blob: ca00039471152eae75bb8321a129a4edfc6f580e [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 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700212void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700213{
214 tlb->mm = mm;
215
Linus Torvalds2b047252013-08-15 11:42:25 -0700216 /* Is it from 0 to ~0? */
217 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700218 tlb->need_flush_all = 0;
Linus Torvalds2b047252013-08-15 11:42:25 -0700219 tlb->start = start;
220 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700221 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700222 tlb->local.next = NULL;
223 tlb->local.nr = 0;
224 tlb->local.max = ARRAY_SIZE(tlb->__pages);
225 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800226 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700227
228#ifdef CONFIG_HAVE_RCU_TABLE_FREE
229 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800230#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700231}
232
233void tlb_flush_mmu(struct mmu_gather *tlb)
234{
235 struct mmu_gather_batch *batch;
236
237 if (!tlb->need_flush)
238 return;
239 tlb->need_flush = 0;
240 tlb_flush(tlb);
241#ifdef CONFIG_HAVE_RCU_TABLE_FREE
242 tlb_table_flush(tlb);
243#endif
244
Peter Zijlstra9547d012011-05-24 17:12:14 -0700245 for (batch = &tlb->local; batch; batch = batch->next) {
246 free_pages_and_swap_cache(batch->pages, batch->nr);
247 batch->nr = 0;
248 }
249 tlb->active = &tlb->local;
250}
251
252/* tlb_finish_mmu
253 * Called at the end of the shootdown operation to free up any resources
254 * that were required.
255 */
256void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
257{
258 struct mmu_gather_batch *batch, *next;
259
260 tlb_flush_mmu(tlb);
261
262 /* keep the page table cache within bounds */
263 check_pgt_cache();
264
265 for (batch = tlb->local.next; batch; batch = next) {
266 next = batch->next;
267 free_pages((unsigned long)batch, 0);
268 }
269 tlb->local.next = NULL;
270}
271
272/* __tlb_remove_page
273 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
274 * handling the additional races in SMP caused by other CPUs caching valid
275 * mappings in their TLBs. Returns the number of free page slots left.
276 * When out of page slots we must call tlb_flush_mmu().
277 */
278int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
279{
280 struct mmu_gather_batch *batch;
281
Shaohua Lif21760b2012-01-12 17:19:16 -0800282 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700283
Peter Zijlstra9547d012011-05-24 17:12:14 -0700284 batch = tlb->active;
285 batch->pages[batch->nr++] = page;
286 if (batch->nr == batch->max) {
287 if (!tlb_next_batch(tlb))
288 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700289 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700290 }
291 VM_BUG_ON(batch->nr > batch->max);
292
293 return batch->max - batch->nr;
294}
295
296#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800297
Peter Zijlstra26723912011-05-24 17:12:00 -0700298#ifdef CONFIG_HAVE_RCU_TABLE_FREE
299
300/*
301 * See the comment near struct mmu_table_batch.
302 */
303
304static void tlb_remove_table_smp_sync(void *arg)
305{
306 /* Simply deliver the interrupt */
307}
308
309static void tlb_remove_table_one(void *table)
310{
311 /*
312 * This isn't an RCU grace period and hence the page-tables cannot be
313 * assumed to be actually RCU-freed.
314 *
315 * It is however sufficient for software page-table walkers that rely on
316 * IRQ disabling. See the comment near struct mmu_table_batch.
317 */
318 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
319 __tlb_remove_table(table);
320}
321
322static void tlb_remove_table_rcu(struct rcu_head *head)
323{
324 struct mmu_table_batch *batch;
325 int i;
326
327 batch = container_of(head, struct mmu_table_batch, rcu);
328
329 for (i = 0; i < batch->nr; i++)
330 __tlb_remove_table(batch->tables[i]);
331
332 free_page((unsigned long)batch);
333}
334
335void tlb_table_flush(struct mmu_gather *tlb)
336{
337 struct mmu_table_batch **batch = &tlb->batch;
338
339 if (*batch) {
340 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341 *batch = NULL;
342 }
343}
344
345void tlb_remove_table(struct mmu_gather *tlb, void *table)
346{
347 struct mmu_table_batch **batch = &tlb->batch;
348
349 tlb->need_flush = 1;
350
351 /*
352 * When there's less then two users of this mm there cannot be a
353 * concurrent page-table walk.
354 */
355 if (atomic_read(&tlb->mm->mm_users) < 2) {
356 __tlb_remove_table(table);
357 return;
358 }
359
360 if (*batch == NULL) {
361 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
362 if (*batch == NULL) {
363 tlb_remove_table_one(table);
364 return;
365 }
366 (*batch)->nr = 0;
367 }
368 (*batch)->tables[(*batch)->nr++] = table;
369 if ((*batch)->nr == MAX_TABLE_BATCH)
370 tlb_table_flush(tlb);
371}
372
Peter Zijlstra9547d012011-05-24 17:12:14 -0700373#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700374
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 * Note: this doesn't free the actual pages themselves. That
377 * has been handled earlier when unmapping all the memory regions.
378 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000379static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
380 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800382 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700383 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000384 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700385 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
Hugh Dickinse0da3822005-04-19 13:29:15 -0700388static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
389 unsigned long addr, unsigned long end,
390 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391{
392 pmd_t *pmd;
393 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700394 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Hugh Dickinse0da3822005-04-19 13:29:15 -0700396 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 do {
399 next = pmd_addr_end(addr, end);
400 if (pmd_none_or_clear_bad(pmd))
401 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000402 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 } while (pmd++, addr = next, addr != end);
404
Hugh Dickinse0da3822005-04-19 13:29:15 -0700405 start &= PUD_MASK;
406 if (start < floor)
407 return;
408 if (ceiling) {
409 ceiling &= PUD_MASK;
410 if (!ceiling)
411 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700413 if (end - 1 > ceiling - 1)
414 return;
415
416 pmd = pmd_offset(pud, start);
417 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000418 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419}
420
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
422 unsigned long addr, unsigned long end,
423 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424{
425 pud_t *pud;
426 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700427 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
Hugh Dickinse0da3822005-04-19 13:29:15 -0700429 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 do {
432 next = pud_addr_end(addr, end);
433 if (pud_none_or_clear_bad(pud))
434 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700435 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 } while (pud++, addr = next, addr != end);
437
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 start &= PGDIR_MASK;
439 if (start < floor)
440 return;
441 if (ceiling) {
442 ceiling &= PGDIR_MASK;
443 if (!ceiling)
444 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700446 if (end - 1 > ceiling - 1)
447 return;
448
449 pud = pud_offset(pgd, start);
450 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000451 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
454/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 * This function frees user-level page tables of a process.
456 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 * Must be called with pagetable lock held.
458 */
Jan Beulich42b77722008-07-23 21:27:10 -0700459void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700460 unsigned long addr, unsigned long end,
461 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462{
463 pgd_t *pgd;
464 unsigned long next;
465
Hugh Dickinse0da3822005-04-19 13:29:15 -0700466 /*
467 * The next few lines have given us lots of grief...
468 *
469 * Why are we testing PMD* at this top level? Because often
470 * there will be no work to do at all, and we'd prefer not to
471 * go all the way down to the bottom just to discover that.
472 *
473 * Why all these "- 1"s? Because 0 represents both the bottom
474 * of the address space and the top of it (using -1 for the
475 * top wouldn't help much: the masks would do the wrong thing).
476 * The rule is that addr 0 and floor 0 refer to the bottom of
477 * the address space, but end 0 and ceiling 0 refer to the top
478 * Comparisons need to use "end - 1" and "ceiling - 1" (though
479 * that end 0 case should be mythical).
480 *
481 * Wherever addr is brought up or ceiling brought down, we must
482 * be careful to reject "the opposite 0" before it confuses the
483 * subsequent tests. But what about where end is brought down
484 * by PMD_SIZE below? no, end can't go down to 0 there.
485 *
486 * Whereas we round start (addr) and ceiling down, by different
487 * masks at different levels, in order to test whether a table
488 * now has no other vmas using it, so can be freed, we don't
489 * bother to round floor or end up - the tests don't need that.
490 */
491
492 addr &= PMD_MASK;
493 if (addr < floor) {
494 addr += PMD_SIZE;
495 if (!addr)
496 return;
497 }
498 if (ceiling) {
499 ceiling &= PMD_MASK;
500 if (!ceiling)
501 return;
502 }
503 if (end - 1 > ceiling - 1)
504 end -= PMD_SIZE;
505 if (addr > end - 1)
506 return;
507
Jan Beulich42b77722008-07-23 21:27:10 -0700508 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 do {
510 next = pgd_addr_end(addr, end);
511 if (pgd_none_or_clear_bad(pgd))
512 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700513 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700515}
516
Jan Beulich42b77722008-07-23 21:27:10 -0700517void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700518 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700519{
520 while (vma) {
521 struct vm_area_struct *next = vma->vm_next;
522 unsigned long addr = vma->vm_start;
523
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700524 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000525 * Hide vma from rmap and truncate_pagecache before freeing
526 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700527 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800528 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700529 unlink_file_vma(vma);
530
David Gibson9da61ae2006-03-22 00:08:57 -0800531 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700532 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700533 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700534 } else {
535 /*
536 * Optimization: gather nearby vmas into one call down
537 */
538 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800539 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700540 vma = next;
541 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800542 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700543 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700544 }
545 free_pgd_range(tlb, addr, vma->vm_end,
546 floor, next? next->vm_start: ceiling);
547 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700548 vma = next;
549 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550}
551
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800552int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
553 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800555 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800556 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700557 if (!new)
558 return -ENOMEM;
559
Nick Piggin362a61a2008-05-14 06:37:36 +0200560 /*
561 * Ensure all pte setup (eg. pte page lock and page clearing) are
562 * visible before the pte is made visible to other CPUs by being
563 * put into page tables.
564 *
565 * The other side of the story is the pointer chasing in the page
566 * table walking code (when walking the page table without locking;
567 * ie. most of the time). Fortunately, these data accesses consist
568 * of a chain of data-dependent loads, meaning most CPUs (alpha
569 * being the notable exception) will already guarantee loads are
570 * seen in-order. See the alpha page table accessors for the
571 * smp_read_barrier_depends() barriers in page table walking code.
572 */
573 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
574
Hugh Dickinsc74df322005-10-29 18:16:23 -0700575 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800576 wait_split_huge_page = 0;
577 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800580 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800581 } else if (unlikely(pmd_trans_splitting(*pmd)))
582 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700583 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800584 if (new)
585 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800586 if (wait_split_huge_page)
587 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700588 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589}
590
Hugh Dickins1bb36302005-10-29 18:16:22 -0700591int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700593 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
594 if (!new)
595 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596
Nick Piggin362a61a2008-05-14 06:37:36 +0200597 smp_wmb(); /* See comment in __pte_alloc */
598
Hugh Dickins1bb36302005-10-29 18:16:22 -0700599 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800600 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700601 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800602 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800603 } else
604 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700605 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800606 if (new)
607 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700608 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609}
610
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800611static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700612{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800613 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
614}
615
616static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
617{
618 int i;
619
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800620 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700621 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800622 for (i = 0; i < NR_MM_COUNTERS; i++)
623 if (rss[i])
624 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700625}
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800628 * This function is called to print an error when a bad pte
629 * is found. For example, we might have a PFN-mapped pte in
630 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700631 *
632 * The calling function must still handle the error.
633 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800634static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
635 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700636{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800637 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
638 pud_t *pud = pud_offset(pgd, addr);
639 pmd_t *pmd = pmd_offset(pud, addr);
640 struct address_space *mapping;
641 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800642 static unsigned long resume;
643 static unsigned long nr_shown;
644 static unsigned long nr_unshown;
645
646 /*
647 * Allow a burst of 60 reports, then keep quiet for that minute;
648 * or allow a steady drip of one report per second.
649 */
650 if (nr_shown == 60) {
651 if (time_before(jiffies, resume)) {
652 nr_unshown++;
653 return;
654 }
655 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800656 printk(KERN_ALERT
657 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800658 nr_unshown);
659 nr_unshown = 0;
660 }
661 nr_shown = 0;
662 }
663 if (nr_shown++ == 0)
664 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800665
666 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
667 index = linear_page_index(vma, addr);
668
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800669 printk(KERN_ALERT
670 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800671 current->comm,
672 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800673 if (page)
674 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800675 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800676 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
677 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
678 /*
679 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
680 */
681 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800682 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
683 vma->vm_ops->fault);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800684 if (vma->vm_file && vma->vm_file->f_op)
Joe Perches071361d2012-12-12 10:19:12 -0800685 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
686 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700687 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030688 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700689}
690
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700691static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800692{
693 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
694}
695
Nick Pigginb5810032005-10-29 18:16:12 -0700696/*
Nick Piggin7e675132008-04-28 02:13:00 -0700697 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800698 *
Nick Piggin7e675132008-04-28 02:13:00 -0700699 * "Special" mappings do not wish to be associated with a "struct page" (either
700 * it doesn't exist, or it exists but they don't want to touch it). In this
701 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700702 *
Nick Piggin7e675132008-04-28 02:13:00 -0700703 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704 * pte bit, in which case this function is trivial. Secondly, an architecture
705 * may not have a spare pte bit, which requires a more complicated scheme,
706 * described below.
707 *
708 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709 * special mapping (even if there are underlying and valid "struct pages").
710 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800711 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700712 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700714 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800716 *
717 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
718 *
Nick Piggin7e675132008-04-28 02:13:00 -0700719 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700720 *
Nick Piggin7e675132008-04-28 02:13:00 -0700721 * This restricts such mappings to be a linear translation from virtual address
722 * to pfn. To get around this restriction, we allow arbitrary mappings so long
723 * as the vma is not a COW mapping; in that case, we know that all ptes are
724 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700725 *
726 *
Nick Piggin7e675132008-04-28 02:13:00 -0700727 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
728 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730 * page" backing, however the difference is that _all_ pages with a struct
731 * page (that is, those where pfn_valid is true) are refcounted and considered
732 * normal pages by the VM. The disadvantage is that pages are refcounted
733 * (which can be slower and simply not an option for some PFNMAP users). The
734 * advantage is that we don't have to follow the strict linearity rule of
735 * PFNMAP mappings in order to support COWable mappings.
736 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800737 */
Nick Piggin7e675132008-04-28 02:13:00 -0700738#ifdef __HAVE_ARCH_PTE_SPECIAL
739# define HAVE_PTE_SPECIAL 1
740#else
741# define HAVE_PTE_SPECIAL 0
742#endif
743struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
744 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800745{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800746 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700747
748 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800749 if (likely(!pte_special(pte)))
750 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700751 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
752 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700753 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800754 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700755 return NULL;
756 }
757
758 /* !HAVE_PTE_SPECIAL case follows: */
759
Jared Hulbertb379d792008-04-28 02:12:58 -0700760 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
761 if (vma->vm_flags & VM_MIXEDMAP) {
762 if (!pfn_valid(pfn))
763 return NULL;
764 goto out;
765 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700766 unsigned long off;
767 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700768 if (pfn == vma->vm_pgoff + off)
769 return NULL;
770 if (!is_cow_mapping(vma->vm_flags))
771 return NULL;
772 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800773 }
774
Hugh Dickins62eede62009-09-21 17:03:34 -0700775 if (is_zero_pfn(pfn))
776 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800777check_pfn:
778 if (unlikely(pfn > highest_memmap_pfn)) {
779 print_bad_pte(vma, addr, pte, NULL);
780 return NULL;
781 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800782
783 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700784 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700785 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800786 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700787out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800788 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800789}
790
791/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 * copy one vm_area from one task to the other. Assumes the page tables
793 * already present in the new task to be cleared in the whole range
794 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 */
796
Hugh Dickins570a335b2009-12-14 17:58:46 -0800797static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700799 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700800 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801{
Nick Pigginb5810032005-10-29 18:16:12 -0700802 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 pte_t pte = *src_pte;
804 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805
806 /* pte contains position in swap or file, so copy. */
807 if (unlikely(!pte_present(pte))) {
808 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700809 swp_entry_t entry = pte_to_swp_entry(pte);
810
Hugh Dickins570a335b2009-12-14 17:58:46 -0800811 if (swap_duplicate(entry) < 0)
812 return entry.val;
813
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 /* make sure dst_mm is on swapoff's mmlist. */
815 if (unlikely(list_empty(&dst_mm->mmlist))) {
816 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700817 if (list_empty(&dst_mm->mmlist))
818 list_add(&dst_mm->mmlist,
819 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 spin_unlock(&mmlist_lock);
821 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800822 if (likely(!non_swap_entry(entry)))
823 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800824 else if (is_migration_entry(entry)) {
825 page = migration_entry_to_page(entry);
826
827 if (PageAnon(page))
828 rss[MM_ANONPAGES]++;
829 else
830 rss[MM_FILEPAGES]++;
831
832 if (is_write_migration_entry(entry) &&
833 is_cow_mapping(vm_flags)) {
834 /*
835 * COW mappings require pages in both
836 * parent and child to be set to read.
837 */
838 make_migration_entry_read(&entry);
839 pte = swp_entry_to_pte(entry);
840 set_pte_at(src_mm, addr, src_pte, pte);
841 }
Christoph Lameter06972122006-06-23 02:03:35 -0700842 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700844 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 }
846
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 /*
848 * If it's a COW mapping, write protect it both
849 * in the parent and the child
850 */
Linus Torvalds67121172005-12-11 20:38:17 -0800851 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700853 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
855
856 /*
857 * If it's a shared mapping, mark it clean in
858 * the child
859 */
860 if (vm_flags & VM_SHARED)
861 pte = pte_mkclean(pte);
862 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800863
864 page = vm_normal_page(vma, addr, pte);
865 if (page) {
866 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700867 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800868 if (PageAnon(page))
869 rss[MM_ANONPAGES]++;
870 else
871 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800872 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700873
874out_set_pte:
875 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800876 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877}
878
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800879int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
880 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
881 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700883 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700885 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700886 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800887 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800888 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
890again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800891 init_rss_vec(rss);
892
Hugh Dickinsc74df322005-10-29 18:16:23 -0700893 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 if (!dst_pte)
895 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700896 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700897 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700898 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700899 orig_src_pte = src_pte;
900 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700901 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 do {
904 /*
905 * We are holding two locks at this point - either of them
906 * could generate latencies in another task on another CPU.
907 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700908 if (progress >= 32) {
909 progress = 0;
910 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100911 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700912 break;
913 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 if (pte_none(*src_pte)) {
915 progress++;
916 continue;
917 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800918 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
919 vma, addr, rss);
920 if (entry.val)
921 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 progress += 8;
923 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700925 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700926 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700927 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800928 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700929 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700930 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800931
932 if (entry.val) {
933 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
934 return -ENOMEM;
935 progress = 0;
936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 if (addr != end)
938 goto again;
939 return 0;
940}
941
942static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
943 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
944 unsigned long addr, unsigned long end)
945{
946 pmd_t *src_pmd, *dst_pmd;
947 unsigned long next;
948
949 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
950 if (!dst_pmd)
951 return -ENOMEM;
952 src_pmd = pmd_offset(src_pud, addr);
953 do {
954 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800955 if (pmd_trans_huge(*src_pmd)) {
956 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800957 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800958 err = copy_huge_pmd(dst_mm, src_mm,
959 dst_pmd, src_pmd, addr, vma);
960 if (err == -ENOMEM)
961 return -ENOMEM;
962 if (!err)
963 continue;
964 /* fall through */
965 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 if (pmd_none_or_clear_bad(src_pmd))
967 continue;
968 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
969 vma, addr, next))
970 return -ENOMEM;
971 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
972 return 0;
973}
974
975static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
976 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
977 unsigned long addr, unsigned long end)
978{
979 pud_t *src_pud, *dst_pud;
980 unsigned long next;
981
982 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
983 if (!dst_pud)
984 return -ENOMEM;
985 src_pud = pud_offset(src_pgd, addr);
986 do {
987 next = pud_addr_end(addr, end);
988 if (pud_none_or_clear_bad(src_pud))
989 continue;
990 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
991 vma, addr, next))
992 return -ENOMEM;
993 } while (dst_pud++, src_pud++, addr = next, addr != end);
994 return 0;
995}
996
997int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
998 struct vm_area_struct *vma)
999{
1000 pgd_t *src_pgd, *dst_pgd;
1001 unsigned long next;
1002 unsigned long addr = vma->vm_start;
1003 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001004 unsigned long mmun_start; /* For mmu_notifiers */
1005 unsigned long mmun_end; /* For mmu_notifiers */
1006 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001007 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008
Nick Piggind9928952005-08-28 16:49:11 +10001009 /*
1010 * Don't copy ptes where a page fault will fill them correctly.
1011 * Fork becomes much lighter when there are big shared or private
1012 * readonly mappings. The tradeoff is that copy_page_range is more
1013 * efficient than faulting.
1014 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001015 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1016 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001017 if (!vma->anon_vma)
1018 return 0;
1019 }
1020
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 if (is_vm_hugetlb_page(vma))
1022 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1023
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001024 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001025 /*
1026 * We do not free on error cases below as remove_vma
1027 * gets called on error from higher level routine
1028 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001029 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001030 if (ret)
1031 return ret;
1032 }
1033
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001034 /*
1035 * We need to invalidate the secondary MMU mappings only when
1036 * there could be a permission downgrade on the ptes of the
1037 * parent mm. And a permission downgrade will only happen if
1038 * is_cow_mapping() returns true.
1039 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001040 is_cow = is_cow_mapping(vma->vm_flags);
1041 mmun_start = addr;
1042 mmun_end = end;
1043 if (is_cow)
1044 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1045 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001046
1047 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 dst_pgd = pgd_offset(dst_mm, addr);
1049 src_pgd = pgd_offset(src_mm, addr);
1050 do {
1051 next = pgd_addr_end(addr, end);
1052 if (pgd_none_or_clear_bad(src_pgd))
1053 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001054 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1055 vma, addr, next))) {
1056 ret = -ENOMEM;
1057 break;
1058 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001060
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001061 if (is_cow)
1062 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001063 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064}
1065
Robin Holt51c6f662005-11-13 16:06:42 -08001066static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001067 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001069 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070{
Nick Pigginb5810032005-10-29 18:16:12 -07001071 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001072 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001073 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001074 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001075 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001076 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001077
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001078again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001079 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001080 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1081 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001082 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 do {
1084 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001085 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001087 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001088
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001090 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001091
Linus Torvalds6aab3412005-11-28 14:34:23 -08001092 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 if (unlikely(details) && page) {
1094 /*
1095 * unmap_shared_mapping_pages() wants to
1096 * invalidate cache without truncating:
1097 * unmap shared but keep private pages.
1098 */
1099 if (details->check_mapping &&
1100 details->check_mapping != page->mapping)
1101 continue;
1102 /*
1103 * Each page->index must be checked when
1104 * invalidating or truncating nonlinear.
1105 */
1106 if (details->nonlinear_vma &&
1107 (page->index < details->first_index ||
1108 page->index > details->last_index))
1109 continue;
1110 }
Nick Pigginb5810032005-10-29 18:16:12 -07001111 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001112 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 tlb_remove_tlb_entry(tlb, pte, addr);
1114 if (unlikely(!page))
1115 continue;
1116 if (unlikely(details) && details->nonlinear_vma
1117 && linear_page_index(details->nonlinear_vma,
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001118 addr) != page->index) {
1119 pte_t ptfile = pgoff_to_pte(page->index);
1120 if (pte_soft_dirty(ptent))
1121 pte_file_mksoft_dirty(ptfile);
1122 set_pte_at(mm, addr, pte, ptfile);
1123 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001125 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001126 else {
1127 if (pte_dirty(ptent))
1128 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001129 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001130 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001131 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001132 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001133 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001134 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001135 if (unlikely(page_mapcount(page) < 0))
1136 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001137 force_flush = !__tlb_remove_page(tlb, page);
1138 if (force_flush)
1139 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 continue;
1141 }
1142 /*
1143 * If details->check_mapping, we leave swap entries;
1144 * if details->nonlinear_vma, we leave file entries.
1145 */
1146 if (unlikely(details))
1147 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001148 if (pte_file(ptent)) {
1149 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1150 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001151 } else {
1152 swp_entry_t entry = pte_to_swp_entry(ptent);
1153
1154 if (!non_swap_entry(entry))
1155 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001156 else if (is_migration_entry(entry)) {
1157 struct page *page;
1158
1159 page = migration_entry_to_page(entry);
1160
1161 if (PageAnon(page))
1162 rss[MM_ANONPAGES]--;
1163 else
1164 rss[MM_FILEPAGES]--;
1165 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001166 if (unlikely(!free_swap_and_cache(entry)))
1167 print_bad_pte(vma, addr, ptent, NULL);
1168 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001169 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001170 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001171
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001172 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001173 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001174 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001175
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001176 /*
1177 * mmu_gather ran out of room to batch pages, we break out of
1178 * the PTE lock to avoid doing the potential expensive TLB invalidate
1179 * and page-free while holding it.
1180 */
1181 if (force_flush) {
Linus Torvalds2b047252013-08-15 11:42:25 -07001182 unsigned long old_end;
1183
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001184 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001185
Linus Torvalds2b047252013-08-15 11:42:25 -07001186 /*
1187 * Flush the TLB just for the previous segment,
1188 * then update the range to be the remaining
1189 * TLB range.
1190 */
1191 old_end = tlb->end;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001192 tlb->end = addr;
Linus Torvalds2b047252013-08-15 11:42:25 -07001193
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001194 tlb_flush_mmu(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001195
1196 tlb->start = addr;
1197 tlb->end = old_end;
1198
1199 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001200 goto again;
1201 }
1202
Robin Holt51c6f662005-11-13 16:06:42 -08001203 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204}
1205
Robin Holt51c6f662005-11-13 16:06:42 -08001206static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001207 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001209 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210{
1211 pmd_t *pmd;
1212 unsigned long next;
1213
1214 pmd = pmd_offset(pud, addr);
1215 do {
1216 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001217 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001218 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001219#ifdef CONFIG_DEBUG_VM
1220 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1221 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1222 __func__, addr, end,
1223 vma->vm_start,
1224 vma->vm_end);
1225 BUG();
1226 }
1227#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001228 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001229 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001230 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001231 /* fall through */
1232 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001233 /*
1234 * Here there can be other concurrent MADV_DONTNEED or
1235 * trans huge page faults running, and if the pmd is
1236 * none or trans huge it can change under us. This is
1237 * because MADV_DONTNEED holds the mmap_sem in read
1238 * mode.
1239 */
1240 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1241 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001242 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001243next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001244 cond_resched();
1245 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001246
1247 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248}
1249
Robin Holt51c6f662005-11-13 16:06:42 -08001250static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001251 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001253 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254{
1255 pud_t *pud;
1256 unsigned long next;
1257
1258 pud = pud_offset(pgd, addr);
1259 do {
1260 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001261 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001263 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1264 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001265
1266 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267}
1268
Al Viro038c7aa2012-03-05 13:25:09 -05001269static void unmap_page_range(struct mmu_gather *tlb,
1270 struct vm_area_struct *vma,
1271 unsigned long addr, unsigned long end,
1272 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273{
1274 pgd_t *pgd;
1275 unsigned long next;
1276
1277 if (details && !details->check_mapping && !details->nonlinear_vma)
1278 details = NULL;
1279
1280 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001281 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 tlb_start_vma(tlb, vma);
1283 pgd = pgd_offset(vma->vm_mm, addr);
1284 do {
1285 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001286 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001288 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1289 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001291 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293
Al Virof5cc4ee2012-03-05 14:14:20 -05001294
1295static void unmap_single_vma(struct mmu_gather *tlb,
1296 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001297 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001298 struct zap_details *details)
1299{
1300 unsigned long start = max(vma->vm_start, start_addr);
1301 unsigned long end;
1302
1303 if (start >= vma->vm_end)
1304 return;
1305 end = min(vma->vm_end, end_addr);
1306 if (end <= vma->vm_start)
1307 return;
1308
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301309 if (vma->vm_file)
1310 uprobe_munmap(vma, start, end);
1311
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001312 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001313 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001314
1315 if (start != end) {
1316 if (unlikely(is_vm_hugetlb_page(vma))) {
1317 /*
1318 * It is undesirable to test vma->vm_file as it
1319 * should be non-null for valid hugetlb area.
1320 * However, vm_file will be NULL in the error
1321 * cleanup path of do_mmap_pgoff. When
1322 * hugetlbfs ->mmap method fails,
1323 * do_mmap_pgoff() nullifies vma->vm_file
1324 * before calling this function to clean up.
1325 * Since no pte has actually been setup, it is
1326 * safe to do nothing in this case.
1327 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001328 if (vma->vm_file) {
1329 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001330 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001331 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1332 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001333 } else
1334 unmap_page_range(tlb, vma, start, end, details);
1335 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336}
1337
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338/**
1339 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001340 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 * @vma: the starting vma
1342 * @start_addr: virtual address at which to start unmapping
1343 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001345 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 * Only addresses between `start' and `end' will be unmapped.
1348 *
1349 * The VMA list must be sorted in ascending virtual address order.
1350 *
1351 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1352 * range after unmap_vmas() returns. So the only responsibility here is to
1353 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1354 * drops the lock and schedules.
1355 */
Al Viro6e8bb012012-03-05 13:41:15 -05001356void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001358 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001360 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001362 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001363 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001364 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001365 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366}
1367
1368/**
1369 * zap_page_range - remove user pages in a given range
1370 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001371 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 * @size: number of bytes to zap
1373 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001374 *
1375 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001377void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 unsigned long size, struct zap_details *details)
1379{
1380 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001381 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001382 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001385 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001386 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001387 mmu_notifier_invalidate_range_start(mm, start, end);
1388 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001389 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001390 mmu_notifier_invalidate_range_end(mm, start, end);
1391 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392}
1393
Jack Steinerc627f9c2008-07-29 22:33:53 -07001394/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001395 * zap_page_range_single - remove user pages in a given range
1396 * @vma: vm_area_struct holding the applicable pages
1397 * @address: starting address of pages to zap
1398 * @size: number of bytes to zap
1399 * @details: details of nonlinear truncation or shared cache invalidation
1400 *
1401 * The range must fit into one VMA.
1402 */
1403static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1404 unsigned long size, struct zap_details *details)
1405{
1406 struct mm_struct *mm = vma->vm_mm;
1407 struct mmu_gather tlb;
1408 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001409
1410 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001411 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001412 update_hiwater_rss(mm);
1413 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001414 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001415 mmu_notifier_invalidate_range_end(mm, address, end);
1416 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417}
1418
Jack Steinerc627f9c2008-07-29 22:33:53 -07001419/**
1420 * zap_vma_ptes - remove ptes mapping the vma
1421 * @vma: vm_area_struct holding ptes to be zapped
1422 * @address: starting address of pages to zap
1423 * @size: number of bytes to zap
1424 *
1425 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1426 *
1427 * The entire address range must be fully contained within the vma.
1428 *
1429 * Returns 0 if successful.
1430 */
1431int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1432 unsigned long size)
1433{
1434 if (address < vma->vm_start || address + size > vma->vm_end ||
1435 !(vma->vm_flags & VM_PFNMAP))
1436 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001437 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001438 return 0;
1439}
1440EXPORT_SYMBOL_GPL(zap_vma_ptes);
1441
Johannes Weiner142762b2010-05-24 14:32:39 -07001442/**
Michel Lespinasse240aade2013-02-22 16:35:56 -08001443 * follow_page_mask - look up a page descriptor from a user-virtual address
Johannes Weiner142762b2010-05-24 14:32:39 -07001444 * @vma: vm_area_struct mapping @address
1445 * @address: virtual address to look up
1446 * @flags: flags modifying lookup behaviour
Michel Lespinasse240aade2013-02-22 16:35:56 -08001447 * @page_mask: on output, *page_mask is set according to the size of the page
Johannes Weiner142762b2010-05-24 14:32:39 -07001448 *
1449 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1450 *
1451 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1452 * an error pointer if there is a mapping to something not represented
1453 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 */
Michel Lespinasse240aade2013-02-22 16:35:56 -08001455struct page *follow_page_mask(struct vm_area_struct *vma,
1456 unsigned long address, unsigned int flags,
1457 unsigned int *page_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458{
1459 pgd_t *pgd;
1460 pud_t *pud;
1461 pmd_t *pmd;
1462 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001463 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001465 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
Michel Lespinasse240aade2013-02-22 16:35:56 -08001467 *page_mask = 0;
1468
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001469 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1470 if (!IS_ERR(page)) {
1471 BUG_ON(flags & FOLL_GET);
1472 goto out;
1473 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001475 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 pgd = pgd_offset(mm, address);
1477 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001478 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
1480 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001481 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001482 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001483 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie632a932013-09-11 14:22:04 -07001484 if (flags & FOLL_GET)
1485 goto out;
Andi Kleenceb86872008-07-23 21:27:50 -07001486 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1487 goto out;
1488 }
1489 if (unlikely(pud_bad(*pud)))
1490 goto no_page_table;
1491
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001493 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001494 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001495 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001497 if (flags & FOLL_GET) {
1498 /*
1499 * Refcount on tail pages are not well-defined and
1500 * shouldn't be taken. The caller should handle a NULL
1501 * return when trying to follow tail pages.
1502 */
1503 if (PageHead(page))
1504 get_page(page);
1505 else {
1506 page = NULL;
1507 goto out;
1508 }
1509 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001510 goto out;
1511 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001512 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1513 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001514 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001515 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001516 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001517 goto split_fallthrough;
1518 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001519 spin_lock(&mm->page_table_lock);
1520 if (likely(pmd_trans_huge(*pmd))) {
1521 if (unlikely(pmd_trans_splitting(*pmd))) {
1522 spin_unlock(&mm->page_table_lock);
1523 wait_split_huge_page(vma->anon_vma, pmd);
1524 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001525 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001526 pmd, flags);
1527 spin_unlock(&mm->page_table_lock);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001528 *page_mask = HPAGE_PMD_NR - 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001529 goto out;
1530 }
1531 } else
1532 spin_unlock(&mm->page_table_lock);
1533 /* fall through */
1534 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001535split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001536 if (unlikely(pmd_bad(*pmd)))
1537 goto no_page_table;
1538
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001539 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540
1541 pte = *ptep;
Hugh Dickins5117b3b2013-02-22 16:36:07 -08001542 if (!pte_present(pte)) {
1543 swp_entry_t entry;
1544 /*
1545 * KSM's break_ksm() relies upon recognizing a ksm page
1546 * even while it is being migrated, so for that case we
1547 * need migration_entry_wait().
1548 */
1549 if (likely(!(flags & FOLL_MIGRATION)))
1550 goto no_page;
1551 if (pte_none(pte) || pte_file(pte))
1552 goto no_page;
1553 entry = pte_to_swp_entry(pte);
1554 if (!is_migration_entry(entry))
1555 goto no_page;
1556 pte_unmap_unlock(ptep, ptl);
1557 migration_entry_wait(mm, pmd, address);
1558 goto split_fallthrough;
1559 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001560 if ((flags & FOLL_NUMA) && pte_numa(pte))
1561 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001562 if ((flags & FOLL_WRITE) && !pte_write(pte))
1563 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001564
Linus Torvalds6aab3412005-11-28 14:34:23 -08001565 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001566 if (unlikely(!page)) {
1567 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001568 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001569 goto bad_page;
1570 page = pte_page(pte);
1571 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001572
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001573 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001574 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001575 if (flags & FOLL_TOUCH) {
1576 if ((flags & FOLL_WRITE) &&
1577 !pte_dirty(pte) && !PageDirty(page))
1578 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001579 /*
1580 * pte_mkyoung() would be more correct here, but atomic care
1581 * is needed to avoid losing the dirty bit: it is easier to use
1582 * mark_page_accessed().
1583 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001584 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001586 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001587 /*
1588 * The preliminary mapping check is mainly to avoid the
1589 * pointless overhead of lock_page on the ZERO_PAGE
1590 * which might bounce very badly if there is contention.
1591 *
1592 * If the page is already locked, we don't need to
1593 * handle it now - vmscan will handle it later if and
1594 * when it attempts to reclaim the page.
1595 */
1596 if (page->mapping && trylock_page(page)) {
1597 lru_add_drain(); /* push cached pages to LRU */
1598 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001599 * Because we lock page here, and migration is
1600 * blocked by the pte's page reference, and we
1601 * know the page is still mapped, we don't even
1602 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001603 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001604 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001605 unlock_page(page);
1606 }
1607 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001608unlock:
1609 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001611 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001613bad_page:
1614 pte_unmap_unlock(ptep, ptl);
1615 return ERR_PTR(-EFAULT);
1616
1617no_page:
1618 pte_unmap_unlock(ptep, ptl);
1619 if (!pte_none(pte))
1620 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001621
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001622no_page_table:
1623 /*
1624 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001625 * has touched so far, we don't want to allocate unnecessary pages or
1626 * page tables. Return error instead of NULL to skip handle_mm_fault,
1627 * then get_dump_page() will return NULL to leave a hole in the dump.
1628 * But we can only make this optimization where a hole would surely
1629 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001630 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001631 if ((flags & FOLL_DUMP) &&
1632 (!vma->vm_ops || !vma->vm_ops->fault))
1633 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001634 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635}
1636
Linus Torvalds95042f92011-04-12 14:15:51 -07001637static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1638{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001639 return stack_guard_page_start(vma, addr) ||
1640 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001641}
1642
Huang Ying0014bd92011-01-30 11:15:47 +08001643/**
1644 * __get_user_pages() - pin user pages in memory
1645 * @tsk: task_struct of target task
1646 * @mm: mm_struct of target mm
1647 * @start: starting user address
1648 * @nr_pages: number of pages from start to pin
1649 * @gup_flags: flags modifying pin behaviour
1650 * @pages: array that receives pointers to the pages pinned.
1651 * Should be at least nr_pages long. Or NULL, if caller
1652 * only intends to ensure the pages are faulted in.
1653 * @vmas: array of pointers to vmas corresponding to each page.
1654 * Or NULL if the caller does not require them.
1655 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1656 *
1657 * Returns number of pages pinned. This may be fewer than the number
1658 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1659 * were pinned, returns -errno. Each page returned must be released
1660 * with a put_page() call when it is finished with. vmas will only
1661 * remain valid while mmap_sem is held.
1662 *
1663 * Must be called with mmap_sem held for read or write.
1664 *
1665 * __get_user_pages walks a process's page tables and takes a reference to
1666 * each struct page that each user address corresponds to at a given
1667 * instant. That is, it takes the page that would be accessed if a user
1668 * thread accesses the given user virtual address at that instant.
1669 *
1670 * This does not guarantee that the page exists in the user mappings when
1671 * __get_user_pages returns, and there may even be a completely different
1672 * page there in some cases (eg. if mmapped pagecache has been invalidated
1673 * and subsequently re faulted). However it does guarantee that the page
1674 * won't be freed completely. And mostly callers simply care that the page
1675 * contains data that was valid *at some point in time*. Typically, an IO
1676 * or similar operation cannot guarantee anything stronger anyway because
1677 * locks can't be held over the syscall boundary.
1678 *
1679 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1680 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1681 * appropriate) must be called after the page is finished with, and
1682 * before put_page is called.
1683 *
1684 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1685 * or mmap_sem contention, and if waiting is needed to pin all pages,
1686 * *@nonblocking will be set to 0.
1687 *
1688 * In most cases, get_user_pages or get_user_pages_fast should be used
1689 * instead of __get_user_pages. __get_user_pages should be used only if
1690 * you need some special @gup_flags.
1691 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08001692long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1693 unsigned long start, unsigned long nr_pages,
1694 unsigned int gup_flags, struct page **pages,
1695 struct vm_area_struct **vmas, int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696{
Michel Lespinasse28a35712013-02-22 16:35:55 -08001697 long i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001698 unsigned long vm_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001699 unsigned int page_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700
Michel Lespinasse28a35712013-02-22 16:35:55 -08001701 if (!nr_pages)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001702 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001703
1704 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1705
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 /*
1707 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001708 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001710 vm_flags = (gup_flags & FOLL_WRITE) ?
1711 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1712 vm_flags &= (gup_flags & FOLL_FORCE) ?
1713 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001714
1715 /*
1716 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1717 * would be called on PROT_NONE ranges. We must never invoke
1718 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1719 * page faults would unprotect the PROT_NONE ranges if
1720 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1721 * bitflag. So to avoid that, don't set FOLL_NUMA if
1722 * FOLL_FORCE is set.
1723 */
1724 if (!(gup_flags & FOLL_FORCE))
1725 gup_flags |= FOLL_NUMA;
1726
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 i = 0;
1728
1729 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001730 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731
1732 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001733 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 pgd_t *pgd;
1736 pud_t *pud;
1737 pmd_t *pmd;
1738 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001739
1740 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001741 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 return i ? : -EFAULT;
1743 if (pg > TASK_SIZE)
1744 pgd = pgd_offset_k(pg);
1745 else
1746 pgd = pgd_offset_gate(mm, pg);
1747 BUG_ON(pgd_none(*pgd));
1748 pud = pud_offset(pgd, pg);
1749 BUG_ON(pud_none(*pud));
1750 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001751 if (pmd_none(*pmd))
1752 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001753 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001755 if (pte_none(*pte)) {
1756 pte_unmap(pte);
1757 return i ? : -EFAULT;
1758 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001759 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001761 struct page *page;
1762
Linus Torvalds95042f92011-04-12 14:15:51 -07001763 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001764 if (!page) {
1765 if (!(gup_flags & FOLL_DUMP) &&
1766 is_zero_pfn(pte_pfn(*pte)))
1767 page = pte_page(*pte);
1768 else {
1769 pte_unmap(pte);
1770 return i ? : -EFAULT;
1771 }
1772 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001773 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001774 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 }
1776 pte_unmap(pte);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001777 page_mask = 0;
Linus Torvalds95042f92011-04-12 14:15:51 -07001778 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 }
1780
Nick Pigginb291f002008-10-18 20:26:44 -07001781 if (!vma ||
1782 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001783 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 return i ? : -EFAULT;
1785
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001786 if (is_vm_hugetlb_page(vma)) {
1787 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001788 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001789 continue;
1790 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001791
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001793 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001794 unsigned int foll_flags = gup_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001795 unsigned int page_increm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796
Ethan Solomita462e00c2007-07-15 23:38:16 -07001797 /*
Ying Han47792802009-01-06 14:40:18 -08001798 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001799 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001800 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001801 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001802 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001803
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001804 cond_resched();
Michel Lespinasse240aade2013-02-22 16:35:56 -08001805 while (!(page = follow_page_mask(vma, start,
1806 foll_flags, &page_mask))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001807 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001808 unsigned int fault_flags = 0;
1809
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001810 /* For mlock, just skip the stack guard page. */
1811 if (foll_flags & FOLL_MLOCK) {
1812 if (stack_guard_page(vma, start))
1813 goto next_page;
1814 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001815 if (foll_flags & FOLL_WRITE)
1816 fault_flags |= FAULT_FLAG_WRITE;
1817 if (nonblocking)
1818 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001819 if (foll_flags & FOLL_NOWAIT)
1820 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001821
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001822 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001823 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001824
Nick Piggin83c54072007-07-19 01:47:05 -07001825 if (ret & VM_FAULT_ERROR) {
1826 if (ret & VM_FAULT_OOM)
1827 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001828 if (ret & (VM_FAULT_HWPOISON |
1829 VM_FAULT_HWPOISON_LARGE)) {
1830 if (i)
1831 return i;
1832 else if (gup_flags & FOLL_HWPOISON)
1833 return -EHWPOISON;
1834 else
1835 return -EFAULT;
1836 }
1837 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001838 return i ? i : -EFAULT;
1839 BUG();
1840 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001841
1842 if (tsk) {
1843 if (ret & VM_FAULT_MAJOR)
1844 tsk->maj_flt++;
1845 else
1846 tsk->min_flt++;
1847 }
Nick Piggin83c54072007-07-19 01:47:05 -07001848
Michel Lespinasse53a77062011-01-13 15:46:14 -08001849 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001850 if (nonblocking)
1851 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001852 return i;
1853 }
1854
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001855 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001856 * The VM_FAULT_WRITE bit tells us that
1857 * do_wp_page has broken COW when necessary,
1858 * even if maybe_mkwrite decided not to set
1859 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001860 * page lookups as if they were reads. But only
1861 * do so when looping for pte_write is futile:
1862 * in some cases userspace may also be wanting
1863 * to write to the gotten user page, which a
1864 * read fault here might prevent (a readonly
1865 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001866 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001867 if ((ret & VM_FAULT_WRITE) &&
1868 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001869 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001870
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001871 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001873 if (IS_ERR(page))
1874 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001876 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001877
Russell Kinga6f36be2006-12-30 22:24:19 +00001878 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001879 flush_dcache_page(page);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001880 page_mask = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001882next_page:
Michel Lespinasse240aade2013-02-22 16:35:56 -08001883 if (vmas) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884 vmas[i] = vma;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001885 page_mask = 0;
1886 }
1887 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1888 if (page_increm > nr_pages)
1889 page_increm = nr_pages;
1890 i += page_increm;
1891 start += page_increm * PAGE_SIZE;
1892 nr_pages -= page_increm;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001893 } while (nr_pages && start < vma->vm_end);
1894 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 return i;
1896}
Huang Ying0014bd92011-01-30 11:15:47 +08001897EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001898
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001899/*
1900 * fixup_user_fault() - manually resolve a user page fault
1901 * @tsk: the task_struct to use for page fault accounting, or
1902 * NULL if faults are not to be recorded.
1903 * @mm: mm_struct of target mm
1904 * @address: user address
1905 * @fault_flags:flags to pass down to handle_mm_fault()
1906 *
1907 * This is meant to be called in the specific scenario where for locking reasons
1908 * we try to access user memory in atomic context (within a pagefault_disable()
1909 * section), this returns -EFAULT, and we want to resolve the user fault before
1910 * trying again.
1911 *
1912 * Typically this is meant to be used by the futex code.
1913 *
1914 * The main difference with get_user_pages() is that this function will
1915 * unconditionally call handle_mm_fault() which will in turn perform all the
1916 * necessary SW fixup of the dirty and young bits in the PTE, while
1917 * handle_mm_fault() only guarantees to update these in the struct page.
1918 *
1919 * This is important for some architectures where those bits also gate the
1920 * access permission to the page because they are maintained in software. On
1921 * such architectures, gup() will not be enough to make a subsequent access
1922 * succeed.
1923 *
1924 * This should be called with the mm_sem held for read.
1925 */
1926int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1927 unsigned long address, unsigned int fault_flags)
1928{
1929 struct vm_area_struct *vma;
1930 int ret;
1931
1932 vma = find_extend_vma(mm, address);
1933 if (!vma || address < vma->vm_start)
1934 return -EFAULT;
1935
1936 ret = handle_mm_fault(mm, vma, address, fault_flags);
1937 if (ret & VM_FAULT_ERROR) {
1938 if (ret & VM_FAULT_OOM)
1939 return -ENOMEM;
1940 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1941 return -EHWPOISON;
1942 if (ret & VM_FAULT_SIGBUS)
1943 return -EFAULT;
1944 BUG();
1945 }
1946 if (tsk) {
1947 if (ret & VM_FAULT_MAJOR)
1948 tsk->maj_flt++;
1949 else
1950 tsk->min_flt++;
1951 }
1952 return 0;
1953}
1954
1955/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001956 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001957 * @tsk: the task_struct to use for page fault accounting, or
1958 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001959 * @mm: mm_struct of target mm
1960 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001961 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001962 * @write: whether pages will be written to by the caller
1963 * @force: whether to force write access even if user mapping is
1964 * readonly. This will result in the page being COWed even
1965 * in MAP_SHARED mappings. You do not want this.
1966 * @pages: array that receives pointers to the pages pinned.
1967 * Should be at least nr_pages long. Or NULL, if caller
1968 * only intends to ensure the pages are faulted in.
1969 * @vmas: array of pointers to vmas corresponding to each page.
1970 * Or NULL if the caller does not require them.
1971 *
1972 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001973 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001974 * were pinned, returns -errno. Each page returned must be released
1975 * with a put_page() call when it is finished with. vmas will only
1976 * remain valid while mmap_sem is held.
1977 *
1978 * Must be called with mmap_sem held for read or write.
1979 *
1980 * get_user_pages walks a process's page tables and takes a reference to
1981 * each struct page that each user address corresponds to at a given
1982 * instant. That is, it takes the page that would be accessed if a user
1983 * thread accesses the given user virtual address at that instant.
1984 *
1985 * This does not guarantee that the page exists in the user mappings when
1986 * get_user_pages returns, and there may even be a completely different
1987 * page there in some cases (eg. if mmapped pagecache has been invalidated
1988 * and subsequently re faulted). However it does guarantee that the page
1989 * won't be freed completely. And mostly callers simply care that the page
1990 * contains data that was valid *at some point in time*. Typically, an IO
1991 * or similar operation cannot guarantee anything stronger anyway because
1992 * locks can't be held over the syscall boundary.
1993 *
1994 * If write=0, the page must not be written to. If the page is written to,
1995 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1996 * after the page is finished with, and before put_page is called.
1997 *
1998 * get_user_pages is typically used for fewer-copy IO operations, to get a
1999 * handle on the memory by some means other than accesses via the user virtual
2000 * addresses. The pages may be submitted for DMA to devices or accessed via
2001 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2002 * use the correct cache flushing APIs.
2003 *
2004 * See also get_user_pages_fast, for performance critical applications.
2005 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08002006long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2007 unsigned long start, unsigned long nr_pages, int write,
2008 int force, struct page **pages, struct vm_area_struct **vmas)
Nick Pigginb291f002008-10-18 20:26:44 -07002009{
Hugh Dickins58fa8792009-09-21 17:03:31 -07002010 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07002011
Hugh Dickins58fa8792009-09-21 17:03:31 -07002012 if (pages)
2013 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07002014 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002015 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07002016 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002017 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07002018
Michel Lespinasse53a77062011-01-13 15:46:14 -08002019 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2020 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07002021}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022EXPORT_SYMBOL(get_user_pages);
2023
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002024/**
2025 * get_dump_page() - pin user page in memory while writing it to core dump
2026 * @addr: user address
2027 *
2028 * Returns struct page pointer of user page pinned for dump,
2029 * to be freed afterwards by page_cache_release() or put_page().
2030 *
2031 * Returns NULL on any kind of failure - a hole must then be inserted into
2032 * the corefile, to preserve alignment with its headers; and also returns
2033 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2034 * allowing a hole to be left in the corefile to save diskspace.
2035 *
2036 * Called without mmap_sem, but after all other threads have been killed.
2037 */
2038#ifdef CONFIG_ELF_CORE
2039struct page *get_dump_page(unsigned long addr)
2040{
2041 struct vm_area_struct *vma;
2042 struct page *page;
2043
2044 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002045 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2046 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002047 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002048 flush_cache_page(vma, addr, page_to_pfn(page));
2049 return page;
2050}
2051#endif /* CONFIG_ELF_CORE */
2052
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002053pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002054 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002055{
2056 pgd_t * pgd = pgd_offset(mm, addr);
2057 pud_t * pud = pud_alloc(mm, pgd, addr);
2058 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002059 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002060 if (pmd) {
2061 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002062 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002063 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002064 }
2065 return NULL;
2066}
2067
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002069 * This is the old fallback for page remapping.
2070 *
2071 * For historical reasons, it only allows reserved pages. Only
2072 * old drivers should use this, and they needed to mark their
2073 * pages reserved for the old functions anyway.
2074 */
Nick Piggin423bad62008-04-28 02:13:01 -07002075static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2076 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002077{
Nick Piggin423bad62008-04-28 02:13:01 -07002078 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002079 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002080 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002081 spinlock_t *ptl;
2082
Linus Torvalds238f58d2005-11-29 13:01:56 -08002083 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002084 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002085 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002086 retval = -ENOMEM;
2087 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002088 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002089 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002090 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002091 retval = -EBUSY;
2092 if (!pte_none(*pte))
2093 goto out_unlock;
2094
2095 /* Ok, finally just insert the thing.. */
2096 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002097 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002098 page_add_file_rmap(page);
2099 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2100
2101 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002102 pte_unmap_unlock(pte, ptl);
2103 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002104out_unlock:
2105 pte_unmap_unlock(pte, ptl);
2106out:
2107 return retval;
2108}
2109
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002110/**
2111 * vm_insert_page - insert single page into user vma
2112 * @vma: user vma to map to
2113 * @addr: target user address of this page
2114 * @page: source kernel page
2115 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002116 * This allows drivers to insert individual pages they've allocated
2117 * into a user vma.
2118 *
2119 * The page has to be a nice clean _individual_ kernel allocation.
2120 * If you allocate a compound page, you need to have marked it as
2121 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002122 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002123 *
2124 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2125 * took an arbitrary page protection parameter. This doesn't allow
2126 * that. Your vma protection will have to be set up correctly, which
2127 * means that if you want a shared writable mapping, you'd better
2128 * ask for a shared writable mapping!
2129 *
2130 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002131 *
2132 * Usually this function is called from f_op->mmap() handler
2133 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2134 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2135 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002136 */
Nick Piggin423bad62008-04-28 02:13:01 -07002137int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2138 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002139{
2140 if (addr < vma->vm_start || addr >= vma->vm_end)
2141 return -EFAULT;
2142 if (!page_count(page))
2143 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002144 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2145 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2146 BUG_ON(vma->vm_flags & VM_PFNMAP);
2147 vma->vm_flags |= VM_MIXEDMAP;
2148 }
Nick Piggin423bad62008-04-28 02:13:01 -07002149 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002150}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002151EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002152
Nick Piggin423bad62008-04-28 02:13:01 -07002153static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2154 unsigned long pfn, pgprot_t prot)
2155{
2156 struct mm_struct *mm = vma->vm_mm;
2157 int retval;
2158 pte_t *pte, entry;
2159 spinlock_t *ptl;
2160
2161 retval = -ENOMEM;
2162 pte = get_locked_pte(mm, addr, &ptl);
2163 if (!pte)
2164 goto out;
2165 retval = -EBUSY;
2166 if (!pte_none(*pte))
2167 goto out_unlock;
2168
2169 /* Ok, finally just insert the thing.. */
2170 entry = pte_mkspecial(pfn_pte(pfn, prot));
2171 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002172 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002173
2174 retval = 0;
2175out_unlock:
2176 pte_unmap_unlock(pte, ptl);
2177out:
2178 return retval;
2179}
2180
Nick Piggine0dc0d82007-02-12 00:51:36 -08002181/**
2182 * vm_insert_pfn - insert single pfn into user vma
2183 * @vma: user vma to map to
2184 * @addr: target user address of this page
2185 * @pfn: source kernel pfn
2186 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002187 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002188 * they've allocated into a user vma. Same comments apply.
2189 *
2190 * This function should only be called from a vm_ops->fault handler, and
2191 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002192 *
2193 * vma cannot be a COW mapping.
2194 *
2195 * As this is called only for pages that do not currently exist, we
2196 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002197 */
2198int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002199 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002200{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002201 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002202 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002203 /*
2204 * Technically, architectures with pte_special can avoid all these
2205 * restrictions (same for remap_pfn_range). However we would like
2206 * consistency in testing and feature parity among all, so we should
2207 * try to keep these invariants in place for everybody.
2208 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002209 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2210 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2211 (VM_PFNMAP|VM_MIXEDMAP));
2212 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2213 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002214
Nick Piggin423bad62008-04-28 02:13:01 -07002215 if (addr < vma->vm_start || addr >= vma->vm_end)
2216 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002217 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002218 return -EINVAL;
2219
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002220 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002221
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002222 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002223}
2224EXPORT_SYMBOL(vm_insert_pfn);
2225
Nick Piggin423bad62008-04-28 02:13:01 -07002226int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2227 unsigned long pfn)
2228{
2229 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2230
2231 if (addr < vma->vm_start || addr >= vma->vm_end)
2232 return -EFAULT;
2233
2234 /*
2235 * If we don't have pte special, then we have to use the pfn_valid()
2236 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2237 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002238 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2239 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002240 */
2241 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2242 struct page *page;
2243
2244 page = pfn_to_page(pfn);
2245 return insert_page(vma, addr, page, vma->vm_page_prot);
2246 }
2247 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2248}
2249EXPORT_SYMBOL(vm_insert_mixed);
2250
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002251/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 * maps a range of physical memory into the requested pages. the old
2253 * mappings are removed. any references to nonexistent pages results
2254 * in null mappings (currently treated as "copy-on-access")
2255 */
2256static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2257 unsigned long addr, unsigned long end,
2258 unsigned long pfn, pgprot_t prot)
2259{
2260 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002261 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
Hugh Dickinsc74df322005-10-29 18:16:23 -07002263 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 if (!pte)
2265 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002266 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 do {
2268 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002269 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 pfn++;
2271 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002272 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002273 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 return 0;
2275}
2276
2277static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2278 unsigned long addr, unsigned long end,
2279 unsigned long pfn, pgprot_t prot)
2280{
2281 pmd_t *pmd;
2282 unsigned long next;
2283
2284 pfn -= addr >> PAGE_SHIFT;
2285 pmd = pmd_alloc(mm, pud, addr);
2286 if (!pmd)
2287 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002288 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 do {
2290 next = pmd_addr_end(addr, end);
2291 if (remap_pte_range(mm, pmd, addr, next,
2292 pfn + (addr >> PAGE_SHIFT), prot))
2293 return -ENOMEM;
2294 } while (pmd++, addr = next, addr != end);
2295 return 0;
2296}
2297
2298static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2299 unsigned long addr, unsigned long end,
2300 unsigned long pfn, pgprot_t prot)
2301{
2302 pud_t *pud;
2303 unsigned long next;
2304
2305 pfn -= addr >> PAGE_SHIFT;
2306 pud = pud_alloc(mm, pgd, addr);
2307 if (!pud)
2308 return -ENOMEM;
2309 do {
2310 next = pud_addr_end(addr, end);
2311 if (remap_pmd_range(mm, pud, addr, next,
2312 pfn + (addr >> PAGE_SHIFT), prot))
2313 return -ENOMEM;
2314 } while (pud++, addr = next, addr != end);
2315 return 0;
2316}
2317
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002318/**
2319 * remap_pfn_range - remap kernel memory to userspace
2320 * @vma: user vma to map to
2321 * @addr: target user address to start at
2322 * @pfn: physical address of kernel memory
2323 * @size: size of map area
2324 * @prot: page protection flags for this mapping
2325 *
2326 * Note: this is only safe if the mm semaphore is held when called.
2327 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2329 unsigned long pfn, unsigned long size, pgprot_t prot)
2330{
2331 pgd_t *pgd;
2332 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002333 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 struct mm_struct *mm = vma->vm_mm;
2335 int err;
2336
2337 /*
2338 * Physically remapped pages are special. Tell the
2339 * rest of the world about it:
2340 * VM_IO tells people not to look at these pages
2341 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002342 * VM_PFNMAP tells the core MM that the base pages are just
2343 * raw PFN mappings, and do not have a "struct page" associated
2344 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002345 * VM_DONTEXPAND
2346 * Disable vma merging and expanding with mremap().
2347 * VM_DONTDUMP
2348 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002349 *
2350 * There's a horrible special case to handle copy-on-write
2351 * behaviour that some programs depend on. We mark the "original"
2352 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002353 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002355 if (is_cow_mapping(vma->vm_flags)) {
2356 if (addr != vma->vm_start || end != vma->vm_end)
2357 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002358 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002359 }
2360
2361 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2362 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002363 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002364
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002365 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366
2367 BUG_ON(addr >= end);
2368 pfn -= addr >> PAGE_SHIFT;
2369 pgd = pgd_offset(mm, addr);
2370 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 do {
2372 next = pgd_addr_end(addr, end);
2373 err = remap_pud_range(mm, pgd, addr, next,
2374 pfn + (addr >> PAGE_SHIFT), prot);
2375 if (err)
2376 break;
2377 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002378
2379 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002380 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002381
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 return err;
2383}
2384EXPORT_SYMBOL(remap_pfn_range);
2385
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07002386/**
2387 * vm_iomap_memory - remap memory to userspace
2388 * @vma: user vma to map to
2389 * @start: start of area
2390 * @len: size of area
2391 *
2392 * This is a simplified io_remap_pfn_range() for common driver use. The
2393 * driver just needs to give us the physical memory range to be mapped,
2394 * we'll figure out the rest from the vma information.
2395 *
2396 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2397 * whatever write-combining details or similar.
2398 */
2399int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2400{
2401 unsigned long vm_len, pfn, pages;
2402
2403 /* Check that the physical memory area passed in looks valid */
2404 if (start + len < start)
2405 return -EINVAL;
2406 /*
2407 * You *really* shouldn't map things that aren't page-aligned,
2408 * but we've historically allowed it because IO memory might
2409 * just have smaller alignment.
2410 */
2411 len += start & ~PAGE_MASK;
2412 pfn = start >> PAGE_SHIFT;
2413 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2414 if (pfn + pages < pfn)
2415 return -EINVAL;
2416
2417 /* We start the mapping 'vm_pgoff' pages into the area */
2418 if (vma->vm_pgoff > pages)
2419 return -EINVAL;
2420 pfn += vma->vm_pgoff;
2421 pages -= vma->vm_pgoff;
2422
2423 /* Can we fit all of the mapping? */
2424 vm_len = vma->vm_end - vma->vm_start;
2425 if (vm_len >> PAGE_SHIFT > pages)
2426 return -EINVAL;
2427
2428 /* Ok, let it rip */
2429 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2430}
2431EXPORT_SYMBOL(vm_iomap_memory);
2432
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002433static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2434 unsigned long addr, unsigned long end,
2435 pte_fn_t fn, void *data)
2436{
2437 pte_t *pte;
2438 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002439 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002440 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002441
2442 pte = (mm == &init_mm) ?
2443 pte_alloc_kernel(pmd, addr) :
2444 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2445 if (!pte)
2446 return -ENOMEM;
2447
2448 BUG_ON(pmd_huge(*pmd));
2449
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002450 arch_enter_lazy_mmu_mode();
2451
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002452 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002453
2454 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002455 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002456 if (err)
2457 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002458 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002459
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002460 arch_leave_lazy_mmu_mode();
2461
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002462 if (mm != &init_mm)
2463 pte_unmap_unlock(pte-1, ptl);
2464 return err;
2465}
2466
2467static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2468 unsigned long addr, unsigned long end,
2469 pte_fn_t fn, void *data)
2470{
2471 pmd_t *pmd;
2472 unsigned long next;
2473 int err;
2474
Andi Kleenceb86872008-07-23 21:27:50 -07002475 BUG_ON(pud_huge(*pud));
2476
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002477 pmd = pmd_alloc(mm, pud, addr);
2478 if (!pmd)
2479 return -ENOMEM;
2480 do {
2481 next = pmd_addr_end(addr, end);
2482 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2483 if (err)
2484 break;
2485 } while (pmd++, addr = next, addr != end);
2486 return err;
2487}
2488
2489static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2490 unsigned long addr, unsigned long end,
2491 pte_fn_t fn, void *data)
2492{
2493 pud_t *pud;
2494 unsigned long next;
2495 int err;
2496
2497 pud = pud_alloc(mm, pgd, addr);
2498 if (!pud)
2499 return -ENOMEM;
2500 do {
2501 next = pud_addr_end(addr, end);
2502 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2503 if (err)
2504 break;
2505 } while (pud++, addr = next, addr != end);
2506 return err;
2507}
2508
2509/*
2510 * Scan a region of virtual memory, filling in page tables as necessary
2511 * and calling a provided function on each leaf page table.
2512 */
2513int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2514 unsigned long size, pte_fn_t fn, void *data)
2515{
2516 pgd_t *pgd;
2517 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002518 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002519 int err;
2520
2521 BUG_ON(addr >= end);
2522 pgd = pgd_offset(mm, addr);
2523 do {
2524 next = pgd_addr_end(addr, end);
2525 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2526 if (err)
2527 break;
2528 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002529
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002530 return err;
2531}
2532EXPORT_SYMBOL_GPL(apply_to_page_range);
2533
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002535 * handle_pte_fault chooses page fault handler according to an entry
2536 * which was read non-atomically. Before making any commitment, on
2537 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002538 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002539 * must check under lock before unmapping the pte and proceeding
2540 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002541 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002542 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002543static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002544 pte_t *page_table, pte_t orig_pte)
2545{
2546 int same = 1;
2547#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2548 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002549 spinlock_t *ptl = pte_lockptr(mm, pmd);
2550 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002551 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002552 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002553 }
2554#endif
2555 pte_unmap(page_table);
2556 return same;
2557}
2558
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002559static 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 -08002560{
2561 /*
2562 * If the source page was a PFN mapping, we don't have
2563 * a "struct page" for it. We do a best-effort copy by
2564 * just copying from the original user address. If that
2565 * fails, we just zero-fill it. Live with it.
2566 */
2567 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002568 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002569 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2570
2571 /*
2572 * This really shouldn't fail, because the page is there
2573 * in the page tables. But it might just be unreadable,
2574 * in which case we just give up and fill the result with
2575 * zeroes.
2576 */
2577 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002578 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002579 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002580 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002581 } else
2582 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002583}
2584
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 * This routine handles present pages, when users try to write
2587 * to a shared page. It is done by copying the page to a new address
2588 * and decrementing the shared-page counter for the old page.
2589 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 * Note that this routine assumes that the protection checks have been
2591 * done by the caller (the low-level page fault routine in most cases).
2592 * Thus we can safely just mark it writable once we've done any necessary
2593 * COW.
2594 *
2595 * We also mark the page dirty at this point even though the page will
2596 * change only once the write actually happens. This avoids a few races,
2597 * and potentially makes it more efficient.
2598 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002599 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2600 * but allow concurrent faults), with pte both mapped and locked.
2601 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002603static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2604 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002605 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002606 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002608 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002610 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002611 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002612 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002613 unsigned long mmun_start = 0; /* For mmu_notifiers */
2614 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615
Linus Torvalds6aab3412005-11-28 14:34:23 -08002616 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002617 if (!old_page) {
2618 /*
2619 * VM_MIXEDMAP !pfn_valid() case
2620 *
2621 * We should not cow pages in a shared writeable mapping.
2622 * Just mark the pages writable as we can't do any dirty
2623 * accounting on raw pfn maps.
2624 */
2625 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2626 (VM_WRITE|VM_SHARED))
2627 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002628 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002629 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002631 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002632 * Take out anonymous pages first, anonymous shared vmas are
2633 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002634 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002635 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002636 if (!trylock_page(old_page)) {
2637 page_cache_get(old_page);
2638 pte_unmap_unlock(page_table, ptl);
2639 lock_page(old_page);
2640 page_table = pte_offset_map_lock(mm, pmd, address,
2641 &ptl);
2642 if (!pte_same(*page_table, orig_pte)) {
2643 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002644 goto unlock;
2645 }
2646 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002647 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002648 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002649 /*
2650 * The page is all ours. Move it to our anon_vma so
2651 * the rmap code will not search our parent or siblings.
2652 * Protected against the rmap code by the page lock.
2653 */
2654 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002655 unlock_page(old_page);
2656 goto reuse;
2657 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002658 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002659 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002660 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002661 /*
2662 * Only catch write-faults on shared writable pages,
2663 * read-only shared pages can get COWed by
2664 * get_user_pages(.write=1, .force=1).
2665 */
David Howells9637a5e2006-06-23 02:03:43 -07002666 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002667 struct vm_fault vmf;
2668 int tmp;
2669
2670 vmf.virtual_address = (void __user *)(address &
2671 PAGE_MASK);
2672 vmf.pgoff = old_page->index;
2673 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2674 vmf.page = old_page;
2675
David Howells9637a5e2006-06-23 02:03:43 -07002676 /*
2677 * Notify the address space that the page is about to
2678 * become writable so that it can prohibit this or wait
2679 * for the page to get into an appropriate state.
2680 *
2681 * We do this without the lock held, so that it can
2682 * sleep if it needs to.
2683 */
2684 page_cache_get(old_page);
2685 pte_unmap_unlock(page_table, ptl);
2686
Nick Pigginc2ec1752009-03-31 15:23:21 -07002687 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2688 if (unlikely(tmp &
2689 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2690 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002691 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002692 }
Nick Pigginb827e492009-04-30 15:08:16 -07002693 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2694 lock_page(old_page);
2695 if (!old_page->mapping) {
2696 ret = 0; /* retry the fault */
2697 unlock_page(old_page);
2698 goto unwritable_page;
2699 }
2700 } else
2701 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002702
David Howells9637a5e2006-06-23 02:03:43 -07002703 /*
2704 * Since we dropped the lock we need to revalidate
2705 * the PTE as someone else may have changed it. If
2706 * they did, we just return, as we can count on the
2707 * MMU to tell us if they didn't also make it writable.
2708 */
2709 page_table = pte_offset_map_lock(mm, pmd, address,
2710 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002711 if (!pte_same(*page_table, orig_pte)) {
2712 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002713 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002714 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002715
2716 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002718 dirty_page = old_page;
2719 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002720
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002721reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002722 flush_cache_page(vma, address, pte_pfn(orig_pte));
2723 entry = pte_mkyoung(orig_pte);
2724 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002725 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002726 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002727 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002728 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002729
2730 if (!dirty_page)
2731 return ret;
2732
2733 /*
2734 * Yes, Virginia, this is actually required to prevent a race
2735 * with clear_page_dirty_for_io() from clearing the page dirty
2736 * bit after it clear all dirty ptes, but before a racing
2737 * do_wp_page installs a dirty pte.
2738 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002739 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002740 */
2741 if (!page_mkwrite) {
2742 wait_on_page_locked(dirty_page);
2743 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002744 /* file_update_time outside page_lock */
2745 if (vma->vm_file)
2746 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002747 }
2748 put_page(dirty_page);
2749 if (page_mkwrite) {
2750 struct address_space *mapping = dirty_page->mapping;
2751
2752 set_page_dirty(dirty_page);
2753 unlock_page(dirty_page);
2754 page_cache_release(dirty_page);
2755 if (mapping) {
2756 /*
2757 * Some device drivers do not set page.mapping
2758 * but still dirty their pages
2759 */
2760 balance_dirty_pages_ratelimited(mapping);
2761 }
2762 }
2763
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002764 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767 /*
2768 * Ok, we need to copy. Oh, well..
2769 */
Nick Pigginb5810032005-10-29 18:16:12 -07002770 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002771gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002772 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773
2774 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002775 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002776
Hugh Dickins62eede62009-09-21 17:03:34 -07002777 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002778 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2779 if (!new_page)
2780 goto oom;
2781 } else {
2782 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2783 if (!new_page)
2784 goto oom;
2785 cow_user_page(new_page, old_page, address, vma);
2786 }
2787 __SetPageUptodate(new_page);
2788
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002789 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002790 goto oom_free_new;
2791
Haggai Eran6bdb9132012-10-08 16:33:35 -07002792 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002793 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002794 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2795
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 /*
2797 * Re-check the pte - we dropped the lock
2798 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002799 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002800 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002801 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002802 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002803 dec_mm_counter_fast(mm, MM_FILEPAGES);
2804 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002805 }
2806 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002807 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002808 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002809 entry = mk_pte(new_page, vma->vm_page_prot);
2810 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002811 /*
2812 * Clear the pte entry and flush it first, before updating the
2813 * pte with the new entry. This will avoid a race condition
2814 * seen in the presence of one thread doing SMC and another
2815 * thread doing COW.
2816 */
Izik Eidus828502d2009-09-21 17:01:51 -07002817 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002818 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002819 /*
2820 * We call the notify macro here because, when using secondary
2821 * mmu page tables (such as kvm shadow page tables), we want the
2822 * new page to be mapped directly into the secondary page table.
2823 */
2824 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002825 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002826 if (old_page) {
2827 /*
2828 * Only after switching the pte to the new page may
2829 * we remove the mapcount here. Otherwise another
2830 * process may come and find the rmap count decremented
2831 * before the pte is switched to the new page, and
2832 * "reuse" the old page writing into it while our pte
2833 * here still points into it and can be read by other
2834 * threads.
2835 *
2836 * The critical issue is to order this
2837 * page_remove_rmap with the ptp_clear_flush above.
2838 * Those stores are ordered by (if nothing else,)
2839 * the barrier present in the atomic_add_negative
2840 * in page_remove_rmap.
2841 *
2842 * Then the TLB flush in ptep_clear_flush ensures that
2843 * no process can access the old page before the
2844 * decremented mapcount is visible. And the old page
2845 * cannot be reused until after the decremented
2846 * mapcount is visible. So transitively, TLBs to
2847 * old page will be flushed before it can be reused.
2848 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002849 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002850 }
2851
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 /* Free the old page.. */
2853 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002854 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002855 } else
2856 mem_cgroup_uncharge_page(new_page);
2857
Haggai Eran6bdb9132012-10-08 16:33:35 -07002858 if (new_page)
2859 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002860unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002861 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002862 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002863 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002864 if (old_page) {
2865 /*
2866 * Don't let another task, with possibly unlocked vma,
2867 * keep the mlocked page.
2868 */
2869 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2870 lock_page(old_page); /* LRU manipulation */
2871 munlock_vma_page(old_page);
2872 unlock_page(old_page);
2873 }
2874 page_cache_release(old_page);
2875 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002876 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002877oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002878 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002879oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002880 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002881 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002883
2884unwritable_page:
2885 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002886 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887}
2888
Peter Zijlstra97a89412011-05-24 17:12:04 -07002889static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 unsigned long start_addr, unsigned long end_addr,
2891 struct zap_details *details)
2892{
Al Virof5cc4ee2012-03-05 14:14:20 -05002893 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894}
2895
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002896static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 struct zap_details *details)
2898{
2899 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 pgoff_t vba, vea, zba, zea;
2901
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002902 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904
2905 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002906 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2908 zba = details->first_index;
2909 if (zba < vba)
2910 zba = vba;
2911 zea = details->last_index;
2912 if (zea > vea)
2913 zea = vea;
2914
Peter Zijlstra97a89412011-05-24 17:12:04 -07002915 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2917 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002918 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 }
2920}
2921
2922static inline void unmap_mapping_range_list(struct list_head *head,
2923 struct zap_details *details)
2924{
2925 struct vm_area_struct *vma;
2926
2927 /*
2928 * In nonlinear VMAs there is no correspondence between virtual address
2929 * offset and file offset. So we must perform an exhaustive search
2930 * across *all* the pages in each nonlinear VMA, not just the pages
2931 * whose virtual address lies outside the file truncation point.
2932 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002933 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002935 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 }
2937}
2938
2939/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002940 * 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 -07002941 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 * @holebegin: byte in first page to unmap, relative to the start of
2943 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002944 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 * must keep the partial page. In contrast, we must get rid of
2946 * partial pages.
2947 * @holelen: size of prospective hole in bytes. This will be rounded
2948 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2949 * end of the file.
2950 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2951 * but 0 when invalidating pagecache, don't throw away private data.
2952 */
2953void unmap_mapping_range(struct address_space *mapping,
2954 loff_t const holebegin, loff_t const holelen, int even_cows)
2955{
2956 struct zap_details details;
2957 pgoff_t hba = holebegin >> PAGE_SHIFT;
2958 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2959
2960 /* Check for overflow. */
2961 if (sizeof(holelen) > sizeof(hlen)) {
2962 long long holeend =
2963 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2964 if (holeend & ~(long long)ULONG_MAX)
2965 hlen = ULONG_MAX - hba + 1;
2966 }
2967
2968 details.check_mapping = even_cows? NULL: mapping;
2969 details.nonlinear_vma = NULL;
2970 details.first_index = hba;
2971 details.last_index = hba + hlen - 1;
2972 if (details.last_index < details.first_index)
2973 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974
Peter Zijlstra97a89412011-05-24 17:12:04 -07002975
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002976 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002977 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2979 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2980 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002981 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982}
2983EXPORT_SYMBOL(unmap_mapping_range);
2984
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002986 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2987 * but allow concurrent faults), and pte mapped but not yet locked.
2988 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002990static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2991 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002992 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002994 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002995 struct page *page, *swapcache;
Hugh Dickins65500d22005-10-29 18:15:59 -07002996 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002998 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002999 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003000 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07003001 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003003 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003004 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07003005
3006 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02003007 if (unlikely(non_swap_entry(entry))) {
3008 if (is_migration_entry(entry)) {
3009 migration_entry_wait(mm, pmd, address);
3010 } else if (is_hwpoison_entry(entry)) {
3011 ret = VM_FAULT_HWPOISON;
3012 } else {
3013 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003014 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02003015 }
Christoph Lameter06972122006-06-23 02:03:35 -07003016 goto out;
3017 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003018 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 page = lookup_swap_cache(entry);
3020 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08003021 page = swapin_readahead(entry,
3022 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 if (!page) {
3024 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003025 * Back out if somebody else faulted in this pte
3026 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003028 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 if (likely(pte_same(*page_table, orig_pte)))
3030 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003031 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07003032 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 }
3034
3035 /* Had to read the page from swap area: Major fault */
3036 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07003037 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003038 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02003039 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01003040 /*
3041 * hwpoisoned dirty swapcache pages are kept for killing
3042 * owner processes (which may be unknown at hwpoison time)
3043 */
Andi Kleend1737fd2009-09-16 11:50:06 +02003044 ret = VM_FAULT_HWPOISON;
3045 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08003046 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02003047 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 }
3049
Hugh Dickins56f31802013-02-22 16:36:10 -08003050 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003051 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07003052
Balbir Singh20a10222007-11-14 17:00:33 -08003053 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003054 if (!locked) {
3055 ret |= VM_FAULT_RETRY;
3056 goto out_release;
3057 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003059 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003060 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3061 * release the swapcache from under us. The page pin, and pte_same
3062 * test below, are not enough to exclude that. Even if it is still
3063 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003064 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003065 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003066 goto out_page;
3067
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003068 page = ksm_might_need_to_copy(page, vma, address);
3069 if (unlikely(!page)) {
3070 ret = VM_FAULT_OOM;
3071 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003072 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08003073 }
3074
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003075 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003076 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003077 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003078 }
3079
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003081 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003083 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003084 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003085 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003086
3087 if (unlikely(!PageUptodate(page))) {
3088 ret = VM_FAULT_SIGBUS;
3089 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 }
3091
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003092 /*
3093 * The page isn't present yet, go ahead with the fault.
3094 *
3095 * Be careful about the sequence of operations here.
3096 * To get its accounting right, reuse_swap_page() must be called
3097 * while the page is counted on swap but not yet in mapcount i.e.
3098 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3099 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003100 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3101 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3102 * in page->private. In this case, a record in swap_cgroup is silently
3103 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003104 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003106 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003107 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003109 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003111 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003112 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003113 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07003116 if (pte_swp_soft_dirty(orig_pte))
3117 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 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);
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07003411 else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3412 pte_mksoft_dirty(entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003414 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003415 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003416 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003417 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003418 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003419 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003420 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003421 get_page(dirty_page);
3422 }
Hugh Dickins42946212005-10-29 18:16:05 -07003423 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003424 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003425
3426 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003427 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003429 if (cow_page)
3430 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003431 if (anon)
3432 page_cache_release(page);
3433 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003434 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435 }
3436
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003437 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003438
Nick Pigginb827e492009-04-30 15:08:16 -07003439 if (dirty_page) {
3440 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003441 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003442
3443 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003444 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003445 unlock_page(dirty_page);
3446 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003447 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003448 /*
3449 * Some device drivers do not set page.mapping but still
3450 * dirty their pages
3451 */
3452 balance_dirty_pages_ratelimited(mapping);
3453 }
3454
3455 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003456 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003457 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003458 } else {
3459 unlock_page(vmf.page);
3460 if (anon)
3461 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003462 }
Nick Piggind00806b2007-07-19 01:46:57 -07003463
Nick Piggin83c54072007-07-19 01:47:05 -07003464 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003465
3466unwritable_page:
3467 page_cache_release(page);
3468 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003469uncharge_out:
3470 /* fs's fault handler get error */
3471 if (cow_page) {
3472 mem_cgroup_uncharge_page(cow_page);
3473 page_cache_release(cow_page);
3474 }
3475 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003476}
Nick Piggind00806b2007-07-19 01:46:57 -07003477
Nick Piggin54cb8822007-07-19 01:46:59 -07003478static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3479 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003480 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003481{
3482 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003483 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003484
Hugh Dickins16abfa02007-10-04 16:56:06 +01003485 pte_unmap(page_table);
3486 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003487}
3488
Jes Sorensenf4b81802006-09-27 01:50:10 -07003489/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490 * Fault of a previously existing named mapping. Repopulate the pte
3491 * from the encoded file_pte if possible. This enables swappable
3492 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003493 *
3494 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3495 * but allow concurrent faults), and pte mapped but not yet locked.
3496 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003498static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003499 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003500 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501{
Hugh Dickins65500d22005-10-29 18:15:59 -07003502 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003504 flags |= FAULT_FLAG_NONLINEAR;
3505
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003506 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003507 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508
Hugh Dickins2509ef22009-01-06 14:40:10 -08003509 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003510 /*
3511 * Page table corrupted: show pte and kill process.
3512 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003513 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003514 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003515 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003516
3517 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003518 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519}
3520
Mel Gorman9532fec2012-11-15 01:24:32 +00003521int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3522 unsigned long addr, int current_nid)
3523{
3524 get_page(page);
3525
3526 count_vm_numa_event(NUMA_HINT_FAULTS);
3527 if (current_nid == numa_node_id())
3528 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3529
3530 return mpol_misplaced(page, vma, addr);
3531}
3532
Mel Gormand10e63f2012-10-25 14:16:31 +02003533int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3534 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3535{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003536 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003537 spinlock_t *ptl;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003538 int current_nid = -1;
3539 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003540 bool migrated = false;
Mel Gormand10e63f2012-10-25 14:16:31 +02003541
3542 /*
3543 * The "pte" at this point cannot be used safely without
3544 * validation through pte_unmap_same(). It's of NUMA type but
3545 * the pfn may be screwed if the read is non atomic.
3546 *
3547 * ptep_modify_prot_start is not called as this is clearing
3548 * the _PAGE_NUMA bit and it is not really expected that there
3549 * would be concurrent hardware modifications to the PTE.
3550 */
3551 ptl = pte_lockptr(mm, pmd);
3552 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003553 if (unlikely(!pte_same(*ptep, pte))) {
3554 pte_unmap_unlock(ptep, ptl);
3555 goto out;
3556 }
3557
Mel Gormand10e63f2012-10-25 14:16:31 +02003558 pte = pte_mknonnuma(pte);
3559 set_pte_at(mm, addr, ptep, pte);
3560 update_mmu_cache(vma, addr, ptep);
3561
3562 page = vm_normal_page(vma, addr, pte);
3563 if (!page) {
3564 pte_unmap_unlock(ptep, ptl);
3565 return 0;
3566 }
3567
Mel Gorman4daae3b2012-11-02 11:33:45 +00003568 current_nid = page_to_nid(page);
Mel Gorman9532fec2012-11-15 01:24:32 +00003569 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
Mel Gormand10e63f2012-10-25 14:16:31 +02003570 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003571 if (target_nid == -1) {
3572 /*
3573 * Account for the fault against the current node if it not
3574 * being replaced regardless of where the page is located.
3575 */
3576 current_nid = numa_node_id();
3577 put_page(page);
3578 goto out;
3579 }
3580
3581 /* Migrate to the requested node */
Mel Gormanb8593bf2012-11-21 01:18:23 +00003582 migrated = migrate_misplaced_page(page, target_nid);
3583 if (migrated)
Mel Gorman4daae3b2012-11-02 11:33:45 +00003584 current_nid = target_nid;
3585
3586out:
Mel Gorman9532fec2012-11-15 01:24:32 +00003587 if (current_nid != -1)
Mel Gormanb8593bf2012-11-21 01:18:23 +00003588 task_numa_fault(current_nid, 1, migrated);
Mel Gormand10e63f2012-10-25 14:16:31 +02003589 return 0;
3590}
3591
3592/* NUMA hinting page fault entry point for regular pmds */
3593#ifdef CONFIG_NUMA_BALANCING
3594static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3595 unsigned long addr, pmd_t *pmdp)
3596{
3597 pmd_t pmd;
3598 pte_t *pte, *orig_pte;
3599 unsigned long _addr = addr & PMD_MASK;
3600 unsigned long offset;
3601 spinlock_t *ptl;
3602 bool numa = false;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00003603 int local_nid = numa_node_id();
Mel Gormand10e63f2012-10-25 14:16:31 +02003604
3605 spin_lock(&mm->page_table_lock);
3606 pmd = *pmdp;
3607 if (pmd_numa(pmd)) {
3608 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3609 numa = true;
3610 }
3611 spin_unlock(&mm->page_table_lock);
3612
3613 if (!numa)
3614 return 0;
3615
3616 /* we're in a page fault so some vma must be in the range */
3617 BUG_ON(!vma);
3618 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3619 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3620 VM_BUG_ON(offset >= PMD_SIZE);
3621 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3622 pte += offset >> PAGE_SHIFT;
3623 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3624 pte_t pteval = *pte;
3625 struct page *page;
Mel Gorman9532fec2012-11-15 01:24:32 +00003626 int curr_nid = local_nid;
3627 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003628 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02003629 if (!pte_present(pteval))
3630 continue;
3631 if (!pte_numa(pteval))
3632 continue;
3633 if (addr >= vma->vm_end) {
3634 vma = find_vma(mm, addr);
3635 /* there's a pte present so there must be a vma */
3636 BUG_ON(!vma);
3637 BUG_ON(addr < vma->vm_start);
3638 }
3639 if (pte_numa(pteval)) {
3640 pteval = pte_mknonnuma(pteval);
3641 set_pte_at(mm, addr, pte, pteval);
3642 }
3643 page = vm_normal_page(vma, addr, pteval);
3644 if (unlikely(!page))
3645 continue;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003646 /* only check non-shared pages */
3647 if (unlikely(page_mapcount(page) != 1))
3648 continue;
Mel Gorman9532fec2012-11-15 01:24:32 +00003649
3650 /*
3651 * Note that the NUMA fault is later accounted to either
3652 * the node that is currently running or where the page is
3653 * migrated to.
3654 */
3655 curr_nid = local_nid;
3656 target_nid = numa_migrate_prep(page, vma, addr,
3657 page_to_nid(page));
3658 if (target_nid == -1) {
3659 put_page(page);
3660 continue;
3661 }
3662
3663 /* Migrate to the requested node */
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003664 pte_unmap_unlock(pte, ptl);
Mel Gormanb8593bf2012-11-21 01:18:23 +00003665 migrated = migrate_misplaced_page(page, target_nid);
3666 if (migrated)
Mel Gorman9532fec2012-11-15 01:24:32 +00003667 curr_nid = target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003668 task_numa_fault(curr_nid, 1, migrated);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003669
3670 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003671 }
3672 pte_unmap_unlock(orig_pte, ptl);
3673
3674 return 0;
3675}
3676#else
3677static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3678 unsigned long addr, pmd_t *pmdp)
3679{
3680 BUG();
Andrew Mortonb3dd2072012-12-17 15:59:24 -08003681 return 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003682}
3683#endif /* CONFIG_NUMA_BALANCING */
3684
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685/*
3686 * These routines also need to handle stuff like marking pages dirty
3687 * and/or accessed for architectures that don't do it in hardware (most
3688 * RISC architectures). The early dirtying is also good on the i386.
3689 *
3690 * There is also a hook called "update_mmu_cache()" that architectures
3691 * with external mmu caches can use to update those (ie the Sparc or
3692 * PowerPC hashed page tables that act as extended TLBs).
3693 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003694 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3695 * but allow concurrent faults), and pte mapped but not yet locked.
3696 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003698static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003699 struct vm_area_struct *vma, unsigned long address,
3700 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701{
3702 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003703 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003705 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003707 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003708 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003709 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003710 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003711 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003712 }
3713 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003714 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003715 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003717 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003718 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003719 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003720 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 }
3722
Mel Gormand10e63f2012-10-25 14:16:31 +02003723 if (pte_numa(entry))
3724 return do_numa_page(mm, vma, address, entry, pte, pmd);
3725
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003726 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003727 spin_lock(ptl);
3728 if (unlikely(!pte_same(*pte, entry)))
3729 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003730 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003732 return do_wp_page(mm, vma, address,
3733 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 entry = pte_mkdirty(entry);
3735 }
3736 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003737 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003738 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003739 } else {
3740 /*
3741 * This is needed only for protection faults but the arch code
3742 * is not yet telling us if this is a protection fault or not.
3743 * This still avoids useless tlb flushes for .text page faults
3744 * with threads.
3745 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003746 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003747 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003748 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003749unlock:
3750 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003751 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752}
3753
3754/*
3755 * By the time we get here, we already hold the mm semaphore
3756 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003757static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3758 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759{
3760 pgd_t *pgd;
3761 pud_t *pud;
3762 pmd_t *pmd;
3763 pte_t *pte;
3764
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003765 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003766 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767
David Rientjes1f1d06c2012-05-29 15:06:23 -07003768retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003770 pud = pud_alloc(mm, pgd, address);
3771 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003772 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773 pmd = pmd_alloc(mm, pud, address);
3774 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003775 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003776 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003777 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003778 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003779 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3780 pmd, flags);
3781 if (!(ret & VM_FAULT_FALLBACK))
3782 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003783 } else {
3784 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003785 int ret;
3786
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003787 barrier();
3788 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003789 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3790
Linus Torvaldse53289c2013-01-09 08:36:54 -08003791 /*
3792 * If the pmd is splitting, return and retry the
3793 * the fault. Alternative: wait until the split
3794 * is done, and goto retry.
3795 */
3796 if (pmd_trans_splitting(orig_pmd))
3797 return 0;
3798
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003799 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003800 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003801 orig_pmd, pmd);
3802
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003803 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003804 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3805 orig_pmd);
3806 /*
3807 * If COW results in an oom, the huge pmd will
3808 * have been split, so retry the fault on the
3809 * pte for a smaller charge.
3810 */
3811 if (unlikely(ret & VM_FAULT_OOM))
3812 goto retry;
3813 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003814 } else {
3815 huge_pmd_set_accessed(mm, vma, address, pmd,
3816 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003817 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003818
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003819 return 0;
3820 }
3821 }
3822
Mel Gormand10e63f2012-10-25 14:16:31 +02003823 if (pmd_numa(*pmd))
3824 return do_pmd_numa_page(mm, vma, address, pmd);
3825
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003826 /*
3827 * Use __pte_alloc instead of pte_alloc_map, because we can't
3828 * run pte_offset_map on the pmd, if an huge pmd could
3829 * materialize from under us from a different thread.
3830 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003831 if (unlikely(pmd_none(*pmd)) &&
3832 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003833 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003834 /* if an huge pmd materialized from under us just retry later */
3835 if (unlikely(pmd_trans_huge(*pmd)))
3836 return 0;
3837 /*
3838 * A regular pmd is established and it can't morph into a huge pmd
3839 * from under us anymore at this point because we hold the mmap_sem
3840 * read mode and khugepaged takes it in write mode. So now it's
3841 * safe to run pte_offset_map().
3842 */
3843 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003845 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846}
3847
Johannes Weiner519e5242013-09-12 15:13:42 -07003848int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3849 unsigned long address, unsigned int flags)
3850{
3851 int ret;
3852
3853 __set_current_state(TASK_RUNNING);
3854
3855 count_vm_event(PGFAULT);
3856 mem_cgroup_count_vm_event(mm, PGFAULT);
3857
3858 /* do counter updates before entering really critical section. */
3859 check_sync_rss_stat(current);
3860
3861 /*
3862 * Enable the memcg OOM handling for faults triggered in user
3863 * space. Kernel faults are handled more gracefully.
3864 */
3865 if (flags & FAULT_FLAG_USER)
3866 mem_cgroup_enable_oom();
3867
3868 ret = __handle_mm_fault(mm, vma, address, flags);
3869
3870 if (flags & FAULT_FLAG_USER)
3871 mem_cgroup_disable_oom();
3872
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003873 if (WARN_ON(task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)))
3874 mem_cgroup_oom_synchronize();
3875
Johannes Weiner519e5242013-09-12 15:13:42 -07003876 return ret;
3877}
3878
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879#ifndef __PAGETABLE_PUD_FOLDED
3880/*
3881 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003882 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003884int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003886 pud_t *new = pud_alloc_one(mm, address);
3887 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003888 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889
Nick Piggin362a61a2008-05-14 06:37:36 +02003890 smp_wmb(); /* See comment in __pte_alloc */
3891
Hugh Dickins872fec12005-10-29 18:16:21 -07003892 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003893 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003894 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003895 else
3896 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003897 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003898 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899}
3900#endif /* __PAGETABLE_PUD_FOLDED */
3901
3902#ifndef __PAGETABLE_PMD_FOLDED
3903/*
3904 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003905 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003907int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003909 pmd_t *new = pmd_alloc_one(mm, address);
3910 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003911 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912
Nick Piggin362a61a2008-05-14 06:37:36 +02003913 smp_wmb(); /* See comment in __pte_alloc */
3914
Hugh Dickins872fec12005-10-29 18:16:21 -07003915 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003917 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003918 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003919 else
3920 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003922 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003923 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003924 else
3925 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003927 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003928 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929}
3930#endif /* __PAGETABLE_PMD_FOLDED */
3931
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932#if !defined(__HAVE_ARCH_GATE_AREA)
3933
3934#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003935static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936
3937static int __init gate_vma_init(void)
3938{
3939 gate_vma.vm_mm = NULL;
3940 gate_vma.vm_start = FIXADDR_USER_START;
3941 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003942 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3943 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003944
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 return 0;
3946}
3947__initcall(gate_vma_init);
3948#endif
3949
Stephen Wilson31db58b2011-03-13 15:49:15 -04003950struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951{
3952#ifdef AT_SYSINFO_EHDR
3953 return &gate_vma;
3954#else
3955 return NULL;
3956#endif
3957}
3958
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003959int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960{
3961#ifdef AT_SYSINFO_EHDR
3962 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3963 return 1;
3964#endif
3965 return 0;
3966}
3967
3968#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003969
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003970static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003971 pte_t **ptepp, spinlock_t **ptlp)
3972{
3973 pgd_t *pgd;
3974 pud_t *pud;
3975 pmd_t *pmd;
3976 pte_t *ptep;
3977
3978 pgd = pgd_offset(mm, address);
3979 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3980 goto out;
3981
3982 pud = pud_offset(pgd, address);
3983 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3984 goto out;
3985
3986 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003987 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003988 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3989 goto out;
3990
3991 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3992 if (pmd_huge(*pmd))
3993 goto out;
3994
3995 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3996 if (!ptep)
3997 goto out;
3998 if (!pte_present(*ptep))
3999 goto unlock;
4000 *ptepp = ptep;
4001 return 0;
4002unlock:
4003 pte_unmap_unlock(ptep, *ptlp);
4004out:
4005 return -EINVAL;
4006}
4007
Namhyung Kim1b36ba82010-10-26 14:22:00 -07004008static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4009 pte_t **ptepp, spinlock_t **ptlp)
4010{
4011 int res;
4012
4013 /* (void) is needed to make gcc happy */
4014 (void) __cond_lock(*ptlp,
4015 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4016 return res;
4017}
4018
Johannes Weiner3b6748e2009-06-16 15:32:35 -07004019/**
4020 * follow_pfn - look up PFN at a user virtual address
4021 * @vma: memory mapping
4022 * @address: user virtual address
4023 * @pfn: location to store found PFN
4024 *
4025 * Only IO mappings and raw PFN mappings are allowed.
4026 *
4027 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4028 */
4029int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4030 unsigned long *pfn)
4031{
4032 int ret = -EINVAL;
4033 spinlock_t *ptl;
4034 pte_t *ptep;
4035
4036 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4037 return ret;
4038
4039 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4040 if (ret)
4041 return ret;
4042 *pfn = pte_pfn(*ptep);
4043 pte_unmap_unlock(ptep, ptl);
4044 return 0;
4045}
4046EXPORT_SYMBOL(follow_pfn);
4047
Rik van Riel28b2ee22008-07-23 21:27:05 -07004048#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004049int follow_phys(struct vm_area_struct *vma,
4050 unsigned long address, unsigned int flags,
4051 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004052{
Johannes Weiner03668a42009-06-16 15:32:34 -07004053 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004054 pte_t *ptep, pte;
4055 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004056
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004057 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4058 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004059
Johannes Weiner03668a42009-06-16 15:32:34 -07004060 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004061 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004062 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07004063
Rik van Riel28b2ee22008-07-23 21:27:05 -07004064 if ((flags & FOLL_WRITE) && !pte_write(pte))
4065 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004066
4067 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07004068 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004069
Johannes Weiner03668a42009-06-16 15:32:34 -07004070 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004071unlock:
4072 pte_unmap_unlock(ptep, ptl);
4073out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004074 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004075}
4076
4077int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4078 void *buf, int len, int write)
4079{
4080 resource_size_t phys_addr;
4081 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08004082 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004083 int offset = addr & (PAGE_SIZE-1);
4084
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004085 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004086 return -EINVAL;
4087
4088 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4089 if (write)
4090 memcpy_toio(maddr + offset, buf, len);
4091 else
4092 memcpy_fromio(buf, maddr + offset, len);
4093 iounmap(maddr);
4094
4095 return len;
4096}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02004097EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004098#endif
4099
David Howells0ec76a12006-09-27 01:50:15 -07004100/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004101 * Access another process' address space as given in mm. If non-NULL, use the
4102 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004103 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004104static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4105 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004106{
David Howells0ec76a12006-09-27 01:50:15 -07004107 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004108 void *old_buf = buf;
4109
David Howells0ec76a12006-09-27 01:50:15 -07004110 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004111 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004112 while (len) {
4113 int bytes, ret, offset;
4114 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004115 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004116
4117 ret = get_user_pages(tsk, mm, addr, 1,
4118 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004119 if (ret <= 0) {
4120 /*
4121 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4122 * we can access using slightly different code.
4123 */
4124#ifdef CONFIG_HAVE_IOREMAP_PROT
4125 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004126 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004127 break;
4128 if (vma->vm_ops && vma->vm_ops->access)
4129 ret = vma->vm_ops->access(vma, addr, buf,
4130 len, write);
4131 if (ret <= 0)
4132#endif
4133 break;
4134 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004135 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004136 bytes = len;
4137 offset = addr & (PAGE_SIZE-1);
4138 if (bytes > PAGE_SIZE-offset)
4139 bytes = PAGE_SIZE-offset;
4140
4141 maddr = kmap(page);
4142 if (write) {
4143 copy_to_user_page(vma, page, addr,
4144 maddr + offset, buf, bytes);
4145 set_page_dirty_lock(page);
4146 } else {
4147 copy_from_user_page(vma, page, addr,
4148 buf, maddr + offset, bytes);
4149 }
4150 kunmap(page);
4151 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004152 }
David Howells0ec76a12006-09-27 01:50:15 -07004153 len -= bytes;
4154 buf += bytes;
4155 addr += bytes;
4156 }
4157 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004158
4159 return buf - old_buf;
4160}
Andi Kleen03252912008-01-30 13:33:18 +01004161
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004162/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004163 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004164 * @mm: the mm_struct of the target address space
4165 * @addr: start address to access
4166 * @buf: source or destination buffer
4167 * @len: number of bytes to transfer
4168 * @write: whether the access is a write
4169 *
4170 * The caller must hold a reference on @mm.
4171 */
4172int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4173 void *buf, int len, int write)
4174{
4175 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4176}
4177
Andi Kleen03252912008-01-30 13:33:18 +01004178/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004179 * Access another process' address space.
4180 * Source/target buffer must be kernel space,
4181 * Do not walk the page table directly, use get_user_pages
4182 */
4183int access_process_vm(struct task_struct *tsk, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 struct mm_struct *mm;
4187 int ret;
4188
4189 mm = get_task_mm(tsk);
4190 if (!mm)
4191 return 0;
4192
4193 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4194 mmput(mm);
4195
4196 return ret;
4197}
4198
Andi Kleen03252912008-01-30 13:33:18 +01004199/*
4200 * Print the name of a VMA.
4201 */
4202void print_vma_addr(char *prefix, unsigned long ip)
4203{
4204 struct mm_struct *mm = current->mm;
4205 struct vm_area_struct *vma;
4206
Ingo Molnare8bff742008-02-13 20:21:06 +01004207 /*
4208 * Do not print if we are in atomic
4209 * contexts (in exception stacks, etc.):
4210 */
4211 if (preempt_count())
4212 return;
4213
Andi Kleen03252912008-01-30 13:33:18 +01004214 down_read(&mm->mmap_sem);
4215 vma = find_vma(mm, ip);
4216 if (vma && vma->vm_file) {
4217 struct file *f = vma->vm_file;
4218 char *buf = (char *)__get_free_page(GFP_KERNEL);
4219 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004220 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004221
Jan Blunckcf28b482008-02-14 19:38:44 -08004222 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004223 if (IS_ERR(p))
4224 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004225 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004226 vma->vm_start,
4227 vma->vm_end - vma->vm_start);
4228 free_page((unsigned long)buf);
4229 }
4230 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004231 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004232}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004233
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004234#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004235void might_fault(void)
4236{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004237 /*
4238 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4239 * holding the mmap_sem, this is safe because kernel memory doesn't
4240 * get paged out, therefore we'll never actually fault, and the
4241 * below annotations will generate false positives.
4242 */
4243 if (segment_eq(get_fs(), KERNEL_DS))
4244 return;
4245
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004246 /*
4247 * it would be nicer only to annotate paths which are not under
4248 * pagefault_disable, however that requires a larger audit and
4249 * providing helpers like get_user_atomic.
4250 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004251 if (in_atomic())
4252 return;
4253
4254 __might_sleep(__FILE__, __LINE__, 0);
4255
4256 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004257 might_lock_read(&current->mm->mmap_sem);
4258}
4259EXPORT_SYMBOL(might_fault);
4260#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004261
4262#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4263static void clear_gigantic_page(struct page *page,
4264 unsigned long addr,
4265 unsigned int pages_per_huge_page)
4266{
4267 int i;
4268 struct page *p = page;
4269
4270 might_sleep();
4271 for (i = 0; i < pages_per_huge_page;
4272 i++, p = mem_map_next(p, page, i)) {
4273 cond_resched();
4274 clear_user_highpage(p, addr + i * PAGE_SIZE);
4275 }
4276}
4277void clear_huge_page(struct page *page,
4278 unsigned long addr, unsigned int pages_per_huge_page)
4279{
4280 int i;
4281
4282 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4283 clear_gigantic_page(page, addr, pages_per_huge_page);
4284 return;
4285 }
4286
4287 might_sleep();
4288 for (i = 0; i < pages_per_huge_page; i++) {
4289 cond_resched();
4290 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4291 }
4292}
4293
4294static void copy_user_gigantic_page(struct page *dst, struct page *src,
4295 unsigned long addr,
4296 struct vm_area_struct *vma,
4297 unsigned int pages_per_huge_page)
4298{
4299 int i;
4300 struct page *dst_base = dst;
4301 struct page *src_base = src;
4302
4303 for (i = 0; i < pages_per_huge_page; ) {
4304 cond_resched();
4305 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4306
4307 i++;
4308 dst = mem_map_next(dst, dst_base, i);
4309 src = mem_map_next(src, src_base, i);
4310 }
4311}
4312
4313void copy_user_huge_page(struct page *dst, struct page *src,
4314 unsigned long addr, struct vm_area_struct *vma,
4315 unsigned int pages_per_huge_page)
4316{
4317 int i;
4318
4319 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4320 copy_user_gigantic_page(dst, src, addr, vma,
4321 pages_per_huge_page);
4322 return;
4323 }
4324
4325 might_sleep();
4326 for (i = 0; i < pages_per_huge_page; i++) {
4327 cond_resched();
4328 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4329 }
4330}
4331#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */