blob: 30991f83d0bf54dc537f1927a10883aa5787dd97 [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>
Dan Williams01c8f1c2016-01-15 16:56:40 -080053#include <linux/pfn_t.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070054#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080057#include <linux/kallsyms.h>
58#include <linux/swapops.h>
59#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090060#include <linux/gfp.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000061#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080062#include <linux/string.h>
Dan Williams0abdd7a2014-01-21 15:48:12 -080063#include <linux/dma-debug.h>
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -070064#include <linux/debugfs.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070065#include <linux/userfaultfd_k.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
Alexey Dobriyan6952b612009-09-18 23:55:55 +040067#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070068#include <asm/pgalloc.h>
69#include <asm/uaccess.h>
70#include <asm/tlb.h>
71#include <asm/tlbflush.h>
72#include <asm/pgtable.h>
73
Jan Beulich42b77722008-07-23 21:27:10 -070074#include "internal.h"
75
Peter Zijlstra90572892013-10-07 11:29:20 +010076#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
77#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080078#endif
79
Andy Whitcroftd41dee32005-06-23 00:07:54 -070080#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070081/* use the per-pgdat data instead for discontigmem - mbligh */
82unsigned long max_mapnr;
83struct page *mem_map;
84
85EXPORT_SYMBOL(max_mapnr);
86EXPORT_SYMBOL(mem_map);
87#endif
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089/*
90 * A number of key systems in x86 including ioremap() rely on the assumption
91 * that high_memory defines the upper bound on direct map memory, then end
92 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
93 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
94 * and ZONE_HIGHMEM.
95 */
96void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Linus Torvalds1da177e2005-04-16 15:20:36 -070098EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Ingo Molnar32a93232008-02-06 22:39:44 +0100100/*
101 * Randomize the address space (stacks, mmaps, brk, etc.).
102 *
103 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
104 * as ancient (libc5 based) binaries can segfault. )
105 */
106int randomize_va_space __read_mostly =
107#ifdef CONFIG_COMPAT_BRK
108 1;
109#else
110 2;
111#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100112
113static int __init disable_randmaps(char *s)
114{
115 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800116 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100117}
118__setup("norandmaps", disable_randmaps);
119
Hugh Dickins62eede62009-09-21 17:03:34 -0700120unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700121unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700122
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200123EXPORT_SYMBOL(zero_pfn);
124
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700125/*
126 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
127 */
128static int __init init_zero_pfn(void)
129{
130 zero_pfn = page_to_pfn(ZERO_PAGE(0));
131 return 0;
132}
133core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100134
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800135
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800136#if defined(SPLIT_RSS_COUNTING)
137
David Rientjesea48cf72012-03-21 16:34:13 -0700138void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800139{
140 int i;
141
142 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700143 if (current->rss_stat.count[i]) {
144 add_mm_counter(mm, i, current->rss_stat.count[i]);
145 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800146 }
147 }
David Rientjes05af2e12012-03-21 16:34:13 -0700148 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800149}
150
151static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
152{
153 struct task_struct *task = current;
154
155 if (likely(task->mm == mm))
156 task->rss_stat.count[member] += val;
157 else
158 add_mm_counter(mm, member, val);
159}
160#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
161#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
162
163/* sync counter once per 64 page faults */
164#define TASK_RSS_EVENTS_THRESH (64)
165static void check_sync_rss_stat(struct task_struct *task)
166{
167 if (unlikely(task != current))
168 return;
169 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700170 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800171}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700172#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800173
174#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
175#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
176
177static void check_sync_rss_stat(struct task_struct *task)
178{
179}
180
Peter Zijlstra9547d012011-05-24 17:12:14 -0700181#endif /* SPLIT_RSS_COUNTING */
182
183#ifdef HAVE_GENERIC_MMU_GATHER
184
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700185static bool tlb_next_batch(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700186{
187 struct mmu_gather_batch *batch;
188
189 batch = tlb->active;
190 if (batch->next) {
191 tlb->active = batch->next;
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700192 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700193 }
194
Michal Hocko53a59fc2013-01-04 15:35:12 -0800195 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700196 return false;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800197
Peter Zijlstra9547d012011-05-24 17:12:14 -0700198 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
199 if (!batch)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700200 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700201
Michal Hocko53a59fc2013-01-04 15:35:12 -0800202 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700203 batch->next = NULL;
204 batch->nr = 0;
205 batch->max = MAX_GATHER_BATCH;
206
207 tlb->active->next = batch;
208 tlb->active = batch;
209
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700210 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700211}
212
213/* tlb_gather_mmu
214 * Called to initialize an (on-stack) mmu_gather structure for page-table
215 * tear-down from @mm. The @fullmm argument is used when @mm is without
216 * users and we're going to destroy the full address space (exit/execve).
217 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700218void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700219{
220 tlb->mm = mm;
221
Linus Torvalds2b047252013-08-15 11:42:25 -0700222 /* Is it from 0 to ~0? */
223 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700224 tlb->need_flush_all = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700225 tlb->local.next = NULL;
226 tlb->local.nr = 0;
227 tlb->local.max = ARRAY_SIZE(tlb->__pages);
228 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800229 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700230
231#ifdef CONFIG_HAVE_RCU_TABLE_FREE
232 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800233#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000234
235 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700236}
237
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700238static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700239{
Will Deacon721c21c2015-01-12 19:10:55 +0000240 if (!tlb->end)
241 return;
242
Peter Zijlstra9547d012011-05-24 17:12:14 -0700243 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100244 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700245#ifdef CONFIG_HAVE_RCU_TABLE_FREE
246 tlb_table_flush(tlb);
247#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000248 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700249}
250
251static void tlb_flush_mmu_free(struct mmu_gather *tlb)
252{
253 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700254
Will Deacon721c21c2015-01-12 19:10:55 +0000255 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700256 free_pages_and_swap_cache(batch->pages, batch->nr);
257 batch->nr = 0;
258 }
259 tlb->active = &tlb->local;
260}
261
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700262void tlb_flush_mmu(struct mmu_gather *tlb)
263{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700264 tlb_flush_mmu_tlbonly(tlb);
265 tlb_flush_mmu_free(tlb);
266}
267
Peter Zijlstra9547d012011-05-24 17:12:14 -0700268/* tlb_finish_mmu
269 * Called at the end of the shootdown operation to free up any resources
270 * that were required.
271 */
272void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
273{
274 struct mmu_gather_batch *batch, *next;
275
276 tlb_flush_mmu(tlb);
277
278 /* keep the page table cache within bounds */
279 check_pgt_cache();
280
281 for (batch = tlb->local.next; batch; batch = next) {
282 next = batch->next;
283 free_pages((unsigned long)batch, 0);
284 }
285 tlb->local.next = NULL;
286}
287
288/* __tlb_remove_page
289 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
290 * handling the additional races in SMP caused by other CPUs caching valid
291 * mappings in their TLBs. Returns the number of free page slots left.
292 * When out of page slots we must call tlb_flush_mmu().
293 */
294int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
295{
296 struct mmu_gather_batch *batch;
297
Will Deaconfb7332a2014-10-29 10:03:09 +0000298 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700299
Peter Zijlstra9547d012011-05-24 17:12:14 -0700300 batch = tlb->active;
301 batch->pages[batch->nr++] = page;
302 if (batch->nr == batch->max) {
303 if (!tlb_next_batch(tlb))
304 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700305 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700306 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800307 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700308
309 return batch->max - batch->nr;
310}
311
312#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800313
Peter Zijlstra26723912011-05-24 17:12:00 -0700314#ifdef CONFIG_HAVE_RCU_TABLE_FREE
315
316/*
317 * See the comment near struct mmu_table_batch.
318 */
319
320static void tlb_remove_table_smp_sync(void *arg)
321{
322 /* Simply deliver the interrupt */
323}
324
325static void tlb_remove_table_one(void *table)
326{
327 /*
328 * This isn't an RCU grace period and hence the page-tables cannot be
329 * assumed to be actually RCU-freed.
330 *
331 * It is however sufficient for software page-table walkers that rely on
332 * IRQ disabling. See the comment near struct mmu_table_batch.
333 */
334 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
335 __tlb_remove_table(table);
336}
337
338static void tlb_remove_table_rcu(struct rcu_head *head)
339{
340 struct mmu_table_batch *batch;
341 int i;
342
343 batch = container_of(head, struct mmu_table_batch, rcu);
344
345 for (i = 0; i < batch->nr; i++)
346 __tlb_remove_table(batch->tables[i]);
347
348 free_page((unsigned long)batch);
349}
350
351void tlb_table_flush(struct mmu_gather *tlb)
352{
353 struct mmu_table_batch **batch = &tlb->batch;
354
355 if (*batch) {
356 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
357 *batch = NULL;
358 }
359}
360
361void tlb_remove_table(struct mmu_gather *tlb, void *table)
362{
363 struct mmu_table_batch **batch = &tlb->batch;
364
Peter Zijlstra26723912011-05-24 17:12:00 -0700365 /*
366 * When there's less then two users of this mm there cannot be a
367 * concurrent page-table walk.
368 */
369 if (atomic_read(&tlb->mm->mm_users) < 2) {
370 __tlb_remove_table(table);
371 return;
372 }
373
374 if (*batch == NULL) {
375 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
376 if (*batch == NULL) {
377 tlb_remove_table_one(table);
378 return;
379 }
380 (*batch)->nr = 0;
381 }
382 (*batch)->tables[(*batch)->nr++] = table;
383 if ((*batch)->nr == MAX_TABLE_BATCH)
384 tlb_table_flush(tlb);
385}
386
Peter Zijlstra9547d012011-05-24 17:12:14 -0700387#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700388
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 * Note: this doesn't free the actual pages themselves. That
391 * has been handled earlier when unmapping all the memory regions.
392 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000393static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
394 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800396 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700397 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000398 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800399 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400}
401
Hugh Dickinse0da3822005-04-19 13:29:15 -0700402static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
403 unsigned long addr, unsigned long end,
404 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405{
406 pmd_t *pmd;
407 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
Hugh Dickinse0da3822005-04-19 13:29:15 -0700410 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 do {
413 next = pmd_addr_end(addr, end);
414 if (pmd_none_or_clear_bad(pmd))
415 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000416 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 } while (pmd++, addr = next, addr != end);
418
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419 start &= PUD_MASK;
420 if (start < floor)
421 return;
422 if (ceiling) {
423 ceiling &= PUD_MASK;
424 if (!ceiling)
425 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700427 if (end - 1 > ceiling - 1)
428 return;
429
430 pmd = pmd_offset(pud, start);
431 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000432 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800433 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
Hugh Dickinse0da3822005-04-19 13:29:15 -0700436static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
437 unsigned long addr, unsigned long end,
438 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439{
440 pud_t *pud;
441 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700442 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
Hugh Dickinse0da3822005-04-19 13:29:15 -0700444 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 do {
447 next = pud_addr_end(addr, end);
448 if (pud_none_or_clear_bad(pud))
449 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700450 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 } while (pud++, addr = next, addr != end);
452
Hugh Dickinse0da3822005-04-19 13:29:15 -0700453 start &= PGDIR_MASK;
454 if (start < floor)
455 return;
456 if (ceiling) {
457 ceiling &= PGDIR_MASK;
458 if (!ceiling)
459 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700461 if (end - 1 > ceiling - 1)
462 return;
463
464 pud = pud_offset(pgd, start);
465 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000466 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467}
468
469/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700470 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 */
Jan Beulich42b77722008-07-23 21:27:10 -0700472void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700473 unsigned long addr, unsigned long end,
474 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475{
476 pgd_t *pgd;
477 unsigned long next;
478
Hugh Dickinse0da3822005-04-19 13:29:15 -0700479 /*
480 * The next few lines have given us lots of grief...
481 *
482 * Why are we testing PMD* at this top level? Because often
483 * there will be no work to do at all, and we'd prefer not to
484 * go all the way down to the bottom just to discover that.
485 *
486 * Why all these "- 1"s? Because 0 represents both the bottom
487 * of the address space and the top of it (using -1 for the
488 * top wouldn't help much: the masks would do the wrong thing).
489 * The rule is that addr 0 and floor 0 refer to the bottom of
490 * the address space, but end 0 and ceiling 0 refer to the top
491 * Comparisons need to use "end - 1" and "ceiling - 1" (though
492 * that end 0 case should be mythical).
493 *
494 * Wherever addr is brought up or ceiling brought down, we must
495 * be careful to reject "the opposite 0" before it confuses the
496 * subsequent tests. But what about where end is brought down
497 * by PMD_SIZE below? no, end can't go down to 0 there.
498 *
499 * Whereas we round start (addr) and ceiling down, by different
500 * masks at different levels, in order to test whether a table
501 * now has no other vmas using it, so can be freed, we don't
502 * bother to round floor or end up - the tests don't need that.
503 */
504
505 addr &= PMD_MASK;
506 if (addr < floor) {
507 addr += PMD_SIZE;
508 if (!addr)
509 return;
510 }
511 if (ceiling) {
512 ceiling &= PMD_MASK;
513 if (!ceiling)
514 return;
515 }
516 if (end - 1 > ceiling - 1)
517 end -= PMD_SIZE;
518 if (addr > end - 1)
519 return;
520
Jan Beulich42b77722008-07-23 21:27:10 -0700521 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 do {
523 next = pgd_addr_end(addr, end);
524 if (pgd_none_or_clear_bad(pgd))
525 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700526 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700528}
529
Jan Beulich42b77722008-07-23 21:27:10 -0700530void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700531 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700532{
533 while (vma) {
534 struct vm_area_struct *next = vma->vm_next;
535 unsigned long addr = vma->vm_start;
536
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700537 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000538 * Hide vma from rmap and truncate_pagecache before freeing
539 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700540 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800541 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700542 unlink_file_vma(vma);
543
David Gibson9da61ae2006-03-22 00:08:57 -0800544 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700545 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700546 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700547 } else {
548 /*
549 * Optimization: gather nearby vmas into one call down
550 */
551 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800552 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700553 vma = next;
554 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800555 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700556 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700557 }
558 free_pgd_range(tlb, addr, vma->vm_end,
559 floor, next? next->vm_start: ceiling);
560 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700561 vma = next;
562 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563}
564
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800565int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
566 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800568 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800569 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700570 if (!new)
571 return -ENOMEM;
572
Nick Piggin362a61a2008-05-14 06:37:36 +0200573 /*
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
577 *
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
585 */
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800588 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800590 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800592 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800593 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800594 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800595 if (new)
596 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700597 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598}
599
Hugh Dickins1bb36302005-10-29 18:16:22 -0700600int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700602 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
603 if (!new)
604 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Nick Piggin362a61a2008-05-14 06:37:36 +0200606 smp_wmb(); /* See comment in __pte_alloc */
607
Hugh Dickins1bb36302005-10-29 18:16:22 -0700608 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700610 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800611 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800612 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800614 if (new)
615 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700616 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617}
618
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800619static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700620{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800621 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
622}
623
624static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
625{
626 int i;
627
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800628 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700629 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800630 for (i = 0; i < NR_MM_COUNTERS; i++)
631 if (rss[i])
632 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700633}
634
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800636 * This function is called to print an error when a bad pte
637 * is found. For example, we might have a PFN-mapped pte in
638 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700639 *
640 * The calling function must still handle the error.
641 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800642static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
643 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700644{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800645 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
646 pud_t *pud = pud_offset(pgd, addr);
647 pmd_t *pmd = pmd_offset(pud, addr);
648 struct address_space *mapping;
649 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800650 static unsigned long resume;
651 static unsigned long nr_shown;
652 static unsigned long nr_unshown;
653
654 /*
655 * Allow a burst of 60 reports, then keep quiet for that minute;
656 * or allow a steady drip of one report per second.
657 */
658 if (nr_shown == 60) {
659 if (time_before(jiffies, resume)) {
660 nr_unshown++;
661 return;
662 }
663 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800664 printk(KERN_ALERT
665 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800666 nr_unshown);
667 nr_unshown = 0;
668 }
669 nr_shown = 0;
670 }
671 if (nr_shown++ == 0)
672 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800673
674 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
675 index = linear_page_index(vma, addr);
676
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800677 printk(KERN_ALERT
678 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800679 current->comm,
680 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800681 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800682 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800683 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800684 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
685 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
686 /*
687 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
688 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700689 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
690 vma->vm_file,
691 vma->vm_ops ? vma->vm_ops->fault : NULL,
692 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
693 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700694 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030695 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700696}
697
698/*
Nick Piggin7e675132008-04-28 02:13:00 -0700699 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800700 *
Nick Piggin7e675132008-04-28 02:13:00 -0700701 * "Special" mappings do not wish to be associated with a "struct page" (either
702 * it doesn't exist, or it exists but they don't want to touch it). In this
703 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700704 *
Nick Piggin7e675132008-04-28 02:13:00 -0700705 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
706 * pte bit, in which case this function is trivial. Secondly, an architecture
707 * may not have a spare pte bit, which requires a more complicated scheme,
708 * described below.
709 *
710 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
711 * special mapping (even if there are underlying and valid "struct pages").
712 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800713 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700714 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
715 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700716 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
717 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800718 *
719 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
720 *
Nick Piggin7e675132008-04-28 02:13:00 -0700721 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700722 *
Nick Piggin7e675132008-04-28 02:13:00 -0700723 * This restricts such mappings to be a linear translation from virtual address
724 * to pfn. To get around this restriction, we allow arbitrary mappings so long
725 * as the vma is not a COW mapping; in that case, we know that all ptes are
726 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700727 *
728 *
Nick Piggin7e675132008-04-28 02:13:00 -0700729 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
730 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700731 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
732 * page" backing, however the difference is that _all_ pages with a struct
733 * page (that is, those where pfn_valid is true) are refcounted and considered
734 * normal pages by the VM. The disadvantage is that pages are refcounted
735 * (which can be slower and simply not an option for some PFNMAP users). The
736 * advantage is that we don't have to follow the strict linearity rule of
737 * PFNMAP mappings in order to support COWable mappings.
738 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800739 */
Nick Piggin7e675132008-04-28 02:13:00 -0700740#ifdef __HAVE_ARCH_PTE_SPECIAL
741# define HAVE_PTE_SPECIAL 1
742#else
743# define HAVE_PTE_SPECIAL 0
744#endif
745struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
746 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800747{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800748 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700749
750 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700751 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800752 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000753 if (vma->vm_ops && vma->vm_ops->find_special_page)
754 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700755 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
756 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700757 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800758 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700759 return NULL;
760 }
761
762 /* !HAVE_PTE_SPECIAL case follows: */
763
Jared Hulbertb379d792008-04-28 02:12:58 -0700764 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
765 if (vma->vm_flags & VM_MIXEDMAP) {
766 if (!pfn_valid(pfn))
767 return NULL;
768 goto out;
769 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700770 unsigned long off;
771 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700772 if (pfn == vma->vm_pgoff + off)
773 return NULL;
774 if (!is_cow_mapping(vma->vm_flags))
775 return NULL;
776 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800777 }
778
Hugh Dickinsb38af472014-08-29 15:18:44 -0700779 if (is_zero_pfn(pfn))
780 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800781check_pfn:
782 if (unlikely(pfn > highest_memmap_pfn)) {
783 print_bad_pte(vma, addr, pte, NULL);
784 return NULL;
785 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800786
787 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700788 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700789 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800790 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700791out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800792 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800793}
794
795/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 * copy one vm_area from one task to the other. Assumes the page tables
797 * already present in the new task to be cleared in the whole range
798 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 */
800
Hugh Dickins570a335b2009-12-14 17:58:46 -0800801static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700803 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700804 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805{
Nick Pigginb5810032005-10-29 18:16:12 -0700806 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 pte_t pte = *src_pte;
808 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809
810 /* pte contains position in swap or file, so copy. */
811 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800812 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700813
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800814 if (likely(!non_swap_entry(entry))) {
815 if (swap_duplicate(entry) < 0)
816 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800817
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800818 /* make sure dst_mm is on swapoff's mmlist. */
819 if (unlikely(list_empty(&dst_mm->mmlist))) {
820 spin_lock(&mmlist_lock);
821 if (list_empty(&dst_mm->mmlist))
822 list_add(&dst_mm->mmlist,
823 &src_mm->mmlist);
824 spin_unlock(&mmlist_lock);
825 }
826 rss[MM_SWAPENTS]++;
827 } else if (is_migration_entry(entry)) {
828 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800829
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800830 rss[mm_counter(page)]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800831
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800832 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 if (pte_swp_soft_dirty(*src_pte))
841 pte = pte_swp_mksoft_dirty(pte);
842 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700843 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700845 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 }
847
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 /*
849 * If it's a COW mapping, write protect it both
850 * in the parent and the child
851 */
Linus Torvalds67121172005-12-11 20:38:17 -0800852 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700854 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 }
856
857 /*
858 * If it's a shared mapping, mark it clean in
859 * the child
860 */
861 if (vm_flags & VM_SHARED)
862 pte = pte_mkclean(pte);
863 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800864
865 page = vm_normal_page(vma, addr, pte);
866 if (page) {
867 get_page(page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800868 page_dup_rmap(page, false);
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800869 rss[mm_counter(page)]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800870 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700871
872out_set_pte:
873 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800874 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875}
876
Jerome Marchand21bda262014-08-06 16:06:56 -0700877static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800878 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
879 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700881 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700883 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700884 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800885 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800886 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
888again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800889 init_rss_vec(rss);
890
Hugh Dickinsc74df322005-10-29 18:16:23 -0700891 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 if (!dst_pte)
893 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700894 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700895 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700896 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700897 orig_src_pte = src_pte;
898 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700899 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 do {
902 /*
903 * We are holding two locks at this point - either of them
904 * could generate latencies in another task on another CPU.
905 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700906 if (progress >= 32) {
907 progress = 0;
908 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100909 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700910 break;
911 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 if (pte_none(*src_pte)) {
913 progress++;
914 continue;
915 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800916 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
917 vma, addr, rss);
918 if (entry.val)
919 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 progress += 8;
921 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700923 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700924 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700925 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800926 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700927 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700928 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800929
930 if (entry.val) {
931 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
932 return -ENOMEM;
933 progress = 0;
934 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 if (addr != end)
936 goto again;
937 return 0;
938}
939
940static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
941 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
942 unsigned long addr, unsigned long end)
943{
944 pmd_t *src_pmd, *dst_pmd;
945 unsigned long next;
946
947 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
948 if (!dst_pmd)
949 return -ENOMEM;
950 src_pmd = pmd_offset(src_pud, addr);
951 do {
952 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -0800953 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800954 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800955 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800956 err = copy_huge_pmd(dst_mm, src_mm,
957 dst_pmd, src_pmd, addr, vma);
958 if (err == -ENOMEM)
959 return -ENOMEM;
960 if (!err)
961 continue;
962 /* fall through */
963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 if (pmd_none_or_clear_bad(src_pmd))
965 continue;
966 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
967 vma, addr, next))
968 return -ENOMEM;
969 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
970 return 0;
971}
972
973static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
974 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
975 unsigned long addr, unsigned long end)
976{
977 pud_t *src_pud, *dst_pud;
978 unsigned long next;
979
980 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
981 if (!dst_pud)
982 return -ENOMEM;
983 src_pud = pud_offset(src_pgd, addr);
984 do {
985 next = pud_addr_end(addr, end);
986 if (pud_none_or_clear_bad(src_pud))
987 continue;
988 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
989 vma, addr, next))
990 return -ENOMEM;
991 } while (dst_pud++, src_pud++, addr = next, addr != end);
992 return 0;
993}
994
995int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
996 struct vm_area_struct *vma)
997{
998 pgd_t *src_pgd, *dst_pgd;
999 unsigned long next;
1000 unsigned long addr = vma->vm_start;
1001 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001002 unsigned long mmun_start; /* For mmu_notifiers */
1003 unsigned long mmun_end; /* For mmu_notifiers */
1004 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001005 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
Nick Piggind9928952005-08-28 16:49:11 +10001007 /*
1008 * Don't copy ptes where a page fault will fill them correctly.
1009 * Fork becomes much lighter when there are big shared or private
1010 * readonly mappings. The tradeoff is that copy_page_range is more
1011 * efficient than faulting.
1012 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001013 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1014 !vma->anon_vma)
1015 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001016
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 if (is_vm_hugetlb_page(vma))
1018 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1019
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001020 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001021 /*
1022 * We do not free on error cases below as remove_vma
1023 * gets called on error from higher level routine
1024 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001025 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001026 if (ret)
1027 return ret;
1028 }
1029
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001030 /*
1031 * We need to invalidate the secondary MMU mappings only when
1032 * there could be a permission downgrade on the ptes of the
1033 * parent mm. And a permission downgrade will only happen if
1034 * is_cow_mapping() returns true.
1035 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001036 is_cow = is_cow_mapping(vma->vm_flags);
1037 mmun_start = addr;
1038 mmun_end = end;
1039 if (is_cow)
1040 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1041 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001042
1043 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 dst_pgd = pgd_offset(dst_mm, addr);
1045 src_pgd = pgd_offset(src_mm, addr);
1046 do {
1047 next = pgd_addr_end(addr, end);
1048 if (pgd_none_or_clear_bad(src_pgd))
1049 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001050 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1051 vma, addr, next))) {
1052 ret = -ENOMEM;
1053 break;
1054 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001056
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001057 if (is_cow)
1058 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001059 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060}
1061
Robin Holt51c6f662005-11-13 16:06:42 -08001062static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001063 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001065 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066{
Nick Pigginb5810032005-10-29 18:16:12 -07001067 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001068 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001069 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001070 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001071 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001072 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001073 swp_entry_t entry;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001074
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001075again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001076 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001077 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1078 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001079 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 do {
1081 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001082 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001084 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001085
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001087 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001088
Linus Torvalds6aab3412005-11-28 14:34:23 -08001089 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 if (unlikely(details) && page) {
1091 /*
1092 * unmap_shared_mapping_pages() wants to
1093 * invalidate cache without truncating:
1094 * unmap shared but keep private pages.
1095 */
1096 if (details->check_mapping &&
1097 details->check_mapping != page->mapping)
1098 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 }
Nick Pigginb5810032005-10-29 18:16:12 -07001100 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001101 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 tlb_remove_tlb_entry(tlb, pte, addr);
1103 if (unlikely(!page))
1104 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001105
1106 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001107 if (pte_dirty(ptent)) {
1108 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001109 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001110 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001111 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001112 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001113 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001114 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001115 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001116 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001117 if (unlikely(page_mapcount(page) < 0))
1118 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001119 if (unlikely(!__tlb_remove_page(tlb, page))) {
1120 force_flush = 1;
Will Deaconce9ec372014-10-28 13:16:28 -07001121 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001122 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001123 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 continue;
1125 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001126 /* If details->check_mapping, we leave swap entries. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 if (unlikely(details))
1128 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001129
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001130 entry = pte_to_swp_entry(ptent);
1131 if (!non_swap_entry(entry))
1132 rss[MM_SWAPENTS]--;
1133 else if (is_migration_entry(entry)) {
1134 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001135
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001136 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001137 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001138 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001139 if (unlikely(!free_swap_and_cache(entry)))
1140 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001141 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001142 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001143
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001144 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001145 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001146
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001147 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001148 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001149 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001150 pte_unmap_unlock(start_pte, ptl);
1151
1152 /*
1153 * If we forced a TLB flush (either due to running out of
1154 * batch buffers or because we needed to flush dirty TLB
1155 * entries before releasing the ptl), free the batched
1156 * memory too. Restart if we didn't do everything.
1157 */
1158 if (force_flush) {
1159 force_flush = 0;
1160 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001161
1162 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001163 goto again;
1164 }
1165
Robin Holt51c6f662005-11-13 16:06:42 -08001166 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167}
1168
Robin Holt51c6f662005-11-13 16:06:42 -08001169static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001170 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001172 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173{
1174 pmd_t *pmd;
1175 unsigned long next;
1176
1177 pmd = pmd_offset(pud, addr);
1178 do {
1179 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001180 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001181 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001182#ifdef CONFIG_DEBUG_VM
1183 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1184 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1185 __func__, addr, end,
1186 vma->vm_start,
1187 vma->vm_end);
1188 BUG();
1189 }
1190#endif
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001191 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001192 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001193 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001194 /* fall through */
1195 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001196 /*
1197 * Here there can be other concurrent MADV_DONTNEED or
1198 * trans huge page faults running, and if the pmd is
1199 * none or trans huge it can change under us. This is
1200 * because MADV_DONTNEED holds the mmap_sem in read
1201 * mode.
1202 */
1203 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1204 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001205 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001206next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001207 cond_resched();
1208 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001209
1210 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211}
1212
Robin Holt51c6f662005-11-13 16:06:42 -08001213static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001214 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001216 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217{
1218 pud_t *pud;
1219 unsigned long next;
1220
1221 pud = pud_offset(pgd, addr);
1222 do {
1223 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001224 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001226 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1227 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001228
1229 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230}
1231
Al Viro038c7aa2012-03-05 13:25:09 -05001232static void unmap_page_range(struct mmu_gather *tlb,
1233 struct vm_area_struct *vma,
1234 unsigned long addr, unsigned long end,
1235 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236{
1237 pgd_t *pgd;
1238 unsigned long next;
1239
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001240 if (details && !details->check_mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 details = NULL;
1242
1243 BUG_ON(addr >= end);
1244 tlb_start_vma(tlb, vma);
1245 pgd = pgd_offset(vma->vm_mm, addr);
1246 do {
1247 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001248 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001250 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1251 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 tlb_end_vma(tlb, vma);
1253}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
Al Virof5cc4ee2012-03-05 14:14:20 -05001255
1256static void unmap_single_vma(struct mmu_gather *tlb,
1257 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001258 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001259 struct zap_details *details)
1260{
1261 unsigned long start = max(vma->vm_start, start_addr);
1262 unsigned long end;
1263
1264 if (start >= vma->vm_end)
1265 return;
1266 end = min(vma->vm_end, end_addr);
1267 if (end <= vma->vm_start)
1268 return;
1269
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301270 if (vma->vm_file)
1271 uprobe_munmap(vma, start, end);
1272
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001273 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001274 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001275
1276 if (start != end) {
1277 if (unlikely(is_vm_hugetlb_page(vma))) {
1278 /*
1279 * It is undesirable to test vma->vm_file as it
1280 * should be non-null for valid hugetlb area.
1281 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001282 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001283 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001284 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001285 * before calling this function to clean up.
1286 * Since no pte has actually been setup, it is
1287 * safe to do nothing in this case.
1288 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001289 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001290 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001291 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001292 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001293 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001294 } else
1295 unmap_page_range(tlb, vma, start, end, details);
1296 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297}
1298
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299/**
1300 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001301 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 * @vma: the starting vma
1303 * @start_addr: virtual address at which to start unmapping
1304 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001306 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 * Only addresses between `start' and `end' will be unmapped.
1309 *
1310 * The VMA list must be sorted in ascending virtual address order.
1311 *
1312 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1313 * range after unmap_vmas() returns. So the only responsibility here is to
1314 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1315 * drops the lock and schedules.
1316 */
Al Viro6e8bb012012-03-05 13:41:15 -05001317void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001319 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001321 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001325 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001326 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327}
1328
1329/**
1330 * zap_page_range - remove user pages in a given range
1331 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001332 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001334 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 *
1336 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001338void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 unsigned long size, struct zap_details *details)
1340{
1341 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001342 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001343 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001346 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001347 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001348 mmu_notifier_invalidate_range_start(mm, start, end);
1349 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001350 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001351 mmu_notifier_invalidate_range_end(mm, start, end);
1352 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353}
1354
Jack Steinerc627f9c2008-07-29 22:33:53 -07001355/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001356 * zap_page_range_single - remove user pages in a given range
1357 * @vma: vm_area_struct holding the applicable pages
1358 * @address: starting address of pages to zap
1359 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001360 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001361 *
1362 * The range must fit into one VMA.
1363 */
1364static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1365 unsigned long size, struct zap_details *details)
1366{
1367 struct mm_struct *mm = vma->vm_mm;
1368 struct mmu_gather tlb;
1369 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001370
1371 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001372 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001373 update_hiwater_rss(mm);
1374 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001375 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001376 mmu_notifier_invalidate_range_end(mm, address, end);
1377 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378}
1379
Jack Steinerc627f9c2008-07-29 22:33:53 -07001380/**
1381 * zap_vma_ptes - remove ptes mapping the vma
1382 * @vma: vm_area_struct holding ptes to be zapped
1383 * @address: starting address of pages to zap
1384 * @size: number of bytes to zap
1385 *
1386 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1387 *
1388 * The entire address range must be fully contained within the vma.
1389 *
1390 * Returns 0 if successful.
1391 */
1392int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1393 unsigned long size)
1394{
1395 if (address < vma->vm_start || address + size > vma->vm_end ||
1396 !(vma->vm_flags & VM_PFNMAP))
1397 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001398 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001399 return 0;
1400}
1401EXPORT_SYMBOL_GPL(zap_vma_ptes);
1402
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001403pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001404 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001405{
1406 pgd_t * pgd = pgd_offset(mm, addr);
1407 pud_t * pud = pud_alloc(mm, pgd, addr);
1408 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001409 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001410 if (pmd) {
1411 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001412 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001413 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001414 }
1415 return NULL;
1416}
1417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001419 * This is the old fallback for page remapping.
1420 *
1421 * For historical reasons, it only allows reserved pages. Only
1422 * old drivers should use this, and they needed to mark their
1423 * pages reserved for the old functions anyway.
1424 */
Nick Piggin423bad62008-04-28 02:13:01 -07001425static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1426 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001427{
Nick Piggin423bad62008-04-28 02:13:01 -07001428 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001429 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001430 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001431 spinlock_t *ptl;
1432
Linus Torvalds238f58d2005-11-29 13:01:56 -08001433 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001434 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001435 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001436 retval = -ENOMEM;
1437 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001438 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001439 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001440 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001441 retval = -EBUSY;
1442 if (!pte_none(*pte))
1443 goto out_unlock;
1444
1445 /* Ok, finally just insert the thing.. */
1446 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001447 inc_mm_counter_fast(mm, mm_counter_file(page));
Linus Torvalds238f58d2005-11-29 13:01:56 -08001448 page_add_file_rmap(page);
1449 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1450
1451 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001452 pte_unmap_unlock(pte, ptl);
1453 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001454out_unlock:
1455 pte_unmap_unlock(pte, ptl);
1456out:
1457 return retval;
1458}
1459
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001460/**
1461 * vm_insert_page - insert single page into user vma
1462 * @vma: user vma to map to
1463 * @addr: target user address of this page
1464 * @page: source kernel page
1465 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001466 * This allows drivers to insert individual pages they've allocated
1467 * into a user vma.
1468 *
1469 * The page has to be a nice clean _individual_ kernel allocation.
1470 * If you allocate a compound page, you need to have marked it as
1471 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001472 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001473 *
1474 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1475 * took an arbitrary page protection parameter. This doesn't allow
1476 * that. Your vma protection will have to be set up correctly, which
1477 * means that if you want a shared writable mapping, you'd better
1478 * ask for a shared writable mapping!
1479 *
1480 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001481 *
1482 * Usually this function is called from f_op->mmap() handler
1483 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1484 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1485 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001486 */
Nick Piggin423bad62008-04-28 02:13:01 -07001487int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1488 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001489{
1490 if (addr < vma->vm_start || addr >= vma->vm_end)
1491 return -EFAULT;
1492 if (!page_count(page))
1493 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001494 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1495 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1496 BUG_ON(vma->vm_flags & VM_PFNMAP);
1497 vma->vm_flags |= VM_MIXEDMAP;
1498 }
Nick Piggin423bad62008-04-28 02:13:01 -07001499 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001500}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001501EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001502
Nick Piggin423bad62008-04-28 02:13:01 -07001503static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001504 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad62008-04-28 02:13:01 -07001505{
1506 struct mm_struct *mm = vma->vm_mm;
1507 int retval;
1508 pte_t *pte, entry;
1509 spinlock_t *ptl;
1510
1511 retval = -ENOMEM;
1512 pte = get_locked_pte(mm, addr, &ptl);
1513 if (!pte)
1514 goto out;
1515 retval = -EBUSY;
1516 if (!pte_none(*pte))
1517 goto out_unlock;
1518
1519 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001520 if (pfn_t_devmap(pfn))
1521 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1522 else
1523 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad62008-04-28 02:13:01 -07001524 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001525 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001526
1527 retval = 0;
1528out_unlock:
1529 pte_unmap_unlock(pte, ptl);
1530out:
1531 return retval;
1532}
1533
Nick Piggine0dc0d82007-02-12 00:51:36 -08001534/**
1535 * vm_insert_pfn - insert single pfn into user vma
1536 * @vma: user vma to map to
1537 * @addr: target user address of this page
1538 * @pfn: source kernel pfn
1539 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001540 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001541 * they've allocated into a user vma. Same comments apply.
1542 *
1543 * This function should only be called from a vm_ops->fault handler, and
1544 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001545 *
1546 * vma cannot be a COW mapping.
1547 *
1548 * As this is called only for pages that do not currently exist, we
1549 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001550 */
1551int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001552 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001553{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001554 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001555 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001556 /*
1557 * Technically, architectures with pte_special can avoid all these
1558 * restrictions (same for remap_pfn_range). However we would like
1559 * consistency in testing and feature parity among all, so we should
1560 * try to keep these invariants in place for everybody.
1561 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001562 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1563 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1564 (VM_PFNMAP|VM_MIXEDMAP));
1565 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1566 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001567
Nick Piggin423bad62008-04-28 02:13:01 -07001568 if (addr < vma->vm_start || addr >= vma->vm_end)
1569 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001570 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001571 return -EINVAL;
1572
Dan Williams01c8f1c2016-01-15 16:56:40 -08001573 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001574
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001575 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001576}
1577EXPORT_SYMBOL(vm_insert_pfn);
1578
Nick Piggin423bad62008-04-28 02:13:01 -07001579int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001580 pfn_t pfn)
Nick Piggin423bad62008-04-28 02:13:01 -07001581{
1582 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1583
1584 if (addr < vma->vm_start || addr >= vma->vm_end)
1585 return -EFAULT;
1586
1587 /*
1588 * If we don't have pte special, then we have to use the pfn_valid()
1589 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1590 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001591 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1592 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001593 */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001594 if (!HAVE_PTE_SPECIAL && pfn_t_valid(pfn)) {
Nick Piggin423bad62008-04-28 02:13:01 -07001595 struct page *page;
1596
Dan Williams01c8f1c2016-01-15 16:56:40 -08001597 page = pfn_t_to_page(pfn);
Nick Piggin423bad62008-04-28 02:13:01 -07001598 return insert_page(vma, addr, page, vma->vm_page_prot);
1599 }
1600 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1601}
1602EXPORT_SYMBOL(vm_insert_mixed);
1603
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001604/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 * maps a range of physical memory into the requested pages. the old
1606 * mappings are removed. any references to nonexistent pages results
1607 * in null mappings (currently treated as "copy-on-access")
1608 */
1609static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1610 unsigned long addr, unsigned long end,
1611 unsigned long pfn, pgprot_t prot)
1612{
1613 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001614 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615
Hugh Dickinsc74df322005-10-29 18:16:23 -07001616 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 if (!pte)
1618 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001619 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 do {
1621 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001622 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 pfn++;
1624 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001625 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001626 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 return 0;
1628}
1629
1630static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1631 unsigned long addr, unsigned long end,
1632 unsigned long pfn, pgprot_t prot)
1633{
1634 pmd_t *pmd;
1635 unsigned long next;
1636
1637 pfn -= addr >> PAGE_SHIFT;
1638 pmd = pmd_alloc(mm, pud, addr);
1639 if (!pmd)
1640 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001641 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 do {
1643 next = pmd_addr_end(addr, end);
1644 if (remap_pte_range(mm, pmd, addr, next,
1645 pfn + (addr >> PAGE_SHIFT), prot))
1646 return -ENOMEM;
1647 } while (pmd++, addr = next, addr != end);
1648 return 0;
1649}
1650
1651static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1652 unsigned long addr, unsigned long end,
1653 unsigned long pfn, pgprot_t prot)
1654{
1655 pud_t *pud;
1656 unsigned long next;
1657
1658 pfn -= addr >> PAGE_SHIFT;
1659 pud = pud_alloc(mm, pgd, addr);
1660 if (!pud)
1661 return -ENOMEM;
1662 do {
1663 next = pud_addr_end(addr, end);
1664 if (remap_pmd_range(mm, pud, addr, next,
1665 pfn + (addr >> PAGE_SHIFT), prot))
1666 return -ENOMEM;
1667 } while (pud++, addr = next, addr != end);
1668 return 0;
1669}
1670
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001671/**
1672 * remap_pfn_range - remap kernel memory to userspace
1673 * @vma: user vma to map to
1674 * @addr: target user address to start at
1675 * @pfn: physical address of kernel memory
1676 * @size: size of map area
1677 * @prot: page protection flags for this mapping
1678 *
1679 * Note: this is only safe if the mm semaphore is held when called.
1680 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1682 unsigned long pfn, unsigned long size, pgprot_t prot)
1683{
1684 pgd_t *pgd;
1685 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001686 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 struct mm_struct *mm = vma->vm_mm;
1688 int err;
1689
1690 /*
1691 * Physically remapped pages are special. Tell the
1692 * rest of the world about it:
1693 * VM_IO tells people not to look at these pages
1694 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001695 * VM_PFNMAP tells the core MM that the base pages are just
1696 * raw PFN mappings, and do not have a "struct page" associated
1697 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001698 * VM_DONTEXPAND
1699 * Disable vma merging and expanding with mremap().
1700 * VM_DONTDUMP
1701 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001702 *
1703 * There's a horrible special case to handle copy-on-write
1704 * behaviour that some programs depend on. We mark the "original"
1705 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001706 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001708 if (is_cow_mapping(vma->vm_flags)) {
1709 if (addr != vma->vm_start || end != vma->vm_end)
1710 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001711 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001712 }
1713
1714 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1715 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001716 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001717
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001718 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
1720 BUG_ON(addr >= end);
1721 pfn -= addr >> PAGE_SHIFT;
1722 pgd = pgd_offset(mm, addr);
1723 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 do {
1725 next = pgd_addr_end(addr, end);
1726 err = remap_pud_range(mm, pgd, addr, next,
1727 pfn + (addr >> PAGE_SHIFT), prot);
1728 if (err)
1729 break;
1730 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001731
1732 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001733 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001734
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 return err;
1736}
1737EXPORT_SYMBOL(remap_pfn_range);
1738
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001739/**
1740 * vm_iomap_memory - remap memory to userspace
1741 * @vma: user vma to map to
1742 * @start: start of area
1743 * @len: size of area
1744 *
1745 * This is a simplified io_remap_pfn_range() for common driver use. The
1746 * driver just needs to give us the physical memory range to be mapped,
1747 * we'll figure out the rest from the vma information.
1748 *
1749 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1750 * whatever write-combining details or similar.
1751 */
1752int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1753{
1754 unsigned long vm_len, pfn, pages;
1755
1756 /* Check that the physical memory area passed in looks valid */
1757 if (start + len < start)
1758 return -EINVAL;
1759 /*
1760 * You *really* shouldn't map things that aren't page-aligned,
1761 * but we've historically allowed it because IO memory might
1762 * just have smaller alignment.
1763 */
1764 len += start & ~PAGE_MASK;
1765 pfn = start >> PAGE_SHIFT;
1766 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1767 if (pfn + pages < pfn)
1768 return -EINVAL;
1769
1770 /* We start the mapping 'vm_pgoff' pages into the area */
1771 if (vma->vm_pgoff > pages)
1772 return -EINVAL;
1773 pfn += vma->vm_pgoff;
1774 pages -= vma->vm_pgoff;
1775
1776 /* Can we fit all of the mapping? */
1777 vm_len = vma->vm_end - vma->vm_start;
1778 if (vm_len >> PAGE_SHIFT > pages)
1779 return -EINVAL;
1780
1781 /* Ok, let it rip */
1782 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1783}
1784EXPORT_SYMBOL(vm_iomap_memory);
1785
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001786static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1787 unsigned long addr, unsigned long end,
1788 pte_fn_t fn, void *data)
1789{
1790 pte_t *pte;
1791 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001792 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001793 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001794
1795 pte = (mm == &init_mm) ?
1796 pte_alloc_kernel(pmd, addr) :
1797 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1798 if (!pte)
1799 return -ENOMEM;
1800
1801 BUG_ON(pmd_huge(*pmd));
1802
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001803 arch_enter_lazy_mmu_mode();
1804
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001805 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001806
1807 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001808 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001809 if (err)
1810 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001811 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001812
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001813 arch_leave_lazy_mmu_mode();
1814
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001815 if (mm != &init_mm)
1816 pte_unmap_unlock(pte-1, ptl);
1817 return err;
1818}
1819
1820static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1821 unsigned long addr, unsigned long end,
1822 pte_fn_t fn, void *data)
1823{
1824 pmd_t *pmd;
1825 unsigned long next;
1826 int err;
1827
Andi Kleenceb86872008-07-23 21:27:50 -07001828 BUG_ON(pud_huge(*pud));
1829
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001830 pmd = pmd_alloc(mm, pud, addr);
1831 if (!pmd)
1832 return -ENOMEM;
1833 do {
1834 next = pmd_addr_end(addr, end);
1835 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1836 if (err)
1837 break;
1838 } while (pmd++, addr = next, addr != end);
1839 return err;
1840}
1841
1842static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1843 unsigned long addr, unsigned long end,
1844 pte_fn_t fn, void *data)
1845{
1846 pud_t *pud;
1847 unsigned long next;
1848 int err;
1849
1850 pud = pud_alloc(mm, pgd, addr);
1851 if (!pud)
1852 return -ENOMEM;
1853 do {
1854 next = pud_addr_end(addr, end);
1855 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1856 if (err)
1857 break;
1858 } while (pud++, addr = next, addr != end);
1859 return err;
1860}
1861
1862/*
1863 * Scan a region of virtual memory, filling in page tables as necessary
1864 * and calling a provided function on each leaf page table.
1865 */
1866int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1867 unsigned long size, pte_fn_t fn, void *data)
1868{
1869 pgd_t *pgd;
1870 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001871 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001872 int err;
1873
1874 BUG_ON(addr >= end);
1875 pgd = pgd_offset(mm, addr);
1876 do {
1877 next = pgd_addr_end(addr, end);
1878 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1879 if (err)
1880 break;
1881 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001882
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001883 return err;
1884}
1885EXPORT_SYMBOL_GPL(apply_to_page_range);
1886
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001888 * handle_pte_fault chooses page fault handler according to an entry which was
1889 * read non-atomically. Before making any commitment, on those architectures
1890 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1891 * parts, do_swap_page must check under lock before unmapping the pte and
1892 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001893 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001894 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001895static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001896 pte_t *page_table, pte_t orig_pte)
1897{
1898 int same = 1;
1899#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1900 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001901 spinlock_t *ptl = pte_lockptr(mm, pmd);
1902 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001903 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001904 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001905 }
1906#endif
1907 pte_unmap(page_table);
1908 return same;
1909}
1910
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001911static 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 -08001912{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001913 debug_dma_assert_idle(src);
1914
Linus Torvalds6aab3412005-11-28 14:34:23 -08001915 /*
1916 * If the source page was a PFN mapping, we don't have
1917 * a "struct page" for it. We do a best-effort copy by
1918 * just copying from the original user address. If that
1919 * fails, we just zero-fill it. Live with it.
1920 */
1921 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001922 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001923 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1924
1925 /*
1926 * This really shouldn't fail, because the page is there
1927 * in the page tables. But it might just be unreadable,
1928 * in which case we just give up and fill the result with
1929 * zeroes.
1930 */
1931 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001932 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001933 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001934 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001935 } else
1936 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001937}
1938
Michal Hockoc20cd452016-01-14 15:20:12 -08001939static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
1940{
1941 struct file *vm_file = vma->vm_file;
1942
1943 if (vm_file)
1944 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
1945
1946 /*
1947 * Special mappings (e.g. VDSO) do not have any file so fake
1948 * a default GFP_KERNEL for them.
1949 */
1950 return GFP_KERNEL;
1951}
1952
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001954 * Notify the address space that the page is about to become writable so that
1955 * it can prohibit this or wait for the page to get into an appropriate state.
1956 *
1957 * We do this without the lock held, so that it can sleep if it needs to.
1958 */
1959static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1960 unsigned long address)
1961{
1962 struct vm_fault vmf;
1963 int ret;
1964
1965 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1966 vmf.pgoff = page->index;
1967 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08001968 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001969 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08001970 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001971
1972 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1973 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1974 return ret;
1975 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1976 lock_page(page);
1977 if (!page->mapping) {
1978 unlock_page(page);
1979 return 0; /* retry */
1980 }
1981 ret |= VM_FAULT_LOCKED;
1982 } else
1983 VM_BUG_ON_PAGE(!PageLocked(page), page);
1984 return ret;
1985}
1986
1987/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07001988 * Handle write page faults for pages that can be reused in the current vma
1989 *
1990 * This can happen either due to the mapping being with the VM_SHARED flag,
1991 * or due to us being the last reference standing to the page. In either
1992 * case, all we need to do here is to mark the page as writable and update
1993 * any related book-keeping.
1994 */
1995static inline int wp_page_reuse(struct mm_struct *mm,
1996 struct vm_area_struct *vma, unsigned long address,
1997 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
1998 struct page *page, int page_mkwrite,
1999 int dirty_shared)
2000 __releases(ptl)
2001{
2002 pte_t entry;
2003 /*
2004 * Clear the pages cpupid information as the existing
2005 * information potentially belongs to a now completely
2006 * unrelated process.
2007 */
2008 if (page)
2009 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2010
2011 flush_cache_page(vma, address, pte_pfn(orig_pte));
2012 entry = pte_mkyoung(orig_pte);
2013 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2014 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2015 update_mmu_cache(vma, address, page_table);
2016 pte_unmap_unlock(page_table, ptl);
2017
2018 if (dirty_shared) {
2019 struct address_space *mapping;
2020 int dirtied;
2021
2022 if (!page_mkwrite)
2023 lock_page(page);
2024
2025 dirtied = set_page_dirty(page);
2026 VM_BUG_ON_PAGE(PageAnon(page), page);
2027 mapping = page->mapping;
2028 unlock_page(page);
2029 page_cache_release(page);
2030
2031 if ((dirtied || page_mkwrite) && mapping) {
2032 /*
2033 * Some device drivers do not set page.mapping
2034 * but still dirty their pages
2035 */
2036 balance_dirty_pages_ratelimited(mapping);
2037 }
2038
2039 if (!page_mkwrite)
2040 file_update_time(vma->vm_file);
2041 }
2042
2043 return VM_FAULT_WRITE;
2044}
2045
2046/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002047 * Handle the case of a page which we actually need to copy to a new page.
2048 *
2049 * Called with mmap_sem locked and the old page referenced, but
2050 * without the ptl held.
2051 *
2052 * High level logic flow:
2053 *
2054 * - Allocate a page, copy the content of the old page to the new one.
2055 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2056 * - Take the PTL. If the pte changed, bail out and release the allocated page
2057 * - If the pte is still the way we remember it, update the page table and all
2058 * relevant references. This includes dropping the reference the page-table
2059 * held to the old page, as well as updating the rmap.
2060 * - In any case, unlock the PTL and drop the reference we took to the old page.
2061 */
2062static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2063 unsigned long address, pte_t *page_table, pmd_t *pmd,
2064 pte_t orig_pte, struct page *old_page)
2065{
2066 struct page *new_page = NULL;
2067 spinlock_t *ptl = NULL;
2068 pte_t entry;
2069 int page_copied = 0;
2070 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2071 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2072 struct mem_cgroup *memcg;
2073
2074 if (unlikely(anon_vma_prepare(vma)))
2075 goto oom;
2076
2077 if (is_zero_pfn(pte_pfn(orig_pte))) {
2078 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2079 if (!new_page)
2080 goto oom;
2081 } else {
2082 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2083 if (!new_page)
2084 goto oom;
2085 cow_user_page(new_page, old_page, address, vma);
2086 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002087
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002088 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002089 goto oom_free_new;
2090
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002091 __SetPageUptodate(new_page);
2092
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002093 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2094
2095 /*
2096 * Re-check the pte - we dropped the lock
2097 */
2098 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2099 if (likely(pte_same(*page_table, orig_pte))) {
2100 if (old_page) {
2101 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002102 dec_mm_counter_fast(mm,
2103 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002104 inc_mm_counter_fast(mm, MM_ANONPAGES);
2105 }
2106 } else {
2107 inc_mm_counter_fast(mm, MM_ANONPAGES);
2108 }
2109 flush_cache_page(vma, address, pte_pfn(orig_pte));
2110 entry = mk_pte(new_page, vma->vm_page_prot);
2111 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2112 /*
2113 * Clear the pte entry and flush it first, before updating the
2114 * pte with the new entry. This will avoid a race condition
2115 * seen in the presence of one thread doing SMC and another
2116 * thread doing COW.
2117 */
2118 ptep_clear_flush_notify(vma, address, page_table);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002119 page_add_new_anon_rmap(new_page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002120 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002121 lru_cache_add_active_or_unevictable(new_page, vma);
2122 /*
2123 * We call the notify macro here because, when using secondary
2124 * mmu page tables (such as kvm shadow page tables), we want the
2125 * new page to be mapped directly into the secondary page table.
2126 */
2127 set_pte_at_notify(mm, address, page_table, entry);
2128 update_mmu_cache(vma, address, page_table);
2129 if (old_page) {
2130 /*
2131 * Only after switching the pte to the new page may
2132 * we remove the mapcount here. Otherwise another
2133 * process may come and find the rmap count decremented
2134 * before the pte is switched to the new page, and
2135 * "reuse" the old page writing into it while our pte
2136 * here still points into it and can be read by other
2137 * threads.
2138 *
2139 * The critical issue is to order this
2140 * page_remove_rmap with the ptp_clear_flush above.
2141 * Those stores are ordered by (if nothing else,)
2142 * the barrier present in the atomic_add_negative
2143 * in page_remove_rmap.
2144 *
2145 * Then the TLB flush in ptep_clear_flush ensures that
2146 * no process can access the old page before the
2147 * decremented mapcount is visible. And the old page
2148 * cannot be reused until after the decremented
2149 * mapcount is visible. So transitively, TLBs to
2150 * old page will be flushed before it can be reused.
2151 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002152 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002153 }
2154
2155 /* Free the old page.. */
2156 new_page = old_page;
2157 page_copied = 1;
2158 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002159 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002160 }
2161
2162 if (new_page)
2163 page_cache_release(new_page);
2164
2165 pte_unmap_unlock(page_table, ptl);
2166 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002167 if (old_page) {
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002168 /*
2169 * Don't let another task, with possibly unlocked vma,
2170 * keep the mlocked page.
2171 */
2172 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2173 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002174 if (PageMlocked(old_page))
2175 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002176 unlock_page(old_page);
2177 }
2178 page_cache_release(old_page);
2179 }
2180 return page_copied ? VM_FAULT_WRITE : 0;
2181oom_free_new:
2182 page_cache_release(new_page);
2183oom:
2184 if (old_page)
2185 page_cache_release(old_page);
2186 return VM_FAULT_OOM;
2187}
2188
Boaz Harroshdd906182015-04-15 16:15:11 -07002189/*
2190 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2191 * mapping
2192 */
2193static int wp_pfn_shared(struct mm_struct *mm,
2194 struct vm_area_struct *vma, unsigned long address,
2195 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2196 pmd_t *pmd)
2197{
2198 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2199 struct vm_fault vmf = {
2200 .page = NULL,
2201 .pgoff = linear_page_index(vma, address),
2202 .virtual_address = (void __user *)(address & PAGE_MASK),
2203 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2204 };
2205 int ret;
2206
2207 pte_unmap_unlock(page_table, ptl);
2208 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2209 if (ret & VM_FAULT_ERROR)
2210 return ret;
2211 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2212 /*
2213 * We might have raced with another page fault while we
2214 * released the pte_offset_map_lock.
2215 */
2216 if (!pte_same(*page_table, orig_pte)) {
2217 pte_unmap_unlock(page_table, ptl);
2218 return 0;
2219 }
2220 }
2221 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2222 NULL, 0, 0);
2223}
2224
Shachar Raindel93e478d2015-04-14 15:46:35 -07002225static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2226 unsigned long address, pte_t *page_table,
2227 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2228 struct page *old_page)
2229 __releases(ptl)
2230{
2231 int page_mkwrite = 0;
2232
2233 page_cache_get(old_page);
2234
2235 /*
2236 * Only catch write-faults on shared writable pages,
2237 * read-only shared pages can get COWed by
2238 * get_user_pages(.write=1, .force=1).
2239 */
2240 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2241 int tmp;
2242
2243 pte_unmap_unlock(page_table, ptl);
2244 tmp = do_page_mkwrite(vma, old_page, address);
2245 if (unlikely(!tmp || (tmp &
2246 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2247 page_cache_release(old_page);
2248 return tmp;
2249 }
2250 /*
2251 * Since we dropped the lock we need to revalidate
2252 * the PTE as someone else may have changed it. If
2253 * they did, we just return, as we can count on the
2254 * MMU to tell us if they didn't also make it writable.
2255 */
2256 page_table = pte_offset_map_lock(mm, pmd, address,
2257 &ptl);
2258 if (!pte_same(*page_table, orig_pte)) {
2259 unlock_page(old_page);
2260 pte_unmap_unlock(page_table, ptl);
2261 page_cache_release(old_page);
2262 return 0;
2263 }
2264 page_mkwrite = 1;
2265 }
2266
2267 return wp_page_reuse(mm, vma, address, page_table, ptl,
2268 orig_pte, old_page, page_mkwrite, 1);
2269}
2270
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002271/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 * This routine handles present pages, when users try to write
2273 * to a shared page. It is done by copying the page to a new address
2274 * and decrementing the shared-page counter for the old page.
2275 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 * Note that this routine assumes that the protection checks have been
2277 * done by the caller (the low-level page fault routine in most cases).
2278 * Thus we can safely just mark it writable once we've done any necessary
2279 * COW.
2280 *
2281 * We also mark the page dirty at this point even though the page will
2282 * change only once the write actually happens. This avoids a few races,
2283 * and potentially makes it more efficient.
2284 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002285 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2286 * but allow concurrent faults), with pte both mapped and locked.
2287 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002289static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2290 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002291 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002292 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293{
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002294 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295
Linus Torvalds6aab3412005-11-28 14:34:23 -08002296 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002297 if (!old_page) {
2298 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002299 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2300 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002301 *
2302 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002303 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002304 */
2305 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2306 (VM_WRITE|VM_SHARED))
Boaz Harroshdd906182015-04-15 16:15:11 -07002307 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2308 orig_pte, pmd);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002309
2310 pte_unmap_unlock(page_table, ptl);
2311 return wp_page_copy(mm, vma, address, page_table, pmd,
2312 orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002313 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002315 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002316 * Take out anonymous pages first, anonymous shared vmas are
2317 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002318 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002319 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002320 if (!trylock_page(old_page)) {
2321 page_cache_get(old_page);
2322 pte_unmap_unlock(page_table, ptl);
2323 lock_page(old_page);
2324 page_table = pte_offset_map_lock(mm, pmd, address,
2325 &ptl);
2326 if (!pte_same(*page_table, orig_pte)) {
2327 unlock_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002328 pte_unmap_unlock(page_table, ptl);
2329 page_cache_release(old_page);
2330 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002331 }
2332 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002333 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002334 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002335 /*
2336 * The page is all ours. Move it to our anon_vma so
2337 * the rmap code will not search our parent or siblings.
2338 * Protected against the rmap code by the page lock.
2339 */
2340 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002341 unlock_page(old_page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002342 return wp_page_reuse(mm, vma, address, page_table, ptl,
2343 orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002344 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002345 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002346 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002347 (VM_WRITE|VM_SHARED))) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002348 return wp_page_shared(mm, vma, address, page_table, pmd,
2349 ptl, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351
2352 /*
2353 * Ok, we need to copy. Oh, well..
2354 */
Nick Pigginb5810032005-10-29 18:16:12 -07002355 page_cache_get(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002356
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002357 pte_unmap_unlock(page_table, ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002358 return wp_page_copy(mm, vma, address, page_table, pmd,
2359 orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360}
2361
Peter Zijlstra97a89412011-05-24 17:12:04 -07002362static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 unsigned long start_addr, unsigned long end_addr,
2364 struct zap_details *details)
2365{
Al Virof5cc4ee2012-03-05 14:14:20 -05002366 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367}
2368
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002369static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 struct zap_details *details)
2371{
2372 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 pgoff_t vba, vea, zba, zea;
2374
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002375 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377
2378 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002379 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2381 zba = details->first_index;
2382 if (zba < vba)
2383 zba = vba;
2384 zea = details->last_index;
2385 if (zea > vea)
2386 zea = vea;
2387
Peter Zijlstra97a89412011-05-24 17:12:04 -07002388 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2390 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002391 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 }
2393}
2394
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002396 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2397 * address_space corresponding to the specified page range in the underlying
2398 * file.
2399 *
Martin Waitz3d410882005-06-23 22:05:21 -07002400 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 * @holebegin: byte in first page to unmap, relative to the start of
2402 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002403 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 * must keep the partial page. In contrast, we must get rid of
2405 * partial pages.
2406 * @holelen: size of prospective hole in bytes. This will be rounded
2407 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2408 * end of the file.
2409 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2410 * but 0 when invalidating pagecache, don't throw away private data.
2411 */
2412void unmap_mapping_range(struct address_space *mapping,
2413 loff_t const holebegin, loff_t const holelen, int even_cows)
2414{
2415 struct zap_details details;
2416 pgoff_t hba = holebegin >> PAGE_SHIFT;
2417 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2418
2419 /* Check for overflow. */
2420 if (sizeof(holelen) > sizeof(hlen)) {
2421 long long holeend =
2422 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2423 if (holeend & ~(long long)ULONG_MAX)
2424 hlen = ULONG_MAX - hba + 1;
2425 }
2426
2427 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 details.first_index = hba;
2429 details.last_index = hba + hlen - 1;
2430 if (details.last_index < details.first_index)
2431 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432
Ross Zwisler0f90cc62015-10-15 15:28:32 -07002433
2434 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002435 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002436 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002438 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439}
2440EXPORT_SYMBOL(unmap_mapping_range);
2441
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002443 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2444 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002445 * We return with pte unmapped and unlocked.
2446 *
2447 * We return with the mmap_sem locked or unlocked in the same cases
2448 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002450static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2451 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002452 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002454 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002455 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002456 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002457 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002459 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002460 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002461 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002463 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002464 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002465
2466 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002467 if (unlikely(non_swap_entry(entry))) {
2468 if (is_migration_entry(entry)) {
2469 migration_entry_wait(mm, pmd, address);
2470 } else if (is_hwpoison_entry(entry)) {
2471 ret = VM_FAULT_HWPOISON;
2472 } else {
2473 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002474 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002475 }
Christoph Lameter06972122006-06-23 02:03:35 -07002476 goto out;
2477 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002478 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 page = lookup_swap_cache(entry);
2480 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002481 page = swapin_readahead(entry,
2482 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 if (!page) {
2484 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002485 * Back out if somebody else faulted in this pte
2486 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002488 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 if (likely(pte_same(*page_table, orig_pte)))
2490 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002491 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002492 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 }
2494
2495 /* Had to read the page from swap area: Major fault */
2496 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002497 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002498 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002499 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002500 /*
2501 * hwpoisoned dirty swapcache pages are kept for killing
2502 * owner processes (which may be unknown at hwpoison time)
2503 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002504 ret = VM_FAULT_HWPOISON;
2505 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002506 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002507 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 }
2509
Hugh Dickins56f31802013-02-22 16:36:10 -08002510 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002511 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002512
Balbir Singh20a10222007-11-14 17:00:33 -08002513 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002514 if (!locked) {
2515 ret |= VM_FAULT_RETRY;
2516 goto out_release;
2517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002519 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002520 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2521 * release the swapcache from under us. The page pin, and pte_same
2522 * test below, are not enough to exclude that. Even if it is still
2523 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002524 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002525 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002526 goto out_page;
2527
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002528 page = ksm_might_need_to_copy(page, vma, address);
2529 if (unlikely(!page)) {
2530 ret = VM_FAULT_OOM;
2531 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002532 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002533 }
2534
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002535 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002536 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002537 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002538 }
2539
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002541 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002543 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002544 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002545 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002546
2547 if (unlikely(!PageUptodate(page))) {
2548 ret = VM_FAULT_SIGBUS;
2549 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 }
2551
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002552 /*
2553 * The page isn't present yet, go ahead with the fault.
2554 *
2555 * Be careful about the sequence of operations here.
2556 * To get its accounting right, reuse_swap_page() must be called
2557 * while the page is counted on swap but not yet in mapcount i.e.
2558 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2559 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002560 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002562 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002563 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002565 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002567 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002568 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002569 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002572 if (pte_swp_soft_dirty(orig_pte))
2573 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002575 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002576 do_page_add_anon_rmap(page, vma, address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002577 mem_cgroup_commit_charge(page, memcg, true, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002578 } else { /* ksm created a completely new copy */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002579 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002580 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002581 lru_cache_add_active_or_unevictable(page, vma);
2582 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002584 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002585 if (mem_cgroup_swap_full(page) ||
2586 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002587 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002588 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002589 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002590 /*
2591 * Hold the lock to avoid the swap entry to be reused
2592 * until we take the PT lock for the pte_same() check
2593 * (to avoid false positives from pte_same). For
2594 * further safety release the lock after the swap_free
2595 * so that the swap count won't change under a
2596 * parallel locked swapcache.
2597 */
2598 unlock_page(swapcache);
2599 page_cache_release(swapcache);
2600 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002601
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002602 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002603 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2604 if (ret & VM_FAULT_ERROR)
2605 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 goto out;
2607 }
2608
2609 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002610 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002611unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002612 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613out:
2614 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002615out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002616 mem_cgroup_cancel_charge(page, memcg, false);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002617 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002618out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002619 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002620out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002621 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002622 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002623 unlock_page(swapcache);
2624 page_cache_release(swapcache);
2625 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002626 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627}
2628
2629/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002630 * This is like a special single-page "expand_{down|up}wards()",
2631 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002632 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002633 */
2634static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2635{
2636 address &= PAGE_MASK;
2637 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002638 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002639
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002640 /*
2641 * Is there a mapping abutting this one below?
2642 *
2643 * That's only ok if it's the same stack mapping
2644 * that has gotten split..
2645 */
2646 if (prev && prev->vm_end == address)
2647 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2648
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002649 return expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002650 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002651 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2652 struct vm_area_struct *next = vma->vm_next;
2653
2654 /* As VM_GROWSDOWN but s/below/above/ */
2655 if (next && next->vm_start == address + PAGE_SIZE)
2656 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2657
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002658 return expand_upwards(vma, address + PAGE_SIZE);
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002659 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002660 return 0;
2661}
2662
2663/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002664 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2665 * but allow concurrent faults), and pte mapped but not yet locked.
2666 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002668static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2669 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002670 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671{
Johannes Weiner00501b52014-08-08 14:19:20 -07002672 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002673 struct page *page;
2674 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676
Linus Torvalds11ac5522010-08-14 11:44:56 -07002677 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002678
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002679 /* File mapping without ->vm_ops ? */
2680 if (vma->vm_flags & VM_SHARED)
2681 return VM_FAULT_SIGBUS;
2682
Linus Torvalds11ac5522010-08-14 11:44:56 -07002683 /* Check if we need to add a guard page to the stack */
2684 if (check_stack_guard_page(vma, address) < 0)
Linus Torvalds9c145c562015-01-29 11:15:17 -08002685 return VM_FAULT_SIGSEGV;
Linus Torvalds11ac5522010-08-14 11:44:56 -07002686
2687 /* Use the zero-page for reads */
Dominik Dingel593befa2014-10-23 12:07:44 +02002688 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
Hugh Dickins62eede62009-09-21 17:03:34 -07002689 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2690 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002691 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002692 if (!pte_none(*page_table))
2693 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002694 /* Deliver the page fault to userland, check inside PT lock */
2695 if (userfaultfd_missing(vma)) {
2696 pte_unmap_unlock(page_table, ptl);
2697 return handle_userfault(vma, address, flags,
2698 VM_UFFD_MISSING);
2699 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002700 goto setpte;
2701 }
2702
Nick Piggin557ed1f2007-10-16 01:24:40 -07002703 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002704 if (unlikely(anon_vma_prepare(vma)))
2705 goto oom;
2706 page = alloc_zeroed_user_highpage_movable(vma, address);
2707 if (!page)
2708 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002709
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002710 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002711 goto oom_free_page;
2712
Minchan Kim52f37622013-04-29 15:08:15 -07002713 /*
2714 * The memory barrier inside __SetPageUptodate makes sure that
2715 * preceeding stores to the page contents become visible before
2716 * the set_pte_at() write.
2717 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002718 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719
Nick Piggin557ed1f2007-10-16 01:24:40 -07002720 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002721 if (vma->vm_flags & VM_WRITE)
2722 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002723
Nick Piggin557ed1f2007-10-16 01:24:40 -07002724 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002725 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002726 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002727
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002728 /* Deliver the page fault to userland, check inside PT lock */
2729 if (userfaultfd_missing(vma)) {
2730 pte_unmap_unlock(page_table, ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002731 mem_cgroup_cancel_charge(page, memcg, false);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002732 page_cache_release(page);
2733 return handle_userfault(vma, address, flags,
2734 VM_UFFD_MISSING);
2735 }
2736
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002737 inc_mm_counter_fast(mm, MM_ANONPAGES);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002738 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002739 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002740 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002741setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002742 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743
2744 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002745 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002746unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002747 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002748 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002749release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002750 mem_cgroup_cancel_charge(page, memcg, false);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002751 page_cache_release(page);
2752 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002753oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002754 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002755oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 return VM_FAULT_OOM;
2757}
2758
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002759/*
2760 * The mmap_sem must have been held on entry, and may have been
2761 * released depending on flags and vma->vm_ops->fault() return value.
2762 * See filemap_fault() and __lock_page_retry().
2763 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002764static int __do_fault(struct vm_area_struct *vma, unsigned long address,
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002765 pgoff_t pgoff, unsigned int flags,
2766 struct page *cow_page, struct page **page)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002767{
2768 struct vm_fault vmf;
2769 int ret;
2770
2771 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2772 vmf.pgoff = pgoff;
2773 vmf.flags = flags;
2774 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002775 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002776 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002777
2778 ret = vma->vm_ops->fault(vma, &vmf);
2779 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2780 return ret;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002781 if (!vmf.page)
2782 goto out;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002783
2784 if (unlikely(PageHWPoison(vmf.page))) {
2785 if (ret & VM_FAULT_LOCKED)
2786 unlock_page(vmf.page);
2787 page_cache_release(vmf.page);
2788 return VM_FAULT_HWPOISON;
2789 }
2790
2791 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2792 lock_page(vmf.page);
2793 else
2794 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2795
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002796 out:
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002797 *page = vmf.page;
2798 return ret;
2799}
2800
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002801/**
2802 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2803 *
2804 * @vma: virtual memory area
2805 * @address: user virtual address
2806 * @page: page to map
2807 * @pte: pointer to target page table entry
2808 * @write: true, if new entry is writable
2809 * @anon: true, if it's anonymous page
2810 *
2811 * Caller must hold page table lock relevant for @pte.
2812 *
2813 * Target users are page handler itself and implementations of
2814 * vm_ops->map_pages.
2815 */
2816void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002817 struct page *page, pte_t *pte, bool write, bool anon)
2818{
2819 pte_t entry;
2820
2821 flush_icache_page(vma, page);
2822 entry = mk_pte(page, vma->vm_page_prot);
2823 if (write)
2824 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002825 if (anon) {
2826 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002827 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002828 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002829 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002830 page_add_file_rmap(page);
2831 }
2832 set_pte_at(vma->vm_mm, address, pte, entry);
2833
2834 /* no need to invalidate: a not-present page won't be cached */
2835 update_mmu_cache(vma, address, pte);
2836}
2837
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002838static unsigned long fault_around_bytes __read_mostly =
2839 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002840
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002841#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002842static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002843{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002844 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002845 return 0;
2846}
2847
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002848/*
2849 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2850 * rounded down to nearest page order. It's what do_fault_around() expects to
2851 * see.
2852 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002853static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002854{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002855 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002856 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002857 if (val > PAGE_SIZE)
2858 fault_around_bytes = rounddown_pow_of_two(val);
2859 else
2860 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002861 return 0;
2862}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002863DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2864 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002865
2866static int __init fault_around_debugfs(void)
2867{
2868 void *ret;
2869
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002870 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2871 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002872 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002873 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002874 return 0;
2875}
2876late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002877#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002878
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002879/*
2880 * do_fault_around() tries to map few pages around the fault address. The hope
2881 * is that the pages will be needed soon and this will lower the number of
2882 * faults to handle.
2883 *
2884 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2885 * not ready to be mapped: not up-to-date, locked, etc.
2886 *
2887 * This function is called with the page table lock taken. In the split ptlock
2888 * case the page table lock only protects only those entries which belong to
2889 * the page table corresponding to the fault address.
2890 *
2891 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2892 * only once.
2893 *
2894 * fault_around_pages() defines how many pages we'll try to map.
2895 * do_fault_around() expects it to return a power of two less than or equal to
2896 * PTRS_PER_PTE.
2897 *
2898 * The virtual address of the area that we map is naturally aligned to the
2899 * fault_around_pages() value (and therefore to page order). This way it's
2900 * easier to guarantee that we don't cross page table boundaries.
2901 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002902static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2903 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2904{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002905 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002906 pgoff_t max_pgoff;
2907 struct vm_fault vmf;
2908 int off;
2909
Jason Low4db0c3c2015-04-15 16:14:08 -07002910 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002911 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2912
2913 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002914 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2915 pte -= off;
2916 pgoff -= off;
2917
2918 /*
2919 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002920 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002921 */
2922 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2923 PTRS_PER_PTE - 1;
2924 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002925 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002926
2927 /* Check if it makes any sense to call ->map_pages */
2928 while (!pte_none(*pte)) {
2929 if (++pgoff > max_pgoff)
2930 return;
2931 start_addr += PAGE_SIZE;
2932 if (start_addr >= vma->vm_end)
2933 return;
2934 pte++;
2935 }
2936
2937 vmf.virtual_address = (void __user *) start_addr;
2938 vmf.pte = pte;
2939 vmf.pgoff = pgoff;
2940 vmf.max_pgoff = max_pgoff;
2941 vmf.flags = flags;
Michal Hockoc20cd452016-01-14 15:20:12 -08002942 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002943 vma->vm_ops->map_pages(vma, &vmf);
2944}
2945
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002946static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2947 unsigned long address, pmd_t *pmd,
2948 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2949{
2950 struct page *fault_page;
2951 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002952 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002953 int ret = 0;
2954
2955 /*
2956 * Let's call ->map_pages() first and use ->fault() as fallback
2957 * if page by the offset is not ready to be mapped (cold cache or
2958 * something).
2959 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08002960 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002961 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2962 do_fault_around(vma, address, pte, pgoff, flags);
2963 if (!pte_same(*pte, orig_pte))
2964 goto unlock_out;
2965 pte_unmap_unlock(pte, ptl);
2966 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002967
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002968 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002969 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2970 return ret;
2971
2972 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2973 if (unlikely(!pte_same(*pte, orig_pte))) {
2974 pte_unmap_unlock(pte, ptl);
2975 unlock_page(fault_page);
2976 page_cache_release(fault_page);
2977 return ret;
2978 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002979 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002980 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002981unlock_out:
2982 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002983 return ret;
2984}
2985
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002986static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2987 unsigned long address, pmd_t *pmd,
2988 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2989{
2990 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07002991 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002992 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002993 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002994 int ret;
2995
2996 if (unlikely(anon_vma_prepare(vma)))
2997 return VM_FAULT_OOM;
2998
2999 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3000 if (!new_page)
3001 return VM_FAULT_OOM;
3002
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003003 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003004 page_cache_release(new_page);
3005 return VM_FAULT_OOM;
3006 }
3007
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003008 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003009 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3010 goto uncharge_out;
3011
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003012 if (fault_page)
3013 copy_user_highpage(new_page, fault_page, address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003014 __SetPageUptodate(new_page);
3015
3016 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3017 if (unlikely(!pte_same(*pte, orig_pte))) {
3018 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003019 if (fault_page) {
3020 unlock_page(fault_page);
3021 page_cache_release(fault_page);
3022 } else {
3023 /*
3024 * The fault handler has no page to lock, so it holds
Yigal Korman0df9d412015-11-16 14:09:15 +02003025 * i_mmap_lock for read to protect against truncate.
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003026 */
Yigal Korman0df9d412015-11-16 14:09:15 +02003027 i_mmap_unlock_read(vma->vm_file->f_mapping);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003028 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003029 goto uncharge_out;
3030 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003031 do_set_pte(vma, address, new_page, pte, true, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003032 mem_cgroup_commit_charge(new_page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07003033 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003034 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003035 if (fault_page) {
3036 unlock_page(fault_page);
3037 page_cache_release(fault_page);
3038 } else {
3039 /*
3040 * The fault handler has no page to lock, so it holds
Yigal Korman0df9d412015-11-16 14:09:15 +02003041 * i_mmap_lock for read to protect against truncate.
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003042 */
Yigal Korman0df9d412015-11-16 14:09:15 +02003043 i_mmap_unlock_read(vma->vm_file->f_mapping);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003044 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003045 return ret;
3046uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003047 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003048 page_cache_release(new_page);
3049 return ret;
3050}
3051
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003052static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003053 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003054 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003056 struct page *fault_page;
3057 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003058 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003059 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003060 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003061 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003062
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003063 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003064 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003065 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066
3067 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003068 * Check if the backing address space wants to know that the page is
3069 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003071 if (vma->vm_ops->page_mkwrite) {
3072 unlock_page(fault_page);
3073 tmp = do_page_mkwrite(vma, fault_page, address);
3074 if (unlikely(!tmp ||
3075 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3076 page_cache_release(fault_page);
3077 return tmp;
3078 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 }
3080
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003081 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3082 if (unlikely(!pte_same(*pte, orig_pte))) {
3083 pte_unmap_unlock(pte, ptl);
3084 unlock_page(fault_page);
3085 page_cache_release(fault_page);
3086 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003087 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003088 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003089 pte_unmap_unlock(pte, ptl);
3090
3091 if (set_page_dirty(fault_page))
3092 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003093 /*
3094 * Take a local copy of the address_space - page.mapping may be zeroed
3095 * by truncate after unlock_page(). The address_space itself remains
3096 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3097 * release semantics to prevent the compiler from undoing this copying.
3098 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003099 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003100 unlock_page(fault_page);
3101 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3102 /*
3103 * Some device drivers do not set page.mapping but still
3104 * dirty their pages
3105 */
3106 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003107 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003108
Johannes Weiner74ec6752015-02-10 14:11:28 -08003109 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003110 file_update_time(vma->vm_file);
3111
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003112 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003113}
Nick Piggind00806b2007-07-19 01:46:57 -07003114
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003115/*
3116 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3117 * but allow concurrent faults).
3118 * The mmap_sem may have been released depending on flags and our
3119 * return value. See filemap_fault() and __lock_page_or_retry().
3120 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003121static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Nick Piggin54cb8822007-07-19 01:46:59 -07003122 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003123 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003124{
3125 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003126 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003127
Hugh Dickins16abfa02007-10-04 16:56:06 +01003128 pte_unmap(page_table);
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003129 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3130 if (!vma->vm_ops->fault)
3131 return VM_FAULT_SIGBUS;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003132 if (!(flags & FAULT_FLAG_WRITE))
3133 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3134 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003135 if (!(vma->vm_flags & VM_SHARED))
3136 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3137 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003138 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003139}
3140
Rashika Kheriab19a9932014-04-03 14:48:02 -07003141static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003142 unsigned long addr, int page_nid,
3143 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003144{
3145 get_page(page);
3146
3147 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003148 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003149 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003150 *flags |= TNF_FAULT_LOCAL;
3151 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003152
3153 return mpol_misplaced(page, vma, addr);
3154}
3155
Rashika Kheriab19a9932014-04-03 14:48:02 -07003156static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gormand10e63f2012-10-25 14:16:31 +02003157 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3158{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003159 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003160 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003161 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003162 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003163 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003164 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003165 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003166 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003167
Mel Gormanc0e7cad2015-02-12 14:58:41 -08003168 /* A PROT_NONE fault should not end up here */
3169 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3170
Mel Gormand10e63f2012-10-25 14:16:31 +02003171 /*
3172 * The "pte" at this point cannot be used safely without
3173 * validation through pte_unmap_same(). It's of NUMA type but
3174 * the pfn may be screwed if the read is non atomic.
3175 *
Mel Gorman4d942462015-02-12 14:58:28 -08003176 * We can safely just do a "set_pte_at()", because the old
3177 * page table entry is not accessible, so there would be no
3178 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003179 */
3180 ptl = pte_lockptr(mm, pmd);
3181 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003182 if (unlikely(!pte_same(*ptep, pte))) {
3183 pte_unmap_unlock(ptep, ptl);
3184 goto out;
3185 }
3186
Mel Gorman4d942462015-02-12 14:58:28 -08003187 /* Make it present again */
3188 pte = pte_modify(pte, vma->vm_page_prot);
3189 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003190 if (was_writable)
3191 pte = pte_mkwrite(pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003192 set_pte_at(mm, addr, ptep, pte);
3193 update_mmu_cache(vma, addr, ptep);
3194
3195 page = vm_normal_page(vma, addr, pte);
3196 if (!page) {
3197 pte_unmap_unlock(ptep, ptl);
3198 return 0;
3199 }
3200
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003201 /* TODO: handle PTE-mapped THP */
3202 if (PageCompound(page)) {
3203 pte_unmap_unlock(ptep, ptl);
3204 return 0;
3205 }
3206
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003207 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003208 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3209 * much anyway since they can be in shared cache state. This misses
3210 * the case where a mapping is writable but the process never writes
3211 * to it but pte_write gets cleared during protection updates and
3212 * pte_dirty has unpredictable behaviour between PTE scan updates,
3213 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003214 */
Mel Gormanbea66fb2015-03-25 15:55:37 -07003215 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003216 flags |= TNF_NO_GROUP;
3217
Rik van Rieldabe1d92013-10-07 11:29:34 +01003218 /*
3219 * Flag if the page is shared between multiple address spaces. This
3220 * is later used when determining whether to group tasks together
3221 */
3222 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3223 flags |= TNF_SHARED;
3224
Peter Zijlstra90572892013-10-07 11:29:20 +01003225 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003226 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003227 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003228 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003229 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003230 put_page(page);
3231 goto out;
3232 }
3233
3234 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003235 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003236 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003237 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003238 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003239 } else
3240 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003241
3242out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003243 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003244 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003245 return 0;
3246}
3247
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003248static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3249 unsigned long address, pmd_t *pmd, unsigned int flags)
3250{
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003251 if (vma_is_anonymous(vma))
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003252 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3253 if (vma->vm_ops->pmd_fault)
3254 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3255 return VM_FAULT_FALLBACK;
3256}
3257
3258static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3259 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3260 unsigned int flags)
3261{
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003262 if (vma_is_anonymous(vma))
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003263 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3264 if (vma->vm_ops->pmd_fault)
3265 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3266 return VM_FAULT_FALLBACK;
3267}
3268
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269/*
3270 * These routines also need to handle stuff like marking pages dirty
3271 * and/or accessed for architectures that don't do it in hardware (most
3272 * RISC architectures). The early dirtying is also good on the i386.
3273 *
3274 * There is also a hook called "update_mmu_cache()" that architectures
3275 * with external mmu caches can use to update those (ie the Sparc or
3276 * PowerPC hashed page tables that act as extended TLBs).
3277 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003278 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3279 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003280 * We return with pte unmapped and unlocked.
3281 *
3282 * The mmap_sem may have been released depending on flags and our
3283 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003285static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003286 struct vm_area_struct *vma, unsigned long address,
3287 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288{
3289 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003290 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291
Christian Borntraegere37c6982014-12-07 21:41:33 +01003292 /*
3293 * some architectures can have larger ptes than wordsize,
3294 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3295 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3296 * The code below just needs a consistent view for the ifs and
3297 * we later double check anyway with the ptl lock held. So here
3298 * a barrier will do.
3299 */
3300 entry = *pte;
3301 barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003303 if (pte_none(entry)) {
Oleg Nesterovb5330622015-09-08 14:58:28 -07003304 if (vma_is_anonymous(vma))
3305 return do_anonymous_page(mm, vma, address,
3306 pte, pmd, flags);
3307 else
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003308 return do_fault(mm, vma, address, pte, pmd,
3309 flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003310 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003311 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003312 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 }
3314
Mel Gorman8a0516e2015-02-12 14:58:22 -08003315 if (pte_protnone(entry))
Mel Gormand10e63f2012-10-25 14:16:31 +02003316 return do_numa_page(mm, vma, address, entry, pte, pmd);
3317
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003318 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003319 spin_lock(ptl);
3320 if (unlikely(!pte_same(*pte, entry)))
3321 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003322 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003324 return do_wp_page(mm, vma, address,
3325 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 entry = pte_mkdirty(entry);
3327 }
3328 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003329 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003330 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003331 } else {
3332 /*
3333 * This is needed only for protection faults but the arch code
3334 * is not yet telling us if this is a protection fault or not.
3335 * This still avoids useless tlb flushes for .text page faults
3336 * with threads.
3337 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003338 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003339 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003340 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003341unlock:
3342 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003343 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344}
3345
3346/*
3347 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003348 *
3349 * The mmap_sem may have been released depending on flags and our
3350 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003352static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3353 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354{
3355 pgd_t *pgd;
3356 pud_t *pud;
3357 pmd_t *pmd;
3358 pte_t *pte;
3359
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003360 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003361 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 pud = pud_alloc(mm, pgd, address);
3365 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003366 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 pmd = pmd_alloc(mm, pud, address);
3368 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003369 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003370 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003371 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003372 if (!(ret & VM_FAULT_FALLBACK))
3373 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003374 } else {
3375 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003376 int ret;
3377
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003378 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003379 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003380 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3381
Mel Gorman8a0516e2015-02-12 14:58:22 -08003382 if (pmd_protnone(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003383 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003384 orig_pmd, pmd);
3385
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003386 if (dirty && !pmd_write(orig_pmd)) {
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003387 ret = wp_huge_pmd(mm, vma, address, pmd,
3388 orig_pmd, flags);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003389 if (!(ret & VM_FAULT_FALLBACK))
3390 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003391 } else {
3392 huge_pmd_set_accessed(mm, vma, address, pmd,
3393 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003394 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003395 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003396 }
3397 }
3398
3399 /*
3400 * Use __pte_alloc instead of pte_alloc_map, because we can't
3401 * run pte_offset_map on the pmd, if an huge pmd could
3402 * materialize from under us from a different thread.
3403 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003404 if (unlikely(pmd_none(*pmd)) &&
3405 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003406 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003407 /* if an huge pmd materialized from under us just retry later */
Dan Williams5c7fb562016-01-15 16:56:52 -08003408 if (unlikely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003409 return 0;
3410 /*
3411 * A regular pmd is established and it can't morph into a huge pmd
3412 * from under us anymore at this point because we hold the mmap_sem
3413 * read mode and khugepaged takes it in write mode. So now it's
3414 * safe to run pte_offset_map().
3415 */
3416 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003418 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419}
3420
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003421/*
3422 * By the time we get here, we already hold the mm semaphore
3423 *
3424 * The mmap_sem may have been released depending on flags and our
3425 * return value. See filemap_fault() and __lock_page_or_retry().
3426 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003427int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3428 unsigned long address, unsigned int flags)
3429{
3430 int ret;
3431
3432 __set_current_state(TASK_RUNNING);
3433
3434 count_vm_event(PGFAULT);
3435 mem_cgroup_count_vm_event(mm, PGFAULT);
3436
3437 /* do counter updates before entering really critical section. */
3438 check_sync_rss_stat(current);
3439
3440 /*
3441 * Enable the memcg OOM handling for faults triggered in user
3442 * space. Kernel faults are handled more gracefully.
3443 */
3444 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003445 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003446
3447 ret = __handle_mm_fault(mm, vma, address, flags);
3448
Johannes Weiner49426422013-10-16 13:46:59 -07003449 if (flags & FAULT_FLAG_USER) {
3450 mem_cgroup_oom_disable();
3451 /*
3452 * The task may have entered a memcg OOM situation but
3453 * if the allocation error was handled gracefully (no
3454 * VM_FAULT_OOM), there is no need to kill anything.
3455 * Just clean up the OOM state peacefully.
3456 */
3457 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3458 mem_cgroup_oom_synchronize(false);
3459 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003460
Johannes Weiner519e5242013-09-12 15:13:42 -07003461 return ret;
3462}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003463EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003464
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465#ifndef __PAGETABLE_PUD_FOLDED
3466/*
3467 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003468 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003470int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003472 pud_t *new = pud_alloc_one(mm, address);
3473 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003474 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475
Nick Piggin362a61a2008-05-14 06:37:36 +02003476 smp_wmb(); /* See comment in __pte_alloc */
3477
Hugh Dickins872fec12005-10-29 18:16:21 -07003478 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003479 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003480 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003481 else
3482 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003483 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003484 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485}
3486#endif /* __PAGETABLE_PUD_FOLDED */
3487
3488#ifndef __PAGETABLE_PMD_FOLDED
3489/*
3490 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003491 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003493int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003495 pmd_t *new = pmd_alloc_one(mm, address);
3496 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003497 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498
Nick Piggin362a61a2008-05-14 06:37:36 +02003499 smp_wmb(); /* See comment in __pte_alloc */
3500
Hugh Dickins872fec12005-10-29 18:16:21 -07003501 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003503 if (!pud_present(*pud)) {
3504 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003505 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003506 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003507 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003508#else
3509 if (!pgd_present(*pud)) {
3510 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003511 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003512 } else /* Another has populated it */
3513 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003515 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003516 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517}
3518#endif /* __PAGETABLE_PMD_FOLDED */
3519
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003520static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003521 pte_t **ptepp, spinlock_t **ptlp)
3522{
3523 pgd_t *pgd;
3524 pud_t *pud;
3525 pmd_t *pmd;
3526 pte_t *ptep;
3527
3528 pgd = pgd_offset(mm, address);
3529 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3530 goto out;
3531
3532 pud = pud_offset(pgd, address);
3533 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3534 goto out;
3535
3536 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003537 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003538 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3539 goto out;
3540
3541 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3542 if (pmd_huge(*pmd))
3543 goto out;
3544
3545 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3546 if (!ptep)
3547 goto out;
3548 if (!pte_present(*ptep))
3549 goto unlock;
3550 *ptepp = ptep;
3551 return 0;
3552unlock:
3553 pte_unmap_unlock(ptep, *ptlp);
3554out:
3555 return -EINVAL;
3556}
3557
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003558static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3559 pte_t **ptepp, spinlock_t **ptlp)
3560{
3561 int res;
3562
3563 /* (void) is needed to make gcc happy */
3564 (void) __cond_lock(*ptlp,
3565 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3566 return res;
3567}
3568
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003569/**
3570 * follow_pfn - look up PFN at a user virtual address
3571 * @vma: memory mapping
3572 * @address: user virtual address
3573 * @pfn: location to store found PFN
3574 *
3575 * Only IO mappings and raw PFN mappings are allowed.
3576 *
3577 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3578 */
3579int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3580 unsigned long *pfn)
3581{
3582 int ret = -EINVAL;
3583 spinlock_t *ptl;
3584 pte_t *ptep;
3585
3586 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3587 return ret;
3588
3589 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3590 if (ret)
3591 return ret;
3592 *pfn = pte_pfn(*ptep);
3593 pte_unmap_unlock(ptep, ptl);
3594 return 0;
3595}
3596EXPORT_SYMBOL(follow_pfn);
3597
Rik van Riel28b2ee22008-07-23 21:27:05 -07003598#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003599int follow_phys(struct vm_area_struct *vma,
3600 unsigned long address, unsigned int flags,
3601 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003602{
Johannes Weiner03668a42009-06-16 15:32:34 -07003603 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003604 pte_t *ptep, pte;
3605 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003606
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003607 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3608 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003609
Johannes Weiner03668a42009-06-16 15:32:34 -07003610 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003611 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003612 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003613
Rik van Riel28b2ee22008-07-23 21:27:05 -07003614 if ((flags & FOLL_WRITE) && !pte_write(pte))
3615 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003616
3617 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003618 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003619
Johannes Weiner03668a42009-06-16 15:32:34 -07003620 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003621unlock:
3622 pte_unmap_unlock(ptep, ptl);
3623out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003624 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003625}
3626
3627int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3628 void *buf, int len, int write)
3629{
3630 resource_size_t phys_addr;
3631 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003632 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003633 int offset = addr & (PAGE_SIZE-1);
3634
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003635 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003636 return -EINVAL;
3637
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003638 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003639 if (write)
3640 memcpy_toio(maddr + offset, buf, len);
3641 else
3642 memcpy_fromio(buf, maddr + offset, len);
3643 iounmap(maddr);
3644
3645 return len;
3646}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003647EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003648#endif
3649
David Howells0ec76a12006-09-27 01:50:15 -07003650/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003651 * Access another process' address space as given in mm. If non-NULL, use the
3652 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003653 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003654static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3655 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003656{
David Howells0ec76a12006-09-27 01:50:15 -07003657 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003658 void *old_buf = buf;
3659
David Howells0ec76a12006-09-27 01:50:15 -07003660 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003661 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003662 while (len) {
3663 int bytes, ret, offset;
3664 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003665 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003666
3667 ret = get_user_pages(tsk, mm, addr, 1,
3668 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003669 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003670#ifndef CONFIG_HAVE_IOREMAP_PROT
3671 break;
3672#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003673 /*
3674 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3675 * we can access using slightly different code.
3676 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003677 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003678 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003679 break;
3680 if (vma->vm_ops && vma->vm_ops->access)
3681 ret = vma->vm_ops->access(vma, addr, buf,
3682 len, write);
3683 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003684 break;
3685 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003686#endif
David Howells0ec76a12006-09-27 01:50:15 -07003687 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003688 bytes = len;
3689 offset = addr & (PAGE_SIZE-1);
3690 if (bytes > PAGE_SIZE-offset)
3691 bytes = PAGE_SIZE-offset;
3692
3693 maddr = kmap(page);
3694 if (write) {
3695 copy_to_user_page(vma, page, addr,
3696 maddr + offset, buf, bytes);
3697 set_page_dirty_lock(page);
3698 } else {
3699 copy_from_user_page(vma, page, addr,
3700 buf, maddr + offset, bytes);
3701 }
3702 kunmap(page);
3703 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003704 }
David Howells0ec76a12006-09-27 01:50:15 -07003705 len -= bytes;
3706 buf += bytes;
3707 addr += bytes;
3708 }
3709 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003710
3711 return buf - old_buf;
3712}
Andi Kleen03252912008-01-30 13:33:18 +01003713
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003714/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003715 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003716 * @mm: the mm_struct of the target address space
3717 * @addr: start address to access
3718 * @buf: source or destination buffer
3719 * @len: number of bytes to transfer
3720 * @write: whether the access is a write
3721 *
3722 * The caller must hold a reference on @mm.
3723 */
3724int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3725 void *buf, int len, int write)
3726{
3727 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3728}
3729
Andi Kleen03252912008-01-30 13:33:18 +01003730/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003731 * Access another process' address space.
3732 * Source/target buffer must be kernel space,
3733 * Do not walk the page table directly, use get_user_pages
3734 */
3735int access_process_vm(struct task_struct *tsk, unsigned long addr,
3736 void *buf, int len, int write)
3737{
3738 struct mm_struct *mm;
3739 int ret;
3740
3741 mm = get_task_mm(tsk);
3742 if (!mm)
3743 return 0;
3744
3745 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3746 mmput(mm);
3747
3748 return ret;
3749}
3750
Andi Kleen03252912008-01-30 13:33:18 +01003751/*
3752 * Print the name of a VMA.
3753 */
3754void print_vma_addr(char *prefix, unsigned long ip)
3755{
3756 struct mm_struct *mm = current->mm;
3757 struct vm_area_struct *vma;
3758
Ingo Molnare8bff742008-02-13 20:21:06 +01003759 /*
3760 * Do not print if we are in atomic
3761 * contexts (in exception stacks, etc.):
3762 */
3763 if (preempt_count())
3764 return;
3765
Andi Kleen03252912008-01-30 13:33:18 +01003766 down_read(&mm->mmap_sem);
3767 vma = find_vma(mm, ip);
3768 if (vma && vma->vm_file) {
3769 struct file *f = vma->vm_file;
3770 char *buf = (char *)__get_free_page(GFP_KERNEL);
3771 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003772 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003773
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02003774 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003775 if (IS_ERR(p))
3776 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003777 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003778 vma->vm_start,
3779 vma->vm_end - vma->vm_start);
3780 free_page((unsigned long)buf);
3781 }
3782 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003783 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003784}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003785
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003786#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02003787void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003788{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003789 /*
3790 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3791 * holding the mmap_sem, this is safe because kernel memory doesn't
3792 * get paged out, therefore we'll never actually fault, and the
3793 * below annotations will generate false positives.
3794 */
3795 if (segment_eq(get_fs(), KERNEL_DS))
3796 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003797 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003798 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003799 __might_sleep(file, line, 0);
3800#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003801 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003802 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02003803#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003804}
David Hildenbrand9ec23532015-05-11 17:52:07 +02003805EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003806#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003807
3808#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3809static void clear_gigantic_page(struct page *page,
3810 unsigned long addr,
3811 unsigned int pages_per_huge_page)
3812{
3813 int i;
3814 struct page *p = page;
3815
3816 might_sleep();
3817 for (i = 0; i < pages_per_huge_page;
3818 i++, p = mem_map_next(p, page, i)) {
3819 cond_resched();
3820 clear_user_highpage(p, addr + i * PAGE_SIZE);
3821 }
3822}
3823void clear_huge_page(struct page *page,
3824 unsigned long addr, unsigned int pages_per_huge_page)
3825{
3826 int i;
3827
3828 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3829 clear_gigantic_page(page, addr, pages_per_huge_page);
3830 return;
3831 }
3832
3833 might_sleep();
3834 for (i = 0; i < pages_per_huge_page; i++) {
3835 cond_resched();
3836 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3837 }
3838}
3839
3840static void copy_user_gigantic_page(struct page *dst, struct page *src,
3841 unsigned long addr,
3842 struct vm_area_struct *vma,
3843 unsigned int pages_per_huge_page)
3844{
3845 int i;
3846 struct page *dst_base = dst;
3847 struct page *src_base = src;
3848
3849 for (i = 0; i < pages_per_huge_page; ) {
3850 cond_resched();
3851 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3852
3853 i++;
3854 dst = mem_map_next(dst, dst_base, i);
3855 src = mem_map_next(src, src_base, i);
3856 }
3857}
3858
3859void copy_user_huge_page(struct page *dst, struct page *src,
3860 unsigned long addr, struct vm_area_struct *vma,
3861 unsigned int pages_per_huge_page)
3862{
3863 int i;
3864
3865 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3866 copy_user_gigantic_page(dst, src, addr, vma,
3867 pages_per_huge_page);
3868 return;
3869 }
3870
3871 might_sleep();
3872 for (i = 0; i < pages_per_huge_page; i++) {
3873 cond_resched();
3874 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3875 }
3876}
3877#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003878
Olof Johansson40b64ac2013-12-20 14:28:05 -08003879#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003880
3881static struct kmem_cache *page_ptl_cachep;
3882
3883void __init ptlock_cache_init(void)
3884{
3885 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3886 SLAB_PANIC, NULL);
3887}
3888
Peter Zijlstra539edb52013-11-14 14:31:52 -08003889bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003890{
3891 spinlock_t *ptl;
3892
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003893 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003894 if (!ptl)
3895 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003896 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003897 return true;
3898}
3899
Peter Zijlstra539edb52013-11-14 14:31:52 -08003900void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003901{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003902 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003903}
3904#endif