kleenceb86872008-07-23 21:27:50 -07002402 BUG_ON(pud_huge(*pud));
2403
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002404 pmd = pmd_alloc(mm, pud, addr);
2405 if (!pmd)
2406 return -ENOMEM;
2407 do {
2408 next = pmd_addr_end(addr, end);
2409 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2410 if (err)
2411 break;
2412 } while (pmd++, addr = next, addr != end);
2413 return err;
2414}
2415
2416static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2417 unsigned long addr, unsigned long end,
2418 pte_fn_t fn, void *data)
2419{
2420 pud_t *pud;
2421 unsigned long next;
2422 int err;
2423
2424 pud = pud_alloc(mm, pgd, addr);
2425 if (!pud)
2426 return -ENOMEM;
2427 do {
2428 next = pud_addr_end(addr, end);
2429 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2430 if (err)
2431 break;
2432 } while (pud++, addr = next, addr != end);
2433 return err;
2434}
2435
2436/*
2437 * Scan a region of virtual memory, filling in page tables as necessary
2438 * and calling a provided function on each leaf page table.
2439 */
2440int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2441 unsigned long size, pte_fn_t fn, void *data)
2442{
2443 pgd_t *pgd;
2444 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002445 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002446 int err;
2447
2448 BUG_ON(addr >= end);
2449 pgd = pgd_offset(mm, addr);
2450 do {
2451 next = pgd_addr_end(addr, end);
2452 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2453 if (err)
2454 break;
2455 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002456
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002457 return err;
2458}
2459EXPORT_SYMBOL_GPL(apply_to_page_range);
2460
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002462 * handle_pte_fault chooses page fault handler according to an entry
2463 * which was read non-atomically. Before making any commitment, on
2464 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002465 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002466 * must check under lock before unmapping the pte and proceeding
2467 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002468 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002469 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002470static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002471 pte_t *page_table, pte_t orig_pte)
2472{
2473 int same = 1;
2474#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2475 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002476 spinlock_t *ptl = pte_lockptr(mm, pmd);
2477 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002478 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002479 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002480 }
2481#endif
2482 pte_unmap(page_table);
2483 return same;
2484}
2485
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002486static 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 -08002487{
2488 /*
2489 * If the source page was a PFN mapping, we don't have
2490 * a "struct page" for it. We do a best-effort copy by
2491 * just copying from the original user address. If that
2492 * fails, we just zero-fill it. Live with it.
2493 */
2494 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002495 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002496 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2497
2498 /*
2499 * This really shouldn't fail, because the page is there
2500 * in the page tables. But it might just be unreadable,
2501 * in which case we just give up and fill the result with
2502 * zeroes.
2503 */
2504 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002505 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002506 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002507 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002508 } else
2509 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002510}
2511
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 * This routine handles present pages, when users try to write
2514 * to a shared page. It is done by copying the page to a new address
2515 * and decrementing the shared-page counter for the old page.
2516 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 * Note that this routine assumes that the protection checks have been
2518 * done by the caller (the low-level page fault routine in most cases).
2519 * Thus we can safely just mark it writable once we've done any necessary
2520 * COW.
2521 *
2522 * We also mark the page dirty at this point even though the page will
2523 * change only once the write actually happens. This avoids a few races,
2524 * and potentially makes it more efficient.
2525 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002526 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2527 * but allow concurrent faults), with pte both mapped and locked.
2528 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002530static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2531 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002532 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002533 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002535 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002537 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002538 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002539 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002540 unsigned long mmun_start = 0; /* For mmu_notifiers */
2541 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542
Linus Torvalds6aab3412005-11-28 14:34:23 -08002543 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002544 if (!old_page) {
2545 /*
2546 * VM_MIXEDMAP !pfn_valid() case
2547 *
2548 * We should not cow pages in a shared writeable mapping.
2549 * Just mark the pages writable as we can't do any dirty
2550 * accounting on raw pfn maps.
2551 */
2552 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2553 (VM_WRITE|VM_SHARED))
2554 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002555 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002556 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002558 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002559 * Take out anonymous pages first, anonymous shared vmas are
2560 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002561 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002562 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002563 if (!trylock_page(old_page)) {
2564 page_cache_get(old_page);
2565 pte_unmap_unlock(page_table, ptl);
2566 lock_page(old_page);
2567 page_table = pte_offset_map_lock(mm, pmd, address,
2568 &ptl);
2569 if (!pte_same(*page_table, orig_pte)) {
2570 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002571 goto unlock;
2572 }
2573 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002574 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002575 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002576 /*
2577 * The page is all ours. Move it to our anon_vma so
2578 * the rmap code will not search our parent or siblings.
2579 * Protected against the rmap code by the page lock.
2580 */
2581 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002582 unlock_page(old_page);
2583 goto reuse;
2584 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002585 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002586 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002587 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002588 /*
2589 * Only catch write-faults on shared writable pages,
2590 * read-only shared pages can get COWed by
2591 * get_user_pages(.write=1, .force=1).
2592 */
David Howells9637a5e2006-06-23 02:03:43 -07002593 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002594 struct vm_fault vmf;
2595 int tmp;
2596
2597 vmf.virtual_address = (void __user *)(address &
2598 PAGE_MASK);
2599 vmf.pgoff = old_page->index;
2600 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2601 vmf.page = old_page;
2602
David Howells9637a5e2006-06-23 02:03:43 -07002603 /*
2604 * Notify the address space that the page is about to
2605 * become writable so that it can prohibit this or wait
2606 * for the page to get into an appropriate state.
2607 *
2608 * We do this without the lock held, so that it can
2609 * sleep if it needs to.
2610 */
2611 page_cache_get(old_page);
2612 pte_unmap_unlock(page_table, ptl);
2613
Nick Pigginc2ec1752009-03-31 15:23:21 -07002614 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2615 if (unlikely(tmp &
2616 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2617 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002618 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002619 }
Nick Pigginb827e492009-04-30 15:08:16 -07002620 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2621 lock_page(old_page);
2622 if (!old_page->mapping) {
2623 ret = 0; /* retry the fault */
2624 unlock_page(old_page);
2625 goto unwritable_page;
2626 }
2627 } else
2628 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002629
David Howells9637a5e2006-06-23 02:03:43 -07002630 /*
2631 * Since we dropped the lock we need to revalidate
2632 * the PTE as someone else may have changed it. If
2633 * they did, we just return, as we can count on the
2634 * MMU to tell us if they didn't also make it writable.
2635 */
2636 page_table = pte_offset_map_lock(mm, pmd, address,
2637 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002638 if (!pte_same(*page_table, orig_pte)) {
2639 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002640 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002641 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002642
2643 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002645 dirty_page = old_page;
2646 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002647
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002648reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002649 flush_cache_page(vma, address, pte_pfn(orig_pte));
2650 entry = pte_mkyoung(orig_pte);
2651 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002652 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002653 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002654 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002655 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002656
2657 if (!dirty_page)
2658 return ret;
2659
2660 /*
2661 * Yes, Virginia, this is actually required to prevent a race
2662 * with clear_page_dirty_for_io() from clearing the page dirty
2663 * bit after it clear all dirty ptes, but before a racing
2664 * do_wp_page installs a dirty pte.
2665 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002666 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002667 */
2668 if (!page_mkwrite) {
2669 wait_on_page_locked(dirty_page);
2670 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002671 /* file_update_time outside page_lock */
2672 if (vma->vm_file)
2673 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002674 }
2675 put_page(dirty_page);
2676 if (page_mkwrite) {
2677 struct address_space *mapping = dirty_page->mapping;
2678
2679 set_page_dirty(dirty_page);
2680 unlock_page(dirty_page);
2681 page_cache_release(dirty_page);
2682 if (mapping) {
2683 /*
2684 * Some device drivers do not set page.mapping
2685 * but still dirty their pages
2686 */
2687 balance_dirty_pages_ratelimited(mapping);
2688 }
2689 }
2690
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002691 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693
2694 /*
2695 * Ok, we need to copy. Oh, well..
2696 */
Nick Pigginb5810032005-10-29 18:16:12 -07002697 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002698gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002699 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700
2701 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002702 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002703
Hugh Dickins62eede62009-09-21 17:03:34 -07002704 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002705 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2706 if (!new_page)
2707 goto oom;
2708 } else {
2709 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2710 if (!new_page)
2711 goto oom;
2712 cow_user_page(new_page, old_page, address, vma);
2713 }
2714 __SetPageUptodate(new_page);
2715
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002716 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002717 goto oom_free_new;
2718
Haggai Eran6bdb9132012-10-08 16:33:35 -07002719 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002720 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002721 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2722
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 /*
2724 * Re-check the pte - we dropped the lock
2725 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002726 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002727 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002728 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002729 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002730 dec_mm_counter_fast(mm, MM_FILEPAGES);
2731 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002732 }
2733 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002734 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002735 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002736 entry = mk_pte(new_page, vma->vm_page_prot);
2737 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002738 /*
2739 * Clear the pte entry and flush it first, before updating the
2740 * pte with the new entry. This will avoid a race condition
2741 * seen in the presence of one thread doing SMC and another
2742 * thread doing COW.
2743 */
Izik Eidus828502d2009-09-21 17:01:51 -07002744 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002745 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002746 /*
2747 * We call the notify macro here because, when using secondary
2748 * mmu page tables (such as kvm shadow page tables), we want the
2749 * new page to be mapped directly into the secondary page table.
2750 */
2751 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002752 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002753 if (old_page) {
2754 /*
2755 * Only after switching the pte to the new page may
2756 * we remove the mapcount here. Otherwise another
2757 * process may come and find the rmap count decremented
2758 * before the pte is switched to the new page, and
2759 * "reuse" the old page writing into it while our pte
2760 * here still points into it and can be read by other
2761 * threads.
2762 *
2763 * The critical issue is to order this
2764 * page_remove_rmap with the ptp_clear_flush above.
2765 * Those stores are ordered by (if nothing else,)
2766 * the barrier present in the atomic_add_negative
2767 * in page_remove_rmap.
2768 *
2769 * Then the TLB flush in ptep_clear_flush ensures that
2770 * no process can access the old page before the
2771 * decremented mapcount is visible. And the old page
2772 * cannot be reused until after the decremented
2773 * mapcount is visible. So transitively, TLBs to
2774 * old page will be flushed before it can be reused.
2775 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002776 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002777 }
2778
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 /* Free the old page.. */
2780 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002781 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002782 } else
2783 mem_cgroup_uncharge_page(new_page);
2784
Haggai Eran6bdb9132012-10-08 16:33:35 -07002785 if (new_page)
2786 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002787unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002788 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002789 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002790 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002791 if (old_page) {
2792 /*
2793 * Don't let another task, with possibly unlocked vma,
2794 * keep the mlocked page.
2795 */
2796 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2797 lock_page(old_page); /* LRU manipulation */
2798 munlock_vma_page(old_page);
2799 unlock_page(old_page);
2800 }
2801 page_cache_release(old_page);
2802 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002803 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002804oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002805 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002806oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002807 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002808 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002810
2811unwritable_page:
2812 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002813 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814}
2815
Peter Zijlstra97a89412011-05-24 17:12:04 -07002816static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 unsigned long start_addr, unsigned long end_addr,
2818 struct zap_details *details)
2819{
Al Virof5cc4ee2012-03-05 14:14:20 -05002820 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821}
2822
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002823static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 struct zap_details *details)
2825{
2826 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 pgoff_t vba, vea, zba, zea;
2828
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002829 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831
2832 vba = vma->vm_pgoff;
2833 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2834 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2835 zba = details->first_index;
2836 if (zba < vba)
2837 zba = vba;
2838 zea = details->last_index;
2839 if (zea > vea)
2840 zea = vea;
2841
Peter Zijlstra97a89412011-05-24 17:12:04 -07002842 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2844 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002845 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 }
2847}
2848
2849static inline void unmap_mapping_range_list(struct list_head *head,
2850 struct zap_details *details)
2851{
2852 struct vm_area_struct *vma;
2853
2854 /*
2855 * In nonlinear VMAs there is no correspondence between virtual address
2856 * offset and file offset. So we must perform an exhaustive search
2857 * across *all* the pages in each nonlinear VMA, not just the pages
2858 * whose virtual address lies outside the file truncation point.
2859 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002860 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002862 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 }
2864}
2865
2866/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002867 * 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 -07002868 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 * @holebegin: byte in first page to unmap, relative to the start of
2870 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002871 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 * must keep the partial page. In contrast, we must get rid of
2873 * partial pages.
2874 * @holelen: size of prospective hole in bytes. This will be rounded
2875 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2876 * end of the file.
2877 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2878 * but 0 when invalidating pagecache, don't throw away private data.
2879 */
2880void unmap_mapping_range(struct address_space *mapping,
2881 loff_t const holebegin, loff_t const holelen, int even_cows)
2882{
2883 struct zap_details details;
2884 pgoff_t hba = holebegin >> PAGE_SHIFT;
2885 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2886
2887 /* Check for overflow. */
2888 if (sizeof(holelen) > sizeof(hlen)) {
2889 long long holeend =
2890 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2891 if (holeend & ~(long long)ULONG_MAX)
2892 hlen = ULONG_MAX - hba + 1;
2893 }
2894
2895 details.check_mapping = even_cows? NULL: mapping;
2896 details.nonlinear_vma = NULL;
2897 details.first_index = hba;
2898 details.last_index = hba + hlen - 1;
2899 if (details.last_index < details.first_index)
2900 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901
Peter Zijlstra97a89412011-05-24 17:12:04 -07002902
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002903 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002904 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2906 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2907 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002908 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909}
2910EXPORT_SYMBOL(unmap_mapping_range);
2911
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002913 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2914 * but allow concurrent faults), and pte mapped but not yet locked.
2915 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002917static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2918 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002919 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002921 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002922 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002923 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002925 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002926 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002927 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002928 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002930 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002931 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002932
2933 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002934 if (unlikely(non_swap_entry(entry))) {
2935 if (is_migration_entry(entry)) {
2936 migration_entry_wait(mm, pmd, address);
2937 } else if (is_hwpoison_entry(entry)) {
2938 ret = VM_FAULT_HWPOISON;
2939 } else {
2940 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002941 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002942 }
Christoph Lameter06972122006-06-23 02:03:35 -07002943 goto out;
2944 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002945 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 page = lookup_swap_cache(entry);
2947 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002948 page = swapin_readahead(entry,
2949 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 if (!page) {
2951 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002952 * Back out if somebody else faulted in this pte
2953 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002955 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 if (likely(pte_same(*page_table, orig_pte)))
2957 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002958 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002959 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 }
2961
2962 /* Had to read the page from swap area: Major fault */
2963 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002964 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002965 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002966 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002967 /*
2968 * hwpoisoned dirty swapcache pages are kept for killing
2969 * owner processes (which may be unknown at hwpoison time)
2970 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002971 ret = VM_FAULT_HWPOISON;
2972 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002973 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 }
2975
Michel Lespinassed065bd82010-10-26 14:21:57 -07002976 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002977
Balbir Singh20a10222007-11-14 17:00:33 -08002978 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002979 if (!locked) {
2980 ret |= VM_FAULT_RETRY;
2981 goto out_release;
2982 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002984 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002985 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2986 * release the swapcache from under us. The page pin, and pte_same
2987 * test below, are not enough to exclude that. Even if it is still
2988 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002989 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002990 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002991 goto out_page;
2992
2993 if (ksm_might_need_to_copy(page, vma, address)) {
2994 swapcache = page;
2995 page = ksm_does_need_to_copy(page, vma, address);
2996
2997 if (unlikely(!page)) {
2998 ret = VM_FAULT_OOM;
2999 page = swapcache;
3000 swapcache = NULL;
3001 goto out_page;
3002 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08003003 }
3004
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003005 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003006 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003007 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003008 }
3009
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003011 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003013 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003014 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003015 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003016
3017 if (unlikely(!PageUptodate(page))) {
3018 ret = VM_FAULT_SIGBUS;
3019 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 }
3021
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003022 /*
3023 * The page isn't present yet, go ahead with the fault.
3024 *
3025 * Be careful about the sequence of operations here.
3026 * To get its accounting right, reuse_swap_page() must be called
3027 * while the page is counted on swap but not yet in mapcount i.e.
3028 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3029 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003030 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3031 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3032 * in page->private. In this case, a record in swap_cgroup is silently
3033 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003034 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003036 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003037 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003039 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003041 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003042 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003043 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 flush_icache_page(vma, page);
3046 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003047 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003048 /* It's better to call commit-charge after rmap is established */
3049 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003051 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003052 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003053 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003054 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003055 if (swapcache) {
3056 /*
3057 * Hold the lock to avoid the swap entry to be reused
3058 * until we take the PT lock for the pte_same() check
3059 * (to avoid false positives from pte_same). For
3060 * further safety release the lock after the swap_free
3061 * so that the swap count won't change under a
3062 * parallel locked swapcache.
3063 */
3064 unlock_page(swapcache);
3065 page_cache_release(swapcache);
3066 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003067
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003068 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003069 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3070 if (ret & VM_FAULT_ERROR)
3071 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 goto out;
3073 }
3074
3075 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003076 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003077unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003078 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079out:
3080 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003081out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003082 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003083 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003084out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003085 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003086out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003087 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003088 if (swapcache) {
3089 unlock_page(swapcache);
3090 page_cache_release(swapcache);
3091 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003092 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093}
3094
3095/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003096 * This is like a special single-page "expand_{down|up}wards()",
3097 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003098 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003099 */
3100static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3101{
3102 address &= PAGE_MASK;
3103 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003104 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003105
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003106 /*
3107 * Is there a mapping abutting this one below?
3108 *
3109 * That's only ok if it's the same stack mapping
3110 * that has gotten split..
3111 */
3112 if (prev && prev->vm_end == address)
3113 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3114
Michal Hockod05f3162011-05-24 17:11:44 -07003115 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003116 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003117 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3118 struct vm_area_struct *next = vma->vm_next;
3119
3120 /* As VM_GROWSDOWN but s/below/above/ */
3121 if (next && next->vm_start == address + PAGE_SIZE)
3122 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3123
3124 expand_upwards(vma, address + PAGE_SIZE);
3125 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003126 return 0;
3127}
3128
3129/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003130 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3131 * but allow concurrent faults), and pte mapped but not yet locked.
3132 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003134static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3135 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003136 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003138 struct page *page;
3139 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141
Linus Torvalds11ac5522010-08-14 11:44:56 -07003142 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003143
Linus Torvalds11ac5522010-08-14 11:44:56 -07003144 /* Check if we need to add a guard page to the stack */
3145 if (check_stack_guard_page(vma, address) < 0)
3146 return VM_FAULT_SIGBUS;
3147
3148 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003149 if (!(flags & FAULT_FLAG_WRITE)) {
3150 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3151 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003152 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003153 if (!pte_none(*page_table))
3154 goto unlock;
3155 goto setpte;
3156 }
3157
Nick Piggin557ed1f2007-10-16 01:24:40 -07003158 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003159 if (unlikely(anon_vma_prepare(vma)))
3160 goto oom;
3161 page = alloc_zeroed_user_highpage_movable(vma, address);
3162 if (!page)
3163 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003164 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003166 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003167 goto oom_free_page;
3168
Nick Piggin557ed1f2007-10-16 01:24:40 -07003169 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003170 if (vma->vm_flags & VM_WRITE)
3171 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003172
Nick Piggin557ed1f2007-10-16 01:24:40 -07003173 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003174 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003175 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003176
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003177 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003178 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003179setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003180 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181
3182 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003183 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003184unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003185 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003186 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003187release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003188 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003189 page_cache_release(page);
3190 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003191oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003192 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003193oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 return VM_FAULT_OOM;
3195}
3196
3197/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003198 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003200 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3201 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 *
3203 * As this is called only for pages that do not currently exist, we
3204 * do not need to flush old virtual caches or the TLB.
3205 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003206 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003207 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003208 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003210static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003211 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003212 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003214 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003215 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003216 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003217 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003220 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003221 struct vm_fault vmf;
3222 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003223 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003224
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003225 /*
3226 * If we do COW later, allocate page befor taking lock_page()
3227 * on the file cache page. This will reduce lock holding time.
3228 */
3229 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3230
3231 if (unlikely(anon_vma_prepare(vma)))
3232 return VM_FAULT_OOM;
3233
3234 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3235 if (!cow_page)
3236 return VM_FAULT_OOM;
3237
3238 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3239 page_cache_release(cow_page);
3240 return VM_FAULT_OOM;
3241 }
3242 } else
3243 cow_page = NULL;
3244
Nick Piggind0217ac2007-07-19 01:47:03 -07003245 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3246 vmf.pgoff = pgoff;
3247 vmf.flags = flags;
3248 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003250 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003251 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3252 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003253 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254
Andi Kleena3b947e2009-09-16 11:50:08 +02003255 if (unlikely(PageHWPoison(vmf.page))) {
3256 if (ret & VM_FAULT_LOCKED)
3257 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003258 ret = VM_FAULT_HWPOISON;
3259 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003260 }
3261
Nick Piggind00806b2007-07-19 01:46:57 -07003262 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003263 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003264 * locked.
3265 */
Nick Piggin83c54072007-07-19 01:47:05 -07003266 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003267 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003268 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003269 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003270
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 /*
3272 * Should we do an early C-O-W break?
3273 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003274 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003275 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003276 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003277 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003278 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003279 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003280 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003281 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003282 /*
3283 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003284 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003285 * to become writable
3286 */
Mark Fasheh69676142007-07-19 01:47:00 -07003287 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003288 int tmp;
3289
Mark Fasheh69676142007-07-19 01:47:00 -07003290 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003291 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003292 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3293 if (unlikely(tmp &
3294 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3295 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003296 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003297 }
Nick Pigginb827e492009-04-30 15:08:16 -07003298 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3299 lock_page(page);
3300 if (!page->mapping) {
3301 ret = 0; /* retry the fault */
3302 unlock_page(page);
3303 goto unwritable_page;
3304 }
3305 } else
3306 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003307 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003308 }
3309 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003310
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 }
3312
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003313 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
3315 /*
3316 * This silly early PAGE_DIRTY setting removes a race
3317 * due to the bad i386 page protection. But it's valid
3318 * for other architectures too.
3319 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003320 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321 * an exclusive copy of the page, or this is a shared mapping,
3322 * so we can make it writable and dirty to avoid having to
3323 * handle that later.
3324 */
3325 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003326 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003327 flush_icache_page(vma, page);
3328 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003329 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003332 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003333 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003334 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003335 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003336 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003337 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003338 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003339 get_page(dirty_page);
3340 }
Hugh Dickins42946212005-10-29 18:16:05 -07003341 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003342 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003343
3344 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003345 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003347 if (cow_page)
3348 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003349 if (anon)
3350 page_cache_release(page);
3351 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003352 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353 }
3354
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003355 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003356
Nick Pigginb827e492009-04-30 15:08:16 -07003357 if (dirty_page) {
3358 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003359 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003360
3361 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003362 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003363 unlock_page(dirty_page);
3364 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003365 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003366 /*
3367 * Some device drivers do not set page.mapping but still
3368 * dirty their pages
3369 */
3370 balance_dirty_pages_ratelimited(mapping);
3371 }
3372
3373 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003374 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003375 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003376 } else {
3377 unlock_page(vmf.page);
3378 if (anon)
3379 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003380 }
Nick Piggind00806b2007-07-19 01:46:57 -07003381
Nick Piggin83c54072007-07-19 01:47:05 -07003382 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003383
3384unwritable_page:
3385 page_cache_release(page);
3386 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003387uncharge_out:
3388 /* fs's fault handler get error */
3389 if (cow_page) {
3390 mem_cgroup_uncharge_page(cow_page);
3391 page_cache_release(cow_page);
3392 }
3393 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003394}
Nick Piggind00806b2007-07-19 01:46:57 -07003395
Nick Piggin54cb8822007-07-19 01:46:59 -07003396static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3397 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003398 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003399{
3400 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003401 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003402
Hugh Dickins16abfa02007-10-04 16:56:06 +01003403 pte_unmap(page_table);
3404 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003405}
3406
Jes Sorensenf4b81802006-09-27 01:50:10 -07003407/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408 * Fault of a previously existing named mapping. Repopulate the pte
3409 * from the encoded file_pte if possible. This enables swappable
3410 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003411 *
3412 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3413 * but allow concurrent faults), and pte mapped but not yet locked.
3414 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003416static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003417 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003418 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419{
Hugh Dickins65500d22005-10-29 18:15:59 -07003420 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003422 flags |= FAULT_FLAG_NONLINEAR;
3423
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003424 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003425 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426
Hugh Dickins2509ef22009-01-06 14:40:10 -08003427 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003428 /*
3429 * Page table corrupted: show pte and kill process.
3430 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003431 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003432 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003433 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003434
3435 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003436 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437}
3438
Mel Gorman9532fec2012-11-15 01:24:32 +00003439int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3440 unsigned long addr, int current_nid)
3441{
3442 get_page(page);
3443
3444 count_vm_numa_event(NUMA_HINT_FAULTS);
3445 if (current_nid == numa_node_id())
3446 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3447
3448 return mpol_misplaced(page, vma, addr);
3449}
3450
Mel Gormand10e63f2012-10-25 14:16:31 +02003451int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3452 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3453{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003454 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003455 spinlock_t *ptl;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003456 int current_nid = -1;
3457 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003458 bool migrated = false;
Mel Gormand10e63f2012-10-25 14:16:31 +02003459
3460 /*
3461 * The "pte" at this point cannot be used safely without
3462 * validation through pte_unmap_same(). It's of NUMA type but
3463 * the pfn may be screwed if the read is non atomic.
3464 *
3465 * ptep_modify_prot_start is not called as this is clearing
3466 * the _PAGE_NUMA bit and it is not really expected that there
3467 * would be concurrent hardware modifications to the PTE.
3468 */
3469 ptl = pte_lockptr(mm, pmd);
3470 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003471 if (unlikely(!pte_same(*ptep, pte))) {
3472 pte_unmap_unlock(ptep, ptl);
3473 goto out;
3474 }
3475
Mel Gormand10e63f2012-10-25 14:16:31 +02003476 pte = pte_mknonnuma(pte);
3477 set_pte_at(mm, addr, ptep, pte);
3478 update_mmu_cache(vma, addr, ptep);
3479
3480 page = vm_normal_page(vma, addr, pte);
3481 if (!page) {
3482 pte_unmap_unlock(ptep, ptl);
3483 return 0;
3484 }
3485
Mel Gorman4daae3b2012-11-02 11:33:45 +00003486 current_nid = page_to_nid(page);
Mel Gorman9532fec2012-11-15 01:24:32 +00003487 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
Mel Gormand10e63f2012-10-25 14:16:31 +02003488 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003489 if (target_nid == -1) {
3490 /*
3491 * Account for the fault against the current node if it not
3492 * being replaced regardless of where the page is located.
3493 */
3494 current_nid = numa_node_id();
3495 put_page(page);
3496 goto out;
3497 }
3498
3499 /* Migrate to the requested node */
Mel Gormanb8593bf2012-11-21 01:18:23 +00003500 migrated = migrate_misplaced_page(page, target_nid);
3501 if (migrated)
Mel Gorman4daae3b2012-11-02 11:33:45 +00003502 current_nid = target_nid;
3503
3504out:
Mel Gorman9532fec2012-11-15 01:24:32 +00003505 if (current_nid != -1)
Mel Gormanb8593bf2012-11-21 01:18:23 +00003506 task_numa_fault(current_nid, 1, migrated);
Mel Gormand10e63f2012-10-25 14:16:31 +02003507 return 0;
3508}
3509
3510/* NUMA hinting page fault entry point for regular pmds */
3511#ifdef CONFIG_NUMA_BALANCING
3512static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3513 unsigned long addr, pmd_t *pmdp)
3514{
3515 pmd_t pmd;
3516 pte_t *pte, *orig_pte;
3517 unsigned long _addr = addr & PMD_MASK;
3518 unsigned long offset;
3519 spinlock_t *ptl;
3520 bool numa = false;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00003521 int local_nid = numa_node_id();
Mel Gormand10e63f2012-10-25 14:16:31 +02003522
3523 spin_lock(&mm->page_table_lock);
3524 pmd = *pmdp;
3525 if (pmd_numa(pmd)) {
3526 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3527 numa = true;
3528 }
3529 spin_unlock(&mm->page_table_lock);
3530
3531 if (!numa)
3532 return 0;
3533
3534 /* we're in a page fault so some vma must be in the range */
3535 BUG_ON(!vma);
3536 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3537 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3538 VM_BUG_ON(offset >= PMD_SIZE);
3539 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3540 pte += offset >> PAGE_SHIFT;
3541 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3542 pte_t pteval = *pte;
3543 struct page *page;
Mel Gorman9532fec2012-11-15 01:24:32 +00003544 int curr_nid = local_nid;
3545 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003546 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02003547 if (!pte_present(pteval))
3548 continue;
3549 if (!pte_numa(pteval))
3550 continue;
3551 if (addr >= vma->vm_end) {
3552 vma = find_vma(mm, addr);
3553 /* there's a pte present so there must be a vma */
3554 BUG_ON(!vma);
3555 BUG_ON(addr < vma->vm_start);
3556 }
3557 if (pte_numa(pteval)) {
3558 pteval = pte_mknonnuma(pteval);
3559 set_pte_at(mm, addr, pte, pteval);
3560 }
3561 page = vm_normal_page(vma, addr, pteval);
3562 if (unlikely(!page))
3563 continue;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003564 /* only check non-shared pages */
3565 if (unlikely(page_mapcount(page) != 1))
3566 continue;
Mel Gorman9532fec2012-11-15 01:24:32 +00003567
3568 /*
3569 * Note that the NUMA fault is later accounted to either
3570 * the node that is currently running or where the page is
3571 * migrated to.
3572 */
3573 curr_nid = local_nid;
3574 target_nid = numa_migrate_prep(page, vma, addr,
3575 page_to_nid(page));
3576 if (target_nid == -1) {
3577 put_page(page);
3578 continue;
3579 }
3580
3581 /* Migrate to the requested node */
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003582 pte_unmap_unlock(pte, ptl);
Mel Gormanb8593bf2012-11-21 01:18:23 +00003583 migrated = migrate_misplaced_page(page, target_nid);
3584 if (migrated)
Mel Gorman9532fec2012-11-15 01:24:32 +00003585 curr_nid = target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003586 task_numa_fault(curr_nid, 1, migrated);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003587
3588 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003589 }
3590 pte_unmap_unlock(orig_pte, ptl);
3591
3592 return 0;
3593}
3594#else
3595static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3596 unsigned long addr, pmd_t *pmdp)
3597{
3598 BUG();
Andrew Mortonb3dd2072012-12-17 15:59:24 -08003599 return 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003600}
3601#endif /* CONFIG_NUMA_BALANCING */
3602
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603/*
3604 * These routines also need to handle stuff like marking pages dirty
3605 * and/or accessed for architectures that don't do it in hardware (most
3606 * RISC architectures). The early dirtying is also good on the i386.
3607 *
3608 * There is also a hook called "update_mmu_cache()" that architectures
3609 * with external mmu caches can use to update those (ie the Sparc or
3610 * PowerPC hashed page tables that act as extended TLBs).
3611 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003612 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3613 * but allow concurrent faults), and pte mapped but not yet locked.
3614 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003616int handle_pte_fault(struct mm_struct *mm,
3617 struct vm_area_struct *vma, unsigned long address,
3618 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619{
3620 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003621 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003623 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003625 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003626 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003627 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003628 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003629 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003630 }
3631 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003632 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003635 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003636 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003637 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003638 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 }
3640
Mel Gormand10e63f2012-10-25 14:16:31 +02003641 if (pte_numa(entry))
3642 return do_numa_page(mm, vma, address, entry, pte, pmd);
3643
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003644 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003645 spin_lock(ptl);
3646 if (unlikely(!pte_same(*pte, entry)))
3647 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003648 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003650 return do_wp_page(mm, vma, address,
3651 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 entry = pte_mkdirty(entry);
3653 }
3654 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003655 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003656 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003657 } else {
3658 /*
3659 * This is needed only for protection faults but the arch code
3660 * is not yet telling us if this is a protection fault or not.
3661 * This still avoids useless tlb flushes for .text page faults
3662 * with threads.
3663 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003664 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003665 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003666 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003667unlock:
3668 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003669 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670}
3671
3672/*
3673 * By the time we get here, we already hold the mm semaphore
3674 */
Nick Piggin83c54072007-07-19 01:47:05 -07003675int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003676 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677{
3678 pgd_t *pgd;
3679 pud_t *pud;
3680 pmd_t *pmd;
3681 pte_t *pte;
3682
3683 __set_current_state(TASK_RUNNING);
3684
Christoph Lameterf8891e52006-06-30 01:55:45 -07003685 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003686 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003688 /* do counter updates before entering really critical section. */
3689 check_sync_rss_stat(current);
3690
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003691 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003692 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693
David Rientjes1f1d06c2012-05-29 15:06:23 -07003694retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696 pud = pud_alloc(mm, pgd, address);
3697 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003698 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 pmd = pmd_alloc(mm, pud, address);
3700 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003701 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003702 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3703 if (!vma->vm_ops)
3704 return do_huge_pmd_anonymous_page(mm, vma, address,
3705 pmd, flags);
3706 } else {
3707 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003708 int ret;
3709
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003710 barrier();
3711 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003712 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3713
Linus Torvaldse53289c2013-01-09 08:36:54 -08003714 /*
3715 * If the pmd is splitting, return and retry the
3716 * the fault. Alternative: wait until the split
3717 * is done, and goto retry.
3718 */
3719 if (pmd_trans_splitting(orig_pmd))
3720 return 0;
3721
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003722 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003723 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003724 orig_pmd, pmd);
3725
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003726 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003727 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3728 orig_pmd);
3729 /*
3730 * If COW results in an oom, the huge pmd will
3731 * have been split, so retry the fault on the
3732 * pte for a smaller charge.
3733 */
3734 if (unlikely(ret & VM_FAULT_OOM))
3735 goto retry;
3736 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003737 } else {
3738 huge_pmd_set_accessed(mm, vma, address, pmd,
3739 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003740 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003741
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003742 return 0;
3743 }
3744 }
3745
Mel Gormand10e63f2012-10-25 14:16:31 +02003746 if (pmd_numa(*pmd))
3747 return do_pmd_numa_page(mm, vma, address, pmd);
3748
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003749 /*
3750 * Use __pte_alloc instead of pte_alloc_map, because we can't
3751 * run pte_offset_map on the pmd, if an huge pmd could
3752 * materialize from under us from a different thread.
3753 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003754 if (unlikely(pmd_none(*pmd)) &&
3755 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003756 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003757 /* if an huge pmd materialized from under us just retry later */
3758 if (unlikely(pmd_trans_huge(*pmd)))
3759 return 0;
3760 /*
3761 * A regular pmd is established and it can't morph into a huge pmd
3762 * from under us anymore at this point because we hold the mmap_sem
3763 * read mode and khugepaged takes it in write mode. So now it's
3764 * safe to run pte_offset_map().
3765 */
3766 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003768 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769}
3770
3771#ifndef __PAGETABLE_PUD_FOLDED
3772/*
3773 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003774 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003776int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003778 pud_t *new = pud_alloc_one(mm, address);
3779 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003780 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781
Nick Piggin362a61a2008-05-14 06:37:36 +02003782 smp_wmb(); /* See comment in __pte_alloc */
3783
Hugh Dickins872fec12005-10-29 18:16:21 -07003784 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003785 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003786 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003787 else
3788 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003789 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003790 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791}
3792#endif /* __PAGETABLE_PUD_FOLDED */
3793
3794#ifndef __PAGETABLE_PMD_FOLDED
3795/*
3796 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003797 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003799int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003801 pmd_t *new = pmd_alloc_one(mm, address);
3802 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003803 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804
Nick Piggin362a61a2008-05-14 06:37:36 +02003805 smp_wmb(); /* See comment in __pte_alloc */
3806
Hugh Dickins872fec12005-10-29 18:16:21 -07003807 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003809 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003810 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003811 else
3812 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003814 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003815 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003816 else
3817 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003819 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003820 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821}
3822#endif /* __PAGETABLE_PMD_FOLDED */
3823
3824int make_pages_present(unsigned long addr, unsigned long end)
3825{
3826 int ret, len, write;
3827 struct vm_area_struct * vma;
3828
3829 vma = find_vma(current->mm, addr);
3830 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003831 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003832 /*
3833 * We want to touch writable mappings with a write fault in order
3834 * to break COW, except for shared mappings because these don't COW
3835 * and we would not want to dirty them for nothing.
3836 */
3837 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003838 BUG_ON(addr >= end);
3839 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003840 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841 ret = get_user_pages(current, current->mm, addr,
3842 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003843 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003845 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846}
3847
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848#if !defined(__HAVE_ARCH_GATE_AREA)
3849
3850#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003851static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852
3853static int __init gate_vma_init(void)
3854{
3855 gate_vma.vm_mm = NULL;
3856 gate_vma.vm_start = FIXADDR_USER_START;
3857 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003858 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3859 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003860
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 return 0;
3862}
3863__initcall(gate_vma_init);
3864#endif
3865
Stephen Wilson31db58b2011-03-13 15:49:15 -04003866struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867{
3868#ifdef AT_SYSINFO_EHDR
3869 return &gate_vma;
3870#else
3871 return NULL;
3872#endif
3873}
3874
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003875int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876{
3877#ifdef AT_SYSINFO_EHDR
3878 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3879 return 1;
3880#endif
3881 return 0;
3882}
3883
3884#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003885
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003886static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003887 pte_t **ptepp, spinlock_t **ptlp)
3888{
3889 pgd_t *pgd;
3890 pud_t *pud;
3891 pmd_t *pmd;
3892 pte_t *ptep;
3893
3894 pgd = pgd_offset(mm, address);
3895 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3896 goto out;
3897
3898 pud = pud_offset(pgd, address);
3899 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3900 goto out;
3901
3902 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003903 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003904 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3905 goto out;
3906
3907 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3908 if (pmd_huge(*pmd))
3909 goto out;
3910
3911 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3912 if (!ptep)
3913 goto out;
3914 if (!pte_present(*ptep))
3915 goto unlock;
3916 *ptepp = ptep;
3917 return 0;
3918unlock:
3919 pte_unmap_unlock(ptep, *ptlp);
3920out:
3921 return -EINVAL;
3922}
3923
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003924static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3925 pte_t **ptepp, spinlock_t **ptlp)
3926{
3927 int res;
3928
3929 /* (void) is needed to make gcc happy */
3930 (void) __cond_lock(*ptlp,
3931 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3932 return res;
3933}
3934
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003935/**
3936 * follow_pfn - look up PFN at a user virtual address
3937 * @vma: memory mapping
3938 * @address: user virtual address
3939 * @pfn: location to store found PFN
3940 *
3941 * Only IO mappings and raw PFN mappings are allowed.
3942 *
3943 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3944 */
3945int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3946 unsigned long *pfn)
3947{
3948 int ret = -EINVAL;
3949 spinlock_t *ptl;
3950 pte_t *ptep;
3951
3952 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3953 return ret;
3954
3955 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3956 if (ret)
3957 return ret;
3958 *pfn = pte_pfn(*ptep);
3959 pte_unmap_unlock(ptep, ptl);
3960 return 0;
3961}
3962EXPORT_SYMBOL(follow_pfn);
3963
Rik van Riel28b2ee22008-07-23 21:27:05 -07003964#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003965int follow_phys(struct vm_area_struct *vma,
3966 unsigned long address, unsigned int flags,
3967 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003968{
Johannes Weiner03668a42009-06-16 15:32:34 -07003969 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003970 pte_t *ptep, pte;
3971 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003972
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003973 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3974 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003975
Johannes Weiner03668a42009-06-16 15:32:34 -07003976 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003977 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003978 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003979
Rik van Riel28b2ee22008-07-23 21:27:05 -07003980 if ((flags & FOLL_WRITE) && !pte_write(pte))
3981 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003982
3983 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003984 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003985
Johannes Weiner03668a42009-06-16 15:32:34 -07003986 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003987unlock:
3988 pte_unmap_unlock(ptep, ptl);
3989out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003990 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003991}
3992
3993int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3994 void *buf, int len, int write)
3995{
3996 resource_size_t phys_addr;
3997 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003998 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003999 int offset = addr & (PAGE_SIZE-1);
4000
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004001 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004002 return -EINVAL;
4003
4004 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4005 if (write)
4006 memcpy_toio(maddr + offset, buf, len);
4007 else
4008 memcpy_fromio(buf, maddr + offset, len);
4009 iounmap(maddr);
4010
4011 return len;
4012}
4013#endif
4014
David Howells0ec76a12006-09-27 01:50:15 -07004015/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004016 * Access another process' address space as given in mm. If non-NULL, use the
4017 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004018 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004019static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4020 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004021{
David Howells0ec76a12006-09-27 01:50:15 -07004022 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004023 void *old_buf = buf;
4024
David Howells0ec76a12006-09-27 01:50:15 -07004025 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004026 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004027 while (len) {
4028 int bytes, ret, offset;
4029 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004030 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004031
4032 ret = get_user_pages(tsk, mm, addr, 1,
4033 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004034 if (ret <= 0) {
4035 /*
4036 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4037 * we can access using slightly different code.
4038 */
4039#ifdef CONFIG_HAVE_IOREMAP_PROT
4040 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004041 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004042 break;
4043 if (vma->vm_ops && vma->vm_ops->access)
4044 ret = vma->vm_ops->access(vma, addr, buf,
4045 len, write);
4046 if (ret <= 0)
4047#endif
4048 break;
4049 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004050 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004051 bytes = len;
4052 offset = addr & (PAGE_SIZE-1);
4053 if (bytes > PAGE_SIZE-offset)
4054 bytes = PAGE_SIZE-offset;
4055
4056 maddr = kmap(page);
4057 if (write) {
4058 copy_to_user_page(vma, page, addr,
4059 maddr + offset, buf, bytes);
4060 set_page_dirty_lock(page);
4061 } else {
4062 copy_from_user_page(vma, page, addr,
4063 buf, maddr + offset, bytes);
4064 }
4065 kunmap(page);
4066 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004067 }
David Howells0ec76a12006-09-27 01:50:15 -07004068 len -= bytes;
4069 buf += bytes;
4070 addr += bytes;
4071 }
4072 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004073
4074 return buf - old_buf;
4075}
Andi Kleen03252912008-01-30 13:33:18 +01004076
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004077/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004078 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004079 * @mm: the mm_struct of the target address space
4080 * @addr: start address to access
4081 * @buf: source or destination buffer
4082 * @len: number of bytes to transfer
4083 * @write: whether the access is a write
4084 *
4085 * The caller must hold a reference on @mm.
4086 */
4087int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4088 void *buf, int len, int write)
4089{
4090 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4091}
4092
Andi Kleen03252912008-01-30 13:33:18 +01004093/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004094 * Access another process' address space.
4095 * Source/target buffer must be kernel space,
4096 * Do not walk the page table directly, use get_user_pages
4097 */
4098int access_process_vm(struct task_struct *tsk, unsigned long addr,
4099 void *buf, int len, int write)
4100{
4101 struct mm_struct *mm;
4102 int ret;
4103
4104 mm = get_task_mm(tsk);
4105 if (!mm)
4106 return 0;
4107
4108 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4109 mmput(mm);
4110
4111 return ret;
4112}
4113
Andi Kleen03252912008-01-30 13:33:18 +01004114/*
4115 * Print the name of a VMA.
4116 */
4117void print_vma_addr(char *prefix, unsigned long ip)
4118{
4119 struct mm_struct *mm = current->mm;
4120 struct vm_area_struct *vma;
4121
Ingo Molnare8bff742008-02-13 20:21:06 +01004122 /*
4123 * Do not print if we are in atomic
4124 * contexts (in exception stacks, etc.):
4125 */
4126 if (preempt_count())
4127 return;
4128
Andi Kleen03252912008-01-30 13:33:18 +01004129 down_read(&mm->mmap_sem);
4130 vma = find_vma(mm, ip);
4131 if (vma && vma->vm_file) {
4132 struct file *f = vma->vm_file;
4133 char *buf = (char *)__get_free_page(GFP_KERNEL);
4134 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004135 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004136
Jan Blunckcf28b482008-02-14 19:38:44 -08004137 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004138 if (IS_ERR(p))
4139 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004140 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004141 vma->vm_start,
4142 vma->vm_end - vma->vm_start);
4143 free_page((unsigned long)buf);
4144 }
4145 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004146 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004147}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004148
4149#ifdef CONFIG_PROVE_LOCKING
4150void might_fault(void)
4151{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004152 /*
4153 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4154 * holding the mmap_sem, this is safe because kernel memory doesn't
4155 * get paged out, therefore we'll never actually fault, and the
4156 * below annotations will generate false positives.
4157 */
4158 if (segment_eq(get_fs(), KERNEL_DS))
4159 return;
4160
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004161 might_sleep();
4162 /*
4163 * it would be nicer only to annotate paths which are not under
4164 * pagefault_disable, however that requires a larger audit and
4165 * providing helpers like get_user_atomic.
4166 */
4167 if (!in_atomic() && current->mm)
4168 might_lock_read(&current->mm->mmap_sem);
4169}
4170EXPORT_SYMBOL(might_fault);
4171#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004172
4173#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4174static void clear_gigantic_page(struct page *page,
4175 unsigned long addr,
4176 unsigned int pages_per_huge_page)
4177{
4178 int i;
4179 struct page *p = page;
4180
4181 might_sleep();
4182 for (i = 0; i < pages_per_huge_page;
4183 i++, p = mem_map_next(p, page, i)) {
4184 cond_resched();
4185 clear_user_highpage(p, addr + i * PAGE_SIZE);
4186 }
4187}
4188void clear_huge_page(struct page *page,
4189 unsigned long addr, unsigned int pages_per_huge_page)
4190{
4191 int i;
4192
4193 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4194 clear_gigantic_page(page, addr, pages_per_huge_page);
4195 return;
4196 }
4197
4198 might_sleep();
4199 for (i = 0; i < pages_per_huge_page; i++) {
4200 cond_resched();
4201 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4202 }
4203}
4204
4205static void copy_user_gigantic_page(struct page *dst, struct page *src,
4206 unsigned long addr,
4207 struct vm_area_struct *vma,
4208 unsigned int pages_per_huge_page)
4209{
4210 int i;
4211 struct page *dst_base = dst;
4212 struct page *src_base = src;
4213
4214 for (i = 0; i < pages_per_huge_page; ) {
4215 cond_resched();
4216 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4217
4218 i++;
4219 dst = mem_map_next(dst, dst_base, i);
4220 src = mem_map_next(src, src_base, i);
4221 }
4222}
4223
4224void copy_user_huge_page(struct page *dst, struct page *src,
4225 unsigned long addr, struct vm_area_struct *vma,
4226 unsigned int pages_per_huge_page)
4227{
4228 int i;
4229
4230 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4231 copy_user_gigantic_page(dst, src, addr, vma,
4232 pages_per_huge_page);
4233 return;
4234 }
4235
4236 might_sleep();
4237 for (i = 0; i < pages_per_huge_page; i++) {
4238 cond_resched();
4239 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4240 }
4241}
4242#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */