blob: 1ce2e2a734fc2b0812845279a69cb87f48588d3e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
Hugh Dickins9a840892009-09-21 17:02:01 -070048#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/rmap.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040050#include <linux/export.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070053#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080054#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070055#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080056#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090059#include <linux/gfp.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000060#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080061#include <linux/string.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Alexey Dobriyan6952b612009-09-18 23:55:55 +040063#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064#include <asm/pgalloc.h>
65#include <asm/uaccess.h>
66#include <asm/tlb.h>
67#include <asm/tlbflush.h>
68#include <asm/pgtable.h>
69
Jan Beulich42b77722008-07-23 21:27:10 -070070#include "internal.h"
71
Peter Zijlstra75980e92013-02-22 16:34:32 -080072#ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
74#endif
75
Andy Whitcroftd41dee32005-06-23 00:07:54 -070076#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070077/* use the per-pgdat data instead for discontigmem - mbligh */
78unsigned long max_mapnr;
79struct page *mem_map;
80
81EXPORT_SYMBOL(max_mapnr);
82EXPORT_SYMBOL(mem_map);
83#endif
84
Linus Torvalds1da177e2005-04-16 15:20:36 -070085/*
86 * A number of key systems in x86 including ioremap() rely on the assumption
87 * that high_memory defines the upper bound on direct map memory, then end
88 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
90 * and ZONE_HIGHMEM.
91 */
92void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
Linus Torvalds1da177e2005-04-16 15:20:36 -070094EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
Ingo Molnar32a93232008-02-06 22:39:44 +010096/*
97 * Randomize the address space (stacks, mmaps, brk, etc.).
98 *
99 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
100 * as ancient (libc5 based) binaries can segfault. )
101 */
102int randomize_va_space __read_mostly =
103#ifdef CONFIG_COMPAT_BRK
104 1;
105#else
106 2;
107#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100108
109static int __init disable_randmaps(char *s)
110{
111 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800112 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100113}
114__setup("norandmaps", disable_randmaps);
115
Hugh Dickins62eede62009-09-21 17:03:34 -0700116unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700117unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700118
119/*
120 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
121 */
122static int __init init_zero_pfn(void)
123{
124 zero_pfn = page_to_pfn(ZERO_PAGE(0));
125 return 0;
126}
127core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100128
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800129
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800130#if defined(SPLIT_RSS_COUNTING)
131
David Rientjesea48cf72012-03-21 16:34:13 -0700132void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800133{
134 int i;
135
136 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700137 if (current->rss_stat.count[i]) {
138 add_mm_counter(mm, i, current->rss_stat.count[i]);
139 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800140 }
141 }
David Rientjes05af2e12012-03-21 16:34:13 -0700142 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800143}
144
145static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
146{
147 struct task_struct *task = current;
148
149 if (likely(task->mm == mm))
150 task->rss_stat.count[member] += val;
151 else
152 add_mm_counter(mm, member, val);
153}
154#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
155#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
156
157/* sync counter once per 64 page faults */
158#define TASK_RSS_EVENTS_THRESH (64)
159static void check_sync_rss_stat(struct task_struct *task)
160{
161 if (unlikely(task != current))
162 return;
163 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700164 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800165}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700166#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800167
168#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
169#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
170
171static void check_sync_rss_stat(struct task_struct *task)
172{
173}
174
Peter Zijlstra9547d012011-05-24 17:12:14 -0700175#endif /* SPLIT_RSS_COUNTING */
176
177#ifdef HAVE_GENERIC_MMU_GATHER
178
179static int tlb_next_batch(struct mmu_gather *tlb)
180{
181 struct mmu_gather_batch *batch;
182
183 batch = tlb->active;
184 if (batch->next) {
185 tlb->active = batch->next;
186 return 1;
187 }
188
Michal Hocko53a59fc2013-01-04 15:35:12 -0800189 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
190 return 0;
191
Peter Zijlstra9547d012011-05-24 17:12:14 -0700192 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
193 if (!batch)
194 return 0;
195
Michal Hocko53a59fc2013-01-04 15:35:12 -0800196 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700197 batch->next = NULL;
198 batch->nr = 0;
199 batch->max = MAX_GATHER_BATCH;
200
201 tlb->active->next = batch;
202 tlb->active = batch;
203
204 return 1;
205}
206
207/* tlb_gather_mmu
208 * Called to initialize an (on-stack) mmu_gather structure for page-table
209 * tear-down from @mm. The @fullmm argument is used when @mm is without
210 * users and we're going to destroy the full address space (exit/execve).
211 */
212void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
213{
214 tlb->mm = mm;
215
216 tlb->fullmm = fullmm;
Dave Hansen1de14c32013-04-12 16:23:54 -0700217 tlb->need_flush_all = 0;
Alex Shi597e1c32012-06-28 09:02:21 +0800218 tlb->start = -1UL;
219 tlb->end = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700220 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700221 tlb->local.next = NULL;
222 tlb->local.nr = 0;
223 tlb->local.max = ARRAY_SIZE(tlb->__pages);
224 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800225 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700226
227#ifdef CONFIG_HAVE_RCU_TABLE_FREE
228 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800229#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700230}
231
232void tlb_flush_mmu(struct mmu_gather *tlb)
233{
234 struct mmu_gather_batch *batch;
235
236 if (!tlb->need_flush)
237 return;
238 tlb->need_flush = 0;
239 tlb_flush(tlb);
240#ifdef CONFIG_HAVE_RCU_TABLE_FREE
241 tlb_table_flush(tlb);
242#endif
243
Peter Zijlstra9547d012011-05-24 17:12:14 -0700244 for (batch = &tlb->local; batch; batch = batch->next) {
245 free_pages_and_swap_cache(batch->pages, batch->nr);
246 batch->nr = 0;
247 }
248 tlb->active = &tlb->local;
249}
250
251/* tlb_finish_mmu
252 * Called at the end of the shootdown operation to free up any resources
253 * that were required.
254 */
255void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
256{
257 struct mmu_gather_batch *batch, *next;
258
Alex Shi597e1c32012-06-28 09:02:21 +0800259 tlb->start = start;
260 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700261 tlb_flush_mmu(tlb);
262
263 /* keep the page table cache within bounds */
264 check_pgt_cache();
265
266 for (batch = tlb->local.next; batch; batch = next) {
267 next = batch->next;
268 free_pages((unsigned long)batch, 0);
269 }
270 tlb->local.next = NULL;
271}
272
273/* __tlb_remove_page
274 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
275 * handling the additional races in SMP caused by other CPUs caching valid
276 * mappings in their TLBs. Returns the number of free page slots left.
277 * When out of page slots we must call tlb_flush_mmu().
278 */
279int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
280{
281 struct mmu_gather_batch *batch;
282
Shaohua Lif21760b2012-01-12 17:19:16 -0800283 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700284
Peter Zijlstra9547d012011-05-24 17:12:14 -0700285 batch = tlb->active;
286 batch->pages[batch->nr++] = page;
287 if (batch->nr == batch->max) {
288 if (!tlb_next_batch(tlb))
289 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700290 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700291 }
292 VM_BUG_ON(batch->nr > batch->max);
293
294 return batch->max - batch->nr;
295}
296
297#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800298
Peter Zijlstra26723912011-05-24 17:12:00 -0700299#ifdef CONFIG_HAVE_RCU_TABLE_FREE
300
301/*
302 * See the comment near struct mmu_table_batch.
303 */
304
305static void tlb_remove_table_smp_sync(void *arg)
306{
307 /* Simply deliver the interrupt */
308}
309
310static void tlb_remove_table_one(void *table)
311{
312 /*
313 * This isn't an RCU grace period and hence the page-tables cannot be
314 * assumed to be actually RCU-freed.
315 *
316 * It is however sufficient for software page-table walkers that rely on
317 * IRQ disabling. See the comment near struct mmu_table_batch.
318 */
319 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320 __tlb_remove_table(table);
321}
322
323static void tlb_remove_table_rcu(struct rcu_head *head)
324{
325 struct mmu_table_batch *batch;
326 int i;
327
328 batch = container_of(head, struct mmu_table_batch, rcu);
329
330 for (i = 0; i < batch->nr; i++)
331 __tlb_remove_table(batch->tables[i]);
332
333 free_page((unsigned long)batch);
334}
335
336void tlb_table_flush(struct mmu_gather *tlb)
337{
338 struct mmu_table_batch **batch = &tlb->batch;
339
340 if (*batch) {
341 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342 *batch = NULL;
343 }
344}
345
346void tlb_remove_table(struct mmu_gather *tlb, void *table)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 tlb->need_flush = 1;
351
352 /*
353 * When there's less then two users of this mm there cannot be a
354 * concurrent page-table walk.
355 */
356 if (atomic_read(&tlb->mm->mm_users) < 2) {
357 __tlb_remove_table(table);
358 return;
359 }
360
361 if (*batch == NULL) {
362 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363 if (*batch == NULL) {
364 tlb_remove_table_one(table);
365 return;
366 }
367 (*batch)->nr = 0;
368 }
369 (*batch)->tables[(*batch)->nr++] = table;
370 if ((*batch)->nr == MAX_TABLE_BATCH)
371 tlb_table_flush(tlb);
372}
373
Peter Zijlstra9547d012011-05-24 17:12:14 -0700374#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700375
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376/*
377 * If a p?d_bad entry is found while walking page tables, report
378 * the error, before resetting entry to p?d_none. Usually (but
379 * very seldom) called out from the p?d_none_or_clear_bad macros.
380 */
381
382void pgd_clear_bad(pgd_t *pgd)
383{
384 pgd_ERROR(*pgd);
385 pgd_clear(pgd);
386}
387
388void pud_clear_bad(pud_t *pud)
389{
390 pud_ERROR(*pud);
391 pud_clear(pud);
392}
393
394void pmd_clear_bad(pmd_t *pmd)
395{
396 pmd_ERROR(*pmd);
397 pmd_clear(pmd);
398}
399
400/*
401 * Note: this doesn't free the actual pages themselves. That
402 * has been handled earlier when unmapping all the memory regions.
403 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000404static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800407 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000409 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700410 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411}
412
Hugh Dickinse0da3822005-04-19 13:29:15 -0700413static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414 unsigned long addr, unsigned long end,
415 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
417 pmd_t *pmd;
418 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 do {
424 next = pmd_addr_end(addr, end);
425 if (pmd_none_or_clear_bad(pmd))
426 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000427 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 } while (pmd++, addr = next, addr != end);
429
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430 start &= PUD_MASK;
431 if (start < floor)
432 return;
433 if (ceiling) {
434 ceiling &= PUD_MASK;
435 if (!ceiling)
436 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 if (end - 1 > ceiling - 1)
439 return;
440
441 pmd = pmd_offset(pud, start);
442 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000443 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444}
445
Hugh Dickinse0da3822005-04-19 13:29:15 -0700446static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
447 unsigned long addr, unsigned long end,
448 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
450 pud_t *pud;
451 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700452 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
Hugh Dickinse0da3822005-04-19 13:29:15 -0700454 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 do {
457 next = pud_addr_end(addr, end);
458 if (pud_none_or_clear_bad(pud))
459 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700460 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 } while (pud++, addr = next, addr != end);
462
Hugh Dickinse0da3822005-04-19 13:29:15 -0700463 start &= PGDIR_MASK;
464 if (start < floor)
465 return;
466 if (ceiling) {
467 ceiling &= PGDIR_MASK;
468 if (!ceiling)
469 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 if (end - 1 > ceiling - 1)
472 return;
473
474 pud = pud_offset(pgd, start);
475 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000476 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477}
478
479/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700480 * This function frees user-level page tables of a process.
481 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 * Must be called with pagetable lock held.
483 */
Jan Beulich42b77722008-07-23 21:27:10 -0700484void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
488 pgd_t *pgd;
489 unsigned long next;
490
Hugh Dickinse0da3822005-04-19 13:29:15 -0700491 /*
492 * The next few lines have given us lots of grief...
493 *
494 * Why are we testing PMD* at this top level? Because often
495 * there will be no work to do at all, and we'd prefer not to
496 * go all the way down to the bottom just to discover that.
497 *
498 * Why all these "- 1"s? Because 0 represents both the bottom
499 * of the address space and the top of it (using -1 for the
500 * top wouldn't help much: the masks would do the wrong thing).
501 * The rule is that addr 0 and floor 0 refer to the bottom of
502 * the address space, but end 0 and ceiling 0 refer to the top
503 * Comparisons need to use "end - 1" and "ceiling - 1" (though
504 * that end 0 case should be mythical).
505 *
506 * Wherever addr is brought up or ceiling brought down, we must
507 * be careful to reject "the opposite 0" before it confuses the
508 * subsequent tests. But what about where end is brought down
509 * by PMD_SIZE below? no, end can't go down to 0 there.
510 *
511 * Whereas we round start (addr) and ceiling down, by different
512 * masks at different levels, in order to test whether a table
513 * now has no other vmas using it, so can be freed, we don't
514 * bother to round floor or end up - the tests don't need that.
515 */
516
517 addr &= PMD_MASK;
518 if (addr < floor) {
519 addr += PMD_SIZE;
520 if (!addr)
521 return;
522 }
523 if (ceiling) {
524 ceiling &= PMD_MASK;
525 if (!ceiling)
526 return;
527 }
528 if (end - 1 > ceiling - 1)
529 end -= PMD_SIZE;
530 if (addr > end - 1)
531 return;
532
Jan Beulich42b77722008-07-23 21:27:10 -0700533 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 do {
535 next = pgd_addr_end(addr, end);
536 if (pgd_none_or_clear_bad(pgd))
537 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700538 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700540}
541
Jan Beulich42b77722008-07-23 21:27:10 -0700542void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700543 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700544{
545 while (vma) {
546 struct vm_area_struct *next = vma->vm_next;
547 unsigned long addr = vma->vm_start;
548
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700549 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000550 * Hide vma from rmap and truncate_pagecache before freeing
551 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700552 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800553 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700554 unlink_file_vma(vma);
555
David Gibson9da61ae2006-03-22 00:08:57 -0800556 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700557 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700558 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700559 } else {
560 /*
561 * Optimization: gather nearby vmas into one call down
562 */
563 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800564 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700565 vma = next;
566 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800567 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700568 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700569 }
570 free_pgd_range(tlb, addr, vma->vm_end,
571 floor, next? next->vm_start: ceiling);
572 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700573 vma = next;
574 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575}
576
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800577int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
578 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800580 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800581 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700582 if (!new)
583 return -ENOMEM;
584
Nick Piggin362a61a2008-05-14 06:37:36 +0200585 /*
586 * Ensure all pte setup (eg. pte page lock and page clearing) are
587 * visible before the pte is made visible to other CPUs by being
588 * put into page tables.
589 *
590 * The other side of the story is the pointer chasing in the page
591 * table walking code (when walking the page table without locking;
592 * ie. most of the time). Fortunately, these data accesses consist
593 * of a chain of data-dependent loads, meaning most CPUs (alpha
594 * being the notable exception) will already guarantee loads are
595 * seen in-order. See the alpha page table accessors for the
596 * smp_read_barrier_depends() barriers in page table walking code.
597 */
598 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
Hugh Dickinsc74df322005-10-29 18:16:23 -0700600 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800601 wait_split_huge_page = 0;
602 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800605 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800606 } else if (unlikely(pmd_trans_splitting(*pmd)))
607 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700608 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800609 if (new)
610 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800611 if (wait_split_huge_page)
612 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614}
615
Hugh Dickins1bb36302005-10-29 18:16:22 -0700616int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700618 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
619 if (!new)
620 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Nick Piggin362a61a2008-05-14 06:37:36 +0200622 smp_wmb(); /* See comment in __pte_alloc */
623
Hugh Dickins1bb36302005-10-29 18:16:22 -0700624 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800625 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700626 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800627 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800628 } else
629 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700630 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800631 if (new)
632 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700633 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634}
635
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800636static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700637{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800638 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
639}
640
641static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
642{
643 int i;
644
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800645 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700646 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800647 for (i = 0; i < NR_MM_COUNTERS; i++)
648 if (rss[i])
649 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700650}
651
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800653 * This function is called to print an error when a bad pte
654 * is found. For example, we might have a PFN-mapped pte in
655 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700656 *
657 * The calling function must still handle the error.
658 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800659static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
660 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700661{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800662 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
663 pud_t *pud = pud_offset(pgd, addr);
664 pmd_t *pmd = pmd_offset(pud, addr);
665 struct address_space *mapping;
666 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800667 static unsigned long resume;
668 static unsigned long nr_shown;
669 static unsigned long nr_unshown;
670
671 /*
672 * Allow a burst of 60 reports, then keep quiet for that minute;
673 * or allow a steady drip of one report per second.
674 */
675 if (nr_shown == 60) {
676 if (time_before(jiffies, resume)) {
677 nr_unshown++;
678 return;
679 }
680 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800681 printk(KERN_ALERT
682 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800683 nr_unshown);
684 nr_unshown = 0;
685 }
686 nr_shown = 0;
687 }
688 if (nr_shown++ == 0)
689 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800690
691 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
692 index = linear_page_index(vma, addr);
693
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800694 printk(KERN_ALERT
695 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800696 current->comm,
697 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800698 if (page)
699 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800700 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800701 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
702 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
703 /*
704 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
705 */
706 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800707 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
708 vma->vm_ops->fault);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800709 if (vma->vm_file && vma->vm_file->f_op)
Joe Perches071361d2012-12-12 10:19:12 -0800710 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
711 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700712 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030713 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700714}
715
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700716static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800717{
718 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
719}
720
Nick Pigginb5810032005-10-29 18:16:12 -0700721/*
Nick Piggin7e675132008-04-28 02:13:00 -0700722 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800723 *
Nick Piggin7e675132008-04-28 02:13:00 -0700724 * "Special" mappings do not wish to be associated with a "struct page" (either
725 * it doesn't exist, or it exists but they don't want to touch it). In this
726 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700727 *
Nick Piggin7e675132008-04-28 02:13:00 -0700728 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
729 * pte bit, in which case this function is trivial. Secondly, an architecture
730 * may not have a spare pte bit, which requires a more complicated scheme,
731 * described below.
732 *
733 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
734 * special mapping (even if there are underlying and valid "struct pages").
735 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800736 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700737 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
738 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700739 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
740 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800741 *
742 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
743 *
Nick Piggin7e675132008-04-28 02:13:00 -0700744 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700745 *
Nick Piggin7e675132008-04-28 02:13:00 -0700746 * This restricts such mappings to be a linear translation from virtual address
747 * to pfn. To get around this restriction, we allow arbitrary mappings so long
748 * as the vma is not a COW mapping; in that case, we know that all ptes are
749 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700750 *
751 *
Nick Piggin7e675132008-04-28 02:13:00 -0700752 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
753 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700754 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
755 * page" backing, however the difference is that _all_ pages with a struct
756 * page (that is, those where pfn_valid is true) are refcounted and considered
757 * normal pages by the VM. The disadvantage is that pages are refcounted
758 * (which can be slower and simply not an option for some PFNMAP users). The
759 * advantage is that we don't have to follow the strict linearity rule of
760 * PFNMAP mappings in order to support COWable mappings.
761 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800762 */
Nick Piggin7e675132008-04-28 02:13:00 -0700763#ifdef __HAVE_ARCH_PTE_SPECIAL
764# define HAVE_PTE_SPECIAL 1
765#else
766# define HAVE_PTE_SPECIAL 0
767#endif
768struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
769 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800770{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800771 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700772
773 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800774 if (likely(!pte_special(pte)))
775 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700776 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
777 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700778 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800779 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700780 return NULL;
781 }
782
783 /* !HAVE_PTE_SPECIAL case follows: */
784
Jared Hulbertb379d792008-04-28 02:12:58 -0700785 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
786 if (vma->vm_flags & VM_MIXEDMAP) {
787 if (!pfn_valid(pfn))
788 return NULL;
789 goto out;
790 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700791 unsigned long off;
792 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700793 if (pfn == vma->vm_pgoff + off)
794 return NULL;
795 if (!is_cow_mapping(vma->vm_flags))
796 return NULL;
797 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800798 }
799
Hugh Dickins62eede62009-09-21 17:03:34 -0700800 if (is_zero_pfn(pfn))
801 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800802check_pfn:
803 if (unlikely(pfn > highest_memmap_pfn)) {
804 print_bad_pte(vma, addr, pte, NULL);
805 return NULL;
806 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800807
808 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700809 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700810 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800811 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700812out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800813 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800814}
815
816/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 * copy one vm_area from one task to the other. Assumes the page tables
818 * already present in the new task to be cleared in the whole range
819 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 */
821
Hugh Dickins570a335b2009-12-14 17:58:46 -0800822static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700824 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700825 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826{
Nick Pigginb5810032005-10-29 18:16:12 -0700827 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 pte_t pte = *src_pte;
829 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
831 /* pte contains position in swap or file, so copy. */
832 if (unlikely(!pte_present(pte))) {
833 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700834 swp_entry_t entry = pte_to_swp_entry(pte);
835
Hugh Dickins570a335b2009-12-14 17:58:46 -0800836 if (swap_duplicate(entry) < 0)
837 return entry.val;
838
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 /* make sure dst_mm is on swapoff's mmlist. */
840 if (unlikely(list_empty(&dst_mm->mmlist))) {
841 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700842 if (list_empty(&dst_mm->mmlist))
843 list_add(&dst_mm->mmlist,
844 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 spin_unlock(&mmlist_lock);
846 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800847 if (likely(!non_swap_entry(entry)))
848 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800849 else if (is_migration_entry(entry)) {
850 page = migration_entry_to_page(entry);
851
852 if (PageAnon(page))
853 rss[MM_ANONPAGES]++;
854 else
855 rss[MM_FILEPAGES]++;
856
857 if (is_write_migration_entry(entry) &&
858 is_cow_mapping(vm_flags)) {
859 /*
860 * COW mappings require pages in both
861 * parent and child to be set to read.
862 */
863 make_migration_entry_read(&entry);
864 pte = swp_entry_to_pte(entry);
865 set_pte_at(src_mm, addr, src_pte, pte);
866 }
Christoph Lameter06972122006-06-23 02:03:35 -0700867 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700869 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 }
871
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 /*
873 * If it's a COW mapping, write protect it both
874 * in the parent and the child
875 */
Linus Torvalds67121172005-12-11 20:38:17 -0800876 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700878 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 }
880
881 /*
882 * If it's a shared mapping, mark it clean in
883 * the child
884 */
885 if (vm_flags & VM_SHARED)
886 pte = pte_mkclean(pte);
887 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800888
889 page = vm_normal_page(vma, addr, pte);
890 if (page) {
891 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700892 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800893 if (PageAnon(page))
894 rss[MM_ANONPAGES]++;
895 else
896 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800897 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700898
899out_set_pte:
900 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800901 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902}
903
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800904int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
905 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
906 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700908 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700910 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700911 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800912 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800913 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
915again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800916 init_rss_vec(rss);
917
Hugh Dickinsc74df322005-10-29 18:16:23 -0700918 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 if (!dst_pte)
920 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700921 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700922 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700923 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700924 orig_src_pte = src_pte;
925 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700926 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 do {
929 /*
930 * We are holding two locks at this point - either of them
931 * could generate latencies in another task on another CPU.
932 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700933 if (progress >= 32) {
934 progress = 0;
935 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100936 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700937 break;
938 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 if (pte_none(*src_pte)) {
940 progress++;
941 continue;
942 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800943 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
944 vma, addr, rss);
945 if (entry.val)
946 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 progress += 8;
948 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700950 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700951 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700952 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800953 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700954 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700955 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800956
957 if (entry.val) {
958 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
959 return -ENOMEM;
960 progress = 0;
961 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 if (addr != end)
963 goto again;
964 return 0;
965}
966
967static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
969 unsigned long addr, unsigned long end)
970{
971 pmd_t *src_pmd, *dst_pmd;
972 unsigned long next;
973
974 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
975 if (!dst_pmd)
976 return -ENOMEM;
977 src_pmd = pmd_offset(src_pud, addr);
978 do {
979 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800980 if (pmd_trans_huge(*src_pmd)) {
981 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800982 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800983 err = copy_huge_pmd(dst_mm, src_mm,
984 dst_pmd, src_pmd, addr, vma);
985 if (err == -ENOMEM)
986 return -ENOMEM;
987 if (!err)
988 continue;
989 /* fall through */
990 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 if (pmd_none_or_clear_bad(src_pmd))
992 continue;
993 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
994 vma, addr, next))
995 return -ENOMEM;
996 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
997 return 0;
998}
999
1000static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1002 unsigned long addr, unsigned long end)
1003{
1004 pud_t *src_pud, *dst_pud;
1005 unsigned long next;
1006
1007 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1008 if (!dst_pud)
1009 return -ENOMEM;
1010 src_pud = pud_offset(src_pgd, addr);
1011 do {
1012 next = pud_addr_end(addr, end);
1013 if (pud_none_or_clear_bad(src_pud))
1014 continue;
1015 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pud++, src_pud++, addr = next, addr != end);
1019 return 0;
1020}
1021
1022int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 struct vm_area_struct *vma)
1024{
1025 pgd_t *src_pgd, *dst_pgd;
1026 unsigned long next;
1027 unsigned long addr = vma->vm_start;
1028 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001029 unsigned long mmun_start; /* For mmu_notifiers */
1030 unsigned long mmun_end; /* For mmu_notifiers */
1031 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001032 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033
Nick Piggind9928952005-08-28 16:49:11 +10001034 /*
1035 * Don't copy ptes where a page fault will fill them correctly.
1036 * Fork becomes much lighter when there are big shared or private
1037 * readonly mappings. The tradeoff is that copy_page_range is more
1038 * efficient than faulting.
1039 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001040 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1041 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001042 if (!vma->anon_vma)
1043 return 0;
1044 }
1045
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 if (is_vm_hugetlb_page(vma))
1047 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1048
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001049 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001050 /*
1051 * We do not free on error cases below as remove_vma
1052 * gets called on error from higher level routine
1053 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001054 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001055 if (ret)
1056 return ret;
1057 }
1058
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001059 /*
1060 * We need to invalidate the secondary MMU mappings only when
1061 * there could be a permission downgrade on the ptes of the
1062 * parent mm. And a permission downgrade will only happen if
1063 * is_cow_mapping() returns true.
1064 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001065 is_cow = is_cow_mapping(vma->vm_flags);
1066 mmun_start = addr;
1067 mmun_end = end;
1068 if (is_cow)
1069 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1070 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001071
1072 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 dst_pgd = pgd_offset(dst_mm, addr);
1074 src_pgd = pgd_offset(src_mm, addr);
1075 do {
1076 next = pgd_addr_end(addr, end);
1077 if (pgd_none_or_clear_bad(src_pgd))
1078 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001079 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1080 vma, addr, next))) {
1081 ret = -ENOMEM;
1082 break;
1083 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001085
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001086 if (is_cow)
1087 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001088 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
Robin Holt51c6f662005-11-13 16:06:42 -08001091static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001092 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001094 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095{
Nick Pigginb5810032005-10-29 18:16:12 -07001096 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001097 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001098 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001099 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001100 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001101 pte_t *pte;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001102 unsigned long range_start = addr;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001103
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001104again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001105 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001106 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1107 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001108 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 do {
1110 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001111 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001113 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001114
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001116 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001117
Linus Torvalds6aab3412005-11-28 14:34:23 -08001118 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 if (unlikely(details) && page) {
1120 /*
1121 * unmap_shared_mapping_pages() wants to
1122 * invalidate cache without truncating:
1123 * unmap shared but keep private pages.
1124 */
1125 if (details->check_mapping &&
1126 details->check_mapping != page->mapping)
1127 continue;
1128 /*
1129 * Each page->index must be checked when
1130 * invalidating or truncating nonlinear.
1131 */
1132 if (details->nonlinear_vma &&
1133 (page->index < details->first_index ||
1134 page->index > details->last_index))
1135 continue;
1136 }
Nick Pigginb5810032005-10-29 18:16:12 -07001137 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001138 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 tlb_remove_tlb_entry(tlb, pte, addr);
1140 if (unlikely(!page))
1141 continue;
1142 if (unlikely(details) && details->nonlinear_vma
1143 && linear_page_index(details->nonlinear_vma,
1144 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001145 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001148 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001149 else {
1150 if (pte_dirty(ptent))
1151 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001152 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001153 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001154 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001155 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001156 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001157 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001158 if (unlikely(page_mapcount(page) < 0))
1159 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001160 force_flush = !__tlb_remove_page(tlb, page);
1161 if (force_flush)
1162 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 continue;
1164 }
1165 /*
1166 * If details->check_mapping, we leave swap entries;
1167 * if details->nonlinear_vma, we leave file entries.
1168 */
1169 if (unlikely(details))
1170 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001171 if (pte_file(ptent)) {
1172 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1173 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001174 } else {
1175 swp_entry_t entry = pte_to_swp_entry(ptent);
1176
1177 if (!non_swap_entry(entry))
1178 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001179 else if (is_migration_entry(entry)) {
1180 struct page *page;
1181
1182 page = migration_entry_to_page(entry);
1183
1184 if (PageAnon(page))
1185 rss[MM_ANONPAGES]--;
1186 else
1187 rss[MM_FILEPAGES]--;
1188 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001189 if (unlikely(!free_swap_and_cache(entry)))
1190 print_bad_pte(vma, addr, ptent, NULL);
1191 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001192 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001193 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001194
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001195 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001196 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001197 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001198
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001199 /*
1200 * mmu_gather ran out of room to batch pages, we break out of
1201 * the PTE lock to avoid doing the potential expensive TLB invalidate
1202 * and page-free while holding it.
1203 */
1204 if (force_flush) {
1205 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001206
1207#ifdef HAVE_GENERIC_MMU_GATHER
Vineet Guptae6c495a2013-07-03 15:03:31 -07001208 tlb->start = range_start;
1209 tlb->end = addr;
Alex Shi597e1c32012-06-28 09:02:21 +08001210#endif
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001211 tlb_flush_mmu(tlb);
Vineet Guptae6c495a2013-07-03 15:03:31 -07001212 if (addr != end) {
1213 range_start = addr;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001214 goto again;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001215 }
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001216 }
1217
Robin Holt51c6f662005-11-13 16:06:42 -08001218 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219}
1220
Robin Holt51c6f662005-11-13 16:06:42 -08001221static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001222 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001224 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225{
1226 pmd_t *pmd;
1227 unsigned long next;
1228
1229 pmd = pmd_offset(pud, addr);
1230 do {
1231 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001232 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001233 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001234#ifdef CONFIG_DEBUG_VM
1235 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1236 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1237 __func__, addr, end,
1238 vma->vm_start,
1239 vma->vm_end);
1240 BUG();
1241 }
1242#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001243 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001244 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001245 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246 /* fall through */
1247 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001248 /*
1249 * Here there can be other concurrent MADV_DONTNEED or
1250 * trans huge page faults running, and if the pmd is
1251 * none or trans huge it can change under us. This is
1252 * because MADV_DONTNEED holds the mmap_sem in read
1253 * mode.
1254 */
1255 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1256 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001257 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001258next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001259 cond_resched();
1260 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001261
1262 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263}
1264
Robin Holt51c6f662005-11-13 16:06:42 -08001265static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001266 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001268 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269{
1270 pud_t *pud;
1271 unsigned long next;
1272
1273 pud = pud_offset(pgd, addr);
1274 do {
1275 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001276 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001278 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1279 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001280
1281 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282}
1283
Al Viro038c7aa2012-03-05 13:25:09 -05001284static void unmap_page_range(struct mmu_gather *tlb,
1285 struct vm_area_struct *vma,
1286 unsigned long addr, unsigned long end,
1287 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288{
1289 pgd_t *pgd;
1290 unsigned long next;
1291
1292 if (details && !details->check_mapping && !details->nonlinear_vma)
1293 details = NULL;
1294
1295 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001296 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 tlb_start_vma(tlb, vma);
1298 pgd = pgd_offset(vma->vm_mm, addr);
1299 do {
1300 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001301 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001303 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1304 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001306 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
Al Virof5cc4ee2012-03-05 14:14:20 -05001309
1310static void unmap_single_vma(struct mmu_gather *tlb,
1311 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001312 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001313 struct zap_details *details)
1314{
1315 unsigned long start = max(vma->vm_start, start_addr);
1316 unsigned long end;
1317
1318 if (start >= vma->vm_end)
1319 return;
1320 end = min(vma->vm_end, end_addr);
1321 if (end <= vma->vm_start)
1322 return;
1323
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301324 if (vma->vm_file)
1325 uprobe_munmap(vma, start, end);
1326
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001327 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001328 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001329
1330 if (start != end) {
1331 if (unlikely(is_vm_hugetlb_page(vma))) {
1332 /*
1333 * It is undesirable to test vma->vm_file as it
1334 * should be non-null for valid hugetlb area.
1335 * However, vm_file will be NULL in the error
1336 * cleanup path of do_mmap_pgoff. When
1337 * hugetlbfs ->mmap method fails,
1338 * do_mmap_pgoff() nullifies vma->vm_file
1339 * before calling this function to clean up.
1340 * Since no pte has actually been setup, it is
1341 * safe to do nothing in this case.
1342 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001343 if (vma->vm_file) {
1344 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001345 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001346 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1347 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001348 } else
1349 unmap_page_range(tlb, vma, start, end, details);
1350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351}
1352
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353/**
1354 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001355 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 * @vma: the starting vma
1357 * @start_addr: virtual address at which to start unmapping
1358 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001360 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 * Only addresses between `start' and `end' will be unmapped.
1363 *
1364 * The VMA list must be sorted in ascending virtual address order.
1365 *
1366 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1367 * range after unmap_vmas() returns. So the only responsibility here is to
1368 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1369 * drops the lock and schedules.
1370 */
Al Viro6e8bb012012-03-05 13:41:15 -05001371void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001373 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001375 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001377 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001378 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001379 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001380 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381}
1382
1383/**
1384 * zap_page_range - remove user pages in a given range
1385 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001386 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 * @size: number of bytes to zap
1388 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001389 *
1390 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001392void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 unsigned long size, struct zap_details *details)
1394{
1395 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001396 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001397 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001400 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001401 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001402 mmu_notifier_invalidate_range_start(mm, start, end);
1403 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001404 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001405 mmu_notifier_invalidate_range_end(mm, start, end);
1406 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407}
1408
Jack Steinerc627f9c2008-07-29 22:33:53 -07001409/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001410 * zap_page_range_single - remove user pages in a given range
1411 * @vma: vm_area_struct holding the applicable pages
1412 * @address: starting address of pages to zap
1413 * @size: number of bytes to zap
1414 * @details: details of nonlinear truncation or shared cache invalidation
1415 *
1416 * The range must fit into one VMA.
1417 */
1418static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1419 unsigned long size, struct zap_details *details)
1420{
1421 struct mm_struct *mm = vma->vm_mm;
1422 struct mmu_gather tlb;
1423 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001424
1425 lru_add_drain();
1426 tlb_gather_mmu(&tlb, mm, 0);
1427 update_hiwater_rss(mm);
1428 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001429 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001430 mmu_notifier_invalidate_range_end(mm, address, end);
1431 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432}
1433
Jack Steinerc627f9c2008-07-29 22:33:53 -07001434/**
1435 * zap_vma_ptes - remove ptes mapping the vma
1436 * @vma: vm_area_struct holding ptes to be zapped
1437 * @address: starting address of pages to zap
1438 * @size: number of bytes to zap
1439 *
1440 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1441 *
1442 * The entire address range must be fully contained within the vma.
1443 *
1444 * Returns 0 if successful.
1445 */
1446int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1447 unsigned long size)
1448{
1449 if (address < vma->vm_start || address + size > vma->vm_end ||
1450 !(vma->vm_flags & VM_PFNMAP))
1451 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001452 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001453 return 0;
1454}
1455EXPORT_SYMBOL_GPL(zap_vma_ptes);
1456
Johannes Weiner142762b2010-05-24 14:32:39 -07001457/**
Michel Lespinasse240aade2013-02-22 16:35:56 -08001458 * follow_page_mask - look up a page descriptor from a user-virtual address
Johannes Weiner142762b2010-05-24 14:32:39 -07001459 * @vma: vm_area_struct mapping @address
1460 * @address: virtual address to look up
1461 * @flags: flags modifying lookup behaviour
Michel Lespinasse240aade2013-02-22 16:35:56 -08001462 * @page_mask: on output, *page_mask is set according to the size of the page
Johannes Weiner142762b2010-05-24 14:32:39 -07001463 *
1464 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1465 *
1466 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1467 * an error pointer if there is a mapping to something not represented
1468 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 */
Michel Lespinasse240aade2013-02-22 16:35:56 -08001470struct page *follow_page_mask(struct vm_area_struct *vma,
1471 unsigned long address, unsigned int flags,
1472 unsigned int *page_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473{
1474 pgd_t *pgd;
1475 pud_t *pud;
1476 pmd_t *pmd;
1477 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001478 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001480 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
Michel Lespinasse240aade2013-02-22 16:35:56 -08001482 *page_mask = 0;
1483
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001484 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1485 if (!IS_ERR(page)) {
1486 BUG_ON(flags & FOLL_GET);
1487 goto out;
1488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001490 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 pgd = pgd_offset(mm, address);
1492 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001493 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
1495 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001496 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001497 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001498 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001499 BUG_ON(flags & FOLL_GET);
1500 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1501 goto out;
1502 }
1503 if (unlikely(pud_bad(*pud)))
1504 goto no_page_table;
1505
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001507 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001508 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001509 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001510 BUG_ON(flags & FOLL_GET);
1511 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1512 goto out;
1513 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001514 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1515 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001516 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001517 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001518 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001519 goto split_fallthrough;
1520 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001521 spin_lock(&mm->page_table_lock);
1522 if (likely(pmd_trans_huge(*pmd))) {
1523 if (unlikely(pmd_trans_splitting(*pmd))) {
1524 spin_unlock(&mm->page_table_lock);
1525 wait_split_huge_page(vma->anon_vma, pmd);
1526 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001527 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001528 pmd, flags);
1529 spin_unlock(&mm->page_table_lock);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001530 *page_mask = HPAGE_PMD_NR - 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001531 goto out;
1532 }
1533 } else
1534 spin_unlock(&mm->page_table_lock);
1535 /* fall through */
1536 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001537split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001538 if (unlikely(pmd_bad(*pmd)))
1539 goto no_page_table;
1540
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001541 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542
1543 pte = *ptep;
Hugh Dickins5117b3b2013-02-22 16:36:07 -08001544 if (!pte_present(pte)) {
1545 swp_entry_t entry;
1546 /*
1547 * KSM's break_ksm() relies upon recognizing a ksm page
1548 * even while it is being migrated, so for that case we
1549 * need migration_entry_wait().
1550 */
1551 if (likely(!(flags & FOLL_MIGRATION)))
1552 goto no_page;
1553 if (pte_none(pte) || pte_file(pte))
1554 goto no_page;
1555 entry = pte_to_swp_entry(pte);
1556 if (!is_migration_entry(entry))
1557 goto no_page;
1558 pte_unmap_unlock(ptep, ptl);
1559 migration_entry_wait(mm, pmd, address);
1560 goto split_fallthrough;
1561 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001562 if ((flags & FOLL_NUMA) && pte_numa(pte))
1563 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001564 if ((flags & FOLL_WRITE) && !pte_write(pte))
1565 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001566
Linus Torvalds6aab3412005-11-28 14:34:23 -08001567 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001568 if (unlikely(!page)) {
1569 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001570 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001571 goto bad_page;
1572 page = pte_page(pte);
1573 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001574
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001575 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001576 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001577 if (flags & FOLL_TOUCH) {
1578 if ((flags & FOLL_WRITE) &&
1579 !pte_dirty(pte) && !PageDirty(page))
1580 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001581 /*
1582 * pte_mkyoung() would be more correct here, but atomic care
1583 * is needed to avoid losing the dirty bit: it is easier to use
1584 * mark_page_accessed().
1585 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001586 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001588 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001589 /*
1590 * The preliminary mapping check is mainly to avoid the
1591 * pointless overhead of lock_page on the ZERO_PAGE
1592 * which might bounce very badly if there is contention.
1593 *
1594 * If the page is already locked, we don't need to
1595 * handle it now - vmscan will handle it later if and
1596 * when it attempts to reclaim the page.
1597 */
1598 if (page->mapping && trylock_page(page)) {
1599 lru_add_drain(); /* push cached pages to LRU */
1600 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001601 * Because we lock page here, and migration is
1602 * blocked by the pte's page reference, and we
1603 * know the page is still mapped, we don't even
1604 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001605 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001606 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001607 unlock_page(page);
1608 }
1609 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001610unlock:
1611 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001613 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001615bad_page:
1616 pte_unmap_unlock(ptep, ptl);
1617 return ERR_PTR(-EFAULT);
1618
1619no_page:
1620 pte_unmap_unlock(ptep, ptl);
1621 if (!pte_none(pte))
1622 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001623
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001624no_page_table:
1625 /*
1626 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001627 * has touched so far, we don't want to allocate unnecessary pages or
1628 * page tables. Return error instead of NULL to skip handle_mm_fault,
1629 * then get_dump_page() will return NULL to leave a hole in the dump.
1630 * But we can only make this optimization where a hole would surely
1631 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001632 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001633 if ((flags & FOLL_DUMP) &&
1634 (!vma->vm_ops || !vma->vm_ops->fault))
1635 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001636 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637}
1638
Linus Torvalds95042f92011-04-12 14:15:51 -07001639static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1640{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001641 return stack_guard_page_start(vma, addr) ||
1642 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001643}
1644
Huang Ying0014bd92011-01-30 11:15:47 +08001645/**
1646 * __get_user_pages() - pin user pages in memory
1647 * @tsk: task_struct of target task
1648 * @mm: mm_struct of target mm
1649 * @start: starting user address
1650 * @nr_pages: number of pages from start to pin
1651 * @gup_flags: flags modifying pin behaviour
1652 * @pages: array that receives pointers to the pages pinned.
1653 * Should be at least nr_pages long. Or NULL, if caller
1654 * only intends to ensure the pages are faulted in.
1655 * @vmas: array of pointers to vmas corresponding to each page.
1656 * Or NULL if the caller does not require them.
1657 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1658 *
1659 * Returns number of pages pinned. This may be fewer than the number
1660 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1661 * were pinned, returns -errno. Each page returned must be released
1662 * with a put_page() call when it is finished with. vmas will only
1663 * remain valid while mmap_sem is held.
1664 *
1665 * Must be called with mmap_sem held for read or write.
1666 *
1667 * __get_user_pages walks a process's page tables and takes a reference to
1668 * each struct page that each user address corresponds to at a given
1669 * instant. That is, it takes the page that would be accessed if a user
1670 * thread accesses the given user virtual address at that instant.
1671 *
1672 * This does not guarantee that the page exists in the user mappings when
1673 * __get_user_pages returns, and there may even be a completely different
1674 * page there in some cases (eg. if mmapped pagecache has been invalidated
1675 * and subsequently re faulted). However it does guarantee that the page
1676 * won't be freed completely. And mostly callers simply care that the page
1677 * contains data that was valid *at some point in time*. Typically, an IO
1678 * or similar operation cannot guarantee anything stronger anyway because
1679 * locks can't be held over the syscall boundary.
1680 *
1681 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1682 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1683 * appropriate) must be called after the page is finished with, and
1684 * before put_page is called.
1685 *
1686 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1687 * or mmap_sem contention, and if waiting is needed to pin all pages,
1688 * *@nonblocking will be set to 0.
1689 *
1690 * In most cases, get_user_pages or get_user_pages_fast should be used
1691 * instead of __get_user_pages. __get_user_pages should be used only if
1692 * you need some special @gup_flags.
1693 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08001694long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1695 unsigned long start, unsigned long nr_pages,
1696 unsigned int gup_flags, struct page **pages,
1697 struct vm_area_struct **vmas, int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698{
Michel Lespinasse28a35712013-02-22 16:35:55 -08001699 long i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001700 unsigned long vm_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001701 unsigned int page_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Michel Lespinasse28a35712013-02-22 16:35:55 -08001703 if (!nr_pages)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001704 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001705
1706 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1707
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 /*
1709 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001710 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001712 vm_flags = (gup_flags & FOLL_WRITE) ?
1713 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1714 vm_flags &= (gup_flags & FOLL_FORCE) ?
1715 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001716
1717 /*
1718 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1719 * would be called on PROT_NONE ranges. We must never invoke
1720 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1721 * page faults would unprotect the PROT_NONE ranges if
1722 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1723 * bitflag. So to avoid that, don't set FOLL_NUMA if
1724 * FOLL_FORCE is set.
1725 */
1726 if (!(gup_flags & FOLL_FORCE))
1727 gup_flags |= FOLL_NUMA;
1728
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 i = 0;
1730
1731 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001732 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733
1734 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001735 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 pgd_t *pgd;
1738 pud_t *pud;
1739 pmd_t *pmd;
1740 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001741
1742 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001743 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 return i ? : -EFAULT;
1745 if (pg > TASK_SIZE)
1746 pgd = pgd_offset_k(pg);
1747 else
1748 pgd = pgd_offset_gate(mm, pg);
1749 BUG_ON(pgd_none(*pgd));
1750 pud = pud_offset(pgd, pg);
1751 BUG_ON(pud_none(*pud));
1752 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001753 if (pmd_none(*pmd))
1754 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001755 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001757 if (pte_none(*pte)) {
1758 pte_unmap(pte);
1759 return i ? : -EFAULT;
1760 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001761 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001763 struct page *page;
1764
Linus Torvalds95042f92011-04-12 14:15:51 -07001765 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001766 if (!page) {
1767 if (!(gup_flags & FOLL_DUMP) &&
1768 is_zero_pfn(pte_pfn(*pte)))
1769 page = pte_page(*pte);
1770 else {
1771 pte_unmap(pte);
1772 return i ? : -EFAULT;
1773 }
1774 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001775 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001776 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 }
1778 pte_unmap(pte);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001779 page_mask = 0;
Linus Torvalds95042f92011-04-12 14:15:51 -07001780 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 }
1782
Nick Pigginb291f002008-10-18 20:26:44 -07001783 if (!vma ||
1784 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001785 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 return i ? : -EFAULT;
1787
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001788 if (is_vm_hugetlb_page(vma)) {
1789 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001790 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001791 continue;
1792 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001793
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001795 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001796 unsigned int foll_flags = gup_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001797 unsigned int page_increm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
Ethan Solomita462e00c2007-07-15 23:38:16 -07001799 /*
Ying Han47792802009-01-06 14:40:18 -08001800 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001801 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001802 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001803 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001804 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001805
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001806 cond_resched();
Michel Lespinasse240aade2013-02-22 16:35:56 -08001807 while (!(page = follow_page_mask(vma, start,
1808 foll_flags, &page_mask))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001809 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001810 unsigned int fault_flags = 0;
1811
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001812 /* For mlock, just skip the stack guard page. */
1813 if (foll_flags & FOLL_MLOCK) {
1814 if (stack_guard_page(vma, start))
1815 goto next_page;
1816 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001817 if (foll_flags & FOLL_WRITE)
1818 fault_flags |= FAULT_FLAG_WRITE;
1819 if (nonblocking)
1820 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001821 if (foll_flags & FOLL_NOWAIT)
1822 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001823
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001824 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001825 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001826
Nick Piggin83c54072007-07-19 01:47:05 -07001827 if (ret & VM_FAULT_ERROR) {
1828 if (ret & VM_FAULT_OOM)
1829 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001830 if (ret & (VM_FAULT_HWPOISON |
1831 VM_FAULT_HWPOISON_LARGE)) {
1832 if (i)
1833 return i;
1834 else if (gup_flags & FOLL_HWPOISON)
1835 return -EHWPOISON;
1836 else
1837 return -EFAULT;
1838 }
1839 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001840 return i ? i : -EFAULT;
1841 BUG();
1842 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001843
1844 if (tsk) {
1845 if (ret & VM_FAULT_MAJOR)
1846 tsk->maj_flt++;
1847 else
1848 tsk->min_flt++;
1849 }
Nick Piggin83c54072007-07-19 01:47:05 -07001850
Michel Lespinasse53a77062011-01-13 15:46:14 -08001851 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001852 if (nonblocking)
1853 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001854 return i;
1855 }
1856
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001857 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001858 * The VM_FAULT_WRITE bit tells us that
1859 * do_wp_page has broken COW when necessary,
1860 * even if maybe_mkwrite decided not to set
1861 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001862 * page lookups as if they were reads. But only
1863 * do so when looping for pte_write is futile:
1864 * in some cases userspace may also be wanting
1865 * to write to the gotten user page, which a
1866 * read fault here might prevent (a readonly
1867 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001868 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001869 if ((ret & VM_FAULT_WRITE) &&
1870 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001871 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001872
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001873 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001875 if (IS_ERR(page))
1876 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001878 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001879
Russell Kinga6f36be2006-12-30 22:24:19 +00001880 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001881 flush_dcache_page(page);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001882 page_mask = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001884next_page:
Michel Lespinasse240aade2013-02-22 16:35:56 -08001885 if (vmas) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 vmas[i] = vma;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001887 page_mask = 0;
1888 }
1889 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1890 if (page_increm > nr_pages)
1891 page_increm = nr_pages;
1892 i += page_increm;
1893 start += page_increm * PAGE_SIZE;
1894 nr_pages -= page_increm;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001895 } while (nr_pages && start < vma->vm_end);
1896 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 return i;
1898}
Huang Ying0014bd92011-01-30 11:15:47 +08001899EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001900
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001901/*
1902 * fixup_user_fault() - manually resolve a user page fault
1903 * @tsk: the task_struct to use for page fault accounting, or
1904 * NULL if faults are not to be recorded.
1905 * @mm: mm_struct of target mm
1906 * @address: user address
1907 * @fault_flags:flags to pass down to handle_mm_fault()
1908 *
1909 * This is meant to be called in the specific scenario where for locking reasons
1910 * we try to access user memory in atomic context (within a pagefault_disable()
1911 * section), this returns -EFAULT, and we want to resolve the user fault before
1912 * trying again.
1913 *
1914 * Typically this is meant to be used by the futex code.
1915 *
1916 * The main difference with get_user_pages() is that this function will
1917 * unconditionally call handle_mm_fault() which will in turn perform all the
1918 * necessary SW fixup of the dirty and young bits in the PTE, while
1919 * handle_mm_fault() only guarantees to update these in the struct page.
1920 *
1921 * This is important for some architectures where those bits also gate the
1922 * access permission to the page because they are maintained in software. On
1923 * such architectures, gup() will not be enough to make a subsequent access
1924 * succeed.
1925 *
1926 * This should be called with the mm_sem held for read.
1927 */
1928int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1929 unsigned long address, unsigned int fault_flags)
1930{
1931 struct vm_area_struct *vma;
1932 int ret;
1933
1934 vma = find_extend_vma(mm, address);
1935 if (!vma || address < vma->vm_start)
1936 return -EFAULT;
1937
1938 ret = handle_mm_fault(mm, vma, address, fault_flags);
1939 if (ret & VM_FAULT_ERROR) {
1940 if (ret & VM_FAULT_OOM)
1941 return -ENOMEM;
1942 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1943 return -EHWPOISON;
1944 if (ret & VM_FAULT_SIGBUS)
1945 return -EFAULT;
1946 BUG();
1947 }
1948 if (tsk) {
1949 if (ret & VM_FAULT_MAJOR)
1950 tsk->maj_flt++;
1951 else
1952 tsk->min_flt++;
1953 }
1954 return 0;
1955}
1956
1957/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001958 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001959 * @tsk: the task_struct to use for page fault accounting, or
1960 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001961 * @mm: mm_struct of target mm
1962 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001963 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001964 * @write: whether pages will be written to by the caller
1965 * @force: whether to force write access even if user mapping is
1966 * readonly. This will result in the page being COWed even
1967 * in MAP_SHARED mappings. You do not want this.
1968 * @pages: array that receives pointers to the pages pinned.
1969 * Should be at least nr_pages long. Or NULL, if caller
1970 * only intends to ensure the pages are faulted in.
1971 * @vmas: array of pointers to vmas corresponding to each page.
1972 * Or NULL if the caller does not require them.
1973 *
1974 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001975 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001976 * were pinned, returns -errno. Each page returned must be released
1977 * with a put_page() call when it is finished with. vmas will only
1978 * remain valid while mmap_sem is held.
1979 *
1980 * Must be called with mmap_sem held for read or write.
1981 *
1982 * get_user_pages walks a process's page tables and takes a reference to
1983 * each struct page that each user address corresponds to at a given
1984 * instant. That is, it takes the page that would be accessed if a user
1985 * thread accesses the given user virtual address at that instant.
1986 *
1987 * This does not guarantee that the page exists in the user mappings when
1988 * get_user_pages returns, and there may even be a completely different
1989 * page there in some cases (eg. if mmapped pagecache has been invalidated
1990 * and subsequently re faulted). However it does guarantee that the page
1991 * won't be freed completely. And mostly callers simply care that the page
1992 * contains data that was valid *at some point in time*. Typically, an IO
1993 * or similar operation cannot guarantee anything stronger anyway because
1994 * locks can't be held over the syscall boundary.
1995 *
1996 * If write=0, the page must not be written to. If the page is written to,
1997 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1998 * after the page is finished with, and before put_page is called.
1999 *
2000 * get_user_pages is typically used for fewer-copy IO operations, to get a
2001 * handle on the memory by some means other than accesses via the user virtual
2002 * addresses. The pages may be submitted for DMA to devices or accessed via
2003 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2004 * use the correct cache flushing APIs.
2005 *
2006 * See also get_user_pages_fast, for performance critical applications.
2007 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08002008long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2009 unsigned long start, unsigned long nr_pages, int write,
2010 int force, struct page **pages, struct vm_area_struct **vmas)
Nick Pigginb291f002008-10-18 20:26:44 -07002011{
Hugh Dickins58fa8792009-09-21 17:03:31 -07002012 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07002013
Hugh Dickins58fa8792009-09-21 17:03:31 -07002014 if (pages)
2015 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07002016 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002017 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07002018 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002019 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07002020
Michel Lespinasse53a77062011-01-13 15:46:14 -08002021 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2022 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07002023}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024EXPORT_SYMBOL(get_user_pages);
2025
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002026/**
2027 * get_dump_page() - pin user page in memory while writing it to core dump
2028 * @addr: user address
2029 *
2030 * Returns struct page pointer of user page pinned for dump,
2031 * to be freed afterwards by page_cache_release() or put_page().
2032 *
2033 * Returns NULL on any kind of failure - a hole must then be inserted into
2034 * the corefile, to preserve alignment with its headers; and also returns
2035 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2036 * allowing a hole to be left in the corefile to save diskspace.
2037 *
2038 * Called without mmap_sem, but after all other threads have been killed.
2039 */
2040#ifdef CONFIG_ELF_CORE
2041struct page *get_dump_page(unsigned long addr)
2042{
2043 struct vm_area_struct *vma;
2044 struct page *page;
2045
2046 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002047 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2048 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002049 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002050 flush_cache_page(vma, addr, page_to_pfn(page));
2051 return page;
2052}
2053#endif /* CONFIG_ELF_CORE */
2054
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002055pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002056 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002057{
2058 pgd_t * pgd = pgd_offset(mm, addr);
2059 pud_t * pud = pud_alloc(mm, pgd, addr);
2060 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002061 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002062 if (pmd) {
2063 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002064 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002065 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002066 }
2067 return NULL;
2068}
2069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002071 * This is the old fallback for page remapping.
2072 *
2073 * For historical reasons, it only allows reserved pages. Only
2074 * old drivers should use this, and they needed to mark their
2075 * pages reserved for the old functions anyway.
2076 */
Nick Piggin423bad62008-04-28 02:13:01 -07002077static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2078 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002079{
Nick Piggin423bad62008-04-28 02:13:01 -07002080 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002081 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002082 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002083 spinlock_t *ptl;
2084
Linus Torvalds238f58d2005-11-29 13:01:56 -08002085 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002086 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002087 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002088 retval = -ENOMEM;
2089 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002090 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002091 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002092 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002093 retval = -EBUSY;
2094 if (!pte_none(*pte))
2095 goto out_unlock;
2096
2097 /* Ok, finally just insert the thing.. */
2098 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002099 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002100 page_add_file_rmap(page);
2101 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2102
2103 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002104 pte_unmap_unlock(pte, ptl);
2105 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002106out_unlock:
2107 pte_unmap_unlock(pte, ptl);
2108out:
2109 return retval;
2110}
2111
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002112/**
2113 * vm_insert_page - insert single page into user vma
2114 * @vma: user vma to map to
2115 * @addr: target user address of this page
2116 * @page: source kernel page
2117 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002118 * This allows drivers to insert individual pages they've allocated
2119 * into a user vma.
2120 *
2121 * The page has to be a nice clean _individual_ kernel allocation.
2122 * If you allocate a compound page, you need to have marked it as
2123 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002124 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002125 *
2126 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2127 * took an arbitrary page protection parameter. This doesn't allow
2128 * that. Your vma protection will have to be set up correctly, which
2129 * means that if you want a shared writable mapping, you'd better
2130 * ask for a shared writable mapping!
2131 *
2132 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002133 *
2134 * Usually this function is called from f_op->mmap() handler
2135 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2136 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2137 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002138 */
Nick Piggin423bad62008-04-28 02:13:01 -07002139int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2140 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002141{
2142 if (addr < vma->vm_start || addr >= vma->vm_end)
2143 return -EFAULT;
2144 if (!page_count(page))
2145 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002146 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2147 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2148 BUG_ON(vma->vm_flags & VM_PFNMAP);
2149 vma->vm_flags |= VM_MIXEDMAP;
2150 }
Nick Piggin423bad62008-04-28 02:13:01 -07002151 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002152}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002153EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002154
Nick Piggin423bad62008-04-28 02:13:01 -07002155static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2156 unsigned long pfn, pgprot_t prot)
2157{
2158 struct mm_struct *mm = vma->vm_mm;
2159 int retval;
2160 pte_t *pte, entry;
2161 spinlock_t *ptl;
2162
2163 retval = -ENOMEM;
2164 pte = get_locked_pte(mm, addr, &ptl);
2165 if (!pte)
2166 goto out;
2167 retval = -EBUSY;
2168 if (!pte_none(*pte))
2169 goto out_unlock;
2170
2171 /* Ok, finally just insert the thing.. */
2172 entry = pte_mkspecial(pfn_pte(pfn, prot));
2173 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002174 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002175
2176 retval = 0;
2177out_unlock:
2178 pte_unmap_unlock(pte, ptl);
2179out:
2180 return retval;
2181}
2182
Nick Piggine0dc0d82007-02-12 00:51:36 -08002183/**
2184 * vm_insert_pfn - insert single pfn into user vma
2185 * @vma: user vma to map to
2186 * @addr: target user address of this page
2187 * @pfn: source kernel pfn
2188 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002189 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002190 * they've allocated into a user vma. Same comments apply.
2191 *
2192 * This function should only be called from a vm_ops->fault handler, and
2193 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002194 *
2195 * vma cannot be a COW mapping.
2196 *
2197 * As this is called only for pages that do not currently exist, we
2198 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002199 */
2200int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002201 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002202{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002203 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002204 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002205 /*
2206 * Technically, architectures with pte_special can avoid all these
2207 * restrictions (same for remap_pfn_range). However we would like
2208 * consistency in testing and feature parity among all, so we should
2209 * try to keep these invariants in place for everybody.
2210 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002211 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2212 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2213 (VM_PFNMAP|VM_MIXEDMAP));
2214 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2215 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002216
Nick Piggin423bad62008-04-28 02:13:01 -07002217 if (addr < vma->vm_start || addr >= vma->vm_end)
2218 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002219 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002220 return -EINVAL;
2221
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002222 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002223
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002224 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002225}
2226EXPORT_SYMBOL(vm_insert_pfn);
2227
Nick Piggin423bad62008-04-28 02:13:01 -07002228int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2229 unsigned long pfn)
2230{
2231 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2232
2233 if (addr < vma->vm_start || addr >= vma->vm_end)
2234 return -EFAULT;
2235
2236 /*
2237 * If we don't have pte special, then we have to use the pfn_valid()
2238 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2239 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002240 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2241 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002242 */
2243 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2244 struct page *page;
2245
2246 page = pfn_to_page(pfn);
2247 return insert_page(vma, addr, page, vma->vm_page_prot);
2248 }
2249 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2250}
2251EXPORT_SYMBOL(vm_insert_mixed);
2252
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002253/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 * maps a range of physical memory into the requested pages. the old
2255 * mappings are removed. any references to nonexistent pages results
2256 * in null mappings (currently treated as "copy-on-access")
2257 */
2258static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2259 unsigned long addr, unsigned long end,
2260 unsigned long pfn, pgprot_t prot)
2261{
2262 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002263 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264
Hugh Dickinsc74df322005-10-29 18:16:23 -07002265 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 if (!pte)
2267 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002268 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 do {
2270 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002271 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 pfn++;
2273 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002274 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002275 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 return 0;
2277}
2278
2279static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2280 unsigned long addr, unsigned long end,
2281 unsigned long pfn, pgprot_t prot)
2282{
2283 pmd_t *pmd;
2284 unsigned long next;
2285
2286 pfn -= addr >> PAGE_SHIFT;
2287 pmd = pmd_alloc(mm, pud, addr);
2288 if (!pmd)
2289 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002290 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 do {
2292 next = pmd_addr_end(addr, end);
2293 if (remap_pte_range(mm, pmd, addr, next,
2294 pfn + (addr >> PAGE_SHIFT), prot))
2295 return -ENOMEM;
2296 } while (pmd++, addr = next, addr != end);
2297 return 0;
2298}
2299
2300static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2301 unsigned long addr, unsigned long end,
2302 unsigned long pfn, pgprot_t prot)
2303{
2304 pud_t *pud;
2305 unsigned long next;
2306
2307 pfn -= addr >> PAGE_SHIFT;
2308 pud = pud_alloc(mm, pgd, addr);
2309 if (!pud)
2310 return -ENOMEM;
2311 do {
2312 next = pud_addr_end(addr, end);
2313 if (remap_pmd_range(mm, pud, addr, next,
2314 pfn + (addr >> PAGE_SHIFT), prot))
2315 return -ENOMEM;
2316 } while (pud++, addr = next, addr != end);
2317 return 0;
2318}
2319
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002320/**
2321 * remap_pfn_range - remap kernel memory to userspace
2322 * @vma: user vma to map to
2323 * @addr: target user address to start at
2324 * @pfn: physical address of kernel memory
2325 * @size: size of map area
2326 * @prot: page protection flags for this mapping
2327 *
2328 * Note: this is only safe if the mm semaphore is held when called.
2329 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2331 unsigned long pfn, unsigned long size, pgprot_t prot)
2332{
2333 pgd_t *pgd;
2334 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002335 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 struct mm_struct *mm = vma->vm_mm;
2337 int err;
2338
2339 /*
2340 * Physically remapped pages are special. Tell the
2341 * rest of the world about it:
2342 * VM_IO tells people not to look at these pages
2343 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002344 * VM_PFNMAP tells the core MM that the base pages are just
2345 * raw PFN mappings, and do not have a "struct page" associated
2346 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002347 * VM_DONTEXPAND
2348 * Disable vma merging and expanding with mremap().
2349 * VM_DONTDUMP
2350 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002351 *
2352 * There's a horrible special case to handle copy-on-write
2353 * behaviour that some programs depend on. We mark the "original"
2354 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002355 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002357 if (is_cow_mapping(vma->vm_flags)) {
2358 if (addr != vma->vm_start || end != vma->vm_end)
2359 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002360 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002361 }
2362
2363 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2364 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002365 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002366
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002367 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
2369 BUG_ON(addr >= end);
2370 pfn -= addr >> PAGE_SHIFT;
2371 pgd = pgd_offset(mm, addr);
2372 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 do {
2374 next = pgd_addr_end(addr, end);
2375 err = remap_pud_range(mm, pgd, addr, next,
2376 pfn + (addr >> PAGE_SHIFT), prot);
2377 if (err)
2378 break;
2379 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002380
2381 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002382 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002383
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 return err;
2385}
2386EXPORT_SYMBOL(remap_pfn_range);
2387
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07002388/**
2389 * vm_iomap_memory - remap memory to userspace
2390 * @vma: user vma to map to
2391 * @start: start of area
2392 * @len: size of area
2393 *
2394 * This is a simplified io_remap_pfn_range() for common driver use. The
2395 * driver just needs to give us the physical memory range to be mapped,
2396 * we'll figure out the rest from the vma information.
2397 *
2398 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2399 * whatever write-combining details or similar.
2400 */
2401int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2402{
2403 unsigned long vm_len, pfn, pages;
2404
2405 /* Check that the physical memory area passed in looks valid */
2406 if (start + len < start)
2407 return -EINVAL;
2408 /*
2409 * You *really* shouldn't map things that aren't page-aligned,
2410 * but we've historically allowed it because IO memory might
2411 * just have smaller alignment.
2412 */
2413 len += start & ~PAGE_MASK;
2414 pfn = start >> PAGE_SHIFT;
2415 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2416 if (pfn + pages < pfn)
2417 return -EINVAL;
2418
2419 /* We start the mapping 'vm_pgoff' pages into the area */
2420 if (vma->vm_pgoff > pages)
2421 return -EINVAL;
2422 pfn += vma->vm_pgoff;
2423 pages -= vma->vm_pgoff;
2424
2425 /* Can we fit all of the mapping? */
2426 vm_len = vma->vm_end - vma->vm_start;
2427 if (vm_len >> PAGE_SHIFT > pages)
2428 return -EINVAL;
2429
2430 /* Ok, let it rip */
2431 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2432}
2433EXPORT_SYMBOL(vm_iomap_memory);
2434
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002435static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2436 unsigned long addr, unsigned long end,
2437 pte_fn_t fn, void *data)
2438{
2439 pte_t *pte;
2440 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002441 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002442 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002443
2444 pte = (mm == &init_mm) ?
2445 pte_alloc_kernel(pmd, addr) :
2446 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2447 if (!pte)
2448 return -ENOMEM;
2449
2450 BUG_ON(pmd_huge(*pmd));
2451
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002452 arch_enter_lazy_mmu_mode();
2453
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002454 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002455
2456 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002457 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002458 if (err)
2459 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002460 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002461
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002462 arch_leave_lazy_mmu_mode();
2463
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002464 if (mm != &init_mm)
2465 pte_unmap_unlock(pte-1, ptl);
2466 return err;
2467}
2468
2469static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2470 unsigned long addr, unsigned long end,
2471 pte_fn_t fn, void *data)
2472{
2473 pmd_t *pmd;
2474 unsigned long next;
2475 int err;
2476
Andi Kleenceb86872008-07-23 21:27:50 -07002477 BUG_ON(pud_huge(*pud));
2478
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002479 pmd = pmd_alloc(mm, pud, addr);
2480 if (!pmd)
2481 return -ENOMEM;
2482 do {
2483 next = pmd_addr_end(addr, end);
2484 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2485 if (err)
2486 break;
2487 } while (pmd++, addr = next, addr != end);
2488 return err;
2489}
2490
2491static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2492 unsigned long addr, unsigned long end,
2493 pte_fn_t fn, void *data)
2494{
2495 pud_t *pud;
2496 unsigned long next;
2497 int err;
2498
2499 pud = pud_alloc(mm, pgd, addr);
2500 if (!pud)
2501 return -ENOMEM;
2502 do {
2503 next = pud_addr_end(addr, end);
2504 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2505 if (err)
2506 break;
2507 } while (pud++, addr = next, addr != end);
2508 return err;
2509}
2510
2511/*
2512 * Scan a region of virtual memory, filling in page tables as necessary
2513 * and calling a provided function on each leaf page table.
2514 */
2515int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2516 unsigned long size, pte_fn_t fn, void *data)
2517{
2518 pgd_t *pgd;
2519 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002520 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002521 int err;
2522
2523 BUG_ON(addr >= end);
2524 pgd = pgd_offset(mm, addr);
2525 do {
2526 next = pgd_addr_end(addr, end);
2527 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2528 if (err)
2529 break;
2530 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002531
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002532 return err;
2533}
2534EXPORT_SYMBOL_GPL(apply_to_page_range);
2535
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002537 * handle_pte_fault chooses page fault handler according to an entry
2538 * which was read non-atomically. Before making any commitment, on
2539 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002540 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002541 * must check under lock before unmapping the pte and proceeding
2542 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002543 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002544 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002545static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002546 pte_t *page_table, pte_t orig_pte)
2547{
2548 int same = 1;
2549#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2550 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002551 spinlock_t *ptl = pte_lockptr(mm, pmd);
2552 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002553 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002554 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002555 }
2556#endif
2557 pte_unmap(page_table);
2558 return same;
2559}
2560
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002561static 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 -08002562{
2563 /*
2564 * If the source page was a PFN mapping, we don't have
2565 * a "struct page" for it. We do a best-effort copy by
2566 * just copying from the original user address. If that
2567 * fails, we just zero-fill it. Live with it.
2568 */
2569 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002570 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002571 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2572
2573 /*
2574 * This really shouldn't fail, because the page is there
2575 * in the page tables. But it might just be unreadable,
2576 * in which case we just give up and fill the result with
2577 * zeroes.
2578 */
2579 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002580 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002581 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002582 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002583 } else
2584 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002585}
2586
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 * This routine handles present pages, when users try to write
2589 * to a shared page. It is done by copying the page to a new address
2590 * and decrementing the shared-page counter for the old page.
2591 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 * Note that this routine assumes that the protection checks have been
2593 * done by the caller (the low-level page fault routine in most cases).
2594 * Thus we can safely just mark it writable once we've done any necessary
2595 * COW.
2596 *
2597 * We also mark the page dirty at this point even though the page will
2598 * change only once the write actually happens. This avoids a few races,
2599 * and potentially makes it more efficient.
2600 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002601 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2602 * but allow concurrent faults), with pte both mapped and locked.
2603 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002605static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2606 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002607 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002608 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002610 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002612 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002613 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002614 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002615 unsigned long mmun_start = 0; /* For mmu_notifiers */
2616 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617
Linus Torvalds6aab3412005-11-28 14:34:23 -08002618 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002619 if (!old_page) {
2620 /*
2621 * VM_MIXEDMAP !pfn_valid() case
2622 *
2623 * We should not cow pages in a shared writeable mapping.
2624 * Just mark the pages writable as we can't do any dirty
2625 * accounting on raw pfn maps.
2626 */
2627 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2628 (VM_WRITE|VM_SHARED))
2629 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002630 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002631 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002633 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002634 * Take out anonymous pages first, anonymous shared vmas are
2635 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002636 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002637 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002638 if (!trylock_page(old_page)) {
2639 page_cache_get(old_page);
2640 pte_unmap_unlock(page_table, ptl);
2641 lock_page(old_page);
2642 page_table = pte_offset_map_lock(mm, pmd, address,
2643 &ptl);
2644 if (!pte_same(*page_table, orig_pte)) {
2645 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002646 goto unlock;
2647 }
2648 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002649 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002650 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002651 /*
2652 * The page is all ours. Move it to our anon_vma so
2653 * the rmap code will not search our parent or siblings.
2654 * Protected against the rmap code by the page lock.
2655 */
2656 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002657 unlock_page(old_page);
2658 goto reuse;
2659 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002660 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002661 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002662 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002663 /*
2664 * Only catch write-faults on shared writable pages,
2665 * read-only shared pages can get COWed by
2666 * get_user_pages(.write=1, .force=1).
2667 */
David Howells9637a5e2006-06-23 02:03:43 -07002668 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002669 struct vm_fault vmf;
2670 int tmp;
2671
2672 vmf.virtual_address = (void __user *)(address &
2673 PAGE_MASK);
2674 vmf.pgoff = old_page->index;
2675 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2676 vmf.page = old_page;
2677
David Howells9637a5e2006-06-23 02:03:43 -07002678 /*
2679 * Notify the address space that the page is about to
2680 * become writable so that it can prohibit this or wait
2681 * for the page to get into an appropriate state.
2682 *
2683 * We do this without the lock held, so that it can
2684 * sleep if it needs to.
2685 */
2686 page_cache_get(old_page);
2687 pte_unmap_unlock(page_table, ptl);
2688
Nick Pigginc2ec1752009-03-31 15:23:21 -07002689 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2690 if (unlikely(tmp &
2691 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2692 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002693 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002694 }
Nick Pigginb827e492009-04-30 15:08:16 -07002695 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2696 lock_page(old_page);
2697 if (!old_page->mapping) {
2698 ret = 0; /* retry the fault */
2699 unlock_page(old_page);
2700 goto unwritable_page;
2701 }
2702 } else
2703 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002704
David Howells9637a5e2006-06-23 02:03:43 -07002705 /*
2706 * Since we dropped the lock we need to revalidate
2707 * the PTE as someone else may have changed it. If
2708 * they did, we just return, as we can count on the
2709 * MMU to tell us if they didn't also make it writable.
2710 */
2711 page_table = pte_offset_map_lock(mm, pmd, address,
2712 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002713 if (!pte_same(*page_table, orig_pte)) {
2714 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002715 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002716 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002717
2718 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002720 dirty_page = old_page;
2721 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002722
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002723reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002724 flush_cache_page(vma, address, pte_pfn(orig_pte));
2725 entry = pte_mkyoung(orig_pte);
2726 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002727 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002728 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002729 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002730 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002731
2732 if (!dirty_page)
2733 return ret;
2734
2735 /*
2736 * Yes, Virginia, this is actually required to prevent a race
2737 * with clear_page_dirty_for_io() from clearing the page dirty
2738 * bit after it clear all dirty ptes, but before a racing
2739 * do_wp_page installs a dirty pte.
2740 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002741 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002742 */
2743 if (!page_mkwrite) {
2744 wait_on_page_locked(dirty_page);
2745 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002746 /* file_update_time outside page_lock */
2747 if (vma->vm_file)
2748 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002749 }
2750 put_page(dirty_page);
2751 if (page_mkwrite) {
2752 struct address_space *mapping = dirty_page->mapping;
2753
2754 set_page_dirty(dirty_page);
2755 unlock_page(dirty_page);
2756 page_cache_release(dirty_page);
2757 if (mapping) {
2758 /*
2759 * Some device drivers do not set page.mapping
2760 * but still dirty their pages
2761 */
2762 balance_dirty_pages_ratelimited(mapping);
2763 }
2764 }
2765
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002766 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768
2769 /*
2770 * Ok, we need to copy. Oh, well..
2771 */
Nick Pigginb5810032005-10-29 18:16:12 -07002772 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002773gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002774 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775
2776 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002777 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002778
Hugh Dickins62eede62009-09-21 17:03:34 -07002779 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002780 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2781 if (!new_page)
2782 goto oom;
2783 } else {
2784 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2785 if (!new_page)
2786 goto oom;
2787 cow_user_page(new_page, old_page, address, vma);
2788 }
2789 __SetPageUptodate(new_page);
2790
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002791 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002792 goto oom_free_new;
2793
Haggai Eran6bdb9132012-10-08 16:33:35 -07002794 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002795 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002796 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2797
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 /*
2799 * Re-check the pte - we dropped the lock
2800 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002801 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002802 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002803 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002804 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002805 dec_mm_counter_fast(mm, MM_FILEPAGES);
2806 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002807 }
2808 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002809 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002810 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002811 entry = mk_pte(new_page, vma->vm_page_prot);
2812 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002813 /*
2814 * Clear the pte entry and flush it first, before updating the
2815 * pte with the new entry. This will avoid a race condition
2816 * seen in the presence of one thread doing SMC and another
2817 * thread doing COW.
2818 */
Izik Eidus828502d2009-09-21 17:01:51 -07002819 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002820 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002821 /*
2822 * We call the notify macro here because, when using secondary
2823 * mmu page tables (such as kvm shadow page tables), we want the
2824 * new page to be mapped directly into the secondary page table.
2825 */
2826 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002827 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002828 if (old_page) {
2829 /*
2830 * Only after switching the pte to the new page may
2831 * we remove the mapcount here. Otherwise another
2832 * process may come and find the rmap count decremented
2833 * before the pte is switched to the new page, and
2834 * "reuse" the old page writing into it while our pte
2835 * here still points into it and can be read by other
2836 * threads.
2837 *
2838 * The critical issue is to order this
2839 * page_remove_rmap with the ptp_clear_flush above.
2840 * Those stores are ordered by (if nothing else,)
2841 * the barrier present in the atomic_add_negative
2842 * in page_remove_rmap.
2843 *
2844 * Then the TLB flush in ptep_clear_flush ensures that
2845 * no process can access the old page before the
2846 * decremented mapcount is visible. And the old page
2847 * cannot be reused until after the decremented
2848 * mapcount is visible. So transitively, TLBs to
2849 * old page will be flushed before it can be reused.
2850 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002851 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002852 }
2853
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 /* Free the old page.. */
2855 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002856 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002857 } else
2858 mem_cgroup_uncharge_page(new_page);
2859
Haggai Eran6bdb9132012-10-08 16:33:35 -07002860 if (new_page)
2861 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002862unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002863 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002864 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002865 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002866 if (old_page) {
2867 /*
2868 * Don't let another task, with possibly unlocked vma,
2869 * keep the mlocked page.
2870 */
2871 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2872 lock_page(old_page); /* LRU manipulation */
2873 munlock_vma_page(old_page);
2874 unlock_page(old_page);
2875 }
2876 page_cache_release(old_page);
2877 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002878 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002879oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002880 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002881oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002882 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002883 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002885
2886unwritable_page:
2887 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002888 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889}
2890
Peter Zijlstra97a89412011-05-24 17:12:04 -07002891static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 unsigned long start_addr, unsigned long end_addr,
2893 struct zap_details *details)
2894{
Al Virof5cc4ee2012-03-05 14:14:20 -05002895 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896}
2897
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002898static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 struct zap_details *details)
2900{
2901 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 pgoff_t vba, vea, zba, zea;
2903
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002904 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
2907 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002908 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2910 zba = details->first_index;
2911 if (zba < vba)
2912 zba = vba;
2913 zea = details->last_index;
2914 if (zea > vea)
2915 zea = vea;
2916
Peter Zijlstra97a89412011-05-24 17:12:04 -07002917 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2919 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002920 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 }
2922}
2923
2924static inline void unmap_mapping_range_list(struct list_head *head,
2925 struct zap_details *details)
2926{
2927 struct vm_area_struct *vma;
2928
2929 /*
2930 * In nonlinear VMAs there is no correspondence between virtual address
2931 * offset and file offset. So we must perform an exhaustive search
2932 * across *all* the pages in each nonlinear VMA, not just the pages
2933 * whose virtual address lies outside the file truncation point.
2934 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002935 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002937 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 }
2939}
2940
2941/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002942 * 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 -07002943 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 * @holebegin: byte in first page to unmap, relative to the start of
2945 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002946 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 * must keep the partial page. In contrast, we must get rid of
2948 * partial pages.
2949 * @holelen: size of prospective hole in bytes. This will be rounded
2950 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2951 * end of the file.
2952 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2953 * but 0 when invalidating pagecache, don't throw away private data.
2954 */
2955void unmap_mapping_range(struct address_space *mapping,
2956 loff_t const holebegin, loff_t const holelen, int even_cows)
2957{
2958 struct zap_details details;
2959 pgoff_t hba = holebegin >> PAGE_SHIFT;
2960 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2961
2962 /* Check for overflow. */
2963 if (sizeof(holelen) > sizeof(hlen)) {
2964 long long holeend =
2965 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2966 if (holeend & ~(long long)ULONG_MAX)
2967 hlen = ULONG_MAX - hba + 1;
2968 }
2969
2970 details.check_mapping = even_cows? NULL: mapping;
2971 details.nonlinear_vma = NULL;
2972 details.first_index = hba;
2973 details.last_index = hba + hlen - 1;
2974 if (details.last_index < details.first_index)
2975 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976
Peter Zijlstra97a89412011-05-24 17:12:04 -07002977
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002978 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002979 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2981 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2982 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002983 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984}
2985EXPORT_SYMBOL(unmap_mapping_range);
2986
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002988 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2989 * but allow concurrent faults), and pte mapped but not yet locked.
2990 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002992static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2993 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002994 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002996 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002997 struct page *page, *swapcache;
Hugh Dickins65500d22005-10-29 18:15:59 -07002998 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003000 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07003001 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003002 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07003003 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003005 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003006 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07003007
3008 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02003009 if (unlikely(non_swap_entry(entry))) {
3010 if (is_migration_entry(entry)) {
3011 migration_entry_wait(mm, pmd, address);
3012 } else if (is_hwpoison_entry(entry)) {
3013 ret = VM_FAULT_HWPOISON;
3014 } else {
3015 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003016 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02003017 }
Christoph Lameter06972122006-06-23 02:03:35 -07003018 goto out;
3019 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003020 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 page = lookup_swap_cache(entry);
3022 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08003023 page = swapin_readahead(entry,
3024 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 if (!page) {
3026 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003027 * Back out if somebody else faulted in this pte
3028 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003030 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 if (likely(pte_same(*page_table, orig_pte)))
3032 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003033 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07003034 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 }
3036
3037 /* Had to read the page from swap area: Major fault */
3038 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07003039 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003040 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02003041 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01003042 /*
3043 * hwpoisoned dirty swapcache pages are kept for killing
3044 * owner processes (which may be unknown at hwpoison time)
3045 */
Andi Kleend1737fd2009-09-16 11:50:06 +02003046 ret = VM_FAULT_HWPOISON;
3047 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08003048 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02003049 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 }
3051
Hugh Dickins56f31802013-02-22 16:36:10 -08003052 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003053 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07003054
Balbir Singh20a10222007-11-14 17:00:33 -08003055 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003056 if (!locked) {
3057 ret |= VM_FAULT_RETRY;
3058 goto out_release;
3059 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003061 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003062 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3063 * release the swapcache from under us. The page pin, and pte_same
3064 * test below, are not enough to exclude that. Even if it is still
3065 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003066 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003067 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003068 goto out_page;
3069
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003070 page = ksm_might_need_to_copy(page, vma, address);
3071 if (unlikely(!page)) {
3072 ret = VM_FAULT_OOM;
3073 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003074 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08003075 }
3076
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003077 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003078 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003079 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003080 }
3081
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003083 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003085 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003086 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003087 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003088
3089 if (unlikely(!PageUptodate(page))) {
3090 ret = VM_FAULT_SIGBUS;
3091 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 }
3093
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003094 /*
3095 * The page isn't present yet, go ahead with the fault.
3096 *
3097 * Be careful about the sequence of operations here.
3098 * To get its accounting right, reuse_swap_page() must be called
3099 * while the page is counted on swap but not yet in mapcount i.e.
3100 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3101 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003102 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3103 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3104 * in page->private. In this case, a record in swap_cgroup is silently
3105 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003106 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003108 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003109 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003111 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003113 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003114 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003115 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 flush_icache_page(vma, page);
3118 set_pte_at(mm, address, page_table, pte);
Hugh Dickins56f31802013-02-22 16:36:10 -08003119 if (page == swapcache)
Johannes Weineraf347702013-02-22 16:32:20 -08003120 do_page_add_anon_rmap(page, vma, address, exclusive);
Hugh Dickins56f31802013-02-22 16:36:10 -08003121 else /* ksm created a completely new copy */
3122 page_add_new_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003123 /* It's better to call commit-charge after rmap is established */
3124 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003126 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003127 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003128 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003129 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003130 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003131 /*
3132 * Hold the lock to avoid the swap entry to be reused
3133 * until we take the PT lock for the pte_same() check
3134 * (to avoid false positives from pte_same). For
3135 * further safety release the lock after the swap_free
3136 * so that the swap count won't change under a
3137 * parallel locked swapcache.
3138 */
3139 unlock_page(swapcache);
3140 page_cache_release(swapcache);
3141 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003142
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003143 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003144 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3145 if (ret & VM_FAULT_ERROR)
3146 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 goto out;
3148 }
3149
3150 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003151 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003152unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003153 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154out:
3155 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003156out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003157 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003158 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003159out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003160 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003161out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003162 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003163 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003164 unlock_page(swapcache);
3165 page_cache_release(swapcache);
3166 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003167 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168}
3169
3170/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003171 * This is like a special single-page "expand_{down|up}wards()",
3172 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003173 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003174 */
3175static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3176{
3177 address &= PAGE_MASK;
3178 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003179 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003180
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003181 /*
3182 * Is there a mapping abutting this one below?
3183 *
3184 * That's only ok if it's the same stack mapping
3185 * that has gotten split..
3186 */
3187 if (prev && prev->vm_end == address)
3188 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3189
Michal Hockod05f3162011-05-24 17:11:44 -07003190 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003191 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003192 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3193 struct vm_area_struct *next = vma->vm_next;
3194
3195 /* As VM_GROWSDOWN but s/below/above/ */
3196 if (next && next->vm_start == address + PAGE_SIZE)
3197 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3198
3199 expand_upwards(vma, address + PAGE_SIZE);
3200 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003201 return 0;
3202}
3203
3204/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003205 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3206 * but allow concurrent faults), and pte mapped but not yet locked.
3207 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003209static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3210 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003211 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003213 struct page *page;
3214 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216
Linus Torvalds11ac5522010-08-14 11:44:56 -07003217 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003218
Linus Torvalds11ac5522010-08-14 11:44:56 -07003219 /* Check if we need to add a guard page to the stack */
3220 if (check_stack_guard_page(vma, address) < 0)
3221 return VM_FAULT_SIGBUS;
3222
3223 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003224 if (!(flags & FAULT_FLAG_WRITE)) {
3225 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3226 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003227 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003228 if (!pte_none(*page_table))
3229 goto unlock;
3230 goto setpte;
3231 }
3232
Nick Piggin557ed1f2007-10-16 01:24:40 -07003233 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003234 if (unlikely(anon_vma_prepare(vma)))
3235 goto oom;
3236 page = alloc_zeroed_user_highpage_movable(vma, address);
3237 if (!page)
3238 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07003239 /*
3240 * The memory barrier inside __SetPageUptodate makes sure that
3241 * preceeding stores to the page contents become visible before
3242 * the set_pte_at() write.
3243 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08003244 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003246 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003247 goto oom_free_page;
3248
Nick Piggin557ed1f2007-10-16 01:24:40 -07003249 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003250 if (vma->vm_flags & VM_WRITE)
3251 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003252
Nick Piggin557ed1f2007-10-16 01:24:40 -07003253 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003254 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003255 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003256
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003257 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003258 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003259setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003260 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261
3262 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003263 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003264unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003265 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003266 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003267release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003268 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003269 page_cache_release(page);
3270 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003271oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003272 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003273oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274 return VM_FAULT_OOM;
3275}
3276
3277/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003278 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003280 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3281 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 *
3283 * As this is called only for pages that do not currently exist, we
3284 * do not need to flush old virtual caches or the TLB.
3285 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003286 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003287 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003288 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003290static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003291 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003292 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003294 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003295 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003296 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003297 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003300 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003301 struct vm_fault vmf;
3302 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003303 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003304
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003305 /*
3306 * If we do COW later, allocate page befor taking lock_page()
3307 * on the file cache page. This will reduce lock holding time.
3308 */
3309 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3310
3311 if (unlikely(anon_vma_prepare(vma)))
3312 return VM_FAULT_OOM;
3313
3314 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3315 if (!cow_page)
3316 return VM_FAULT_OOM;
3317
3318 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3319 page_cache_release(cow_page);
3320 return VM_FAULT_OOM;
3321 }
3322 } else
3323 cow_page = NULL;
3324
Nick Piggind0217ac2007-07-19 01:47:03 -07003325 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3326 vmf.pgoff = pgoff;
3327 vmf.flags = flags;
3328 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003330 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003331 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3332 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003333 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
Andi Kleena3b947e2009-09-16 11:50:08 +02003335 if (unlikely(PageHWPoison(vmf.page))) {
3336 if (ret & VM_FAULT_LOCKED)
3337 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003338 ret = VM_FAULT_HWPOISON;
3339 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003340 }
3341
Nick Piggind00806b2007-07-19 01:46:57 -07003342 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003343 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003344 * locked.
3345 */
Nick Piggin83c54072007-07-19 01:47:05 -07003346 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003347 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003348 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003349 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003350
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 /*
3352 * Should we do an early C-O-W break?
3353 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003354 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003355 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003356 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003357 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003358 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003359 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003360 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003361 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003362 /*
3363 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003364 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003365 * to become writable
3366 */
Mark Fasheh69676142007-07-19 01:47:00 -07003367 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003368 int tmp;
3369
Mark Fasheh69676142007-07-19 01:47:00 -07003370 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003371 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003372 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3373 if (unlikely(tmp &
3374 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3375 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003376 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003377 }
Nick Pigginb827e492009-04-30 15:08:16 -07003378 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3379 lock_page(page);
3380 if (!page->mapping) {
3381 ret = 0; /* retry the fault */
3382 unlock_page(page);
3383 goto unwritable_page;
3384 }
3385 } else
3386 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003387 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003388 }
3389 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003390
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 }
3392
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003393 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394
3395 /*
3396 * This silly early PAGE_DIRTY setting removes a race
3397 * due to the bad i386 page protection. But it's valid
3398 * for other architectures too.
3399 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003400 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 * an exclusive copy of the page, or this is a shared mapping,
3402 * so we can make it writable and dirty to avoid having to
3403 * handle that later.
3404 */
3405 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003406 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003407 flush_icache_page(vma, page);
3408 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003409 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003412 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003413 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003414 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003415 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003416 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003417 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003418 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003419 get_page(dirty_page);
3420 }
Hugh Dickins42946212005-10-29 18:16:05 -07003421 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003422 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003423
3424 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003425 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003427 if (cow_page)
3428 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003429 if (anon)
3430 page_cache_release(page);
3431 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003432 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 }
3434
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003435 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003436
Nick Pigginb827e492009-04-30 15:08:16 -07003437 if (dirty_page) {
3438 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003439 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003440
3441 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003442 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003443 unlock_page(dirty_page);
3444 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003445 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003446 /*
3447 * Some device drivers do not set page.mapping but still
3448 * dirty their pages
3449 */
3450 balance_dirty_pages_ratelimited(mapping);
3451 }
3452
3453 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003454 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003455 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003456 } else {
3457 unlock_page(vmf.page);
3458 if (anon)
3459 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003460 }
Nick Piggind00806b2007-07-19 01:46:57 -07003461
Nick Piggin83c54072007-07-19 01:47:05 -07003462 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003463
3464unwritable_page:
3465 page_cache_release(page);
3466 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003467uncharge_out:
3468 /* fs's fault handler get error */
3469 if (cow_page) {
3470 mem_cgroup_uncharge_page(cow_page);
3471 page_cache_release(cow_page);
3472 }
3473 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003474}
Nick Piggind00806b2007-07-19 01:46:57 -07003475
Nick Piggin54cb8822007-07-19 01:46:59 -07003476static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3477 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003478 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003479{
3480 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003481 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003482
Hugh Dickins16abfa02007-10-04 16:56:06 +01003483 pte_unmap(page_table);
3484 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003485}
3486
Jes Sorensenf4b81802006-09-27 01:50:10 -07003487/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 * Fault of a previously existing named mapping. Repopulate the pte
3489 * from the encoded file_pte if possible. This enables swappable
3490 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003491 *
3492 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3493 * but allow concurrent faults), and pte mapped but not yet locked.
3494 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003496static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003497 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003498 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499{
Hugh Dickins65500d22005-10-29 18:15:59 -07003500 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003502 flags |= FAULT_FLAG_NONLINEAR;
3503
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003504 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003505 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506
Hugh Dickins2509ef22009-01-06 14:40:10 -08003507 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003508 /*
3509 * Page table corrupted: show pte and kill process.
3510 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003511 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003512 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003513 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003514
3515 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003516 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517}
3518
Mel Gorman9532fec2012-11-15 01:24:32 +00003519int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3520 unsigned long addr, int current_nid)
3521{
3522 get_page(page);
3523
3524 count_vm_numa_event(NUMA_HINT_FAULTS);
3525 if (current_nid == numa_node_id())
3526 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3527
3528 return mpol_misplaced(page, vma, addr);
3529}
3530
Mel Gormand10e63f2012-10-25 14:16:31 +02003531int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3532 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3533{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003534 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003535 spinlock_t *ptl;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003536 int current_nid = -1;
3537 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003538 bool migrated = false;
Mel Gormand10e63f2012-10-25 14:16:31 +02003539
3540 /*
3541 * The "pte" at this point cannot be used safely without
3542 * validation through pte_unmap_same(). It's of NUMA type but
3543 * the pfn may be screwed if the read is non atomic.
3544 *
3545 * ptep_modify_prot_start is not called as this is clearing
3546 * the _PAGE_NUMA bit and it is not really expected that there
3547 * would be concurrent hardware modifications to the PTE.
3548 */
3549 ptl = pte_lockptr(mm, pmd);
3550 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003551 if (unlikely(!pte_same(*ptep, pte))) {
3552 pte_unmap_unlock(ptep, ptl);
3553 goto out;
3554 }
3555
Mel Gormand10e63f2012-10-25 14:16:31 +02003556 pte = pte_mknonnuma(pte);
3557 set_pte_at(mm, addr, ptep, pte);
3558 update_mmu_cache(vma, addr, ptep);
3559
3560 page = vm_normal_page(vma, addr, pte);
3561 if (!page) {
3562 pte_unmap_unlock(ptep, ptl);
3563 return 0;
3564 }
3565
Mel Gorman4daae3b2012-11-02 11:33:45 +00003566 current_nid = page_to_nid(page);
Mel Gorman9532fec2012-11-15 01:24:32 +00003567 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
Mel Gormand10e63f2012-10-25 14:16:31 +02003568 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003569 if (target_nid == -1) {
3570 /*
3571 * Account for the fault against the current node if it not
3572 * being replaced regardless of where the page is located.
3573 */
3574 current_nid = numa_node_id();
3575 put_page(page);
3576 goto out;
3577 }
3578
3579 /* Migrate to the requested node */
Mel Gormanb8593bf2012-11-21 01:18:23 +00003580 migrated = migrate_misplaced_page(page, target_nid);
3581 if (migrated)
Mel Gorman4daae3b2012-11-02 11:33:45 +00003582 current_nid = target_nid;
3583
3584out:
Mel Gorman9532fec2012-11-15 01:24:32 +00003585 if (current_nid != -1)
Mel Gormanb8593bf2012-11-21 01:18:23 +00003586 task_numa_fault(current_nid, 1, migrated);
Mel Gormand10e63f2012-10-25 14:16:31 +02003587 return 0;
3588}
3589
3590/* NUMA hinting page fault entry point for regular pmds */
3591#ifdef CONFIG_NUMA_BALANCING
3592static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3593 unsigned long addr, pmd_t *pmdp)
3594{
3595 pmd_t pmd;
3596 pte_t *pte, *orig_pte;
3597 unsigned long _addr = addr & PMD_MASK;
3598 unsigned long offset;
3599 spinlock_t *ptl;
3600 bool numa = false;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00003601 int local_nid = numa_node_id();
Mel Gormand10e63f2012-10-25 14:16:31 +02003602
3603 spin_lock(&mm->page_table_lock);
3604 pmd = *pmdp;
3605 if (pmd_numa(pmd)) {
3606 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3607 numa = true;
3608 }
3609 spin_unlock(&mm->page_table_lock);
3610
3611 if (!numa)
3612 return 0;
3613
3614 /* we're in a page fault so some vma must be in the range */
3615 BUG_ON(!vma);
3616 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3617 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3618 VM_BUG_ON(offset >= PMD_SIZE);
3619 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3620 pte += offset >> PAGE_SHIFT;
3621 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3622 pte_t pteval = *pte;
3623 struct page *page;
Mel Gorman9532fec2012-11-15 01:24:32 +00003624 int curr_nid = local_nid;
3625 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003626 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02003627 if (!pte_present(pteval))
3628 continue;
3629 if (!pte_numa(pteval))
3630 continue;
3631 if (addr >= vma->vm_end) {
3632 vma = find_vma(mm, addr);
3633 /* there's a pte present so there must be a vma */
3634 BUG_ON(!vma);
3635 BUG_ON(addr < vma->vm_start);
3636 }
3637 if (pte_numa(pteval)) {
3638 pteval = pte_mknonnuma(pteval);
3639 set_pte_at(mm, addr, pte, pteval);
3640 }
3641 page = vm_normal_page(vma, addr, pteval);
3642 if (unlikely(!page))
3643 continue;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003644 /* only check non-shared pages */
3645 if (unlikely(page_mapcount(page) != 1))
3646 continue;
Mel Gorman9532fec2012-11-15 01:24:32 +00003647
3648 /*
3649 * Note that the NUMA fault is later accounted to either
3650 * the node that is currently running or where the page is
3651 * migrated to.
3652 */
3653 curr_nid = local_nid;
3654 target_nid = numa_migrate_prep(page, vma, addr,
3655 page_to_nid(page));
3656 if (target_nid == -1) {
3657 put_page(page);
3658 continue;
3659 }
3660
3661 /* Migrate to the requested node */
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003662 pte_unmap_unlock(pte, ptl);
Mel Gormanb8593bf2012-11-21 01:18:23 +00003663 migrated = migrate_misplaced_page(page, target_nid);
3664 if (migrated)
Mel Gorman9532fec2012-11-15 01:24:32 +00003665 curr_nid = target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003666 task_numa_fault(curr_nid, 1, migrated);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003667
3668 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003669 }
3670 pte_unmap_unlock(orig_pte, ptl);
3671
3672 return 0;
3673}
3674#else
3675static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3676 unsigned long addr, pmd_t *pmdp)
3677{
3678 BUG();
Andrew Mortonb3dd2072012-12-17 15:59:24 -08003679 return 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003680}
3681#endif /* CONFIG_NUMA_BALANCING */
3682
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683/*
3684 * These routines also need to handle stuff like marking pages dirty
3685 * and/or accessed for architectures that don't do it in hardware (most
3686 * RISC architectures). The early dirtying is also good on the i386.
3687 *
3688 * There is also a hook called "update_mmu_cache()" that architectures
3689 * with external mmu caches can use to update those (ie the Sparc or
3690 * PowerPC hashed page tables that act as extended TLBs).
3691 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003692 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3693 * but allow concurrent faults), and pte mapped but not yet locked.
3694 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003696int handle_pte_fault(struct mm_struct *mm,
3697 struct vm_area_struct *vma, unsigned long address,
3698 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699{
3700 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003701 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003703 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003705 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003706 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003707 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003708 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003709 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003710 }
3711 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003712 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003715 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003716 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003717 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003718 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719 }
3720
Mel Gormand10e63f2012-10-25 14:16:31 +02003721 if (pte_numa(entry))
3722 return do_numa_page(mm, vma, address, entry, pte, pmd);
3723
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003724 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003725 spin_lock(ptl);
3726 if (unlikely(!pte_same(*pte, entry)))
3727 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003728 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003730 return do_wp_page(mm, vma, address,
3731 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732 entry = pte_mkdirty(entry);
3733 }
3734 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003735 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003736 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003737 } else {
3738 /*
3739 * This is needed only for protection faults but the arch code
3740 * is not yet telling us if this is a protection fault or not.
3741 * This still avoids useless tlb flushes for .text page faults
3742 * with threads.
3743 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003744 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003745 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003746 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003747unlock:
3748 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003749 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750}
3751
3752/*
3753 * By the time we get here, we already hold the mm semaphore
3754 */
Nick Piggin83c54072007-07-19 01:47:05 -07003755int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003756 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757{
3758 pgd_t *pgd;
3759 pud_t *pud;
3760 pmd_t *pmd;
3761 pte_t *pte;
3762
3763 __set_current_state(TASK_RUNNING);
3764
Christoph Lameterf8891e52006-06-30 01:55:45 -07003765 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003766 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003768 /* do counter updates before entering really critical section. */
3769 check_sync_rss_stat(current);
3770
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003771 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003772 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773
David Rientjes1f1d06c2012-05-29 15:06:23 -07003774retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776 pud = pud_alloc(mm, pgd, address);
3777 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003778 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779 pmd = pmd_alloc(mm, pud, address);
3780 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003781 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003782 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3783 if (!vma->vm_ops)
3784 return do_huge_pmd_anonymous_page(mm, vma, address,
3785 pmd, flags);
3786 } else {
3787 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003788 int ret;
3789
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003790 barrier();
3791 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003792 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3793
Linus Torvaldse53289c2013-01-09 08:36:54 -08003794 /*
3795 * If the pmd is splitting, return and retry the
3796 * the fault. Alternative: wait until the split
3797 * is done, and goto retry.
3798 */
3799 if (pmd_trans_splitting(orig_pmd))
3800 return 0;
3801
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003802 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003803 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003804 orig_pmd, pmd);
3805
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003806 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003807 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3808 orig_pmd);
3809 /*
3810 * If COW results in an oom, the huge pmd will
3811 * have been split, so retry the fault on the
3812 * pte for a smaller charge.
3813 */
3814 if (unlikely(ret & VM_FAULT_OOM))
3815 goto retry;
3816 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003817 } else {
3818 huge_pmd_set_accessed(mm, vma, address, pmd,
3819 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003820 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003821
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003822 return 0;
3823 }
3824 }
3825
Mel Gormand10e63f2012-10-25 14:16:31 +02003826 if (pmd_numa(*pmd))
3827 return do_pmd_numa_page(mm, vma, address, pmd);
3828
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003829 /*
3830 * Use __pte_alloc instead of pte_alloc_map, because we can't
3831 * run pte_offset_map on the pmd, if an huge pmd could
3832 * materialize from under us from a different thread.
3833 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003834 if (unlikely(pmd_none(*pmd)) &&
3835 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003836 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003837 /* if an huge pmd materialized from under us just retry later */
3838 if (unlikely(pmd_trans_huge(*pmd)))
3839 return 0;
3840 /*
3841 * A regular pmd is established and it can't morph into a huge pmd
3842 * from under us anymore at this point because we hold the mmap_sem
3843 * read mode and khugepaged takes it in write mode. So now it's
3844 * safe to run pte_offset_map().
3845 */
3846 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003848 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849}
3850
3851#ifndef __PAGETABLE_PUD_FOLDED
3852/*
3853 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003854 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003856int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003858 pud_t *new = pud_alloc_one(mm, address);
3859 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003860 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
Nick Piggin362a61a2008-05-14 06:37:36 +02003862 smp_wmb(); /* See comment in __pte_alloc */
3863
Hugh Dickins872fec12005-10-29 18:16:21 -07003864 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003865 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003866 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003867 else
3868 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003869 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003870 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871}
3872#endif /* __PAGETABLE_PUD_FOLDED */
3873
3874#ifndef __PAGETABLE_PMD_FOLDED
3875/*
3876 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003877 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003879int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003881 pmd_t *new = pmd_alloc_one(mm, address);
3882 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003883 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884
Nick Piggin362a61a2008-05-14 06:37:36 +02003885 smp_wmb(); /* See comment in __pte_alloc */
3886
Hugh Dickins872fec12005-10-29 18:16:21 -07003887 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003889 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003890 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003891 else
3892 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003894 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003895 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003896 else
3897 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003899 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003900 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901}
3902#endif /* __PAGETABLE_PMD_FOLDED */
3903
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904#if !defined(__HAVE_ARCH_GATE_AREA)
3905
3906#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003907static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
3909static int __init gate_vma_init(void)
3910{
3911 gate_vma.vm_mm = NULL;
3912 gate_vma.vm_start = FIXADDR_USER_START;
3913 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003914 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3915 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003916
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 return 0;
3918}
3919__initcall(gate_vma_init);
3920#endif
3921
Stephen Wilson31db58b2011-03-13 15:49:15 -04003922struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923{
3924#ifdef AT_SYSINFO_EHDR
3925 return &gate_vma;
3926#else
3927 return NULL;
3928#endif
3929}
3930
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003931int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932{
3933#ifdef AT_SYSINFO_EHDR
3934 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3935 return 1;
3936#endif
3937 return 0;
3938}
3939
3940#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003941
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003942static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003943 pte_t **ptepp, spinlock_t **ptlp)
3944{
3945 pgd_t *pgd;
3946 pud_t *pud;
3947 pmd_t *pmd;
3948 pte_t *ptep;
3949
3950 pgd = pgd_offset(mm, address);
3951 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3952 goto out;
3953
3954 pud = pud_offset(pgd, address);
3955 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3956 goto out;
3957
3958 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003959 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003960 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3961 goto out;
3962
3963 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3964 if (pmd_huge(*pmd))
3965 goto out;
3966
3967 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3968 if (!ptep)
3969 goto out;
3970 if (!pte_present(*ptep))
3971 goto unlock;
3972 *ptepp = ptep;
3973 return 0;
3974unlock:
3975 pte_unmap_unlock(ptep, *ptlp);
3976out:
3977 return -EINVAL;
3978}
3979
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003980static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3981 pte_t **ptepp, spinlock_t **ptlp)
3982{
3983 int res;
3984
3985 /* (void) is needed to make gcc happy */
3986 (void) __cond_lock(*ptlp,
3987 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3988 return res;
3989}
3990
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003991/**
3992 * follow_pfn - look up PFN at a user virtual address
3993 * @vma: memory mapping
3994 * @address: user virtual address
3995 * @pfn: location to store found PFN
3996 *
3997 * Only IO mappings and raw PFN mappings are allowed.
3998 *
3999 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4000 */
4001int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4002 unsigned long *pfn)
4003{
4004 int ret = -EINVAL;
4005 spinlock_t *ptl;
4006 pte_t *ptep;
4007
4008 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4009 return ret;
4010
4011 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4012 if (ret)
4013 return ret;
4014 *pfn = pte_pfn(*ptep);
4015 pte_unmap_unlock(ptep, ptl);
4016 return 0;
4017}
4018EXPORT_SYMBOL(follow_pfn);
4019
Rik van Riel28b2ee22008-07-23 21:27:05 -07004020#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004021int follow_phys(struct vm_area_struct *vma,
4022 unsigned long address, unsigned int flags,
4023 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004024{
Johannes Weiner03668a42009-06-16 15:32:34 -07004025 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004026 pte_t *ptep, pte;
4027 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004028
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004029 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4030 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004031
Johannes Weiner03668a42009-06-16 15:32:34 -07004032 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004033 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004034 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07004035
Rik van Riel28b2ee22008-07-23 21:27:05 -07004036 if ((flags & FOLL_WRITE) && !pte_write(pte))
4037 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004038
4039 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07004040 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004041
Johannes Weiner03668a42009-06-16 15:32:34 -07004042 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004043unlock:
4044 pte_unmap_unlock(ptep, ptl);
4045out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004046 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004047}
4048
4049int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4050 void *buf, int len, int write)
4051{
4052 resource_size_t phys_addr;
4053 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08004054 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004055 int offset = addr & (PAGE_SIZE-1);
4056
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004057 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004058 return -EINVAL;
4059
4060 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4061 if (write)
4062 memcpy_toio(maddr + offset, buf, len);
4063 else
4064 memcpy_fromio(buf, maddr + offset, len);
4065 iounmap(maddr);
4066
4067 return len;
4068}
4069#endif
4070
David Howells0ec76a12006-09-27 01:50:15 -07004071/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004072 * Access another process' address space as given in mm. If non-NULL, use the
4073 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004074 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004075static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4076 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004077{
David Howells0ec76a12006-09-27 01:50:15 -07004078 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004079 void *old_buf = buf;
4080
David Howells0ec76a12006-09-27 01:50:15 -07004081 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004082 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004083 while (len) {
4084 int bytes, ret, offset;
4085 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004086 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004087
4088 ret = get_user_pages(tsk, mm, addr, 1,
4089 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004090 if (ret <= 0) {
4091 /*
4092 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4093 * we can access using slightly different code.
4094 */
4095#ifdef CONFIG_HAVE_IOREMAP_PROT
4096 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004097 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004098 break;
4099 if (vma->vm_ops && vma->vm_ops->access)
4100 ret = vma->vm_ops->access(vma, addr, buf,
4101 len, write);
4102 if (ret <= 0)
4103#endif
4104 break;
4105 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004106 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004107 bytes = len;
4108 offset = addr & (PAGE_SIZE-1);
4109 if (bytes > PAGE_SIZE-offset)
4110 bytes = PAGE_SIZE-offset;
4111
4112 maddr = kmap(page);
4113 if (write) {
4114 copy_to_user_page(vma, page, addr,
4115 maddr + offset, buf, bytes);
4116 set_page_dirty_lock(page);
4117 } else {
4118 copy_from_user_page(vma, page, addr,
4119 buf, maddr + offset, bytes);
4120 }
4121 kunmap(page);
4122 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004123 }
David Howells0ec76a12006-09-27 01:50:15 -07004124 len -= bytes;
4125 buf += bytes;
4126 addr += bytes;
4127 }
4128 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004129
4130 return buf - old_buf;
4131}
Andi Kleen03252912008-01-30 13:33:18 +01004132
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004133/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004134 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004135 * @mm: the mm_struct of the target address space
4136 * @addr: start address to access
4137 * @buf: source or destination buffer
4138 * @len: number of bytes to transfer
4139 * @write: whether the access is a write
4140 *
4141 * The caller must hold a reference on @mm.
4142 */
4143int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4144 void *buf, int len, int write)
4145{
4146 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4147}
4148
Andi Kleen03252912008-01-30 13:33:18 +01004149/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004150 * Access another process' address space.
4151 * Source/target buffer must be kernel space,
4152 * Do not walk the page table directly, use get_user_pages
4153 */
4154int access_process_vm(struct task_struct *tsk, unsigned long addr,
4155 void *buf, int len, int write)
4156{
4157 struct mm_struct *mm;
4158 int ret;
4159
4160 mm = get_task_mm(tsk);
4161 if (!mm)
4162 return 0;
4163
4164 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4165 mmput(mm);
4166
4167 return ret;
4168}
4169
Andi Kleen03252912008-01-30 13:33:18 +01004170/*
4171 * Print the name of a VMA.
4172 */
4173void print_vma_addr(char *prefix, unsigned long ip)
4174{
4175 struct mm_struct *mm = current->mm;
4176 struct vm_area_struct *vma;
4177
Ingo Molnare8bff742008-02-13 20:21:06 +01004178 /*
4179 * Do not print if we are in atomic
4180 * contexts (in exception stacks, etc.):
4181 */
4182 if (preempt_count())
4183 return;
4184
Andi Kleen03252912008-01-30 13:33:18 +01004185 down_read(&mm->mmap_sem);
4186 vma = find_vma(mm, ip);
4187 if (vma && vma->vm_file) {
4188 struct file *f = vma->vm_file;
4189 char *buf = (char *)__get_free_page(GFP_KERNEL);
4190 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004191 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004192
Jan Blunckcf28b482008-02-14 19:38:44 -08004193 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004194 if (IS_ERR(p))
4195 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004196 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004197 vma->vm_start,
4198 vma->vm_end - vma->vm_start);
4199 free_page((unsigned long)buf);
4200 }
4201 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004202 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004203}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004204
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004205#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004206void might_fault(void)
4207{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004208 /*
4209 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4210 * holding the mmap_sem, this is safe because kernel memory doesn't
4211 * get paged out, therefore we'll never actually fault, and the
4212 * below annotations will generate false positives.
4213 */
4214 if (segment_eq(get_fs(), KERNEL_DS))
4215 return;
4216
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004217 /*
4218 * it would be nicer only to annotate paths which are not under
4219 * pagefault_disable, however that requires a larger audit and
4220 * providing helpers like get_user_atomic.
4221 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004222 if (in_atomic())
4223 return;
4224
4225 __might_sleep(__FILE__, __LINE__, 0);
4226
4227 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004228 might_lock_read(&current->mm->mmap_sem);
4229}
4230EXPORT_SYMBOL(might_fault);
4231#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004232
4233#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4234static void clear_gigantic_page(struct page *page,
4235 unsigned long addr,
4236 unsigned int pages_per_huge_page)
4237{
4238 int i;
4239 struct page *p = page;
4240
4241 might_sleep();
4242 for (i = 0; i < pages_per_huge_page;
4243 i++, p = mem_map_next(p, page, i)) {
4244 cond_resched();
4245 clear_user_highpage(p, addr + i * PAGE_SIZE);
4246 }
4247}
4248void clear_huge_page(struct page *page,
4249 unsigned long addr, unsigned int pages_per_huge_page)
4250{
4251 int i;
4252
4253 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4254 clear_gigantic_page(page, addr, pages_per_huge_page);
4255 return;
4256 }
4257
4258 might_sleep();
4259 for (i = 0; i < pages_per_huge_page; i++) {
4260 cond_resched();
4261 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4262 }
4263}
4264
4265static void copy_user_gigantic_page(struct page *dst, struct page *src,
4266 unsigned long addr,
4267 struct vm_area_struct *vma,
4268 unsigned int pages_per_huge_page)
4269{
4270 int i;
4271 struct page *dst_base = dst;
4272 struct page *src_base = src;
4273
4274 for (i = 0; i < pages_per_huge_page; ) {
4275 cond_resched();
4276 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4277
4278 i++;
4279 dst = mem_map_next(dst, dst_base, i);
4280 src = mem_map_next(src, src_base, i);
4281 }
4282}
4283
4284void copy_user_huge_page(struct page *dst, struct page *src,
4285 unsigned long addr, struct vm_area_struct *vma,
4286 unsigned int pages_per_huge_page)
4287{
4288 int i;
4289
4290 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4291 copy_user_gigantic_page(dst, src, addr, vma,
4292 pages_per_huge_page);
4293 return;
4294 }
4295
4296 might_sleep();
4297 for (i = 0; i < pages_per_huge_page; i++) {
4298 cond_resched();
4299 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4300 }
4301}
4302#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */