blob: 6953d3926e01e8ad1370304064ec2a0e94ad1ebb [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>
50#include <linux/module.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070053#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080054#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070055#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080056#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090059#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Alexey Dobriyan6952b612009-09-18 23:55:55 +040061#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062#include <asm/pgalloc.h>
63#include <asm/uaccess.h>
64#include <asm/tlb.h>
65#include <asm/tlbflush.h>
66#include <asm/pgtable.h>
67
Jan Beulich42b77722008-07-23 21:27:10 -070068#include "internal.h"
69
Andy Whitcroftd41dee32005-06-23 00:07:54 -070070#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070071/* use the per-pgdat data instead for discontigmem - mbligh */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned long num_physpages;
80/*
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 * and ZONE_HIGHMEM.
86 */
87void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89EXPORT_SYMBOL(num_physpages);
90EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Ingo Molnar32a93232008-02-06 22:39:44 +010092/*
93 * Randomize the address space (stacks, mmaps, brk, etc.).
94 *
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
97 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100 1;
101#else
102 2;
103#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100104
105static int __init disable_randmaps(char *s)
106{
107 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800108 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100109}
110__setup("norandmaps", disable_randmaps);
111
Hugh Dickins62eede62009-09-21 17:03:34 -0700112unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700113unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
122}
123core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100124
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800125
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800126#if defined(SPLIT_RSS_COUNTING)
127
KAMEZAWA Hiroyukia3a2e762010-04-06 14:34:42 -0700128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800129{
130 int i;
131
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
136 }
137 }
138 task->rss_stat.events = 0;
139}
140
141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142{
143 struct task_struct *task = current;
144
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
149}
150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153/* sync counter once per 64 page faults */
154#define TASK_RSS_EVENTS_THRESH (64)
155static void check_sync_rss_stat(struct task_struct *task)
156{
157 if (unlikely(task != current))
158 return;
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
161}
162
163unsigned long get_mm_counter(struct mm_struct *mm, int member)
164{
165 long val = 0;
166
167 /*
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
170 */
171 val = atomic_long_read(&mm->rss_stat.count[member]);
172 /*
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
175 */
176 if (val < 0)
177 return 0;
178 return (unsigned long)val;
179}
180
181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182{
183 __sync_task_rss_stat(task, mm);
184}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700185#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800186
187#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190static void check_sync_rss_stat(struct task_struct *task)
191{
192}
193
Peter Zijlstra9547d012011-05-24 17:12:14 -0700194#endif /* SPLIT_RSS_COUNTING */
195
196#ifdef HAVE_GENERIC_MMU_GATHER
197
198static int tlb_next_batch(struct mmu_gather *tlb)
199{
200 struct mmu_gather_batch *batch;
201
202 batch = tlb->active;
203 if (batch->next) {
204 tlb->active = batch->next;
205 return 1;
206 }
207
208 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209 if (!batch)
210 return 0;
211
212 batch->next = NULL;
213 batch->nr = 0;
214 batch->max = MAX_GATHER_BATCH;
215
216 tlb->active->next = batch;
217 tlb->active = batch;
218
219 return 1;
220}
221
222/* tlb_gather_mmu
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
226 */
227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228{
229 tlb->mm = mm;
230
231 tlb->fullmm = fullmm;
232 tlb->need_flush = 0;
233 tlb->fast_mode = (num_possible_cpus() == 1);
234 tlb->local.next = NULL;
235 tlb->local.nr = 0;
236 tlb->local.max = ARRAY_SIZE(tlb->__pages);
237 tlb->active = &tlb->local;
238
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800241#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700242}
243
244void tlb_flush_mmu(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 if (!tlb->need_flush)
249 return;
250 tlb->need_flush = 0;
251 tlb_flush(tlb);
252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
253 tlb_table_flush(tlb);
254#endif
255
256 if (tlb_fast_mode(tlb))
257 return;
258
259 for (batch = &tlb->local; batch; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
261 batch->nr = 0;
262 }
263 tlb->active = &tlb->local;
264}
265
266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294 struct mmu_gather_batch *batch;
295
296 tlb->need_flush = 1;
297
298 if (tlb_fast_mode(tlb)) {
299 free_page_and_swap_cache(page);
300 return 1; /* avoid calling tlb_flush_mmu() */
301 }
302
303 batch = tlb->active;
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
307 return 0;
308 }
309 VM_BUG_ON(batch->nr > batch->max);
310
311 return batch->max - batch->nr;
312}
313
314#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800315
Peter Zijlstra26723912011-05-24 17:12:00 -0700316#ifdef CONFIG_HAVE_RCU_TABLE_FREE
317
318/*
319 * See the comment near struct mmu_table_batch.
320 */
321
322static void tlb_remove_table_smp_sync(void *arg)
323{
324 /* Simply deliver the interrupt */
325}
326
327static void tlb_remove_table_one(void *table)
328{
329 /*
330 * This isn't an RCU grace period and hence the page-tables cannot be
331 * assumed to be actually RCU-freed.
332 *
333 * It is however sufficient for software page-table walkers that rely on
334 * IRQ disabling. See the comment near struct mmu_table_batch.
335 */
336 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
337 __tlb_remove_table(table);
338}
339
340static void tlb_remove_table_rcu(struct rcu_head *head)
341{
342 struct mmu_table_batch *batch;
343 int i;
344
345 batch = container_of(head, struct mmu_table_batch, rcu);
346
347 for (i = 0; i < batch->nr; i++)
348 __tlb_remove_table(batch->tables[i]);
349
350 free_page((unsigned long)batch);
351}
352
353void tlb_table_flush(struct mmu_gather *tlb)
354{
355 struct mmu_table_batch **batch = &tlb->batch;
356
357 if (*batch) {
358 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
359 *batch = NULL;
360 }
361}
362
363void tlb_remove_table(struct mmu_gather *tlb, void *table)
364{
365 struct mmu_table_batch **batch = &tlb->batch;
366
367 tlb->need_flush = 1;
368
369 /*
370 * When there's less then two users of this mm there cannot be a
371 * concurrent page-table walk.
372 */
373 if (atomic_read(&tlb->mm->mm_users) < 2) {
374 __tlb_remove_table(table);
375 return;
376 }
377
378 if (*batch == NULL) {
379 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
380 if (*batch == NULL) {
381 tlb_remove_table_one(table);
382 return;
383 }
384 (*batch)->nr = 0;
385 }
386 (*batch)->tables[(*batch)->nr++] = table;
387 if ((*batch)->nr == MAX_TABLE_BATCH)
388 tlb_table_flush(tlb);
389}
390
Peter Zijlstra9547d012011-05-24 17:12:14 -0700391#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700392
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393/*
394 * If a p?d_bad entry is found while walking page tables, report
395 * the error, before resetting entry to p?d_none. Usually (but
396 * very seldom) called out from the p?d_none_or_clear_bad macros.
397 */
398
399void pgd_clear_bad(pgd_t *pgd)
400{
401 pgd_ERROR(*pgd);
402 pgd_clear(pgd);
403}
404
405void pud_clear_bad(pud_t *pud)
406{
407 pud_ERROR(*pud);
408 pud_clear(pud);
409}
410
411void pmd_clear_bad(pmd_t *pmd)
412{
413 pmd_ERROR(*pmd);
414 pmd_clear(pmd);
415}
416
417/*
418 * Note: this doesn't free the actual pages themselves. That
419 * has been handled earlier when unmapping all the memory regions.
420 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000421static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
422 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800424 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700425 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000426 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700427 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
431 unsigned long addr, unsigned long end,
432 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
434 pmd_t *pmd;
435 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700436 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 do {
441 next = pmd_addr_end(addr, end);
442 if (pmd_none_or_clear_bad(pmd))
443 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000444 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 } while (pmd++, addr = next, addr != end);
446
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447 start &= PUD_MASK;
448 if (start < floor)
449 return;
450 if (ceiling) {
451 ceiling &= PUD_MASK;
452 if (!ceiling)
453 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 if (end - 1 > ceiling - 1)
456 return;
457
458 pmd = pmd_offset(pud, start);
459 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000460 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461}
462
Hugh Dickinse0da3822005-04-19 13:29:15 -0700463static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
464 unsigned long addr, unsigned long end,
465 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466{
467 pud_t *pud;
468 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700469 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 do {
474 next = pud_addr_end(addr, end);
475 if (pud_none_or_clear_bad(pud))
476 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700477 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 } while (pud++, addr = next, addr != end);
479
Hugh Dickinse0da3822005-04-19 13:29:15 -0700480 start &= PGDIR_MASK;
481 if (start < floor)
482 return;
483 if (ceiling) {
484 ceiling &= PGDIR_MASK;
485 if (!ceiling)
486 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700488 if (end - 1 > ceiling - 1)
489 return;
490
491 pud = pud_offset(pgd, start);
492 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000493 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494}
495
496/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700497 * This function frees user-level page tables of a process.
498 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 * Must be called with pagetable lock held.
500 */
Jan Beulich42b77722008-07-23 21:27:10 -0700501void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700502 unsigned long addr, unsigned long end,
503 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504{
505 pgd_t *pgd;
506 unsigned long next;
507
Hugh Dickinse0da3822005-04-19 13:29:15 -0700508 /*
509 * The next few lines have given us lots of grief...
510 *
511 * Why are we testing PMD* at this top level? Because often
512 * there will be no work to do at all, and we'd prefer not to
513 * go all the way down to the bottom just to discover that.
514 *
515 * Why all these "- 1"s? Because 0 represents both the bottom
516 * of the address space and the top of it (using -1 for the
517 * top wouldn't help much: the masks would do the wrong thing).
518 * The rule is that addr 0 and floor 0 refer to the bottom of
519 * the address space, but end 0 and ceiling 0 refer to the top
520 * Comparisons need to use "end - 1" and "ceiling - 1" (though
521 * that end 0 case should be mythical).
522 *
523 * Wherever addr is brought up or ceiling brought down, we must
524 * be careful to reject "the opposite 0" before it confuses the
525 * subsequent tests. But what about where end is brought down
526 * by PMD_SIZE below? no, end can't go down to 0 there.
527 *
528 * Whereas we round start (addr) and ceiling down, by different
529 * masks at different levels, in order to test whether a table
530 * now has no other vmas using it, so can be freed, we don't
531 * bother to round floor or end up - the tests don't need that.
532 */
533
534 addr &= PMD_MASK;
535 if (addr < floor) {
536 addr += PMD_SIZE;
537 if (!addr)
538 return;
539 }
540 if (ceiling) {
541 ceiling &= PMD_MASK;
542 if (!ceiling)
543 return;
544 }
545 if (end - 1 > ceiling - 1)
546 end -= PMD_SIZE;
547 if (addr > end - 1)
548 return;
549
Jan Beulich42b77722008-07-23 21:27:10 -0700550 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 do {
552 next = pgd_addr_end(addr, end);
553 if (pgd_none_or_clear_bad(pgd))
554 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700555 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700557}
558
Jan Beulich42b77722008-07-23 21:27:10 -0700559void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700560 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700561{
562 while (vma) {
563 struct vm_area_struct *next = vma->vm_next;
564 unsigned long addr = vma->vm_start;
565
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700566 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000567 * Hide vma from rmap and truncate_pagecache before freeing
568 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700569 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800570 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700571 unlink_file_vma(vma);
572
David Gibson9da61ae2006-03-22 00:08:57 -0800573 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700574 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700575 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700576 } else {
577 /*
578 * Optimization: gather nearby vmas into one call down
579 */
580 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800581 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700582 vma = next;
583 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800584 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700585 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700586 }
587 free_pgd_range(tlb, addr, vma->vm_end,
588 floor, next? next->vm_start: ceiling);
589 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700590 vma = next;
591 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592}
593
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800594int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
595 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800597 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800598 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700599 if (!new)
600 return -ENOMEM;
601
Nick Piggin362a61a2008-05-14 06:37:36 +0200602 /*
603 * Ensure all pte setup (eg. pte page lock and page clearing) are
604 * visible before the pte is made visible to other CPUs by being
605 * put into page tables.
606 *
607 * The other side of the story is the pointer chasing in the page
608 * table walking code (when walking the page table without locking;
609 * ie. most of the time). Fortunately, these data accesses consist
610 * of a chain of data-dependent loads, meaning most CPUs (alpha
611 * being the notable exception) will already guarantee loads are
612 * seen in-order. See the alpha page table accessors for the
613 * smp_read_barrier_depends() barriers in page table walking code.
614 */
615 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
616
Hugh Dickinsc74df322005-10-29 18:16:23 -0700617 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800618 wait_split_huge_page = 0;
619 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800622 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800623 } else if (unlikely(pmd_trans_splitting(*pmd)))
624 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700625 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800626 if (new)
627 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800628 if (wait_split_huge_page)
629 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700630 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631}
632
Hugh Dickins1bb36302005-10-29 18:16:22 -0700633int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700635 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
636 if (!new)
637 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638
Nick Piggin362a61a2008-05-14 06:37:36 +0200639 smp_wmb(); /* See comment in __pte_alloc */
640
Hugh Dickins1bb36302005-10-29 18:16:22 -0700641 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800642 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700643 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800644 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800645 } else
646 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700647 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800648 if (new)
649 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700650 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651}
652
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800653static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700654{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800655 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
656}
657
658static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
659{
660 int i;
661
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800662 if (current->mm == mm)
663 sync_mm_rss(current, mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800664 for (i = 0; i < NR_MM_COUNTERS; i++)
665 if (rss[i])
666 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700667}
668
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800670 * This function is called to print an error when a bad pte
671 * is found. For example, we might have a PFN-mapped pte in
672 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700673 *
674 * The calling function must still handle the error.
675 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800676static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
677 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700678{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800679 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
680 pud_t *pud = pud_offset(pgd, addr);
681 pmd_t *pmd = pmd_offset(pud, addr);
682 struct address_space *mapping;
683 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800684 static unsigned long resume;
685 static unsigned long nr_shown;
686 static unsigned long nr_unshown;
687
688 /*
689 * Allow a burst of 60 reports, then keep quiet for that minute;
690 * or allow a steady drip of one report per second.
691 */
692 if (nr_shown == 60) {
693 if (time_before(jiffies, resume)) {
694 nr_unshown++;
695 return;
696 }
697 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800698 printk(KERN_ALERT
699 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800700 nr_unshown);
701 nr_unshown = 0;
702 }
703 nr_shown = 0;
704 }
705 if (nr_shown++ == 0)
706 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800707
708 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
709 index = linear_page_index(vma, addr);
710
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800711 printk(KERN_ALERT
712 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800713 current->comm,
714 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800715 if (page)
716 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800717 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800718 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
719 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
720 /*
721 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
722 */
723 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800724 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800725 (unsigned long)vma->vm_ops->fault);
726 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800727 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800728 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700729 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800730 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700731}
732
KOSAKI Motohiroca16d142011-05-26 19:16:19 +0900733static inline int is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800734{
735 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
736}
737
Hugh Dickins62eede62009-09-21 17:03:34 -0700738#ifndef is_zero_pfn
739static inline int is_zero_pfn(unsigned long pfn)
740{
741 return pfn == zero_pfn;
742}
743#endif
744
745#ifndef my_zero_pfn
746static inline unsigned long my_zero_pfn(unsigned long addr)
747{
748 return zero_pfn;
749}
750#endif
751
Nick Pigginb5810032005-10-29 18:16:12 -0700752/*
Nick Piggin7e675132008-04-28 02:13:00 -0700753 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800754 *
Nick Piggin7e675132008-04-28 02:13:00 -0700755 * "Special" mappings do not wish to be associated with a "struct page" (either
756 * it doesn't exist, or it exists but they don't want to touch it). In this
757 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700758 *
Nick Piggin7e675132008-04-28 02:13:00 -0700759 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
760 * pte bit, in which case this function is trivial. Secondly, an architecture
761 * may not have a spare pte bit, which requires a more complicated scheme,
762 * described below.
763 *
764 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
765 * special mapping (even if there are underlying and valid "struct pages").
766 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800767 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700768 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
769 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700770 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
771 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800772 *
773 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
774 *
Nick Piggin7e675132008-04-28 02:13:00 -0700775 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700776 *
Nick Piggin7e675132008-04-28 02:13:00 -0700777 * This restricts such mappings to be a linear translation from virtual address
778 * to pfn. To get around this restriction, we allow arbitrary mappings so long
779 * as the vma is not a COW mapping; in that case, we know that all ptes are
780 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700781 *
782 *
Nick Piggin7e675132008-04-28 02:13:00 -0700783 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
784 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700785 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
786 * page" backing, however the difference is that _all_ pages with a struct
787 * page (that is, those where pfn_valid is true) are refcounted and considered
788 * normal pages by the VM. The disadvantage is that pages are refcounted
789 * (which can be slower and simply not an option for some PFNMAP users). The
790 * advantage is that we don't have to follow the strict linearity rule of
791 * PFNMAP mappings in order to support COWable mappings.
792 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800793 */
Nick Piggin7e675132008-04-28 02:13:00 -0700794#ifdef __HAVE_ARCH_PTE_SPECIAL
795# define HAVE_PTE_SPECIAL 1
796#else
797# define HAVE_PTE_SPECIAL 0
798#endif
799struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
800 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800801{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800802 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700803
804 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800805 if (likely(!pte_special(pte)))
806 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700807 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
808 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700809 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800810 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700811 return NULL;
812 }
813
814 /* !HAVE_PTE_SPECIAL case follows: */
815
Jared Hulbertb379d792008-04-28 02:12:58 -0700816 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
817 if (vma->vm_flags & VM_MIXEDMAP) {
818 if (!pfn_valid(pfn))
819 return NULL;
820 goto out;
821 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700822 unsigned long off;
823 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700824 if (pfn == vma->vm_pgoff + off)
825 return NULL;
826 if (!is_cow_mapping(vma->vm_flags))
827 return NULL;
828 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800829 }
830
Hugh Dickins62eede62009-09-21 17:03:34 -0700831 if (is_zero_pfn(pfn))
832 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800833check_pfn:
834 if (unlikely(pfn > highest_memmap_pfn)) {
835 print_bad_pte(vma, addr, pte, NULL);
836 return NULL;
837 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800838
839 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700840 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700841 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800842 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700843out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800844 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800845}
846
847/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 * copy one vm_area from one task to the other. Assumes the page tables
849 * already present in the new task to be cleared in the whole range
850 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 */
852
Hugh Dickins570a335b2009-12-14 17:58:46 -0800853static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700855 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700856 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
Nick Pigginb5810032005-10-29 18:16:12 -0700858 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 pte_t pte = *src_pte;
860 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 /* pte contains position in swap or file, so copy. */
863 if (unlikely(!pte_present(pte))) {
864 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700865 swp_entry_t entry = pte_to_swp_entry(pte);
866
Hugh Dickins570a335b2009-12-14 17:58:46 -0800867 if (swap_duplicate(entry) < 0)
868 return entry.val;
869
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 /* make sure dst_mm is on swapoff's mmlist. */
871 if (unlikely(list_empty(&dst_mm->mmlist))) {
872 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700873 if (list_empty(&dst_mm->mmlist))
874 list_add(&dst_mm->mmlist,
875 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 spin_unlock(&mmlist_lock);
877 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800878 if (likely(!non_swap_entry(entry)))
879 rss[MM_SWAPENTS]++;
880 else if (is_write_migration_entry(entry) &&
Christoph Lameter06972122006-06-23 02:03:35 -0700881 is_cow_mapping(vm_flags)) {
882 /*
883 * COW mappings require pages in both parent
884 * and child to be set to read.
885 */
886 make_migration_entry_read(&entry);
887 pte = swp_entry_to_pte(entry);
888 set_pte_at(src_mm, addr, src_pte, pte);
889 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700891 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 }
893
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 /*
895 * If it's a COW mapping, write protect it both
896 * in the parent and the child
897 */
Linus Torvalds67121172005-12-11 20:38:17 -0800898 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700900 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 }
902
903 /*
904 * If it's a shared mapping, mark it clean in
905 * the child
906 */
907 if (vm_flags & VM_SHARED)
908 pte = pte_mkclean(pte);
909 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800910
911 page = vm_normal_page(vma, addr, pte);
912 if (page) {
913 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700914 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800915 if (PageAnon(page))
916 rss[MM_ANONPAGES]++;
917 else
918 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800919 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700920
921out_set_pte:
922 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800923 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924}
925
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800926int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
927 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
928 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700930 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700932 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700933 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800934 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800935 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
937again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800938 init_rss_vec(rss);
939
Hugh Dickinsc74df322005-10-29 18:16:23 -0700940 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 if (!dst_pte)
942 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700943 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700944 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700945 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700946 orig_src_pte = src_pte;
947 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700948 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 do {
951 /*
952 * We are holding two locks at this point - either of them
953 * could generate latencies in another task on another CPU.
954 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700955 if (progress >= 32) {
956 progress = 0;
957 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100958 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700959 break;
960 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 if (pte_none(*src_pte)) {
962 progress++;
963 continue;
964 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800965 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
966 vma, addr, rss);
967 if (entry.val)
968 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 progress += 8;
970 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700972 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700973 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700974 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800975 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700976 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700977 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800978
979 if (entry.val) {
980 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
981 return -ENOMEM;
982 progress = 0;
983 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 if (addr != end)
985 goto again;
986 return 0;
987}
988
989static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
990 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
991 unsigned long addr, unsigned long end)
992{
993 pmd_t *src_pmd, *dst_pmd;
994 unsigned long next;
995
996 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
997 if (!dst_pmd)
998 return -ENOMEM;
999 src_pmd = pmd_offset(src_pud, addr);
1000 do {
1001 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001002 if (pmd_trans_huge(*src_pmd)) {
1003 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001004 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001005 err = copy_huge_pmd(dst_mm, src_mm,
1006 dst_pmd, src_pmd, addr, vma);
1007 if (err == -ENOMEM)
1008 return -ENOMEM;
1009 if (!err)
1010 continue;
1011 /* fall through */
1012 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 if (pmd_none_or_clear_bad(src_pmd))
1014 continue;
1015 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1019 return 0;
1020}
1021
1022static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1024 unsigned long addr, unsigned long end)
1025{
1026 pud_t *src_pud, *dst_pud;
1027 unsigned long next;
1028
1029 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030 if (!dst_pud)
1031 return -ENOMEM;
1032 src_pud = pud_offset(src_pgd, addr);
1033 do {
1034 next = pud_addr_end(addr, end);
1035 if (pud_none_or_clear_bad(src_pud))
1036 continue;
1037 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038 vma, addr, next))
1039 return -ENOMEM;
1040 } while (dst_pud++, src_pud++, addr = next, addr != end);
1041 return 0;
1042}
1043
1044int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 struct vm_area_struct *vma)
1046{
1047 pgd_t *src_pgd, *dst_pgd;
1048 unsigned long next;
1049 unsigned long addr = vma->vm_start;
1050 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001051 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
Nick Piggind9928952005-08-28 16:49:11 +10001053 /*
1054 * Don't copy ptes where a page fault will fill them correctly.
1055 * Fork becomes much lighter when there are big shared or private
1056 * readonly mappings. The tradeoff is that copy_page_range is more
1057 * efficient than faulting.
1058 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001059 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +10001060 if (!vma->anon_vma)
1061 return 0;
1062 }
1063
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (is_vm_hugetlb_page(vma))
1065 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1066
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001067 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001068 /*
1069 * We do not free on error cases below as remove_vma
1070 * gets called on error from higher level routine
1071 */
1072 ret = track_pfn_vma_copy(vma);
1073 if (ret)
1074 return ret;
1075 }
1076
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001077 /*
1078 * We need to invalidate the secondary MMU mappings only when
1079 * there could be a permission downgrade on the ptes of the
1080 * parent mm. And a permission downgrade will only happen if
1081 * is_cow_mapping() returns true.
1082 */
1083 if (is_cow_mapping(vma->vm_flags))
1084 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1085
1086 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 dst_pgd = pgd_offset(dst_mm, addr);
1088 src_pgd = pgd_offset(src_mm, addr);
1089 do {
1090 next = pgd_addr_end(addr, end);
1091 if (pgd_none_or_clear_bad(src_pgd))
1092 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001093 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1094 vma, addr, next))) {
1095 ret = -ENOMEM;
1096 break;
1097 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001099
1100 if (is_cow_mapping(vma->vm_flags))
1101 mmu_notifier_invalidate_range_end(src_mm,
1102 vma->vm_start, end);
1103 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104}
1105
Robin Holt51c6f662005-11-13 16:06:42 -08001106static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001107 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001109 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110{
Nick Pigginb5810032005-10-29 18:16:12 -07001111 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001112 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001113 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001114 spinlock_t *ptl;
1115 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001116
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001117again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001118 init_rss_vec(rss);
Hugh Dickins508034a2005-10-29 18:16:30 -07001119 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001120 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 do {
1122 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001123 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001125 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001126
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001128 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001129
Linus Torvalds6aab3412005-11-28 14:34:23 -08001130 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 if (unlikely(details) && page) {
1132 /*
1133 * unmap_shared_mapping_pages() wants to
1134 * invalidate cache without truncating:
1135 * unmap shared but keep private pages.
1136 */
1137 if (details->check_mapping &&
1138 details->check_mapping != page->mapping)
1139 continue;
1140 /*
1141 * Each page->index must be checked when
1142 * invalidating or truncating nonlinear.
1143 */
1144 if (details->nonlinear_vma &&
1145 (page->index < details->first_index ||
1146 page->index > details->last_index))
1147 continue;
1148 }
Nick Pigginb5810032005-10-29 18:16:12 -07001149 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001150 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 tlb_remove_tlb_entry(tlb, pte, addr);
1152 if (unlikely(!page))
1153 continue;
1154 if (unlikely(details) && details->nonlinear_vma
1155 && linear_page_index(details->nonlinear_vma,
1156 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001157 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001160 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001161 else {
1162 if (pte_dirty(ptent))
1163 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001164 if (pte_young(ptent) &&
1165 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001166 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001167 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001168 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001169 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001170 if (unlikely(page_mapcount(page) < 0))
1171 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001172 force_flush = !__tlb_remove_page(tlb, page);
1173 if (force_flush)
1174 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 continue;
1176 }
1177 /*
1178 * If details->check_mapping, we leave swap entries;
1179 * if details->nonlinear_vma, we leave file entries.
1180 */
1181 if (unlikely(details))
1182 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001183 if (pte_file(ptent)) {
1184 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1185 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001186 } else {
1187 swp_entry_t entry = pte_to_swp_entry(ptent);
1188
1189 if (!non_swap_entry(entry))
1190 rss[MM_SWAPENTS]--;
1191 if (unlikely(!free_swap_and_cache(entry)))
1192 print_bad_pte(vma, addr, ptent, NULL);
1193 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001194 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001195 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001196
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001197 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001198 arch_leave_lazy_mmu_mode();
Hugh Dickins508034a2005-10-29 18:16:30 -07001199 pte_unmap_unlock(pte - 1, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001200
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001201 /*
1202 * mmu_gather ran out of room to batch pages, we break out of
1203 * the PTE lock to avoid doing the potential expensive TLB invalidate
1204 * and page-free while holding it.
1205 */
1206 if (force_flush) {
1207 force_flush = 0;
1208 tlb_flush_mmu(tlb);
1209 if (addr != end)
1210 goto again;
1211 }
1212
Robin Holt51c6f662005-11-13 16:06:42 -08001213 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214}
1215
Robin Holt51c6f662005-11-13 16:06:42 -08001216static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001217 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001219 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220{
1221 pmd_t *pmd;
1222 unsigned long next;
1223
1224 pmd = pmd_offset(pud, addr);
1225 do {
1226 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001227 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001228 if (next-addr != HPAGE_PMD_SIZE) {
1229 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 split_huge_page_pmd(vma->vm_mm, pmd);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001231 } else if (zap_huge_pmd(tlb, vma, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001232 continue;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 /* fall through */
1234 }
Peter Zijlstra97a89412011-05-24 17:12:04 -07001235 if (pmd_none_or_clear_bad(pmd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001237 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1238 cond_resched();
1239 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001240
1241 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242}
1243
Robin Holt51c6f662005-11-13 16:06:42 -08001244static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001245 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001247 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248{
1249 pud_t *pud;
1250 unsigned long next;
1251
1252 pud = pud_offset(pgd, addr);
1253 do {
1254 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001255 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001257 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1258 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001259
1260 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261}
1262
Robin Holt51c6f662005-11-13 16:06:42 -08001263static unsigned long unmap_page_range(struct mmu_gather *tlb,
1264 struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001266 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267{
1268 pgd_t *pgd;
1269 unsigned long next;
1270
1271 if (details && !details->check_mapping && !details->nonlinear_vma)
1272 details = NULL;
1273
1274 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001275 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 tlb_start_vma(tlb, vma);
1277 pgd = pgd_offset(vma->vm_mm, addr);
1278 do {
1279 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001280 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001282 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1283 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001285 mem_cgroup_uncharge_end();
Robin Holt51c6f662005-11-13 16:06:42 -08001286
1287 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288}
1289
1290#ifdef CONFIG_PREEMPT
1291# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1292#else
1293/* No preempt: go for improved straight-line efficiency */
1294# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1295#endif
1296
1297/**
1298 * unmap_vmas - unmap a range of memory covered by a list of vma's
1299 * @tlbp: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 * @vma: the starting vma
1301 * @start_addr: virtual address at which to start unmapping
1302 * @end_addr: virtual address at which to end unmapping
1303 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1304 * @details: details of nonlinear truncation or shared cache invalidation
1305 *
Hugh Dickinsee39b372005-04-19 13:29:15 -07001306 * Returns the end address of the unmapping (restart addr if interrupted).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001308 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001310 * We aim to not hold locks for too long (for scheduling latency reasons).
1311 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 * return the ending mmu_gather to the caller.
1313 *
1314 * Only addresses between `start' and `end' will be unmapped.
1315 *
1316 * The VMA list must be sorted in ascending virtual address order.
1317 *
1318 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1319 * range after unmap_vmas() returns. So the only responsibility here is to
1320 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1321 * drops the lock and schedules.
1322 */
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001323unsigned long unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 struct vm_area_struct *vma, unsigned long start_addr,
1325 unsigned long end_addr, unsigned long *nr_accounted,
1326 struct zap_details *details)
1327{
Hugh Dickinsee39b372005-04-19 13:29:15 -07001328 unsigned long start = start_addr;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001329 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001331 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 unsigned long end;
1334
1335 start = max(vma->vm_start, start_addr);
1336 if (start >= vma->vm_end)
1337 continue;
1338 end = min(vma->vm_end, end_addr);
1339 if (end <= vma->vm_start)
1340 continue;
1341
1342 if (vma->vm_flags & VM_ACCOUNT)
1343 *nr_accounted += (end - start) >> PAGE_SHIFT;
1344
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001345 if (unlikely(is_pfn_mapping(vma)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001346 untrack_pfn_vma(vma, 0, 0);
1347
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 while (start != end) {
Robin Holt51c6f662005-11-13 16:06:42 -08001349 if (unlikely(is_vm_hugetlb_page(vma))) {
Andi Kleena137e1c2008-07-23 21:27:43 -07001350 /*
1351 * It is undesirable to test vma->vm_file as it
1352 * should be non-null for valid hugetlb area.
1353 * However, vm_file will be NULL in the error
1354 * cleanup path of do_mmap_pgoff. When
1355 * hugetlbfs ->mmap method fails,
1356 * do_mmap_pgoff() nullifies vma->vm_file
1357 * before calling this function to clean up.
1358 * Since no pte has actually been setup, it is
1359 * safe to do nothing in this case.
1360 */
Peter Zijlstra97a89412011-05-24 17:12:04 -07001361 if (vma->vm_file)
Andi Kleena137e1c2008-07-23 21:27:43 -07001362 unmap_hugepage_range(vma, start, end, NULL);
Andi Kleena137e1c2008-07-23 21:27:43 -07001363
Robin Holt51c6f662005-11-13 16:06:42 -08001364 start = end;
1365 } else
Peter Zijlstra97a89412011-05-24 17:12:04 -07001366 start = unmap_page_range(tlb, vma, start, end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 }
1368 }
Peter Zijlstra97a89412011-05-24 17:12:04 -07001369
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001370 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001371 return start; /* which is now the end (or restart) address */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @address: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of nonlinear truncation or shared cache invalidation
1380 */
Hugh Dickinsee39b372005-04-19 13:29:15 -07001381unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 unsigned long size, struct zap_details *details)
1383{
1384 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001385 struct mmu_gather tlb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 unsigned long end = address + size;
1387 unsigned long nr_accounted = 0;
1388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001390 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001391 update_hiwater_rss(mm);
Hugh Dickins508034a2005-10-29 18:16:30 -07001392 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001393 tlb_finish_mmu(&tlb, address, end);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001394 return end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395}
1396
Jack Steinerc627f9c2008-07-29 22:33:53 -07001397/**
1398 * zap_vma_ptes - remove ptes mapping the vma
1399 * @vma: vm_area_struct holding ptes to be zapped
1400 * @address: starting address of pages to zap
1401 * @size: number of bytes to zap
1402 *
1403 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1404 *
1405 * The entire address range must be fully contained within the vma.
1406 *
1407 * Returns 0 if successful.
1408 */
1409int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1410 unsigned long size)
1411{
1412 if (address < vma->vm_start || address + size > vma->vm_end ||
1413 !(vma->vm_flags & VM_PFNMAP))
1414 return -1;
1415 zap_page_range(vma, address, size, NULL);
1416 return 0;
1417}
1418EXPORT_SYMBOL_GPL(zap_vma_ptes);
1419
Johannes Weiner142762b2010-05-24 14:32:39 -07001420/**
1421 * follow_page - look up a page descriptor from a user-virtual address
1422 * @vma: vm_area_struct mapping @address
1423 * @address: virtual address to look up
1424 * @flags: flags modifying lookup behaviour
1425 *
1426 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1427 *
1428 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1429 * an error pointer if there is a mapping to something not represented
1430 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001432struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001433 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434{
1435 pgd_t *pgd;
1436 pud_t *pud;
1437 pmd_t *pmd;
1438 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001439 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001441 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001443 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1444 if (!IS_ERR(page)) {
1445 BUG_ON(flags & FOLL_GET);
1446 goto out;
1447 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001449 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 pgd = pgd_offset(mm, address);
1451 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001452 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
1454 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001455 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001456 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001457 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001458 BUG_ON(flags & FOLL_GET);
1459 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1460 goto out;
1461 }
1462 if (unlikely(pud_bad(*pud)))
1463 goto no_page_table;
1464
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001466 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001467 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001468 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001469 BUG_ON(flags & FOLL_GET);
1470 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1471 goto out;
1472 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001473 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001474 if (flags & FOLL_SPLIT) {
1475 split_huge_page_pmd(mm, pmd);
1476 goto split_fallthrough;
1477 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001478 spin_lock(&mm->page_table_lock);
1479 if (likely(pmd_trans_huge(*pmd))) {
1480 if (unlikely(pmd_trans_splitting(*pmd))) {
1481 spin_unlock(&mm->page_table_lock);
1482 wait_split_huge_page(vma->anon_vma, pmd);
1483 } else {
1484 page = follow_trans_huge_pmd(mm, address,
1485 pmd, flags);
1486 spin_unlock(&mm->page_table_lock);
1487 goto out;
1488 }
1489 } else
1490 spin_unlock(&mm->page_table_lock);
1491 /* fall through */
1492 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001493split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001494 if (unlikely(pmd_bad(*pmd)))
1495 goto no_page_table;
1496
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001497 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
1499 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001500 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001501 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001502 if ((flags & FOLL_WRITE) && !pte_write(pte))
1503 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001504
Linus Torvalds6aab3412005-11-28 14:34:23 -08001505 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001506 if (unlikely(!page)) {
1507 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001508 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001509 goto bad_page;
1510 page = pte_page(pte);
1511 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001512
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001513 if (flags & FOLL_GET)
1514 get_page(page);
1515 if (flags & FOLL_TOUCH) {
1516 if ((flags & FOLL_WRITE) &&
1517 !pte_dirty(pte) && !PageDirty(page))
1518 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001519 /*
1520 * pte_mkyoung() would be more correct here, but atomic care
1521 * is needed to avoid losing the dirty bit: it is easier to use
1522 * mark_page_accessed().
1523 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001524 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001526 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001527 /*
1528 * The preliminary mapping check is mainly to avoid the
1529 * pointless overhead of lock_page on the ZERO_PAGE
1530 * which might bounce very badly if there is contention.
1531 *
1532 * If the page is already locked, we don't need to
1533 * handle it now - vmscan will handle it later if and
1534 * when it attempts to reclaim the page.
1535 */
1536 if (page->mapping && trylock_page(page)) {
1537 lru_add_drain(); /* push cached pages to LRU */
1538 /*
1539 * Because we lock page here and migration is
1540 * blocked by the pte's page reference, we need
1541 * only check for file-cache page truncation.
1542 */
1543 if (page->mapping)
1544 mlock_vma_page(page);
1545 unlock_page(page);
1546 }
1547 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001548unlock:
1549 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001551 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001553bad_page:
1554 pte_unmap_unlock(ptep, ptl);
1555 return ERR_PTR(-EFAULT);
1556
1557no_page:
1558 pte_unmap_unlock(ptep, ptl);
1559 if (!pte_none(pte))
1560 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001561
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001562no_page_table:
1563 /*
1564 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001565 * has touched so far, we don't want to allocate unnecessary pages or
1566 * page tables. Return error instead of NULL to skip handle_mm_fault,
1567 * then get_dump_page() will return NULL to leave a hole in the dump.
1568 * But we can only make this optimization where a hole would surely
1569 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001570 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001571 if ((flags & FOLL_DUMP) &&
1572 (!vma->vm_ops || !vma->vm_ops->fault))
1573 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001574 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575}
1576
Linus Torvalds95042f92011-04-12 14:15:51 -07001577static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1578{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001579 return stack_guard_page_start(vma, addr) ||
1580 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001581}
1582
Huang Ying0014bd92011-01-30 11:15:47 +08001583/**
1584 * __get_user_pages() - pin user pages in memory
1585 * @tsk: task_struct of target task
1586 * @mm: mm_struct of target mm
1587 * @start: starting user address
1588 * @nr_pages: number of pages from start to pin
1589 * @gup_flags: flags modifying pin behaviour
1590 * @pages: array that receives pointers to the pages pinned.
1591 * Should be at least nr_pages long. Or NULL, if caller
1592 * only intends to ensure the pages are faulted in.
1593 * @vmas: array of pointers to vmas corresponding to each page.
1594 * Or NULL if the caller does not require them.
1595 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1596 *
1597 * Returns number of pages pinned. This may be fewer than the number
1598 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1599 * were pinned, returns -errno. Each page returned must be released
1600 * with a put_page() call when it is finished with. vmas will only
1601 * remain valid while mmap_sem is held.
1602 *
1603 * Must be called with mmap_sem held for read or write.
1604 *
1605 * __get_user_pages walks a process's page tables and takes a reference to
1606 * each struct page that each user address corresponds to at a given
1607 * instant. That is, it takes the page that would be accessed if a user
1608 * thread accesses the given user virtual address at that instant.
1609 *
1610 * This does not guarantee that the page exists in the user mappings when
1611 * __get_user_pages returns, and there may even be a completely different
1612 * page there in some cases (eg. if mmapped pagecache has been invalidated
1613 * and subsequently re faulted). However it does guarantee that the page
1614 * won't be freed completely. And mostly callers simply care that the page
1615 * contains data that was valid *at some point in time*. Typically, an IO
1616 * or similar operation cannot guarantee anything stronger anyway because
1617 * locks can't be held over the syscall boundary.
1618 *
1619 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1620 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1621 * appropriate) must be called after the page is finished with, and
1622 * before put_page is called.
1623 *
1624 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1625 * or mmap_sem contention, and if waiting is needed to pin all pages,
1626 * *@nonblocking will be set to 0.
1627 *
1628 * In most cases, get_user_pages or get_user_pages_fast should be used
1629 * instead of __get_user_pages. __get_user_pages should be used only if
1630 * you need some special @gup_flags.
1631 */
Nick Pigginb291f002008-10-18 20:26:44 -07001632int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001633 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001634 struct page **pages, struct vm_area_struct **vmas,
1635 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636{
1637 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001638 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639
Peter Zijlstra9d737772009-06-25 11:58:55 +02001640 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001641 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001642
1643 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1644
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 /*
1646 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001647 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001649 vm_flags = (gup_flags & FOLL_WRITE) ?
1650 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1651 vm_flags &= (gup_flags & FOLL_FORCE) ?
1652 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 i = 0;
1654
1655 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001656 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657
1658 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001659 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 pgd_t *pgd;
1662 pud_t *pud;
1663 pmd_t *pmd;
1664 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001665
1666 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001667 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 return i ? : -EFAULT;
1669 if (pg > TASK_SIZE)
1670 pgd = pgd_offset_k(pg);
1671 else
1672 pgd = pgd_offset_gate(mm, pg);
1673 BUG_ON(pgd_none(*pgd));
1674 pud = pud_offset(pgd, pg);
1675 BUG_ON(pud_none(*pud));
1676 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001677 if (pmd_none(*pmd))
1678 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001679 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001681 if (pte_none(*pte)) {
1682 pte_unmap(pte);
1683 return i ? : -EFAULT;
1684 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001685 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001687 struct page *page;
1688
Linus Torvalds95042f92011-04-12 14:15:51 -07001689 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001690 if (!page) {
1691 if (!(gup_flags & FOLL_DUMP) &&
1692 is_zero_pfn(pte_pfn(*pte)))
1693 page = pte_page(*pte);
1694 else {
1695 pte_unmap(pte);
1696 return i ? : -EFAULT;
1697 }
1698 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001699 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001700 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 }
1702 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001703 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 }
1705
Nick Pigginb291f002008-10-18 20:26:44 -07001706 if (!vma ||
1707 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001708 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 return i ? : -EFAULT;
1710
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001711 if (is_vm_hugetlb_page(vma)) {
1712 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001713 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001714 continue;
1715 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001716
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001718 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001719 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Ethan Solomita462e00c2007-07-15 23:38:16 -07001721 /*
Ying Han47792802009-01-06 14:40:18 -08001722 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001723 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001724 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001725 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001726 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001727
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001728 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001729 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001730 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001731 unsigned int fault_flags = 0;
1732
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001733 /* For mlock, just skip the stack guard page. */
1734 if (foll_flags & FOLL_MLOCK) {
1735 if (stack_guard_page(vma, start))
1736 goto next_page;
1737 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001738 if (foll_flags & FOLL_WRITE)
1739 fault_flags |= FAULT_FLAG_WRITE;
1740 if (nonblocking)
1741 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001742 if (foll_flags & FOLL_NOWAIT)
1743 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001744
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001745 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001746 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001747
Nick Piggin83c54072007-07-19 01:47:05 -07001748 if (ret & VM_FAULT_ERROR) {
1749 if (ret & VM_FAULT_OOM)
1750 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001751 if (ret & (VM_FAULT_HWPOISON |
1752 VM_FAULT_HWPOISON_LARGE)) {
1753 if (i)
1754 return i;
1755 else if (gup_flags & FOLL_HWPOISON)
1756 return -EHWPOISON;
1757 else
1758 return -EFAULT;
1759 }
1760 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001761 return i ? i : -EFAULT;
1762 BUG();
1763 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001764
1765 if (tsk) {
1766 if (ret & VM_FAULT_MAJOR)
1767 tsk->maj_flt++;
1768 else
1769 tsk->min_flt++;
1770 }
Nick Piggin83c54072007-07-19 01:47:05 -07001771
Michel Lespinasse53a77062011-01-13 15:46:14 -08001772 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001773 if (nonblocking)
1774 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001775 return i;
1776 }
1777
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001778 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001779 * The VM_FAULT_WRITE bit tells us that
1780 * do_wp_page has broken COW when necessary,
1781 * even if maybe_mkwrite decided not to set
1782 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001783 * page lookups as if they were reads. But only
1784 * do so when looping for pte_write is futile:
1785 * in some cases userspace may also be wanting
1786 * to write to the gotten user page, which a
1787 * read fault here might prevent (a readonly
1788 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001789 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001790 if ((ret & VM_FAULT_WRITE) &&
1791 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001792 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001793
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001794 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001796 if (IS_ERR(page))
1797 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001799 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001800
Russell Kinga6f36be2006-12-30 22:24:19 +00001801 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001802 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001804next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 if (vmas)
1806 vmas[i] = vma;
1807 i++;
1808 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001809 nr_pages--;
1810 } while (nr_pages && start < vma->vm_end);
1811 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 return i;
1813}
Huang Ying0014bd92011-01-30 11:15:47 +08001814EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001815
Nick Piggind2bf6be2009-06-16 15:31:39 -07001816/**
1817 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001818 * @tsk: the task_struct to use for page fault accounting, or
1819 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001820 * @mm: mm_struct of target mm
1821 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001822 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001823 * @write: whether pages will be written to by the caller
1824 * @force: whether to force write access even if user mapping is
1825 * readonly. This will result in the page being COWed even
1826 * in MAP_SHARED mappings. You do not want this.
1827 * @pages: array that receives pointers to the pages pinned.
1828 * Should be at least nr_pages long. Or NULL, if caller
1829 * only intends to ensure the pages are faulted in.
1830 * @vmas: array of pointers to vmas corresponding to each page.
1831 * Or NULL if the caller does not require them.
1832 *
1833 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001834 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001835 * were pinned, returns -errno. Each page returned must be released
1836 * with a put_page() call when it is finished with. vmas will only
1837 * remain valid while mmap_sem is held.
1838 *
1839 * Must be called with mmap_sem held for read or write.
1840 *
1841 * get_user_pages walks a process's page tables and takes a reference to
1842 * each struct page that each user address corresponds to at a given
1843 * instant. That is, it takes the page that would be accessed if a user
1844 * thread accesses the given user virtual address at that instant.
1845 *
1846 * This does not guarantee that the page exists in the user mappings when
1847 * get_user_pages returns, and there may even be a completely different
1848 * page there in some cases (eg. if mmapped pagecache has been invalidated
1849 * and subsequently re faulted). However it does guarantee that the page
1850 * won't be freed completely. And mostly callers simply care that the page
1851 * contains data that was valid *at some point in time*. Typically, an IO
1852 * or similar operation cannot guarantee anything stronger anyway because
1853 * locks can't be held over the syscall boundary.
1854 *
1855 * If write=0, the page must not be written to. If the page is written to,
1856 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1857 * after the page is finished with, and before put_page is called.
1858 *
1859 * get_user_pages is typically used for fewer-copy IO operations, to get a
1860 * handle on the memory by some means other than accesses via the user virtual
1861 * addresses. The pages may be submitted for DMA to devices or accessed via
1862 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1863 * use the correct cache flushing APIs.
1864 *
1865 * See also get_user_pages_fast, for performance critical applications.
1866 */
Nick Pigginb291f002008-10-18 20:26:44 -07001867int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001868 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001869 struct page **pages, struct vm_area_struct **vmas)
1870{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001871 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001872
Hugh Dickins58fa8792009-09-21 17:03:31 -07001873 if (pages)
1874 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001875 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001876 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001877 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001878 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001879
Michel Lespinasse53a77062011-01-13 15:46:14 -08001880 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1881 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001882}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883EXPORT_SYMBOL(get_user_pages);
1884
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001885/**
1886 * get_dump_page() - pin user page in memory while writing it to core dump
1887 * @addr: user address
1888 *
1889 * Returns struct page pointer of user page pinned for dump,
1890 * to be freed afterwards by page_cache_release() or put_page().
1891 *
1892 * Returns NULL on any kind of failure - a hole must then be inserted into
1893 * the corefile, to preserve alignment with its headers; and also returns
1894 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1895 * allowing a hole to be left in the corefile to save diskspace.
1896 *
1897 * Called without mmap_sem, but after all other threads have been killed.
1898 */
1899#ifdef CONFIG_ELF_CORE
1900struct page *get_dump_page(unsigned long addr)
1901{
1902 struct vm_area_struct *vma;
1903 struct page *page;
1904
1905 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001906 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1907 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001908 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001909 flush_cache_page(vma, addr, page_to_pfn(page));
1910 return page;
1911}
1912#endif /* CONFIG_ELF_CORE */
1913
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001914pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001915 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001916{
1917 pgd_t * pgd = pgd_offset(mm, addr);
1918 pud_t * pud = pud_alloc(mm, pgd, addr);
1919 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001920 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001921 if (pmd) {
1922 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001923 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001924 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001925 }
1926 return NULL;
1927}
1928
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001930 * This is the old fallback for page remapping.
1931 *
1932 * For historical reasons, it only allows reserved pages. Only
1933 * old drivers should use this, and they needed to mark their
1934 * pages reserved for the old functions anyway.
1935 */
Nick Piggin423bad62008-04-28 02:13:01 -07001936static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1937 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001938{
Nick Piggin423bad62008-04-28 02:13:01 -07001939 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001940 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001941 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001942 spinlock_t *ptl;
1943
Linus Torvalds238f58d2005-11-29 13:01:56 -08001944 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001945 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001946 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001947 retval = -ENOMEM;
1948 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001949 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001950 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001951 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001952 retval = -EBUSY;
1953 if (!pte_none(*pte))
1954 goto out_unlock;
1955
1956 /* Ok, finally just insert the thing.. */
1957 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001958 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001959 page_add_file_rmap(page);
1960 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1961
1962 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001963 pte_unmap_unlock(pte, ptl);
1964 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001965out_unlock:
1966 pte_unmap_unlock(pte, ptl);
1967out:
1968 return retval;
1969}
1970
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001971/**
1972 * vm_insert_page - insert single page into user vma
1973 * @vma: user vma to map to
1974 * @addr: target user address of this page
1975 * @page: source kernel page
1976 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001977 * This allows drivers to insert individual pages they've allocated
1978 * into a user vma.
1979 *
1980 * The page has to be a nice clean _individual_ kernel allocation.
1981 * If you allocate a compound page, you need to have marked it as
1982 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001983 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001984 *
1985 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1986 * took an arbitrary page protection parameter. This doesn't allow
1987 * that. Your vma protection will have to be set up correctly, which
1988 * means that if you want a shared writable mapping, you'd better
1989 * ask for a shared writable mapping!
1990 *
1991 * The page does not need to be reserved.
1992 */
Nick Piggin423bad62008-04-28 02:13:01 -07001993int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1994 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001995{
1996 if (addr < vma->vm_start || addr >= vma->vm_end)
1997 return -EFAULT;
1998 if (!page_count(page))
1999 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08002000 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07002001 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002002}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002003EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002004
Nick Piggin423bad62008-04-28 02:13:01 -07002005static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2006 unsigned long pfn, pgprot_t prot)
2007{
2008 struct mm_struct *mm = vma->vm_mm;
2009 int retval;
2010 pte_t *pte, entry;
2011 spinlock_t *ptl;
2012
2013 retval = -ENOMEM;
2014 pte = get_locked_pte(mm, addr, &ptl);
2015 if (!pte)
2016 goto out;
2017 retval = -EBUSY;
2018 if (!pte_none(*pte))
2019 goto out_unlock;
2020
2021 /* Ok, finally just insert the thing.. */
2022 entry = pte_mkspecial(pfn_pte(pfn, prot));
2023 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002024 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002025
2026 retval = 0;
2027out_unlock:
2028 pte_unmap_unlock(pte, ptl);
2029out:
2030 return retval;
2031}
2032
Nick Piggine0dc0d82007-02-12 00:51:36 -08002033/**
2034 * vm_insert_pfn - insert single pfn into user vma
2035 * @vma: user vma to map to
2036 * @addr: target user address of this page
2037 * @pfn: source kernel pfn
2038 *
2039 * Similar to vm_inert_page, this allows drivers to insert individual pages
2040 * they've allocated into a user vma. Same comments apply.
2041 *
2042 * This function should only be called from a vm_ops->fault handler, and
2043 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002044 *
2045 * vma cannot be a COW mapping.
2046 *
2047 * As this is called only for pages that do not currently exist, we
2048 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002049 */
2050int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002051 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002052{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002053 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002054 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002055 /*
2056 * Technically, architectures with pte_special can avoid all these
2057 * restrictions (same for remap_pfn_range). However we would like
2058 * consistency in testing and feature parity among all, so we should
2059 * try to keep these invariants in place for everybody.
2060 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002061 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2062 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2063 (VM_PFNMAP|VM_MIXEDMAP));
2064 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2065 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002066
Nick Piggin423bad62008-04-28 02:13:01 -07002067 if (addr < vma->vm_start || addr >= vma->vm_end)
2068 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002069 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002070 return -EINVAL;
2071
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002072 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002073
2074 if (ret)
2075 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2076
2077 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002078}
2079EXPORT_SYMBOL(vm_insert_pfn);
2080
Nick Piggin423bad62008-04-28 02:13:01 -07002081int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2082 unsigned long pfn)
2083{
2084 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2085
2086 if (addr < vma->vm_start || addr >= vma->vm_end)
2087 return -EFAULT;
2088
2089 /*
2090 * If we don't have pte special, then we have to use the pfn_valid()
2091 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2092 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002093 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2094 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002095 */
2096 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2097 struct page *page;
2098
2099 page = pfn_to_page(pfn);
2100 return insert_page(vma, addr, page, vma->vm_page_prot);
2101 }
2102 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2103}
2104EXPORT_SYMBOL(vm_insert_mixed);
2105
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002106/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 * maps a range of physical memory into the requested pages. the old
2108 * mappings are removed. any references to nonexistent pages results
2109 * in null mappings (currently treated as "copy-on-access")
2110 */
2111static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2112 unsigned long addr, unsigned long end,
2113 unsigned long pfn, pgprot_t prot)
2114{
2115 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002116 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Hugh Dickinsc74df322005-10-29 18:16:23 -07002118 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 if (!pte)
2120 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002121 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 do {
2123 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002124 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 pfn++;
2126 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002127 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002128 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 return 0;
2130}
2131
2132static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2133 unsigned long addr, unsigned long end,
2134 unsigned long pfn, pgprot_t prot)
2135{
2136 pmd_t *pmd;
2137 unsigned long next;
2138
2139 pfn -= addr >> PAGE_SHIFT;
2140 pmd = pmd_alloc(mm, pud, addr);
2141 if (!pmd)
2142 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002143 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 do {
2145 next = pmd_addr_end(addr, end);
2146 if (remap_pte_range(mm, pmd, addr, next,
2147 pfn + (addr >> PAGE_SHIFT), prot))
2148 return -ENOMEM;
2149 } while (pmd++, addr = next, addr != end);
2150 return 0;
2151}
2152
2153static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2154 unsigned long addr, unsigned long end,
2155 unsigned long pfn, pgprot_t prot)
2156{
2157 pud_t *pud;
2158 unsigned long next;
2159
2160 pfn -= addr >> PAGE_SHIFT;
2161 pud = pud_alloc(mm, pgd, addr);
2162 if (!pud)
2163 return -ENOMEM;
2164 do {
2165 next = pud_addr_end(addr, end);
2166 if (remap_pmd_range(mm, pud, addr, next,
2167 pfn + (addr >> PAGE_SHIFT), prot))
2168 return -ENOMEM;
2169 } while (pud++, addr = next, addr != end);
2170 return 0;
2171}
2172
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002173/**
2174 * remap_pfn_range - remap kernel memory to userspace
2175 * @vma: user vma to map to
2176 * @addr: target user address to start at
2177 * @pfn: physical address of kernel memory
2178 * @size: size of map area
2179 * @prot: page protection flags for this mapping
2180 *
2181 * Note: this is only safe if the mm semaphore is held when called.
2182 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2184 unsigned long pfn, unsigned long size, pgprot_t prot)
2185{
2186 pgd_t *pgd;
2187 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002188 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 struct mm_struct *mm = vma->vm_mm;
2190 int err;
2191
2192 /*
2193 * Physically remapped pages are special. Tell the
2194 * rest of the world about it:
2195 * VM_IO tells people not to look at these pages
2196 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08002197 * VM_RESERVED is specified all over the place, because
2198 * in 2.4 it kept swapout's vma scan off this vma; but
2199 * in 2.6 the LRU scan won't even find its pages, so this
2200 * flag means no more than count its pages in reserved_vm,
2201 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08002202 * VM_PFNMAP tells the core MM that the base pages are just
2203 * raw PFN mappings, and do not have a "struct page" associated
2204 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002205 *
2206 * There's a horrible special case to handle copy-on-write
2207 * behaviour that some programs depend on. We mark the "original"
2208 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002210 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002211 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002212 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002213 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002214 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002215
Linus Torvalds6aab3412005-11-28 14:34:23 -08002216 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002218 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002219 if (err) {
2220 /*
2221 * To indicate that track_pfn related cleanup is not
2222 * needed from higher level routine calling unmap_vmas
2223 */
2224 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002225 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002226 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002227 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002228
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 BUG_ON(addr >= end);
2230 pfn -= addr >> PAGE_SHIFT;
2231 pgd = pgd_offset(mm, addr);
2232 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 do {
2234 next = pgd_addr_end(addr, end);
2235 err = remap_pud_range(mm, pgd, addr, next,
2236 pfn + (addr >> PAGE_SHIFT), prot);
2237 if (err)
2238 break;
2239 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002240
2241 if (err)
2242 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2243
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 return err;
2245}
2246EXPORT_SYMBOL(remap_pfn_range);
2247
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002248static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2249 unsigned long addr, unsigned long end,
2250 pte_fn_t fn, void *data)
2251{
2252 pte_t *pte;
2253 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002254 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002255 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002256
2257 pte = (mm == &init_mm) ?
2258 pte_alloc_kernel(pmd, addr) :
2259 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2260 if (!pte)
2261 return -ENOMEM;
2262
2263 BUG_ON(pmd_huge(*pmd));
2264
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002265 arch_enter_lazy_mmu_mode();
2266
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002267 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002268
2269 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002270 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002271 if (err)
2272 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002273 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002274
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002275 arch_leave_lazy_mmu_mode();
2276
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002277 if (mm != &init_mm)
2278 pte_unmap_unlock(pte-1, ptl);
2279 return err;
2280}
2281
2282static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2283 unsigned long addr, unsigned long end,
2284 pte_fn_t fn, void *data)
2285{
2286 pmd_t *pmd;
2287 unsigned long next;
2288 int err;
2289
Andi Kleenceb86872008-07-23 21:27:50 -07002290 BUG_ON(pud_huge(*pud));
2291
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002292 pmd = pmd_alloc(mm, pud, addr);
2293 if (!pmd)
2294 return -ENOMEM;
2295 do {
2296 next = pmd_addr_end(addr, end);
2297 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2298 if (err)
2299 break;
2300 } while (pmd++, addr = next, addr != end);
2301 return err;
2302}
2303
2304static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2305 unsigned long addr, unsigned long end,
2306 pte_fn_t fn, void *data)
2307{
2308 pud_t *pud;
2309 unsigned long next;
2310 int err;
2311
2312 pud = pud_alloc(mm, pgd, addr);
2313 if (!pud)
2314 return -ENOMEM;
2315 do {
2316 next = pud_addr_end(addr, end);
2317 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2318 if (err)
2319 break;
2320 } while (pud++, addr = next, addr != end);
2321 return err;
2322}
2323
2324/*
2325 * Scan a region of virtual memory, filling in page tables as necessary
2326 * and calling a provided function on each leaf page table.
2327 */
2328int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2329 unsigned long size, pte_fn_t fn, void *data)
2330{
2331 pgd_t *pgd;
2332 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002333 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002334 int err;
2335
2336 BUG_ON(addr >= end);
2337 pgd = pgd_offset(mm, addr);
2338 do {
2339 next = pgd_addr_end(addr, end);
2340 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2341 if (err)
2342 break;
2343 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002344
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002345 return err;
2346}
2347EXPORT_SYMBOL_GPL(apply_to_page_range);
2348
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002350 * handle_pte_fault chooses page fault handler according to an entry
2351 * which was read non-atomically. Before making any commitment, on
2352 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002353 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002354 * must check under lock before unmapping the pte and proceeding
2355 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002356 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002357 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002358static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002359 pte_t *page_table, pte_t orig_pte)
2360{
2361 int same = 1;
2362#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2363 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002364 spinlock_t *ptl = pte_lockptr(mm, pmd);
2365 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002366 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002367 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002368 }
2369#endif
2370 pte_unmap(page_table);
2371 return same;
2372}
2373
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002374static 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 -08002375{
2376 /*
2377 * If the source page was a PFN mapping, we don't have
2378 * a "struct page" for it. We do a best-effort copy by
2379 * just copying from the original user address. If that
2380 * fails, we just zero-fill it. Live with it.
2381 */
2382 if (unlikely(!src)) {
2383 void *kaddr = kmap_atomic(dst, KM_USER0);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002384 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2385
2386 /*
2387 * This really shouldn't fail, because the page is there
2388 * in the page tables. But it might just be unreadable,
2389 * in which case we just give up and fill the result with
2390 * zeroes.
2391 */
2392 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002393 clear_page(kaddr);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002394 kunmap_atomic(kaddr, KM_USER0);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002395 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002396 } else
2397 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002398}
2399
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 * This routine handles present pages, when users try to write
2402 * to a shared page. It is done by copying the page to a new address
2403 * and decrementing the shared-page counter for the old page.
2404 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 * Note that this routine assumes that the protection checks have been
2406 * done by the caller (the low-level page fault routine in most cases).
2407 * Thus we can safely just mark it writable once we've done any necessary
2408 * COW.
2409 *
2410 * We also mark the page dirty at this point even though the page will
2411 * change only once the write actually happens. This avoids a few races,
2412 * and potentially makes it more efficient.
2413 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002414 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2415 * but allow concurrent faults), with pte both mapped and locked.
2416 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002418static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2419 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002420 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002421 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002423 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002425 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002426 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002427 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428
Linus Torvalds6aab3412005-11-28 14:34:23 -08002429 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002430 if (!old_page) {
2431 /*
2432 * VM_MIXEDMAP !pfn_valid() case
2433 *
2434 * We should not cow pages in a shared writeable mapping.
2435 * Just mark the pages writable as we can't do any dirty
2436 * accounting on raw pfn maps.
2437 */
2438 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2439 (VM_WRITE|VM_SHARED))
2440 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002441 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002442 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002444 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002445 * Take out anonymous pages first, anonymous shared vmas are
2446 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002447 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002448 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002449 if (!trylock_page(old_page)) {
2450 page_cache_get(old_page);
2451 pte_unmap_unlock(page_table, ptl);
2452 lock_page(old_page);
2453 page_table = pte_offset_map_lock(mm, pmd, address,
2454 &ptl);
2455 if (!pte_same(*page_table, orig_pte)) {
2456 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002457 goto unlock;
2458 }
2459 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002460 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002461 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002462 /*
2463 * The page is all ours. Move it to our anon_vma so
2464 * the rmap code will not search our parent or siblings.
2465 * Protected against the rmap code by the page lock.
2466 */
2467 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002468 unlock_page(old_page);
2469 goto reuse;
2470 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002471 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002472 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002473 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002474 /*
2475 * Only catch write-faults on shared writable pages,
2476 * read-only shared pages can get COWed by
2477 * get_user_pages(.write=1, .force=1).
2478 */
David Howells9637a5e2006-06-23 02:03:43 -07002479 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002480 struct vm_fault vmf;
2481 int tmp;
2482
2483 vmf.virtual_address = (void __user *)(address &
2484 PAGE_MASK);
2485 vmf.pgoff = old_page->index;
2486 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2487 vmf.page = old_page;
2488
David Howells9637a5e2006-06-23 02:03:43 -07002489 /*
2490 * Notify the address space that the page is about to
2491 * become writable so that it can prohibit this or wait
2492 * for the page to get into an appropriate state.
2493 *
2494 * We do this without the lock held, so that it can
2495 * sleep if it needs to.
2496 */
2497 page_cache_get(old_page);
2498 pte_unmap_unlock(page_table, ptl);
2499
Nick Pigginc2ec1752009-03-31 15:23:21 -07002500 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2501 if (unlikely(tmp &
2502 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2503 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002504 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002505 }
Nick Pigginb827e492009-04-30 15:08:16 -07002506 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2507 lock_page(old_page);
2508 if (!old_page->mapping) {
2509 ret = 0; /* retry the fault */
2510 unlock_page(old_page);
2511 goto unwritable_page;
2512 }
2513 } else
2514 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002515
David Howells9637a5e2006-06-23 02:03:43 -07002516 /*
2517 * Since we dropped the lock we need to revalidate
2518 * the PTE as someone else may have changed it. If
2519 * they did, we just return, as we can count on the
2520 * MMU to tell us if they didn't also make it writable.
2521 */
2522 page_table = pte_offset_map_lock(mm, pmd, address,
2523 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002524 if (!pte_same(*page_table, orig_pte)) {
2525 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002526 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002527 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002528
2529 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002531 dirty_page = old_page;
2532 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002533
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002534reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002535 flush_cache_page(vma, address, pte_pfn(orig_pte));
2536 entry = pte_mkyoung(orig_pte);
2537 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002538 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002539 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002540 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002541 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002542
2543 if (!dirty_page)
2544 return ret;
2545
2546 /*
2547 * Yes, Virginia, this is actually required to prevent a race
2548 * with clear_page_dirty_for_io() from clearing the page dirty
2549 * bit after it clear all dirty ptes, but before a racing
2550 * do_wp_page installs a dirty pte.
2551 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002552 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002553 */
2554 if (!page_mkwrite) {
2555 wait_on_page_locked(dirty_page);
2556 set_page_dirty_balance(dirty_page, page_mkwrite);
2557 }
2558 put_page(dirty_page);
2559 if (page_mkwrite) {
2560 struct address_space *mapping = dirty_page->mapping;
2561
2562 set_page_dirty(dirty_page);
2563 unlock_page(dirty_page);
2564 page_cache_release(dirty_page);
2565 if (mapping) {
2566 /*
2567 * Some device drivers do not set page.mapping
2568 * but still dirty their pages
2569 */
2570 balance_dirty_pages_ratelimited(mapping);
2571 }
2572 }
2573
2574 /* file_update_time outside page_lock */
2575 if (vma->vm_file)
2576 file_update_time(vma->vm_file);
2577
2578 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
2581 /*
2582 * Ok, we need to copy. Oh, well..
2583 */
Nick Pigginb5810032005-10-29 18:16:12 -07002584 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002585gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002586 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
2588 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002589 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002590
Hugh Dickins62eede62009-09-21 17:03:34 -07002591 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002592 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2593 if (!new_page)
2594 goto oom;
2595 } else {
2596 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2597 if (!new_page)
2598 goto oom;
2599 cow_user_page(new_page, old_page, address, vma);
2600 }
2601 __SetPageUptodate(new_page);
2602
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002603 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002604 goto oom_free_new;
2605
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 /*
2607 * Re-check the pte - we dropped the lock
2608 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002609 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002610 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002611 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002612 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002613 dec_mm_counter_fast(mm, MM_FILEPAGES);
2614 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002615 }
2616 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002617 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002618 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002619 entry = mk_pte(new_page, vma->vm_page_prot);
2620 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002621 /*
2622 * Clear the pte entry and flush it first, before updating the
2623 * pte with the new entry. This will avoid a race condition
2624 * seen in the presence of one thread doing SMC and another
2625 * thread doing COW.
2626 */
Izik Eidus828502d2009-09-21 17:01:51 -07002627 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002628 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002629 /*
2630 * We call the notify macro here because, when using secondary
2631 * mmu page tables (such as kvm shadow page tables), we want the
2632 * new page to be mapped directly into the secondary page table.
2633 */
2634 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002635 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002636 if (old_page) {
2637 /*
2638 * Only after switching the pte to the new page may
2639 * we remove the mapcount here. Otherwise another
2640 * process may come and find the rmap count decremented
2641 * before the pte is switched to the new page, and
2642 * "reuse" the old page writing into it while our pte
2643 * here still points into it and can be read by other
2644 * threads.
2645 *
2646 * The critical issue is to order this
2647 * page_remove_rmap with the ptp_clear_flush above.
2648 * Those stores are ordered by (if nothing else,)
2649 * the barrier present in the atomic_add_negative
2650 * in page_remove_rmap.
2651 *
2652 * Then the TLB flush in ptep_clear_flush ensures that
2653 * no process can access the old page before the
2654 * decremented mapcount is visible. And the old page
2655 * cannot be reused until after the decremented
2656 * mapcount is visible. So transitively, TLBs to
2657 * old page will be flushed before it can be reused.
2658 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002659 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002660 }
2661
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 /* Free the old page.. */
2663 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002664 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002665 } else
2666 mem_cgroup_uncharge_page(new_page);
2667
Hugh Dickins920fc352005-11-21 21:32:17 -08002668 if (new_page)
2669 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002670unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002671 pte_unmap_unlock(page_table, ptl);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002672 if (old_page) {
2673 /*
2674 * Don't let another task, with possibly unlocked vma,
2675 * keep the mlocked page.
2676 */
2677 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2678 lock_page(old_page); /* LRU manipulation */
2679 munlock_vma_page(old_page);
2680 unlock_page(old_page);
2681 }
2682 page_cache_release(old_page);
2683 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002684 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002685oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002686 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002687oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002688 if (old_page) {
2689 if (page_mkwrite) {
2690 unlock_page(old_page);
2691 page_cache_release(old_page);
2692 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002693 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002696
2697unwritable_page:
2698 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002699 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700}
2701
Peter Zijlstra97a89412011-05-24 17:12:04 -07002702static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 unsigned long start_addr, unsigned long end_addr,
2704 struct zap_details *details)
2705{
Peter Zijlstra97a89412011-05-24 17:12:04 -07002706 zap_page_range(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707}
2708
2709static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2710 struct zap_details *details)
2711{
2712 struct vm_area_struct *vma;
2713 struct prio_tree_iter iter;
2714 pgoff_t vba, vea, zba, zea;
2715
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 vma_prio_tree_foreach(vma, &iter, root,
2717 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718
2719 vba = vma->vm_pgoff;
2720 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2721 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2722 zba = details->first_index;
2723 if (zba < vba)
2724 zba = vba;
2725 zea = details->last_index;
2726 if (zea > vea)
2727 zea = vea;
2728
Peter Zijlstra97a89412011-05-24 17:12:04 -07002729 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2731 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002732 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 }
2734}
2735
2736static inline void unmap_mapping_range_list(struct list_head *head,
2737 struct zap_details *details)
2738{
2739 struct vm_area_struct *vma;
2740
2741 /*
2742 * In nonlinear VMAs there is no correspondence between virtual address
2743 * offset and file offset. So we must perform an exhaustive search
2744 * across *all* the pages in each nonlinear VMA, not just the pages
2745 * whose virtual address lies outside the file truncation point.
2746 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 list_for_each_entry(vma, head, shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002749 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 }
2751}
2752
2753/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002754 * 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 -07002755 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 * @holebegin: byte in first page to unmap, relative to the start of
2757 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002758 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 * must keep the partial page. In contrast, we must get rid of
2760 * partial pages.
2761 * @holelen: size of prospective hole in bytes. This will be rounded
2762 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2763 * end of the file.
2764 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2765 * but 0 when invalidating pagecache, don't throw away private data.
2766 */
2767void unmap_mapping_range(struct address_space *mapping,
2768 loff_t const holebegin, loff_t const holelen, int even_cows)
2769{
2770 struct zap_details details;
2771 pgoff_t hba = holebegin >> PAGE_SHIFT;
2772 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2773
2774 /* Check for overflow. */
2775 if (sizeof(holelen) > sizeof(hlen)) {
2776 long long holeend =
2777 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2778 if (holeend & ~(long long)ULONG_MAX)
2779 hlen = ULONG_MAX - hba + 1;
2780 }
2781
2782 details.check_mapping = even_cows? NULL: mapping;
2783 details.nonlinear_vma = NULL;
2784 details.first_index = hba;
2785 details.last_index = hba + hlen - 1;
2786 if (details.last_index < details.first_index)
2787 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788
Peter Zijlstra97a89412011-05-24 17:12:04 -07002789
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002790 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2792 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2793 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2794 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002795 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796}
2797EXPORT_SYMBOL(unmap_mapping_range);
2798
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002799int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2800{
2801 struct address_space *mapping = inode->i_mapping;
2802
2803 /*
2804 * If the underlying filesystem is not going to provide
2805 * a way to truncate a range of blocks (punch a hole) -
2806 * we should return failure right now.
2807 */
Al Viroacfa4382008-12-04 10:06:33 -05002808 if (!inode->i_op->truncate_range)
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002809 return -ENOSYS;
2810
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002811 mutex_lock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002812 down_write(&inode->i_alloc_sem);
2813 unmap_mapping_range(mapping, offset, (end - offset), 1);
2814 truncate_inode_pages_range(mapping, offset, end);
Nick Piggind00806b2007-07-19 01:46:57 -07002815 unmap_mapping_range(mapping, offset, (end - offset), 1);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002816 inode->i_op->truncate_range(inode, offset, end);
2817 up_write(&inode->i_alloc_sem);
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002818 mutex_unlock(&inode->i_mutex);
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002819
2820 return 0;
2821}
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -08002822
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002824 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2825 * but allow concurrent faults), and pte mapped but not yet locked.
2826 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002828static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2829 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002830 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002832 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002833 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002834 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002836 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002837 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002838 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002839 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002841 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002842 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002843
2844 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002845 if (unlikely(non_swap_entry(entry))) {
2846 if (is_migration_entry(entry)) {
2847 migration_entry_wait(mm, pmd, address);
2848 } else if (is_hwpoison_entry(entry)) {
2849 ret = VM_FAULT_HWPOISON;
2850 } else {
2851 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002852 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002853 }
Christoph Lameter06972122006-06-23 02:03:35 -07002854 goto out;
2855 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002856 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 page = lookup_swap_cache(entry);
2858 if (!page) {
Hugh Dickinsa5c9b692009-06-23 12:36:58 -07002859 grab_swap_token(mm); /* Contend for token _before_ read-in */
Hugh Dickins02098fe2008-02-04 22:28:42 -08002860 page = swapin_readahead(entry,
2861 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 if (!page) {
2863 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002864 * Back out if somebody else faulted in this pte
2865 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002867 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 if (likely(pte_same(*page_table, orig_pte)))
2869 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002870 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002871 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 }
2873
2874 /* Had to read the page from swap area: Major fault */
2875 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002876 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002877 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002878 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002879 /*
2880 * hwpoisoned dirty swapcache pages are kept for killing
2881 * owner processes (which may be unknown at hwpoison time)
2882 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002883 ret = VM_FAULT_HWPOISON;
2884 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002885 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 }
2887
Michel Lespinassed065bd82010-10-26 14:21:57 -07002888 locked = lock_page_or_retry(page, mm, flags);
Balbir Singh20a10222007-11-14 17:00:33 -08002889 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002890 if (!locked) {
2891 ret |= VM_FAULT_RETRY;
2892 goto out_release;
2893 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002895 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002896 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2897 * release the swapcache from under us. The page pin, and pte_same
2898 * test below, are not enough to exclude that. Even if it is still
2899 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002900 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002901 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002902 goto out_page;
2903
2904 if (ksm_might_need_to_copy(page, vma, address)) {
2905 swapcache = page;
2906 page = ksm_does_need_to_copy(page, vma, address);
2907
2908 if (unlikely(!page)) {
2909 ret = VM_FAULT_OOM;
2910 page = swapcache;
2911 swapcache = NULL;
2912 goto out_page;
2913 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002914 }
2915
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002916 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002917 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002918 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002919 }
2920
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002922 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002924 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002925 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002926 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002927
2928 if (unlikely(!PageUptodate(page))) {
2929 ret = VM_FAULT_SIGBUS;
2930 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 }
2932
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002933 /*
2934 * The page isn't present yet, go ahead with the fault.
2935 *
2936 * Be careful about the sequence of operations here.
2937 * To get its accounting right, reuse_swap_page() must be called
2938 * while the page is counted on swap but not yet in mapcount i.e.
2939 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2940 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002941 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2942 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2943 * in page->private. In this case, a record in swap_cgroup is silently
2944 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002945 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002947 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002948 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002950 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002952 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002953 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002954 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 flush_icache_page(vma, page);
2957 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002958 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002959 /* It's better to call commit-charge after rmap is established */
2960 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002962 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002963 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002964 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002965 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002966 if (swapcache) {
2967 /*
2968 * Hold the lock to avoid the swap entry to be reused
2969 * until we take the PT lock for the pte_same() check
2970 * (to avoid false positives from pte_same). For
2971 * further safety release the lock after the swap_free
2972 * so that the swap count won't change under a
2973 * parallel locked swapcache.
2974 */
2975 unlock_page(swapcache);
2976 page_cache_release(swapcache);
2977 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002978
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002979 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002980 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2981 if (ret & VM_FAULT_ERROR)
2982 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 goto out;
2984 }
2985
2986 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002987 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002988unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002989 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990out:
2991 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002992out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08002993 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002994 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002995out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002996 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002997out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002998 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002999 if (swapcache) {
3000 unlock_page(swapcache);
3001 page_cache_release(swapcache);
3002 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003003 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004}
3005
3006/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003007 * This is like a special single-page "expand_{down|up}wards()",
3008 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003009 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003010 */
3011static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3012{
3013 address &= PAGE_MASK;
3014 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003015 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003016
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003017 /*
3018 * Is there a mapping abutting this one below?
3019 *
3020 * That's only ok if it's the same stack mapping
3021 * that has gotten split..
3022 */
3023 if (prev && prev->vm_end == address)
3024 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3025
Michal Hockod05f3162011-05-24 17:11:44 -07003026 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003027 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003028 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3029 struct vm_area_struct *next = vma->vm_next;
3030
3031 /* As VM_GROWSDOWN but s/below/above/ */
3032 if (next && next->vm_start == address + PAGE_SIZE)
3033 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3034
3035 expand_upwards(vma, address + PAGE_SIZE);
3036 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003037 return 0;
3038}
3039
3040/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003041 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3042 * but allow concurrent faults), and pte mapped but not yet locked.
3043 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003045static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3046 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003047 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003049 struct page *page;
3050 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052
Linus Torvalds11ac5522010-08-14 11:44:56 -07003053 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003054
Linus Torvalds11ac5522010-08-14 11:44:56 -07003055 /* Check if we need to add a guard page to the stack */
3056 if (check_stack_guard_page(vma, address) < 0)
3057 return VM_FAULT_SIGBUS;
3058
3059 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003060 if (!(flags & FAULT_FLAG_WRITE)) {
3061 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3062 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003063 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003064 if (!pte_none(*page_table))
3065 goto unlock;
3066 goto setpte;
3067 }
3068
Nick Piggin557ed1f2007-10-16 01:24:40 -07003069 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003070 if (unlikely(anon_vma_prepare(vma)))
3071 goto oom;
3072 page = alloc_zeroed_user_highpage_movable(vma, address);
3073 if (!page)
3074 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003075 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003077 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003078 goto oom_free_page;
3079
Nick Piggin557ed1f2007-10-16 01:24:40 -07003080 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003081 if (vma->vm_flags & VM_WRITE)
3082 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003083
Nick Piggin557ed1f2007-10-16 01:24:40 -07003084 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003085 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003086 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003087
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003088 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003089 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003090setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003091 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092
3093 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003094 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003095unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003096 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003097 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003098release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003099 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003100 page_cache_release(page);
3101 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003102oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003103 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003104oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 return VM_FAULT_OOM;
3106}
3107
3108/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003109 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003111 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3112 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 *
3114 * As this is called only for pages that do not currently exist, we
3115 * do not need to flush old virtual caches or the TLB.
3116 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003117 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003118 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003119 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003121static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003122 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003123 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003125 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003126 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003127 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 int anon = 0;
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003130 int charged = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003131 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003132 struct vm_fault vmf;
3133 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003134 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003135
Nick Piggind0217ac2007-07-19 01:47:03 -07003136 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3137 vmf.pgoff = pgoff;
3138 vmf.flags = flags;
3139 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003141 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003142 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3143 VM_FAULT_RETRY)))
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003144 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145
Andi Kleena3b947e2009-09-16 11:50:08 +02003146 if (unlikely(PageHWPoison(vmf.page))) {
3147 if (ret & VM_FAULT_LOCKED)
3148 unlock_page(vmf.page);
3149 return VM_FAULT_HWPOISON;
3150 }
3151
Nick Piggind00806b2007-07-19 01:46:57 -07003152 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003153 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003154 * locked.
3155 */
Nick Piggin83c54072007-07-19 01:47:05 -07003156 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003157 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003158 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003159 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003160
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 /*
3162 * Should we do an early C-O-W break?
3163 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003164 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003165 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003166 if (!(vma->vm_flags & VM_SHARED)) {
Nick Piggin54cb8822007-07-19 01:46:59 -07003167 anon = 1;
Nick Piggind00806b2007-07-19 01:46:57 -07003168 if (unlikely(anon_vma_prepare(vma))) {
Nick Piggind0217ac2007-07-19 01:47:03 -07003169 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07003170 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07003171 }
Nick Piggin83c54072007-07-19 01:47:05 -07003172 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3173 vma, address);
Nick Piggind00806b2007-07-19 01:46:57 -07003174 if (!page) {
Nick Piggind0217ac2007-07-19 01:47:03 -07003175 ret = VM_FAULT_OOM;
Nick Piggin54cb8822007-07-19 01:46:59 -07003176 goto out;
Nick Piggind00806b2007-07-19 01:46:57 -07003177 }
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003178 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003179 ret = VM_FAULT_OOM;
3180 page_cache_release(page);
3181 goto out;
3182 }
3183 charged = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003184 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003185 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003186 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003187 /*
3188 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003189 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003190 * to become writable
3191 */
Mark Fasheh69676142007-07-19 01:47:00 -07003192 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003193 int tmp;
3194
Mark Fasheh69676142007-07-19 01:47:00 -07003195 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003196 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003197 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3198 if (unlikely(tmp &
3199 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3200 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003201 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003202 }
Nick Pigginb827e492009-04-30 15:08:16 -07003203 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3204 lock_page(page);
3205 if (!page->mapping) {
3206 ret = 0; /* retry the fault */
3207 unlock_page(page);
3208 goto unwritable_page;
3209 }
3210 } else
3211 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003212 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003213 }
3214 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003215
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 }
3217
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003218 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219
3220 /*
3221 * This silly early PAGE_DIRTY setting removes a race
3222 * due to the bad i386 page protection. But it's valid
3223 * for other architectures too.
3224 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003225 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 * an exclusive copy of the page, or this is a shared mapping,
3227 * so we can make it writable and dirty to avoid having to
3228 * handle that later.
3229 */
3230 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003231 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003232 flush_icache_page(vma, page);
3233 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003234 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003237 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003238 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003239 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003240 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003241 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003242 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003243 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003244 get_page(dirty_page);
3245 }
Hugh Dickins42946212005-10-29 18:16:05 -07003246 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003247 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003248
3249 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003250 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 } else {
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07003252 if (charged)
3253 mem_cgroup_uncharge_page(page);
Nick Piggind00806b2007-07-19 01:46:57 -07003254 if (anon)
3255 page_cache_release(page);
3256 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003257 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258 }
3259
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003260 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003261
3262out:
Nick Pigginb827e492009-04-30 15:08:16 -07003263 if (dirty_page) {
3264 struct address_space *mapping = page->mapping;
3265
3266 if (set_page_dirty(dirty_page))
3267 page_mkwrite = 1;
3268 unlock_page(dirty_page);
3269 put_page(dirty_page);
3270 if (page_mkwrite && mapping) {
3271 /*
3272 * Some device drivers do not set page.mapping but still
3273 * dirty their pages
3274 */
3275 balance_dirty_pages_ratelimited(mapping);
3276 }
3277
3278 /* file_update_time outside page_lock */
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003279 if (vma->vm_file)
3280 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003281 } else {
3282 unlock_page(vmf.page);
3283 if (anon)
3284 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003285 }
Nick Piggind00806b2007-07-19 01:46:57 -07003286
Nick Piggin83c54072007-07-19 01:47:05 -07003287 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003288
3289unwritable_page:
3290 page_cache_release(page);
3291 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003292}
Nick Piggind00806b2007-07-19 01:46:57 -07003293
Nick Piggin54cb8822007-07-19 01:46:59 -07003294static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3295 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003296 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003297{
3298 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003299 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003300
Hugh Dickins16abfa02007-10-04 16:56:06 +01003301 pte_unmap(page_table);
3302 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003303}
3304
Jes Sorensenf4b81802006-09-27 01:50:10 -07003305/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 * Fault of a previously existing named mapping. Repopulate the pte
3307 * from the encoded file_pte if possible. This enables swappable
3308 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003309 *
3310 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3311 * but allow concurrent faults), and pte mapped but not yet locked.
3312 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003314static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003315 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003316 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317{
Hugh Dickins65500d22005-10-29 18:15:59 -07003318 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003320 flags |= FAULT_FLAG_NONLINEAR;
3321
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003322 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003323 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
Hugh Dickins2509ef22009-01-06 14:40:10 -08003325 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003326 /*
3327 * Page table corrupted: show pte and kill process.
3328 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003329 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003330 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003331 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003332
3333 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003334 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335}
3336
3337/*
3338 * These routines also need to handle stuff like marking pages dirty
3339 * and/or accessed for architectures that don't do it in hardware (most
3340 * RISC architectures). The early dirtying is also good on the i386.
3341 *
3342 * There is also a hook called "update_mmu_cache()" that architectures
3343 * with external mmu caches can use to update those (ie the Sparc or
3344 * PowerPC hashed page tables that act as extended TLBs).
3345 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003346 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3347 * but allow concurrent faults), and pte mapped but not yet locked.
3348 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003350int handle_pte_fault(struct mm_struct *mm,
3351 struct vm_area_struct *vma, unsigned long address,
3352 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353{
3354 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003355 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003357 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003359 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003360 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003361 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003362 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003363 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003364 }
3365 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003366 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003367 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003369 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003370 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003371 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003372 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 }
3374
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003375 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003376 spin_lock(ptl);
3377 if (unlikely(!pte_same(*pte, entry)))
3378 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003379 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003381 return do_wp_page(mm, vma, address,
3382 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 entry = pte_mkdirty(entry);
3384 }
3385 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003386 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003387 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003388 } else {
3389 /*
3390 * This is needed only for protection faults but the arch code
3391 * is not yet telling us if this is a protection fault or not.
3392 * This still avoids useless tlb flushes for .text page faults
3393 * with threads.
3394 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003395 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003396 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003397 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003398unlock:
3399 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003400 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401}
3402
3403/*
3404 * By the time we get here, we already hold the mm semaphore
3405 */
Nick Piggin83c54072007-07-19 01:47:05 -07003406int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003407 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408{
3409 pgd_t *pgd;
3410 pud_t *pud;
3411 pmd_t *pmd;
3412 pte_t *pte;
3413
3414 __set_current_state(TASK_RUNNING);
3415
Christoph Lameterf8891e52006-06-30 01:55:45 -07003416 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003417 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003419 /* do counter updates before entering really critical section. */
3420 check_sync_rss_stat(current);
3421
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003422 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003423 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 pud = pud_alloc(mm, pgd, address);
3427 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003428 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429 pmd = pmd_alloc(mm, pud, address);
3430 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003431 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003432 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3433 if (!vma->vm_ops)
3434 return do_huge_pmd_anonymous_page(mm, vma, address,
3435 pmd, flags);
3436 } else {
3437 pmd_t orig_pmd = *pmd;
3438 barrier();
3439 if (pmd_trans_huge(orig_pmd)) {
3440 if (flags & FAULT_FLAG_WRITE &&
3441 !pmd_write(orig_pmd) &&
3442 !pmd_trans_splitting(orig_pmd))
3443 return do_huge_pmd_wp_page(mm, vma, address,
3444 pmd, orig_pmd);
3445 return 0;
3446 }
3447 }
3448
3449 /*
3450 * Use __pte_alloc instead of pte_alloc_map, because we can't
3451 * run pte_offset_map on the pmd, if an huge pmd could
3452 * materialize from under us from a different thread.
3453 */
Mel Gormancc036382011-04-27 15:26:56 -07003454 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003455 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003456 /* if an huge pmd materialized from under us just retry later */
3457 if (unlikely(pmd_trans_huge(*pmd)))
3458 return 0;
3459 /*
3460 * A regular pmd is established and it can't morph into a huge pmd
3461 * from under us anymore at this point because we hold the mmap_sem
3462 * read mode and khugepaged takes it in write mode. So now it's
3463 * safe to run pte_offset_map().
3464 */
3465 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003467 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468}
3469
3470#ifndef __PAGETABLE_PUD_FOLDED
3471/*
3472 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003473 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003475int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003477 pud_t *new = pud_alloc_one(mm, address);
3478 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003479 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480
Nick Piggin362a61a2008-05-14 06:37:36 +02003481 smp_wmb(); /* See comment in __pte_alloc */
3482
Hugh Dickins872fec12005-10-29 18:16:21 -07003483 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003484 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003485 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003486 else
3487 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003488 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003489 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490}
3491#endif /* __PAGETABLE_PUD_FOLDED */
3492
3493#ifndef __PAGETABLE_PMD_FOLDED
3494/*
3495 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003496 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003498int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003500 pmd_t *new = pmd_alloc_one(mm, address);
3501 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003502 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
Nick Piggin362a61a2008-05-14 06:37:36 +02003504 smp_wmb(); /* See comment in __pte_alloc */
3505
Hugh Dickins872fec12005-10-29 18:16:21 -07003506 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003508 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003509 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003510 else
3511 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003513 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003514 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003515 else
3516 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003518 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003519 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520}
3521#endif /* __PAGETABLE_PMD_FOLDED */
3522
3523int make_pages_present(unsigned long addr, unsigned long end)
3524{
3525 int ret, len, write;
3526 struct vm_area_struct * vma;
3527
3528 vma = find_vma(current->mm, addr);
3529 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003530 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003531 /*
3532 * We want to touch writable mappings with a write fault in order
3533 * to break COW, except for shared mappings because these don't COW
3534 * and we would not want to dirty them for nothing.
3535 */
3536 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003537 BUG_ON(addr >= end);
3538 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003539 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 ret = get_user_pages(current, current->mm, addr,
3541 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003542 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003544 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545}
3546
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547#if !defined(__HAVE_ARCH_GATE_AREA)
3548
3549#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003550static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551
3552static int __init gate_vma_init(void)
3553{
3554 gate_vma.vm_mm = NULL;
3555 gate_vma.vm_start = FIXADDR_USER_START;
3556 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003557 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3558 gate_vma.vm_page_prot = __P101;
Roland McGrathf47aef52007-01-26 00:56:49 -08003559 /*
3560 * Make sure the vDSO gets into every core dump.
3561 * Dumping its contents makes post-mortem fully interpretable later
3562 * without matching up the same kernel and hardware config to see
3563 * what PC values meant.
3564 */
3565 gate_vma.vm_flags |= VM_ALWAYSDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 return 0;
3567}
3568__initcall(gate_vma_init);
3569#endif
3570
Stephen Wilson31db58b2011-03-13 15:49:15 -04003571struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572{
3573#ifdef AT_SYSINFO_EHDR
3574 return &gate_vma;
3575#else
3576 return NULL;
3577#endif
3578}
3579
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003580int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581{
3582#ifdef AT_SYSINFO_EHDR
3583 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3584 return 1;
3585#endif
3586 return 0;
3587}
3588
3589#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003590
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003591static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003592 pte_t **ptepp, spinlock_t **ptlp)
3593{
3594 pgd_t *pgd;
3595 pud_t *pud;
3596 pmd_t *pmd;
3597 pte_t *ptep;
3598
3599 pgd = pgd_offset(mm, address);
3600 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3601 goto out;
3602
3603 pud = pud_offset(pgd, address);
3604 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3605 goto out;
3606
3607 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003608 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003609 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3610 goto out;
3611
3612 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3613 if (pmd_huge(*pmd))
3614 goto out;
3615
3616 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3617 if (!ptep)
3618 goto out;
3619 if (!pte_present(*ptep))
3620 goto unlock;
3621 *ptepp = ptep;
3622 return 0;
3623unlock:
3624 pte_unmap_unlock(ptep, *ptlp);
3625out:
3626 return -EINVAL;
3627}
3628
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003629static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3630 pte_t **ptepp, spinlock_t **ptlp)
3631{
3632 int res;
3633
3634 /* (void) is needed to make gcc happy */
3635 (void) __cond_lock(*ptlp,
3636 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3637 return res;
3638}
3639
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003640/**
3641 * follow_pfn - look up PFN at a user virtual address
3642 * @vma: memory mapping
3643 * @address: user virtual address
3644 * @pfn: location to store found PFN
3645 *
3646 * Only IO mappings and raw PFN mappings are allowed.
3647 *
3648 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3649 */
3650int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3651 unsigned long *pfn)
3652{
3653 int ret = -EINVAL;
3654 spinlock_t *ptl;
3655 pte_t *ptep;
3656
3657 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3658 return ret;
3659
3660 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3661 if (ret)
3662 return ret;
3663 *pfn = pte_pfn(*ptep);
3664 pte_unmap_unlock(ptep, ptl);
3665 return 0;
3666}
3667EXPORT_SYMBOL(follow_pfn);
3668
Rik van Riel28b2ee22008-07-23 21:27:05 -07003669#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003670int follow_phys(struct vm_area_struct *vma,
3671 unsigned long address, unsigned int flags,
3672 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003673{
Johannes Weiner03668a42009-06-16 15:32:34 -07003674 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003675 pte_t *ptep, pte;
3676 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003677
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003678 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3679 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003680
Johannes Weiner03668a42009-06-16 15:32:34 -07003681 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003682 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003683 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003684
Rik van Riel28b2ee22008-07-23 21:27:05 -07003685 if ((flags & FOLL_WRITE) && !pte_write(pte))
3686 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003687
3688 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003689 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003690
Johannes Weiner03668a42009-06-16 15:32:34 -07003691 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003692unlock:
3693 pte_unmap_unlock(ptep, ptl);
3694out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003695 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003696}
3697
3698int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3699 void *buf, int len, int write)
3700{
3701 resource_size_t phys_addr;
3702 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003703 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003704 int offset = addr & (PAGE_SIZE-1);
3705
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003706 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003707 return -EINVAL;
3708
3709 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3710 if (write)
3711 memcpy_toio(maddr + offset, buf, len);
3712 else
3713 memcpy_fromio(buf, maddr + offset, len);
3714 iounmap(maddr);
3715
3716 return len;
3717}
3718#endif
3719
David Howells0ec76a12006-09-27 01:50:15 -07003720/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003721 * Access another process' address space as given in mm. If non-NULL, use the
3722 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003723 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003724static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3725 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003726{
David Howells0ec76a12006-09-27 01:50:15 -07003727 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003728 void *old_buf = buf;
3729
David Howells0ec76a12006-09-27 01:50:15 -07003730 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003731 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003732 while (len) {
3733 int bytes, ret, offset;
3734 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003735 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003736
3737 ret = get_user_pages(tsk, mm, addr, 1,
3738 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003739 if (ret <= 0) {
3740 /*
3741 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3742 * we can access using slightly different code.
3743 */
3744#ifdef CONFIG_HAVE_IOREMAP_PROT
3745 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003746 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003747 break;
3748 if (vma->vm_ops && vma->vm_ops->access)
3749 ret = vma->vm_ops->access(vma, addr, buf,
3750 len, write);
3751 if (ret <= 0)
3752#endif
3753 break;
3754 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003755 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003756 bytes = len;
3757 offset = addr & (PAGE_SIZE-1);
3758 if (bytes > PAGE_SIZE-offset)
3759 bytes = PAGE_SIZE-offset;
3760
3761 maddr = kmap(page);
3762 if (write) {
3763 copy_to_user_page(vma, page, addr,
3764 maddr + offset, buf, bytes);
3765 set_page_dirty_lock(page);
3766 } else {
3767 copy_from_user_page(vma, page, addr,
3768 buf, maddr + offset, bytes);
3769 }
3770 kunmap(page);
3771 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003772 }
David Howells0ec76a12006-09-27 01:50:15 -07003773 len -= bytes;
3774 buf += bytes;
3775 addr += bytes;
3776 }
3777 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003778
3779 return buf - old_buf;
3780}
Andi Kleen03252912008-01-30 13:33:18 +01003781
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003782/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003783 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003784 * @mm: the mm_struct of the target address space
3785 * @addr: start address to access
3786 * @buf: source or destination buffer
3787 * @len: number of bytes to transfer
3788 * @write: whether the access is a write
3789 *
3790 * The caller must hold a reference on @mm.
3791 */
3792int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3793 void *buf, int len, int write)
3794{
3795 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3796}
3797
Andi Kleen03252912008-01-30 13:33:18 +01003798/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003799 * Access another process' address space.
3800 * Source/target buffer must be kernel space,
3801 * Do not walk the page table directly, use get_user_pages
3802 */
3803int access_process_vm(struct task_struct *tsk, unsigned long addr,
3804 void *buf, int len, int write)
3805{
3806 struct mm_struct *mm;
3807 int ret;
3808
3809 mm = get_task_mm(tsk);
3810 if (!mm)
3811 return 0;
3812
3813 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3814 mmput(mm);
3815
3816 return ret;
3817}
3818
Andi Kleen03252912008-01-30 13:33:18 +01003819/*
3820 * Print the name of a VMA.
3821 */
3822void print_vma_addr(char *prefix, unsigned long ip)
3823{
3824 struct mm_struct *mm = current->mm;
3825 struct vm_area_struct *vma;
3826
Ingo Molnare8bff742008-02-13 20:21:06 +01003827 /*
3828 * Do not print if we are in atomic
3829 * contexts (in exception stacks, etc.):
3830 */
3831 if (preempt_count())
3832 return;
3833
Andi Kleen03252912008-01-30 13:33:18 +01003834 down_read(&mm->mmap_sem);
3835 vma = find_vma(mm, ip);
3836 if (vma && vma->vm_file) {
3837 struct file *f = vma->vm_file;
3838 char *buf = (char *)__get_free_page(GFP_KERNEL);
3839 if (buf) {
3840 char *p, *s;
3841
Jan Blunckcf28b482008-02-14 19:38:44 -08003842 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003843 if (IS_ERR(p))
3844 p = "?";
3845 s = strrchr(p, '/');
3846 if (s)
3847 p = s+1;
3848 printk("%s%s[%lx+%lx]", prefix, p,
3849 vma->vm_start,
3850 vma->vm_end - vma->vm_start);
3851 free_page((unsigned long)buf);
3852 }
3853 }
3854 up_read(&current->mm->mmap_sem);
3855}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003856
3857#ifdef CONFIG_PROVE_LOCKING
3858void might_fault(void)
3859{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003860 /*
3861 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3862 * holding the mmap_sem, this is safe because kernel memory doesn't
3863 * get paged out, therefore we'll never actually fault, and the
3864 * below annotations will generate false positives.
3865 */
3866 if (segment_eq(get_fs(), KERNEL_DS))
3867 return;
3868
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003869 might_sleep();
3870 /*
3871 * it would be nicer only to annotate paths which are not under
3872 * pagefault_disable, however that requires a larger audit and
3873 * providing helpers like get_user_atomic.
3874 */
3875 if (!in_atomic() && current->mm)
3876 might_lock_read(&current->mm->mmap_sem);
3877}
3878EXPORT_SYMBOL(might_fault);
3879#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003880
3881#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3882static void clear_gigantic_page(struct page *page,
3883 unsigned long addr,
3884 unsigned int pages_per_huge_page)
3885{
3886 int i;
3887 struct page *p = page;
3888
3889 might_sleep();
3890 for (i = 0; i < pages_per_huge_page;
3891 i++, p = mem_map_next(p, page, i)) {
3892 cond_resched();
3893 clear_user_highpage(p, addr + i * PAGE_SIZE);
3894 }
3895}
3896void clear_huge_page(struct page *page,
3897 unsigned long addr, unsigned int pages_per_huge_page)
3898{
3899 int i;
3900
3901 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3902 clear_gigantic_page(page, addr, pages_per_huge_page);
3903 return;
3904 }
3905
3906 might_sleep();
3907 for (i = 0; i < pages_per_huge_page; i++) {
3908 cond_resched();
3909 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3910 }
3911}
3912
3913static void copy_user_gigantic_page(struct page *dst, struct page *src,
3914 unsigned long addr,
3915 struct vm_area_struct *vma,
3916 unsigned int pages_per_huge_page)
3917{
3918 int i;
3919 struct page *dst_base = dst;
3920 struct page *src_base = src;
3921
3922 for (i = 0; i < pages_per_huge_page; ) {
3923 cond_resched();
3924 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3925
3926 i++;
3927 dst = mem_map_next(dst, dst_base, i);
3928 src = mem_map_next(src, src_base, i);
3929 }
3930}
3931
3932void copy_user_huge_page(struct page *dst, struct page *src,
3933 unsigned long addr, struct vm_area_struct *vma,
3934 unsigned int pages_per_huge_page)
3935{
3936 int i;
3937
3938 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3939 copy_user_gigantic_page(dst, src, addr, vma,
3940 pages_per_huge_page);
3941 return;
3942 }
3943
3944 might_sleep();
3945 for (i = 0; i < pages_per_huge_page; i++) {
3946 cond_resched();
3947 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3948 }
3949}
3950#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */